Datasets:

---

spiralworks

/

openreview-structured-forums

like 0

Follow

Spiral Works 3

[ "Won-Seok Choi", "Dong-Sig Han", "Suhyung Choi", "Hyeonseo Yang", "Byoung-Tak Zhang" ]

[ "metric learning", "environment recognition", "bayesian inference", "self-supervised learning" ]

{ "note_id": [ "xT44vNAeN9", "lh3BM0fgqi", "cjNZWZeXpS", "bHfrq1h593", "ZiQVHiMfaK", "X6546HUDzV", "X4RYM4Rne9", "VVxHNxYBMQ", "QkSjzNg06j", "QLK4n40JY5", "Pp2ILdcmQX", "Ozad3T9Zea", "Gy0ViEK6YO", "C1u2agBOuI", "46suI9rrfr" ], "note_type": [ "official_review", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "decision", "official_comment", "meta_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment" ], "note_created": [ 1730693493566, 1733206606194, 1732166261889, 1732166606254, 1730987743800, 1730868873308, 1737523791745, 1732617255443, 1734479944642, 1732166776944, 1733219366758, 1732547778845, 1732165969045, 1730690488960, 1732166166908 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission6783/Reviewer_WjYy" ], [ "ICLR.cc/2025/Conference/Submission6783/Reviewer_WjYy" ], [ "ICLR.cc/2025/Conference/Submission6783/Authors" ], [ "ICLR.cc/2025/Conference/Submission6783/Authors" ], [ "ICLR.cc/2025/Conference/Submission6783/Reviewer_d5mr" ], [ "ICLR.cc/2025/Conference/Submission6783/Reviewer_9vYe" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission6783/Authors" ], [ "ICLR.cc/2025/Conference/Submission6783/Area_Chair_GJtW" ], [ "ICLR.cc/2025/Conference/Submission6783/Authors" ], [ "ICLR.cc/2025/Conference/Submission6783/Authors" ], [ "ICLR.cc/2025/Conference/Submission6783/Reviewer_d5mr" ], [ "ICLR.cc/2025/Conference/Submission6783/Authors" ], [ "ICLR.cc/2025/Conference/Submission6783/Reviewer_8Td2" ], [ "ICLR.cc/2025/Conference/Submission6783/Authors" ] ], "structured_content_str": [ "{\"summary\": \"The paper proposes the Object-Based Sub-Environment Recognition (OBSER) framework, a novel Bayesian framework for measuring object-environment and environment-environment relationships using a feature extractor trained with metric learning. The key idea is the introduction of a statistically separable (EDS) function and using it to perform (i) object-object similarity, which involves obtaining the closest class of objects from a list given a query object, (ii) object-environment recognition, which involves retrieving the closest environment to a given object and (iii) environment-environment recognition, which defines the difference between two sub-environments. Experiments to recognize environments are done on two datasets, the ImageNet based dataset, and a dataset of curated environments from Minecraft.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": [\"The primary motivation behind the paper is sound. Indeed, environment recognition using relationships between objects and environments is an interesting problem in embodied agents.\", \"The results do illustrate the claim that higher difference between epsilon and delta values lead to a better accuracy score. This is reflected in models for both Tables 1 and 2.\"], \"weaknesses\": [\"Not clear what objects and environment are: By reading the paper starting from the introduction, it is not clear what objects and environments mean. The authors show some examples of objects and biomes as environment in the Minecraft example, but none for the ImageNet dataset. This makes it difficult to understand the contribution of the work.\", \"Results are difficult to interpret: There are mainly just two results described in the paper in Tables 1 and 2, which are the respective classification accuracies for ImageNet and Minecraft datasets. It is difficult to interpret these tables. For instance, are the differences between metric and self-supervised learning methods the main observation, or the relationship between EDS values and different models?\", \"Real-world applications are unclear: While the motivation of the work is sound, it is not straightforward to interpret how this paper contributes to real-world object and scene recognition. The paper does not contain any examples of real-world object scenes or object recognition. The only examples provided are for Minecraft, which is still a simulation environment, and not reflective of human-centric objects and environments.\", \"[Minor] Unnecessary math: There is a lot of mathematical terminology introduced in Sections 3-5 (besides the main contribution in Section 4.1), which are not necessary to be present in the main paper.\", \"[Minor] Many details are missing from the main paper and are present in the supplementary. The authors should consider transferring some details about the implementation from supplementary to the main paper. For any details, the reader has to constantly switch between paper and supplementary, which is not a good user-experience.\"], \"questions\": [\"Is there any justification for the type of classifiers provided in Tables 1 and 2? In Table 1, linear, mean and KNN classifiers are used, while only mean and KNN classifiers are used in Table 2. Also, why the difference between the shots of mean vs KNN (1,3,5 vs 3,5,7).\", \"For the ImageNet dataset, what are some examples of objects and environments? It is unclear from reading the paper.\", \"In conclusion, the authors mention that by integrating the proposed method with embodied object recognition or navigation modules, inference accuracy can be improved. Can the authors provide some justification with a real-world use-case about what is the intuition behind this?\", \"What classification accuracies are mentioned in Tables 1 and 2? Is it the object-object similar classes?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"details_of_ethics_concerns\": \"No ethical concerns observed.\", \"rating\": \"5\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"The authors have partially addressed my concerns by providing some examples from the Replica dataset. While I do appreciate the effort put forth by the authors, I still feel that the real-world applications of this work is limited, as noted by other reviewers also. For instance the results of metric learning model SupCon appear to be one of the best for ImageNet (another realistic dataset) and Minecraft, but amongst the worst for Replica (at least on the unseen case). This makes it difficult to root for which backbone best benefits from the OBSER framework.\\n\\nBased on the author responses, and looking at the other reviews, I will raise my rating to 5.\"}", "{\"title\": \"Response to Reviewer WjYy (1/1)\", \"comment\": \"Dear Reviewer WjYy,\\n\\nWe appreciate your recognition of the motivation behind our work and the validity of the results supporting our claims. Here are our responses to your thoughtful comments.\\n\\n\\u201csome examples of objects and environments in ImageNet\\u201d : Conceptually, a sub-environment with ImageNet can be interpreted as separated \\u201croom\\u201d or \\u201cplace\\u201d, and an object as \\u201cobserved target\\u201d. We validated the proposed framework with a subset of ImageNet with different data distributions. With the following steps, we generated these artificial environments.\\n\\n- First, we randomly selected 10 (or 40) classes in ImageNet for each seed.\\n- With a given class distribution (rho), We sampled data of each class from the dataset.\\n- We analyzed the framework with gathered data as empirical distribution.\\n\\nAlso, to validate our framework in a more realistic environment, we additionally conducted experiments with the Replica environment, which is a 3D indoor environment. In this case, a sub-environment becomes a room, and an agent can observe objects in each room.\\n\\n\\u201cmainly just two results\\u201d : Since we visualized that all three relationships (object-object, object-environment, and environment-environment) can be estimated with the OBSER framework in the forms of tables (1,2), graphs (5,6), and qualitative results with demos (7-9), we cannot agree that the main results are just two.\\n\\n\\u201cFor instance,\\u2026 different models?\\u201d, \\u201cany justification for the type of classifiers provided in Tables 1 and 2\\u201d : Tables 1 and 2 show the relationship between EDS values and classification accuracy because the classification task highly depends on the object-object relationship. For classification tasks, we chose the KNN classifier and mean classifier to show the influence of concentration and separability. We chose mean classifier rather than linear probing because mean classifier can be used without explicit label information (this means we can compute mean classifier accuracy directly with any set of labels). If the representations are not separated, both KNN and mean classification accuracy are low. If the representations are separated but not concentrated, then KNN accuracy is low, while mean classification accuracy is high. Also, using top-1 accuracy with the KNN classifier is often noisy (because it only considers the nearest neighbor), so we used (3, 5, 7) instead of (1, 3, 5).\\n\\n\\u201cReal-world applications are unclear\\u201d, \\u201csome justification with a real-world use-case about what is the intuition behind this\\u201d : As mentioned above, We conducted additional experiments with photorealistic environment. We chose the object retrieval task in multiple sub-environments (in this case, rooms), which can be interpreted as visual navigation. We design chained inference with all three relationships. With a given query, the framework first retrieves the most relevant room in its memory. With the memory, the framework finds the most similar room and finds the queried object. As a result, we found that our framework can find the queried object by searching with a limited number of relevant rooms (1 or 3) rather than iterating all rooms (48 or 35) for exploration.\\n\\n\\u201cUnnecessary math\\u201d : To make sure that a proposed method is correct, we think that theoretical support to guarantee its correctness is essential. For this reason, we introduced the EDS function to show that the precision of inference with the framework is guaranteed with an optimized feature extractor with low epsilon and high delta. So, we argue that our theoretical attempt is necessary and justified.\\n\\nFinally, we have moved several contents in the main paper and the appendix for more readability.\"}", "{\"title\": \"Response to Reviewer 8Td2 (1/1)\", \"comment\": \"Dear Reviewer 8Td2,\\n\\nWe are thoroughly thankful for your positive and insightful feedback. Here is our response to your comments.\\n\\n\\u201cThe application of OBSER is not clear.\\u201d : To show the applicability of our framework in a more realistic environment, we additionally conducted experiments with a Replica environment, which is a 3D indoor environment. In this environment, we build chained inference, which contains all three relationships with the OBSER framework. The object retrieval task, which can be interpreted as visual navigation, is used for evaluation. When the object query is given, the framework first finds the most relevant room in its memory and finds the most similar room in an environment with the memory. The framework retrieves the object most similar to the given query. As a result, we found that the proposed framework can successfully retrieve the result by exploring a limited number of relevant rooms (1 or 3), which is much smaller than the total number of rooms (48 or 35). This implies that sub-environment recognition can be effectively applied to agents for efficient inferences.\\n\\n\\u201cWhat if the category of possible objects is unknown\\u201d : In our experiments, the framework didn\\u2019t have access to any category information from the environment. Our method is based on metric learning (and SSL), so the framework doesn\\u2019t require explicit class information.\\n\\nHowever, when the problem is open-set, which contains unseen objects that are not used for training, metric learning-based models, such as SupCon, may fail in some situations. In metric learning, the model is trained with positive pairs (and negatives) of the data which is selected with explicit class. However, SSL models can be robust in such environments because they are trained only with inductive biases in the data domain. In the Replica environment, we conducted with an \\u201cunseen\\u201d setting, and in this situation, the framework can only be accessible to rooms in \\u201capartment_0\\u201d for episodic memory. In this setting, SupCon underperforms MoCov3 and DINOs, and we think this phenomenon occurs for this reason.\"}", "{\"summary\": \"The authors propose a framework for an agent to infer its sub-environment through the measurements of object-object, object-environment, and environment-environment relationships. The authors validate their framework in Minecraft, while also providing some preliminary results on the ImageNet dataset. The paper is well structured, and the authors show several relevant results, including the relevance of statistically separable EDS functions to achieve accurate measures for their downstream environment inference.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"The paper is generally well structured. The authors explain each part of their proposed method in detail, including for example the relevance of statistically separable EDS functions to achieve accurate measures for their downstream environment inference, and the empirical implication of hyperparameter choice for downstream inference (e.g. the choice of Tau for KL divergence in Figure 6.).\", \"weaknesses\": \"It is hard to quickly have a notion of which parts of the proposed methods, exactly, are novel. The authors use several existing methodologies in their proposed framework, but fail to appropriately specify which of these, in particular, are novel propositions or implications. It is also hard to connect the motivation of this work to the tasks and results shown. In particular, the emphasis on the motivation for \\\"real-world\\\" applications, and complex natural environments is lost by the simplicity of the test settings (e.g. virtual world or fixed datasets).\", \"questions\": [\"It would be worthwhile to adjust the tone of the claims in the paper to better align with the results shown. The results may show interesting results in a \\\"simulated environment, towards more complex environmental settings\\\" perhaps even eventually leading to real-world, but as far as this work goes there is a wide gap between simulated and real-world settings, since no robotic experiments were provided. Below are some of the most relevant parts, strongly suggesting (non-existing) results in real-world settings\", \"The abstract\", \"The introduction should reflect this (3rd claim)\", \"Figure 1: The caption should be updated (it is not, in fact, a real-world agent)\", \"Title in 6.2 should change\", \"Section 4: Which of these are new propositions and which of these are derived from existing work? This should be made very explicit.\", \"Missing y-label in Fig. 5 and Fig. 6\", \"English should be improved throughout:\", \"e.g. \\\"which computes the kernel density accumulated with class-wise distribution.\\\", or \\\"We utilized pretrained weights for every models.\\\" etc.\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The paper presents a Bayesian framework to recognize sub-environments within complex, dynamic environments. OBSER enables agents, like robots, to identify sub-environments based on objects present, facilitating task-driven navigation and inference in open-world scenarios. The framework introduces EDS function to improve the robustness of feature representations and utilizes metric learning for object-object, object-environment, and environment-environment relationships.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"OBSER provides a holistic approach to sub-environment recognition, measuring three relationships\\u2014object-object, object-environment, and environment-environment\\u2014which enables better contextual awareness.\\n\\nThe introduction of the EDS function to assess separability and concentration offers a robust way to manage feature representations, addressing a gap in object-based environmental recognition.\", \"weaknesses\": \"I find that this method may be challenging to implement for embodied robots. First, constructing episodic memory seems crucial for task completion success, yet several questions arise: (1) How was this memory constructed? (2) How could it be constructed effectively with limited experience? (3) How can retrieval be managed efficiently as memory size increases?\\n\\nMost importantly, I am uncertain about how the object-object, object-environment, and environment-environment relationships contribute to embodied tasks. Without ablation studies or proof, it\\u2019s hard to determine the critical importance of these relationships.\\n\\nAre there other baselines with which EDS could be compared? The paper would benefit from broader comparisons with other state-of-the-art environment recognition frameworks to better highlight OBSER's distinct advantages and limitations in context.\\n\\nThe diversities of Minecraft environment and objects seems limited.\\n\\nOBSER's reliance on object distribution might limit its effectiveness in sub-environments where objects are scarce or ambiguous, which could impact performance in less structured real-world spaces.\", \"questions\": \"Please refer to the weakness.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"title\": \"Response to Reviewer d5mr (Modification of Figure 1)\", \"comment\": \"Dear Reviewer d5mr,\\n\\nWe are very pleased that our explanations and revisions have addressed your concerns. Your comment has been immensely helpful in improving the clarity and novelty of our work.\\n\\nAs you mentioned, we have modified Figure 1 and its caption to include more information about the task and inference with sub-environment recognition. Please refer to Figure 1 in the latest version of the paper to ensure that your concern is fully addressed.\\n\\nIf you have further suggestions, please let us know during the discussion phase.\\n\\nBest regards,\\n\\nAuthors of submission 6783\"}", "{\"metareview\": \"The paper received borderline ratings (6,6,5,5). The reviewers identified several weaknesses, such as challenges in pinpointing novel aspects, limited diversity in Minecraft environments, unclear real-world applications, and a lack of clarity in the definitions of objects and environments. The author provided responses to the reviewers. The AC checked the paper, the reviewers and the responses. The AC did not find the responses convincing enough. For example, the AC is in agreement with reviewer WjYy that the real world applications are limited. Also, the authors did not adequately address the questions posed by Reviewer 9vYe \\u201c(1) How was this memory constructed? (2) How could it be constructed effectively with limited experience? (3) How can retrieval be managed efficiently as memory size increases?\\u201d. Due to these issues, the AC recommends rejection.\", \"additional_comments_on_reviewer_discussion\": \"Reviewer d5mr mentioned that their concerns were addressed by the rebuttal. Reviewer WjYy appreciated the new experiments but they still had concerns regarding the real world applicability. The AC checked the follow up response by the authors but did not find it convincing enough. The AC checked the responses to reviewer 9vYe and 8Td2, but could not find good answers to some of the questions (some examples mentioned above). Therefore, the AC decided that the paper needs major revision and recommended rejection.\"}", "{\"title\": \"Response to all reviewers\", \"comment\": \"Dear Reviewers,\\n\\nFirst, we would like to express our gratitude for the positive consideration of our work. The reviewers expressed that the motivation of our work is sound (9vYe, WjYy), the paper is well-structured (d5mr), and the experiments support our claims well (d5mr, 9vYe, WjYy). We also appreciate the insightful comments to improve our paper. To announce important changes in our paper, we share this information in a separate comment instead of replying to each comment.\\n\\nOur research topic is to propose a new sub-environment recognition framework with the metric learning-based method. We designed the sub-environment recognition with three fundamental relationships (object-object, object-environment, and environment-environment). In this work, we focus on expressing and validating such relationships via kernel density estimation with a feature extractor trained with metric learning and SSL, rather than solving embodied tasks with these relationships. We believe that combining our framework with embodied agents is a promising research topic. For this reason, we have found that several expressions in the previous revision can be read unintendedly, which may lead to a misunderstanding of the focus of our paper. To clarify our claims, we adjust expressions such as \\\"embodied agent\\\" throughout the paper (especially in Introduction and Conclusion).\\n\\nNevertheless, we claim that our method can be applied to real-world situations, so we perform additional experiments with Replica, which is a 3D photorealistic indoor environment introduced in [1]. In these experiments, we designed chained inference with all three relationships. We set each room of each environment as the sub-environments and apply the OBSER framework to solve the object retrieval tasks in multiple sub-environments. As a result, we found that our method can reliably retrieve the queried objects with only a limited number of relevant rooms (1 and 3) for search, which is much smaller than the total number of rooms (48 in the seen setting and 35 in the unseen setting). This suggests that with sub-environment recognition, the agent can efficiently retrieve the goal by reducing unnecessary exploration of irrelevant sub-environments. Also, we found that the metric learning model (SupCon) performs worse than SSL models in unseen settings, which implies that SSL models are robust in less-structed (curated) but more general environments. Please refer to Section 6.2 and Appendix D in the revised version for more details.\\n\\n[1] Straub, Julian, et al. \\\"The Replica dataset: A digital replica of indoor spaces.\\\"\\u00a0*arXiv preprint arXiv:1906.05797* (2019).\"}", "{\"title\": \"Response to Reviewer WjYy (Metric learning and SSL)\", \"comment\": \"Dear Reviewer WjYy,\\n\\nWe sincerely thank you for recognizing our effort to enhance the novelty of the OBSER framework through its application to a photo-realistic (real-world-like) environment. We would like to share our perspective regarding the concerns you have raised.\\n\\n- As mentioned in LN 522, we observed that SSL models perform more robustly than metric learning model (SupCon) in the Replica environment, especially in unseen settings. In Appendix C.3.2 and C.4, we also observed similar results with the Minecraft environment. We claim that in this phenomenon, self-supervised learning methodologies enable representations to encompass more information than the class information defined by metric learning, thereby facilitating robust sub-environment recognition in cases where data is less structured or newly occurred (unseen).\\n- We built episodic memory of sub-environments differently for experiments with each environment. We designed the ImageNet experiment to show the relationship between EDS values and measures of sub-environment recognition. Therefore, in ImageNet, we sampled data from the dataset using arbitrary class distributions.\\n- On the other hand, in Minecraft and Replica, we gathered data with ego-centric observations to demonstrate the real-world agents: we first divided the environment into sub-environments, gathered scene observation, and extracted object observations to form episodic memory. The gathered data often contains ambiguities, such as obscured observations and ambiguous objects, and dealing with such ambiguities is important to achieve real-world agents. We claim that the results of Minecraft and Replica experiments are more essential than the results of ImageNet in showing the applicability of the OBSER framework to the real world.\\n- In summary, metric learning models enable numerically more precise inference in well-defined data and environments but may face difficulties in generalizing to less-structured data and environments. In contrast, SSL models may produce relatively noisy measurements for the inference but demonstrate greater strengths in generalization (LN 1628-1635, Appendix C.4). We claim that using the SSL-based OBSER framework is more effective than metric learning-based models for achieving real-world agents.\\n\\nTo ensure that our observation is effectively conveyed to the readers, we will incorporate more details about it in Section 6 (Experiments) and Section 7 (Conclusion). We hope this sufficiently addresses your concerns and helps readers better understand our work.\\n\\nOnce again, we deeply appreciate your interest and commitment to our work.\\n\\nBest regards,\\n\\nAuthors of Submission 6783\"}", "{\"comment\": \"The authors mostly address my concerns.\\n\\nAs a minor comment, Figure 1 is still not very expressive, and the caption is not very informative or obvious given the figure.\"}", "{\"title\": \"Response to Reviewer d5mr (1/1)\", \"comment\": \"Dear Reviewer d5mr,\\n\\nWe are thoroughly thankful for your thoughtful and positive feedback. We appreciate that you found the structure of the paper clear and appreciated the detailed explanations of our methodology, including the statistical and empirical aspects of our approach. Here is our response to your review.\\n\\n\\n\\\"which parts of the proposed methods, exactly, are novel\\\". \\\"connect the motivation of this work to the tasks and results shown\\\": Our main goal of this work is introducing sub-environment recognition with three fundamental relationships and suggesting metric learning (or SSL) based approaches to estimate these relationships. To make sure the proposed framework can be applied to solve tasks in a realistic environment, we conduct additional experiments with object-retrieval tasks in a Replica environment. We found that with sub-environment recognition, an agent can more efficiently explore by prioritizing the order of exploration from relevant sub-environment to irrelevant ones.\\n\\n\\n\\\"complex natural environments is lost by the simplicity of the test settings\\\": We choose Replica environment as an additional environment because the environment is photorealistic environment with high quality 3D meshes (MP3D environment is much larger, but somewhat noisy). We gather ego-centric scene observations to extract object observations for episodic memories, making the problem harder because the data is less structured. However, as a result of the experiments, we found that our proposed framework can reliably infer in such environments. For more details, please refer to Section 6.2 and Appendix D.\\n\\n\\n\\\"adjust the tone of the claims\\\": We have found that some expressions you mentioned might reduce the clarity of our claims. Thus, we have adjusted them to focus on our main research topic. You can find the difference in the abstract, introduction, several captions, and conclusion.\\n\\n\\n\\\"Section 4: Which of these are new propositions \\\": We newly define the two important properties of representations (epsilon and delta) in kernel densities. We also \\\"rederive\\\" the optimization of the EDS function, minimizing the KL divergence, to find an upper bound in terms of epsilon and delta. With this upper bound, we can guarantee the optimization of such loss can bring the optimization of the EDS function. In previous works [1,2], the authors derive the InfoNCE from classifier or contrastive learning, which is different from distribution matching. Optimization with KL divergence is introduced in [3], but they used an alternative way to derive the optimization problem.\\n\\n\\nWe also have added a y-axis for several figures and modified expressions for more readability.\\n\\n\\n[1] Oord, Aaron van den, Yazhe Li, and Oriol Vinyals. \\\"Representation learning with contrastive predictive coding.\\\"\\u00a0*arXiv preprint arXiv:1807.03748* (2018).\\n\\n[2] Khosla, Prannay, et al. \\\"Supervised contrastive learning.\\\"\\u00a0*Advances in neural information processing systems* 33 (2020): 18661-18673.\\n\\n[3] Choi, Won-Seok, et al. \\\"DUEL: Duplicate Elimination on Active Memory for Self-Supervised Class-Imbalanced Learning.\\\"\\u00a0*Proceedings of the AAAI Conference on Artificial Intelligence*. Vol. 38. No. 10. 2024.\"}", "{\"summary\": \"This paper proposes the Object-Based Sub-Environment Recognition (OBSER) framework. OBSER identifies sub-environments with three relationships: object-object, object-environment, and environment-environment relationship. The effectiveness of OBSER is measured with the proposed statistically separable (EDS) function in the Minecraft environment.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": [\"OBSER identifies sub-environments with three relationships between objects and environments, and exhibits better distinguishability in terms of EDS compared to other off-the-shelf vision models.\"], \"weaknesses\": \"- The application of OBSER is not clear. I'm not sure how OBSER will facilitate downstream tasks, e.g. decision agents in Minecraft like DreamerV3[1], Voyager [2], or GITM [3].\\n\\n[1] Mastering Diverse Domains through World Models\\n\\n[2] An Open-Ended Embodied Agent with Large Language Models\\n\\n[3] Generally Capable Agents for Open-World environments via Large Language Models with Text-based Knowledge and Memory\", \"questions\": [\"What if the category of possible objects is unknown, e.g. in the open-set setting?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"1\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to Reviewer 9vYe (1/1)\", \"comment\": \"Dear Reviewer 9vYe,\\n\\nWe appreciate that you highlighted our intention of introducing the OBSER framework and EDS function. We are also thankful for the insightful comments about applying our framework to the embodied agents. Here is our response.\\n\\n\\u201cI find that this method may be challenging to implement for embodied robots\\u201d: In this paper, we focus on the inference of the sub-environment recognition, so we assumed that the framework can work independently with embodiments when it can reach the observations. For the Minecraft experiment, we deployed randomly exploring agents to gather ego-centric observations. In the Replica experiment, we randomly gathered scene observations. We think both are commonly used methods for gathering environmental information with embodied agents.\\n\\n\\u201climited, or excessive experience\\u201d : Our framework uses self-supervised learning models (and metric learning) and data augmentation can be used to enhance the sample efficiency. This method has enhanced performance in offline RL [1]. Also, our method uses representation, not data itself, and time complexity (and also space complexity) only takes O((M+N)^2*D) with memory size M, N, and dimension D for KL divergence estimation. Even when the experience is huge to harm the memory or computation cost, we can reduce unnecessary (or duplicated) data from the memory.\\n\\n\\u201chow the object-object, object-environment, and environment-environment relationships contribute to embodied tasks\\u201d : To show the applicability of the OBSER framework in embodied tasks, we validate the chained inference of the OBSER framework for object retrieval tasks, such as visual navigation. We use the Replica environment and set each room as a sub-environment. As a result, we found that retrieval with a small number of relevant rooms (top-1 or top-3) can bring a similar performance to one with all rooms (48 or 35). This indicates that sub-environment recognition can enhance the efficiency of both inference and exploration.\\n\\n\\u201cThe diversities of Minecraft environment and objects seems limited.\\u201d : In the Minecraft environment, we chose common biomes in \\u201coverworld,\\u201d the main map of the game. There are some similar biomes that contain the same objects. However, this makes it harder to infer the sub-environments accurately. The proposed framework can successfully detect such minor differences (details) between biomes.\\n\\n\\u201cobjects are scarce or ambiguous, \\u2026 in less structured real-world space\\u201d : As mentioned in the above responses, the proposed framework also shows robustness and efficiency with the object retrieval task in photo-realistic environments. For this experiment, we used randomly gathered observations which are less structured and contain ambiguous data, instead of curated ones.\\n\\n[1] Schwarzer, Max, et al. \\\"Data-efficient reinforcement learning with self-predictive representations.\\\"\\u00a0*arXiv preprint arXiv:2007.05929* (2020).\"}" ] }

[ "Bin Xie", "Hao Tang", "Bin Duan", "Dawen Cai", "Yan Yan", "Gady Agam" ]

[ "SAM", "Auto-prompt", "Medical Image Segmentation" ]

{ "note_id": [ "dRSdzOuWmP", "Y0N4NU5z1i", "CiOJkt1x5u", "A2JHhsBo1d", "8964J7OJGl" ], "note_type": [ "official_review", "comment", "official_review", "official_review", "official_review" ], "note_created": [ 1730672718354, 1731704760268, 1730542675182, 1730303231580, 1730552357305 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission10824/Reviewer_rzag" ], [ "ICLR.cc/2025/Conference/Submission10824/Authors" ], [ "ICLR.cc/2025/Conference/Submission10824/Reviewer_Qp4u" ], [ "ICLR.cc/2025/Conference/Submission10824/Reviewer_f2fR" ], [ "ICLR.cc/2025/Conference/Submission10824/Reviewer_pPir" ] ], "structured_content_str": [ "{\"summary\": \"This work introduces MaskSAM, a prompt-free adaptation of SAM, incorporating a learnable classifier token and a 3D depth-MLP adapter. This works aims to address the meaningful challenge of adapting the foundation model of SAM. However, the adapter design is fairly common in prior works, and the contribution seems to be limited, with only marginal improvements in both experimental results and methodology.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. This work is well written and easy to understand.\\n2. This work aims to solve a meaningful problem to adapt SAM, by introducing learnable classification tokenizer and extend to 3D embedding.\\n3. The proposed method is simple and effective to some extent.\", \"weaknesses\": \"1. My primary concern is the lack of novelty. As outlined in the paper, numerous studies have already explored adaptations of this approach for medical image segmentation. This work does not present a clearly innovative idea or substantial differentiation from existing benchmarks, making its contribution appear more incremental than original.\\n\\n2. The improvement in experimental performance also seems minimal. As shown in Table 1 and Table 2, the proposed method does not demonstrate significantly better results compared to existing approaches.\", \"questions\": \"1. Please include a figure and description that clearly illustrate the high-level novelty of this work.\\n\\n2. For Figure 3, consider adding caption details to make the comparison more understandable.\\n\\n3. In Tables 1 and 2, the improvement appears marginal. Please provide a description to explain this result.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"withdrawal_confirmation\": \"I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors.\"}", "{\"summary\": \"The study proposes MaskSAM, an automatic prompt-free model framework for medical image segmentation. This work aims to improve the limitations of the classic Segment Anything Model (SAM), adapting it to meet the demands of multi-class labeling and 3D information in medical imaging.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. MaskSAM achieves multi-class medical image segmentation without the need for additional prompts by generating auxiliary classifier tokens, binary masks, and bounding boxes, marking progress in addressing SAM's need for multi-class labeling.\\n\\n2. The paper introduces 3D depth-convolution and depth-MLP adapters, extending SAM's 2D architecture to 3D medical imaging, which is innovative for handling volumetric data such as CT and MRI.\\n\\n3. Experimental results on several medical datasets (AMOS2022, ACDC, Synapse) demonstrate that MaskSAM outperforms existing self-supervised or supervised methods, such as nnUNet, in terms of Dice score.\", \"weaknesses\": \"1. It is recommended that the authors consider simplifying the model\\u2019s modules without compromising performance to reduce computational complexity and improve practical applicability.\\n\\n2. Conduct more detailed ablation studies on the contributions of different modules to better demonstrate the impact of the 3D adapters and prompt-free generation framework.\\n\\n3. Explore MaskSAM's performance on other types of medical data (e.g., pathology slide images or other 2D image types) to verify its generalization capability across different modalities.\\n\\n4. It is suggested to include more interpretative images and visualization methods to help readers understand the model\\u2019s feature-capturing capability and classification decisions.\", \"questions\": \"None.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The paper introduces MaskSAM, a novel adaptation framework of the \\\"Segment Anything Model\\\" (SAM) for medical image segmentation, overcoming SAM's limitations in this field. MaskSAM addresses the unique challenges of medical image segmentation by creating a prompt-free solution that supports 3D medical images and generates semantic labels for masks. The main contributions of MaskSAM include:\", \"prompt_free_mask_classification\": \"Unlike the original SAM, MaskSAM can predict semantic labels without additional prompts. It achieves this through a new prompt generator integrated with the SAM image encoder to produce auxiliary binary masks and bounding boxes as prompts, enhancing the adaptability to complex medical images.\", \"3d_adaptation\": \"MaskSAM adapts the SAM\\u2019s 2D architecture to 3D medical images, incorporating custom-designed adapters like the 3D depth-convolution adapter (DConvAdapter) for image embeddings and the 3D depth-MLP adapter (DMLPAdapter) for prompt embeddings, which extract 3D features crucial for accurate medical segmentation.\", \"soundness\": \"2\", \"presentation\": \"1\", \"contribution\": \"2\", \"strengths\": [\"1. Prompt-Free Method for Medical Image Segmentation:\", \"MaskSAM introduces a prompt-free segmentation scheme, which overcomes the original SAM model's dependence on prompts, making it more efficient in medical image segmentation tasks.\", \"Through a specially designed Prompt Generator, MaskSAM is able to generate auxiliary binary masks and bounding boxes, solving the problem that the SAM model cannot generate semantic labels in multi-category medical images. This improvement improves the adaptability of the model, making it more suitable for complex multi-category medical images.\", \"2. Enhanced 3D image adaptation capabilities:\", \"Medical images (such as CT and MRI scans) are usually 3D data. MaskSAM successfully extends the SAM model from 2D images to 3D images through innovative 3D adapters (DConvAdapter and DMLPAdapter), enhancing its applicability to medical images. These adapters significantly improve the model's ability to capture three-dimensional spatial information by adding deep convolutions and deep MLPs, making it perform well in complex 3D medical image segmentation tasks.\", \"3. Maintain the original SAM model structure and avoid model reconstruction:\", \"When adapting to medical images, MaskSAM retains the structure of the original SAM model without making major changes to the model framework. This design retains the zero-shot segmentation capability of the SAM model and achieves adaptation to medical images through a lightweight adapter module. This approach makes it easier for MaskSAM to migrate to the medical field without destroying the advantages of the SAM model, and has high scalability.\"], \"weaknesses\": \"1.High computational cost: The article mentions that MaskSAM introduces multiple adapters and generators on the basis of retaining the SAM structure, but does not elaborate on its computational resource requirements and model inference speed. In practical applications, this complex structure may bring high computational costs, especially when processing 3D medical images, which requires a lot of computing resources. Therefore, in some low-computing environments, the practical application value of MaskSAM may be affected.\\n\\n2.Limited applicability and innovation of the method: Although MaskSAM proposes a promptless method suitable for medical images, the core of the method is still based on the existing SAM framework, and the innovation is mainly focused on the adjustment and adaptation of SAM. The implementation of 3D adaptation (e.g., DConvAdapter and DMLPAdapter) is innovative but relatively incremental innovation on the basis of 3D data processing and does not demonstrate higher-level generality or versatility.\\n\\n3.Clarity and structure issues: The current structure and presentation of the paper is somewhat confusing, which affects the readability and understanding of the MaskSAM framework. For example, Sections 2.6 and 2.7 have the same title, which may lead to confusion about the distinction and importance of these sections. In addition, the location of various adapters (e.g., DConvAdapter and DMLPAdapter) is not clearly described in the context of the architecture, making it unclear to the reader what their specific role and integration points are in the model. Providing a clear and coherent overview with clearly defined section titles and visual illustrations of the adapter locations would greatly enhance understanding and readability.\\n\\n4.Generalization and scope of experiments: While the paper introduces the hint generator as a key innovation, its impact and generalization potential are still limited due to the focus on a single dataset or a narrow set of datasets. The effectiveness of MaskSAM on medical images from different domains that it has not seen has not been fully demonstrated, which is crucial to the value of establishing a generalizable hint-free segmentation model. By extending the experiments to multiple datasets from different medical imaging domains, this paper can strongly validate the broader applicability of the hint generator and justify its inclusion as a meaningful enhancement.\", \"questions\": \"1. The content of the article is too confusing. (For example: Why are the titles of Sections 2.6 and 2.7 the same? There is no clear description of where the various adapters are placed)\\n\\n2. You can conduct experiments on multiple fields of the same medical type to prove the domain generalization ability of MaskSAM. If you only verify it on a dataset of the same domain, then the significance of Prompt Generator is limited.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper presents a novel Segment Anything variant called \\u201cMaskSAM.\\u201d The proposed prompt generator alleviates the SAM\\u2019s dependence on manual prompts during the testing phase, and the re-designed adapters effectively leverage 3D information in medical images. Experiments on public datasets demonstrate that this framework achieves good performance.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"1. The writing is well and easy to follow.\\n2. The idea of auto-prompting the segment anything model is interesting.\\n3. The proposed adapters can effectively utilize the 3D information in the image.\", \"weaknesses\": \"1. The claim \\u201cMaskSAM is the first prompt-free SAM-based framework that retains the full structure of the original SAM\\u201d is inaccurate. There are many related works that maintain the complete structure of SAM (including prompt encoder and mask decoder) and focus on auto prompting, such as:\\n* AlignSAM [*1], which leverages reinforcement learning to generate prompts.\\n* PerSAM [*2] and Matcher [*3], which leverages feature similarity to generate prompts.\\n2. As this work focuses on auto prompting, it is better to visualize the class-aware bboxes generated by the model.\\n3. The compared methods are a bit outdated. In Table 2, the latest non-SAM method is UNETR, which is proposed in 2022. In fact, there have been some high-impact efforts in the last two years that achieved good results on the Synapse dataset. For example, UCTNet [*4] and UNETR++ [*5] achieve the mean DSC of 89.44% and 87.22% respectively. Considering that the total number of parameters of MaskSAM should also be much larger than that of UCTNet and UNETR++, the superiority of MaskSAM is unclear.\\n4. Which variant of SAM is this study based on? (ViT-B, ViT-L, or ViT-H)\\n5. Since the authors have added many additional trainable components to SAM, it is necessary to report the trainable parameters of different SAM-based methods.\\n6. (Optional) SAM2 [*6], the successor to SAM, was proposed at the end of July. It is encouraged to discuss some SAM2 variants that are also designed for medical image segmentation, such as [*7, *8, *9].\\n\\n[*1] AlignSAM: Aligning Segment Anything Model to Open Context via Reinforcement Learning. CVPR 2024\\n\\n[*2] Personalize Segment Anything Model with One Shot. ICLR 2024\\n\\n[*3] Matcher: Segment Anything with One Shot Using All-Purpose Feature Matching. ICLR 2024\\n\\n[*4] UCTNet: Uncertainty-guided CNN-Transformer hybrid networks for medical image segmentation. Pattern Recognition 2024\\n\\n[*5] UNETR++: Delving Into Efficient and Accurate 3D Medical Image Segmentation. IEEE Transactions on Medical Imaging 2024\\n\\n[*6] Sam 2: Segment Anything in Images and Videos. arXiv\\n\\n[*7] Medical sam 2: Segment medical images as video via segment anything model 2. arXiv\\n\\n[*8] Biomedical SAM 2: Segment Anything in Biomedical Images and Videos. arXiv\\n\\n[*9] SAM2-UNet: Segment Anything 2 Makes Strong Encoder for Natural and Medical Image Segmentation. arXiv\", \"questions\": \"See weakness.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"details_of_ethics_concerns\": \"None.\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}" ] }

[ "Muhammad Uzair Khattak", "Muhammad Ferjad Naeem", "Jameel Hassan Abdul Samadh", "Muzammal Naseer", "Federico Tombari", "Fahad Shahbaz Khan", "Salman Khan" ]

[ "Video Large Multi-modal Models", "Complex Reasoning", "Prompting for Multi-modal models" ]

{ "note_id": [ "dYWqTjazeE", "Rf9wZVzVHl", "IddE2qfyIZ", "GiWd6ivXMi", "EwYM2LC2if" ], "note_type": [ "official_review", "official_review", "official_review", "official_review", "comment" ], "note_created": [ 1730453046310, 1729499598236, 1729552757440, 1730277741879, 1732121841198 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission6697/Reviewer_sfTd" ], [ "ICLR.cc/2025/Conference/Submission6697/Reviewer_yDR5" ], [ "ICLR.cc/2025/Conference/Submission6697/Reviewer_SmJP" ], [ "ICLR.cc/2025/Conference/Submission6697/Reviewer_D1DJ" ], [ "ICLR.cc/2025/Conference/Submission6697/Authors" ] ], "structured_content_str": [ "{\"summary\": \"In this paper, the authors present a Complex Video Reasoning and Robustness Evaluation Suite (CVRR-ES) for assessing the performance of Video-LMMs on 11 diverse real-world video dimensions. Moreover, by evaluating 11 recent Video-LMMs, they further develop a training-free Dual-Step Contextual Prompting (DSCP) technique to boost enhance their performance on the proposed benchmark.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"1 Topic is good. Video understanding is an important problem in the multimodal research.\\n\\n2 Contribution is good. Instead of investigating the exsiting comprehension abilities, the authors propose to focus on reasoning capabilities over complex videos and robustness to user prompts. Both aspects are important parts for video understanding evaluation.\", \"weaknesses\": \"1 The build-up of this benchmark is similar to MVBench. In Figure 1, It is not correct that, MVbench does not contain \\\"in the wild\\\" and \\\"contextual dependency\\\". Moreover, MVbench contains the reasoning QA. Please further make the clarification on the difference, the robustness to user prompt? More explanation should be added.\\n\\n2 Dual Step Contextual Prompting is straightforward.\\n\\n3 The firgures are way to small. They are not quite easy to see.\", \"questions\": \"Please see weakness.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper proposes a new complex video reasoning and robustness benchmark, CVRR-ES, to assess Video-LLMs. CVRR-ES includes 2,400 high-quality open-ended question-answer pairs, spanning 214 high-quality videos, covering 11 video evaluation dimensions. This work evaluates 11 models, including closed-source and open-source Video-LLMs and establishes a human baseline. It also introduces a training-free dual-step contextual prompting method, DSCP, to enhance the performance of Video-LLMs.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"1\", \"strengths\": [\"The motivation and issues discussed in the paper are valuable. By constructing the CVRR-ES to assess the reasoning ability and robustness of existing Video-LLMs on real-world videos, the benchmark is very detailed in the design of video evaluation dimensions.\", \"The construction of a human baseline provides a good reference for the evaluation of Video-LLMs.\"], \"weaknesses\": [\"I express skepticism about whether the number of videos in the benchmark can achieve a robust assessment. The CVRR-ES benchmark includes only 214 videos, with the shortest video being just 2 seconds. Upon reviewing several videos from the anonymous link, I noticed a significant proportion of short videos. I question whether such short videos can adequately cover 11 categories. Moreover, current work that focuses solely on designing Video-LLMs, without specifically constructing evaluation benchmarks, provides a much larger number of assessment videos than the 214 included in CVRR-ES, for example, Tarsier [1].\", \"As mentioned in the previous question, the distribution of videos of different lengths within the benchmark is crucial for the assessment of reasoning ability and robustness, and the paper does not provide relevant explanations. The authors should include a table showing the distribution of video lengths across the dataset, and explain how they ensured a balanced representation of different video lengths across the 11 categories.\", \"In the motivation, it is mentioned that the goal is to build human-centric AI systems. Does the paper's reflection on this point merely consist of providing a human baseline? I think that offering more fine-grained visual examples would be more helpful for human-AI comparisons.\", \"I think that the contribution of the DSCP is somewhat overstated and lacks novelty. Such prompt engineering-based methods have already been applied in many works for data generation, model evaluation, and other stages. The introduction and ablation experiments of this technology in the paper seem redundant.\", \"The discussion on DSCP occupies a significant portion of the experimental analysis. I think that the current analysis provided in the paper lacks insight and does not fully reflect the value of CVRR-ES, especially in terms of human-machine comparison.\", \"The phrase should be \\\"there exist a few limitations\\\" instead of \\\"there exist few limitations\\\" in line 520.\", \"The paper does not provide prompt templates for all the closed-source and open-source Video-LLMs used, which will influence the reproducibility.\", \"The problems discussed in this paper are valuable, but the most crucial aspects of benchmark construction and evaluation are not entirely convincing. Instead, a significant amount of space is dedicated to introducing the DSCP method. I don't think it meets the acceptance standards of ICLR yet. I will consider modifying the score based on the feedback from other reviewers and the authors' responses.\", \"***\", \"[1] Wang J, Yuan L, Zhang Y. Tarsier: Recipes for Training and Evaluating Large Video Description Models[J]. arXiv preprint arXiv:2407.00634, 2024.\"], \"questions\": \"Please see the \\\"Weaknesses\\\".\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"details_of_ethics_concerns\": \"None\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The authors introduce a new video-lmm benchmark, aiming to target specifically reasoning vs perception. They also propose a CoT prompting technique, which was shown improves performance on benchmarks.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"1\", \"strengths\": [\"shows that correct prompting on video benchmarks can improve performance\", \"proposes a human-verified benchmark (not fully automatic pipeline), which focuses on CoT.\"], \"weaknesses\": [\"lacks comparison to many popular benchmarks: Video-MME, PerceptionTest, MLVU, LongVideoBench, EgoSchema, TempCompass.\", \"note that some of these benchmarks have a higher overlap with the proposal\", \"Uses LLM as a judge. This makes evaluation expensive, especially at scale.\"], \"questions\": [\"can you calculate the correlation between scores on this benchmark to existing ones (e.g., r^2) - so we can see how different the proposed benchmark is from existing ones?\", \"Has a multiple-choice alternative been tested, and if yes, why did you opt not to use this?\", \"how did you verify that this benchmark requries a 'higher' level of reasoning compared to existing benchmarks?\", \"does DSCP impact performance more on the proposed benchmark compared to existing benchmarks?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper introduces a novel benchmark suite, the Complex Video Reasoning and Robustness Evaluation Suite (CVRR-ES), designed to comprehensively assess the robustness and reasoning capabilities of Video Large Language Multimodal Models (Video-LMMs) in complex video understanding and real-world contexts. The authors also propose a training-free Dual-Step Contextual Prompting (DSCP) method, which leverages a two-stage prompting strategy to enhance model reasoning and robustness. By providing contextual cues, DSCP helps guide models toward more accurate responses and reduces misinterpretations of confusing or misleading queries. Experimental results demonstrate that current Video-LMMs often display overly affirmative responses when handling misleading or negatively framed text queries. Additionally, the experiments reveal challenges in these models' understanding of emotional and social contexts, as well as incomplete actions. The DSCP technique substantially improves model performance across these challenging dimensions.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"1\", \"strengths\": \"The benchmark\\u2019s 11 evaluation dimensions are challenging and effectively assess model capabilities in complex scenarios.\\nThe 10 principled instructions in DSCP significantly enhance the ability of Video LMMs to handle these complex tasks more effectively.\", \"weaknesses\": \"The contribution to the field of Video LLMs is somewhat limited. Much recent work in the community has focused on evaluating reasoning capabilities, with benchmarks like MVBench already addressing reasoning alongside a broader range of evaluation angles. Given this context, focusing primarily on complex reasoning as a contribution may have limited impact on advancing the field and may not fully meet the standards expected for ICLR.\\nUsing GPT-3.5 as the evaluation LLM also appears somewhat outdated. \\nFurthermore, achieving the complex reasoning that the paper discusses requires fine-grained understanding of video content, which typically demands inputting a substantial number of video frames into the model. With a few frames, handling complex reasoning tasks becomes challenging. However, the paper does not specify the number of frames input to the open-source models, and intuitively, increasing the number of frames could enhance Video LLMs\\u2019 understanding and reasoning over video content.\\nLastly, while DSCP is presented as a novel method, it is relatively basic, relying on a two-stage text prompt that appears somewhat rough and may not constitute a substantial contribution.\", \"questions\": \"What was the number of frames inputted for the open-source models during evaluation?\\nCould the authors provide a summary of the complex reasoning evaluation dimensions? The 11 dimensions seem quite detailed; grouping some of them into broader, more general capabilities could make the benchmark easier to understand and more accessible for users, potentially aiding wider adoption.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"details_of_ethics_concerns\": \"N/A\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"withdrawal_confirmation\": \"I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors.\"}" ] }

[ "Mikołaj Małkiński", "Szymon Pawlonka", "Jacek Mańdziuk" ]

[ "Multimodal Large Language Models", "Abstract Visual Reasoning", "Bongard Problems" ]

{ "note_id": [ "z35NroXjUC", "r0LGpAiNOa", "nol0SPLqI3", "mTW3O5wjC0", "m7d6q0ZwlT", "m7AO6Q9U9o", "jXOYlGsmIE", "dDm4xnYU34", "ZUKMMzpCMB", "Z5w4MHB2wD", "YdN5qnlawl", "XpVPGgLY3x", "UfG7540BAU", "NftHtEqDF9", "I5Xk759eyw", "Fn8iyfEtZJ", "FjgFkCFOlY", "BUxgkiawya", "BHXw8kjt1R", "AtQrnGrOpX", "9C5WMy2hYa", "4Pv9MOc7a6", "3gOdXuQ0uz", "2So1lhTcgk", "2QAdGemI61" ], "note_type": [ "official_comment", "official_comment", "official_comment", "meta_review", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review", "official_comment", "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment" ], "note_created": [ 1732405295302, 1732404240751, 1733216249246, 1734748116955, 1733118893123, 1732405507306, 1732784603569, 1730793978388, 1731215332502, 1730250249968, 1732901276454, 1737523771528, 1733022414231, 1732405107974, 1732404167368, 1732404484879, 1732405384073, 1732404701118, 1731024879967, 1732405215294, 1732660806345, 1732585956720, 1732404963885, 1733216117992, 1732660836134 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission6466/Authors" ], [ "ICLR.cc/2025/Conference/Submission6466/Authors" ], [ "ICLR.cc/2025/Conference/Submission6466/Authors" ], [ "ICLR.cc/2025/Conference/Submission6466/Area_Chair_qeHQ" ], [ "ICLR.cc/2025/Conference/Submission6466/Reviewer_VpcT" ], [ "ICLR.cc/2025/Conference/Submission6466/Authors" ], [ "ICLR.cc/2025/Conference/Submission6466/Reviewer_VpcT" ], [ "ICLR.cc/2025/Conference/Submission6466/Reviewer_VpcT" ], [ "ICLR.cc/2025/Conference/Submission6466/Reviewer_VQyh" ], [ "ICLR.cc/2025/Conference/Submission6466/Reviewer_9XwS" ], [ "ICLR.cc/2025/Conference/Submission6466/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission6466/Reviewer_YYwM" ], [ "ICLR.cc/2025/Conference/Submission6466/Authors" ], [ "ICLR.cc/2025/Conference/Submission6466/Authors" ], [ "ICLR.cc/2025/Conference/Submission6466/Authors" ], [ "ICLR.cc/2025/Conference/Submission6466/Authors" ], [ "ICLR.cc/2025/Conference/Submission6466/Authors" ], [ "ICLR.cc/2025/Conference/Submission6466/Reviewer_YYwM" ], [ "ICLR.cc/2025/Conference/Submission6466/Authors" ], [ "ICLR.cc/2025/Conference/Submission6466/Authors" ], [ "ICLR.cc/2025/Conference/Submission6466/Reviewer_YYwM" ], [ "ICLR.cc/2025/Conference/Submission6466/Authors" ], [ "ICLR.cc/2025/Conference/Submission6466/Authors" ], [ "ICLR.cc/2025/Conference/Submission6466/Authors" ] ], "structured_content_str": [ "{\"title\": \"Response to the Reviewer VpcT\", \"comment\": \"> To further understand the limitations, do you think the same model with different sizes (LLaVA 1.6 7b/13b/70b) will highlight some observations? I think it can reveal how scaling law will help with this problem.\\n\\nThank you for this excellent suggestion. We conducted an experiment to investigate how scaling the number of model parameters impacts AVR capabilities. To this end, we explored diverse model sizes including proprietary and open-access models. Firstly, we evaluated GPT-4o mini and Gemini 1.5 Flash, which are smaller alternatives of the already used GPT-4o and Gemini 1.5 Pro models, resp. Secondly, we employed several larger variants from the InternVL2 and LLaVA-NeXT model families. Overall, the set of considered models includes InternVL2-8B, InternVL2-26B, InternVL2-40B, InternVL2-Llama3-76B, LLaVA-v1.6 Vicuna-13B, LLaVA-v1.6 34B, LLaVA-NeXT 72B, and LLaVA-NeXT 110B. We conducted experiments on all 4 datasets using two solution strategies, including **Direct**, which is an intuitive baseline, and **Descriptive**, identified as the most effective strategy in the main experiments.\\n\\nResults and detailed analysis are presented in the newly introduced Appendix D. In summary, model scaling yields consistent improvements in AVR performance, especially for open-access models. Nonetheless, proprietary models exhibit strong performance even at smaller sizes, with GPT-4o mini performing at least as good as the best variant across open-access models in all but one cases.\\n\\nWhile these results show that model scaling can be indeed helpful in boosting AVR capabilities, they also suggest that large model size is not critical for strong performance in these tasks. Consequently, these results suggest that simply scaling the model size may be insufficient to achieve stronger abstract reasoning capabilities. It is highly relevant to investigate this aspect in future research, e.g., by incorporating AVR datasets into model training. Once again, we would like to thank you for this insightful comment.\\n\\n> Can you list the key reasons why MLLMs fail in BPs? This should be provided in the conclusion section. You need to expand \\\"broader limitations in their reasoning abilities\\\" in Line 537-538.\\n\\nThank you for this suggestion. In response, we have restructured Section 6 to expand \\u201cbroader limitations in their reasoning abilities\\u201d to specific insights drawn from our experiments.\\n\\nFurthermore, our evaluation revealed that models may fail to provide precise answers, especially when compared to human performance. For example, Fig. 12b highlights a case in which GPT-4o generated an almost correct concept (*\\u201c[...] Right: All images feature women in non-wedding attire, wearing dresses or suits of various colors other than white.\\u201d*), but overlooked a conflicting detail (one image featuring a woman in a white suit). This emphasizes the nuanced nature of BPs, where attention to details is critical for correctly recognizing the underlying concept. Developing robust MLLMs that consistently solve such tasks remains an open problem.\"}", "{\"title\": \"Response to the Reviewer VQyh\", \"comment\": \"> The new BP dataset created in the paper is quite difficult for the models. [...]\\n\\nThank you for your valuable suggestion. To evaluate the difficulty and validity of the proposed dataset, during the rebuttal period, we conducted a study with 30 human participants, as detailed in the newly introduced Appendix B. As shown in Fig. 7, humans solved an average of 39.2 problems, achieving 65% accuracy. Performance varied across participants, with the number of correctly solved problems ranging from 23 to 59, demonstrating that dedicated individuals are capable of solving almost all problems, and that in principle all 60 problems are solvable. Notably, the lowest number of problems solved by a human participant (23) exceeded the number of problems solved by all models in total (22, see \\u201cSolved by any model\\u201d in Fig. 22). In addition, Fig. 8 illustrates variability in problem difficulty: 22 problems were solved by at least 25 respondents, while 10 were solved by fewer than 10 test-takers. Notably, several problems easily solved by humans (e.g., 91, 92, 95, and 96) were not solved by any model, highlighting the need for further advances in this area.\\n\\n> It would be nice to show some human score for BPs. Is there any attempt toward that?\\n\\nHuman performance has also been evaluated for Bongard HOI and Bongard-OpenWorld, where participants achieved an average accuracy of 91% on both datasets. In comparison, the 65% accuracy observed in our study highlights that Bongard-RWR poses a greater challenge even for humans. Nonetheless, the high scores achieved by certain participants demonstrate that all tasks in Bongard-RWR are solvable with dedicated effort, making it a valuable and challenging testbed for advancing research in this area.\"}", "{\"title\": \"Response to Reviewer VpcT\", \"comment\": \"Dear Reviewer, thank you for your comments.\\n\\nWe believe that Bongard-RWR is not merely an incremental augmentation of Bongard HOI and Bongard-OpenWorld, as we use a fundamentally different methodology to construct the dataset. While BPs in Bongard HOI and Bongard-OpenWorld are sampled automatically from a larger image dataset to represent concepts grounded in the real-world, we construct each BP manually, or in certain cases rely on a semi-automated process followed by manual review of the obtained matrices. Most importantly, **Bongard-RWR focuses on concepts not covered in Bongard HOI and Bongard-OpenWorld, making it orthogonal to these prior works**.\\n\\nThe sample size of Bongard-RWR is indeed smaller than that of the remaining datasets, however, **the matrices qualitatively differ from existing datasets, by representing abstract concepts with real-world images**. We show that the introduced BPs pose a significant challenge for the contemporary methods, which we view as a novel and valuable direction to explore in the future work.\"}", "{\"metareview\": \"The AC acknowledges the authors\\u2019 efforts in exploring a potential direction for this field and providing dataset statistics. However, the paper's primary contribution, the dataset, lacks sufficient novelty, after careful checking and syncing with reviwers' commets, as it appears to primarily bridge synthetic and realistic Bongard problems through a data augmentation technique, building incrementally on existing datasets like Bongard HOI and Bongard-OpenWorld. Furthermore, the limited sample size of 100 raises methodological concerns regarding potential bias and insufficient statistical power, with no detailed justification provided for its adequacy. While the work has merit, these limitations lead to a reject recommendation for now.\", \"additional_comments_on_reviewer_discussion\": \"Reviewers engaged in discussion and with further concerns.\"}", "{\"comment\": \"Thank you for your response. However, it does not **directly** address my concerns regarding the dataset\\u2019s motivation and size. While I acknowledge the manual effort involved and the challenges of capturing fine-grained abstract concepts, the justification provided does not sufficiently differentiate Bongard-RWR from existing datasets like Bongard HOI and Bongard-OpenWorld, making it appear more like an incremental augmentation. Additionally, the small sample size of 100 remains a significant concern, as the justification does not convincingly demonstrate its adequacy for achieving the research objectives. Considering these issues alongside the feedback from other reviewers, I find it difficult to support acceptance of this paper.\"}", "{\"title\": \"Response to all Reviewers\", \"comment\": \"Dear Reviewers, we are grateful for your insightful comments and suggestions, which helped us to improve the paper. Following the received feedback, during the rebuttal time we improved the presentation of results in Section 5 by redesigning Table 1 and introducing Figs. 4 and 5. Additionally, we summarized key findings in Section 6, discussed limitations and opportunities for future work (Appendix A), conducted a human study including 30 participants to assess Bongard-RWR difficulty (Appendix B), moved the detailed discussion of binary classification tasks and the table with full results from the main paper to Appendix C, and performed a scaling law experiment (Appendix D). Labels of figures, tables, and sections mentioned in this response refer to the revised version of the manuscript. Specific comments are addressed in the individual responses to each Reviewer. We hope the provided responses adequately address the Reviewers' concerns.\"}", "{\"title\": \"Response to the Rebuttal\", \"comment\": \"Dear authors,\\n\\nI would like to thank you for your efforts to show the potential direction of this field and the statistics of the dataset. However, I am still not satisfied with the contribution part. The reason I am reiterating this aspect is that the dataset is the major contribution of this paper, thus it is critical to show what is new compared with previous datasets. It appears that the motivation is just bridging the Bongard problem in two domains (synthetic BPs and realistic BPs) by adapting them into another retrieved realistic domain, which I would regard as an augmentation technique to enrich the realistic data for the Bongard problems, as an increment to Bongard HOI and Bongard-OpenWorld. \\n\\nAdditionally, the sample size of 100 raises methodological concerns. Given the potential for various sources of bias, this limited sample may not provide sufficient statistical power to draw robust conclusions. I would encourage the authors to provide a more detailed justification for why this sample size is adequate for the research objectives.\\n\\nBest,\\nReviewer VpcT\"}", "{\"summary\": \"This paper investigates the Bongard Problems (BPs) as a case study in Multimodal LLMs. This problem represents a foundational problem in Abstract visual reasoning (AVR). This paper gathered a new dataset Bongard-RWR and used multiple strategies to test the proprietary and open-sourced MLLMs on both synthetic and real-world BPs. The experiments show that there is still a gap for MLLMs to do AVR, but certain strategies.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. The problem of Abstract visual reasoning (AVR) is a popular problem and has gained great attention lately. The perspective of Bongard Problems is novel and meaningful.\\n2. The experiments of this paper are comprehensive on various MLLMs and 4 datasets (one of them is newly collected by this paper).\\n3. The dataset is already open-sourced, which may contribute to the research community in the future.\", \"weaknesses\": \"1. The contribution of the proposed Bongard-RWR dataset is not clear enough, although the authors briefly mentioned in Line 311-312. The previous datasets already covered both synthetic and real-world settings. How do the proposed datasets differentiate from these datasets? It is unclear why the authors find previous datasets insufficient or lacking in meaningfulness due to differences in concepts.\\n2. While the paper evaluates current MLLMs, it fails to offer a clear direction for enhancing their AVR capabilities or to present any new insights beyond highlighting general limitations. The results of different strategies might only reflect the training data (instruction tuning data) for these models. \\n3. Since the newly gathered dataset is considered as a main contribution, there are no clear statistics of this dataset provided.\", \"questions\": [\"Can you provide the rationale behind the proposed strategies in the experiments? Does it relate to the theory of AVR from a cognitive science perspective? Can you summarize the high-level idea of the design?\", \"To further understand the limitations, do you think the same model with different sizes (LLaVA 1.6 7b/13b/70b) will highlight some observations? I think it can reveal how scaling law will help with this problem.\", \"Can you list the key reasons why MLLMs fail in BPs? This should be provided in the conclusion section. You need to expand \\\"broader limitations in their reasoning abilities\\\" in Line 537-538.\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The paper focuses on the capability of multimodal large language models (MLLM) to solve Bongard Problems (BP). The authors propose a new BP dataset, where the two sides of the images are distinguished by abstract visual concepts. The authors design a set of evaluation methods for MLLMs, including binary classifications and solving BP using different generation strategies. The results on the proposed BP dataset and three other BP datasets show that the abstract visual reasoning capabilities of MLLMs are limited in solving BP, especially in the proposed ones.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"The paper provides a comprehensive evaluation of the capabilities of vision language models on Bongard Problems (BPs). The paper includes eight models and 10 metrics to evaluate them on 4 types of BPs, including one created by the authors. I believe it will be very helpful for future study of BPs.\", \"weaknesses\": [\"The presentation of the paper needs to be improved. Despite numerous settings, new generation strategies and new dataset, the authors choose to present the results in just one table. It is almost impossible to interpret the table and verify the claims made by the authors without checking back and forth. I suggest reviewer break down the table into small ones, where each of them provides a clear message to the readers. If that takes more space than the current version, some of the experiment results can be moved to appendix.\", \"The observation from the results is currently not very interesting. The major claim made by the authors is that the capability of the VLMs on BPs are limited. Not many intuitions behind are obtained, part of it is because of the poor presentation, and the other part is probably because the observations are not very consistent across different models. I suggest the authors dive deep into the details and conduct additional experiments if necessary. Overall, I believe the contribution of the paper can be improved.\", \"The new BP dataset created in the paper is quite difficult for the models. I believe a comprehensive human study is necessary in this case to ensure the validity of the dataset. It is great that the dataset is released, but after a quick scan of the dataset, I personally find it very hard to distinguish the left and the right images for some BPs. Therefore, I suggest the authors to report human performance on the tasks and report consistency across human annotators.\", \"------\", \"### Post-rebuttal comments:\", \"The authors added a human study of the new BP dataset and re-organized the experiment section, which are important improvements for the paper. Therefore, I have raised my score.\", \"I agree with other reviewers that the dataset and the analysis can be made more rigorous and comprehensive. Therefore, I believe this paper has not reached the bar of NeurIPS.\"], \"questions\": \"It would be nice to show some human score for BPs. Is there any attempt toward that?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper investigates the abstract reasoning capabilities of multimodal large language models (MLLMs) using Bongard Problems (BPs), which require analogy-making between sets of images. The authors test several MLLMs on both synthetic and real-world BP datasets, highlighting significant limitations in their performance, especially with abstract, synthetic problems. Despite some improvement with real-world images, the models still struggle with effectively incorporating new information and handling complex visual reasoning tasks. To address this, the paper introduces a new dataset, Bongard-RWR, translating synthetic BP concepts into real-world images.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": [\"The paper studies an interesting topic of Bongard Problems (BPs). It introduces a comprehensive evaluation of MLLMs using BPs, which are notoriously challenging in abstract reasoning. By testing both synthetic and real-world BP datasets, the authors highlight critical weaknesses in these models, providing valuable insight into their current capabilities. This positions BPs as a powerful benchmark for future MLLM advancements.\", \"This paper creates Bongard-RWR to bridge the gap between abstract synthetic reasoning and real-world tasks. This dataset allows for more meaningful comparisons of MLLM capabilities across domains. Also, this dataset enables future work to pinpoint the source of reasoning failures more precisely.\", \"The paper is well written and easy to understand.\"], \"weaknesses\": \"I didn't see major weaknesses.\", \"questions\": [\"Some open-source models like InternVL2 and Phi-3.5V perform well on many benchmarks but struggle, nearing chance-level performance on Bongard Problems, while LLava-1.6 and Pixtral perform okay. What do the authors think causes these discrepancies?\", \"How closely does abstract visual reasoning (AVR) reflect real-world visual reasoning (VR)? For instance, if a model excels in AVR, can we expect it to perform well in VR, and vice versa? Some potential correlations experiments between AVR and VR performance might give some insights.\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to Reviewer VpcT\", \"comment\": \"We appreciate the reviewer\\u2019s comments and concerns about the contribution of Bongard-RWR and its sample size. While we acknowledge that a larger dataset would provide greater statistical power, our approach was shaped by two key factors.\\n\\nFirstly, Bongard-RWR was designed with significant manual effort, as detailed in Section 4.1. Of the 100 synthetic BPs considered, only 12 were automatically represented in the real-world domain using Algorithm 1. The remaining problems required manual construction, including the translation of abstract concepts to the real-world domain, image selection, and, in some cases, photographing manually constructed scenes. This approach ensures that the dataset captures fine-grained abstract concepts, such as \\u201cEnds of the curve are parallel vs. Ends of the curve are perpendicular\\u201d or \\u201cExtensions of segments cross at one point vs. Extensions of segments do not cross at one point\\u201d, which are challenging to represent with real-world images. In contrast, datasets like Bongard HOI and Bongard-OpenWorld, which were generated automatically using existing data sources, focus on coarse-grained concepts such as \\u201cA person jumping on a surfboard. vs. Not a person jumping on a surfboard\\u201d. While this automated process enables larger sample sizes, it limits the representation of fine-grained abstract concepts.\\n\\nSecondly, the cost of MLLM inference needs to be accounted for. As detailed in Section 3.2, we employ an MLLM voting committee to evaluate generated answers, which comprises 4 proprietary models. The total number of inferences scales with the number of datasets, problem instances per dataset, models evaluated, and generation strategies employed. Scaling the experiments to much larger datasets is currently infeasible within our research budget.\\n\\nWe hope that future advancements in optimizing MLLM inference costs will enable exploration of larger datasets. Until then, we believe that Bongard-RWR should provide a useful supplement to the existing real-world Bongard datasets (Bongard HOI and Bongard-OpenWorld) enabling testing of the abstract reasoning capabilities of machine learning models against detailed and nuanced aspects of the real-world scenes. We hope that the topic may attract other researchers to join this research path and contribute to the proposed version of the Bongard-RWR \\u2013 the dataset will be made freely available for research purposes.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"title\": \"Response and further concerns\", \"comment\": \"Thank you for your response. Regarding the first two questions, I understand that they might be challenging to address at this stage. However, for the last question, I would need a more detailed explanation. I have carefully reviewed your reply as well as Reviewer VpcT's follow-up, and it seems that the main issue remains unresolved. Specifically, the prior datasets already represent a significant contribution to this area, while your dataset appears to be a minor augmentation (like Reviewer VpcT proposed). However, the way your paper is written gives the impression that your contribution is overstated. Reviewer VpcT observed this problem and summaried into his/her/their comment.\\n\\nBased on Reviewer VpcT's comment, I want highlight another problem: the data selection process for your dataset remains unclear. As mentioned in your paper, the GPT-4o data generation process involves manual intervention at several steps. This raises concerns about how these manually selected images align with the design standards of Bongard problems. As I pointed out earlier, a potential risk is that readers might question whether your dataset was manually selected to emphasize the weaknesses of MLLMs, particularly in terms of the data are sampling from conditional probability distributions. This could contradict the foundational principles of Bongard problems and potentially undermine the validity of your conclusions.\\n\\nTo effectively explore MLLM performance and facilitate meaningful comparisons between human and MLLM capabilities, it would be more appropriate to use a widely collected dataset in general domain (e.g., sampling natural and symbolic images in Bongard HOI and Bongard-OpenWorld, and other datasets) for Bongard problems and evaluate the models within that distribution. This contradicts your current conclusion.\"}", "{\"title\": \"Response to the Reviewer VpcT\", \"comment\": \"> Since the newly gathered dataset is considered as a main contribution, there are no clear statistics of this dataset provided.\\n\\nEach problem instance in Bongard-RWR consists of 12 images (6 per side) and a text label, which limits the scope of traditional statistical summaries of the dataset itself. Instead, our analysis focused on evaluating the performance of different models and solution strategies, as detailed in Section 4. For instance, we noted several qualitative findings such as: *\\u201cThis suggests that correctly identifying concepts expressed in Bongard-RWR likely requires more advanced reasoning abilities, even in the relatively simpler binary classification setting.\\u201c* or *\\u201c[...] the weak results on Bongard-RWR suggest that the discrepancy is more related to the specific underlying concepts than the visual domain as such [...]\\u201d*.\", \"we_provide_further_information_about_bongard_rwr_in_the_supplementary_materials\": \"1. Appendix F, Fig. 13 outlines the structure of the dataset based on the approach taken to construct each problem instance.\\n1. Appendix F, Fig. 14 compares selected samples from Bongard-RWR to their synthetic counterparts.\\n1. Appendix G contains detailed performance metrics across models and strategies, with Bongard-RWR results presented in Figs. 21 and 22.\\n1. Appendix H directly compares model performance on Bongard-RWR to synthetic BPs (Bongard, 1970).\\n\\nAdditionally, in response to the Reviewers\\u2019 feedback, we have conducted a study involving human participants to relate model performance against a human benchmark. The study involved 30 human participants, as detailed in the newly introduced Appendix B. \\n\\nAs shown in Fig. 7, humans solved an average of 39.2 problems, achieving 65% accuracy. Performance varied across participants, with the number of correctly solved problems ranging from 23 to 59, demonstrating that dedicated individuals are capable of solving almost all problems, and that in principle all 60 problems are solvable. Notably, the lowest number of problems solved by a human participant (23) exceeded the number of problems solved by all models in total (22, see \\u201cSolved by any model\\u201d in Fig. 22). In addition, Fig. 8 illustrates variability in problem difficulty: 22 problems were solved by at least 25 respondents, while 10 were solved by fewer than 10 test-takers. Notably, several problems easily solved by humans (e.g., 91, 92, 95, and 96) were not solved by any model, highlighting the need for further advances in this area.\\n\\nWe hope that the above multi-faceted analysis offers a comprehensive view of the introduced dataset. If there are any particular statistics the Reviewer would find valuable that are missing in the paper, we would greatly appreciate further guidance.\"}", "{\"title\": \"Response to the Reviewer VQyh\", \"comment\": \"> The presentation of the paper needs to be improved. [...]\\n\\nThank you for this suggestion. In response, we have restructured our presentation to improve clarity and accessibility. The detailed table and discussion of binary classification tasks have been moved to Appendix C. A concise summary of binary classification results is now provided in a single paragraph accompanied by the newly introduced Fig. 4. Additionally, we have introduced Table 1, which highlights the performance of Direct, Descriptive, and Contrastive strategies, and Fig. 5, which illustrates the impact of the -direct and -iterative variants of the Descriptive and Contrastive strategies. We hope this revised format simplifies interpretation of results and improves the clarity of the insights drawn from our experiments.\\n\\n> The observation from the results is currently not very interesting. [...]\\n\\nWe hope that our response to the previous comment addresses the first part regarding poor presentation. In particular, we believe that Fig. 4 (top) highlights a potential bias in MLLMs agreeing or disagreeing with presented concepts, which emphasizes the importance of considering natural language generation setups for evaluating MLLM abstract reasoning capabilities. Fig. 4 (bottom) illustrates that while MLLMs perform well in tasks involving real-world concepts (e.g., Bongard HOI and Bongard-OpenWorld), they struggle with abstract concepts (e.g., synthetic BPs and Bongard-RWR). We further show that MLLMs do not benefit from human-like approaches to solving BPs (see \\u201cContrastive reasoning\\u201d), struggle to effectively utilize the context window (see \\u201cIterative reasoning\\u201d and Descriptive-iterative in Fig. 5), and require further work to consistently integrate text and vision modalities at the answer generation step (see \\u201cMultimodal answer generation\\u201d and Descriptive-direct in Fig. 5).\\n\\nWe agree that some observations vary across different models, but we believe this is an expected outcome given the diversity of MLLMs in terms of the data mixtures used during pre-training, fine-tuning and learning from human feedback, the architectures of their text and vision encoders, and the approaches for combining multimodal embeddings. The unique nature of BPs may further contribute to variability in model performance, as these analogy-based reasoning tasks are largely absent in standard MLLM data sources, which primarily include visual question answering or image captioning tasks.\\n\\nTo further investigate this variability, during the rebuttal time, we conducted an experiment exploring the impact of scaling model parameters on AVR capabilities. We explored diverse model sizes including both proprietary and open-access models. Firstly, we evaluated GPT-4o mini and Gemini 1.5 Flash, which are smaller alternatives of the already used GPT-4o and Gemini 1.5 Pro models, resp. Secondly, we employed several larger variants from the InternVL2 and LLaVA-NeXT model families. Overall, the set of considered models includes InternVL2-8B, InternVL2-26B, InternVL2-40B, InternVL2-Llama3-76B, LLaVA-v1.6 Vicuna-13B, LLaVA-v1.6 34B, LLaVA-NeXT 72B, and LLaVA-NeXT 110B. We conducted experiments on all 4 datasets using two solution strategies, including Direct, which is an intuitive baseline, and Descriptive, identified as the most effective strategy in the main experiments.\\n\\nResults and detailed analysis are presented in the newly introduced Appendix C. In summary, model scaling yields consistent improvements in AVR performance, especially for open-access models. Nonetheless, proprietary models exhibit strong performance even at smaller sizes, with GPT-4o mini performing at least as good as the best variant across open-access models in all but one cases.\\n\\nWhile these results show more consistent trends within a single model architecture, some inconsistencies remain, as certain smaller variants outperform their larger counterparts. This suggests that future efforts to improve MLLM abstract reasoning capabilities should devote larger focus to this aspect, e.g., by incorporating AVR datasets into model training.\"}", "{\"title\": \"Response to the Reviewer YYwM\", \"comment\": \"> [...] As we know, Human IQ test has a long history in psychometrics and only a few of works try to explore the different between MLLM and human in this topic [1].\\n\\nThank you for pointing out this highly relevant study. As it was published after the ICLR submission deadline, we were unable to include it in the initial version of our submission. However, based on your valuable suggestion, we have now incorporated a discussion of this work in the newly added Appendix A, within the paragraph \\u201cIncorporating proposed strategies to enhance abstract reasoning abilities\\u201d.\\n\\n> It is worth to highlight that BP is only a subtask of AVR test. [...]\\n\\nWe focused our study on BPs because they fundamentally require solvers to articulate answers in natural language, making them a valuable testbed for evaluating the reasoning capabilities of MLLMs. In contrast, tasks like Raven\\u2019s Progressive Matrices pose a discriminative challenge, which can make it difficult to distinguish whether a model\\u2019s solution is a result of true understanding or educated guess. Nevertheless, we agree that BPs represent only a subset of AVR tasks and that exploring a broader range of challenges is essential for advancing the field. To address this limitation, we have added a discussion in the newly included paragraph \\u201cGoing Beyond Bongard Problems\\u201d in Appendix A.\\n\\n> [...] The multi-image reasoning capbility of most of open-sourced MLLMs are not eligible in these tasks.\\n\\nWe specifically selected open-source models that support multi-image reasoning, as confirmed in their documentation:\\n1. https://huggingface.co/mistralai/Pixtral-12B-2409 \\u201cYou can also pass multiple images per message [...]\\u201c\\n1. https://huggingface.co/OpenGVLab/InternVL2-8B \\u201cInternVL 2.0 is trained with an 8k context window and utilizes training data consisting of long texts, multiple images, and videos, [...]\\u201d\\n1. https://huggingface.co/microsoft/Phi-3.5-vision-instruct \\u201cThe model provides uses for general purpose AI systems and applications with visual and text input capabilities which require: [...] 6. Multiple image comparison 7. Multi-image or video clip summarization [...]\\u201d\\n1. https://huggingface.co/docs/transformers/main/en/model_doc/llava_next#multi-image-inference \\u201cLLaVa-Next can perform inference with multiple images as input, [...]\\u201c\\n\\nOut of the 10 solution strategies employed in our work, 4 (Images to Sides, Contrastive, Contrastive-iterative, and Contrastive-direct) explicitly involve processing multiple images in a single forward pass. However, we recognize that multi-image training data is significantly less available compared to standard (image, text) pairs, which may partially explain the weaker performance of the Contrastive strategies compared to Descriptive ones. We believe that advancing multi-image training datasets is a critical step towards improving model performance in this regime.\\n\\n> For a better understanding, the user study between human and MLLM is also important for this task.\\n\\nThank you for your valuable suggestion. To evaluate the difficulty and validity of the proposed dataset, during the rebuttal period, we conducted a study with 30 human participants, as detailed in the newly introduced Appendix B. As shown in Fig. 7, humans solved an average of 39.2 problems, achieving 65% accuracy. Performance varied across participants, with the number of correctly solved problems ranging from 23 to 59, demonstrating that dedicated individuals are capable of solving almost all problems, and that in principle all 60 problems are solvable. Notably, the lowest number of problems solved by a human participant (23) exceeded the number of problems solved by all models in total (22, see \\u201cSolved by any model\\u201d in Fig. 22). In addition, Fig. 8 illustrates variability in problem difficulty: 22 problems were solved by at least 25 respondents, while 10 were solved by fewer than 10 test-takers. Notably, several problems easily solved by humans (e.g., 91, 92, 95, and 96) were not solved by any model, highlighting the need for further advances in this area.\\n\\nThese findings highlight significant gaps between human and model-based performance on this dataset, demonstrating the need for further development of MLLM AVR abilities.\\n\\n> Lacking an in-detailed analysis of the 100 samples in your datatset, e.g. the difficulty of each image pair.\\n\\nTo address this concern, we conducted an analysis based on participant feedback from our human study. Participants rated the overall difficulty of all problems on a scale from 1 to 10, resulting in an average difficulty of 7.6, indicating that the dataset is perceived as quite challenging. In addition, Fig. 8 illustrates variability in problem difficulty: 22 problems were solved by at least 25 respondents, while 10 were solved by fewer than 10. Notably, several problems easily solved by humans (e.g., 91, 92, 95, and 96) were not solved by any model, highlighting the need for further advances in this area.\"}", "{\"title\": \"Response to the Reviewer 9XwS\", \"comment\": \"> Some open-source models like InternVL2 and Phi-3.5V perform well on many benchmarks but struggle, nearing chance-level performance on Bongard Problems, while LLava-1.6 and Pixtral perform okay. What do the authors think causes these discrepancies?\\n\\nThe open models considered in our study differ in several key aspects, including the data mixtures used during pre-training, fine-tuning and learning from human feedback, the architectures of their text and vision encoders, and the approaches for combining multimodal embeddings. These variations make it challenging to pinpoint a single factor driving the observed differences in performance. Furthermore, the nature of BPs is quite distinct from tasks typically found in standard MLLM data sources. For example, training datasets often include tasks like visual question answering or image captioning, while analogy-based reasoning tasks like BPs receive little to no coverage, which may contribute to variability in model performance.\\n\\nNevertheless, both Pixtral and LLaVA-NeXT introduce important improvements in their vision encoders. In particular, Pixtral processes images at their original resolution and aspect ratio, offering greater flexibility in the number of tokens used for image processing. LLaVA-NeXT introduces the AnyRes algorithm, which supports high-resolution image processing by segmenting images into different grid configurations. We hypothesize that high-fidelity vision encoders are critical for MLLMs to excel in solving BPs.\\n\\n> How closely does abstract visual reasoning (AVR) reflect real-world visual reasoning (VR)? For instance, if a model excels in AVR, can we expect it to perform well in VR, and vice versa? Some potential correlations experiments between AVR and VR performance might give some insights.\\n\\nThank you for this thought-provoking question. The primary distinction between AVR and VR lies in task formulation. VR tasks are typically explicit and grounded in clear objectives, such as object detection or segmentation. For example, a common VR task in visual question answering (VQA) might involve an image of a room and a corresponding question such as \\u201cWhat stands next to the table on the left side?\\u201d. In contrast, AVR tasks focus on abstract analogies, which require identification of similarities and differences across images based on an underlying pattern expressed in a more conceptual and abstract manner.\\n\\nDespite these differences, AVR and VR share key similarities. AVR tasks fundamentally require reasoning across a set of images, which overlaps with certain VR benchmarks focused on multi-image reasoning. For instance, the MMMU [1] and MMIU [2] benchmarks contain tasks that test capabilities of reasoning across multiple images. From this perspective, datasets and strategies used in our work can serve as benchmarks to evaluate multi-image VR capabilities of MLLMs.\\n\\nOur experiments reveal an important relationship between AVR and VR performance. Proprietary MLLMs pre-trained on diverse datasets and fine-tuned on various VR tasks demonstrate relatively strong performance in our AVR experiments. However, good VR performance does not necessarily guarantee strong AVR capabilities, as evidenced by the results of open-access models. For instance, InternVL2-8B reported a relatively strong performance of 51.8% on MMMU [3], which is a good result even when compared to a much larger model such as Claude 3.5 Sonnet, which achieved 68.3% [4]. In contrast, InternVL2-8B performed significantly worse than Claude 3.5 Sonnet in our experiments (e.g. on Bongard HOI using the Descriptive strategy the models scored 2% and 44%, resp.). This discrepancy indicates that strong VR performance does not necessarily transfer to AVR problems.\\n\\nVice versa, excelling in AVR does not necessarily imply strong performance in VR. For instance, a supervised classifier trained specifically to solve synthetic BPs is unlikely to generalize to real-world scenarios. Assessing model\\u2019s applicability to VR tasks requires employing AVR datasets that span both synthetic and real-world domains, such as those considered in our work. A model that performs well across such diverse setup is more likely to demonstrate generalizable reasoning abilities applicable to the broader VR domain.\\n\\n[1] Yue, Xiang, et al. \\\"MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGI.\\\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024.\\n\\n[2] Meng, Fanqing, et al. \\\"MMIU: Multimodal Multi-image Understanding for Evaluating Large Vision-Language Models.\\\" arXiv preprint arXiv:2408.02718 (2024).\\n\\n[3] OpenGVLab. \\u201cInternVL2-8B Model card.\\u201d https://huggingface.co/OpenGVLab/InternVL2-8B Accessed: 2024-11-21.\\n\\n[4] Anthropic. \\u201cClaude 3.5 Sonnet.\\u201d https://www.anthropic.com/news/claude-3-5-sonnet Accessed: 2024-11-21.\"}", "{\"title\": \"Response to the Reviewer YYwM\", \"comment\": \"> What is your insight on these proposed generation strategies in Figure 2? (Why do you choose these tasks?)\\n\\nWe chose these strategies to systematically explore and evaluate the capabilities and limitations of MLLMs in solving BPs, drawing inspiration from: (1) specificity of the task, (2) MLLM training setups, and (3) human problem-solving approaches.\\n\\nWe began with the **Direct** strategy, the simplest and most intuitive approach, in which the model directly formulates an answer based on the provided image. It serves as a baseline to understand how MLLMs handle the task without any explicit intermediate reasoning steps. As discussed in Section 5 (\\u201cGenerative capabilities in the Direct setting\\u201d), models overally underperform in this setting, which raised the need for formulating task-specific strategies to explore the limits of model abstract reasoning performance.\\n\\nTo this end, we explored a task decomposition approach, where the reasoning process is broken down into smaller subtasks, whose solutions are later combined to solve the main task.\\n\\nThe **Descriptive** strategy focuses on generating descriptions of individual images. This approach was inspired by the abundant amount of (image, text) pairs found in MLLM training datasets, which we expected would enable models to generate accurate image descriptions that facilitate abstract reasoning. As shown in Section 5 (\\u201cIndependent image description\\u201d) the Descriptive strategy is generally preferred to other options, signifying the importance of selecting solution strategies that align with model training methods.\\n\\nConversely, the **Contrastive** strategy focuses on identifying differences between image pairs composed of images from the opposite sides of the problem. Inspired by psychological insights into human problem-solving, particularly the role of contrast in highlighting differences (Nussli et al., 2009), we wanted to explore if MLLMs can benefit from such human-based strategies. As discussed in Section 5 (\\u201cContrastive reasoning\\u201d), this strategy leads to worse results than Descriptive, highlighting clear differences in human and MLLM approaches to solving abstract reasoning tasks.\\n\\nNext, we considered **-iterative** variants of the **Descriptive** and **Contrastive** strategies. This design was motivated by the way humans iteratively refine their understanding as they incorporate new information. In addition, it compensates for the lack of memory in MLLMs by including past questions and responses within the reasoning context. As presented in Section 5 (\\u201cIterative reasoning\\u201d), the results indicate that contemporary models struggle to effectively leverage additional information from the context window, raising the need for improving multi-step reasoning capabilities of MLLMs.\\n\\nEncouraged by the relatively strong performance of task decomposition methods compared to the Direct strategy, we hypothesized that combining these approaches could further enhance the abstract reasoning capabilities of MLLMs. To this end, we introduced **-direct** variants of the **Descriptive** and **Contrastive** strategies, allowing to cross-check individual observations against the entire matrix at the final answer generation step. However, the experiments presented in Section 5 (\\u201cMultimodal answer generation\\u201d) indicate that MLLMs face challenges in fully utilizing this additional multi-modal context to improve their reasoning performance.\\n\\n> There are also some works from the observation in psychological experiments like MindSet [1], which could be the data sources for BP tasks or AVR tasks.\\n\\nThank you for bringing this relevant reference to our attention. We agree that evaluating MLLMs using the proposed strategies on the MindSet: Vision toolbox represents a valuable direction for future research. Such an investigation could provide deeper insights into object perception and concept identification in MLLMs, further clarifying their alignment with human perceptual reasoning. We have included a discussion of this idea in Appendix A under the paragraph \\u201cFine-grained analysis of MLLM perception\\u201d.\\n\\n> As your works show some similar observations to previous benchmark works like VCog-Bench, and Fran\\u00e7ois Chollet's book [...]\\n\\nThank you for pointing out the relevance of these works. We agree that they are closely related to our study and have already cited them in Section 2: *\\u201cRecent research concerning the evaluation of abstract reasoning skills of LLMs concentrates around the Abstraction and Reasoning Corpus (ARC) task (Chollet, 2019).\\u201d* and *\\u201cCao et al. (2024) proposed a suite of AVR tasks to compare MLLM and human performance.\\u201d* Given the breadth of our study, it was challenging to expand the discussion of these works within the main body of the paper. However, we have added a detailed discussion in the newly introduced paragraph \\u201cGoing Beyond Bongard Problems\\u201d in Appendix A to address these works more comprehensively.\"}", "{\"summary\": \"This paper focuses on exploring one limitation of MLLM in visual understanding: there is still a big gap between MLLMs and humans in the IQ test / AVR test from psychometrics. They choose Bongard Problems to conduct the case study. Firstly, they build a new benchmark with some Bongard Problem samples. Then, they compare some SOTA MLLMs and VLMs with this multi-image reasoning task. The results suggest that the weak performance of MLLMs on Bongard Problems is not due to the domain specificity, but rather comes from their multi-image reasoning capability.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. This is an important problem for MLLM nowadays. I like the motivation of your paper. As we know, Human IQ test has a long history in psychometrics and only a few of works try to explore the different between MLLM and human in this topic [1].\\n\\n2. They have open-sourced their dataset containing 100 samples. I have a quick review of the dataset repo. The data structures and descriptions are easy to follow. \\n\\n3. Compare with most of visual reasoning benchmark papers, this paper use less data but it has a solid cognition background. I happen to agree with some viewpoints of the authors, which are also highlighted by Google DeepMind recently.\", \"references\": \"[1] Galatzer-Levy, Isaac R., et al. \\\"The Cognitive Capabilities of Generative AI: A Comparative Analysis with Human Benchmarks.\\\" arXiv preprint arXiv:2410.07391 (2024).\", \"weaknesses\": \"1. It is worth to highlight that BP is only a subtask of AVR test. Compared to RAVEN, BP is not widely used by human's IQ test recently. I think you should discuss this limitation and consider exploring a board scope of AVR tasks in the future work.\\n\\n2. The Evaluation settings in Figure 3 is unclear. It seems that most of the settings proposed by you can only use to test the closed-sourced models like Claude 3.5 and GPT-4o. The multi-image reasoning capbility of most of open-sourced MLLMs are not eligible in these tasks.\\n\\n3. For a better understanding, the user study between human and MLLM is also important for this task.\\n\\n4. Lacking an in-detailed analysis of the 100 samples in your datatset, e.g. the difficulty of each image pair.\", \"questions\": \"1. What is your insight on these proposed generation strategies in Figure 2? (Why do you choose these tasks?)\\n\\n2. There are also some works from the observation in psychological experiments like MindSet [1], which could be the data sources for BP tasks or AVR tasks.\\n\\n3. As your works show some similar observations to previous benchmark works like VCog-Bench, and Fran\\u00e7ois Chollet's book \\\"On the measure of intelligence, you can put some further discussions in the main body of your paper.\", \"post_rebuttal\": \"I take a quick view of the authors' reply and other reviews' score, I decide to decrease my final score to 4. Though the authors choose a good topic as motivations, the main pain points (over claiming, lacking data, lacking detailed definition) are difficult to solve. This, we all agree this paper doesn't match the bar of ICLR.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to the Reviewer VpcT\", \"comment\": \"> Can you provide the rationale behind the proposed strategies in the experiments? Does it relate to the theory of AVR from a cognitive science perspective? Can you summarize the high-level idea of the design?\\n\\nWe chose these strategies to systematically explore and evaluate the capabilities and limitations of MLLMs in solving BPs, drawing inspiration from: (1) specificity of the task, (2) MLLM training setups, and (3) human problem-solving approaches.\\n\\nWe began with the **Direct** strategy, the simplest and most intuitive approach, in which the model directly formulates an answer based on the provided image. It serves as a baseline to understand how MLLMs handle the task without any explicit intermediate reasoning steps. As discussed in Section 5 (\\u201cGenerative capabilities in the Direct setting\\u201d), models overally underperform in this setting, which raised the need for formulating task-specific strategies to explore the limits of model abstract reasoning performance.\\n\\nTo this end, we explored a task decomposition approach, where the reasoning process is broken down into smaller subtasks, whose solutions are later combined to solve the main task.\\n\\nThe **Descriptive** strategy focuses on generating descriptions of individual images. This approach was inspired by the abundant amount of (image, text) pairs found in MLLM training datasets, which we expected would enable models to generate accurate image descriptions that facilitate abstract reasoning. As shown in Section 5 (\\u201cIndependent image description\\u201d) the Descriptive strategy is generally preferred to other options, signifying the importance of selecting solution strategies that align with model training methods.\\n\\nConversely, the **Contrastive** strategy focuses on identifying differences between image pairs composed of images from the opposite sides of the problem. Inspired by psychological insights into human problem-solving, particularly the role of contrast in highlighting differences (Nussli et al., 2009), we wanted to explore if MLLMs can benefit from such human-based strategies. As discussed in Section 5 (\\u201cContrastive reasoning\\u201d), this strategy leads to worse results than Descriptive, highlighting clear differences in human and MLLM approaches to solving abstract reasoning tasks.\\n\\nNext, we considered **-iterative** variants of the **Descriptive** and **Contrastive** strategies. This design was motivated by the way humans iteratively refine their understanding as they incorporate new information. In addition, it compensates for the lack of memory in MLLMs by including past questions and responses within the reasoning context. As presented in Section 5 (\\u201cIterative reasoning\\u201d), the results indicate that contemporary models struggle to effectively leverage additional information from the context window, raising the need for improving multi-step reasoning capabilities of MLLMs.\\n\\nEncouraged by the relatively strong performance of task decomposition methods compared to the Direct strategy, we hypothesized that combining these approaches could further enhance the abstract reasoning capabilities of MLLMs. To this end, we introduced **-direct** variants of the **Descriptive** and **Contrastive** strategies, allowing to cross-check individual observations against the entire matrix at the final answer generation step. However, the experiments presented in Section 5 (\\u201cMultimodal answer generation\\u201d) indicate that MLLMs face challenges in fully utilizing this additional multi-modal context to improve their reasoning performance.\"}", "{\"title\": \"Response to Reviewer YYwM regarding further concerns (1 / 2)\", \"comment\": \"Dear Reviewer YYwM, thank you for reviewing the revised paper and our response to the Reviewers.\\n\\n> Human study\\n\\nWe fully agree that human studies focused on understanding human cognition and psychometrics necessitate additional methodological rigor and clear selection criteria. However, given the limited time available during the rebuttal period, it was not feasible to conduct a more comprehensive study on a larger scale. It is worth noting that while the details of the human study on Bongard-OpenWorld are not reported, Bongard-HOI employed 35 participants, which is comparable to the scale of our study conducted within this constrained timeline.\\n\\nConsecutively, our human study was not intended to analyze human cognition in depth or investigate the specific strategies humans employ when solving BPs. Instead, our primary objective was to provide a reference point for human performance, contextualizing the results of MLLMs and demonstrating usability and solvability of the dataset. The average human accuracy of 65% (with the best individual performance reaching 98.3%) significantly exceeds the best performance achieved by MLLMs, where the highest result (Claude 3.5 Sonnet using the Descriptive strategy) was 22%, solving only 13 out of 60 problems. This comparison shows a substantial gap between human and MLLM performance, highlighting the challenges posed by Bongard-RWR. While we acknowledge that our study is insufficient to draw detailed insights regarding human cognition, we believe it effectively serves its intended purpose of establishing a reference for assessing MLLM results.\\n\\nOur observation regarding human approaches to solving BPs is grounded in prior research, see \\u201cContrastive reasoning\\u201d in Section 5: *\\u201cHumans often rely on direct comparisons between image panels from different categories to highlight differences __(Nussli et al., 2009)__\\u201d*. Investigating the specific cognitive processes or strategies humans use to solve Bongard-RWR was beyond the scope of our work. We recognize this as an interesting direction for future research and appreciate the Reviewer\\u2019s suggestion to approach this in greater depth.\\n\\n> Insights from Your Rebuttal\\n\\nThank you for bringing this highly relevant work to our attention. We were unaware of (Kian et al., 2024) during the paper submission, as it was accepted for publication on 10th July, 2024 ([OpenReview link](https://openreview.net/forum?id=eDWcNqiQWW)). According to the ICLR FAQ, such works are considered concurrent to ICLR submissions. We will thoroughly investigate this paper and incorporate its findings in our future works.\\n\\nWhile we do not explore all the suggested topics in the same depth, our experiments already align with some of the methods highlighted in (Kian et al., 2024). For example, we utilize few-shot learning in Prompts 6 and 7, where in-context examples of expected solutions are provided. In addition, the Descriptive and Contrastive strategies, together with their -direct and -iterative variants, can be considered as implementations of chain-of-thought reasoning, where partial descriptions of BP images form subsequent observations summarized at the final answer generation step. While we do not explicitly perform model training or RLHF fine-tuning, we recognize the value of these approaches and agree they form promising directions for future exploration.\\n\\nNevertheless, in this paper we have aimed to provide a broad contribution by considering 3 datasets from the literature, introducing a new dataset, evaluating 8 models (both proprietary and open-access), exploring 7 natural language generation strategies and 3 binary classification tasks. Moreover, we conducted a scaling law experiment and a human study to contextualize MLLM performance. While we acknowledge that incorporating suggested experiments would provide additional insights, we hope the breadth and depth of our current contributions may be considered valuable on their own.\"}", "{\"title\": \"Response and further concerns\", \"comment\": \"Dear Authors of ICLR 2025 Submission #6466,\\n\\nThank you for your response and for updating your manuscript. I have carefully reviewed your revised version and the replies from the other reviewers. I would like to raise some new concerns, and I am considering adjusting my score based on these points.\\n\\n**Human Study**\\n\\nI am happy you follow Reviewer VQyh's and my suggestions to conduct the human study, but the current version is not satisfied.\\n\\nThe methodology of your cognition and psychometrics human study raises several concerns regarding its rigor. For a robust human study, it is essential to first obtain IRB approval from your institution. Additionally, participants should be recruited based on clear criteria (e.g., data selection process, age range, gender distribution, IQ levels). Moreover, your results seem to contradict your previous statements. You mention that \\\"Humans often rely on direct comparisons between image panels from different categories to highlight differences, whereas the tested methods perform better when making comparisons on text-based image descriptions...\\\" However, your human study shows that participants could solve only 65% of the problems in your dataset, and it remains unclear how humans approach solving the Bongard problems in your dataset.\\n\\nYou have not provided evidence to demonstrate whether the suboptimal human performance is due to your data design or the inherent difficulty of the problems. This issue appears to be more related to HCI and psychometrics, and it necessitates a rigorously designed experiment to investigate it thoroughly.\\n\\n**Insights from Your Rebuttal**\\n\\nI have concerns regarding the insights you proposed in your rebuttal. You state that MLLMs still face challenges even after applying all your improved strategies. Identifying challenges faced by MLLMs is not novel in itself. For instance, Kian's paper, \\\"The Curious Case of Nonverbal Abstract Reasoning with Multi-modal Large Language Models,\\\" explores similar issues, and there is significant overlap with your insights.\\n\\nTo strengthen your statement, I suggest re-evaluating the essence of solving the Bongard Problem based on Kian's findings, including adding experiments for few-shot learning, chain-of-thought, training-testing experiments, RLHF finetuning. This could provide a deeper understanding and highlight the unique contributions of your work.\\n\\n**Comparison with Existing Datasets**\\n\\nReviewer VpcT raised an important question that I had not initially noticed. You state, \\\"Note, however, that a direct performance comparison on synthetic Bongard datasets versus real-world Bongard HOI and Bongard-OpenWorld datasets is not meaningful, as these datasets depict different concepts.\\\" This prompted me to examine the data from Bongard HOI and Bongard-OpenWorld.\\n\\nFrom my perspective, there does not seem to be a significant difference between their datasets and yours. Your logic appears to be that because MLLMs perform well on Bongard HOI and Bongard-OpenWorld, these datasets are less meaningful, whereas the poor performance on your dataset proves yours value. This reasoning is not rigorous and raises several concerns. For example, readers might suspect that your data were selectively chosen to highlight the weaknesses of MLLMs, potentially contradicting the definition of Bongard Problems, which could undermine the validity of your conclusions. \\n\\nI recommend that you address the concerns raised by Reviewer VpcT thoroughly. In my view, this is a core issue with your paper, and your current response does not adequately resolve it. I am open to further discuss these problems with you, \\u00a0Reviewer VQyh, and Reviewer VpcT during the rebuttal. Let's wait for other reviewers' reply.\\n \\u00a0 \\u00a0 \\n \\nReviewer YYwM\"}", "{\"title\": \"Response to the Reviewer VpcT\", \"comment\": \"> The contribution of the proposed Bongard-RWR dataset is not clear enough, [...]\\n\\nIn preliminary experiments, we observed that while MLLMs perform well on Bongard HOI and Bongard-OpenWorld, they struggle with synthetic BPs. For instance, compare the best results in Images to Sides in Fig. 4 on synthetic BPs (75) to Bongard HOI (99) and Bongard-OpenWorld (96), or the max score in Table 1 on synthetic BPs (21) to Bongard HOI (45) and Bongard-OpenWorld (57). This discrepancy prompted us to investigate the reasons behind these differences.\\n\\nWe identified two key factors that distinguish synthetic BPs from Bongard HOI and Bongard-OpenWorld. **Firstly, synthetic BPs feature simple 2D geometric shapes, while Bongard HOI and Bongard-OpenWorld consist of real-world images.** A potential explanation might be that the lack of training data on synthetic images could contribute to the lower performance on synthetic BPs. **Secondly, synthetic BPs focus on abstract concepts (e.g., \\u201cOne line vs. Two lines\\u201d in Fig. 1a), while Bongard HOI and Bongard-OpenWorld involve concepts grounded in the real world (e.g., \\u201cA person jumping on a surfboard vs. Not a person jumping on a surfboard\\u201d in Fig. 1c).** This led us to hypothesize that MLLMs may find abstract concepts harder to recognize, especially when instantiated in various forms (see \\u201c$D^X_{i1}, D^X_{i2}, D^X_{i3}$\\u201d in Section 4.1).\\n\\nThis motivated us to create the Bongard-RWR dataset, which bridges the gap between these two domains by representing abstract concepts with real-world images. As shown in Section 5, Bongard-RWR presents a major challenge to contemporary MLLMs, even though it relies on real-world images. This indicates that abstract concepts are inherently more difficult for models to recognize than concepts grounded in the real world. We believe that analyzing model performance across synthetic and real-world concepts opens a promising avenue for future work to fully explore this hypothesis. We have outlined ideas for extending this line of research in the newly added paragraph \\u201cCross-domain analysis of MLLM perception\\u201d in Appendix A.\\n\\n> While the paper evaluates current MLLMs, it fails to offer a clear direction for enhancing their AVR capabilities or to present any new insights beyond highlighting general limitations. [...]\\n\\nThank you for this comment. To better capture the current model limitations, we have restructured the presentation of results in Section 5. In particular, Fig. 4 (top) highlights a potential bias in MLLMs agreeing or disagreeing with presented concepts, which emphasizes the importance of considering natural language generation setups for evaluating MLLM abstract reasoning capabilities. Fig. 4 (bottom) illustrates that while MLLMs perform well in tasks involving real-world concepts (e.g., Bongard HOI and Bongard-OpenWorld), they struggle with abstract concepts (e.g., synthetic BPs and Bongard-RWR). We further show that MLLMs do not benefit from human-like approaches to solving BPs (see \\u201cContrastive reasoning\\u201d), struggle to effectively utilize the context window (see \\u201cIterative reasoning\\u201d and Descriptive-iterative in Fig. 5), and require further work to consistently integrate text and vision modalities at the answer generation step (see \\u201cMultimodal answer generation\\u201d and Descriptive-direct in Fig. 5).\\n\\nWe believe that these insights provide clear recommendations for utilizing contemporary MLLMs to solve AVR tasks, and more generally, multi-image reasoning tasks. Specifically, the **Descriptive** strategy leads to the strongest performance, particularly when combined with its **-direct** variant in models that present stronger capabilities in integrating text and vision modalities, such as GPT-4o.\\n\\nAnother direction for enhancing AVR capabilities emerges from the scaling law experiment discussed below in response to another comment. In summary, these results indicate that simply scaling model size may be insufficient to achieve stronger abstract reasoning capabilities. Future efforts are required to explicitly address this aspect, e.g., by incorporating AVR datasets into model training.\\n\\nIn the revision we discuss additional ideas for future work in the newly introduced Appendix A.\"}", "{\"title\": \"Response to Reviewer YYwM\", \"comment\": \"First of all, thank you for your thorough review of our paper and pointing out important directions to explore in the future work.\\n\\n> Specifically, the prior datasets already represent a significant contribution to this area, while your dataset appears to be a minor augmentation (like Reviewer VpcT proposed). However, the way your paper is written gives the impression that your contribution is overstated. Reviewer VpcT observed this problem and summaried into his/her/their comment.\\n\\nWe acknowledge the contributions of Bongard HOI and Bongard-OpenWorld and aimed to make it clear in the paper that Bongard-RWR connects synthetic BPs with real-world BPs introduced in these two datasets, instead of trying to replace them. In particular, we motivate and summarize the dataset\\u2019s contribution as: *\\\"To further examine the main difficulties faced by MLLMs in solving both types of BPs (synthetic and real world ones) we introduce a focused dataset of BPs (Bongard-RWR) comprising real-world images that represent concepts from synthetic BPs using real world images. Thanks to relying on the same abstract concepts as synthetic BPs, Bongard-RWR facilitates direct comparisons of the MLLMs performance in both domains.\\\"* and *\\\"To delve deeper into the performance discrepancies between synthetic and real-world domains, we introduced Bongard- RWR, a new BP dataset designed to represent concepts from synthetic BPs via real-world images.\\\"*\\n\\nWe also think that Bongard-RWR is not merely an augmentation of these two datasets, as we use a fundamentally different methodology to construct the dataset. While BPs in Bongard HOI and Bongard-OpenWorld are sampled automatically from a larger image dataset to represent concepts grounded in the real-world, we construct each BP manually, or in certain cases rely on a semi-automated process followed by manual review of the obtained matrices. Most importantly, Bongard-RWR focuses on concepts not covered in Bongard HOI and Bongard-OpenWorld, making it orthogonal to these prior works.\\n\\nFinally, we emphasize that Bongard-RWR is only one aspect of the broader contributions presented in our study, as detailed in our previous response: *\\\"[...] in this paper we have aimed to provide a broad contribution [...]\\\"*. We hope that the overall breadth and depth of our paper constitutes a valuable study on the reasoning limitations of MLLMs.\\n\\n> readers might question whether your dataset was manually selected to emphasize the weaknesses of MLLMs, particularly in terms of the data are sampling from conditional probability distributions. This could contradict the foundational principles of Bongard problems and potentially undermine the validity of your conclusions. \\n\\nWe appreciate the concern and want to clarify that our dataset was not manually curated to emphasize the weaknesses of MLLMs. To construct matrices in Bongard-RWR, we focused on selecting diverse images that accurately reflect concepts used in synthetic BPs via real-world images. We did not take into account model, nor human, performance in solving these tasks. In fact, the original synthetic BPs were likewise designed manually and the set of BPs was further extended by individual contributors, which follows the approach taken to design Bongard-RWR.\\n\\n> To effectively explore MLLM performance and facilitate meaningful comparisons between human and MLLM capabilities, it would be more appropriate to use a widely collected dataset in general domain (e.g., sampling natural and symbolic images in Bongard HOI and Bongard-OpenWorld, and other datasets) for Bongard problems and evaluate the models within that distribution. This contradicts your current conclusion.\\n\\nThe primary goal of our research was not to compare human and MLLM capabilities. Instead, we introduced a new dataset to investigate why MLLMs perform poorly on synthetic BPs compared to Bongard-HOI and Bongard-OpenWorld. A study focused solely on Bongard-HOI and Bongard-OpenWorld would be inconclusive, as these datasets do not capture the same concepts as synthetic BPs. The human study was conducted solely to serve as a general baseline. However, we agree that the robustness of the presented results could be further supported by a larger sample size. In the future work we will consider constructing additional instantiations of the selected matrices to ensure model performance is not heavily influenced by the individual selected images.\"}", "{\"title\": \"Response to Reviewer YYwM regarding further concerns (2 / 2)\", \"comment\": \"> Comparison with Existing Datasets\\n\\nWe respectfully disagree with the notion that we consider MLLM results on Bongard HOI and Bongard-OpenWorld to be less meaningful. On the contrary, we believe these datasets play an essential role in evaluating model performance on concepts grounded in the real-world. The contributions of Bongard HOI and Bongard-OpenWorld are complementary to our work.\\n\\nThe development of Bongard-RWR was motivated by observed discrepancies in model performance between synthetic BPs and real-world datasets including Bongard HOI and Bongard-OpenWorld, as outlined in our response to the Reviewer VpcT: *\\u201cIn preliminary experiments, [\\u2026] This discrepancy prompted us to investigate the reasons behind these differences.\\u201d* Through this investigation, we identified two main ideas differentiating synthetic BPs from Bongard HOI and Bongard-OpenWorld, as further explained in the mentioned response: *\\u201cWe identified two key factors that distinguish synthetic BPs from Bongard HOI and Bongard-OpenWorld. [...]\\u201d*.\\n\\nBongard-RWR was specifically designed to fill this gap, by introducing BPs comprising real-world images presenting abstract concepts. It is not intended to serve as a more meaningful dataset but rather to complement synthetic BPs, Bongard HOI and Bongard-OpenWorld.\\n\\n> I recommend that you address the concerns raised by Reviewer VpcT thoroughly.\\n\\nIn our response to Reviewer VpcT, we provided a detailed explanation of the motivation behind Bongard-RWR, outlined key insights and recommendations from our experiments, extended the dataset statistics with data from the human study, explained the rationale for the introduced strategies, analyzed the influence of model scaling on abstract reasoning performance, and expanded the paper\\u2019s conclusions. If there are any specific aspects of our response that require further clarification, we would be happy to address them in more detail.\"}" ] }

[ "Sheng Fang", "Yongjin Liu", "Wei Yao", "Chengming Yu", "Jin Zhang" ]

[ "bilevel optimization", "quasi-Newton", "convergence analysis", "Hessian-free" ]

{ "note_id": [ "x3Pn9Clpjo", "uKqiozJ5aO", "sId5h8uXLM", "rnqqo59xjk", "qytRhPA4u6", "mXDbAFZnlt", "lBWgo4ZYhX", "jRV0fxyLyN", "hLKvnqaeTi", "hIYBU7lgnn", "ZEwwCDPJ3I", "YjorkLiSaF", "WLsXCLchir", "VBgKULeNlL", "Ur7BN8C4tM", "SmtsX5yJcx", "SVVO07fM6F", "S5MIUjCNjk", "QlGfuJCnU3", "QZijiQX7oP", "QWtoG0Canl", "Q5VymJZYyU", "O7pk9hmQUE", "Ii0UlJKirr", "Id7cpWFoWh", "HgyBilVGM5", "EVqifIZyFL", "E5hWmRHECt", "DsLaQOL2nC", "7K2Ngy6LUG", "5kqVGfpOYe", "5fsYeuoluY", "1QhQzfV5YP", "0iYLJomgeD" ], "note_type": [ "official_comment", "official_review", "official_comment", "official_review", "meta_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "decision", "official_comment", "official_comment" ], "note_created": [ 1732033612993, 1730264207563, 1732033548620, 1730248338541, 1734688330107, 1732521149194, 1733307932043, 1732033459328, 1733308067447, 1733308137323, 1730400739554, 1730382279598, 1732524983462, 1732518746047, 1732716671185, 1732203081196, 1730696017989, 1732519616173, 1732544449295, 1732515640614, 1732033385340, 1732033673906, 1732791624027, 1732033727480, 1732519679023, 1733021796292, 1732544184117, 1732518318383, 1732515934577, 1732525069620, 1732544392198, 1737523841654, 1733308002964, 1732202840174 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission7492/Authors" ], [ "ICLR.cc/2025/Conference/Submission7492/Reviewer_ozni" ], [ "ICLR.cc/2025/Conference/Submission7492/Authors" ], [ "ICLR.cc/2025/Conference/Submission7492/Reviewer_XUKY" ], [ "ICLR.cc/2025/Conference/Submission7492/Area_Chair_zoCa" ], [ "ICLR.cc/2025/Conference/Submission7492/Authors" ], [ "ICLR.cc/2025/Conference/Submission7492/Authors" ], [ "ICLR.cc/2025/Conference/Submission7492/Authors" ], [ "ICLR.cc/2025/Conference/Submission7492/Authors" ], [ "ICLR.cc/2025/Conference/Submission7492/Authors" ], [ "ICLR.cc/2025/Conference/Submission7492/Reviewer_5z6s" ], [ "ICLR.cc/2025/Conference/Submission7492/Reviewer_yCCW" ], [ "ICLR.cc/2025/Conference/Submission7492/Authors" ], [ "ICLR.cc/2025/Conference/Submission7492/Authors" ], [ "ICLR.cc/2025/Conference/Submission7492/Reviewer_ozni" ], [ "ICLR.cc/2025/Conference/Submission7492/Reviewer_yCCW" ], [ "ICLR.cc/2025/Conference/Submission7492/Reviewer_E1d6" ], [ "ICLR.cc/2025/Conference/Submission7492/Authors" ], [ "ICLR.cc/2025/Conference/Submission7492/Authors" ], [ "ICLR.cc/2025/Conference/Submission7492/Authors" ], [ "ICLR.cc/2025/Conference/Submission7492/Authors" ], [ "ICLR.cc/2025/Conference/Submission7492/Authors" ], [ "ICLR.cc/2025/Conference/Submission7492/Reviewer_yCCW" ], [ "ICLR.cc/2025/Conference/Submission7492/Authors" ], [ "ICLR.cc/2025/Conference/Submission7492/Authors" ], [ "ICLR.cc/2025/Conference/Submission7492/Reviewer_5z6s" ], [ "ICLR.cc/2025/Conference/Submission7492/Authors" ], [ "ICLR.cc/2025/Conference/Submission7492/Authors" ], [ "ICLR.cc/2025/Conference/Submission7492/Authors" ], [ "ICLR.cc/2025/Conference/Submission7492/Authors" ], [ "ICLR.cc/2025/Conference/Submission7492/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission7492/Authors" ], [ "ICLR.cc/2025/Conference/Submission7492/Reviewer_yCCW" ] ], "structured_content_str": [ "{\"title\": \"Discussion on the impact of Reviewer yCCW\\u2019s debug in the toy experiment (Part 3/3).\", \"comment\": \"(2)Since the implementations of AID-TN, AID-BIO, and AMIGO-CG in the above experiments used the closed-form expression $f_{yy} = A$ for the Hessian, $\\\\textbf{it would not be}$ $\\\\textbf{ fair to compare them with other algorithms that do not use the explicit Hessian information.}$ Therefore, $\\\\textbf{we also conducted the toy experiment}$ $\\\\textbf{using the more practical torch.autograd.grad for computing Hessian-vector products.}$ Note that this approach is typical in the machine learning literature, as the closed form of the Hessian is generally difficult to compute in practical applications. After performing a grid search to fine-tune the hyperparameters for step sizes tested across [0.01, 0.02, 0.03, 0.04, 0.05, 0.06, 0.07, 0.08, 0.09, 0.1, 0.2, 0.3, 0.4, 0.5], TN/CG iterations evaluated at [1, 10, 20, 30, 40, 50], and inner gradient iterations between 1 and 10, the new results\\u2014using torch.autograd.grad for computing the Hessian-vector product instead of the closed Hessian form\\u2014are presented in Figure 2 of the single-page PDF at the anonymous link and in the following tables.\\n\\n\\n\\n| Time (s) for \\\\|x - x^*\\\\| / \\\\|x^*\\\\| | 0.5 | 1.0 | 1.5 |\\n|-----------------------------------------|------------|------------|------------|\\n| **qNBO(SR1)** | 1.0E-06 | 5.8E-07 | 1.0E-06 |\\n| **qNBO(BFGS)** | 5.0E-03 | 2.7E-04 | 4.5E-05 |\\n| **BOME** | 3.5E-01 | 3.3E-01 | 3.2E-01 |\\n| **SHINE-OPA** | 6.9E-02 | 5.5E-02 | 6.9E-02 |\\n| **AID-BIO** | 1.4E+00 | 7.1E-01 | 4.0E-01 |\\n| **AMIGO-CG** | 1.4E+00 | 7.1E-01 | 4.0E-01 |\\n| **PZOBO** | 3.4E+00 | 3.7E+00 | 4.1E+00 |\\n| **F2SA** | 3.5E-01 | 3.3E-01 | 3.2E-01 |\\n| **AID-TN** | 1.9E+00 | 1.0E+00 | 5.2E-01 |\\n|\\n\\n\\n| Time (s) for \\\\|y - y^*\\\\| / \\\\|y^*\\\\| | **0.5** | **1.0** | **1.5** |\\n|-------------------|------------|------------|------------|\\n| **qNBO(SR1)** | 1.0E-06 | 4.8E-07 | 2.3E-06 |\\n| **qNBO(BFGS)** | 7.9E-03 | 3.0E-04 | 6.3E-05 |\\n| **BOME** | 2.7E-01 | 2.7E-01 | 2.7E-01 |\\n| **SHINE-OPA** | 1.0E-01 | 4.8E-02 | 4.8E-02 |\\n| **AID-BIO** | 4.1E+00 | 3.3E+00 | 2.7E+00 |\\n| **AMIGO-CG** | 4.1E+00 | 3.3E+00 | 2.7E+00 |\\n| **PZOBO** | 4.1E+00 | 4.2E+00 | 4.5E+00 |\\n| **F2SA** | 5.2E-01 | 4.8E-01 | 4.4E-01 |\\n| **AID-TN** | 2.5E+00 | 1.3E+00 | 6.7E-01 |\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n\\n|Time (s) for $\\\\|d_x - \\\\nabla \\\\Phi\\\\|$ | 0.5 | 1.0 | 1.5 |\\n|-------------------|------------|------------|------------|\\n| **qNBO(SR1)** | 8.6E-05 | 2.3E-05 | 7.0E-05 |\\n| **qNBO(BFGS)** | 4.3E-02 | 2.2E-02 | 1.6E-02 |\\n| **BOME** | - | - | - |\\n| **SHINE-OPA** | 4.3E+00 | 1.7E+00 | 1.9E+00 |\\n| **AID-BIO** | 5.6E+01 | 7.8E+01 | 6.9E+01 |\\n| **AMIGO-CG** | 5.6E+01 | 7.8E+01 | 8.9E+01 |\\n| **PZOBO** | 2.2E+02 | 2.2E+02 | 2.4E+02 |\\n| **F2SA** | 1.3E+01 | 1.1E+01 | 1.1E+01 |\\n| **AID-TN** | 3.0E+01 | 1.5E+01 | 9.1E+00 |\", \"these_results_show_that\": \"(i) for AID-BIO and AMIGO-CG, 50 CG iterations yielded the best results, while a single CG iteration caused divergence, resulting in much worse performance compared to the scenarios using the closed-form Hessian; (ii) compared to other algorithms, both qNBO (BFGS) and qNBO (SR1) exhibit smaller hypergradient errors and produce iterates that are closer to the optimal solutions. $\\\\textbf{Therefore, after implementing the debugging suggestions from Reviewer yCCW and using the more practical torch.autograd.grad for}$\\n\\n$\\\\textbf{Hessian-vector products, the results for the toy experiment align with the conclusions of this work.}$\\n\\nTo facilitate the reproduction of the results, the specific hyperparameter settings are as follows:\\n\\n$\\\\textbf{AID-BIO/AMIGO-CG:}$ The maximum number of outer iterations is $K = 5000$, CG iterations is $N = 50$, the inner gradient iteration is $T = 1$, the inner step size is $\\\\beta = 0.01$, and the outer step size is $\\\\alpha = 0.01$.\\n\\n$\\\\textbf{AID-TN:}$ The maximum number of outer iterations is $K = 5000$, TN iterations is $N = 50$, the inner gradient iteration is $T = 1$, the inner step size is $\\\\beta = 0.01$, and the outer step size is $\\\\alpha = 0.2$.\\n\\nThe hyperparameters of qNBO (SR1) are the same as those used in the above experiment, and the settings for the other algorithms remain unchanged from those in the paper.\"}", "{\"summary\": \"This work studies the bilevel optimization problem, where the lower-level problem is strongly convex and the upper-level objective function is nonconvex. The authors introduce a novel framework called qNBO, which approximates the hypergradient by employing quasi-Newton methods for both evaluating the lower-level solution and estimating the inverse Hessian-vector product. When the quasi-Newton method used is BFGS, they establish a non-asymptotic convergence for the proposed algorithm. Finally, empirical experiments demonstrate the superior performance of qNBO compared to other bilevel optimization algorithms.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1.The new bilevel algorithms are well motivated. By employing quasi-Newton recursion schemes, they incorporate two subroutines to accelerate hypergradient approximation and avoid incorrect inversion in solving the bilevel optimization problem.\\n\\n2.Several machine learning tasks, including hyperparameter optimization in logistic regression, data hyper-cleaning, and few-shot meta-learning, are evaluated, demonstrating the superior performance of the proposed methods. Additionally, various ablation studies are provided.\\n\\n3.A convergence rate guarantee is established when the quasi-Newton method used is BFGS.\", \"weaknesses\": \"My main concern is that:\\n\\nAlthough quasi-Newton methods are used to estimate the inverse Hessian-vector product, the proposed algorithms still require computing Jacobian-vector products, which can be computationally expensive in large-scale cases. Could this computation pose a potential problem?\", \"questions\": \"For both the theory and experiments, here are a few things that need to be addressed.\\n\\n1.In Remark 3.4, the authors stated that the qNBO (SR1) algorithm lacks convergence guarantees without specific corrections used to achieve numerical stability in the general case. It was also mentioned that the performance of qNBO (SR1) on the 20News dataset was omitted due to its ineffectiveness in the experiment in Section 4.2. What are the underlying reasons for these issues? A bit more discussion on these points would be helpful.\\n\\n2.From the first terms on the right hand sides of both (6) and (7), it seems that Theorems 3.3 and 3.6 require a lower bound assumption for $\\\\Phi(x)$. Is this correct? Additionally, in the proof sketch in Appendix D.2, $\\\\tilde{\\\\nabla}\\\\Phi$ is mentioned but not defined. It would be helpful to specify its definition. \\n\\n3.In the experiments, how to select the step sizes and the number of inner iterations for the proposed algorithms?\\n\\n4.In Table 1, why does qNBO (BFGS) save so much time compared to SHINE for data hyper-cleaning?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Discussion on the impact of Reviewer yCCW\\u2019s debug in the toy experiment (Part 2/3).\", \"comment\": \"To address these issues, we have re-conducted the toy experiment, as detailed below:\\n\\n(1)$\\\\textbf{By retaining the closed-form expression $f_{yy} = A$ of the Hessian in the implementations of AID-TN, AID-BIO, and AMIGO-CG}$, we performed a grid search to fine-tune the hyperparameters of all algorithms, including inner loop iterations and step sizes. As noted by Reviewer yCCW, AID-BIO and AMIGO-CG achieve a relative error of $10^{-5}$ on the upper-level variable (compared to the values reported in the paper: $10^{-2}$ for AID-BIO and $10^0$ for AMIGO-CG) after setting the inner loop iterations for CG steps and gradient steps to 1 while maintaining the outer step size at 0.01.\\n\\nBy further conducting a grid search over TN/CG iterations in the range [1, 5, 10, 15, 20], inner gradient iterations between 1 and 10, and step sizes within [0.1, 0.2, 0.3, 0.4, 0.5], $\\\\textbf{the relative error of AID-BIO and AMIGO-CG can be reduced to below $10^{-6}$ within 1.5 seconds when the outer step size is adjusted to $\\\\alpha = 0.2$}$. At the same time, after fine-tuning the hyperparameters of qNBO (SR1), $\\\\textbf{the new implementation of qNBO (SR1) can also reduce the error to below $10^{-6}$ within 1.5 seconds}$. \\n\\nThe updated results are presented in Figure 1 of a single-page PDF (containing only figures and available at the anonymous link: https://drive.google.com/file/d/15x5Fp1XYlo1DKrKs4RcquXU6khqv9Y1l/view?usp=drive_link) and in the following tables.\\n\\n| Time (s) for \\\\|x - x^*\\\\| / \\\\|x^*\\\\| | 0.5 | 1.0 | 1.5 |\\n|----------------------------------------|----------|----------|----------|\\n| **qNBO(SR1)** | 1.8E-6 | 4.0E-7 | 4.3E-7 |\\n| **qNBO(BFGS)** | 5.6E-3 | 2.8E-4 | 4.0E-5 |\\n| **AID-BIO** | 9.2E-7 | 7.1E-7 | 7.1E-7 |\\n| **AMIGO-CG** | 1.6E-5 | 7.1E-7 | 7.1E-7 |\\n| **AID-TN** | 2.4E-1 | 2.4E-1 | 2.4E-1 |2.4E-1 |\\n\\n\\n| Time (s) for \\\\|y - y^*\\\\| / \\\\|y^*\\\\| | 0.5 | 1.0 | 1.5 |\\n|----------------------------------------|-----------|-----------|-----------|\\n| **qNBO(SR1)** | 1.7E-06 | 7.2E-07 | 8.5E-07 |\\n| **qNBO(BFGS)** | 7.9E-03 | 3.0E-04 | 6.3E-05 |\\n| **AID-BIO** | 1.7E-06 | 1.1E-06 | 1.1E-06 |\\n| **AMIGO-CG** | 2.7E-05 | 1.2E-06 | 1.2E-06 |\\n| **AID-TN** | 3.1E-01 | 3.1E-01 | 3.1E-01 |\\n\\n| Time (s) for $\\\\|d_x - \\\\nabla \\\\Phi\\\\|$ | 0.5 | 1.0 | 1.5 |\\n|----------------------------------------|-----------|-----------|-----------|\\n| **qNBO(SR1)** | 1.4E-05 | 7.2E-06 | 8.5E-06 |\\n| **qNBO(BFGS)** | 2.5E-02 | 1.2E-02 | 8.7E-03 |\\n| **AID-BIO** | 1.9E-05 | 1.2E-05 | 1.2E-05 |\\n| **AMIGO-CG** | 3.6E-04 | 1.2E-05 | 1.2E-05 |\\n| **AID-TN** | 4.1E+00 | 4.3E+00 | 4.0E+00 |\\n\\n\\nNote that the reason the performance of the AID-BIO and AMIGO-CG algorithms differs slightly in the updated results is due to incomplete memory release. To facilitate reproduction of the results, the specific hyperparameter settings are as follows:\\n\\n$\\\\textbf{AID-BIO/AMIGO-CG:}$ The maximum number of outer iterations is K = 5000, the CG iterations is $N = 1$, the inner gradient iteration is $T = 1$, the inner step size is $\\\\beta= 0.01$ and $\\\\textbf{the outer step size is $\\\\alpha = 0.2$}$.\\n\\n$\\\\textbf{AID-TN:}$ The maximum number of outer iterations is $K = 5000$, the TN iteration is $N = 1$, the inner gradient iteration is $T = 1$, the inner step size is $\\\\beta = 0.01$ and $\\\\textbf{the outer step size is $\\\\alpha = 0.2$}$.\\n\\n$\\\\textbf{qNBO (SR1):}$ The maximum number of outer iterations is $K = 5000$, the inner iterations are $T = 6$, $P = 9$, $Q_k = 25$, the inner step sizes are $\\\\beta = 0.01$, $\\\\gamma = 1$, the initial matrix is $H_0 = I$, and the outer step size is $\\\\alpha = 0.5$.\"}", "{\"summary\": \"The paper presents a bilevel optimization framework, QNBO, which utilizes quasi-Newton methods to address the computational challenges associated with hierarchical learning tasks. The authors integrate quasi-Newton recursion schemes for efficiently approximating inverse Hessian-vector products, specifically using BFGS and SR1 methods. The authors argue that QNBO offers superior convergence and computational efficiency, substantiated through theoretical proofs and numerical experiments.\", \"soundness\": \"4\", \"presentation\": \"4\", \"contribution\": \"4\", \"strengths\": \"1. The QNBO framework presents an innovative approach by combining bilevel optimization with quasi-Newton methods to improve hypergradient approximation accuracy.\\n\\n2. The paper provides a non-asymptotic convergence analysis for QNBO, especially for BFGS, detailing convergence under both quadratic and general settings. \\n\\n3. The authors validate QNBO on real-world tasks such as hyperparameter optimization, meta-learning, and data hyper-cleaning. QNBO\\u2019s performance is benchmarked against established algorithms, including SHINE and PZOBO, with performance gains observed in multiple scenarios.\\n\\n4. I did not notice any major issues with the clarity of the presentation.\", \"weaknesses\": \"1. The SR1 method, despite its potential in quasi-Newton methods, is noted for its numerical instability in general functions without correction strategies.\\n\\n2. QNBO offers a flexible choice between subroutine A and B based on performance and computational cost. However, the conditions for selecting one over the other are not rigorously outlined.\", \"questions\": \"1. Given the success of stochastic adaptations in other optimization frameworks, could the authors discuss potential challenges or advantages of incorporating stochastic quasi-Newton methods into QNBO? Specifically, how might this adaptation impact convergence rates or computational efficiency in practical implementations?\\n\\n2. The initial matrix $H_0$ plays a significant role in quasi-Newton methods. Here, a scalar multiple of the identity matrix is used. Could the authors provide specific results or theoretical analysis on how different choices of $H_0$ might affect QNBO\\u2019s convergence rate and computational efficiency?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"metareview\": \"This paper introduces qNBO, a novel bilevel optimization framework that leverages quasi-Newton methods to enhance hypergradient approximation. The authors provide a convergence analysis and demonstrate strong empirical performance across various machine learning tasks. While some areas require further investigation, such as the stability of the SR1 method and the stochastic setting, the paper's innovative approach and promising results justify its acceptance.\", \"additional_comments_on_reviewer_discussion\": \"There was lengthy back and forth with Reviewer yCCW who had identified some implementation issues, and helped solve it with the authors. This was tremendously helpful.\"}", "{\"title\": \"Response to concerns raised in Weaknesses.\", \"comment\": \">The SR1 method, despite its potential in quasi-Newton methods, is noted for its numerical instability in general functions without correction strategies.\\n\\n$\\\\textbf{Response:}$\\nThank you for your concern. The main drawback of SR1 updating is that the denominator in its update formula\\n$$ H_{t+1}=H_t+\\\\frac{(s_t-H_t g_t)( s_t-H_t g_t)^T}{( s_t-H_t g_t)^T g_t}$$\\ncan vanish. In fact, even for convex quadratic functions, there may be steps where no symmetric rank-one updating formula satisfies the secant equation. Such situations can lead to numerical instabilities and even a breakdown of the SR1 algorithm (see Section 6.2 of [1]). \\n\\nTo address these issues, a correction strategy can be applied to improve the numerical stability of SR1. The goal of such strategies is to keep the Hessian approximations under control. For detailed discussions on this topic, see Section 4.1 of Ye et al. (2023), which employs a correction strategy introduced in [2]. However, the correction strategy relies on a Hessian-based constant, which varies during iterations of $(x, y)$ in the bilevel setting. This introduces the need for an adaptive correction strategy in qNBO (SR1) to handle the dynamic nature of the Hessian in the bilevel setting. Developing such an adaptive strategy presents a distinct challenge in the bilevel optimization and could be an interesting future extension.\\n\\n\\n[1]Nocedal, Jorge, and Stephen J. Wright, Numerical optimization. New York, NY: Springer New York, 1999. \\n\\n[2]Rodomanov, Anton, and Yurii Nesterov, Greedy quasi-Newton methods with explicit superlinear convergence. SIAM Journal on Optimization (2021): 785-811.\\n\\n\\n\\n\\n>QNBO offers a flexible choice between subroutine A and B based on performance and computational cost. However, the conditions for selecting one over the other are not rigorously outlined.\\n\\n\\n$\\\\textbf{Response:}$\\nThanks for the concern. Unfortunately, at this point, we do not have a rigorous selection strategy for choosing between the two options: $Q_k=1$ or $Q_k>1$ in the iteration of $u_{k+1}$. Theoretically, if the ULI assumption described in (Ramzi et al., 2022) holds, then $Q_k=1$ would suffice, and we can share $\\\\\\\\{s_t, g_t\\\\\\\\}_{t=0}^{T-1}$ with subroutine $\\\\mathcal{A}$. However, this assumption is quite strong (cf. Ramzi et al., 2022).\\n\\nTo avoid potential incorrect inversion, we adopt $Q_k-1>0$ quasi-Newton steps in subroutine $\\\\mathcal{B}$ for the direction $\\\\nabla_y F(x_k, y_{k+1})$. Thanks to Reviewer 5z6s\\u2019s suggestion on employing a warm-start strategy for $u_{k+1}$ , we have conducted additional experiments incorporating this strategy. See Appendix C.5.1 (Figure 6) of the revised paper for details. The empirical results show that qNBO performs better when applying the warm-start strategy for $u$. A rigorous analysis would require a more detailed characterization of the relationships between quasi-Newton iterations and their sensitivity to variations in $(x, y)$, which would be an interesting future extension.\"}", "{\"title\": \"Thank you for your constructive feedback and for increasing the score. We have invested significant effort in conducting the meta-learning experiment. Below is our detailed response (part 1/2).\", \"comment\": \">Comment 1: Meta learning experiment\\n\\n$\\\\textbf{Response:}$\\nThe meta-learning experiment was conducted to assess whether qNBO could be applied to more complex machine learning tasks and achieve comparable or superior performance. To the best of our knowledge, PZOBO is regarded as the baseline algorithm for this type of meta-learning experiment, while other algorithms developed after PZOBO have not been applied to such experiments. Consequently, we focused primarily on comparing qNBO with PZOBO across various datasets (miniImageNet, Omniglot, and FC100).\\n\\nWe also attempted to test BOME but faced challenges when applying it to the miniImageNet dataset. The highest accuracy achieved by BOME was only 22.6%, compared to 59.8% for PZOBO and 62.1% for qNBO. Since BOME has not been previously applied to this type of meta-learning, its performance has not been publicly reported. Therefore, we chose not to include this result to maintain rigor.\\n\\nThank you for your suggestion. We tested SHINE in the context of meta-learning. Due to the large size of the datasets used in our experiments (miniImageNet, FC100, and Omniglot\\u2014specific details are provided in Appendix C4.1, page 21 of the paper), we initially employed a constant step size of 1 for the inner iteration, as an alternative option in Algorithm 1 of SHINE, instead of the time-consuming line-search. For testing, we selected the smaller FC100 dataset, where the highest test accuracy achieved by SHINE was only 33% (compared to 47.01% for PZOBO and 47.3% for qNBO). Incorporating the line-search increased the highest test accuracy to 37.5%, but this approach was computationally expensive and still underperformed compared to our proposed algorithm. Since meta-learning experiments are more complex and involve larger datasets than data cleaning tasks, we dedicated considerable time to fine-tuning the hyperparameters and testing the experiments over the past few days. We appreciate your understanding.\"}", "{\"title\": \"Discussion on the impact of Reviewer yCCW\\u2019s debug in the toy experiment (Part 1/3).\", \"comment\": \"We greatly appreciate Reviewer yCCW\\u2019s insightful debugging of AID-TN, AID-BIO, and AMIGO-CG in our toy experiment. The suggestions are both accurate and helpful, aligning with the code used in other experiments. Taking this opportunity, we further reviewed the toy experiment and identified two additional directions for improvement in the toy experiment.\\n\\nFirst, we need to fine-tune the hyperparameters of all algorithms (e.g., inner loop iterations and step sizes) to ensure a fair comparison. Previously, the inner loop setting (10 steps) for AID-TN, AID-BIO, and AMIGO-CG was adopted directly from other experiments without optimization.\\n\\nSecond, in the toy experiment, we used the closed-form expression $f_{yy} = A$ for the Hessian in the implementation of AID-TN, AID-BIO, and AMIGO-CG. Specifically, the code ICLR2025\\\\qNBO\\\\qNBO\\\\toy\\\\toy3 (lines 750, 814, and 876) calls the function f_yy1 (lines 99\\u2013100 of toy3) to directly return the Hessian matrix $A$. However, a more practical approach would involve computing Hessian-vector products using torch.autograd.grad. It is important to note that only the implementations of AID-TN, AID-BIO, and AMIGO-CG require the Hessian.\"}", "{\"comment\": \"We are glad to have addressed your concerns. Thank you so much for your constructive suggestions and support!\"}", "{\"comment\": \"Thank you so much for your support! We appreciate your constructive suggestions in enhancing the quality of our work.\"}", "{\"summary\": \"This paper proposes an algorithmic framework, qNBO, which leverages quasi-Newton methods to solve bilevel optimization problems in the nonconvex-strongly-convex setting. Some practical implementations are also provided under the theoretical framework. It establishes non-asymptotic convergence for the BFGS adaptation within the framework and provides numerical experiments demonstrating the superior performance of the proposed algorithms across various machine-learning applications.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. The paper is well written and easy to follow. Quasi-Newton type methods for solving bilevel optimization have not been well studied, even in the nonconvex-strongly-convex setting. This work provides a quite general algorithmic framework, allowing any quasi-Newton method to be applied in qNBO.\\n2. A convergence rate and complexity analysis are provided for qNBO (BFGS). Technical derivations seem to be nontrivial. The authors incorporate the superlinear convergence of BFGS into the non-asymptotic convergence framework commonly used in bilevel optimization literature. \\n3. Experiments validate the promise of the proposed methods.\", \"weaknesses\": \"1. For the theory, the results in Theorems 3.3 and 3.6 are limited to the setting where $Q_k=k+1$. As a result, qNBO is a double-loop algorithm. Is it possible to design a single-loop version? Recent progress has been made toward single-loop bilevel optimization algorithms, especially in the nonconvex-strongly-convex setting, by using a warm-start strategy.\\n2. For the experiments, since the value of $Q_k$ affects the running time, it would be beneficial to empirically demonstrate how increasing $Q_k$ influences performance gains. Note that the numerical results in Fig. 1(d) illustrate only the impact of $Q_k$ as the outer iteration $k$ varies.\\n3. Page 6: Do Theorems 3.3 and 3.6 require the assumption that an optimal solution $x^*$ exists? Note that $\\\\Phi(x^*)$ appears in both (6) and (7). \\n4. Page 7: Why is it stated that qNBO is more efficient than AID-BIO? Note that the notation $\\\\tilde{\\\\mathcal{O}}$ omits the $\\\\log\\\\frac{1}{\\\\epsilon}$ term. \\n5. In Appendix C.5, the authors report various ablation studies on the parameters $T$ and $Q$, but they do not specify which experiment these ablation studies are based on.\", \"questions\": \"Please see the questions in the weakness part. Here, I want to add some minor questions:\\n\\n1. Page 6: In Eq. (6) of Theorem 3.3, the constant $M_{f_{xy}}$ can be computed, as $f$ takes the quadratic form given in (5).\\n2. Page 8: In Eq. (9), ``${a_i}\\u2019$, ${b_i}\\u2019$\\u201d should be ``$a\\u2019_i$, $b\\u2019_i$\\u201d.\\n\\nI think this paper provides a good attempt to use second-order optimization to accelerate bilevel optimization. Overall, I like the approach. I am raising my score to 8 if the questions and the typos can be addressed well in the rebuttal.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"Summary:\\n\\nThe paper considers bilevel optimization problems with strongly convex lower-level objectives and proposes a framework for performing hypergradient approximation by using recursive quasi-Newton schemes to avoid issues in Hessian inversion arising in common BO algorithm. \\nThe proposed algorithms, qNBO (BFGS) and qNBO (SR1), take inspiration from an existing algorithm 'SHINE' (Ramzi et al 2022). SHINE uses quasi-newton for solving the lower-level and reuses the approximate inverse hessian to perform hypergradient computation. \\nHowever, such approximate hessian is only accurate when multiplied with a specific vector: the gradient of the inner objective. \\nHence, it might yield inaccurate solutions when multiplied by the gradient of the outer objective, as required by implicit differentiation. Although SHINE propose a way to steer the approximation towards the upper-level, it is unlear to what extend the method enjoys good quantitive convergence rates. This paper addresses such limitation by introducing a separate quasi-newton based estimation of the inverse-hessian along the gradient of the outer objective. This results, in principle, in a more accurate estimation of the hypergradient. \\nThe paper then provides finite time convergence guarantees and illustrate the method on a toy example, a regression task and a meta-learning task.\", \"soundness\": \"1\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": [\"The motivation and approach in the paper are quite interesting, it makes sense to accelerate convergence of the inner problem using quasi-newton approaches. It is also nice that the method provide convergence rates for their algorithm, although I think the analysis do not seem to capture the benefits of qn for computing the hyper-gradient (when computing the iterates u_k).\", \"The paper is overall clearly written and well explained, which makes it easy to read.\", \"The method seems to give quite an improvement compared to SHINE in terms of convergence speed (no full finite time convergence was provided for SHINE). While the experiments seem in favor of the proposed method, I think the implementation of SHINE contains some errors after checking the code (see weakness section).\", \"I appreciate that the authors provided the code to reproduce the results, this allows for a better assessment for the results. Although there were bugs in it, that unfortunately change some conclusions (see the weakness section), I think this should be seen as an opportunity to improve the experimental study.\"], \"weaknesses\": \"- Experiments: The results in the toy experiment looked suspicious to me: AID-TN, AID-BIO and AmigoCG seemed to work unreasonably worse on a quite simple example where they are supposed to perform quite well. I decided to check the implementation provided in the supplementary and found a number of bugs that explain these results. Please refer to the questions section for the details of these bugs.\\nI think these can be easily fixed. However, after fixing these bugs, the results do not exactly match the conclusions of the current paper: it is possible to solve the toy problem faster/more accurately using AID-BIO/AmigoCG using a single step of CG (instead of 10). More precisely, after fixing the bugs: AID-BIO/AmigoCG reach 10^-4 relative error on the upper variable (instead the numbers reported in the paper: 10^-2 for AID-BIO and 10^0 for AmigoCG), and further go below 10^-5 (better than qNBO) when using only a single CG step and single gradient step. AID-TN reaches 10^-1, instead of 10^0. I do not exclude other bugs on other methods (the implementation of SHINE already looks incorrect (see questions section)), so I strongly encourage the authors to test the implementation with care. \\n\\n\\n\\n\\n- Convergence results: \\nWhile the convergence analysis exploits the super-linear convergence of quasi-newton for the inner-level iterates, the iterates used for computing the hyper-gradient do not seem to benefit from such super-linear rate. In fact, the convergence of these iterates is sub-linear: 1/Q where Q is the number of quasi-Newton iterates. This might explain the worse dependence of the algorithm on the number of gradient queries on the lower loss epsilon^{-2} compared to AID-BIO epsilon^{-1}. Still there is an improved dependence on the conditioning for the vector-jacobian complexity (kappa^3) vs AID-BIO (kappa^{3.5}). However, it is unclear how this compensates the increased complexity for the gradient, especially given the results of the toy experiments (without the bugs). \\nOn a positive note, the result is an improvement over the guarantees provided for SHINE as they are quantitative.\", \"ablation_studies\": \"methods such as AID-BIO do not exclude using other solvers for the lower-level problem. One could then use quasi-newton methods and still apply CG for the conjugate gradient. How do you think this would perform?\\n\\n\\nAlthough I think the proposed direction is interesting, there is a number of problems with the current version of the paper that need to be addressed. I have many reasons not to trust the empirical results and that would require checking all the details of the code for all experiments. It is unclear to me if the rebuttal period would be enough for that.\", \"questions\": \"Some of the figures in the appendix seem to be consitent with what the python file toy3.py produces. However, it contains a number of bugs for various implementations:\", \"1__aid_tn\": \"a) the implementation of the function TN is wrong, it should be:\\n\\t \\tv = v - inner_lr * output\\n\\t\\treturn inner_lr * p\", \"instead_of\": \"ogrady = F_y(x,y)\", \"2__aid_bio\": \"Same issue as in AID-TN: Conjugate gradient should be called before computing the cross-derivative for the same reasons, otherwise this results in unnecessarily inaccurate hyper-gradients which slow performance:\\n\\t\\tv = cg(fyy,Fy,v,10)\\n\\t\\tfyx=f_xy(x,y,v)\", \"3__amigocg\": \"In the deterministic case, this corresponds exactly to AID-BIO algorithm, so they should give the exact outputs. Similarly to AID-BIO, cg should be called before the cross-derivative. Additionally, there is no need to evalulate f_x(x,y) and f_y(x,y) for hyper-gradient computation as this unnecessarily slows the method. More importantly, there should be a minus sign instead of a + in front of the implicit gradient:\\n\\t\\tx_grad=Fx-w\", \"4__shine\": \"The implementation is also incorrect though I did not have time to fix it myself. The OPA correction involves only the cross derivatives of the inner-level function not the outer level function:\\n\\t\\togrady = f_xy(x,y)\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to concerns raised in Questions (Part 1/2).\", \"comment\": \">Given the success of stochastic adaptations in other optimization frameworks, could the authors discuss potential challenges or advantages of incorporating stochastic quasi-Newton methods into QNBO? Specifically, how might this adaptation impact convergence rates or computational efficiency in practical implementations?\\n\\n$\\\\textbf{Response:}$ \\nThank you for your thoughtful questions. qNBO incorporates a generic decoupling structure, wherein all search directions are linear with respect to the objective functions. This structure builds upon existing methods such as HOAG (Pedregosa, 2016), AID-BIO (Ji et al., 2021), AMIGO (Arbel & Mairal, 2022), and SOBA/SABA (Dagr\\u00e9ou et al., 2022), enabling qNBO to extend effectively from deterministic to stochastic settings.\\n\\n\\nTo elaborate, qNBO consists of three components, with the first two employing quasi-Newton recursion schemes. A straightforward stochastic adaptation involves replacing deterministic quasi-Newton recursion schemes with stochastic variants, such as K-BFGS [1] or Stochastic Block BFGS [2]. Additionally, in Part 3 of qNBO, deterministic gradients are replaced with stochastic gradients throughout the iterative process, consistent with the transition from deterministic to stochastic methods discussed by Dagr\\u00e9ou et al. (2022). \\n\\nStochastic quasi-Newton methods could be better equipped than stochastic gradient methods to handle ill-conditioning. Incorporating these methods into qNBO offers the advantage of mitigating the adverse effects of high non-linearity and ill-conditioning in the lower-level objective function by leveraging second-order information. Another potential benefit is the improvement of the constants involved in the sublinear convergence rates of stochastic methods. However, the full potential of the stochastic quasi-Newton schemes discussed (and possibly others) is not yet known. \\n\\nFurthermore, this straightforward extension may compromise convergence rates and computational efficiency under the same smoothness assumptions as in the deterministic setting. For an intuitive comparison (without using quasi-Newton methods), refer to the analysis of SOBA (Dagr\\u00e9ou et al., 2022) and its improvement by MA-SOBA (Chen et al., 2023, Optimal Algorithms for Stochastic Bilevel Optimization under Relaxed Smoothness Conditions).\\n\\nIn our opinion, the main challenges in maintaining comparable convergence rates and computational efficiency in a stochastic setting include:\\n\\n$\\\\textbf{Constructing effective estimators:}$ How can we develop unbiased or biased estimators in Part 3 of qNBO by integrating techniques such as variance reduction or momentum (e.g., moving averages)? Potential candidates for these estimators include SGD, SAGA, SARAH, STORM, PAGE, and even Adam. However, analyzing this extension, particularly its convergence properties, requires significant effort. For recent progress (without using quasi-Newton methods), see SRBA (Dagr\\u00e9ou et al., 2024) and SPABA (Chu et al., 2024).\\n\\n$\\\\textbf{Analyzing convergence rates and complexity:}$ How can we evaluate the proposed stochastic algorithms in a bilevel setting while incorporating noisy second-order (curvature) information? Addressing these difficulties may require theoretical breakthroughs that go beyond existing first-order techniques, leaving this an unresolved challenge and an interesting future extension.\\n\\n\\n[1] Goldfarb, D., Ren, Y., & Bahamou, A. (2020). Practical quasi-newton methods for training deep neural networks. \\n\\n[2] Gower, R., Goldfarb, D., & Richt\\u00e1rik, P. (2016). Stochastic block BFGS: Squeezing more curvature out of data.\"}", "{\"title\": \"Response to concerns raised in Questions.\", \"comment\": \">In Remark 3.4, the authors stated that the qNBO (SR1) algorithm lacks convergence guarantees without specific corrections used to achieve numerical stability in the general case. It was also mentioned that the performance of qNBO (SR1) on the 20News dataset was omitted due to its ineffectiveness in the experiment in Section 4.2. What are the underlying reasons for these issues? A bit more discussion on these points would be helpful.\\n\\n$\\\\textbf{Response:}$ \\nThank you for your concern. The SR1 update formula is given by:\\n$$ H_{t+1}=H_t+\\\\frac{(s_t-H_t g_t)( s_t-H_t g_t)^T}{( s_t-H_t g_t)^T g_t}.$$\\nHowever, even for convex quadratic functions, there are instances where no symmetric rank-one updating formula satisfies the secant equation. This issue arises when $ s_t-H_t g_t \\\\neq 0$ but $( s_t-H_t g_t)^T g_t =0$ (see Section 6.2 of [1]). Such scenarios can lead to numerical instabilities and even a breakdown of the SR1 algorithm. To improve the numerical stability of the SR1 method, a correction strategy should be applied. For discussions on this topic, refer to Section 4.1 of Ye et al. (2023). \\n\\n\\n[1]Nocedal, Jorge, and Stephen J. Wright, Numerical optimization[M]. New York, NY: Springer New York, 1999. \\n\\n\\n>From the first terms on the right hand sides of both (6) and (7), it seems that Theorems 3.3 and 3.6 require a lower bound assumption for $\\\\Phi(x)$. Is this correct? Additionally, in the proof sketch in Appendix D.2, $\\\\tilde{\\\\nabla}\\\\Phi$ is mentioned but not defined. It would be helpful to specify its definition.\\n\\n$\\\\textbf{Response:}$\\nYes, your understanding is correct. As noted in the conclusions of the proofs of Theorems 3.3 and 3.6 (Lines 1669 and 1942 of the original version), $\\\\Phi(x^*)$ can be replaced with any lower bound of $\\\\Phi$. This correction has been incorporated into the revised version. To clarify, we have added the standard lower bound assumption for $\\\\Phi(x)$ in the revised version. \\n\\nThank you for pointing out this notation issue. $\\\\tilde{\\\\nabla} \\\\Phi$ represents the approximate hypergradient, defined as $\\\\tilde{\\\\nabla} \\\\Phi := \\\\nabla_{x} F(x_k, y_{k+1}) - [\\\\nabla^2_{xy} f(x_k, y_{k+1})]^T u_{k+1}$. This definition has been included in Line 165 of the revised paper. \\n\\n>In the experiments, how to select the step sizes and the number of inner iterations for the proposed algorithms?\\n\\n$\\\\textbf{Response:}$ The implementations and hyperparameter settings for the proposed algorithms in all numerical experiments are detailed in Section C of the Appendix.\\n\\nBased on our experience, the warm-up step size $\\\\beta$ for the inner loop $\\\\mathcal{A}$ is typically set to 0.1. However, for logistic regression, a step size of $0.0001/(j+1)$ has been found to be more effective. The step size $\\\\gamma$ was tuned within ${1, 0.5, 0.1}$. Specifically, we set $\\\\gamma = 1$ in the toy example, while $\\\\gamma = 0.1$ was more effective in other experiments. The number of inner iterations, $T$, significantly influences the parameters $\\\\omega$ and $\\\\tau$, which are critical for satisfying the convergence conditions. Adjustments to $T$ are made based on the condition number of the objective function in the bilevel problem. Experiments, including ablation studies on $T$ from the data hyper-cleaning experiment (Figure 9 in the Appendix of the original version), demonstrate that smaller values of $T$ lead to greater efficiency.\\n\\n>In Table 1, why does qNBO (BFGS) save so much time compared to SHINE for data hyper-cleaning?\\n\\n$\\\\textbf{Response:}$ \\nThe reason is that qBNO employs a constant step size strategy instead of the time-consuming line search used in SHINE for the inner solver. To clarify, consider the data hyper-cleaning experiment on the MNIST dataset. The SHINE-OPA algorithm requires 23.19 seconds to achieve a test accuracy of 91.51%. (The 20.25 seconds reported in Table 1 of the paper is the average time across ten runs.) Notably, the line search method in the lower-level solver accounts for 19.61 seconds of the total time.\"}", "{\"title\": \"Reply to rebuttals\", \"comment\": \"The authors totally deal with my concerns, so I keep my score.\"}", "{\"title\": \"Code snippet using autograd\", \"comment\": \"Code using autograd\", \"in_the_implementation_of_aid_bio_please_replace_the_following\": \"~~~ \\n fyy=f_yy1(x,y)\\n v = cg(fyy,Fy,v,1)\\n fyx= f_xy(x,y,v)\\n~~~\", \"by\": \"~~~ \\n fyy_func=f_yy_func(x,y)\\n v= cg_func(fyy_func,Fy,v,1)\\n fyx_func = f_xy_func(x,y)\\n fyx= fyx_func(v)\\n~~~ \\n\\nHere ```cg_func(A,b,x,num_steps)``` is similar to the implemented function ```cg(A,b,x,num_steps)``` , except that the first argument is now a function that takes a vector and performs Hessian-vector multiplication using autograd. This means that ```cg_func``` only needs to replace the two matrix-vector products ```A @ x``` and ```A @ p``` by evaluations ```A(x)``` and ```A(p)```. \\n\\nThe functions ```f_yy_func``` and ```f_xy_func``` return function handles and are given by:\\n\\n~~~ \\ndef f_yy_func(x,y):\\n\\n def f_yy(z):\\n val = f(x,y)\\n\\n grad = autograd.grad(outputs=val, \\n inputs=y, \\n grad_outputs=None, \\n retain_graph=True,\\n create_graph=True, \\n only_inputs=True,\\n allow_unused=True)\\n\\n hvp = autograd.grad(outputs=grad, inputs=y, grad_outputs=z) \\n return hvp[0]\\n return f_yy\\n~~~ \\n\\n~~~ \\ndef f_xy_func(x,y):\\n def f_xy(z):\\n val = f(x,y)\\n\\n grad = autograd.grad(outputs=val, \\n inputs=y, \\n grad_outputs=None, \\n retain_graph=True,\\n create_graph=True, \\n only_inputs=True,\\n allow_unused=True)\\n\\n hvp = autograd.grad(outputs=grad, inputs=x, grad_outputs=z) \\n return hvp[0]\\n return f_xy\\n~~~\"}", "{\"summary\": \"This paper integrates a quasi-Newton approach into a bilevel optimization algorithm to approximate the inverse Hessian-vector product.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"3\", \"strengths\": \"The proposed method avoids costly second-order computations for approximating the Hessian-vector product while achieving a comparable convergence rate.\", \"weaknesses\": \"The paper is challenging to follow, with a vague explanation of the quasi-Newton recursion scheme. The design of the proposed method appears overly complex compared to existing methods, and there is limited explanation regarding the validity of $u_{k+1}$ as an accurate approximation for the Hessian-vector product.\", \"questions\": \"Can this method be effectively applied in a stochastic setting while maintaining comparable convergence rates and computational efficiency? The required adjustments in iteration rounds and step size suggest potential difficulties in achieving the same convergence speed as existing methods.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to the concerns raised in Weaknesses.\", \"comment\": \">For the theory, the results in Theorems 3.3 and 3.6 are limited to the setting where $Q_k=k+1$. As a result, qNBO is a double-loop algorithm. Is it possible to design a single-loop version? Recent progress has been made toward single-loop bilevel optimization algorithms, especially in the nonconvex-strongly-convex setting, by using a warm-start strategy.\\n\\n$\\\\textbf{Response:}$ \\nThank you for the suggestion. A warm-start strategy is reasonable based on the Lipschitz continuity of $u^*(x, y)$ under the given assumptions. We have incorporated experiments using the warm-start strategy you suggested for $u$. As shown in Appendix C.1 (Figure 6 on Page 17) of the revised paper, the empirical results demonstrate that qNBO(BFGS) performs better when applying the warm-start strategy for $u_k$ (denoted as $Q_k = \\\\text{ws} $ in the Figure 6). A rigorous analysis would require a more explicit characterization of the relationships between quasi-Newton iterations and their sensitivity with respect to $(x, y)$, which would be an interesting future extension. In our opinion, this could serve as a first step toward developing a single-loop version of qNBO. Thank you for your insightful question; we will continue to explore this idea.\\n\\n\\n\\n>For the experiments, since the value of $Q_k$ affects the running time, it would be beneficial to empirically demonstrate how increasing $Q_k$ influences performance gains. Note that the numerical results in Fig. 1(d) illustrate only the impact of $Q_k$ as the outer iteration $k$ varies.\\n\\n$\\\\textbf{Response:}$\\nThank you for the suggestion. In the original submitted version (Figure 10 on Page 22 in Appendix C.5), an ablation study on the iteration number $Q_k$ of qNBO (BFGS) in the data hyper-cleaning experiment was performed using running times. We have added additional ablation study experiments based on running times for the toy experiment; see Appendix C.5.1 (Page 17, Figure 6) in the revised version. As shown in Figure 6, when $Q_k$ is constant, an increase in $Q_k$ initially slows the decline in various indicators over time but ultimately leads to a smaller minimal error. However, $Q_k = k+1$ demonstrates superior performance in terms of both the rate of decline and the minimal error achieved, even when its initial number of steps is fewer than that of constant $Q_k$. Furthermore, employing a warm-start strategy for $u_k$ improves the performance of qNBO (BFGS), highlighting the potential advantages of this approach, as noted by the reviewer.\\n\\n\\n\\n>Page 6: Do Theorems 3.3 and 3.6 require the assumption that an optimal solution $x^\\u2217$ exists? Note that $\\\\Phi(x^*)$ appears in both (6) and (7).\\n\\n$\\\\textbf{Response:}$ Thank you for your questions. Theorems 3.3 and 3.6 do not assume the existence of optimal solutions. As noted at the conclusions of their proofs (Lines 1669 and 1942 of the original version), $\\\\Phi(x^*)$ can be replaced with any lower bound of $\\\\Phi$. Thus, we add the lower bound assumption on $\\\\Phi(x)$ (Assumption 3.3 in the revised paper) and replace $\\\\Phi(x^*)$ with $\\\\inf_x \\\\Phi(x)$ in Theorems 3.4 and 3.7 and the corresponding proofs in the appendix of the revised version. \\n\\n>Page 7: Why is it stated that qNBO is more efficient than AID-BIO? Note that the notation $\\\\tilde{\\\\mathcal{O}}$ omits the $\\\\log(1/\\\\epsilon)$ term.\\n\\n$\\\\textbf{Response:}$ Thank you for your concern. Your understanding is correct\\u2014it is inaccurate to claim that qNBO is more efficient than AID-BIO in terms of matrix-vector complexity. According to Table 1 of AID-BIO (Ji et al., 2021), the number of Jacobian-vector products for AID-BIO is of the order $\\\\mathcal{O}(\\\\kappa^{3}\\\\epsilon^{-1})$, and the number of Hessian-vector products is of the order $\\\\mathcal{O}(\\\\kappa^{3.5}\\\\epsilon^{-1})$. When the dimension of the lower-level variable exceeds that of the upper-level variable (e.g., the model parameter\\u2019s dimension is larger than that of the hyperparameter), Hessian-vector products become more computationally intensive. Consequently, the matrix-vector complexity for AID-BIO is $\\\\mathcal{O}(\\\\kappa^{3.5}\\\\epsilon^{-1})$, compared to $\\\\tilde{\\\\mathcal{O}}(\\\\kappa^{3}\\\\epsilon^{-1})$ for qNBO. Thus, qNBO is not more efficient than AID-BIO in terms of $\\\\epsilon$-order; it is only more efficient in the $\\\\kappa$ term within matrix-vector complexity. This correction has been included in the revised version.\\n\\n>In Appendix C.5, the authors report various ablation studies on the parameters $T$ and $Q$, but they do not specify which experiment these ablation studies are based on.\\n\\n$\\\\textbf{Response:}$ The ablation studies described in Appendix C.5 are based on the data hyper-cleaning experiment. This information has been added to Appendix C.5.2 of the revised version of the paper. Thank you for bringing this to our attention.\"}", "{\"title\": \"We greatly appreciate your thorough and valuable feedback. Below is our response to the concerns raised in your new comments (part 3/3).\", \"comment\": \">However, all the other experiments are missing the AID-BIO/AMIGO-CG which are the strongest baseline according to the toy experiments. Why are some algorithms missing in those comparisons? Why is the meta-learning experiment compares only with PZOBO?\\n\\n$\\\\textbf{Response:}$ Based on your suggestion, we included AID-BIO/AMIGO-CG as comparison algorithms in the logistic regression and data hyper-cleaning experiments. The results are as follows:\\n\\n\\nMNIST (data hyper-cleaning)\\n| Time (s) for test accuracy | 1 | 5 | 10 |\\n|----------------------------------------|----------|----------|----------|\\n| **AID-BIO** | 0.901 | 0.912 | 0.913 |\\n| **BFGS** | 0.896 | 0.913 | 0.917 |\\n| **SR1** | 0.899 | 0.912 | 0.914 |\\n\\nFashion MNIST (data hyper-cleaning)\\n| Time (s) for test accuracy | 1 | 5 | 10 |\\n|----------------------------------------|----------|----------|----------|\\n| **AID-BIO** | 0.818 | 0.825 | 0.823 |\\n| **BFGS** | 0.826 | 0.829 | 0.83 |\\n| **SR1** | 0.82 | 0.828 | 0.831 |\\n\\nReal-sim (logistic regression)\\n| Time (s) for test loss | 1 | 2 | 5 |\\n|----------------------------------------|----------|----------|----------|\\n| **AID-BIO** | 8.1E2 | 6.4E2 | 6E2 |\\n| **BFGS** | 1.7E3 | 3.8E2 | 2.1E2 |\\n| **SR1** | 4.3E2 | 2.4E2 | 2.2E2 |\\n\\n\\n20news group (logistic regression)\\n| Time (s) for test loss | 1 | 2 | 5 |\\n|----------------------------------------|----------|----------|----------|\\n| **AID-BIO** | 1.7E2 | 1.2E2 | 1.1E2 |\\n| **BFGS** | 1.6E3 | 1.2E2 | 8.1E1 |\\n| **SR1** | 3.3E2 | 2.7E2 | 2.2E2 |\\n\\nSimilarly, we have updated the above results in the latest version of the paper. For more details, please refer to Figures 2-3 of the revised paper. Additionally, we would like to point out that the AMIGO-CG algorithm mentioned in the original paper has been corrected to the AMIGO algorithm, which uses the stochastic gradient descent method to update the auxiliary variable $v$. Since AID-BIO shares the same iteration format as the AMIGO-CG algorithm, we treat them as a single algorithm for comparison.\\n\\nWhy the only competitor in the meta-learning experiment is PZOBO? \\n(1)The comparison between PZOBO and MAML and ANIL is conducted in [1]. The results which are shown in Figure 5 of [1] indicate that PZOBO has better performances than MAML and ANIL.\\n(2)In our research, we primarily focused on assessing whether qNBO could be applied to more complex machine learning experiments. Thus, in the meta-learning experiment, we only compared qNBO with PZOBO, currently recognized as the best algorithm [1] for this type of meta-learning. Due to the extensive duration required for meta-learning experiments and our constrained timeline, along with our focus on other research priorities, we did not explore these algorithms further. We hope the reviewer understand the constraints and focus of our study.\\n\\n[1] Sow D, Ji K, Liang Y. On the convergence theory for hessian-free bilevel algorithms[J]. Advances in Neural Information Processing Systems, 2022, 35: 4136-4149.\\n\\n\\n\\n\\n>As a result, the final complexity has a degraded dependence in the gradient queries to the lower loss. It is unclear how this degraded complexity is compensated by the improved dependence on the conditioning for vector-jacobian products.\\n\\n\\n\\n\\n$\\\\textbf{Response:}$\\nThank you for your thoughtful feedback. Yes, your understanding is correct\\u2014there are unsatisfactory aspects in the theoretical analysis of this work, as stated in the theoretical comparisons at the end of Section 3. In theory, qNBO is not more efficient than AID-BIO in terms of $\\\\epsilon$-order; it is only more efficient with respect to the $\\\\kappa$ term within matrix-vector complexity. This degraded complexity appears to result from the cold-start strategy for $u_{k+1}$. Following Reviewer 5z6s\\u2019s suggestion to employ a warm-start strategy for $u_{k+1}$, we conducted additional experiments incorporating this approach. Details are provided in Figure 6 of the revised paper. The empirical results demonstrate that qNBO achieves better performance when the warm-start strategy is applied to $u$. A rigorous analysis would require a more detailed characterization of the relationships between quasi-Newton iterations and their sensitivity to variations in $(x, y)$, which we consider an interesting direction for future research.\"}", "{\"title\": \"Response to the concerns raised regarding the design of the proposed method and its stochastic adaptation (part 1/2).\", \"comment\": \">Weaknesses: The paper is challenging to follow, with a vague explanation of the quasi-Newton recursion scheme. The design of the proposed method appears overly complex compared to existing methods, and there is limited explanation regarding the validity of $u_{k+1}$ as an accurate approximation for the Hessian-vector product.\\n\\n$\\\\textbf{Response:}$\\nThank you for your comment. Appendix B provides a comprehensive explanation of the quasi-Newton recursion scheme, including the algorithmic framework and key details omitted from the main text for brevity. Below, we elaborate on the necessity of subroutine $\\\\mathcal{B}$ and and present explicit lemmas to justify why $u_{k+1}$ serves as an accurate approximation of the Hessian-vector product.\\n\\nIn quasi-Newton methods, the convergence of the quasi-Newton matrix $H^{-1}$ to the true Hessian matrix $\\\\nabla_{yy} f$ depends on strong assumptions [1], which are often not satisfied in practice [2]. Nevertheless, even without these assumptions, the quasi-Newton matrix $H^{-1}$ can still converge to the true Hessian $\\\\nabla_{yy} f$ along specific directions [1,3]. Accordingly, in Step 2 of Algorithm 1, we apply the quasi-Newton recursion along the direction $\\\\nabla_y F$ to compute $u_{k+1}$ , which serves as an approximation of the Hessian-vector product $[\\\\nabla_{yy} f]^{-1} \\\\nabla_y F$ . Recent studies [4,5,6] have explicitly characterized the accuracy of the quasi-Newton matrix in approximating the Hessian along specific directions. Building on these findings, Lemmas D.16 (Lemma D.18 in the revised paper) and D.21 (Lemma D.23 in the revised paper\\uff09in Appendix D demonstrate how $u = H^{-1} \\\\nabla_y F$ effectively approximates $[\\\\nabla_{yy} f]^{-1} \\\\nabla_y F$ .\\n\\n\\n[1] Nocedal, Jorge, and Stephen J. Wright, Numerical optimization[M]. New York, NY: Springer New York, 1999.\\n\\n\\n[2] Mannel, F. On the convergence of the Broyden-like matrices[J]. 2020.\\n\\n[3] Dennis J E, Mor\\u00e9 J J. A characterization of superlinear convergence and its application to quasi-Newton methods[J]. Mathematics of computation, 1974, 28(126): 549-560.\\n\\n[4] Rodomanov A, Nesterov Y. Rates of superlinear convergence for classical quasi-Newton methods[J]. Mathematical Programming, 2022: 1-32.\\n\\n[5] Rodomanov A, Nesterov Y. New results on superlinear convergence of classical quasi-Newton methods[J]. Journal of optimization theory and applications, 2021, 188: 744-769.\\n\\n[6] Jin Q, Mokhtari A. Non-asymptotic superlinear convergence of standard quasi-Newton methods[J]. Mathematical Programming, 2023, 200(1): 425-473.\"}", "{\"title\": \"The issues encountered with AID-TN, AID-BIO, and AMIGO-CG in the toy experiment do not impact the outcomes of other experiments.\", \"comment\": \"We sincerely thank Reviewer yCCW for the thorough and valuable feedback on our code. Before addressing the issues identified with AID-TN, AID-BIO, and AMIGO-CG in the toy experiment, we first assessed their potential impact on other experiments.\\n\\nOur analysis indicates that these issues do not affect the outcomes of other experiments. This conclusion is based on the fact that $\\\\textbf{the toy experiment and other experiments use separate algorithm files.}$ Specifically, all the code for the algorithms in the toy experiment is contained in the file ICLR2025\\\\qNBO\\\\qNBO\\\\toy, which was rewritten exclusively for the toy experiment. Notably, the algorithms AID-TN, AID-BIO, and AMIGO-CG in the toy experiment utilize a closed-form Hessian (discussed later). In contrast, their implementations in other experiments rely on code from the file ICLR2025\\\\other algorithms\\\\solvers, which originates from the open-source implementation cited in the paper.\\n\\n\\nThe following provides the specific file paths for thoroughly checking the bugs mentioned by Reviewer yCCW for AID-TN, AID-BIO, and AMIGO-CG during the toy experiment. $\\\\textbf{Notably, these bugs are absent from the codes in}$ ICLR2025\\\\other algorithms\\\\solvers $\\\\textbf{used for other experiments.}$\\n\\n$\\\\textbf{AID-TN:}$ From ICLR2025\\\\other algorithms\\\\solvers\\\\stocbio.py (lines 84\\u201395) and ICLR2025\\\\other algorithms\\\\benchmark_utils\\\\hessian_approximation.py (lines 116\\u2013121), it can be verified that the code of AID-TN used in other experiments aligns with Reviewer yCCW\\u2019s debug on AID-TN for the toy experiment.\\n \\n$\\\\textbf{AID-BIO:}$ From ICLR2025\\\\other algorithms\\\\solvers\\\\stocbio.py (lines 84\\u201395), it can be verified that the code of AID-BIO used in other experiments aligns with Reviewer yCCW\\u2019s debug on AID-BIO for the toy experiment.\\n\\n$\\\\textbf{AMIGO-CG:}$ From ICLR2025\\\\other algorithms\\\\solvers\\\\amigo.py (lines 87\\u201398 and lines 99\\u2013102), it can be verified that the code of AMIGO-CG used in other experiments aligns with Reviewer yCCW\\u2019s debug on AMIGO-CG for the toy experiment.\"}", "{\"title\": \"The implementation of SHINE is correct.\", \"comment\": \"SHINE incorporates two different OPA (Outer Problem Awareness) corrections as described in Ramzi et al. (2022).\\u00a0$\\\\textbf{The first correction involves the cross-derivatives of the inner-level function $f_{xy}(x, y)$ when the upper-level variable $x$ is one dimentianal}$ (i.e., $f_{xy}(x, y)$ is a vector). This correction is outlined in equation (5) of Ramzi et al. (2022) and explained above it. It corresponds to the term $f_{xy}[f_{yy}]^{-1}$ in the computation of the hypergradient and requires $x$ to be one-dimensional.\\n\\n$\\\\textbf{The second OPA correction, described in equation (8) of Ramzi et al. (2022), involves $F_y(x,y)$}$, which is a vector regardless of the dimensionality of $x$. This correction corresponds to $[f_{yy}]^{-1}F_y$ in the computation of the hypergradient. $\\\\textbf{This is the version we used in both the toy experiment and other experiments.}$\", \"the_specific_file_paths_for_verifying_the_implementation_of_shine_are_as_follows\": \"$\\\\textbf{SHINE:}$ From ICLR2025\\\\qNBO\\\\qNBO\\\\logisticregression\\\\hoag\\\\hoag.py (lines 308 and 383) and ICLR2025\\\\qNBO\\\\qNBO \\\\logisticregression\\\\hoag\\\\lbfgs.py (line 84), it can be confirmed that the code for SHINE aligns with equation (8) of Ramzi et al. (2022). Additionally, it matches the original SHINE open-source code (lines 147\\u2013148 in https://github.com/zaccharieramzi/hoag/blob/shine/hoag/hoag.py and lines 80\\u201383 in https://github.com/zaccharieramzi/hoag/blob/shine/hoag/lbfgs.py).\"}", "{\"title\": \"Thank you. The response addresses my major concerns.\", \"comment\": [\"It is good to see that you have fixed the implementations. I have raised my score to 5 since I believe that the method is interesting, but that there is still some work to do in order to ensure the experimental comparaisons are correctly conducted. In particular:\", \"For the meta learning experiment, the justification for restricting the comparaison to PZOBO is unconvincing, especially that many of the considered methods appeared around the same time as PZOBO. The paper should at least compare with SHINE, since it is quite related.\", \"Effect of conditioning: In the reported results comparing qNBO with AID-BIO for higher conditioning number, the number of iterations (T, N) are fixed to 1 for AID-BIO while T=15 and Q_k=25 for qNBO. Compared to the better conditionned problem, T was increased from 6 to 15 for qNBO, but not for AID-BIO. How were these choices made? To have a fair comparaison, one most perform a grid search on these hyper-parameters for each method using a similar budget for each method. Was this the case? For instance, have you tried running AID-BIO with T=N=14 on this problem?\"]}", "{\"title\": \"Additional experiments to explore why 1 step CG iteration, as recommended by Reviewer yCCW, performs better than 10 steps.\", \"comment\": \"We compared the performance of AID-BIO with different numbers of CG iterations ($T=1, 10$). The results are presented in Figure 3 of the single-page PDF at the anonymous link and in the following tables. Using the number of outer-loop iterations of the algorithm on the x-axis, we observed that both CG iteration strategies exhibited nearly identical performance. This suggests that, due to the closed form of the Hessian provided in the code, a single CG iteration is sufficient to reach the optimal solution. Additional CG iterations do not improve computational accuracy, making $T=1$ more time-efficient than $T=10$.\\n\\n| Iteration for \\\\|x - x^*\\\\| / \\\\|x^*\\\\| | T=1 | T=10 |\\n|-----------|-------------|--------------|\\n| 100 | 4.9E-01 | 4.9E-01 |\\n| 200 | 4.3E-02 | 4.3E-02 |\\n| 300 | 4.1E-03 | 4.2E-03 |\\n| 400 | 4.2E-04 | 4.3E-04 |\\n| 500 | 4.5E-05 | 4.5E-05 |\\n\\n| Iteration for \\\\|y - y^*\\\\| / \\\\|y^*\\\\| | T=1 | T=10 |\\n|-----------|-------------|--------------|\\n| 100 | 6.3E-01 | 6.3E-01 |\\n| 200 | 5.5E-02 | 5.6E-02 |\\n| 300 | 5.4E-03 | 5.4E-03 |\\n| 400 | 5.5E-04 | 5.6E-04 |\\n| 500 | 5.8E-05 | 5.9E-05 |\"}", "{\"title\": \"Response to the concerns raised in Questions.\", \"comment\": \">Page 6: In Eq. (6) of Theorem 3.3, the constant Mfxy can be computed, as f takes the quadratic form given in (5).\\n\\n$\\\\textbf{Response:}$ Thank you for your suggestion. Since $\\\\nabla^2_{xy} f(x, y) = -I$ for the lower-level objective function given in Eq. (5), the constant $M_{f_{xy}} = 1$. We have made the corresponding revisions in the paper. \\n\\n\\n>Page 8: In Eq. (9), ${a_i}\\u2019$, ${b_i}\\u2019$\\u201d should be ai\\u2032, bi\\u2032\\u201d.\\n\\n$\\\\textbf{Response:}$ \\nThank you for pointing this out. The corrections to Eq. (9) have been made in the revised version.\"}", "{\"title\": \"Thanks for the response\", \"comment\": \"I thank the authors for the detailed responses. I went through the responses and other reviewers' comments. I am satisfied with the answers to my questions, so I increase my score accordingly.\\n\\nBest,\\nReviewer\"}", "{\"title\": \"We greatly appreciate your thorough and valuable feedback. Below is our response to the concerns raised in your new comments (part 1/3).\", \"comment\": \">Hessian-vector product using autograd\\n\\n$\\\\textbf{Response:}$\\nThank you very much for your reply. Upon carefully reviewing our code, we discovered an error in the iteration variable used during the CG iteration. Specifically, the original code was incorrectly updating the variable $x$ instead of $v$. After correcting this by updating $v$ instead of $x$, our results aligned perfectly with yours. We appreciate your valuable feedback.\", \"below_are_the_updated_performance_metrics\": \"| Time (s) for $\\\\parallel \\\\mathbf{x} - \\\\mathbf{x}^*\\\\parallel / \\\\parallel \\\\mathbf{x}^*\\\\parallel $ | 0.5 | 1.0 | 1.5 |\\n|---------------------------------------------------------------|----------|----------|----------|\\n| **qNBO (SR1)** | 1.8E-6 | 4.0E-7 | 4.3E-7 |\\n| **qNBO (BFGS)** | 5.6E-3 | 2.8E-4 | 4.0E-5 |\\n| **AID-BIO** | 1.0E-4 | 1.0E-6 | 1.0E-6 |\\n| **AMIGO-CG** | 1.7E-4 | 1.0E-6 | 1.0E-6 |\\n\\n| Time (s) for $\\\\parallel \\\\mathbf{y} - \\\\mathbf{y}^*\\\\parallel / \\\\parallel \\\\mathbf{y}^*\\\\parallel $ | 0.5 | 1.0 | 1.5 |\\n|---------------------------------------------------------------|-----------|-----------|-----------|\\n| **qNBO (SR1)** | 1.7E-6 | 7.2E-7 | 8.5E-7 |\\n| **qNBO (BFGS)** | 7.9E-3 | 3.0E-4 | 6.3E-5 |\\n| **AID-BIO** | 2.5E-4 | 1.0E-6 | 1.0E-6 |\\n| **AMIGO-CG** | 4.1E-4 | 1.0E-6 | 1.0E-6 |\\n\\n| Time (s) for $\\\\parallel d_x - \\\\nabla \\\\Phi\\\\parallel $ | 0.5 | 1.0 | 1.5 |\\n|---------------------------------------------------------------|-----------|-----------|-----------|\\n| **qNBO (SR1)** | 1.4E-5 | 7.2E-6 | 8.5E-6 |\\n| **qNBO (BFGS)** | 2.5E-2 | 1.2E-2 | 8.7E-3 |\\n| **AID-BIO** | 1.8E-3 | 8.0E-6 | 8.0E-6 |\\n| **AMIGO-CG** | 2.9E-3 | 8.0E-6 | 8.0E-6 |\\n\\nWe have updated the paper to incorporate these revised results. For more details, please refer to Figure 1 in the main text and the details on experiments section in Appendix C of the revised paper.\"}", "{\"title\": \"Response to concerns raised in Weaknesses.\", \"comment\": \">Although quasi-Newton methods are used to estimate the inverse Hessian-vector product, the proposed algorithms still require computing Jacobian-vector products, which can be computationally expensive in large-scale cases. Could this computation pose a potential problem?\\n\\n$\\\\textbf{Response:}$ \\nThank you for the interesting question. Apart from the computation of Jacobian-vector products, all other operations in qNBO involve inner products and vector summations. Consequently, when Jacobian-vector products are not computationally expensive, qNBO can significantly reduce computational costs.\\n\\nModern automatic differentiation frameworks, such as PyTorch and JAX, make the computational complexity of Jacobian-vector products comparable to that of gradient computation. Alternatively, recent methods in the bilevel optimization literature offer different strategies for estimating Jacobian-vector products. For instance, PZOBO (Sow et al., 2022b) computes these products using differences between two optimization-based trajectories, while FdeHBO (Yang et al., 2023) and HJFBiO (Huang, 2024) employ finite-difference approximations. Additionally, ZDSBA (Aghasi & Ghadimi, 2024) uses Gaussian smoothing as an effective alternative.\"}", "{\"title\": \"Response to the concerns raised regarding the design of the proposed method and its stochastic adaptation (part 2/2).\", \"comment\": \">Questions: Can this method be effectively applied in a stochastic setting while maintaining comparable convergence rates and computational efficiency? The required adjustments in iteration rounds and step size suggest potential difficulties in achieving the same convergence speed as existing methods.\\n\\n$\\\\textbf{Response:}$ \\nThank you for your thoughtful questions. This work introduces a flexible algorithmic framework, qNBO, designed to improve hypergradient approximation. qNBO incorporates a generic decoupling structure, wherein all search directions are linear with respect to the objective functions. This structure builds upon existing methods such as HOAG (Pedregosa, 2016), AID-BIO (Ji et al., 2021), AMIGO (Arbel & Mairal, 2022), and SOBA/SABA (Dagr\\u00e9ou et al., 2022), enabling qNBO to extend effectively from deterministic to stochastic settings.\\n\\n\\nTo elaborate, qNBO consists of three components, with the first two employing quasi-Newton recursion schemes. A straightforward stochastic adaptation involves replacing deterministic quasi-Newton recursion schemes with stochastic variants, such as K-BFGS [1] or Stochastic Block BFGS [2]. Additionally, in Part 3 of qNBO, deterministic gradients are replaced with stochastic gradients throughout the iterative process, consistent with the transition from deterministic to stochastic methods discussed by Dagr\\u00e9ou et al. (2022).\\n\\nHowever, this straightforward extension may compromise convergence rates and computational efficiency under the same smoothness assumptions as in the deterministic setting. For an intuitive comparison (without using quasi-Newton methods), refer to the analysis of SOBA (Dagr\\u00e9ou et al., 2022) and its improvement by MA-SOBA (Chen et al., 2023, Optimal Algorithms for Stochastic Bilevel Optimization under\\nRelaxed Smoothness Conditions).\\n\\nAddressing your question about maintaining comparable convergence rates and computational efficiency in a stochastic setting, we consider this a promising research direction. The main challenges in achieving this include:\\n\\n$\\\\textbf{Constructing effective estimators:}$ How can we develop unbiased or biased estimators in Part 3 of qNBO by integrating techniques such as variance reduction or momentum (e.g., moving averages)? Potential candidates for these estimators include SGD, SAGA, SARAH, STORM, PAGE, and even Adam. However, analyzing this extension, particularly its convergence properties, requires significant effort. For recent progress (without using quasi-Newton methods), see SRBA (Dagr\\u00e9ou et al., 2024) and SPABA (Chu et al., 2024).\\n\\n$\\\\textbf{Analyzing convergence rates and complexity:}$ How can we evaluate the proposed stochastic algorithms in a bilevel setting while incorporating noisy second-order (curvature) information?Addressing these difficulties may require theoretical breakthroughs that go beyond existing first-order techniques, leaving this an unresolved challenge and an interesting future extension.\\n\\n\\n[1] Goldfarb, D., Ren, Y., & Bahamou, A. (2020). Practical quasi-newton methods for training deep neural networks. \\n\\n[2] Gower, R., Goldfarb, D., & Richt\\u00e1rik, P. (2016). Stochastic block BFGS: Squeezing more curvature out of data.\"}", "{\"title\": \"Response to concerns raised in Questions (Part 2/2).\", \"comment\": \">The initial matrix $H_0$ plays a significant role in quasi-Newton methods. Here, a scalar multiple of the identity matrix is used. Could the authors provide specific results or theoretical analysis on how different choices of $H_0$ might affect QNBO\\u2019s convergence rate and computational efficiency?\\n\\n\\n$\\\\textbf{Response:}$ \\nThank you for your thoughtful questions. As discussed in Chapter 6 of [1], the initial matrix $H_0$ is often set as a scalar multiple of the identity matrix, but there is no good general strategy for selecting this scalar. We have conducted experiments to evaluate the impact of the scalar multiple of $H_0$ on the performance of qNBO algorithms in toy and data hyper-cleaning experiments. The results indicate the following: (1) different choices of the scalar multiple for $H_0$ do not affect the convergence rate or computational efficiency of qNBO (SR1) in the toy experiment; and (2) for qNBO (BFGS), the scalar multiple of $H_0$ should not be too small in the data hyper-cleaning experiment. These findings are presented in Figures 1 and 2 of a single-page PDF (containing only figures and available at the anonymous link: https://drive.google.com/file/d/1rKXLVTyE-_iSna_8xBpSknMCSYmqMmlQ/view?usp=sharing).\\n\\nIn theory, for the convergence result in the quadratic case, the condition $H_0 = LI$ should be explicitly stated in the theorem\\u2019s description (see Theorem 3.4 in the revised paper). For the general case, we have added a new theorem for qNBO (BFGS) in the Appendix (Theorem D.25 of the revised paper) to justify the selection of $H_0 = LI$. The proof of this theorem builds on the local convergence result of Rodomanov and Nesterov (2021b) (see Lemma D.3 of the revised paper).\\n\\n[1]Nocedal, Jorge, and Stephen J. Wright, Numerical optimization. New York, NY: Springer New York, 1999.\"}", "{\"title\": \"We greatly appreciate your thorough and valuable feedback. Below is our response to the concerns raised in your new comments (part 2/3).\", \"comment\": \"> The reason why a single step was sufficient is probably because the hessian is very-well conditionned\\n\\n\\n$\\\\textbf{Response:}$ \\nFollowing your suggestion, we calculated the condition number of matrix $A$ and found it to be 1.98. To further investigate the impact of the condition number on the number of CG convergence steps, we selected condition numbers for $A$ as 10, 50, and 100. The results are as follows:\\n\\n$\\\\kappa = 10$\\n\\n| Time (s) for $\\\\parallel x - x^*\\\\parallel / \\\\parallel x^*\\\\parallel$ | 0.5 | 1.0 | 1.5 |\\n|----------------------------------------|----------|----------|----------|\\n| **AID-BIO(1)** | 2.6E-4 | 1.0E-6 | 1.0E-6 |\\n| **AID-BIO(5)** | 4E-2 | 7.4E-4 | 1.7E-6 |\\n\\n$\\\\kappa = 50$\\n| Time (s) for $\\\\parallel x - x^* \\\\parallel / \\\\parallel x^*\\\\parallel$ | 0.5 | 1.0 | 1.5 |\\n|----------------------------------------|----------|----------|----------|\\n| **AID-BIO(1)** | 1.2E-2 | 9.7E-4 | 7.8E-5 |\\n| **AID-BIO(5)** | 3.2E-1 | 7.2E-2 | 8.3E-2 |\\n\\n\\n\\n$\\\\kappa = 100$\\n| Time (s) for $\\\\parallel x - x^*\\\\parallel / \\\\parallel x^*\\\\parallel$ | 0.5 | 1.0 | 1.5 |\\n|----------------------------------------|----------|----------|----------|\\n| **AID-BIO(1)** | 3E-3 | 7E-6 | 1.0E-6 |\\n| **AID-BIO(5)** | 3E-2 | 7.2E-4 | 6.1E-5 |\\n\\n\\nFrom the analysis above, it is evident that the condition number significantly impacts the efficiency of the algorithm. However, in most cases, using $P=1$ for the CG iteration remains the optimal choice.\\nYour conjecture also inspired us to further explore comparisons between AID-BIO and qNBO (SR1) under larger condition numbers. The results indicate that qNBO (SR1) has a distinct advantage.\\n\\nTime (s) for $\\\\parallel d_x - \\\\nabla \\\\Phi\\\\parallel $ ($\\\\kappa = 200$ ):\\n\\n| Method | 0.5 | 1.0 | 1.5 |\\n|-------------------|--------|--------|--------|\\n| **AID-BIO(1)** | 2.8 | 2.6 | 2.6 |\\n| **AID-BIO(5)** | 2.7 | 2.6 | 2.6 |\\n| **qNBO (SR1)** | 1.2E-2 | 7.8E-3 | 7.8E-3 |\\n\\nPlease refer to the link for the corresponding image results. https://drive.google.com/file/d/11KDfI1dcpDmYpf6K1C12OKteC41NuPz5/view?usp=sharing\", \"the_following_is_the_hyperparameter_selection_of_the_algorithm\": \"$\\\\textbf{AID-BIO/AMIGO-CG:}$ The CG iterations is $N = 1$, the inner gradient iteration is $T = 1$, the inner step size is $\\\\beta= 0.01$ and $\\\\textbf{the outer step size is $\\\\alpha = 0.01$}$.\\n\\n$\\\\textbf{qNBO (SR1):}$ The inner iterations are $T = 14$, $P = 1$, $Q_k = 25$, the inner step sizes are $\\\\beta = 0.01$, $\\\\gamma = 1$, the initial matrix is $H_0 = I$, and the outer step size is $\\\\alpha = 0.5$.\\n\\nWe have updated all the codes and included them in the supplementary material. If you are interested, please feel free to download and review them. For toy experiments, refer to ICLR2025\\\\qNBO\\\\qNBO\\\\toy\\\\README.md. For logistic regression experiments, see ICLR2025\\\\qNBO\\\\qNBO\\\\logisticregression\\\\README.md. For data hyper-cleaning experiments, refer to ICLR2025\\\\qNBO\\\\qNBO\\\\Dataclean\\\\README.md.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}", "{\"title\": \"Thank you for your constructive feedback and for increasing the score. We have invested significant effort in conducting the meta-learning experiment. Below is our detailed response (part 2/2).\", \"comment\": \">Comment 2: Effect of conditioning\\n\\n$\\\\textbf{Response:}$\\nIn the experiment with a condition number $\\\\kappa = 200$, we conducted a grid search for AID-BIO. The search range for the outer step size was [0.01, 0.02, 0.03, 0.04, 0.05, 0.1, 0.2, 0.3, 0.4, 0.5]. For the inner gradient iteration number $T$ and the conjugate gradient (CG) iteration number $N$, the search range was [1, 5, 10].\\n\\nCompared to other hyperparameter selections, $T = N = 1$ may exhibit a slight disadvantage in speed when the ordinate is $|| d_x - \\\\nabla \\\\Phi ||$ or $|| y - y^* || / || y^* ||$. However, it demonstrates significant advantages in both accuracy and speed when the ordinate is $|| x - x^* || / || x^* ||$. Below are some of the results:\\n| Time (s) for $\\\\parallel x - x^*\\\\parallel / \\\\parallel x^*\\\\parallel$ | 0.5 | 1.0 | 1.5 |\\n|-------------------------------------------------------------------|----------|----------|----------|\\n| **T=N=1** | 2.6E-1 | 1.4E-2 | 2.1E-3 |\\n| **T=1,N=10** | 2.8E-1 | 1.4E-2 | 1.0E-2 |\\n| **T=10,N=1** | 4.8E-1 | 2.98E-1 | 1.9E-1 |\\n| **T=10,N=10** | 1.9E-1 | 1.9E-1 | 1.9E-1 |\\n| **qNBO(SR1)** | 8.7E-3 | 8.7E-3 | 8.7E-3 |\\n\\n| Time (s) for $\\\\parallel y - y^*\\\\parallel / \\\\parallel y^*\\\\parallel$ | 0.5 | 1.0 | 1.5 |\\n|-------------------------------------------------------------------|----------|----------|----------|\\n| **T=N=1** | 7.6E-1 | 6.8E-1 | 7.0E-1 |\\n| **T=1,N=10** | 9.5E-1 | 7.1E-1 | 7.0E-1 |\\n| **T=10,N=1** | 7.8E-1 | 6.8E-1 | 7.0E-1 |\\n| **T=10,N=10** | 6.8E-1 | 6.8E-1 | 6.8E-1 |\\n| **qNBO(SR1)** | 1.1E-2 | 1.1E-2 | 1.1E-2 |\\n\\n| Time (s) for $\\\\|\\\\| d_x-\\\\nabla \\\\Phi \\\\|\\\\|$ | 0.5 | 1.0 | 1.5 |\\n|-------------------------------------------------------------------|----------|----------|----------|\\n| **T=N=1** | 4.6E+00 | 2.7E+00 | 2.8E+00 |\\n| **T=1,N=10** | 5.7E+00 | 3.1E+00 | 3.2E+00 |\\n| **T=10,N=1** | 5.5E+00 | 2.6E+00 | 2.6E+00 |\\n| **T=10,N=10** | 2.6E+00 | 2.6E+00 | 2.6E+00 |\\n| **qNBO(SR1)** | 1.7E-1 | 1.7E-1 | 1.7E-1 |\\n\\nTherefore, overall, $T = N = 1$ is the optimal strategy. Based on your suggestion, we conducted an additional experiment with $T = N = 14$ under the condition number $\\\\kappa = 200$. The results are presented below:\\n| Time (s) for $\\\\parallel x - x^*\\\\parallel / \\\\parallel x^*\\\\parallel$ | 0.5 | 1.0 | 1.5 |\\n|-------------------------------------------------------------------|----------|----------|----------|\\n| **T=N=1** | 2.6E-1 | 1.4E-2 | 2.1E-3 |\\n| **T=N=14** | 1.9E-1 | 1.9E-1 | 1.9E-1 |\\n| **qNBO(SR1)** | 8.7E-3 | 8.7E-3 | 8.7E-3 |\\n\\n| Time (s) for $\\\\parallel y - y^*\\\\parallel / \\\\parallel y^*\\\\parallel$ | 0.5 | 1.0 | 1.5 |\\n|-------------------------------------------------------------------|----------|----------|----------|\\n| **T=N=1** | 7.6E-1 | 6.8E-1 | 7.0E-1 |\\n| **T=N=14** | 6.8E-1 | 6.8E-1 | 6.8E-1 |\\n| **qNBO(SR1)** | 1.1E-2 | 1.1E-2 | 1.1E-2 |\\n\\n\\n| Time (s) for $\\\\|\\\\| d_x-\\\\nabla \\\\Phi \\\\|\\\\|$ | 0.5 | 1.0 | 1.5 |\\n|----------------------------------------|----------|----------|----------|\\n| **T=N=1** | 4.6E+00 | 2.7E+00 | 2.8E+00 |\\n| **T=N=14** | 2.8E+00 | 2.7E+00 | 2.7E+00 |\\n| **qNBO(SR1)** | 1.7E-1 | 1.7E-1 | 1.7E-1 |\\n\\nDue to the random matrix $A$ generated in the toy experiment, the results above slightly differ from the previous response in specific values, but the orders of magnitude remain the same. Additionally, in the previous response, the result for qNBO(SR1) with $\\\\kappa = 200$ and $\\\\|\\\\| d_x-\\\\nabla \\\\Phi \\\\|\\\\|$ at $T = 14$ was reported as 1.2E-2 and 7.8E-3. However, due to an oversight, the correct values should have been 1.7E-1 (the results in the figure are correct), as shown above. We apologize for the confusion.\"}", "{\"title\": \"Thank you for the response. However, there are still major flaws in the paper and the new response\", \"comment\": \"Thank you for the response and the additional results. While I appreciate the effort, I do not find the response convincing, in particular the new experiments with Hessian-vector product. As detailed below.\\n\\n**The implementation of SHINE is correct**\\nThat's good news, and it is good that the proposed method outperforms SHINE.\\n\\n**Impact of using Hessian-vector product using autograd**\\nWhile I agree that it makes more sense to use the autograd based implementation of the Hessian-vector product, \\nI am not convinced by the results presented in the anonymous link. In particular figure 2 using autograd Hessian. \\nUsing autograd should not change the trajectories of the iterates (although it could affect speed). \\nThis means that one doesn't suddenly need to use 50 CG steps instead of 1 just because autograd is used. \\nIn fact, I have modified the code of the toy experiment to use autograd and get a small slow-down for AID-BIO and AMIGO-CG when using the same hyper-parameters as when the full hessian is computed. \\nBoth methods are still quite fast and reach a lower error near 10^-6 within 1.5s. \\nThis is in striking contrast with the results of figure 2 which shows an error of the order of 10^-1. \\nAutograd should not result in significant slow down here, so I suspect there is an error in the implementation. \\nPlease find a code snipet for AID-BIO to use autograd in a separate comment. \\nTherefore, I maintain that the conclusions from the toy experiments, when using correct implementation, are different from those in the paper.\\nTo me this suggests this example is not a good illustration for the strength of the proposed algorithm. \\n\\n**The reason why 1 CG iteartion was sufficient**\", \"i_do_not_agree_with_the_explanation\": \"\\\"due to the closed form of the Hessian provided in the code\\\". That's not the reason why a single step was enough. It doesn't matter if one provides A directly to perform Hessian vector-product of the form A@x, or to perform autograd to compute this same vector product: this should result in the same number up to numerical precision.\\nThe reason why a single step was sufficient is probably because the hessian is very-well conditionned (conditioning number $\\\\kappa$ close to 1) which implies that it requires few CG steps to converge, since the convergence rate is in $((\\\\sqrt(\\\\kappa)-1)/(\\\\sqrt(\\\\kappa)+1))^K$, so a few iterations (small K) results in a small error already. \\n\\n\\n\\n\\n**The other experiments are not impacted by the bugs in the toy experiments**\\nAlso glad to see that this is not the case. However, all the other experiments are missing the AID-BIO/AMIGO-CG which are the strongest baseline according to the toy experiments. \\nWhy are some algorithms missing in those comparisons? Why is the meta-learning experiment compares only with PZOBO? \\nThe experiments might not be impacted by the bug, but there are still not very convincing due to the missing baselines. \\n\\n**Analysis**\\nBesides the numerical experiments, there is still an unsatisfactory aspect in the theoretical analysis which is not addressed by the authors: \\nThe hyper-gradient do not seem to benefit from such super-linear rate, in fact the analysis suggests a very slow rate of 1/Q where Q is the number of quasi-Newton updates, which is even worse than gradient descent. \\nAs a result, the final complexity has a degraded dependence in the gradient queries to the lower loss. It is unclear how this degraded complexity is compensated by the improved dependence on the conditioning for vector-jacobian products.\"}" ] }

[ "Yingbo Zhou", "Yutong Ye", "Pengyu Zhang", "Xiao Du", "Mingsong Chen" ]

[ "Inherent Dependencies", "Interactive Components", "Time Series Forecasting" ]

{ "note_id": [ "xk4QSKUeNC", "vvRoTS10Sz", "rwmLVQ7puE", "p2EkIXXL3l", "mW9a9AuZIQ", "kgnqoPZLig", "iVlj3jVa5a", "heb1wiZ4UC", "h87UBjW4nO", "eUi2mydxVt", "eTMt53BYdi", "aa4UpflMhW", "Xxb0H7tcWX", "XlVVn8hKrR", "WQLtJqjWxq", "UprJnSLSm0", "UXiCTxyXxq", "ThS9NxjTxH", "QyLj1Mj4qk", "PXOVdgd2ti", "OZvc4uSpQc", "JOXfN2Su36", "DF0W0TWSTn", "A3Xaz65Bqh", "9ysdVlEK71", "4uPPJmRYQn", "4sYTTu4jiP", "0WhAKhv4zP" ], "note_type": [ "official_comment", "official_review", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "meta_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "decision" ], "note_created": [ 1732588089145, 1730972434399, 1730673590932, 1732506085552, 1732206271309, 1733034684773, 1732950603586, 1732200080822, 1731267119588, 1732179457939, 1732506103398, 1730806070671, 1733004692742, 1732506065559, 1732194525428, 1732653201735, 1732583243074, 1732950485654, 1730577148308, 1732194476954, 1732950402003, 1734676691605, 1732200034177, 1732206897701, 1732950446596, 1732684163247, 1732216502651, 1737523737050 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission5979/Authors" ], [ "ICLR.cc/2025/Conference/Submission5979/Reviewer_eGh2" ], [ "ICLR.cc/2025/Conference/Submission5979/Reviewer_ioRG" ], [ "ICLR.cc/2025/Conference/Submission5979/Authors" ], [ "ICLR.cc/2025/Conference/Submission5979/Authors" ], [ "ICLR.cc/2025/Conference/Submission5979/Authors" ], [ "ICLR.cc/2025/Conference/Submission5979/Authors" ], [ "ICLR.cc/2025/Conference/Submission5979/Authors" ], [ "ICLR.cc/2025/Conference/Submission5979/Reviewer_WMfY" ], [ "ICLR.cc/2025/Conference/Submission5979/Authors" ], [ "ICLR.cc/2025/Conference/Submission5979/Authors" ], [ "ICLR.cc/2025/Conference/Submission5979/Reviewer_AH3t" ], [ "ICLR.cc/2025/Conference/Submission5979/Reviewer_ioRG" ], [ "ICLR.cc/2025/Conference/Submission5979/Authors" ], [ "ICLR.cc/2025/Conference/Submission5979/Authors" ], [ "ICLR.cc/2025/Conference/Submission5979/Reviewer_WMfY" ], [ "ICLR.cc/2025/Conference/Submission5979/Reviewer_pkJh" ], [ "ICLR.cc/2025/Conference/Submission5979/Authors" ], [ "ICLR.cc/2025/Conference/Submission5979/Reviewer_pkJh" ], [ "ICLR.cc/2025/Conference/Submission5979/Authors" ], [ "ICLR.cc/2025/Conference/Submission5979/Authors" ], [ "ICLR.cc/2025/Conference/Submission5979/Area_Chair_vkrf" ], [ "ICLR.cc/2025/Conference/Submission5979/Authors" ], [ "ICLR.cc/2025/Conference/Submission5979/Authors" ], [ "ICLR.cc/2025/Conference/Submission5979/Authors" ], [ "ICLR.cc/2025/Conference/Submission5979/Authors" ], [ "ICLR.cc/2025/Conference/Submission5979/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ] ], "structured_content_str": [ "{\"comment\": \"Once again, we sincerely thank you for your prompt response.\\n\\nIn addition to the above statements, we provide more visual results in __Figures 9-12 of the revised manuscript__ to help you understand the seasonal and trend components learned by our method.\\n\\nWe hope these visualizations will further enhance your understanding and appreciation of our work.\"}", "{\"summary\": \"This paper introduces a Transformer-based model (TwinsFormer) to improve time series forecasting by capturing long-term dependencies. The key point of the model is to enable the interaction of decomposed time series components, which is realized by employing a dual-stream approach with an attention module and a feed-forward network to strengthen dependencies between trend and seasonal parts. Experiments demonstrate that TwinsFormer outperforms previous state-of-the-art methods in long-term and short-term tasks. Further analysis reveals that the interactive strategy can be a plug-and-play module to existing Transformer-based methods.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": [\"The motivations are well-summarized and the paper's writing is easy to follow.\", \"There is a good summary of the differences from previous works, such as the decomposition of observed time series rather than the time series embeddings, which I think is reasonable.\"], \"weaknesses\": [\"This work is not innovative in the decomposition, which I think is worth further exploration. The claim \\\"TwinsFormer is the first attempt to consider interactions between decomposed components on Transformer for time series forecasting\\\" can be overstated (e.g., TimeMixer).\", \"How about the results of Table 1 if using a longer lookback length? I think the lookback window of lookback=96 is too short to obtain the robustly decomposed trend and seasonal components. These results would greatly influence my evaluation of this work.\", \"Whether the ablation in Table 4 can further improve the performance on more recent Transformer-based forecasters, for example, PatchTST, iTransformer, and Crossformer.\", \"Does the model adopt channel independence? Are the interactions between decomposed components carried out independently within the channel or among multiple channels?\", \"The efficient version of TwinsFormer is trained with 20% variates and prediction for all variates, which is the same as iTransformer. I don't think it is an innovative part to place these experiments. Similarly, other model analyses basically follow the previous work (such as TimeMixer), so I think Sections 4.2 and 4.3 lack originality and do not provide enough insight to readers. It should undergo a major revision to delve more into the perspective of time series decomposition.\"], \"questions\": [\"It seems that the model only decomposed input series at the beginning. Could you verify that the subsequent TwinsBlock can still adequately deal with trend and seasonal components? Does it matter if you randomly swap the Treand-Seasonal input for a TwinsBlock in the middle of the model?\", \"Can the author provide more concrete showcases: How does Twinsformer cope with the intricate interactions between decomposed components to precisely unravel the inherent dependencies as mentioned in Figure 2?\", \"Why does Figure 1 compute MAE instead of MSE?\"], \"updated_after_rebuttal\": \"Thanks for the responses. My concern is partially addressed, and I have adjusted my score to 5.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The paper proposed the TwinsFormer architecture for time series forecasting. TwinsFormer decomposes the observed time series into trend and seasonal components and designs a dual-branch Transformer architecture that models their interaction at each layer. For the seasonal branch, the author applies an FFN + attention architecture that is similar to iTransformer. The only difference is that author subtracts the attention output from the input embeddings to capture the seasonal variations. For the trend branch, the autor adopts a convolution-based structure called Interactive Module (IM). The author compared TwinsFormer with a few baselines on 9 datasets and also conducted ablation study on the design choices. Experiments show that TwinsFormer consistently outperforms baselines. In addition, the author applied the dual-branch design in other transformer-based time-series forecasting models and observed consistent performance boost.\", \"soundness\": \"3\", \"presentation\": \"4\", \"contribution\": \"2\", \"strengths\": \"The paper is well-written and it is easy to understand the design and motivation of TwinTransformer. The author also conducted experiments and ablation studies on a variety of datasets to verify the conclusions.\", \"weaknesses\": \"The contribution is marginal because several previous papers (e.g., DESTformer: A Transformer Based on Explicit Seasonal\\u2013Trend Decomposition for Long-Term Series Forecasting, https://www.mdpi.com/2076-3417/13/18/10505) have tried to decompose the Transformer into trend and seasonal component and fuse the encoded features at each layer. Actually, if we closely insepct Table 3, Variant #2 performs similarly as TwinsFormer. This may indicate that part of the dual-branch design (e.g., the choice of using Interactive Module IM) is not that beneficial compared with the high-level choice of separately modeling the seasonal and trend features. In addition, the performance boost of TwinsFormer may come from adopting iTransformer in the seasonal branch. We can compare the iTransformer column and TwinsTransformer column in Table 1, 2, 6, 7 and notice that iTransformer performs well in most benchmarks.\", \"questions\": \"What's the variance of the performance numbers in Table 3? Are they really significant?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Dear Reviewer ioRG,\\n\\nSorry to bother you again.\\n\\nGiven the importance of timely communication and the rebuttal phase nearing the end, we would like to know if we have addressed your concerns. If you have any remaining questions, please let us know. We are looking forward to your valuable reply.\\n\\nThank you for your efforts in our paper.\\n \\nBest regards,\\n\\nAuthors\"}", "{\"title\": \"Response to Reviewer ioRG (Part 1)\", \"comment\": \"Thank you for your time in reviewing our paper and the detailed comments. The followings are our responses to your concerns and questions:\\n\\n__W1__: Difference to DESTformer. \\n\\nDESTformer is an interesting work that fuses the encoded features of decomposed components at each layer of the model via the addition operation. However, our method and DESTformer are completely different. 1) Since we encode the historical time series of each variate into an embedding, our seasonal-trend decomposition acts in the observed time series space rather than the embedding space of DESTformer. 2) DESTformer is designed to alleviate the information utilization bottleneck of existing decomposition-based transformers, while TwinsFormer is the first attempt to explore the interactions between decomposed components to break through the limitation of the seasonal-trend decomposition.\\n\\n__W2__: Different performance to iTransformer.\\n\\nIn model variate \\u2461 of Table 3, we feed trend components to the seasonal branch and feed seasonal components to the trend branch. \\nThe similar performance to TwinsFormer only indicates that swapping inputs for different branches has relatively little impact on model performance.\\nTo further illustrate the role of our interactive module in our framework, we disable our interactive module in the trend branch as the model variate \\u2467.\\nSpecifically, we feed the seasonal components obtained by the moving average kernel to the seasonal branch and treat the trend components obtained by the moving average kernel as the output of the trend branch.\\nThe results of \\u2467 on ECL, Traffic, PEMS03, and PEMS07 datasets are listed below:\\n| TwinsFormer | ECL | Traffic | PEMS03 | PEMS07 |\\n|:-------:|:------------:|:------------:|:------------:|:------------:|\\n| 96 | 0.139\\\\|0.233 | 0.382\\\\|0.260 | 0.065\\\\|0.169 | 0.060\\\\|0.158 |\\n| 192 | 0.158\\\\|0.252 | 0.392\\\\|0.267 | 0.086\\\\|0.196 | 0.079\\\\|0.181 |\\n| 336 | 0.172\\\\|0.267 | 0.410\\\\|0.276 | 0.121\\\\|0.234 | 0.104\\\\|0.209 |\\n| 720 | 0.200\\\\|0.293 | 0.442\\\\|0.292 | 0.165\\\\|0.276 | 0.132\\\\|0.236 |\\n| avg | 0.167\\\\|0.262 | 0.406\\\\|0.273 | 0.109\\\\|0.219 | 0.094\\\\|0.196 |\\n\\n| \\u2467 | ECL | Traffic | PEMS03 | PEMS07 |\\n|:-------:|:------------:|:------------:|:------------:|:------------:|\\n| 96 | 0.156\\\\|0.248 | 0.396\\\\|0.268 | 0.071\\\\|0.177 | 0.062\\\\|0.162 |\\n| 192 | 0.172\\\\|0.264 | 0.412\\\\|0.275 | 0.099\\\\|0.211 | 0.085\\\\|0.191 |\\n| 336 | 0.188\\\\|0.278 | 0.423\\\\|0.292 | 0.169\\\\|0.282 | 0.170\\\\|0.283 |\\n| 720 | 0.225\\\\|0.325 | 0.449\\\\|0.308 | 0.197\\\\|0.306 | 0.197\\\\|0.301 |\\n| avg | 0.185\\\\|0.279 | 0.420\\\\|0.286 | 0.134\\\\|0.244 | 0.129\\\\|0.234 |\\n\\nWithout the interactive module, the performance degradations of model variate \\u2467 are obvious, even worse than iTransformer.\\nThese degradations indicate that despite iTransformer providing a good backbone for our approach, our interactive module plays a key role in further improving forecasting performance.\\n\\n__Q1__: Considering that an untrainable linear transformation (i.e., the moving average kernel) cannot reflect the non-linear pattern of the observed time series, we propose a Transformer-based interaction framework to better capture the intrinsic dependencies for time series forecasting. \\n__Since we utilize the model to learn the interactions between decomposed components for time series forecasting, it is significant to analyze the impact of the inputs and operations of key modules on the model performance__.\\nThe results in Table 3 are obtained with the same fixed random seed, which means the result is the same in each run under the same hyperparameters.\\nSince the variance is too small, we compute the standard deviation of ablation studies with five random seeds as below:\"}", "{\"comment\": \"Anyway, thank you for your prompt reply.\\n\\nIn this paper, we __delve into__ the trend-seasonal decomposition and find that the seasonal and trend components obtained by the trend-seasonal decomposition __cannot accurately capture__ the inherent dependencies of the time series due to the __untrainable linear transformation__ (i.e., the moving average kernel). \\n\\nTherefore, we propose a __first Transformer-based framework__ that explicitly explores inherent dependencies by __learning implicit and progressive interactions__ between __different components__ for time series forecasting.\\n\\nConcretely, we design an interactive module around the attention module and the feed-forward network to __learn more robust and reliable decomposed representations__ based on residual and interactive learning.\\n\\nTechnically, we __only utilize__ the trend-seasonal decomposition to __initialize__ our seasonal and trend components, and our dual-stream interaction framework can enhance the __transformation and decoupling__ of seasonal and trend.\\n\\nTo better understand the learning process of the seasonal and trend components, we provide more visualization results in __Figures 10-12 of the revised manuscript__.\\n\\nExperimentally, TwinsFormer achieves __state-of-the-art performances__ in both long-term and short-term forecasting tasks, and is __plug-and-play__ into existing Transformer-based methods with __negligible extra computational overhead__.\\n\\nWe understand your concern regarding the novelty of our method. To better address this issue, we would greatly appreciate it if you could provide more detailed feedback on the specific aspects of our approach that you find lacking in novelty. Your insights would help us understand how we can further enhance the originality and impact of our work.\\n\\nWe look forward to your further feedback and are committed to addressing any concerns raised by you.\"}", "{\"comment\": \"Dear Reviewer pkJh,\\n\\nAs the author/reviewer discussion will close soon, we would like to know if our response has addressed your concerns and questions. If you have any further concerns or suggestions for the paper or our rebuttal, please let us know. We would be happy to engage in further discussion.\\n\\nThank you again for the time and effort you dedicated to reviewing this work.\\n\\nSincerely,\\n\\nAuthors\"}", "{\"title\": \"Response to Reviewer AH3t (Part 2)\", \"comment\": \"| ETTm1 | TwinsFormer | PatchTST | DLinear | FEDformer |\\n|-------|:-------------:|:-------------:|:-------------:|:-------------:|\\n| 96 | 0.027\\\\|0.126 | 0.026\\\\|0.123 | 0.028\\\\|0.123 | 0.033\\\\|0.140 |\\n| 192 | 0.042\\\\|0.154 | 0.040\\\\|0.151 | 0.045\\\\|0.156 | 0.058\\\\|0.186 |\\n| 336 | 0.055\\\\|0.180 | 0.053\\\\|0.174 | 0.061\\\\|0.182 | 0.084\\\\|0.231 |\\n| 720 | 0.076\\\\|0.209 | 0.073\\\\|0.206 | 0.080\\\\|0.210 | 0.102\\\\|0.250 |\\n| avg | 0.050\\\\|0.167 | 0.048\\\\|0.164 | 0.054\\\\|0.168 | 0.069\\\\|0.202 |\\n\\n| ETTm2 | TwinsFormer | PatchTST | DLinear | FEDformer |\\n|-------|:-------------:|:-------------:|:-------------:|:-------------:|\\n| 96 | 0.064\\\\|0.185 | 0.065\\\\|0.187 | 0.063\\\\|0.183 | 0.067\\\\|0.198 |\\n| 192 | 0.095\\\\|0.232 | 0.093\\\\|0.231 | 0.092\\\\|0.227 | 0.102\\\\|0.245 |\\n| 336 | 0.123\\\\|0.265 | 0.121\\\\|0.266 | 0.119\\\\|0.261 | 0.130\\\\|0.279 |\\n| 720 | 0.170\\\\|0.317 | 0.172\\\\|0.322 | 0.175\\\\|0.320 | 0.178\\\\|0.325 |\\n| avg | 0.113\\\\|0.250 | 0.113\\\\|0.252 | 0.112\\\\|0.248 | 0.119\\\\|0.262 |\\n\\nThese experimental results show that our method can achieve competitive performance in univariate forecasting tasks.\"}", "{\"summary\": \"The paper presents TwinsFormer, a Transformer-based framework designed for time series forecasting. The primary innovation lies in its ability to model interactions between the trend and seasonal components of time series data via an interactive dual-stream design.\", \"soundness\": \"3\", \"presentation\": \"4\", \"contribution\": \"4\", \"strengths\": [\"The proposed framework addresses the challenge of accurately capturing the intricate temporal dependencies between trend and seasonal components in time series forecasting. The dual-stream design and interaction mechanism are novel, offering an efficient way to learn the dependencies between these components while maintaining computational efficiency.\", \"The paper provides a clear and rigorous explanation of why traditional independent decomposition models (which separate trend and seasonal components) might miss crucial interactions.\", \"The authors highlight that their interactive strategy can be applied to existing Transformer-based architectures with minimal computational overhead.\", \"Extensive experiments across 13 real-world datasets, including both short-term and long-term forecasting scenarios, demonstrate the effectiveness of TwinsFormer. The paper shows that it outperforms several state-of-the-art models, such as iTransformer, TimeMixer, and Crossformer, across a variety of metrics (MSE, MAE).\"], \"weaknesses\": [\"The paper uses a simple moving average kernel for decomposing the time series into trend and seasonal components, which assumes that the trend and seasonal components are captured well by linear operations. Could more sophisticated decomposition techniques (such as wavelet or non-linear methods) potentially offer better results?\", \"The paper doesn\\u2019t provide detailed analysis regarding the sensitivity of the model\\u2019s performance to hyperparameters, especially those related to the decomposition mechanism (such as kernel size).\", \"The dual-stream design, although innovative, introduces significant architectural complexity. While the paper mentions that the model is computationally efficient, it does not provide a comprehensive analysis of the computational cost for both training and inference across different model sizes (e.g., small, medium, and large).\", \"The paper introduces a subtraction mechanism in the seasonal branch of the model to eliminate redundant information. How about other operations like concatenation or addition?\"], \"questions\": \"Check Weaknesses\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Many thanks to Reviewer WMfY for providing thorough and insightful comments. The following are our responses to your specific questions and concerns.\\n\\n__W1__: Thanks for your scientific rigor. We compare the performance (__MSE|MAE__) of the moving average kernel (__MAK__) with Wavelet (__Wav__) and Non-linear (__Non__) decomposition designs on three benchmarks below:\\n| MAK | Weather | ECL | Traffic |\\n|:--------:|:-------------:|:-------------:|:-------------:|\\n| 96 | 0.161\\\\|0.201 | 0.139\\\\|0.233 | 0.382\\\\|0.260 |\\n| 192 | 0.211\\\\|0.248 | 0.158\\\\|0.252 | 0.392\\\\|0.267 |\\n| 336 | 0.266\\\\|0.291 | 0.172\\\\|0.267 | 0.410\\\\|0.276 |\\n| 720 | 0.347\\\\|0.343 | 0.200\\\\|0.293 | 0.442\\\\|0.292 |\\n| avg | 0.246\\\\|0.271 | 0.167\\\\|0.262 | 0.406\\\\|0.273 |\\n\\n| Wav | Weather | ECL | Traffic |\\n|:--------:|:-------------:|:-------------:|:-------------:|\\n| 96 | 0.160\\\\|0.200 | 0.138\\\\|0.232 | 0.381\\\\|0.260 |\\n| 192 | 0.210\\\\|0.248 | 0.157\\\\|0.251 | 0.390\\\\|0.265 |\\n| 336 | 0.265\\\\|0.290 | 0.171\\\\|0.266 | 0.409\\\\|0.275 |\\n| 720 | 0.345\\\\|0.345 | 0.200\\\\|0.292 | 0.438\\\\|0.290 |\\n| avg | 0.245\\\\|0.271 | 0.167\\\\|0.260 | 0.405\\\\|0.273 |\\n\\n| Non | Weather | ECL | Traffic |\\n|:--------:|:-------------:|:-------------:|:-------------:|\\n| 96 | 0.158\\\\|0.198 | 0.136\\\\|0.230 | 0.380\\\\|0.258 |\\n| 192 | 0.207\\\\|0.246 | 0.155\\\\|0.250 | 0.388\\\\|0.263 |\\n| 336 | 0.263\\\\|0.287 | 0.171\\\\|0.267 | 0.405\\\\|0.270 |\\n| 720 | 0.343\\\\|0.342 | 0.199\\\\|0.293 | 0.435\\\\|0.286 |\\n| avg | 0.243\\\\|0.268 | 0.165\\\\|0.260 | 0.402\\\\|0.269 |\\n\\nThese results show that our method has good compatibility with various decomposition techniques.\\n\\n__W2__: Thanks for your valuable comments. Since the existing decomposition designs uniformly use the kernel size of 25 by default, we ignore the decomposition mechanism's hyperparameters. We provide the results (__MSE|MAE__) related to the kernel size (__KS__) and embedding dimension (__Dim__) on ECL and Traffic below:\\n| KS | ECL | Traffic |\\n|:--:|:-------------:|:-------------:|\\n| 5 | 0.139\\\\|0.234 | 0.385\\\\|0.263 |\\n| 15 | 0.138\\\\|0.233 | 0.382\\\\|0.261 |\\n| 25 | 0.139\\\\|0.233 | 0.382\\\\|0.260 |\\n| 35 | 0.138\\\\|0.233 | 0.383\\\\|0.261 |\\n| 45 | 0.139\\\\|0.233 | 0.383\\\\|0.261 |\\n| 55 | 0.140\\\\|0.234 | 0.381\\\\|0.260 |\\n| 65 | 0.139\\\\|0.233 | 0.385\\\\|0.262 |\\n| 75 | 0.140\\\\|0.234 | 0.383\\\\|0.261 |\\n\\n| Dim | ECL | Traffic |\\n|:----:|:-------------:|:-------------:|\\n| 128 | 0.152\\\\|0.246 | 0.412\\\\|0.284 |\\n| 256 | 0.142\\\\|0.237 | 0.394\\\\|0.268 |\\n| 512 | 0.139\\\\|0.233 | 0.382\\\\|0.260 |\\n| 1024 | 0.139\\\\|0.233 | 0.381\\\\|0.260 |\\n\\n__W3__: To comprehensively analyze the computational cost, we treat TwinsFormer with __20%__, __60%__, and __100%__ variates as small, medium, and large model sizes, respectively.\\n| Models | Training Times (s/Epoch) | Inference Times (s) | GPU (GB) | Paramenter (MB) | FLOPs (GB) |\\n|:------------------:|:------------------------------:|:-------------------------:|:---------------------:|:---------------:|:----------:|\\n| iTransformer | 52.25 | 14.28 | 7.50 | 6.11 | 8.63 |\\n| PatchTST | 175.36 | 22.45 | 9.95 | 3.58 | 33.52 |\\n| Crossformer | 283.47 | 66.19 | 11.77 | 46.90 | 229.09 |\\n| Ours (100%) | 84.32 | 20.23 | 8.30 | 8.46 | 30.46 |\\n| Ours (60%) | 47.71 | 15.27 | 4.17 | 5.39 | 13.21 |\\n| Ours (20%) | 11.24 | 10.33 | 1.66 | 3.21 | 2.86 |\\n\\n__We can observe that our TwinsFormer requires less memory and runs faster than other Transformer-based models.__\\n\\n__W4__: According to __Equation (1)__ of the manuscript, we can find that the seasonal components are obtained by subtracting the trend components from the original time series in the trend-seasonal decomposition. \\nTo better learn the interactions between seasonal and trend components, we believe that the subtraction mechanism is more in line with the principle of trend-seasonal decomposition and further facilitates the decoupling of seasonal and trend components.\\nSince __the concatenation__ will change the channel dimension of the encoded features and bring additional computational overhead, existing Transformer-based models use the addition to implement the skip connection by default.\\nTherefore, Transformer-based methods do not use the concatenation to implement the skip connection.\\nIn __Table 3__ of the manuscript, the model variant \\u2462 shows the __performance__ of replacing the subtraction mechanism with __the addition__.\", \"title\": \"Response to Reviewer WMfY\"}", "{\"comment\": \"Dear Reviewer pkJh,\\n\\nSorry to bother you again.\\n\\nGiven the importance of timely communication and the rebuttal phase nearing the end, we would like to know if we have addressed your concerns. If you have any remaining questions, please let us know. We are looking forward to your valuable reply.\\n\\nThank you for your efforts in our paper.\\n \\nBest regards,\\n\\nAuthors\"}", "{\"summary\": \"This paper proposed TwinsFormer for time series forecasting that uses dual streams to interactively process trend and seasonal components, enhancing dependency learning between them. By integrating these interactions within the attention and feed-forward layers, TwinsFormer improves forecasting accuracy without adding significant computational cost.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"(1) The proposed TwinsBlock is novel and well justified by rationality analysis, ablation and visualization analysis.\\n\\n(2) The proposed TwinsFormer demonstrates better performance than baseline models in most datasets,\\n\\n(3) The paper is well-organized and easy to follow.\", \"weaknesses\": \"(1) The improvements of TwinsFormer over baselines in the Main results (Tables 1 and 2) are quite marginal. Generally, the improvement over the second-best baseline is less than **0.02** for MSE and MAE metrics.\\n\\n(2) Given the marginal improvements over baselines shown in Tables 1 and 2, it is essential to include the variance of repeated experiments for a more accurate assessment.\\n\\n(3) In Equation (9), the authors justify their design with the assumption of \\\"replacing [.] with +\\\". Given that '+' appears to be a more appropriate choice based on your analysis, why not simply sum up \\\\( X_t \\\\), \\\\( g(X_s) \\\\), and \\\\( h(X_s, g(X_s)) \\\\)? Is there any evidence to support concatenation as the optimal design choice?\\n\\n(4) It seems that all evaluations are currently conducted on multivariate forecasting tasks. Including evaluations on univariate forecasting tasks would provide a more comprehensive assessment for TwinsFormer.\", \"questions\": \"(1) Is the improvement in Tables 1 and 2 statistically significant? Could you report the confidence level using a statistical hypothesis test ,e.g. T-test?\\n\\n(2) Does the proposed TwinFormer still work well for univariate forecasting tasks?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thanks for the rebuttal. I still think the method lacks novelty and won't change the score.\"}", "{\"comment\": \"Dear Reviewer eGh2,\\n\\nSorry to bother you again.\\n\\nGiven the importance of timely communication and the rebuttal phase nearing the end, we would like to know if we have addressed your concerns. If you have any remaining questions, please let us know. We are looking forward to your valuable reply.\\n\\nThank you for your efforts in our paper.\\n \\nSincerely,\\n\\nAuthors\"}", "{\"title\": \"Response to Reviewer eGh2 (Part 2)\", \"comment\": \"__Q2__: As seen in __Figure 11-12 of the revised paper__, we provide concrete showcases on the ECL and Traffic benchmarks, clearly illustrating the distinctions between the decomposed components learned by the moving average kernel and those derived from our interaction mechanisms.\\n\\n__Q3__: In Tables 1 and 2 of the manuscript, we provide the full average performance among all the benchmarks in both MSE and MAE.\\nThese results show that TwinsFormer performs better in MAE for all the benchmarks, so we plot the MAE results instead of MSE results in Figure 1 to better __highlight the superiority__ of our method.\"}", "{\"comment\": \"Thanks for your responses. I keep my score as 8.\"}", "{\"title\": \"Thanks\", \"comment\": \"Thanks for the explanation. I have increased score.\"}", "{\"comment\": \"Dear Reviewer ioRG,\\n\\nAs the author/reviewer discussion will close soon, we would like to know if our response has addressed your concerns and questions. If you have any further concerns or suggestions for the paper or our rebuttal, please let us know. We would be happy to engage in further discussion.\\n\\nThank you again for the time and effort you dedicated to reviewing this work.\\n\\nSincerely,\\n\\nAuthors\"}", "{\"summary\": \"The paper proposes a method to improve transformer based time series forecasting. A dual stream architecture is proposed, where the multi-variate time series is decomposed into trend and seasonality components. Then the learned embedding for each of these components are then combined through an interactive module, to give the final combined representation. To my understanding, the key differentiator from existing works that do decomposition, is that, Twinsformer learns a better interaction between the decomposed components leading to a better downstream forecasting performance. Empirical performance is quite detailed on long term and short term forecasting tasks, and compared to the baselines, the performance is quite strong. Ablation studies appear comprehensive.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": [\"Originality & Significance\", \"strong empirical performance\", \"interesting idea to address interaction among different components, but not very novel\"], \"weaknesses\": [\"the motivation is not very convincing\", \"\\\"inherent dependencies\\\" are not well defined\", \"lines 79-82, motivation on channel 11 is not clear, how does this make a case for \\\"inherent dependencies\\\", or what that is?\", \"lines 191-194 - it seems the goal has changed to address the problem of limitations in \\\"multi-variate correlation\\\" which is addressed by Twinsformer - makes it further unclear what the goal is\", \"many of the design choices also do not seem to have a clear motivation, and seem heuristic, and are post-hoc justified in ablation rather than a clear reason for selection\"], \"questions\": [\"Please help clarify the motivations and the key problem being addressed through the proposed module, as it seems rather ad-hoc and heuristic\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to Reviewer eGh2 (Part 1)\", \"comment\": \"Thanks for reviewing our paper. The following are our responses to your specific questions and concerns:\\n\\n__W1__: In the manuscript, we highlight the __interactions between decomposed components (i.e., seasonal and trend components)__ rather than the new design for the decomposition.\\nSpecifically, we employ the moving average kernel to decompose the time series data into seasonal and trend components, and then design a Transformer-based dual-stream framework to make different components learn from each other and capture better inherent dependencies for the final forecasting.\\nIn addition, we supplement the experimental results of wavelet and non-linear decompositions as described by Reviewer WMfY, which illustrates the compatibility of our dual-stream framework for more decomposition designs.\\nAs an __MLP-based__ framework, __TimeMixer__ introduced the decomposable multiscale mixing module to learn the __interactions among different scale seasonal (or trend) components__ rather than the __interactions between seasonal and trend components__.\\nExisting decomposition designs treat the seasonal and trend components as two unrelated variables, ignoring the interactions between the decomposed components.\\nTherefore, __TwinsFormer is the first attempt to consider interactions between decomposed components on Transformer__ for time series forecasting. \\n\\n__W2:__: Indeed, different lookback window lengths affect the robustness of the decomposed trend and seasonal components.\\nWe __analyzed the lookback length sensitivity__ in lines 408-427 of our manuscript, and __Figure 4__ clearly illustrates that our TwinsFormer can effectively capture inherent dependencies from a longer lookback window. \\n\\n__W3__: Generally, the most important module of Transformer-based forecasters is the attention mechanism.\\nIn Table 4 of our manuscript, we mainly analyze the impact of different attention mechanisms on the performance of our framework.\\nFor __PatchTST__ and __Crossformer__, the time series fed to their attention modules is obtained by __patching and segmentation operations__, which cannot capture the seasonal and trend characteristics of the time series.\\nAs for __iTranformer__, its success lies in learning inherent dependencies from the dimension of variates rather than exploring new attention mechanisms.\\nWithout the decomposition and interactive modules, our backbone is consistent with iTransformer. \\nTherefore, we did not analyze the attention compatibility and performance promotion on PatchTST, iTransformer, and Crossformer in Table 4.\\n\\n__W4__: In our work, we adopt __channel independence__ to decompose the lookback length of each channel into seasonal and trend components, and then encode the decomposed components into corresponding embeddings.\\nWe utilize the attention mechanism in the seasonal branch to learn the multivariate correlation among different seasonal embeddings.\\nMeanwhile, we treat the residual seasonal signals obtained by the subtraction mechanism as the supervision information to perform interactive learning with the trend embeddings in the trend branch.\\nIn other words, our interactions between decomposed components among __multiple channels__.\\n\\n__W5__: Both iTransformer and TimeMixer are excellent works, and __we adopted some of their experimental formats because their approaches to model analysis are highly representative and widely recognized__ in the field.\\nMoreover, our dual-stream Transformer-based framework incorporates both decomposition design and a plug-and-play interactive module, necessitating an exploration of the individual components' roles and the generalization capabilities of our module.\\nWhile the analysis format may resemble existing methods, our primary focus is on the __unique contributions and effectiveness__ of our dual-stream Transformer-based framework. \\n\\n__Q1__: We regard the decomposed components obtained by the moving average kernel as the initial decomposed components, and then feed the encoded seasonal and trend embeddings to the TwinsBlock. \\nTo verify our TwinsBlock can still adequately deal with the trend and seasonal components, we supplement additional visualization results in __Figure 10 of the revised paper__. \\nIn Table 3 of the manuscript, we conducted the ablation study on swapping seasonal and trend components (i.e. __the model variant__ \\u2461), and the results show that __feeding the seasonal components to the attention module is more beneficial for model performance__.\\nRecall __Figure__ 2 of the manuscript, the trend components are very smooth over a certain time interval, while the seasonal components can better reflect the temporal patterns of the observed time series, which also __suggests that the seasonal components can better represent the multivariate correlations__.\\nThus, we do not need to consider randomly swapping the trend-seasonal inputs to a TwinsBlock in the middle of the model.\"}", "{\"comment\": \"Dear Reviewer eGh2,\\n\\nAs the author/reviewer discussion will close soon, we would like to know if our response has addressed your concerns and questions. If you have any further concerns or suggestions for the paper or our rebuttal, please let us know. We would be happy to engage in further discussion.\\n\\nThank you again for the time and effort you dedicated to reviewing this work.\\n\\nSincerely,\\n\\nAuthors\"}", "{\"metareview\": \"This paper introduces TwinsFormer, a Transformer-based model for time-series forecasting that emphasizes modeling interactions between seasonal and trend components. Given the limited and relatively inactive discussion, AC has carefully reconsidered the paper\\u2019s contributions.\\n\\nAC finds the primary shortcoming is the lack of in-depth analysis (beyond Figure 2) from the perspective of time series decomposition. The paper does not sufficiently explain why interactions between seasonal and trend components are crucial, nor how ignoring these interactions leads to inaccuracies in current forecasting models. As Reviewer eGh2 suggests, this can not be achieved by merely replicating analysis experiments following previous literature. This gap in motivation and clarity may also explain why several reviewer concerns remain unresolved. For instance, Reviewer pkJh found the concept of \\u201cinherent dependencies\\u201d too vague without more detailed exposition. \\n\\nIn conclusion, the performance gains are undoubtedly acknowledged. We believe that this paper can be accepted if further providing robust justification of both its results and its underlying motivations.\", \"additional_comments_on_reviewer_discussion\": \"Three reviewers (eGh2, ioRG, pkJh) questioned the novelty of the approach. Although the authors responded, they did not adequately address these concerns or convince the reviewers to recommend acceptance (the final scores are 8/6/5/5/5).\"}", "{\"title\": \"Response to Reviewer AH3t (Part 1)\", \"comment\": \"Many thanks to Reviewer AH3t for providing a detailed review and insightful questions. We address your concerns here:\\n\\n__W1__: The metrics used in time series forecasting, Mean Squared Error (MSE) and Mean Absolute Error (MAE), are highly sensitive to small changes, especially when the baseline models are already performing at a high level. \\nA difference of less than 0.02 in these metrics can make a significant improvement in practical applications, particularly in domains where even minor errors can have substantial consequences, such as financial and traffic forecasting.\\nFurthermore, All experimental results were obtained using the same random seed, meaning that running the experiments with the same parameters consistently yields identical results. \\nAdditionally, the results in Tables 1 and 2 are __the average results__ of four different forecasting lengths, and we provide the full experimental results in __Tables 6-7__ of the manuscript.\\n\\n__W2\\\\&Q1__: To ensure the fairness of the experiment, all experimental results are obtained by running under the same random seed.\\nMeanwhile, we reported the __standard deviations__ of TwinsFormer performance in __Table 8__ of the manuscript. \\nWe also provide statistical tests on ECL and all subsets of ETT and PEMS with five runs of random seeds. \\nThe performance is averaged from four prediction lengths. \\nThe standard deviations and T-test statistics (__MSE|MAE__) of TwinsFormer and iTransformer are listed below, showing that the performance is on par with the previous SOTA iTransformer within the margin of error.\\n| | ETT | ECL | PEMS |\\n|:------------:|:----------------------------------:|:----------------------------------:|:----------------------------------:|\\n| iTransformer | 0.383 $\\\\pm$ 0.001 \\\\| 0.399 $\\\\pm$ 0.002 | 0.178 $\\\\pm$ 0.001 \\\\| 0.270 $\\\\pm$ 0.001 | 0.122 $\\\\pm$ 0.002 \\\\| 0.224 $\\\\pm$ 0.001 |\\n| TwinsFormer | 0.372 $\\\\pm$ 0.001 \\\\| 0.392 $\\\\pm$ 0.001 | 0.167$\\\\pm$ 0.001 \\\\| 0.262 $\\\\pm$ 0.001 | 0.112 $\\\\pm$ 0.001 \\\\| 0.214 $\\\\pm$ 0.002 |\\n| T-test | 2.158\\\\|2.412 | -1.179\\\\|-1.196 | -0.763\\\\|-0.787 |\\n\\n__W3__: In the rationality analysis of Section 3, we formulate the model to provide a formal expression to facilitate the understanding of the practical significance of our dual-stream structure and its compatibility with the decomposition design.\\nReplacing concatenation with addition operations is intended to simplify the formula of the interactions between decomposed components rather than highlight which operation is the optimal design choice.\\nTo better illustrate the __difference between the addition and concatenation operations__, we compare the average performance (__MSE|MAE__) of the two operations on three datasets as below.\\n| Type | Weather | ECL | Traffic |\\n|:----:|:------------:|:------------:|:------------:|\\n| [] | 0.246\\\\|0.271 | 0.167\\\\|0.262 | 0.406\\\\|0.273 |\\n| + | 0.249\\\\|0.274 | 0.172\\\\|0.265 | 0.410\\\\|0.280 |\\n\\nThe results indicate that the performance of '+' is inferior to that of '[]', which may be due to the addition operation inhibiting the decoupling of seasonal and trend components.\\n\\n__W4\\\\&Q2__: The performance (__MSE|MAE__) of univariate time series forecasting on all the subsets of ETT is presented below.\\nAs iTransformer, TimeMixer, and Crossformer do not offer training hyperparameters and results for univariate forecasting tasks, we compare our method with PatchTST, DLinear, and FEDformer. \\n| ETTh1 | TwinsFormer | PatchTST | DLinear | FEDformer |\\n|-------|:-------------:|:-------------:|:-------------:|:-------------:|\\n| 96 | 0.057\\\\|0.179 | 0.059\\\\|0.189 | 0.056\\\\|0.180 | 0.079\\\\|0.215 |\\n| 192 | 0.072\\\\|0.210 | 0.074\\\\|0.215 | 0.071\\\\|0.204 | 0.104\\\\|0.245 |\\n| 336 | 0.080\\\\|0.219 | 0.076\\\\|0.220 | 0.098\\\\|0.244 | 0.119\\\\|0.270 |\\n| 720 | 0.079\\\\|0.228 | 0.087\\\\|0.236 | 0.189\\\\|0.359 | 0.142\\\\|0.299 |\\n| avg | 0.072\\\\|0.209 | 0.074\\\\|0.215 | 0.104\\\\|0.247 | 0.111\\\\|0.257 |\\n\\n| ETTh2 | TwinsFormer | PatchTST | DLinear | FEDformer |\\n|-------|:-------------:|:-------------:|:-------------:|:-------------:|\\n| 96 | 0.129\\\\|0.275 | 0.131\\\\|0.284 | 0.131\\\\|0.279 | 0.128\\\\|0.271 |\\n| 192 | 0.178\\\\|0.329 | 0.171\\\\|0.329 | 0.176\\\\|0.329 | 0.185\\\\|0.330 |\\n| 336 | 0.210\\\\|0.346 | 0.171\\\\|0.336 | 0.209\\\\|0.367 | 0.231\\\\|0.378 |\\n| 720 | 0.221\\\\|0.378 | 0.223\\\\|0.380 | 0.276\\\\|0.426 | 0.278\\\\|0.420 |\\n| avg | 0.185\\\\|0.332 | 0.174\\\\|0.332 | 0.198\\\\|0.350 | 0.206\\\\|0.350 |\"}", "{\"title\": \"Response to Reviewer ioRG (Part 2)\", \"comment\": \"| | ECL | Traffic | PEMS03 | PEMS07 |\\n|:-------:|:------------:|:------------:|:------------:|:------------:|\\n|Ours|0.167$\\\\pm$0.001\\\\|0.262$\\\\pm$0.001|0.406$\\\\pm$0.001\\\\|0.273$\\\\pm$0.002|0.109$\\\\pm$0.002\\\\|0.219$\\\\pm$0.001|0.094$\\\\pm$0.002\\\\|0.196 $\\\\pm$ 0.001|\\n\\u2460|0.176$\\\\pm$0.001\\\\|0.272$\\\\pm$0.002|0.417$\\\\pm$0.002\\\\|0.282$\\\\pm$0.001|0.116$\\\\pm$0.001\\\\|0.226$\\\\pm$0.002|0.102$\\\\pm$0.001\\\\|0.204$\\\\pm$0.002|\\n|\\u2461|0.172$\\\\pm$0.001\\\\|0.265$\\\\pm$0.001|0.413$\\\\pm$0.001 \\\\|0.277$\\\\pm$0.001|0.114$\\\\pm$0.002\\\\|0.224$\\\\pm$0.001|0.101$\\\\pm$0.002 \\\\|0.204$\\\\pm$0.002|\\n| \\u2462 |0.180$\\\\pm$0.002\\\\|0.275$\\\\pm$0.001|0.416$\\\\pm$0.002\\\\|0.283$\\\\pm$0.001|0.118$\\\\pm$0.001\\\\|0.228$\\\\pm$0.002| 0.102$\\\\pm$0.002\\\\|0.207$\\\\pm$0.001| \\n|\\u2463 |0.185$\\\\pm$0.001\\\\|0.278$\\\\pm$0.002|0.418$\\\\pm$0.001\\\\|0.283$\\\\pm$0.002|0.122$\\\\pm$0.001\\\\|0.232$\\\\pm$0.001| 0.105$\\\\pm$0.001\\\\|0.210$\\\\pm$0.002| \\n|\\u2464 |0.183$\\\\pm$0.002\\\\|0.277$\\\\pm$0.001|0.413$\\\\pm$0.001\\\\|0.278$\\\\pm$0.001|0.118$\\\\pm$0.001\\\\|0.229$\\\\pm$0.001| 0.103$\\\\pm$0.001\\\\|0.208$\\\\pm$0.001|\\n|\\u2465 |0.176$\\\\pm$0.001\\\\|0.271$\\\\pm$0.001|0.412$\\\\pm$0.001\\\\|0.281$\\\\pm$0.001|0.121$\\\\pm$0.002\\\\|0.228$\\\\pm$0.002| 0.104$\\\\pm$0.002\\\\|0.210$\\\\pm$0.001|\\n| \\u2466 | 0.176$\\\\pm$0.002\\\\|0.268$\\\\pm$0.001|0.413$\\\\pm$0.002\\\\|0.278$\\\\pm$0.001|0.118$\\\\pm$0.001\\\\|0.228$\\\\pm$0.001| 0.107$\\\\pm$0.001\\\\|0.215$\\\\pm$0.002|\\n|\\u2467|0.185$\\\\pm$0.002\\\\|0.279$\\\\pm$0.002|0.420$\\\\pm$0.002\\\\|0.286$\\\\pm$0.001|0.134$\\\\pm$0.002\\\\|0.244$\\\\pm$0.001| 0.129$\\\\pm$0.002\\\\|0.234$\\\\pm$0.001 |\\n\\nThese results exhibit that the performance of TwinsFormer is stable in ablation studies.\"}", "{\"comment\": \"Dear Reviewer AH3t,\\n\\nAs the author/reviewer discussion will close soon, we would like to know if our response has addressed your concerns and questions. If you have any further concerns or suggestions for the paper or our rebuttal, please let us know. We would be happy to engage in further discussion.\\n\\nThank you again for the time and effort you dedicated to reviewing this work.\\n\\nSincerely,\\n\\nAuthors\"}", "{\"comment\": \"Thank you very much for your feedback. We sincerely appreciate your valuable comments and the recognition you have given to our work.\"}", "{\"title\": \"Response to Reviewer pkJh\", \"comment\": \"Thanks for your time and effort in reviewing our paper. We are pleased to see that you identified the experimental results. However, we guess that there are some misunderstandings about the novelty of our work, which we try to clarify in our following responses to specific questions and concerns:\\n\\nGenerally, __time series forecasting__ aims to predict future temporal variations based on historical values of time series, where the primary challenge is how to effectively capture __inherent dependencies__ from historical data. \\nInherent dependencies refer to the __correlations and patterns__ within the data that are crucial for accurate prediction, such as temporal dependencies, seasonal and trend patterns, autocorrelation, multivariate correlation, and so on.\\nMore specifically, __temporal dependencies__ are the influence of past values on future values. For example, today's weather can be predicted based on yesterday's weather conditions.\\n__Seasonal patterns__ denote repetitive patterns that occur at regular intervals, such as daily, weekly, or yearly cycles.\\n__Trend patterns__ are long-term upward or downward movements in the data.\\n__Autocorrelation__ is the correlation between a time series and a lagged version of itself.\\n__Multivariate correlation__ is the correlation among multiple time series.\\n\\nIn time series forecasting, trend-seasonal decomposition is a common technique to tackle intricate temporal patterns.\\nIn lines 053-084 of our manuscript, we point out that __existing decomposition designs__ generally utilize two __independent branches__ to highlight seasonal and trend properties separately for final prediction.\\nConsidering trend-seasonal decomposition is an __untrainable linear transformation__ (i.e., moving average kernel), the decomposed trend and seasonal components __cannot__ accurately characterize the inherent dependencies of the raw time series.\\n\\nTo intuitively understand the limitation of the seasonal-trend decomposition, we provide the trend-seasonal decomposition visualization of two channels (i.e., variates) on the Electricity and Traffic datasets.\\nAs seen in Figure 2 of our manuscript, the \\\"Observed\\\" denotes the raw time series, and the \\\"Trend\\\" and \\\"Seasonal\\\" are decomposed components from the raw time series by the moving average kernel.\\n__Comparatively__, the trend and seasonal components exhibit distinct characteristics for the observed time series, where accurately capturing the inherent dependencies from the seasonal or trend components __alone is impossible__.\\nTaking __channel 11__ on the Electricity dataset as an example, the long-term upward or downward movements roughly depicted by __the trend component__ of channel 11 __cannot reflect the periodic patterns__ of raw channel 11, while the repetitive patterns depicted by __the seasonal component__ of channel 11 are __not actual temporal patterns__ of channel 11.\\n__Based on the observation, we believe the interactions between decomposed components are important for time series forecasting.__ \\n\\nBenefiting from the ability to model __long-term dependencies (i.e., autocorrelation and multivariate correlation)__ by the attention module, Transformer-based methods have shown significant success in time series forecasting. \\n__To learn the interactions between decomposed components, we take the attention mechanism as the basic module to connect the seasonal and trend components.__\\nTo avoid excessive computational overhead caused by the attention mechanism, we must consider how to obtain the multivariate correlations for different components.\\nIn __Figure 2__ of the manuscript, we observe that the seasonal components can better reflect the temporal patterns of observed time series, so we use the seasonal components to learn the multivariate correlation of time series data.\\nBased on __Equation (1)__ of the manuscript, we can find that the seasonal component is obtained by subtracting the trend component from the original time series, so we replace the addition operation of the existing Transformer-based architecture with a subtraction mechanism.\\nIn __lines 243-269__ of our manuscript, we provide a rationality analysis for our dual-stream framework.\\n\\n\\n__In summary__, by delving into the trend-seasonal decomposition process and combining the attention mechanism's remarkable ability to learn long-term dependencies, our proposed Transformer-based dual-stream interaction framework is __traceable rather than heuristic__.\\nBased on our interactive architecture, we can learn __inherent dependencies beyond the multivariate correlation__ to improve time series forecasting.\\nRegarding the __design choices__, we acknowledge that some may initially appear ad-hoc. \\nHowever, we would like to clarify that these choices were made based on a __combination__ of empirical evidence and theoretical considerations. While some decisions were indeed informed by preliminary experiments, they were also __guided by established principles__ in the field.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}" ] }

[ "Srivathsan Amruth", "Niranjan Gopinath" ]

[ "Regularisation", "Supervised Learning", "Neural Network Architecture Paradigms" ]

{ "note_id": [ "ymVvH97h0G", "xHydU0Qzmp", "w5x4wE4spi", "uZPaNb6Ntq", "tGJEiKzIYZ", "sFF8AFDElY", "ruGYhJaK9i", "mjQ6hVXWbE", "kwSR23WPjP", "eECOguZoRB", "bVstD8J2iX", "am0C1ioEU9", "YW86lHhPnm", "RzoEeo7ta4", "RUtNB8dpdd", "RBXIh62czX", "HocMUEYVoy", "ECYuZR9YFD", "CrlA82vksg" ], "note_type": [ "official_comment", "official_comment", "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_review", "official_review", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "meta_review" ], "note_created": [ 1732542594192, 1733003789648, 1737523897482, 1731605782005, 1732542442756, 1731599525533, 1733160979207, 1731607538908, 1730664836791, 1731598995995, 1730940561245, 1729349381941, 1732240201998, 1730702772770, 1733161473986, 1732511256484, 1732514132284, 1732586025994, 1734761898447 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission8258/Authors" ], [ "ICLR.cc/2025/Conference/Submission8258/Reviewer_a7nU" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission8258/Authors" ], [ "ICLR.cc/2025/Conference/Submission8258/Authors" ], [ "ICLR.cc/2025/Conference/Submission8258/Authors" ], [ "ICLR.cc/2025/Conference/Submission8258/Authors" ], [ "ICLR.cc/2025/Conference/Submission8258/Authors" ], [ "ICLR.cc/2025/Conference/Submission8258/Reviewer_aCCL" ], [ "ICLR.cc/2025/Conference/Submission8258/Authors" ], [ "ICLR.cc/2025/Conference/Submission8258/Reviewer_a7nU" ], [ "ICLR.cc/2025/Conference/Submission8258/Reviewer_ZFUN" ], [ "ICLR.cc/2025/Conference/Submission8258/Reviewer_ZFUN" ], [ "ICLR.cc/2025/Conference/Submission8258/Reviewer_pjrF" ], [ "ICLR.cc/2025/Conference/Submission8258/Authors" ], [ "ICLR.cc/2025/Conference/Submission8258/Area_Chair_YhWL" ], [ "ICLR.cc/2025/Conference/Submission8258/Reviewer_pjrF" ], [ "ICLR.cc/2025/Conference/Submission8258/Reviewer_aCCL" ], [ "ICLR.cc/2025/Conference/Submission8258/Area_Chair_YhWL" ] ], "structured_content_str": [ "{\"comment\": \"Thank you for your kind review once again! Totally agree on al the points you presented here and we will take another look into how to re-implement what is suggested properly for another revision of this paper in the future. Once again thank you for taking the time to kindly look through and share your thoughts on how it can be improved, we really appreciate it!\"}", "{\"comment\": \"In response to the authors' first point, I would like to clarify that they have completely misunderstood my point. My critique is not that the authors \\\"did not propose Propositions 2-5.\\\" Rather, it is that including informal proof sketches of well-known statements (Propositions 2-5) is unnecessary and detracts from the overall presentation of the paper. If the authors wish to retain these informal proof sketches, they would be more appropriate in the Appendix. I do not see why this constitutes a discussion on the definition of a \\\"Proposition.\\\"\", \"this_was_my_original_concern\": [\"Some **contents of the paper is unnecessary**, which greatly diminishes the quality of the paper. E.g.,\", \"Extensive description of MNIST\", \"Extensive description of classification problem\", \"Informal proof sketches of Proposition 2-5\", \"Second, the authors have referred me to look at the discussion with Reviewer ZFUN. I agree with the reviewer that \\\"all results should be presented for completeness\\\" and more comprehensive experimentation is needed for completeness (e.g., proper hyper parameter sweeps, etc.). Since my main concerns still remain, I retain my score.\"]}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"comment\": \"Dear Reviewer, thank you for your comments.\", \"addressing_the_questions\": \"1) What data distribution assumption is used in the experiment to implement the proposed regularization? . You mention in your first line of the review \\\"more specifically the central limit theorem\\\". I think this is self explanatory. \\n\\n2) Where is the assumption used? This is the defining assumption of a \\u2018perceptron\\u2019 network. The whole idea was to keep the simplest network to be able to test validity. I think this question highlights a fundamental misunderstanding in terms of the most basic part of the paper already as this assumption is supposed to be trivial (doesn\\u2019t necessarily imply \\u2018easy\\u2019 but rather elementary that it doesn't need a complex or elaborate argument if one has some experience in theoretical ML ) as it would be difficult to continue understanding the rest of the paper clearly if this is something you are not too familiar with as the basis of most ideas are built from this school of thought. So maybe why there is a stark difference in viewpoints .\", \"regarding_weaknesses\": \"1) \\\"Don't see any novel architecture here\\\" != \\\"novel regularization technique\\\" (abstract line 2) \\n\\n\\n2) The phrasing suggests several misconceptions, as the questions contain multiple semantic errors. Addressing from back to front :\\n\\n\\\"Assume output of the model is Bernoulli\\\" -> Wrong. The output of the model is some probability distribution. Specifically for the toy example presented it is the LC of Bernoullis.\\n\\n\\\" MNIST ... continuous data \\\" -> another misconception since the MNIST data set is originally 8 bit $(i \\\\in \\\\mathbb{Z}, \\\\ 0 \\\\leq i \\\\leq 255)$ and it's normalized to [0,1]. This normalization doesn't make it continuous. It is still discrete. \\n\\n\\\"This does not make sense for practical data\\\" -> again wrong. The generating function is assumed to be a LC of Bernoullis which are in theory able to approximate many distributions. Also if by practical you mean anything stored on a computer, then it is even more so as it is discrete (refer to section 2.5 propositions).\\n\\n\\\"assume the data follows a linear combination of Bernoulli distributions\\\" -> \\\"follow\\\" does not necessarily mean \\\"generated by\\\".\\n\\n\\n\\n\\n\\n3) Again there are multiple false claims made here.\\n\\n\\\" For data in practice ... hard to compute its characteristic function ... do not know its true distribution \\\" -> This is the key misconception that you may have. As described in the first part of the paper the PFINNs are constructed in such a way that we assume SOME structure regarding an assumption about it's true distribution.\\nFurthermore, the idea of ML in general is to derive a function/MAPPING with the help of heuristics as a closed form solution is difficult/impossible to find. In our case , the PFINNs we use have the goal of finding a MAPPING from the data to a preset probability function(be it a PMF or approximate PDF or even a singularity) that approximately generates the data (the learning is to find the best-fit parameters). This can be seen from the toy example provided. Since the practitioner constructs the architecture based on THIS apriori knowledge, this means it is easy for a person with sufficient knowledge to construct the limiting characteristic function (in our case the Normal characteristic function we use as the target for regularizing against) for the appropriate setting.\\n\\n\\\" This makes the proposed regularization method invalid.\\\" -> This is therefore a flawed claim.\\n\\n\\\"How do you update the parameters through the regularization?\\\" -> This is stated in the paper.\\n\\n\\n\\n4) Agree on the briefness, \\n\\nProfessionalism -> We think this is a slippery argument because we would say instead of vaguely suggesting \\u2018high quality papers\\u2019 to actually provide some concrete examples. This way there is some basis for discussion.\\n\\nLooking forward to your reply. I think there is lots to learn from discussion and maybe you can elucidate your points further to see if there is something that hopefully both of us can take away.\"}", "{\"comment\": \"Thank you for your kind response once again! Especially appreciate the advice of the toy dataset , will have to spend sometime engineering it properly and observing the changes to present a holistic and try to put that into a revision of this paper in the future. Thank you once again for taking the time to read and review!\"}", "{\"comment\": \"Dear Reviewer, thank you for your review.\\n\\nAlso with regards to your comment on the section 2.5 , semantically, a proposition is a statement that can either be true or false, but not both. It is a declarative sentence that makes a claim about a mathematical object or concept, and its truth value (whether it is true or false) is determined based on logical reasoning, definitions, and previously established facts or axioms.\\n\\nExample of which \\\"1 + 1 = 2\\\" is a proposition, and it is true based on the usual set theoretic axioms. This proposition doesnt carry the same idea of \\\"proposing\\\" which is what you are implying we did.\\n\\nOtherwise, since you don't have any questions and there is a signifcant overlap with your comment with the Reviewer ZFUN, I would invite you look at our discussion there.\", \"more_specifically_let_me_highlight_the_parts_that_are_relavent\": \"\\\"This was a point of contention which we had when trying to distill the numerical results to be presented. From our discussions we decided to fix the parameter instead because it feels very much like p-value hacking just to squeeze the best possible results for a given setting to show where one method shines since it is really data dependant and from what we found, it quite difficult to explain exactly why it works as it is a non convex problem to be explained ; We do have results where the method obliterates the usual L1,L2 when hidden layers are added and on some specific datasets but I feel would be disingenuous to present it in this light as a \\u2018shiny new best\\u2019 alternative because the truth is rather it is supposed to be something that can be \\u2018possibly\\u2019 used depending on context. That is why the choice of fixed lambda and common datasets were used to present a more honest view regarding it. The empirical idea regarding the presentation of the numerics was more for showing it is a stable algorithm that works and works relatively okay in the wild on standard datasets with untuned parameters.\\n\\nPerhaps however, if you feel that a revised version where there are significant differences post-tuning to be shown would be better then we can adjust the numeric section accordingly again with the relevant results.\\n\\nWith regards to this your insight on the mean is spot on, infact after reviewing the manuscript again now we realised that the unfortunate auto placement of the table makes it seem like it was the main focus that the \\u2018mean \\u2018is the best of the 4 but it was meant to just be a general observation , that entire paragraph was just a general qualitative discussion which wasn't meant to be brief outline of the table and not meant to be a main driving/selling point of the method. Maybe the choice of word \\\"attractive\\\" was also counterintuitive (since we were actually trying to downplay how important it was) with regards to what we were trying say. Because, the real intended take away we wanted for the reader to have was the final point where the loss landscape , which is primarily data dependant can drastically change everything so having an alternative STABLE tool set is key; more so not that this alternative tool is \\u2018better\\u2019 but rather a new option that could be useful in a practitioners tool kit when the other options are not working well. Would you say maybe we can jettison that specific paragraph of describing the table results in a revised version to maybe drive the point clearer ?\\n\\nThank you for highlighting the redundance of the MNIST and classification sections. We just placed them for completeness but if it hinders the reading flow we are happy to remove it. (Likewise with regards to redundancy of section 5 which just served as a motivation for why we need numerical methods to solve a continuous function)\\n\\nWe hope these changes would help to rectify portions of the presentation and soundness aspects.\\n\\nOtherwise, to get some help with the revision process, could we take some of your time to explain the rationale for you rating of the contribution and soundness, since we feel that this approach yields some novel findings which have been supported by the relevant proofs with the help of numerical tools to validate that it is functioning as expected.\\n\\nSpecifically, our contribution lies in embedding prior probabilistic knowledge into the regularization process. This is the key idea that we want to share through the paper. To better explain, as we mentioned earlier, the primary goal of the numerical experiments was not to achieve higher accuracy scores, but rather to demonstrate the validity of our proof of concept. The numerics serve to confirm that the approach works as intended, rather than being focused solely on performance metrics. Hence as you observed, sometimes the non-regularize setting works best and to find a measure \\\"generalisation\\\" would be a difficult task. The metric we thought was sufficient to show the generalisation instead in numerics was a good enough and stable performance across a diverse range of datasets as presented.\\\"\"}", "{\"comment\": \"Thank you for the clarification. The rationale for that was just due to a misunderstanding because of the choice of wording in your original comment where \\\"Proposition 2-5 which are not proposed by the paper but already well-known\\\". The comment to shift it towards the appendix is appreciated and more clear as to what was the intention.\\n\\nOtherwise thank you for your suggestions. We will consider them for another iteration of the paper.\"}", "{\"comment\": \"Dear Reviewer, thank you for your kind and thoughtful comments.\", \"regarding_the_weaknesses\": \"1) We agree strongly with your point. As mentioned in reply to ZFUN \\\" the idea presented in the remaining motivation was specifically with regards to the regularization target but didn't focus on why the characteristic function was used. The advantage of it comes specifically from the idea that it addresses directly distributional behavior and also calculating convolutions becomes a simple step of point wise algebra. Since this paper was cut down from a larger manuscript we realized the motivation part might have been chopped a bit too aggressively, maybe would you think addressing this would help in the revision? \\\"\\n\\n2) This is also a good question. The idea is more so understanding/assuming a global viewpoint would allow to smooth any microcomplexities that may arise from the usual learning process. A simple example to visualise this is if you imagine drawing some random samples from something we imagine to be a uniform distribution, then you have a non-leveled histogram but by trying to \\\"preserve\\\" the uniform distribution properties, we can allow for this jagged histogram to be smoothed out.\\n\\n3) Only singular forms and you idea of having the combined forms was also something that was on our mind but we did not add those into the table since it was getting too big (the permutations make it a \\\"too many columns to comfortably read\\\" table) . However, we do have the combined results from our tests. Would you recommend we put it in our revision for additional information, and if so how do you think we can neatly present it? \\n\\n4) The idea was more so choice of \\\"punishment metric \\\" where instead of using something like the L-p regularisation, this can be a handy and more \\\"explainable choice\\\" of regularization. More concretely, the regularisation of L2 penalizes large weights by adding a term proportional to the squared magnitude of the model's parameters to the loss function but heuristically, our regularisation punishes the deviation of the generated distribution through assessing how non-similar it is to some global distribution assumption we postulate it is supposed to take (this global assumption is essentially what we use to construct the design of the NN in the first place).\", \"regarding_your_overall_comment\": \"Thank you for your kind words and it was nice to see you enjoyed the manuscript. We also totally see your point of view where it would have been nice to see where the model does well. I think as per what we addressed in Reviewer ZFUN's comment also, maybe our view on wanting the reader not to take away that this model is the \\\"best\\\" model by cherry picking examples may have worked against us because we ended up not showing examples where it shines despite it being difficult to explain exactly why but would it help if we add a section regarding that and putting back the results where it significantly outperformed the others in certain datasets in the revised manuscript?\\n\\nThank you once again!\"}", "{\"summary\": \"This paper proposes to incorporate probability rules into the neural network architectures, more specifically the central limit theorem. They use a linear model on MNIST as an example and propose to regularize the distance between the characteristic function of the data distribution and the normal distribution.\", \"soundness\": \"1\", \"presentation\": \"1\", \"contribution\": \"1\", \"strengths\": \"Incorporating probability rules into the neural network architectures is a good idea. The classification task in machine learning is essentially learning a prediction rule $\\\\mathbb{P}(Y|X)$. Incorporating probability rules into the model may facilitate the learning of this prediction rule.\", \"weaknesses\": [\"The described model in Sec 2 looks like just neural networks used in common practice. I don't see any novel architecture here.\", \"The authors assume the data follows a linear combination of Bernoulli distributions. This does not make sense for practical data. For example, the MNIST data, which is given as an example in Sec 2, is continuous data in $[0, 1]$ and is not Bernoulli. Or does the authors mean to assume the output of the model is Bernoulli?\", \"In line 315, the authors claim \\\"for a general class of PFINNs, one only needs to adjust the modeling of the random variable presented in Definition 3 to reformulate the equation in Proposition 1 accordingly.\\\" However, for data in practice, it is hard to compute its characteristic function as we do not know its true distribution and the distribution is what the model is trying to learn in some sense. This makes the proposed regularization method invalid. Even if we can compute the characteristic function of data distribution, the regularization is not a function of weight parameters. How do you update the parameters through the regularization?\", \"The writing is not professional. The paper spends a lot of space introducing the setup of the model and MNIST dataset. The dataset and network architecture are quite common and can be introduced briefly. For example, the dataset can be represented generally as $\\\\\\\\{x_i, y_i \\\\\\\\}_{i=1}^N$. Some sentences are not professional in scientific writing, such as, \\\"To explore the existence of PFINNs as (neuro)symbolic AI and as hybrid systems would require significantly more than 9 pages in a manuscript. (line 052)\\\" and \\\"Maybe other potential questions may merit consideration? (line 171)\\\". I would suggest the authors read some high-quality papers and learn their writing styles.\"], \"questions\": [\"What data distribution assumption is used in the experiment to implement the proposed regularization?\", \"Where is the assumption 1 used in the paper?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"1\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Dear Reviewer, thank you for your kind comments.\\n\\nRegarding the question about motivation, thank you for highlighting this, the idea presented in the remaining motivation was specifically with regards to the regularization target but didn't focus on why the characteristic function was used. The advantage of it comes specifically from the idea that it addresses directly distributional behavior and also calculating convolutions becomes a simple step of point wise algebra. Since this paper was cut down from a larger manuscript we realized the motivation part might have been chopped a bit too aggressively, maybe would you think addressing this would help in the revision? \\n\\nRegarding the weaknesses mentioned of why not use multiple lambda parameters -> This was a point of contention which we had when trying to distill the numerical results to be presented. From our discussions we decided to fix the parameter instead because it feels very much like p-value hacking just to squeeze the best possible results for a given setting to show where one method shines since it is really data dependant and from what we found, it quite difficult to explain exactly why it works as it is a non convex problem to be explained ; We do have results where the method obliterates the usual L1,L2 when hidden layers are added and on some specific datasets but I feel would be disingenuous to present it in this light as a \\u2018shiny new best\\u2019 alternative because the truth is rather it is supposed to be something that can be \\u2018possibly\\u2019 used depending on context. That is why the choice of fixed lambda and common datasets were used to present a more honest view regarding it. The empirical idea regarding the presentation of the numerics was more for showing it is a stable algorithm that works and works relatively okay in the wild on standard datasets with untuned parameters. \\n\\nPerhaps however, if you feel that a revised version where there are significant differences post-tuning to be shown would be better then we can adjust the numeric section accordingly again with the relevant results. \\n\\n\\nWith regards to this your insight on the mean is spot on, infact after reviewing the manuscript again now we realised that the unfortunate auto placement of the table makes it seem like it was the main focus that the \\u2018mean \\u2018is the best of the 4 but it was meant to just be a general observation , that entire paragraph was just a general qualitative discussion which wasn't meant to be brief outline of the table and not meant to be a main driving/selling point of the method. Maybe the choice of word \\\"attractive\\\" was also counterintuitive (since we were actually trying to downplay how important it was) with regards to what we were trying say. Because, the real intended take away we wanted for the reader to have was the final point where the loss landscape , which is primarily data dependant can drastically change everything so having an alternative STABLE tool set is key; more so not that this alternative tool is \\u2018better\\u2019 but rather a new option that could be useful in a practitioners tool kit when the other options are not working well. Would you say maybe we can jettison that specific paragraph of describing the table results in a revised version to maybe drive the point clearer ?\\n\\nThank you for highlighting the redundance of the MNIST and classification sections. We just placed them for completeness but if it hinders the reading flow we are happy to remove it. (Likewise with regards to redundancy of section 5 which just served as a motivation for why we need numerical methods to solve a continuous function)\\n\\nWe hope these changes would help to rectify portions of the presentation and soundness aspects.\\n\\n\\nOtherwise, to get some help with the revision process, could we take some of your time to explain the rationale for you rating of the contribution and soundness, since we feel that this approach yields some novel findings which have been supported by the relevant proofs with the help of numerical tools to validate that it is functioning as expected. \\n\\nSpecifically, our contribution lies in embedding prior probabilistic knowledge into the regularization process. This is the key idea that we want to share through the paper. To better explain, as we mentioned earlier, the primary goal of the numerical experiments was not to achieve higher accuracy scores, but rather to demonstrate the validity of our proof of concept. The numerics serve to confirm that the approach works as intended, rather than being focused solely on performance metrics. Hence as you observed, sometimes the non-regularize setting works best and to find a measure \\\"generalisation\\\" would be a difficult task. The metric we thought was sufficient to show the generalisation instead in numerics was a good enough and stable performance across a diverse range of datasets as presented.\"}", "{\"summary\": \"This paper proposes a characteristic function-based regularization method for contextual regularization. They compute the characteristic function for a linear combination of Bernoulli random variables and discretize it for use as a regularization term. They empirically evaluate the proposed regularization approach compared to existing baselines across 5 different classification datasets. Their results demonstrate that integrating this method improves performance on the considered benchmark supervised classification tasks.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"This paper investigates a novel regularization technique based on characteristic functions of datapoints, getting motivation from physics-informed neural networks.\", \"weaknesses\": [\"Some contents of the paper is unnecessary, which greatly diminishes the quality of the paper (e.g., extensive description of MNIST, and classification problem (Sections 2.1-2.2), informal proof sketches of Proposition 2-5 which are not proposed by the paper but already well-known)\", \"Gains achieved by the method are weak. The authors state \\\"It is generally observed that the mean for the regularization we proposed, throughout 4 out of 5 datasets, achieve the highest mean\\\". However, as shown in Table 1, for those 4 out of 5 datasets, the performance of their method often times just match or is only slightly better (~0.0001) compared to no regularization at all (None column).\"], \"questions\": \"NA\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"details_of_ethics_concerns\": \"NA\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper proposes regularization via the characteristic function of datapoints to train networks, similar to physics-informed neural networks.\\n\\nThey derive the characteristic function for a linear combination of Bernoulli random variables, and discretize this function to use as regularization.\\n\\nThe perform experiments on a flattened version of MNIST with a linear model with a softmax activation.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"Proposes a new approach to perform regularization via a discretized version of the characteristic function\", \"weaknesses\": \"I believe that a large amount of the content in this paper is unnecessary background. For instance, much of section 2 focuses on introducing the dataset of MNIST and the setting of a classification problem \\u2013 all of which are standard and used in the field.\\n\\nFor section 5, the informal proof sketches are describing results that are not all that relevant to the paper and some of which are already known facts. For instance, it is already known that the set of reals are complete. Most of these propositions are not new and follow the content in (https://www.lix.polytechnique.fr/~bournez/load/MPRI/Cours-2024-MPRI-partie-I-goodMPRI.pdf)\\n\\nIn the experiments, the authors fix $\\\\lambda$ to 0.01 for all methods, while I believe that this should be tuned for each method individually on held-out validation data. \\n\\nFurthermore, I have some reservations about the authors' empirical results. There seems to be almost no difference between regularizing with $\\\\psi_{\\\\inf}$ and standard training without any regularization. While the authors claim the best mean performance across 4 of the 5 tasks, this roughly equivalent performance with standard training without any regularization makes up 3 of those best-performing tasks. Thus, it\\u2019s unclear if this regularization is beneficial in general and just is essentially not performing any regularization.\", \"questions\": \"How is the characteristic function related to the motivations in section 3 (specifically with regards to the infinite series of inquiries about the input image)?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Reviewer Response\", \"comment\": \"Thank you for the clarifications!\\n\\n> **The advantage of it comes specifically from the idea that it addresses directly distributional behavior and also calculating convolutions becomes a simple step of point wise algebra. Since this paper was cut down from a larger manuscript we realized the motivation part might have been chopped a bit too aggressively, maybe would you think addressing this would help in the revision?**\\n\\nYes I think this would help improve the motivation and should be included in the main text. \\n\\n> **This was a point of contention which we had when trying to distill the numerical results to be presented. From our discussions we decided to fix the parameter instead because it feels very much like p-value hacking just to squeeze the best possible results for a given setting to show where one method shines since it is really data dependant and from what we found, it quite difficult to explain exactly why it works as it is a non convex problem to be explained ; We do have results where the method obliterates the usual L1,L2 when hidden layers are added and on some specific datasets but I feel would be disingenuous to present it in this light as a \\u2018shiny new best\\u2019 alternative because the truth is rather it is supposed to be something that can be \\u2018possibly\\u2019 used depending on context.**\", \"a_couple_of_comments_on_this\": \"(1) I think that all the results should be presented for completeness and that the current state of the experiments section is a bit narrow in scope and hard to make any sort of conclusive statements. (2) A proper round of hyperparameter sweeps could be done here: a set of values of the regularization parameter used for each method, which is tuned over a fixed set of validation data. This would be a much stronger experimental setting than just fixing a single regularization parameter, especially when it is the nature and impact of the regularization which is being currently studied. The current experimental results show almost no difference when compared to performing optimization without any regularization, so experiments demonstrating where this method would be beneficial are crucial.\\n\\n> **Because, the real intended take away we wanted for the reader to have was the final point where the loss landscape , which is primarily data dependant can drastically change everything so having an alternative STABLE tool set is key; more so not that this alternative tool is \\u2018better\\u2019 but rather a new option that could be useful in a practitioners tool kit when the other options are not working well.**\\n\\nIf this is the takeaway from the paper, then it is important to show results where this tool set indeed works well and while other methods fail.\\n\\n> **Otherwise, to get some help with the revision process, could we take some of your time to explain the rationale for you rating of the contribution and soundness, since we feel that this approach yields some novel findings which have been supported by the relevant proofs with the help of numerical tools to validate that it is functioning as expected.**\\n\\nOverall, my main concerns, which still remain, are exactly that there do not seem to be any conclusive benefits from the proposed approach and that there are no empirical demonstrations of cases where it outperforms standard regularization. While the goal is \\\"embedding prior probabilistic knowledge into the regularization process\\\" and \\\"validity of our proof of concept\\\", this doesn't seem sufficient to me as a contribution, given that it exactly matches standard optimization without any regualrization.\"}", "{\"summary\": \"The authors propose a novel form of regularization which regularizes the model output probabilities towards that of a characteristic function defined over a sum of Bernoulli random variables.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": [\"The idea is novel and provides a new route of regularization which has not been considered before.\", \"The presentation and derivation is clear, and well done.\"], \"weaknesses\": [\"The overall motivation of using characteristic function regularization is not clear.\", \"The abstract states \\u201cimproves performance \\u2026 by preserving essential distributional properties\\u2026\\u201d -> How does the preservation of such properties aid in generalization?\", \"The abstract states that the method is meant to be used in conjunction with existing regularization methods. Were the results presented results utilizing multiple forms of regularization (such as $L_2 + \\\\psi_2$) or were the only singular forms of regularization?\", \"In the conclusion, the author state the follwoing: \\u201cintegrating these techniques can offer a probability theory based perspective on model architecture construction which allows assembling relevant regularization mechanisms.\\u201d \\u2014> I do not see how this can be done after reading the work. can you give a concrete example of how the results presented in this work may give any insight into model architecture construction?\", \"## Overall\", \"While I found the work interesting and captivating to read, after finishing the manuscript I am left wondering what possible benefit the regularization provides over existing methods. The results are somewhat ambiguous and I find they do not demonstrate why or when a clear benefit can be achieved by applying the given regularization method. If the authors could provide some insight as to when and why the method would be successful, I think it would go a long way in demonstrating the real-world usefulness of characteristic function regularization. Even if this could be demonstrated in a synthetic toy setting, it could provide interesting insights.\"], \"questions\": \"Questions are covered above in the weaknesses section.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Dear Reviewer, thank you for your kind clarification.\\n\\nRegarding the point you mentioned of the 784 dimensional vector the idea is that you map it into a 10 dimensional space which is assumed to be independent. But as is I think the 2 points regarding that would be clarified by the 2nd point you mentioned regarding that.\\n\\nWould take into account your comments regarding some of the wording for a future iteration of the paper. There seems to be some communication gap due to word choices and totally understandable with regards to some of the points you mentioned.\"}", "{\"title\": \"The author-reviewer discussion period is ending soon\", \"comment\": \"Dear reviewers,\\n\\nIf you haven\\u2019t done so already, please engage in the discussion as soon as possible. Specifically, please acknowledge that you have thoroughly reviewed the authors' rebuttal and indicate whether your concerns have been adequately addressed. Your input during this critical phase is essential\\u2014not only for the authors but also for your fellow reviewers and the Area Chair\\u2014to ensure a fair evaluation.\\nBest wishes,\\nAC\"}", "{\"title\": \"Reviewer Response\", \"comment\": \"Thank you for your response. Overall I agree with the main concern of reviewer ZFUN in that there has been no clear benefit demonstrated for the proposed method. It is a very interesting idea, but the results do not highlight the unique characteristics of the proposed method which make it superior for a given situation.\\n\\nI understand that it may be hard to find such cases where the model excels among real datasets. However, in my initial review, I suggested engineering a toy dataset which would highlight the benefits of the method. If there is a truly novel and useful case for this kind of regularization over traditional methods like L2, then it should be possible to engineer an artificial setting which clearly demonstrates these benefits in a controlled setting, which would also show why and how L2 regularization would have trouble.\\n\\nThis would go a long way in demonstrating to the reader the necessity of usefulness of the proposed method.\"}", "{\"comment\": [\"About the assumption: usually you don't need to make such an assumption if you do not use it in any theorem. But of course, you can still write it anyway.\", \"novel architecture: because the title and abstract write \\\"We first define Probability Function Informed Neural Networks as a class of universal function approximations ...\\\" and the title of Sec 2 reads \\\"MATHEMATICAL CONSTRUCTION OF A TYPE OF SIMPLE PFINN\\\", I expected to see some novel PFINN architectures. But there is not.\", \"\\\"Assume output of the model is Bernoulli\\\". I am saying this because, in line 152, you write \\\"We will utilize this approximation under the assumption that each node in the output layer corresponds to a random variable $\\\\alpha_i$ that follows a Bernoulli distribution:\\\" At this point, the proposed method is still not quite clear to me. For the toy example, if the output distribution is the LC of Bernoullis, why are you assuming the data distribution is the LC of Bernoullis?\", \"MNIST is discrete: ok let's say the data is discrete. But the data is a 784-dimensional vector. Can you define a Bernoulli random variable for a high-dimensional vector? I don't think the coordinates are independent.\", \"\\\"the PFINNs we use have the goal of finding a MAPPING from the data to a preset probability function\\\".\", \"I didn't say this in the paper. If this is the case, it's more reasonable to say you are learning a mapping from data distribution to an LC of Bernoullis and assume the output is an LC of Bernoullis.\", \"high quality papers. See this paper as an example: Gradient Descent Provably Optimizes Over-parameterized Neural Networks\", \"Overall, I think the proposed method might be novel, but there are a lot to improve, especially the writing.\"]}", "{\"metareview\": \"This paper proposes a regularization technique for neural network training, but it is not ready for publication in its current form. The approach lacks clear motivation and fails to demonstrate any substantial empirical benefits. Additionally, the reviewers have raised significant concerns about the clarity of the presentation. For example, the paper allocates excessive space to unnecessary details, which obscures its contributions. I strongly recommend that the authors carefully address the reviewers\\u2019 feedback, focusing on improving the clarity and presentation while reevaluating the overall merit of the proposed method.\", \"additional_comments_on_reviewer_discussion\": \"Most of the concerns raised by the reviewers were not addressed during the rebuttal period; actually, the reviewers became more certain in their original evaluation, and this did not change until the end of the reviewer discussion phase.\"}" ] }

[ "Zizhao Chen", "Mustafa Omer Gul", "Yiwei Chen", "Gloria Geng", "Anne Wu", "Yoav Artzi" ]

[ "continual learning", "natural language processing", "interactive learning", "reinforcement learning" ]

{ "note_id": [ "yegdNxvDeA", "wOjLomsSed", "uSJs3vkP5j", "qLBGilVlUj", "l1PeoIG3gx", "jgmtqPqunt", "i8BVIT7VHX", "cqtogsWXpN", "UMPbwMU6zB", "PbfwhjDzDu", "OaGmfG6P3T", "MNaziH9ItQ", "LdwK3EfuWC", "JkSrdvSerU", "IJllESHwsZ", "HynNm57Sos", "HPPpSNng1m", "HAA51h9OUc", "DsbpbJgz3K", "DRjIe1N2f8", "A3H3uH2TLl", "6siUU5Ihen", "4ogPHCCReK", "2EzOSFMUGZ", "0hTbWPh8G4" ], "note_type": [ "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "meta_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "decision", "official_review", "official_review", "official_comment" ], "note_created": [ 1732165109231, 1732711166068, 1732650770428, 1732064876332, 1732591573366, 1730614466557, 1732715212697, 1732740363573, 1732681100449, 1730692887536, 1732061497770, 1732062117640, 1732227247695, 1732060476063, 1732570957642, 1733956769422, 1732649128761, 1732682270676, 1732063131683, 1732741005287, 1732684920381, 1737523610897, 1730480543926, 1730701547815, 1732164438405 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission3970/Authors" ], [ "ICLR.cc/2025/Conference/Submission3970/Reviewer_Aocp" ], [ "ICLR.cc/2025/Conference/Submission3970/Reviewer_Gzkj" ], [ "ICLR.cc/2025/Conference/Submission3970/Authors" ], [ "ICLR.cc/2025/Conference/Submission3970/Authors" ], [ "ICLR.cc/2025/Conference/Submission3970/Reviewer_Gzkj" ], [ "ICLR.cc/2025/Conference/Submission3970/Reviewer_aUKE" ], [ "ICLR.cc/2025/Conference/Submission3970/Authors" ], [ "ICLR.cc/2025/Conference/Submission3970/Authors" ], [ "ICLR.cc/2025/Conference/Submission3970/Reviewer_XxNx" ], [ "ICLR.cc/2025/Conference/Submission3970/Authors" ], [ "ICLR.cc/2025/Conference/Submission3970/Authors" ], [ "ICLR.cc/2025/Conference/Submission3970/Authors" ], [ "ICLR.cc/2025/Conference/Submission3970/Authors" ], [ "ICLR.cc/2025/Conference/Submission3970/Reviewer_Aocp" ], [ "ICLR.cc/2025/Conference/Submission3970/Area_Chair_icSk" ], [ "ICLR.cc/2025/Conference/Submission3970/Reviewer_aUKE" ], [ "ICLR.cc/2025/Conference/Submission3970/Authors" ], [ "ICLR.cc/2025/Conference/Submission3970/Authors" ], [ "ICLR.cc/2025/Conference/Submission3970/Authors" ], [ "ICLR.cc/2025/Conference/Submission3970/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission3970/Reviewer_Aocp" ], [ "ICLR.cc/2025/Conference/Submission3970/Reviewer_aUKE" ], [ "ICLR.cc/2025/Conference/Submission3970/Authors" ] ], "structured_content_str": [ "{\"title\": \"Initial Response Part 2\", \"comment\": \"> Scalability: Heavy reliance on human feedback: The proposed framework relies heavily on human feedback. Although the authors have countered the problem of annotating the responses, the problem of getting good interaction data still remains a crucial problem, making it hard to scale.\\n\\nKey to our approach is that it relies on the interaction the system has with users during its deployment, as it completes tasks. So these interactions are *already taking place*. Our approach adds no overhead in terms of interaction or annotation on top of regular interactions with users. This is the most important property of our work: the signal arises naturally from the interactions the system already has with human users. This is different from contemporary RLHF methods, which rely on getting post-hoc preference annotations from third-party annotators. This is why our approach naturally scales. As people use the system, we get more signals, and the system autonomously learns from them \\u2013 without any annotation effort. The annotations we conduct are for evaluation purposes only, and are *not* part of the learning approach. \\n\\n> Cost: Another issue with scalability is the cost associated with getting quality interaction data. As mentioned by authors, this small experiment took over a month to collect data and costed $11k USD (line 347). This also makes it harder to use the proposed method in real time.\\n\\nThe costs we list are for our experiments. We need interactions with users, so we use MTurk. This is what causes the high experimental costs. A company that deploys our approach will not suffer these costs, because the interaction will arise naturally from its models being used (i.e., in a product like ChatGPT). The reason the process took a month is because we worked with a group of workers on MTurk, and their availability was limited (also, training took time because we don\\u2019t have very strong GPUs). This time scale is completely immaterial to deploying our approach in a real environment \\u2013 it\\u2019s completely a *byproduct of an academic experiment*. \\n\\n## Response to Minor Issues Raised\\n\\n- We revised the introduction of MultiRef in the intro to give a better understanding of its main principles. \\n- LoRA implementation details are in Appendix B.3 and Appendix D. \\n- We update Figure 4 with different markers and line styles for HH and Control.\\n- Repeated variables (t): time and turn indeed refer to the same thing; we have updated to use \\u201cturn\\u201d consistently throughout. \\n- We fixed the typo.\\n\\n## Response to Questions Raised\\n\\n> How do you address overfitting given the extensive reuse of training data at each fine-tuning step?\\n\\nWe clarify why this is not an issue above. \\n\\n> Can you provide comparison of models across different evaluation metrics?\\n\\nWe are adding them. Please see the detailed answer above. \\n\\n> Have you tested the framework on other LLMs or tasks to confirm generalizability? How might the framework\\u2019s usability vary with stronger LLMs that provides less interaction data?\\n\\nPlease see the discussion above regarding the costs and feasibility within academic constraints. Furthermore, IDEFICS2-8B was by far the best open-weight multimodal LLM available when we conducted the experiments. Of course, now there are better models (it\\u2019s a fast paced space), with which we just expect faster and even more efficient learning. \\n\\n> Have you considered other optimization technique like Direct Preference Optimization (DPO) which uses the binary labeled data while fine-tuning the LLM?\\n\\nDPO required paired preferences. We don\\u2019t have this kind of data, and we can\\u2019t get it without additional annotation or overhead on the interaction. This is why we experiment with KTO, a related method that only requires the kind of single-example feedback we can decode (i.e., unpaired preference). \\n\\nWe will update the new PDF soon but want to kick-start an engaging discussion first. Please let us know if any questions remain, and consider raising the overall score if our responses are helpful.\"}", "{\"comment\": \"Many thanks for this. I understand the overall goal a bit better now. I feel like this entire narrative should be clarified in the paper to make sure that the reader understands:\\n\\n1. I why you selected this dataset as your reference and not any of the other that I've listed above. I believe that having this systematically reported in a table would be really valuable for a reader; \\n2. make sure that you justify the notion of learning continually in interaction and why this task is so well suited for it (in addition to the important point that you ideally don't need a task success measure);\\n3. clarify how you prevent overfitting and you assess task generalisation once your agent is deployed. Potentially reporting in the limitations section how this approach can generalise to the other tasks that I've reported might be useful for discussing future work. \\n\\nOverall, I believe that this paper represents a good contribution however, as highlighted by other reviewers, requires some edits to make sure that the narrative flows nicely and the goals and contributions are clear. As a result of my current perception of the paper, I've increased the score for contribution to the field.\"}", "{\"comment\": [\"Thank you for your detailed responses and for addressing my comments. While I understand the constraints of conducting research in academic settings, I still remain concerned about the risk of overfitting and scalability.\", \"The model's continual fine-tuning over limited data raises concerns about overfitting to patterns specific to the MULTIREF scenario. Although the authors argue that consistent task improvements demonstrate generalization, these improvements may stem from the model being repeatedly evaluated on a narrow and static task. This does not necessarily indicate true learning or generalization, as the model's gains could reflect a form of memorization or adaptation to the MULTIREF setup rather than the development of robust capabilities transferable to other domains. Validation in diverse settings is critical to confirm the broader applicability of RESPECT.\", \"While I acknowledge the cost and logistical challenges of testing RESPECT across varied scenarios, scalability is a critical requirement for any framework aiming for real-world deployment. The reliance on high-quality human interaction data makes scaling difficult, particularly in diverse or open-ended domains where such data may be sparse, noisy, or expensive to obtain. This limitation could hinder broader adoption, especially for applications requiring diverse language understanding or complex real-world reasoning.\", \"The observed plateau in task performance raises significant questions about the framework's ability to sustain improvement over extended rounds. While I appreciate the authors' explanation regarding hyperparameter expressivity, this further underscores the need for adaptive learning techniques that can handle evolving data distributions and maintain progress. Additionally, the continual fine-tuning approach introduces risks of catastrophic forgetting or confusion in learning. As new rounds overwrite prior knowledge, the model may improve on the immediate task while losing general capabilities or robustness. This is a fundamental concern in continual learning setups and warrants further investigation to ensure meaningful and sustained learning over time.\"]}", "{\"title\": \"Initial Response Part 2\", \"comment\": \"> The related work cites some interesting work related to using AI-generated feedback for improvement. However, the authors do not provide a baseline where this is explored for this game. For instance, the method proposed by Yuan et al 2024.\\n\\nWe do not use AI-generated feedback. This is an important aspect of our approach, because it means we do not rely on the ability of the model (or another model) to judge performance. The signal comes from followup human utterances. It\\u2019s decoded by the LLM. AI-generated feedback uses a model to judge the model output (i.e., actions). We don\\u2019t do that. We will highlight this difference in the related work, and note this separate thread of research. \\n\\n## Response to Additional Suggestions\\n\\nWe will review the manuscript to revise as needed. In the meantime, here are some answers: \\n\\n- Figure 4 Center shows the trend in the average number of dialogue turns over different rounds.\\n- We use Empirica to pair workers. Pairing is completely random \\u2013 same for pairing humans with each other and our model. This is to best simulate deployment. We do not make any effort to optimize for successful interactions, for example, because that would introduce biases into our deployment. More details on MTurk setup are in Appendix A3.1. \\n- We did not include the baseline term in REINFORCE implementation, see Eq (2). We opted for vanilla REINFORCE for its simplicity. PPO, for example, would have required a value function, and would likely be a poor fit for our offline setup. However, there\\u2019s room to try other techniques in future work. KTO represents a more complex and recent method, and our mixed results show the challenges of deploying it in our setup, highlighting an important area for future work. \\n- Typo is fixed.\\n- We will make a pass to make the use of terms (LLM, MLLMs, etc) consistent. \\n- Figure 1 and feedback type: the thumbs up in Figure 1 suggests positive feedback. We add explanations in Figure 1 caption.\\n\\n## Response to Raised Questions\\n\\n> It's not clear to me how the authors complete their fine-tuning considering that most of the training regimes are not designed for dialogue data specifically. For instance, if you have a dialogue of 4 turns, do you simply treat this as a single example or do you derive many examples for it? This is an important detail which I don't think is specified in the section that describes the training regime. \\n\\nPlease see our response to the relevant point above. Related details are specified in Section 3 on MultiRef.\\n\\n> What kind of REINFORCE implementation did you use? Did you adopt a baseline term? I think it's important to report more detail to aid reproducibility\\n\\nPlease see in the additional suggestions above.\\n\\n> Considering that the action space of the model is very limited, have you considered a form of token masking to improve the performance of your algorithms?\\n\\nModels learned the output format well from data alone without token masking. Post-hoc error analysis also revealed that outputs are valid. Errors are genuinely selecting a valid but wrong token. Constrained decoding helped KTO in round 3 a bit to counter the instability of B-KTO, but it wasn\\u2019t the model\\u2019s only problem (it just selected bad targets). It was not necessary otherwise.\\n\\nPlease let us know if you have further questions, and consider raising the overall score if our responses are helpful.\"}", "{\"comment\": \"Our use of the MultiRef scenario is a laboratory proxy for studying real-world human-model interactions. ReSpect is designed for such scenarios, where there is no game-completion signal for success. Imagine interactions between a human user and ChatGPT. There is no easily accessible ground truth task success or \\u201cgame\\u201d points for answering how to plan a family outing or touching up a paragraph (without external annotation \\u2013 which we do not assume). Instead, we propose to learn from implicit rewards humans emitted during the interaction. This gives us granular feedback and enables us to learn from even incomplete interactions, and without annotation effort. MultiRef allows us to study this in an academic setting because we can deploy multiple rounds, collect live interactions, and evaluate the policy exactly.\"}", "{\"summary\": [\"The paper presents RESPECT, a framework for LLMs to learn from implicit user feedback in multi-turn interactions. Rather than relying on external annotations, RESPECT enables models to retrospectively analyze past interactions and learn from cues like rephrased requests, signs of user approval, or frustration.\", \"This approach is applied in MULTIREF, a new multi-turn reference game where users instruct the model to select abstract shapes (tangrams), and the model gradually improves its accuracy based on decoded feedback signals.\", \"The study compares three learning strategies: supervised learning, REINFORCE, and KTO, finding that models using only positive feedback perform best.\"], \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": [\"The use of continual learning in the RESPECT framework demonstrates strong potential for developing LLMs that improve continuously from real-world interactions.\", \"The retrospective aspect of RESPECT is particularly compelling, as it enables models to learn from user corrective feedbacks.\"], \"weaknesses\": [\"The experiments are confined to the MULTIREF scenario with abstract tangram shapes. This limited scope raises questions about the generalizability of RESPECT to other domains. Applying RESPECT to diverse settings, such as conversational agents could demonstrate its robustness and adaptability across a broader range of applications, particularly those involving complex language or high-stakes interactions.\", \"There's a risk that the model might overfit to specific patterns of implicit feedback rather than truly improving at the task.\", \"The paper does not compare RESPECT to other established methods for learning from implicit feedback or continual learning. Without such comparisons, it's difficult to assess the relative merits of this approach For example, methods in RLHF using preference modeling or utility maximization strategies could serve as useful baselines.\", \"The feedback decoder relies on the model's ability to interpret implicit signals correctly. However, there's no guarantee that the model's interpretation aligns with the human's intended feedback. The paper would benefit from a more thorough analysis of cases where feedback may be misinterpreted and how this affects learning.\", \"While the paper shows improvement over six rounds for B-SUP, this may not be sufficient to fully understand long-term learning dynamics. The observed plateau and temporary decrease in performance warrant further investigation. Extended experiments over more rounds could provide insights into whether the approach continues to improve or stabilizes at a certain level.\"], \"questions\": [\"How well do you expect RESPECT to generalize to other domains or tasks beyond MultiRef? Have you tested it in any other scenarios?\", \"Have you considered ways to mitigate the impact of feedback misinterpretation on learning?\", \"Have you considered any potential ethical implications of learning from implicit human feedback, such as privacy concerns?\", \"The paper mentions that negative feedback signals are generally underutilized due to challenges in integrating them effectively. Would a more nuanced approach to weighting or categorizing negative feedback improve the model\\u2019s performance? eg some negative feedback could carry more importance than others. For instance, if the user strongly corrects an action (e.g., \\\"No, that's completely wrong\\\"), this feedback could be weighted more heavily than a milder form of dissatisfaction (e.g., \\\"Not quite right\\\"). Assigning different weights would allow the model to learn more from severe mistakes than minor ones.\", \"What are the computational costs of implementing RESPECT, especially the retrospective analysis of past interactions?\", \"In Figure 3, there appears to be a formatting issue or typo. It says \\\"positive or negative positive, neutral, or negative feedback,\\\" which seems confusing. Likely, this is unintended and should read either \\\"positive, neutral, or negative feedback\\\" or \\\"positive or negative feedback\\\".\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you for your response and further clarification on data usage and utility of framework. Based on our discussion, I would like to keep my scores.\\n\\nThe current setup does not effectively depict continual learning, which is the essence of the paper. As clarified by the authors, the new dataset is being used to fine-tune the base (original) model ($\\\\theta_{0}$) rather than the model from the previous step ($\\\\theta_{p}$). At each step, the model learns new information instead of building upon past knowledge in a continuous manner. This approach represents isolated learning rather than continual learning. The observed performance improvement is likely due to the increasing dataset size after each round, providing more samples for learning. This raises questions about the paper's core contribution.\"}", "{\"comment\": \"These are relatively minor edits, and we agree they will help the paper. Thanksgiving is upon us now (and most authors are offline), but we will incorporate these changes. The PDF is not updateable after today. We will include 1+2 in your list in a section that compares MultiRef to the datasets you listed, and do a pass over the intro to strengthen the justification. We will add an explicit discussion of limitations (something that got dropped in the last minute as we going for submission), among other things discussing generalization.\\n\\nGiven that we have converged on the needed edits and clarified all issues, we hope you will consider updating your score. I think that we both agree that this is an important contribution, and a perspective on learning that will benefit the ICLR program. \\n\\nThanks (and happy holiday, if you are celebrating)!\"}", "{\"comment\": \"Thank you for engaging. Below are our follow up to your questions:\\n\\n**Data usage and clarification on Figure 1**\\n\\nThis is a good point, and we see why Figure 1 can be confusing. We will fix this. Our intention is that the overarching $\\\\theta_{\\\\rho+1} \\\\leftarrow \\\\theta_{\\\\rho}$ indicates *the model in deployment* is updated from $\\\\theta_{\\\\rho}$ to $\\\\theta_{\\\\rho+1}$, and not referring to the training process.\\n\\n> This brings up the question, how is it any different than fine-tuning from scratch, all that is changing is the dataset. The model is not improving on the learned information, all that is happening is the model is learning new information. And the reason we see improvement over rounds is because the dataset size is being increased compared to the last round. Maybe I am missing something here and would like authors to clarify my misunderstanding.\\n\\nThe system improves, because the next model is better. The difference lies exactly in the dataset - it is non-stationary across continual learning rounds. The dataset $D_{\\\\le \\\\rho}$ is *not* collected by simply letting humans interact with the same model for more iterations, yielding homogeneous data points as assumed by static datasets. Instead, $D_{\\\\le \\\\rho}$ grows by including more interactions between humans and **newer/better** models (that were trained on existing interactions). There is a stark distribution shift in data, as illustrated in Figure 4 and linguistic analysis: increased task performance, fewer turns per interactions, reduced vocabulary size, etc. \\n\\nWe opted to train from scratch because it simplifies the setup. This is similar to [Kojima et al. 2021](https://arxiv.org/abs/2108.04812) and [Suhr and Artzi 2023](https://arxiv.org/abs/2212.09710). Fine-tuning complicates the setup because of repeated optimization of the model. There is one important exception: KTO. We follow as close as possible to the conventional KTO recipe, so there we fine-tune the same model again and again. \\n\\nWhy is continual learning so important for our study of implicit conversation signals? There are complex dynamics between the signals we use and how the system evolves by learning from them. For example, it could be the case that as the model gets better at the task, our feedback decoding process becomes less effective. We show this is not the case, and that the signals are present throughout the system\\u2019s lifetime (even as it improves) and our decoding approach remains effective and stable. Our setup also mimics the development-deploy cycles in practice, where the data collected in production often correspond to multiple prior model checkpoints.\\n\\n**Usability of the framework with better multimodal models**\\n\\nYes, we also expect models to get better. This only strengthens the importance of our approach. The interaction data we use arises from users using the system, and our approach shows an avenue to improve systems in exactly the scenario you raise: when we can\\u2019t get a lot of data to fine-tune in advance. What we show is that you can deploy such (M)LLMs, and have them improve from their interactions, even if they start bad. \\n\\nStrong models don\\u2019t make our approach redundant, because even if models perform perfectly when released, the world keeps changing, and humans keep changing how they use models. Our approach allows models to learn and adapt continually over their lifetime. Therefore, there will always be room to improve and to continue learning (just like humans do in a changing world \\ud83d\\ude42). \\n\\nOf course there is the middle ground where the model is better but still not perfect. Suppose by using IDEFICS3, we started at a MultiRef policy with a task success rate of roughly 50% (approximately IDEFICS2 at round 1) instead of 31% (round 0). Now we learned that our overall framework can support at least 30% improvements in the long run, without any additional annotation! We learned that the feedback decoder will keep working really well despite distribution shifts due to stronger models. We learned that towards later rounds we may need to tune hyperparameters, add model expressivity, increase the number of new interactions added per round, or with a learned feedback decoder to recover more subtle feedback (as we pointed out in discussion). All of these improvements, insights, and future directions on the promising topic of \\u201clearning from implicit conversational feedback\\u201d, are only revealed *after* this research. That\\u2019s the value of our work and we believe sharing them contributes to the knowledge base of this community, even if MultiRef (or any other prototypical task scenarios) becomes trivial for multimodal models eventually in the future.\\n\\nAgain thank you for engaging. Do let us know if we clarified on continual learning and addressed your concerns about usability with better multimodal models, and please consider raising the overall score if you find our answers helpful.\"}", "{\"summary\": \"The paper proposed a method to train a model with implicit human-in-the-loop feedback for a referential game. The proposed method first translate implicit human natural language feedback and quantize them into positive, (neutral), and negative labels, and then use the feedback to fine-tune the language model for decision making.\\n\\nThe experiment is situated in a referential game, where human is serving as a speaker to describe a subset of tangrams, and the model is serving as the listener to pick out the objects the human was describing.\", \"the_paper_experimented_with_three_learning_methods\": \"supervised learning, REINFORCE, and KTO. The models were initialized with pre-trained IDEFICS2-8B weights and fine-tuned with LoRA. Each model setting was fine-tuned with 0/1/2/3 rounds (B-SUP for 6 rounds), before being deployed in the online setting to have human-bot evaluation.\\n\\nThe paper observed that supervised learning method with binary quantization provided best performance, and that the feedback decoder's performance is relative stable across rounds and is consistent with human evaluation. The paper also observed that the human language is getting simpler with smaller vocabulary size and reduced utterance length across the rounds.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. The paper proposed a learning method, RESPECT that utilizes implicit human-in-the-loop feedback for explicit action improvement\\n2. The paper experimented with 3 learning methods: supervised learning, REINFORCE, and KTO\\n3. The paper conducted thorough experiments in a multimodal referential game\\n4. The paper conducted pre-training as well as online testing for iterative model improvement and evaluation\\n5. The paper is very well structured and well-written. The paper analyzed in detail about learning strategy tradeoffs, feedback label selections, feedback decoder evaluation, and language analysis.\", \"weaknesses\": [\"The paper wishes to highlight the contribution on 'continual learning' and model's iterative improvement with human's online feedback, but the actual experiments conducted is slightly misleading. The authors were careful to distinguish the differences between 'round' and 'turn.\", \"In the setup, each 'round' includes multiple 'turns' of interactions between a human and the bot.\", \"The model is retrained after each 'round', with the history of all previous 'rounds'\", \"After fine-tuning at the end of each 'round', the model is fixed and deployed for evaluation\", \"The main difference between proposed method versus the classic fine-tuning is the increasing context length during each fine-tuning round. It is unclear what the intended benefit is for the increasing context history?\"], \"for_example\": \"- Interaction history could help personalize the message or have a better understanding the counter party's message, if the bot was interacting with the exact same human. It was unclear if the bots were interacting with the same human users across different rounds.\\n- Interaction history could help the bot understand the task goal, through multiple rounds of probing and try-and-error (like RL), only if the bot was not briefed on what the task goal (referential game) was. According to the experiment setup (Figure 1, 2, 3), the model seems to have prior knowledge of the exact task goal.\\n\\nBeyond the examples illustrated above, the improved performance demonstrated in the paper might just be a result of fine-tuning with more data, and the expensive online evaluation among different turns showcased model's intermediate checkpoint performance. \\n\\nNevertheless, the paper proposed a new method that turns implicit human feedback into explicit rewards that could help improve model's performance. It is the 'Continual Learning' aspects that lacks sufficient support.\", \"questions\": \"Ln 425-426: According to Section 4.2, RL includes both positive and negative rewards. What might be the reasons that including extra rewards would 'encourage a uniform distribution'?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Initial Response Part 1\", \"comment\": \"We thank the reviewer for their helpful comments. We address the concerns below, and following this discussion, we will upload the edits (the changes are not yet in the PDF).\\n\\n## Response to Raised Weaknesses\\n\\n> The experiments are confined to the MULTIREF scenario with abstract tangram shapes.\\n\\n> How well do you expect RESPECT to generalize to other domains or tasks beyond MultiRef? Have you tested it in any other scenarios? (Question 1)\\n\\nIt\\u2019s important to consider the research trade-offs and implications of asking for more scenarios. We deployed multiple variants of our approach to interact with humans over thousands of interactions. This is the only way to study such methods. However, it\\u2019s also very expensive and time consuming. Yes, there may be insights to be gained by more scenarios, but would this kind of research be feasible in an academic setting? Where budgets are limited and teams are tiny. The implications of raising the costs even higher is pushing academic research out. The problem we study is important. As a research community, we must strike the right balance to make progress. Our research contributes insights and methods to drive this important direction forward. \\n\\nAs we discuss in the related work, [Don-Yehiya et al.](https://arxiv.org/abs/2407.10944) showed similar signals exist in scale (unlike us, they didn\\u2019t experiment with continual learning and a model bootstrapping itself). We believe their conclusion complements our work, showing that general scenarios elicit similar signals. \\n\\n\\n> There's a risk that the model might overfit to specific patterns of implicit feedback rather than truly improving at the task.\\n\\nThis is an empirical question. Our experiments empirically show consistent improvement in task performance over time, showing the model is improving at the task. We also extensively evaluate the feedback decoding. It\\u2019s not perfect, but very effective. There are few false negatives classified by the LLM feedback decoder (see the confusion matrix \\u2013 Figure 6). False negatives are higher though, suggesting there is room to improve further (only making the approach more effective). The task success rate improved significantly (by 51%) though, even with this imperfect reward decoder. \\n\\n> The paper does not compare RESPECT to other established methods for learning from implicit feedback or continual learning. \\n\\nWe are not certain what other implicit feedback methods the reviewer is referring to. We are happy to review any specific approach. \\n\\nIt\\u2019s important to note that continual learning is at times used to describe scenarios where models are adapted to new tasks. We use it in the sense of improving a model on its original task over time. So methods for domain adaptation are not relevant.\", \"with_regard_to_rlhf_ish_methods\": \"we experimented with KTO (a utility maximization method), and presented extensive deployment results with it. General RLHF methods and DPO-like methods are not applicable to our domain because they rely on paired preference data, but we have feedback on single outputs. This is why we used KTO, which allows using single-example feedback.\\n\\n> The feedback decoder relies on the model's ability to interpret implicit signals correctly. However, there's no guarantee that the model's interpretation aligns with the human's intended feedback. The paper would benefit from a more thorough analysis of cases where feedback may be misinterpreted and how this affects learning.\\n\\nWe conduct extensive human evaluation to quantitatively assess how well the interpreted signals align with human intended feedback and their impact on learning in Section 6. To summarize: \\n\\n- The LLM feedback decoder yields negligible false positives (<2%), and consistently reaches >60% accuracy considering all classes (diagonals in confusion matrix). This also suggests the training data for supervised systems are fairly clean (because they rely on positive signals only).\\n- The false negative rates are higher than false positives (around 15% for binary decoders, and 20% for ternary decoders throughout rounds). As a result, systems using ternary decoders recovered fewer data points as we discard predicted neutrals.\\n- To assess the impact of the quality of feedback decoders on learning, we adapt both binary decoders and ternary decoders for each optimization strategy. The deployment results (Figure 4) reveal that binary decoders have a slight edge over ternary decoders with B-SUP v.s. T-SUP, suggesting that more data is perhaps more important than cleaner data for supervised learning at least. The benefit is negligible for learning with RL systems.\\n- The human\\u2019s intended feedback is fairly unambiguous, supported by 93% agreement rate among three annotators.\\n\\nTo be continued in Part 2.\"}", "{\"title\": \"Initial Response Part 2\", \"comment\": \"> While the paper shows improvement over six rounds for B-SUP, this may not be sufficient to fully understand long-term learning dynamics. The observed plateau and temporary decrease in performance warrant further investigation. Extended experiments over more rounds could provide insights into whether the approach continues to improve or stabilizes at a certain level.\\n\\nIt\\u2019s important to consider the cost of each round of experiments, and the balance between the cost and insights. Each round costs $2000 for all six systems. Our experiments already show >50% task performance improvement, which tells us there is dramatic learning over time. There\\u2019s always the possibility that the next round will show even high performance, or a plateau. But this doesn\\u2019t change the fundamental answer to the research questions. \\n\\nWe discussed the plateau in Line 374 and additional experiments in Appendix D. To summarize: we observe the performance was capped by the LoRA adapters and conducted additional experiments with more expressive adapters, which recovered monotonous improvement in task success rate. We suspect the overall slowdown is because the optimal hyper parameters might have changed along continual learning (i.e., as the amount of data increased and/or the distribution shifted). \\n\\n## Response to Raised Questions\\n\\n> Have you considered ways to mitigate the impact of feedback misinterpretation on learning?\\n\\nSee the discussion of the feedback decoder above. In addition, there\\u2019s more room to develop better reward decoders. For example, given some investment in data annotation, one could fine-tune a separate LLM to decode the feedback (we focused on showing the LLM bootstrapping from its own interactions without such annotations, but this is a possible direction). \\n\\n> Have you considered any potential ethical implications of learning from implicit human feedback, such as privacy concerns?\\n\\nDeployment of our approach suffers the same risks as approaches that fine-tune on interaction data. This is something that companies do all the time (with annotations on the data), and they deploy different techniques to filter out the data to not include sensitive/private information. We will note this in the paper. \\n\\n> The paper mentions that negative feedback signals are generally underutilized due to challenges in integrating them effectively. Would a more nuanced approach to weighting or categorizing negative feedback improve the model\\u2019s performance? ...\\n\\nThis is a great idea for future work. We will note it in the discussion. This is the first work in this space, so we opted for simplicity. \\n\\n> What are the computational costs of implementing RESPECT, especially the retrospective analysis of past interactions?\\n\\nThe computational cost of retrospective analysis is minimal compared to retraining policy models. Here, prompting-based retrospection takes less than 3 minutes on 300 interactions (around 2400 actions) on a single A6000 GPU, without inference optimization. \\n\\n> In Figure 3, there appears to be a formatting issue or typo. It says \\\"positive or negative positive, neutral, or negative feedback,\\\" which seems confusing. Likely, this is unintended and should read either \\\"positive, neutral, or negative feedback\\\" or \\\"positive or negative feedback\\\".\\n\\nThis is not a typo. Figure 3 presents prompt templates for both the binary and ternary feedback decoders by color coding. We will make sure it\\u2019s clearer in the caption. \\n\\n\\nPlease let us know if you have further questions, and consider raising the overall score if our responses are helpful.\"}", "{\"title\": \"General Response\", \"comment\": \"We thank the reviewers for their comments. We appreciate the reviewers warm words about the work as \\u201cthe ideas \\u2026 is really interesting\\u201d, the interaction scenario being \\u201cvery useful for the future development of this domain\\u201d and this work \\u201cdemonstrates strong potential for developing LLMs that improve continuously from real-world interactions\\u201d. We fully respond to each review in a thread following their review. Below are answers to critical general questions:\\n\\n**Why continual learning?**\\n\\nContinual learning is at times used to describe scenarios where models are adapted to new tasks. We use it in the sense of improving a model on its original task over time. Studying our approach in a continual learning setting allows us to observe if the reward decoding is robust as the system improves and the data distribution shifts. We indeed show the decoding is robust and stable. Studying this kind of dynamics, as we do (with thousands of real-time human interactions), is critical to better understand the methods that learn from feedback. \\n\\n**Why one model and one task?** \\n\\nAll our experiments are human-in-the-loop, and all model interactions are conducted with humans. This is critical to reflect the true potential of interactions with humans for learning. However, this significantly complicates our studies and makes them very costly. This is a cost that goes beyond the cost of buying compute hardware, which researchers often consider, and is less influenced by just trying another task (if one that fits exists) or model (if an open-weights one exists). Our experiments cost >$11k, and this is without pilot experiments and studies. Such crowdsourcing funds are gone (it\\u2019s not hardware you can use again), and are very hard to raise. It\\u2019s important to balance utility vs. costs. Our studies provide clear answers to our research questions: multi-turn interaction data contains useful implicit signals, our method can decode it robustly, and it can learn from it rapidly. \\n\\nIn addition, one must also consider the engineering cost of building the scenarios for other tasks. MultiRef is an interactive platform, where humans are paired with each other or with models. \\n\\nAs one considers asking for more models and more tasks, it\\u2019s critical to consider what it means for academic research on such problems. If the bar is unnecessarily high for such studies, it simply says one shouldn\\u2019t study them in academic settings. This will be a terrible loss for academia, as interactions are increasingly core to how models are used and trained. \\n\\n**Why not DPO?**\\n\\nOur scenario doesn\\u2019t support DPO, because it does not provide paired preference data, as DPO requires. Our reward decoding process provides a reward for a single output. Paired preferences require two alternatives. This adds overhead to the interaction, which our approach does not. We experimented with KTO though.\"}", "{\"title\": \"Clarification on Continual Learning and answer questions\", \"comment\": \"We thank the reviewer for their helpful comments. We address the concerns below, and following this discussion we will edit the paper to make sure the continual learning aspect comes through clearly (the changes are not yet in the PDF).\\n\\n## Response to Raised Weaknesses\\n\\nWe define **continual learning** as the model improving over time on its task through interaction with human users. Our data collection and evaluation are both integral parts of our deployment. At each round, we deploy the current model to interact with humans. Each system has its own set of interactions, and each system is trained on the interactions that **it** had with the human users, so its own performance improvement influences its continual learning. In this process, the model continually improves over time from the signals obtained from its own interactions with humans. \\n\\nIn this sense, whether the task has multiple turns (like in our case) or not is not a factor of the setup. You could think of the model learning within single interactions (improving from earlier in the interaction to later), but this is not within the scope of our work. Our work is also not about adaptation, because we don\\u2019t maintain user-specific models. This is possible with our approach, but not within the scope of our experiments. \\n\\nThe term continual learning is used broadly, for example for domain adaptation. We will add a clarification very early in the paper to how we use it. We apologize for the confusion. Please let us know if there is anything we can add to clarify the setup. \\n\\nThere is also no direct relation between the progression of rounds and the content length (context in the sense of the LLM prompt, see Figure 11). The context is longer in later turns in the same interaction because the history is longer (more past turns), but this is true both in early and later rounds of the continual learning setup. \\n\\nWhy is continual learning so important for our study of implicit conversation signals? There are complex dynamics between the signals we use and how the system evolves by learning from them. For example, it could be the case that as the model gets better at the task, our feedback decoding process becomes worse. We show this is not the case, and that the signals are present throughout the system\\u2019s lifetime (even as it improves) and our decoding approach remains effective and stable. Our setup also mimics the development-deploy cycles in practice, where the data collected in production often correspond to multiple prior model checkpoints.\\n\\n## Response to Raised Questions\\n\\n> Ln 425-426: According to Section 4.2, RL includes both positive and negative rewards. What might be the reasons that including extra rewards would 'encourage a uniform distribution'?\\n\\nIf we understand correctly, you are asking about why the inclusion of negative rewards \\u201cencourages a uniform distribution\\u201d. This was observed by [Kojima et al. 2021](https://arxiv.org/abs/2108.04812). Essentially, by pushing down the probability of one output (the output with a negative reward), the model increases the probability of **all** other outputs. Because it\\u2019s a very large output space, this essentially pushes towards a uniform distribution. This is similar to unlearning objectives, which aim for the same (i.e., uniform distribution). \\n\\n\\nPlease let us know if any questions remain, and consider raising the overall score if our responses are helpful.\"}", "{\"title\": \"Response\", \"comment\": \"Thanks for the clarifications; these are really useful indeed. However, I would like to touch upon a point that does not convince me entirely:\\n\\n> We computed this (below). However, it\\u2019s very important to note that this is not a baseline, but an alternative that uses privileged information that our approach does not have access to, because our approach has no information about task success, but only the human utterances that the model was exposed to. This is a really important design decision, which is critical for applicability to real-life scenarios.\\n\\nI wouldn't consider this as privileged information. If you're playing a game, you always know at the end if you won or lost. So I consider your design choice as an artificial constraint that your model doesn't need to have. So technically using the task success as an additional reward is a totally legit decision which is not accessing *privileged* information.\\n\\nHappy to reconsider my score once this is clarified further. I also thank the authors for adding the additional experiment.\"}", "{\"metareview\": \"The authors present a way to farm supervision signals from LMs in multiturn interactions and use this as training data to improve the model. The novelty of this approach beyond reward-modeling / RLHF / RLAIF is not very clear. While authors present this as continual learning (which could be a way to distinguish from RLAIF i.e. continuing to learn post deployment), it still requires parameter updates and is trained over the whole dataset each time.\\n\\nI agree there is something novel here in that humans interacting with the system are not expressly asked to rate the interactions, it's more implicit. But this novelty is not motivated clearly or sufficient for a main conference paper.\", \"additional_comments_on_reviewer_discussion\": \"Significant engagement from reviewers -- sticking points are scalability if human oversight and clarity in scoping the contribution.\"}", "{\"comment\": \"Thank you for your response and addressing most of my queries. I would like the authors to further clarify on the framework and its usability with respect to their response.\\n\\n> We train from scratch in every round (6 rounds in total \\u2013 not every interaction step) for most systems. Training from scratch means: at every round, we fine-tune the initial IDEFICS2-8B parameters. We are not using the data to optimize again and again the same parameters. The measures we took to avoid overfitting are standard, just like any fine-tuning process, and the risk of overfitting is not different from any supervised learning scenario. Our evaluation is on unseen games (every round brings completely new games), showing strong generalization and improvement.\\n\\nI am a bit confused. My initial understanding was the framework finetunes the test LLM model (and not the base LLM model) based on all the data collected from past interactions. This flow has been depicted in figure 1 (on the very first page) of the paper. It clearly shows that $\\\\theta_{p}$ is updated to $\\\\theta_{p+1}$ based on all the aggregated data $D_{p}$. But now the authors claim that this is not the case\\u2014at each round, they fine-tune the initial IDEFICS2-8B parameters. This would mean that $\\\\theta_{0}$ is updated to $\\\\theta_{p+1}$. This brings up the question, how is it any different than fine-tuning from scratch, all that is changing is the dataset. The model is not improving on the learned information, all that is happening is model is learning new information. And the reason we see improvement over rounds is because the dataset size is being increased compared to the last round. Maybe I am missing something here and would like authors to clarify my misunderstanding.\\n\\n> > Have you tested the framework on other LLMs or tasks to confirm generalizability? How might the framework\\u2019s usability vary with stronger LLMs that provides less interaction data?\\n\\n> Please see the discussion above regarding the costs and feasibility within academic constraints. Furthermore, IDEFICS2-8B was by far the best open-weight multimodal LLM available when we conducted the experiments. Of course, now there are better models (it\\u2019s a fast paced space), with which we just expect faster and even more efficient learning.\\n\\nI understand the author's point on cost within academic constraints. My question was more stemmed on the usability of the framework with better multimodal models. As we have seen this trend with LLMs, we can expect better multimodal modals in few years. With those models we might not have enough interaction data for a particular task to fine-tune. Now, the authors says that with better models we can expect faster and efficient learning, but it is based on the fact that model is not performing well (the only way to get interaction data). If it performing well already then we can't expect expect faster and efficient learning. I would like to know authors take on this.\"}", "{\"title\": \"Response Part 1\", \"comment\": \"Thank you for engaging. We respond to your concerns individually below. We swapped the order because some explanation from bullet point 2 can help with bullet point 1.\\n\\n> Scalability is a critical requirement for any framework aiming for real-world deployment. The reliance on high-quality human interaction data makes scaling difficult, particularly in diverse or open-ended domains where such data may be sparse, noisy, or expensive to obtain\\u2026.\", \"on_scaling_data\": \"We agree that scalability is important. Scalability is actually the *strength* of learning from human-model deployment data. Key to our approach is that it relies on the interaction the system has with users during its deployment, as it completes tasks. So these interactions are already taking place. Our approach adds no overhead in terms of interaction or annotation on top of regular interactions with users. This is the most important property of our work: the signal arises naturally from the interactions the system already has with human users. This is different from contemporary RLHF methods, which rely on getting post-hoc preference annotations from third-party annotators. **This is why our approach is a fundamental improvement on current training recipes (e.g., RLHF), and this is why it naturally scales.** As people use the system, we get more signals, and the system autonomously learns from them \\u2013 without any annotation effort.\", \"on_high_quality_data\": \"[Don-Yehiya et al.](https://arxiv.org/abs/2407.10944) shows that similar learning signals exist in a large-scale real-world dataset [LMSYS-CHAT-1M](https://arxiv.org/abs/2309.11998) which covers general diverse domains. As we discuss in the related work section, their work complements our focus on continual learning, showing the prevalence of these signals in the interactions people have with existing LLMs. So, the signal is already out there, and it\\u2019s already at a very large scale. Together we think our research strengthens each other and provides justifications for further research in this novel and promising field and broader adaptation.\\n\\n\\n> These improvements may stem from the model being repeatedly evaluated on a narrow and static task. \\u2026 Validation in diverse settings is critical to confirm the broader applicability of RESPECT.\\n\\nWe agree a diversity of tasks is a natural and important next step. Please see above how the breadth of applicability is strengthened from concurrent work by [Don-Yehiya et al.](https://arxiv.org/abs/2407.10944). We contribute to the long-term applicability with non-stationary data distribution, and the potential of bootstrapping, ruling out concerns for distilling by learning from interactions of stronger models.\\n\\nIt\\u2019s important to note that although the domain is scoped, it doesn\\u2019t mean no learning is happening. It\\u2019s very common to fine-tune models on specific domains and tasks, and this kind of process is generally viewed as true learning.\\n\\nScoping down the task is necessary to study the long term effects of human-in-the-loop scenarios in an academic lab. We did not choose a general task with free-form text generation precisely due to the amount of data it can take per round to see an improvement, especially when we are interested in the long-term dynamics that demands multiple rounds of human-model deployment. MultiRef enables us to collect 330 live human-model interactions per round, and see a difference. Is there overfitting to this task? Yes, of course. We certainly do not expect our best MultiRef policy to answer a grade school math problem. The point of this work is not to instill tangram reasoning into IDEFICS2 and deliver model checkpoints, but to demonstrate the feasibility and effectiveness of bootstrapping new skills in LLMs from interactions in deployment, annotation free.\\n\\nWhat\\u2019s the alternative? Collecting 1 million conversations (like LMSYS-CHAT-1M) for generality? That\\u2019s collecting 24 such large-scale datasets by the end of our continual learning (not even considering pilot runs). We would love to see if ReSpect works there, but an experiment at that scale perhaps is only possible in industry. Even in industry smaller scale proofs of concepts are mandatory to secure sponsorships for broad deployments, and that is our work: showing it is possible in the long run to bootstrap from interactions alone. Our work does just this: enables the next step of scaling up. \\n\\nTo be continued in Part 2\"}", "{\"title\": \"Initial Response Part 1\", \"comment\": \"We thank the reviewer for their helpful comments. We address the concerns below, and following this discussion, we will upload the updated version (the changes are not yet in the PDF).\\n\\n## Response to Raised Weaknesses \\n\\n> Although I appreciate the rationale behind using tangrams, I wonder whether the authors could have tested this approach in more realistic reference game that are well-known in the community such as 20Q game [1], GuessWhat?! [2] or Photobook [3]. \\n\\nMultiRef has several key attributes of language interaction that are not well exposed by the other scenarios you list (although we do consider them interesting and related). MultiRef has a combinatorial solution space. This elicits the multi-turn interaction, with gradual instruction by the speaker and gradual solution construction by the listener. The impact of listener's actions is immediately seen by the speaker. The gradual construction and immediate visibility are essential to elicit the natural implicit feedback we observe. These are properties that are common in natural interactions. \\n\\n20Q and GuessWhat both are focused on binary questions, in contrast to our instructions that lead to more complex action space. They also both feature a much smaller solution space, which is guessed in one goal after the data is shared, not allowing for the immediate visibility of action impact after each turn. PhotoBook is more related to our scenario, but it also doesn\\u2019t provide real-time action visibility, and actions are only visible after both sides submit. More important: PhotoBook is a collaborative scenario where two equal partners negotiate a common ground. MultiRef is an instructional scenario, where the speaker instructs the listener. \\n\\nTherefore, while the scenarios you list are relevant, their focus is different, and they don\\u2019t expose the kind of real-time dynamic behavior we observe in real-time multi-turn interactions \\u2013 and this is what MultiRef is about. \\n\\nWe will add a discussion of the related scenarios. \\n\\n> It's not clear to me what is the rationale behind using KTO compared to DPO which is more established (e.g., used by Meta for Llama 3.2 tuning). Considering that the authors have access to positive and negative examples, I wonder whether they should try DPO instead considering that has been tested more for VLMs\\n\\nDPO requires paired preferences, that\\u2019s **two** potential outputs for exactly the same input, one marked as winner and the other as loser. We don\\u2019t have this kind of data, so DPO is not applicable. This is why we chose KTO. \\n\\n> It's not clear to me how the authors complete their fine-tuning considering that most of the training regimes are not designed for dialogue data specifically. See Question 1 as well for details.\\n\\nWe create an example for each utterance. Each such example includes the set of images, their selection status, the history of the interaction up to the current utterance, and the utterance. The model output is the action generated by the model. The learning algorithm sees each such \\u201cexample\\u201d as separate. This data formulation is identical to what the model \\u201csees\\u201d during deployment.\\n\\n> I think the authors are missing a simpler baseline that is fine-tuned using the final reward (i.e., whether you win or not the game) as a reward as was done in previous work [4]. \\n\\nWe computed this (below). However, it\\u2019s very important to note that this is not a baseline, but an alternative that **uses privileged information that our approach does not have access to**, because our approach has no information about task success, but only the human utterances that the model was exposed to. This is a really important design decision, which is critical for applicability to real-life scenarios. \\n\\nWe simulate Interaction-Oracle learning on B-SUP datasets. We report the exact match rate on an offline validation set of 344 turns from human-human games (not exactly in the distribution of human-bot games, but this is informative).\\n\\n| Round | B-SUP | Interaction-Oracle\\n| ----------- | --------- | --------- \\n| 1 | 43.3% | 43.8%\\n| 2 | 45.9% | 43.3%\\n| 3 | 49.1% | 46.2%\\n| 4 | 48.5% | 47.3%\\n| 5 | 53.1% | 47.1%\\n| 6 | 52.0% | 47.7%\\n\\nProviding clean and granular progress rewards achieves higher validation rates except for the first round (even though B-SUP has no idea about task success). The model picks up intermediate mistakes in the process by filtering solely based on the final task success. We also overestimate the performance of Interaction-Oracle with offline simulations here, because they were trained on interactions collected from a better model (potential model distillation) \\u2013 so please consider these numbers an overestimate. \\n\\nTo be continued in Part 2.\"}", "{\"comment\": \"We are sorry, but this is wrong. You follow a narrow definition of continual learning, which misses the whole point of what is continual learning by focusing on a marginal implementation detail (that should be empirically driven, and this is how we decided it). The system is continually learning by interacting with humans, because the data is coming from the current generation of the system interacting with users (this is how the data is collected). Whether we fine-tune the most recent model, or just train with the aggregated data so far from the initial parameters is immaterial to if continual learning takes place or not. Our choice simplifies optimization problems and avoids tuning regularization methods like KL or rehearsal. Your narrow definition is misaligned with a lot of existing work (see the most recent EMNLP best paper award and many other papers).\\n\\n> The observed performance improvement is likely due to the increasing dataset size\\n\\nThe new data comes from newer generations of the systems interacting with people. It does not appear form thin air or additional annotation. Even if this data is not from a different distribution (which is clearly the case because performance increases over time), the source of the data matters.\"}", "{\"title\": \"Response Part 2\", \"comment\": \"> The observed plateau in task performance raises significant questions about the framework's ability to sustain improvement over extended rounds. \\u2026 This further underscores the need for adaptive learning techniques that can handle evolving data distributions and maintain progress.\\n\\nWe agree with the importance of addressing catastrophic forgetting and new adaptive learning techniques. In fact, we will be exhilarated to see new breakthroughs in that area. However, that is simply not the research question of this work: we *do not* propose a general continual learning adaptation technique that solves or aims to solve catastrophic forgetting once and for all. We identify an underappreciated learning signal (implicit human feedback from interactions themselves), and what long-term dynamics it creates when we learn from rollouts of shifting distributions from its own. Continual learning is a lens through which we study the impact of bootstrapping by training on interactions from deployment, and to mimic the develop-deploy cycles of conversation models in production, but continual learning itself is not a research target for its own merit in this work.\\n\\nAs a work that *implements (not studies)* continual learning, we took measures to avoid catastrophic forgetting in the face of data distribution shifts (e.g. better task performance, reduced vocabulary sizes, shorter utterances) via data accumulation and rehearsal. To clarify, we did not do \\u201ccontinual fine-tuning\\u201d. For each round $\\\\rho$, we load the same checkpoint from IDEFICS2-8b and train on all data so far $D_{\\\\le \\\\rho}$ for supervised and RL systems. Training from scratch is a common technique to overcome the loss of model elasticity in continual learning setups (exactly the problem as you pointed out). This is similar to [Kojima et al. 2021](https://arxiv.org/abs/2108.04812) and [Suhr and Artzi 2023](https://arxiv.org/abs/2212.09710). Empirically these techniques also worked well in MultiRef: we observe strong improvement of an absolute 51% after 6 rounds. Addressing catastrophic forgetting in continual learning is an important research topic, but tangential to our contribution on learning from implicit feedback signals.\\n\\nLastly, regarding the plateau, it is not due to catastrophic forgetting, because of the aforementioned guardrails, and the fact that we focus on improving one single hard task. We have discussed extensively why it occurred (model expressivity, hyperparameters for continual learning), what helped fix it partially (better LoRA adapters), pointed to future directions such as learning a feedback decoder to uncover more subtle feedback (more nuanced decoder as you pointed out previously). All of these insights, analysis, and takeaways are only made possible *after* this research, i.e., deploying an imperfect framework, for 6 rounds, collecting thousands of human-model interactions, witnessing a significant improvement from 31% to 82%. The insights and lessons from this project are a novel contribution to the community in its own right.\\nWhat does it take to achieve the perfect 100%? There is always a point of diminishing returns in a research project and we have offered all the insights for future work in the discussion section. We will leave the last 18% as an exercise for thoughtful readers :) Regardless the outcome of new endeavours, they won't change our answer to the research question, that we *can already* effectively learn from interactions in deployments, through multiple rounds and without a single expert annotation. \\n\\nAgain thank you for engaging. Let us know if we have clarified why overfitting to MultiRef is from an economical experiment design, and that data scalability is actually a strength of learning from human-model interactions in deployment. Please consider raising the overall score if you find our answers helpful.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"summary\": \"In this paper, the authors propose RESPECT, a new method for refining multimodal language models (in this case vision and language models) from interaction data automatically generated by a model while interacting with another agent when solving a referential task.\\n\\nTo tackle this problem, the authors propose a new benchmark called MultiREF which requires agents to manipulate tangrams, specific abstract shapes that are well-known in the community for their ability to elicit interesting communicative grounding phenomena due to their intrinsic ambiguity. \\n\\nBased on this dataset, the authors focus on a very specific training regime which alternates two phases: 1) retrospection = decoding implicit feedback from past interactions by means of a classifier which derives feedback labels (i.e., positive, neutral, negative); 2) learning = refining the model using the feedback received from the previous stage. Because the authors simplify the prediction task to a classification task, they argue that Step 1) can be simply performed by a carefully prompted model to perform a simple binary/three-way classification task. For the second step, the authors test different learning strategies such as a) supervised learning from positive data only; b) Online reinforcement learning (using REINFORCE) with a hand-crafted reward function which leverages the labels derived by the classifier from Step 1. c) Kahneman-Tversky Optimisation (KTO) as a form of reinforcement learning from feedback (in this case, AI feedback). \\n\\nThe authors set up a really complex evaluation with real users that interact with the system in real-time. In their evaluation, they start from IDEFICS2-8B model as their initialisation and use a frozen IDEFICS2-8B as the feedback decoder. From their evaluation, seems that there is still a long way to go to develop robust training regimes that can facilitate the type of adaptation required for these interactive tasks. In fact, the supervised learning variant seems to be the most robust which relies purely on positive examples and ignores negative ones.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"4\", \"strengths\": \"1. Interesting evaluation that tests the system with real users over a period of 4 weeks. This represents a great effort to showcase the strengths and weaknesses of the different training regimes.\\n2. Very interesting idea to simplify the task of the \\\"critic\\\" to fixed labels that can be used for very specific training regimes\\n3. The authors test different training regimes that are well known in the community such as Supervised Learning, REINFORCE and KTO\", \"weaknesses\": \"1. Although I appreciate the rationale behind using tangrams, I wonder whether the authors could have tested this approach in more realistic reference game that are well-known in the community such as 20Q game [1], GuessWhat?! [2] or Photobook [3]. I feel like this would have given a much broader perspective on the robustness and reliability of the proposed training regimes for more complex language generation tasks\\n\\n2. It's not clear to me what is the rationale behind using KTO compared to DPO which is more established (e.g., used by Meta for Llama 3.2 tuning). Considering that the authors have access to positive and negative examples, I wonder whether they should try DPO instead considering that has been tested more for VLMs\\n\\n3. It's not clear to me how the authors complete their fine-tuning considering that most of the training regimes are not designed for dialogue data specifically. See Question 1 as well for details.\\n\\n4. I think the authors are missing a simpler baseline that is fine-tuned using the final reward (i.e., whether you win or not the game) as a reward as was done in previous work [4]. \\n\\n5. The related work cites some interesting work related to using AI-generated feedback for improvement. However, the authors do not provide a baseline where this is explored for this game. For instance, the method proposed by Yuan et al 2024.\\n\\n\\n## Additional suggestions to improve the manuscript\\n\\nI would suggest the author refine their manuscript by clarifying some aspects of their methodology by answering my questions below. At the same time, I would suggest them to implement the following changes:\\n\\n- Extend the analysis of the paper with average dialogue turns over the different rounds; this can shed light on the ability of the models to improve their gameplay strategy over time\\n\\n- Clarify how you pair different users on MTurk to maximise usage. For example, there are dedicated methods to do so (e.g., [5]) but it's not clear to me how you've arranged this. \\n\\n- Provide more details regarding the REINFORCE baseline\\n\\n- Typo on Line 244 \\\"process transformers\\\" \\n\\n- It's confusing to see that you call your models (which are Vision and Language Models) as LLMs\\n\\n- I would suggest improving Figure 1 to show the type of feedback that the system is leveraging for the improvement.\\n\\n## References\\n\\n[1]: Bertolazzi, L., Mazzaccara, D., Merlo, F., & Bernardi, R. (2023, September). Chatgpt\\u2019s information seeking strategy: Insights from the 20-questions game. In Proceedings of the 16th International Natural Language Generation Conference (pp. 153-162).\\n\\n[2]: De Vries, H., Strub, F., Chandar, S., Pietquin, O., Larochelle, H., & Courville, A. (2017). Guesswhat?! visual object discovery through multi-modal dialogue. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (pp. 5503-5512).\\n\\n[3]: Haber, J., Baumg\\u00e4rtner, T., Takmaz, E., Gelderloos, L., Bruni, E., & Fern\\u00e1ndez, R. (2019, July). The PhotoBook Dataset: Building Common Ground through Visually-Grounded Dialogue. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (pp. 1895-1910).\\n\\n[4]: Strub, F., De Vries, H., Mary, J., Piot, B., Courvile, A., & Pietquin, O. (2017, August). End-to-end optimization of goal-driven and visually grounded dialogue systems. In Proceedings of the 26th International Joint Conference on Artificial Intelligence (pp. 2765-2771).\\n\\n[5]: G\\u00f6tze, J., Paetzel-Pr\\u00fcsmann, M., Liermann, W., Diekmann, T., & Schlangen, D. (2022, June). The slurk Interaction Server Framework: Better Data for Better Dialog Models. In Proceedings of the Thirteenth Language Resources and Evaluation Conference (pp. 4069-4078).\", \"questions\": \"1. It's not clear to me how the authors complete their fine-tuning considering that most of the training regimes are not designed for dialogue data specifically. For instance, if you have a dialogue of 4 turns, do you simply treat this as a single example or do you derive many examples for it? This is an important detail which I don't think is specified in the section that describes the training regime.\\n\\n2. What kind of REINFORCE implementation did you use? Did you adopt a baseline term? I think it's important to report more detail to aid reproducibility\\n\\n3. Considering that the action space of the model is very limited, have you considered a form of token masking to improve the performance of your algorithms?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The author proposes a novel framework, Retrospective learning from past interactions (RESPECT), for improving the LLMs based on signals from past interactions via retrospection. They also contributed with a task, Multi-turn Grounded Interaction Scenario (MULTIREF), a conversational interaction scenario where two partners, a speaker and a listener, coordinate on the selection of a set of items. They further fine-tuned an LLM with different optimization techniques with the data collected from the proposed task and framework.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": [\"The idea of learning from past mistakes is really interesting and the proposed framework, RESPECT, doesn't depend on the optimization strategy. As highlighted in the paper, this framework can be used with various optimization strategy (Supervised Learning, Reinforcement Learning, Utility Maximization).\", \"The paper also contributed with a new task, MULTIREF, which will be very useful for the future development of this domain.\"], \"weaknesses\": [\"**Major:**\", \"**Excessive use of training data:** The proposed method relies heavily on data. The model is fine-tuned at each step with all the interaction data acquired from past steps. Now, although the authors mention that they are taking measures to avoid overfitting (lines 246-248), this much repeated data usage would eventually result in overfitting.\", \"**Lack of metric evaluation:** Although the authors showcases various observations and results through plots and confusion matrix, they lack the tables for comparing different metics. Having those results will significantly boost the paper quality.\", \"**Lack of generalizability of proposed method:**\", \"**Across different LLMs:** The proposed framework is only tested on one LLM (IDEFICS2-8B). Now, although this framework can be applied over other LLMs, it is unclear whether it will boost their performance or not. One reason authors might have seen such improvement is because the tested LLM is bad at that particular task. If we have a very good LLM then this framework might not help much as we will have less interaction data to fine-tune model on.\", \"**Across different tasks:** Another interesting extension of the proposed method can be over different tasks. Currently authors have only tested over a particular task but it would be interesting to see if it can be extended over other tasks like summarization (authors have highlighted this as future work in discussion).\", \"**Scalability:**\", \"**Heavy reliance on human feedback:** The proposed framework relies heavily on human feedback. Although the authors have countered the problem of annotating the responses, the problem of getting good interaction data still remains a crucial problem, making it hard to scale.\", \"**Cost:** Another issue with scalability is the cost associated with getting quality interaction data. As mentioned by authors, this small experiment took over a month to collect data and costed $11k USD (line 347). This also makes it harder to use the proposed method in real time.\", \"**Minor:**\", \"Adding information about MULTIREF in the introduction will help in understanding the contribution of the paper.\", \"Information on LoRA configuration used for fine-tuning is missing.\", \"Very similar labels are used for control and HH data in figure 4, making it hard to interpret. Maybe changing it with something else will make it more interpretable.\", \"**Repeated variable:** Authors have repeated the use of variable t. At line 90 it represents time while at line 141 it represents turn and at line 145 it is again used as time.\", \"**Typo:** Fullstop (.) missing in line 422. \\\"(supervised vs. RL/KTO) Overall\\\" &#8594; \\\"(supervised vs. RL/KTO). Overall\\\"\"], \"questions\": \"1. How do you address overfitting given the extensive reuse of training data at each fine-tuning step?\\n2. Can you provide comparison of models across different evaluation metrics?\\n3. Have you tested the framework on other LLMs or tasks to confirm generalizability? How might the framework\\u2019s usability vary with stronger LLMs that provides less interaction data?\\n4. Have you considered other optimization technique like Direct Preference Optimization (DPO) which uses the binary labeled data while fine-tuning the LLM?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Initial Response Part 1\", \"comment\": \"Thank you for your feedback. We have modified our manuscript to incorporate it (uploading soon following this discussion). Regarding the concerns you raised:\\n\\n## Response to Raised Weaknesses \\n\\n> Excessive use of training data: The proposed method relies heavily on data. The model is fine-tuned at each step with all the interaction data acquired from past steps. Now, although the authors mention that they are taking measures to avoid overfitting (lines 246-248), this much repeated data usage would eventually result in overfitting.\\n\\nWe train from **scratch** in every round (6 rounds in total \\u2013 not every interaction step) for most systems. Training from scratch means: at every round, we fine-tune the initial IDEFICS2-8B parameters. We are not using the data to optimize again and again the same parameters. The measures we took to avoid overfitting are standard, just like any fine-tuning process, and the risk of overfitting is not different from any supervised learning scenario. Our evaluation is on unseen games (every round brings completely new games), showing strong generalization and improvement. \\n\\n> Lack of metric evaluation: Although the authors showcases various observations and results through plots and confusion matrix, they lack the tables for comparing different metics. Having those results will significantly boost the paper quality.\\n\\nWe will add 8 tables in Appendix E accompanying Figure 4,5,6. Figures 4,5,6 illustrate the trends. We included specific numbers in the Results section with analysis. The confusion matrix is already labeled with numbers so we will highlight the key observations in the caption as well.\\n\\n> Lack of generalizability ... across different LLMs: The proposed framework is only tested on one LLM (IDEFICS2-8B). Now, although this framework can be applied over other LLMs, it is unclear whether it will boost their performance or not. One reason authors might have seen such improvement is because the tested LLM is bad at that particular task. If we have a very good LLM then this framework might not help much as we will have less interaction data to fine-tune model on.\\n\\nOur goal is to show implicit conversational signals can drive learning. These signals exist beyond our setup (as we discuss in related work). The poor initial performance is a feature of the experimental design \\u2013 it allows us to show an effect. Other LLMs might be good at the task, but they will be bad at other tasks, and our approach will allow them to improve over time. Our research question is: Can implicit signals from interactions be leveraged and sustain continuous improvements in deployment? To show its effectiveness, we carefully selected the multi-turn reference scenario from CogSci and tangrams that are unfamiliar to modern pretrained vision LMs (shown in Ji et al.), such that (a) the scenario can be deployed in an academic lab and (b) progress is measurable with a reasonable amount of data. This methodology leads to our evaluation with one model and one task.\\n\\nOf course, it is always useful to try different models, but it is also important to consider the costs of such experiments \\u2013 with humans in the loop. What are the implications of demanding many LLMs and many tasks with such studies? It will simply be impossible to run such research in academic labs. Even industry labs are not likely to just put millions of dollars into a speculative new approach without a proof of concept. This is what our research is. \\n\\n> Lack of generalizability ... across different tasks: Another interesting extension of the proposed method can be over different tasks. Currently authors have only tested over a particular task but it would be interesting to see if it can be extended over other tasks like summarization (authors have highlighted this as future work in discussion).\\n\\nWe agree with the reviewer on experimenting with different scenarios. As we discussed in our related work, a concurrent work by [Don-Yehiya et al.](https://arxiv.org/abs/2407.10944) showed these signals exist in lmsys-chat-1m. So this tells us that these signals are general. We show how they can be used for an LLM to improve its own behavior. Similar to trying different models, more tasks have extreme engineering and experimental implications \\u2013 this simply makes this kind of work impossible, even though it\\u2019s an important and impactful research avenue. \\n\\nTo be continued in Part 2.\"}" ] }

[ "Yakir Oz", "Gilad Yehudai", "Gal Vardi", "Itai Antebi", "michal Irani", "Niv Haim" ]

[ "data reconstruction", "memorization", "privacy" ]

{ "note_id": [ "vpIFbtEPK3", "uSJfgBnKWO", "pxVfNoKVzB", "ouo0bkJH7c", "niSVgFh6Rh", "mttVB8JhBG", "fzN93QADWg", "eJ0njrP42t", "bWzF9xAORu", "SqDkPNrGm7", "R2Foooim5d", "KoMayPlzPE", "Ip7Rl5qcUh", "FewtO7Ild4", "AAABDrvaJF", "8U2WZmm5zX" ], "note_type": [ "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "decision", "meta_review", "official_review", "official_comment" ], "note_created": [ 1732439186388, 1730103579557, 1732045592122, 1732709008562, 1732045276151, 1732435662022, 1730684085603, 1732045205876, 1732045375253, 1733070741386, 1733070684220, 1730211307040, 1737523458112, 1734360180226, 1731165454902, 1732582660070 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission1560/Reviewer_c8yx" ], [ "ICLR.cc/2025/Conference/Submission1560/Reviewer_VVGm" ], [ "ICLR.cc/2025/Conference/Submission1560/Authors" ], [ "ICLR.cc/2025/Conference/Submission1560/Reviewer_VVGm" ], [ "ICLR.cc/2025/Conference/Submission1560/Authors" ], [ "ICLR.cc/2025/Conference/Submission1560/Authors" ], [ "ICLR.cc/2025/Conference/Submission1560/Reviewer_hzzB" ], [ "ICLR.cc/2025/Conference/Submission1560/Authors" ], [ "ICLR.cc/2025/Conference/Submission1560/Authors" ], [ "ICLR.cc/2025/Conference/Submission1560/Authors" ], [ "ICLR.cc/2025/Conference/Submission1560/Authors" ], [ "ICLR.cc/2025/Conference/Submission1560/Reviewer_c8yx" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission1560/Area_Chair_umNp" ], [ "ICLR.cc/2025/Conference/Submission1560/Reviewer_qRiu" ], [ "ICLR.cc/2025/Conference/Submission1560/Reviewer_qRiu" ] ], "structured_content_str": [ "{\"comment\": \"Thank you for author\\u2018s response. Considering author\\u2018s feedback and the limitations of the paper, I will maintain my score.\"}", "{\"summary\": \"This paper proposes a new method for training data reconstruction. Different from other works, this paper proposes to recover the image embedding first, and then employs a inversion network to reconstruct the images. Especially, this paper proposes to use the clustering-based approach for effectively identifying training samples. Experiments show the performance of the proposed method.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"The proposed method has show better performance compared with baselines, on the datasets of Food-101 and iNaturalist.\", \"weaknesses\": \"1. The experiments are only conducted on the CLIP and DINO which has served as the common components for the image diffusion models. However, I think the proposed network should be suitable for various networks. I wonder the training data reconstruction with other basic networks, like ResNet. And the authors should explain why they focused on transformer-based models in the experimental section.\\n\\n2. This paper claims to be suitable for the reconstruction of high-resolution images, while the experiments are only conducted on the datasets with low resolution like 224 (why 224x224 fits that definition of \\\"high resolution\\\" in this paper?). I wonder the performance with higher resolution like 512.\\n\\n3. There is no quantitative metric to compare the reconstruction effects between the proposed method and baselines. For example, using the reconstructed images can lead to a model with the similar accuracy of the original model?\\n\\n4. The presentation of this paper is not good. For example, the texts in Fig.9 has be out of the width constraint, which could be resized.\", \"questions\": \"1. What is the performance of the proposed method with more types of networks besides the transformer?\\n\\n2. What is the reconstruction performance on images with higher resolution?\\n\\n3. Can the authors provide more reliable quantitative metrics?\", \"flag_for_ethics_review\": \"['Yes, Privacy, security and safety']\", \"details_of_ethics_concerns\": \"The proposed method can be utilized to steal the training data of trained models, influencing the privacy requirement.\", \"rating\": \"5\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": [\"We thank the reviewer for the thorough feedback.\", \"Answers for \\u201cWeaknesses\\u201d part:\", \"This is a good point. Previous works were successful in reconstructing training data only from very small-scale networks (MLPs and CNNs) and small-scale datasets (MNIST and CIFAR), see lines 33-36. In this work, we focused on popular backbone models for transfer learning. As we report in lines 494-497, we attempted our method on larger CNNs (e.g., VGG), and it wasn\\u2019t as successful. Reconstructing data from Transformer-based backbones proved more successful. In this respect, it is important to note that transformer-based backbones are very popular for feature extraction for transfer learning tasks in practice, and we believe that reconstructing data from transformer-based feature extractors is an important contribution on its own. We emphasize that our work is the first to reconstruct training images from real-world networks, rather than small-scale and unrealistic models as in previous works. We agree that exploring data reconstruction from other backbones is an important research direction, but beyond the scope of this paper.\", \"Thank you for pointing this out. Please note that all previous works on data reconstruction from trained models were done on much lower-resolution datasets such as MNIST (28x28) or CIFAR (32x32). Our work is the first to reconstruct training images from higher-resolution images (224x224), comparable to Imagenet-size images. Additionally, note that most standard backbones (like the ones we use), are trained at this resolution, and that typically, input images are first resized (and center-cropped) to this same resolution (224 pixels). Thus, this is a standard resolution. We agree that data reconstruction from higher resolution models is important, but this is beyond the scope of our paper.\", \"We make a great effort to establish a quantitative method for comparing reconstruction quality. We kindly refer the reviewer to the \\u201cQuantitative Evaluation\\u201d paragraph in lines 346-400. In particular, we compare the reconstruction quality between 6 different metrics (Figure 5). We additionally compare to activation maximization method (Appendix A.10) which is applicable in our setting. We emphasize that since this is the first work to successfully reconstruct training data from higher resolution images than CIFAR, it is difficult to compare to other methods, as there are none besides activation maximization. The reviewer\\u2019s suggestion is very good, we attempted this but it wasn\\u2019t successful. A possible reason may be, that quite surprisingly, the output of the network on reconstructed images can be very different from network output on corresponding original images, even though they are visually similar. There seems to be an inherent scaling issue in the reconstruction optimization. It is an interesting direction for future research.\", \"Thank you for your comment. We will resize Figure 9 to fit the width constraints. If there are other issues with the presentation we would be happy to fix them.\"], \"questions\": [\"The KKT-based reconstruction is known to be successful for different types of networks (see [Haim et al. 2022], [Buzaglo et al. 2023]). Our method works for networks that can be \\u201cinverted\\u201d similarly to CLIP/DINO-ViT/ViT, hence we currently don\\u2019t know how to extend it directly to ResNets or other types of networks. This is an active field of research that we also work on, although it seems a more difficult task.\", \"As we mentioned, although it would be interesting to test even higher-resolution images, this is currently beyond the scope of our paper. The resolution 224X224 is used in most standard backbones (like the ones we use), and it is the Imagenet resolution. We think it is surprising by itself that it is possible to reconstruct images beyond a resolution of 32x32, which is the highest resolution achieved by previous works in the field. (also see our answer to weaknesses (2)).\", \"Please see our answer to weakness (3).\"]}", "{\"title\": \"Review comments\", \"comment\": \"After reading the response from the authors, some concerns have not be addressed, including the experiments on more types of backbones, experiments with higher resolutions. Therefore, I think this paper still needs to make further explorations, making the contribution of this paper more solid. I will keep my original rating.\"}", "{\"comment\": [\"We thank the reviewer for the feedback.\", \"The main contribution is in demonstrating the vulnerability to training-data reconstruction of models that are trained in a transfer learning manner using embeddings of widely-used backbone models. This is a very popular approach for training classifiers on tasks with limited data, and exposing this vulnerability should be of major concern to practitioners. We provide an abundance of experimental evidence to support our claims. Another major contribution is the clustering-based reconstruction approach, which is very important in making such reconstruction attacks useful in practical settings. Note that many works in the field are using the training data in order to validate their attack. Our clustering-based approach alleviates this requirement.\", \"Thank you for pointing this out. Please note that all previous works on data reconstruction from trained models were done on much lower-resolution datasets such as MNIST (28x28) or CIFAR (32x32). Our work is the first to reconstruct training images from higher-resolution images (224x224), comparable to Imagenet-size images. Moreover, note that most standard backbones (like the ones we use) are trained at this resolution, and that typically, input images are first resized (and center-cropped) to this same resolution (224 pixels). Thus, this is a standard resolution. We agree that data reconstruction from higher-resolution models is important, but this is beyond the scope of our paper.\", \"We compare our method to two activation maximization methods (Appendix A.10) which are applicable in our setting. Since this is the first work to successfully reconstruct training data from higher resolution images than CIFAR, there are no other existing methods that we are aware of, and designed for direct comparison in such settings.\"]}", "{\"title\": \"Response to all the reviewers\", \"comment\": \"Again, we thank the reviewers for their comments. We hope that our response has addressed all the issues raised by the reviewers and that they will consider updating their scores accordingly. If there are remaining questions, we would of course be happy to address those before the public discussion phase ends.\\n\\nMany thanks, The authors\"}", "{\"summary\": \"This paper studies the training data reconstruction problem. Different from previous methods, this paper studies the data reconstruction from models trained in a transfer learning approach, and claims the first approach to reconstruct images from the latent features. The experimental section showcases comprehensive results to verify the efficacy of proposed method in different datasets and backbone networks.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"1\", \"strengths\": \"1. This paper is easy to follow.\\n2. I appreciate the comprehensive experiments performed to analyze the effectiveness of the proposed approach.\", \"weaknesses\": \"1. The technique contribution is limited. This paper aims to reconstruct images from latent features (embeddings), however, the key components used for this purpose are borrowed from previous works. Specifically, it use \\u2018\\u2019Reconstructing training data from trained neural networks\\u2019\\u2019 to reconstruct embeddings, followed by \\u2018\\u2019Splicing vit features for semantic appearance transfer\\u2019\\u2019 to convert embeddings into RGB images.\\n2. The resolution of reconstructed images are still low (224x224).\\n3. The experiment part contains results of this method, but without any comparison with other approaches.\", \"questions\": \"See weakness\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"We thank the reviewer for the thorough and positive feedback.\", \"weaknesses\": [\"While we agree that some results look similar to class average, many are very similar to the original image itself (Fig. 3). We also think that such a comparison should be with respect to other techniques. In appendix A.10 we compare our results with two activation-maximization baselines on the same models. Note that the results from these techniques are far from being similar to any of the original images or their class representatives.\", \"Indeed, the choice of the backbone is significant to the success of the reconstruction, and especially for this reason we focused our experiments on very popular backbones for vision tasks (DINO-ViT, ViT and CLIP).\", \"The reviewer is correct in noticing that the inversion process is a bottleneck of our method. However, our approach is independent of\", \"the specific choice of the inversion technique, so that better inversion methods will certainly achieve better results. Additionally, encoder-based inversion techniques that we use for CLIP, are not so computationally expensive as the optimization-based techniques.\"], \"questions\": [\"We compared our reconstruction method to two different activation maximization reconstruction schemes. The results are in Appendix A.10 and are significantly inferior to our reconstruction scheme. We additionally compared six different metrics to evaluate the reconstruction results (Figure 5), and in Figure 6, we show a high correlation between an image being close to the margin and having a high reconstruction score, which is predicted by the theory. Given that this is the first work to reconstruct images with a resolution higher than CIFAR from trained classifiers, activation maximization remains the primary method available for comparison in this context.\", \"Regarding diffusion-based generative models, works that focus on data reconstruction from such modes (e.g., Carlini et al. 2023) have a very different approach. They are also highly dependent on knowing the training data for their evaluation (e.g., specifically targeting highly duplicated samples in the training set, that are searched for prior to the attack, and then using their text prompts from the training data in order to generate image samples). While such works are of prime importance, our submission focuses on classifier models and emphasizes privacy risks in transfer learning tasks, which are widely used in practice.\"]}", "{\"comment\": [\"We thank the reviewer for the feedback.\", \"Thank you for pointing this out. In this work, we focus on models trained on embeddings of common backbone models, where the backbone models\\u2019 parameters are fixed (the so-called \\u201cfeature-extractor\\u201d approach for transfer learning). We believe that this approach, while limited, is still the most common training approach used in practice. To support this, please see Kim et al. [1] who analyzed multiple papers on transfer learning for medical data and found that using backbones as feature-extractors is the most popular approach for transfer learning. We agree with the reviewer that exploring data reconstruction from models trained with other transfer learning approaches (e.g., fine-tuning backbone parameters) is of great importance and value to the community, however it is beyond the scope of this paper. Whether the current approach can be generalized to such settings is an interesting direction for future works.\", \"The main contribution is in demonstrating the vulnerability to training-data reconstruction of models that are trained in a transfer learning manner using embeddings of widely-used backbone models. This is a very popular approach for training classifiers on tasks with limited data, and exposing this vulnerability should be of major concern to practitioners. We provide an abundance of experimental evidence to support our claims. Another major contribution is the clustering-based reconstruction approach, which is very important in making such reconstruction attacks useful in practical settings. Note that many works in the field are using the training data in order to validate their attack. Our clustering-based approach alleviates this requirement.\", \"Right. We will fix it in the final version. Thanks.\", \"[1] \\u201cTransfer learning for medical image classification: a literature review\\u201d, Kim et al. 2022\"]}", "{\"comment\": \"Thank you for your response. The paper already shows results for ViT, Dino-ViT, Dino-ViT2 and CLIP which are extremely popular backbones for transfer learning, and with the respective standard resolution for these backbones. We would appreciate if the reviewer elaborate on the type of experiments that are required.\"}", "{\"comment\": \"We thank the reviewer for their openness to increasing the score and we highly appreciate their efforts in reviewing prior literature. Regarding reconstruction from segmentation models: [Buzaglo et al. (2023)](https://arxiv.org/abs/2307.01827) showed that data reconstruction may be possible even for models trained with other losses (not limited to classification) with weight decay term. However, the feasibility of such attacks in the context of segmentation models requires extensive investigation. We believe it would be beneficial to begin by testing these tasks on small-scale models before progressing to large-scale models like SAM, even in the context of transfer learning.\"}", "{\"summary\": \"This paper demonstrates the reconstruction of high-resolution training images from models trained using a transfer learning approach, as well as the reconstruction of non-visual data. Moreover, it introduces a novel clustering-based approach for effectively identifying training samples without prior knowledge of the training images.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. The writing is well done and clear.\\n\\n2. They present the weaknesses and limitations of their method in detail.\\n\\n3. The experiments consider various commonly used pre-trained feature extractors such as CLIP, demonstrating the effectiveness of their method.\", \"weaknesses\": \"1. The method is limited to specific cases where a fixed feature extractor and some MLP layers serve as the classifier. It cannot generalize to other more common transfer learning scenarios, such as fine-tuning an entire classifier or certain layers of a classifier.\\n\\n2. The introduced method lacks innovation. Specifically, the core contributions of reconstructing embedding vectors in Section 3.1 and mapping embedding vectors in Section 3.2 to the image domain either originate from other works or involve only simple modifications, such as changing the MSE loss to the cosine similarity loss. Please clearly explain the novel aspects of the method introduced.\\n\\n3. The format of Figure 9 is incorrect as it exceeds the page boundary.\", \"questions\": \"My confusion is the aforementioned weaknesses. If there are any misunderstandings, please point them out.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"metareview\": \"The paper conducts optimization-based data inversion from pre-trained models with transfer learning. The objective is to reconstruct the training data given the pre-trained encoder and classifiers using clustering techniques. The main contribution is in demonstrating the vulnerability to training-data reconstruction of models trained in a transfer learning manner using embeddings of widely-used backbone models.\", \"strengths\": [\"Operates in the embedding space, making it more scalable.\", \"The use of clustering is novel.\", \"Highlights privacy risks in models trained with sensitive data.\", \"Contains comprehensive experiments.\"], \"weaknesses\": [\"The data reconstructed is the expected data rather than the original data, thus the scope should be adjusted.\", \"Limited usability due to the required encoder/classifier setup and limited verification on CLIP and DINO.\", \"Limited technical contribution as it uses existing techniques.\", \"Given the limited technical contribution, the empirical contribution must be substantial and the discussions insightful to make them the main empirical contribution of the paper. However, the paper has a limited empirical contribution and requires major experiments to fully addressed the concerns raised by the reviewers. Thus, I recommend the rejection of the paper.\"], \"additional_comments_on_reviewer_discussion\": \"Reviewer qRiu recommended the acceptance of the paper based on the fact that the method works on the embedding space instead of the image space and deems this novel. However, this is a common approach in the field and is not a new contribution. Moreover, the reviewer refers to the privacy risks highlighted in the paper. The weaknesses the reviewer cites include toning down the scope and the limited applicability of the proposal, as dense models may not work with it. After the rebuttal, the reviewer mentions that their major concerns were addressed except for evaluations on fine-grained problems such as segmentation.\\n\\nReviewer hzzB mentioned that the contribution is limited since the key components used in the proposal were taken from previous work and noted that the resolution (224x224) is still low. The authors replied by emphasizing their main contributions and stated that data reconstruction models typically work with smaller resolutions. The reviewer did not respond to the authors.\\n\\nReviewer c8yx mentioned that the method is limited to specific cases and cannot generalize to common transfer learning scenarios. The reviewer also highlighted that the method is not innovative since it relies on existing works. The authors replied to the reviewer by stating that their setup is common and that the main contribution is the demonstration of the vulnerability to training-data reconstruction. The reviewer acknowledged the review but maintained a borderline reject score.\\n\\nReviewer VVGm noted that the experiments were only evaluated on CLIP and DINO and that other networks should have been evaluated as well. Similarly to reviewer hzzB, this reviewer mentioned that 224x224 shouldn\\u2019t be considered as high resolution. After the rebuttal from the authors, the reviewer still complained about the missing experiments on different setups and considered that the paper needs more exploration to be more solid, maintaining the borderline reject score.\\n\\nThe reviewers did not respond to my post-rebuttal discussion.\\n\\nOverall, the paper addresses a significant problem, but the results are not significant enough as mentioned by reviewers c8yx and VVGm. Moreover, given the limited technical contribution, the empirical contribution must be substantial and the discussions insightful to make them the main empirical contribution of the paper. This is not the case, as noted by the three negative reviews. The comments from the positive reviewer qRiu are not enough to outweigh these drawbacks. Thus, I recommend the rejection of the paper.\"}", "{\"summary\": \"This paper explores optimization-based **data inversion** techniques from pre-trained models with transfer learning, in which we could reconstruct the training data simply given the pre-trained encoder and classifier itself. It extends the model-inversion techniques with improved designs like loss, generative decoder. The clustering-based approach further demonstrates the possibility of high-quality reconstructions to the original data. Experiments on two common datasets reveals the potential privacy risks associated with models trained on sensitive data.\", \"soundness\": \"4\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": [\"[**Novelty**]\", \"Working in embedding space instead of directly reconstructing images is an innovative approach, and makes it more scalable to different models on both visual and non-visual data\", \"the adoption of clustering and averaging technique is novel, which identifies high-quality reconstructions when training data is not available\", \"[**Significance**]\", \"it highlights the significant privacy risks associated with models trained with sensitive data in a transfer learning setup, in which high-resolution data reconstruction in real-world conditions emphasizes the critical need for privacy-preserving mechanisms with today's pre-trained models.\", \"[**Completeness & Clarity**]\", \"High-quality visualizations, including comparisons of reconstructed images to original data and plots that illustrate reconstruction metrics, add depth to the evaluation, making it easier to interpret the results.\", \"The writing is clear, which effectively lays out the motivation and approach, and well explains the limitations\"], \"weaknesses\": [\"On significance, while I like the simple paradigm of solving x when f is known within f(x)=y, I kind feel that reconstructed data samples are not exactly matching with the actual training data, especially when the training data is not available. It is more close to averaged per-category data when training data is not available, the author might want to turn down their scope.\", \"Another thinking is that the current method may only work with classifiers-based model, for more more fine-grained training data reconstruction from segmentation model like SAM might be more desired. Also the reconstruction quality varies significantly with the choice of backbone model (DINO, CLIP, etc.), which affects the novelty of the clustering approach by making it model-dependent.\", \"As the author also mentioned, the inversion process used to map embeddings back to images is computationally expensive, which could hinder scalability\"], \"questions\": [\"I am also wondering whether we have any baselines in this line of data inversion techniques\", \"a quick question, for diffusion-based generative models, wondering whether it is also a more realistic concern to reveal training data directly.\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"details_of_ethics_concerns\": \"NA\", \"rating\": \"8\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thanks for the reply, I think my major concerns are well addressed, except for the part on **\\\"more fine-grained training data reconstruction from segmentation model like SAM\\\"**. I also checked all previous papers on training data reconstruction like [Deconstructing Data Reconstruction](https://arxiv.org/abs/2307.01827), and [Reconstructing Training Data from Trained\\nNeural Networks](https://arxiv.org/pdf/2206.07758), the improvements proposed in this paper largely boost the images quality, which demonstrates immediate potential of data reconstruction at the original quality. I would like to call out that, compared to all previous works, the improvement here is significant. Thus I'm leaning towards slightly increasing my score after the discussion.\"}" ] }

[ "Gang Li", "Wendi Yu", "Yao Yao", "Wei Tong", "Yingbin Liang", "Qihang Lin", "Tianbao Yang" ]

[ "Model Developmental Safety", "Continual Learning", "Vision-Language Models", "Constrained Optimization" ]

{ "note_id": [ "z7mqIAFmvZ", "y8Rt3CU3Cb", "wLr82wQ7vI", "tIo9bXw3Ug", "tFI9CaQzaO", "qu4pt8ss68", "oElRLNJmZ8", "j0P0teea75", "f9j2FNOXt9", "b6mbw2ufSl", "YesSUdhbnJ", "X0i6121OuJ", "TKvNz3kjbb", "I51CfmtGmp", "HmlNcUYzh4", "DnpycLnOaN", "DKYePg0pio", "CVQ12DOBrb", "CDETjk1gGY", "8x9rFsLgUd", "4oAlmPzsG3", "2ZFelMQGMm" ], "note_type": [ "official_comment", "official_review", "meta_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_review" ], "note_created": [ 1733072645434, 1730211167598, 1734418713314, 1732422097738, 1732576779962, 1732767135902, 1733201838917, 1732422053045, 1732422958667, 1730657344571, 1732422711786, 1737523891334, 1732422783002, 1733204250083, 1732716739018, 1732510354866, 1733073256556, 1732768816331, 1732423242193, 1730318810579, 1733205391407, 1730873394162 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission8156/Authors" ], [ "ICLR.cc/2025/Conference/Submission8156/Reviewer_1P9i" ], [ "ICLR.cc/2025/Conference/Submission8156/Area_Chair_cHwn" ], [ "ICLR.cc/2025/Conference/Submission8156/Authors" ], [ "ICLR.cc/2025/Conference/Submission8156/Reviewer_SvDs" ], [ "ICLR.cc/2025/Conference/Submission8156/Authors" ], [ "ICLR.cc/2025/Conference/Submission8156/Reviewer_1P9i" ], [ "ICLR.cc/2025/Conference/Submission8156/Authors" ], [ "ICLR.cc/2025/Conference/Submission8156/Authors" ], [ "ICLR.cc/2025/Conference/Submission8156/Reviewer_VmzU" ], [ "ICLR.cc/2025/Conference/Submission8156/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission8156/Authors" ], [ "ICLR.cc/2025/Conference/Submission8156/Authors" ], [ "ICLR.cc/2025/Conference/Submission8156/Reviewer_VmzU" ], [ "ICLR.cc/2025/Conference/Submission8156/Reviewer_1P9i" ], [ "ICLR.cc/2025/Conference/Submission8156/Authors" ], [ "ICLR.cc/2025/Conference/Submission8156/Authors" ], [ "ICLR.cc/2025/Conference/Submission8156/Authors" ], [ "ICLR.cc/2025/Conference/Submission8156/Reviewer_3RP8" ], [ "ICLR.cc/2025/Conference/Submission8156/Authors" ], [ "ICLR.cc/2025/Conference/Submission8156/Reviewer_SvDs" ] ], "structured_content_str": [ "{\"comment\": \"Dear Reviewer,\\n\\nThank you for your valuable feedback on our paper. As the author-reviewer discussion phase is nearing its end, please feel free to reach out if you have any remaining concerns or suggestions on our paper. We would greatly appreciate it.\"}", "{\"summary\": \"The paper formulates the safety multi-stage development problem using a comprehensive mathematical framework, offering a detailed analysis of its application on CLIP with a theoretically derived, task-dependent head. The authors propose an efficient constrained optimisation algorithm, which is empirically validated through extensive experiments.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": [\"Introducing the concept of model developmental safety is highly valuable, particularly in the context of large language models (LLMs), where the continual development often strains prior safety and alignment constraints. This concept is timely and impactful.\", \"The paper provides a robust guarantee for model developmental safety (MDS) of CLIP, underpinned by a detailed convergence analysis.\", \"Leveraging theoretical insights, the authors apply LoRA-based, task-dependent heads to effectively reduce the value of $\\\\delta$, with empirical validation provided in Appendix A.5.3.\", \"The proposed method demonstrates impressive performance improvements over baselines, notably in terms of the safety ratio, showcasing its effectiveness and robustness.\"], \"weaknesses\": \"Applying the model developmental safety (MDS) framework to vision-language models like CLIP for image classification is an interesting approach; however, it may not fully showcase the safety-critical nature of MDS. Since image classification in CLIP carries relatively low safety risk, especially compared with application in the safety of Large Language Models (LLM).\\n\\nDue to this, It\\u2019s challenging to distinguish this work from conventional Continual Learning (CL) approaches, despite the explanations in the related work section. To clarify the unique contribution, it could be beneficial to either emphasise scenarios where safety risks in CLIP are more evident or explore a more safety-critical application domain. For instance, focusing on multiple cycles of model development within LLMs\\u2014which frequently involved fine-tuning and are urgently required to ensuring safety and alignment\\u2014may better align with MDS objectives and make the safety focus more explicit and practical.\", \"questions\": [\"In Eq. (2), the concept of DevSafety seems to be defined as the worst-case performance drop of protected tasks. Could you please elaborate on how this definition differs from similar metrics, such as the forgetting measure commonly used in continual learning? Or is the primary aim of DevSafety indeed to achieve zero forgetting?\", \"he continual learning (CL) baselines included seem somewhat dated, with the most recent stemming from 2018 (Castro et al., 2018). It would strengthen the paper\\u2019s claims to compare the proposed method with more recent baselines mentioned in the related work.\", \"For clarity, it might be helpful to define $ s(\\\\mathrm{\\\\mathbf{x}}; \\\\mathrm{\\\\mathbf{w}})$ at its first mention (L158), rather than waiting until L179, as this could enhance readability and comprehension for readers.\", \"Providing a brief discussion of the limitations and potential directions for future work would be valuable, helping readers understand the broader impact and next steps for this research.\", \"The literature review on Safe Reinforcement Learning (SafeRL) at Line 134, while informative, may fall slightly outside the main scope of this article. You might consider clarifying its relevance to the paper\\u2019s focus, or potentially removing this section to maintain a more concise scope.\", \"> Francisco M Castro, Manuel J Mar\\u00edn-Jim\\u00e9nez, Nicol\\u00e1s Guil, Cordelia Schmid, and Karteek Alahari. End-to-end incremental learning. In Proceedings of the European conference on computer vision (ECCV), pp. 233\\u2013248, 2018.\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"metareview\": \"This paper introduces Model Developmental Safety (MDS), a safety-centric framework to ensure zero-forgetting of protected capabilities during iterative model development, particularly in safety-critical domains.\\n\\nAfter the rebuttal period, this paper receives mixed ratings. Reviewers 3RP8 and 1P9i still have remaining concerns, such as the lack of broader validation across diverse safety-critical domains and issues with the paper's presentation. All reviewers give borderline scores. Given the current weaknesses of the paper, the paper is rejected from the highly competitive ICLR conferences. The authors are encouraged to improve their manuscript according to the reviewers' suggestions, and submit it to the next venue.\", \"additional_comments_on_reviewer_discussion\": \"After the rebuttal period, this paper receives mixed ratings. Reviewers 3RP8 and 1P9i still have remaining concerns, such as the lack of broader validation across diverse safety-critical domains and issues with the paper's presentation. All reviewers give borderline scores. Given the current weaknesses of the paper, the paper is rejected from the highly competitive ICLR conferences. The authors are encouraged to improve their manuscript according to the reviewers' suggestions, and submit it to the next venue.\"}", "{\"title\": \"Part 2\", \"comment\": \"**RQ6:** Why is DevSafety measured by 'acc', while in Equation (2), it is defined by measuring the difference of empirical loss between the new and old models.\\n\\n**A:** The \\\"acc\\\" corresponds to the zero-one loss in Equation (2). We use difference of accuracy instead of difference of cross-entropy loss to measure DevSafety is because that accuracy is what people care in practice for classification tasks. \\n\\n**RQ7:** The author may want to elaborate on 'mild conditions' in lines 273-274.\\n\\n**A:** As shown in page 97 proposition 2.1 in [2], assume function $F$ and $h_k$ are continuous, the optimal solution of the penalty form will converge to the optimal solution of constrained form when penalty parameter goes infinity. Our analysis (Theorem 1) also stated the conditions such that our algorithm for solving the penalized form with a large enough $\\\\beta=O(1/\\\\epsilon)$ finds an $\\\\epsilon$-level approximated KKT-condition. \\n\\n[2] Bertsekas, Dimitri P. Constrained optimization and Lagrange multiplier methods. Academic press, 2014.\\n\\n\\n**RQ8:** What is the insight of leveraging the moving average estimators in lines 290-291?\\n\\n**A:** This is motivated by the work [3], whose insight is to utilize the historical data such that the contrastive learning does not require a very large batch size (e.g. 32,768 for OpenAI CLIP) to achieve a satisfactory result. \\n\\n[3] Yuan, Zhuoning, et al. \\\"Provable stochastic optimization for global contrastive learning: Small batch does not harm performance.\\\" International Conference on Machine Learning. PMLR, 2022.\\n\\n**RQ9:** In lines 461-464, how should we interpret Figure 2 that the development safety has been achieved?\\n\\n**A:** Since the x-axis represents DevSafety (acc), developmental safety is achieved when a point is located **on or to the right** of the vertical dotted line, i.e., DevSafety(acc)$\\\\geq 0$. Similarly, the target is improved if the point is also located above the horizontal dotted line.\"}", "{\"comment\": \"I want to thank the author for the detailed response. Most of my concerns have been addressed, and thus, I am willing to increase my score. Please let me know when the result of RQ5 is ready, and I will consider it to decide whether it is worth further increasing the final score.\\n\\nI would suggest the author include the reply of RQ4 as a discussion for future work in camera-ready since it is crucial for the community to appreciate the scope of the present paper and inspire future study.\"}", "{\"comment\": \"Thank you for your helpful suggestions. We incorporated the discussion of RQ4 in the revision. Moreover, we've finished the experiments with a recent replay-based baseline, namely Co$^2$L[1]. Following their paper, we tune their $\\\\tau$ in {0.05, 0.1}, $\\\\kappa$ in {0.1, 0.2}, $\\\\kappa^*$ in {0.01, 0.1}, $\\\\lambda$ in {0.1, 1, 10}. The results are presented below. We can see that even recent SOTA continual learning method still fails to ensure model developmental safety, with a zero retention ratio across all the tasks. This is anticipated as conventional continual learning focuses on trading off between protect tasks and target tasks, without ensuring zero-forgetting. We included the results in the revision.\\n\\n| Method | Measures | 100 | 1k | 2k | 4k |\\n| --------- | ---------------------- | ----------------------- | ----------------------- | ----------------------- | ----------------------- |\\n| Base | RetentionRatio//DevSafety | 100%//0.00(0.0000) | 100%//0.00(0.0000) | 100%//0.00(0.0000) | 100%//0.00(0.0000) |\\n| | Target Tunnel | 0.1064(0.0000) | 0.1064(0.0000) | 0.1064(0.0000) | 0.1064(0.0000) |\\n| **Co$^2$L** | RetentionRatio//DevSafety | 0.00%//-0.1407(0.0043) | 0.00%//-0.1252(0.0061) | 0.00%//-0.0821(0.0029) | 0.00%//-0.0479(0.0039) |\\n| | Target Tunnel | 0.6808(0.0460) | 0.8936(0.0626) | 0.8936(0.0301) | 0.8723(0.0000) |\\n| RM | RetentionRatio//DevSafety | 0.00%//-0.1021(0.0022) | 0.00%//-0.0969(0.0036) | 0.00%//-0.0955(0.0057) | 0.00%//-0.0897(0.0068) |\\n| | Target Tunnel | 0.9574(0.0233) | 0.8894(0.0340) | 0.8808(0.0170) | 0.8681(0.0085) |\\n| Ours | RetentionRatio//DevSafety | 40.00%//-0.0050(0.0076) | 60.00%//-0.0001(0.0043) | 100.00%//0.0105(0.0053) | 100.00%//0.0186(0.0058) |\\n| | Target Tunnel | 0.9362(0.0699) | 0.8723(0.0233) | 0.9106(0.0159) | 0.8723(0.0233) |\\n\\n\\n| Method | Measures | 100 | 1k | 2k | 4k |\\n| --------- | ---------------------- | ---------------------- | ----------------------- | ----------------------- | ----------------------- |\\n| Base| RetentionRatio//DevSafety | 100%//0.00(0.0000) | 100%//0.00(0.0000) | 100%//0.00(0.0000) | 100%//0.00(0.0000) |\\n| | Target Foggy | 0.3953(0.0000) | 0.3953(0.0000) | 0.3953(0.0000) | 0.3953(0.0000) |\\n| **Co$^2$L** | RetentionRatio//DevSafety | 0.00%//-0.0686(0.0064) | 0.00%//-0.1217(0.0383) | 0.00%//-0.1305(0.0183) | 0.00%//-0.0721(0.0154) |\\n| | Target Foggy | 0.7132(0.0109) | 0.6047(0.0380) | 0.6357(0.0110) | 0.6357(0.0290) |\\n| RM | RetentionRatio//DevSafety | 0.00%//-0.0418(0.0062) | 0.00%//-0.0173(0.0054) | 0.00%//-0.0159(0.0034) | 20.00%//-0.0124(0.0091) |\\n| | Target Foggy | 0.5674(0.0378) | 0.5023(0.0186) | 0.4419(0.0658) | 0.2279(0.0174) |\\n| Ours | RetentionRatio//DevSafety | 0.00%//-0.0241(0.0082) | 60.00%//-0.0009(0.0044) | 100.00%//0.0044(0.0033) | 100.00%//0.0061(0.0047) |\\n| | Target Foggy | 0.5721(0.0406) | 0.4930(0.0174) | 0.4326(0.0186) | 0.4279(0.0316) |\\n\\n\\n| Method | Measures | 100 | 1k | 2k | 4k |\\n| --------- | ---------------------- | ---------------------- | ----------------------- | ---------------------- | ----------------------- |\\n| Base | Retention Ratio//DevSafety | 100%//0.00(0.0000) | 100%//0.00(0.0000) | 100%//0.00(0.0000) | 100%//0.00(0.0000) |\\n| | Target Overcast | 0.7361(0.0000) | 0.7361(0.0000) | 0.7361(0.0000) | 0.7361(0.0000) |\\n| **Co$^2$L** | Retention Ratio//DevSafety | 0.00%//-0.0138(0.0099) | 0.00%//-0.0072(0.0032) | 0.00%//-0.0095(0.0043) | 0.00%//-0.0137(0.0052) |\\n| | Target Tunnel | 0.5916(0.0417) | 0.8369(0.0049) | 0.8396(0.0055) | 0.8507(0.0172) |\\n| RM | Retention Ratio//DevSafety | 0.00%//-0.2932(0.0365) | 0.00%//-0.3016(0.0228) | 0.00%//-0.2444(0.0120) | 0.00%//-0.2634(0.0105) |\\n| | Target Overcast | 0.9787(0.0050) | 0.9730(0.0028) | 0.9588(0.0041) | 0.9647(0.0023) |\\n| Ours | Retention Ratio//DevSafety | 0.00%//-0.0655(0.0249) | 20.00%//-0.0043(0.0037) | 60.00%//0.0012(0.0029) | 100.00%//0.0046(0.0016) |\\n| | Target Overcast | 0.8789(0.0464) | 0.7827(0.0225) | 0.7562(0.0167) | 0.7525(0.0366) |\\n\\n[1] Cha, Hyuntak, Jaeho Lee, and Jinwoo Shin. \\\"Co2l: Contrastive continual learning.\\\" Proceedings of the IEEE/CVF International conference on computer vision. 2021.\"}", "{\"title\": \"Thank You for Addressing My Concerns\", \"comment\": \"Apologies for the delayed response. I have carefully reviewed your further reply and the additional experiments addressing RQ4 and RQ1. Regarding RQ2, I acknowledge the potential of this work to extend to other domains, as elaborated in the theoretical analysis. However, I feel there remains a gap between a \\u201cconceptually extendable and inspiring\\u201d work and one that is fully \\u201cimplemented and validated\\u201d within the target application.\\n\\nRegrettably, I am unable to raise my score at this stage. That said, I firmly believe this work holds significant promise and would merit a much higher score once it has been thoroughly implemented and validated in more safety-critical applications.\"}", "{\"comment\": \"We thank the reviewer for dedicating the time to provide a comprehensive review. Below we would like to address raised concerns and questions.\\n\\n**RQ1:** Only strictly preserving the model's original performance is not enough. For instance, strictly maintaining the performance of tasks that are **not good enough may not** bring more benefits to improving the safety of the existing learning-enabled applications.\\n\\n**A:** We agree with the reviewer! However, we emphasize that the protected tasks are specified by the user and hence the user can choose which **existing essential abilities** to be protected. Thus, if one tasks' performance is very poor, one may not include such task as a protected one. In addition, our framework does not prevent the new model from becoming better than the old model on protected tasks. Indeed, from our results you can see that in many cases the performance on protected tasks are also improved while we improve the performance of a target task. For example, in Figure 1, the performance of a protected class \\\"partly cloudy\\\" is also significantly improved and that of other protected classes are slightly improved in Round 1. Similarly, in the right figure of Figure 4 we can also see that the new model also improves many protected classes. \\n\\n**RQ2:** Other than continual learning, other paradigms like data engines that consider the whole machine learning cycles [a, b, c] to achieve the safe development of the learning-based system, such as [c], the reviewer may suggest the author include some discussion of this paradigm.\\n\\n**A:** Thank you for pointing out these works! Automatic data engine paradigm is an important research direction for enhancing existing learning-enable systems by iteratively providing self-improved data. For example, [c] mines the vast amount of unlabeled data to increase the performance of detecting rare or unseen categories in object detection for autonomous driving systems. We have used a similar approach to mine a vast amount of unlabeled data on the internet to retrieve related data of the target task mentioned in line243 and detailed in Appendix A.2. However, the model updater module of [c] does not consider how to prevent zero-forgetting on protected tasks. We included the discussion with automatic data engine paradigm in Appendix A.2.\\n\\n\\n**RQ3:** With experiments conducted on CLIP models, can the proposed algorithm be extended to other models?\\n\\n**A:** The proposed constrained optimization framework is generic and it is not tied to any kinds of models. Although the algorithms are developed specifically for the contrastive loss as the objective, it can be easily extended to other losses. Indeed, the contrastive loss can be replaced by any loss function. For example, if we consider LLMs, the objective can be a supervised finetuning loss of a target task. The key of our framework is how to handle the constraints. It can be also a standard cross-entropy loss for learning a lightweight model. We hope our work can inspire researchers in safety-critical application domain for more exploration.\\n\\n\\n\\n**RQ4:** Only experiments with classification tasks in autonomous driving and scene recognition are included in the paper. How about other tasks, like 2D and 3D object detections for perception and tasks for motion prediction?\\n\\n**A:** Thanks for the reviewer\\u2019s constructive comments on improving our paper. Note that classficaiton task is the foundamental, ubiquitous and one of the most important tasks in learning-based system, so we take it as the demonstration of our proposed framework. Since the focus of this paper is to introduce and demonstrate the general constrained optimization framework for model development safety, we leave further exploration on 2D and 3D object detections and tasks for motion prediction for furthur exploration.\\n\\n**RQ5:** The author may want to compare with other replay-based methods.\\n\\n**A**: Thank you for the suggestion! We are working on adding a recent replay-based contrastive continual learning baseline [1]. However, given limited time we are still working on the experiments. We would like to emphasize that all the baselines we compared are continual learning baselines, with FLYP standing for direct-replay methods, GEM as a typical continual learning method, WCCL and RM as the regularized-based continual learning baselines tailored to our setting. We expect that continual learning can not ensure model developmental safety as they aim to trade off between protect tasks and target tasks.\\n\\n[1] Cha, Hyuntak, Jaeho Lee, and Jinwoo Shin. \\\"Co2l: Contrastive continual learning.\\\" Proceedings of the IEEE/CVF International conference on computer vision. 2021.\"}", "{\"comment\": \"Thank you for your constructive comments. We've revised our paper to address raised concerns.\\n\\n**RQ1:** The use of the terms \\\"safety\\\" / \\\"safety-centric\\\" in this paper is confusing because it doesn\\u2019t engage with the broader ethical and operational safety considerations commonly associated with the term.\\n\\n**A:** We thank the reviewer for the constructive comments. We agree that some usages of term \\\"safety\\\" in the paper might be confusing, such as \\\"safety-centric method\\\", \\\"safety of safety\\\". To address ambiguity, we replaced \\\"safety-centric method\\\" with \\\"retention-centric method\\\", \\\"safety of safety\\\" with \\\"retention of safety\\\", \\\"safety ratio\\\" with \\\"retention ratio\\\", since our work, as summarized by the reviewer, focuses on designing constraints to retain protected task performance. But we prefer to keep the term \\\"model developmental safety\\\" as our work underscores the importance of strictly preserving existing protected ability in favor of potential safe applications and development efficiency in the model development process. Other words like \\\"stability\\\", \\\"preservation\\\" are wildly adopted in the existing continual learning literature but these works just mitigate the forgetting but not achieve strictly perservation or zero-forgetting. I believe the term \\\"model developmental safety (MDS)\\\" will be helpful for readers to identify the difference between our work and exisiting literature for iterative model development process. Furthermore, to prevent any ambiguity around the term \\\"model developmental safety\\\", we define it at the beginning of the paper, and revise the paper to ensure that every mention of \\\"safety\\\" is preceded by \\\"developmental\\\", when referring to MDS, to clarify its meaning for readers.\\n\\n\\n**RQ2:** CLIP is only one kind of Vision Language Model (VLM), 'aka' is inappropriately used in the paper and other representative variants such as LLaVA and BLIP that can generate languages are not evaluated in the paper.\\n\\n**A:** Thanks for pointing out the inappropriate use of \\\"aka\\\". To avoid confusion, we have revised the paper with \\\"\\u2026 we study how to develop a pretrained vision-language model, specifically the CLIP model, \\u2026\\\" . Note that the focus of the paper is to propose a constrained optimization framework to strictly preserve the existing protected capabilities while improving target task performance in iterative model development process. To demonstrate the proposed framework, we apply the framework to develop a CLIP model for acquiring new capabilities or improving existing capabilities of image classification. Our framework owns the potential to be applied to other variants of VLMs such LLaVA and BLIP with corresponding adaption with objective and constraint design, but as they are not the focus of this paper, we leave them for further exploration.\\n\\n\\n\\n**RQ3:** The relationship between this paper and knowledge/representation editing on VLM/LLM.\\n\\n**A:** We thank the reviewer for the constructive suggestion. Knowledge/representation editing on VLM/LLM is related to our work but has a different focus. Knowledge/representation editing, emerging in the era of large foundation models, aims to efficiently modify the behavior of LLMs with minimal impact on unrelated inputs[2], such as to update stale facts, eliminate unintended biases, reduce undesired hallucinations, etc. With knowledge editing **minimizing** the impact on unrelated contents, our proposed framework is a general framework and aims to **strictly preserve** protected abilities in favor of potential safe applications and development efficiency in iterative model development process. On the other hand, our proposed framework may be applied to addressing the challenge faced by knowledge/representation editing, by formulating the modified part as the objective (target task) and regrading the unrelated parts as constraints (protect task) to ensure zero-forgetting on unrelated parts (i.e., complete locality in knowledge/representation editing). We have incoporated the discussion about knowledge/representation editing in the revision.\\n\\n\\n\\n[2] EasyEdit: An Easy-to-use Knowledge Editing Framework for Large Language Models. arXiv:2308.07269\"}", "{\"summary\": \"This paper aims to study the problem of improving model accuracy on new categories while ensuring that accuracy on fixed existing categories does not degrade. It formulates the problem as an inequality-constrained optimization problem and proposes an algorithm to solve it. Overall, I believe the paper's core innovation lies in introducing a new optimization algorithm for solving non-convex constraint problems.\", \"soundness\": \"4\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1.The paper provides extensive theoretical analysis of the optimization algorithm to ensure its correctness.\\n\\n2.Experiments are conducted on large-scale datasets to verify the general performance of the method.\\n\\n3.The appendix effectively supplements details of the methodology and experiments.\", \"weaknesses\": \"1.The article is based on the premise of safety, proposing the assumption of strictly maintaining the original model performance unchanged. However, in real-world applications, classification tasks do not always require extremely high accuracy, and some fluctuation in accuracy is acceptable in certain scenarios. Given that the classification task discussed in the article is not an extreme case, I believe that the strict maintenance assumption proposed may be overly rigid for the actual tasks accomplished by the CLIP model.\\n\\n2.The abstract and introduction are somewhat misleading, as catastrophic forgetting encompasses a broad range of phenomena beyond the classification issues discussed in the paper, including the ability to recognize image content. The \\u201cprotected capabilities of the old model\\u201d described in the introduction may cause ambiguity.\\n\\n3.Evaluating model performance solely using the safety ratio metric is insufficient. The issue with the safety ratio metric is that, if a model update method results in an imperceptible decrease in accuracy on existing categories while significantly increasing accuracy on new categories, such a scenario might be acceptable to a certain extent. However, this situation would be rated poorly with this metric. To differentiate these cases from methods that cause significant performance declines on existing categories, it is necessary to include data on the change in recognition accuracy for existing categories after training.\\n\\n4.When evaluating the ability to protect classification accuracy across multiple categories, the safety ratio is not provided, and I cannot find any points in Figure 2 that obviously exceed the DevSafety (acc) boundary of 0. This raises doubts as to whether the improvement in dressing room classification accuracy was accompanied by declines in certain other categories, making me skeptical of the authors\\u2019 conclusion that old performance remains consistent in multi-task scenarios.\", \"questions\": \"Please refer to Weaknesses.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"We thank the reviewer for acknowledging the contribution of our work. Below, we would like to answer the questions raised.\\n\\n\\n**RQ1:** Classification tasks do not always require extremely high accuracy, the strict maintenance assumption proposed may be overly rigid.\\n\\n**A:** We politely disagree with the reviewer that \\\"classification tasks do not always require extremely high accuracy\\\". In autonomous driving, if the weather condition is foggy, but the system identifies it as sunny, it could cause a wrong decision and may result in accidents. Similarly in medical diagnosis, if a patient is misclassified he/she may suffer from life risk. Please also note that our constrained optimization framework is generic, it can be potentially extended to other scenarios with different loss functions or models.\\n\\n\\n**RQ2:** The \\u201cprotected capabilities of the old model\\u201d described in the introduction may cause ambiguity, as catastrophic forgetting encompasses a broad range of phenomena beyond the classification issues discussed in the paper.\\n\\n**A:** The introduction intends to be general. Indeed, the \\\"protected capabilities of the old model\\\" means the general capabilities of models not just the classification issues. As discussed in section 3, it may also be coding ability of LLMs or detection ability of objective detection models. The classification ability of models is just the one we measured in our experiments.\\n\\n\\n\\n**RQ3:** A method may result in an imperceptible decrease in accuracy on existing categories while significantly increasing accuracy on new categories, such a scenario might be acceptable to a certain extent. It is necessary to include performance changes for protected tasks after training.\\n\\n**A:** Our evaluation is based on the motivation of this paper, i.e., preserving the performance of protected tasks while improving that of a target task. One might argue that this might not be needed in some scenarios that can tolerate some performance drop of protected tasks. However, this is not the problem we addressed in the paper. As we have argued in the paper, preserving the performance of protected asks are very important in some safety-critical applcations (e.g., autonomous driving, medical diagnosis). Nevertheless, we also include the DevSafety(acc) numbers for each method in Appendix A.5 in the revision and presented below for your reference, which directly show the largest decrease over all the protected tasks. We can see that baselines usually lead to 3-10 percent decrease when targeting Tunnel and 1.5-7 percent decrease when targeting Foggy. \\n\\n\\n**RQ4:** I cannot find any points in Figure 2 that obviously exceed the DevSafety (acc) boundary of 0.\\n\\n\\n**A:** As long as Devsafety is larger than or **equal** to zero, it achieves the model developmental safety. In Figure 2, as long as the points are on the vertical line or at the right of the vertical line it means that model developmental safety is achieved. In Figure 4, as long as the points are above the red horizontal time, it means that model developmental safety is achieved. This indeed happens for our method, which means we preserve the performance of other classes while improving the dressing room classification accuracy.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"title\": \"Results for DevSafety(acc) numbers\", \"comment\": \"| | | 100 | 1k | 2k | 4k |\\n| --------- | ---------------------- | ----------------------- | ----------------------- | ----------------------- | ----------------------- |\\n| Base(Ref) | SafetyRatio//DevSafety | 100%//0.00(0.0000) | 100%//0.00(0.0000) | 100%//0.00(0.0000) | 100%//0.00(0.0000) |\\n| | Target Tunnel | 0.1064(0.0000) | 0.1064(0.0000) | 0.1064(0.0000) | 0.1064(0.0000) |\\n| FLYP | SafetyRatio//DevSafety | 0.00%//-0.0398(0.0067) | 0.00%//-0.0660(0.0126) | 0.00%//-0.0647(0.0123) | 0.00%//-0.0774(0.0069) |\\n| | Target Tunnel | 0.9361(0.0330) | 0.9702(0.0318) | 0.9915(0.0170) | 0.9659(0.0170) |\\n| WCCL | SafetyRatio//DevSafety | 0.00%//-0.0836(0.0164) | 0.00%//-0.0756(0.0090) | 0.00%//-0.0673(0.0103) | 0.00%//-0.0893(0.0089) |\\n| | Target Tunnel | 0.9957(0.0085) | 0.6000(0.1002) | 0.6553(0.0282) | 0.6383(0.0485) |\\n| GEM | SafetyRatio//DevSafety | 0.00%//-0.1019(0.0267) | 0.00%//-0.1034(0.0153) | 0.00%//-0.1301(0.0169) | 0.00%//-0.0873(0.0231) |\\n| | Target Tunnel | 0.8255(0.1214) | 0.5915(0.2020) | 0.6085(0.0768) | 0.3915(0.1819) |\\n| RM | SafetyRatio//DevSafety | 0.00%//-0.1021(0.0022) | 0.00%//-0.0969(0.0036) | 0.00%//-0.0955(0.0057) | 0.00%//-0.0897(0.0068) |\\n| | Target Tunnel | 0.9574(0.0233) | 0.8894(0.0340) | 0.8808(0.0170) | 0.8681(0.0085) |\\n| Ours | SafetyRatio//DevSafety | 40.00%//-0.0050(0.0076) | 60.00%//-0.0001(0.0043) | 100.00%//0.0105(0.0053) | 100.00%//0.0186(0.0058) |\\n| | Target Tunnel | 0.9362(0.0699) | 0.8723(0.0233) | 0.9106(0.0159) | 0.8723(0.0233) |\\n\\n| | | 100 | 1k | 2k | 4k |\\n| --------- | ---------------------- | ---------------------- | ----------------------- | ----------------------- | ----------------------- |\\n| Base(Ref) | SafetyRatio//DevSafety | 100%//0.00(0.0000) | 100%//0.00(0.0000) | 100%//0.00(0.0000) | 100%//0.00(0.0000) |\\n| | Target Foggy | 0.3953(0.0000) | 0.3953(0.0000) | 0.3953(0.0000) | 0.3953(0.0000) |\\n| FLYP | SafetyRatio//DevSafety | 0.00%//-0.0590(0.0140) | 20.00%//-0.0281(0.0167) | 0.00%//-0.0254(0.0101) | 0.00%//-0.0201(0.0105) |\\n| | Target Foggy | 0.5721(0.0315) | 0.5209(0.0581) | 0.5302(0.0228) | 0.4977(0.0186) |\\n| WCCL | SafetyRatio//DevSafety | 0.00%//-0.0504(0.0123) | 0.00%//-0.0259(0.0080) | 20.00%//-0.0141(0.0111) | 0.00%//-0.0132(0.0076) |\\n| | Target Foggy | 0.3395(0.0865) | 0.2186(0.0186) | 0.2093(0.0208) | 0.2000(0.0114) |\\n| GEM | SafetyRatio//DevSafety | 0.00%//-0.0695(0.0099) | 0.00%//-0.0339(0.0053) | 0.00%//-0.0424(0.0060) | 0.00%//-0.0424(0.0060) |\\n| | Target Foggy | 0.3349(0.0865) | 0.2837(0.0271) | 0.2558(0.0000) | 0.2558(0.0000) |\\n| RM | SafetyRatio//DevSafety | 0.00%//-0.0418(0.0062) | 0.00%//-0.0173(0.0054) | 0.00%//-0.0159(0.0034) | 20.00%//-0.0124(0.0091) |\\n| | Target Foggy | 0.5674(0.0378) | 0.5023(0.0186) | 0.4419(0.0658) | 0.2279(0.0174) |\\n| Ours | SafetyRatio//DevSafety | 0.00%//-0.0241(0.0082) | 60.00%//-0.0009(0.0044) | 100.00%//0.0044(0.0033) | 100.00%//0.0061(0.0047) |\\n| | Target Foggy | 0.5721(0.0406) | 0.4930(0.0174) | 0.4326(0.0186) | 0.4279(0.0316) |\"}", "{\"title\": \"Thank you for your response!\", \"comment\": \"We are glad to hear that you agree our work holds significant promise.\\n\\nWhile we believe more experiments can strengthen the paper, we would like to point out that (i) we have already conducted extensive experiments including 5 baselines, 4 target tasks and 2 datasets, but also ablation studies on the proposed algorithm; (ii) our experiments do include the tasks in safety-critical application in autonomous driving.\"}", "{\"comment\": \"Thanks for the author's response. I have also read the comments from the other reviewers, and I tend to maintain my rating.\"}", "{\"title\": \"Thank you for your prompt reply\", \"comment\": \"Thank you to the authors for their efforts in improving the paper. However, some of my concerns, particularly regarding RQ1, RQ2, and RQ4, remain insufficiently addressed. As such, I regret that I am unable to raise my score at this time.\"}", "{\"comment\": \"Dear Reviewer,\\n\\nThank you for the valuable time you have dedicated to reviewing our paper. We have included additional experimental results with a recent replay-based baseline to further validate the effectiveness of our method. As the author-reviewer discussion phase is coming to a close, we would greatly appreciate it if you could let us know whether they address your concerns or if further clarification is needed.\"}", "{\"comment\": \"**Q:** Concerns regarding RQ1 and RQ4 remain insufficiently addressed.\\n\\n**A**: We thank the reviewer for the prompt response. Regarding R4, we've finished the experiments with a recent replay-based baseline, namely Co$^2$L[1]. Following their paper, we tune their $\\\\tau$ in {0.05, 0.1}, $\\\\kappa$ in {0.1, 0.2}, $\\\\kappa^*$ in {0.01, 0.1}, $\\\\lambda$ in {0.1, 1, 10}. The results are presented below. We can see that even recent SOTA continual learning method still fails to ensure model developmental safety indicated by that Retention Ratio is zero and DevSafty measure is less than zero. This is anticipated as conventional continual learning focuses on trading off between protect tasks and target tasks, without ensuring zero-forgetting. We included the results in the revision. The experiments further highlight the distinction between conventional continual learning and our approach, demonstrating that existing continual learning methods cannot achieve the model developmental safety explored in this paper, as related to RQ1.\\n\\n| Method | Measures | 100 | 1k | 2k | 4k |\\n| --------- | ---------------------- | ----------------------- | ----------------------- | ----------------------- | ----------------------- |\\n| Base | RetentionRatio//DevSafety | 100%//0.00(0.0000) | 100%//0.00(0.0000) | 100%//0.00(0.0000) | 100%//0.00(0.0000) |\\n| | Target Tunnel | 0.1064(0.0000) | 0.1064(0.0000) | 0.1064(0.0000) | 0.1064(0.0000) |\\n| **Co$^2$L** | RetentionRatio//DevSafety | 0.00%//-0.1407(0.0043) | 0.00%//-0.1252(0.0061) | 0.00%//-0.0821(0.0029) | 0.00%//-0.0479(0.0039) |\\n| | Target Tunnel | 0.6808(0.0460) | 0.8936(0.0626) | 0.8936(0.0301) | 0.8723(0.0000) |\\n| RM | RetentionRatio//DevSafety | 0.00%//-0.1021(0.0022) | 0.00%//-0.0969(0.0036) | 0.00%//-0.0955(0.0057) | 0.00%//-0.0897(0.0068) |\\n| | Target Tunnel | 0.9574(0.0233) | 0.8894(0.0340) | 0.8808(0.0170) | 0.8681(0.0085) |\\n| Ours | RetentionRatio//DevSafety | 40.00%//-0.0050(0.0076) | 60.00%//-0.0001(0.0043) | 100.00%//0.0105(0.0053) | 100.00%//0.0186(0.0058) |\\n| | Target Tunnel | 0.9362(0.0699) | 0.8723(0.0233) | 0.9106(0.0159) | 0.8723(0.0233) |\\n\\n\\n| Method | Measures | 100 | 1k | 2k | 4k |\\n| --------- | ---------------------- | ---------------------- | ----------------------- | ----------------------- | ----------------------- |\\n| Base| RetentionRatio//DevSafety | 100%//0.00(0.0000) | 100%//0.00(0.0000) | 100%//0.00(0.0000) | 100%//0.00(0.0000) |\\n| | Target Foggy | 0.3953(0.0000) | 0.3953(0.0000) | 0.3953(0.0000) | 0.3953(0.0000) |\\n| **Co$^2$L** | RetentionRatio//DevSafety | 0.00%//-0.0686(0.0064) | 0.00%//-0.1217(0.0383) | 0.00%//-0.1305(0.0183) | 0.00%//-0.0721(0.0154) |\\n| | Target Foggy | 0.7132(0.0109) | 0.6047(0.0380)| 0.6357(0.0110) | 0.6357(0.0290) |\\n| RM | RetentionRatio//DevSafety | 0.00%//-0.0418(0.0062) | 0.00%//-0.0173(0.0054) | 0.00%//-0.0159(0.0034) | 20.00%//-0.0124(0.0091) |\\n| | Target Foggy | 0.5674(0.0378) | 0.5023(0.0186) | 0.4419(0.0658) | 0.2279(0.0174) |\\n| Ours | RetentionRatio//DevSafety | 0.00%//-0.0241(0.0082) | 60.00%//-0.0009(0.0044) | 100.00%//0.0044(0.0033) | 100.00%//0.0061(0.0047) |\\n| | Target Foggy | 0.5721(0.0406) | 0.4930(0.0174) | 0.4326(0.0186) | 0.4279(0.0316) |\\n\\n\\n| Method | Measures | 100 | 1k | 2k | 4k |\\n| --------- | ---------------------- | ---------------------- | ----------------------- | ---------------------- | ----------------------- |\\n| Base | Retention Ratio//DevSafety | 100%//0.00(0.0000) | 100%//0.00(0.0000) | 100%//0.00(0.0000) | 100%//0.00(0.0000) |\\n| | Target Overcast | 0.7361(0.0000) | 0.7361(0.0000) | 0.7361(0.0000) | 0.7361(0.0000) |\\n| **Co$^2$L** | Retention Ratio//DevSafety | 0.00%//-0.0138(0.0099) | 0.00%//-0.0072(0.0032) | 0.00%//-0.0095(0.0043) | 0.00%//-0.0137(0.0052) |\\n| | Target Tunnel | 0.5916(0.0417) | 0.8369(0.0049) | 0.8396(0.0055) | 0.8507(0.0172) |\\n| RM | Retention Ratio//DevSafety | 0.00%//-0.2932(0.0365) | 0.00%//-0.3016(0.0228) | 0.00%//-0.2444(0.0120) | 0.00%//-0.2634(0.0105) |\\n| | Target Overcast | 0.9787(0.0050) | 0.9730(0.0028) | 0.9588(0.0041) | 0.9647(0.0023) |\\n| Ours | Retention Ratio//DevSafety | 0.00%//-0.0655(0.0249) | 20.00%//-0.0043(0.0037) | 60.00%//0.0012(0.0029) | 100.00%//0.0046(0.0016) |\\n| | Target Overcast | 0.8789(0.0464) | 0.7827(0.0225) | 0.7562(0.0167) | 0.7525(0.0366) |\\n\\n[1] Cha, Hyuntak, Jaeho Lee, and Jinwoo Shin. \\\"Co2l: Contrastive continual learning.\\\" Proceedings of the IEEE/CVF International conference on computer vision. 2021.\"}", "{\"comment\": \"We thank the reviewer for acknowledging the value of our work and providing helpful comments. Below we would like to answer the remaining questions.\\n\\n**RQ1:** It\\u2019s challenging to distinguish this work from conventional Continual Learning (CL) approaches.\\n\\n**A:** Our proposed framework differs from conventional continual learning in two folds. (a) **Learning setting**: In typical settings for continual learning, models are always trained to learn new tasks, with limited access to previously trained data. In contrast, our work may be utilized to either learn new tasks or improve existing tasks, with sufficient number of old data for ensuring zero-forgetting on protected tasks. (b) **Goal**: the goal of continual learning is to have a good average performance for all the learned tasks, our work prioritze preserving some protected tasks while improving the target tasks. Therefore, our work exhibits substantial distinction with continual learning.\\n\\n**RQ2:** Image classification in CLIP carries relatively low safety risk. It could be beneficial to explore a more safety-critical application domain.\\n\\n\\n\\n**A:** Note that experiments in our paper like weather detection and scene recognition are directly related to safety-critical autonomous driving systems. In these scenarios, enhancing the detection of one type of weather at the expense of reduced performance for other types could pose significant safety risks. Moreover, our proposed retention-centric optimization framework is generic and CLIP model with classification task is a demonstration, it can be easily extended to other losses or other models, such as a supervised finetuning loss for LLMs or a standard cross-entropy loss for learning a lightweight model. We hope our work can inspire researchers in safety-critical application domain for further exploration.\\n\\n\\n**RQ3:** How does the definition of DevSafety differ from similar metrics, such as the forgetting measure commonly used in continual learning? \\n\\n**A:** As pointed out by reviewer, DevSafety is defined to take worst-case of all the protected tasks to measure if all the protected tasks are strictly preserved. In contrast, forgetting measure commonly used in continual learning is defined as average performance drop of protected tasks. Note that average performance doesn't drop doesn't mean each individual protected task performance doesn't drop, e.g., some protected tasks get better and some protected tasks get worse. Due to the intrinsic nature of each protected task, such as each task is associated with detecting one kind of disease for medical diagnosis, it may lead to potential unsafe deployment even when the average performance doesn't drop. So, the primary aim of DevSafety is indeed to achieve zero forgetting for safety-critical applications. \\n\\n\\n\\n**RQ4:** It would strengthen the paper\\u2019s claims to compare the proposed method with more recent baselines mentioned in the related work.\\n\\n**A:** Thank you for your suggestion. We are working on adding a recent replay-based contrastive continual learning baseline [1]. However, given limited time we are still working on the experiments. We would like to emphasize that all the baselines we compared are continual learning baselines, with FLYP standing for direct-replay methods, GEM as a typical continual learning method, WCCL and RM as the regularized-based continual learning baselines tailored to our setting. We expect that continual learning can not ensure model developmental safety as they focus on trading off between protect tasks and target tasks.\\n\\n[1] Cha, Hyuntak, Jaeho Lee, and Jinwoo Shin. \\\"Co2l: Contrastive continual learning.\\\" Proceedings of the IEEE/CVF International conference on computer vision. 2021.\\n\\n**RQ5:** Suggestions about the definition of s(x;w), discussion of future directions, literature review on Safe Reinforcement Learning (SafeRL) \\n\\n**A:** We thank the reviewer for helpful suggestions for improving the paper. We have revised the writing accordingly. We'd like to mention that the reason for deferring the definition of s(x;w) to line 179 is that the definition in line 179 is the special form for CLIP models, so we present the special explicit form of s(x;w) after introducing CLIP models to avoid confusion. As suggested, we have included more discussion of further directions of this work. Please refer to our new version of the paper.\"}", "{\"summary\": \"The paper proposes **model developmental safety** to argue the importance of handling catastrophic\\nforgetting with a constrained optimization framework. The proposed method is evaluated to ensure the development of CLIP models. The experiments cover datasets from self-driving to scene classification.\", \"soundness\": \"3\", \"presentation\": \"1\", \"contribution\": \"2\", \"strengths\": \"1. The theoretical analysis of the framework in Section 5 is sound and comprehensive.\\n2. The evaluation of ensuring CLIP models' continual development is fair.\\n3. The visualization of the learning trajectories is well-presented.\", \"weaknesses\": \"1. Although the authors tried to address the term ambiguity in Section 2 (with AI Safety), the use of the terms \\\"safety\\\" / \\\"safety-centric\\\" in this paper is often overstated because it doesn\\u2019t engage with the broader ethical and operational safety considerations commonly associated with the term. Even further, in Line 132, the paper writes \\\"safety of safety\\\", which is not rigorously explained. In fact, the paper focuses on designing constraints to preserve task performance, which, while essential, diverges from widely understood safety principles in deep learning models. An alternative term such as **\\\"developmental stability\\\"**, **\\\"continual stability\\\"**, or **\\\"capability preservation\\\"** can more clearly represent the framework's intentions without abusing the term of safety.\\n\\n2. The empirical evaluation is lacking. Note that the Vision Language Model (VLM) is a general category of foundation models, and CLIP is only one example of this category. Other representative variants such as LLaVA and BLIP that can generate languages are not evaluated in the paper. In the abstract, the paper writes that *\\\"...we study how to develop a pretrained vision-language model (aka the CLIP model)...\\\"*, which may mislead future readers since \\\"aka\\\" is wrongly used here.\\n\\n3. Considering the paper's motivation to ensure stable continual development **without harming protected capabilities**, it largely overlaps with the task of **knowledge/representation editing** [1,2,3,4,5,6] on VLM/LLM. However, few pieces of related literature are discussed in the paper. Authors may consider discussing the main advantages of their proposed framework regarding this existing line of research.\\n\\n[1] Mass-Editing Memory in a Transformer. arXiv:2210.07229.\\n\\n[2] EasyEdit: An Easy-to-use Knowledge Editing Framework for Large Language Models. arXiv:2308.07269\\n\\n[3] KEBench: A Benchmark on Knowledge Editing for Large Vision-Language Models. arXiv:2403.07350.\\n\\n[4] Representation Engineering: A Top-Down Approach to AI Transparency. arXiv:2310.01405.\\n\\n[5] PaCE: Parsimonious Concept Engineering for Large Language Models. arXiv:2406.04331.\\n\\n[6] Reducing Hallucinations in Vision-Language Models via Latent Space Steering. arXiv.2410.15778.\", \"questions\": \"Please address my concerns stated in the weakness section. Also, please revise or re-consider all uses of \\\"aka\\\" in the paper (e.g., Line 29, Line 124) as they may lead to unnecessary confusion.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Thank you!\", \"comment\": \"Dear Reviewer SvDs:\\n\\nPlease check below our new experimental results for addressing RQ5 on comparing with one recent continual contrastive learning method. \\n\\nThank you for your time! \\n\\nRegards\\nAuthors\"}", "{\"summary\": \"This paper focuses on the model deployment cycle for a learning-enabled system. The author proposes a concept called \\\"model developmental safety\\\" (MDS) to measure whether the learning-enabled system can strictly maintain the performance, i.e., zero forgetting, of the old tasks for safety-critical domains. The author proposes an efficient constrained optimization algorithm tailored to finetune the pretrained CLIP model that takes the MDS as the data-dependent constraint, providing a statistical guarantee for achieving MDS. Experiments have been conducted on BDD100k from autonomous driving scenarios and Places365 for scene recognition.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"(1) The proposed \\\"model developmental safety\\\" (MDS) concept seems interesting and relevant to safety-critical applications, though many concerns remain, which will be elaborated on in the Weakness section.\\n\\n(2) The proposed constrained optimization algorithm is sound for fine-tuning CLIP by retaining old data to achieve MDS; its effectiveness has also been validated by comparison with other methods.\", \"weaknesses\": \"(1) The motivation and necessity of the MDS is not sound enough. First, the MDS can be viewed as a more strict version of preventing catastrophic forgetting, i.e., maintaining \\\"zero forgetting\\\" during continual learning. The author claimed in lines 065-069 that zero forgetting is crucial for many safety-critical applications when considering the whole deployment cycle of the learning-enabled cycle, which is reasonable. However, only strictly preserving the model's original performance is not enough. For instance, strictly maintaining the performance of tasks that are not good enough may not bring more benefits to improving the safety of the existing learning-enabled applications. The review may suggest that the author calibrate their statement.\\n\\nMoreover, other than the traditional paradigm of continual learning, there also exist other paradigms like data engines that consider the whole machine learning cycles [a, b, c] to achieve the safe development of the learning-based system, where [c] provides an automatic self-improved data engine for safety-critical application, i.e., autonomous driving. Different from the present work, [c] does not need to retain old data to maintain the performance; instead, it mines the vast amount of unlabeled data to increase the performance of long-tailed or new tasks while maintaining the performance of the old tasks. Moreover, [c] validates the self-improved data engine on object detection tasks, which is more challenging and safety-critical in autonomous driving and classification. The reviewer may suggest the author include some discussion of other learning paradigms, like automatic data engines, given that they have similar motivations and targeted applications.\\n\\n(2) The proposed algorithm seems too restricted to the pretrained CLIP model, making it hard to evaluate the applicability of the proposed method for safety-critical real-world applications. Although the proposed method is sound for fine-tuning the pretrained CLIP to achieve MDP, the proposed constrained optimization seems too restricted, making the reviewer wonder whether it has sufficient applicability for other foundation models. The development of the algorithm mainly depends on the CLIP model and the contrastive loss, while the contrastive loss is not the only choice for training the foundation model. The author may want to elaborate on how the proposed algorithm can be extended to different kinds of foundation models.\\n\\nMoreover, there is a practical concern that the foundation model like CLIP may not satisfy the requirement for the real-time latency of safety-critical applications like autonomous driving, as the real-world intelligent system is also integrated with many different components. The author may want to show that the proposed algorithm can also apply to lightweight models other than foundation models to validate the applicability of the proposed method.\\n\\n(3) In the experiment, the author only considered the classification task in autonomous driving and scene recognition. However, many safety-critical applications that are more challenging and underperformed [d] will benefit more from MDS, e.g., 2D and 3D object detections for perception and tasks for motion prediction. The author may want to have more case studies other than classification to show the generality of the proposed algorithm.\\n\\n(4) For the comparison methods, the author only compared with the GEM proposed in 2017, while many other replay-based methods [e, f] have been proposed in recent years that can achieve state-of-the-art performance in the continual learning literature. The author may want to compare with those methods.\", \"minor\": \"(1) The reviewer wonders why DevSafety is measured by 'acc', while in Equation (2), it is defined by measuring the difference of empirical loss between the new and old models.\\n\\n(2) The author may want to elaborate on 'mild conditions' in lines 273-274. \\n\\n(3) What is the insight of leveraging the moving average estimators in lines 290-291? \\n\\n(4) Typo in line 312: 'proected' -> 'protected'\\n\\n(5) In lines 461-464, how should we interpret Figure 2 that the development safety has been achieved?\", \"reference\": \"[a] NEIL: Extracting Visual Knowledge from Web Data. ICCV 2013\\n\\n[b] Never-ending learning. Communications of the ACM 2018\\n\\n[c] AIDE: An Automatic Data Engine for Object Detection in Autonomous Driving. CVPR 2024\\n\\n[d] End-to-end Autonomous Driving: Challenges and Frontiers. TPAMI 2024\\n\\n[e] A Comprehensive Survey of Continual Learning: Theory, Method and Application. TPAMI 2024\\n\\n[f] Class-Incremental Learning: A Survey. TPAMI 2024\", \"questions\": \"Please refer to the Weaknesses section.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"details_of_ethics_concerns\": \"N/A.\", \"rating\": \"6\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}" ] }

[ "Bokai Lin", "Zihao Zeng", "Zipeng Xiao", "Siqi Kou", "TianQi Hou", "Xiaofeng Gao", "Hao Zhang", "Zhijie Deng" ]

[ "Inference Optimization", "KV Cache Compression", "Low-rank Projection" ]

{ "note_id": [ "ucGZ5DXXs4", "u6v8Kmf2KU", "qTiymQKo70", "mRDwIdQdY9", "hb96iqGShB", "hLK2qFrCBO", "hI5fvHNy6A", "cgBglCeU3n", "a673Dw9ckJ", "YMUC369HPU", "XprtDzDL74", "WK8qt7nAB1", "TAZKh2kssu", "S7l92uUKMv", "RNVfPQeSuX", "RMub2UGETq", "Q5CqEt76zb", "OH9BZhoCQJ", "LKnBYxsNbh", "JpQn5pxRTj", "Ig4P4AEuZB", "GrVYcxStul", "Fozt5LK0fJ", "DocN4zQc4U", "Dl21cA2RKT", "CqfwRZ2qZd", "BAEox3ZQ4w", "1KUndkBM2w", "14VFq4tlaP", "0vf8hoA4cq" ], "note_type": [ "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "meta_review", "official_review", "decision", "official_comment", "official_comment", "official_review", "official_review" ], "note_created": [ 1730691366943, 1732115182254, 1732529737989, 1732112964232, 1732501179377, 1732113162419, 1732278598912, 1732113513126, 1732669036847, 1732115237719, 1732849957612, 1732113249160, 1732624591482, 1732242623016, 1732115144922, 1732112674857, 1732113357816, 1732674953182, 1732478422082, 1730706783982, 1732115211188, 1732112823997, 1732737765889, 1734541125259, 1730729851767, 1737523950284, 1732529638999, 1732529442821, 1730590944941, 1730693154454 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission8952/Reviewer_76BC" ], [ "ICLR.cc/2025/Conference/Submission8952/Authors" ], [ "ICLR.cc/2025/Conference/Submission8952/Authors" ], [ "ICLR.cc/2025/Conference/Submission8952/Authors" ], [ "ICLR.cc/2025/Conference/Submission8952/Authors" ], [ "ICLR.cc/2025/Conference/Submission8952/Authors" ], [ "ICLR.cc/2025/Conference/Submission8952/Authors" ], [ "ICLR.cc/2025/Conference/Submission8952/Authors" ], [ "ICLR.cc/2025/Conference/Submission8952/Authors" ], [ "ICLR.cc/2025/Conference/Submission8952/Authors" ], [ "ICLR.cc/2025/Conference/Submission8952/Authors" ], [ "ICLR.cc/2025/Conference/Submission8952/Authors" ], [ "ICLR.cc/2025/Conference/Submission8952/Reviewer_LgQe" ], [ "ICLR.cc/2025/Conference/Submission8952/Reviewer_76BC" ], [ "ICLR.cc/2025/Conference/Submission8952/Authors" ], [ "ICLR.cc/2025/Conference/Submission8952/Authors" ], [ "ICLR.cc/2025/Conference/Submission8952/Authors" ], [ "ICLR.cc/2025/Conference/Submission8952/Authors" ], [ "ICLR.cc/2025/Conference/Submission8952/Reviewer_9USD" ], [ "ICLR.cc/2025/Conference/Submission8952/Reviewer_XwVW" ], [ "ICLR.cc/2025/Conference/Submission8952/Authors" ], [ "ICLR.cc/2025/Conference/Submission8952/Authors" ], [ "ICLR.cc/2025/Conference/Submission8952/Reviewer_XwVW" ], [ "ICLR.cc/2025/Conference/Submission8952/Area_Chair_kKYk" ], [ "ICLR.cc/2025/Conference/Submission8952/Reviewer_LgQe" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission8952/Authors" ], [ "ICLR.cc/2025/Conference/Submission8952/Authors" ], [ "ICLR.cc/2025/Conference/Submission8952/Reviewer_9USD" ], [ "ICLR.cc/2025/Conference/Submission8952/Reviewer_uec1" ] ], "structured_content_str": [ "{\"summary\": \"The paper presents a novel method to efficiently compress Key-Value (KV) cache for large language models (LLMs) like LLaMA2-7B and Mistral-7B, which can be a bottleneck in terms of storage and memory transfer. The authors propose using low-rank orthogonal projection matrices, initialized with PCA and further fine-tuned with a distillation objective, to compress the feature dimension of the KV cache. The novel Matryoshka training strategy allows adaptive selection of compression levels for different layers and heads, balancing model performance and compression rate. Experiments demonstrate high data efficiency, with the method achieving over 90% performance while compressing the KV cache by 60% on average.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"3\", \"strengths\": [\"The paper introduces a novel method to compress KV cache by focusing on the feature dimension. By employing low-rank projection matrices combined with orthogonality constraints, the authors efficiently reduce the KV cache size without requiring retraining from scratch, allowing the compression mechanism to be integrated directly into pre-trained models.\", \"The proposed training strategy to fine-tune orthogonal projection matrices effectively preserves model performance while allowing adaptive compression rates, providing a flexible approach to balance resource usage.\", \"The use of heterogeneous compression rates across different layers and heads is well-motivated and effectively demonstrated\"], \"weaknesses\": [\"Although the paper briefly mentions other KV cache compression methods, such as sharing KV headers across layers and merging redundant tokens, it lacks a detailed comparison to highlight the advantages of the feature-dimension based compression. Including experimental comparisons or a more thorough discussion of the advantages and disadvantages of each approach would strengthen the contribution and clarify the unique benefits of the proposed method.\", \"The justification for using predefined schedules for Matryoshka strategy and the heterogeneous compression rates with greedy search could be made stronger with more theoretical backing or detailed analysis. For example, are the results sensitive to different schedules? Is the greedy search algo. deterministic and with the grantee to converge, and how's the scalability?\", \"The Figure 4 seems to indicate the Matryoshka strategy is much more important than greedy search (yellow and green lines are closed in right), and Orthogonal Constraint has less effect when Cache Utilization<0.5. More discussion and analysis on these findings are encouraged.\"], \"questions\": \"see weakness\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"General Response (Part 3/4)\", \"comment\": [\"[Combination with KIVI @76BC, @XwVW, @9USD]: In Appendix E, we add the results of the combination of our method with KIVI [3], also presented below:\"], \"table_2\": \"Results of Combination of Distilled MatryoshkaKV Projections and KIVI (2bit KV cache quantization) on Six Benchmarks. We use uniform compression levels for inference here for simplicity.\\n| **Model** | **Budget** | **Method** | **HLSG** | **ARC-C** | **ARC-E** | **PIQA** | **WG** | **CSQA** | **Avg.** |\\n|-------------------|------------|------------|----------|-----------|-----------|----------|---------|----------|----------|\\n| **LLaMA2 7B-base** | 100.0% | MKV | 70.89 | 36.95 | 53.26 | 76.39 | 61.56 | 67.08 | 60.98 |\\n| | | MKV+KIVI | 69.76 | 35.93 | 51.98 | 76.55 | 61.48 | 66.26 | 60.49 |\\n| | 87.5% | MKV | 70.87 | 36.95 | 51.15 | 76.17 | 61.80 | 64.95 | 60.47 |\\n| | | MKV+KIVI | 70.45 | 36.61 | 50.79 | 76.44 | 61.01 | 64.13 | 59.91 |\\n| | 75.0% | MKV | 69.30 | 33.90 | 54.67 | 75.90 | 61.09 | 63.23 | 60.08 |\\n| | | MKV+KIVI | 68.62 | 32.20 | 54.85 | 76.06 | 60.30 | 63.23 | 59.68 |\\n| | 62.5% | MKV | 67.25 | 36.27 | 53.62 | 75.52 | 59.27 | 64.46 | 59.39 |\\n| | | MKV+KIVI | 66.56 | 35.25 | 51.68 | 75.41 | 59.43 | 61.59 | 58.33 |\\n| | 50.0% | MKV | 65.08 | 33.56 | 52.03 | 74.81 | 57.54 | 60.36 | 56.98 |\\n| | | MKV+KIVI | 63.25 | 32.54 | 51.15 | 74.43 | 57.38 | 59.46 | 56.35 |\\n| | 37.5% | MKV | 61.02 | 29.83 | 49.21 | 73.45 | 55.64 | 55.36 | 54.09 |\\n| | | MKV+KIVI | 57.11 | 28.81 | 48.85 | 71.71 | 55.64 | 50.37 | 52.08 |\\n| | 25.0% | MKV | 50.61 | 25.76 | 45.33 | 69.64 | 54.30 | 43.90 | 47.96 |\\n| | | MKV+KIVI | 48.12 | 27.80 | 42.86 | 67.85 | 53.59 | 40.54 | 46.78 |\\n\\nThe results show that our MatryohskaKV can be easily combined with KV quantization techniques and achieve a higher compression rate.\\n\\n- [Baseline @LgQe, @9USD]: In Appendix F, we add some baselines such as ASVD [4] and compare with our MatryoshkaKV\\uff0c with the results listed below:\", \"table_3\": \"Comparison between our MatryoshkaKV and baseline ASVD. We use uniform compression levels for inference here for simplicity.\\n| **Model** | **Budget** | **Method** | **HLSG** | **ARC-C** | **ARC-E** | **PIQA** | **WG** | **CSQA** | **Avg.** |\\n|-----------------------|------------|------------|----------|-----------|-----------|----------|--------|----------|----------|\\n| LLaMA2 | 100.0% | baseline | 74.00 | 35.93 | 50.97 | 78.50 | 61.64 | 65.93 | 61.16 |\\n| | 95% | ASVD | 71.12 | 36.95 | 52.20 | 76.28 | 62.35 | 66.67 | 60.92 |\\n| | | MKV | 72.59 | 36.27 | 53.09 | 76.44 | 62.43 | 66.75 | 61.25 |\\n| | 90% | ASVD | 70.45 | 34.92 | 52.03 | 75.63 | 61.72 | 64.70 | 60.06 |\\n| | | MKV | 72.30 | 36.61 | 54.50 | 76.50 | 62.90 | 65.93 | 62.03 |\\n| | 85% | ASVD | 67.23 | 35.93 | 50.26 | 74.86 | 60.38 | 62.16 | 59.29 |\\n| | | MKV | 72.33 | 35.93 | 53.26 | 76.33 | 61.80 | 64.78 | 61.13 |\"}", "{\"title\": \"Sincerely looking forward to the further discussions\", \"comment\": \"Dear reviewer,\\n\\nWe kindly inquire if our response has addressed your concerns. If so, we would greatly appreciate your reconsideration of our work and potential score adjustment.\\n\\nIf you have any additional questions or suggestions, we would be happy to have further discussions.\\n\\nBest regards,\\n\\nThe Authors\"}", "{\"title\": \"Response to Reviewer uec1\", \"comment\": \"We are grateful to the author for providing us with numerous suggestions regarding theory and analysis.\\n\\n**W1, W2:** The paper lacks rigorous theoretical analysis of why their proposed MatryoshkaKV method works better than PCA-based approaches. \\n\\nWe conduct an in-depth analysis of a single head within a single layer in Appendix A. We find that to achieve the minimum error, it is essential to jointly optimize the projection matrices for both K and V. \\nFor example, if the principal components of K and V just lie in each other's orthogonal complement, the approximation error would be extremely large. \\nThis shows that methods like PCA, which do not consider the interaction between K and V, are suboptimal. \\nUnfortunately, due to the nonlinear relationship (softmax function) in the calculation between K and V within the attention mechanism, certain mathematical methods such as GSVD (Generalized Singular Value Decomposition) are inapplicable for jointly optimizing the projection matrices of both K and V.\\n\\nMoreover, the optimal solution also varies with the input data distribution. \\nGiven the difficulty in modeling the distribution of all corpora globally, we believe that using a data-driven approach for optimization is a reasonable way to minimize the error of the model after KV cache compression on most tasks. \\nThus, we make these orthogonal matrices trainable to obtain optimal results. \\nBy directly fine-tuning the projection matrices to maximize the data likelihood, we can better adapt to different data distributions and task requirements, thereby enhancing the overall performance of the model.\\n\\n**Q1:** Why does the Matryoshka training strategy work better than static compression ratios? What's the theoretical justification?\\n\\nThe static strategy has a significant drawback in that it suffers from poor generalization. \\nLow-rank projections fail to adapt and perform effectively at compression ratios that are not encountered during the training phase, as proven by results in Figure 4. \\nIn contrast, our proposed strategy, which **employs randomly sampled compression ratios on each attention head during the training process**, offers several notable advantages:\\n- The randomly sampled compression ratios during training endow the low-rank projections with the ability to generalize across all possible compression ratios.\\n- Setting various compression ratios across different attention heads effectively decouples the working ratios of different heads. This means that during the testing phase, we have the flexibility to set different ratios for different heads, which aligns with the observed phenomenon of anisotropy in different heads as reported in some related works we've discussed in Section 5.4.\\n\\n**Q2:** How sensitive is the method to the choice of sampling schedule for compression rates during training?\\n\\nThanks for the question. We clarify **our training approach is not sensitive to the selection of the schedule**. \\nWe have added experiments using other schedules in **Table 4 in General Reply**, where the results do not change significantly.\\n\\n**Q3:** Why is PCA initialization critical for convergence? Could other initialization strategies work?\\n\\nThanks. We identify the importance of PCA initialization empirically. \\nWe have attempted initialization with *randn_init*, *kaiming_init*, and *xaiver_init*, yet none of these methods have proven to be effective.\\nThis is due to the introduction of the orthogonality constraint upon the projection matrices, which makes the optimization process probably unstable and suffers from a cold start. \\n\\n**W3:** Limited evaluation on very long sequence tasks.\\n\\n**Our approach to KV cache compression is on the feature dim axis rather than the sequence length axis, thus, it applies to both short and long texts**.\\nWe have supplemented experiments on Longbench to support this view. Our results are displayed in **Table 1 in General Reply**.\\n\\nDue to GPU memory constraints, we set the maximum length of training samples to 2048 during CPT. \\nAccordingly, we truncate LongBench samples to 2048 to calculate perplexity. \\n**On average, the perplexity only increases by 0.40, 0.76, and 1.30 at 75%, 50%, and 37.5% KV cache budget**.\\nMeanwhile, the experiments of combining our method with H2O can also prove that we can achieve a higher compression ratio for long context, which further demonstrates that **our method can effectively relieve the I/O bottleneck for long context**.\\n\\nWe will add the results of 4K length (the maximum length for LLaMA2-7B-base training samples) in the updated version.\"}", "{\"title\": \"Response to Reviewer 9USD\", \"comment\": \"We group 8 task-dependent samples into a single batch and perform a unified forward pass to measure the deviation of the model's output from the original. A total of approximately 160 forward passes are required, taking around 3\\u20134 minutes on a single 40GB A100 GPU.\\n\\nOn one hand, we update the compression ratios of multiple heads in parallel (we updated the compression ratios of 32 heads in one go), which significantly reduces the time of greedy search while maintaining its effectiveness.\\n\\nOn the other hand, since the adaptive rate for a specific task remains constant, these adaptive rates can be pre-calculated. \\nThus, when testing on the same task again, the greedy search operation is no longer necessary.\"}", "{\"title\": \"Response to Reviewer 76BC (Part 1/2)\", \"comment\": \"We would like to express our sincere gratitude for your in-depth and insightful comments on our paper.\\nYour feedback has provided us with valuable directions for improvement, and we truly appreciate the time and effort you have dedicated to reviewing our work.\\n\\n**W1:** It lacks a detailed comparison to highlight the advantages of the feature-dimension based compression\\n\\nThanks for the question. We clarify that there are no specific advantages of feature-dimension based compression compared to works on other dimensions. Nevertheless, the exploration of compression in the feature dimension has been largely uncharted previously due to the inherent difficulties of compressing features of deep models. \\n\\nThe enabling of the combination of feature-dimension compression with works on other dimensions is another important contribution of this work. \\nWe have demonstrated the ability to combine our MatryoshkaKV with the Group Query Attention (GQA) method in Section 5, by equipping Mistral-7V-v0.3-base with our trainable projections. \\n\\nAdditionally, we combined our methods with H2O (token eviction and merging)[1] and KIVI (KV quantization)[2], and the results are presented in **Table 1 in General Reply** and **Table 2 in General Reply**. \\n\\n**W2:** The justification for using predefined schedules for the Matryoshka strategy and the heterogeneous compression rates with greedy search could be made stronger with more theoretical backing or detailed analysis.\\n\\nThanks for the comment. We clarify that **our training approach is not highly sensitive to the selection of the schedule**. We have added experiments using other schedules in **Table 4 in General Reply**. It can be observed that the results do not change significantly. \\n\\nAs for the greedy search for the adaptive compression levels algorithm, **it is only employed after the training process to obtain heterogeneous compression rates**. During the inference phase, this rate remains unchanged.\\nTherefore, it has no connection with whether the training converges or not. During the training, we merely **randomly mask the feature dimension of each head**. We have found that normal convergence can be achieved in this way.\\n\\n\\n[1]: Zhang et al., \\\"H2O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models\\\", NeurIPS 2023.\\n\\n[2]: Liu et al., \\\"KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache\\\", ICML 2024.\\n\\n[3]: Aditya Kusupati, et al. Matryoshka representation learning. Advances in Neural Information Processing Systems, 2022. \\n\\n[4]: DeepSeek-AI\\u00a0at al., \\\"DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model\\\", arXiv 2024.\"}", "{\"title\": \"Reponse to Reviewer 76BC\", \"comment\": \"Thank you for your meticulous review, which has played a pivotal role in improving our work.\"}", "{\"title\": \"Response to Reviewer 9USD (Part 2/2)\", \"comment\": \"**W2 (Q3):** It would be ideal to have a task-agnostic KV cache compression scheme. If the second stage of training can just be limited to projection tuning\\uff1f\\n\\nWe clarify our experiments on continual pretraining are what you desired. We add the SFT experiments to only prove that our method can effectively handle certain specific or more challenging tasks (such as GSM8K). \\nGenerally, LLMs also require task-specific SFT for handling such tasks.\\n\\nMoreover, the advantage of our 2-stage training pipeline is that, **compared to CPT, it has lower costs and requires less time**. \\nFor example, for our SFT on GSM8K, the first stage only consumes 0.5 A100(40G) GPU $\\\\times$ hour, and the second stage only consumes 1.5 A100(40G) GPU $\\\\times$ hour. \\nThis is much less than the cost of the CPT scenario, which consumes 30 A100 GPU $\\\\times$ hour. \\nDespite the lower cost, satisfactory performance can be achieved on more difficult tasks (such as GSM8K). \\nCompared with the normally LoRA-finetuned Llama2-7b (Lora rank is set as 8), we can achieve approximately 90% accuracy with only 50% of the KV cache.\\n\\nWe have attempted to tune merely orthogonal projections, but we find that this would significantly affect the convergence of the model. \\nWe believe that for some difficult downstream tasks such as GSM8K, perhaps the model requires a higher degree of freedom to fit this data distribution.\\n\\n**W3:** Baselines & Using the approach concurrently with other techniques.\\n\\nThanks for your advice on supplementing baselines. \\n\\nWe first clarify that our PCA baseline is roughly identical to Eigen-Attention [2], and the reproduction details are described in Section 3.2).\\nThe results of the PCA baseline in our paper are similar to those in Eigen-Attention [2] (see Table 2 in [2] and Table 1 in ours).\\nWe will make it clear in the upcoming revisions. \\n\\nBesides, **we have incorporated ASVD [3] as our baseline, the results are supplemented in Table 3 in General Reply**. \\n\\nAs ASVD only furnishes checkpoints for three cache budgets, namely 85%, 90%, and 95%, we have thus compared our method with ASVD within these three specific budget scenarios.\\nThe results exhibit that our MatyoshkaKV performs better than ASVD at all three cache budget levels.\\n\\n\\n**In Section 5, we apply MatryoshkaKV to Mistral-v0.3-7B-base, and it evidently demonstrates enhanced compression capabilities when used together with Group Query Attention (GQA). \\nMoreover, we also combine our methods with H2O(token eviction and merging) and KIVI (a KV quantization technique)** \\n\\n\\n \\nFor the combination with H2O, our results are supplemented in **Table 1 in General Reply**. According to the results, when MatryoshkaKV and H2O are used concurrently, the perplexity on long contexts only increases by 1.02 at a 10% KV cache budget.\\nAdditionally, if we compress by 50% on both the sequence length and feature dimension axes (with an actual cache usage rate of 25%), we can achieve an average accuracy of 55.85 on 6 benchmarks, which is 91.32% of the baseline. \\nThis is much better than the effect of only using 25% separately on these two individual axes.\\n\\nFor the combination with KIVI, our results are supplemented in **Table 2 in General Reply**. The combination of MatryoshkaKV and KIVI doesn't lead to a significant decline in accuracy on six benchmarks mentioned in Section 5, showing that MatryoshkaKV can be integrated with KV quantization to achieve a higher compression ratio.\\n\\n**W4:** Results on the more recent family of Llama models (Llama 3/3.1) are missing.\\n\\nThank you for the comment. First of all, we would like to clarify that we have covered multiple architectures, including LLaMA2-7B-base and Mistral-7B-v0.3-base. \\nThe difference between LlaMA3 and LLaMA2 is not significant, which mainly lies in whether GQA is used.\\nWe've already demonstrated that our method can be compatible with GQA by applying our MatryoshkaKV to Mistral-7B-v0.3-base.\\nTherefore, it can be presumed that our method basically works for LLaMA3 as well. \\nAdditionally, related works such as MiniCache [4] have also conducted experiments on LlaMA2 and Mistral.\\nWe will attempt to include the results of Llama3 in the final version.\\n\\n[1]: Roberts et al., \\\"TensorNetwork: A Library for Physics and Machine Learning\\\", arXiv 2019\\n\\n\\n[2]: Saxena et al., \\\"Eigen Attention: Attention in Low-Rank Space for KV Cache Compression.\\\", ArXiv 2024.\\n\\n[3]: Yuan et al., \\\"ASVD: Activation-aware Singular Value Decomposition for Compressing Large Language Models\\\", ArXiv 2024\\n\\n[4]: Liu, A., Liu, J., Pan, Z., He, Y., Haffari, G., & Zhuang, B. \\\"MiniCache: KV Cache Compression in Depth Dimension for Large Language Models\\\", NeurIPS 2024.\"}", "{\"title\": \"Thanks for your feedback\", \"comment\": \"Thank you for recognizing our work. Your invaluable suggestions have played a crucial role in enhancing the quality of our manuscript.\"}", "{\"title\": \"General Response (Part 1/4)\", \"comment\": [\"We thank the reviewers for their thoughtful feedback. We are encouraged that the reviewers found our MatryoshkaKV has comprehensive evaluations (@LgQe, @76BC) and robust performance (@LgQe, @XwVW, @uec1, @9USD), being flexible (@LgQe, @XwVW, @76BC) and novel for KV cache compression (@uec1, @76BC).\", \"**We have made a revision of our previous paper and highlighted the modified parts in red**. The following are the modifications we have made to our paper:\", \"In the experiment part of Section 5, specifically in the subsection of CPT, **we add numerous experimental details**, including:\", \"[Clarification @9USD]: The training dataset used in CPT and its source.\", \"[Consumption @LgQe]: The runtime metrics for the experimental part.\", \"[Clarification @LgQe]: The $\\\\Delta r$ used in the greedy search for adaptive compression rates before the inference stage.\", \"**We provide a more detailed explanation of our method:**\", \"[Clarification @LgQe]: In Section 4, we clarify the whole pipeline of our method.\", \"[Analysis @uec1]: In Appendix A, we refine some mathematical formulae.\"]}", "{\"title\": \"Thanks for your feedback\", \"comment\": \"We sincerely appreciate your thoughtful feedback and the improved score. Your meticulous review has been invaluable in significantly enhancing the quality of our work.\"}", "{\"title\": \"Response to Reviewer 76BC (Part 2/2)\", \"comment\": \"**W3 (part 1):** Figure 4 seems to indicate the Matryoshka strategy is much more important than the greedy search\\n\\nThanks for your comments. \\nMatryoshka strategy **employs randomly sampled compression ratios on each attention head during the training process**.\\nIt ensures the hierarchization of the orthogonal matrices and decouples the compression ratios of different heads during training, enabling users to arbitrarily choose KV cache compression ratios for tasks of different difficulties. \\nAs shown in Figure 4 right, without the Matryoshka strategy, our orthogonal projections fail to adapt and perform effectively at compression ratios that are not encountered during the training phase.\\n\\nOur greedy search algorithm is to set different ratios for different heads, which aligns with the observed phenomenon of anisotropy in different heads as reported in some related works we've discussed in Section 5.4.\\n\\n**It is noted that without the Matryoshka strategy, it is impossible for us to find the optimal ratio on each attention head through greedy search**. Therefore, the two need to be used in combination. \\nOn the one hand, each head may **only** perform well at the trained static compression ratio. \\nOn the other hand, the compression ratios between each head are also mutually coupled. \\nThis makes it impossible for us to use the greedy search algorithm to find the **independent** compression ratio on **each** attention head.\\n\\nUniform compression rates indeed work well in some cases, but not for all tasks. \\nAs demonstrated in Table 3 of Appendix D, on certain challenging tasks like CSQA and ARCC, employing a uniform compression rate can sometimes result in a significant drop in accuracy. \\nFor instance, on ARCC at a 50% cache budget, after applying the greedy search algorithm, the accuracy rises from 34.34 to 36.61. \\nThis indicates that in practice, the flexible identification of compression levels for different heads may be necessary.\\n\\n\\n**W3 (part 2):** Orthogonal Constraint has less effect when Cache Utilization<0.5.\\n\\nWhen orthogonality is not maintained, the model can still learn primary and secondary information and transform the cache features into spaces with reduced dimensions.\\n\\n\\nHowever, without orthogonality between each column vector within the orthogonal matrices, the space spanned by the cache feature may not be fully utilized due to a lower rank.\\nThis can lead to a noticeable degradation in performance, especially at higher cache ratios. \\n\\nTo sum up, orthogonality ensures that the full-rank space is effectively utilized, supporting consistent performance across different cache budgets.\\n\\n[1]: Zhang et al., \\\"H2O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models\\\", NeurIPS 2023.\\n\\n[2]: Liu et al., \\\"KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache\\\", ICML 2024.\\n\\n[3]: Aditya Kusupati, et al. Matryoshka representation learning. Advances in Neural Information Processing Systems, 2022. \\n\\n[4]: DeepSeek-AI\\u00a0at al., \\\"DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model\\\", arXiv 2024.\"}", "{\"title\": \"Score change\", \"comment\": \"Thanks for the detailed rebuttal. I appreciate the author's effort. I will raise my score.\"}", "{\"comment\": \"Thanks for the detailed and careful rebuttal. I appreciate the author's effort. I will keep my score and positive support on the paper.\"}", "{\"title\": \"General Response (Part 4/4)\", \"comment\": [\"[Schedule Choice @LgQe, @uec1]: In Appendix G, we include experiments that demonstrate how our choice of predefined compression rate schedules during training impacts the results, which are also listed here:\"], \"table_4\": \"Accuracy of our MatryoshkaKV after CPT on six benchmarks. We use uniform compression levels for inference here for simplicity. Different hyper-parameters are compared. In the table we donate the schedule $ M_{1} = \\\\frac{i}{4} d (i = 1,2,3,4) $ , and the schedule $ M_{2} = \\\\frac{i}{8} d (i = 1,2, ...,8) $. We use uniform compression levels for inference here for simplicity.\\n| **Model** | **Budget** | **Method** | **HLSG** | **ARC-C** | **ARC-E** | **PIQA** | **WG** | **CSQA** | **Avg.** |\\n|-------------------|------------|------------|----------|-----------|-----------|----------|--------|----------|----------|\\n| LLaMA2 7B-base | 100.0% | $\\\\mathcal{M}_{1}$ | 72.03 | 36.61 | 52.56 | 76.71 | 61.64 | 67.16 | 62.07 |\\n| | | $\\\\mathcal{M}_{2}$ | 72.05 | 37.29 | 52.38 | 76.66 | 61.72 | 67.32 | 61.24 |\\n| | 87.5% | $\\\\mathcal{M}_{1}$ | 72.03 | 37.29 | 53.09 | 76.28 | 62.75 | 65.77 | 62.18 |\\n| | | $\\\\mathcal{M}_{2}$ | 72.22 | 35.93 | 52.20 | 76.28 | 62.12 | 65.27 | 60.67 |\\n| | 75.0% | $\\\\mathcal{M}_{1}$ | 70.79 | 34.92 | 53.62 | 76.88 | 60.54 | 65.03 | 61.31 |\\n| | | $\\\\mathcal{M}_{2}$ | 70.98 | 34.58 | 55.20 | 76.77 | 61.56 | 63.64 | 60.46 |\\n| | 62.5% | $\\\\mathcal{M}_{1}$ | 69.03 | 32.88 | 52.91 | 74.86 | 59.19 | 64.54 | 59.69 |\\n| | | $\\\\mathcal{M}_{2}$ | 69.22 | 37.29 | 55.73 | 75.22 | 59.35 | 64.21 | 60.17 |\\n| | 50.0% | $\\\\mathcal{M}_{1}$ | 66.34 | 32.88 | 53.09 | 74.97 | 58.25 | 62.49 | 58.59 |\\n| | | $\\\\mathcal{M}_{2}$ | 66.62 | 34.24 | 52.91 | 75.46 | 58.41 | 62.00 | 58.27 |\\n| | 37.5% | $\\\\mathcal{M}_{1}$ | 61.55 | 31.19 | 49.91 | 73.83 | 56.27 | 52.09 | 53.78 |\\n| | | $\\\\mathcal{M}_{2}$ | 62.38 | 32.20 | 50.26 | 73.34 | 56.67 | 55.28 | 55.02 |\\n| | 25.0% | $\\\\mathcal{M}_{1}$ | 50.91 | 26.10 | 44.97 | 68.39 | 52.72 | 38.33 | 46.38 |\\n| | | $\\\\mathcal{M}_{2}$ | 51.91 | 27.46 | 44.44 | 69.64 | 54.54 | 44.39 | 48.73 |\\n\\n- [Inference speed @LgQe, @9USD]: We evaluate the inference speed of our LLM equipped with MatryoshkaKV. During the inference process with a batch size of 32, our current implementation consumes a slightly faster time than the baseline full-KV model.\", \"table_5\": \"The inference speed of our MatryoshkaKV under the uniform compression rate during the inference process with a batch size of 32.\\n| | LLaMA2-7B-base | 100% | 87.5% | 75% | 62.5% | 50% | 37.5% | 25% |\\n|------|------|------|------|------|------|------|------|------|\\n| Tokens per second | 33.65 | 34.12 | 34.08 | 34.90 | 35.27 | 36.42 | 36.75 | 37.22 |\\n\\n[1]: Saxena et al., \\\"Eigen Attention: Attention in Low-Rank Space for KV Cache Compression.\\\", ArXiv 2024.\\n\\n[2]: Zhang et al., \\\"H2O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models\\\", NeurIPS 2023.\\n\\n[3]: Liu et al., \\\"KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache\\\", ICML 2024.\\n\\n[4]: Yuan et al., \\\"ASVD: Activation-aware Singular Value Decomposition for Compressing Large Language Models\\\", ArXiv 2024\"}", "{\"title\": \"Reponse to Reviewer LgQe\", \"comment\": \"We would like to express our sincere gratitude for your careful review and the valuable concern you raised regarding the time consumption of our work. Your feedback has provided us with an opportunity to further clarify and enhance the description of our method.\\n\\n**W1 (Q1):** Lack of Runtime Evaluation.\\n\\nWe first clarify that in the continual pretraining (CPT) experiments, the process of our approach is (1) obtaining the PCA initialization based on a small subset of a general corpus, (2) training our model on the CPT corpus, (3) searching for the heterogeneous compression levels for various heads with a small calibration dataset (5 - 10 samples) on the specific task of concern, and (4) performing inference on that task given the identified compression levels. \\nThe time for steps (1-3) can be substantially amortized by multiple downstream tasks, and thus should not be a concern. Besides, the training time for our approach on the CPT corpus is 8 hours using 4 A100 GPUs. \\nRegarding the wall-clock time for per inference step, our current implementation consumes a similar time as the baseline full-KV model:\\n\\nThe inference speed of our MatryoshkaKV under the uniform compression rate during the inference process with a batch size of 32.\\n| | LLaMA2-7B-base | 100% | 87.5% | 75% | 62.5% | 50% | 37.5% | 25% |\\n|------|------|------|------|------|------|------|------|------|\\n| Tokens per second | 33.65 | 34.12 | 34.08 | 34.90 | 35.27 | 36.42 | 36.75 | 37.22 |\\n\\nWe clarify our paper mainly focuses on reducing storage and memory transfer costs instead of runtime, and we have proven that our approach indeed consumes significantly less storage cost than the baseline. \\nWe will make these points clear in the revised revision. \\n\\n**W2:** Missing State-of-the-Art Comparisons.\\n\\nSorry for the missing comparisons. We first clarify that our PCA baseline is roughly identical to Eigen-Attention [1] (the details are described in Section 3.2) and has a similar performance as Eigen-Attention [1] (see Table 2 in [1] and Table 1 in ours), so we take our PCA baseline as a surrogate of Eigen-Attention [1]. Our existing results have reflected the superiority of our method. \\n\\nRegarding HeadKV, we first clarify that it has not been open-sourced. It uses low-rank orthogonal matrices and merges multiple heads into one. There are certain similarities between HeadKV and our method: both use orthogonal matrices and have basically the same parameterization. However, the dimensions in which we and HeadKV perform compression are inherently different, and HeadKV also does not further optimize the orthogonal matrices, hence still suffers from suboptimal performance due to the confounding issues. We have added these discussions to the revision and promise to add HeadKV to the empirical comparison once HeadKV is open-sourced.\\n\\nBesides, we have **compared to another related work, ASVD [2]**, the results are supplemented in **Table 3 in General Reply**.\\n\\nASVD notices the low-rank property of LLM parameters and compresses the KV cache while compressing the model parameters. Since it only releases checkpoints for three cache budgets of 85%, 90%, and 95%, we mainly compare our method with them in these three budgets. The results exhibit that our MatyoshkaKV performs better than ASVD at all three cache budget levels.\\n\\n\\n**Q2:** How is $\\\\Delta r$ determined in the greedy search algorithm? And what values are used in experiments? \\n\\nWe simply choose $\\\\Delta r = d / 8$ in the greedy search algorithm, because our predefined compression rate schedule is { $ \\\\frac{i}{8} (i=1,2, ...,8) $ }, These values can be further tuned for better empirical results, left as future work. \\n\\n**Q3:** MKV seems to work well with uniform compression rates. Does this always apply to all tasks? \\n\\nWe clarify uniform compression rates do not always work well for all tasks. \\nAs shown in Table 3 of Appendix D, on some challenging tasks such as CSQA and ARCC, using a uniform compression rate may lead to an accuracy drop of more than 2 percentage points.\\nSpecifically, at a 50% cache budget on ARCC, after applying the greedy search algorithm, the accuracy increases from 34.34 to 36.61.\\nThus, the flexible identification of the compression levels for various heads can be necessary in practice. \\nAnd, we would like to comment that the greedy search-based identification algorithm can be efficient enough in practice, using a calibration set of only 5 - 10 samples. \\n\\n**Q4:**: k\\u2192r?\\n\\nWe sincerely apologize for this error. We have already corrected it. \\n\\n[1]: Saxena et al., \\\"Eigen Attention: Attention in Low-Rank Space for KV Cache Compression.\\\", ArXiv 2024.\\n\\n[2]: Yuan et al., \\\"ASVD: Activation-aware Singular Value Decomposition for Compressing Large Language Models\\\", ArXiv 2024\"}", "{\"title\": \"Response to Reviewer 9USD (Part 1/2)\", \"comment\": \"Thank you for your careful reading and many pertinent suggestions.\\nWe are also very grateful for your advice on supplementary experiments. \\nWe have revised the unclear parts in the paper as quickly as possible and presented the results according to your experimental suggestions.\\n\\n**W1 (Q2):** Which calibration dataset was used for the results presented in Table 1? The necessity of training every time?\\n\\nFirstly, we would like to clarify that in the continual pretraining scenario, **our approach encompasses two datasets: a training set, Redpajama, which is task-agnostic and utilized for training our orthogonal matrices via single-time continue pre-training (CPT), and a calibration set which is task-dependent and employed solely for the greedy search to attain adaptive compression rates (training-free) before inference on the specific task**. \\n\\nSecondly, the calibration set is randomly sampled from the downstream dataset and composed of merely 5-10 samples. We will have a more clear description in the revised version.\\n\\n**Q1:** Can the authors clarify how they achieve performance benefits with heterogeneous ranks across the head dimension? \\n\\nIndeed, what we have done based on padding currently results in an actual time that is even slightly worse than that of the uniform approach. \\nHowever, this can be resolved through system-level optimizations (e.g., flattening the caches of different heads for storage, see [AdaKV](https://github.com/FFY0/AdaKV)), which is not the main technical focus of this paper. \\nThe upper bound of the theoretical acceleration of our approach exists.\\n\\nOn the other hand, in practice, our uniform strategy is also quite powerful. \\nIt can also achieve an average accuracy of 55.02 on 6 benchmarks with a 37.5% cache budget, which is 89.96% of the baseline. \\nThe uniform strategy doesn't require those system implementations and takes about the same amount of time as the original attention in real-world scenarios as listed below.\\n\\nThe inference speed of our MatryoshkaKV under the uniform compression rate during the inference process with a batch size of 32.\\n| | LLaMA2-7B-base | 100% | 87.5% | 75% | 62.5% | 50% | 37.5% | 25% |\\n|------|------|------|------|------|------|------|------|------|\\n| Tokens per second | 33.65 | 34.12 | 34.08 | 34.90 | 35.27 | 36.42 | 36.75 | 37.22 |\\n\\nWe clarify our paper mainly focuses on reducing storage and memory transfer costs instead of runtime, and we have proven that our approach indeed consumes significantly less storage cost than the baseline. \\nWe will make these points clear in the revised revision. \\n\\n\\n[1]: Roberts et al., \\\"TensorNetwork: A Library for Physics and Machine Learning\\\", arXiv 2019.\\n\\n[2]: Saxena et al., \\\"Eigen Attention: Attention in Low-Rank Space for KV Cache Compression.\\\", ArXiv 2024.\\n\\n[3]: Yuan et al., \\\"ASVD: Activation-aware Singular Value Decomposition for Compressing Large Language Models\\\", ArXiv 2024\\n\\n[4]: Liu, A., Liu, J., Pan, Z., He, Y., Haffari, G., & Zhuang, B. \\\"MiniCache: KV Cache Compression in Depth Dimension for Large Language Models\\\", NeurIPS 2024.\"}", "{\"title\": \"Kindly requesting reconsideration\", \"comment\": \"Dear reviewer,\\n\\nThe revision deadline is approaching, and we sincerely hope to receive any concerns or questions you may have beforehand. If there are any issues, please let us know at your earliest convenience so we can address them promptly.\\n\\nBest regards,\\n\\nThe Authors\"}", "{\"title\": \"Updated Rating\", \"comment\": \"The authors have resolved most of my concerns, so I am happy to increase my score.\", \"i_have_one_question_related_to_the_task_dependent_greedy_search\": \"does it mean that whenever a model is used for inference on a new task, it needs to go through the greedy search for adaptive compression rates? Can the authors clarify the increase in inference latency/run-time due to this step?\"}", "{\"summary\": \"The authors study the problem of KV cache compression. While existing work focuses on compression along the layer number, the head number, and the sequence length, the authors work on the feature dimension. While PCA is the most intuitive approach, it does not provide good enough performance. Instead, the authors propose to directly tune the orthogonal projection matrices with a distillation objective using an elaborate Matryoshka training strategy.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. The paper is easy to follow.\\n2. Much stronger performance than the PCA baseline when the compression ratio is low.\", \"weaknesses\": \"1. It is unclear whether the novelty of the paper is significant.\\n2. The paper does not compare with the methods that compress the other dimensions. Thus, it is unclear whether the proposed method is more effective. It is also unclear whether the proposed method can be combined with the others while maintaining its effectiveness.\", \"questions\": \"1. Can the author provide more insight so that the novelty is more than just the direct application of the Matryoshka training strategy proposed by the other paper?\\n2. Can the author compare with one to two baselines that compress the other dimensions to show compressing the feature dimension is more effective? Or can the author show that they can be combined together to further improve the performance?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"General Response (Part 2/4)\", \"comment\": \"**Simultaneously, we add some experiments requested by the reviewers:**\\n- [Combination with H2O @76BC, @uec1, @XwVW, @9USD]: In Appendix E, we add the results of the combination of our method with H20 [2] and show perplexity on LongBench, which are also listed below.\", \"table_1\": \"Results of Combination of Distilled MatryoshkaKV Projections and H2O across Seven Benchmarks. We use uniform compression levels for inference here for simplicity. The first and second columns indicate the individual compression rates along two axes. If H2O uses 20% cache on the sequence length axis and MatryoshkaKV uses 50% cache on the feature dimension axis, the overall cache utilization is 10%.\\n| **H\\u2082O** | **MKV** | **LongBench** | **HLSG** | **ARC-C** | **ARC-E** | **PIQA** | **WG** | **CSQA** | **Avg.** |\\n|---------|---------|---------------|----------|-----------|-----------|----------|--------|----------|----------|\\n| 100% | 100% | 4.17 | 72.05 | 37.29 | 52.38 | 76.66 | 61.72 | 67.32 | 61.24 |\\n| | 87.5% | 4.44 | 72.22 | 35.93 | 52.20 | 76.28 | 62.12 | 65.27 | 60.67 |\\n| | 75.0% | 4.57 | 70.98 | 34.58 | 55.20 | 76.77 | 61.56 | 63.64 | 60.46 |\\n| | 62.5% | 4.70 | 69.22 | 37.29 | 55.73 | 75.22 | 59.35 | 64.21 | 60.17 |\\n| | 50.0% | 4.93 | 66.62 | 34.24 | 52.91 | 75.46 | 58.41 | 62.00 | 58.27 |\\n| | 37.5% | 5.47 | 62.38 | 32.20 | 50.26 | 73.34 | 56.67 | 55.28 | 55.02 |\\n| | 25.0% | 7.66 | 51.91 | 27.46 | 44.44 | 69.64 | 54.54 | 44.39 | 48.73 |\\n| 75% | 100% | 4.18 | 70.71 | 36.61 | 52.38 | 76.55 | 60.54 | 66.50 | 60.55 |\\n| | 87.5% | 4.44 | 71.42 | 35.25 | 53.09 | 76.33 | 59.91 | 64.62 | 60.74 |\\n| | 75.0% | 4.57 | 70.31 | 34.34 | 54.14 | 76.39 | 59.27 | 62.90 | 59.94 |\\n| | 62.5% | 4.70 | 68.47 | 36.27 | 54.32 | 75.41 | 58.48 | 63.96 | 59.89 |\\n| | 50.0% | 4.94 | 66.00 | 32.54 | 51.50 | 75.63 | 57.30 | 61.43 | 57.46 |\\n| | 37.5% | 5.47 | 61.50 | 32.88 | 49.21 | 73.01 | 55.09 | 55.12 | 54.63 |\\n| | 25.0% | 7.67 | 51.32 | 27.80 | 44.09 | 69.37 | 53.59 | 44.55 | 48.47 |\\n| 50% | 100% | 4.20 | 68.72 | 33.22 | 52.20 | 76.12 | 56.67 | 64.78 | 58.62 |\\n| | 87.5% | 4.46 | 67.89 | 34.58 | 51.85 | 76.28 | 55.88 | 62.00 | 58.13 |\\n| | 75.0% | 4.59 | 66.01 | 35.59 | 53.79 | 75.41 | 54.54 | 62.00 | 58.05 |\\n| | 62.5% | 4.73 | 63.59 | 34.92 | 51.32 | 75.68 | 55.25 | 60.52 | 57.04 |\\n| | 50.0% | 4.96 | 61.33 | 36.10 | 50.74 | 73.67 | 55.57 | 57.67 | 55.85 |\\n| | 37.5% | 5.50 | 59.26 | 29.83 | 49.91 | 73.61 | 53.04 | 54.14 | 53.29 |\\n| | 25.0% | 7.71 | 49.44 | 26.44 | 41.80 | 68.72 | 52.96 | 43.24 | 46.94 |\\n| 20% | 100% | 4.40 | 61.55 | 25.76 | 41.27 | 73.29 | 53.28 | 47.01 | 49.98 |\\n| | 87.5% | 4.65 | 61.36 | 30.51 | 39.86 | 73.72 | 52.09 | 49.06 | 50.94 |\\n| | 75.0% | 4.79 | 60.29 | 28.47 | 38.62 | 72.75 | 53.12 | 50.45 | 50.62 |\\n| | 62.5% | 4.93 | 58.77 | 26.78 | 39.86 | 70.84 | 52.72 | 49.30 | 50.58 |\\n| | 50.0% | 5.19 | 56.39 | 26.78 | 38.10 | 71.22 | 51.62 | 49.16 | 49.66 |\\n| | 37.5% | 5.74 | 52.12 | 23.39 | 34.22 | 68.50 | 52.17 | 41.44 | 44.82 |\\n| | 25.0% | 8.01 | 43.22 | 21.02 | 31.92 | 63.93 | 51.38 | 33.09 | 40.75 |\\n\\n\\nAccording to the results, by concurrently using MatryoshkaKV and H2O, the perplexity on long contexts increases by merely 1.02 at 10% KV cache budget. \\nAdditionally, if we compress by 50% on both the sequence length and feature dimension axes (with an actual cache usage rate of 25%), we can achieve an average accuracy of 55.85 on 6 benchmarks, which is 91.32% of the baseline. \\nThis is much better than the effect of only using 25% separately on these two individual axes.\"}", "{\"title\": \"Reponse to Reviewer XwVW\", \"comment\": \"We sincerely thank the reviewer for the time to read our paper. We are glad you thought our paper was easy to follow.\\n\\n**Q1 (W1):** It is unclear whether the novelty of paper is significant.\\n\\nThanks for the comment. We would like to clarify that solving the KV cache compression problem with the proposed trainable orthogonal projection and Matryoshka training strategy is non-trivial. \\nTo achieve KV compression, many previous works are based on token eviction and merging. In contrast, we focus on **compression in the feature dimension**, which is completely orthogonal and compatible with them. The corresponding results are displayed in our response to Q2(W2).\\n\\nBy tuning the orthogonal projection with the Matryoshka strategy, we have addressed the issue of training-free PCA to lead to sub-optimal solutions. The Matryoshka strategy ensures the hierarchy of the columns in the orthogonal matrices during training, enabling users to arbitrarily choose KV cache compression ratios for tasks of different difficulties. We have also experimentally demonstrated the importance of this strategy in the Ablation part of Section 5.3. \\n\\nWe kindly ask the reviewer to re-evaluate the contribution of this paper and welcome more specific comments on this.\\n\\n**Q2 (W2):** Can author show MatryoshkaKV can be combined to further improve performance?\\n\\nThanks for the kind suggestion. In particular, the application of MatryoshkaKV to Mistral-v0.3-7B-base in Section 5 can serve as a combination of our method with other dimensions of KV compression because Mistral-v0.3-7B-base uses the Group Query Attention (GQA), which has already compressed the KV in head-number axis. \\n\\nAdditionally, we **combine our methods with H2O, which operates on the sequence axis**, and the results are supplemented in **Table 1 in General Reply**. We see our approach can be effectively combined with H2O, the perplexity on long contexts increases by merely 1.02 at 10% KV cache budget.\\nAdditionally, if we compress by 50% on both the sequence length and feature dimension axes (with an actual cache usage rate of 25%), we can achieve an average accuracy of 55.85 on 6 benchmarks, which is 91.32% of the baseline. \\nThis is much better than the effect of only using 25% separately on these two individual axes.\\n\\nMoreover, the results of the **combination of MatryoshkaKV and KIVI (a KV quantization technique)** are also supplemented in **Table 2 in General Reply**.\\n\\nThe results also reflect our MatryoshkaKV can be used in synergy with KV cache techniques on other dimensions.\"}", "{\"comment\": \"Hi,\\n\\nI have read the response and I raised my score to 6 due to the experiment showing that MatryoshkaKV can be combined to further improve performance with other compression methods.\\n\\nSincerely yours,\\nReviewer XwVW\"}", "{\"metareview\": \"Summary:\\nMatryoshkaKV introduces a novel method for compressing Key-Value (KV) cache in large language models along the feature dimension using trainable orthogonal projection matrices. The method combines PCA initialization with knowledge distillation and a Matryoshka training strategy to enable adaptive compression rates across different layers and attention heads. The approach achieves significant compression (up to 95%) while maintaining over 90% of the original model's accuracy.\", \"main_strengths\": [\"Novel approach to KV cache compression focusing on the previously unexplored feature dimension\", \"Strong empirical results showing significant compression with minimal performance degradation\", \"Compatibility with other compression techniques (demonstrated with GQA, H2O, and KIVI)\", \"Practical implementation requiring minimal hyperparameter tuning\", \"Comprehensive ablation studies validating key components\"], \"main_weaknesses\": [\"Limited theoretical analysis explaining why the method outperforms PCA-based approaches\", \"Runtime evaluation could be more comprehensive\", \"Current implementation of heterogeneous ranks may not provide runtime benefits without system-level optimizations\", \"Task-specific calibration requirement for optimal compression rates\"], \"additional_comments_on_reviewer_discussion\": \"Outcomes from Author-Reviewer Discussion:\", \"the_authors_have_addressed_several_key_concerns_through_their_responses\": [\"Provided runtime metrics showing comparable inference speed to baseline\", \"Demonstrated compatibility with other compression methods\", \"Clarified the calibration process (5-10 samples per task)\", \"Added comparisons with additional baselines (ASVD)\", \"Explained the practical implications of heterogeneous compression rates\", \"Reviewer Agreement/Disagreement:\", \"Initial ratings ranged from 5 to 8, with most reviewers increasing their scores after author responses. Consensus emerged around accepting the paper, acknowledging its practical value despite some theoretical limitations.\"], \"suggestions_for_improvement\": [\"Strengthen theoretical analysis of why the method outperforms PCA\", \"Add system-level optimizations for heterogeneous compression\", \"Include more comprehensive runtime evaluations\", \"Clarify the calibration process and its practical implications\", \"Consider additional experiments with newer model architectures\"]}", "{\"summary\": \"The paper introduces MatryoshkaKV, a method for compressing the Key-Value (KV) cache in large language models (LLMs) to reduce memory during inference. The method begins with PCA for initial dimensionality reduction but addresses PCA\\u2019s limitations by tuning projection matrices through knowledge distillation and applying Matryoshka training strategy to enable adaptive compression, allowing the model to balance performance and compression. Furthermore, this paper demonstrates effectiveness with high compression rates while maintaining relatively high accuracy across various LLMs on both CPT and SFT tasks.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. The proposed Matryoshka training strategy effectively preserves hierarchical structures in orthogonal matrices inherited from PCA at various compression levels, ensuring robust performance across dimensions.\\n\\n2. Greedy search algorithm effectively adapts to differing sparsity in each \\n$\\ud835\\udc4a_\\ud835\\udc58$ and $\\ud835\\udc4a_\\ud835\\udc63$ matrix, showcasing flexibility in compression rates across layers.\\n\\n3. There are comprehensive MKV evaluations across cache budgets, which reveals substantial improvements, particularly under extremely low cache budget.\", \"weaknesses\": \"1. Lack of Runtime Evaluation: The absence of runtime metrics makes it challenging to assess the practical benefits of this method fully (see Questions).\\n\\n2. Missing State-of-the-Art Comparisons: Unusually, the paper doesn\\u2019t thoroughly compare to existing state-of-the-art methods. Although it mentioned the other methods may collapse under 60% cache budget (lines 126-131), a comparison with Eigen-Attention and HeadKV at different cache budgets and tasks in terms of both performance and runtime would strengthen the evaluation.\", \"questions\": \"1. Although the paper mentions it only needs processing 2 million training tokens (line 104), it does not clarify the runtime for each base model and task. Please provide the runtime details for the KV compression process, including PCA initialization, the greedy search for compression level selection, and fine-tuning in both CPT and SFT tasks. Notably, since both the greedy search and compression levels rely on outputs from the original model, this could potentially double the training and inference times.\\n\\n2. How is $\\\\Delta\\ud835\\udc5f$ determined in the greedy search algorithm? And what values are used in the experiments? \\n\\n3. In Fig.4, MKV seems to work well with uniform compression rates. Does this always apply to all tasks? \\n\\n3. Line 96: k\\u2192r?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}", "{\"title\": \"Sincerely looking forward to the further discussions\", \"comment\": \"Dear Reviewer,\\n\\nWe hope that our clarifications and detailed explanations of our methods have adequately addressed your concerns. If our responses have resolved your queries, we kindly hope you might reconsider adjusting your score.\\nShould you have any further questions or suggestions, we would be more than happy to engage in additional discussions to improve our work further.\\n\\nThank you for your time and thoughtful review.\\n\\nBest regards,\\n\\nThe Authors\"}", "{\"title\": \"Thanks for your thoughtful and comprehensive feedback\", \"comment\": \"We sincerely thank you for taking the time to reassess our work. Your valuable suggestions have been instrumental in helping us improve and refine our manuscript.\"}", "{\"summary\": \"This work proposes a KV cache compression technique for efficient LLM inference through low-rank projections along the feature dimension. Using principal component analysis (PCA) over key and value matrices to obtain low-rank counterparts leads to performance loss, especially at high compression ratios. To tackle this, the authors propose to tune the projection matrices for each layer and attention head with a distillation loss. Results are presented for Llama-2 (7b) and Mistral-v0.3 (7b) models.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. This work targets KV cache compression, a crucial problem for efficient LLM inference at large sequence lengths. Results show extensive improvements over vanilla PCA for a variety of downstream tasks.\", \"weaknesses\": \"1. The training to obtain orthogonal projection matrices involves a KL divergence loss, which ensures that the KV compressed model performance stays close to the original model. However, this makes the strategy task-dependent by using a form of calibration dataset for the downstream task itself, leading to the necessity of training every time one needs a compressed KV cache for performing inference on certain task(s). Additionally, there is no clarification on the calibration dataset used for continual pretraining experiments in Section 5.1.\\n2. For results in Section 5.2, the authors propose a 2-stage training pipeline: 1. LoRA (standard fine-tuning) 2. Updating projection matrices and LoRA parameters jointly. While the first stage is generally employed to improve the performance of LLMs on downstream tasks, the second stage is the associated overhead with this form of KV cache compression. Additionally, the joint update in this stage implies the need to compress the KV cache specifically for each downstream task. It would be ideal to have a task-agnostic KV cache compression scheme.\\n3. Comparisons with baselines are missing and/or somewhat ambiguous. Is the PCA baseline in Table 1 the same as Eigen Attention [1]? Another missing potential baseline is ASVD [2], which also involves training-free low-rank projection to reduce the KV cache footprint. The authors clarify that a variety of works compress the KV cache by targeting the sequence length or channel dimension, but don't demonstrate the possibility of using their approach concurrently with such techniques [3,4].\\n4. Results on the more recent family of Llama models (Llama 3/3.1) are missing but crucial to establish the effectiveness of this approach.\\n\\n[1] Saxena et al., \\\"Eigen Attention: Attention in Low-Rank Space for KV Cache Compression.\\\", ArXiv 2024.\\n\\n[2] Yuan et al., \\\"ASVD: Activation-aware Singular Value Decomposition for Compressing Large Language Models\\\", ArXiv 2024.\\n\\n[3] Zhang et al., \\\"H2O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models\\\", NeurIPS 2023.\\n\\n[4] Liu et al., \\\"KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cache\\\", ICML 2024.\", \"questions\": \"1. Can the authors clarify how they achieve performance benefits with heterogeneous ranks across the head dimension? Different dimensions across the heads may require some form of padding before concatenating them during actual inference, so it would be great to see some hardware performance numbers as well.\\n2. Which calibration dataset was used for the results presented in Table 1? \\n3. For the SFT setup described in Section 5.2, it would be interesting to see if the second stage of training can just be limited to projection tuning instead of the proposed joint tuning with LoRA parameters as well.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper proposes MatryoshkaKV, a method to compress the key-value (KV) cache in large language models (LLMs) along the feature dimension using trainable orthogonal projection matrices. As LLMs grow in size, the KV cache can become a bottleneck in storage and memory transfer. Previous approaches have focused on compressing the cache along the layer, head, and sequence length dimensions. This work explores compressing along the feature dimension.\\n\\nThe authors first investigate using PCA to obtain orthogonal projection matrices for dimensionality reduction of the keys and values in each attention head. While this works well at moderate compression levels without needing training, performance degrades quickly at higher compression.\\n\\nTo improve on this, they propose MatryoshkaKV which tunes the orthogonal projection matrices end-to-end using a knowledge distillation objective and a special \\\"Matryoshka\\\" training strategy that enables adaptively searching for optimal compression rates per layer and head at inference time. The orthogonality of the projections is enforced using a Cayley parameterization.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"The paper tackles the problem of KV cache compression in LLMs from a new angle by focusing on the feature dimension. While prior work has explored compressing along the layer, head, and sequence length dimensions, this work shows that significant compression gains can also be achieved along the feature axis. This opens up a promising new direction for efficient LLM inference.\\n\\nThe MatryoshkaKV method demonstrates impressive performance in experiments. It can compress KV caches by 60-75% on average while retaining over 90% of the full model's accuracy. This is a significant improvement over the PCA baseline, especially at high compression rates. The results hold across both continual pre-training and supervised fine-tuning settings, showing the approach is robust and widely applicable.\", \"weaknesses\": \"The paper lacks rigorous theoretical analysis of why their proposed MatryoshkaKV method works better than PCA-based approaches\\n\\nWhile they provide some error analysis in Appendix A, it's relatively brief and doesn't fully explain the theoretical underpinnings of their method's superior performance\\n\\nLimited evaluation on very long sequence tasks where KV cache compression would be most valuable\", \"questions\": \"Why does the Matryoshka training strategy work better than static compression ratios? What's the theoretical justification?\\n\\nHow sensitive is the method to the choice of sampling schedule for compression rates during training?\\n\\nWhy is PCA initialization critical for convergence? Could other initialization strategies work?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}" ] }

[ "Dan Ley", "Suraj Srinivas", "Shichang Zhang", "Himabindu Lakkaraju" ]

[ "generalized", "group", "data attribution", "efficiency", "training data", "influence", "tracin", "trak" ]

{ "note_id": [ "xMTuaoV78A", "xExRaNsTpN", "wt1pdnmMUy", "wK2TTlA261", "pJN4k57y22", "nArbEcKZTt", "kvHt68v1ZX", "hYZ2YENYKK", "fwXDgMnd4l", "fsX4e2zn0Q", "dO8YwkELye", "QjIPX8wRvV", "QFq1XVhE43", "NwzxgivZtl", "NQkmhSEvTi", "IWti56kMyG", "BUqGmvmGbC", "A41dGWOl63", "8gnTfKcvBG", "8Ae5zwIZ2M", "4bpqn8ncXo" ], "note_type": [ "official_comment", "official_comment", "meta_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "decision", "official_review", "official_comment", "official_review", "official_comment", "official_comment" ], "note_created": [ 1732305769175, 1732302325634, 1734545133932, 1733172183929, 1732300814758, 1733200567378, 1732576913790, 1732602777851, 1730705514532, 1733174454219, 1730234829057, 1732553350605, 1732317687643, 1733174996246, 1733171273711, 1737524187154, 1729735507104, 1733173971542, 1730675580379, 1732295082383, 1733172566190 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission12355/Authors" ], [ "ICLR.cc/2025/Conference/Submission12355/Reviewer_cACW" ], [ "ICLR.cc/2025/Conference/Submission12355/Area_Chair_CY2j" ], [ "ICLR.cc/2025/Conference/Submission12355/Authors" ], [ "ICLR.cc/2025/Conference/Submission12355/Authors" ], [ "ICLR.cc/2025/Conference/Submission12355/Reviewer_X7t6" ], [ "ICLR.cc/2025/Conference/Submission12355/Reviewer_X7t6" ], [ "ICLR.cc/2025/Conference/Submission12355/Reviewer_VGjW" ], [ "ICLR.cc/2025/Conference/Submission12355/Reviewer_VGjW" ], [ "ICLR.cc/2025/Conference/Submission12355/Authors" ], [ "ICLR.cc/2025/Conference/Submission12355/Reviewer_X7t6" ], [ "ICLR.cc/2025/Conference/Submission12355/Reviewer_hzLg" ], [ "ICLR.cc/2025/Conference/Submission12355/Authors" ], [ "ICLR.cc/2025/Conference/Submission12355/Authors" ], [ "ICLR.cc/2025/Conference/Submission12355/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission12355/Reviewer_cACW" ], [ "ICLR.cc/2025/Conference/Submission12355/Authors" ], [ "ICLR.cc/2025/Conference/Submission12355/Reviewer_hzLg" ], [ "ICLR.cc/2025/Conference/Submission12355/Authors" ], [ "ICLR.cc/2025/Conference/Submission12355/Authors" ] ], "structured_content_str": [ "{\"title\": \"Rebuttal by Authors\", \"comment\": \"Thank you for your thoughtful review of our paper. Please see our response to the weaknesses and the questions below.\\n\\n> W1: large scale experiments\\n\\nWe are actively working on running experiments on ImageNet. This requires significant computational resources and running time. We will update the paper with these results as soon as they are available.\\n\\n> W2: Lack of K-Means analysis and runtimes.\\n\\nYes, we totally agree that an analysis of K-means runtime and providing more details about the code will enhance the paper, especially for reproducibility. Our K-means is based on a public implementation https://github.com/subhadarship/kmeans_pytorch. Our code is available in the supplementary material.\\n\\nIn analyzing computational costs, we would like to first highlight the distinction between two types of costs for any attribution method:\\n\\n1. The offline training time cost, which refers to all computational overhead incurred during the preprocessing and model preparation phase. This includes data processing, model training, and any preparatory steps needed before deployment. Many retraining-based attribution methods like TRAK inherently have significant training time costs, as they require training multiple model checkpoints for ensemble. \\n\\n2. The online serving time cost, which refers to the computational cost incurred during the actual attribution score computation for each test point. \\n\\nThe difference between these two types of costs is that the offline training time cost is incurred only once, and will be amortized over many test points, while the online serving time cost is incurred once for each attribution computation.\\n\\nThe K-means clustering in our approach belongs entirely to the offline training time cost category since it can be performed offline as a preprocessing step. The clustering is completed before seeing any test points and does not impact the online serving. Therefore, we choose to compare the serving time cost of attribution methods for practical consideration.\\n\\nNevertheless, we provide a detailed K-means runtime analysis below. Following the notation in [1], we let n be the number of training data points, p be the number of model parameters, d be the (hidden) feature dimension, and m be the number of test data points we want to compute the attribution for. Additionally, we let T be the number of training iterations (for TracIn), and J be the number of checkpoints to be ensembled (for TRAK). The results are shown in the Table below.\\n\\n|Time Complexity|IF|TracIn|TRAK|\\n|-|-|-|-|\\n|Original|O(mnp)|O(mnpT)|O(mnpJ) + O(npTJ)|\\n|GGDA|O(mKp) + O(Knd)|O(mKpT) + O(Knd)| O(mKpJ) + O(npTJ) +O(Knd)|\\n\\nWe start by explaining IF. Its run time is O(np) for each test data point [1], and O(mnp) in total for the entire test set. The GGDA version of IF will include the K-means computation, but only once. K-means has time complexity O(Knd), but using GGDA can bring down the per test instance time complexity to O(Kp). Therefore, GGDA-IF has time complexity O(mKp) + O(Knd). \\n\\nSimilarly, TracIn has time complexity O(mnpT). TRAK has a serving time complexity O(mnpJ), but also additional model training time O(npTJ). \\n\\nFor all cases, the GGDA runtime will be significantly better when K << n, which is the setting we adopt. Also, as m gets larger, the efficiency gain of GGDA will be further enhanced.\\n\\nWe also show that GGDA is much more efficient empirically even counting K-means time. In the tables below, we show the GGDA total attribution time (column 3) for IF-LiSSA, TracIn, and TRAK. The runtime of the original DA methods is bolded (row 1 with group size 1, i.e., n = K), whereas GGDA can be more than ten times faster even with K-means for larger K. \\n\\n> IF-LiSSA\\n|GroupSize(n/K)|MeanGroupingTime|MeanAttributionTime|MeanTotalTime|\\n|-|-|-|-|\\n|1|0.00|612.90|**612.90**|\\n|4|169.03|194.84|363.88|\\n|16|127.63|77.62|205.26|\\n|64|92.25|37.82|130.06|\\n|256|58.87|29.34|88.21|\\n|1024|28.54|27.27|55.80|\\n\\n> TracIn\\n|GroupSize(n/K)|MeanGroupingTime|MeanAttributionTime|MeanTotalTime|\\n|-|-|-|-|\\n|1|0.00|430.17|**430.17**|\\n|4|169.03|134.48|303.51|\\n|16|127.63|57.25|184.89|\\n|64|92.25|41.02|133.27|\\n|256|58.87|36.82|95.69|\\n|1024|28.54|35.60|64.13|\\n\\n> TRAK\\n|GroupSize(n/K)|MeanGroupingTime|MeanAttributionTime|MeanTotalTime|\\n|-|-|-|-|\\n|1|0.00|4317.43|**4317.43**|\\n|4|172.80|1463.22|1636.01|\\n|16|132.48|617.34|749.82|\\n|64|93.77|405.86|499.63|\\n|256|57.68|349.92|407.59|\\n|1024|28.54|335.64|364.18|\\n\\n[1] Hammoudeh, Z., & Lowd, D. (2024). Training data influence analysis and estimation: A survey. Machine Learning, 113(5), 2351-2403.\\n\\n> Remark\\n\\nThank you once again for your insightful suggestions and comments, which have been instrumental in enhancing the quality of our paper. We believe that we have addressed all the concerns. If there is any aspect that you feel has not been fully resolved, we would be happy to provide further information. If you are satisfied with our response, we would truly appreciate your consideration in raising your evaluation score.\"}", "{\"comment\": \"I have changed my score correspondingly.\"}", "{\"metareview\": \"**Summary:** The paper proposes group data attribution to replace individual data attribution for more computationally efficient yet robust training. The method uses k-means clustering and other heuristics to group the data.\\n\\n**Strengths:** The paper is well-written. The problem and method are clearly described, and the experimental results are clearly presented.\\n\\n**Weaknesses:** Based on the reviews, the main shortcomings are the novelty, lack of theoretical analysis, and larger-scale experiments. The reviewers claim some of the claimed contributions are already known, and the construction lifts the individual data attribution to groups. Reviewer VGjW claims that this is equivalent to summing the attribution of individual points for each group. In terms of analysis, the paper lacks any guarantees or in-depth study of the k-means step. Finally, the results are mainly presented on smaller models/datasets.\\n\\n**Decision:** The current version of the paper lacks sufficient novelty, theoretical and runtime analysis, and larger-scale results to be ready for publication. I recommend rejection for the current version.\", \"additional_comments_on_reviewer_discussion\": \"The authors provide additional experimental results and try to address some of the concerns raised by the reviewers. They claim that summing individual-level scores does not always best represent the group scores. However, I am not particularly convinced this is true based on their response. The authors need to clarify this in the paper and maybe perform an ablation.\"}", "{\"comment\": \"Thank you for your patience and for engaging in our paper's discussion. Please see our global response on ImageNet results. We do observe that GGDA continues to maintain significant speedups in this large scale setting.\\n\\nWith regards to grouping/kmeans timing, the computation of penultimate layer activation gradients for the entire training set was ~1 hour / 3600 seconds and ~ 175 seconds, 955 seconds, and 3800 seconds for grad-K-Means grouping on group sizes 1024, 256 and 64 respectively. These times are small in comparison to the overall 88,000 seconds runtime for standard DA (group size 1) and may also be considered as pre-processing steps as previously mentioned to Reviewer X7t6.\\n\\nRegarding evaluation, we are unable to perform retraining based evaluations in time, since training even the ResNet-18 on ImageNet requires significant runtime (>30 hours per retraining). We kindly refer to our ResNet-18 results for CIFAR-10 (Figure 2 in the pdf), where GGDA remains within +/- 0.5% test accuracy w.r.t. DA (lower test accuracy is better in the plot).\"}", "{\"title\": \"Rebuttal by Authors\", \"comment\": \"Thank you for your thoughtful review of our paper. We would like to address the weaknesses you pointed out and provide clarification on the questions raised.\\n\\n> W1: Property function is not a contribution\\n\\nWe agree that the property function is not a novel contribution of our work. We simply wanted to highlight that our method is capable of handling general property functions as an additional feature. We have changed the writing in our revised version to clarify this point and avoid any potential misunderstanding about our claimed contributions.\\n\\n\\n> W2a: Computational advantage doesn\\u2019t apply to Hessian computation\\n\\nThis is a very astute observation regarding the Hessian! It is true that modifying the Hessian estimation is critical to reducing the computational complexity of attribution estimation. While our draft has neglected to mention this in detail, our implementation also subsequently involves Hessian approximations, where we denote the Hessian as $$H_{\\\\theta} = \\\\sum_z \\\\nabla^2 \\\\ell(z)$$, thus rendering it to be a sum over number of groups. We mention this for TRAK, where we refer to the subsequent Fisher approximation as a \\u201cbatched\\u201d Fisher approximation. We perform similar \\u201cbatched\\u201d Hessian approximations for influence functions using the LiSSA estimator to ensure that per-sample computations are never performed. We again apologize for neglecting to discuss these batched Hessian approximations in our draft!\\n\\n\\n> W2b: Whether batched approximation is good\\n\\nWe agree with the reviewer\\u2019s intuition that clustering algorithms are required to ensure a small approximation error. We emphasize that this is exactly the purpose of our group selection strategies! In particular, we found the gradient-based clustering to be highly effective, pointing to exactly the approximation benefit described by the reviewer. \\n\\n\\n> Q1: TracIn definition is not standard\\n\\nThanks for bringing this to our notice! Our definition in equation (3) does deviate from the one proposed in the TracIn paper with a constant term involving the learning rate pre-multiplier. If a constant learning rate schedule is used throughout training, our variant is equivalent to the one proposed in the TracIn paper. \\n\\nPlease see the discussion in Appendix A.2 and A.3 for a complete discussion of the TracIn variant we employ. \\n\\n> Q2: Grad-K-Means requires individual gradients \\n\\nWe agree that a full gradient computation will brings down efficiency. Therefore, we choose to use gradients wrt last layer activations and NOT gradients wrt parameters as we discuss in line 354. Please also note that, as discussed in the paper, it is hard to efficiently compute batched versions of gradients wrt parameters, but not so for gradients wrt activations, whose batched versions can be trivially computed. While we agree that any grouping method must scale with $n$, our grouping method in practice is cheap and does not involve computing per-sample gradients wrt weights. Furthermore, these groups only need to be computed once offline, and do not need to be recomputed for subsequent influence estimations for new test data points / model properties.\\n\\n> Q3: Typos\\n\\nThank you for pointing out the typos. We have fixed them in the revised version of the paper.\\n\\n> Remark\\n\\nThank you once again for your insightful suggestions and comments, which have been instrumental in enhancing the quality of our paper. We believe that we have addressed all the concerns. If there is any aspect that you feel has not been fully resolved, we would be happy to provide further information. If you are satisfied with our response, we would truly appreciate your consideration in raising your evaluation score.\"}", "{\"comment\": \"Thank you for the additional details on efficiency experiments. However, with only these results, I cannot adjust the evaluation score. I look forward to a future version with more comprehensive experiments, including retraining-based evaluations.\"}", "{\"comment\": \"Thank you for your detailed response. I have updated my score accordingly. However, I am still awaiting results from larger-scale experiments.\"}", "{\"title\": \"Response\", \"comment\": \"Thanks to the authors for their responses, and await the new experiments. I still am not entirely convinced that the theoretical analysis is revealing much new insight, but I also don't think that every paper needs to have a deep theoretical component - I just want to understand conceptually whether/how this is different from just treating each cluster of examples as if each were just a single \\\"large\\\" example (after all, the loss function decomposes linearly into individual loss terms).\\n\\nTo make things more concrete---suppose we were in a language modeling setting, this feels like just going from token-level to document-level attributions, which is clearly going to be more computationally efficient. Is the key just that the authors have proposed to cluster the examples in a specific way in order to best approximate the individual-level attributions?\"}", "{\"summary\": \"This paper proposes Generalized Group Data Attribution - a method for combining individual data attribution (scores indicating the influence of single training points for single test predictions) into group data attributions (scores indicating the influence of groups of training points for model properties). The resulting attributions are faster to estimate and enable a variety of downstream applications.\", \"soundness\": \"4\", \"presentation\": \"3\", \"contribution\": \"1\", \"strengths\": \"The paper is clearly written and addresses an important problem, namely the resource-intensive nature of many data attribution methods. The proposed solution is clearly explained, and the writing is clear and concise.\", \"weaknesses\": \"In my opinion, the main weakness of this paper is the novelty and depth of the investigation. As far as I can tell, the paper proposes turning a point-to-point data attribution method into a group-to-group data attribution method by effectively summing the corresponding individual attributions. This does not seem so fundamental a contribution---e.g., the fact that this reduces sample complexity from O(# points) to O(# groups) seems to follow directly by construction, as without loss of generality one can just call each group a \\\"datapoint.\\\"\\n\\nI think that a more in-depth investigation of the mechanism by which individual attributions are combined could strengthen the paper---for example, are there weighting schemes that improve performance? Are robust estimators (e.g., the median) qualitatively different than taking the average? I also think that the application section could be more fleshed out - the dataset pruning results are the most interesting to me: further investigation into the source of success of GGDA (variance reduction? Soft thresholding? Etc.) would have improved the analysis.\", \"questions\": \"See weaknesses above. Also - is there any intuition for why grad-k-means works so well as a clustering method?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you for your suggestions for helping us improve our paper. We would like to kindly let you know that we have gathered additional results for GGDA on ImageNet. Please see our global response for the results. We do observe that GGDA continues to maintain significant speedups in this large-scale setting.\"}", "{\"summary\": \"The paper \\u201cGeneralized Group Data Attribution\\u201d introduces the GGDA framework, designed to enhance data attribution efficiency by grouping training data points. GGDA aggregates training data into groups instead of handling individual points, significantly improving computational efficiency while maintaining comparable accuracy. It extends popular attribution methods like Influence Functions, TracIn, and TRAK to group-based settings, making them suitable for large-scale datasets.\\n\\nExtensive experiments on various datasets and models validate GGDA\\u2019s performance in tasks like dataset pruning and noisy label detection, demonstrating its effectiveness and scalability. However, the paper could benefit from more experiments on real-world large-scale datasets, as well as a deeper theoretical analysis of the K-Means grouping strategy, which plays a critical role in enhancing attribution efficiency yet lacks detailed discussion in the theoretical framework.\\n\\nOverall, GGDA shows promise for data attribution, but needs further validation for large-scale dataset use and a deeper theoretical analysis of the K-Means grouping strategy.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"The paper's experimental section introduces K-Means clustering in gradient space as part of the grouping strategy. This innovative design improves attribution accuracy. The approach demonstrates significant advantages in different attribution tasks, such as dataset pruning and noisy label detection, validating its applicability across various scenarios.\", \"weaknesses\": \"1. **Absence of Large-Scale Dataset Experiments**: The experiments primarily focus on small to medium-scale datasets, leaving out truly large-scale datasets (e.g., billion-level data). To better demonstrate GGDA\\u2019s scalability, future work should incorporate experiments on large-scale datasets and report both computational efficiency and attribution performance in such scenarios.\\n2. **Lack of K-Means Analysis**: K-Means plays a vital role in the proposed method's effectiveness, but the paper lacks theoretical analysis and runtime details for this component. This omission limits the evaluation of its feasibility and efficiency. Providing more detailed descriptions in the appendix or code repository would enhance reproducibility for researchers.\", \"questions\": \"1. Do the authors plan to conduct experiments on large-scale datasets (e.g., billion-level data) in future work to further validate the scalability and attribution performance of GGDA?\\n2. The paper lacks details on K-Means runtime and theoretical analysis. The authors could add these implementation details in the appendix to facilitate reproducibility, especially considering that K-Means in the gradient space can be time-consuming when the gradient space is large.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Thank you\", \"comment\": \"Thank you, I am waiting for larger-scale experiments to justify more strongly the claims of the paper and to consider raising the score.\"}", "{\"title\": \"Rebuttal by Authors\", \"comment\": \"Thank you for your thoughtful review of our paper. Please see our response to the weaknesses and the questions below.\\n\\n> W1: large-scale experiments on ImageNet\\n\\nWe are actively working on running experiments on ImageNet. This requires significant computational resources and running time. We will update the paper with these results as soon as they are available.\\n\\n> W2: individual vs batch gradient speed, and efficient implementation using vmap.\\n\\nWe agree that our statement about batched gradient computation and per-sample gradient computation wasn't clear enough. What we meant was that computing gradients of a batch of B samples is much more efficient than looping through each of those B samples and computing gradients one at a time, because the computation of a larger B is not too much slower than the case of B = 1.\\n\\nRegarding your suggestion about using efficient implementations like vmap. We were not using it but we really appreciate the suggestion. We indeed found that using vmap can speed up both per-sample gradient and batched gradients, which we illustrate in the table below. For this experiment, we consider the run time (ms) for computing gradients for 8192 training data points in four different cases, where we compare per-sample vs. batch and vmap vs. non-vmap.\\n\\n| Per-sample Grad | Per-sample Grad (vmap)| Batched Grad (B = 64) | Batched Grad (vmap) (B = 64)|\\n|-|-|-|-|\\n| 6760 | 12.5 (540x)| 560 (12x) | 2.23 (3031x)|\\n\\nWe would like to emphasize that 1) Although we were not using vmap, our comparison in the paper was fair because all implementations didn't use it 2) Optimizations like vmap are orthogonal to what we are proposing and can be combined with our GGDA method as well. \\n\\nWe will consider rerunning our experiments with vmap in our next version. We really appreciate the suggestion.\\n\\n> W3: Table \\u00b1 symbols\\n\\nThank you for seeking clarification here. The \\u00b1 symbols in the tables do indeed indicate multiple training trials across 10 different seeds, while keeping groups and attributions fixed. Table 1 presents Logistic Regression (LR) results, where there is no variation across training trials due to convexity (hence, \\u00b1 0.0). We have updated the table captions to clarify these points.\", \"w4\": \"Add baselines to pruning tables.\\n\\nThank you for the suggestion. The baseline of no removal is added to both Table 1 and Table 2, where we see a clear improvement with GGDA.\\n\\n> W5: Rationale for clustering by activation gradient vs activations\\n\\nWe investigate both clustering strategies and find that gradient k-means outperforms activation k-means in most cases. The effectiveness of gradient k-means can be attributed to that it naturally aligns with how attribution values are calculated for most methods, particularly influence functions. This alignment helps ensure that points grouped together are likely to have similar attribution values.\\n\\nHowever, this alignment is not guaranteed, which is why we investigate multiple grouping strategies in our work. The key insight is that an ideal grouping strategy should cluster together points with similar attribution values. The degree of this similarity can vary depending on the specific dataset and attribution method being used.\\n\\n> Q1: Why does grouping attribution outperform single-point attribution on data pruning?\\n\\nWe hypothesize that GGDA's superior performance in pruning stems from two key factors:\\n\\nFirst, while per-sample attributors excel at identifying highly important datapoints, they struggle with accurately estimating the least important points, which is crucial for pruning tasks. This means that while the ranking of datapoint importance is reliable for important points, it becomes less accurate for unimportant ones. We believe this may be due to numerical instabilities in Hessian estimation, particularly affecting influence function methods. The use of groups in GGDA helps mitigate this by smoothing out independent estimation errors across different points.\\n\\nSecond, the effectiveness of GGDA relies heavily on appropriate group selection strategies. For semantically coherent data groups, the attribution values of all points within the groups should be uniform. This makes the identification of coherent data groups crucial for accurate attribution estimation. In contrast, bad grouping creates semantically incoherent sets of samples, making it difficult to estimate accurate attributions for the group as a whole. This highlights why proper group selection is crucial for realizing the benefits of our approach.\\n\\n> Remark\\n\\nThank you once again for your insightful suggestions and comments, which have been instrumental in enhancing the quality of our paper. We believe that we have addressed all the concerns. If there is any aspect that you feel has not been fully resolved, we would be happy to provide further information. If you are satisfied with our response, we would truly appreciate your consideration in raising your evaluation score.\"}", "{\"comment\": \"To clarify our contributions further. We generally agree with your observation that moving from individual-level to group-level attributions\\u2014such as transitioning from token-level to document-level in a language modeling context\\u2014can naturally lead to computational efficiency gains. However, we would like to emphasize that summing individual-level scores does not always best represent the group scores. Specifically, even though the loss function can be linearly decomposed into individual loss terms, the contribution score of a group may not. For example for LOO, leaving one subset out doesn't equal leaving each individual out and then summation, which may overlook certain interactions or context-dependent effects that are better captured when scores are computed directly at the group level.\\n\\nIn our approach, rather than relying on a summation of individual attributions, we directly compute the group scores. This methodology allows us to retain the efficiency benefits of group-level computation while, in some cases, providing a more representative attribution for the group as a whole. For instance, by clustering examples based on their semantic similarity, our framework is designed to approximate group-level behavior more effectively, capturing nuances that may be missed by simply aggregating individual scores.\\n\\nWe hope this clarification addresses your conceptual question and highlights the value of our approach. Thank you again for your thoughtful engagement with our work\\u2014we look forward to incorporating your insights as we refine our manuscript.\"}", "{\"title\": \"Additional ImageNet Results\", \"comment\": \"As requested, we have performed additional experiments on **ImageNet** to demonstrate GGDAs efficiency claims, using models of increasing size: *ResNet-18*, *ResNet-34*, and *ResNet-50*. Experiments are performed using NVIDIA's H100 80GB HBM3.\\n\\n**TracIn Attribution Times (in seconds) for 1024 Training Examples:**\\n\\n| Group Size | ResNet-18 | ResNet-34 | ResNet-50 |\\n| - | - | - | - |\\n| 1 | 6.50 | 10.90 | 14.48 |\\n| 4 | 1.64 | 3.04 | 3.24 |\\n| 16 | 0.77 | 1.00 | 1.32 |\\n| 64 | 0.56 | 0.66 | 1.12 |\\n| 256 | 0.42 | 0.58 | 1.02 |\\n\\nGrouping datapoints enables a) batched gradient computation and b) batched dot products between gradients and the test property gradient vector. We demonstrate speedups of up to 15x with group size 256.\\n\\n**Influence (Fisher) Attribution Times (in seconds) for 1024 Training Examples:**\\n\\n| Group Size | ResNet-18 | ResNet-34 | ResNet-50 |\\n| - | - | - | - |\\n| 1 | 12.0 | 17.79 | 22.48 |\\n| 4 | 5.06 | 8.86 | 10.41 |\\n| 16 | 3.75 | 6.56 | 7.85 |\\n| 64 | 3.46 | 6.26 | 7.50 |\\n| 256 | 3.46 | 6.23 | 7.53 |\\n\\nHere we demonstrate 3-4x speedups on attribution time in the group setting. We also compute attribution times (in seconds) for the *entire* ImageNet training set for the ResNet-18 model. \\n\\n| Group Size | TracIn | Influence (Fisher) |\\n| - | - | - |\\n| 1 | 88217 | 88231 | \\n| 4 | 22708 | 21364 |\\n| 16 | 17529 | 17496 | \\n\\nWe observe the same significant speedups (about 5x faster for group size 16) to support our Hessian speed-up claims for influence based methods. Unfortunately, we are unable to perform evaluations on the resulting attribution scores since this requires retraining multiple models, each utilizing >30 hours of training time.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"summary\": \"This paper proposes to generalize the traditional data attribution method to consider attributing to (1) groups of data points instead of individual data points, and (2) the property function that captures model behavior beyond the test loss.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"Overall\\n1. The motivation is clear.\\n2. The definitions and the problem formulation are consistent throughout.\\n3. The experimental results are presented cleanly.\", \"weaknesses\": \"The main weaknesses are the novelty and soundness of both proposals:\\n1. The second proposal, \\\"property function,\\\" is already known in the literature as \\\"target function.\\\" For example, see [1] in statistics, and [2] in data attribution.\\n2. The first proposal, considering groups of data points, doesn't seem to have the claimed computational gain by looking at the analysis for the Influence Function.\\n - From Section 4, the analysis is true obviously since there are no conceptual differences from the influence function (together with the target function concept mentioned above). While I'm not sure about the computational complexity advantage claimed in the paper (batch gradient computation $\\\\\\\\approx$ per gradient computation), from my understanding, for whatever algorithm is used to approximate the iHVP (inverse-Hessian vector product) computation, constructing $H_{\\\\\\\\theta} = \\\\\\\\sum_{i=1}^{n} \\\\\\\\nabla_\\\\\\\\theta^2 \\\\\\\\ell(x_i)$ will inevitably scale with $n$. Hence, the claimed computational advantage is not there at least until Line 266. \\n - It's better to bring the idea of batched $H_{\\\\\\\\theta}$ as in the TRAK paragraph for $\\\\\\\\hat{F}_{\\\\\\\\theta}^{\\\\\\\\text{batched}}$ to the Influence Function analysis to demonstrate the claimed advantage.\\n\\n However, even with the batched approximation of $H_{\\\\\\\\theta}$, some theoretical justification is lacking. Whether this will be a good approximation is unclear to me. Additionally, as described in Line 319, some clustering algorithms are needed in order to obtain a small approximation error on $\\\\\\\\nabla_{\\\\\\\\theta}\\\\\\\\ell(x_i)$, which I suppose will make the algorithm scale with $n$ again.\\n\\nOverall, the second proposal (*property function*) already exists in the literature, while for the first proposal (*grouped data points*), I'm not convinced by the claimed computational efficiency gain for the Influence function and TRAK when Hessian or (empirical) FIM are involved. Without a justification for an efficient and good approximation of the Hessian (and its inverse with iHVP computation), such an extension is trivial from the linearity of IF, TRAK, and related influence-function-based methods.\\n\\n**Reference**\\n- [1]: [An Automatic Finite-Sample Robustness Metric: When Can Dropping a Little Data Make a Big Difference?](http://arxiv.org/abs/2011.14999)\\n- [2]: [Most Influential Subset Selection: Challenges, Promises, and Beyond](https://arxiv.org/abs/2409.18153)\", \"questions\": \"See Weaknesses. Additionally:\\n1. In Section 4, the definition of TracIn is not standard, at least deviating from the original paper and even ignoring the scaling. It only sums over batches that contain a particular training sample $x_i$, not over all iterations unless we're considering full-batch training. I think this should be mentioned.\\n2. The *Gradient K-Means* grouping method in the experiment suffers from the issue I raise above (2), where when we need to consider individual gradients, the claimed computational efficiency goes away.\", \"some_minor_suggestions_in_writing\": \"1. Line 120, replace $\\\\mathcal{D} = \\\\\\\\{(x_0, y_0), (x_1, y_1), \\\\\\\\ldots (x_n y_n)\\\\\\\\}$ by $\\\\\\\\mathcal{D} = \\\\\\\\{(x_0, y_0), (x_1, y_1), \\\\\\\\ldots , (x_n, y_n)\\\\\\\\}$ (two \\\",\\\" are missing).\\n2. Line 217 and 266, replace $k << n$ by $k \\\\\\\\ll n$.\\n3. Line 232, replace - with --- without spaces at the beginning (before TRAK) and the ending (after TracIn).\\n4. Line 252, replace `\\\\citep{}` with `\\\\citet{}`.\\n5. Line 276~288, tracein should all be TracIn?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you for your patience and for engaging in our paper's discussion. Please see our global response on ImageNet results. We do observe that GGDA continues to maintain significant speedups in this large scale setting.\\n\\nWith regards to grouping/kmeans timing, the computation of penultimate layer activation gradients for the entire training set was ~1 hour / 3600 seconds and ~ 175 seconds, 955 seconds, and 3800 seconds for grad-K-Means grouping on group sizes 1024, 256 and 64 respectively. These times are small in comparison to the overall 88,000 seconds runtime for standard DA (group size 1) and may also be considered as pre-processing steps.\"}", "{\"summary\": \"This paper addresses data attribution estimation, which assesses the contribution of a training sample to a model\\u2019s generalization according to a downstream performance metric. While data attribution is beneficial for tasks like data pruning and correcting mislabeled samples, it is often computationally impractical, as its demands scale linearly with the number of training samples. To tackle this, the authors propose Generalized Group Data Attribution (GGDA), which shifts attribution from individual samples to groups of samples. They demonstrate that K-means clustering on activation gradients is an effective heuristic for forming these groups. The authors reframe traditional attribution metrics, including the Leave-One-Out and gradient-based metrics, and apply GGDA to dataset pruning and noisy-label identification in small-scale experiments on MNIST, CIFAR-10, HELOC, and TRAC.\\n\\n## Claims\\n1.\\tGGDA can be applied to any sample-based data attribution method.\\n2.\\tIt trades attribution fidelity for computational efficiency.\\n3.\\tGGDA significantly speeds up data attribution.\\n4.\\tIt is effective for noisy-label identification and data pruning.\\n5.\\tGGDA enables practical applications for large-scale machine learning.\", \"soundness\": \"1\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"The paper is well-written, clearly defining introduced concepts, and is well-motivated, as improving computational efficiency in data attribution is valuable for large-scale machine learning. The authors investigate a generally applicable approach to enhance the computational efficiency of data attribution methods, as claimed. The use a variety of data (tabular, image, text) modalities to validate their approach in downstream supervised learning tasks.\", \"weaknesses\": \"Weaknesses\\n\\n1.\\tThe experimental datasets (e.g., MNIST, CIFAR-10) are relatively small, calling into question GGDA\\u2019s scalability claims for large-scale ML. Can the method be tested on a larger dataset like ImageNet? Does it maintain an effective compute-fidelity tradeoff as sample size increases?\\n2.\\tIn Section 4, line 272, the authors claim computational advantages for group data attribution. However, in line 265, they note that \\u201ca single batched gradient computation is roughly equivalent in runtime to individual per-sample gradients.\\u201d Do the results in Tables 1, 2, and 3 use the best available per-sample data attribution methods? Are implementation of individual per-sample gradient-based methods batched, for example via vmap functionals?\\n3.\\tTables lack clarity regarding \\u00b1 symbols. Do these indicate multiple trials with different seeds? Are groups recomputed for each trial? Why are the values \\u00b10.0 in Table 1?\\n4.\\tTables 1 and 2 do not include baselines for no data removal.\\n5.\\tThe rationale for clustering by activation gradient, rather than activations alone, is unclear. Aren\\u2019t gradients inherently dependent on activations? Could further intuition be provided?\", \"questions\": \"1.\\tIt is a bit surprising that group data attribution improves the data selection fidelity. How does GGDA achieve that, and could this improvement be due to the group selection heuristic rather than the method itself?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Rebuttal by Authors\", \"comment\": \"Thank you for your thoughtful review. Please see our response to the weaknesses and the questions below.\\n\\n> W1: Limited novelty\\n\\nWe want to clarify a potential misunderstanding here. While using groups instead of individual data points may seem conceptually straightforward, there are significant technical challenges and novel contributions in our work:\\n\\nFirst, determining effective grouping strategies is non-trivial. We systematically investigate and compare four different grouping approaches, each with distinct theoretical and practical implications. Such comprehensive analysis of grouping strategies is novel to our knowledge.\\n\\nSecond, extending existing attribution methods to perform well with groups requires careful mathematical consideration and non-trivial algorithmic adjustments. For instance, methods like Influence Functions (IF) need specific modifications for accurate and efficient gradient computation. The complexity lies in properly handling group-level effects while maintaining the theoretical guarantees of the original methods, which we detail in Section 4 along with the modifications needed for other methods like TracIn and TRAK.\\n\\nFurthermore, a key theoretical insight of our work is that the computational benefits do not come simply from summing individual attributions. Rather, the critical innovation is our ability to compute group attributions directly, reducing complexity from $O(points)$ to $O(groups)$ **without performing intermediate per-sample calculations**. This property holds for IF-style attributions due to their linearity, though notably not for Leave-One-Out (LOO) attributions where group effects cannot be decomposed into sums of individual effects.\\n\\n> W2: Investigation of combining individual attributions\\n\\nWe appreciate the questions about weighting schemes and robust estimators for combining individual attributions. However, we want to clarify our group attribution process. We don't combine attributions of individual points - instead, we directly compute attributions at the group level. This is a key advantage of our approach, avoiding the need to first compute individual attributions (which incurs significant computational cost).\\n\\nAs derived in our appendix, for IF specifically, the mathematically derived way to compute group attributions is through the summation of individual influences within the group. Importantly, the loss function applied to a group of points should not be mean reduced. The theoretical derivation for TRAK requires more non-trivial and more careful consideration. As such, these are not empirical choices, but follow directly from theoretical foundations.\\n\\n> W3: Why does GGDA work so well for pruning?\\n\\nWe hypothesize that GGDA's superior performance in pruning stems from two key factors:\\n\\nFirst, while per-sample attributors are good at identifying highly important data points, they struggle with accurately estimating the least important points, which is crucial for pruning tasks. This means that while the ranking of data point importance is reliable for important points, it becomes less accurate for unimportant ones. We believe this may be due to numerical instabilities in Hessian estimation, particularly affecting IF methods. The use of groups in GGDA helps mitigate this by smoothing out independent estimation errors across different points.\\n\\nSecond, the effectiveness of GGDA relies heavily on appropriate group selection strategies.\\nFor semantically coherent data groups, the attribution values of all points within the groups should be uniform. This makes the identification of coherent data groups crucial for accurate attribution estimation. In contrast, bad grouping creates semantically incoherent sets of samples, making it difficult to estimate accurate attributions for the group as a whole.\\n\\nWe believe that further investigation into these aspects presents an interesting direction for future research.\\n\\n> Q1: Why does grad-k-means work so well?\\n\\nThe effectiveness of gradient k-means can be attributed to it naturally aligning with how attribution values are calculated for most methods, particularly IF. This alignment helps to ensure that points grouped together are likely to have similar attribution values.\\n\\nHowever, this alignment is not guaranteed, which is why we investigate multiple grouping strategies in our work. The key insight is that an ideal grouping strategy should cluster together points with similar attribution values. The degree of this similarity can vary depending on the specific dataset and attribution method being used.\\n\\n> Remark \\n\\nThank you once again for your insightful suggestions and comments. We are actively incorporating the above explanations to enhance the quality of our paper. If there is any aspect that you feel has not been fully resolved, we would be happy to provide further information. If you are satisfied with our response, we would truly appreciate your consideration in raising your evaluation score.\"}", "{\"comment\": \"Thank you for your patience, for engaging in our paper's discussion and for raising your evaluation score. Please see our global response on ImageNet results. We do observe that GGDA continues to maintain significant speedups in this large scale setting.\\n\\nWith regards to grouping/kmeans timing, the computation of penultimate layer activation gradients for the entire training set was ~1 hour / 3600 seconds and ~ 175 seconds, 955 seconds, and 3800 seconds for grad-K-Means grouping on group sizes 1024, 256 and 64 respectively. These times are small in comparison to the overall 88,000 seconds runtime for standard DA (group size 1) and may also be considered as pre-processing steps as previously mentioned.\\n\\nRegarding evaluation, we are unable to perform retraining based evaluations in time, since training even the ResNet-18 on ImageNet requires significant runtime (>30 hours per retraining). We kindly refer to our ResNet-18 results for CIFAR-10 (Figure 2 in the pdf), where GGDA remains within +/- 0.5% test accuracy w.r.t. DA (lower test accuracy is better in the plot).\"}" ] }

[ "Ogechi Blessing Onuoha", "Bjørn Sand Jensen", "Jan Paul Siebert", "Frank E. Pollick" ]

[ "geolocation", "multimodal", "indoor", "deep learning", "dataset benchmark", "geolocalization" ]

{ "note_id": [ "zcnkdqwRO0", "rjxKfZP0cr", "nyA38tY7cQ", "nuX0FlVh8V", "mzwUkGWqor", "kWEOTlfnzT", "isAXdtEElH", "gI83cvKXpR", "fnGB8ElTvY", "asiZkNEGSY", "WWsuVHXGau", "QTqeZwe06x", "Q3XLP0RWUD", "JIDYGX45Rs", "HG7vztcjDq", "C5OhLpS5rR", "4i54Br5F9Y" ], "note_type": [ "official_comment", "official_review", "official_comment", "comment", "official_review", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review" ], "note_created": [ 1732810203861, 1729515225191, 1732791132916, 1736424212931, 1730766304295, 1732790095434, 1730806903383, 1732809460068, 1732790059704, 1732790049636, 1729699078514, 1732960279312, 1732805821924, 1732790137426, 1732790104870, 1732790111295, 1729952054164 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission11640/Reviewer_jNt5" ], [ "ICLR.cc/2025/Conference/Submission11640/Reviewer_jNt5" ], [ "ICLR.cc/2025/Conference/Submission11640/Reviewer_d91o" ], [ "ICLR.cc/2025/Conference/Submission11640/Authors" ], [ "ICLR.cc/2025/Conference/Submission11640/Reviewer_d91o" ], [ "ICLR.cc/2025/Conference/Submission11640/Authors" ], [ "ICLR.cc/2025/Conference/Submission11640/Reviewer_gNbp" ], [ "ICLR.cc/2025/Conference/Submission11640/Reviewer_jNt5" ], [ "ICLR.cc/2025/Conference/Submission11640/Authors" ], [ "ICLR.cc/2025/Conference/Submission11640/Authors" ], [ "ICLR.cc/2025/Conference/Submission11640/Reviewer_fgUm" ], [ "ICLR.cc/2025/Conference/Submission11640/Reviewer_G1Xc" ], [ "ICLR.cc/2025/Conference/Submission11640/Reviewer_fgUm" ], [ "ICLR.cc/2025/Conference/Submission11640/Authors" ], [ "ICLR.cc/2025/Conference/Submission11640/Authors" ], [ "ICLR.cc/2025/Conference/Submission11640/Authors" ], [ "ICLR.cc/2025/Conference/Submission11640/Reviewer_G1Xc" ] ], "structured_content_str": [ "{\"comment\": \"Select more than three geolocalization methods (e.g. PIGEON) for experimentation, and refer to the supplementary materials of CLIP to explore the performance of this dataset on other tasks. This would greatly enhance the impact of this work.\"}", "{\"summary\": \"This paper introduces a new dataset for indoor geolocation and proposes a sampling method to obtain a test set. Additionally, it demonstrates the fine-tuning of GeoCLIP on the proposed dataset.\", \"soundness\": \"2\", \"presentation\": \"1\", \"contribution\": \"3\", \"strengths\": \"1) Authors focus on a meaningful and niche (relative to large-scale outdoor scenarios) problem.\\n2) Building datasets is almost always beneficial, especially large-scale datasets.\", \"weaknesses\": \"1) The writing of this article needs to be improved, and the logic within each section is somewhat confusing. For example, the turn on L49 is not rigorous and only shows that you know little about the datasets for other tasks.\\n2) The article lacks citations for many key statements.\", \"l41\": \"such as seasonal changes, time of day, weather conditions, and human-induced modifications.\\n3) As I said in the \\u201cStrengths\\u201d, indoor geolocation (localization) is a meaningful research topic. However, the author lacks a rigorous literature review. I can name many famous indoor localization datasets without thinking: Baidu Mall(CVPR'17), InLoc(CVPR'18), NL-Indoor(CVPR'21). (Although the task objectives of these datasets are different from yours, I think they should be discussed.)\\n4) The experiments based on the proposed dataset are very limited and do not demonstrate the significance of the dataset.\\n5) Although you crawled the text descriptions of images, the word \\\"multimodal\\\" in the title is hardly seen again in the paper.\", \"questions\": \"None.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thanks for your rebuttal. I agree with your point that broken links may constitute a relatively small fraction in the future, but this largely depends on how long the \\u201cfuture\\u201d is defined. Furthermore, even the absence of a single image in the training set can potentially affect the reproducibility of every model\\u2019s results. Missing images in the test/validation set are equally critical, as they fundamentally alter the evaluation of models.\\n\\nI do not think that providing MD5 checksums is an effective solution to address this \\u2018data missing\\u2019 issue. From the perspective of datasets, the potential unavailability of data is a critical weakness that can severely limit their application in future research.\"}", "{\"withdrawal_confirmation\": \"I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors.\", \"comment\": \"Dear Reviewers,\\n\\nThank you for taking the time to review our manuscript and providing valuable feedback. Your insights have been immensely helpful in identifying areas where we can further enhance the clarity, impact, and utility of our work.\\n\\nAfter careful consideration, we have decided to withdraw the paper to address your comments more comprehensively. We are committed to refining our contributions to better demonstrate our dataset's utility and strengthen the overall impact of the research.\\n\\nThank you once again for your time and thoughtful input.\"}", "{\"summary\": \"The paper introduces INDOOR-3.6M, a large dataset specifically for indoor image geolocation, addressing the challenges posed by indoor environments that lack the rich landmarks of outdoor spaces. INDOOR-3.6M includes 3.6 million globally diverse, geotagged indoor images, accompanied by metadata for enhanced model training. Alongside this, the authors provide INDOOR-15K, a benchmark dataset for evaluating indoor-specific geolocation models, and propose a sampling strategy to ensure balanced geographic diversity. They also propose a sampling method, which combines population density and land area to ensure balanced geographic representation within the INDOOR-15K benchmark dataset.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": [\"The presentation is well-structured, making the paper clear and easy to understand.\", \"Unlike other datasets, the proposed dataset is easily accessible and distributable, as demonstrated in the supplementary materials.\", \"The proposed dataset fills a critical gap in the research community, providing much-needed resources for indoor geolocation tasks.\"], \"weaknesses\": [\"In line 231, the paper mentions collecting images from the internet that contain latitude and longitude coordinates. Is there a human review mechanism to ensure the accuracy and reliability of the geographic information in these images? Additionally, even if the images themselves are under a CC license, is there a protocol to blur potential privacy-sensitive information within the images, such as faces, intimate clothing, etc.?\", \"Could the release of this dataset lead to illegal applications, such as using images (e.g., from social media) to obtain the user locations (even if they don\\u2019t want others to know their location), thereby introducing security risks?\", \"In line 490, the proposed dataset only provides URLs. These URLs may become inaccessible over time, especially for sites where links frequently change, such as booking websites (with some hotels even removing pages). How does the dataset plan to address the potential issue of broken or inaccessible URLs?\", \"Is it reasonable to determine geographic information solely from images? Have experiments been conducted for more complex cases? For example, (1) many hotel chains use standardized decor, so to what extent can images alone reliably confirm the location of, say, the same hotel brand in the U.S. versus China? (2) Different individuals may have distinct decor and layout styles that may be more indicative of personal taste than geographic location. For instance, a Chinese staff working in the U.S. might have a TV studio with a Chinese interior style. To what extent can models accurately recognize the geographic location in such cases?\", \"The paper lacks a detailed description and explanation of the proposed IndoorGeoCLIP model. Are the authors only fine-tuning the existing GeoCLIP model using their proposed dataset? If so, this contribution might be seen as insufficiently significant and could be considered merely a necessary experiment to support the dataset.\", \"The experiments seem insufficient. The authors should consider using more existing methods to evaluate the proposed Indoor15K benchmark; currently, Table 3 includes only GeoCLIP, which seems inadequate. Additionally, in Table 3, wouldn\\u2019t it be more intuitive to label IndoorGeoCLIP directly as \\u201cGeoCLIP (fine-tuning)\\u201d?\", \"I acknowledge the authors\\u2019 introduction of an interesting sampling strategy, but they have not demonstrated its advantages through appropriate ablation studies. I also believe this sampling method could be beneficial for other geolocation datasets (both indoor and outdoor), and further verification of its performance and feasibility would be valuable.\", \"In the supplementary material, the download_images.py code essentially functions as an automated data-scraping script that downloads images through URLs. It raises the question of whether the authors have obtained proper authorization from all websites involved in the dataset to conduct automated data scraping.\"], \"questions\": \"Refer to Weaknesses.\", \"flag_for_ethics_review\": \"['Yes, Discrimination / bias / fairness concerns', 'Yes, Legal compliance (e.g., GDPR, copyright, terms of use)']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"> I acknowledge the authors\\u2019 introduction of an interesting sampling strategy, but they have not demonstrated its advantages through appropriate ablation studies.\\n\\nWe have included our ablation study to evaluate the sampling strategy\\u2019s effectiveness, as described in Appendix A.1 (lines 702\\u2013751) of the revised manuscript. These highlight our methodology's advantages in improving geographic representation and performance.\\n\\n> In the supplementary material, the download_images.py code essentially functions as an automated data-scraping script that downloads images through URLs. It raises the question of whether the authors have obtained proper authorization from all websites involved in the dataset to conduct automated data scraping.\\n\\nThe download_images.py script serves as a tool for downloading *publicly* accessible images via URLs provided in the dataset metadata. As stated in our ethics section, this dataset is intended solely for research purposes and downloading these images should . We strongly discourage its use in any way or for any applications that may breach privacy, violate ethical standards, or contravene legal norms.\"}", "{\"summary\": \"This paper introduces INDOOR-3.6M, a large-scale dataset of geotagged indoor imagery for indoor geolocation tasks. Recognizing the limitations of existing geolocation models in indoor environments, the authors propose a new sampling methodology to ensure geographic diversity and balance in their dataset. They also introduce INDOOR-15K, a benchmark dataset specifically designed for evaluating indoor geolocation models. Lastly, they showcase the dataset's utility by introducing IndoorGeoCLIP, a fine-tuned version of the GeoCLIP model, which demonstrates superior performance compared to the baseline GeoCLIP on their test set.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": [\"This paper addresses a relevant gap in image geolocation by creating a dataset specifically for indoor environments. Existing datasets primarily focus on outdoor scenes, making them unsuitable for the unique challenges of indoor geolocation. The creation of the dataset, along with the development of a specialized indoor geolocation benchmark (INDOOR-15K) are novel contributions.\", \"The dataset seems carefully curated and includes metadata, such as scene classification and object segmentation, which add to its potential applications. The use of the Places365 ResNet indoor/outdoor image classifier to filter images is good. Further, the effort to mitigate data leakage concerns by selecting images captured after 2017 for the INDOOR-15K benchmark demonstrates a commitment to quality and reliable evaluation.\", \"The paper is well-written and structured logically, making it easy to understand. The authors clearly articulate the challenges of indoor geolocation and provide a good overview of existing datasets and their limitations.\", \"The INDOOR-3.6M dataset has the potential for numerous interesting applications. The most direct is the development and evaluation of more robust and accurate indoor geolocation models. The dataset can also facilitate research in related areas, such as indoor navigation, scene understanding, and place recognition.\"], \"weaknesses\": [\"While the paper introduces IndoorGeoCLIP as a specialized model fine-tuned on their dataset, the evaluation is limited. Exploring and comparing the performance of other state-of-the-art geolocation models or techniques on INDOOR-15K would strengthen the analysis. Additionally, the authors could include a more in-depth error analysis to identify the specific challenges posed by indoor geolocation, and what it can be used for, to guide future research.\", \"The paper primarily focuses on creating and describing the dataset. It lacks a thorough demonstration of the dataset's usefulness beyond the fine-tuning of GeoCLIP. Further experiments and analyses showcasing the dataset's application in tasks like place recognition, or indoor navigation would strengthen the paper significantly.\", \"The reliance on URLs to online sources for data access could lead to unreliable data availability, where links might become broken over time, limiting the dataset's long-term usability. The authors should consider providing alternative access methods, such as potentially mirroring the links.\"], \"questions\": \"N/A\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you for your reply!\\n\\nTo be honest, Indoor-3.6M was the paper I was most eager to review during the bidding phase. However, I am somewhat disappointed overall (in terms of figures, presentation, experiments, and other aspects).\\n\\nI agree with your first two points, and this might indeed be just my personal opinion.\\n\\nHowever, regarding the third point, I maintain my view that adding experiments would enhance your contribution and help other researchers better understand this work. If you had provided additional experiments, I might have raised my score to 5 or 6. Unfortunately, you did not.\\n\\nI chose to lower my confidence level to reflect an increase to a score of 4.\"}", "{\"title\": \"Response to Concerns on Dataset Scale, Accuracy Improvement, and Sampling Strategy\", \"comment\": \"> it's unclear from Table1 that if the proposed dataset provide more indoor data than some of the other much larger scale mixed dataset like YFCC100M.\\n\\nThank you for raising this point. While YFCC100M is significantly larger in overall scale, it was curated in 2016 and may not reflect recent changes in indoor environments. INDOOR-3.6M, on the other hand, includes images uploaded as recently as 2024, ensuring the dataset captures contemporary indoor spaces. This temporal relevance, combined with its exclusive focus on indoor imagery, provides a unique advantage over older, mixed-environment datasets like YFCC100M, ensuring relevance for real-world applications, such as human trafficking investigations, where up-to-date imagery is critical.. \\n\\n> Can the fine-tuned models achieve better accuracy on the indoor subset of the other datasets? \\n\\nWe show a table showing the peformance of the finetuned models on the other datasets in Appendix A.2\\n\\n> The sampling weight of each country is determined by the weighted sum of it's population and land area. How the final weight (population * 0.3 + area * 0.7) determined? The sampling weight seems to assume that the scene visual diversity of the contries are linearly propotion to the population and land area. Is there any support for this assumption? Perhaps some measure using CLIP distance can provide some insight.\", \"the_use_of_population_and_land_area_as_proxies_for_scene_visual_diversity_is_based_on_the_idea_that\": \"- Population: Countries with larger populations are likely to have more diverse human-made structures and indoor environments, reflecting varied cultural, economic, and functional needs.\\n- Land Area: Larger countries typically encompass more diverse geographic regions, which can translate into a wider variety of architectural and interior styles.\\nWhile these factors are not perfect predictors of scene visual diversity, they offer a practical and scalable heuristic for balancing the dataset across countries. We acknowledge that this assumption could benefit from further validation. We will clarify this rationale and acknowledge the limitations of our current approach in the manuscript.\\n\\nIt is worth noting that our sampling strategy is the first to address geographic bias without requiring additional data collection, making it a practical and resource-efficient solution. By leveraging existing data, we propose an accessible approach to mitigating over- or under-representation of certain regions. Future work will explore advanced methods, perhaps including visual diversity metrics such as CLIP feature distances, to further optimize sampling and refine our methodology.\\n\\n\\n> In Table1, what is the column \\\"Benchmark\\\" actually stand for? \\n\\nThe \\\"Benchmark\\\" column in Table 1 indicates whether the dataset provides a dedicated test or evaluation set specifically designed to benchmark the performance of geolocation models. While this often implies a clear train/test split, it may also include datasets that are predefined test sets for standardized evaluation without an explicitly defined training data. We have clarified this definition in the revised manuscript to avoid any confusion.\\n\\n> Why the correctness of the city/contry/continent prediction is determined by a distance threshold \\n\\nOur choice of distance thresholds for evaluating geolocation predictions follows established practices in prior work on image geolocation, allowing direct comparison with state-of-the-art methods. Using distance thresholds as a metric ensures standardized evaluation and facilitates meaningful benchmarking across datasets and models. In addition, distance thresholds offer a gradient of accuracy that reflects the model's true geographical precision during evaluation, avoiding unfair assessments of models whose predictions are geographically close to the true location but may fall outside strict administrative boundaries.\\n\\n\\n> Scene labels, segmentations, and object detection results are provided by the dataset. What is the purpose of them? Are they serve as some additional conditions for the geolocation task?\\n\\nThese annotations enhance the dataset\\u2019s versatility and support further research in geolocation and related areas. They provide supplementary cues that can improve geolocation accuracy, such as segmentation masks to exclude irrelevant regions and scene labels or detected objects to identify localized furniture, signage, or cultural artifacts. Real-world geolocation experts, such as those in Europol\\u2019s \\u201cTrace an Object\\u201d initiative, frequently rely on objects in scenes to infer locations \\u200b(See https://www.europol.europa.eu/media-press/newsroom/news/new-trace-object-uploads-fresh-leads-needed-in-child-sexual-abuse-cold-cases and https://www.europol.europa.eu/stopchildabuse). By including these features, the dataset enables researchers to explore models that mimic such expert strategies.\"}", "{\"title\": \"Response to Concerns Regarding Data Collection, Dataset Reliability, Model Contributions, and Ethical Implications\", \"comment\": \"> Is there a human review mechanism to ensure the accuracy and reliability of the geographic information in these images? Additionally, is there a protocol to blur potential privacy-sensitive information within the images, such as faces, intimate clothing, etc.?\\n\\nWe appreciate the reviewer\\u2019s concern about the accuracy of geotags and the potential privacy implications of the dataset. We address these concerns: \\n1. Reliability of Geotags: The GPS tags of images sourced from Booking.com and Wikidata are highly reliable, as they are typically reviewed and verified by humans before being published. While the geotags from Flickr may be noisier, their inclusion is intentional to reflect real-world data conditions. Such variability provides an opportunity for researchers to develop robust models that can handle noisy geotags effectively, which is critical for real-world applications.\\n2. Privacy and Accessibility: The images included in the dataset are already publicly accessible and hosted on their original platforms, ensuring that we do not distribute sensitive or restricted content. Since these images are \\u201cin the wild,\\u201d they are already accessible to anyone under their respective licensing terms. Our approach merely aggregates links to these images, respecting their original hosting context. \\n3. Omission of Privacy-Sensitive Content: We provide segmentation mask labels for each image. These labels enable researchers to programmatically identify and exclude images containing people or other sensitive elements if required for their specific use case. This will allow researchers to customize the dataset to align with the ethical guidelines and privacy considerations for their specific application. \\n\\n> Could the release of this dataset lead to illegal applications, thereby introducing security risks? \\n\\n We understand your concern about potential misuse and have explicitly outlined in the manuscript that the dataset is intended for research purposes and we strongly discourage any applications that could lead to privacy violations or unethical outcomes.\\n\\n> How does the dataset plan to address the potential issue of broken or inaccessible URLs? \\n\\nWhile we recognize that some URLs may become unavailable over time, we anticipate this affecting only a small fraction of the dataset given its scale and diversity. Additionally, the accompanying metadata and contextual information remain valuable resources for research, even if some images are no longer accessible. To ensure data integrity, we will include the MD5 checksum of each image to the dataset to ensure that downloaded images are identical to the ones in our proposed dataset.\\n\\n> Is it reasonable to determine geographic information solely from images?\\n\\nWe appreciate the reviewer\\u2019s observations on the challenges of inferring geographic information from images. While humans may struggle to pinpoint exact locations, they excel at estimating approximate regions using semantic reasoning and data association [1,2], highlighting the feasibility of geolocation as a computational task.\\nCrowdsourced geolocation campaigns (see https://x.com/Europol/status/895978347263668224 and https://www.reddit.com/r/TraceAnObject/comments/lel9f9/tao17390_07feb2021_can_you_identify_this_hotel/), demonstrate how subtle visual markers enable humans to locate scenes, providing a basis for training models to replicate and enhance these capabilities. Prior work, like Stylianou et al.\\u2019s Hotels50K, shows that even in standardized environments, localized visual cues support accurate geolocation. Our INDOOR-3.6M dataset builds on this by offering geographically diverse images with features like objects, text, and scene-specific details that aid geolocation by capturing regionally distinctive cues.\\nWhile standardized decor or culturally specific styles pose challenges for models, INDOOR-3.6M is specifically designed to address these complexities, enabling models to learn nuanced visual cues and improve geolocation accuracy, even in ambiguous scenarios.\\n\\n> The paper lacks a detailed description and explanation of the proposed IndoorGeoCLIP model. \\n> wouldn\\u2019t it be more intuitive to label IndoorGeoCLIP directly as \\u201cGeoCLIP (fine-tuning)\\u201d? \\n\\nYour suggestion to label IndoorGeoCLIP as \\u201cGeoCLIP (fine-tuning)\\u201d for clarity is valid and we have implemented this in our revised manuscript. Regarding the inclusion of additional geolocation methods for benchmarking, we primarily focused on GeoCLIP as it is the current state-of-the-art for environment agnostic geolocation. \\n\\n\\n1. Hays, James, and Alexei A. Efros. \\\"Large-scale image geolocalization.\\\" Multimodal location estimation of videos and images (2015): 41-62.\\n2. Kohler, Rachel. Supporting Open Source Investigative Journalism with Crowdsourced Image Geolocation. Diss. Virginia Tech, 2017.\"}", "{\"summary\": \"This paper introduces a novel image geolocation dataset tailored for indoor scenes, addressing the limitations of existing datasets and establishing a benchmark for evaluating indoor image geolocation algorithms. Specifically:\\n1. The paper presents a dataset that covers a wide variety of indoor scenes and includes rich multimodal metadata, which is expected to advance the field of indoor image geolocation;\\n2. The paper proposes an innovative sampling method to obtain geographically representative samples from datasets with geographic bias; \\n3. The paper provides a standardized evaluation framework for fair assessment and comparison of research progress in indoor geolocation research.\", \"soundness\": \"1\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"i) The dataset is collected from three different sources, covering a variety of indoor scenes, thus filling the gap in indoor image geolocation datasets and providing rich multimodal information;\\nii) A sampling method that integrates population density and land area is proposed to ensure that the distribution of GPS points across regions is geographically representative;\\niii) A new benchmark dataset specifically for indoor geolocation is introduced to address the limitation of existing benchmark datasets, which primarily focus on outdoor environments.\", \"weaknesses\": \"(1) The contribution of the dataset does not seem particularly prominent, especially regarding the limitations of existing datasets mentioned in the Introduction (such as insufficient diversity, imbalanced distribution, and blurred boundaries between indoor and outdoor environments), which have not been significantly addressed;\\n(2) Compared to some of the datasets in Table 1, the proposed dataset does not show a clear advantage in terms of scale. Additionally, I would like to know the amount of indoor scene data within mixed-scene datasets (e.g., YFCC100M). It seems feasible to separate indoor and outdoor images in such datasets using image classification methods.\\n(3) The experiments in Table 3 do not adequately demonstrate the superiority of the proposed dataset for this task. It is recommended to supplement the results by providing the performance of IndoorGeoCLIP on the three datasets listed in Table 2, to further substantiate the advantages of the proposed dataset.\", \"questions\": \"1.The paper demonstrates the performance of the IndoorGeoCLIP model on various levels of geolocation tasks (such as street-level, city-level, and country-level), but it lacks comparative experiments with other classic indoor geolocation methods, making it difficult to comprehensively validate the superiority of IndoorGeoCLIP. Explicitly,\\ni) Lack of Baseline Model Comparisons: Apart from GeoCLIP, the experiments lack comparisons with other geolocation models, making it insufficient to illustrate the relative advantages of IndoorGeoCLIP in indoor scenes.\\nii ) Insufficient Ablation Studies: The experiments only show the performance changes of the GeoCLIP model before and after fine-tuning, without conducting ablation analyses on the contributions of the dataset's multimodal features (such as textual and visual data) to geolocation.\\n2. Some of the images lack sufficient clarity.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Appreciate authors effort in the rebuttal.\\n\\nI'm still not fully convinced that the number of people and the land area are good indicator of the diversity of a country. It's possible that a smaller country have more different kind of people and culture comparing to a larger country due to some historical reason.\\n\\nI understand that previous work may use distance threshold for evaluation. It is still good to propose another metrics by using the actual continent/country/region to judge if the prediction is correct.\\n\\nBy also checking the other review comments, I'm still lean to negative rating.\"}", "{\"comment\": \"Thank you for your response. However, it seems that my concerns are not fully addressed, and the explanation is not entirely convincing. I also hope you can take into account some of the issues mentioned in the weaknesses.\"}", "{\"title\": \"Response to suggestions for Enhancing Dataset Utility and Accessibility\", \"comment\": \"> While the paper introduces IndoorGeoCLIP as a specialized model fine-tuned on their dataset, the evaluation is limited.\\n\\nWe acknowledge the importance of a comprehensive evaluation and comparison with other geolocation models. We primarily focused on comparing with GeoCLIP to establish a baseline using our proposed dataset as GeoCLIP is the current state-of-the-art for environment agnostic geolocation. \\n\\n> The paper primarily focuses on creating and describing the dataset. It lacks a thorough demonstration of the dataset's usefulness beyond the fine-tuning of GeoCLIP. Further experiments and analyses showcasing the dataset's application in tasks like place recognition, or indoor navigation would strengthen the paper significantly.\\n\\nWe also appreciate your suggestion to expand on the dataset's potential applications in tasks like place recognition etc. While it is true that datasets specialized for tasks like place recognition or indoor navigation exist, our primary focus is on indoor geolocation\\u2014a task for which no large-scale, geographically diverse dataset currently exists. INDOOR-3.6M was specifically designed to address this critical gap by providing the necessary scale, geographic diversity, and multimodal metadata to advance indoor geolocation research.\\nTasks like place recognition and navigation typically rely on datasets optimized for those purposes, but such datasets are often geographically or contextually limited. While our dataset includes multimodal metadata that can support broader applications, we believe that the most significant and immediate contribution of INDOOR-3.6M lies in the indoor geolocation domain, a domain where existing datasets fall short. \\n\\n> The reliance on URLs to online sources for data access could lead to unreliable data availability, where links might become broken over time, limiting the dataset's long-term usability. The authors should consider providing alternative access methods, such as potentially mirroring the links.\\n\\nFinally, we note your concern regarding URL-based data access and carefully chose this approach to balance data availability, ethical considerations, and potential misuse risks. By linking to the original hosts of the images, we respect copyright and data protection regulations while ensuring compliance with privacy laws. This approach aligns with datasets such as LAION-5B dataset[1] (https://laion.ai/faq/), OpenImages Dataset (https://storage.googleapis.com/openimages/web/download_v7.html#df-image-information) [2] which also provide URL-based linking, and face similar challenges.\\nImportantly, hosting images solely on their original platforms introduces an additional safeguard against misuse. These platforms have established infrastructures to monitor and regulate access to their content, providing a layer of deterrence to bad actors. This is particularly critical given the potential harmful applications of geotagged indoor imagery. By ensuring that access to the images occurs through trusted hosts, we reduce the risk of the dataset being exploited for unethical purposes.\\nWhile we recognize that some URLs may become unavailable over time, we anticipate this affecting only a small fraction of the dataset given its scale and diversity. Additionally, the accompanying metadata and contextual information remain valuable resources for research, even if some images are no longer accessible. To ensure data integrity, we will include the MD5 checksum of each image to the dataset to ensure that the downloaded images are identical to the ones in our proposed dataset.\\n\\n1. Schuhmann, Christoph, et al. \\\"Laion-5b: An open large-scale dataset for training next generation image-text models.\\\" Advances in Neural Information Processing Systems 35 (2022): 25278-25294.\\n2. Kuznetsova, Alina, et al. \\\"The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale.\\\" International journal of computer vision 128.7 (2020): 1956-1981.\"}", "{\"title\": \"Response to reviewer's comments\", \"comment\": \"> The paper demonstrates the performance of the IndoorGeoCLIP model on various levels of geolocation tasks (such as street-level, city-level, and country-level), but it lacks comparative experiments with other classic indoor geolocation methods, making it difficult to comprehensively validate the superiority of IndoorGeoCLIP. Explicitly, i) Lack of Baseline Model Comparisons: Apart from GeoCLIP, the experiments lack comparisons with other geolocation models, making it insufficient to illustrate the relative advantages of IndoorGeoCLIP in indoor scenes. ii ) Insufficient Ablation Studies: The experiments only show the performance changes of the GeoCLIP model before and after fine-tuning, without conducting ablation analyses on the contributions of the dataset's multimodal features (such as textual and visual data) to geolocation.\\n\\nWe focused on GeoCLIP as the primary baseline for comparison because it represents the state-of-the-art for hybrid geolocation tasks, encompassing both indoor and outdoor environments. GeoCLIP\\u2019s robustness and strong performance in mixed-environment geolocation make it a natural choice to establish a benchmark for our dataset. By fine-tuning GeoCLIP on INDOOR-3.6M, we were able to adapt a leading model to the unique challenges of indoor geolocation, demonstrating the effectiveness of our dataset while maintaining alignment with state-of-the-art methodologies.\\nOur emphasis on the image as the primary input for geolocation tasks reflects real-world constraints where visual content is often the only resource available. In practical applications, such as forensic investigations, textual metadata is rarely accessible, making image-based models crucial.\\nThe inclusion of textual metadata alongside visual data expand the utility of INDOOR-3.6M beyond geolocation tasks. The rich textual annotations make the dataset a valuable resource for other computer vision challenges, such as scene understanding, multimodal representation learning, and image-to-text generation. \\nWe acknowledge that some images in INDOOR-3.6M may lack sufficient clarity, stemming from variations in resolution, lighting, and photographic quality. This was a deliberate choice to reflect the diverse types of imagery encountered in real-life geolocation applications, where input images often vary widely in quality.\\n\\n> Some of the images lack sufficient clarity.\\n\\nIncluding such images ensures that the dataset captures the realistic challenges faced by geolocation systems in practical use cases, such as forensic investigations, surveillance, and emergency response. These scenarios frequently involve suboptimal images, such as those taken with low-resolution devices or in poor lighting conditions. Training models on a dataset that includes such variability helps enhance their robustness and generalizability, enabling them to perform reliably regardless of the quality of the input image.\"}", "{\"title\": \"Response to Reviewer Comments\", \"comment\": \"> The article lacks citations for many key statements. L41: such as seasonal changes, time of day, weather conditions, and human-induced modifications.\\n\\nThe mentioned factors\\u2014seasonal changes, time of day, weather conditions, and human-induced modifications\\u2014stem from the authors\\u2019 experience and knowledge about the task. Additionally, these factors are mentioned in related work, such as Pramanick et al.[1], which explicitly mentions how seasonal variations and time of day can impact geolocation accuracy. We have included this citation in the revised manuscript to strengthen the statement.\\n\\n>As I said in the \\u201cStrengths\\u201d, indoor geolocation (localization) is a meaningful research topic. However, the author lacks a rigorous literature review. I can name many famous indoor localization datasets without thinking: Baidu Mall(CVPR'17), InLoc(CVPR'18), NL-Indoor(CVPR'21). (Although the task objectives of these datasets are different from yours, I think they should be discussed.)\\n\\nThank you for highlighting the importance of a rigorous literature review and for mentioning datasets like Baidu Mall, InLoc, and NL-Indoor. While these datasets are indeed significant in their respective domains, they focus on specific tasks which differ from the scope of our work. Given these distinctions, we have chosen to focus our review on works and datasets directly relevant to global geolocation, i.e identifying where in the world an indoor image was captured, to ensure a clear and focused narrative in the manuscript. This approach allows us to better contextualize our contributions within the specific domain of geolocation research, rather than discussing datasets with fundamentally different objectives.\\n\\n> The experiments based on the proposed dataset are very limited and do not demonstrate the significance of the dataset.\\nWe have updated the experiments section with of the revised manuscript which we hope can be kindly considered for this review.\\n\\n> The experiments based on the proposed dataset are very limited and do not demonstrate the significance of the dataset.\\n\\nWe selected GeoCLIP as the primary baseline due to its state-of-the-art performance in hybrid geolocation tasks, making it an ideal candidate to adapt to the unique challenges of indoor geolocation. By fine-tuning GeoCLIP on INDOOR-3.6M, we showcased how our dataset enhances the capabilities of a leading model, setting a benchmark for indoor-specific geolocation. Our focus on discussing visual data for geolocation aligns with real-world scenarios where images are often the sole resource, such as in forensic investigations. \\n\\n1. Pramanick, Shraman, et al. \\\"Where in the world is this image? transformer-based geo-localization in the wild.\\\" European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2022.\"}", "{\"summary\": \"This works introduce a new large-scale dataset to train and evaluate the tasks of global locations prediction from images. The train set consist of 3.6M image-text pairs collected from the internet on a global scale. The test set is constructed by sampling the images using the population and the land area of each contry. Finetuning the previous sota, GeoCLIP, on the proposed dataset achieves much better geolocation accuracy on the indoor scenes.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": [\"This paper propose a new geolocation dataset dedicating for indoor scenes. The evaluation results shows that the previous sota can perform even better on the indoor scenes after finetuning on the proposed dataset.\", \"The discussions of many aspects of the geolocation tasks and the challenge of collecting the dataset is comprehensive.\"], \"weaknesses\": [\"The significance is somewhat unclear to me:\", \"Regarding dataset scale, it's unclear from Table1 that if the proposed dataset provide more indoor data than some of the other much larger scale mixed dataset like YFCC100M.\", \"The accuracy improvement for indoor geolocation task is limited to the proposed dataset. Can the fine-tuned models achieve better accuracy on the indoor subset of the other datasets?\", \"What is the accuracy merit by the proposed dataset additional to the existing resource? Say if we train on a combined indoor data from existing datasets (e.g., Im2GPS, YFCC100M, MP-16, Hotels50k), what is the additional accuracy boost by adding the proposed INDOOR-3.6M?\", \"The sampling strategy in Sec.4.1 looks ad-hoc to me:\", \"The sampling weight of each contry is determine by the weighted sum of it's population and land area. How the final weight (population * 0.3 + area * 0.7) determined? As it listed as one of the main contribution, I expect more insights. Perhaps an analysis of the accuracy on different contries by varying the weighting can show some support for the goal of the sampling.\", \"The sampling weight seems to assume that the scene visual diversity of the contries are linearly propotion to the population and land area. Is there any support for this assumption? Perhaps some measure using CLIP distance can provide some insight.\"], \"questions\": \"In Table1, what is the column \\\"Benchmark\\\" actually stand for? Is it saying that if the dataset has a train/test split?\\n\\nWhy the correctness of the city/contry/continent prediction is determined by a distance threshold (Table 2 and 3)?\\n\\nScene labels, segmentations, and object detection results are provided by the dataset. What is the purpose of them? Are they serve as some additional conditions for the geolocation task?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}" ] }

[ "Fiza Husain", "Anjaly Parayil", "Ayush Choure", "Rujia Wang", "Chetan Bansal" ]

[ "Heterogeneous Graphs", "Graph Neural Networks", "Recommendation System" ]

{ "note_id": [ "d4nxnuGGWg", "UZzXXPOJgH", "SRtuxeop5f", "ECnf9q3H2U", "1kiGbwWhip" ], "note_type": [ "official_review", "official_review", "comment", "official_review", "official_review" ], "note_created": [ 1730250338897, 1730731027670, 1732280551513, 1730540549396, 1730639642571 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission13965/Reviewer_LpRQ" ], [ "ICLR.cc/2025/Conference/Submission13965/Reviewer_8mcU" ], [ "ICLR.cc/2025/Conference/Submission13965/Authors" ], [ "ICLR.cc/2025/Conference/Submission13965/Reviewer_S1qQ" ], [ "ICLR.cc/2025/Conference/Submission13965/Reviewer_oxVf" ] ], "structured_content_str": [ "{\"summary\": \"Cloud service providers (e.g., Azure, AWS, GCP) aim to ensure the continuous availability of their cloud services. The cloud service monitors, also known as watchdogs, continuously monitor the status of cloud services, tracking various metrics and logs to detect anomalies. In this work, the authors represent a cloud service network as a heterogeneous monitor entity graph. In the context of this paper, the heterogeneous graph consists of three types of nodes: monitor, metric (such as device usage or latency), and dimension (like an attribute of a device). It also contains three types of edges representing the connections between these node types. The paper utilizes the heterogeneous graph to derive the embedding vectors for each node in the monitor entity graph. These embedding vectors are then used to compute a composite loss function, which combines BCE Loss, Ranking Loss, and Diversity Loss. (However, the paper does not provide detailed information on the inputs for these losses.)\", \"soundness\": \"1\", \"presentation\": \"1\", \"contribution\": \"1\", \"strengths\": [\"The authors proposed the DiRaGNN framework for the dimension recommendation and ranking problem in the context of cloud services.\", \"The authors address the computational challenges of processing large-scale heterogeneous graphs by employing neighborhood sampling and subgraph generation.\", \"The author evaluates the diversity loss and ranking loss through an ablation analysis, comparing three model variants and visualizing how the relevance of top-ranked dimensions changes across configurations using rank stability plots. However, providing additional quantitative metrics for the rank stability plots would further help reviewers understand the differences between the model variants.\"], \"weaknesses\": [\"The paper writing is poor. Almost every part of the paper lacks critical information that would help the reviewer understand the work, making the entire paper filled with ambiguity. Moreover, the authors did not clearly introduce or define what the \\u201cdimension ranking problem\\u201d (or the \\u201centity ranking problem in a cloud setting\\u201d) is. It was only after reading a cited paper, Intelligent Monitoring Framework for Cloud Services: A Data-Driven Approach, that I could finally understand what the authors meant by the \\u201cdimension recommendation and ranking problem.\\u201d\", \"The font size in the figures is too small (Figure 1, Figure 3b, Figure 4, and Figure 5), making it very difficult to read. Additionally, the figures are poorly designed, looking almost like children\\u2019s doodles. (For example, in Figure 1a, the elements in the blocks are not properly aligned, and in Figure 1b, the word \\u201cdimension\\u201d is even tilted rather than being properly displayed in the horizontal direction.)\", \"While the paper introduces the adopted DiRaGNN framework and loss functions, the authors fail to explain the source of the datasets used. Furthermore, the ground truth for the dimension ranking task for monitors is neither defined nor explained. Without these details, it becomes difficult for the reviewer to assess the capability of the proposed DiRaGNN framework, relying solely on the reported metric scores.\", \"The content of subsubsection 3.2.1 (MESSAGE PASSING MECHANISM) actually describes the concept of a Heterogeneous Graph Attention Network (referenced below). However, the authors do not cite any relevant papers to support this approach.\", \"In the framework overview shown in Figure 4, the DiRaGNN framework includes both a classifier and link prediction. However, the paper does not explain how the loss is calculated through these two components, making it difficult to understand the framework's workflow.\", \"-The authors only conduct experiments on a single dataset. Performing experiments on additional datasets would help demonstrate the generalizability of the proposed approach.\", \"The paper\\u2019s presentation could be improved; figures are not vector graphics, and some tables appear poorly formatted.\"], \"questions\": [\"There is a family of graph neural networks called heterogeneous graph neural networks. In addition to the well-known Heterogeneous GAT (HGAT), other models include HGT and HetGNN (referenced below). These methods can learn node representations for heterogeneous graphs and inherently handle multiple edge types. Have you ever tried using these methods as baseline models to compare with DiRaGNN?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"1\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The paper presents \\\"DiRaGNN,\\\" an attention-enhanced entity ranking model designed to address the challenges of sparsity in structural and engagement information for entity ranking problems within graph networks. DiRaGNN leverages a transformer-inspired multi-head attention mechanism to focus on heterogeneous neighbors and their attributes. The model employs a multi-faceted loss function that includes diversity-aware and ranking losses to improve recommendation relevance and reduce popularity bias. Experimental results show that DiRaGNN significantly improves entity ranking accuracy, achieving a 39.7% increase in mean reciprocal rank (MRR) over existing approaches.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. The proposed model is explicitly designed to tackle the issue of sparse interactions, which is a common limitation in entity ranking tasks for practical consideration.\\n2. The paper is generally written with good clarity and thus is easy to follow.\", \"weaknesses\": \"1. The technical contribution of this work appears limited, particularly in the design of the DiRaGNN model, which could be seen as fitting within the general framework of graph transformers. The authors are encouraged to emphasize their unique technical innovations to better distinguish their approach.\\n2. Some illustrations, such as Figure 4, are difficult to interpret. The authors are advised to improve the visual clarity of these figures to enhance readability.\\n3. The process of neighborhood sampling and subgraph generation lacks sufficient detail. For instance, the \\\"carefully designed edge splitting strategy\\\" should be elaborated upon. Additionally, for the multi-hop subgraph sampling method, it is unclear whether it was achieved through random walks or simple node sampling.\\n4. In terms of evaluation, only one medium-sized dataset was used, which may limit the generalizability of the results. Moreover, the paper does not specify the number of repeated experiments or the number of seeds used, which is crucial for demonstrating the robustness of the performance.\", \"questions\": \"Please see the above weaknesses for details.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"withdrawal_confirmation\": \"I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors.\"}", "{\"summary\": \"In this paper, the authors propose an attention-enhanced entity ranking model aimed at improving dimension recommendation and ranking systems. However, the technique employed is not particularly innovative, as there are several existing studies incorporating transformer architectures into heterogeneous graphs. Furthermore, the authors do not provide strong experimental results in comparison to recently proposed methods.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"1\", \"strengths\": \"1. This paper introduces a diverse ranking approach for GNNs by incorporating multi-head attention to capture long-range dependencies within the graph structure.\\n2. Overall, the paper is well-written.\", \"weaknesses\": \"1. The paper\\u2019s key novelty is unclear, given that previous work has already integrated transformer architectures into GNNs.\\n2. The experimental results are limited, as the authors only include SAGEConv and TransformerConv as baselines and use a single heterogeneous graph dataset (without a publicly accessible link).\\n3. The intuition behind the proposed methods, particularly concerning the monitor entity graph, is not clearly explained.\", \"questions\": \"My main concerns are: (i) the distinctions between the proposed method and existing attention-based graph methods, such as [1][2][3]; (ii) the underlying intuition for the proposed approach, especially the specific design choices related to the monitor entity graphs; and (iii) the limited experimental scope, which includes only one private dataset and two baseline methods, making it difficult to assess the proposed method\\u2019s advantages.\\n\\nIn particular, incorporating attention mechanisms into graphs has been extensively studied in prior work [1][2][3]. I strongly recommend that the authors provide a detailed discussion of how their approach differs from previous methods. Additionally, it is crucial to evaluate the method across various graph datasets, as real-world graphs often exhibit diverse structures that can better test the generalization capabilities of the proposed model. Thus, I suggest that the authors compare their method on some publicly available graph datasets.\\n\\n[1] Heterogeneous Graph Attention Network\\n[2] Graph Transformer Networks\\n[3] NodeFormer: A Scalable Graph Structure Learning Transformer for Node Classification\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"details_of_ethics_concerns\": \"NA\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper focuses on the problem of recommending dimensions for monitor creation in the cloud setting, where an attention-enhanced entity ranking model is proposed. The authors illustrate the characteristics of the monitor entity graph, then study a set of loss functions to improve the recommendation quality, and finally empirical results show the improvements over classic baselines (e.g., SAGEConv).\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"1\", \"strengths\": \"This paper is easy to read and well organized.\\n\\nIn section 2 and 3, the authors spend 2.5 pages to formulate the research problem and illustrate the characteristics of the monitor entity graph, which might be helpful for the beginners to interpret the topics of this work.\", \"weaknesses\": \"1. The section of related works is missing. I highly recommend the authors to compare this work with recent advances in the filed of representation learning over heterogeneous graphs and emphasize the technique contributions.\\n\\n2. It seems that DIAGNN is a combination of existing works, and thus the contribution looks incremental. Meanwhile, it is not clear to me which part of the networks is proposed to solve the issue of graph sparsity and why it works.\\n\\n3. The training objective, including CE loss, TOP1-max Ranking loss and diversity loss, is not new. Also, the idea of subgraph sampling with negative training examples is standard in task of node classification and recommender systems. The authors are encouraged to clarify the difference to existing works.\", \"questions\": \"Please check the comments above.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}" ] }

[ "Dongyoung Kim", "Kimin Lee", "Jinwoo Shin", "Jaehyung Kim" ]

[ "large language model", "alignment", "preference" ]

{ "note_id": [ "zo7oUiV6i5", "suhfXaKiDI", "pzkeDZ2vjs", "lEU8y27dR8", "k6CmFMJrpZ", "iHsqZ9lJaq", "c7p2tOYNK6", "bLtR7A7MAy", "bB8t5MfWnU", "Y1LJ8N9PuS", "WY9w2L3VMo", "RrM1XmWurf", "RiBf5j8M1O", "OC8VfzsYfx", "N6tFH0xOTm", "KgjlmYwwBr", "KIDwgXd4hr", "CR652znOR8" ], "note_type": [ "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "meta_review", "official_review", "decision", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment" ], "note_created": [ 1732513745895, 1732030742163, 1732202652578, 1730230547465, 1732040184195, 1734608691976, 1730292818408, 1737524224825, 1732498557719, 1730697744865, 1732031241031, 1732616140295, 1732031075317, 1732717000648, 1732510985199, 1732031463362, 1732498623496, 1732031116165 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission12928/Authors" ], [ "ICLR.cc/2025/Conference/Submission12928/Authors" ], [ "ICLR.cc/2025/Conference/Submission12928/Authors" ], [ "ICLR.cc/2025/Conference/Submission12928/Reviewer_dZTX" ], [ "ICLR.cc/2025/Conference/Submission12928/Reviewer_BPdQ" ], [ "ICLR.cc/2025/Conference/Submission12928/Area_Chair_thUJ" ], [ "ICLR.cc/2025/Conference/Submission12928/Reviewer_NqES" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission12928/Authors" ], [ "ICLR.cc/2025/Conference/Submission12928/Reviewer_BPdQ" ], [ "ICLR.cc/2025/Conference/Submission12928/Authors" ], [ "ICLR.cc/2025/Conference/Submission12928/Reviewer_dZTX" ], [ "ICLR.cc/2025/Conference/Submission12928/Authors" ], [ "ICLR.cc/2025/Conference/Submission12928/Authors" ], [ "ICLR.cc/2025/Conference/Submission12928/Reviewer_NqES" ], [ "ICLR.cc/2025/Conference/Submission12928/Authors" ], [ "ICLR.cc/2025/Conference/Submission12928/Authors" ], [ "ICLR.cc/2025/Conference/Submission12928/Authors" ] ], "structured_content_str": [ "{\"comment\": \"Dear Reviewer NqES,\\n\\nWe are glad to hear that we have addressed most of your concerns. Also, thank you for raising the score!\\n\\nPlease don't hesitate if you have any further questions.\\n\\nBest regards, \\nAuthors\"}", "{\"title\": \"Response to Reviewer dZTX\", \"comment\": \"Dear Reviewer dZTX,\\n\\nWe sincerely appreciate your thoughtful comments. We have carefully considered each of your questions and provide detailed responses below.\\n\\n---\\n\\n**[Q1] How do you choose the initial seed data? Are the data randomly chosen as seeds? Are there differences in human preferences across data? How do you handle such differences?** \\n\\nThe initial seed data is selected through random sampling from the UltraFeedback dataset, which could include diverse human preferences. We do not explicitly address these preference variations, but we observe that SPA remains effective across different random seed selections and seed sizes, as demonstrated in Tables 3 and 4. This robustness stems from SPA\\u2019s use of direct preference judgment, which leverages the LLM\\u2019s intrinsic knowledge about human preferences derived from the overall seed data, rather than relying on specific data instances like in-context learning approaches.\\n\\nOn the other hand, we\\u2019d like to mention that there is a field of research that aims to better capture diverse human preferences in LLM\\u2019s responses [1,2]. Since SPA can easily incorporate the modifications in the reward modeling or training objective, we believe that the proposed SPA can be jointly applicable to these existing works.\\n\\n---\\n\\n**[Q2] In section 4.1, can you elaborate on the sampling process used to generate the responses y1 and y2? In particular, does the model use specific diversity techniques to ensure diversity of reactions, or does it rely purely on randomness in the generation process?** \\n\\nThe responses $y_1$ and $y_2$ are sampled purely based on randomness without applying specific diversity techniques, as described in lines 195-196 of the original draft. However, our framework is flexible and can incorporate various diversity sampling techniques if desired. Given the demonstrated effectiveness of diversity techniques in iterative preference learning (e.g., sampling N>2 responses, then choosing best/worst as the preferred/dispreferred responses [3,4]), we believe that incorporating such methods could further enhance the performance of our framework by enhancing the diversity of the responses.\\n\\n---\\n\\n**[Q3] How does it ensure the reliability of the expanded data?** \\n\\nFirst, the proposed direct preference judgment to label the preference between responses (in Section 4.1) yields more reliable data expansion. This is because it leverages intrinsic reward signals directly derived from the target LLM, which are continuously refined through iterative updates. In contrast, prior methods that expand data rely on preference labels generated via implicit prompting or fixed external reward models, limiting their reliability and adaptability. \\n\\nAdditionally, the proposed self-refinement mechanism in Section 4.2 further strengthens the reliability, by addressing labeling noise within the iterative preference learning framework. This self-refinement step uses a logit interpolation to approximate outputs from a more strongly aligned LLM, enabling effective noise detection. By reducing labeling noise in the expanded preference data, this technique enhances the overall reliability and accuracy of the generated data. \\n\\nWe highlight that the enhanced reliability of the expanded data with these components is demonstrated through our experimental results. For example, our SPA approach significantly outperforms prior methods, with improvements such as a 3.52% increase in length-controlled (LC) win rate on AlpacaEval 2.0. Moreover, SPA achieves superior alignment performance using only 3.3% of the seed preference data, compared to DPO method that uses the full 100% seed data. \\n\\n---\\n\\n[1] Zhou et al., Beyond One-preference-for-all: Multi-objective Direct Preference Optimization., arXiv:2310 \\n[2] Pitis et al., Improving Context-aware Preference Modeling for Language Models., arXiv:2407 \\n[3] Wu et al., Self-play Preference Optimization for Language Model Alignment., arXiv:2405 \\n[4] Rosset et al., Direct Nash Optimization: Teaching Language Models to Self-improve with General Preferences., arXiv:2404\\n\\n---\\n\\nIf you have any further questions/concerns, please do not hesitate to let us know. \\n\\nThank you very much, \\nAuthors\"}", "{\"comment\": \"Dear Reviewer BPdQ,\\n\\nWe are happy to hear that our rebuttal addressed your concerns well. Also, we appreciate your support for our work. If you have any further questions or suggestions, please do not hesitate to let us know.\\n\\nBest regards, \\nAuthors\"}", "{\"summary\": \"This paper proposes SPA, an approach to enhance the alignment performance of large language models by using minimal human-annotated preference data.\\nIntroduce a confidence-based refinement of preference labels to reduce the risk of noise in preference learning with generated data.\\nIt is experimentally verified that a tiny percentage of preference data (3.3%) achieves results comparable to or exceeding those obtained using the entire data and the existing optimal baseline method in the AlpacaEval 2.0 evaluation.\", \"soundness\": \"4\", \"presentation\": \"4\", \"contribution\": \"3\", \"strengths\": \"1. The overall idea is relatively simple, but the method achieves good performance and has great potential with limited human labeling.\\n2. The paper is well-structured and presents a rigorous methodology, with comprehensive experimental validation that supports the claims made about SPA\\u2019s effectiveness.\", \"weaknesses\": \"This is a very solid piece of work. The proposed method is simple yet effective. I don't have any particular concerns or issues with it.\", \"questions\": \"1. How do you choose the initial seed data? Are the data randomly chosen as seeds? Are there differences in human preferences across data? How do you handle such differences?\\n2. In section 4.1, can you elaborate on the sampling process used to generate the responses y1 and y2? In particular, does the model use specific diversity techniques to ensure diversity of reactions, or does it rely purely on randomness in the generation process?\\n3. How does it ensure the reliability of the expanded data?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Dear authors, thank you for your response! The added information about length control and gold labels addresses all of my (minor) readability related concerns with the manuscript; I do not have any additional questions or concerns with this work at this time. Naturally, I will be maintaining my score.\"}", "{\"metareview\": \"This work presents a novel framework, called Spread Preference Annotation with direct preference judgment (SPA), aimed at reducing the high costs associated with collecting large preference datasets for alignment. Overall, all reviewers agreed that this work is novel and important, and they gave very high scores to it. I believe this approach is both simple and effective and makes clearly contribution to LLM Alignment. The achieved results are promising, for example, with only 3.3% of the ground-truth labels this method still achieved superior alignment performance. Based on these, I would like to suggest accept this work.\", \"additional_comments_on_reviewer_discussion\": \"The reviewers pointed out some minor issues, such as data and formatting. In the rebuttal, the authors have addressed the concerns of the reviewers, and all reviewers are satisfied with it. Reviewer BPdQ even raised his/her score to 10 to appreciate the novelty of this work. The reviewers are all positive towards this work.\"}", "{\"summary\": \"This paper introduces an efficient LLM alignment method, namely Spread Preference Annotation, aiming at reducing the demand for human-labeled preference data. By leveraging the inherent preference of current aligned model, this work generates new preference data points and conducts alignment iteratively. To reduce the potential noise caused by distribution shift, this work incorporates a self-refinement mechanism on preference labels, where this approximate a more strongly aligned model to better identify noise through a linearly extrapolated prediction method. Through experiments this paper proves that SPA can achieve better alignment performance with much less data.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. This work inherits the idea of self-rewarding, but leverages the inherent preference of current aligned model in an intuitive way, which sounds novel to me.\\n2. The reduction on data usage seems promising, and the performance is robust.\", \"weaknesses\": \"1. 'De-coupled noise preference detection' is not stated clearly enough in section 4.2. Based on my understanding, $z_{\\\\tilde{\\\\theta}}$ is used to substitute for $z_{\\\\theta}$ in the 'Self-refinement' part, which is also supported in Algorithm 1. If I am correct, I think it would be easier for readers to understand if the final usage of the approximated logits and labels are stated in the main text.\\n2. Lacks some explanatory discussion on why this method can work on such a small subsets and even perform better than DPO with the full dataset (Details in questions part).\\n\\nIf the author can address my concern in weakness/questions and provide some insightful discussion, I am willing to raise my score.\", \"questions\": \"1. What is the model used for the 'LLM-as-judge' method in Table 2 ? Have you tried using ${\\\\pi}_{i-1}$ in this baseline?\\n2. There are lines out of the border in references and appendix.\\n3. (corresponding to weakness 2) Why this method can work with only 3.3% of total data? Does this method elicit the latent human preference knowledge from pre-trained model, or the 3.3% subset is already enough to define the human preference in UltraFeedback dataset? (Note that this is not a fatal question, feel free to provide any discussion or hypothesis).\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Oral)\"}", "{\"title\": \"Gentle reminder: The interactive discussion period will end in less than two days\", \"comment\": \"Dear Reviewer NqES,\\n\\nThank you again for your time and efforts in reviewing our paper.\\n\\nAs the discussion period draws close, we kindly remind you that two days remain for further comments or questions. We would appreciate the opportunity to address any additional concerns you may have before the discussion phase ends.\\n\\nThank you very much.\\n\\nBest regards, \\nAuthors\"}", "{\"summary\": \"This work proposes SPA, a framework to lower the high costs of collecting large preference datasets for alignment. SPA uses an LLM's logits to generate pairwise preference data, without reward model learning or in-context learning, which is then used for preference learning for LLMs. The authors show the practical usefulness of SPA by aligning mistral 7B on 3.3% of Ultrafeedback preference data to achieve strong performance compared to state-of-the-art methods on AlpacaEval 2.0 and MT-Bench\", \"soundness\": \"4\", \"presentation\": \"4\", \"contribution\": \"3\", \"strengths\": \"1. The paper is well-written and very easy to follow throughout. The paper contextualizes itself within the alignment literature well, covering the fundamentals of pairwise preference learning (Bradley-Terry modeling) to direct alignment algorithms like DPO (Section 3)\\n2. SPA is able to use only a small seed preference dataset to then directly score preference labels using the implicit reward model learned by DPO (Section 4.1). Since these predictions can be noisy, the authors introduce a novel self-refinement denoising technique using a confidence prediction (eq 9) to smooth the preference label (eq 10)\\n3. Reproducibility: the authors provide implementation details and hyperparameters in Section 5 (*L311-321*). The authors will open-source the code and models after acceptance, which is appreciated. Lastly, because the modification to the DPO objective is minimal, the authors mention only a few lines of change to the DPO codebase, which is another advantage for practical utility of SPA\\n4. The authors compare to popular categories of baselines: iterative DPO methods, LLM-as-judge methods, and explicit reward-modeling + RLHF methods (*L288 - L291*) and achieve strong results on AlpacaEval 2.0 and MT-Bench\\n5. The authors show SPA extends beyond Mistral to other popular LLMs like Phi and Llama (Table 5) and is robust, in the win rate variance sense, to the seed of the initial preference data (Table 4)\", \"weaknesses\": \"No major weaknesses, mainly minor clarifications:\\n1. Can the authors provide a little more description about the \\\"length control\\\" aspect of AlpacaEval 2.0 in the main paper? This setting is used in nearly all results, but is not explained clearly where first introduced (Section 5.2)\\n2. What is \\\"gold label\\\" (Table 1, 5)? Is this the Ultrafeedback preference data? Please make this explicit in the writeup\", \"questions\": \"Minor Typos\\n1. \\\"Lengh-control\\\" in Figure 2\\n2. \\\"additional codes\\\" , L267\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"10\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to Reviewer BPdQ\", \"comment\": \"Dear Reviewer BPdQ,\\n\\nWe sincerely appreciate your thoughtful comments. We have carefully considered each of your questions and provide detailed responses below.\\n\\n---\\n\\n **[W1] Can the authors provide a little more description about the \\\"length control\\\" aspect of AlpacaEval 2.0 in the main paper? This setting is used in nearly all results, but is not explained clearly where first introduced (Section 5.2)** \\n\\nThank you for the constructive feedback. The length-controlled (LC) win rate in AlpacaEval 2.0 [1] is a newly introduced evaluation metric designed to reduce bias toward longer responses when using LLMs as judges [2,3]. To achieve this, a regression model is trained to separate the contributions of response length and quality, based on data from leaderboard submissions. The LC win rate then estimates win rates by neutralizing the effect of response length, focusing purely on quality. As demonstrated in [1], this metric correlates more closely with human evaluation [4] than the standard win rate. We have added these details about this \\u201clength control\\u201d aspect of AlpacaEval 2.0 in the revised draft (lines 304-305). \\n\\n---\\n\\n**[W2] What is \\\"gold label\\\" (Table 1, 5)? Is this the Ultrafeedback preference data? Please make this explicit in the writeup** \\n\\nThe term \\u201cgold label\\u201d refers to the \\u201cground-truth preference label\\u201d provided by the UltraFeedback dataset. In our experiments, we use the UltraFeedback dataset in two distinct ways. First, a portion serves as the seed preference data, $\\\\mathcal{D} = \\\\{(x, y_l, y_w)\\\\}$, directly using its ground-truth labels. Second, another portion of the dataset is used solely as a prompt set, $X_i = \\\\{x\\\\}$, by discarding the original labels $y_l$ and $y_w$ to allow to use it for new data generation at each iteration; thus, this portion does not utilize \\u201cgold labels.\\u201d Therefore, in Tables 1 and 5, we denote \\u201cgold label\\u201d to clarify the amount of ground-truth preference data utilized, ensuring a fair comparison between methods that rely on different quantities of labeled data. We explicitly mention this in the revised draft (lines 294-295). \\n\\n---\\n\\n**[Q1] Typos.** \\n\\nThank you for the careful reading and pointing out the typo! We have corrected this in the revised draft. \\n\\n---\\n\\n[1] Dubois et al., Length-Controlled AlpacaEval: A Simple Way to Debias Automatic Evaluators., arXiv:24.04 \\n[2] Wang et al., How Far Can Camels Go? Exploring the State of Instruction Tuning on Open Resources., NeurIPS 2023 Datasets and Benchmarks Track \\n[3] Zheng et al., Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena., NeurIPS 2023 Datasets and Benchmarks Track \\n[4] Chiang et al., Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference., arXiv:24.03 \\n\\n---\\n\\nIf you have any further questions/concerns, please do not hesitate to let us know. \\n\\nThank you very much, \\nAuthors\"}", "{\"comment\": \"Dear authors, thank you for your response! I do not have any additional questions or concerns with this work at this time. Naturally, I will be maintaining my score.\"}", "{\"title\": \"Response to Reviewer NqES (1/2)\", \"comment\": \"Dear Reviewer NqES,\\n\\nWe sincerely appreciate your thoughtful comments. We have carefully considered each of your questions and provide detailed responses below.\\n\\n---\\n\\n**[W1] 'De-coupled noise preference detection' is not stated clearly enough in section 4.2. Based on my understanding, z\\u03b8~ is used to substitute for z\\u03b8 in the 'Self-refinement' part, which is also supported in Algorithm 1. If I am correct, it would be easier for readers to understand if the final usage of the approximated logits and labels are stated in the main text.** \\n\\nYou\\u2019re correct; $z_{\\\\widetilde{\\\\theta}}$ (Eq. 11) is used to substitute for $z_{\\\\theta}$ in the self-refinement step (Eq. 10). Following your suggestion, we have explicitly stated this in Section 4.2 (lines 256-257 in the revised draft) for clarity. \\n\\n---\\n\\n**[W2, Q3] Lacks some explanatory discussion on why this method can work on such a small subsets and even perform better than DPO with the full dataset. Does this method elicit the latent human preference knowledge from pre-trained model, or the 3.3% subset is already enough to define the human preference in UltraFeedback dataset?**\\n \\nThank you for the insightful question. As noted in your first conjecture, SPA is indeed able to elicit latent preference knowledge from the pre-trained LLM, allowing it to perform effectively even with a limited number (e.g., 3.3%) of the labeled preference data. Specifically, this effectiveness is achieved through two main techniques in SPA: \\n- Direct Preference Judgment (Section 4.1) enables efficient and reliable data expansion, by leveraging intrinsic reward signals directly derived from the target LLM, which are continuously refined through iterative updates.\\n- Self-Refinement Mechanism (Section 4.2) further enhances the reliability of expanded data, by addressing labeling noise through the iterative preference learning. By reducing labeling noise in the expanded preference data, it enhances the overall reliability and accuracy of the generated data.\", \"we_also_highlight_that_our_experiments_support_this_claim_well\": [\"SPA without seed data (Figure 4 and Table 8): it is assumed that there is no seed preference data (i.e., 0%), and the instruction-tuned LLM is directly used to generate preference data without initial DPO step (1st line in Algorithm 1). Here, the proposed SPA continuously improves the alignment performance, which supports the effectiveness from the elicitation\", \"SPA with varying number of seed data (Table 3): SPA is applied with varying number of seed preference data (0.8% to 10%), and SPA is consistently effective and the improvement grows with increased seed data. It opposes the second part of the conjecture as it implies that the 3.3% subset does not fully capture the UltraFeedback dataset\\u2019s human preference knowledge. Remarkably, SPA yields better alignment performance than DPO with full data, if the given seed data is sufficient (e.g., >= 1.7%) to provide the effective guidance to elicit LLM\\u2019s intrinsic knowledge about human preference.\", \"Overall, these results indicate that the effectiveness of SPA is from eliciting LLM\\u2019s intrinsic knowledge about human preference rather than learning the given seed preference data well.\"]}", "{\"comment\": \"Dear Reviewer dZTX,\\n\\nWe are happy to hear that our rebuttal addressed your questions well. Please let us know if you have any further questions.\\n\\nThank you very much.\\n\\nBest regards, \\nAuthors\"}", "{\"title\": \"Response to author replies.\", \"comment\": \"Thanks for your replies which addresses most of my concerns. As a result, I raise my score to 8 as stated in my review.\"}", "{\"title\": \"General Response\", \"comment\": \"Dear reviewers and AC,\\n\\nWe sincerely appreciate your valuable time and effort spent reviewing our manuscript.\\n\\nAs reviewers highlighted, we propose a simple (dZTX), yet novel (NqES, BPdQ) method for LLM alignment, that shows strong empirical results (all reviewers) on the comprehensive experiments (dZTX) with clear writing (dZTX, BPdQ). \\n\\nWe appreciate your constructive feedback on our manuscript. In response to the comments, we have carefully revised and enhanced the manuscript as follows:\\n\\n- Detailed explanation for easier understanding of refinement (Section 4.2)\\n- Definition of gold label in Tables 1 and 5, and details about length-controlled win rate (Section 5.1)\\n- Removing lines out of the border (References and Appendix) \\n- New experiments and corresponding discussions regarding LLM-as-judge using previous iteration\\u2019s model (Appendix D and Tabel 11)\\n- Resolving typos (Figure 2 and line 267)\\n\\nIn the revised manuscript, these updates are temporarily highlighted in $\\\\text{\\\\color{blue}blue}$ for your convenience to check.\\n\\nWe sincerely believe that these updates may help us better deliver the benefits of the proposed SPA to the ICLR community.\\n\\nThank you very much, \\nAuthors.\"}", "{\"title\": \"Gentle reminder: The interactive discussion period will end in less than two days\", \"comment\": \"Dear Reviewer dZTX,\\n\\nThank you again for your time and efforts in reviewing our paper.\\n\\nAs the discussion period draws close, we kindly remind you that two days remain for further comments or questions. We would appreciate the opportunity to address any additional concerns you may have before the discussion phase ends.\\n\\nThank you very much.\\n\\nBest regards, \\nAuthors\"}", "{\"title\": \"Response to Reviewer NqES (2/2)\", \"comment\": \"**[Q1] What is the model used for the 'LLM-as-judge' method in Table 2 ? Have you tried using \\u03c0i\\u22121 in this baseline?**\\n\\nFor the \\u2018LLM-as-Judge\\u2019 method in Table 2, we used a fixed model fine-tuned specifically for evaluating preferences between responses. Specifically, the training dataset for this model was constructed by converting seed preference data into pairwise comparison prompts, following approaches similar to [1,2]. Then, the common SFT model (used for initialization in other baselines like DPO and SPA) was further fine-tuned on this dataset via supervised learning. More details, such as the training process and the pairwise comparison prompts, are included in Appendix B.2 of the original draft. \\n \\nRegarding the use of $\\\\pi_{i-1}$ in this baseline, we conducted new experiments; at the 2nd iteration, the evaluation model was initialized with the resulting model from the 1st iteration and fine-tuned with the same preference evaluation dataset. The evaluation results, as denoted by LLM-as-Judge (Iter. 2, prev. init), on AlpacaEval 2.0 are presented below along with other methods at the 2nd iteration. While this approach yielded somewhat improved alignment compared to the fixed model, SPA still significantly outperformed this baseline. This underscores that SPA\\u2019s superior performance arises from its novel preference evaluation techniques rather than the specific evaluation model used. We have added these results and the corresponding discussion to Appendix D of the revised draft.\\n\\n\\\\begin{array}{c|ccc}\\n\\\\hline\\n\\\\text{AlpacaEval 2.0} & \\\\text{LC Win Rate (\\\\\\\\%)} &\\\\text{Win Rate (\\\\\\\\%)} \\n\\\\newline \\\\hline \\n\\\\text{LLM-as-judge (Iter. 1)} & 8.88 & 8.01 \\\\newline \\\\hline\\n\\\\text{LLM-as-judge (Iter. 2, orig)} & 9.49 & 8.46 \\\\newline \\n\\\\text{LLM-as-judge (Iter. 2, prev. init)} & 9.74 & 10.09 \\\\newline \\\\hline\\n\\\\text{SPA (Iter. 2, ours)} & \\\\textbf{15.46} & \\\\textbf{19.91} \\\\newline \\\\hline\\n\\\\end{array}\\n\\n---\\n\\n**[Q2] There are lines out of the border in references and appendix.** \\n\\nThank you for the careful reading and pointing out the editorial problems. We have corrected these in the revised draft (p.13 and p.15). \\n\\n---\\n\\n[1] Bai et al., Constitutional AI: Harmlessness from AI Feedback., Anthropic 2022 \\n[2] Lee et al., Aligning Large Language Models by On-Policy Self-Judgment., ACL 2024 \\n\\n---\\n\\nIf you have any further questions/concerns, please do not hesitate to let us know. \\n\\nThank you very much, \\nAuthors\"}" ] }

[ "FNU Hairi", "Yang Jiao", "Tianchen Zhou", "Haibo Yang", "Chaosheng Dong", "Fan Yang", "Michinari Momma", "Yan Gao", "Jia Liu" ]

[ "Multi-Objective Reinforcement Learning", "Actor-Critic Algorithm" ]

{ "note_id": [ "vHns00SmxB", "sBIjJptNbp", "rgeEuUOwmh", "reNzCx11OX", "qjtC7L1NnS", "pdTtz1sbJT", "ljasxr2isF", "lazASTASQH", "j7Bi9uEbxc", "gl89P8btrv", "fgod9xO2jm", "eMKow7TZfY", "drX8NPuIok", "cCYwEJ4dis", "Z7qpdYBBfD", "YsI1klJ6kS", "Wn3mMoJefw", "Vki9IVtRkJ", "U0Zuz2rbNE", "TzeHufx6LE", "S0mNoKntbg", "R618QiO1KD", "PHQNLLILDR", "JOdTnKYeUH", "H9E0UA3hdO", "GOoJoOLzBJ", "FO860D1jVt", "EVNwrQyVQL", "EDRIL5NrS9", "ED0U5B4AMj", "BWfSQJFoFy", "9wkbDro8sT", "59c7cAQta0" ], "note_type": [ "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "meta_review", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment" ], "note_created": [ 1732591842216, 1732149062836, 1732591143197, 1733222381678, 1732237868833, 1734782761578, 1733222168590, 1732242335532, 1732238675545, 1730931917048, 1732217816519, 1733222460152, 1732149951877, 1731219627865, 1733222515170, 1732261485982, 1732148925236, 1730696995518, 1733197275632, 1732674625802, 1732238610357, 1731007461632, 1732686536181, 1732237809692, 1732148724906, 1737524209018, 1732261552273, 1732217732008, 1732217215290, 1732591730344, 1732149140775, 1730543336906, 1732149850731 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission12698/Authors" ], [ "ICLR.cc/2025/Conference/Submission12698/Authors" ], [ "ICLR.cc/2025/Conference/Submission12698/Authors" ], [ "ICLR.cc/2025/Conference/Submission12698/Authors" ], [ "ICLR.cc/2025/Conference/Submission12698/Authors" ], [ "ICLR.cc/2025/Conference/Submission12698/Area_Chair_QjUf" ], [ "ICLR.cc/2025/Conference/Submission12698/Authors" ], [ "ICLR.cc/2025/Conference/Submission12698/Reviewer_JPv7" ], [ "ICLR.cc/2025/Conference/Submission12698/Authors" ], [ "ICLR.cc/2025/Conference/Submission12698/Reviewer_P2zK" ], [ "ICLR.cc/2025/Conference/Submission12698/Authors" ], [ "ICLR.cc/2025/Conference/Submission12698/Authors" ], [ "ICLR.cc/2025/Conference/Submission12698/Authors" ], [ "ICLR.cc/2025/Conference/Submission12698/Reviewer_P7fw" ], [ "ICLR.cc/2025/Conference/Submission12698/Authors" ], [ "ICLR.cc/2025/Conference/Submission12698/Authors" ], [ "ICLR.cc/2025/Conference/Submission12698/Authors" ], [ "ICLR.cc/2025/Conference/Submission12698/Reviewer_JPv7" ], [ "ICLR.cc/2025/Conference/Submission12698/Reviewer_P7fw" ], [ "ICLR.cc/2025/Conference/Submission12698/Reviewer_P2zK" ], [ "ICLR.cc/2025/Conference/Submission12698/Authors" ], [ "ICLR.cc/2025/Conference/Submission12698/Reviewer_n72g" ], [ "ICLR.cc/2025/Conference/Submission12698/Authors" ], [ "ICLR.cc/2025/Conference/Submission12698/Authors" ], [ "ICLR.cc/2025/Conference/Submission12698/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission12698/Authors" ], [ "ICLR.cc/2025/Conference/Submission12698/Authors" ], [ "ICLR.cc/2025/Conference/Submission12698/Authors" ], [ "ICLR.cc/2025/Conference/Submission12698/Authors" ], [ "ICLR.cc/2025/Conference/Submission12698/Authors" ], [ "ICLR.cc/2025/Conference/Submission12698/Reviewer_qYKr" ], [ "ICLR.cc/2025/Conference/Submission12698/Authors" ] ], "structured_content_str": [ "{\"title\": \"Response for Reviewer JPv7's Follow-up Comments (Continued)\", \"comment\": \">**Follow-up Comment 2:** As your reply to the first comment stated, this work \\\"focuses on systematically exploring the Pareto Stationary front using a weighted-Chebyshev formulation.\\\" The introduction of the preference weighting is the main contribution.\\nHowever, your experiment design does not show the advantage of including such weighting. For the experiment in Table 1, the effect of emphasizing \\\"Comment\\\" or \\\"Dislike\\\" objectives can be more persuasive. \\n\\n**Our Response:** Thanks for your comment and we would like to further clarify on this. The rationale behind our weight choice $[0.2,0.2,0.2,0,0.4]^{\\\\top}$ in Table 1 is inspired by the Kuaishou dataset for short video streaming, which typically emphasize on \\\"WatchTime\\\" with higher weight (the most important performance metric in short video streaming), and on the other hand ignoring \\\"Dislike\\\" (not important for short video streaming in practice). Essentially, this yields a 4-objective MORL problem. For the remaining objectives, we chose equal weight 0.2 for each of them to highlight no particular preference within this subset of objectives. \\n\\nIn this rebuttal period, to further illustrate our WC-MOAC framework's effectiveness in exploring the Pareto stantionarity front, we have conducted additional experiments with more variations of the $\\\\mathbf{p}$ vector, some of which have placed more emphasis on \\\"Comment\\\" and \\\"Dislike\\\" as you suggested. Please see the new $\\\\mathbf{p}$ vector setting in Table 3 and the associated new plots in Fig. 2 in the revised manuscript. Here, the settings include larger weights for \\\"Comment\\\" and \\\"Dislike\\\" than those in Table 1 following your suggestion, and we thank the reviewer's comment that helps enrich our experiments.\\n\\n______________\\n\\n>**Follow-up Comment 3:** As replied to Comments 7 and 8, the preference vector is not trained. What is the systematic way of deciding the weighting in your approach? As TSCAC, does your method also require domain knowledge? \\n\\n**Our Response:** Thanks for your comments and questions. Please see our point-to-point responses below: \\n\\n1. **Systematic Way of Weighting:** To systematically explore the Pareto stationary solutions, the decision-maker can simply enumerate and choose the vector $\\\\mathbf{p}$ in the $M$-dimensional standard simplex (i.e., the set $\\\\{ p_i\\\\geq 0, i=1,\\\\ldots,M | \\\\sum_{i=1}^{M} p_i =1\\\\}$). For example, one possible approach will be using a grid search on the $M$-dimensional standard simplex for trying different $\\\\mathbf{p}$ vectors. The chosen $\\\\mathbf{p}$ vector is then taken by Algorithm 1 as input. \\n\\n2. **Not Requiring Domain Knowledge:** Unlike TSCAC, our approach does **not** require domain knowledge, which is actually a *salient feature* of our proposed method. The reason is that our goal is to explore the Pareto stationarity front with as many $\\\\mathbf{p}$-vector choices as possible. This implies that we are at the *\\\"exact oppositie\\\"* of getting domain knolwedge to determine $\\\\mathbf{p}$.\\n\\nAdditionally, as we mentioned above, we have added new empirical results in this rebuttal period with more weight vectors $\\\\mathbf{p}$ in Appendix A.2 Section. In Figure 1\\\\(c) and Figure 2(b), it can be seen that WC-MOAC can explore Pareto (stationary) solution in that has smaller dislike values (or equivalently, higher negative dislike values) than SDMGrad. Also from Figure 1\\\\(c) and 2(b), in terms of \\\"Comment\\\" objective, WC-MOAC explore Pareto (stationary) solution that is no worse than SDMGrad.\\n______________\\n\\n>**Follow-up Comment4:** Meanwhile, for the experiment shown in Figure 1, concentrating on click, dislike, and watch do not differ much.\\n\\n**Our Response:** Thanks for your comment. However, we are not exactly sure if we fully understand this comment. If our guess is correct, we suspect there may be some misunderstanding in reading Figure 1(a), where the results of \\\"Click,\\\" \\\"Dislike,\\\" and \\\"WatchTime\\\" are quite similar under different $\\\\mathbf{p}$ vectors. We would like to clarify Figure 1(a) in here. First of all, Figure 1(a) is **not** showing the performance of our proposed WC-MOAC method. Rather, Figure 1(a) illustrates the performance of SDMGrad over varying weight vector $\\\\mathbf{p}$. We note that SDMGrad is a **baseline method** for comparison, which is based on linear scalarization. The fact that the SDMGrad method concentrates (i.e., having a small footprint) around \\\"WatchTime\\\", \\\"Like\\\" and some concentrations on \\\"Click\\\" in Figure 1(a) shows exactly it is **not** suitable for Pareto stationarity front exploration. \\n\\nOn the other hand, **our proposed WC-MOAC method** is shown in Figure 1(b), which shows that WC-MOAC explores a much **larger** Pareto sationary solution footprint without much concentration in all objectives. For ease of comparison, in Figure 1\\\\(c), we compared the explored footprints from these two methods to compare the Pareto front explorations. Again, it confirms that WC-MOAC explores significantly more Pareto solutions.\\n______________\"}", "{\"title\": \"Response to Reviewer P7fw (Continued)\", \"comment\": \">**Your Comment 6:** In Section 3, the problem formulation about MOMDP (Lines 181-187) appears very similar to the second paragraph of Section 3.1 of (Zhou et al., 2024).\\n\\n**Our Response:** MDP is the **standard language** for defining RL problems, and MOMDP is a natural extension of MDP in the multi-objective setting. Similar to our response to your Comments 3--4, **no** paper owns an exclusive right to the MOMDP language for defining MORL problems.\\nAs a result, it shouldn't be a surprise to see similar, if not the entirely same, MOMDP language in all papers that study infinite horizon RL/MORL problems with ergodic property. To see the commonality of these MOMDP formulations, see the beginning of Section 3 of paper (Qiu et al (2021)), Section 2.1 of (Roijers et al (2018)), Section 2.1 of paper (Zhang et al (2018)) and Section 3.1 of (Hairi et al (2022)), just to name a few.\\n\\nQiu, S., Yang, Z., Ye, J., and Wang, Z. On finite-time convergence of actor-critic algorithm. IEEE Journal on Selected Areas in Information Theory, 2(2):652\\u2013664, 2021.\\n\\nRoijers, D. M., Steckelmacher, D., and Nowe \\u0301, A. Multi-objective reinforcement learning for the expected utility of the return. In Proceedings of the Adaptive and Learn- ing Agents workshop at FAIM, volume 2018, 2018.\\n\\nHairi, F., Liu, J., and Lu, S. Finite-time convergence and sample complexity of multi-agent actor-critic reinforcement learning with average reward. In International Con- ference on Learning Representations, 2022.\\n\\nZhang, K., Yang, Z., Liu, H., Zhang, T., and Basar, T. Fully decentralized multi-agent reinforcement learning with networked agents. In International Conference on Ma- chine Learning, pp. 5872\\u20135881. PMLR, 2018.\\n\\n________________\\n>**Your Comment 7:** The part on \\u201cLearning Goal and Optimality in MORL\\u201d (Lines 200-215) appears almost the same as the \\u201cProblem Statement\\u201d in Section 3.1 of (Zhou et al., 2024). Moreover, even the footnote #3 in this paper almost goes verbatim compared to the footnote #1 in (Zhou et al., 2024).\\n\\n**Our Response:** Both our paper and (Zhou et al. (2024)) study infinite-horizon multi-objective reinforcement learning (MORL) problem either under accumulated discounted reward and average reward settings. We see similarities. However, the wordings are significantly disctint. More importantly, the solution approaches and goals in these two works are completely different (Pareto stationarity front exploration vs. identifying only a Pareto stationary point). Also, the footnotes in Section 3.1 in (Zhou et al., 2024) and the footnote #3 in this paper are common language that appear in many papers in the literature. Again, **no** paper owns an exclusive right to these common sentences.\\n\\nMoreover, we would like to protest the claim that just because studying the same MORL setting would consitute the so-called \\\"plagiarism or dual submission.\\\" In academic research, a long line of works could all be dedicated to stuying a challenging problem setting or even a common problem (e.g., trying to solve or just make progress in proving/disproving a famous conjecture).\\n\\n________________\\n\\n>**Your Comment 8:** The preliminaries about the policy gradient for MORAL (Lines 274-296) also largely resembles Section 3.2. Specifically, several sentences about Lemma 2 and Assumption 2 of this paper are exactly the same as those in Lemma 1 and Assumption 2 of (Zhou et al., 2024).\\n\\n**Our Response:** First of all, in our paper, we don't have a MORAL concept and we assume you are referring to MORL (multi-objective reinforcement learning). Second, it is simply **not** true that sentences in Lemma 2 of our paper are *\\\"exactly the same\\\"* as those in Lemma 1 of (Zhou et al. (2024)) and Assumption 2 with Assumption 2 in (Zhou et al. (2024)) for that matter.\\n\\nMore importantly, Lemma 2 and Assumption 2 (correspondingly Lemma 1 and Assumption 2 in (Zhou et al., 2024)) are standard and basic concepts in the context of actor-critic framework and the policy gradient theorem for RL, which can be found in many classic papers in RL (e.g., (Sutton et al. (1999)). Moreover, we have also properly cited the sources in the subsequent remark after Lemma 2. Also, Assumption 2 is also a collection of standard assumptions in critic analysis with linear approximation in the literature, for which we have provided references for it. We never claimed that Lemma 1 and Assumption 2 are our own contributions.\\n\\nRichard S Sutton, David A McAllester, Satinder P Singh, Yishay Mansour, et al. Policy gradient methods for reinforcement learning with function approximation. In NIPs, volume 99, pp. 1057\\u2013 1063. Citeseer, 1999.\\n\\n________________\"}", "{\"title\": \"Our Follow-up Response to Comment 6\", \"comment\": \">**Your Comment 6 (Also Question 4):** The paper also raises questions regarding the sensitivity of the algorithm's performance to variations in $p$ and whether there are recommended heuristics for selecting optimal values.\\n\\n**Our Follow-up Response to Comment 6(Also Question 4):** Thanks for your patience. We have finished additional experiments with more variations of $\\\\mathbf{p}$. We have attached radar charts that include new results in Appendix A.2 Section. Specifically, in addtion to the 5 one-hot vectors, we have included $\\\\mathbf{p}$ vectors that takes the following values:\\n| click | like | comment | dislike | watchtime |\\n| -------- | ------- |------- |------- |------- |\\n| 0.85 | 0.05 | 0.05 | 0 | 0.05 |\\n| 0.7 | 0.1 | 0.1 | 0 | 0.1 |\\n| 0.55 | 0.15 | 0.15 | 0 | 0.15 |\\n| 0.4 | 0.2 | 0.2 | 0 | 0.2 |\\n|0.05 | 0.05 | 0.85 | 0.0001 | 0.05 |\\n|0.10 | 0.10 | 0.70 | 0.0001 | 0.10 |\\n|0.15 | 0.15 | 0.55 | 0.0001 | 0.15 |\\n\\nFrom the empirical results in Figure 2(a) in Appendix A.2 Section, we can see that with additional weight vectors $\\\\mathbf{p}$, WC-MOAC is exploring **more Pareto stationary solutions** compared to WC-MOAC with only one-hot vectors as the weight vectors. In Figure 2(b), it further shows that with more $\\\\mathbf{p}$ vectors, WC-MOAC explores even wider Pareto footprints. This further confirms our theoretical prediction as well as strengthens the empirical observation that, with increasing number of weight/explore vectors $\\\\mathbf{p}$, WC-MOAC possess the potential to explore more Pareto stationary points.\\n\\nWith the above new experimental results on more variations of $\\\\mathbf{p}$, we would like to add an additional remark that the terminology \\\"sensitivity\\\" in the reviewer's Comment 6 is somewhat irrelavent to the above experiments with additional $\\\\mathbf{p}$ varations. The reason is that sensitivity is typically referring to the cases with a target choice of $\\\\mathbf{p}$, but there could be imperfection/inaccuracy/errors in picking such a $\\\\mathbf{p}$ in practice. Thus, one would like to study how \\\"sensitive\\\" the system will be affected by the deviation of the real choice of the $\\\\mathbf{p}$ from the target choice of $\\\\mathbf{p}$. However, in this work, we do *not* have any target choice of $\\\\mathbf{p}$. Instead, the goal of our WC-MOAC method is to explore Pareto Stationary front *by applying as many different $\\\\mathbf{p}$ vectors in the algorithm as possible*. Hence, our new results above are showing \\\"exploration capabiltiy\\\" rather than \\\"sensitivity.\\\" \\n\\n\\nLastly, regarding how to choose $\\\\mathbf{p}$ to systematically explore the Pareto stationary solutions, the decision-maker can simply enumerate and choose the vector $\\\\mathbf{p}$ in the $M$-dimensional standard simplex (i.e., the set $\\\\{ p_i\\\\geq 0, i=1,\\\\ldots,M | \\\\sum_{i=1}^{M} p_i =1\\\\}$). For example, one possible approach will be using a grid search on the $M$-dimensional standard simplex for trying different $\\\\mathbf{p}$ vectors. The chosen $\\\\mathbf{p}$ vector is then taken by Algorithm 1 as input.\\n\\nWe hope that our clarifications above have addressed the reviewer's question on the selection of $\\\\mathbf{p}$. Please let us know if any question remains.\"}", "{\"title\": \"Response to Reviewer P7Fw's Follow-up Comments (Continued):\", \"comment\": \">**Your Follow-up Comment 6:**\\n>> Lastly, our empirical results strongly suggest the nature of exploration Pareto stationarity points by looking at Figure 1, which does not have any counterpart in (Zhou et al. (2024)). The key takeaways from this paper is very different from that of (Zhou et al. (2024)).\\n>\\n> I can understand that the empirical results are shown to corroborate the nature of exploration of Pareto stationary points. That being said, if one directly compares the results of MC-MOAC and those of the MOAC in [Zhou et al., 2024] in terms of Like, Comment, Dislike, and WatchTime, the differences appear rather not significant. It is not immediately clear to me why the empirical results (cf. Table 1 and Figure 1) clearly supports the claim. While I understand in general that the works published within 3 months of submission are not necessarily needed in the comparison, a comparison with MOAC can serve as a really helpful ablation study and hence nicely support the claim of the paper. Please correct me if I missed anything.\\n\\n**Our Response:** Thank you for your question. In the following table, we provide a direct comparison of values of WC-MOAC in Our Table 1 and that of MOAC in (Zhou et al.(2024)) over the aforementioned 4 objectives. The 3rd-5th row of values are based on improvements over Behavior-Clone(BC) in Table 1 of our paper. We also remark that for dislike objective, the more decrease indicates the better performance.\\n| Objective | Like($10^{-2}$) | Comment($10^{-3}$) | Dislike($10^{-4}$) | WatchTime |\\n| -------- | ------- |------- |------- |------- |\\n| MOAC | 1.312 | 3.266 | 1.486 | 1.307 |\\n| WC-MOAC | 1.329 | 3.092 | 1.339 | 1.375 |\\n| MOAC over BC | 6.57% | 1.27% | -35.5% | 1.71% |\\n| WC-MOAC over BC | 7.96% | -4.12% | -41.88% | 7.00% |\\n| WC-MOAC over MOAC | 1.39% | -5.39% | 6.38% | 5.29% |\\n\\nWe can see that among the four objectives \\\"Like,\\\" \\\"Comment,\\\" \\\"Dislike,\\\" and\\n\\\"WatchTime,\\\" 3 of them (\\\"Like,\\\" \\\"Dislike,\\\" and \\\"WatchTime\\\") favors WC-MOAC over MOAC and only \\\"Comment\\\" favors MOAC over WC-MOAC. More specifically, WC-MOAC improved over MOAC by 5.29% in watchtime, 6.38% in dislike and 1.39% in likes. WC-MOAC is worse than MOAC by 5.39% in the \\\"Comment\\\" objective. Overall, it shows WC-MOAC with the particular choice of $\\\\mathbf{p}$ does find a Pareto solution that possesses 3:1 favorable objectives than MOAC.\\n\\nWe would also like to point out that, sometimes, 1% increase doesn\\u2019t necessarily mean \\\"not significant\\\". On the contrary, it often has a significant implication in many real-world systems. For example, according to [1], \\\"*the revenue from Kuaishou streaming business increased by 10.4% to RMB39.1 billion in 2023, from RMB35.4 billion in 2022, benefiting from consistent enrichment of content supply and continuous optimization of our live streaming ecosystem and algorithms*\\\" [1]\\\". With a 5.29% improvement on watchtime objective, it is a potentially significant improvement considering the population of user base and total revenue. \\n\\n[1] https://ir.kuaishou.com/news-releases/news-release-details/kuaishou-technology-announces-fourth-quarter-and-full-year-2023/\\n\\nComing back to the main point of our paper, with the help of weight-vector $\\\\mathbf{p}$, by exploring Pareto Stationary front, WC-MOAC enables finding a more preferable solution than MOAC, which merely guarantees finding a Pareto stationary point.\\n_________________\"}", "{\"title\": \"Response to Reviewer n72g's Comments (Continued)\", \"comment\": \">**Your Comment 7:** Page 9 Theorem 4. Some of the quantities, such as $\\\\lambda_A$, $L_J$, $R_{w}$, $\\\\gamma$ are from the appendix, it is better to add some explanation for these quantities in the statement of the theorem.\\n\\n**Our Response:** Thank you for the suggestion. $\\\\lambda_A$ is defined in Line 771 and this quantity is introduced to ensure the convergence for the critic component for all policies. $R_w$ is defined in Lemma Line 782 for the average reward setting and in Line 805 for the discounted setting, which is an upper bound constant for $\\\\ell_2$ norm for the critic parameters. $L_J$ is the Lipschitz constant defined in Line 425 in Assumption 3. $\\\\mathbf{\\\\gamma}=(\\\\gamma^1,\\\\cdots,\\\\gamma^M)^{\\\\top}$ is a vector that concatenates all discount factors $\\\\gamma^1$ to $\\\\gamma^M$ for all objectives in discounted setting. We will provide these clarifications in our revision.\\n\\n________________\\n>**Your Comment 8:** Page 9 line 467. You mention that state-of-the-art sample complexity for single-objective RL is bu Xu et al. And your Corollary 5 achieves the same complexity for MORL, which is a great achievement. Actually, under some special structure like linear quadratic problem, the one can show a sample complexity of , as proved by Zhou and Lu in \\u201cSingle Timescale Actor-Critic Method to Solve the Linear Quadratic Regulator with Convergence Guarantees\\u201d. Please consider adding it as a remark to enhance completeness of the paper.\\n\\n**Our Response:** Thank you for pointing out this excellent reference. We are more than happy to add discussion on sample complexity of the speacial structured RL problems in the revision as you suggested.\"}", "{\"metareview\": \"This paper proposes the Weighted-Chebyshev Multi-Objective Actor-Critic (WC-MOAC) algorithm for Multi-Objective Reinforcement Learning (MORL). The authors aim to systematically explore the Pareto Stationary Front using an integration of Weighted-Chebyshev scalarization and momentum-based updates within the actor-critic framework. There are significant concerns regarding its originality and overlap with prior work, particularly with the paper by Zhou et al. (2024). These concerns include structural, algorithmic, and even textual similarities, which undermine the contribution of this work as a standalone advance.\\n\\nThis meta-review was reviewed by the senior area chair and the SAC confirms the text and decision.\", \"additional_comments_on_reviewer_discussion\": \"During the rebuttal period, reviewers raised significant concerns regarding the overlap of the submission with Zhou et al. (2024), particularly in algorithmic design, theoretical guarantees, and textual content. Reviewers also emphasized the lack of novelty in the weighted-Chebyshev formulation, arguing it was a minor extension of existing methods. I read the paper and agree with Reviewers P7fw and JPv7 that this paper has concerning similarities to Zhou et al. (2024), even with text overlap flagging ethical concerns.\"}", "{\"title\": \"Response to Reviewer P7Fw's Follow-up Comments:\", \"comment\": \"Thank you for your comments. Based on the comments, our reponses are as follows.\\n>**Your Follow-up Comment 1:** I sincerely thank the authors for the detailed response and for answering my technical questions. Below let us focus more on the Comments 1-8.\\n\\n\\n**Our Response:** We are glad to see that you have accepted our technical contributions and technical differences compared to (Zhou et al.(2024)) in our previous responses to your Comments 9-11.\\n\\n_____________\\n>**Your Follow-up Comment 2:** Comment 1 (regarding the pseudo code of WC-MOAC and MOAC), Comment 2 (regarding the theoretical results), and Comments 3-4 (regarding the introduction): I can understand that both algorithms share the actor-critic architecture, and the two papers are both focused on the convergence to the stationarity. However, from a reviewer\\u2019s viewpoint, given the high similarity (in terms of design, paper structure, and even wording), it appears almost impossible to argue that this paper is completed independently from [Zhou et al., 2024] or the two are just concurrent works. Moreover, as far as I understand, such similarity in both structure and wording shall be viewed as paraphrasing, which is not allowed if there is no due credit or acknowledgement given to the prior works (i.e., [Zhou et al., 2024]). Furthermore, as mentioned by Reviewer JPv7, without a proper comparison with or mention of [Zhou et al., 2024], the readers can be easily misguided by the overclaimed contributions.\\n\\n**Our Reponse:** In Section 2 on related work, we have specifically compared and contrasted our work to (Zhou et al. (2024)). Specifically, we have stated that (Zhou et al. (2024)) falls into the no-preference category (See Line 139) and is not designed for Pareto stationarity front exploration. But to address the reviewer's concern, we will provide a much more thorough comparison to (Zhou et al.(2024)) and give more explicit credit to (Zhou et al. (2024)) in our revisions. \\n\\n____________\\n>**Your Follow-up Comment 3:** Comment 5 (regarding the Key Contributions): Just like the comment above, my concern about paraphrasing in this part still remains unsolved. Moreover, apart from the structure and wording, the novelty and contributions can be overestimated without properly mentioning [Zhou et al., 2024]. Specifically:\\nBased on the current writing of the paper, the readers can get easily misled that WC-MOAC is a totally new algorithm in the MORL literature. However, under a comparison with MOAC, WC-MOAC can indeed appear somewhat incremental.\\n\\n**Our Response:** We understand that it is the reviewer's right to claim our work is \\\"incremental\\\". However, even if our work is indeed \\\"incremental\\\", which we respectfully disagree, it is fundamentally different from \\\"plagiarism\\\" and \\\"dual submission\\\". These harsh accusations are harmful to all authors' academic reputatation and even career-threating. Therefore, we sincerely hope the reviewer can remove the flag even if the reviewer still leans toward a rejection.\\n_____________\\n>**Your Follow-up Comment 4:** \\n>> The second bullet in our \\\"Key Contributions\\\" addresses, even with such weight vector $\\\\mathbf{p}$, it achieves sample complexity that is independent of $M$, where $M$ is the number of objectives.\\n>\\n> This is also one key feature that has been shown by MOAC. Then, this would not be a convincingly new feature for people to choose WC-MOAC over MOAC.\\n\\n**Our Response:** In our revisions, we are happy to tone down our wording in the bullets of key contributions to important remark and cite (Zhou et al.(2024)). However, we do want emphasize that there does **not** exist such an issue of *\\\"choosing WC-MOAC over MOAC or vice versa\\\"*. This is because this paper studies the systematic exploration of the Pareto Stationay front, while (Zhou et al.(2024)) is only designed to guarantee the convergence to a Pareto stationary point. In other words, these two works aim at two completely different goals and not competing with each other. Therefore, it is important for WC-MOAC to establish/maintain an $M$-independence property in the **new problem** of systematically exploring Pareto stationarity front, which is **not** the goal of (Zhou et al. (2024)).\\n________________\"}", "{\"title\": \"Misguiding Paper Writing With Over-stated Contributions\", \"comment\": \"Thanks a lot for your detailed replies! We can now focus on my previous Comments 1, 2, 7 and 8.\\n\\nThe main issue with the paper's organization is that others can strongly overestimate the contributions. \\n\\nAs stated in Reviewer P2zK's summary, \\\"The WC-MOAC algorithm is designed to combine multi-temporal-difference (TD) learning in the critic phase with ... multi-gradient descent in the actor phase, effectively managing the complexities of non-convexity and scalability in multi-objective problems. The primary theoretical result of this work is a finite-time sample complexity bound, which is independent of the number of objectives ... This independence ... is a notable theoretical advance.\\\"\", \"reviewer_qykr02_stated\": \"\\\" Central to the discussion is the concept of (local) (weak) Pareto optimality, which serves as the foundational solution criterion for balancing multiple objectives. The proposed training architecture employs a single actor coupled with multiple critics, where each critic is specifically trained to approximate the value function corresponding to a distinct objective.\\\"\\n\\nNone of these are your novel contributions, which Zhou and colleagues have introduced. The paper writing is strongly misleading.\\n\\nAs your reply to the first comment stated, this work \\\"focuses on systematically exploring the Pareto Stationary front using a weighted-Chebyshev formulation.\\\" The introduction of the preference weighting is the main contribution.\\n\\nHowever, your experiment design does not show the advantage of including such weighting. For the experiment in Table 1, the effect of emphasizing \\\"Comment\\\" or \\\"Dislike\\\" objectives can be more persuasive. As replied to Comments 7 and 8, the preference vector is not trained. What is the systematic way of deciding the weighting in your approach? As TSCAC, does your method also require domain knowledge? Meanwhile, for the experiment shown in Figure 1, concentrating on click, dislike, and watch do not differ much.\"}", "{\"title\": \"Reponse to Reviewer P2zK's Comments (Continued):\", \"comment\": \">**Your Comment 6 (Also Question 4):** The paper also raises questions regarding the sensitivity of the algorithm's performance to variations in $p$ and whether there are recommended heuristics for selecting optimal values.\\n\\n**Our Response:** Thanks for your comments and suggestions. We are currently working hard on running new simulations. We will attach the new results as soon as we are finished. \\n\\n________________\\n>**Your Comment 7 (Also Question 5):** Lastly, in relation to multi-temporal difference learning, were there specific stability challenges encountered, particularly when combining it with weighted-Chebyshev updates? If so, what mechanisms or parameters were introduced to maintain convergence stability?\\n\\n**Our Response:** Thanks for your questions. First of all, we want to get a clarification on what the \\\"stability\\\" is referring to in the reviewer's first question. If it's referring to convergence, then we have following response:\\n\\nIn each iteration of the critic component, the multiple-temporal difference (TD) learning is evaluating the current *stationary policy* on all $M$ objectives. Due to the fact that given a state-action pair $(s,a)\\\\in\\\\mathcal{S}\\\\times\\\\mathcal{A}$, the reward from objective $i$ (i.e. $r^{i}(s,a)$) and the reward from objective $j$ (i.e., $r^{j}(s,a)$) is independent of each other for any $i,j\\\\in [M]$ and $i\\\\neq j$, we can utilize existing single-objective TD learning analysis, where a negative Lyapunov drift argument can be constructed to ensure convergence (i.e., achieving stability).\\n\\nIf the terminology \\\"stability\\\" in the reviwer's question means \\\"robustness under non-stationary policies\\\" (e.g., non-stationary polices required in RL problems under partially observed Markov decision processes (POMDP)), then this remains a very challenging open problem, which deserves a dedicated paper on this topic. We are very interested in this open problem in our future studies and we thank the reviewer for pointing out this research direction.\"}", "{\"summary\": \"This paper presents the weighted-Chebyshev multi-objective actor-critic (WC-MOAC) algorithm to address the challenge of systematically exploring Pareto-stationary solutions in multi-objective reinforcement learning under multiple non-convex reward objectives. The WC-MOAC algorithm is designed to combine multi-temporal-difference (TD) learning in the critic phase with weightedChebyshev scalarization and multi-gradient descent in the actor phase, effectively managing the complexities of non-convexity and scalability in multi-objective problems. The primary theoretical result of this work is a finite-time sample complexity bound, $O\\\\left(\\\\epsilon^{-2} p_{\\\\min }^{-2}\\\\right)$, which is independent of the number of objectives $M$, where $p_{\\\\min }$ denotes the minimum entry of the weight vector $p$ in the scalarization. This independence from $M$ is a notable theoretical advance. The empirical evaluation, conducted on the KuaiRand offline dataset, indicates that the WC-MOAC algorithm surpasses baseline methods in performance, highlighting its robustness and potential for realworld applications in multi-objective reinforcement learning.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"3\", \"strengths\": \"I feel the paper's originality comes from its novel application of the weighted-Chebyshev scalarization technique within the framework of multi-objective reinforcement learning, accompanied by theoretical guarantees on finite-time sample complexity. This is particularly noteworthy given the limited prior work addressing non-convex reward objectives in this field. The technical rigor is evident, as the authors present a well-founded theoretical analysis that establishes finite-time sample complexity guarantees. This analysis is both rigorous and practically relevant, especially in environments that require systematic exploration of Pareto-stationary solutions. The empirical validation further strengthens the contribution, with experimental results on the KuaiRand dataset showing that the proposed algorithm consistently outperforms baseline approaches. This demonstrates the algorithm's robustness and effectiveness in practical scenarios. Additionally, the research fills an important gap in multi-objective reinforcement learning, with potential applications across various domains requiring multi-objective optimization. Its relevance to real-world problems enhances the significance of the work.\", \"weaknesses\": \"While the paper provides a comprehensive theoretical foundation, certain methodological aspects could be clarified further. For example, the integration of the multi-gradient descent update, which computes a dynamic weighting vector $\\\\lambda_t$ that balances exploration with convergence, could benefit from a more detailed discussion on its rationale and practical implementation steps. Additionally, the empirical evaluation is limited to a single dataset, the KuaiRand offline dataset, which raises questions about the algorithm's generalizability. Expanding the experimental analysis to include diverse datasets or multi-objective environments would provide deeper insights into the algorithm's robustness across varied applications.\", \"questions\": \"The paper would benefit from further clarification on several key points. First, could the authors discuss the impact of the preference vector $p$ on the algorithm's performance, especially in environments with highly non-convex reward landscapes? As the convergence rate depends on $p_{\\\\min }$, understanding the selection of $p$ would be useful for real-world applications. Additionally, what limitations, if any, might arise in adapting the proposed algorithm to online settings, given that the study's empirical evaluation is restricted to offline data? The paper also raises questions regarding the sensitivity of the algorithm's performance to variations in $p$ and whether there are recommended heuristics for selecting optimal values. Lastly, in relation to multi-temporaldifference learning, were there specific stability challenges encountered, particularly when combining it with weighted-Chebyshev updates? If so, what mechanisms or parameters were introduced to maintain convergence stability?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"details_of_ethics_concerns\": \"N/A.\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to Reviewer JPv7's comments (Continued)\", \"comment\": \">**Your Comment 8 (Also Question 4):** For the experiment in Figure 1, would TSCA work better by switching the main objective? More experiment designs to show the necessity of preference would be better.\\n\\n**Our Response:** In principle, it is possible to switch objects as suggested to get potentially different results for TSCAC algorithm. However, it does require priori knowleges on what should constitute reasonable contraints as it is a constraint actor-critic algorithm. On the other hand, it shows the advantage of our approach, which explores systematically, as it doesn't require additional domain knowledge which serves as constraints.\\n\\n________________\\n>**Your Comment 9 (Also Question 5):** The paper focuses on the finite-time convergence results for multi-objective actor-critic. What is the advantage of using actor-critic algorithms?\\n\\n**Our Response:** If one were to compare actor-critic(AC) with other algorithms, notably Q-learning. We have the following comparisons:\\n1. on policy vs off policy: AC generally speaking is an on-policy method. In constrast to Q-learning, which is an off-policy algorithm, the Q-function of the sampling policy doesn't align with the current Q-values being computed. As a result, AC provides a better understanding and intuition for the current policy for both single-objective RL and MORL.\\n2. Continuous vs discrete state-action space (Grondman et al. (2012)): As policy for AC is parameterized, it is natural for AC to extend to continuous state-action space in comparison with Q-learning. For Q-learning to work properly, Q functions are typically required to be approximated due to $\\\\max$ operator, which is not well-defined in non-compact continuous space. This should hold true for both single-object and multi-objective scenarios.\\n \\nGrondman, Ivo, et al. \\\"A survey of actor-critic reinforcement learning: Standard and natural policy gradients.\\\" IEEE Transactions on Systems, Man, and Cybernetics, part C (applications and reviews) 42.6 (2012): 1291-1307.\\n\\n3. Since AC is a policy gradient based algorithm, it is natural to incorporate multiple-gradient based approach, which is a standard approach for finding Pareto stationary points.\\n\\n________________\\n\\nWe hope our responses have addressed the reviewer's concerns, and we are happy to answer any further questions. If the concerns are addressed, we would highly appreciate a re-rating of our work.\"}", "{\"title\": \"Response to Reviewer P7Fw's Follow-up Comments (Continued):\", \"comment\": \"> **Your Follow-up Comment 7:** Comments 6-8 (regarding the writing of the problem formulation and preliminaries): I understand that MDP and MOMDP are standard problem formulations in RL. However, there are various ways to describe the same setting. As far as I know, it is a standard practice to describe the problem formulation that tailors to the need of each specific paper and hence shall be written in the authors\\u2019 own words. Again, as mentioned above, this similarity in both structure and wording shall be viewed as paraphrasing and is not allowed if there is no due credit or acknowledgement given to the prior works. Therefore, I would like to emphasize that \\u201cusing the similar formulations\\u201d is not the issue, and the high similarity in terms of structure and wording is where the concerns arise. I hope that the authors can understand my concerns.\\n\\n**Our Response:** In terms of formulations, we have already genuinely described them using our own words. In our humble opinion, this is strictly following academic fair use. Moreover, we want to mention such standard phrasing always brings similarity, as the intention is to precisely describe the same mathematical models that have been widely accepted in the literature. Additionally, we have provided many literature referencs that uses the same assumptions and similar preliminaries. In addition, MDP and MOMDP are standard concepts, not new contributions of any related work in this field. One may find such simialirites in Section 3 of paper (Qiu et al (2021)), Section 2.1 of (Roijers et al (2018)), Section 2.1 of paper (Zhang et al (2018)) and Section 3.1 of (Hairi et al (2022)). Also, one may find similar preliminaries in (Qiu et al (2021)) and (Xu et al.(2020)). \\n\\nQiu, S., Yang, Z., Ye, J., and Wang, Z. On finite-time convergence of actor-critic algorithm. IEEE Journal on Selected Areas in Information Theory, 2(2):652\\u2013664, 2021.\\n\\nRoijers, D. M., Steckelmacher, D., and Nowe \\u0301, A. Multi-objective reinforcement learning for the expected utility of the return. In Proceedings of the Adaptive and Learn- ing Agents workshop at FAIM, volume 2018, 2018.\\n\\nHairi, F., Liu, J., and Lu, S. Finite-time convergence and sample complexity of multi-agent actor-critic reinforcement learning with average reward. In International Con- ference on Learning Representations, 2022.\\n\\nZhang, K., Yang, Z., Liu, H., Zhang, T., and Basar, T. Fully decentralized multi-agent reinforcement learning with networked agents. In International Conference on Ma- chine Learning, pp. 5872\\u20135881. PMLR, 2018.\\n\\nTengyu Xu, Zhe Wang, and Yingbin Liang. Improving sample complexity bounds for (natural) actor-critic algorithms. arXiv preprint arXiv:2004.12956, 2020\\n\\nFor example, in Section 2.1 of (Zhang et al.(2018)), the definition is *\\\"A Markov decision processes is characterized by a quadruple $M = \\\\langle S , A, P , r \\\\rangle$, where S is a finite state space, A is a finite action space, $P (s\\u2032 | s, a) : \\\\mathcal{S} \\u00d7 {A} \\u00d7 {S} \\u2192 [0, 1]$ is a state transition probability\\\" from state s to s\\u2032 determined by action a, and $R(s, a) : S \\u00d7 A \\u2192 R$ is a reward function defined by $R(s, a) = \\\\mathbb{E}[r_{t+1} | s_t = s, a_t = a]$, with $r_{t+1}$ being the instantaneous reward at time t. Policy of the agent is a mapping $\\\\pi : \\\\mathcal{S} \\u00d7 \\\\mathcal{A} \\\\rightarrow [0, 1]$, representing the probability of choosing action a at state s. The objective of the agent is to find the optimal policy that maximizes the expected time-average reward, notably, long-term return, which is given by $J(\\\\pi)$:\\n$$J(\\u03c0) = \\\\lim_{T} \\\\frac{1}{T}\\\\sum_{t=0}^{T-1}\\\\mathbb{E}(r_{t+1}) = \\\\sum_{s\\\\in\\\\mathcal{S}}d_{\\\\pi}(s)\\\\sum_{a\\\\in\\\\mathcal{A}}\\\\pi(s,a)R(s,a), $$\\nwhere $d_{\\\\pi}(s) = \\\\lim_{t} \\\\mathbb{P}(s_t = s|\\\\pi)$ is the stationary distribution of the Markov chain under policy $\\\\pi$.\\\"*\"}", "{\"title\": \"Response to Reviewer P7fw (Continued)\", \"comment\": \"(continued response to Comment 11):\\n\\n4) **Theorem 3 in (Zhou et al. (2022)):** Theorem 3 in (Zhou et al. (2022)) actually highlights the **new** contribution of our analysis since we don't have the $\\\\sum_{i=1}^{m}|\\\\hat{\\\\lambda_k^{i}}-\\\\lambda_{k-1}^{i}| $ coefficient terms in front of $O(\\\\frac{1}{n})$, where $n$ is the same as $T$ in this work. The difference is that their convergence metric is $O(m)$ in the worst case because of the $\\\\sum_{i=1}^{m}|\\\\hat{\\\\lambda_k^{i}}-\\\\lambda_{k-1}^{i}|$ coefficient terms, where $m$ is the number of objectives in (Zhou et al.(2022)). Similarly, the subsequent terms in Theorem 3 of (Zhou et al. (2022)) are not only at least $O(m)$-dependent, but also depend on the accumulative learning rates $\\\\sum_{k}\\\\eta_k$ and momentum rates $\\\\sum_{k}(1-\\\\alpha_k)$. These key differences highlights novelty in our theoretical analaysis on pseudo weight $q_t:=\\\\frac{\\\\lambda_t\\\\odot \\\\mathbf{p}}{\\\\langle \\\\lambda_t, \\\\mathbf{p}\\\\rangle}$ term.\\n\\n________________\\n\\nWe hope our responses have addressed the reviewer's concerns on ethics and technical novelty, and we are happy to answer any further questions. If the concerns are addressed, we would highly appreciate a re-evaluation of our work.\"}", "{\"summary\": \"This paper proposes weighted-Chebyshev multi-objective actor-critic (WC-MOAC) to find Pareto-stationary policies for MORL with sample complexity guarantees. The proposed WC-MOAC combines two techniques, namely the (stochastic) multiple gradient approach and the momentum-based updates of the dual variables. The WC-MOAC has a sample complexity of $\\\\tilde{O}(\\\\epsilon^{-2})$ in finding an epsilon Pareto-stationary policy. Finally, simulation results on an offline dataset is provided.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"1\", \"strengths\": [\"The obtained sample complexity is independent of the number of objective functions and has a nice scaling with $M$ (However, there is a severe concern stated below).\", \"This paper handles two formulations, namely, discounted-reward and average-reward MDPs, simultaneously (despite that the algorithm and the analysis are agnostic to the reward setting in principle).\", \"Overall the paper is well-organized and easy to follow, with the concepts, definitions, and theoretical results clearly explained.\"], \"weaknesses\": [\"The major issue with this paper is that it is very similar to the paper \\u201cFinite-Time Convergence and Sample Complexity of Actor-Critic Multi-Objective Reinforcement Learning\\u201d by (Zhou et al., 2024) published in ICML 2024, in terms of both the algorithm design, paper writing, and the claimed contributions. The flow of the paper exactly follows (Zhou et al., 2024), and most of the paragraphs are just paraphrased versions of (Zhou et al., 2024). For example:\", \"The pseudo code of WC-MOAC are almost the same (almost verbatim) as those of the MOAC algorithm (cf. Algorithms 1 and 2 in (Zhou et al., 2024)).\", \"The theoretical result of WC-MOAC in Theorem 4 and Corollary 5 appear almost the same as the Theorem 5 and Corollary 6 in (Zhou et al., 2024).\", \"The 1st paragraph of Introduction (Lines 34-46) appears to be paraphrased from the first two paragraphs of the Introduction of (Zhou et al., 2024).\", \"The 2nd paragraph of Introduction (Lines 47-64) appears to be paraphrased from the third paragraph of the Introduction of (Zhou et al., 2024).\", \"The paragraphs about the \\u201cKey Contributions\\u201d (Lines 97-124) appear to directly follow the \\u201cMain Contributions\\u201d of the Introduction of (Zhou et al., 2024).\", \"In Section 3, the problem formulation about MOMDP (Lines 181-187) appears very similar to the second paragraph of Section 3.1 of (Zhou et al., 2024).\", \"The part on \\u201cLearning Goal and Optimality in MORL\\u201d (Lines 200-215) appears almost the same as the \\u201cProblem Statement\\u201d in Section 3.1 of (Zhou et al., 2024). Moreover, even the footnote #3 in this paper almost goes verbatim compared to the footnote #1 in (Zhou et al., 2024).\", \"The preliminaries about the policy gradient for MORAL (Lines 274-296) also largely resembles Section 3.2. Specifically, several sentences about Lemma 2 and Assumption 2 of this paper are exactly the same as those in Lemma 1 and Assumption 2 of (Zhou et al., 2024).\", \"The description about Assumption 3 and Lemma 3 in this paper (Lines 424-437) appears to exactly follow the Assumption 3 and Lemma of (Zhou et al., 2024).\"], \"questions\": [\"In addition to the above, here are some further technical questions:\", \"Technically: One of my main technical concerns is the motivation for finding a Pareto-stationary policy (under the assumption that the state and action spaces are finite) in the specific context of MORL. Specifically, while it is indeed difficult to find the whole Pareto front in MORL, it is actually not hard to find one or some Pareto-optimal policies by adapting reducing MORL to single-objective RL and finding the convex coverage set (e.g., (Yang et al., 2019)). For example, based on (Chen and Magulari, 2022), one can use a policy-based method with off-policy TD learning (under linear function approximation) to find an epsilon-optimal solution for single-objective RL with a sample complexity of $\\\\tilde{O}(1/\\\\epsilon^2)$. There are also several other recent works like (Lan 2021; Fatkhullin et al., 2023; Liu et al., 2020; Chen et al., 2022) that can find an epsilon-optimal policy with sample complexity guarantees. To adapt these results to MORL, one can use linear scalarization and thereby find one Pareto-optimal policy (specific to some preference vector). As a result, it remains not totally clear why it is theoretically appealing to design an algorithm for finding only a Pareto-stationary policy if we can already find Pareto-optimal policies (despite that Pareto-stationarity is indeed a widely adopted concept in the MOO literature).\", \"Another concern is the novelty in terms of algorithm and convergence analysis. Specifically, the WC-MOAC algorithm appears to be a direct application of the MOAC algorithm (Zhou et al. 2024) and also similar to the MOO algorithm CR-MOGM of (Zhou et al. 2022), which is the enhanced (stochastic) MGDA method (e.g., (Desideri, 2012; Liu and Vicente, 2021)) with the momentum update of the dual variable vector, to the setting of MORL (with the multi-objective critic learned by standard TD updates). Under a properly learned critic, then the stochastic multiple gradients can have a sufficiently low bias such that it enables similar convergence guarantees as in the general MOO. This is also shown in Theorem 11 (as a direct result of Lemma 10). As a result, the sample complexity and the convergence analysis of WC-MOAC essentially resemble those of CR-MOGM in (Zhou et al. 2022) for the general non-convex case, cf. Theorem 3 and Appendix E of (Zhou et al. 2022).\"], \"references\": [\"(Yang et al., 2019) Runzhe Yang, Xingyuan Sun, and Karthik Narasimhan, \\u201cA Generalized Algorithm for Multi-Objective Reinforcement Learning and Policy Adaptation,\\u201d NeurIPS 2019.\", \"(Zhou et al., 2022) Shiji Zhou, Wenpeng Zhang, Jiyan Jiang, Wenliang Zhong, Jinjie Gu, Wenwu Zhu, \\u201cOn the Convergence of Stochastic Multi-Objective Gradient Manipulation and Beyond,\\u201d NeurIPS 2022.\", \"(Chen and Magulari, 2022) Zaiwei Chen and Siva Theja Maguluri, \\u201cSample Complexity of Policy-Based Methods under Off-Policy Sampling and Linear Function Approximation,\\u201d AISTATS 2022.\", \"(Lan 2021) Guanghui Lan, \\u201cPolicy Mirror Descent for Reinforcement Learning: Linear Convergence, New Sampling Complexity, and Generalized Problem Classes,\\u201d Mathematical programming, 2021.\", \"(Fatkhullin et al., 2023) Ilyas Fatkhullin, Anas Barakat, Anastasia Kireeva, Niao He, \\u201cStochastic Policy Gradient Methods: Improved Sample Complexity for Fisher-non-degenerate Policies,\\u201d ICML 2023.\", \"(Liu et al., 2020) Yanli Liu, Kaiqing Zhang, Tamer Basar, Wotao Yin, \\u201cAn Improved Analysis of (Variance-Reduced) Policy Gradient and Natural Policy Gradient Methods,\\u201d NeurIPS 2020.\", \"(Chen et al., 2022) Zaiwei Chen, Sajad Khodadadian, and Siva Theja Maguluri, \\\"Finite-sample analysis of off-policy natural actor\\u2013critic with linear function approximation,\\\" IEEE Control Systems Letters, 2022.\"], \"flag_for_ethics_review\": \"['Yes, Research integrity issues (e.g., plagiarism, dual submission)']\", \"details_of_ethics_concerns\": [\"One major issue with this paper is that it is very similar to the paper \\u201cFinite-Time Convergence and Sample Complexity of Actor-Critic Multi-Objective Reinforcement Learning\\u201d by (Zhou et al., 2024) published in ICML 2024, in terms of both the algorithm design, paper writing, and the claimed contributions. The major issue with this paper is that it is very similar to the paper \\u201cFinite-Time Convergence and Sample Complexity of Actor-Critic Multi-Objective Reinforcement Learning\\u201d by (Zhou et al., 2024) published in ICML 2024, in terms of both the algorithm design, paper writing, and the claimed contributions. The flow of the paper exactly follows (Zhou et al., 2024), and most of the paragraphs are just paraphrased versions of (Zhou et al., 2024). For example:\", \"The pseudo code of WC-MOAC are almost the same (almost verbatim) as those of the MOAC algorithm (cf. Algorithms 1 and 2 in (Zhou et al., 2024)).\", \"The theoretical result of WC-MOAC in Theorem 4 and Corollary 5 appear almost the same as the Theorem 5 and Corollary 6 in (Zhou et al., 2024).\", \"The 1st paragraph of Introduction (Lines 34-46) appears to be paraphrased from the first two paragraphs of the Introduction of (Zhou et al., 2024).\", \"The 2nd paragraph of Introduction (Lines 47-64) appears to be paraphrased from the third paragraph of the Introduction of (Zhou et al., 2024).\", \"The paragraphs about the \\u201cKey Contributions\\u201d (Lines 97-124) appear to directly follow the \\u201cMain Contributions\\u201d of the Introduction of (Zhou et al., 2024).\", \"In Section 3, the problem formulation about MOMDP (Lines 181-187) appears very similar to the second paragraph of Section 3.1 of (Zhou et al., 2024).\", \"The part on \\u201cLearning Goal and Optimality in MORL\\u201d (Lines 200-215) appears almost the same as the \\u201cProblem Statement\\u201d in Section 3.1 of (Zhou et al., 2024). Moreover, even the footnote #3 in this paper almost goes verbatim compared to the footnote #1 in (Zhou et al., 2024).\", \"The preliminaries about the policy gradient for MORAL (Lines 274-296) also largely resembles Section 3.2. Specifically, several sentences about Lemma 2 and Assumption 2 of this paper are exactly the same as those in Lemma 1 and Assumption 2 of (Zhou et al., 2024).\", \"The description about Assumption 3 and Lemma 3 in this paper (Lines 424-437) appears to exactly follow the Assumption 3 and Lemma of (Zhou et al., 2024).\", \"Based on the above, there could be some research integrity issues that would require further attention.\"], \"rating\": \"1\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to Reviewer P7Fw's Follow-up Comments (Continued):\", \"comment\": \"(Continuing response to follow-up comment 7)\\n\\nSimilarly in Section 3 of (Qiu et al.(2021)), it states *\\\"The infinite-horizon average reward reinforcement learning problem [24], [25] is modeled as an average reward Markov Decision Process (MDP). Suppose that $\\\\mathcal{S}$ and $\\\\mathcal{A}$ are the finite state space and finite action space respectively. The policy $\\\\pi$ is defined as a function that $\\\\pi : \\\\mathcal{A}\\u00d7\\\\mathcal{S} \\\\rightarrow [0,1]$ such that $\\\\pi(a|s)$ is the probability of choosing action $a \\\\in \\\\mathcal{A}$ at state $s \\\\in \\\\mathcal{S}$. From a practical perspective, the policy $\\\\pi$ is usually parameterized by $\\\\theta\\\\in\\\\Theta$ in a nonconvex form and then we denote the parameterized policy as $\\\\pi_{\\\\theta}$ . An agent takes $a \\u223c \\\\pi_{\\\\theta}(\\u00b7 | s)$ at state s. Letting $P(s'| a, s)$ be the probability of an agent moving from state $s$ to state $s'$ with an action $a$, we can have a Markov transition probability induced by $\\\\pi_{\\\\theta}$ as$P^{\\\\pi_{\\\\theta}} (s\\u2032 | s) = \\\\sum_{a\\\\in\\\\mathcal{A}} P(s\\u2032 | a, s)\\\\pi_{\\\\theta}(a | s)$, which is the probability of moving from state $s$ to state $s\\u2032$. At each time $\\u03c4$, we use a tuple $(s_\\u03c4,a_\\u03c4,s_{\\u03c4+1},r_{\\u03c4+1})$ to denote that an agent at state $s_\\u03c4$ chooses an action $a_\\u03c4$ and transitions to the next state $s_{\\u03c4+1}$ with a reward $r_{\\u03c4+1} :=r(s_\\u03c4,a_\\u03c4,s_{\\u03c4+1})$,where $r:S\\u00d7A\\u00d7S\\\\rightarrow\\\\mathbb{R}$ is a map to reward values. We assume $|r| \\u2264 r_{\\\\max}$. Next, we make the following assumption on the policy $\\u03c0_{\\\\theta}$ and the probability $P(s\\u2032 | a, s)$.\", \"assumption_1\": \"We assume that the parameterization of $\\\\pi_{\\\\theta}$ and $P(s\\u2032|a,s)$ guarantee that the Markov chain decided by $P^{\\\\pi_{\\\\theta}} (s\\u2032 | s)$ for any $\\\\theta\\\\in\\\\Theta$ is irreducible and aperiodic. Then, there can always exist a unique stationary distribution for any $\\\\theta\\\\in\\\\Theta$, which is denoted as $d_{\\\\pi_{\\\\theta}}(s)$ for any state $s\\\\in\\\\mathcal{S}$.\\\"*\\n\\nIt is clear from the above two examples of MDP definitions that there are significant wording overlaps in related work in the literature. Nonetheless, we are happy to adopt the suggestion of citing (Zhou et al.(2024)) in the formulations to indicate the similarity in problem setting.\"}", "{\"title\": \"Response to Reviewer qYKr's Comments\", \"comment\": \">**Your Comment 1 (Also Weakness 1):** The major weakness might lie in the novelty of the proposed method. To the knowledge of the reviewer, Although the theoretical analysis partially answers the challenges mentioned in the introduction, the proposed method seems to largely based on (D\\u00e9sid\\u00e9ri, 2012) and (Momma et al., 2022), with the multi-objective optimization gradients replaced with policy gradients.\\n\\n**Our Response:** Thanks for your comments. We agree with the reviewer that the actor component of the proposed MOAC approach is indeed based on the MGDA approach in (D\\u00e9sid\\u00e9ri, 2012) and (Momma et al. (2022)). However, we do want to emphsize that the adaptation of these techniques to the MORL setting is by no means straightforward due to the unique challenges in MORL, which requires additional new algorithmic techniques (e.g., the momentum-based approach to eliminate systematic biases). Moreover, the theoretical finite-time convergence proof and analysis are new, which cannot be found in (D\\u00e9sid\\u00e9ri, 2012) and (Momma et al. (2022)). In particular, the Pareto stationary convergence rate's dependency on the weight vector $\\\\mathbf{p}$ is a new result in the literature. In addition, the empirical studies in Figure 1 does indicate that our proposed WC-MOAC approach in Algorithm 1 is able to explore a much larger Pareto stationarity front footprint than that of the conventional approach based on linear scalarization.\\n\\n________________\\n>**Your Comment 2 (Also Question 1):** In Theorem 4, what is the different between |w*-w| and zeta? Intuitively, both these two terms are related to the error of value functions. The reviewer is confused because the definition of w* also needs clarification. The optimality of the critic network depends on both $\\\\phi(s)$ and w. Different state representation phi might correspond to different optimal w. Is $\\\\phi$ function fixed?\\n\\n**Our Response:** Thank you for your comment. $\\\\mathbf{w}^{i,*} - \\\\mathbf{w}_t^i$ and $\\\\zeta_{\\\\text{approx}}$ have different physical meansings, which are stated as follows:\\n\\n**1) The meaning of $\\\\zeta_{\\\\text{approx}}$:** The quantity $\\\\zeta_{\\\\text{approx}}$ represents the *fundamental inability* of the linear function approximation apporach for accurately quantifying the $i$-th value function (i.e., $V^{i}(s)\\\\approx\\\\phi(s)^{\\\\top} \\\\mathbf{w}^{i}$). Mathematically, $\\\\zeta_{\\\\text{approx}}=\\\\max_{i\\\\in[M]}V^{i}(\\\\cdot) - \\\\mathbf{w}^{i,\\\\*}\\\\Phi$ characterizes the maximum gap between the ground truth value function $V^{i}$ and best linear approximation given the feature matrix $\\\\Phi$ and the optimal linear approximation solution $\\\\mathbf{w}^{i,*}$ among all $M$ objectives. \\n\\n**2) The meaning of $\\\\mathbf{w}^{i,\\\\*} - \\\\mathbf{w}_t^i$:** The quantity $\\\\mathbf{w}^{i,\\\\*} - \\\\mathbf{w}^i_t$ represents the finite-time convergence gap compared to the optimal linear function approximation solution $\\\\mathbf{w}^{i,*}$, which itself has a gap $\\\\zeta_{\\\\text{approx}}$ to the ground truth value function $V^i(\\\\cdot)$ of the $i$-th objective.\\n\\nFor ease of explaination, let's use the discounted reward setting in here as a concrete example. The value function $V^{i}(s)$ for the $i$-th objective at $s\\\\in \\\\mathcal{S}$ follows the canonical definition of the discounted reward setting. However, in some MDPs, due to large state space $\\\\mathcal{S}$, approximation is used to cope with the high dimensionality challenge. In this paper, we applied linear approximation assuming the feature vector mapping $\\\\phi(s)$ is *given*. How to choose a suitable feature mapping $\\\\phi(\\\\cdot)$ is an interesting problem by itself, but beyond the scope of this paper. This suggests for each objective $i$, the critic will maintain a $\\\\mathbf{w}^{i}$ parameter to evaluate the $i$-th objective. $\\\\mathbf{w}^{i,\\\\*}$ is used to denote the optimal approximation under such linear approximation. Please see Line 854 in the supplementary material for the precise defition of $\\\\mathbf{w}^{i,\\\\*}$ under the discounted reward setting. In Theorem 4, the term $\\\\mathbf{w}^{i}-\\\\mathbf{w}^{i,*}$ refers to the error in convergence from critic component of Algorithm 1 and the optimal linear approximation parameter for $i$-th objective.\\n\\n________________\"}", "{\"title\": \"Response to Reviewer P7fw (Continued)\", \"comment\": \"> **Your Comment 2:** The theoretical result of WC-MOAC in Theorem 4 and Corollary 5 appear almost the same as the Theorem 5 and Corollary 6 in (Zhou et al., 2024).\\n\\n**Our Response:** Similar to most theoretical work on RL, in Theorem 4 of our paper, we proved the convergence to the stationarity under each given $\\\\mathbf{p}$-exploration vector. The key differences in the convergence results compared to (Zhou et al., 2024) stem from the effect of the $\\\\mathbf{p}$-vector. This weight vector affects the convergence in a fashion that is inversely proportional to $p_{\\\\min}^2$, where $p_{\\\\min}$ is minimum entry of the weight vector. What this **new** result entails is that the convergence rate of exploring the Pareto stationary front (also the Pareto front under convex settings) could be affected by $p_{\\\\min}$. In other words, the convergence rate could potentially be slower when $p_{\\\\min}$ is small.\\nFurthermore, in Corollary 5, the sample complexity of the Pareto stationarity front exploration is also characterized by $p_{\\\\min}$ when designing learning rate $\\\\eta_t$. Specifically, by carefully setting the learning rate $\\\\eta_t=\\\\frac{p^{2}_{\\\\min}}{t^{2}}$, one can achieve a $\\\\mathbf{p}$-independent sample complexity. In contrast, in (Zhou et al., 2024), there is no such Pareto stationarity front exploration in the convergence their result. In addition, the goal in (Zhou et al., 2024) is focused on converging to only one of potentially infinite Pareto Stationary point, which is a much weaker goal compared to ours.\\n\\n________________\\n> **Your Comments 3 and 4:** The 1st paragraph of Introduction (Lines 34-46) appears to be paraphrased from the first two paragraphs of the Introduction of (Zhou et al., 2024). The 2nd paragraph of Introduction (Lines 47-64) appears to be paraphrased from the third paragraph of the Introduction of (Zhou et al., 2024).\\n\\n**Our Reponse:** Essentially Comment 3 and Comment 4 are referring to the same blocks of paragrahphs, so we address them together.\", \"we_want_to_point_out_the_following_aspects\": \"**1) Clear Difference in Strutural Positioning and Wording:** In this paper, the paragraph is more narrative, introduced RL in a broad sense and moving to specific examples; Whereas in (Zhou et al. (2024)), the referred two paragraphs are more formal, which details the squential steps of RL process and transitioning the need for MORL.\\n\\n**2) Clear Difference in Level of Details:** Even though the passage in this paper is concise and general, with newer references for example generative AI motivated literature (Franceschelli & Musolesi, 2024); Whereas (Zhou et al. (2024)) goes into more details and (lack of) rigours of MORL, with details of RL.\\n\\n**3) Common Knowledge in the Field:** Many of the terms and ideas are established knowledge in RL. In order for our paper to be self-contained, it is necessary to restate these backgrounds to motivated our problem. We believe that **no** paper possesses exclusive rights to these \\\"general knowledge\\\" in RL, and our works strictly follows the rules of fair use.\\n\\n________________\\n\\n>**Your comment 5:** The paragraphs about the \\u201cKey Contributions\\u201d (Lines 97-124) appear to directly follow the \\u201cMain Contributions\\u201d of the Introduction of (Zhou et al., 2024).\\n\\n**Our Response:** Here, we would like to compare \\\"Key Contributions\\\" in this paper with those in (Zhou et al., 2024) bullet-by-bullet:\\n\\n* In the first bullet of our \\\"Key Contributions,\\\"\\\" we mentioned the impacts of the weight/exploration vector $\\\\mathbf{p}$ on the sample complexity. More importantly, we proposed a completely different WC-MOAC algorithm, which systematically explores Pareto stionarity front assisted by the $\\\\mathbf{p}$-vectors. These are clearly **not in** (Zhou et al.(2024)).\\n\\n* The second bullet in our \\\"Key Contributions\\\" addresses, even with such weight vector $\\\\mathbf{p}$, it achieves sample complexity that is independent of $M$, where $M$ is the number of objectives.\\n\\n* Similarly, third bullet in our \\\"Key Contributions\\\" showed the benefit of momentum approach in the weighted-Chebyshev Pareto stationarity exploration.\\n\\n* Lastly, our empirical results strongly suggest the nature of exploration Pareto stationarity points by looking at Figure 1, which does *not* have any counterpart in (Zhou et al. (2024)). The key takeways from this paper is very different from the that of (Zhou et al. (2024)).\"}", "{\"summary\": \"The paper tackles the multiobjective reinforcement learning problem. It builds on top of the actor-critic algorithm using multi-gradient\\ndescent algorithm (MGDA), named multiobjective actor-critic (MOAC), designed by Zhou et al. (2024) and leverages the framework introduced by Momma et al. (2022) to optimize a multiobjective weighted by a preference vector, denoted by p. The difference from Zhou et al. (2024) is the introduction of preference weighting, which allows exploration of the Pareto Stationary front solutions. The proposed WC-MOAC algorithm maintains the sample complexity $\\\\tilde{O}(\\\\epsilon^{-2})$, shown in Zhou et al. (2024), after the addition of the preference weighting. Meanwhile, the algorithm is tested on a large-scale real-world dataset from the recommendation logs of the short video streaming mobile app Kuaisho.\\n\\nZhou, T., Hairi, F. N. U., Yang, H., Liu, J., Tong, T., Yang, F., ... & Gao, Y. (2024). Finite-time convergence and sample complexity of actor-critic multi-objective reinforcement learning. ICML.\\n\\nMomma, M., Dong, C., & Liu, J. (2022, June). A multi-objective/multi-task learning framework induced by pareto stationarity. In International Conference on Machine Learning (pp. 15895-15907). PMLR.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"The methodology and the algorithm are clearly stated. The assumptions and the main theorem are well presented.\", \"weaknesses\": \"The contribution can be stated more directly. For example, the abstract mentions that the algorithm fills the gap \\\"to systematically explore the Pareto-stationary solutions\\\". It would be more apparent if the paper explained why it is needed to explore these solutions and how the exploration is achieved.\\n\\nThe paper overstates the contribution without a clear acknowledgement of previous work. The paper should introduce the work by Zhou et al. (2024) with more details, and it is also necessary to make a comparison to clarify the novelty of the work. Some examples of overstated contributions examples are listed below. (1) It is claimed in the abstract that \\\"the performance of our WC-MOAC algorithm significantly outperforms other baseline MORL approaches.\\\" However, the performance drops from the behaviour clone baseline for two objectives.\\u00a0 (2) Line 100 states, \\\"Collectively, our results provide the first building block toward a theoretical foundation for MORL.\\\" Zhou et al. (2024) gave a sample complexity bound of the same order.\\u00a0 (3) Line 113 states, \\\"To mitigate the cumulative systematic bias injected from the WC-scalarization weight direction and finite-length state-action trajectories, we propose a momentum-based mechanism in WC-MOAC.\\\" What is the difference in the mechanism from Zhou et al. (2024)?\\u00a0\\n\\nThe clarity and correctness of writing can also be improved. Please refer to the question section for details.\\u00a0\\n\\nThe experiment design is not fully convincing, and the result is hard to understand. Please refer to the question section for details. More experiments can be designed to show the advantages of using the preference.\", \"questions\": \"1. Line 113 states, \\u201c To mitigate the cumulative systematic bias injected from the WC-scalarization weight direction and finite-length state-action trajectories, we propose a momentum-based mechanism.\\u201d What is this systematic bias? Also, why does this momentum mechanism remove the dependence on the number of objectives M?\\n\\n2. Where is Lemma 1 cited from? In Qiu et al. (2024) Proposition 4.2, they state that a stochastic policy has to maximize linearized scalarization for all weight vectors p to be a weakly Pareto optimal policy. However, Lemma 1 of the paper only requires maximizing the infinity norm of the scalarization for some weight vector.\\n\\n3. For the experiment in Table 1, how is the preference vector chosen?\\u00a0\\n\\n4. For the experiment in Figure 1, would\\u00a0TSCA work better by switching the main objective? More experiment designs to show the necessity of preference would be better.\\n\\n5. The paper focuses on the finite-time convergence results for multi-objective actor-critic. What is the advantage of using actor-critic algorithms?\\n\\nQiu, S., Zhang, D., Yang, R., Lyu, B., & Zhang, T. (2024). Traversing pareto optimal policies: Provably efficient multi-objective reinforcement learning.\\u00a0arXiv preprint arXiv:2407.17466.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"I sincerely thank the authors for the detailed response and for answering my technical questions. Below let us focus more on the Comments 1-8.\\n\\n- **Comment 1 (regarding the pseudo code of WC-MOAC and MOAC), Comment 2 (regarding the theoretical results), and Comments 3-4 (regarding the introduction)**: I can understand that both algorithms share the actor-critic architecture, and the two papers are both focused on the convergence to the stationarity. However, from a reviewer\\u2019s viewpoint, given the high similarity (in terms of design, paper structure, and even wording), it appears almost impossible to argue that this paper is completed independently from [Zhou et al., 2024] or the two are just concurrent works. Moreover, as far as I understand, such similarity in both structure and wording shall be viewed as paraphrasing, which is not allowed if there is no due credit or acknowledgement given to the prior works (i.e., [Zhou et al., 2024]). Furthermore, as mentioned by Reviewer JPv7, without a proper comparison with or mention of [Zhou et al., 2024], the readers can be easily misguided by the overclaimed contributions. \\n\\n- **Comment 5 (regarding the Key Contributions)**: Just like the comment above, my concern about paraphrasing in this part still remains unsolved. Moreover, apart from the structure and wording, the novelty and contributions can be overestimated without properly mentioning [Zhou et al., 2024]. Specifically:\\n\\n> In the first bullet of our \\\"Key Contributions,\\\"\\\" we mentioned the impacts of the weight/exploration vector on the sample complexity. More importantly, we proposed a completely different WC-MOAC algorithm, which systematically explores Pareto stionarity front assisted by the $\\\\mathbf{p}$-vectors. These are clearly not in (Zhou et al.(2024)).\\n\\nBased on the current writing of the paper, the readers can get easily misled that WC-MOAC is a totally new algorithm in the MORL literature. However, under a comparison with MOAC, WC-MOAC can indeed appear somewhat incremental.\\n \\n> The second bullet in our \\\"Key Contributions\\\" addresses, even with such weight vector $\\\\mathbf{p}$, it achieves sample complexity that is independent of $M$, where $M$ is the number of objectives.\\n\\nThis is also one key feature that has been shown by MOAC. Then, this would not be a convincingly new feature for people to choose WC-MOAC over MOAC.\\n\\n> Similarly, the third bullet in our \\\"Key Contributions\\\" showed the benefit of the momentum approach in the weighted-Chebyshev Pareto stationarity exploration.\\n\\nAs far as I know, the momentum-based design has already been proposed by MOAC (cf. Equation (10) in [Zhou et al., 2024]), and its benefit has been shown by the MOAC paper. \\n\\n> Lastly, our empirical results strongly suggest the nature of exploration Pareto stationarity points by looking at Figure 1, which does not have any counterpart in (Zhou et al. (2024)). The key takeaways from this paper is very different from that of (Zhou et al. (2024)).\\n\\nI can understand that the empirical results are shown to corroborate the nature of exploration of Pareto stationary points. That being said, if one directly compares the results of MC-MOAC and those of the MOAC in [Zhou et al., 2024] in terms of Like, Comment, Dislike, and WatchTime, the differences appear rather not significant. It is not immediately clear to me why the empirical results (cf. Table 1 and Figure 1) clearly supports the claim. While I understand in general that the works published within 3 months of submission are not necessarily needed in the comparison, a comparison with MOAC can serve as a really helpful ablation study and hence nicely support the claim of the paper. Please correct me if I missed anything. \\n\\n\\n- **Comments 6-8 (regarding the writing of the problem formulation and preliminaries)**: I understand that MDP and MOMDP are standard problem formulations in RL. However, there are various ways to describe the same setting. As far as I know, it is a standard practice to describe the problem formulation that tailors to the need of each specific paper and hence shall be written in the authors\\u2019 own words. Again, as mentioned above, this similarity in both structure and wording shall be viewed as paraphrasing and is not allowed if there is no due credit or acknowledgement given to the prior works. Therefore, I would like to emphasize that \\u201cusing the similar formulations\\u201d is not the issue, and the high similarity in terms of structure and wording is where the concerns arise. I hope that the authors can understand my concerns.\"}", "{\"comment\": \"Thank you for your detailed responses and additional experiments. However, I encourage the authors to address the integrity issues raised by Reviewer P7fw and similar concerns of Reviewer JPv7 to ensure a more comprehensive and novel submission.\"}", "{\"title\": \"Reponse to Reviewer P2zK's Comments:\", \"comment\": \">**Your Comment 1 (Also Weakness 1):** While the paper provides a comprehensive theoretical foundation, certain methodological aspects could be clarified further. For example, the integration of the multi-gradient descent update, which computes a dynamic weighting vector $\\\\lambda_t$ that balances exploration with convergence, could benefit from a more detailed discussion on its rationale and practical implementation steps.\\n\\n**Our Response:** Thanks for your comments. In terms of computing weight vector $\\\\lambda_t$, it is obtained from solving Eq. (10), whose solution $\\\\hat{\\\\lambda}^{*}_t$ is then used in the momentum computation in Eq.(11). In Eq.(10), the $\\\\lambda_t$-solution balances two aspects: \\n\\n1) the first term corresponds to multiple-gradient descent approach, which is an approach to ensure achieving Pareto stationarity upon convergence. \\n\\n2) The second term $\\\\lambda^{\\\\top}(\\\\mathbf{P}\\\\odot(\\\\mathbf{J}^{*}_{ub}-\\\\mathbf{J}(\\\\theta)))$ corresponds to the weighted-Chebyshev scarlization formulation, which induces Pareto stationarity front exploration as explained in the paper. \\n\\n3) The hypber-parameter $u>0$ is used to balance the trade-off between i) achieving a Pareto stationary solution and ii) systematically exploring the Pareto stationarity front. \\n\\n________________\\n>**Your Comment 2 (Also Weakness 2):** Additionally, the empirical evaluation is limited to a single dataset, the KuaiRand offline dataset, which raises questions about the algorithm's generalizability. Expanding the experimental analysis to include diverse datasets or multi-objective environments would provide deeper insights into the algorithm's robustness across varied applications.\\n\\n**Our Response:** Thank you for your suggestions. In this rebuttal period, we will provide more experimental settings. We are still working hard on conducting further experiments and will provide you an upate as soon as we are finished. Unfortunately, adding more datasets during this rebuttal period is quite challenging due to i) the limited amount of time and ii) the KuaiRand is the only suitable large-scale dataset for MORL that we can find at the current stage. We hope these resource limitations will not negatively affect the reviewer's evaluation toward our work.\\n\\n________________\\n\\n\\n>**Your Comment 3 (Also Question 1):** The paper would benefit from further clarification on several key points. First, could the authors discuss the impact of the preference vector $\\\\mathbf{p}$ on the algorithm's performance, especially in environments with highly non-convex reward landscapes? \\n\\n**Our Response:** We thank the reviewer for this insightful question. For settings with non-convex accumulated reward landscapes (i.e., $J^{i}(\\\\theta)$ is nonconvex in variable $\\\\theta$ for some reward function $i\\\\in [M]$), our algorithm converges to a Pareto stationary point for *any* given weight vector $\\\\mathbf{p}$ suggested by Theorem 4. Note that Pareto stationarity is a necessary condition for Pareto optimality. Similar to single-objective non-convex optimization for machine learning problems, it is often acceptable to find a Pareto stationary solution in the non-convex multi-objective settings.\\n\\n________________\\n>**Your Comment 4 (Also Question 2):** As the convergence rate depends on $p_{\\\\min}$, understanding the selection of $p$ would be useful for real-world applications. \\n\\n**Our Response:** Thanks for your comments and questions. Our Theorem 4 suggests that the smaller miminum enntry in vector $\\\\mathbf{p}$ (i.e. smaller $p_{\\\\min}$-value), the slower the convergence is. This implies that it might take more iterations to explore the Pareto stationarity front with a smaller $p_{\\\\min}$-value. \\nOn the other hand, in order to systematically explore the Pareto stationary solutions, the user can vary and enumerate the vector $\\\\mathbf{p}$ in the $M$-dimensional standard simplex, which is then taken by Algorithm 1 as input. For example, one possible approach will be using a grid search for trying different $\\\\mathbf{p}$ vectors. \\n\\n________________\\n>**Your Comment 5 (Also Question 3):** Additionally, what limitations, if any, might arise in adapting the proposed algorithm to online settings, given that the study's empirical evaluation is restricted to offline data? \\n\\n**Our Response:** In fact, the proposed Algorithm 1 is an online algorithm. Since KuaiShou dataset used in the empirical studies is an offline dataset, we have adapted our algorithm to the offline MORL setting. In principle, Algorithm 1 should perform better in online settings due to the fact that, in online setting, the algrotihm can sample more diverse data to evaluate and improve upon the current policy. In contrast, in offline MORL, the algorithm needs to work on the online dataset collected by a behavior policy, which may only have a limtied coverage.\\n\\n________________\"}", "{\"summary\": \"This paper investigates multi-objective reinforcement learning (MORL) using a weighted-Chebyshev multi-objective actor-critic (WC-MOAC) framework. The authors implement a multi-temporal-difference (TD) learning method for the critic and apply a weighted-Chebyshev multi-gradient-descent algorithm (MGDA) for the policy gradient in the actor. The algorithm is shown to achieve a sample complexity of $\\\\mathcal{O}(\\\\epsilon^{-2})$ to reach an $\\\\epsilon$-Pareto stationary point. The authors also provide empirical validation using a real-world dataset.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"4\", \"strengths\": \"The paper is well written with clear motivation and explanation. The authors\\u2019 theoretical analysis provides an important contribution to MORL. I would suggest the paper be accepted after minor revision.\", \"weaknesses\": \"Some of the details are not clear enough and can be improved. I have put them in Questions.\", \"questions\": \"1.\\tPage 5 eq (2). On the left of \\u201c:=\\u201d is a function of x, while on the right you have a scalar. Please make the definition consistent.\\n2.\\tPage 5 Lemma 1. You do not provide a proof for Lemma 1. Please reference the exact theorem or proposition from the cited work that supports Lemma 1.\\n3.\\tLemma 1 states existence of the vector p, which is directly related to a weak Pareto-optimal. However, it is not clear how $p$ is determined in practice within the algorithm. Does Theorem 4 hold for arbitrary p or for the specific p provided in Lemma 1? It would be helpful to include a more detailed explanation of p\\u2019s role in the algorithm and whether it influences the theoretical guarantees.\\n4.\\tPage 7. It will be clearer if you specify what arguments you are optimizing. In eq (7), could you clarify whether the optimization is over $\\\\lambda$, or both $\\\\lambda$ and $\\\\theta$? It appears from my understanding that eq (9) optimizes over both $\\\\lambda$ and $\\\\theta$, whereas eq (10) involves only $\\\\lambda$.\\n5.\\tPage 7 line 340, \\u201cgradient\\u201d should be \\u201cJacobian\\u201d.\\n6.\\tPage 8 Def 3 line 419. $\\\\lambda$ is already given. Why do you have $\\\\min_{\\\\lambda}$?\\n7.\\tPage 9 Theorem 4. Some of the quantities, such as $\\\\lambda_A$, $L_J$, $R_w$, $\\\\bf{\\\\gamma}$ are from the appendix, it is better to add some explanation for these quantities in the statement of the theorem.\\n8.\\tPage 9 line 467. You mention that state-of-the-art sample complexity for single-objective RL is $\\\\mathcal{O}(\\\\epsilon^{-2})$ bu Xu et al. And your Corollary 5 achieves the same complexity for MORL, which is a great achievement. Actually, under some special structure like linear quadratic problem, the one can show a sample complexity of $\\\\mathcal{O}(\\\\epsilon^{-1})$, as proved by Zhou and Lu in \\u201cSingle Timescale Actor-Critic Method to Solve the Linear Quadratic Regulator with Convergence Guarantees\\u201d. Please consider adding it as a remark to enhance completeness of the paper.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Follow-up Response to Reviewer P2zK\", \"comment\": \">**Your Comment:** I encourage the authors to address the integrity issues raised by Reviewer P7fw and similar concerns of Reviewer JPv7 to ensure a more comprehensive and novel submission.\\n\\n**Our Response:** Thanks for your comment. Yes, we have already carefully responsed to all the comments to Reviewer P7fw and Reviewer JPv7. As the integrity accusation is a very serious issue, we have taken serious effort to address their comments. We strongly encourage reviewer P2zK to go through our point-to-point responses to gain an independent assessment. Meanwhile, we provide the following summary to the above two reviewers' response in here:\\n\\nThe primary concern only revolves around perceived *cosmetic* similarities between our current submission and (Zhou et al., 2024). Specifically, we want to empahsize the **critical differences** between our work and (Zhou et al., 2024) in the following aspects:\\n\\n**1) Differences in Research Goals:** Our research focuses on systematically exploring the Pareto stationarity front in a multi-objective problem using a weighted-Chebyshev formulation. In comparison, the goal of (Zhou et al., 2024) is to achieve only a Pareto stationary solution rather than exploring the full Pareto stationarity front.\\n \\n**2) Differences in Algorithmic Designs:** In this paper, we have integrated a weighted-Chebyshev technique with the multi-gradient descent approach (MGDA) for multi-objective optimization, a technique entirely absent in (Zhou et al., 2024). Our paper devotes Section 4.2 (about 1.5 pages) to describing this approach in detail, illustrating how we use the weighted-Chebyshev method to achieve systematic exploration. This technique leverages an exploration/weight vector $\\\\mathbf{p}$, extensively explained in our derivations. In contrast, (Zhou et al., 2024) is based only on a direct adaptation from MGDA. As a result, there is *nothing* in (Zhou et al., 2024) that corresponds to Section 4.2 in this paper.\\n\\n**3) New and Unique Theoretical Contributions:** Our Theorem 4 and Corollary 5 highlight the effects of the weight vector $\\\\mathbf{p}$, particularly through its minimum entry $p_{\\\\min}$, on convergence rate to Pareto stationarity and corresponding sample complexities for each exploration. While the convergence rate results have similar convergence rate order, the coefficients in Theorem 4 are distinctly different from those in (Zhou et al., 2024). Again, merely having the same convergence rates does **not** mean the algorithms and the analysis are the same. For example, both the gradient descent (GD) method and the SVRG stochastic gradient method have the same $O(1/T)$. Clearly, calling SVRG and GD the same method is absurd. Therefore, we don't believe any sensible researchers would think Theorem 4 and Corollary 5 are copying from (Zhou et al., 2024).\\n\\n**4) Differences in Experimental Results:** In Figure 1 and Figure 2, we have presented the Pareto footprint comparisons between our proposed algorithm and the conventional linear scalarization approach (i.e., comparing the Pareto front explored by our method via varying exploration/weight vector $\\\\mathbf{p}$ with that achieved by the conventional linear scalarization approach). In contrast, there is no such experimental comparistive studies in (Zhou et al., 2024).\\n\\nGiven the serious nature of the accusations, we again strongly encourage reviewer P2zK to go through the point-to-point responses to both reviewers to gain an independent assessment. **We are also happy to answer any further questions Reviewer P2zK may have in this aspect**. Thanks!\"}", "{\"title\": \"Response to Reviewer n72g's Comments\", \"comment\": \">**Your Comment 1:** Page 5 eq (2). On the left of \\u201c:=\\u201d is a function of x, while on the right you have a scalar. Please make the definition consistent.\\n\\n**Our Response:** Thank you for pointing out this typo. We will revise the definition as follows in the revision:\\n$$WC_{\\\\mathbf{p}}(\\\\mathbf{F}(\\\\cdot)) := \\\\min_{\\\\mathbf{x}}\\\\max_{i\\\\in[M]}\\\\{ p_i f_i(\\\\mathbf{x})\\\\}= \\\\min_{\\\\mathbf{x}} \\\\\\\\| \\\\mathbf{p} \\\\odot \\\\mathbf{F}(\\\\mathbf{x}) \\\\\\\\|_{\\\\infty}.$$\\n\\n________________\\n>**Your Comment 2:** Page 5 Lemma 1. You do not provide a proof for Lemma 1. Please reference the exact theorem or proposition from the cited work that supports Lemma 1.\\n\\n**Our Response:** Thanks for your comments. Lemma 1 is cited from Proposition 4.7 and its proof on Page 42 in (Qiu et al. (2024)). Specifically, on Page 42 of the proof of Proposition 4.7, Part 1) and Part 2) collectively provide an \\\"if-and-only-if\\\" statement in Lemma 1 of this paper. We will provide more clarification on this in our revision.\\n\\nQiu, S., Zhang, D., Yang, R., Lyu, B., & Zhang, T. (2024). Traversing pareto optimal policies: Provably efficient multi-objective reinforcement learning. arXiv preprint arXiv:2407.17466.\\n________________\\n>**Your Comment 3:** Lemma 1 states existence of the vector $\\\\mathbf{p}$, which is directly related to a weak Pareto-optimal. However, it is not clear how $\\\\mathbf{p}$ is determined in practice within the algorithm. Does Theorem 4 hold for arbitrary p or for the specific p provided in Lemma 1? It would be helpful to include a more detailed explanation of p\\u2019s role in the algorithm and whether it influences the theoretical guarantees.\\n\\n**Our Response:** Thanks for your question. Again, we would like to emphasize that Lemma 1 is an \\\"if-and-only-if\\\" statement, which implies that there is a one-to-one mapping between a WC-solution and a Pareto-optimal solution and further suggest a systematical way to explore the Pareto-front (assuming the WC-scalaization problem can be solved to optimality). As a result, when applying Algorithm 1, one does *not* need to carefully pick the vector $\\\\mathbf{p}$. Rather, the decision-maker is supposed to vary and enumerate the vector $\\\\mathbf{p}$ in the $M$-dimensional standard simplex, which is then taken by Algorithm 1 as input to systematically explore the Pareto front. Also, Theorem 4 holds for arbitrary $\\\\mathbf{p}$ whose entries are strictly positive (but one would typically use $\\\\mathbf{p}$ in the $M$-dimensional standard simplex to avoid arbitrariness in scaling). For those $\\\\mathbf{p}$-vectors with some entries being 0, one can simply discard those zero entries and reformulate the problem in a multi-objective RL problem with lower dimension, which only retains those non-zero entries in the original $\\\\mathbf{p}$.\\n________________\\n>**Your Comment 4:** Page 7. It will be clearer if you specify what arguments you are optimizing. In eq (7), could you clarify whether the optimization is over $\\\\mathbf{\\\\lambda}$, or both $\\\\lambda$ and $\\\\mathbf{\\\\theta}$? It appears from my understanding that eq (9) optimizes over both $\\\\lambda$ and $\\\\theta$, whereas eq (10) involves only $\\\\lambda$.\\n\\n**Our Response:** Thanks for your question. In Eq.(7), we are optimizing over $\\\\lambda$ for any given $\\\\theta$, where $\\\\pi_{\\\\theta}$ denotes the current policy. Similarly, in Eq.(9), we are also optimizing over $\\\\lambda$ for any given $\\\\theta$. The rationale is that, given the current policy evaluation from the critic component, we want to leverage and integrate multi-policy-gradient descent (first term in Eq.(10)) and the weighted-Chebyshev scalarization (second term in Eq.(10)) to guide the policy update direction to achieve a balance between i) converging to a Pareto stationary solution and 2) systematically exploring the Pareto stationarity front.\\n\\n________________\\n>**Your Comment 5:** Page 7 line 340, \\u201cgradient\\u201d should be \\u201cJacobian\\u201d.\\n\\n**Our Response:** Thanks for your question. Here, matrix $\\\\mathbf{G}$ we are referring to is \\n$$\\\\mathbf{G}=-[\\\\nabla_{\\\\theta} J^{1}(\\\\theta), \\\\cdots, \\\\nabla_{\\\\theta} J^{i}(\\\\theta), \\\\cdots, \\\\nabla_{\\\\theta} J^{M}(\\\\theta)].$$\\nIn other words, each column $i\\\\in [M]$ is the gradient of $J_{\\\\text{ub}}^{i,\\\\*}-J^{i}(\\\\theta)$ with respect to $\\\\mathbf{\\\\theta}$, where $J_{\\\\text{ub}}^{i,\\\\*}$ is an upper bound constant. Hence, $\\\\mathbf{G}$ is related to but not exactly the Jacobian (more precisely, it is negative Jacobian matrix). We will further clarify this in our revision.\\n\\n________________\\n\\n>**Your Comment 6:** Page 8 Def 3 line 419. $\\\\lambda$ is already given. Why do you have $\\\\min_{\\\\lambda}$?\\n\\n**Our Response:** Thank you for pointing out this typo out. We will revise it in revision and change it to $\\\\\\\\| \\\\nabla_{\\\\theta} \\\\mathbf{J}_{\\\\theta}(\\\\theta) \\\\lambda \\\\\\\\|_2^2 \\\\leq \\\\epsilon$.\\n\\n________________\"}", "{\"title\": \"Response to Reviewer P7fw\", \"comment\": \"We appreciate the reviewer's constructive comments and valuable insights. The detailed point-by-point responses are as follows:\\n\\n-------------\\n\\n> **Your Comment 1:** The pseudo code of WC-MOAC are almost the same (almost verbatim) as those of the MOAC algorithm (cf. Algorithms 1 and 2 in (Zhou et al., 2024)).\\n\\n**Our Response:** In this paper and (Zhou et al., 2024), the algorithms both utilized actor-critic framework, which is a widely used and standard RL framework. If merely following the actor-critic framework violates academic integridty, then we are not sure how the reviwer would view the thousands of papers published in the RL literatrue that adopted the actor-critic framework. Further, in terms of the algorithms, the following are the major differences between our work and (Zhou et al., 2024): \\n1) Algorithm 1 in this paper is presented as a single-oracle, which includes both Critic and Actor components; whereas in (Zhou et al., 2024), the critic component is presented as a subroutine, which is called in its main Algorithm 2.\\nIn this paper, we have leveraged a weight/exploration vector $\\\\mathbf{p}$ to explore the Pareto stationarirty front (a necessary condition for Pareto Front). This is reflected by two major aspects (see next two points), which are not present in (Zhou et al., 2024). \\n\\n2) In the second column of Line 395, $\\\\hat{\\\\mathbf{\\\\lambda}}^{*}_t$ is the solution of Eq. (10), which is an objective function that carefully balances the Weighted-Chebyshev exploration controlled by the $\\\\mathbf{p}$-weight vector and the mutpile-gradient descent term. If one were to check Eq. (10) in this paper and Eq. (9) in (Zhou et al. (2014)), it is apparent that they are completely **different**.\\n\\n3) Leveraging a weight vector $\\\\mathbf{p}$, the update for $\\\\mathbf{g_t}=\\\\mathbf{G_t}(\\\\mathbf{p}\\\\odot \\\\mathbf{\\\\lambda_t})$ in this paper, which enables the exploration of the Pareto front (also see Figure 1 for the exploration effect). In contrast, there is **no such Hadamard multiplied $\\\\mathbf{p}$ term** for the policy update in (Zhou et al., 2024), which is in the form of $\\\\mathbf{g}_{t}=\\\\mathbf{G}_t(\\\\mathbf{p}\\\\odot \\\\mathbf{\\\\lambda}_t)$. We also note that the solution of Eq. (10) yields a completely different quantity from that of Eq. (9) in (Zhou et al. (2014)). As a reuslt of such a difference, the algorithm in (Zhou et al. (2024)) can only guarantee convergence to a Pareto stationary point, but lacks the capability of systematic Pareto stationarity exploration.\\n\\n4) In the actor component of this paper, we require the computation of score function $\\\\mathbf{\\\\psi}_{t,l}$ by Eq.(12) (see approximately in Line 388 in the second column). This is clearly **not** in the Pseudo code of (Zhou et al., 2024).\\n\\n5) In Algorithm 1, it is required that $w^{i}_k=w^{i}_t$ for synchronization of the critic parameters (see the first column in Line 398-399). Note that this is **absent** in (Zhou et al. 2024).\\n\\n6) Among the input of Algorithm 1 (Lines 380-381), our algorithm requires a weighted/exploration vector $\\\\mathbf{p}$ in addition to standard input parameters for actor-critic with linear approximations.\\n\\nAgain, we want to emphasize that the actor-critic algorithmic framework in the RL literature will share a lot of structural similarities, including (Zhou et al. (2014)). This also indicates the popularity of the actor-critic framework for solving RL problems. The similar notations are due to the convention in the literature for the ease of understanding.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"title\": \"Response to Reviewer qYKr's Comments (Continued)\", \"comment\": \">**Your Comment 3 (Also Question 2):** In terms of the proposed method, what is the major novelty compared to previous methods in MOO literature? The reviewer understands the discussions already presented in the paper, but is just wondering whether the proposed method is replacing the gradients in MOO with policy gradients given by different critics.\\n\\n**Our Response:** Yes, as we mentioned in our response to your Comment 1, the adaptation of the MOO techniques to the MORL setting calls for many novel algorithmic designs due to the uniques challenges in MORL. In addition to the momentum-based technique in our reponse to your Comment 1, in the actor component of our WC-MOAC approach, the policy gradient is updated through a judious integration of the multi-gradient descent and *weighted-Chebyshev* approahces. Moreover, this new algorithmic design for MORL necessitates new finite-time convergence proof in our theoretical Pareto stationary convergence analysis, which leads to the *new* insights and knowledge of the dependency on the weight vector $\\\\mathbf{p}$ in the WC-MOAC's Pareto stationary convergence. \\n\\n>**Your Comment 4 (Also Question 3):** It seems that the detailed experiment description is not specified even in the appendix. What is the horizon of this game? What is the size of the state set?\\n\\n**Our Response:** Thanks for your question. In this work, we consider infinite time horizon. The state size is 1218, which is indicated in Table 2 in the Appendix.\\n________________\\n\\n>**Your Comment 5 (Also Question 4):** Is the weight vector $\\\\mathbf{p}$ predefined or learned?\\n\\n**Our Response:** Thanks for your question. The weight vector $\\\\mathbf{p}$ is a given input in each run of Algorithm 1. To further understand the role of $\\\\mathbf{p}$, note first that Lemma 1 implies that there is a one-to-one mapping between a WC-solution and a Pareto-optimal solution and further suggest a systematical way to explore the Pareto-front (assuming the WC-scalaization problem can be solved to optimality) by varying $\\\\mathbf{p}$ as input to Algorithm 1. Thus, when applying Algorithm 1, the decision-maker is supposed to vary and enumerate all possible vectors $\\\\mathbf{p}$ in the $M$-dimensional standard simplex. We hope this clarifies the role of $\\\\mathbf{p}$ in our WC-MOAC approach.\\n________________\"}", "{\"title\": \"Response to Reviewer JPv7's comments (Continued)\", \"comment\": \">**Your Comment 4 (Also Weakness 4):** Line 113 states, \\\"To mitigate the cumulative systematic bias injected from the WC-scalarization weight direction and finite-length state-action trajectories, we propose a momentum-based mechanism in WC-MOAC.\\\" What is the difference in the mechanism from Zhou et al. (2024)?\\n\\n**Our Response:** Indeed, momenteum approach-wise is the same approach in this paper and (Zhou et al. (2024)). However, WC inspired Eq.(10) and $\\\\mathbf{g_t}$ update have incurred significant differences in the following senses.\\nThe novelty in the analysis stems from involving Hadamard product by formulating the problem in a weighted-Chebyshev problem, where as in (Zhou et al (2024)), there's no such operand. In contrast to pure MGDA approach, where the $\\\\mathbf{\\\\lambda_t}$ is used to weigh multiple gradients from objectives, incorporating WC aspect, due to aforementioned Hadamard product, now requires the analysis to construct a pseudo weight $\\\\mathbf{q_t}:=\\\\frac{\\\\mathbf{\\\\lambda_t}\\\\odot \\\\mathbf{p}}{\\\\langle \\\\mathbf{\\\\lambda_t}, \\\\mathbf{p}\\\\rangle}$ and utilize properties of such pseudo weight. Please see Line 907-970, 1160-1210, where we carefully handled such pseudo weight. As a result of this analysis, we observed that $p_{\\\\min}$ is a cruicial quantity in characterizing the convergence to stationarity. As one can easily observe that coefficients in front of the cruicial terms are significantly different from (Zhou et al. (2024)). For example, the cofficient in front of $1/T$ term is $16L_Jr_{\\\\max}(1+\\\\frac{2}{p_{\\\\min}^{2}}\\\\sum_{t=1}^{T}\\\\eta_t)$ in this work and $18L_Jr_{\\\\max}(1+2\\\\sum_{t=1}^{T}\\\\eta_t)$ in (Zhou et al. (2024)). \\n\\nIn fact, the WC-approach proposed in this paper is a more general approach than that of (Zhou et al. (2024)). When $u=0$ and $\\\\mathbf{p}=(1,\\\\cdots,1)^{\\\\top}$ (for simplicity, ignoring the scaling), the results in our work imply those in (Zhou et al. (2024)). To see this, Eq.(10) reduces to the $\\\\min_{\\\\lambda}||\\\\mathbf{K_p}\\\\lambda||^{2}$, which is equivalent to MGDA where the Quadratic programming is to solve $\\\\min_{\\\\lambda}||\\\\mathbf{G}_t\\\\mathbf{\\\\lambda}||^{2}$ where $\\\\mathbf{G}_t=[\\\\mathbf{g}^{i}_t,\\\\cdots, \\\\mathbf{g}^{M}_t]$ and $\\\\mathbf{g}^{i}_t$ is defined in the second column of Line 394. In algorithm 1, the update of $\\\\mathbf{g}_t$ will reduce to $\\\\mathbf{g}_t= \\\\mathbf{G}_t\\\\mathbf{\\\\lambda}_t$. \\n\\n________________\\n>**Your Comment 5 (Also Question 1):** Line 113 states, \\u201c To mitigate the cumulative systematic bias injected from the WC-scalarization weight direction and finite-length state-action trajectories, we propose a momentum-based mechanism.\\u201d What is this systematic bias? Also, why does this momentum mechanism remove the dependence on the number of objectives M?\\n\\n**Our Response:** Thank you for the insightful question. To see the advantage of momentum approch in obtaining an $M$-independent result, first observe the right-hand-side (RHS) of the inequality in Eq. (18), which consists of two terms. For the first term on the right-hand-side of the inequality in Eq. (18), it further goes through a telescoping operation (see Appendix C.2 Line 1173). The so-called systematic bias term is the summand term on the Left-hand-side of in the equation of Line 1173, i.e. $\\\\mathbb{E}[\\\\mathbf{q_t^{\\\\top}}\\\\mathbf{J}(\\\\mathbf{\\\\theta_{t+1}})]-\\\\mathbf{q_{t}}\\\\mathbf{J}(\\\\mathbf{\\\\theta_t})$. Here, the contribution of momentum-based $\\\\bf{\\\\lambda_t}$ as in Eq. (11) helps with providing bound on $\\\\mathbf{q_{t+1}}-\\\\mathbf{q_{t}}$ via Line 1187 and 1192. \\n\\n________________\\n>**Your Comment 6 (Also Question 2)** Where is Lemma 1 cited from? In Qiu et al. (2024) Proposition 4.2, they state that a stochastic policy has to maximize linearized scalarization for all weight vectors p to be a weakly Pareto optimal policy. However, Lemma 1 of the paper only requires maximizing the infinity norm of the scalarization for some weight vector.\\n\\n**Our Response:** Lemma 1 is cited from Proposition 4.7 and its proof within in (Qiu et al. (2024)). In Page 42 of proof for Proposition 4.7, part 1) and part 2) provide if and only if statement in Lemma 1 of this paper.\\n\\n________________\\n>**Your Comment 7 (Also Question 3):** For the experiment in Table 1, how is the preference vector chosen? \\n\\n**Our Response:** The particular $\\\\mathbf{p}$ vector we chose for the result in Table 1 is $(0.2,0.2,0.2,0,0.4)^{\\\\top}$. The rationale behind the weight choice is inpired by Kuaishou dataset and its application to emphasize on WatchTime with higher weight, and on the other hand ignoring dislike, essentially making it a 4-objective problem. Then for the remaining objectives, we chose uniform weights to highlight no particular preference within this subset of objectives.\\n\\n________________\"}", "{\"title\": \"Response to Reviewer JPv7's comments\", \"comment\": \">**Your Comment 1 (Also Weakness 1):** The contribution can be stated more directly. For example, the abstract mentions that the algorithm fills the gap \\\"to systematically explore the Pareto-stationary solutions\\\". It would be more apparent if the paper explained why it is needed to explore these solutions and how the exploration is achieved.\\n\\n**Our Response:** When decision-maker aims for optimizing in a multi-objective setting, at the beginning there's not always a clear priori on what the preference for a particular Pareto Optimal solution among the set of Pareto Front(PF). Considering such potential uncertainty in priori, exploring the PF has been an on-going effort on multi-objective problem settings [1]. On the other hand, once PF is entirely characterizes, it allows a Posteriori selection of a preferable Pareto solution, giving better insight to a decision-maker [2]. Towards this effort, we are proposing a WC-MOAC algorithm that enables systematically exploration of Pareto solution in MORL context, which is under-investigated.\\n\\n[1] Kristof Van Moffaert and Ann Now\\u00e9. Multi-objective reinforcement learning using sets of pareto dominating policies. The Journal of Machine Learning Research, 15(1):3483\\u20133512, 2014.\\n\\n[2] Simone Parisi, Matteo Pirotta, and Marcello Restelli. Multi-objective reinforcement learning through continuous pareto manifold approximation. Journal of Artificial Intelligence Research, 57:187\\u2013227, 2016.\\n\\n________________\\n\\n>**Your comment (Also Weakness 2):** The paper overstates the contribution without a clear acknowledgement of previous work. The paper should introduce the work by Zhou et al. (2024) with more details, and it is also necessary to make a comparison to clarify the novelty of the work.\\n\\n**Our Response:** To highlight the contribution of this work, we provide following comparisons with (Zhou et al. (2024)).\\n\\n**Objective Differences**: Our work focuses on systematically exploring the Pareto Stationary front using a weighted-Chebyshev formulation. This objective is distinctly different from that of (Zhou et al., 2024), which aims to find a Pareto stationary point rather than exploring the full Pareto stationary front.\\n \\n**Methodological Innovation**: We have incorporated a weighted-Chebyshev formulation, a technique which is entirely absent in (Zhou et al., 2024). Our paper devotes Section 4.2 to describing this approach in detail, illustrating how we use the weighted-Chebyshev method to achieve systematic exploration. This technique leverages an exploration/weight vector $\\\\mathbf{p}$, extensively explained in our derivations.\\n\\n**Unique Theoretical Contributions**: Our Theorem 4 and Corollary 5 highlight the effects of the weight vector $\\\\mathbf{p}$, particularly through its minimum entry, $p_{\\\\min}$, on convergence rate to Pareto stationarity and corresponding sample complexities for each exploration. While some order-wise results appear similar, the coefficients in Theorem 4 are different from those in (Zhou et al., 2024). \\n\\n**Empirical Evidence**: In Figure 1 of Experiment Section, we present the Pareto footprint, i.e., the Pareto fronts explored by our method via varying exploration/weight vector $\\\\mathbf{p}$ and a popular linear scalarization approach. In contrast, (Zhou et al., 2024) includes no such comparative analysis.\\n\\n________________\\n>**Your Comment 3(Also Weakness 3):** Line 100 states, \\\"Collectively, our results provide the first building block toward a theoretical foundation for MORL.\\\" Zhou et al. (2024) gave a sample complexity bound of the same order. \\n\\n**Our Response:** We are happy to remove this statement and acknowledge the contribution from the previous literature including (Zhou et al. (2024)). For the second point, we want to emphasize that proving the same sample complexity doesn't mean it's not a new result in the context of leveraging the weight-Chebyshev. In particular, reiterating the point in our response to your comment 2, while some order-wise results appear similar, the coefficients in Theorem 4 are different from those in (Zhou et al., 2024). \\n\\n________________\"}", "{\"title\": \"Response for Reviewer JPv7's Follow-up Comments\", \"comment\": \"Thank you very much for the prompt reply. Here are our point-to-point response for your follow-up comments.\\n>**Follow-up comment 1:**\\nThe main issue with the paper's organization is that others can strongly overestimate the contributions.\\nAs stated in Reviewer P2zK's summary, \\\"The WC-MOAC algorithm is designed to combine multi-temporal-difference (TD) learning in the critic phase with ... multi-gradient descent in the actor phase, effectively managing the complexities of non-convexity and scalability in multi-objective problems. The primary theoretical result of this work is a finite-time sample complexity bound, which is independent of the number of objectives ... This independence ... is a notable theoretical advance.\\\"\", \"reviewer_qykr02_stated\": \"\\\" Central to the discussion is the concept of (local) (weak) Pareto optimality, which serves as the foundational solution criterion for balancing multiple objectives. The proposed training architecture employs a single actor coupled with multiple critics, where each critic is specifically trained to approximate the value function corresponding to a distinct objective.\\\"\\nNone of these are your novel contributions, which Zhou and colleagues have introduced. The paper writing is strongly misleading.\\n\\n**Our Reseponse:** We agree with the reviewer that the convergence error bound in (Zhou et al.(2024)) in also independent of $M$, the number of objectives. However, we do want to emphasize that these seemingly similar $M$-independent results are for two **different** algorithms. Specifically, in (Zhou et al.(2024)), $M$-independence result is for a pure *MGDA-type approach* in their the actor component. In contrast, in our work, the $M$-independence result is established for a **hybrid weighted-Chebyshev MGDA** approach for the actor component. Consequently, the proof and theoretical analysis for our $M$-independence result is quite different from that in (Zhou et al. (2024)). Here, we want to outline how we overcome two *new and unique* challenges in our proof of the $M$-independence result for the hybrid weighted-Chebyshev MGDA approach, which stems from two terms on the right-hand-side (RHS) of the inequality in Eq. (18):\\n\\n1) For the first term on the RHS of the inequality in Eq. (18), we note that it will further go through a telescoping operation (see Appendix C.2 Line 1173). However, due to the WC-component $\\\\lambda^{\\\\top}(\\\\mathbf{p} \\\\odot \\\\left(\\\\mathbf{J_{\\\\text{ub}}}^{*}-\\\\mathbf{J} (\\\\mathbf{\\\\theta}) \\\\right))$ (cf. Problem (10)) of the actor update, it requires a *clever* construction of **pseudo weights** defined as $\\\\mathbf{q_{t}}:=\\\\frac{\\\\mathbf{p}\\\\odot \\\\lambda_t}{\\\\langle \\\\mathbf{p},\\\\lambda_t\\\\rangle}$, as the naive approach for MGDA doesn't apply here. Note that such a pseudo weight construction is *not* needed in (Zhou et al. (2024)). The analysis in this work further requires applying the properties of this pseudo weight. Please see our *proofs in Line 948-1008, 1197-1247*, where we carefully handled such a pseudo weight. As a result of this analysis, we are able to show that $p_{\\\\min}$ is a key quantity in affecting the convergence to Pareto stationarity with $M$-independence. \\n\\n2) For the second term on the RHS of the inequality in Eq. (18), it requires establishing the $\\\\mathbb{E}[\\\\\\\\|\\\\sum_{i=1}^{M}\\\\lambda^{i}(w_t^{i}-w^{i,\\\\*})\\\\\\\\|^{2}]\\\\le\\\\max_{i\\\\in[M]}\\\\mathbb{E}[\\\\\\\\|w_t^{i}-w^{i,\\\\*}\\\\\\\\|^{2}]$ for any $\\\\sum_{i=1}^{M}\\\\lambda^{i}=1$ and $\\\\lambda^{i}\\\\ge 0$ for all $i\\\\in[M]$. This **proof analysis from Line 1076- Line 1102**, which utlizes the independence of evaluations of objective $i$ and $j$ given trajectory for any $i\\\\neq j$ and $i,j\\\\in[M]$, are new and different from (Zhou et al. (2024)). \\n\\nRegarding Reviewer qYKr02's comments (i.e., *\\\"Central to the discussion is the concept of (local) (weak) Pareto optimality, which serves as the foundational solution criterion for balancing multiple objectives. The proposed training architecture employs a single actor coupled with multiple critics, where each critic is specifically trained to approximate the value function corresponding to a distinct objective.\\\"*), we want to clarify that the actor-critic (AC) architecture itself is indeed not our contribution. However, we want to point out that the hybrid weighted Chebyshev MGDA-type technique for dynamic $\\\\mathbf{\\\\lambda}$-weighting in the actor component of our WC-MOAC framework and its subsequent theoretical and empirical analysis are our **new** contributions. \\n\\nWith the above clarifications, we are happy to re-position our paper and tone down some of the contribution statements to avoid similar misunderstandings arising. We thank for the reviewer's comments that help us improve the quality of our paper. \\n______________\"}", "{\"title\": \"Response to Reviewer P7fw (Continued)\", \"comment\": \">**Your Comment 9:** The description about Assumption 3 and Lemma 3 in this paper (Lines 424-437) appears to exactly follow the Assumption 3 and Lemma of (Zhou et al., 2024).\\n\\n**Our Reponse:** The Assumption 3 and Lemma 3 are standard assumptions widely used in RL and optimization in general. For example, Assumption 3 is listed as Assumption 1 and Proposition 1 in (Xu et al. (2020)) and references within, Assumption 5 of (Qiu et al. (2021)). Lemma 3 is Assumption 2 in (Xu et al (2020)) and Assumption 1 in (Qiu et al. (2021)). In fact, Assumption 3 (Lipschitz smoothness of the value function) is one of the most basic assumptions that establish the convergece of most policy gradient approaches.\\n\\nQiu, S., Yang, Z., Ye, J., and Wang, Z. On finite-time convergence of actor-critic algorithm. IEEE Journal on Selected Areas in Information Theory, 2(2):652\\u2013664, 2021.\\n\\nTengyu Xu, Zhe Wang, and Yingbin Liang. Improving sample complexity bounds for (natural) actor-critic algorithms. arXiv preprint arXiv:2004.12956, 2020.\\n\\n________________\\n>**Your Comment 10 (Also Question 1):** One of my main technical concerns is the motivation for finding a Pareto-stationary policy (under the assumption that the state and action spaces are finite) in the specific context of MORL. Specifically, while it is indeed difficult to find the whole Pareto front in MORL, it is actually not hard to find one or some Pareto-optimal policies by adapting reducing MORL to single-objective RL and finding the convex coverage set (e.g., (Yang et al., 2019)). For example, based on (Chen and Magulari, 2022), one can use a policy-based method with off-policy TD learning (under linear function approximation) to find an epsilon-optimal solution for single-objective RL with a sample complexity of $\\\\mathcal{O}(\\\\epsilon^{-2})$.There are also several other recent works like (Lan 2021; Fatkhullin et al., 2023; Liu et al., 2020; Chen et al., 2022) that can find an epsilon-optimal policy with sample complexity guarantees. To adapt these results to MORL, one can use linear scalarization and thereby find one Pareto-optimal policy (specific to some preference vector). As a result, it remains not totally clear why it is theoretically appealing to design an algorithm for finding only a Pareto-stationary policy if we can already find Pareto-optimal policies (despite that Pareto-stationarity is indeed a widely adopted concept in the MOO literature).\\n\\n**Our Response:** It is indeed true that finding a Pareto optimal point is not hard and can be achieved via the linear scalarization approach. However, as stated in (Yang et al. (2019)), the CCS is only a subset of the Pareto frontier, containing the solutions on its outer convex boundary. Similarly, by Proposition 4.2 in (Qiu et al. (2024)), the linear scalarization (LS) approach cannot guarantee exploring the full Pareto front. When Pareto front is non-convex, it only provides limited exploration (more precisely, the convex hull of the Pareto front). \\n\\nMorever, from Proposition 4.7 and its proof on Page 42 in (Qiu et al. (2024)), the weighted-Chebyshev approach can guarantee exploring all Pareto-optimal solutions. This observation motivates us to design a weighted-Chebyshev-based Pareto front exploration approach for MOAC (i.e., WC-MOAC). Although our theoretical results only guarantees convergence to Pareto stationary points due to the fundamental intractability of the non-convex MORL setting, our empirical comparison results on the footprints of exploration for both LS in Figure 1(a) and weighted-Chebyshev approach in Figure 1(b) show that WC-based exploration indeed achieves a larger collection of Pareto-stationary solutions than that of the LS approach in Figure \\\\(c).\\n\\nShuang Qiu, Dake Zhang, Rui Yang, Boxiang Lyu, and Tong Zhang. Traversing pareto optimal policies: Provably efficient multi-objective reinforcement learning, 2024. URL https://arxiv.org/abs/2407.17466.\\n\\n________________\"}", "{\"summary\": \"This paper studies multi-objective reinforcement learning by extending existing multi-objective optimization techniques within the framework of actor-critic methodologies. Central to the discussion is the concept of (local) (weak) Pareto optimality, which serves as the foundational solution criterion for balancing multiple objectives. The proposed training architecture employs a single actor coupled with multiple critics, where each critic is specifically trained to approximate the value function corresponding to a distinct objective. During the training process, the actor is updated using a weighted gradient approach that integrates the estimated values from each critic. These weights are supposed to balance the influence of each objective, guiding the actor towards policies that achieve a desirable trade-off among the competing objectives. The authors rigorously demonstrate that their method is capable of converging to locally Pareto optimal policies. This convergence is guaranteed up to an error bound, which is linked to the accuracy of the critic's value function estimations.\", \"soundness\": \"3\", \"presentation\": \"4\", \"contribution\": \"2\", \"strengths\": \"1. The manuscript is exceptionally well-written, presenting complex concepts in a clear and accessible manner. It seems that the authors have meticulously structured the paper, which significantly enhances readability. Additionally, the discussion of related works is thorough, providing a comprehensive overview of existing approaches in multi-objective reinforcement learning and actor-critic methodologies.\\n\\nThe experimental component of the study appears to be conducted on a large-scale real-world short video platform. (However, the manuscript lacks detailed descriptions of the experimental setup, making it challenging to fully assess the validity the results. Clarification on the this concern is detailed in the section below.)\", \"weaknesses\": \"1. The major weakness might lie in the novelty of the proposed method. To the knowledge of the reviewer, Although the theoretical analysis partially answers the challenges mentioned in the introduction, the proposed method seems to largely based on (D\\u00e9sid\\u00e9ri, 2012) and (Momma et al., 2022), with the multi-objective optimization gradients replaced with policy gradients.\\n\\n2. See questions below.\", \"questions\": \"1. In Theorem 4, what is the different between |w*-w| and zeta? Intuitively, both these two terms are related to the error of value functions. The reviewer is confused because the definition of w* also needs clarification. The optimality of the critic network depends on both phi(s) and w. Different state representation phi might correspond to different optimal w. Is phi function fixed?\\n\\n2. In terms of the proposed method, what is the major novelty compared to previous methods in MOO literature? The reviewer understands the discussions already presented in the paper, but is just wondering whether the proposed method is replacing the gradients in MOO with policy gradients given by different critics.\\n\\n3. It seems that the detailed experiment description is not specified even in the appendix. What is the horizon of this game? What is the size of the state set?\\n\\n4. Is the weight vector p predefined or learned?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to Reviewer P7fw (Continued)\", \"comment\": \">**Your Comment 11 (Also Question 2):** Another concern is the novelty in terms of algorithm and convergence analysis. Specifically, the WC-MOAC algorithm appears to be a direct application of the MOAC algorithm (Zhou et al. 2024) and also similar to the MOO algorithm CR-MOGM of (Zhou et al. 2022), which is the enhanced (stochastic) MGDA method (e.g., (Desideri, 2012; Liu and Vicente, 2021)) with the momentum update of the dual variable vector, to the setting of MORL (with the multi-objective critic learned by standard TD updates). Under a properly learned critic, then the stochastic multiple gradients can have a sufficiently low bias such that it enables similar convergence guarantees as in the general MOO. This is also shown in Theorem 11 (as a direct result of Lemma 10). As a result, the sample complexity and the convergence analysis of WC-MOAC essentially resemble those of CR-MOGM in (Zhou et al. 2022) for the general non-convex case, cf. Theorem 3 and Appendix E of (Zhou et al. 2022).\\n\\n**Our Response:** Due to the length of this comment, we organize our response in four parts:\\n\\n1) **Direction Application of (Zhou et al. 2024):** It's simply **not** true to say that our work is a direct application of (Zhou et al. (2024)). On the contrary, our work is far more general than (Zhou et al. (2024)) because the goal in (Zhou et al. (2024)) is only limited to finding a mere Pareto Stationary point; while our work aims to systematically explore the Pareto stationarity front. From this perspective, our paper can be considered a general approach that includes the results in (Zhou et al. (2024)) as a special case. To see this, note that, when $u=0$ and $\\\\mathbf{p}=(1,\\\\cdots,1)^{\\\\top}$ (for simplicity, ignoring the scaling) in our work, the results actually imply the those in (Zhou et al. (2024)). This is because Eq.(10) reduces to the $\\\\min_{\\\\lambda}||\\\\mathbf{K_p}\\\\lambda||^{2}$, which is equivalent to MGDA where the Quadratic programming is to solve $\\\\min_{\\\\lambda}||\\\\mathbf{G_t}\\\\lambda||^{2}$ where $\\\\mathbf{G_t}=[\\\\mathbf{g}^{i}_t,\\\\cdots, \\\\mathbf{g}^{M}_t]$ and $\\\\mathbf{g}^{i}_t$ is defined in the second column of Line 394. In Algorithm 1, the update of $\\\\mathbf{g}_t$ will reduce to $\\\\mathbf{g}_t= \\\\mathbf{G}_t\\\\lambda_t$. In other words, the WC-approach proposed in this paper is a far more general approach than that of (Zhou et al. (2024)). Please also see our response to your Comment 2 for further details.\\n\\n2) **Novelty in Algorithm and Analysis:** The novelty in the convergence analysis stems from analysis of involving Hadamard product by formulating the problem in a weighted-Chebyshev problem. In contrast, there is **no** such operand in (Zhou et al (2024)). In particular, unlike the pure MGDA approach where the $\\\\mathbf{\\\\lambda_t}$ is used to weigh multiple gradients of the objectives, due to the new WC component in our work, we need to construct a pseudo weight $q_t:=\\\\frac{\\\\lambda_t\\\\odot \\\\mathbf{p}}{\\\\langle \\\\lambda_t, \\\\mathbf{p}\\\\rangle}$ due to aforementioned Hadamard product and utilize properties of such pseudo weight. Please see our proofs in Lines 907-970 and Lines 1160-1210, where we carefully handled such pseudo weight, as comparison with naive MGDA. As a result of this analysis, we observed that $p_{\\\\min}$ is a cruicial quantity in characterizing the stationarity convergence. As one can easily observe that coefficients in front of the cruicial terms are significantly different from (Zhou et al. (2024)). Just to name one example, the coefficient in front of $\\\\mathcal{O}(1/T)$ term is $16L_J r_{\\\\max}(1+\\\\frac{2}{p_{\\\\min}^{2}}\\\\sum_{t=1}^{T}\\\\eta_t)$ whereas $18L_Jr_{\\\\max}(1+2\\\\sum_{t=1}^{T}\\\\eta_t)$. To claim that they are *\\\"almost the same\\\"*, it is unfair to say the least.\\n\\n3) **Standard TD Updates:** We agree that Theorem 11 considers the standard TD learning. However, it has nothing to do with stochastic multiple gradient apporach as suggested by the reviewer. Specifically, in our work, TD learning is not designed to optimize the Mean Square Bellmann Error, where it may be possible to formulate the policy evaluation into such multiple gradient approach. However, this problem is beyond the scope of this paper, which is more appropriate to investigate in a separate paper. In this work, MGDA is incorporated into the actor component in our algorithmic design as we described in details at Step 2 of Section 4.2.\"}" ] }

[ "Sheryl Hsu", "Omar Khattab", "Chelsea Finn", "Archit Sharma" ]

[ "LLMs", "Reinforcement Learning", "Information Retrieval" ]

{ "note_id": [ "wFTVJhp0wH", "vmCIzJe2ox", "kU55FYvLK7", "hvtHXFBuxW", "eBtI7J75JE", "a3iiEzPO1T", "JLah0zFXwh", "Aw85hl9mP5", "7Bfie4tqu9", "6ASUmT4XM6", "5JLgKxONfv", "5BYIVRgRmF", "4MQZ69BeQs", "2oJwcIqXob" ], "note_type": [ "official_review", "meta_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_review", "decision", "official_review", "official_comment", "official_comment" ], "note_created": [ 1730743043033, 1734748833676, 1732604243536, 1732510520301, 1732264989340, 1732264413491, 1730453134839, 1732263994057, 1732263835924, 1729566480411, 1737523938377, 1730702939818, 1732264939837, 1732604598440 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission8870/Reviewer_NW9U" ], [ "ICLR.cc/2025/Conference/Submission8870/Area_Chair_MKpG" ], [ "ICLR.cc/2025/Conference/Submission8870/Reviewer_cwbp" ], [ "ICLR.cc/2025/Conference/Submission8870/Reviewer_9WtD" ], [ "ICLR.cc/2025/Conference/Submission8870/Authors" ], [ "ICLR.cc/2025/Conference/Submission8870/Authors" ], [ "ICLR.cc/2025/Conference/Submission8870/Reviewer_cwbp" ], [ "ICLR.cc/2025/Conference/Submission8870/Authors" ], [ "ICLR.cc/2025/Conference/Submission8870/Authors" ], [ "ICLR.cc/2025/Conference/Submission8870/Reviewer_9WtD" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission8870/Reviewer_v1Yv" ], [ "ICLR.cc/2025/Conference/Submission8870/Authors" ], [ "ICLR.cc/2025/Conference/Submission8870/Area_Chair_MKpG" ] ], "structured_content_str": [ "{\"summary\": \"In multi-hop question answering, a model needs to perform multiple retrieval steps before arriving at an answer. Since the reward (answer) is not known until the last step, this problem lends itself well to reinforcement learning (RL). The paper proposes to optimize one component (the question generator for retrieval) in the multi-hop QA pipeline using RL. The method uses direct supervision (gold retrieved docs) to first train a reward function. Then a diverse set of queries are sampled from the model using varied prompts. The rewards for these generations are fed into an RL algorithm (IPO) to improve the query generator. The method improves substantially on pure SFT methods on HotpotQA and HoVer.\", \"soundness\": \"3\", \"presentation\": \"4\", \"contribution\": \"3\", \"strengths\": [\"Well written paper, clear presentation\", \"Results are strong and convincingly support the claim that using RL to improve query generator works better than pure SFT.\"], \"weaknesses\": [\"A bit difficult to judge the novelty of the contribution. Has similar methods been used for single-hop QA or RAG in general? If so, the novelty here might be marginal, especially since the paper relies on direct supervision for the reward model.\", \"The two multi-hop datasets used are not natural and somewhat out-of-date. It would be great to see the methods usefulness on more relevant tasks and benchmarks. The long-form generation attempt in the appendix is interesting, and could perhaps be developed more and moved into the main text?\"], \"questions\": [\"Has similar methods been used for single-hop QA or RAG in general?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"metareview\": \"This submission introduces LeReT, a reinforcement learning framework that helps LLMS improve their search queries. The motivation is that LLMs often struggle with formulating the right queries. LeReT aims to address this by optimizing query quality through trial and error. The experiments show that the proposed method enhances the retrieval accuracy by up to about 30% and improving downstream evaluations by 17%.\", \"the_reviewers_identified_its_strengths_as\": [\"the empirical results suggest that RL improves query generation for multi-hop queries and retrieval accuracy.\", \"the iterative training proposed brings performance improvement by enhancing retrieval and grounding.\", \"the proposed framework is compatible with various retrieval systems, making possible applications.\", \"It also received concerns on\", \"novelty, the reviewers find that the approach seems incremental, with relatively limited novelty compared to existing RAG or QA methods.\", \"scalability, by relying on direct supervision, limiting scalability and requiring high computational resources for multi-hop retrieval.\", \"experiments, the experimental setup could be improved by employing latest datasets and comparisons with recent works.\", \"After the rebuttal, reviewers 9WtD and cwbp updated their ratings to 6, and reviewers v1Yv and NW9U unfortunately did not engaged with the authors' rebuttal. Overall this submission receives the borderline ratings of 6, 6, 6, 6. Overall, the proposed method is empirically simple, flexible, and can be applied to any off-the-shelf retriever, making it a promising technique for enhancing LLM pipelines. Of course, as pointed out by the reviewers, the technical novelty does not form as the strength of this work. Given the relatively possible ratings and possible applications, this work at its current format receives an acceptance recommendation.\"], \"additional_comments_on_reviewer_discussion\": \"After the rebuttal, reviewers 9WtD and cwbp updated their ratings to 6, and reviewers v1Yv and NW9U unfortunately did not engaged with the authors' rebuttal. Overall this submission receives the borderline ratings of 6, 6, 6, 6.\"}", "{\"comment\": \"This response has mostly address my concerns so I gave a score raise.\"}", "{\"comment\": \"The authors have largely addressed my concerns, and I acknowledge the reliability of the motivation and methodology of this paper. Therefore, I have decided to raise my score.\\nHowever, I would like to add that I do not think it is appropriate to refer to Iterative-LeReT as online RL training. Additionally, training PPO using both process rewards at each hop and a final reward for the conclusion seems to me to be a more direct approach.\"}", "{\"comment\": \"> On lines 218\\u2013219, are there any concrete examples illustrating the differences between BFRS-generated queries and high-temperature sampling queries?\\n\\nIt is hard to empirically observe any differences between queries, especially in single examples. However, overall few shot prompting seems to lead to more consistent queries while high-temperature sampling will sometimes lead to differently formatted outputs like SQL queries instead of a phrase you might type into a retriever.\\n\\n> Regarding line 260, what are the examples used in few-shot prompting? Why is few-shot prompting not feasible during testing? Is it due to considerations of inference efficiency?\\n\\nThe examples used in few-shot prompting are question and query pairs that result in good retrievals selected using DSPy\\u2019s bootstrapping. Few-shot prompting is feasible during testing, however part of our recipe is to perform context distillation before preference optimization via IPO. Additionally, we use different sets of few shot prompts and want to teach the model to output the best query generated by any of the few shot prompts. To do this during test time would involve ensembling across all the few shot prompts, and as demonstrated by our empirical results in Table 1, preference optimization leads to much larger gains than ensembling across few-shot prompts. It would be interesting for future work to combine few-shot prompting with preference-optimized models from LeReT.\\n\\n> As for section 4.3, are there any details about how rewards are computed?\\n\\nYes, for direct supervision we use the average precision of the retrieved documents versus correct documents. For indirect supervision, we experimented with the F1 score of the generated answer as detailed in Appendix B.5 and recently also experimented with providing the generator with the correct answer and asking it which set of retrieved documents would be more helpful.\\n\\n> For Iterative-LeReT and the discussion in Section 4.3, this multi-hop search process seems compatible with online RL training, such as using PPO, by incorporating process rewards at each hop and a final reward for the conclusion. Have the authors explored online RL training?\\n\\nIterative-LeReT can be seen as a form of online RL training. LeReT is compatible with any RL algorithm and we agree that further exploration into PPO or other algorithms could be fruitful.\"}", "{\"comment\": \"We thank reviewer cwbp for their thoughtful comments and suggestions!\\n\\n> The novelty assertion of the proposed method lacks clarity. Regarding the related work spanning from line 139 to 148, it remains ambiguous as to how the proposed method differentiates itself from those other methods. I comprehend that the proposed approach employs diverse query generation and IPO for preference learning. However, these seem to be more of incremental enhancements within an existing framework rather than representing a distinct novelty.\\n\\nThank you for raising this point and we have further updated our related work to clarify our contributions. Recent work has focused on creating better retrievers (RAG[1], DPR[2], ColBERT[3], Baleen[4]), prompt-based query generation techniques (Query2Doc[5], Rethinking with Retrieval[6], DSPy[7]), and downstream answer generation methods (Chain of Note[8], Copy is All You Need[9], Tree of Clarifications[10]). In comparison, our work demonstrates how RL can be used to improve modern retrieval systems effectively. We propose prompt-driven diverse query generation, which enables RL to effectively improve LLMs\\u2019 retrieval capabilities. Our empirical results show that RL substantially improves retrieval and downstream performance beyond baselines (see Table 1).\\n\\nBeyond any specific framework, LeReT is broadly generalizable and can be applied on top of many various existing methods. LeReT can be used with prompting-based approaches such as hypothetical document generation by collecting data on which hypothetical documents / more broadly the output of the prompt lead to better results and teaching the model to do that. In addition, as LeReT treats the retriever as a black box, it can be used with ongoing improvements to actual retrieval systems.\\n\\n1. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. Patrick Lewis et al. NeurIPS 2020.\\n2. Dense Passage Retrieval for Open-Domain Question Answering. Vladimir Karpukhin et al. EMNLP 2020.\\n3. ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT. Omar Khattab et al. SIGIR 2020.\\n4. Baleen: Robust Multi-Hop QA Dataset via Retrieval Augmentation. Xiang Lisa Li et al. NeurIPS 2021.\\n5. Query2Doc: Query Expansion with Large Pre-trained Language Models. Jiaxin Mao et al. SIGIR 2021.\\n6. Rethinking with Retrieval: Faithful Large Language Model Inference. Hangfeng He et al. ACL 2023.\\n7. DSPy: Compiling Declarative Language Model Calls into State-of-the-Art Pipelines. Omar Khattab et al. ICLR 2024.\\n8. Chain-of-Note: Enhancing Robustness in Retrieval-Augmented Language Models. Wenhao Yu et al. EMNLP 2024.\\n9. Copy Is All You Need. Tian Lan et al. ICLR 2023.\\n10. Tree of Clarifications: Answering Ambiguous Questions with Retrieval-Augmented Large Language Models. Gangwoo Kim et al. EMNLP 2023.\\n\\n> Also, the experiments lack comparisons with the most relevant recent works. Only the basic few-shot prompt baseline is compared.\\n\\nThanks for raising this concern. We have added comparisons to Query2Doc [1], and have added a discussion around other related methods for retrieval. We believe our contribution is compatible / orthogonal to several other works, thus we compare to key prior algorithmic works. What other baselines do you think are meaningful to compare LeReT to?\\n\\n[1] Query2doc: Query Expansion with Large Language Models. Liang Wang, Nan Yang, Furu Wei. Empirical Methods in Natural Language Processing (EMNLP) 2023.\\n\\n> Line 76 \\\"If LLMs can observe the retrieved documents for different search queries, they can learn which queries lead to better outcomes.\\\" What supports this claim? And if this is true, then why do you need the direct or indirect supervision to teach LLM \\\"how to query\\\"?\\n\\nWe apologize for the ambiguity and have revised the paper accordingly (line 45). We meant that by observing the retrieved documents for different search queries, we can generate a reward for the query, and train the model to output queries that get a high-reward.\"}", "{\"summary\": \"The paper presents LeReT, a framework for improving LLM retrieval and answer grounding using reinforcement learning. It uses prompt-driven diverse query generation, model optimization with preference-based RL, and reward labeling for retrieved documents. Experiments on HotpotQA and HoVer datasets show significant improvements in retrieval and downstream generation compared to baselines. Analysis reveals the importance of diverse few-shot prompting and LeReT's applicability across different retrievers. Limitations include the use of direct supervision, and future work may focus on indirect supervision and tool updating.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": [\"LeReT is applicable to general retrieval-augmented generation (RAG) systems and can adapt to different retrievers.\", \"LeReT significantly improves retrieval accuracy. Compared to the unadopted Llama and Gemma instruction models, the recall rate increases by 9-22% on HotPotQA and 27-29% on HoVer.\", \"It can be used iteratively: applying LeReT for two iterations shows that the model performance after the second iteration is better than that of the standard non-iterative LeReT.\"], \"weaknesses\": [\"The novelty assertion of the proposed method lacks clarity. Regarding the related work spanning from line 139 to 148, it remains ambiguous as to how the proposed method differentiates itself from those other methods. I comprehend that the proposed approach employs diverse query generation and IPO for preference learning. However, these seem to be more of incremental enhancements within an existing framework rather than representing a distinct novelty.\", \"Also, the experiments lack comparisons with the most relevant recent works. Only the basic few-shot prompt baseline is compared.\"], \"questions\": [\"Line 76 \\\"If LLMs can observe the retrieved documents for different search queries, they can learn which queries lead to better outcomes.\\\" What supports this claim? And if this is true, then why do you need the direct or indirect supervision to teach LLM \\\"how to query\\\"?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"We thank reviewer v1Yv for their thoughtful comments and insights, particularly into the scalability and generalization of our work!\\n\\n> Primarily relies on direct supervision for labeling relevant documents, which may limit its scalability in cases where explicit relevance labels are unavailable.\\n\\nWe agree that adapting LeReT for indirect supervision is an exciting direction and would improve the scalability of the method. A couple of salient points:\\n- As the data scaling curves in Appendix A.1 suggest, a lot of the performance gains happen with relatively few data points, so LeReT should provide meaningful performance improvements even when limited data is available.\\n- The trade-off between comparing final answers (i.e. indirect supervision) and retrieved documents (i.e. direct supervision) is often not clear. There are situations where judging between two sets of documents might be easier than comparing the generated answers. For example, annotators could judge whether a LLM is sourcing medical advice from reliable sites without necessarily having the medical knowledge to evaluate the generated medical information.\\n\\n> Requires extensive computation due to multi-hop retrieval and diverse query sampling, making it resource-intensive.\\n> The need for sampling across multiple hops is computationally intensive and less parallelizable, reducing scalability.\\n\\nWe agree that LeReT is resource-intensive. However, as demonstrated by the scaling curve in Appendix A.1, the majority of the improvement is realized with relatively few preference pairs. Moreover, the trained model can be deployed cheaply compared to other methods like few-shot prompting that require additional tokens during inference, Query2doc which requires multiple calls to the LLM, and Tree-of-Clarification which makes many recursive calls to the generator. One-time training cost may be very favorable in several use cases given the low cost at the time of inference.\\n\\n> While iterative training improves retrieval accuracy, does it potentially overfit the model to specific multi-hop tasks, thereby impacting generalizability in other retrieval-augmented scenarios?\\n> What is involved in adapting LeReT to new domains or types of queries that it was not originally trained on?\\n\\nWe agree that overfitting is a potential concern. Empirically, we see from looking at generated queries that general things such as learning not to output SQL queries but instead phrases can be generalized across domains. We also test a model trained on Hover on HotpotQA (and vice versa) and find that while these models do not perform as well as models trained on the dataset, they outperform few-shot prompting. Full experimental results are available in the now updated Appendix B.6. Algorithmically, IPO is more resistant to overfitting. We find that we can generalize our retrieval models from one domain to another. Building on modern LLMs enables such generalization to other domains as pre-training happens on extremely broad and diverse datasets.\"}", "{\"comment\": \"We thank reviewer NW9U for their thorough summary and feedback!\\n\\n> A bit difficult to judge the novelty of the contribution. Has similar methods been used for single-hop QA or RAG in general? If so, the novelty here might be marginal, especially since the paper relies on direct supervision for the reward model.\\n\\nThanks for raising this point and we recognize that there are many contributions on this topic. We believe [1] is closest to our framework, making comparable assumptions related to direct supervision as our work does. We have updated the related work with a discussion around this. There are several important differences that are crucial to LeReT, especially given the advances in LLMs since 2017. First, [1] does not use a generative LM but a query selection architecture (understandably, since it is 2017). Importantly, a straightforward application of RL similar to [1] for modern generative LLMs would not perform better than few-shot prompting with current LLMs, as seen in Table 5 when comparing few-shot and LeReT @ temp 2.0. Our proposed prompt-driven diverse query generation is what allows reinforcement learning to be effective and improve current LLMs for retrieval. At a fundamental level, our strategy leverages advances in modern LLMs to improve the exploration for the RL problem and leverage the generative nature of these LLMs, substantially improving the downstream performance beyond what few-shot prompting can achieve.\\n\\n[1] Task-Oriented Query Reformulation with Reinforcement Learning. Rodrigo Nogueira, Kyunghyun Cho. Association for Computational Linguistics (ACL) Anthology, 2017.\\n\\n> The two multi-hop datasets used are not natural and somewhat out-of-date. It would be great to see the methods usefulness on more relevant tasks and benchmarks. The long-form generation attempt in the appendix is interesting, and could perhaps be developed more and moved into the main text?\\n\\nWe agree that it would be exciting to see LeReT applied to other datasets. At the time, HotpotQA and HoVeR were the only two multi-hop datasets with document annotations. Tackling long-form generation requires a concerted effort around data generation, which we believe is beyond the scope of this project. We are excited about future work that tackles this direction!\"}", "{\"summary\": \"The paper addresses the issue of hallucination in language models by enhancing their capability to generate effective queries for retrieving relevant facts, thereby improving the accuracy of model responses. The proposed approach, LeReT, generates diverse search queries by incorporating few-shot examples. It constructs comparative queries guided by the supervision of retrieval quality and optimizes the model using the IPO algorithm. The study empirically validates the LeReT framework, exploring various methods of collecting retrieval rewards and assessing performance across different retrievers.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"3\", \"strengths\": \"1. The motivation that enhancing RAG performance through learning to generate better search queries is clear and reasonable.\\n2. The experiments present the improvement in both retrieval quality and finally performance, which validate the soundness of proposed algorithm.\", \"weaknesses\": \"1. The writing and presentation of results in this work could be enhanced. On line 254, I think it should be, \\\"In this work, we consider XXX.\\\"\\n\\n2. Some baselines (https://aclanthology.org/2023.emnlp-main.585.pdf, https://openreview.net/pdf?id=vDvFT7IX4O, and https://aclanthology.org/2023.acl-long.99.pdf) could be added to both the experiments and the related works sections. Although these works may not directly focus on multi-hop retrieval, they fit broadly within the same topic as this work, i.e., \\\"query expansion.\\\" These works focus on generating better search queries to enhance RAG performance.\\n\\n3. Relying on annotated golden documents, referred to as \\\"direct supervision\\\" in this work, limits the vision of the study. In more general scenarios, collecting \\\"indirect supervision\\\" is more feasible. Moreover, signals from \\\"indirect supervision\\\" are the ultimate indicators for downstream tasks.\", \"questions\": \"1. On lines 218\\u2013219, are there any concrete examples illustrating the differences between BFRS-generated queries and high-temperature sampling queries?\\n\\n2. Regarding line 260, what are the examples used in few-shot prompting? Why is few-shot prompting not feasible during testing? Is it due to considerations of inference efficiency?\\n\\n3. As for section 4.3, are there any details about how rewards are computed?\\n\\n4. For Iterative-LeReT and the discussion in Section 4.3, this multi-hop search process seems compatible with online RL training, such as using PPO, by incorporating process rewards at each hop and a final reward for the conclusion. Have the authors explored online RL training?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}", "{\"summary\": \"This paper presents a framework called \\\"Learning to Retrieve by Trying\\\" (LeReT), which aims to improve the grounding of large language models (LLMs) through reinforcement learning (RL)-based retrieval. LeReT enables LLMs to generate queries, learn from trial-and-error, and enhance retrieval by using diverse few-shot prompts combined with preference-based reinforcement learning. LeReT's flexibility makes it adaptable across different retrieval systems, with potential applicability in broader retrieval-augmented generation (RAG) contexts.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. Introduces a unique reinforcement learning framework to improve retrieval accuracy in LLMs, especially for complex multi-hop queries.\\n2. Demonstrates the effectiveness of iterative training in enhancing the retrieval and grounding abilities of LLMs.\\n3. Compatible with various retrieval systems, including ColBERTv2 and Azure AI Search, indicating its broad applicability.\", \"weaknesses\": \"1. Primarily relies on direct supervision for labeling relevant documents, which may limit its scalability in cases where explicit relevance labels are unavailable.\\n2. Requires extensive computation due to multi-hop retrieval and diverse query sampling, making it resource-intensive.\\n3. The need for sampling across multiple hops is computationally intensive and less parallelizable, reducing scalability.\", \"questions\": \"1. While iterative training improves retrieval accuracy, does it potentially overfit the model to specific multi-hop tasks, thereby impacting generalizability in other retrieval-augmented scenarios?\\n2. What is involved in adapting LeReT to new domains or types of queries that it was not originally trained on?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"details_of_ethics_concerns\": \"N/A\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"We thank reviewer 9WtD for their detailed feedback and suggestions, especially around baselines!\\n\\nWe have revised the paper and improved the writing.\\n\\n> Some baselines (https://aclanthology.org/2023.emnlp-main.585.pdf, https://openreview.net/pdf?id=vDvFT7IX4O, and https://aclanthology.org/2023.acl-long.99.pdf) could be added to both the experiments and the related works sections. Although these works may not directly focus on multi-hop retrieval, they fit broadly within the same topic as this work, i.e., \\\"query expansion.\\\" These works focus on generating better search queries to enhance RAG performance.\\n\\nWe appreciate reviewer 9WtD bringing our attention to these related works, and we have updated our related works section appropriately. We have added Query2Doc[1] as a baseline in Table 1. To summarize the discussion here: Tree of Clarifications[2] is a promising method, as it performs question clarification after retrieval in order to improve answer generation. Since LeReT mainly focuses on improving retrieval, the work is complementary to our work and Tree of Clarifications could be run on LeReT-improved retrievals. While HyDE seems quite effective, we find that it requires an additional contrastive encoder to find similar documents to the generated hypothetical document. Given that LeReT is targeting the common use case of a black box retriever like the Bing API or Azure AI Search where it is not possible to access an encoder as required by HyDE, it is not possible to compare the two methods. Additionally, the Query2doc baseline already measures the performance gains from hypothetical documents in the black box retriever setting. We would also like to emphasize that LeReT, our reinforcement-learning based framework for improving retrieval, can be combined with these prior works. For example, LeReT could be used to sample hypothetical documents and queries generated by Query2doc or HyDE and fine tune the model to produce better documents and queries. We are excited for future work to explore these directions.\\n\\n1. Query2doc: Query Expansion with Large Language Models. Liang Wang, Nan Yang, Furu Wei. Empirical Methods in Natural Language Processing (EMNLP) 2023.\\n2. Tree of Clarifications: Answering Ambiguous Questions with Retrieval-Augmented Large Language Models. Gangwoo Kim, Sungdong Kim, Byeongguk Jeon, Joonsuk Park, Jaewoo Kang. Empirical Methods in Natural Language Processing (EMNLP) 2023.\\n3. Precise Zero-Shot Dense Retrieval without Relevance Labels. Luyu Gao, Xueguang Ma, Jimmy Lin, Jamie Callan. Association for Computational Linguistics (ACL) Anthology, 2023.\\n\\n> Relying on annotated golden documents, referred to as \\\"direct supervision\\\" in this work, limits the vision of the study. In more general scenarios, collecting \\\"indirect supervision\\\" is more feasible. Moreover, signals from \\\"indirect supervision\\\" are the ultimate indicators for downstream tasks.\\n\\nWe agree that using LeReT with indirect supervision would allow the method to be used in more general scenarios. However, we find in the data scaling curves in Appendix A.1 that relatively few data points are needed to achieve a lot of the performance gains, so even with limited data LeReT is likely to be effective. In addition, the trade-off in feasibility between collecting \\u201cindirect supervision\\u201d and \\u201cdirect supervision\\u201d is often not clear. There are scenarios where collecting direct supervision by judging between two sets of documents might be easier than collecting indirect supervision by comparing the generated answers. For example, annotators could judge whether a LLM is sourcing medical advice from reliable sites without necessarily having the medical knowledge to evaluate the generated medical information.\"}", "{\"comment\": \"Dear Reviewer v1Yv the ICLR discussion period is extended. Could you please take a look at the authors' rebuttal and other reviews, and see whether you would like to update your ratings? The authors would greatly appreciate your consideration and responses.\"}" ] }

[ "Shouheng Li", "Floris Geerts", "Dongwoo Kim", "Qing Wang" ]

[ "gnn", "expressivity", "generalization" ]

{ "note_id": [ "xsGxBMHTUz", "wNCUKIutac", "vK4SpIBKCN", "v2S2jx4brD", "upynOs9P1s", "qSLQ9jEZjv", "piL97sHcaf", "pbiZzQ3GIf", "mk13LupdAR", "kjgdyhwxWp", "kQdAGacffc", "kCE5ZmKjoe", "ihdbPTY9xi", "bfmMCohXjX", "agY2VYIm4c", "YhbInl8FjZ", "UADQUyyC9F", "TAXlnciBmO", "T7bXrQwLN0", "QL6sXMbYhm", "NYGoX6ZTh3", "NYAiT7LRSu", "JthT669LCq", "Hqsr9SJKiw", "G627A1iZDV", "EoaToXjME3", "C3ePHRsYfr", "9yc3ZGL5jE", "8o76vubvIx", "4NnPBWDGUC", "3rF37FESFz", "35bsaSWpdY" ], "note_type": [ "official_comment", "official_review", "official_review", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "meta_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "decision", "official_comment" ], "note_created": [ 1732688418333, 1730341600913, 1730548781352, 1730663915144, 1732442983189, 1732544761790, 1732442303179, 1732443098013, 1732442836169, 1732750494536, 1732443300565, 1733038958273, 1732562603411, 1732443213262, 1732764478397, 1732443499378, 1730013376125, 1732442320544, 1734965676248, 1732442259304, 1732443840289, 1733113386279, 1732442274232, 1732443164831, 1732808494581, 1733074216849, 1732442226558, 1732442888013, 1732442548604, 1732442357421, 1737523908860, 1732697522488 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission8440/Reviewer_31fD" ], [ "ICLR.cc/2025/Conference/Submission8440/Reviewer_31fD" ], [ "ICLR.cc/2025/Conference/Submission8440/Reviewer_yEuE" ], [ "ICLR.cc/2025/Conference/Submission8440/Reviewer_iUbN" ], [ "ICLR.cc/2025/Conference/Submission8440/Authors" ], [ "ICLR.cc/2025/Conference/Submission8440/Reviewer_yEuE" ], [ "ICLR.cc/2025/Conference/Submission8440/Authors" ], [ "ICLR.cc/2025/Conference/Submission8440/Authors" ], [ "ICLR.cc/2025/Conference/Submission8440/Authors" ], [ "ICLR.cc/2025/Conference/Submission8440/Authors" ], [ "ICLR.cc/2025/Conference/Submission8440/Authors" ], [ "ICLR.cc/2025/Conference/Submission8440/Authors" ], [ "ICLR.cc/2025/Conference/Submission8440/Authors" ], [ "ICLR.cc/2025/Conference/Submission8440/Authors" ], [ "ICLR.cc/2025/Conference/Submission8440/Reviewer_UQK4" ], [ "ICLR.cc/2025/Conference/Submission8440/Authors" ], [ "ICLR.cc/2025/Conference/Submission8440/Reviewer_UQK4" ], [ "ICLR.cc/2025/Conference/Submission8440/Authors" ], [ "ICLR.cc/2025/Conference/Submission8440/Area_Chair_XdA7" ], [ "ICLR.cc/2025/Conference/Submission8440/Authors" ], [ "ICLR.cc/2025/Conference/Submission8440/Authors" ], [ "ICLR.cc/2025/Conference/Submission8440/Authors" ], [ "ICLR.cc/2025/Conference/Submission8440/Authors" ], [ "ICLR.cc/2025/Conference/Submission8440/Authors" ], [ "ICLR.cc/2025/Conference/Submission8440/Authors" ], [ "ICLR.cc/2025/Conference/Submission8440/Reviewer_UQK4" ], [ "ICLR.cc/2025/Conference/Submission8440/Authors" ], [ "ICLR.cc/2025/Conference/Submission8440/Authors" ], [ "ICLR.cc/2025/Conference/Submission8440/Authors" ], [ "ICLR.cc/2025/Conference/Submission8440/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission8440/Authors" ] ], "structured_content_str": [ "{\"title\": \"Reply to the Authors\", \"comment\": \"Thanks the authors for their responses. My previous concerns are adequately addressed. I have increased my score to 8.\"}", "{\"summary\": \"The authors in this paper study the connection between expressivity and generalization ability of GNNs in the graph classification task. They first show that the generalization ability of a graph encoder can be upper bounded by that of graph more encoders with higher expressivity . After that, they establish generalization bounds for GNNs from a optimal transfer viewpoint. Concretely, the derived bound contain the Wasserstein distance of embeddings between two random subsets sampled from training nodes within the same class. In this way, the generalization ability of GNNs in the graph classification task can be depicted by the concentration of intra-class embeddings and the separation of inter-class embeddings. Finally, the authors also provide empirical bounds that adopting sampling to computed the Wasserstein distance and verify their theoretic results on real-world datasets. The experimental results generally support their derived bounds.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"3\", \"strengths\": \"The problem studied in this paper is pivotal in the theory of GNNs, i.e., revealing the connection between expressivity and generalization ability of GNNs. This paper makes a non-trivial progress in this problem, and the derived results are convinced and sound. As for the generalization bounds for GNNs in graph classification task, the derived results extend previous results (Chuang et al. 2021) to graph data, which also provides a new perspective to understand the generalization ability of GNNs. And, the experimental results show that the numerical values of the derived empirical bounds are significant sharper than that of VC bounds.\", \"weaknesses\": \"- The derived generalization bounds for GNNs contain the Lipschitz constants of graph encoders. These constants could be easily evaluated for shallow models, e.g., GCN with two or three layers, yet it may be difficulty to evaluate for more complex GNNs such as graph transformer [1,2].\\n- The derived bounds do not take the optimization into consideration and thus could not reflect the impact of optimization algorithms on generalization ability. It has been shown that the training trajectory could also affect the generalization ability of models [3]. Although this observation is not yet verified for GNNs, I believe that improving the generalization bounds by involving the analysis of training dynamics could be a promising direction.\\n\\n\\n[1] Chen et al., Structure-Aware Transformer for Graph Representation Learning. ICML 2022.\\n\\n[2] Ramp\\u00e1\\u0161ek et al., Recipe for a General, Powerful, Scalable Graph Transformer. NeurIPS 2022.\\n\\n[3] Fu et al., Learning Trajectories are Generalization Indicators. NeurIPS 2023.\", \"questions\": \"Q1: In the experiments you only compare your bounds with VC bounds. It is encouraged to also compare with PAC-Bayesian bounds presented in (Ju et al. 2023).\", \"q2\": \"Providing some examples to illustrate some definitions could help the readers better understand their meanings, e.g., the definitions of homomorphism and graph invariant in Section 3. You could add some figures to demonstrate these definitions more intuitively.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper studies the generalization of graph neural networks. The paper is based on three technical tools. The first one is margin bound, which is common in deriving generalization bound. The second is the Wasserstein distance, which is implicitly linked to the Lipschitzness property of a neural network. The third one, which is the major novelty in this paper, it to bridge between a continuous GNN to a discrete graph embedding that has more separation power, such as 1-WL or homomorphism vectors. Combining the three types of techniques yields a meaningful generalization bound. The authors gave a concrete case showing that how improved expressivity benefits/harms generalization, which is quite interesting. They also conduct experiments to support the theory.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"Overall the theoretical result is sound and help gain better understanding for how expressivity is related to generalization in GNNs. The paper is clearly written, with rigorous math notations and theorems. While not checking the proof, I feel the results are correct and make sense to me. The concrete example is quite interesting and makes this paper potentially useful in practice.\", \"weaknesses\": \"One major weekness of this paper is that, similar to most of the previous generalization papers, the bound given by this paper is quite sophisticated, intractable to compute, and potentially lose. It may be hard to draw the conclusion that this **upper bound** indeed reflects the real case. Another weakness is the readability. The authors did not explain the insights behind this bound clearly. I think it's not intuitive to understand why generalization error can be bounded by the discrete classifier with better separation power. What does the better separation power play a role in the bound? Also, why do the authors rely on Wasserstein distance? I should acknowledge that I am not an expert in generalization theory, but it would be nice to make the paper friendly to the expressivity community as well.\", \"questions\": \"See the comments above.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This work present a practical framework for analyzing the generalization property of GNNs via their expressivity. The main result is an extension of previous works to graph representation learning. The theoretical results are accompanied by case studies and experimental verifications to demonstrate the tightness of the proposed bound.\", \"soundness\": \"3\", \"presentation\": \"4\", \"contribution\": \"3\", \"strengths\": \"1. The proposed framework is empirically effective to estimate the generalization of GNNs. The bound is given by the regularity and expressivity of the graph encoder itself, which are accessible measures in practice. This method has the potential to assess the generalization of real-world models.\\n\\n2. The presentation of the paper is clean and easy to follow.\", \"weaknesses\": \"1. The verification and discussion are limited to MPNNs. Empirically, higher-order information and/or attention mechanism are widely adopted to enhance GNNs. While the results in this work are applicable to these models in a straightforward way, it does not provide insights how these architectures improve the B/S constants in the bound.\", \"questions\": \"Is it possible to do a priori analysis on how a specific message aggregation scheme, e.g., k-WL or attention, can quantitively influence the terms in the bound?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Thanks\", \"comment\": \"> S1 Overall the theoretical result is sound and help gain better understanding for how expressivity is related to generalization in GNNs. The paper is clearly written, with rigorous math notations and theorems. While not checking the proof, I feel the results are correct and make sense to me. The concrete example is quite interesting and makes this paper potentially useful in practice.\\n\\nThank you for the feedback. We are glad you liked our presentation and found the results and example interesting.\"}", "{\"title\": \"Thank you\", \"comment\": \"Thank you for the very detailed response! I am generally satisfied with the response and am happy to see the paper accepted. That is to say, I am not an expert in generalization and is not familiar with the key related works and concepts, so I will not change the confidence level.\"}", "{\"title\": \"Practical insights\", \"comment\": \"> W3 The practical insights are limited. Section 6 is quite confusing. This section shows that different graphs can lead to opposite effects on inter-class separation by the expressive power. Then, what should we do in practice based on this knowledge? Moreover, Section 6 seems irrelevant to intra-class concentration. Then, is the case study of Section 6 meaningful enough?\\n\\nThank you for your comments. Our work makes two key practical contributions:\\n- We propose a generalization bound that empirically quantifies the generalization ability of graph encoders (Theorem 5.1, particularly the sample-based version in Theorem D.2, Appendix D).\\n- The lower bound in Proposition 5.2 identifies two main factors influencing generalization: intra-class concentration and inter-class separation.\\n\\nThe purpose of Section 6 is to illustrate how these factors, in relation to graph encoder expressivity, affect generalization.\\n\\n- **Intra-class concentration:** In Table 1, we analyze the intra-class concentration for a class $c$, where the graphs in this class are assumed to cluster around two representative graphs $G$ and $G'$. The intra-class concentration is estimated using the Wasserstein distance $\\\\mathcal W_1(\\\\lambda_\\\\sharp(G), \\\\lambda_\\\\sharp(G'))$, shown in the last column of Table 1. The results reveal that, for the same graphs, increasing the expressive power of graph encoders can have opposite effects on intra-class concentration, demonstrating the nuanced behavior of expressive models. \\n- **Inter-class separation:** The final paragraph of Section 6 addresses inter-class separation, measured by the Wasserstein distance $\\\\max_{c, c' \\\\in \\\\mathcal Y, c \\\\neq c'} \\\\mathcal W_1(\\\\lambda_\\\\sharp(\\\\mu_c), \\\\lambda_\\\\sharp(\\\\mu_{c'}))$. We show that inter-class separation can influence the generalization bound in two opposite directions, depending on the distribution of class representations. \\n\\nBoth factors\\u2014**intra-class concentration** and **inter-class separation**\\u2014are thus central to Section 6. We would appreciate further clarification on the reviewer\\u2019s comment that \\u201cSection 6 seems irrelevant to intra-class concentration\\u201d to respond more effectively.\\n\\nFinally, our results also provide insights into hyperparameter selection. For instance, we analyze how graph homomorphism patterns or layers influence generalization. The generalization ability of a GNN can be evaluated through its graph encoder (e.g., 1-WL, k-WL, F-WL) without training the model, enabling efficient and explainable hyperparameter selection. Indeed, such discrete graph encoders are fast to run and easy to analyze. This insight is illustrated in Figure 1, where different choices of graph patterns and layers lead to bounds that closely reflect actual generalization gaps. This approach offers practical guidance for model design and hyperparameter tuning.\"}", "{\"title\": \"Computability of bound\", \"comment\": \"> W1 One major weakness of this paper is that, similar to most of the previous generalization papers, the bound given by this paper is quite sophisticated, intractable to compute,\\n\\nThank you for the comment. We want to clarify that alongside the theoretical bound on the generalization gap in Theorem 5.1, we also provide a sample-based version of the bound which is computable. This bound can be found in Appendix D and it is the bound we use in our experiments. We have made this more clear in the paper (after Theorem 5.1) by saying\\n\\u201cMore specifically, we show in Appendix D how an efficient to compute sample-based bound can be used instead of the theoretical bound presented in Theorem 5.1. Importantly, we use this practical bound in our experiments.\\u201d\\nTo make matters clearer, in the new Appendix E, we explain how the bounds in our experiments are computed, including VC based bounds, PAC-Bayesian-based bounds and our bounds. In short, we compute our empirical bounds based on Thm D.2, where we use samples from the training set to estimate $\\\\text{Lip}(f)$ and $\\\\Delta(\\\\lambda_\\\\sharp(\\\\mu_c))$. The most computational expensive step is to compute the Wasserstein distance, which has complexity $\\\\mathcal O((\\\\frac{m_c}{2n})^3)$ but is tractable in practice. We have added pseudocode to compute the bound in Algorithm 1 in Appendix E.\"}", "{\"title\": \"Lipschitz and optimization-based\", \"comment\": \"> W1. The derived generalization bounds for GNNs contain the Lipschitz constants of graph encoders. These constants could be easily evaluated for shallow models, e.g., GCN with two or three layers, yet it may be difficult to evaluate for more complex GNNs such as graph transformers [1,2].\\n\\nThank you for the comments. While our generalization bound involves Lipschitz constants, we want to clarify that they do not directly correspond to the graph encoders.\\n- **Decoder Component:** The term $\\\\text{Lip}(\\\\rho_\\\\psi(\\\\cdot, c))$ represents the Lipschitz constant of the graph decoders. This is typically known, as decoders often rely on standard functions like Softmax, or it can be estimated empirically, as suggested in [4].\\n- **Encoder Component:** The term $\\\\text{Lip}(f)$ , where $f$ maps the graph encoder $\\\\lambda$ to the encoder $\\\\phi$, can also be empirically estimated. Specifically, we compute:\\n$$\\\\max_{G,H}\\\\frac{d_{\\\\mathcal Y}(\\\\phi(G),\\\\phi(H))}{d_{\\\\mathcal Y\\u2019}(\\\\lambda(G),\\\\lambda(H))},$$\\nwhere the maximum is taken over the sample set. This approach only requires the embedding distance ratio, not the exact Lipschitz constants of the graph encoder.\\n\\nRegarding graph transformers, their ability to distinguish graphs depends heavily on the structural encodings they use (e.g., subtree extractors, subgraph extractors, WL trees, etc.). Whenever graph transformers (such as in [1,2]) have known expressivity bounds, our framework can evaluate their generalization, as demonstrated in our analysis. Indeed, we can use the combinatorial graph embeddings determined by their expressiveness analysis.\\n\\n> W2 The derived bounds do not take the optimization into consideration and thus could not reflect the impact of optimization algorithms on generalization ability. It has been shown that the training trajectory could also affect the generalization ability of models [3]. Although this observation has not yet been verified for GNNs, I believe that improving the generalization bounds by involving the analysis of training dynamics could be a promising direction.\\n\\n\\nThe reviewer is absolutely right that trajectory information and the inclusion of (stochastic) gradient descent based optimization may result in a more fine-grained generalization analysis. The provided reference [3] may be an interesting starting point to explore this direction in the graph setting. We note that in the graph setting, Franks et al [5] explore similar ideas, but based on the margin-based analysis by Ji and Telgarsky [6]. However, that analysis works only for deep linear models, without linearities. \\n\\n[1] Chen et al., Structure-Aware Transformer for Graph Representation Learning. ICML 2022.\\n\\n[2] Ramp\\u00e1\\u0161ek et al., Recipe for a General, Powerful, Scalable Graph Transformer. NeurIPS 2022.\\n\\n[3] Fu et al., Learning Trajectories are Generalization Indicators. NeurIPS 2023.\\n\\n[4] Chuang et al., Measuring Generalization with Optimal Transport. NeurIPS 2021\\n\\n[5] Franks et al., Weisfeiler-Leman at the margin: When more expressivity matters. ICML 2024.\\n\\n[6] Ji and Telgarsky, Gradient Descent Aligns the Layers of Deep Linear Networks. ICLR 2019.\"}", "{\"comment\": \"Whenever you have a chance, we\\u2019d appreciate your feedback on our rebuttal - thank you.\"}", "{\"title\": \"Intuition\", \"comment\": \"> W4. is the readability. The authors did not explain the insights behind this bound clearly. I think it's not intuitive to understand why generalization error can be bounded by the discrete classifier with better separation power. What does the better separation power play a role in the bound? Also, why do the authors rely on Wasserstein distance?\\n\\nWe thank the reviewer for the comments, which we next address in turn.\\n\\n*Impact of separation power.* We show that the generalization of GNNs is influenced by a discrete graph encoder that bounds the expressivity of the GNN. The influence, facilitated by inter-class separation and intra-class concentration of the discrete graph embeddings (Section 5), is possible because of a connection between the embeddings of the encoder and the GNN (Lemma 4.2, Prop 4.3, Cor 4.5). Specifically, the Wasserstein distance of the GNN embeddings is bounded by the embedding of the more powerful graph encoder by a factor $B/S$ (or $\\\\text{Lip}(f)$). This result shows that, intuitively, if the encoder separates graphs in a way that aligns well with class distribution, i.e. concentrated within each class and well-separated between classes (measured by Wasserstein distance), the GNN can potentially generalize well. \\n\\nIn other words, the generalization ability of a GNN can be analyzed by looking at the graph encoder (1-WL, k-WL, F-WL, etc) that bounds it. Such discrete graph encoders are often fast to run and easy to analyze. Hence this insight can be used to guide model design and hyperparameter selection without actually training the model.\\n\\n*Choice of Wasserstein distance.* For the choice of Wasserstein distance, we follow the work by Chuang et al. [1], where a generalization bound is obtained in terms of k-variance [2], which in turn is defined in terms of the Wasserstein distance. We adopted that distance as well. One of the key motivations for using it is that it is known to capture the variance well for clustered distributions. In our setting, the distributions are over the learned features, which are - hopefully - clustered based on their label. Intuitively, it is used to better capture the structural properties of the feature distributions, as mentioned in Chuang et al. There are, of course, also mathematical reasons. For example, the choice of Wasserstein distance in k-variance implies that k-variance shares many properties with the standard notion of variance (see [2] Proposition 2 for details). A final motivation is that the Wasserstein distance based bound was shown by Chuang et al. to predict generalization well [1].\\n\\n\\n[1] Chuang et al., Measuring Generalization with Optimal Transport. NeurIPS 2021\\n\\n[2] Solomon et al, k-Variance: A Clustered Notion of Variance. SIAM Journal on Mathematics of Data Science, 2022.\"}", "{\"title\": \"Further clarification?\", \"comment\": \"As authors can only respond until December 3 by ICLR, we kindly ask the reviewer to let us know if any further clarification is needed. Thanks.\"}", "{\"comment\": \"Thanks for reading our responses and your support.\"}", "{\"title\": \"Alignment with reality\", \"comment\": \">W3 It may be hard to draw the conclusion that this upper bound indeed reflects the real case.\\n\\nIn the newly added correlation matrices in Table 2, it can be seen our bound correlates well with empirical gaps. We have also added a new plot (Figure 2) in Appendix F which shows how our bounds change with loss gaps across layers. We also note in Figure 1 that our bound can well reflect the influence of homomorphism patterns on loss gaps. This highlights the potential of using our bound to guide model design and hyperparameter selection.\\n\\nThat said, the concern raised by the reviewer is a general concern related to the theory and practice of generalization. It is, undoubtedly, one of the main open questions in deep learning (and graph learning) to find complexity measures that accurately predict generalization error. In fact, one of the outcomes of the NeurIPS 2020 Competition: Predicting Generalization in Deep Learning was that there is a need for more rigorous theoretical bounds [1]. The work by Chuang et al. [2], on which our bound is based, shows that they outperform other bounds in most (six out of eight) benchmark tasks from the Competition mentioned earlier. Although those tasks are related to general deep learning, their k-Variance bound is one of the best ones for predicting generalization errors [2]. This somewhat justifies why we base our work on the k-Variance-based bounds.\\n\\nOf course, more research is needed to align theory and practice even better, both for graph learning and for general deep learning. It is indeed hard to draw the conclusion that theoretical bounds reflect the real world. However, as mentioned already, we do observe experimentally that it aligns well with the real generalization error (see experimental section, tables ...)\\n\\n[1] Jiang et al., NeurIPS 2020 Competition:Predicting Generalization in Deep Learning\\n\\n[2] Chuang et al., Measuring Generalization with Optimal Transport. NeurIPS 2021\"}", "{\"comment\": \"Thank you for the response. I have further questions if my understanding is correct.\\n\\n1. Weakness 3: Now I get that Section 6 is related to intra-class concentration. The authors lead me to reading Figure 1, but I cannot follow. What are the differences between the figures in the first row? In the discussion, the authors mention the results for ENZYMES and PROTEINS. Which figures in Figure 1 correspond to these two datasets? The caption does not mention these two datasets. \\n\\n2. Weakness 4: I don't think Figure 2 shows a similar trend between your derived bound and the generalization gap. There is no clear trend in the generalization gap shown in Figure 2 since the generalization gap differences between different layers are small in these datasets. At least for PROTEINS, SIDER, and MUTAG, I think the trends of the generalization gap are constant. Then, VC bound is also constant in these three cases and is overall smaller than your bound in PROTEINS and MUTAG.\"}", "{\"title\": \"Impact of expressivity\", \"comment\": \"We thank the reviewer for the very positive assessment of the paper.\\n\\n\\n> W. The verification and discussion are limited to MPNNs. Empirically, higher-order information and/or attention mechanism are widely adopted to enhance GNNs. While the results in this work are applicable to these models in a straightforward way, it does not provide insights how these architectures improve the B/S constants in the bound. Is it possible to do a priori analysis on how a specific message aggregation scheme, e.g., k-WL or attention, can quantitatively influence the terms in the bound?\\n\\nWe thank the reviewer for this interesting question, which is a question we considered while writing the paper. In general, it is difficult to draw conclusions since the influence of architectural design of the graph model is dependent on the graph data distribution. However, with some minor assumptions we can provide some insights, as follows.\\n\\nWe first explain why the influence on generalization is data-dependent. While theoretically the k-WL is 1-separating considering all graphs, for a given data distribution, it can have a greater separation than 1-WL and hence a larger $S$ which makes $B/S$ potentially smaller. However, the Wasserstein distance term in the bound for k-WL is not smaller than 1-WL. This is because k-WL can separate all graphs 1-WL can separate and increase the differences in color histogram, e.g. if the color histograms of G and H differs by 1 in 1-WL (only one vertex has a different color), there are at least $\\\\min(|V_G|, |V_H|)^{k-1}$ vertex tuples would have different colors in k-WL. So it boils down to the trade off between the decrease in B/S and the increase in intra-class Wasserstein distance.\\n\\nSuppose now that we fix $B$ and assume that we map graphs into $[0,B]$. That is, we assume all embeddings to be $B$-bounded (e.g., using normalized versions of k-WL, k-MPNNs etc). Suppose that we work in a discrete setting and that there are $m_k$ distinguishable graphs in G using k-WL. Then, maximal separability is $B/m_k$ when all $m_k$ are spread evenly over $[0,B]$. This results in a factor $(B/(B/m_k))=m_k$. Since $m_k$ increases with $k$, a smaller $k$ is preferred. Also a small $k$ also keeps the intra-class Wasserstein distance term small so it is advantageous to choose the smallest $k$.\\n\\nIn terms of attention GNNs, for Graph Attention Networks (GAT) which uses attention in the feature update function and fall into the category of MPNNs. So our results regarding MPNNs apply to it. The second type is graph transformers which adopts the transformer architecture and uses structural encodings to capture graph structure information. Graph transformers\\u2019 ability to distinguish graphs largely depends on the structure encodings they use (subtree extractor, subgraph extractor, WL trees, etc.). Some of these encodings have known expressivity bounds (e.g. 1-WL) [1]. For these our framework can be used to evaluate their generalization similarly to our analysis in the paper.\\n\\n[1] Chen et al., Structure-Aware Transformer for Graph Representation Learning. ICML 2022.\"}", "{\"summary\": \"This paper theoretically studies the relationship between the generalization of GNN and the expressive power of GNN in terms of the variance of graph structures. This work especially shows a trade-off between intra-class concentration and inter-class separation, which is the key insight into the generalization gap. Some experiments are conducted to support the theory.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"3\", \"strengths\": \"1. Some theoretical results seem solid and strong.\\n2. The problem to study is very interesting and important to the community to bridge the gap between the generalization and expressive power of GNN.\", \"weaknesses\": \"1. The theoretical proof is not complete. The proof of Corollary 4.5 and Proposition 5.2 are missing.\\n2. The paper is full of equations. Section 3 is not easy to follow when introducing so many terminologies with any examples. \\n3. The practical insights are limited. Section 6 is quite confusing. This section shows that different graphs can lead to opposite effects on inter-class separation by the expressive power. Then, what should we do in practice based on this knowledge? Moreover, Section 6 seems irrelevant to intra-class concentration. Then, is the case study of Section 6 meaningful enough?\\n4. Results in Table 2 look weak. I cannot see that the proposed bound is better than VC bound in simulating the trend of accuracy or loss gap.\", \"minor\": \"1. at the end of line 482, it shall be \\\"Figure 2(b)\\\" instead of \\\"Figure 2(a)\\\"?\\n2. Please resize Figure 2 to make it look nicer.\", \"questions\": \"In Figure 2 (a), why not show both the loss gap and the bound for each dataset? In Figure 2 (b), why not show both $W_1$ and $Lip(f)$ for each dataset?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Results\", \"comment\": \"> W4 Results in Table 2 look weak. I cannot see that the proposed bound is better than VC bound in simulating the trend of accuracy or loss gap.\\n\\nThe generalization bound based on the VC-dimension depends monotonically on the VC dimension (see Appendix E). However, the VC dimension itself is monotonically non-decreasing with respect to the number of layers and becomes constant after a certain depth. This implies that the VC-dimension does not effectively capture a model's generalization ability, as deeper models often generalize well, as demonstrated in Table 2.\\nIn contrast, our proposed bound is non-monotonic and provides a better indication of trends in generalization gaps. While these trends may not be immediately evident from the raw numbers in Table 2, we address this by adding correlation matrices between the bounds and empirical gaps alongside the table. These matrices highlight that our bounds correlate much more strongly with empirical gaps compared to other bounds.\\nAdditionally, in Figure 2 of Appendix G, we plot our bound against the loss gap, providing a visual demonstration of the strong correlation between our bounds and empirical gaps. Notably, our bound shows weaker correlation on the ENZYMES dataset. This is due to ENZYMES being a small dataset containing only 600 graphs across 5 classes, resulting in relatively small sample sizes per class ($m_c$ in Theorem D.2). This leads to greater uncertainty in bounding the generalization gap, manifesting in larger bound values and higher variance.\"}", "{\"metareview\": \"This paper studies expressivity and generalization of graph neural networks (GNNs). The generalization bound is characterized by the ratio between the concentration of intra-class embeddings and the separation of inter-class embeddings measured by the Wasserstein distance. The theoretical analysis is supported by some numerical experiments.\\n\\nThe analysis is novel and provides a non-trivial insight. The analysis successfully characterizes the relation between expressivity and generalization with solid and convincing mathematical reasoning. This is a new perspective in the literature. \\nMoreover, the empirical experiments well justifies the theoretical analysis, for example, the derived bounds are well aligned with the actual loss more than the VC bounds. \\n\\nFor these reasons, this paper can be accepted by ICLR.\", \"additional_comments_on_reviewer_discussion\": \"The reviewers pointed out some concerns about readability and missing derivation of some equations. However, those concerns were properly addressed by the authors. The paper is overall supported by the reviewers.\"}", "{\"title\": \"Proof details.\", \"comment\": \"> W1. The theoretical proof is not complete. The proof of Corollary 4.5 and Proposition 5.2 are missing.\\n\\nWe revised the text so that the proofs of Corollary 4.5 and Proposition 5.2 are now more clearly explained. That is, for Corollary 4.5 we now write \\u201cIndeed, Proposition 4.3 implies $\\\\text{Lip}(f)=B/S$. Combined with Proposition 4.4., this gives $\\\\mathcal W_1(\\\\phi_\\\\sharp'(\\\\nu), \\\\phi_\\\\sharp'(\\\\nu'))\\\\leq \\\\text{Lip}(f) \\\\cdot \\\\mathcal W_1(\\\\phi_\\\\sharp(\\\\nu), \\\\phi_\\\\sharp(\\\\nu'))= (B/S)\\\\cdot \\\\mathcal W_1(\\\\phi_\\\\sharp(\\\\nu), \\\\phi_\\\\sharp(\\\\nu'))$\\u201c instead of simply saying it is a consequence of the Proposition 4.3 and 4.4. \\n\\nSimilarly, for Proposition 5.2., instead of simply saying that the lower bound for $1/\\\\gamma$ and equation $(\\\\dagger)$ results in that proposition, we say \\u201cBy replacing $1/\\\\gamma$ in $(\\\\dagger)$ by this bound, we obtainProposition 5.2, see Appendix D for details\\u201d and also have detailed the proof in appendix D. The changes are highlighted in the revision.\"}", "{\"title\": \"Thank you\", \"comment\": \"We sincerely thank the reviewers for taking the time to read our submission and for providing positive feedback. If there are any additional questions or clarifications needed, please do not hesitate to let us know. Should our responses adequately address your concerns, we would greatly appreciate your consideration of a higher score. Thank you once again for your thoughtful reviews.\"}", "{\"comment\": \"Thank you for taking the time to evaluate our responses.\\n\\nWe would like to point out that the correlation matrices in Table 2 clearly show that our results exhibit a stronger correlation between empirical gaps and bounds compared to other bounds across all datasets, except for ENZYMES. There is a reason why our bound has a weaker correlation on ENZYMES. This is because ENZYMES contains 6 classes but only 600 graphs, resulting in a small $m_c$ that increases the bound. \\n\\nOur bound does provide a valid upper bound on the loss gap (as detailed in Theorem 5.1) with theoretical guarantees. These guarantees are consistent with those commonly found in classical learning theory.\\n\\nWe would also like to emphasize that, to the best of our knowledge, our bound is unique in its ability to adapt to the distribution of the underlying graph properties through the use of the k-Variance. This adaptation establishes a more nuanced connection between generalization and expressiveness than what has been offered by previous bounds. We believe this novel perspective is a significant step forward in understanding the interplay between these two important aspects.\"}", "{\"title\": \"Examples\", \"comment\": \"> W2 The paper is full of equations. Section 3 is not easy to follow when introducing so many terminologies with any examples.\\n\\nThanks for your suggestion. We have added examples in Appendix B for the homomorphism and 1-Wasserstein distance, hopefully it makes the concepts more clear.\"}", "{\"title\": \"Tightness of bounds\", \"comment\": \"> W2 The bound is potentially loose.\\n\\nCompared with the VC bound and PAC-Bayesian bound (newly added in revision in response to reviewer Reviewer 31fD), our bound does not grow with layers and thus is more useful in practice. Also, our bound is able to reflect the influence of normalization which cannot be captured by the other two bounds. \\nThe tightness of our bound varies on datasets. There are two major dataset-dependent factors in Theorem D.2 that influence the sharpness of the bound: $m_c$ (samples in each class), and $K$ (number of classes). ENZYMES has 6 classes but only has 600 graphs, leading to a small $m_c$ that increases the bound. In contrast, SIDER (1427 graphs 2 classes), resulting in a smaller and tighter bound. Because of this, to evaluate the bound, it would make more sense to observe how the bound reflects loss gaps across different layers and choices of homomorphism patterns. This also aligns well with real-world scenarios where a dataset is often given and a practitioner is interested in the influence of parameters and model choices on generalization. In the newly added correlation matrices in Table 2, we note our bound correlates well with empirical gaps.\"}", "{\"comment\": \"Thank you very much for reply and your two follow-up questions, which we now answer in turn.\\n\\n> Weakness 3: Now I get that Section 6 is related to intra-class concentration. The authors lead me to reading Figure 1, but I cannot follow. What are the differences between the figures in the first row? In the discussion, the authors mention the results for ENZYMES and PROTEINS. Which figures in Figure 1 correspond to these two datasets? The caption does not mention these two datasets.\\n\\nThanks for pointing out our oversight. The figures\\u2019 titles were left out due to our mistake. For that figure: Top row from left to right: PROTEINS 4-layer MPNN, PROTEINS 6-layer MPNN, ENZYMES 4-layer MPNN, ENZYMES 6-layer MPNN. Thanks for catching this. We will update the figure and caption accordingly. (The deadline for submitting a revised version has passed so we cannot upload our changed revision at this point in time.)\\n\\n> Weakness 4: I don't think Figure 2 shows a similar trend between your derived bound and the generalization gap. There is no clear trend in the generalization gap shown in Figure 2 since the generalization gap differences between different layers are small in these datasets. At least for PROTEINS, SIDER, and MUTAG, I think the trends of the generalization gap are constant. Then, VC bound is also constant in these three cases and is overall smaller than your bound in PROTEINS and MUTAG.\\n\\nThank you for your comment. As the reviewer rightly observes, the generalization gaps in our experiments only fluctuate slightly: For SIDER, the gap decreases from 0.037 to 0.034; For ENZYMES, it initially decreases from 0.248 at layer 1 to 0.235 at layer 4, then increases to 0.264 at layer 6; For PROTEINS and BACE, it fluctuates within a small range of 0.03 to 0.02. While these changes in loss appear minor, we want to emphasize that they correspond to **significant variations** in test accuracy, reaching up to 6 percentage points.\\n\\nRegarding the \\u201ctrend,\\u201d the key point we aim to highlight in Figure 2 is that our bound does not necessarily increase with the number of layers. This demonstrates that our approach is **more adaptable** to the intrinsic properties of the underlying graphs. In contrast, the VC bound is computed using an upper bound on the VC dimension, as described in [1]. This upper bound represents the number of distinguishable graphs by MPNNs with L layers. Notably, the VC bound is non-decreasing with L and becomes constant for sufficiently large L. This is for example nicely illustrated by the BACE results.\\n\\nFor the comparison with the VC bound, it is crucial to clarify that we use an (overly) **optimistic approximation** of the VC dimension. Specifically, as detailed in Appendix E and mentioned earlier, the VC bound is defined in terms of the VC dimension. Theoretically, the VC dimension for L-layer MPNNs is determined by the number of distinguishable graphs of a given size, a number that is exceedingly large and computationally infeasible to determine. Moreover, for the VC-based generalization bound to apply, the sample size must exceed the VC dimension, which would require unrealistically large datasets in our setting. Instead, we approximate the VC dimension by considering the number of distinguishable graphs within the training set. Consequently, our estimated VC dimension is smaller than the sample size, leading to a lower generalization bound than what would result from using the actual VC dimension. \\n\\nIn fact, the computational complexity of estimating the VC dimension underscores a key advantage of our bound: it is more straightforward to compute while remaining effective.\\n\\n[1] Morris, Geerts,T\\u00f6nshoff, Grohe. WL meet VC. ICML 2023.\\n\\nWe hope that these answers clarify things a bit more.\"}", "{\"comment\": \"I think the emphasized contribution that the derived bound does not necessarily increase with the number of layers is weak. It does not mean a strong result, i.e., the bound provably matches the trend of the loss gap with theoretical guarantee. The empirical results are also not supportive enough since the trend of the loss gap is not clear enough in many datasets. Due to the above reason, I prefer to keep my score of 5.\"}", "{\"comment\": \"> S1 Some theoretical results seem solid and strong. The problem to study is very interesting and important to the community to bridge the gap between the generalization and expressive power of GNN.\\n\\nWe thank the reviewer for the positive comments and appreciate that the reviewers finds the results strong and the problem interesting.\"}", "{\"title\": \"Questions\", \"comment\": \"> Q1: In the experiments, you only compare your bounds with VC bounds. It is encouraged to also compare with PAC-Bayesian bounds presented in (Ju et al. 2023).\\n\\nThank you for the suggestion. We have now included PAC-Bayesian bounds in our comparison, see Table 2 in the paper, and Table 3 in the appendix. We observe that the PAC bound grows exponentially with layers and shows a negative correlation with empirical gap. In contrast, our bound better reflects loss gap changes and are positively correlated. In Appendix E, we detail how the PAC-Bayesian bounds are computed based on Ju et al. 2023.\\n\\n> Q2: Providing some examples to illustrate some definitions could help the readers better understand their meanings, e.g., the definitions of homomorphism and graph invariant in Section 3. You could add some figures to demonstrate these definitions more intuitively.\\n\\nThank you for the suggestion. We now provide a new appendix B containing additional examples to demonstrate concepts in Section 3, hopefully it helps to make things clearer.\"}", "{\"title\": \"Interesting and non-trivial\", \"comment\": \"> S1 The problem studied in this paper is pivotal in the theory of GNNs, i.e., revealing the connection between expressivity and generalization ability of GNNs. This paper makes a non-trivial progress in this problem, and the derived results are convinced and sound.\\n\\nWe thank the reviewer for the positive comments.\\n\\n\\n> S2 As for the generalization bounds for GNNs in the graph classification task, the derived results extend previous results (Chuang et al. 2021) to graph data, which also provides a new perspective to understand the generalization ability of GNNs. The experimental results show that the numerical values of the derived empirical bounds are significantly sharper than those of VC bounds.\\n\\nWe are pleased to see that the reviewer gained a new perspective for understanding generalization based on our work.\"}", "{\"title\": \"Minor\", \"comment\": \"> Minor: 1. at the end of line 482, it shall be \\\"Figure 2(b)\\\" instead of \\\"Figure 2(a)\\\"? 2. Please resize Figure 2 to make it look nicer.\\n\\nThank you for the suggestion and pointing out this typo. We resized Figure 2, as requested. We also combined Figure 1 with Figure 2 to save some space.\\n\\n> Minor 2. In Figure 2 (a), why not show both the loss gap and the bound for each dataset? In Figure 2 (b), why not show both W1 and Lip(f) for each dataset?\\n\\nIn the original submission, we presented both the loss gap (bars) and bounds (lines) in Figure 2(a), as well as W1 (bars) and Lip(f) (lines) in Figure 2(b). We understand the reviewer\\u2019s confusion, particularly because Figure 1 used an all-bar representation. In the revision, we have unified the representation to consistently use bars and provided a clearer explanation of the plot legends. Additionally, we combined Figures 1 and 2 to save space and improve clarity. We appreciate the reviewer for highlighting this issue.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}", "{\"comment\": \"We are glad the reviewer is satisfied with our responses. The improved score is greatly appreciated. Thank you!\"}" ] }

[ "Yuchen Guan", "Runxi Cheng", "Kang Liu", "Chun Yuan" ]

[ "Knowledge Distillation", "Kendall's tau Coefficient", "Ranking Loss" ]

{ "note_id": [ "yyTbSSKJQA", "pdTtQmmIF3", "lp4d2JQ5g3", "jlQVLqe9qr", "il8trRuTIh", "hhpyIvpior", "d7j3J9uV09", "by5Cf7lKtf", "a5g6y1Vuoh", "ZGV5o1ydzl", "SFQM3n4bmh", "RIVlgEDWSZ", "HfwpRmfSck", "FT4Vax3GIR", "DGybrQVnyD", "ApADq0cZIj", "8fvbDyieL4", "7AkKNEn8WW", "4PuTiuWimI", "2oJyklrSKN" ], "note_type": [ "official_review", "official_review", "official_comment", "official_comment", "official_comment", "meta_review", "official_review", "official_comment", "official_comment", "decision", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment" ], "note_created": [ 1731084251217, 1730653478684, 1733054468977, 1732785078946, 1732195163537, 1734676239125, 1730620100515, 1733313407311, 1732132902741, 1737524081826, 1733054514593, 1732785225304, 1730644644122, 1733133480433, 1732785179202, 1733054558493, 1732732078097, 1732470434114, 1732132871825, 1732195274473 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission10857/Reviewer_pmPS" ], [ "ICLR.cc/2025/Conference/Submission10857/Reviewer_6ebV" ], [ "ICLR.cc/2025/Conference/Submission10857/Authors" ], [ "ICLR.cc/2025/Conference/Submission10857/Authors" ], [ "ICLR.cc/2025/Conference/Submission10857/Authors" ], [ "ICLR.cc/2025/Conference/Submission10857/Area_Chair_m9jg" ], [ "ICLR.cc/2025/Conference/Submission10857/Reviewer_n8Cd" ], [ "ICLR.cc/2025/Conference/Submission10857/Authors" ], [ "ICLR.cc/2025/Conference/Submission10857/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission10857/Authors" ], [ "ICLR.cc/2025/Conference/Submission10857/Authors" ], [ "ICLR.cc/2025/Conference/Submission10857/Reviewer_eu6q" ], [ "ICLR.cc/2025/Conference/Submission10857/Authors" ], [ "ICLR.cc/2025/Conference/Submission10857/Authors" ], [ "ICLR.cc/2025/Conference/Submission10857/Authors" ], [ "ICLR.cc/2025/Conference/Submission10857/Authors" ], [ "ICLR.cc/2025/Conference/Submission10857/Authors" ], [ "ICLR.cc/2025/Conference/Submission10857/Authors" ], [ "ICLR.cc/2025/Conference/Submission10857/Authors" ] ], "structured_content_str": [ "{\"summary\": \"The paper proposes an enhancement to the knowledge distillation process by introducing a plug-and-play ranking loss based on Kendall\\u2019s \\u03c4 Coefficient, which aims to mitigate the limitations of Kullback-Leibler (KL) divergence. The proposed ranking loss addresses issues like the neglect of low-probability channels and the inability of KL divergence to fully capture inter-class relationships. Extensive experiments on CIFAR-100, ImageNet, and COCO datasets demonstrate the effectiveness of the approach, showing consistent improvements when applied to various teacher-student architecture combinations in CNN and Vision Transformer (ViT) models.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"3\", \"strengths\": \"1. Novelty and contribution: The use of Kendall\\u2019s \\u03c4 ranking loss in the context of knowledge distillation appears to be novel and provides a promising way to complement traditional KL divergence-based losses. The ranking-based approach helps the student model better capture inter-class relationships.\\n2. Plug-and-Play nature: The ranking loss is designed to be plug-and-play, which increases its practicality. It can be easily integrated into existing logit-based distillation frameworks without modifying the underlying architecture.\\n3. Intensive experiments: The paper provides a wide range of experiments on different datasets and architecture combinations, demonstrating the robustness and generalizability of the proposed ranking loss.\\n4. Addressing suboptimal points: The paper provides convincing arguments about how ranking loss helps in avoiding suboptimal solutions often seen in KL divergence optimization. The experimental results back up these claims, particularly in the analysis of accuracy and loss curves.\", \"weaknesses\": \"1. limited ablation study on hyperparameters. The author only discuss the effect of hyper-parameter k in the ranking loss. there is limited analysis of how sensitive the model is to different values of \\u03b1, \\u03b2, and \\u03b3 in the overall loss function.\\n2. Relation with other different distillation loss is not clear. The paper gives some explanations on why ranking loss works through its gradient form. However, I think since this loss is not used for its own. The author should discuss its relation with KD loss. KD constrains the logits after the softmax, however, ranking loss gives the constraint before the softmax, is it really necessary? I am not convinced by this.\\n3. Some of the derivations involving the ranking loss (e.g., differentiable form of Kendall\\u2019s \\u03c4 coefficient) are challenging to follow due to their dense notation and lack of intermediate steps. Please consider adding more explanation or flowchart to increase the readablity.\", \"questions\": \"1. For the experiments involving different values of k and the comparison of different ranking loss forms, have you considered the effect of different initializations of the student model? The stability and sensitivity of the results with respect to different initial conditions could provide additional insights.\\n2. RKD is another important loss for distillation. Have you ever tried to combine with RKD (relation knowledge distillation).\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper highlights the issues with using KL divergence in knowledge distillation and introduces a ranking loss based on Kendall's \\u03c4, which can be integrated into existing methods, enhances low-probability channel focus, and maintains inter-class relationships. Experimental results across various datasets and model architectures demonstrate that this approach consistently enhances performance.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. The proposed method is designed for straightforward integration into existing logit-based distillation frameworks, increasing its relevance and utility.\\n2. Multiple experiments conducted on a variety of datasets and architectures provide evidence of the proposed approach's effectiveness\", \"weaknesses\": \"1. The KL divergence optimization is a relatively common scheme for the logit distillation task. Could the authors elaborate on the main novelty of this integration?\\n2. More ablation experiments and analysis are required for discussion; please see the Questions.\", \"questions\": \"1. What strategies could be implemented to minimize the computational overhead associated with the proposed ranking loss?\\n2. This article mentioned that the proposed method balances the model\\u2019s attention to larger and smaller-valued channels. Could the ranking loss also offer advantages in scenarios with class imbalance?\\n3. Are there any adverse effects when combining the proposed method with others? Could you provide relevant ablation experiments?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Looking forward to your reply (less than 48 hours)\", \"comment\": \"Dear reviewer 6ebV,\\n\\nWe hope our responses have adequately addressed your previous concerns. The discussion period is approaching the end in 48 hours. If you have any additional comments or questions, please feel free to share them. Your feedback is greatly appreciated.\\n\\nSincerely,\\n\\nAuthors\"}", "{\"title\": \"Looking forward to your reply\", \"comment\": \"Dear Reviewer 6ebV,\\n\\nThank you very much again for your time and effort! We would greatly appreciate it if you could take a little time to check our response. We are willing to further address any remaining concerns.\\n\\nSincerely,\\n\\nAuthors\"}", "{\"comment\": \"We sincerely thank you for the in-depth feedback! We highly value each of your comments, and all your concerns are addressed point by point:\\n\\n---\\n\\n**Weakness1. Limited Ablation Study on Hyperparameters.:**\\n\\n- We perform sensitivity analyses for $\\\\alpha-\\\\gamma$ and $\\\\beta-\\\\gamma$, and the results are presented in **Tables R1.1 and R1.2**, and the images are added to the **Figure 9 in Appendix in the paper**. As $\\\\alpha$ increases, a $\\\\gamma$ equal to $\\\\alpha$ has better performance, and as $\\\\beta$ increases, larger $\\\\gamma$ generally has better performance. Our method achieves decent performance on various settings, illustrating its robustness and generalization ability.\\n- In addition, we have also performed ablation on the parameter $\\\\gamma$, and the results are shown in Table 6 of the paper.\\n\\n *Table R1.1: Sensitivity of $\\\\alpha-\\\\gamma$. Teacher is WRN-40-2 and student is WRN-40-1.*\\n | $\\\\alpha-\\\\gamma$ | $\\\\gamma=0.1$ | $\\\\gamma=0.3$ | $\\\\gamma=0.5$ | $\\\\gamma=0.7$ | $\\\\gamma=0.9$ |\\n | :---: | :---: | :---: | :---: | :---: | :---: |\\n | $\\\\alpha=0.1$ | 74.17 | 73.97 | 73.74 | 73.54 | 73.33 |\\n | $\\\\alpha=0.3$ | 74.25 | 74.80 | 74.68 | 74.28 | 73.95 |\\n | $\\\\alpha=0.5$ | 74.44 | 74.43 | 74.80 | 74.66 | 74.31 |\\n | $\\\\alpha=0.7$ | 74.44 | 74.33 | 74.29 | 74.33 | 74.78 |\\n | $\\\\alpha=0.9$ | 74.15 | 74.52 | 74.07 | 74.71 | 74.49 |\\n\\n *Table R1.2: Sensitivity of $\\\\beta-\\\\gamma$. Teacher is WRN-40-2 and student is WRN-40-1.*\\n | $\\\\beta-\\\\gamma$ | $\\\\gamma=0.1$ | $\\\\gamma=0.3$ | $\\\\gamma=0.5$ | $\\\\gamma=0.7$ | $\\\\gamma=0.9$ |\\n | :---: | :---: | :---: | :---: | :---: | :---: |\\n | $\\\\beta=0.1$ | 74.15 | 74.52 | 74.07 | 74.71 | 74.49 |\\n | $\\\\beta=0.3$ | 74.38 | 74.15 | 74.48 | 74.76 | 74.66 |\\n | $\\\\beta=0.5$ | 74.47 | 74.84 | 74.06 | 74.55 | 74.29 |\\n | $\\\\beta=0.7$ | 74.23 | 74.51 | 74.41 | 74.65 | 74.75 |\\n | $\\\\beta=0.9$ | 74.11 | 74.31 | 74.72 | 74.41 | 74.81 |\\n\\n---\\n\\n**Weakness2. Relation with Other Different Distillation Loss is Not Clear:**\\n\\n- The difference between ranking loss and KD loss is not whether they are applied before-softmax or after-softmax. In fact, they are both essentially constraints on logits, as discussed in **Sec. 4.2.2**. The role of softmax is mainly to scale logits uniformly, while our ranking loss does not require uniform scaling to function. We do not apply softmax to the ranking loss because softmax does not alter the ranking of the channels and may increase computational overhead during gradient calculation, which is unnecessary for ranking loss.\\n\\n- KD loss focuses more on intra-channel matching, while ranking loss emphasizes inter-channel relationship matching. Additionally, KD loss generally prioritizes large-value channels (as discussed in **Eq. 1 of the paper**), whereas ranking loss considers all channels in a balanced manner. We believe that ranking loss provides rich inter-class knowledge, which complements KD loss by facilitating a more effective optimization process.\\n\\n---\\n\\n**Weakness3. Dense Notation and Lack of Intermediate Steps:**\\n\\n- We appreciate your reminding. We add intermediate steps and more instructions to the derived formulas of **Section4.1** and **Appendix A.4** in the updated pdf to help understanding.\\n\\n---\\n\\n**Question1. Ablation of Different Initializations:**\\n\\n- In fact, we have conducted ablation on different initializations, and we believe that different teacher-students have a greater impact on the optimal $k$ value than the impact of initialization. We present the ablation experiments with different random initialization seeds in **Table R1.3**, our method works well when k is greater than 2 in general.\\n \\n *Table R1.3: Ablation on Different Initialization. Teacher is ResNet32\\u00d74 and student is ResNet8\\u00d74*\\n | | seed=152 | seed=386 | seed=2347 |\\n | :---: | :---: | :---: | :--: |\\n | $k=0.1$ | 74.01 | 73.80 | 73.54 |\\n | $k=0.5$ | 74.24 | 74.48 | 74.45 |\\n | $k=1$ | 74.74 | 74.74 | 74.44 |\\n | $k=2$ | 75.35 | 75.37 | 74.98 |\\n | $k=4$ | **75.59** | **75.65** | **75.37** |\\n | $k=6$ | 75.44 | 75.51 | 75.31 |\\n\\n---\\n\\n**Question2. Ranking Loss with RKD:**\\n\\n- We add RKD[1] method to the ranking loss and performed ablation experiments, the results are shown in Table R1.4, which also illustrates that the improvement brought by our approach is consistent and general.\\n\\n *Table R1.4: Experiments on Combining Other Metods.*\\n | Teacher -> Student | ResNet32\\u00d74 -> ResNet8\\u00d74 | ResNet50 -> MobileNet-V2 | WRN-40-2 -> WRN-40-1 |\\n | :---: | :---: | :---: | :---: |\\n | RKD | 72.61 | 59.70 | 72.29|\\n | RKD+Ours | 73.97 | 69.39 | 73.93|\\n\\n[1] Relational knowledge distillation. CVPR 2019\"}", "{\"metareview\": \"The paper introduces a plug-and-play ranking loss based on Kendall\\u2019s \\u03c4 Coefficient to enhance existing KL divergence-based distillation methods. The authors highlight that gradients provided by KL divergence are influenced by channel scale and often overlook low-probability channels. Experiments conducted on CIFAR-100, ImageNet, and COCO demonstrate consistent improvements across various teacher-student combinations.\\n\\nThis paper received mixed scores. While the experiments demonstrate that the proposed ranking loss improves classification performance on several baseline models, there are significant concerns that need to be addressed.\\n\\nFirstly, the paper lacks sufficient comparisons and explanations regarding prior methods. For instance, Figure 2 in this paper bears a strong resemblance to LSKD [1], yet there is no comparison or detailed analysis with LSKD on ImageNet. Instead, only results on CIFAR-100 are provided, which is less convincing for a robust evaluation. In addition, the proposed ranking loss shares a similar intent with the intra-class relations introduced in DIST [2]. However, the paper lacks the necessary analysis and experiments to demonstrate whether the ranking loss can further enhance DIST or achieve better performance compared to combining MLKD with DIST. This absence raises concerns about the novelty and the competitive positioning of the proposed method.\\n\\n[1] Sun, Shangquan, et al. \\\"Logit standardization in knowledge distillation.\\\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024.\\n\\n[2] Huang, Tao, et al. \\\"Knowledge distillation from a stronger teacher.\\\" Advances in Neural Information Processing Systems 35 (2022): 33716-33727.\\n\\nSecondly, there is an inconsistency between the paper\\u2019s claims and its proposed method. The authors critique KL divergence for overlooking low-probability channels. However, the proposed ranking loss focuses on enforcing ranking consistency between the teacher and student models, without clearly demonstrating how it addresses the role or importance of low-probability channels.\\n\\nOverall, these gaps raise concerns about the clarity of the paper\\u2019s contributions.\", \"additional_comments_on_reviewer_discussion\": \"Both Reviewer eu6q and 6ebV raised concerns regarding the core intention of the proposed method. The main objective of the method is to enforce ranking consistency between the teacher and student models. However, there is no clear evidence that it increases the emphasis on information from lower-probability channels. In the rebuttal, the authors provided an additional Figure 7 in the appendix, which illustrates that the gradient of the original KL divergence is affected by channel scale. While this observation is straightforward from Eq. 1, it fails to demonstrate that the proposed ranking loss enhances attention on channels with smaller logits. Furthermore, the explanation of the gradients is unclear.\\n\\nAdditionally, the ACs noted that the paper lacks sufficient comparisons and explanations regarding prior methods. For instance, Figure 2 in this paper strongly resembles Figure 2 in LSKD (https://arxiv.org/pdf/2403.01427). However, no comparison or detailed analysis with LSKD is provided on ImageNet, which undermines the robustness of the evaluation. Without such comparisons, the claims of the proposed method remain less convincing.\"}", "{\"summary\": \"This paper addresses the problem of knowledge distillation by highlighting the limitations of traditional KL divergence. The proposed method introduces an auxiliary loss based on Kendall\\u2019s \\u03c4 Coefficient, which enhances the learning of inter-class relationships and low-probability channels. Experiments conducted on three image classification datasets demonstrate the effectiveness of this approach.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": [\"The proposed method is straightforward and can be seamlessly integrated with logits-based knowledge distillation techniques.\", \"Experiments are conducted using both CNNs and ViTs across three different datasets. The ablation studies offer valuable insights into the proposed method.\"], \"weaknesses\": [\"Some claims lack adequate justification. For instance, it remains unclear how the proposed method resolves the suboptimal problem depicted in Figure 1. Including visual comparisons of logits with and without the ranking loss could enhance clarity and understanding.\", \"The proposed method includes multiple hyperparameters; however, the observed performance improvements are limited. Furthermore, the proposed method is evaluated against several straightforward baseline methods for knowledge distillation.\"], \"questions\": \"See weakness.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Thank You for Your Valuable Feedback\", \"comment\": \"Dear Reviewer pmPS,\\n\\nThank you very much for raising the score! We greatly appreciate the time and effort you have taken to help us improve our work. We will continue to refine and enhance our work accordingly.\\n\\nBest regards,\\n\\nAuthors\"}", "{\"comment\": \"We sincerely thank you for the in-depth feedback! We highly value each of your comments, and all your concerns are addressed point by point:\\n\\n---\\n\\n**Weakness1. More Clear Indication of Increased Emphasis on Smaller Channels:**\\n\\n- We add a new figure showing the gradients obtained with different channel probabilities, shown in **Figure 7 in the appendix in the paper**, to confirm our claim. The figure shows that the gradient of KL divergence is significantly affected by the channel scale, providing a smaller gradient at smaller channels. In contrast, the gradient provided by the ranking loss is relatively independent of the channel scale, providing a gradient comparable to the larger channel at smaller channels. This illustrates the increased emphasis on low-probability channels in our method.\\n\\n---\\n\\n**Weakness2. More Visualizations:**\\n\\n- As mentioned in the previous paragraph, we add to the paper a visual analysis of the gradient provided by the ranking loss and KL divergence for different logits, presented in **Figure 7 in the paper**. In addition, we include a comparative visualization of the alignment of student and teacher ranking, shown in **Figure 8 in the paper**. The visualization of the gradient shows that the proposed method improves the attention to smaller channels, while the ranking alignment shows that the proposed method aligns larger channels and smaller channels in a balanced way.\\n\\n---\\n\\n**Weakness3. Explanation of LSKD:**\\n\\n- LSKD is a method to improve the temperature parameter of students and teachers. By modifying the temperature parameter in KD loss, LSKD alleviates the problem that KD loss is too strict on students, which is a different idea from our method of providing auxiliary information for KD loss to help learning.\\n- Our method outperforms LSKD on most experiments. Both LSKD and our method are plug-and-play methods, and the LSKD reported in the paper is actually also MLKD+LSKD. Despite the excellent performance of LSKD, our method still outperforms LSKD on most student-teacher structures (7 vs. 2) in **Table 1 and Table2 in the paper**.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"title\": \"Looking forward to your reply (less than 48 hours)\", \"comment\": \"Dear reviewer eu6q,\\n\\nWe hope our responses have adequately addressed your previous concerns. The discussion period is approaching the end in 48 hours. If you have any additional comments or questions, please feel free to share them. Your feedback is greatly appreciated.\\n\\nSincerely,\\n\\nAuthors\"}", "{\"title\": \"Looking forward to your reply\", \"comment\": \"Dear Reviewer n8Cd,\\n\\nThank you very much again for your time and effort! We would greatly appreciate it if you could take a little time to check our response. We are willing to further address any remaining concerns.\\n\\nSincerely,\\n\\nAuthors\"}", "{\"summary\": \"This study presents an auxiliary ranking loss based on Kendall\\u2019s Tao Coefficient to improve knowledge distillation. The proposed ranking loss addresses the issue of KL divergence\\u2019s neglect of low-probability channels by incorporating inter-class relationship information and enhancing focus on low-probability channels. It can be integrated into any logit-based distillation method and demonstrates consistent optimization objectives with KL divergence. Experiments on three datasets across various CNN and ViT teacher-student combinations show that the ranking loss effectively improves performance across multiple baselines.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. This paper proposes a plug-and-play ranking loss to address the suboptimization issues in knowledge distillation optimization.\\n2. This paper demonstrates that Kullback-Leibler divergence is influenced by channel scale.\", \"weaknesses\": \"1. The paper claims that the proposed ranking loss primarily addresses KL divergence's tendency to overlook low-probability channels. However, based on the proposed formula, the main objective appears to be enforcing ranking consistency between the teacher and student models, with no clear indication of increased emphasis on information from smaller channels. It is recommended that the author explain this aspect.\\n2. In the experimental section, it is recommended to include visualization experiments to highlight the primary contribution\\u2014improved attention to low-probability channels. \\n3. Since LSKD shows superior performance in Tables 1 and 2, further explanation of this result is advised.\", \"questions\": \"Please refer to the Strengths and Weaknesses.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Kind Reminder: Deadline for Discussion Period Approaching (About 24 Hours Left)\", \"comment\": \"Dear Chairs and Reviewers,\\n\\nThank you all for your time and consideration. We truly appreciate your valuable feedback and constructive discussions.\\n\\nWe would like to express our sincere gratitude to Reviewer pmPS for your valuable feedback and for raising your score.\\n\\nWe would also like to kindly remind Reviewers 6ebV, eu6q, and n8Cd that the discussion period is nearing its conclusion (in about 24 hours). We would greatly appreciate your responses. If you have any further questions or concerns, please feel free to let us know.\\n\\nWe are looking forward to your reply.\\n\\nWarm regards,\\n\\nThe authors\"}", "{\"title\": \"Looking forward to your reply\", \"comment\": \"Dear Reviewer eu6q,\\n\\nThank you very much again for your time and effort! We would greatly appreciate it if you could take a little time to check our response. We are willing to further address any remaining concerns.\\n\\nSincerely,\\n\\nAuthors\"}", "{\"title\": \"Looking forward to your reply (less than 48 hours)\", \"comment\": \"Dear reviewer n8Cd,\\n\\nWe hope our responses have adequately addressed your previous concerns. The discussion period is approaching the end in 48 hours. If you have any additional comments or questions, please feel free to share them. Your feedback is greatly appreciated.\\n\\nSincerely,\\n\\nAuthors\"}", "{\"comment\": \"We sincerely appreciate your reply!\\n\\n---\\n\\nWe apologize for any misunderstanding that may have occurred in our previous response. The following is a further explanation of Weakness 2, which may address your concerns:\\n\\n**W2.1 The question of applying constraints before and after the softmax function**\\n- **We emphasize that although the KL divergence is computed on the distributions obtained after softmax, consistency in these distributions is equivalent to a linear alignment of the logits before the softmax, as shown in Eq.12 in the paper.** Therefore, from the perspective of logits, the KL loss after softmax can be seen as a linear alignment constraint on the logits (before softmax). The ranking loss is used to constrain the ranking of logits (before softmax) among channels and can act as an auxiliary to the KD loss at the logit level. \\n\\n**W2.2 Whether the ranking loss is redundant**\\n- Based on our understanding, you may consider that softmax contains inter-channel relationships and therefore overlaps with the ranking loss. To address this concern, we conducted experiments where the ranking loss is applied after the softmax function, as shown in **Table R1.5**. The results indicate that the ranking loss applied after softmax still brings significant performance improvements. This suggests that the inter-channel relationships emphasized by ranking loss are not captured by softmax, indicating that the relational knowledge provided by softmax is insufficient. Our method delivers enhancements both when applied before and after softmax.\\n\\n- Regarding your comment that *\\\"If the KD loss is small enough, the ranking loss will also be small,\\\"* we would like to point out that due to the capacity differences between the teacher and student, it is challenging for the student to mimic the teacher's outputs completely. Therefore, achieving a small enough KD loss is difficult. In **Figure 3 of the paper**, we present the loss curves for both KD and KD+Ours. Although both methods exhibit small KD losses in the later stages of training (right figure), the ranking loss computed from the outputs of the KD method remains significantly higher than that of the KD+Ours (middle figure). This indicates that even towards the end of the training, the KD loss in the traditional KD method is still not small enough to ensure a minimal ranking loss. **Figure 8 in the Appendix** reinforces this point, where the channel ranking of KD is still not aligned after training (left figure). This scenario can be analyzed with the toy case in **Figure 2 of the paper**, where **student1** has channel values closer to teacher, resulting in a smaller KD loss, however, the channel ranking is still incorrect, which leads to misclassification. This illustrates that merely minimizing the KD loss without considering channel ranking may not be sufficient for optimal model performance.\\n\\n *Table R1.5: More Ablation Experiments.*\\n | Teacher -> Student | ResNet32\\u00d74 -> ResNet8\\u00d74 | ResNet50 -> MobileNet-V2 | WRN-40-2 -> WRN-40-1 | ResNet32\\u00d74 -> SHN-V1 | WRN-40-2 -> SHN-V1 |\\n | :---: | :---: | :---: | :---: | :---: | :---: |\\n | KD | 73.33 | 67.35 | 73.54 | 74.07 | 74.83 |\\n | KD+Ours(after softmax) | 74.01 | 69.37 | 74.07 | 75.02 | 75.53 |\", \"title\": \"Further explanation of Weakness 2\"}", "{\"title\": \"General Response\", \"comment\": \"We sincerely appreciate all the reviewers for their thoughtful feedback and their time. We are encouraged to see that they found the insight in our paper interesting (Rev. pmPS, eu6q), and found found the experiments to be well executed (pmPS, 6ebV, n8Cd) with valuable ablations (Rev. n8Cd). We are also delighted to see our method's effectiveness and genaralization ability are noticed by all reviewers. We have made numerous updates to the submission, hope these can adress your concerns.\\n\\nOne issue shared by multiple reviewers is the visualization of smaller channels receiving attention. We visualize the gradients and sequential alignments of all channels using different loss to illustrate smaller channels are emphasized in our method. The visualization can be seen in **Figure 7 and Figure 8 in the paper**.\\n\\nAnother widely concerned issue is the ablation experiment combined with other methods. We combine the feature-based method RKD and the logit-based but without KL divergence method DIST and improved their performance, which are shown in **G1.1 and G1.2**, to further demonstrate the effectiveness and generalization abllity of our method.\\n\\n*Table G1.1: Experiments on Combining Other Metods.*\\n| Teacher -> Student | ResNet32\\u00d74 -> ResNet8\\u00d74 | ResNet50 -> MobileNet-V2 | WRN-40-2 -> WRN-40-1 |\\n| :---: | :---: | :---: | :---: |\\n| RKD | 72.61 | 59.70 | 72.29|\\n| RKD+Ours | 73.97 | 69.39 | 73.93|\\n\\n*Table G1.2: Experiments on Combining Other Metods.*\\n| Teacher -> Student | ResNet32\\u00d74 -> WRN-16-2 | ResNet32\\u00d74 -> WRN-40-2 | WRN-40-2 -> SHN-V1 | VGG13 -> VGG8 | WRN-40-2 -> WRN-40-1 |\\n| :---: | :---: | :---: | :---: | :---: | :---: |\\n| DIST[1] | 75.58 | 78.02 | 76.00 | 73.80 | 74.73 |\\n| DIST[1]+Ours | 75.85(+0.27) | 78.54(+0.52) | 76.23(+0.23) | 74.06(+0.26) | 74.86(+0.13) |\\n\\nAdditionally, I would like to bring your attention that there are less than 4 days until the end of discussion period, please feel free to let us know if you have any further questions, we will try our best to address your concerns.\\n\\nWe are looking forward to your reply.\"}", "{\"comment\": \"We sincerely thank you for the in-depth feedback! We highly value each of your comments, and all your concerns are addressed point by point:\\n\\n---\\n\\n**Weakness1. The Main Novelty:**\\n\\n- Our novelty can be summarized in three points:\\n 1. We propose to apply ranking loss to logit distillation for the first time and propose a generic plug-and-play loss. It can be applied to various logit-based methods or methods that use KL divergence / modified KL and improve performance.\\n 2. We present a detailed mathematical analysis of an long-standing problem in logit distillation (KL divergence neglects small-valued channels) and demonstrate, both mathematically and visually, that ranking loss can help solve this problem.\\n 3. We conduct experiments and ablations in various settings such as ViT structure and detection task to verify the effectiveness and generalization of our method.\\n\\n---\\n\\n**Weakness2. More Ablation Experiments and Analysis:**\\n\\nQuestion1. Minimize Computational Overhead:\\n\\n- Since the computational overhead of the ranking loss is highly correlated with the number of channels, the best way to minimize the computational overhead is to reduce the channels involved in loss calculation. In **Figure 5 in the paper**, using only the top/min 50% of the channels to calculate the ranking loss achieves similar results as using the full channel, while the overhead of the ranking loss is only about 25% of the full channel case (the channel pairs involved in the calculation are changed from $ \\\\frac{C(C-1)}{2} $ to $ \\\\frac{\\\\frac{C}{2}(\\\\frac{C}{2}-1)}{2} $). Even only use 30% of the channels with about 9% computational overhead can also achieve a good performance.\\n\\nQuestion2. Class Imbalance Scenarios:\\n\\n- In our current research, our method primarily addresses the imbalance in the value of channel outputs, the effectiveness is rooted in addressing value-related imbalances rather than those caused by amount of data. Therefore, this method is not designed for class imbalance scenarios.\\n- However, the proposed method has the potential to help solve tasks in class imbalance scenarios. By capturing the ranking relationship, our method can provide more inter-class information for imbalanced classes, helping to learn more from a small amount of data and improve the performance.\\n\\n\\nQuestion3. Ablation Study on Combining Methods:\\n\\n- According to our experiments combined with other methods, which show in **Tables 1 and 2 in the paper**, the proposed ranking loss improves the performance when combined with various methods (e.g., KD, DKD, MLKD) and we have not identified adverse effects of proposed method currently. \\n- Our method is a general, simple, plug-and-play method, and as far as we know, it can combine well with most methods. We propose the auxiliary supervision from a new perspective and bring additional performance improvements. Although in a few methods may be limited, there are still improvements.\\n- In fact, compared to the linear alignment of the traditional distillation loss, our loss focuses on the ranking alignment, which can be characterized by a relatively softer nature and therefore more robust. Moreover, the mathematical derivation of **Section 4.2.2 in the paper** also shows that our method will not have a large adverse effect on logit distillation.\\n- To further verify the generality of our method, we add ranking loss on a distillation method that does not use KL divergence and improve the performance, results are presented in **Table R2.1**.\\n\\n *Table R2.1: Experiments on Combining Other Metods.*\\n | Teacher -> Student | ResNet32\\u00d74 -> WRN-16-2 | ResNet32\\u00d74 -> WRN-40-2 | WRN-40-2 -> SHN-V1 | VGG13 -> VGG8 | WRN-40-2 -> WRN-40-1 |\\n | :---: | :---: | :---: | :---: | :---: | :---: |\\n | DIST[1] | 75.58 | 78.02 | 76.00 | 73.80 | 74.73 |\\n | DIST[1]+Ours | 75.85(+0.27) | 78.54(+0.52) | 76.23(+0.23) | 74.06(+0.26) | 74.86(+0.13) |\\n\\n[1] Knowledge distillation from a stronger teacher, NeurIPS 2022\"}", "{\"comment\": \"We sincerely thank you for the in-depth feedback! We highly value each of your comments, and all your concerns are addressed point by point:\\n\\n---\\n\\n**Weakness1. Some Claims Lack Adequate Justification:**\\n\\n- In fact, we want to use **Figure 1 in the paper** to illustrate smaller channel is a large proportion in logit distillation, and the suboptimal problem is actually shown **in Figure 2 in the paper**. We add a visualization of teacher-student ranking alignment, presented in **Figure 8 in the paper**, to prove that our method contributes to align ranking across all channels. The part in the bottom left corner of the figure illustrates that our method helps the top ranked channels to find their correct position, which helps to solve the suboptimal problem in Figure 2.\\n- We include visual comparisons of logits in **Figure 10 in the paper**. In addition, **Figure 3 in the paper** already shows that the channels output by our method have smaller KL loss, indicating that the channels are more aligned, and the newly added Figure 8 can also prove this point to some extent.\\n\\n**Weakness2. Other Limitations:**\\n\\n- We only introduce 2 hyperparameters: Our method only introduces two hyperparameters, weight and steepness coefficient, and the remaining hyperparameters are baseline settings, which we do not make any modifications. In addition, the weight of our method are set to inherit from baselines in experiments to ensure generalization and robustness.\\n- Our method exhibits consistent improvements: Our method is plug-and-play and can be combined with various methods (e.g. KD, CTKD, DKD, MLKD) to improve the performance, as shown in **Table 1 and Table 2 in the paper**. In comparative experiments, our method outperforms MLKD+LSKD, a hihglight of CVPR2024, demonstrating the effectiveness of our method.\\n- We conduct experiments with multiple SOTA baselines: In **Table 1 and Table 2 in the paper**, we combine our method with KD, CTKD, DKD, MLKD. Where KD is the classical method of logit distillation, DKD is the SOTA of 2022 and MLKD is the SOTA of 2023.\\n- Moreover, we conduct experiments with a method that does not use KL divergence and improve the performance, shown in **Table R2.1**, to further illustrate the generalization of our method.\\n\\n *Table R2.1: Experiments on Combining Other Metods.*\\n | Teacher -> Student | ResNet32\\u00d74 -> WRN-16-2 | ResNet32\\u00d74 -> WRN-40-2 | WRN-40-2 -> SHN-V1 | VGG13 -> VGG8 | WRN-40-2 -> WRN-40-1 |\\n | :---: | :---: | :---: | :---: | :---: | :---: |\\n | DIST[1] | 75.58 | 78.02 | 76.00 | 73.80 | 74.73 |\\n | DIST[1]+Ours | 75.85(+0.27) | 78.54(+0.52) | 76.23(+0.23) | 74.06(+0.26) | 74.86(+0.13) |\\n\\n[1] Knowledge distillation from a stronger teacher, NeurIPS 2022\"}" ] }

[ "Damiano Sgarbossa", "Cyril Malbranke", "Anne-Florence Bitbol" ]

[ "proteins", "protein sequence", "protein language model", "computational biology", "generative model", "protein engineering", "protein fitness prediction", "protein design" ]

{ "note_id": [ "y8e6YiLaFs", "wqbsKov5lX", "wkb3rNlKkX", "u9BqeUaoPk", "klcZZrp0yF", "enuP7inO96", "V3HOLu3SK4", "UEwiI62VKC", "TJO31l4s3q", "QJ0Nqp7xVp", "NewZ2HqCFz", "Kl6bIbEVQD", "J8v9OeOxye", "E8PSrbWmuY", "E6XLfdYZaW", "95Lr9EWLop", "6gZmni9KIF", "6QpSlJeOnG", "3ftrzc0lLN", "2zuI0QwMUp", "1srmfWXXKk" ], "note_type": [ "official_review", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "decision", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "meta_review", "official_comment", "official_comment", "official_comment" ], "note_created": [ 1729538301896, 1732056625384, 1732515298531, 1730554703595, 1732335052755, 1732053149414, 1730000290740, 1732055758867, 1732056907765, 1737523799830, 1732899200847, 1730688438971, 1732054959007, 1732530174592, 1732593480363, 1732054871643, 1732899112109, 1734752801266, 1732459541465, 1732530277912, 1732052945198 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission6891/Reviewer_Ua6N" ], [ "ICLR.cc/2025/Conference/Submission6891/Authors" ], [ "ICLR.cc/2025/Conference/Submission6891/Reviewer_Ua6N" ], [ "ICLR.cc/2025/Conference/Submission6891/Reviewer_SxMF" ], [ "ICLR.cc/2025/Conference/Submission6891/Reviewer_biCn" ], [ "ICLR.cc/2025/Conference/Submission6891/Authors" ], [ "ICLR.cc/2025/Conference/Submission6891/Reviewer_biCn" ], [ "ICLR.cc/2025/Conference/Submission6891/Authors" ], [ "ICLR.cc/2025/Conference/Submission6891/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission6891/Authors" ], [ "ICLR.cc/2025/Conference/Submission6891/Reviewer_tE4Y" ], [ "ICLR.cc/2025/Conference/Submission6891/Authors" ], [ "ICLR.cc/2025/Conference/Submission6891/Authors" ], [ "ICLR.cc/2025/Conference/Submission6891/Reviewer_tE4Y" ], [ "ICLR.cc/2025/Conference/Submission6891/Authors" ], [ "ICLR.cc/2025/Conference/Submission6891/Authors" ], [ "ICLR.cc/2025/Conference/Submission6891/Area_Chair_8Dox" ], [ "ICLR.cc/2025/Conference/Submission6891/Authors" ], [ "ICLR.cc/2025/Conference/Submission6891/Authors" ], [ "ICLR.cc/2025/Conference/Submission6891/Authors" ] ], "structured_content_str": [ "{\"summary\": \"The paper presents ProtMamba, a homology-aware but alignment-free protein language model. It's based on the Mamba architecture and trained on concatenated homologous sequences. Results show its effectiveness in various tasks like sequence generation and fitness prediction.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": [\"The authors propose a new training strategy, which effectively harnesses evolutionary information from homologous sequences without relying on MSA.\", \"The architecture based on Mamba blocks allows for handling extremely long contexts, which is beneficial for protein modeling as concatenating homologous sequences often results in long inputs.\", \"The results are comprehensive and prove the effectiveness of the proposed model.\"], \"weaknesses\": [\"I am not certain whether combining protein language with mamba can be regarded as \\\"novel\\\", but it is ok for me since such combination is not explored yet. An interesting aspect of this study is the training paradigm, which might provide insights for future studies. Nevertheless, a disappointing point is that almost no ablation study of the method can be found.\", \"The mask strategy (span-mask) is similar to that of T5 (you'd better add the missing reference: Raffel et al, Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer). There are some most related strategies that are not ablated, such as:\", \"what if training with token-level mask instead of span-mask? Token-level mask means something like: \\\"a b c <m1> <m2> <m3> g h <eos> <m1> d <m2> e <m3> f\\\".\", \"what if no masking strategy is used and the model is trained in an autoregressive fashion?\", \"what if doing mask-prediction without observing subsequent tokens, which means the input is \\\"a b c <m1> <m2> <m3> g h\\\" while the target is \\\"b c d e f g h <eos>\\\"?\", \"I am not requesting the authors to ablate all of the above. However, for an AI conference, that would be very interesting and no ablation is unacceptable (in my opinion).\", \"The authors claim that the incorporation of position embedding and the concatenation strategy are important. I think the authors can present some results for comparison, for example, a model without position embedding and a model with the addition strategy (of the position embedding).\", \"Absence of comparison with a transformer (with FlashAttention) in the same setting.\", \"Have you tried to scale up the parameter of the architecture?\"], \"questions\": \"NA\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response (1/1)\", \"comment\": \"We thank the reviewer for their valuable feedback on our paper. We performed multiple novel ablations and comparisons that we report in the main comment of the rebuttal. We also address the reviewer\\u2019s specific comments and concerns below:\\n1. **Positional encoding:** We performed ablations on the positional encoding and reported the results in the common answer to all reviewers. These ablations show that there is not much difference between the different ways of using positional encoding and not using positional encoding. The reason why we decided to employ positional encodings is that this allows us to control the length of the patches generated using the FIM objective during inference. Concretely, as shown in Fig 1 of the paper, by artificially skipping from position N and N+3 (after the mask token) in the main sequence, we are able to give the model the instruction of generating 3 FIM tokens as output. This would not be possible if we used no positional encoding. To summarize, positional encodings do not deteriorate performance and allow us to perform inpainting with length control.\\n2. **Benchmarking against a transformer:** The reason why we did not benchmark ProtMamba against a vanilla transformer is that a transformer is not able to handle such a long context during training. Even a small transformer model with a much more limited context length would require at least an order of magnitude more compute and training time than a Mamba model. For this reason, we decided to rather provide a comparison with PoET, which is a published transformer model that takes homology into account.\\n3. **ProteinGym benchmark:** We agree with the reviewer that MSA information is still very important for performance on the ProteinGym benchmark. Accordingly, in our case, ProtMamba with retrieval improves over ProtMamba. We also show that ProtMamba without retrieval performs nearly as good as MSA-based methods like MSA-Transformer, with a fraction of their compute, and without using MSAs, which are not always available / not always have a good quality (as shown in point 6 of our answer). In addition, we believe that the usefulness of ProtMamba goes beyond this benchmark. Indeed, ProtMamba is a generative model and can use homology information to conditionally generate novel sequences, while the MSA-based models in Table 1 cannot do so. Furthermore, in our common answer to reviewers, we reported the performance of ProtMamba on the conditional generation task, showing that it outperforms the MSA-based state of the art method EvoDiff. This confirms the interest of an MSA-free model like ProtMamba. \\n4. **Estimation of compute:** We estimated the compute by using the formula: \\nCompute = (Training time) \\u00d7 (# of GPUs) \\u00d7 (Peak FLOP/s of GPU) \\u00d7 (Utilization rate, in our case ~80%).\\n5. **Poisson distribution for masking:** Our choice was based on the necessity of having a low number of masks, given that in the typical use-cases of the model (e.g. protein design), we expect that one will usually not want to mask more than a few regions (e.g. binding pocket) of the sequence to inpaint them. Besides, we do not expect this choice to have a strong impact on the training process.\\n6. **Performance on some specific ProteinGym datasets:** We further analyzed the performance of the model on the dataset pointed out by the reviewer, namely VRPI_BPT7_Tsuboyama_2023_2WNM, and on others. We noticed that in all ProteinGym datasets in which ProtMamba without retrieval outperforms ProtMamba with retrieval, all the other methods that are MSA-independent also perform better than the MSA-dependent ones. This suggests that in these cases, alignments are not very useful to score variants, possibly because of issues in these alignments. This constitutes another example of the importance of using an MSA-free but homology aware model.\"}", "{\"comment\": \"I have read the response and appreciate the authors' efforts to conduct further experiments and revise the manuscript. I would like to increase my score.\\n\\nIt should be noted that a score of 6 results from the training strategy, rather than adopting an Mamba architecture. I believe that the training strategy is orthogonal to the Mamba framework. I am of the opinion that a Transformer can perform the same task (even better), considering that there are numerous large language models (LLMs) capable of handling long sequences (128k), and their size is larger than that of the model in this study. I strongly suggest that the authors add a Transformer architecture for comparison.\\n\\nAdditionally, increasing the model size should also be an important future work. \\n\\nThank you.\"}", "{\"summary\": \"The paper proposed a Mamba-based protein language model, ProtMamba, using concatenated sequences from protein families, with a FIM training objective. This approach allows for faster training and inference speeds. Experiments demonstrate ProtMamba\\u2019s versatility across protein fitness prediction and context-conditioned generation.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": [\"**Novelty**: This is one of the first works to incorporate state-space models (SSMs) in protein language modeling, utilizing the Mamba architecture for efficient long-context handling.\", \"**Innovative Input Design**: The input consists of a concatenation of unaligned homologous sequences separated by CLS tokens, with a carefully designed masking strategy. This design effectively leverages long homology contexts, maximizing the model\\u2019s ability to capture evolutionary information. Training with a Fill-In-The-Middle (FIM) objective enables flexible application to tasks like mutational effect prediction.\", \"**Comprehensive Model Implementation and Training**: The authors have put substantial effort into implementing and training ProtMamba, incorporating techniques inspired by DNA modeling, such as callback mechanisms and sequence length warmup.\"], \"weaknesses\": \"- **Performance**: ProtMamba does not show significant performance improvements over strong baselines such as ESM-2 and Tranception, which may limit its competitive edge.\\n\\n- **Additional Comparisons**: Including comparisons with other baseline models, such as PoET-205M[1], SaProt[2], ProtHyena[3], or PTM-Mamba[4] would provide a fair evaluation and offer a more comprehensive view of ProtMamba\\u2019s strengths and weaknesses.\\n\\n- One advantage of Mamba is its faster generation capability compared to transformer-based models. The authors could extend ProtMamba\\u2019s use cases by addressing protein sequence generative tasks, such as unconditional generation. A more detailed discussion in Section 3.4 comparing ProtMamba to other generative models would strengthen the paper. You could follow the setting and metrics in PROTEINBENCH[5] paper.\\n\\n\\n[1] Truong Jr, T., & Bepler, T. (2023). Poet: A generative model of protein families as sequences-of-sequences. Advances in Neural Information Processing Systems, 36, 77379-77415.\\n\\n[2] Su, J., Han, C., Zhou, Y., Shan, J., Zhou, X., & Yuan, F. (2023). Saprot: Protein language modeling with structure-aware vocabulary. bioRxiv, 2023-10.\\n\\n[3] Zhang, Y. (2024). Prothyena: A fast and efficient foundation protein language model at single amino acid resolution. bioRxiv, 2024-01.\\n\\n[4] Peng, Z., Schussheim, B., & Chatterjee, P. (2024). PTM-Mamba: A PTM-aware protein language model with bidirectional gated Mamba blocks. bioRxiv.\\n\\n[5] Ye, F., Zheng, Z., Xue, D., Shen, Y., Wang, L., Ma, Y., ... & Gu, Q. (2024). ProteinBench: A Holistic Evaluation of Protein Foundation Models. arXiv preprint arXiv:2409.06744.\", \"questions\": [\"I\\u2019m curious about the impact of the number of context sequences on ProtMamba\\u2019s performance. For example, how does performance change with 0, 5, or more sequences, and is there a threshold beyond which additional context sequences no longer contribute to model performance? In your experiments on scaling FIM perplexity with the number of context sequences, it seems perplexity stabilizes with around 30 context sequences. Could you elaborate on this?\", \"Inference Efficiency: Could you report on ProtMamba\\u2019s efficiency during inference? Additionally, how does ProtMamba\\u2019s performance, memory usage, and computational efficiency scale with increasing context length compared to transformer-based models, like PoET?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you, authors, for your responses. However, I believe that both the Transformer and Transformer++ models are capable of handling such input sizes. The original Mamba paper noted: \\\"All attention models were only tested up to a sequence length of 2^14 = 16384 due to memory limitations.\\\" So at least for first stage of 2^11 in your case, it is possible to test them. For any long-context modeling problem, I believe it is essential to benchmark against other architectures. Examples in the bioi-related field include HyenaDNA or the recently developed Evo-1.\"}", "{\"title\": \"Response to all reviewers (2/2)\", \"comment\": \"3. **Comparison with other models on the ProteinGym benchmark:** We report a more detailed comparison with PoET and other models on the ProteinGym benchmark, together with a comparison of inference time and number of parameters between PoET and ProtMamba. Unfortunately, this comparison is partially incomplete as the training time and FLOPs for PoET are not available.\\nWe furthermore add a comparison on the inference time of PoET and ProtMamba, showing that ProtMamba has a 85-fold improvement with respect to PoET (single) and a 1300-fold improvement with respect to PoET (ensemble) on the complete ProteinGym benchmark. \\n\\n| **Model type** | | **Model** | | **#params** | | **Spearman $\\\\rho$** | | **Time** |\\n|-----------------------|-|-------------------------------|-|-------------|-|-----------------|-|-----------------|\\n| **Alignment-based** | | Site-Independent | | - | | 0.359 | | - |\\n| | | GEMME | | - | | 0.455 | | - |\\n| |\\n| **PLM** | | Tranception L (w/o R) | | 700M | | 0.374 | | - |\\n| | | ESM-2 | | 150M | | 0.387 | | - |\\n| | | ESM-2 | | 650M | | 0.414 | | - |\\n| |\\n| **Homology-aware** | | ProtMamba (single) | | 107M | | 0.406 | | **7m** |\\n| **PLM** | | ProtMamba AR (single) | | 107M | | 0.367 | | 1h 39m |\\n| | | PoET (single) | | 201M | | 0.447 | | 9h 51m |\\n| | | PoET (ensemble) | | 201M | | 0.470 | | 148h |\\n| |\\n| **Alignment + PLM** | | ProtMamba (w/ R) | | 107M | | 0.432 | | 10m |\\n| | | MSA-Transformer | | 100M | | 0.421 | | - |\\n| | | Tranception L (w/ R) | | 700M | | 0.434 | | - |\\n| | | VespaG | | 3B | | 0.458 | | - |\\n| |\\n| **Structure-aware** | | ESM-IF1 | | 142M | | 0.422 | | - |\\n| | | SaProt | | 650M | | 0.457 | | - |\\n| | | ProSST | | 110M | | **0.507** | | - |\\n\\n[1] Sarah Alamdari et al. Protein generation with evolutionary diffusion: sequence is all you need.\"}", "{\"summary\": \"The paper introduces **ProtMamba**, a novel protein language model that is homology-aware but alignment-free, addressing the limitations of traditional multiple sequence alignments (MSAs) in protein modeling. ProtMamba is built on the Mamba architecture, which enables it to handle very long sequences by efficiently processing concatenated homologous protein sequences. The model is trained using a hybrid of autoregressive modeling and Fill-in-the-Middle (FIM) objectives, making it highly versatile for tasks like protein sequence generation and mutational fitness prediction. ProtMamba demonstrates competitive performance on benchmarks like ProteinGym, outperforming similar-sized models in terms of efficiency and predictive accuracy. Additionally, the model excels in sequence generation tasks, producing novel sequences with structural properties comparable to natural proteins.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": [\"ProtMamba is Homology-aware yet alignment-free\", \"Mamba architecture adaptation efficiently handle long contexts\", \"Explores hybrid training scheme for pLMs\", \"ProtMamba strong results on ProteinGym for mutational fitness prediction\", \"ProtMamba can generate reasonable sequences given homology context/sequences\"], \"weaknesses\": \"1. Why use position encoding for ProtMamba? Given the recurrent nature of Mamba, positional information should theoretically be learned implicitly, which is why the original Mamba model does not employ explicit position encodings. The authors claim this is a significant modification, yet they fail to provide any experiments or ablation studies to demonstrate how this change improves model performance. I would like to see a comparison of with positional encoding (PE) vs. without PE, and additionally, a comparison of standard PE (additive) vs. the concatenation method used in ProtMamba. Authors should provide more details about PE implementation and comparison.\\n\\n2. Although the motivation for using a long-context language model like Mamba is compelling, the paper does not benchmark a vanilla transformer at any context length, which is a significant weakness. Without this comparison, it is difficult to argue whether ProtMamba is the optimal architecture for this task in terms of performance. While it is clear that state space models (SSMs) like Mamba will likely outperform transformers in terms of efficiency, the lack of performance benchmarks makes it hard to assess if ProtMamba achieves the best results. \\n\\n3. Regarding the ProteinGym benchmark, when incorporating MSA or homology sequences, ProtMamba does not outperform MSA-based models. This challenges the fundamental premise of the paper that a homology-aware, MSA-free model should perform better and eliminate the need for MSAs. The lack of superior performance compared to MSA-based models suggests that ProtMamba's approach may not fully leverage homology information as effectively as MSA based models.\", \"questions\": \"1. How do the authors compute the total number of tokens and FLOPs for training? Can you provide more details on the implementation, such as whether you use any packages for FLOP calculations or how you approximate them?\\n\\n2. Why did you choose to use a Poisson distribution for masking instead of other distributions?\\n\\n3. In Figure 12, why does the model, in some DMS experiments, perform better without retrieval (MSA/homology sequences) than with MSA? This is especially surprising given that the model was trained over long contexts of homologous sequences. For instance, in VRPI_BPT7_Tsuboyama_2023_2WNM, there is a significant difference. Could you provide some explanations for this discrepancy?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response (1/1)\", \"comment\": \"We thank the reviewer for their valuable feedback on our paper. We performed multiple novel ablations and comparisons that we report in the main comment of the rebuttal. We also address their comments and concerns below:\\n1. **Performance comparisons:** In Table 1, which is present both in the common answer to all reviewers and in the new version of the paper, we show that even if ProtMamba has much fewer parameters and was trained for less time, we still outperform both ESM2 and Tranception on the ProteinGym benchmark. The only models that perform better than ProtMamba are TranceptEVE (which uses an EVE model trained on the single MSAs) and PoET (which is twice larger than ProtMamba and much slower). Furthermore, all of the models we compared with are transformer-based, and therefore their inference times are 20 to 200 fold larger than those of ProtMamba. We added a column in Table 1 to compare inference times.\\n2. **Additional comparisons:** We added a comparison with PoET and SaProt in Table 1. We did not include ProtHyena and PTM-Mamba because they have a different training objective. To the best of our knowledge, they were not tested on ProteinGym because of this.\\n3. **Generative capabilities:** In the global answer to all reviewers we reported a comparison of conditional generation between ProtMamba and EvoDiff, which we believe to be the state of the art model for this task. We show that ProtMamba strongly outperforms EvoDiff. Note that we did not compare to Progen2 and DPLM because they are not designed for homolog-conditioned generation (see our main response for more details). Besides, we did not perform completely unconditional generation (without homolog information) because it is not what ProtMamba was designed for.\\n4. **Proteinbench:** We thank the reviewer for pointing out the interesting Proteinbench benchmark. We could not include it in our analysis because it was published after the ICLR deadline. Furthermore, to the best of our knowledge, the code in the ProteinBench GitHub repository is not available yet (as of 19 Nov 2024). We will be happy to test ProtMamba on Proteinbench later. Also, we benchmarked ProtMamba generative abilities against other similar models, see point 3.\\n5. **Impact of the number of context sequences on ProtMamba\\u2019s performance:** We show in Figure 2 the scaling of perplexity with the number of sequences in the context. We would like to clarify that perplexity does not stabilize after 30 sequences, but continues to decrease as a power law. To better show this, we added in the supplement the same figure in a log-log scale (the figure can be found in the pdf named: \\\"rebuttal_attachment\\\" in new supplementary material file).\\n6. **Efficiency and memory usage during inference:** As ProtMamba is an autoregressive recurrent model, there is no limitation in memory usage in inference. Conversely, transformer-based language models like PoET have a limitation on context length due to memory usage. \\nRegarding time complexity during inference, we show that ProtMamba has a 85-fold improvement on the time to score variants in ProteinGym with respect to PoET with no ensembling. If one uses PoET with ensembling (to reach its reported performance) then the improvement is 1300-fold. See the table below for a comparison between ProtMamba and PoET on the time taken to score all variants in ProteinGym.\\nOther comparisons between Mamba-based and Transformer-based language models (like PoET) in terms of efficiency were studied in [1], which we now cite.\\n\\n| **Model** \\t|**Time to score variants** |\\n|-------------------------------|------------|\\n| ProtMamba (single) \\t|7m \\t|\\n| PoET (single) \\t| 9h 51m \\t|\\n| PoET (ensemble) \\t| 148h \\t|\\n\\n[1] R. Waleffe et al. An empirical study of mamba-based language models.\"}", "{\"title\": \"Response (1/1)\", \"comment\": \"We thank the reviewer for their valuable feedback on our paper. We performed multiple novel ablations and comparisons that we report in the main comment of the rebuttal. We also address the reviewer\\u2019s specific comments and concerns below:\\n1. **Span mask strategy:** We referenced the two previous papers from which we built our span mask strategy. We also cited the relevant reference suggested by the reviewer in the new version of the paper.\\n2. **Ablations on the masking strategy:** We performed the requested ablations (together with several others) and we reported their results in the common answer to all reviewers. In particular, we performed an ablation using a fully autoregressive model, and showed its performance compared to the others. Beyond these ablations, a purely autoregressive model, even if interesting, does not fulfill all our objectives. Indeed, the reason why we use FIM is that this allows us to perform inpainting of protein sequences, which would not be possible using a purely autoregressive model. Finally, we believe that the idea of having a mask size of 1 is quite interesting, and we performed some ablations showing that it is also good in a FIM objective. We leave exploring this direction in more detail to future works.\\n3. **More conclusive results on protein generation:** We would like to highlight that, based on other reviewers\\u2019 comments, we added a comparison with other models on the conditional protein generation task (discussed in the general rebuttal) that may be of interest for the reviewer.\\n4. **Positional encoding:** We performed ablations on the positional encoding and reported the results in the common answer to all reviewers. These ablations show that there is not much difference between the different ways of using positional encoding and not using positional encoding. The reason why we decided to employ positional encodings is that this allows us to control the length of the patches generated using the FIM objective during inference. Concretely, as shown in Fig 1 of the paper, by artificially skipping from position N and N+3 (after the mask token) in the main sequence, we are able to give the model the instruction of generating 3 FIM tokens as output. This would not be possible if we used no positional encoding. To summarize, positional encodings do not deteriorate performance and allow us to perform inpainting with length control.\\n5. **Comparison with a transformer model:** The reason why we did not benchmark ProtMamba against a vanilla transformer is that a transformer is not able to handle such a long context during training. Even a small transformer model with a much more limited context length would require at least an order of magnitude more compute and training time than a Mamba model. For this reason, we decided to rather provide a comparison with PoET, which is a published transformer model that takes homology into account.\\n6. **Model scaling:** We did not scale up the parameter count of the model yet because of hardware limitations. We believe that this work is important as a proof of concept, and we plan to train a larger model in a subsequent work.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"title\": \"Additional comparison with Transformer and Hyena baselines.\", \"comment\": \"Thank you again for your valuable suggestions. Below, we provide additional results about comparisons between other models (which was requested by reviewer **biCn** too). We will report these results in the supplementary material in later revisions.\\n\\n- **Training Setup:** We conducted a comparative evaluation of Mamba, Hyena, and GPT2 using flash attention, focusing on small-scale models, with our training framework and data. Each model was designed to have approximately 14M parameters, consisting of 8 layers with 512-dimensional representations (384 for GPT2). All models were trained with a context size of 32,000 tokens for 72 hours on single H100 GPUs. Note that we also started training a larger GPT2 model with 115M parameters. However, during the first 24 hours, the loss was decreasing much more slowly than with Mamba and Hyena, so we decided to focus on small models instead (the time necessary to converge would have been too high for the timeframe of this rebuttal).\\n\\n- **Evaluation:** We tested the perplexity of the models on a held-out validation set with two context lengths: 16,000 tokens (short context) and 128,000 tokens (long context).\\n1. For short contexts, we report the evaluation rate (number of sequences evaluated per second) and validation perplexity (PPL) to demonstrate efficiency under standard settings.\\n2. For long contexts, we report the evaluation rate and validation perplexity for GPT2 and Mamba. Hyena, however, could not be evaluated with context lengths beyond 32,000 tokens, as it is limited to context lengths defined during training.\\n\\n- **Comments on the comparison:** The table below shows that Hyena has a much higher perplexity than the other two architectures. As expected, GPT2 is better at modeling short contexts than Mamba but its performance decreases with longer contexts (not seen in training). Moreover, Mamba reaches a performance similar to that of GPT2 at longer contexts. We also show that GPT2 is 16- to 128-fold slower than Mamba during evaluation, which can be limiting for many applications (e.g. variant effect prediction, see the comparison with PoET in Table 1 of the paper). Furthermore, the small GPT2 model has an inference time which is still 4- to 34-fold larger than the larger ProtMamba model that we trained (107M).\\n\\n- **Future Directions with Transformer models:** The original approach (training protocol, dataset construction) presented in this paper is flexible and can be implemented with other architectures. In particular, using GPT2 seems indeed a promising direction for improving performance. However, this comes with substantially higher computational costs (in inference time in particular), particularly for longer contexts, where GPT2 is less efficient than Mamba. Another promising direction involves building on the recent emergence of hybrid models that mix state-space models and attention-based methods.\\n\\n| Model \\t| Training Time | Tokens Seen | Short Context PPL (\\u2193) | Eval Rate: Short Context (\\u2191) | Long Context PPL (\\u2193) |Eval Rate: Long Context (\\u2191) |\\n|--------------------|---------------|-------------|--------------------|--------------------------|-------------------|--------------------------|\\n| Hyena (14M) \\t| 72h \\t| 81.9B \\t| 14.42 \\t| 17 $s^{-1}$ \\t| Fail \\t| Fail \\t|\\n| Mamba (14M) \\t| 72h \\t| 76.8B \\t| 12.48 \\t| 33 $s^{-1}$ \\t| 11.89 \\t| 3.84 $s^{-1}$\\t|\\n| GPT2 (15M) \\t| 72h \\t| 33.3B \\t| 10.68 \\t| 2 $s^{-1}$ \\t| 11.72 \\t| 0.03 $s^{-1}$ \\t|\\n| |\\n| ProtMamba (107M) | 400h \\t| 190B \\t| 9.69 \\t| 8.33 $s^{-1}$ \\t| 9.26 \\t| 1.04 $s^{-1}$ \\t|\"}", "{\"summary\": \"The paper introduces ProtMamba, a state-space protein language model that is trained on sets of homologous sequences concatenated together. The model is trained to both generate sequences from scratch and to infill sequences using a fill-in-the-middle objective. For fitness prediction on ProteinGym, the method is shown to perform on-par with much larger models that explicitly use a multiple sequence alignment. For a narrower dataset of chorismate mutase activities, they demonstrate that they can apply prompt engineering and the FIM objective to improve fitness prediction. Finally, they perform a limited evaluation of the model\\u2019s autoregressive generation capabilities and show that the top 10% of generated sequences have some properties similar to natural proteins.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"3\", \"strengths\": \"Prompting protein language models with sequences of homologous proteins (rather than an MSA) is an exciting direction for retrieval-augmented models. Given the inefficiencies of long-context transformers, using a state-space model for this objective is a natural idea to explore.\\n\\nThe choice to couple standard autoregressive language modeling with a fill-in-the-middle objective is an interesting one that has been relatively unexplored for protein language modeling, and the authors show its value for a fitness prediction task involving chorismate mutases.\\n\\nThe authors also provide a limited demonstration that prompting the model with high-activity sequences can improve its ability to perform fitness prediction.\", \"weaknesses\": \"Major points:\", \"general_clarity\": \"I have a hard time following all the details of the paper because of the copious references to supplementary figures to support central claims in the main text.\", \"table_1\": \"Indicating the top performers for each evaluation in bold would improve the readability of the table.\", \"lines_208_212\": \"More recent work [2, 3, 4] suggests that it is beneficial to mask as much as 50% of a sequence. I would like to see ablations that evaluate different masking fractions, rather than just results for the somewhat arbitrary choice of 20%.\", \"figure_4\": \"There are major loss spikes and periods where the training loss actually increases. The authors should comment on the overall training stability of ProtMamba with some analysis of the gradient norms during training. This is important for a reader to decide whether they would choose to adopt Mamba over a transformer.\\n\\nMinor points\", \"lines_89_95\": \"Modern attention implementations like FlashAttention have linear memory complexity, though they still have quadratic time complexity [5]. The authors should update the text to reflect this fact.\", \"figures_6_7\": \"It is unclear to me what L denotes, and why the 150 < L < 250 line extends so much further than the other 2 in Figure 6.\\n\\n[1] Truong Jr. & Bepler. PoET: A generative model of protein families as sequences-of-sequences. NeurIPS, 2023.\\n[2] Wettig et al. Should You Mask 15% in Masked Language Modeling? EACL, 2023.\\n[3] Tay et al. UL2: Unifying Language Learning Paradigms. arXiv, 2022.\\n[4] Hayes et al. Simulating 500 million years of evolution with a language model. bioRxiv, 2024.\\n[5] Dao. FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning. ICLR, 2023.\", \"questions\": \"ProteinGym\\u2019s performance metrics are computed by averaging together the Spearman correlations for all assays with the same (UniProt ID, Function) pair, computing the average-of-averages for each function, and then averaging over functions. When computing the depth-based (and other) averages, I believe the UniProt IDs are averaged first as well, though not the functions. Can the authors confirm that they use the appropriate hierarchical averages to compute results for ProtMamba in Table 1?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response (2/2)\", \"comment\": \"6. **Training instabilities:** While Mamba is known to have training instabilities that cannot be fixed by clipping the gradient norm [1, 2], this problem was recently fixed in Mamba2 [3]. Consistently, our very early tests with Mamba2 do not feature such instabilities. Thus, we expect this issue to vanish in the near future. Practically, in the current version of ProtMamba, we used gradient clipping and restarted the training from a previous checkpoint to handle training instabilities (see Methods).\\n7. **Linear memory complexity of FlashAttention:** We updated the main text according to the reviewer\\u2019s comments on FlashAttention. While the reviewer is correct that FlashAttention has a linear complexity in terms of memory, it nevertheless still has a quadratic complexity in time, which is the main bottleneck of training Transformers with respect to Mamba models. We clarified this point in the new version of the paper.\\n8. **Table 1 formatting:** In the new version of the paper we did a major update in the format of Table 1, where we include many more methods. We used a gray background to highlight top performers, as suggested by the reviewer. We shared the new table 1 also in the common response to all reviewers.\\n9. **Meaning of L in Figures 6-7:** L is the average length of the sequences in the family considered. We now explicitly mention this in the figures\\u2019 captions. The 150 < L < 250 line extends more than others because the distribution of cluster sizes is not uniform. Those clusters with L<150 tend to be more shallow (fewer homologs) than those with 150<L<250, leading to a concatenated length smaller than the maximum context length (131k tokens); instead, for L>250 the limit on the context length (131k tokens) was reached when concatenating sequences. This is the reason why the curves with intermediate lengths (150<L<250) are longer than the other cases.\\n10. **Method used to compute the scores on ProteinGym:** We confirm that we used the same hierarchical method of computing the averages of the Spearman correlation scores as detailed in the ProteinGym paper. We added a section in the Supplement where we describe the methodology to score the variants. We checked that for known models, we find the same Spearman correlation as the one reported in the ProteinGym benchmark.\\n\\n[1] R. Waleffe et al. An empirical study of mamba-based language models.\\n\\n[2] E. Nguyen et al. Sequence modeling and design from molecular to genome scale with Evo.\\n\\n[3] T. Dao et al. Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality\"}", "{\"comment\": \"We thank the reviewer for raising their score, and for their constructive feedback that allowed us to significantly improve our paper with respect to the initial version.\\n\\nOur choice of Mamba over Transformer-based architectures is primarily motivated by computational efficiency. While Transformers with FlashAttention can manage longer sequences with a linear memory-complexity, their inherent quadratic time-complexity in sequence length still imposes significant computational overhead during training and inference. In contrast, Mamba\\u2019s linear time complexity enables it to handle long sequences with far greater efficiency, making it more practical for scenarios where computational resources or time are constrained. Concretely, we show that we are able to obtain a 85 to 1300-fold speedup with respect to PoET, a transformer model trained on a similar objective. We believe that building efficient models can lead to their broader use in the community, and mitigate their carbon footprint.\\n\\nTo address the reviewer\\u2019s concerns, we started training some experiments to compare ProtMamba with a small GPT2 transformer (with Flash-Attention) matching the size of the small Mamba model used in our ablations. The results of these experiments will be included in the final version of the paper and as a comment as soon as possible.\\n\\nWe fully agree with the reviewer that increasing model size is an important direction for future work. Scaling ProtMamba would likely improve its performance across tasks, making it more competitive with larger language models that can handle extended sequences. In line with this, we also plan to explore the recently introduced Mamba2 architecture, which incorporates attention layers and has demonstrated improved performance over Mamba1.\"}", "{\"comment\": \"I have read the authors' response and updated manuscript, and am happy to increase my score in light of the additional analyses performed by the authors.\"}", "{\"title\": \"Response (1/2)\", \"comment\": \"We thank the reviewer for their valuable feedback on our paper. We performed multiple novel ablations and comparisons that we report in the main comment of the rebuttal. We also address their comments and concerns below:\\n1. **General clarity:** Following the reviewer\\u2019s remark, we removed some non-crucial references to the supplement from the main text. We also added a new table that brings back some of the key results in supplementary in the main paper. Finally, we renumbered supplementary figures as \\u201cS1, S2\\u2026\\u201d for added clarity.\\n2. **Comparison with PoET (Table 1):** For completeness, we added the PoET official results to Table 1 which we also shared in the common response to all reviewers above and here.\\nAs mentioned in the main answer to all reviewers, even if PoET performs better than ProtMamba on the benchmark, an important result is that we can reach competitive performance using a fraction of the training FLOPs, a fraction of the training data and a fraction of the inference time (85 to 1300 fold reduction on the inference time on ProteinGym) as shown in the \\\"Time\\\" column (the time needed to evaluate all variants in ProteinGym). \\nUnfortunately, it is not possible to compare the results with PoET using the exact same prompt as used in ProtMamba because of the context length limitations and inference costs of PoET.\\n\\n| **Model** \\t| **#Params** | **Spearman $\\\\rho$** | **Time** |\\n|-------------------------------|-------------|------------|------------|\\n| ProtMamba (w/ R) \\t| 107M \\t| 0.432 \\t| 10m \\t|\\n| ProtMamba (single) \\t| 107M \\t| 0.406 \\t| 7m \\t|\\n| PoET (single) \\t| 201M \\t| 0.447 \\t| 9h 51m \\t|\\n| PoET (ensemble) \\t| 201M \\t| 0.470 \\t| 148h \\t|\\n\\n3. **More conclusive results on protein generation:** We would like to highlight that, based on other reviewers comments, we added a comparison with other models on the conditional protein generation task (discussed in the general response to all reviewers and in section 3.4 of the updated paper) that may be of interest for the reviewer.\\n4. **Comparison between using FIM loss and autoregressive log likelihood for ProteinGym (Table 1):** We report the comparison between using the standard autoregressive log likelihood and using the FIM loss on the ProteinGym benchmark both here and in the common response to all reviewers above. We find that using FIM yields a better performance, which demonstrates its relevance for variant effect prediction. We also note that the FIM technique allows to score all the 20 possible amino-acids at each position at the same time, while the autoregressive one cannot.\\n\\n| **Model** \\t| **#Params** | **Spearman $\\\\rho$** | **Time** |\\n|-------------------------------|-------------|------------|------------|\\n| ProtMamba AR (single)\\t| 107M \\t| 0.361 \\t| 1h 39m \\t|\\n| ProtMamba (single) \\t| 107M \\t| 0.406 \\t| 7m\\t|\\n\\n5. **Choice of masking fraction:** We thank the reviewer for raising this interesting point. We chose a masking fraction of 20% in line with the majority of the literature in Masked Language Modeling that uses ~15%. Note that our model is not strictly equivalent to a model trained with MLM: as it is trained to predict residues autoregressively, both in FIM and not in FIM, it has seen a range of masking fractions. We now propose an \\u201cablation\\u201d using 50% as a masking fraction, and obtain a minor performance improvement, as reported in the general rebuttal. We will take this into account for the new versions.\"}", "{\"title\": \"Additional comparison with Transformer and Hyena baselines.\", \"comment\": \"Thank you again for your valuable suggestions. Below, we provide additional results about comparisons between other models (which was requested by reviewer **Ua6N** too). We will report these results in the supplementary material in later revisions.\\n\\n- **Training Setup:** We conducted a comparative evaluation of Mamba, Hyena, and GPT2 using flash attention, focusing on small-scale models, with our training framework and data. Each model was designed to have approximately 14M parameters, consisting of 8 layers with 512-dimensional representations (384 for GPT2). All models were trained with a context size of 32,000 tokens for 72 hours on single H100 GPUs. Note that we also started training a larger GPT2 model with 115M parameters. However, during the first 24 hours, the loss was decreasing much more slowly than with Mamba and Hyena, so we decided to focus on small models instead (the time necessary to converge would have been too high for the timeframe of this rebuttal).\\n\\n- **Evaluation:** We tested the perplexity of the models on a held-out validation set with two context lengths: 16,000 tokens (short context) and 128,000 tokens (long context).\\n1. For short contexts, we report the evaluation rate (number of sequences evaluated per second) and validation perplexity (PPL) to demonstrate efficiency under standard settings.\\n2. For long contexts, we report the evaluation rate and validation perplexity for GPT2 and Mamba. Hyena, however, could not be evaluated with context lengths beyond 32,000 tokens, as it is limited to context lengths defined during training.\\n\\n- **Comments on the comparison:** The table below shows that Hyena has a much higher perplexity than the other two architectures. As expected, GPT2 is better at modeling short contexts than Mamba but its performance decreases with longer contexts (not seen in training). Moreover, Mamba reaches a performance similar to that of GPT2 at longer contexts. We also show that GPT2 is 16- to 128-fold slower than Mamba during evaluation, which can be limiting for many applications (e.g. variant effect prediction, see the comparison with PoET in Table 1 of the paper). Furthermore, the small GPT2 model has an inference time which is still 4- to 34-fold larger than the larger ProtMamba model that we trained (107M).\\n\\n- **Future Directions with Transformer models:** The original approach (training protocol, dataset construction) presented in this paper is flexible and can be implemented with other architectures. In particular, using GPT2 seems indeed a promising direction for improving performance. However, this comes with substantially higher computational costs (in inference time in particular), particularly for longer contexts, where GPT2 is less efficient than Mamba. Another promising direction involves building on the recent emergence of hybrid models that mix state-space models and attention-based methods.\\n\\n| Model \\t| Training Time | Tokens Seen | Short Context PPL (\\u2193) | Eval Rate: Short Context (\\u2191) | Long Context PPL (\\u2193) |Eval Rate: Long Context (\\u2191) |\\n|--------------------|---------------|-------------|--------------------|--------------------------|-------------------|--------------------------|\\n| Hyena (14M) \\t| 72h \\t| 81.9B \\t| 14.42 \\t| 17 $s^{-1}$ \\t| Fail \\t| Fail \\t|\\n| Mamba (14M) \\t| 72h \\t| 76.8B \\t| 12.48 \\t| 33 $s^{-1}$ \\t| 11.89 \\t| 3.84 $s^{-1}$\\t|\\n| GPT2 (15M) \\t| 72h \\t| 33.3B \\t| 10.68 \\t| 2 $s^{-1}$ \\t| 11.72 \\t| 0.03 $s^{-1}$ \\t|\\n| |\\n| ProtMamba (107M) | 400h \\t| 190B \\t| 9.69 \\t| 8.33 $s^{-1}$ \\t| 9.26 \\t| 1.04 $s^{-1}$ \\t|\"}", "{\"metareview\": \"This paper introduces a \\\"homology-aware\\\" but \\\"alignment-free\\\" model called ProtMamba that combines protein language model and the Mamba architecture.\\nInstead of directly utilizing multiple sequence alignment (MSA), ProtMamba leverages protein language model to leverage individual homologous protein sequences without the need for their alignment.\\nThe reviewers commend this is an interesting research direction with some novelty.\\nHowever, ProtMamba doesn't always lead to significant performance improvement and does not convincingly demonstrate the advantage of the proposed scheme over other existing approaches, some of them with smaller size and utilize traditional attention-based approach.\\nOverall, while the work is promising, additional benchmarks against other existing schemes and providing further rationale and justification regarding the advantages of the proposed model architecture over other existing alternatives (esp., smaller and simpler models) would be needed to further strengthen the current work.\", \"additional_comments_on_reviewer_discussion\": \"The authors have actively engaged with the reviewers during the discussion period.\\nThe authors have provided additional experimental results and explanations that have addressed the initial concerns of the reviewers to some extent.\\nMost reviewers have responded to the authors and engaged in the discussion.\\nThe AC finds that the additional evidence provided by the authors is useful, but generally agrees with the reviewers regarding the need for additional experiments, comparison, and discussions/analysis to strengthen the manuscript.\"}", "{\"comment\": \"We thank the reviewer for their additional comments, and for suggesting a more explicit comparison to other architectures, such as transformer and Hyena-based ones. We agree that HyenaDNA and Evo-1 provide very exciting ways of addressing long context, although their objectives are quite different from those of ProtMamba (genome-level modeling versus protein family-level modeling). As the reviewer pointed out, vanilla transformers can handle sequence lengths up to 2^14 tokens in specific implementations. However, for the task of training on large protein families, which often exceed this limit, transformers still face significant computational and memory constraints.\\n\\nTo address the reviewer\\u2019s concerns, we started training some experiments to compare ProtMamba with both a small GPT2 transformer (with Flash-Attention) and a small Hyena model (matching the size of the small Mamba model used in our ablations). The results of these experiments will be included in the final version of the paper, pending their completion.\\n\\nWe would like to already emphasize that, whatever the results of these comparisons on limited context lengths, the limitations of transformer-based models on context lengths larger than those used in training were demonstrated in PoET, specifically in Figure 4 of [1]. We believe that their results are a strong argument to justify the use of other architectures for long context tasks.\\n\\nLastly, while benchmarking against transformers for limited context lengths provides an important point of comparison, we believe that by focusing on homology-aware sequence modeling without alignment, ProtMamba addresses a pressing need for flexible, efficient models in protein science. Its utility in both predictive tasks and conditional generation highlights its versatility.\\n\\n[1] Truong Jr, T., & Bepler, T. (2023). Poet: A generative model of protein families as sequences-of-sequences. Advances in Neural Information Processing Systems, 36, 77379-77415. https://arxiv.org/abs/2306.06156\"}", "{\"title\": \"Additional comparison with PoET on shorter prompts\", \"comment\": \"We would also like to add a comparison between ProtMamba and PoET where we use exactly the same prompt, i.e. using a short context of 12k amino acid tokens. We obtain:\\n\\n| **Model** \\t| **#params** | **Spearman \\u03c1** \\t| **Time** \\t|\\n|-------------------------------|-------------|--------------|-----------------|\\n| ProtMamba (12k) \\t| 107M \\t| 0.381 \\t| 7m \\t|\\n| ProtMamba (single) \\t| 107M \\t| 0.406 \\t| 7m \\t|\\n| PoET (single) \\t| 201M \\t| 0.447 \\t| 9h 51m \\t|\\n\\nThis result shows that ProtMamba is not as good as a larger transformer model when using a short context length. It also shows that leveraging more homologs is very beneficial for the model\\u2019s performance on this task. More detailed information on how the context length impacts the performance of ProtMamba is included in figures S6, S7 and S8 in the updated paper.\"}", "{\"title\": \"Response to all reviewers (1/2)\", \"comment\": \"We thank the reviewers for their valuable feedback on our paper. We addressed all their questions with official comments on the individual reviews. We report here a brief summary of the main analyses and ablation that we performed in response and that we added to the new version of the paper. We provided a revised version of the manuscript with changes highlighted in blue.\\n\\n1. **Ablations on the training regime:** We trained different models (14M parameters d_model=512, n_layers=8 and 107M parameters like the original ProtMamba) for N=50k steps, reaching a total number of training tokens of T=10B. We report the perplexity computed on the validation set in the following table that we added to the new version of the paper. We signal with \\\"Fail\\\" the ablations where the model fails (e.g. a model trained with just the FIM loss has a very high AR loss on the full sequence).\\n\\n| **Perplexity** | **14M Parameters** \\t| \\t| | | **107M Parameters** \\t| \\t|\\n|---|---|---|-|-|---|---|\\n| \\t| **Autoregressive** \\t| **FIM** \\t| | | **Autoregressive** \\t| **FIM** \\t|\\n||\\n| Only FIM from scratch \\t| Fail \\t| 13.90 \\u00b1 0.34 | | | Fail \\t| 15.59 \\u00b1 0.27 |\\n| AR only \\t| **12.58** \\u00b1 0.31 \\t| 18.03 \\u00b1 0.25 | | | 11.05 \\u00b1 0.36 \\t| Fail \\t|\\n| No positional encoding \\t| 13.01 \\u00b1 0.30 \\t| 16.71 \\u00b1 0.47 | | | 12.31 \\u00b1 0.37 \\t| 17.20 \\u00b1 0.58 |\\n| Additive positional encoding | 12.72 \\u00b1 0.31 \\t| 13.60 \\u00b1 0.33 | | | 12.58 \\u00b1 0.38 \\t| 13.81 \\u00b1 0.31 |\\n| One mask, one token \\t| 12.76 \\u00b1 0.31 \\t| 15.54 \\u00b1 0.29 | | | 11.04 \\u00b1 0.33 \\t| 16.60 \\u00b1 0.36 |\\n| Masking fraction 50% \\t| 13.02 \\u00b1 0.31 \\t| **13.44** \\u00b1 0.33 | | | **10.94** \\u00b1 0.36 \\t| **11.59** \\u00b1 0.35 |\\n||\\n| ProtMamba \\t| 13.00 \\u00b1 0.30 \\t| 13.89 \\u00b1 0.32 | | | 11.35 \\u00b1 0.33 \\t| 12.62 \\u00b1 0.30 |\", \"main_observations_from_the_ablations\": [\"Raising the masking fraction to 50% has a minor positive effect on the model, we will take this into account in future versions.\", \"The use of positional embeddings does not deteriorate performance on the AR loss, it actually improves it on the FIM loss allowing us to perform inpainting with length control during inference.\", \"There is no substantial difference between sum and concatenation of the positional embeddings in small models. In large models, concatenation is slightly better.\", \"Purely autoregressive training hinders the FIM capabilities of the model.\", \"Using one mask per token degrades the FIM capabilities of the model (which are needed when one wants to inpaint a sequence).\", \"We are still performing ablations on a vanilla GPT2 Transformer.\", \"2. **Generative abilities and comparison with other state of the art models:** We compare the generative capabilities of ProtMamba with other state of the art models for conditioned generation, namely EvoDiff-MSA, MSA-Transformer and Potts models. Specifically, we now report a comparison of the pLDDT (using ESMFold) and scPerplexity (using ProteinMPNN) of 250 novel protein sequences generated using ProtMamba, each from a different cluster in our test set. We compare these values to the same scores measured on 250 novel protein sequences generated by EvoDiff-msa, MSA-Transformer and Potts models, retrieved from the Zenodo archive associated to the EvoDiff paper [1], and which were generated each from a different cluster of the EvoDiff validation set. We find that ProtMamba strongly outperforms all these models on conditioned generation, and obtains scores comparable to those of natural sequences, see table below.\", \"| Model | ProtMamba | \\t| EvoDiff | \\t| MSA Transformer\\t| \\t| Potts \\t| \\t| Natural |\", \"|---|---|---|---|---|---|---|---|---|---|\", \"| pLDDT (\\u2191) | **0.75 \\u00b1 0.13** | \\t| 0.60 \\u00b1 0.16 | \\t| 0.54 \\u00b1 0.18 | \\t| 0.56 \\u00b1 0.14 | \\t| 0.77 \\u00b1 0.13 |\", \"| scPerplexity (\\u2193) | **2.63 \\u00b1 0.45** | \\t| 3.17 \\u00b1 0.58 | \\t| 3.37 \\u00b1 0.64 | \\t| 3.17 \\u00b1 0.51 | \\t| 2.66 \\u00b1 0.49 |\"], \"some_observations_on_other_known_models_that_we_did_not_include_in_this_comparison\": [\"We did not perform this analysis using PoET because the authors did not release the code to sample from their model (they did share a script to score the variants).\", \"We did not perform this analysis using other state of the art de novo generative models like ProGen2 or DPLM because it is not possible to generate novel sequences conditioned on homologs using them, contrary to ProtMamba and EvoDiff. For this reason, we considered EvoDiff as the state of the art model for conditional generation. Conceptually, it is tempting to compare ProGen2\\u2019s control tag-conditioned generation or DPLM2\\u2019s structure-conditioned generation to ProtMamba\\u2019s homolog-conditioned generation. However, while these approaches are related, quantitative comparisons would be challenging because the exact conditioning would differ.\"]}" ] }

[ "Xing Lei", "Zifeng Zhuang", "Xuetao Zhang", "Donglin Wang" ]

[ "Goal-Conditioned Reinforcement Learning", "Data Augmentation", "Stitching Property" ]

{ "note_id": [ "qWMP01YLQy", "SCPenZso2p", "L5l2BLeeX6", "GHL4nOwYCP", "D748nSLPMX", "8hSg35mzU9", "7ctZCKYrKT", "5IqY7XJ3EH", "23WzBqWuKh", "08fcNTiIuC" ], "note_type": [ "official_comment", "official_comment", "meta_review", "official_comment", "official_review", "official_review", "official_review", "official_comment", "official_review", "decision" ], "note_created": [ 1733148917834, 1733147986468, 1734747226797, 1733108182158, 1730541473352, 1730634717082, 1730871712590, 1733203119073, 1730493035858, 1737523619253 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission4106/Authors" ], [ "ICLR.cc/2025/Conference/Submission4106/Authors" ], [ "ICLR.cc/2025/Conference/Submission4106/Area_Chair_4faq" ], [ "ICLR.cc/2025/Conference/Submission4106/Authors" ], [ "ICLR.cc/2025/Conference/Submission4106/Reviewer_KRBU" ], [ "ICLR.cc/2025/Conference/Submission4106/Reviewer_TUes" ], [ "ICLR.cc/2025/Conference/Submission4106/Reviewer_hWfc" ], [ "ICLR.cc/2025/Conference/Submission4106/Reviewer_KRBU" ], [ "ICLR.cc/2025/Conference/Submission4106/Reviewer_zjNR" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ] ], "structured_content_str": [ "{\"title\": \"Author Responses\", \"comment\": \"We thank the reviewer for the detailed review. Below, we address the raised concerns and questions.\\n\\n1. The introduction lacks sufficient emphasis on motivation, such as the advantages and necessity of SL compared to TD under the goal-conditional setting. It would be better to discuss the importance of SL-based methods, like OCBC, in detail in the introduction.\\n\\nThank you for your valuable suggestions on the paper. We have revised the introduction based on your feedback, providing a detailed discussion on the importance of SL-based methods, such as OCBC. The specific changes can be found in lines 50-57 of the revised version.\\n\\n2. Following Weakness 1, the experimental results also show a significant gap compared to TD-based algorithms (Table 1). It is still helpful to discuss this experiment phenomenon after Table 1. and 3. As shown in Figure 5, the performance of GCReinSL is inferior to the advanced TGDA method in some higher-dimensional tasks. Could the author discuss this result in detail?\\n\\nFirst, sequence modeling methods such as DT, EDT, and Reinformer exhibit limited stitching capabilities. Second, after carefully reviewing the code, we found that the limited improvement over IQL might be due to using the original parameter settings of Reinformer in the goal-conditioned RL context. To address this, we adjusted the training steps, learning rate, and sequence length to better suit the goal-conditioned RL context. As shown in the revised version's Table 1, GCReinSL for DT significantly improves the stitching capability of sequence modeling (with a 230-point increase compared to Reinformer), substantially narrowing the gap with IQL.\"}", "{\"title\": \"Author Responses\", \"comment\": \"Thank you for the detailed review and for the suggestions for improving the work. Below, I have carefully addressed your comments and concerns.\", \"for_question_1_and_question_2\": \"We have addressed these two issues, as detailed in the revised version under Section 3.1 and Section 4.3.\", \"for_question_3\": \"First, sequence modeling methods such as DT, EDT, and Reinformer exhibit limited stitching capabilities. Second, after carefully reviewing the code, we found that the limited improvement over IQL might be due to using the original parameter settings of Reinformer in the goal-conditioned RL context. To address this, we adjusted the training steps, learning rate, and sequence length to better suit the goal-conditioned RL context. As shown in the revised version's Table 1, GCReinSL for DT significantly improves the stitching capability of sequence modeling (with a 230-point increase compared to Reinformer), substantially narrowing the gap with IQL.\", \"for_question_4\": \"It is important to note that this context pertains to goal-conditioned RL (as clarified in the revised version, it is explicitly not a return-conditioned RL setting). Therefore, leveraging the relationship between the Q-function and probabilities described in Theorem 3.1, we first estimate probabilities using a VAE and then maximize the Q-function. In a return-conditioned RL setting, the returns from the offline dataset can be directly utilized instead. Notably, using a VAE as a probabilities estimator is a well-established approach, as demonstrated in works such as [1], [2], and [3].\\n\\n[1]Fujimoto, Scott, David Meger, and Doina Precup. \\\"Off-policy deep reinforcement learning without exploration.\\\" ICML, 2019.\\n\\n[1] Zhou, Wenxuan, Sujay Bajracharya, and David Held. \\\"Plas: Latent action space for offline reinforcement learning.\\\" CoRL, 2021.\\n\\n[1] Wu, Jialong, et al. \\\"Supported policy optimization for offline reinforcement learning.\\\" NIPS, 2022.\"}", "{\"metareview\": \"This paper studies reinforcement learning via supervised learning and explores how to equip supervised learning with trajectory stitching ability. The authors propose Goal-Conditioned Reinforced Supervised Learning (GCReinSL), which emphasizes maximizing the Q-function during training to estimate the maximum expected return within the distribution, subsequently guiding optimal action selection during inference. However, based on the reviewers\\u2019 feedback and the discussion with the authors, several concerns remain unresolved. Firstly, the title appears to be overclaiming. More importantly, there is a lack of mathematical rigor in the proofs, with inconsistent notations and numerous errors undermining the theoretical analysis\\u2019s credibility. Secondly, the experimental results reveal a significant performance gap compared to TD-based algorithms, contradicting the authors\\u2019 claim of significantly narrowing this gap. The authors did not provide effective responses to address these critical issues during the rebuttal phase. Therefore, I recommend rejection of this paper and encourage the authors to significantly revise their work for future submissions.\", \"additional_comments_on_reviewer_discussion\": \"The reviewers noted that the title of the paper is overclaiming. Furthermore, there is a lack of mathematical rigor in the proofs, with inconsistent notations and numerous errors, which undermines the credibility of the theoretical analysis. Additionally, the experimental results reveal a significant gap compared to TD-based algorithms, contradicting the authors\\u2019 claim of significantly narrowing this gap. The authors did not provide detailed or satisfactory responses to address these issues during the rebuttal phase.\"}", "{\"title\": \"Author Responses\", \"comment\": [\"Thank you for the detailed review and for the suggestions for improving the work.\", \"We have carefully reviewed the issues you pointed out and made the corresponding revisions (the modified sections are highlighted in light blue. If a section title is highlighted in light blue, it indicates that the entire section has been revised). Additionally, there are a few points we would like to address.\", \"An in-distribution Q-function refers to a Q-function learned on an offline dataset that does not produce out-of-distribution (OOD) values.\", \"In Theorem 4.1, $Q^m$ refers to the predicted Q-values output by the model. The intention is to clarify that $Q^m$ represents the direct output of the model itself. For instance, in the case of Decision Transformer (DT), it similarly produces the corresponding return values as part of its own output.\", \"The term \\\"OOD\\\" mentioned in line 220 has already been defined earlier in line 193.\", \"I have addressed all the remaining issues based on your suggestions. If there are any points that are still unclear or require further clarification, I welcome further discussion.\"]}", "{\"summary\": \"This paper studies the stitching property for SL within goal-conditioned offline RL problems. This stitching property is commonly obtained in TD-based algorithms and fails in SL. This paper proposes the GCReinSL, which enhances SL-based approaches by incorporating maximize Q-function. Equipped with the GCReinSL framework, the previous outcome-conditioned behavioral cloning (OCBC) algorithms exhibit the switching property and achieve better performance under the goal-conditioned setting.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"1. The difference in trajectory stitching property between SL and TD does exist and is valuable for studying to improve the generalization performance of SL.\\n2. The method that incorporates maximizing Q-function is natural, and the experiments show effectiveness.\", \"weaknesses\": \"1. The introduction lacks sufficient emphasis on motivation, such as the advantages and necessity of SL compared to TD under the goal-conditional setting. It would be better to discuss the importance of SL-based methods, like OCBC, in detail in the introduction.\\n2. Following Weakness 1, the experimental results also show a significant gap compared to TD-based algorithms (Table 1). It is still helpful to discuss this experiment phenomenon after Table 1.\", \"questions\": \"Please see the Weakness part.\\n1. As shown in Figure 5, the performance of GCReinSL is inferior to the advanced TGDA method in some higher-dimensional tasks. Could the author discuss this result in detail?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The paper aims to enhance the effectiveness of supervised learning (SL) methods in reinforcement learning (RL) by introducing a framework called Goal-Conditioned Reinforced Supervised Learning (GCReinSL). The authors propose that traditional SL-based RL methods, such as outcome-conditioned behavioral cloning (OCBC), lack trajectory stitching capabilities, which are critical for integrating suboptimal trajectories into optimal paths\\u2014a feature common in temporal-difference (TD) learning. To address this, the authors introduce Q-conditioned maximization, positing that the objective in goal-conditioned RL is equivalent to the Q-function, thereby allowing SL methods to maximize expected returns.\\n\\nThe paper presents GCReinSL as a solution to bridge the gap between SL and TD learning by embedding Q-function maximization into SL-based methods. The proposed approach is evaluated on various goal-conditioned offline RL tasks, such as Pointmaze and Antmaze, and compared against other methods like IQL, CQL, and other sequence modeling techniques. The authors claim that GCReinSL improves stitching performance and generalization across unseen goal-state pairs in offline RL datasets.\", \"soundness\": \"2\", \"presentation\": \"1\", \"contribution\": \"1\", \"strengths\": [\"The paper tackles the relevant challenge of bridging the gap between supervised learning (SL) and temporal-difference (TD) learning, especially focusing on trajectory stitching\\u2014a key limitation of SL-based RL methods.\", \"The paper\\u2019s focus on goal-conditioned RL is timely and aligns with practical applications in areas like robotics and offline RL.\"], \"weaknesses\": [\"**Inconsistent and Incomplete Notations**:\", \"The mathematical notations are poorly defined and inconsistent, \\bfor example, equations like Eq. (3), which omits necessary terms such as the expectation over the initial state distribution. These issues create significant barriers to understanding the proposed approach.\", \"**Lack of Theoretical Rigor**:\", \"Theorem 4.1 and its proof are presented in a sloppy and non-rigorous manner. Important terms are either undefined or unclear.\", \"**Underwhelming Empirical Performance**:\", \"The proposed method, GCReinSL, underperforms significantly compared to existing methods like IQL and CQL, particularly in the more challenging Antmaze datasets. The results fail to justify the claimed advantages of sequence modeling approaches over TD-based methods.\", \"I have spent a significant amount of time and effort thoroughly reviewing this paper, but the conceptual, theoretical, and empirical weaknesses, along with poor clarity, lead me to come to a conclusion that the paper is not ready for publication.\", \"**Lack of Clarity**:\", \"The paper is riddled with errors and unclear explanations, making it challenging to read (see below). The poor writing quality detracts from the overall presentation and makes it difficult to follow the core ideas.\", \"### **Comments & questions**\", \"Line 56-57: How come \\u201cthe objective in goal-conditioned RL is equivalent to the Q-function\\u201d? How does a function \\u201cQ-function\\u201d is equivalent to an objective (either learning or optimization)? Perhaps, the authors intend to say \\u201cthe objective function?\\u201d Objectives and objective functions are two different things.\", \"Line 67\\u201d What is \\u201cin-distribution Q-function?\\u201d Please properly define what you mean by \\u201cin-distribution\\u201d (perhaps with respect to offline data set?)\", \"Line 69: Is it \\u201cpredicted Q-function\\u201d or \\u201cestimated Q-function\\u201d?\", \"Line 70: What is \\u201cthe current maximum Q-function?\\u201d What do you mean by current? By \\u201cmaximum Q-function\\u201d do you mean max of Q-functions or max_a Q(s,a)? The latter is NOT maximum Q-function\\u2026 is maximum Q-value.\", \"In Eq.(3), expectation with respect to initial state distribution is NOT included in $J(\\\\pi)$? Then, shouldn\\u2019t $J(\\\\pi)$ be also function of s? Also, in Eq.(2), when you define conditional \\u201cdiscounted state occupancy distribution,\\u201d $\\\\pi$ is used as a goal-conditioned policy $\\\\pi(a, | s, g)$ is conditioned on a pair of state and goal. Now, in (3), the authors are using trajectory-wise policy in (3), particularly the conditional distribution within the expectation of (3), which is not consistent with the definition of (2) The authors should avoid any unnecessary overloading, and be clear about what $\\\\pi$ they use. Perhaps, the authors need to appropriately define the relationship between $\\\\pi(\\\\tau | g)$ and $\\\\pi(a, | s, g)$.\", \"In any real-world environment, how does an agent observe rewards as \\u201cthe (exact) probability of reaching the goal at the next time stee\\u201d as defined in Eq.(4)? Do the trajectories defined in 171-172 contain these probability rewards as offline data?\", \"The conditional distribution \\u201cp^\\\\pi_+(s_{t+} = g | s_0 = s,a)$ in Eq.(6) conditioned on state and action pair has been properly defined.\", \"Line 216-217, what is $\\\\hat{Q}_t$? Is it $\\\\hat{Q}_t = \\\\hat{Q}(s_t, a_t)$? Please be precise.\", \"Line 216-219, the explanation is not clear. For example, the authors explain that OCBC methods will reach the state $g_1$ rather then $g$ since since Q-function is still zero. But, there are no explonations as to why\\u2026 They say that \\u201c$\\\\hat{Q}_t = 1$ is impossible to obtain given $\\\\hat{Q}_0 = 0$. Please clearly explain why.\", \"Line 235: When introducing the \\u201cLatent Variable Model,\\u201d please define what $\\\\psi$ is.\", \"What is $p(z|s)$ in Eq.(8)? The authors only mentioned $p(z|s,a) = \\\\mathcal{N}(0,I)$ as a prior.\", \"Line 246-247: In \\u201cwe can approximate the probability $p^{\\\\pi(\\\\cdot | \\\\cdot ,g)}(s_{t+} = g | s, a)$ in Eq.(6) by $\\u2212\\\\mathcal{L}_\\\\text{ELBO}$\\u201d, how?\", \"What is $t$ in the expectation $E_t$ in Eq.(10)? There is NO mention of $t$ in the entire section of 4.3.\", \"In Theorem 4.1, after defining $\\\\textbf{SG}= (s, g, a, Q)$, the authors write $Q(\\\\textbf{SG}, a)$ for $Q_\\\\text{max}$. So, $Q(\\\\textbf{SG}, a) = Q(s, g, a, Q, a) $? At this point (along with the previous unclear presentation), the paper becomes very difficult to read\\u2026\", \"In Theorem 4.1, so $\\\\mathbf{Q}^m$ is a policy?\", \"In Theorem 4.1, $\\\\pi_{\\\\theta}^* = \\\\arg \\\\min \\\\mathcal{L}^m_Q$ what is the $\\\\arg \\\\min$ over? All of sudden, the authors introduce $\\\\theta$ which I suppose is the parameter for the policy, then they need to define it. Also, if $\\\\pi_{\\\\theta}^*$ is function of $\\\\theta$, but the loss $\\\\mathcal{L}^m_Q$ does not depend on the parameter of a policy? The authors are very sloppy about the notations and presentations throughout the paper, yet even one of their main results (Theorem 4.1) fails to provide meaningful contributions particularly with its non-rigorous presentation.\", \"In the proof of Theorem 4.1 in Appendix A.2, aren\\u2019t $\\\\mathbf{Q}^m$ a vector (or even matrix)? How do you define inequality between vectors, is it element-wise? Then, the authors should be explicit about that. What do you mean by \\u201call Q-values from the offline dataset\\u201d in 917? All true Q-values? The proof of Theorem 4.1 makes difficult to validate its claim.\", \"After reading the proof of Theorem 4.2 in Appendix A.3 and also considering the statement itself, I do not think Theorem 4.2. is a proper mathematical theorem. \\u201cRemark\\u201d (or corollary at best) would be an adequate category.\", \"### **Minor errors**\"], \"line_23\": \"\\u201cby incorporating maximize\\u201d => \\u201cby incorporating maximizing?\\u201d\\nLine 53 Acronym \\u201cDT\\u201d is used without properly defining what it is first.\", \"line_54_55\": \"The citation \\u201cZhuang et al. (2024)\\u201d should be used with \\\\citep\", \"line_75\": \"Acronym \\u201cRvS is used without properly defining what it is first.\", \"line_145\": \"\\u201cmaximise\\u201d or \\u201cmaximize\\u201d like the other expressions in the text? It would be good to maintain the style consistency.\", \"line_162\": \"\\u201cQ-function are\\u201d => \\u201cQ-functions are\\u201d\", \"line_173\": \"Perhaps, \\u201cIn each $\\\\tau_i$ for $i in 1, \\u2026, N$\\u201c would be more clear.\", \"line_177\": \"\\u201cprovide present\\u201d => \\u201cpresent\\u201d\", \"line_220\": \"\\u201cOOD\\u201d => \\u201cout-of-distribution (OOD)\\u201d please define acronyms when first introduced.\", \"line_262\": \"\\u201cAfter estimate the Q-function\\u201d => \\u201cAfter estimating the Q-function\\u201d\", \"line_274_275\": \"\\u201cmore weights to the $Q$ larger than\\u201d is missing $\\\\hat{Q}$.\", \"questions\": \"See the questions above.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"1\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper proposes a novel method named Goal-Conditioned Reinforced Supervised Learning (GCReinSL) aiming to address the limitation of outcome-conditioned behavioral cloning (OCBC) methods in reinforcement learning tasks. Current supervised learning methods in RL lack the capability of trajectory stitching which allows the algorithms to effectively combine data from suboptimal trajectories to achieve better performance. This paper leverages expectile regression for Q-function estimation and demonstrates through theoretical analysis and experiments that this augmentation enables OCBC methods to solve the stitching problem. Experimental results on offline datasets show that GCReinSL outperforms existing goal-conditioned SL methods.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": [\"The paper solves the stitching problem in OCBC methods by introducing Q-conditioned maximization, which allows the algorithm to combine data from suboptimal trajectories.\", \"The paper provides theoretical and empirical analysis to demonstrate the effectiveness of the proposed method in enhancing OCBC methods.\", \"The motivation for using executive regression for Q-function estimation is well-explained and aligns with the goal of estimating the maximum expected return without out-of-distribution issues.\"], \"weaknesses\": [\"The method needs to learn a conditional variational autoencoder which could introduce additional computational overhead and complexity.\", \"The proposed method is constrained by the goal-conditioned formulation which may limit its application.\"], \"questions\": [\"Can you provide more insights into the computational cost of adding the conditional variational autoencoder and how it scales with the size of the dataset?\", \"Can this method adapt to the return-conditioned formulation? What are the challenges and limitations of this approach?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thanks for the author's reply and additional experiments. By carefully choosing the parameters of GCReinSL, the revision paper improves the performance of the proposed method.\\n\\n However, I still have some concerns. As shown in line 53 of the revision paper, one motivation for designing the OCBC method is that the TD-based method is highly sensitive to hyperparameters. However, as shown in the reply, the GCReinSL seems also to be a parameter-sensitive method, which can not fully support the discussion in the introduction. It would be better to discuss how to choose the hyperparameters of the GCReinSL. I will maintain my score.\"}", "{\"summary\": \"This paper studies reinforcement learning via surpervised learning and explores how to endow SL with trajectory stitching ability. Goal-Conditioned Reinforced Supervised Learning (GCReinSL) is proposed which emphasizes the maximization of the Q-function during the training phase to estimate the maximum expected return within the distribution, subsequently guiding optimal action selection during the inference process.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"The paper is relatively well-written. Experiment results are solid.\", \"weaknesses\": \"The idea is not novel, basically a combination of tricks in exsiting literature. The expriments results can not support the title. See my questions for more details.\", \"questions\": \"1. It might be confusing to use $\\\\pi$ to denote the probabiliy over the trajectory in $\\\\pi(\\\\tau|s,g)$ (3) and also to denote the policy $\\\\pi(a|s,g)$.\\n\\n2. In section 4.3, what is the $\\\\pi$ in the probability distribution? at first, it was $p^{\\\\pi}$ in line 234, then it becomes $p^{\\\\pi(\\\\cdot|\\\\cdot|g)}$ in line 245. Is it the behavior policy collecting the offline dataset?\\n\\n3. For the Antmaze taks and the results in Table 1. The DT, EDT and Reinformer almost do not work. GCReinSL improves the performance from approximately 0 to about 10 (with large variance), there is a huge gap compared to RL method (about 50-80). Say, the improvement is about 10, and the initial gap is 50-80. Is it proper to claim 'significantly narrowing the gap with TD learning methods such as IQL'? I see this experiment as a example that SL would fail catastrophically, even with your maximum Q conditioning trick.\\n\\n4. I have a question about the maximum Q conditioning trick. Different from the Return conditioned supervised learning methods such as Reinformer, for which they can directly access to the return in the dataset, in the goal conditioned supervised learning, the Q-function is estimated from VAE, and then the maximization is performed on the estimated Q-function. I guess the estimation error is hard to control as it may come from multiple sources: 1) how you evaluate that the VAE obtain decent estimation of the goal probability? 2) how you sure that the expectile regression gives a proper maximum in distribution Q value? Theorem 4.1 is not a accurate quantification of the return you get as it only considers the ideal case m goes to 1 and it does not consider how the maximum value is cover in the offline dataset.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}" ] }

[ "Mateusz Pach", "Koryna Lewandowska", "Jacek Tabor", "Bartosz Michał Zieliński", "Dawid Damian Rymarczyk" ]

[ "xai", "interpretability", "prototypical parts" ]

{ "note_id": [ "ytXUeATpfQ", "yNO5ePaqle", "wS4mEqwhu9", "wG4iN4El3q", "umztjawYYE", "tdH6OUUzZk", "euB784j0rm", "bid0A5GPFT", "bHyXYonvpo", "X6bK7F0o2z", "WjPFVoNe2s", "Uqw4oNwb30", "USxI5SFIHA", "SbOhGqyso3", "LfV7E91rFL", "LRwQ7ksX5U", "Kydg8p3P6R", "HB9is8Lxi6", "GMan0iRRXx", "FIunVysmJR", "DayXG6p7hS", "CsxZvYM0H0", "Ck1szKWb7j", "AIyjYRqj2A", "9o7ElZbTFg", "4JAn53qJ06", "0kSeWoa1Ws", "0RuuCW6gws" ], "note_type": [ "official_comment", "official_comment", "official_comment", "official_comment", "meta_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_review", "decision", "official_comment", "official_review", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment" ], "note_created": [ 1732723714619, 1732640145513, 1732724400071, 1732270813112, 1734673325617, 1732540286761, 1732608231481, 1732270569607, 1732540677796, 1732270673974, 1732640247163, 1732520106959, 1730659491863, 1732271107475, 1732640076438, 1732540590531, 1731115428361, 1737523485884, 1732553002755, 1730529539524, 1732642286605, 1732270517599, 1730560247480, 1732271144835, 1732270999174, 1732611238469, 1732720380091, 1732540782029 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission2112/Reviewer_na4o" ], [ "ICLR.cc/2025/Conference/Submission2112/Authors" ], [ "ICLR.cc/2025/Conference/Submission2112/Authors" ], [ "ICLR.cc/2025/Conference/Submission2112/Authors" ], [ "ICLR.cc/2025/Conference/Submission2112/Area_Chair_xBkS" ], [ "ICLR.cc/2025/Conference/Submission2112/Authors" ], [ "ICLR.cc/2025/Conference/Submission2112/Reviewer_Zw4S" ], [ "ICLR.cc/2025/Conference/Submission2112/Authors" ], [ "ICLR.cc/2025/Conference/Submission2112/Authors" ], [ "ICLR.cc/2025/Conference/Submission2112/Authors" ], [ "ICLR.cc/2025/Conference/Submission2112/Authors" ], [ "ICLR.cc/2025/Conference/Submission2112/Reviewer_9WXj" ], [ "ICLR.cc/2025/Conference/Submission2112/Reviewer_na4o" ], [ "ICLR.cc/2025/Conference/Submission2112/Authors" ], [ "ICLR.cc/2025/Conference/Submission2112/Authors" ], [ "ICLR.cc/2025/Conference/Submission2112/Authors" ], [ "ICLR.cc/2025/Conference/Submission2112/Reviewer_Zw4S" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission2112/Reviewer_na4o" ], [ "ICLR.cc/2025/Conference/Submission2112/Reviewer_9WXj" ], [ "ICLR.cc/2025/Conference/Submission2112/Reviewer_na4o" ], [ "ICLR.cc/2025/Conference/Submission2112/Authors" ], [ "ICLR.cc/2025/Conference/Submission2112/Reviewer_9gD3" ], [ "ICLR.cc/2025/Conference/Submission2112/Authors" ], [ "ICLR.cc/2025/Conference/Submission2112/Authors" ], [ "ICLR.cc/2025/Conference/Submission2112/Reviewer_9WXj" ], [ "ICLR.cc/2025/Conference/Submission2112/Authors" ], [ "ICLR.cc/2025/Conference/Submission2112/Authors" ] ], "structured_content_str": [ "{\"title\": \"Thank you for your clarification\", \"comment\": \"Thank you for your clarification. I have raised my score to 8: accept, good paper.\"}", "{\"comment\": \"Thank you for your valuable feedback and for raising the score of our manuscript. We are pleased to hear that we were able to address most of your concerns.\\n\\nTo better clarify the unique advantages of our approach compared to shallow network visualization methods, we have revised the section \\\"Usage of low-level vision features for image classification\\\" in the Related Works.\\n\\nWe greatly appreciate your insights, which have helped us improve the manuscript.\"}", "{\"comment\": \"We sincerely appreciate your reassessment and are thrilled that the revisions addressed your concerns. Thank you for your time and constructive feedback!\"}", "{\"comment\": \"**Q1. My main concern is that I did not see prototype projections in this work. Without prototype projections, how could you conclusively visualize prototypes using training images? The closest training images to a prototype could still be far away from the prototype in the latent space.**\\n\\nOur work builds on PIPNet\\u2019s definition of prototypical parts, which is why it lacks projection, which can lead to less faithful visualizations. Despite this drawback, PIPNet-based architectures have been successfully applied in various works De Santi et al. (2024a;b); Nauta et al. (2023a), and further developed, e.g. Wang et al. (2024) improving the interpretability. \\n\\nMoreover, LucidPPN introduces a key difference in the definition of prototypical parts compared to PIPNet. While PIPNet employs Softmax across channels in the latent feature map, LucidPPN uses the sigmoid activation function. The sigmoid function allows each channel\\u2019s activation to be learned independently, not influenced by the relative activations of other channels. While, Softmax normalization can distort activations by emphasizing values that are only relatively high compared to others, even if they are low in absolute terms. \\n\\nTo build an intuition for this statement, let us consider $i$-th pixel of a feature map with activation values $z_i = [\\u22122, 5, \\u22120.2, \\u22120.1]$, and $j$-th pixel with $z_j = [10, 300, 30, 10]$, the Softmax output $\\\\theta$ for both would be $\\\\theta_i = \\\\theta_j = [0, 1, 0, 0]$. This implies that PIPNet would treat both pixels as equally important, despite the activations differing by a factor of 60. In contrast, with sigmoid activation $\\\\sigma$ used in LucidPPN, the outputs would be $\\\\sigma_i = [0.1192, 0.9933, 0.4502, 0.4750]$ and $\\\\sigma_j = [1.0000, 1.0000, 1.0000, 1.0000]$, preserving the distinction in activation magnitudes. As a result, one can easily verify if the image patches selected for visualization are faithful because such patches should have a resemblance score close to 1.\\n\\n\\n**Q2. During training, are the segmentation masks from PDiscoNet aligned with the ShapeTexNet feature maps or the aggregated feature maps?** \\n\\nLoss $L_D$ is applied only to the ShapeTexNet feature maps as we directly align them with masks from PDiscoNet. Indirectly, it also causes alignment of masks with the aggregated feature maps which are computed from the ShapeTexNet feature maps. To the Supplementary Materials (Figure 14), we added an image illustrating this process more concisely.\\n\\n\\n**Q3. I am also not clear as to why binary cross entropy is used instead of multi-class cross entropy for training?** \\n\\nThe intuition behind BCE usage is rooted from multilabel classification. To some degree ShapeTexNet operates in a multilabel setting from the prototypical parts perspective as they may match multiple classes. Hence, to enable multiple classes having high similarity to the same prototypical parts, we use sigmoid instead of softmax when computing the feature maps. This necessitates a shift from Cross-Entropy (CE) to Binary Cross-Entropy (BCE) because CE would then solely maximize the activation of the correct class while ignoring crucial signals from negative classes. Another reason behind our choice is to make it easier to verify the faithfulness of visualizations (see the answer about prototype projection).\"}", "{\"metareview\": \"In this paper, a Lucid Prototypical Parts Network (LucidPPN) prototypical parts network is presented, which has two branches: a ShapeTexNet and a ColorNet. Given an input image, the ShapeTexNet is a convolutional neural network (CNN) that takes a gray-scale version of the image as input and outputs a set of feature maps, and the ColorNet is another CNN that takes a down-sampled version of the image as input and outputs another set of feature maps. Evaluation is carried out on 4 commonly used fine-grained classification benchmarks (CUB-200-2011, Stanford Cars, Stanford Dogs, and Oxford Flowers), and found the LucidPPN models achieved competitive test accuracy compared to other interpretable models.\", \"additional_comments_on_reviewer_discussion\": \"All the reviewers lean to accept the paper. Thanks for the good job.\"}", "{\"comment\": \"As noted in the Review, we agree that there are techniques to visualize the shallow layers of neural networks, and these methods provide a certain level of interpretability. However, LucidPPN focuses on high-level concepts (represented with prototypical parts) from deeper layers that encode complex information. Our goal is to enhance the transparency of these concepts by disentangling the color from the remaining visual features. While there is a wealth of research on prototypical parts-based interpretability (e.g., Chen et al., 2019; Nauta et al., 2021; 2023b; Rymarczyk et al., 2021; 2022; 2023; Wang et al., 2024), none of these works aim to introduce an inherently interpretable mechanism into the network at the level of low-level visual features.\\n\\nRegarding accuracy, we note that the results for the late fusion method (LucidPPN) and the earliest fusion method (single branch) are provided in Table 4. For the CUB dataset, LucidPPN scored 81.5% while single branch 86.6%. To investigate the influence of earlier fusion on the accuracy, we also experimented with fusion applied after the second block of ConvNeXt Tiny. This configuration achieved an accuracy of 84.1%, indicating that earlier fusion can indeed enhance the model's accuracy. However, this comes at the cost of explanation granularity: when fusion occurs earlier, it becomes difficult to disentangle the influence of shape with texture and color on prototypical parts.\\n\\nWe kindly ask the Reviewer to evaluate whether our responses address their concerns. If not, we would appreciate clarification on two points. First, regarding shallow layer visualization, could you specify the techniques or works you had in mind so we can reference them more precisely? Second, in terms of fusion analysis, what specific types of evaluations or comparisons would you find most informative? If no further concerns remain, we kindly request a reevaluation of your score.\"}", "{\"comment\": \"Thanks for the response ! I got the explanation for limited improvement. Considering the overall quality, I vote for borderline accept.\"}", "{\"comment\": \"### **References**\\n\\nJ. Adebayo et al. Sanity checks for saliency maps. NeurIPS 2018. \\n\\nA. Bontempelli et al. Concept-level debugging of part-prototype networks. arXiv 2022. \\n\\nC. Chen et al. This looks like that: deep learning for interpretable image recognition. NeurIPS 2019. \\n\\nL. A. De Santi et al. Patch-based intuitive multimodal prototypes network (pimpnet) for alzheimer\\u2019s disease classification. arXiv 2024a. \\n\\nL. A. De Santi et al. Pipnet3d: Interpretable detection of alzheimer in mri scans. arXiv 2024b. \\n\\nJ. He et al. Partimagenet: A large, high-quality dataset of parts. ECCV 2022.\\n\\nQ. Huang et al. Evaluation and improvement of interpretability for self-explainable part-prototype networks. ICCV 2023.\\n\\nS. SY Kim et al. Hive: Evaluating the human interpretability of visual explanations. ECCV 2022. \\n\\nP. W. Koh et al. Concept bottleneck models. ICML 2020.\\n\\nC. Ma et al. Interpretable image classification with adaptive prototype-based vision transformers. arXiv 2024a. \\n\\nC. Ma et al. This looks like those: Illuminating prototypical concepts using multiple visualizations. NeurIPS 2024b. \\n\\nA. Nagrani et al. Attention bottlenecks for multimodal fusion. NeurIPS 2021. \\n\\nM. Nauta et al. Neural prototype trees for interpretable fine-grained image recognition. CVPR 2021.\\n\\nM. Nauta et al. Interpreting and correcting medical image classification with pip-net. ECAI 2023a. \\n\\nM. Nauta et al. Pip-net: Patch-based intuitive prototypes for interpretable image classification. CVPR 2023b. \\n\\nC. Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature machine intelligence 2019. \\n\\nC. Rudin et al. Interpretable machine learning: Fundamental principles and 10 grand challenges. Statistic Surveys 2022. \\n\\nD. Rymarczyk et al. Protopshare: Prototypical parts sharing for similarity discovery in interpretable image classification. ACM SIGKDD 2021. \\n\\nD. Rymarczyk et al. Interpretable image classification with differentiable prototypes assignment. ECCV 2022. \\n\\nD. Rymarczyk et al. Icicle: Interpretable class incremental continual learning. ICCV 2023\\n\\nR. Tomsett et al. Sanity checks for saliency metrics. AAAI 2020.\\n\\nBS Wang et al. Mcpnet: An interpretable classifier via multi-level concept prototypes. CVPR 2024.\\n\\nM. Xue et al. Protopformer: Concentrating on prototypical parts in vision transformers for interpretable image recognition. arXiv 2022.\"}", "{\"comment\": \"Dear Reviewer na4o,\\n\\nAs the deadline for the discussion period is approaching quickly, we would like to kindly remind the reviewer that we are waiting for your response.\\n\\nIn particular, we have provided point-by-point responses to all of your questions to address your concerns and provided the revision that reflects such changes. Therefore, your timely feedback and change in the score if applicable would be highly appreciated.\\n\\nBest,\\n\\nAuthors\"}", "{\"comment\": \"**1. The Section 3 has a lot of paragraphs but lacks subheadings, making it difficult to follow the logical flow of the different parts.**\\n\\nWe agree that Section 3 was dense. Therefore, we have revised it by introducing subsections that clarify the methodology of LucidPPN more clearly.\\n\\n**2. There was no noticeable advantage in accuracy. Why?**\\n\\nWe answer this question in **shared remarks** in paragraphs *There was no noticeable advantage in accuracy. Why?* and *The improvements demonstrated by the proposed method appear to be limited because its performance on some instances is lower than that of the compared methods.*\"}", "{\"comment\": \"We sincerely thank you for your valuable insights, which have significantly contributed to enhancing our manuscript.\"}", "{\"comment\": \"Thanks for the detailed response from the authors. While some concerns have been addressed, there remain several points that may require further clarification.\\n\\n1.\\tAbout the contributions:\\nThe authors' response on the novelty and contribution of analyzing \\\"color,\\\" \\\"shape,\\\" and \\\"texture\\\" didn\\u2019t seem to fully resolve the concern. Visualizing the shallow layers of the neural network has indeed demonstrated the network's focus on the features of \\\"color,\\\" \\\"shape,\\\" and \\\"texture,\\\" and has visualized \\\"shape\\\" and \\\"texture\\\" more clearly, which contributes to making the network processing more transparent from these perspectives. However, the authors state that \\\"conventional neural networks often entangle visual features in ways that make it difficult to disentangle and present them in an understandable format for users\\\", which may not be entirely consistent with the explanatory results of existing methods that visualize shallow layers of networks. These existing methods also provide a certain level of interpretability. \\n\\n2.\\tAbout performance:\\nWe agree with the authors' explanation that the performance decrease may due to the delayed fusion of texture and shape features with color. This delay prevents the network from effectively associating shape and texture with color in the early layers, leading to incomplete detection of certain features. However, there is a lack of experimental results and analysis to support this explanation. For instance, comparisons between shallow fusion and deep fusion under the same conditions would help verify whether the accuracy drop is indeed caused by the location of fusions. Given that the performance decline is likely due to insufficient attention to important features (such as texture and shape), the limitations of the interpretability method itself, and other factors, further experiments related to this topic would be beneficial. A more comprehensive analysis of the limitations and the true advantages of the method may be expected.\"}", "{\"summary\": \"In this paper, the authors proposed a Lucid Prototypical Parts Network (LucidPPN), a novel prototypical parts network that separates color prototypes from other visual features. A LucidPPN has two branches: a ShapeTexNet and a ColorNet. Given an input image, the ShapeTexNet is a convolutional neural network (CNN) that takes a gray-scale version of the image as input and outputs a set of feature maps, and the ColorNet is another CNN that takes a down-sampled version of the image as input and outputs another set of feature maps. Since the last layer of both the ShapeTexNet and the ColorNet is a 1x1 convolutional layer with KM filters, we can interpret the last convolutional layer as a prototype layer with KM prototypes, where K is the number of prototypes per class and M is the number of classes, and the output of the last layer as prototype activation maps. The output feature maps (aka prototype activation maps) from the ShapeTexNet and the ColorNet are fused using element-wise products, and then max-pooled to yield a prototype similarity score for each prototype. The predicted class score is simply an average of the prototype similarity scores over all prototypes of the class. In a LucidPPN, each of the K prototypes in each of the M classes corresponds to consistent image parts (e.g., the first prototype of each class corresponds to head of a bird, etc.). This is achieved by aligning the fused output feature maps (prototype activation maps) with segmentation masks produced by a pre-trained PDiscoNet (an object part segmentation model) using a prototypical-object part correspondence loss. In addition to a loss function to improve the classification accuracy of the entire model, the authors also introduced a loss function to improve the classification accuracy of the ShapeTexNet alone and to disentangle color from other visual features. The authors evaluated their LucidPPN models on 4 commonly used fine-grained classification benchmarks (CUB-200-2011, Stanford Cars, Stanford Dogs, and Oxford Flowers), and found that their LucidPPN models achieved competitive test accuracy compared to other interpretable models. The authors also did a user study to evaluate the influence of disentangling color from other visual attributes on interpretability.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": [\"Originality: The paper introduced a novel idea of disentangling color from shape and texture, so that the visual attribute of each prototype is more clearly defined (compared to prior work).\", \"Quality: The authors did show that their LucidPPN could maintain a reasonable accuracy while providing less ambiguous prototypes.\", \"Clarity: The paper is clearly written.\", \"Significance: Interpretability is a significant area of research in machine learning.\"], \"weaknesses\": [\"Quality: There seems to be no prototype projection in this work. Without prototype projections, it is unclear if the prototypes can be faithfully visualized using training images (because the closest training images to a prototype could still be far away from the prototype in the latent space).\", \"Clarity: Page 6, Lines 314-315. I am confused as to whether you are aligning the segmentation masks from PDiscoNet with prototype activation maps from the ShapeTexNet or the aggregated feature maps.\"], \"questions\": [\"My main concern is that I did not see prototype projections in this work. Without prototype projections, how could you conclusively visualize prototypes using training images? The closest training images to a prototype could still be far away from the prototype in the latent space.\", \"During training, are the segmentation masks from PDiscoNet aligned with the ShapeTexNet feature maps or the aggregated feature maps?\", \"I am also not clear as to why binary cross entropy is used instead of multi-class cross entropy for training?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"details_of_ethics_concerns\": \"N/A.\", \"rating\": \"8\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"**W1. While analyzing \\u201dcolor,\\u201d \\u201dshape,\\u201d and \\u201dtexture\\u201d offers a valuable perspective, these features have been extensively studied in the field of visual perception. Given that the shallow layers of deep networks are capable of extracting low-level features, the necessity for additional processing and analysis from prototypical parts raises concerns on the novelty and contribution of this work.**\\n\\nIt is true that shallow layers of neural networks are capable of extracting low-level features. The goal of LucidPPN, however, is to make this processing more transparent to the user, which aligns with the broader objective of inherently interpretable models Chen et al. (2019); Rudin (2019). Thanks to LucidPPN we can analyze which colors were important for classification. Conventional neural networks often entangle visual features in ways that make it difficult to disentangle and present them in an understandable format for users. This is why we believe our work offers a novel contribution to the field of interpretable AI, particularly in the context of prototypical parts.\\n\\n\\n**W2. The improvements demonstrated by the proposed method appear to be limited because its performance on some instances is lower than that of the compared methods**\\n\\nWe respond to this comment in **shared remarks** in paragraphs *The improvements demonstrated by the proposed method appear to be limited because its performance on some instances is lower than that of the compared methods.* and *There was no noticeable advantage in accuracy. Why?*\\n\\n\\n**W3. The organization of the experimental section appears somewhat unbalanced. While the results and visualizations presented are commendable, an excessive amount of content is relegated to the appendix, which may hinder the reader\\u2019s ability to grasp key insights and maintain a coherent narrative.**\\n\\nWe agree and in the revised version of the manuscript, we have reorganized the experimental section. However, due to space constraints and to adhere to the ICLR template, some content has been moved to the appendix.\"}", "{\"comment\": \"To answer this question, we provide an additional section \\\"Faithfullness of patch visualizations\\\" in the Supplementary Materials. It contains Figure 20 with a distribution of the sigmoid function values obtained for patches used in prototype visualization. For LucidPPN trained on the CUB dataset (blue curve), 61.04% of those patches have values above 0.9, which indicates that prototype visualizations are relatively faithful.\\n\\nMoreover, we show that higher faithfulness can be obtained when training with an additional loss component $L_C$ that punishes the model if the sigmoid function value for a given prototype is smaller than $1$ for all samples in the batch (see green and yellow curves in Figure 20).\"}", "{\"comment\": \"Dear Reviewer 9gD3,\\n\\nAs the deadline for the discussion period is approaching quickly, we would like to kindly remind the reviewer that we are waiting for your response.\\n\\nIn particular, we have provided point-by-point responses to all of your questions to address your concerns and provided the revision that reflects such changes. Therefore, your timely feedback and change in the score if applicable would be highly appreciated.\\n\\nBest, \\n\\nAuthors\"}", "{\"summary\": \"Summary Of Contributions:\\n1.Introduction of LucidPPN: This novel architecture separates color features from other visual components during inference, enabling clearer identification of feature importance in the decision-making process.\\n2.Consistent Object-Part Mapping: A mechanism ensures that prototypes within each class consistently correspond to the same object parts, improving interpretability.\\n3.Enhanced Visualization Method: A more intuitive visualization type is introduced, optimized for fine-grained classification.\\n4.Comprehensive Analysis: The paper provides an in-depth examination of LucidPPN's usefulness and limitations, particularly identifying cases where color may or may not be a critical feature in fine-grained classification.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"1.The LucidPPN in the paper consists of two branches, one for color and the other for shape/texture, which effectively decouples different features. This method can reduce the ambiguity of traditional prototype networks and enable users to better understand the reasons behind the model's decisions.\\n\\n2.Compared to existing methods, LucidPPN achieves a more detailed analysis of Prototypical Parts, making it easier for users to understand the features that the model is focusing on.\\n\\n3.Through user studies, it was proven that the explanations provided by LucidPPN are clearer and easier for users to understand than those of other models such as PIP-Net. This empirical result helps to enhance the persuasiveness of the method.\", \"weaknesses\": \"Weakness\\n\\n1.The Section 3 has a lot of paragraphs but lacks subheadings, making it difficult to follow the logical flow of the different parts.\\n\\n2.There was no noticeable advantage in accuracy. The model was compared on four datasets in total, and its accuracy was lower than that of PIP-Net on two of the datasets, especially on the CUB dataset, where its accuracy was lower than that of all three methods, and no explanation was given for this gap.\", \"questions\": \"Concerns:\\n1.It is recommended to add subheadings to each key step or method description to make it easier for readers to understand and locate the content.\\n\\n2.Consider further improving the accuracy of LucidPPN to enhance its explainability while maintaining a minimal loss of performance.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}", "{\"title\": \"Thank you for your response\", \"comment\": \"Thank you for your response!\\n\\nI see why you chose sigmoid (instead of softmax) to be applied to the latent feature maps.\\n\\n\\\"As a result, one can easily verify if the image patches selected for visualization are faithful because such patches should have a resemblance score close to 1.\\\"\\n\\nIn order to establish that your prototype visualizations are faithful, did you verify that every image patch selected for the visualization of a prototype has a self-resemblance score close to 1 with the prototype itself? Is it true? What would you do if you found a prototype whose visualized patch did not have a self-resemblance score close to 1?\"}", "{\"summary\": \"The manuscript presents the Lucid Prototypical Parts Network (LucidPPN), designed to identify key visual features\\u2014specifically color, shape, and texture\\u2014based on the prototypical parts networks. The proposed LucidPPN utilizes a non-color branch to process grayscale images alongside a color branch that focuses on color information, thereby clarifying the model's decisions based on these visual attributes. Experimental results demonstrate that the proposed method exhibits advantages over baseline approaches and generates more interpretable prototype parts.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"(1)\\tThe methodology is well-structured, with intuitive design in the separation of color and non-color network branches, making it accessible and easy to understand.\\n\\n(2)\\tThe experiments are comprehensive, with a substantial number of visualization results provided in the appendix, enhancing the manuscript's depth.\", \"weaknesses\": \"(1)\\tWhile analyzing \\\"color,\\\" \\\"shape,\\\" and \\\"texture\\\" offers a valuable perspective, these features have been extensively studied in the field of visual perception. Given that the shallow layers of deep networks are capable of extracting low-level features, the necessity for additional processing and analysis from prototypical parts raises concerns on the novelty and contribution of this work.\\n\\n(2)\\tThe improvements demonstrated by the proposed method appear to be limited because its performance on some instances is lower than that of the compared methods. For example, in Table 1, the proposed method underperforms other prototypical parts networks on some datasets. While color, shape, and texture are indeed significant visual features in interpretability, they may not be sufficiently critical in this context.\\n\\n(3)\\tThe organization of the experimental section appears somewhat unbalanced. While the results and visualizations presented are commendable, an excessive amount of content is relegated to the appendix, which may hinder the reader\\u2019s ability to grasp key insights and maintain a coherent narrative.\", \"questions\": \"(1)\\tThe manuscript focuses on interpretability through the lenses of color, shape, and texture. However, other low-level features such as edges, contrast, and spatial frequency are also relevant. Have alternative low-level features also been considered in the analysis?\\n\\n(2)\\tThe datasets utilized in the experiments are relatively small in size. How will the proposed method perform on larger datasets, such as ImageNet? Some insights into performance scalability would be beneficial.\\n\\n(3)\\tThe manuscript primarily presents visualization results for the prototypical parts identified by the proposed method. How do these results compare with other prototypical parts-based models? A comparative analysis would enhance the understanding of the method's effectiveness.\\n\\n(4)\\tIn global feature visualizations, such as Figure 14, the manuscript illustrates the ability of the proposed method to detect shape and color. How does this compare with traditional edge detection operators (e.g., Sobel) for shape extraction and color feature extraction methods (e.g., color histogram)? Additionally, how does it compare with the direct visualizations of shallow layer attention to texture and color using techniques like Grad-CAM?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"details_of_ethics_concerns\": \"N/A\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Thank you for your clarification\", \"comment\": \"Thank you for your clarification. I have raised my score to 6: marginally above the acceptance threshold.\\n\\nTo further improve the paper, it would be interesting to see what happens if you remove all prototypes whose self-resemblance score is not close to 1.\\n\\nAlso, it would be helpful to add a discussion on why you chose sigmoid (instead of softmax) and binary cross entropy (instead of multi-class cross entropy) for your method.\"}", "{\"comment\": \"We sincerely thank the Reviewers for their positive and encouraging feedback. They have recognized that our work can *serve as a significant inspiration for future research* (R9gD3), introduces *a novel idea of disentangling color from shape and texture* (R-na4o), and addresses a *significant field of research in machine learning* (R-na4o).\\n\\nThe clarity and interpretability of LucidPPN have been particularly appreciated. Reviewers have highlighted that it *can reduce the ambiguity of traditional prototype networks and enable users to better understand the reasons behind the model\\u2019s decisions* (R-Zw4S) and that the model makes *it easier for users to understand the features that the model is focusing on* (R-Zw4S). Furthermore, it was emphasized that *the explanations provided by LucidPPN are clearer and easier for users to understand.* (R-Zw4S). \\n\\nLucidPPN has also been noted for its presentation, described as *clearly written* (R-na4o), and *easy to follow* (R-9gD3). Reviewers appreciated that *this paper provides sufficient cases and visualizations to validate the semantic information of the learned prototypes* (R9gD3) and that *the methodology is well-structured, with intuitive design* (R-9WXj). The experiments were recognized as *comprehensive, with a substantial number of visualization results.* (R-9WXj). \\n\\nWe have carefully addressed the Reviewers\\u2019 comments and incorporated their suggestions to strengthen our manuscript. We kindly ask the Reviewers to consider increasing their rating if they find our responses satisfactory. Responses to remarks shared among Reviewers are provided below, and followed by replies to specific comments. Additionally, we have attached a revised version of the work with all changes highlighted in blue.\\n\\n\\n### **Shared remarks**\\n\\n**(R-Zw4S, R-jXW9) There was no noticeable advantage in accuracy. Why?**\\nThe primary goal of this work was not to surpass PIPNet in accuracy but to reduce the ambiguity of prototypical parts through color disentanglement and correspondence to semantic parts of classified objects. Multiple works Adebayo et al. (2018); Huang et al. (2023); Kim et al. (2022); Ma et al. (2024b) show that explanations are ambiguous for a user and can cause overconfidence. That is why one should consider user study as the main result that shows that LucidPPN enabled significantly better user scores than PIPNet, even on the CUB dataset where LucidPPN\\u2019s accuracy was lower. \\n\\n**(R-Zw4S, R-jXW9) The improvements demonstrated by the proposed method appear to be limited because its performance on some instances is lower than that of the compared methods.**\\nThe accuracy drop stems from a late-stage fusion of texture and shape features with color. This delay prevents the network from correlating shape and texture with color effectively in earlier layers, causing some features to go undetected. It can be seen as a multimodal scenario, where early fusion (in our case PIPNet) achieves higher accuracy than late fusion (in our case LucidPPN), just like in Nagrani et al. (2021). Nonetheless, increasing the disambiguation of prototypical parts can improve accuracy over PIPNet, like in 3 of the 5 datasets, including PartImageNet added in the rebuttal phase.\"}", "{\"summary\": \"This paper propose to disentangle color prototypes from other visual features in ProtoPNets, by introducing a novel network architecture, named LucidPPN. The proposed method clarifies feature importance and aligns prototypical parts with object semantics, enhancing interpretability. Experiments show that LucidPPN achieves competitive accuracy while producing clearer and less ambiguous explanations for users.\", \"soundness\": \"4\", \"presentation\": \"4\", \"contribution\": \"4\", \"strengths\": [\"This paper explicitly decouple prototypes into specific semantic types, such as color and shape, whereas existing methods have overlooked this aspect of information. And I believe this paper could serve as a significant inspiration for future research.\", \"This paper provides sufficient cases and visualizations to validate the semantic information of the learned prototypes.\", \"The paper is well-written and easy to follow.\", \"The authors provide code for reproducibility check.\"], \"weaknesses\": \"[Major]\\n\\n1. **Quantitative evaluation of the interpretability:** In previous work, Huang et al. [1] have discussed the inconsistency of traditional ProtoPNets. Does this issue exists within the proposed method? Please provide qualitative or quantitative evaluations.\\n2. **Experiments:** Please supplement the missing results for baseline methods on datasets like DOGS and FLOWERS in Table 1, as adapting to these datasets, which were not covered in the original papers, seems quite straightforward.\\n3. **Experiments:** This paper only implement the proposed method on several CNNs. However, vision Transformers are introduced to the realm of CV for several years, and have also been implemented as the backbone of ProtoPNets [2]. Please provide additional experimental results using ViT [3-4] or even CLIP [5] as the backbone.\\n4. **Related Work:** In XAI, introducing human understandable semantics as evidences for prediction has been explored by concept bottleneck models (CBMs) [6]. What is the relationship between the proposed method and CBMs. Can concepts be introduced into the realm of ProtoPNet for higher interpretability?\\n\\n\\n> [1] Huang, Qihan, et al. \\\"Evaluation and improvement of interpretability for self-explainable part-prototype networks.\\\" Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023.\\n> \\n> [2] Xue, Mengqi, et al. \\\"Protopformer: Concentrating on prototypical parts in vision transformers for interpretable image recognition.\\\" arXiv preprint arXiv:2208.10431 (2022).\\n>\\n> [3] Dosovitskiy, Alexey, et al. \\\"An image is worth 16x16 words: Transformers for image recognition at scale.\\\" International Conference on Learning Representations. 2021.\\n>\\n> [4] Touvron, Hugo, et al. \\\"Training data-efficient image transformers & distillation through attention.\\\" International conference on machine learning. PMLR, 2021.\\n>\\n> [5] Radford, Alec, et al. \\\"Learning transferable visual models from natural language supervision.\\\" International conference on machine learning. PMLR, 2021.\\n>\\n> [6] Koh, Pang Wei, et al. \\\"Concept bottleneck models.\\\" International conference on machine learning. PMLR, 2020.\\n\\n[Minor]\\n\\n1. **Experiments:** What is the computational cost of inference and training? Please provide a comparison with baseline methods, including metrics such as training time, FLOPs, and memory usage.\", \"questions\": \"My questions are listed in \\\"Weaknesses\\\" section.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"**Q1. The manuscript focuses on interpretability through the lenses of color, shape, and texture. However, other low-level features such as edges, contrast, and spatial frequency are also relevant. Have alternative low-level features also been considered in the analysis?**\\n\\nThis work represents the first step toward disentangling low-level features. As mentioned in the Limitations Section, we plan to explore low-level features in more detail in future work. For now, we focus on extracting color information and correlating prototypical parts with semantic object parts. These contributions already make the work comprehensive. Introducing additional low-level feature extraction and integration at this stage would complicate the model further and make it more difficult to communicate.\\n\\n\\n**Q2. The datasets utilized in the experiments are relatively small in size. How will the proposed method perform on larger datasets, such as ImageNet? Some insights into performance scalability would be beneficial.** \\n\\nMost benchmarking of prototypical part-based methods has been conducted on fine-grained datasets such as CUB and Stanford Cars Chen et al. (2019); Nauta et al. (2021; 2023b); Rymarczyk et al. (2021; 2022; 2023); Wang et al. (2024). Scaling these architectures to ImageNet-sized datasets is an important but orthogonal research direction that remains unsolved. However, to assess whether LucidPPN generalizes to broader classification tasks (beyond fine-grained datasets), we added results on PartImageNet He et al. (2022) in the Supplementary Materials. For this dataset, LucidPPN achieves an accuracy of 84.1%, outperforming PIPNet, which achieves 82.8%.\\n\\n**Q3. The manuscript primarily presents visualization results for the prototypical parts identified by the proposed method. How do these results compare with other prototypical parts-based models? A comparative analysis would enhance the understanding of the method\\u2019s effectiveness.**\\n\\nIn Supplementary Figures 15-19, we added explanations from various prototypical-part-based methods. Additionally, our user studies demonstrate the effectiveness of our explanations from the user\\u2019s perspective.\\n\\n\\n**Q4. In global feature visualizations, such as Figure 14, the manuscript illustrates the ability of the proposed method to detect shape and color. How does this compare with traditional edge detection operators (e.g., Sobel) for shape extraction and color feature extraction methods (e.g., color histogram)? Additionally, how does it compare with the direct visualizations of shallow layer attention to texture and color using techniques like Grad-CAM?**\\n\\nEdge detectors and color histograms are not trainable and do not represent high-level features, unlike prototypical parts. Even our low-level color prototypical representation captures a higher-level concept, such as a \\u201dred tail,\\u201d which is then further decomposed into its components\\u2014color (red) and shape/texture (tail). Regarding GradCAM, we want to highlight that post-hoc methods can often be unreliable, as demonstrated in multiple studies Adebayo et al. (2018); Kim et al. (2022); Rudin (2019); Tomsett et al. (2020). This underscores the need for developing inherently interpretable models Rudin (2019); Rudin et al. (2022) such as our LucidPPN.\"}", "{\"comment\": \"**W1. In previous work, Huang et al. (2023) have discussed the inconsistency of traditional ProtoPNets. Does this issue exist within the proposed method? Please provide qualitative or quantitative evaluations.**\\n\\nTo answer this question we have calculated consistency and stability of our method and compared it in Supplementary Table 11. One can observe that LucidPPN achieves a consistency score comparable to the method proposed by Huang et al. (2023) while outperforming other prototypical-parts-based methods. Regarding stability, LucidPPN demonstrates results on par with other methods. This improvement is likely due to the correspondence of prototypical parts to semantic parts of the classified objects.\\n\\n**W2. Please supplement the missing results for baseline methods on datasets like DOGS and FLOWERS in Table 1, as adapting to these datasets, which were not covered in the original papers, seems quite straightforward.** \\n\\nThank you for your comment. We have added results for baselines on additional datasets, except for the ProtoTree. This exception is due to the model\\u2019s tendency to exhibit instability during training on these datasets. Upon reviewing the issues section of the ProtoTree GitHub repository [https://github.com/M-Nauta/ProtoTree/issues](https://github.com/M-Nauta/ProtoTree/issues), we noticed that others have faced similar challenges in applying this model to different datasets.\\n\\n**W3. This paper only implement the proposed method on several CNNs. However, vision Transformers are introduced to the realm of CV for several years, and have also been implemented as the backbone of ProtoPNets Xue et al. (2022). Please provide additional experimental results using ViT or even CLIP as the backbone.**\\n\\nThank you for your comment. Unfortunately, we are unable to run LucidPPN with a ViT backbone for the following reasons: \\n* **Incompatibility with PIPNet-Based Prototypical Part Definition**: ProtoPFormer Xue et al. (2022) is built on the ProtoPNet-based definition of prototypical parts Chen et al. (2019), whereas our method relies on the PIPNet-based definition Nauta et al. (2023b). Currently, there is no adaptation of the PIPNet-based definition for the ViT backbone. This limitation is why we opted for the ConvNeXt backbone, which has demonstrated comparable performance to ViTs. \\n* **Challenges with Self-Attention**: Adapting a ViT backbone to the PIPNet-based definition is not straightforward due to the nature of self-attention. Unlike convolutions, self-attention lacks the properties of locality and a direct correspondence between the input and feature map Chen et al. (2019). This discrepancy makes it difficult to visualize prototypical parts faithfully. \\n* **Orthogonal Research Direction**: Adapting the ViT backbone to prototypical parts represents a separate research direction. Both ProtoPFormer and recent works in this area Ma et al. (2024a), which were unavailable at the time of submission, highlight that integrating a ViT backbone with prototypical parts is a non-trivial task requiring substantial architectural/training changes. For these reasons, we chose to use the ConvNeXt backbone in our work.\\n\\n\\n**W4. In XAI, introducing human understandable semantics as evidence for prediction has been explored by concept bottleneck models (CBMs) Koh et al. (2020). What is the relationship between the proposed method and CBMs. Can concepts be introduced into the realm of ProtoPNet for higher interpretability?** \\n\\nThank you for pointing out the connection between CBMs and our work. I\\u2019d like to clarify that both concept bottlenecks and prototypical parts can be considered concept-based models Bontempelli et al. (2022). However, there is a key distinction: concept bottlenecks use predefined intermediate classes (named concepts) that are directly associated with the image. While, prototypical parts, aim to identify relevant classification concepts during model training without any additional labels. A potential future research direction could involve combining concept bottlenecks with prototypical parts.\\n\\n**w1. What is the computational cost of inference and training? Please provide a comparison with baseline methods, including metrics such as training time, FLOPs, and memory usage.**\\n\\nIn Supplementary Table 12 we provide information about training time, GFLOPs needed, and average memory usage during training for LucidPPN, PIPNet, ProtoPool, and ProtoPNet. One can observe that LucidPPN is faster and uses less memory than PIPNet. However, ProtoPNet and ProtoPool require much less memory to train while having longer training times.\"}", "{\"comment\": \"Thanks for the prompt response.\\n\\nThe concerns regarding accuracy have been largely addressed. We agree with the authors on their explanation that by decoupling color, shape, and texture in the early stages and fusing them later for more accurate interpretation, there is an inherent trade-off in terms of accuracy.\\n \\nRegarding concerns about the contribution, some doubts still remain. While we acknowledge that the authors have made valuable contributions to the prototype network by introducing inherently interpretable mechanisms at the low-level feature level, similar research [1] on shallow network visualizations has already provided substantial insights into the impact of features such as color, shape, and texture on classification network results. This somewhat limits the scope of the contribution of the proposed method in the context of low-level feature interpretation. The authors are still expected to further clarify the unique advantages of their approach compared to shallow network visualization in the revised version, which would definitely better highlight the contribution and innovation of the manuscript.\\n\\nBut overall, after the above rounds of feedback and discussion, I think this work have adequate technical merits and contributions, and I would be happy to raise the score to 6 (marginally above accept).\\n \\n1. Zeiler, M. D. (2014). Visualizing and Understanding Convolutional Networks. In European Conference on Computer Vision.\"}", "{\"comment\": \"Thank you for your valuable feedback and for raising the score.\\n\\nIt is indeed interesting to see the effects of pruning the prototypes with less faithful representation (those with resemblance scores < 0.9). Therefore, we investigate it in the newly added section \\\"Pruning prototypes with less faithful visualizations\\\" of the Supplementary Materials.\\nIt contains Table 13, which shows that LucidPPN accuracy after pruning drops only by around 2% (from 81.6% to 79.3%). However, interestingly, the accuracy stays the same for $L_C=0.05$.\\nIt suggests that combination of $L_C$ and pruning allows to enforce high resemblance scores (>0.9) of visualized patches without sacrificing on the accuracy.\\n\\nWhen it comes to the discussion on choosing sigmoid and binary cross entropy, we added it as section \\\"Reason behind using the Binary Cross Entropy with Sigmoid instead of the Cross Entropy with Softmax\\\" of the Supplementary Materials.\"}", "{\"comment\": \"Dear Reviewer Zw4S,\\n\\nAs the deadline for the discussion period is approaching quickly, we would like to kindly remind the reviewer that we are waiting for your response.\\n\\nIn particular, we have provided point-by-point responses to all of your questions to address your concerns and provided the revision that reflects such changes. Therefore, your timely feedback and change in the score if applicable would be highly appreciated.\\n\\nBest,\\n\\nAuthors\"}" ] }

[ "Shizhen Zhao", "Xin Wen", "Jiahui Liu", "Chuofan Ma", "Chunfeng Yuan", "XIAOJUAN QI" ]

[ "Long-Tail Learning; Granularity; Category extrapolation" ]

{ "note_id": [ "wD2PGYghc9", "tKZt7F49yd", "pevZpZCNAC", "eZnfwFoJQX", "FnRxG2S06H", "9FIqEfw9T1" ], "note_type": [ "official_review", "comment", "official_review", "official_comment", "official_review", "official_review" ], "note_created": [ 1730696592664, 1731647930072, 1730369389916, 1731647915289, 1730711688595, 1730868495553 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission1815/Reviewer_DAEc" ], [ "ICLR.cc/2025/Conference/Submission1815/Authors" ], [ "ICLR.cc/2025/Conference/Submission1815/Reviewer_U6mt" ], [ "ICLR.cc/2025/Conference/Submission1815/Authors" ], [ "ICLR.cc/2025/Conference/Submission1815/Reviewer_86cN" ], [ "ICLR.cc/2025/Conference/Submission1815/Reviewer_2XU9" ] ], "structured_content_str": [ "{\"summary\": \"This paper presents a strategy for classification in an imbalanced-class setting, which relies on expanding the training dataset via searching the internet for examples of classes related to rare classes in the existing dataset. The paper proposes a method for downloading additional data for nearby classes and a strategy for training from this additional data. The paper reports results on ImageNet-LT and iNaturalist, and shows that the proposed method can improve over state-of-the-art methods for long-tail recognition without using external data.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"1.\\tThe method for the paper is well described and easy to understand.\\n2.\\tThe authors show that the method can be applied on top of a number of prior approaches, displaying the versatility of the method as an add-on to existing approaches.\\n3.\\tIt\\u2019s nice to see that approach also helps over large-scale pre-trained models like CLIP, a common concern for long-tail methods.\", \"weaknesses\": \"1.\\tGenerally, it appears to me that the key improvement from the method is from downloading additional data similar to the rare classes. This is a fairly different setting than most long-tail approaches, which attempt to tackle the problem of learning only from the data you have. Given this, it\\u2019s difficult to compare this to most prior methods reported in the paper, and I would expect more thorough baselines to make up for this.\\na)\\tFor example, the authors highlight Iscen et. al., 2023, as related work, but I would have appreciated a more thorough comparison to Iscen et. al., both in terms of their method of obtaining additional data, and in terms of experimental results.\\nb)\\tAdditionally, I would have appreciated a baseline which simply downloads more images by searching for the rare class from the internet directly. \\n2.\\tI\\u2019m confused about the focus on granularity in the writing. It seems the real win here is from getting more training data \\u2013 which is totally fine (with appropriate baselines). But the writing + experiments around granularity paint a different picture.\\na)\\tFor example, Table 1 aims to show that a finer granularity helps with dataset imbalance. But a confounding factor here is the dataset size, which the comparison doesn\\u2019t control for. Similarly, to my understanding, Figure 3 is more about the distribution of classes, and not necessarily about the granularity. I would buy granularity if the proposed method used the same data, but with more labels or more fine-grained labels. Otherwise, this appears to be a story about more data, not granularity.\", \"questions\": \"I'd like the authors to address my concerns in Weaknesses above around the baseline comparisons (to Iscen et al, and to a simpler baseline), and the narrative around granularity.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"withdrawal_confirmation\": \"I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors.\"}", "{\"summary\": \"The paper proposes a new method for addressing long-tail learning problem via incorporating open-set data that belong to neighbor categories. The method is based on an interesting observation that datasets with finer granularity suffer less from the class imbalance issue. Authors of the paper first conduct a controlled experiment to support their observation, then describe the pipeline for querying neighbor categories from GPT-4 and filtering the web-crawled data. A neighbor-silencing re-balancing loss is proposed to balance the training, and the inference time classifier is obtained via masking out all the weights corresponding to neighbor categories. Experiments are conducted on three class-imbalanced benchmarks and the proposed method generally outperforms various previous approaches by a large margin.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. The authors provide a novel observation of the relation between dataset granularity and model long-tail performance which, although somewhat counter intuitive, is supported by a well-constructed controlled experiment.\\n\\n2. Experiment results are strong, showing that the proposed method outperforms baselines by a large margin.\\n\\n3. The paper is well organized and easy to follow.\", \"weaknesses\": \"1. The reviewer has several questions regarding the rationale of the method and the results of the experiments. They are not necessarily weaknesses, but the reviewer expects them to be addressed. Please refer to the question part.\\n\\n2. The reviewer suggests that the figure 3 should also demonstrate the exact accuracy of the head and tail categories in addition to their relative accuracy gap. Because it is possible that the more fine-grained dataset is more difficult to learn, and thus the major reason for the smaller accuracy gap is the significant accuracy degradation in head categories, which is undesirable. The experiment should clarify this point.\\n\\n3. There are details that need improvement:\\n- In section 2.1, notations $\\\\theta_f$ and $\\\\theta_w$ should be explained.\\n\\n- There is a citation error in section 2.1, line 160.\\n\\n- Appendix D.1 suggests that there is a missing section 3.3. In fact, the reviewer would like to see the discussion that should have appeared in section 3.3.\", \"questions\": \"1. The rationale of the method is unclear. Since it is feasible to craw data from the internet, why not directly augment the tail categories, or at least incorporate the tail categories into the querying list for the searching engine together with the neighbor categories?\\n2. Since the setting does not consider performance on auxiliary categories, an alternative method would be simply treating all the data from neighbor categories as the augmentations for the target category (i.e., they share the same label). In this way, the overwhelming problem naturally disappears and therefore no need to introduce the weight $\\\\lambda_s$. The reviewer would like to see the performance comparison against this variant. \\n3. Comparing the results in table 2 and table 3, the reviewer wonders the discrepancy between the two results based on CLIP pre-trained model. Why the results are different?\\n4. Does the baseline in ablation study (table 6) use auxiliary categories? If it does, the reviewer suppose that the results should maintain consistent with the results in table 3. (Overall accuracy should be 77.9 instead of 79.6).\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"We sincerely appreciate the reviewers for dedicating their time and effort to reviewing our work and for recognizing the potential contributions we may have made. The insightful feedback, constructive comments, and suggestions provided by the reviewers have significantly enhanced the quality and clarity of our work. We will incorporate these valuable suggestions into the revised version to further improve the content.\"}", "{\"summary\": \"This paper finds that finer-granularity datasets perform better in the face of imbalance. Using this finding, the authors propose using LLMs to define additional classes and gathering additional web images for these classes to augment the dataset. The authors also introduce a neighbor-silencing loss to account for these additional classes. This new method outperforms the baseline across different model types (scratch, CLIP, DINOv2).\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": [\"Empirical initial experiments that motivate the solution for the bigger picture problem\", \"Simple solution that is potentially scalable\", \"Strong improvement over the baseline numbers for a variety of image models\"], \"weaknesses\": \"The implementation details (4.2) were a little unclear so I didn't fully understand the experimental setup. This leads to the following concerns.\\n\\nI'm unsure if the results in the experimental section are a fair comparison. The baseline seems to be trained on less data, so the gains cannot be completely attributed to the new method. Similar concerns can be raised for the comparison with other methods. Regarding the fair comparison paragraph in 4.3, these methods were not designed for auxiliary classes so maybe you would need to collect additional images for the long-tail classes for the other methods and show that your method does better even though you are allocating resources towards auxiliary classes instead of the actual long-tail class.\\n\\nIt also seems like the new dataset has a higher proportion of long-tail (or long-tail neighbor) images. This changes the distribution of your training set. It would be good to make clear what part of the improvements are from using the auxiliary classes versus this change in distribution.\\n\\nI would be happy to adjust my rating once these details are clarified or addressed.\", \"questions\": \"What do you do with existing images in the original dataset of auxiliary classes? Sometimes images can also belong to multiple classes. Would improving your pipeline with respect to these issues improve downstream performance?\\n\\nLong-tail classes are often rare for a reason. What if it is difficult to find some of their auxiliary classes on the internet?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper addresses the challenge of training deep learning models on long-tail data distributions, where limited sample diversity hinders the learning of robust features, especially for tail classes. The authors propose a novel approach to overcome this issue by increasing dataset granularity. They introduce open-set auxiliary classes, which are visually similar to existing classes, to enhance feature representation learning for both head and tail classes. To generate these auxiliary categories, the method leverages large language models (LLMs) to search for and retrieve relevant images. Additionally, a neighbor-silencing loss is introduced to prevent auxiliary classes from disrupting the model\\u2019s focus on class discrimination within the target dataset. The approach is validated through extensive experiments on long-tail benchmarks, demonstrating superior performance over existing methods.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. This work approaches long-tail learning from a novel angle by proposing to increase dataset granularity as a way to address data imbalance, differing from traditional re-sampling or re-weighting methods. The authors observe that datasets with finer granularity are typically less affected by imbalance, leading them to introduce category extrapolation to enhance granularity.\\n\\n2. On several standard long-tail benchmarks, the proposed method outperforms other strong baselines using the same data volume, showcasing its effectiveness in the long-tail learning setting. Comprehensive ablation studies further demonstrate the impact of each component of the approach.\\n\\n3. The paper is well organized and clearly written, making it easy for readers to understand the motivation, methodology, and experimental results.\", \"weaknesses\": \"1.Although the proposed method shows performance improvements, its resource costs, dataset construction time, and human resource costs are significantly higher. For example, the introduced data volume is nearly ten times that of the original dataset, indicating that researchers applying this method to other domains or datasets would need to invest substantial time and resources, making the method difficult for the community to adopt.\\n\\n2.The authors expand the dataset by adding multiple times more data to introduce auxiliary categories. However, if these resources were instead directed toward expanding existing categories, could similar performance gains be achieved with less data? Discussing this would help the community and readers better understand the advantages and limitations of the method.\\n\\n3.The paper lacks discussion of using large language models (LLMs) for addressing the long-tail issue in this field, as seen in methods like LTGC[1]. Additionally, exploring generative models as an alternative approach for auxiliary category expansion could offer promising solutions?\\n\\n4.The work lacks data visualizations for the added categories, which limits the demonstration of the effectiveness of the proposed pipeline. \\n\\n[1]Zhao Q, Dai Y, Li H, et al. LTGC: Long-tail Recognition via Leveraging LLMs-driven Generated Content[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024: 19510-19520.\", \"questions\": \"1.In Table 2, why are the results on the iNaturalist18 dataset for the baseline and the proposed method trained from scratch higher than those based on CLIP?\\n\\n2.What are the specific values for the neighbor category ratios across head, medium, and tail classes for different datasets?\\n\\n3.Could you provide more visualizations of the additional categories across different datasets? I\\u2019m particularly curious if enough additional categories are generated for fine-grained datasets like iNaturalist. For the semantic level, the authors designed a DINO-based method, but does the web-crawling engine also introduce noisy labels?\\n\\n4.In the limitations section, it mentions that \\u201cif the model has not seen or is unfamiliar with our query, this step will fail.\\u201d What is the failure rate across different datasets, and which categories tend to fail?\\n\\n5.Compared to web-crawling for images, have you considered using generative models like Stable Diffusion 3.0 for image generation?\\n\\n6.After expanding the categories, 4.1M, 1.1M, and 3.6M data points were collected. What image resolutions were used for training, and what was the training time on a 4-card RTX 3090 setup?\\n\\n7.What value was set for the $\\\\gamma_1$ and $\\\\gamma_2$ threshold? Does it vary across different datasets?\\n\\n8.How does filtering the data using DINOv2, rather than CLIP, affect models initialized in different ways?\\n\\n9.For iNaturalist18, how many images are approximately generated per auxiliary category? If a generated category name is the same as an existing category in the dataset or is a synonym, does it impact the results?\\n\\n10.Is the data open source?\\n\\n11.Typo: DIONOv2 should be corrected to DINOv2.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}" ] }

[ "Hedda Cohen Indelman", "Tamir Hazan" ]

[ "representation properties", "Gumbel-Softmax", "Gumbel-Argmax", "minimality", "completeness", "discrete probabilistic models" ]

{ "note_id": [ "vxrufWmryF", "otJGfJDurX", "ip8XBBUFfU", "fEawzvcZfR", "ao3yQvR6Xj", "WtwVnEp3qj", "Wh07hKj8Xy", "Uy78PE0Y8y", "UjOQpwpqUP", "SPjBWp4L0e", "Miwmlwg9hB", "LrLlZCPKCZ", "Hlx408ERF3", "Fw2KJTDSxM", "5VTNLVNdtX", "3RnhJCdzrV" ], "note_type": [ "official_review", "official_comment", "decision", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "meta_review", "official_comment", "official_comment" ], "note_created": [ 1730602820334, 1731871624667, 1737523640970, 1732487803372, 1732487911309, 1730890719273, 1732130936435, 1732065386339, 1732625318154, 1731655202438, 1731220029388, 1732626312606, 1732131040777, 1734557589232, 1732133089478, 1731657363664 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission4453/Reviewer_P6qr" ], [ "ICLR.cc/2025/Conference/Submission4453/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission4453/Reviewer_e4WY" ], [ "ICLR.cc/2025/Conference/Submission4453/Reviewer_e4WY" ], [ "ICLR.cc/2025/Conference/Submission4453/Reviewer_J5XZ" ], [ "ICLR.cc/2025/Conference/Submission4453/Authors" ], [ "ICLR.cc/2025/Conference/Submission4453/Reviewer_e4WY" ], [ "ICLR.cc/2025/Conference/Submission4453/Reviewer_P6qr" ], [ "ICLR.cc/2025/Conference/Submission4453/Reviewer_P6qr" ], [ "ICLR.cc/2025/Conference/Submission4453/Reviewer_e4WY" ], [ "ICLR.cc/2025/Conference/Submission4453/Reviewer_J5XZ" ], [ "ICLR.cc/2025/Conference/Submission4453/Authors" ], [ "ICLR.cc/2025/Conference/Submission4453/Area_Chair_vRDf" ], [ "ICLR.cc/2025/Conference/Submission4453/Authors" ], [ "ICLR.cc/2025/Conference/Submission4453/Reviewer_e4WY" ] ], "structured_content_str": [ "{\"summary\": \"The paper characterizes the the ability of particular parametrized families of probability mass functions (PMFs) to fit an arbitrary PMF on a finite probability space. The main result is a set of conditions under which the Perturbed-Softmax and Perturbed-Argmax families enjoy completeness (whether the parameter-to-distribution map is onto) and/or minimality (whether the parameter-to-distribution map is one-to-one).\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": [\"The paper is concise and overall well-written (see Weaknesses for feedback). In particular, the proofs in the main text and footnotes are helpful and executed well; usually, this type of presentation is hard to get right.\", \"The mathematical analysis is elegant and general enough to account for any perturbation distribution.\", \"The experimental testbed is generally helpful, although I do not believe the provide much insight regarding the theory (see Weaknesses).\", \"The appendices are helpful for making the paper self-contained overall, but the proof appendices should not be theorem-proof lists. Try to make them self-contained as well; concretely, please reintroduce the theorem, add proof outlines, discuss any techniques that are novel to this paper versus those that are adapted from others with citations, etc)/\"], \"weaknesses\": [\"The earlier part of the presentation could be improved, in that components from the Related Work and Preliminaries could be incorporated into the Introduction. For example, I don't believe the paper mentions one of the major motivations of the Gumbel-Softmax as an approximation to the discrete sampling operation in VAEs that can be backpropagated through. Similarly, while online learning methods are stated as a motivation, the reader should be able to know where the perturbed-argmax/softmax operations appear in their methods, even without having a background in online learning. Equations for the perturbed operations could appear much earlier in the text, so that the Gumbel-softmax/argmax, Gaussian-softmax/argmax, and their perturbed variants are not so abstract.\", \"Unless I may be misunderstanding, their seem to be some technical errors in the main text proofs. For example, the proof of Theorem 4.1 does not use the $h_i$ function, so it must be incomplete. This assumption is in fact used in Theorem 5.1, but not included in the theorem statement.\", \"I do not believe the current experiments align well with rest of the paper. In particular, what is shown is that Gaussian perturbations result in a predictor that learns faster with respect to validation loss. However, the rest of the paper is about the representation capabilities of parametrized models, and if I understand correctly, the Gumbel and Gaussian perturbed models have the same representation properties. Moreover, this difference between Gaussian and Gumbel performance seems to already be understood in theory based on these concentration properties shown earlier in the paper. I felt that an experiment which included a model that did not satisfy completeness conditions (within the same setup as 6.2) and shows the bias of this model at the learned minimizer would be much more illuminating.\"], \"questions\": [\"Should line 160 only consider $p \\\\in \\\\operatorname{ri}(\\\\Delta)$?\", \"Should Theorem 3.1 include the sub-Gaussianity parameter?\", \"Should line 240 say \\\"Gumbel\\\" instead of \\\"exponential family\\\"?\", \"In Theorem 4.1, what does \\\"whose cumulative distribution decays to zero\\\" mean? Can you make this precise, as CDFs do not decay to zero?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Regarding the correctness of Theorem 5.1\", \"comment\": \"Respectfully, Theorem 5.1 as stated is **not false**: the theorem relies on the connection between Equation 14 $(ri(\\\\Delta) \\\\subseteq E_{\\\\gamma} [\\n \\\\arg\\\\max(\\\\theta + \\\\gamma)] \\\\subseteq \\\\Delta)$ and Equation 13 $(\\\\partial f(\\\\theta) = E_{\\\\gamma} [\\\\arg \\\\max(\\\\theta+\\\\gamma)])$. Equation 13 considers the expected argmax and sub-gradient of the expected max-value (defined in Equation 12 $(f(\\\\theta) = E_{\\\\gamma} [\\\\max_{i}${$\\\\theta_i+ \\\\gamma_i$}])). In lines 728-729 we define the sub-gradient explicitly, which we quote here as well: A sub-gradient $p \\\\in \\\\partial f(\\\\theta)$ satisfies $f(\\\\tau) \\\\ge f(\\\\theta) + \\\\langle p, \\\\tau - \\\\theta \\\\rangle$ for every $\\\\tau \\\\in \\\\Theta$.\\n\\nAs a test case, let us consider the example given by Reviewer e4WY, which considers a point mass at zero. For simplicity, let's assume that $d=2$, i.e., $\\\\theta = (\\\\theta_1,\\\\theta_2)$. In this case: \\n\\n$f(\\\\theta)$ = $\\\\max${$\\\\theta_1, \\\\theta_2 $\\\\}\\n\\n$\\\\partial f(\\\\theta) = \\\\arg \\\\max$($\\\\theta_1, \\\\theta_2)$ \\n\\n$(p_1,p_2) \\\\in \\\\partial f(\\\\theta) $ $ \\\\text{ iff }$ $\\\\forall (\\\\tau_1,\\\\tau_2) \\\\max $ {$\\\\tau_1, \\\\tau_2 $} $\\\\ge \\\\max${$\\\\theta_1, \\\\theta_2$} + $\\\\langle p, \\\\tau - \\\\theta \\\\rangle$.\\n\\nIt is clear that if $\\\\theta_1 \\\\ne \\\\theta_2$ then the maximal argument is uniquely defined, and $\\\\partial f(\\\\theta)$ is the point-mass distribution. However, Danskin's Theorem states that when $\\\\theta_1 = \\\\theta_2$, then $p = (p_1,p_2)$ span the probability simplex.\\n\\nTo make this surprising result more intuitive, let's consider the case when $\\\\theta_1 > \\\\theta_2$. In this case $\\\\max${$\\\\theta_1,\\\\theta_2$} $= \\\\theta_1$ and the point-mass $p = (1,0)$ is the (sub-)gradient. To see that, one can verify that \\n$\\\\max$ {$\\\\tau_1, \\\\tau_2$} $\\\\ge \\\\tau_1 = \\\\theta_1 + (\\\\tau_1 - \\\\theta_1) = \\\\theta_1 + \\\\langle (1,0), \\\\tau - \\\\theta \\\\rangle $.\\n\\nHowever, when $\\\\theta_1 = \\\\theta_2$, for every $p = (p_1,p_2)$ that is a probability distribution (i.e., non-negative and sum up to unity) the following two equations hold: \\n\\n$\\\\max${$\\\\tau_1, \\\\tau_2 $} $\\\\ge \\\\langle p, \\\\tau \\\\rangle$\\n\\n$\\\\max${$\\\\theta_1, \\\\theta_2 $} $= \\\\langle p, \\\\theta \\\\rangle$\\n\\nand combining these two equations we get that any $p = (p_1,p_2)$ that is a probability distribution is also a sub-gradient, i.e., $p \\\\in \\\\partial f(\\\\theta)$: \\n\\n$\\\\max${$\\\\tau_1, \\\\tau_2$} $ \\\\ge \\\\langle p, \\\\tau \\\\rangle = \\\\max${$\\\\theta_1, \\\\theta_2$}$ + \\\\langle p, \\\\tau \\\\rangle - \\\\langle p, \\\\theta \\\\rangle $.\\n\\nThe general setting is described by [Danskin's theorem](https://en.wikipedia.org/wiki/Danskin%27s_theorem), where the compact set is the discrete set {$1,...,d$}. One can verify it by following the example in Sec 4.5, page 247 of the book Convex Analysis and Optimization, by Bertsekas and Ozdagler, 2003. \\n\\nWe also refer to Proposition 5.5 in the paper for an analysis of the Perturb-Argmax probability distribution when considering the case of $\\\\Theta = \\\\mathbb{R}^2$ and $\\\\gamma = (\\\\gamma_1,\\\\gamma_2)$ a vector of uniformly distributed discrete random variables. Concretely, as illustrated in Figure 3, the sub-differential of the expected max value spans the probability simplex.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"comment\": \"Point number 1: I wasn't asking so much about the citation as the fact that I don't see what the fundamental theorem of extreme value statistics has to do with the claim you make here. I still don't; if it were true that \\\"the representation properties of completeness and minimality of the softmax operation are identical to the properties of the Gumbel-Argmax probability distribution\\\", why would the rest of your paper be necessary?\", \"point_number_2\": \"Having thought about this more, I think we're both right. :)\\n\\nLet's think first about Perturb-Argmax; I'll change your notation to make things clearer. Here, you want to estimate the mean of a function $h(n)$ where $n$ is the integer from your discrete distribution, obtained as $n \\\\in \\\\operatorname{argmax} \\\\theta + \\\\gamma$; call this $n(\\\\theta, \\\\gamma)$, so the overall problem is $$\\\\min\\\\_\\\\theta \\\\operatorname*{\\\\mathbb{E}}\\\\_{x \\\\sim D}\\\\left[ \\\\operatorname*{\\\\mathbb{E}}\\\\_{\\\\gamma \\\\sim g} h(x, n(\\\\theta, \\\\gamma)) \\\\right] = \\\\min\\\\_\\\\theta \\\\operatorname*{\\\\mathbb{E}}\\\\_{x \\\\sim D}\\\\left[ \\\\operatorname*{\\\\mathbb{E}}\\\\_{n \\\\sim p_{\\\\theta,g}} \\\\left[ h(x, n) \\\\right] \\\\right].$$\\nIn practice, this would be estimated with Monte Carlo samples over both $x$ and $n$. Thus, for any $\\\\theta, g$ and $\\\\theta', g'$ such that the corresponding distributions over $\\\\mathbb N$ are the same, the Monte Carlo convergence rate will be identical; it doesn't matter at all if this is based on Gaussians, Gumbels, or whatever else. This is the case I was thinking about in my initial review.\\n\\nWhen using Perturb-Softmax, rather than obtaining a \\\"hard\\\" sample $n$, you get a \\\"soft\\\" sample of probabilities $\\\\hat p = \\\\operatorname{softmax}(\\\\theta + \\\\gamma)$. Here, equivalence between distributions means that $\\\\mathbb{E}\\\\_\\\\gamma \\\\hat p$ is the same, but the distributions of $\\\\hat p$ between equivalent, say, Gumbel-Softmax and Gaussian-Softmax distributions will be different. Since when using temperature 1 softmax is 1-Lipschitz, indeed we have that the function $\\\\gamma \\\\mapsto \\\\operatorname{softmax}(\\\\theta + \\\\gamma)$ is also 1-Lipschitz, and so when we want to estimate the mean of some function of $\\\\hat p$, the convergence rates can indeed be meaningfully different in terms of the Lipschitz properties of $h(\\\\hat p)$. That the worst-case bound over $h$ is sub-exponential rather than sub-gaussian is indeed suggestive. It would be nice to have a more thorough study of this, though: in the cases where Gaussian-Softmax performs better than Gumbel-Softmax, is this indeed what's happening? I *expect* (but it should be confirmed) that samples from Gumbel-Softmax have $\\\\hat p$ \\\"closer\\\" to one-hot, e.g. lower-entropy, than Gaussian-Softmax; it is reasonably intuitive that this would give slower-converging Monte Carlo behavior, but this is something that I think your paper needs to study more carefully.\"}", "{\"comment\": \"> We find it hard to address the comment that the work requires more empirical evidence, as it is general and not substantial.\\n\\nI gave some specific suggestions in my review of the kinds of empirical evidence that might be helpful to being more convincing in the benefits of Gaussian-Softmax over Gumbel-Softmax in practice. Your experiments are very minimal.\"}", "{\"summary\": \"The paper studies the Gumbel-max distributions in the mathematical view. The authors especially investigate the completeness, identifiability, and minimality of the distributions, i.e., the Purturb-Maxes. Finally, the advantage of Gaussian-softmax, which is an example of the Purturb-softmax, is demonstrated in the experiment section.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"The paper is clearly written and well-organized.\\n\\nThe work is theoretically grounded.\", \"weaknesses\": \"Even though this is a theoretical paper, the work requires empirical evidence more than the ones suggested in the manuscript.\\n\\nAlso, it is not clear how the current experiments are related to the theories provided in the main body.\\n\\nThe proofs can be moved to the appendix.\", \"questions\": \"In practice, the Gumbel-softmax is more widely utilized than the Gumbel-max, and the temperature parameter is naturally added to the distribution setting. However, it seems that it is dropped in the theoretical analysis. Why is that?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Regarding the discussion, definitions and conntribution\", \"comment\": \"Regarding the mathematical discussion and definition:\\n\\n1. Line 195 and the fundamental theorem of extreme value statistics.\\nWhile not explicitly phrased, we attribute understanding of the equivalence between the softmax distribution and the Gumbel-Argmax to Julius Gumbel, hence we referred to the 1954 notes. We agree that Fisher and Tippett have had the same understanding and we readily agree their work should be cited. Indeed, the same understanding also appears in Choice Theory literature (for example, Luce, 1959). Our goal was to reference the origin of the theory, though not Fisher and Tippett, nor Gumbel have cited this theorem explicitly. We will also cite Luce's work for completeness.\\n\\n2. A convergence rates analysis motivates our study of the Gaussian-Softmax and Gumbel-Softmax distributions' representation properties. Their different convergence rates and equivalent representation properties justify fitting probabilities with Gaussian instead of Gumbel random perturbations. The faster convergence of Gaussian-Softmax may hint at their efficient statistical nature that emerges from our experiments (i.e., their improved convergence when using the same number of samples.) \\n\\nWe agree that our main contribution is not primarily the theoretical study of the representation properties of Gumbel-Softmax and Gumbel-Argmax probability models. A better phrasing would be that our framework for investigating the representation properties of these models allows for extending it to identify the representation properties for a wide range of perturbation models. As for the completeness property of Gumbel-Softmax arising in the zero temperature limit in previous work (as mentioned by the reviewer) --- it is an insightful comment that we will add to the work and thank the reviewer for it. Our theoretical investigation also allows identifying the conditions under which a set of parameters $\\\\Theta$ itself is complete, rather than a property achieved by a convergence argument. \\n\\nRegarding \\\"easy-to-fix\\\" inaccuracies. Thank you for spotting these inaccuracies, we will surely correct them.\"}", "{\"comment\": \"I of course agree that the entire probability simplex is a subgradient of the argmax when the $\\\\\\\\theta$s are equal; this is incontrovertible and probably the most common case of subgradients that aren't gradients. (In retrospect, this should have been clear that this is where it came from.)\\n\\nThis abstraction about the connection to duality, though, I think obscures a very important point about the interpretation of the argmax distributions. Let's consider $\\\\gamma$ a point mass at zero in the two-dimensional case. Then, $\\\\\\\\{ \\\\\\\\mathrm{arg\\\\\\\\,max} \\\\\\\\, \\\\\\\\theta : \\\\\\\\theta \\\\\\\\in \\\\\\\\mathbb R^2 \\\\\\\\}$ can contain three kinds of settings: $\\\\\\\\theta\\\\_1 > \\\\\\\\theta\\\\_2$ (a point mass on variable 1), $\\\\\\\\theta\\\\_1 < \\\\\\\\theta\\\\_2$ (a point mass on variable 2), or $\\\\\\\\theta\\\\_1 = \\\\\\\\theta\\\\_2$ (not clearly defined). For the conclusion of your theorem to be true, the argmax operator must break ties not either according to some fixed rule (which would give only a point mass) or uniformly (which would give a uniform distribution over the two variables), but with _arbitrary probabilities_ that are not specified anywhere in the model. This is not reasonable.\\n\\nThat said, I'm now more satisfied that if you simply require the $\\\\gamma$ distributions to be nonatomic, then I have no reason to doubt either theorem.\"}", "{\"title\": \"Response to Rebuttal\", \"comment\": \"Thank you for addressing most of my questions. I maintain my score.\"}", "{\"title\": \"Comment to Reviewer e4WY\", \"comment\": \"On Theorem 5.1, I believe they meant to write the set of all expectations (for $\\\\theta \\\\in \\\\Theta$) in between $\\\\operatorname{ri}(\\\\Delta)$ and $\\\\Delta$? These components are quite important; thank you for pointing them out.\"}", "{\"summary\": \"The paper studies representation properties of \\\"Perturb-Softmax\\\" and \\\"Perturb-Argmax\\\" distributions, generalizations of the Gumbel-Softmax/Gumbel-Max distributions used in a significant line of prior machine learning work on discrete random variables. They show theorems establishing general representation results, and experiments arguing that perturbing with a Gaussian rather than a Gumbel yields improved learning performance.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"The idea of alternatives to the Gumbel trick is interesting. It indeed relates to FTPL (and similarly, I think, to various non-exponential mechanisms used in differential privacy), but in this context to my knowledge the Gumbel version is the only one that has been thoroughly considered.\\n\\nThe experimental results are promising that learning with Gaussian noise may be preferable to Gumbel noise.\", \"weaknesses\": [\"While the setting of this paper is interesting, the presentation is a little bit disjointed:\", \"Your conclusion says that \\\"Our main contribution is a theoretical study of the representation properties of Gumbel-Softmax and Gumbel-Argmax probability models.\\\" I think this is not really true; the basic derivation of Gumbel-Argmax shows that it is obviously equivalent to general softmax parameterizations of discrete distributions, in which case the completeness and minimality properties are well-known. While I don't think either of the two Gumbel-Softmax papers made explicit arguments about completeness, both included a temperature parameter, and the equivalence of Gumbel-Softmax to Gumbel-Argmax in the limit of low temperatures \\u2013\\u00a0explicitly stated as Proposition 1 part (c) of Maddison et al. \\u2013 immediately and obviously implies completeness there. Thus most of these results were implicit in the literature already and the new ones are not _especially_ interesting in the machine learning context.\", \"What is quite interesting, though, is the claim that perturbing with other forms of noise can yield the same kind of completeness (and, less relevantly for learning, minimality) properties. The experiments with Gaussian-Softmax also indicate that other kinds of noise are worth exploring, although it would be nice to have a better understanding of _why_ (see discussion about the \\\"convergence rates\\\" argument below) or of how much improvement can be seen in a wider variety of settings. So, I think the paper should be somewhat reframed as being about the generalized Perturb-*max models, and how other choices than Gumbel can be better.\", \"To this end: while the experiments here are definitely not _nothing_, they also contain no detail other than \\\"here is the final approximation quality,\\\" and they're also not of an especially exciting scale in 2024. Ideally, I'd like to see some better poking at why optimization worked better in a simple case, e.g. one of the fixed discrete target distributions (is the parameter vector $\\\\\\\\theta$ to get a good approximation simpler? if not, is something about the optimization landscape simpler?). It would also be nice to have experiments for a more modern model using this as a component; just take one of the many recent models citing these papers and swap out the Gumbel.\", \"But, before doing that, a vital issue:\", \"Theorem 5.1 as stated here is **plainly false**. In the theorem statement, there is no constraint on the form of random variable $\\\\gamma$. First, they are presumably intended to be iid, but maybe this doesn't matter. So, consider using the random variable $\\\\gamma_i$ which is a point mass at 0. Then the Perturb-Argmax distribution is $\\\\\\\\mathbb{E}_\\\\\\\\gamma[ \\\\\\\\mathrm{arg\\\\\\\\,max}\\\\\\\\; \\\\\\\\theta + \\\\\\\\gamma ] = \\\\\\\\mathrm{arg\\\\\\\\,max}\\\\\\\\; \\\\\\\\theta$; depending on exactly how we define argmax's behaviour in the case of ties, this can _only_ produce one-hot distributions. Since the structure of your proof is that one-hots are in the closure of $\\\\\\\\mathcal P$ and that this closure is convex, presumably you didn't notice this when adapting the proof of Theorem 4.1, since indeed that first property is certainly true. Presumably, then, the claim that the closure of $\\\\\\\\mathcal P$ is convex is false in this case. I don't know where the flaw is, but something must be wrong. In particular, since Theorem 4.1 follows a similar structure, it is particularly concerning that there is some unknown flaw that might also apply to that theorem's proof. While it's not immediately obvious to me what the problem is, I'm not so confident in how to carefully apply Rockafellar's results that you use (which are, I think, stated for $\\\\\\\\mathbb R^d$) to the probability simplex, and so am not confident that it's a mistake in Theorem 5.1 specifically. I cannot see arguing for anything other than rejecting the paper unless this mistake is identified and corrected, and a full correct proof provided during the rebuttal process.\", \"Perhaps relatedly, the mathematical discussion and definition in several aspects of this paper is quite sloppy. Here are a few reasonably important ones:\", \"Line 195 claims that \\\"The fundamental theorem of extreme value statistics asserts the equivalence between the softmax distribution in Equation 4 and the Gumbel-Argmax distribution in Equation 2.\\\" I don't see how this is in any way true. You only cited a 60-page lecture series which does not use this phrase to identify a theorem; presumably, you mean a [version of this theorem](https://en.wikipedia.org/wiki/Fisher%E2%80%93Tippett%E2%80%93Gnedenko_theorem), which states that the maximum of $n$ iid values can have one three asymptotic behaviours as $n \\\\\\\\to \\\\\\\\infty$, one of which is the Gumbel distribution. I don't see at all what this has to do with the distribution of a _soft_ max of a finite number of values. Luckily, though, you don't seem to use this seemingly-incorrect claim again.\", \"Section 3.4 \\\"Convergence rates\\\" compares the upper bounds on concentration of Lipschitz functions of a Gaussian variable and a Gumbel variable, and since the former is faster than the latter, vaguely imply that this means that Gaussian-Softmax is better than Gumbel-Softmax. But the purpose of the remainder of your paper is that you would use the two to identify _the exact same discrete distribution_, and thus estimating $\\\\mathbb E\\\\_{\\\\\\\\gamma \\\\\\\\sim g}[ f(\\\\\\\\theta, x, \\\\\\\\gamma) ]$ by Monte Carlo samples of $\\\\\\\\gamma$ will be _identical_ between equivalent choices of $\\\\\\\\theta$ with different distributions for $\\\\\\\\gamma$. These differences in upper bounds (which are worst-case over functions $\\\\\\\\gamma \\\\\\\\mapsto f(\\\\\\\\theta, x, \\\\\\\\gamma)$) are not relevant here.\", \"And here are a few examples that are easy to fix, but indicative that perhaps you were not as careful in writing as you should have been:\", \"On line 172 you say that the probability density function of the Gumbel is $\\\\\\\\hat g(t) = e^{-e^{-(t+c)}}$ where $c \\\\\\\\approx 0.5772$. This is not true; this is the _cumulative_ density function, specifically of a Gumbel with mean $c$ and scale $1$. While scale $1$ is standard for the Gumbel-max trick, there is absolutely no reason to use mean $c$; it doesn't break anything, but by far the more typical choice would be to use mean $0$.\", \"Theorem 4.1 says \\\"let $\\\\gamma = (\\\\\\\\gamma\\\\_1, \\\\dots, \\\\\\\\gamma\\\\_d)$ be a vector of random variables whose cumulative distribution decays to zero as $\\\\\\\\gamma$ approaches $\\\\\\\\pm \\\\infty$.\\\" First, presumably you mean that the $\\\\\\\\gamma_i$ are iid, and it is the distribution of each of those that has the appropriate decay property. Secondly: it is trivially true of any real-valued random variable that its cumulative distribution function decays to zero as you approach $-\\\\\\\\infty$, and to _one_ as you approach $\\\\\\\\infty$. Perhaps you meant that the measure of sets $(-\\\\\\\\infty, -t)$ and $(t, \\\\\\\\infty)$ should approach zero as $t \\\\\\\\to \\\\\\\\infty$? But this is again automatically true of any real-valued variable. I actually have no idea what you mean here.\", \"Line 719: \\\"A multivariate function is differentiable if its directional derivative is the same in every direction $v \\\\\\\\in \\\\\\\\mathbb R^d$, namely $\\\\\\\\nabla f(\\\\\\\\theta) = \\\\\\\\nabla_v f(\\\\\\\\theta)$ for every $v \\\\\\\\in \\\\\\\\mathbb R^d$.\\\" This is not at all true, as should be clear from the fact that your equation is equating a vector to a scalar. Rather, the directional derivatives should be _consistent_, i.e. $\\\\\\\\nabla f(\\\\\\\\theta) \\\\cdot v = \\\\\\\\nabla_v f(\\\\\\\\theta)$ for all $v$.\"], \"minor_points_of_terminology\": [\"Your definition of identifiability is different than the one I've always heard before, and is also used e.g. [by Wikipedia](https://en.wikipedia.org/wiki/Identifiability), which is exactly what you call \\\"minimal.\\\" I don't see much reason in giving a special name to the bijection case here, and particularly in giving one that is so widely used to mean something slightly different; I would strongly encourage you to change the terminology to \\\"complete\\\" (or perhaps \\\"universal\\\" would be more common in learning contexts) when the map from parameters to probability distributions is (almost) a surjection, and \\\"identifiable\\\" when it is an injection.\", \"Typically one would put `\\\\appendix` at the start of the appendix, which would name the appendix A rather than 8.\", \"Line 713: \\\"Convexity is a one-dimensional property.\\\" This is a strange statement; while you can define it in one dimension and extend it to other cases as you do here, there are also (many) direct definitions of convexity that work for any input vector space.\", \"Trivial typos, etc that I happened to notice:\", \"line 160: you wrote $softmax$ instead of $\\\\\\\\mathrm{softmax}$\", \"\\\"subsetset\\\" on line 209\", \"Footnote 3: \\\"dominant convergence theorem\\\" should be \\\"dominated convergence theorem\\\"\"], \"questions\": [\"How can Theorem 5.1 be corrected? Does the mistake also apply to Theorem 4.1?\", \"Do you have any insights as to why Gaussian-Softmax seems to be more learnable in these settings than Gumbel-Softmax? Can those insights be generalized to help guide to other noise distributions that might be even better?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"After reading other reviewers' comments and the author's feedback, I decided to maintain my score.\"}", "{\"comment\": \"We are happy you found our work well-written and theoretically grounded.\\n\\nThe experiments empirically demonstrate the theoretical benefits of learning with Gaussian perturbation models rather than Gumbel perturbation models. We find it hard to address the comment that the work requires more empirical evidence, as it is general and not substantial. \\n\\nOur framework easily extends to include the temperature scaling as in the Gumbel-Softmax models with temperature, as outlined in Equation 11. Importantly, our theoretical investigation allows us to identify the conditions under which a set of parameters $\\\\Theta$ itself is complete. In the previous work (Maddison et al., 2017; Jang et al., 2017), completeness depends on the temperature hyper-parameter tuning.\"}", "{\"metareview\": \"This work investigates the representation properties of the Gumbel-Softmax and Gumbel-Argmax probability distributions, which are used for learning discrete tokens and structures in generative and discriminative models, respectively. The study identifies the conditions under which these distributions are complete (able to represent any probability distribution) and minimal (able to represent a distribution uniquely), using convexity and differentiability. The analysis extends to general probability models like Perturb-Softmax and Perturb-Argmax, concluding that certain parameter sets allow for complete and minimal representations, with experimental results validating the faster convergence of Gaussian-Softmax compared to Gumbel-Softmax.\\n\\nThe problem addressed by the paper is clearly important and practical. Although the paper is meant to be theoretical, the experimental results are really minimal and lack some important comparisons as the reviewers have pointed out. There are also a few concerns raised by Reviewer e4WY that remain not fully satisfactorily addressed in the rebuttal. So the paper needs one more round of revision.\", \"additional_comments_on_reviewer_discussion\": \"The rebuttal has been noted by the reviewers and have been taken into account by the AC in the recommendation of acceptance/rejection.\"}", "{\"comment\": \"Thank you for your positive review.\\nPlease note that the condition on function $h_i$ does appear in Theorem 4.1's proof in Appendix 8.3. The condition that the function $h_i(\\\\theta) = \\\\theta_i - max_{j \\\\ne i} \\\\theta_j$ is an unbounded continuous function over $\\\\Theta$ allows establishing that zero-one distributions are limit points of probabilities in the closure $\\\\cal P$. We will make this proof idea explicit in the main text. Further, thank you for noticing that the condition on the function $h_i(\\\\theta)$ is explicit in the proof of the completeness of Perturb-Argmax, but is missing from the conditions of Theorem 5.1. We will add this condition explicitly.\\n\\nIndeed, the Gumbel-Softmax is instrumental in the variational auto-encoder model, as mentioned in lines 120-121.\\n\\nTheorem 3.1 includes the sub-Gaussian parameter in the assumption that $\\\\|\\\\nabla f(\\\\gamma) \\\\|^2 \\\\le \\\\sigma^2$. \\nThe \\\"whose cumulative distribution decays to zero\\\" is an inaccurate phrasing, we meant that its density function decays to zero as $\\\\gamma$ approaches $\\\\pm \\\\infty$.\\nThe convergence rate referenced in line 240 is indeed that of the Gumbel distribution, though we meant to convey that there are other distributions in the exponential family with an exponential convergence rate as the Gumbel.\"}", "{\"comment\": \"I\\u2019m not sure what you mean by that. I read the theorem statement as saying that for any choice of $\\\\\\\\gamma$ distribution, the set of distributions achievable by any $\\\\\\\\theta$ (these expectations) is indeed between $\\\\\\\\mathrm{ri}(\\\\\\\\Delta)$ and $\\\\\\\\Delta$. But when $\\\\\\\\gamma$ is a point mass at zero, the set does not contain $\\\\mathrm{ri}(Delta)$; it\\u2019s only a subset of the boundary of $\\\\Delta$, nothing in the relative interior at all!\\n\\nIt might be that the proof works as long as $\\\\\\\\gamma$ is continuous or something like that, but I don\\u2019t know; we\\u2019ll see if the authors do when they\\u2019re ready to respond. :)\"}" ] }

[ "Kamel Alrashedy", "Pradyumna Tambwekar", "Zulfiqar Haider Zaidi", "Megan Langwasser", "Wei Xu", "Matthew Gombolay" ]

[ "Code Generation", "Self-refinement" ]

{ "note_id": [ "zXRX2FqBqH", "yin9reVOX2", "sxz8YSD3Eu", "rUMNFYjcGW", "qVXuKs4POA", "pqDx5L17AH", "p0mdXXutYF", "ituoDLvuUi", "dhe9mMCJ6e", "bAxN1TiW06", "VIdS4fp2xD", "V4FKSpaOm3", "UDGan9mzgL", "S8wKqopZmF", "S3Rw2QZbZ0", "Os7DzdAeDz", "OGrLioN0g8", "GJXrXVOh18", "G42X8B13Pi", "E1JXo0h5u4", "7HeWZZPbgn", "6KnbOMM6W6", "2I2fkD1mfH", "24uGknHcPm" ], "note_type": [ "official_comment", "meta_review", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "decision", "official_comment", "official_comment" ], "note_created": [ 1732217330819, 1734499953035, 1732520069954, 1730699885061, 1732558858140, 1732717067810, 1733092241141, 1732217175401, 1730824864656, 1732216976888, 1730653702908, 1732559463623, 1733115280451, 1732217003346, 1732560348804, 1733092181593, 1730690684377, 1733091877229, 1732217500899, 1732559600013, 1732217443340, 1737524235557, 1732571575544, 1732580655523 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission13116/Authors" ], [ "ICLR.cc/2025/Conference/Submission13116/Area_Chair_3AfN" ], [ "ICLR.cc/2025/Conference/Submission13116/Reviewer_9oaM" ], [ "ICLR.cc/2025/Conference/Submission13116/Reviewer_Ghys" ], [ "ICLR.cc/2025/Conference/Submission13116/Authors" ], [ "ICLR.cc/2025/Conference/Submission13116/Reviewer_jRaW" ], [ "ICLR.cc/2025/Conference/Submission13116/Authors" ], [ "ICLR.cc/2025/Conference/Submission13116/Authors" ], [ "ICLR.cc/2025/Conference/Submission13116/Reviewer_jRaW" ], [ "ICLR.cc/2025/Conference/Submission13116/Authors" ], [ "ICLR.cc/2025/Conference/Submission13116/Reviewer_9oaM" ], [ "ICLR.cc/2025/Conference/Submission13116/Authors" ], [ "ICLR.cc/2025/Conference/Submission13116/Authors" ], [ "ICLR.cc/2025/Conference/Submission13116/Authors" ], [ "ICLR.cc/2025/Conference/Submission13116/Authors" ], [ "ICLR.cc/2025/Conference/Submission13116/Authors" ], [ "ICLR.cc/2025/Conference/Submission13116/Reviewer_8EJ1" ], [ "ICLR.cc/2025/Conference/Submission13116/Authors" ], [ "ICLR.cc/2025/Conference/Submission13116/Authors" ], [ "ICLR.cc/2025/Conference/Submission13116/Authors" ], [ "ICLR.cc/2025/Conference/Submission13116/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission13116/Reviewer_8EJ1" ], [ "ICLR.cc/2025/Conference/Submission13116/Reviewer_Ghys" ] ], "structured_content_str": [ "{\"title\": \"Response to Reviewer 8EJ1\", \"comment\": \"We thank this reviewer for their time and valuable feedback to our submission.\\n\\n## \\u201cSuccess-Rate\\u201d is misleading \\nWe thank the reviewer for highlighting the potential misinterpretation of the term \\u201csuccess-rate.\\u201d We will change \\u201csuccess-rate\\u201d to \\u201ccompile-rate\\u201d as per the reviewer\\u2019s suggestion. \\n\\n## 3D-Premise execution step \\nYes, 3D-Premise goes through the code-execution step in the same manner as CADCodeVerify and the Geometric Solver baseline. The baselines only differ in approach to ours in the Code-Refinement Step (Section 3.3). Both 3D-Premise and the Geometric solver approach also go through N-steps of code-repair to engender a consistent comparison with CADCodeVerify. \\n\\n## Refine-1 causes 3D-Premise to worsen compile-rate \\n\\nWe hypothesize that 3D-Premise leads to a reduction in the compile-rate, because off-the-shelf LLMs struggle to infer the changes they need to make to CAD scripting code to edit the 3D object, based on the image in isolation. In contrast, CADCodeVerify includes both generated images of the object as well as the associated text-feedback, computed via self-verification. This text-feedback likely provides \\u201cinstructions\\u201d to enable the VLM to better interpret the changes it needs to make to the code based on the image, and better aligns with the types of feedback the model has been trained on. Recent work on visual programming adopts a similar approach wherein they extract textual interpretations from visual features rather than utilizing the images in isolation [1] \\n\\n[1] - Gao, Minghe, et al. \\\"De-fine: De composing and Re fin ing Visual Programs with Auto-Feedback.\\\" Proceedings of the 32nd ACM International Conference on Multimedia. 2024. \\n\\n## Geometric Solver Feedback \\n\\nYes, the output from the geometric solver is passed back into the model as feedback to refine the initial generated code. We verbalize the raw outputs from the Geometric solver, prior to passing it into the LLM as feedback (Section B.2 in the appendix). Both the Geometric solver method and 3D-Premise utilize two feedback steps, similar to our approach. \\n\\n## Generated, Refine-1, Refine-2 definitions \\n\\nPlease pardon any confusion regarding the terminology used in the experiments section. We have added these defintions to the first paragraph of Section 6 -- \\u201cGenerated refers to the object generated by the LLM after the Code-Execution Step. Refine-1 refers to the object after the first step of refinement, and Refine-2 refers to the object after the second step of refinement. \\u201c \\n\\n## Additional Questions/Typos \\n\\n- The number in the introduction is correct (5.5%). We will update the abstract to correct this typo \\n- Mistake in the equation on line 177 has been corrected. \\n- We will remove \\u201cupper-limit/upper-bound\\u201d while referencing the Geometric solver baseline. \\n- The three experiments discussed between lines 305-310 reference the 2x3 configuration of code generation approaches across the axes of model-type (GPT-4, Gemini, CodeLlama) and prompt-type (Zero-shot/Few-shot).\"}", "{\"metareview\": \"In this paper, the authors introduce CADCodeVerify, a novel approach that uses VLMs for iterative verification and refinement of CAD code generation. To evaluate CADCodeVerify, the authors also propose CADPrompt, a benchmark for CAD code generation consisting of 200 natural language prompts paired with expert-annotated scripting code for 3D objects. Under this benchmark, the experiments include comparisons with existing methods as well as with human experts, demonstrating that the proposed method is superior.\\n\\nHowever, there are a limited number of technical contributions. The main contribution seems to be the novel mechanism of using off-the-shelf LLMs and VLMs to automatically generate high-quality CAD code. Though the experiments show superior results, the small scale of the experiments makes it difficult to conclude that the pipeline is universally applicable. In addition, the method's reliance on natural language input may not account for precise numerical specifications, which is another limitation. Therefore, I recommend accepting the paper but as a poster.\", \"additional_comments_on_reviewer_discussion\": \"The reviewers primarily expressed concerns about the following points:\\n1. The initial submission did not provide sufficient comparisons with existing methods in the field.\\n2. The distinctions between the proposed method and existing approaches, particularly 3D-Premise, were not clearly articulated.\\n3. Important details of the proposed method were not adequately explained.\\n4. The nature of the natural language input used by the method was ambiguous.\\n5. The size of the dataset is small.\\n6. Some design choices in the proposed method appeared to be empirical without sufficient justification (e.g., reasons why the results are superior to those of other methods, the number of refinements).\\n\\nDuring the rebuttal phase, concerns 1, 2, and 3 were satisfactorily addressed by the authors. However, points 4, 5, and 6 remain as limitations of the work. Despite these issues, the reviewers agree that the results presented are positive and show promise.\"}", "{\"comment\": \"Thanks for the clarifications. I will keep my score\"}", "{\"summary\": \"The paper introduces CADCodeVerify, a approach to iteratively verify and improve 3D objects generated from CAD code using Vision-Language Models (VLMs). The method involves generating CAD scripting code from natural language prompts, executing the code to render a 3D object, and refining the code based on visual feedback through a question generation and answering process, to correct deviations from the initial specifications. The approach is evaluated using CADPrompt, a benchmark dataset of 200 3D objects with corresponding natural language prompts and expert-annotated code. The approach is evaluated on GPT-4, Gemini 1.5 Pro and CodeLLama models.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": [\"The paper comes with a new benchmark suite (CADPrompt) which is a well-curated, crowd-sourced benchmark with annotations and quality checks, which is a valuable resource for assessing CAD code generation and refinement methods.\", \"CADCodeVerify uses an interesting novel idea to eliminate the need for human involvement by generating validation questions and answering them using VLMs.\", \"The geometric solver-based baseline is very interesting and gives an upper estimate of the self-refinement process; it would be interesting to explore if this solver could also be used as a metric for evaluation.\"], \"weaknesses\": [\"The fundamental differences between CADCodeVerify and 3D-Premise are unclear. For example, it\\u2019s not specified whether 3D-Premise uses execution-error-based repair. Also, both approaches seem to use totally different prompts, so it is not clear if it is just a matter of better prompting or something fundamental (such as the question-answer based method)\", \"The paper would be stronger if the approach was also evaluated on the 3D premise dataset\", \"Some other details are missing (see below questions) regarding ambiguity in NL input and chain-of-thought based extension to 3D-Premise.\"], \"questions\": \"1. Would providing images of the generated object from multiple viewpoints improve 3D-Premise\\u2019s performance? What about using a chain of thought prompting (for e.g. explicitly ask the model to reason whether each criterion in the initial task are satisfied)? This will be similar to CADCodeVerify, but doing 1 VLM call instead of doing 3 VLM calls in the verify step to reduce the inference time.\\n\\n2. In Figure 3, does 3D-Premise also repair code based on compile error messages?\\n\\n3. How do you handle ambiguity in inputs in ground truth solutions? For example, if the NL input does not specify a particular size of hole, there could be many ground truth solutions which might impact the cloud-distance based metrics. \\n\\n4. Since Point Cloud distance and Hausdorff distance are noisy metrics, how does the generated CAD program compare on the geometric metrics obtained using the geometric solver?\", \"minor\": [\"Consider renaming success rates to compile rates or something similar to convey that is only compilation success rate and not overall task success rate.\", \"For figures 5,6 (running examples), it will be useful to also see the repair steps.\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to Reviewer jRaW\", \"comment\": \"Dear jRaW,\\n\\nWe hope you have had a chance to review our rebuttal. As today is the last day of the discussion period, we wanted to check-in to see if there was anything you wanted to discuss with us regarding our paper. Please let us know if there are any outstanding concerns that you would like us to address or discuss towards improving your assessment of our paper.\"}", "{\"comment\": \"Dear authors,\\n\\nThank you for your rebuttal and the clarifications. I am satisfied with the answers about compilation difficulty and semantic complexity metrics. Generally, I remain positive that the paper has value. I have follow-up comments/questions on the other answers. Time might lack to address them but you might want to consider them in potential revisions of your paper.\\n\\n**Object/mesh complexity**\\n\\nOnce again, higher polygon counts do not necessarily indicate greater complexity, nor does lower complexity guarantee fewer polygons. For instance, spherical objects often have higher polygon counts but are not structurally complex\\u2026 Anyway, are you suggesting that the cadquery commands used in the generated code aim to avoid unnecessary face subdivisions or excessive polygons generation while accurately capturing the geometry of the objects?\\n\\n**Key advances over 3D-premise**\\n\\n3D-Premise does not require expert-defined questions and refines fully automatically, using the image input along with the (simplified) instruction \\u201cDoes the object correspond to the image\\u201d for feedback. It also utilizes Blender3D for code execution but does not include a code repair stage. Are you referring to a \\u201c...code repair stage to reduce syntax errors\\u201d?\\n\\nStill, the positioning could be more clear and explicit also because it is a baseline that you implemented yourself. \\n\\n**Performance Insights and Limitations**\\n\\nThank you for this response. While \\\"excels\\\" might be too strong a term, it appears that our proposed question-based refinement has enhanced the LLMs' performance in these tasks. However, I recommend exploring a more collaborative human/machine approach in future work to facilitate the application of such support in real-world and industrial contexts.\\n\\n**Methodology Justifications**\\n\\nCould you elaborate on the rationale behind these observations, specifically the notion of diminishing returns after two iterations and the importance of bounding questions to maintain efficiency? It would be helpful to understand the underlying reasoning or empirical evidence supporting these approaches, as well as how they contribute to optimizing the overall process.\"}", "{\"title\": \"Response to Reviewer jRaW\", \"comment\": \"Please feel free to share any additional feedback or questions\\u2014we are happy to discuss further and work toward improving your evaluation of our paper.\"}", "{\"title\": \"Response to Reviewer Ghys\", \"comment\": \"We thank this reviewer for their time and thorough review of our submission. We hope to address your concerns as follows:\\n\\n## Does 3D-Premise also repair code based on compile error messages \\n\\nYes, our implementation of 3D-Premise also leverages the code execution step presented in Section 3.2. For both the Geometric Solver and 3D-Premise baselines, all components of the CAD Code Generation process are kept consistent besides the method of generating feedback for code-refinement. \\n\\n## Handling ambiguity of language-descriptions \\n\\nIn our work, we adopted extensive review procedures to ensure that the descriptions are consistent with the target object as described in Section 4.1. In paragraph two of our limitations section, we note that the same object can be described in different ways depending on the specifier, and conversely, multiple object configurations may satisfy a single NL description. However, such incongruities are inherent to any approach combining freeform language with parametric 3D designs. This problem could be alleviated by providing K possible outputs for every description to compute a top-k metric, however, we leave that exploration to future work. \\n\\n## Point-Cloud distance and Hausdorff distance are noisy \\n\\nTo account for potential noise within the PC or Hausdorff distance measurement, we also included a third metric which evaluates overlap in the objects in the 3D-space (i.e., ntersection over ground truth (IoGT)) instead of computing a point-to-point measurement. All three of these metrics are well established measurements utilized in prior work to evaluate 3D Generations [1,2,3]. By showing consistent results across these three metrics, we believe that we have provided a reliable measurement of the competency of various VLMs and CADCodeVerify for the CADPrompt benchmark. If this reviewer believes that our paper would benefit from including the additional geometric-solver measurement proposed, we will compute that for the camera-ready version of the paper. \\n\\n \\n\\n[1] - Yuan, Zeqing, et al. \\\"3D-PreMise: Can Large Language Models Generate 3D Shapes with Sharp Features and Parametric Control?.\\\" arXiv preprint arXiv:2401.06437 (2024). \\n\\n[2] - Sun, Yongbin, et al. \\\"Pointgrow: Autoregressively learned point cloud generation with self-attention.\\\" Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 2020. \\n\\n[3] - Vahdat, Arash, et al. \\\"Lion: Latent point diffusion models for 3d shape generation.\\\" Advances in Neural Information Processing Systems 35 (2022): 10021-10039. \\n\\n## Renaming Success-Rates \\n\\nWe thank the reviewer for their suggestion, and we will replace \\u201cSuccess-Rate\\u201d with \\u201cCompile-rate\\u201d throughout the paper.\"}", "{\"summary\": \"The paper entitled \\u201cGENERATING CAD CODE WITH VISION-LANGUAGE MODELS FOR 3D DESIGNS\\u201d proposes an approach to generate and verify CAD 3D objects from natural language. The proposed approach introduces a verification step called CADCodeVerify. CADCodeVerify use Vison Language Machine (VLM) to verify CAD objects against generated visual questions. The hypothesis is that the question serves as visual properties the object should satisfy to meet the structural specification. The key contribution of this approach is the use of a refinement step. From the language prompt that describes the object in natural language, a first CAD description code (CADQuery) is generated. The visual questions are also generated from this language prompt input. The object (CADQuery) is verified against visual properties questions using a VLM that generates feedback for each questions. This feedback is used to refined the CADQuery code.\\n\\nTo summarise, the contributions of this work is two fold: 1) an automated CAD generation approach that does not necessitate humain interaction; 2) a dataset called CADPrompt that comprises 200 CAD objects. Expermental results shows when implemented using GPT-4 LLM, we observe improvements in object generation accuracy and an increased success rate compared to other LLMs.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"3\", \"strengths\": [\"Novelty and Originality: The integration of question/answering based VLM to refine the object quality.\", \"Evaluation: The increase of success rate of compilable outputs at the end of the refinement shows that the method can improve the generation of CAD code from language prompt.\", \"Soundness: The technical approach is sound and could be applied to a lot of CAD or CAV tasks. The approach is a specification refinement method using generative AI. From an initial specification, an initial object plus query are generated. Then, feedback from query satisfaction are used to improve the object. A dataset is provided and metrics are determined to rank the quality of the generated object. This has applications beyond CAD specifically.\"], \"weaknesses\": [\"Dataset Scope: CADPrompt could have been better introduce, showing the most complex objects in terms of complexity and difficulty metrics.\", \"Complexity Metric for Objects: This paper measures object complexity by counting vertices and faces. However, using metrics like bounding boxes or decomposed bounding volumes could better reflect structural complexity. Vertices mainly define shape details, not true complexity\\u2014a simple shape like a cube can have many polygons, while complex shapes like an aircraft wing might need fewer. In physics engines, object complexity is often defined by the volume structure the object requires. Lower polygons can shape more complex structures.\", \"Difficulty Metric for Objects: The current method of measuring complexity through word and line count for the natural language descriptions and Python code may not fully capture their complexity. A semantic-based complexity metric, where each instruction has a complexity value, would better reflect how nuanced or contextually challenging the description and code are, which could impact model performance.\", \"Clarity and Readability: Certain sections are overly technical and hard to understand without referencing the appendix. Simplifying these sections would improve accessibility for readers. Terminological inconsistencies between LLM used for language to code and code with feedback to code and VLM for image and questions to answers also introduce potential confusion, which could be resolved for clarity. Certain sections are overly technical and hard to understand without referencing the appendix. Simplifying these sections could enhance accessibility for readers. For example, consider moving equations 1-7, which aren't essential for understanding the main contributions, to the appendix. Conversely, evaluation metric equations (8-10), such as those for point cloud distances, Hausdorff distance, and IoGT, should remain in the main text. Additionally, bringing valuable visuals currently in the appendix (Figures 5, 11, 12, and 14) into the main text would ease the understanding and improve work illustration. The paper would also benefit from a stronger focus on scientific insights and hypothesis testing over too much detailled descriptions. For instance, providing context around why a fixed number of five verification questions is used, without adapting to object complexity, would clarify the rationale. Similarly, explaining why only two refinement iterations are applied, regardless of complexity, would better illustrate the scientific intuitions driving these choices. Furthermore, details on how AI context like 7, 9, and 10, have been engineered such as crafted based on expert knowledge or through authors tuning would be valuable. Shifting focus to the \\\"why\\\" rather than the \\\"what\\\" would help readers grasp the paper challenges.\", \"Positioning: the paper\\u2019s positioning within the state of the art could be made clearer, with more explicit distinctions from existing work. The paper integrates the related work 3D-premise in the experiments, but didn't explicitly state the approach difference. The feedback is based on question answering rather than the initial description of the object.\"], \"questions\": [\"Q1: I appreciate the framework\\u2019s contribution, but could you provide a bit more on how this specifically differs from or builds upon prior work (3D-Premise)? Is it an improvement or an entire new approach?\", \"Q2: Can you tell us more about the quality and diversity of the CADPrompt dataset? I\\u2019m curious about how it\\u2019s tailored to the goals of your approach. A bit more on why it\\u2019s a good benchmark\\u2014perhaps a breakdown of object types, complexity levels, or any specific challenges it presents?\", \"Q3: The experimental results look interesting, but it would be good to understand more in details where the approach performs well and where it might have limitations. Do you have any insight to share about this?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to Reviewer jRaW Part(1/2)\", \"comment\": \"We thank this reviewer for their measured and insightful feedback. We hope our response will sufficiently address the concerns raised.\\n\\n## Compilation Difficulty Metric \\nIt appears that the reviewer may have misunderstood the difficulty metric employed in our paper. Our \\u201cdifficulty\\u201d metric defined in Section 4.3 is not computed as the \\u201cword and line count for the natural language descriptions and Python code\\u201d as stated by this reviewer. Instead, our difficulty metric is \\u201cmeasure of how difficult it is for a set of three language models (i.e., GPT-4, Gemini, and CodeLlama) to generate code for a given 3D object across two prompting methods (i.e., zero- and few-shot prompting), for a total of six attempts to generate compilable code\\u201d as stated in Section 4.3. This metric is one of three metrics we use for analyzing the models\\u2019 performances. \\n\\n## Complexity metric for Objects (Mesh Complexity) \\nIn 3D-modeling software such as Blender or Unity, meshes with a higher polygon count allow for rendering objects in higher levels of detail, implying greater expressivity. While there may be objects with a lower number of faces or vertices which are semantically more complex or more difficult to manufacture, measuring the number of faces and vertices provides insight into the level of detail expressed in the 3D-object. \\n\\n## Semantic Complexity metric \\n\\nTo better understand the challenge proposed by the objects in CADPrompt, we conducted a semantic evaluation as suggested by the reviewer. We recruited an independent mechanical engineer to semantically rate the complexity of 200 examples in our dataset. The annotator utilized a four-point scale as follows: \\n\\n- (Simple) - The object is extremely basic, with few features. It may \\nconsist of one geometric shape. \\n\\n- (Moderately Complex) - The object has a moderate amount of detail, \\nwith a few distinct features or components \\n\\n- (Complex) - The object is complex, with many interconnected parts, \\nfine details, or intricate shapes. \\n\\n- (Very Complex) - The object is highly intricate, with many \\ncomponents, detailed textures, or complex shapes. It may have a large \\nnumber of fine details, interlocking parts, or unique geometric features \\n \\nThe breakdown of our dataset according to this independent reviewer\\u2019s evaluation was as follows: Simple - 17, Moderate Complexity - 39, Complex - 87, and Very Complex - 57. When comparing the semantic complexity with our quantitative metrics, we observe that all our measures of \\u201ccomplexity\\u201d are generally aligned with this semantic complexity measurement (Please refer to Table 1 in the updated version.). \\n\\n## Q1 \\u2013 Differences from 3D-Premise \\n\\n3D-Premise presented the first exploration of leveraging the vision-capabilities of VLMs for CAD Code Refinement. CADCodeVerify offers three key advancements over 3D-Premise. \\n\\n1. Our approach adopts a \\u201cSelf-Verification\\u201d method wherein the VLM is prompted to generate its own set of validation questions rather than answering a fixed set of questions decided by an expert, thereby removing the human-in-the-loop expertise. \\n\\n2. 3D-Premise does not include the code execution stage (Section 3.2), which is crucial for reducing the number of syntax errors in the code. \\n\\n3. 3D-Premise only offers qualitative insights regarding the performance of their proposed method to integrate visual feedback into CAD code refinement. In our approach, we conduct a comprehensive and systematic evaluation of CADCodeVerify's capabilities using our CADPrompt benchmark, establishing a foundation for standardized research on CAD code generation. \\n\\n## Q3 - More details regarding where the approach performs well and its limitations. \\nWe conducted human-in-the-loop experiments to compare its performance with CADCodeVerify. Human feedback was used to provide precise instructions on the exact changes needed for the 3D object. The results demonstrate a slight improvement compared to CADCodeVerify in terms of Point Cloud distance and Hausdorff distance, with performance metrics improving from 0.137 and 0.445 to 0.120 and 0.397, respectively (See Table 4). These findings suggest that CADCodeVerify delivers feedback that closely resembles gold-standard feedback. \\n\\nSome 3D objects are highly complex, making it challenging for LLMs to generate them. Additionally, the feedback provided by both CADCodeVerify and human reviewers is often insufficient to refine these objects due to their complexity. Notably, CADCodeVerify demonstrates a significant improvement in correcting 3D objects with structural errors (see Figure 8). In future work, we plan to fine-tune LLMs for domain-specific tasks in CAD code generation, aiming to enhance the quality of CAD outputs significantly.\"}", "{\"summary\": \"This paper proposes a technique called CADCodeVerify that uses feedback generated from VLMs to refine CAD code generated by LLMs in an iterative manner. This is a novel technique proposed to address a reasonably difficult and seemingly important problem. The paper evaluates this technique against an existing baseline called 3D-Premise and also with a baseline using feedback from a geometric solver. CADCodeVerify shows better performance against both baselines.\", \"soundness\": \"4\", \"presentation\": \"4\", \"contribution\": \"3\", \"strengths\": \"This paper is written really well. As someone who doesn't work much with CAD code, it was immediately clear what the limitations were and how they intended to address them. The technique itself is novel for the application and is well described with appropriate examples. I also appreciate the effort taken for creating the benchmark tasks over which the systems were evaluated.\", \"weaknesses\": \"One main weakness I can see is the lack of baselines. However, this is not a major issue to me since I don't know of other systems that target this specific problem. However, I can think of other techniques like chain of thought prompting, or techniques using some form of reasoning, against which we can compare CADCodeVerify. Right now, the only true baseline seems to be 3D-Premise, and another baseline that uses an alternate form of feedback generation, though I could be mistaken here.\", \"questions\": \"1. In Fig 1, within the CADCodeVerify block, I see step 3.c generating feedback which then produces a new object. My question is: what exactly is the input to the VLLM? Is it the feedback, that then generates the code that compiles to produce the new object, or is it the new object itself? I suspect it is the former, but the figure indicates it to be the latter.\\n2. In the evaluation, do you consider naive or basic prompting? For example, just a zero-shot or few-shot prompt to an LLM that produces code? Is that what the \\\"Generated\\\" field in Table 2 is?\\n3. Are there any insights as to why CADCodeVerify performs better than baselines? For instance, what is it in the proposed technique that allows it to outperform 3D-premise? And why is VLLM feedback more useful than geometric feedback, even when the latter needs the presence of ground truths? One would imagine a technique needing GT labels for providing feedback would be more accurate or equivalent to a technique operating without the labels.\\n4. Maybe I missed this in the text, but in Fig 3, what is Refine 1 and Refine 2?\\n5. Are there any limitations as to the complexity and difficulty of the benchmark tasks? I understand they vary in complexity and difficulty, but at the same time CADCodeVerify does quite well, even reaching around 90% accuracy in the hardest case. Is there a class of problems this benchmark is not testing, or are the hard problems simply not hard enough?\\n6. How representative is the benchmark of the objects actually being drawn in the industry today? One major challenge with code generation has always been that it has never been representative of IRL code. Do we see the same challenges here?\\n7. What dataset was 3D-Premise originally evaluated on, and why have you not evaluated CADCodeVerify over that? What is different about your dataset that sets it apart from the one 3D-Premise used?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to Reviewer 8EJ1\", \"comment\": \"Dear Reviewer 8EJ1,\\n\\nWe hope you\\u2019ve had an opportunity to review our responses to your valuable feedback. As the discussion period concludes today, we wanted to follow up to see if there are any remaining concerns or points of clarification that you would like us to address. Please don\\u2019t hesitate to share any additional feedback or questions\\u2014we are happy to discuss further and work towards improving your evaluation of our work.\\n\\nThank you again for your time and effort in reviewing our paper.\"}", "{\"title\": \"Response to Reviewer Ghys\", \"comment\": \"Dear Ghys,\\n\\nThank you for pointing out that some of your concerns were not fully addressed in our initial response. We sincerely apologize for this oversight and greatly appreciate your detailed feedback. Below, we have provided thorough answers to the remaining questions you raised. \\n\\n## 3D-Premise with Multiple Viewpoints and Chain-of-Thought Prompting: \\n\\nWe conducted additional experiments on 3D-Premise, incorporating multiple viewpoints of images during the refinement process. Specifically, we included four images taken from different angles (0\\u00b0, 90\\u00b0, 180\\u00b0, and 270\\u00b0). Furthermore, we performed experiments using chain-of-thought prompting techniques for 3D-Premise. Additionally, we combined both multiple viewpoints and chain-of-thought prompting in our experiments. \\n\\nThese experiments were conducted using GPT-4 in a few-shot setting. As demonstrated in the table below, our approach, CADCodeVerify, continues to outperform the new baselines. \\n\\n| **Feedback Mechanism** | **IoGT \\u2191** | **Point Cloud dist. \\u2193** | **Hausdorff dist. \\u2193** | **Compile Rate \\u2191** |\\n|--------------------------------------|---------------------|-------------------------|--------------------------|---------------------|\\n| **Generated** | 0.939 (0.030) | 0.155 (0.140) | 0.494 (0.368) | 96.0% |\\n| **3D-Premise** | 0.942 (0.033) | 0.137 (0.155) | 0.446 (0.396) | 91.0% |\\n| **3D-Premise - Multiple viewpoints** | 0.941 (0.035) | 0.132 (0.147) | 0.437 (0.395) | 91.0% |\\n| **3D-Premise - CoT** | 0.941 (0.031) | 0.131 (0.142) | 0.432 (0.409) | 92.0% |\\n| **3D-Premise - CoT & Multiple viewpoints** | 0.941 (0.031) | 0.150 (0.162) | 0.477 (0.367) | 90.0% |\\n| **CADCodeVerify (Ours)** | **0.944 (0.028)** | **0.127 (0.135)** | **0.419 (0.356)** | **96.5%** |\\n| **Geometric solver*** | **0.944 (0.031)** | **0.103 (0.152)** | **0.399 (0.433)** | 95.5% |\\n\\n\\n## Ambiguity in inputs \\nWe agree that ambiguous intents (e.g., phrases like \\u201clarge rectangle\\u201d or \\u201cslightly smaller\\u201d) can lead the model to generate solutions that align with the intent but differ from the ground truth parameters, such as height, width, dimensions, or volume. This could limit the utility of evaluation metrics for comparing methods. \\n\\nTo address this limitation, we propose a human evaluation protocol where outputs are manually rated on a scale from 0 to 100, focusing on their adherence to the intent rather than strict correspondence to the ground truth. This approach provides a more nuanced assessment of model performance and its practical applicability. \\n\\nAs a proof of concept, we conducted a human evaluation on one experiment \\u2014specifically, the GPT-4 few-shot setting (200 objects), which demonstrates the highest performance in Table 2 of the paper. Due to time constraints during the rebuttal period, one of the authors conducted this evaluation. Future work includes human evaluations, with 30 participants each assessing 15 generated objects, the results of this experiment will be incorporated into a camera-ready version\", \"the_evaluation_methodology_was_as_follows\": \"1. The evaluator read the natural language description of the 3D object (input). \\n\\n2. For each refinement approach\\u20143D-Premise, Geometric Solver, and CADCodeVerify\\u2014the evaluator assigned a score from 0 to 100 based on the following criteria: \\n\\n- 0: The refined code did not compile or failed to render a 3D object. \\n\\n- 10: There was a logic error, resulting in implausible configurations of 3D objects that did not resemble real-world contexts. \\n\\n- 10\\u201350: The refined 3D object differed from the input description. \\n\\n- 50\\u201375: The refined 3D object was slightly different from the input description but exhibited some similarities. \\n\\n- 75\\u201390: The refined 3D object was mostly similar to the input description, with minor differences. \\n\\n- 90\\u2013100: The refined 3D object matched the input description exactly. \\n\\n \\n\\nThe table below summarizes the human evaluation results for each refinement approach, reporting the mean, median, standard deviation (SD), minimum, and maximum scores. \\n\\nFeedback Mechanism | Mean | Median | SD | Minimum | Maximum\\n-------------------|--------|--------|-------|---------|---------\\n**3D-Premise** | 68.95 | 90.00 | 35.79 | 0.00 | 100.00\\n**Geometric solver** | 67.52 | 80.00 | 34.91 | 0.00 | 100.00\\n**CADCodeVerify** | 74.52 | 90.00 | 32.64 | 0.00 | 100.00\\n\\nWe hope these experiments address the reviewer\\u2019s concerns. Please feel free to share any additional feedback or questions\\u2014we would be glad to discuss further and work towards enhancing your evaluation of our work.\"}", "{\"title\": \"Response to Reviewer jRaW Part(2/2)\", \"comment\": \"## Clarity/Readability\\nWe thank the reviewer for their suggestions regarding improving the clarity/readability of our paper. We have updated the text in the methodology/methods section of the paper to improve clarity/readability. While we were unable to move Figure 9, of the appendix, to the main paper due to space constraints, we have added an abridged version of this figure to the main paper (Figure 5, in the updated version of the paper). \\n\\nRegarding the equations Section 3, we chose to utilize equations in our description to provide a concise and modular explanation of each component of our approach in a consistent manner to prior work on code generation and refinement. If the reviewer strongly feels as though the equations should not be in the main paper, we can move them to the appendix in the camera-ready version \\n\\n## Justification for choices in methodology \\n\\n- Why were two refinement steps used? -- This information can be found in Section B.3 in the Appendix. We did not observe any improvement after the second stage of refinement. Therefore, we restricted the number of refinement steps to two. This result is consistent with prior works on code-refinement which also showed that the effectiveness of code-refinement wanes after two iterations of feedback [1,2] \\n\\n- Why do we use a five verification questions without considering object complexity? -- We do not use a fixed number of questions per object. As stated in the \\u201cQuestion-Answer Generation\\u201d subsection in Section 3.3, \\u201cCADCodeVerify generates between two to five questions per example.\\u201d Therefore, we prompt the model to generate as many questions as it deems appropriate, with five selected as a reasonable upper bound to constrain the outputs. We can add additional examples of when CADCodeVerify generates a different number of questions (i.e., not five) to the appendix if the reviewer would like to see those. \\n\\n[1] - Madaan, Aman, et al. \\\"Self-refine: Iterative refinement with self-feedback.\\\" Advances in Neural Information Processing Systems 36 (2024). \\n\\n[2] - Chen, Hailin, et al. \\\"Personalised distillation: Empowering open-sourced llms with adaptive learning for code generation.\\\" arXiv preprint arXiv:2310.18628 (2023).\"}", "{\"title\": \"Response to Reviewer 9oaM\", \"comment\": \"Dear 9oaM,\\n\\nThank you for your valuable feedback and response. As we approach the end of the discussion period today, we welcome any final thoughts or concerns about our paper that you would like us to address. Please don\\u2019t hesitate to share any additional feedback or questions\\u2014we are happy to discuss further and work toward improving your evaluation of our paper.\"}", "{\"title\": \"Response to Reviewer jRaW Part (2/2)\", \"comment\": \"## Performance Insights and Limitations\\n\\nThank you very much for your valuable feedback. We agree that your recommendation to explore a collaborative human-machine approach is an intriguing idea, and we plan to pursue it in a follow-up paper. For this rebuttal, we conducted preliminary experiments as a proof of concept to demonstrate how humans and machines can collaborate to provide feedback. These experiments were compared against two baselines: (1) the human-in-the-loop approach (Human) and (2) our proposed approach, CADCodeVerify (Machine). The details of these experiments are as follows: \\n\\n \\nThe results of these experiments are presented below, demonstrating that the collaborative human-machine approach outperforms both baselines for the GPT-4 few-shot setting: \\n\\n\\n \\n\\n- **Human-in-the-loop (Human):** In this experiment, a human (the author) directly provides feedback to LLMs to refine the generated object. \\n\\n- **CADCodeVerify (Machine):** Our proposed approach, CADCodeVerify, delivers feedback to the LLMs automatically without human intervention. \\n\\n- **Collaborative human-machine (Human and Machine):** The machine first generates questions based on language prompts and creates four images of the generated object from different angles (0\\u00b0, 90\\u00b0, 180\\u00b0, 270\\u00b0). The human selects the best viewing angle, and the machine answers the questions using the chosen image. The human then verifies and corrects the machine's responses. Based on these verified answers, the machine generates feedback, and also human adds additional feedback. The combined feedback from both the human and machine is used to refine the 3D object through iterative improvements. \\n\\n| **Feedback Mechanism** | **IoGT \\u2191** | **Point Cloud dist. \\u2193** | **Hausdorff dist. \\u2193** | **Compile Rate \\u2191** |\\n|--------------------------------------|---------------------|-------------------------|--------------------------|---------------------|\\n| **Generated** | 0.930 (0.043) | 0.156 (0.138) | 0.495 (0.287) | 98.5% |\\n| **CADCodeVerify (Machine)** | **0.948 (0.036)** | 0.137 (0.136) | 0.445 (0.302) | 98.5% |\\n| **Human-in-the-Loop (Human)** | 0.944 (0.032) | 0.120 (0.140) | 0.397 (0.354) | **99.0%** |\\n| **Collaborative Human and Machine** | 0.947 (0.025) | **0.087 (0.178)** | **0.33 (0.552)** | **99.0%** |\\n\\n\\n## Methodology Justifications \\n\\n### **Rationale for limiting to two Iterations:**\\n\\nAs noted in Table 5, 6 and 7 (in the Appendix), the iterative refinement process with CADCodeVerify demonstrated no significant improvement in metrics like Point Cloud Distance and Compile Rate, Hausdorff Distance beyond two iterations. This aligns with prior studies in iterative code refinement (e.g., Madaan et al., 2023), suggesting that diminishing returns after two iterations are common in such scenarios. \\n\\nEmpirically, we report the results for both refine-1 and refine-2 across all approaches and three LLMs: GPT-4 (Table 5), Gemini (Table 6), and CodeLlama (Table 7). In many cases, refine-2 performs worse than refine-1 or shows only a slight improvement. \\n\\n \\n\\n### **Importance of Bounding Questions:** \\n\\nBounding the number of validation questions between two to five ensures that the refinement process is directed towards addressing the most critical discrepancies in the generated designs. This avoids overloading the refinement loop with excessive or irrelevant feedback, which could dilute its effectiveness. \\n\\n \\n\\nEmpirically, we randomly selected 100 examples and conducted two experiments using GPT-4. The details of these experiments are as follows: \\n\\n- **Zero-shot QA generation:** In this experiment, questions were generated in a zero-shot manner without providing any demonstration examples or restricting the number of questions. \\n\\n- **CADCodeVerify:** This is our proposed approach, where few-shot examples were provided to the LLMs, along with instructions to generate between two and five questions. \\n\\nThe results, as shown in the table below, indicate that LLMs perform better when questions are generated using few-shot demonstrations and the number of questions is limited. \\n\\n| **Ablation Study** | **IoGT \\u2191** | **PLC distances \\u2193** | **Hausdorff dist. \\u2193** | **Compile Rate \\u2191** |\\n|-----------------------------|---------------------|---------------------------|---------------------------|---------------------|\\n| **Generated** | 0.909 (0.062) | 0.156 (0.150) | 0.491 (0.348) | 96.5% |\\n| **Zero-shot QA generation** | **0.919 (0.049)** | 0.141 (0.112) | 0.471 (0.280) | **98.0%** |\\n| **CADCodeVerify (Ours)** | **0.919 (0.045)** | **0.126 (0.122)** | **0.444 (0.308)** | 97.5% |\"}", "{\"summary\": \"This work contributes a new method for VLM-based CAD code synthesis, and a new dataset for evaluating CAD programming models. It builds on past work on code repair for CAD synthesis \\u2013 the core idea is that instead of just providing the current render during repair (as prior work 3D-Premise did), the LLM also comes up with *verification questions* based on the task description (not using the current render) which it then answers (using the current render). These Q/A pairs are summarized into feedback for what needs changing, which is provided to the repair model alongside the current render to prompt code repair.\\n\\nThe dataset, CADPrompt, was constructed by selecting objects from a dataset used in prior work and annotating them with language prompts along with Python CAD code. They evaluate on a range of models (GPT4, Gemini, CodeLLama) and compare to several baselines (no refinement, 3D-Premise, and Geometric Solver). The last of these baselines, Geometric Solver, is also a contribution \\u2013 it's a baseline that gets to \\\"cheat\\\" and compare the ground truth model to the current rendered model in terms of a range of geometric feaures. Their methods generally outperforms the baselines.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": [\"The contribution of a language-to-CAD dataset is useful for the community, and hiring CAD experts to write the CAD code for it means it's likely high quality.\", \"The comparison is run across a range of LLMs including open source ones (CodeLlama)\", \"They generally surpass the 3D-Premise baseline and score fairly close to the geometric solver baseline that gets to cheat and use the ground truth CAD model in calculating its feedback. They score particularly well on the more difficult problems, and benefit more from feedback than 3D-Premise. They show improvement across three different similarity metrics in addition to compilation success rate.\"], \"weaknesses\": [\"The term \\\"success rate\\\" is a bit of a misleading name for something that means \\\"compilation rate\\\" or \\\"compilation success rate\\\" \\u2013 it gives me the impression that the model succeeded at solving the task, not that it generated an arbitrary piece of code that compiled. For example in the abstract \\\"increasing the success rate of the compiled program\\\" or in the intro \\\"5.5% increase in successful object generation\\\" all gave me this impression until I dug into it.\", \"Also, the abstract says 5.0% and the intro says 5.5% \\u2013 which is correct?\", \"Does 3D-Premise get to do the Code Execution step described in 3.2, where if compilation fails it retries with compiler feedback up to N times? If not, it seems important to do this comparison to disentangle how much of the benefit is coming from compiler feedback versus the Q/A contribution.\", \"The improvements are modest but still reasonable \\u2013 I would defer to a reviewer more familiar with these metrics on this. Looking at GPT-4 Few-shot in Table 2 (which looks like overall the best model), the IoGT metric difference is .942 vs .944, the Point Cloud distance is .127 vs .137, and the Hausdorff distance is .446 vs .419. These are not huge differences, but still a notable step towards the Geometric Solver performance.\", \"It's confusing that \\\"Refine 1\\\" causes 3D-Premise to decrease in performance in Figure 3 \\u2013 see full note in \\\"Questions\\\" section, but clarification on this would be helpful to understanding the results.\"], \"questions\": [\"Is the Geometric Solver output given to an LLM during the feedback phase? I assume so given how it varies depending on which LLM is used in Table 2. Does it get to do two rounds of feedback like the CADCodeVerify model does?\", \"\\\"Refine 1\\\", \\\"Refine 2\\\", and \\\"Generated\\\" should be defined somewhere, right now the first place they appear is in Figure 2 headings (which are not searchable text in the PDF) and then when discussing results, but they aren't clearly defined.\", \"Is \\\"Generated\\\" the step right after Section 3.1 finishes or right after Section 3.2 finishes \\u2013 i.e. is \\\"Generated\\\" before or after the compiler-error based verification step?\", \"Are \\\"Refine-1\\\" and \\\"Refine-2\\\" the two rounds of refinement (Section 3.3) or is the first one the code repair from Section 3.2?\", \"Why does \\\"Refine-1\\\" go *down* so much in Figure 3 for 3D-Premise? This is very surprising to me and would be helpful to clarify. Refinement ought to help, and it did help them in that prior work. While it's not as strong of an effect, refinement also seems to hurt all the methods a bit in some of the other plots (like the Complex plot) \\u2013 why would rounds of refinement hurt instead of help?\", \"These graphs are showing the compilation rate. It'd be helpful to see changes in the other metrics from Table 2 that have to do with correctness instead of whether or not the code compiles \\u2013 those seem key to understanding whether feedback is helping.\"], \"minor_fixes\": [\"The CADQuery compiler $\\\\phi$ is deterministic, right? So the \\\"$\\\\sim$\\\" should be an \\\"=\\\" on line 177\", \"The description of the \\\"number of compilable outputs\\\" discussed in 305-310 was confusing to me until I saw the caption of Figure 11 in the appendix. It would be helpful to move part of this description to the main paper \\u2013 it wasn't clear to me that this meant the number of experiments where the code compiled successfully. And to clarify \\u2013 are those 6 experiments the six large rows in Table 2 (that vary between language model used and not between method used)?\", \"In discussing the geometric solver the paper notes \\\"This feedback method serves as an upper limit for CAD code refinement, as it conveys the precise geometric differences between the generated 3D object and the ground truth.\\\" While I think this is a useful baseline, it's not literally an upper limit for performance from code refinement feedback \\u2013 as shown for example by how it's occasionally outperformed in the experiments. It's a great baseline to have, I just would suggest against using the term \\\"upper limit\\\" for something that's not actually an upper bound.\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to Reviewer jRaW Part (1/2)\", \"comment\": \"Thank you for following up with your questions. We greatly appreciate your valuable feedback and the time you\\u2019ve taken to engage with us. We have addressed your concerns and are committed to ensuring that all your questions are fully answered. Please let us know if there\\u2019s any additional information we can provide to support your assessment.\\n\\n## Object/mesh complexity \\n\\nThank you for your follow-up question regarding mesh complexity. We agree with the reviewer that classifying 3D objects as simply \\\"simple\\\" or \\\"complex\\\" is imperfect. This limitation is evident in our dataset. For instance, while a circle is a simple shape conceptually, it consists of 73 vertices and 140 faces. In response, we have decided to move the discussion of mesh complexity in Section 4.3 to the Appendix. \\n\\nTo address the reviewer\\u2019s concern, we generated an additional way to classify the complexity as \\u201csemantic complexity\\u201d. Following the reviewer's suggestion, we recruited an independent mechanical engineer to semantically rate the complexity of 200 objects, categorizing them as Simple, Moderately Complex, Complex, or Very Complex. We found the semantic complexity measure to be highly effective, as demonstrated by the dataset split based on semantic complexity shown in Table 8 of the Appendix. Further details on semantic complexity are provided in our previous response and in the paper on line 288. \\n\\n \\n\\n## Key advances over 3D-premise \\n\\nThank you for bringing this to our attention. We have incorporated a code repair stage into the 3D-Premise approach, and all three refinement approaches share the same code repair stage, as outlined in Equation 2. \\n\\nFurthermore, since 3D-Premise serves as our baseline, we conducted additional experiments to enhance its performance, including: \\n\\n- Incorporating multiple viewpoints by utilizing four images captured from different angles (0\\u00b0, 90\\u00b0, 180\\u00b0, and 270\\u00b0). \\n\\n- Applying chain-of-thought prompting within the 3D-Premise framework.\", \"the_results_of_these_experiments_are_provided_in_the_table_below_for_the_gpt_4_few_shot_setting\": \"| **Feedback Mechanism** | **IoGT \\u2191** | **Point Cloud dist. \\u2193** | **Hausdorff dist. \\u2193** | **Compile Rate \\u2191** |\\n|--------------------------------------|---------------------|-------------------------|--------------------------|---------------------|\\n| **Generated** | 0.939 (0.030) | 0.155 (0.140) | 0.494 (0.368) | 96.0% |\\n| **3D-Premise** | 0.942 (0.033) | 0.137 (0.155) | 0.446 (0.396) | 91.0% |\\n| **3D-Premise - Multiple viewpoints** | 0.941 (0.035) | 0.132 (0.147) | 0.437 (0.395) | 91.0% |\\n| **3D-Premise - CoT** | 0.941 (0.031) | 0.131 (0.142) | 0.432 (0.409) | 92.0% |\\n| **3D-Premise - CoT & Multiple viewpoints** | 0.941 (0.031) | 0.150 (0.162) | 0.477 (0.367) | 90.0% |\\n| **CADCodeVerify (Ours)** | **0.944 (0.028)** | **0.127 (0.135)** | **0.419 (0.356)** | **96.5%** |\\n| **Geometric solver*** | **0.944 (0.031)** | **0.103 (0.152)** | **0.399 (0.433)** | 95.5% |\"}", "{\"title\": \"Response to Reviewer 9oaM Part (2/2)\", \"comment\": \"## How representative are the objects in your dataset?\\n\\nWe have conducted an additional experiment to understand the breakdown of the semantic complexity of the objects in our dataset. We recruited an independent annotator (who is a mechanical engineer) to annotate the structural complexity of each of the 200 objects in CADPrompt according to the following scale: \\n\\n- (Simple) - The object is extremely basic, with few features. It may \\nconsist of one geometric shape. \\n\\n- (Moderately Complex) - The object has a moderate amount of detail, \\nwith a few distinct features or components \\n\\n- (Complex) - The object is complex, with many interconnected parts, \\nfine details, or intricate shapes. \\n\\n- (Very Complex) - The object is highly intricate, with many \\ncomponents, detailed textures, or complex shapes. It may have a large \\nnumber of fine details, interlocking parts, or unique geometric features \\n\\nThe breakdown of our dataset according to this independent reviewer\\u2019s evaluation was as follows: Simple - 17, Moderate Complexity - 39, Complex - 87, and Very Complex - 57. Please refer to Table 1 in the updated version. \\n\\n \\n\\nIn terms of real-world representativeness, real-world manufacturable designs generally fall under the \\u201cvery complex\\u201d category and upwards. Currently we observe that existing LLMs can easily generate simple/moderate objects through existing in-context learning methods. However, consistently generating very complex objects at high-degrees of precision is still an unsolved problem. CADPrompt is a test-bed which provides a stepping-stone to test the competency of CAD Code Generation methods prior to deployment on real-world objects. \\n\\n## Why didn\\u2019t we evaluate on the dataset used by 3D-Premise \\n\\nThe dataset provided by 3D-Premise was not made publicly available by the authors, which prompted us to create CADPrompt. Furthermore, 3D-Premise's dataset was only comprised of 57 examples, with insufficient information regarding how the annotations and objects were collected and validated, which impeded our ability to reproduce a similar dataset. Leveraging LLMs for CAD model generation is still a nascent area of research, with the first approach introduced as recently as May 2023. As such, there is a dearth of standardized testing methods and benchmark to competently evaluate the capabilities of CAD Code Generation methodologies. Through CADPrompt, we provide a much needed, publicly available benchmark to standardize future work on CAD Code Generation.\"}", "{\"title\": \"Response to Reviewer Ghys\", \"comment\": \"Dear Ghys,\\n\\nThank you for your valuable feedback and engagement with our paper. As the discussion period is coming to an end, we wanted to check if there are any unresolved concerns or additional clarifications you would like us to address. We are eager to ensure that all your questions are fully answered and that we can provide any further information to support your assessment.\"}", "{\"title\": \"Response to Reviewer 9oaM Part (1/2)\", \"comment\": \"We sincerely thank the reviewers for their thoughtful feedback and valuable insights, which have greatly improved the quality of our work.\\n\\n## What is the input to the VLLM during refinement \\n\\nOur feedback generation process is comprised of three steps (Figure 1). (a) The LLM is prompted to generate questions based on the prompt to verify whether the generated object corresponds to the prompt. (b) The LLM then analyzes the generated object and answers the set of validation questions from step-a. (c) The answers to these questions are then summarized concisely as feedback to refine the initially generated object. Therefore, to answer the reviewer\\u2019s question, the generated object is passed back into the model to generate language-based feedback. This language-based feedback is then provided to the LLM to update the initial code. \\n\\n## Do we consider naive or basic prompting \\n\\nYes, the results in the \\u201cGenerated\\u201d phase are the results with just a basic prompt. We experiment with a zero-shot and a few-shot prompt as shown in table-2. \\n\\n## Main weakness is a lack of baselines \\n\\nTo the best of our knowledge, the only pre-existing method for CAD Code Refinement is 3D-Premise. By showing that we outperform 3D-Premise across all four metrics, we show that CADCodeVerify is the new state-of-the-art. We further develop a second baseline which compares our method which provides structural feedback, to a method which provides precise feedback (Geometric Solver) regarding the specific parametric differences between the two objects. Our methodology also integrates in-context learning methods such as Chain-of-thought reasoning, few-shot exemplars, and self-verification. If the reviewer has specific baseline suggestions in mind, we are willing to implement and include them in the paper. However, we believe our work makes a significant technical contribution and is thoroughly evaluated against existing methods. \\n\\n## Why CADCodeVerify performs better than 3D-Premise \\n\\n3D-Premise simply prompts the model to utilize an image of the generated object and update the code based on any discrepancies identified. Whereas CADCodeVerify first prompts the model to \\\"self-verify\\\" and produce actionable feedback, which is provided to the VLM in conjunction with the image of the object to better isolate the changes that need to be made to the object. Recent work on visual programming has also shown that when utilizing visual feedback to update or refine code, it is helpful to process the images and produce textual-information about the image to integrate into the refinement feedback [1]. \\n\\n[1] - Gao, Minghe, et al. \\\"De-fine: De composing and Re fin ing Visual Programs with Auto-Feedback.\\\" Proceedings of the 32nd ACM International Conference on Multimedia. 2024. \\n\\n \\n\\n## Why CADCodeVerify performs better than the Geometric Solver \\n\\nWe also found this result to be interesting! The Geometric solver provides high-level geometric feedback about the object (See Figure 6 in the appendix), including details about the volume, surface-area, height, width, etc. However, the feedback from the geometric solver does not provide any insights regarding the structural correctness of the object. CADCodeVerify on the other hand explicitly focuses on correcting structural errors in the object relative to the prompt provided. The first example shown in figure 2 provides a depiction of this. The generated prism has very similar geometric properties with regards to the categories expressed by the geometric solver, therefore it is unable to adequately discern and correct the errors in the generation. CADCodeVerify on the other hand identifies that the square cutout should be moved \\u201cto the top-edge of the rectangle\\u201d as stated in the original prompt and is able to produce feedback to address this discrepancy. \\n\\n## Refine 1 / 2 definitions \\n\\n\\u201cRefine-1\\u201d and \\u201cRefine-2\\u201d correspond to the results after the first and second refinement steps.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}", "{\"comment\": \"> We hypothesize that 3D-Premise leads to a reduction in the compile-rate, because off-the-shelf LLMs struggle to infer the changes they need to make to CAD scripting code to edit the 3D object, based on the image in isolation.\\n\\nAh, and I see now how in Table 2 the compile rate for 3D-Premise is lower than that of Generated, which is showing this same trend that appears in Figure 4. And in Table 2 the *distance metrics* do improve for 3D-Premise over Generated \\u2013 so self repair is breaking some code so that it no longer compiles, but is possibly improving the distance metrics on the code that it does help on.\\n\\nThank you for the responses and clarifications \\u2013 I will maintain my rating.\"}", "{\"title\": \"Reply to authors\", \"comment\": \"Thanks for rebuttal. However, it seems like not all of my questions were answered. For e.g. Would providing images of the generated object from multiple viewpoints improve 3D-Premise\\u2019s performance? What about using a chain of thought prompting (for e.g. explicitly ask the model to reason whether each criterion in the initial task are satisfied)? This will be similar to CADCodeVerify, but doing 1 VLM call instead of doing 3 VLM calls in the verify step to reduce the inference time.\\n\\nFor ambiguity in inputs, my question was more if this affects the evaluation metrics (for e.g. the model could have found a solution is the correct with respect to the intent, but different from the ground truth object, and hence affecting the evaluation metrics). Also, from your manual reviews, how often do you find the intents ambiguous. For example, I see several instances from the examples in the paper (Figure 2) that use phrases like large rectangle (without a specific dimension) or slightly smaller (rather than specifying exactly how much smaller). If significant number of intents in your dataset are of this form, I think it greatly impacts the results and the conclusions of this paper (essentially making them not so useful).\"}" ] }

[ "Filippos Christianos", "Georgios Papoudakis", "Thomas Coste", "Jianye HAO", "Jun Wang", "Kun Shao" ]

[ "vision-language model", "multi-modal", "android control", "app agent" ]

{ "note_id": [ "wsplJO81WE", "phT27Y79oZ", "oznHwyZaq9", "ltYxHJW2O1", "lLHkfKW1CS", "XeUm94mxHZ", "QdfrBqi3qg", "OwErMWJiFK", "NioiYjmrUh", "Gpc0kaR3px", "Ex1yeZqOfb", "AqEd9n5Ho0", "9XxdqE2Srh", "8TiP2Wy1gR", "2tdmjwIAzX" ], "note_type": [ "official_comment", "official_comment", "official_review", "official_review", "official_review", "official_comment", "official_comment", "official_comment", "meta_review", "official_comment", "official_comment", "official_review", "decision", "official_comment", "official_comment" ], "note_created": [ 1732639843004, 1732035176174, 1730409000443, 1730814709528, 1730580861971, 1732035167927, 1732838092026, 1732546167904, 1734323575261, 1732035180120, 1732298519241, 1730673523189, 1737524002863, 1732035173982, 1732120444172 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission9741/Reviewer_skyy" ], [ "ICLR.cc/2025/Conference/Submission9741/Authors" ], [ "ICLR.cc/2025/Conference/Submission9741/Reviewer_skyy" ], [ "ICLR.cc/2025/Conference/Submission9741/Reviewer_mpP2" ], [ "ICLR.cc/2025/Conference/Submission9741/Reviewer_j9ZA" ], [ "ICLR.cc/2025/Conference/Submission9741/Authors" ], [ "ICLR.cc/2025/Conference/Submission9741/Authors" ], [ "ICLR.cc/2025/Conference/Submission9741/Authors" ], [ "ICLR.cc/2025/Conference/Submission9741/Area_Chair_GxbY" ], [ "ICLR.cc/2025/Conference/Submission9741/Authors" ], [ "ICLR.cc/2025/Conference/Submission9741/Reviewer_skyy" ], [ "ICLR.cc/2025/Conference/Submission9741/Reviewer_Wgue" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission9741/Authors" ], [ "ICLR.cc/2025/Conference/Submission9741/Reviewer_Wgue" ] ], "structured_content_str": [ "{\"comment\": \"Thank you for addressing this point. I am raising my score to a 8. I think this paper is well written, spells out its contributions clearly, provides careful analysis with ablations, and makes a new contribution towards actually deploying models on-device.\", \"note\": \"> and T3A is used for text generation In the AitW dataset\\n\\nThere is a missing period after generation.\"}", "{\"comment\": \"We would like to thank the reviewer for their review and address their concerns below.\\n\\n1. Positional Encodings for Nesting: \\n\\nThe use of positional embeddings allows us to capture the relative positioning of UI elements, helping AcT understand which elements are nearby or grouped together. Without positional embeddings, the order of elements on the screen would be less meaningful. In our approach, nesting information is implicitly encoded in the UI icons themselves, as these icons often have overlapping pixels. By combining this spatial information with the positional embeddings, AcT is able to effectively distinguish between nested and non-nested elements. We attempted to incorporate depth embeddings into the nested structure, combining these with positional embeddings to precisely represent the nesting of UI elements. However, this integration did not yield any noticeable improvement in performance. We hope this clarifies the role of positional embeddings in handling nested UI elements. \\n\\n2. Distinct Action Types: \\n\\n*We have updated the paper to reflect this more clearly, and have added a link to the relevant appendix in Section 3.3.*\\n\\nThe distinct action types are outlined in Appendix A of the paper, specifically Table 5. Though there are only 10 distinct action types in AndroidControl, and 8 in AitW, there are a combined 11 distinct actions. The distinct actions are: (1) open-app, (2) click, (3) long-press, (4) input-text, (5-8) scroll (right/left/up/down), (9) navigate-home, (10) navigate-back, and (11) wait. open-app, wait, and long-press do not feature in AitW, while navigate-home does not feature in AndroidControl.\\n\\n3. Florence Fine-Tuned or Not?\\n\\nYes, it is fine-tuned. The Florence2 baseline refers to the same Florence2 model used within our LiMAC framework. Practically, we fine-tune Florence2 only once for each dataset, and we use it both as part of LiMAC as well as to compare against it.\\n\\n4. End-to-End Accuracy:\\n\\n*We have improved the wording of this explanation in our paper, section 4.2.*\\n\\nOverall accuracy is dependent on both the action type and the action specifications. When predicting a \\u201cwait\\u201d action, if the action type is predicted correctly, this immediately yields a positive result for that timestep. However, when predicting actions with additional parameters, such as \\u201cinput-text\\u201d, even if the action type is predicted correctly, the overall result may be negative if the VLM fails to fill the \\u201ctext\\u201d input field correctly. Using an example to try and illustrate this, take an episode with 10 steps, where the correct action is \\u201cwait\\u201d for 5 steps and \\u201cinput-text\\u201d for the other 5. Predicting \\u201cwait\\u201d for all 10 steps will yield an overall accuracy of 0.5. However, predicting \\u201cinput-text\\u201d for all 10 steps will yield an accuracy of 0.5 * (probability of the VLM to give the correct text input). If this VLM accuracy is 80%, the overall accuracy will be 0.5 * 0.8 = 0.4.\", \"minor_issues\": \"Thank you for pointing these out! These have now been fixed.\\n\\n5. Running on Smartphone: \\n\\nWe have not yet explored deploying the model on smartphones. Instead, we focus on training and comparing agents in simulated environments, with an understanding that limited computational capacity is a real and key constraint in mobile devices. By leveraging open-source datasets, we aim to improve performance and efficiency in these settings. The smaller size of AcT and its faster response times, however, highlight its promising potential for smartphone applications in the future.\"}", "{\"summary\": \"This work introduces LiMAC, which is an architecture for training smaller neural networks that could fit on-device for UI control. LiMAC processes UI elements as embeddings, uses a contrasting learning approach for click actions, a gated architecture for selectively invoking a fine-tuning VLM to generate text content, and it shows ablations of the architecture.\\n\\nThe work evaluates several different variances of LiMAC and compares them to multiple baselines, including those using large, proprietary models. In addition to comparing accuracy, they also compare inference time, which is important for on-device applications.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"The work tackles an under-explored problem of considering compute/inference costs for UI control agents. To get practical agents, the community needs this type of work.\\nThe performance of the LiMAC agent is strong compared to the baselines in the paper. They apply InfoNCE loss in an interesting way for screen understanding and automation. The work includes a large number of baselines, which is appreciated. The authors take off-the-shelve VLMs and fine-tune them for device control, as part of the baselines and main results. And the work include ablation analyses of their results.\\nThe authors explain their methodology well and do rigorous evaluation.\", \"weaknesses\": \"The results are not that much better compared to the baseline of fine-tuning an off-the-shelf model, particularly for Florence2 as shown in Table 1 (70.8 vs 72.2 for AitW and 57 vs. 63.1 for AndCnrl). It's also hard to tease out if the improved performance is from architectural breakthroughs or rather just from adding more parameters by incorporating the LiMAC network.\\n\\nOn a related note, the claims of the superiority of the proposed architecture would be strengthened with online evaluation by using model-based evaluation on AitW as in DigiRL: https://arxiv.org/abs/2406.11896 or using an online benchmark (e.g., AndroidWorld https://arxiv.org/abs/2405.14573).\\n\\nThe architecture of LiMAC is not particularly novel. While the contrastive loss is interesting, the other parts such as representing UIs using embeddings of UI elements is not novel (past examples: Mapping Natural Language Instructions to Mobile UI Action Sequences: https://arxiv.org/pdf/2307.10088, Android in the Wild: A Large-Scale Dataset for\", \"android_device_control_https\": \"//arxiv.org/pdf/2307.10088). While the gated architecture with the VLM is a sensible engineering decision, it is not novel from a research perspective.\", \"questions\": \"Please see weaknesses.\\n\\n## Comments\\n(You do not need to respond to these; they are intended to be helpful)\\n\\nIn future work, you can\\u2026\\n* Do actual on-device implementation. I suspect it may be non-trivial\\n* You could report mobile-specific performance metrics (battery impact, memory usage, etc.)\\n* Analyze real-world latency measurements on real phone, which would be compelling\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The paper introduces Lightweight Multi-modal App Control (LiMAC), a framework designed for efficient mobile app control by combining a small Action Transformer (AcT) and a fine-tuned vision-language model (VLM). LiMAC processes user goals and smartphone states, making real-time decisions through AcT for action types like clicking or scrolling. For complex tasks requiring text input, it leverages the VLM to generate appropriate content. Evaluation of LiMAC on AndroidControl and Android-in-the-Wild datasets, LiMAC shows superior accuracy and speed over traditional VLMs and foundation models like GPT-4o, achieving up to 42% higher action accuracy and reducing task completion time by 30-fold. This approach emphasizes efficient model use on resource-limited devices, while future improvements may incorporate reinforcement learning for enhanced performance.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": [\"LiMAC combines a small Action Transformer (AcT) with a fine-tuned vision-language model (VLM). This hybrid approach is tailored to the computational constraints of mobile devices, achieving efficient and accurate control without relying on large, resource-intensive models. The AcT independently handles common actions, while the VLM is selectively employed for complex natural language tasks, optimizing both resource usage and response time.\", \"LiMAC\\u2019s modular structure supports the integration of different models for specialized tasks, such as using AcT for action type prediction and click targeting, while optionally substituting modules for specific needs (e.g., Qwen2-VL for text generation tasks). This flexibility enables developers to adapt LiMAC easily for varied app control tasks, optimizing specific components without overhauling the entire architecture\\u200b.\", \"The AcT component employs a contrastive objective to identify UI elements for click actions, using cosine similarity and a learnable temperature parameter. This approach is advantageous in handling class imbalance and dynamically varying UI elements across episodes. The use of contrastive learning allows AcT to focus on directional alignment in the embedding space, facilitating precise UI element targeting even in dense or complex layouts\"], \"weaknesses\": [\"Although the paper evaluates LiMAC on two datasets, both datasets are relatively specific to Android applications, potentially limiting the generalizability of results to other operating systems (e.g., iOS) or app control tasks with distinct interface designs.\", \"The paper does not provide an extensive scalability analysis of LiMAC\\u2019s architecture as task complexity or the number of available UI elements increases, which may impact its robustness in more complex or densely populated app environments.\", \"LiMAC operates within a fixed action space, which could restrict its adaptability to applications requiring more dynamic or unconventional actions not included in its current design. This might hinder its flexibility in expanding to novel app interaction scenarios.\", \"The evaluation is conducted on simulated datasets without testing LiMAC\\u2019s performance on actual mobile devices. This limits understanding of its practical usability, particularly concerning latency, responsiveness, and potential challenges from hardware constraints and real-world conditions.\"], \"questions\": \"See Weaknesses.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The paper presents a new framework, called LiMAC, for lightweight and efficient Android navigation. Combining text and image embeddings, an action transformer predicts which of ten actions needs to be taken, and, depending on this output, a fine-tuned VLM may be queried to help with more complex tasks, such as text input. A major advantage of this setup is its significantly reduced inference time requirements compared to models such as SeeAct, while attaining higher overall performance on the Android-in-the-Wild and AndroidControl datasets. The authors apply ablation studies to demonstrate the usefulness of CLIP fine-tuning and the image embeddings, as well as the robustness of their setup to the absence of UI tree descriptions. In all, LiMAC is just a 520M-parameter addition, allowing it to run efficiently directly on Android devices.\", \"soundness\": \"4\", \"presentation\": \"4\", \"contribution\": \"3\", \"strengths\": \"The paper is incredibly clear and well-written. I am not an expert on Android or UI agents, but it was obvious what the contribution was, and why it was important. The small number of parameters in LiMAC make it clear how this advances the ability of users to run advanced Android control models directly on devices. The way the information is encoded is also very thoroughly described. It is somewhat difficult for me to evaluate novelty, but, assuming the related works section does not have any glaring omissions, is a novel approach to solving this problem. The figures are very clear and well-made, and efficiently convey the achitectural design to the reader. Finally, the ablation experiments are well-run, and justify the use of the multiple submodules, as well as demonstrating robustness to a lack of a UI hierarchy description, implying that this method may be applicable out-of-the-box to completely novel Android environments.\", \"weaknesses\": \"I am not an expert in this area, so it is difficult for me to point out major weaknesses\\u2014the paper clearly states a contribution, and is very self-contained. However, there are several minor things that were unclear to me, where the paper might benefit from more detail:\\n\\n1. The paper mentions that positional encodings are used to represent the nesting of UI elements. How is this done, exactly?\\n2. It would be good to know what the \\\"ten distinct action types\\\" are\\u2014it seems like the same few examples are given at several points in the paper, and only a few are focused on. Are the rest omitted because they're not very interesting and \\\"just work\\\", or are very similar to each other, or something else?\\n3. In Table 1, is Florence2 fine-tuned or the base model? (I know that in LiMAC with Florence2, Florence2 would be fine-tuned, but it's not fully clear whether this is the case for Florence2 alone).\\n4. How, exactly, is \\\"end-to-end accuracy\\\" calculated? Why would predicting \\\"wait\\\"s rather than \\\"input-text\\\"s increase overall accuracy? I understand that you try to explain this on lines 369\\u2013403, but a clear definition of the accuracy calculation would make this paragraph make more sense.\", \"minor_issues\": [\"On line 118, \\\"that treat\\\" should probably be \\\"that they treat\\\"\", \"On line 142, it's a bit weird that the full architecture link links back to the current section\"], \"questions\": \"Please see the weaknesses section, parts 1\\u20134.\\n\\n5. You mention that LiMAC \\\"address[es] the computational constraints inherent to smartphone\\\". Have you tried running it on a smartphone? If so, how well did it go?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"We would like to thank the reviewer for their review and reply to their concerns below.\\n\\n1. Android Focus:\\n\\nWe focus on Android because it offers two open-source datasets and has a significant share of the global market as a mobile phone OS. These factors make our findings highly relevant and broadly applicable. We also want to emphasise that our methodology is flexible and can be adapted to other operating systems, such as iOS, once similar datasets become available.\\n\\n2. UI Element Scalability:\\n\\n*We have included a study in Appendix D4 (Figure 5) that examines how the number of UI elements affects the ratio of successful to unsuccessful action predictions.*\\n\\nIn the evaluated datasets, the number of UI elements per observation varies from 1 to 290, with most observations having fewer than 100 UI elements. While it's possible for an observation to exceed this range, potentially leading to out-of-distribution issues, we believe that the current span adequately encompasses the bulk of practical scenarios encountered in app control. This ensures that our model is capable of managing a diverse range of tasks.\\n\\n3. Fixed Action Space: \\n\\n*We have updated our description of the action space in Appendix A*\\n\\nWhile the action space is fixed, it is the same as the AitW and AndroidControl datasets (where it is also fixed) and aligns with prior work on app agents and emulators, such as SeeAct [1], AndroidWorld [2], AppAgent [3], MobileAgent [4], DigiRL [5]. These actions are grounded by how a human interacts with the phone (e.g., click, or swipe). This fixed action space covers the vast majority of actions available on Android devices, and we believe it provides a robust representation of real-world app control behaviour. \\n\\n4. Simulated Data vs. Real-World Devices: \\n\\nWe acknowledge the concern that the approach may face challenges in real-world settings. However, this is a broader issue in the field, not unique to our approach and requires significant amounts of work. By leveraging open-source datasets, we aim to improve performance and efficiency in the simulated setting as a foundation. The smaller size of AcT and its faster response times highlight its promising potential for on-device smartphone applications in the future.\\n\\n[1] Zheng et al. Gpt-4V(ision) is a Generalist Web Agent, if Grounded.\\n\\n[2] Rawles et al. Androidworld: A Dynamic Benchmarking Environment for Autonomous Agents.\\n\\n[3] Yang et al. Appagent: Multimodal Agents as Smartphone Users.\\n\\n[4] Wang et al. Mobile-agent: Autonomous Multi-modal Mobile Device Agent with Visual Perception.\\n\\n[5] Bai et al. DigiRL: Training In-The-Wild Device-Control Agents with Autonomous Reinforcement Learning.\"}", "{\"comment\": \"Thank you for your comment, and thank you for spotting the typo, we will fix it in the next version.\"}", "{\"comment\": \"We thank the reviewer for their valuable feedback.\\n\\n*We now include a paragraph with clarifications on the memory footprint and the use of the term \\u201cLightweight\\u201d in Sec 3, paragraph 2.*\\n\\nWe acknowledge the concern about memory usage, as our approach does require an additional model to be loaded. While we did not claim a smaller memory footprint, we have now clarified this in Section 3 of the paper.\\n\\nThat being said, there is also some nuance to this discussion, as memory operates differently from compute: its cost is small as long as the device has enough capacity. \\nMost modern devices can accommodate transformer-based models up to ~7B parameters [1, 2], though models at the higher end of this range often require aggressive quantization (which impacts performance). Our small VLM + 500M model should be within that limit, even allowing overhead for the OS and open apps. Our use of \\\"lightweight\\\" refers instead to the fact that we i) kept the overall parameter size manageable and likely to fit on a typical modern device (no need for unloading models), leaving some space for the operating system and other apps, ii) are more efficient in our computations (e.g., no auto-regressive generation in AcT), iii) are using less parameters on average for computation (which should translate to less required floating point operations), and iv) simply having faster inference speed. We have now clarified this in Sec 3.\\n\\nWe also thank the reviewer for mentioning ActionBert, as it\\u2019s indeed relevant and could be explored in future work.\\n\\n[1] Li et al. Large Language Model Performance Benchmarking on Mobile Platforms: A Thorough Evaluation.\\n\\n[2] Laskaridis et al. MELTing point: Mobile Evaluation of Language Transformers.\"}", "{\"metareview\": \"The paper introduces Lightweight Multi-modal App Control (LiMAC), a lightweight and modular to take actions on mobile user interfaces. The framework consists of a small Action Transformer (AcT) which predicts which action (e.g. click, scroll, type) needs to be taken, and, depending on this output, a fine-tuned VLM may be queried to help with more complex tasks, such as text input or opening an app. AcT leverages multi-modal embeddings for each UI element in the current screenshot to help predict the next action. Overall, the paper presents a focused contribution about making lightweight UI interaction agents. The experiments performed show improvements over existing methods both in terms of accuracy and inference cost. The interactive AI community will find insights from this paper useful towards building on-device systems.\\n\\nReviewers consistently mentioned that the paper is well-written and the experiments as well as ablations are sound. All reviewers agreed that using a light-weight action transformer for simpler actions, while offloading more complex actions to a fine-tuned VLM is a novel design choice and has useful practical applications such as on-device deployment. \\n\\nSome reviewers (Wgue, mpP2, skyy) mentioned that the datasets used for evaluation are limited (specific to Android). and will benefit from online evaluations. Authors responded to these concerns by saying that mobile navigation datasets are limited to Android, and the infrastructure to run online evaluation is still lacking and previous works have only managed to run small-scale online evaluation studies. Instead, they stick to offline evaluations but utilise bigger training and evaluation datasets. \\n\\nAdditionally, as pointed out by one reviewer (Wgue), one of the limitations of the current work is the lack of focus or experiments about recovering from incorrect action (which is a very important skill for UI interaction). While this is left for future work, recovering from error (planning) might require bigger models and the paper will benefit from discussions around how their lightweight approach can be incorporated when a stronger bigger model is needed for more complex part of UI Navigation (which is reasoning / planning).\", \"additional_comments_on_reviewer_discussion\": \"During the course of the rebuttal period, the authors addressed some concerns and provided more clarifications around novelty, experiment design and results:\\n\\n1. Novelty: While the use of embeddings for individual UI elements has been explored before, the overall architecture of using a lightweight model for simpler decision making and offloading complex tasks to VLM is novel. Additionally, low level implementation differ (using hidden states of transformer rather than the embedding directly). \\n\\n2. Clarifications: The authors provided various clarifications about the technical details of the approach which makes the paper stronger. Adding clarifications about position encodings, action types, metrics, etc in the main manuscript will improve clarity. Additional experiments / results during the rebuttal should also be added in the appendix. \\n\\n3. Finally, I encourage authors to update the discussion section to include analysis of failure modes in the main paper.\"}", "{\"comment\": \"We would like to thank the reviewer for their review and address their concerns below.\\n\\n1. Accuracy Improvement:\\n\\nWhile we acknowledge that the improvement in accuracy of LiMAC compared to our fine-tuned Florence does not fully solve the problem, we still believe it represents a meaningful and valuable step forward (and in most cases with large improvements in computational efficiency/speed). We have carefully evaluated LiMAC and provided a thorough analysis of its strengths and weaknesses, which we hope highlights the contribution of our work in advancing the field. Regarding the \\u201cmore parameters\\u201d, while the overall framework does have more parameters, we believe that statement does not paint the full picture, as these parameters do not simply compose a larger network (i.e., not all parameters are called at the same time). For example, if one was to consider action type selection as its own separate task, our 520M AcT model outperformed larger models. Moreover, the parameters of the VLM are only used when the VLM is called, which only occurs for open-app or input-text actions, making up 13.9% of the total actions for AndroidControl and 10.6% for AitW respectively.\\n\\n2. Online Evaluation: \\n\\nWe definitely agree that online evaluation is a valuable addition and currently missing from our work, and is an aspect that we aim to address in the future. We believe that fully autonomous, online, and generalisable app agent control, is far from solved [1,2]. Nevertheless, the focus of our study differs from that of DigiRL [3]. In DigiRL, the authors train separate models for different goal categories (General and Webshop) and evaluate on a relatively small subset of goals (the first 96 from the train and test set). In contrast, our work utilises the AndroidControl dataset, which contains more than 13K distinct goals, as well as our subset of AitW, with just under 9K goals. Generalising fine-tuned models in online settings remains an open problem, as highlighted in [1,2] which demonstrate that fine-tuned models perform poorly on a large set of online tasks. \\n\\n3. Novelty: \\n\\n*We have added an acknowledgement of [4] and [5] in Sec 3.1, discussing that they have also explored embedding individual UI elements.* \\n\\nWe would argue that the architecture as a whole is novel even if some of the individual components have been explored before. Indeed, Li et al. [4] (which we thank the reviewer for bringing to our attention) and Rawles et al (AitW) [5] have explored embedding individual UI elements, however we still believe we look deeper into the matter (e.g., using fine-tuned CLIP for the embedding accompanied with ablation studies, using concatenations of embeddings of CLIP and BERT, etc). Furthermore, as the reviewer noted, we use the hidden states of the transformer corresponding to these embeddings for a contrastive learning objective (which to the best of our knowledge has not been explored before). The gated architecture is obviously not a novel idea, and it is pretty natural to the task, but we still believe reporting its capabilities in this context is valuable.\\n\\nWe hope this clarifies our approach and contributions, and we thank the reviewer again for their constructive suggestions. We also appreciate the comments on the on-device implementation.\\n\\n[1] Chen et al. SPA-Bench: A Comprehensive Benchmark for SmartPhone Agent Evaluation\\n\\n[2] Zhang et al. LlamaTouch: A Faithful and Scalable Testbed for Mobile UI Task Automation\\n\\n[3] Bai et al. DigiRL: Training In-The-Wild Device-Control Agents with Autonomous Reinforcement Learning.\\n\\n[4] Li et. al. Mapping Natural Language Instructions to Mobile UI Action Sequences \\n\\n[5] Rawles et al Android in the Wild: A Large-Scale Dataset for Android Device Control\"}", "{\"comment\": \"Thank you for taking the time to respond and clarify.\\n\\nI agree you have provided careful analysis of LiMAC, which is appreciated. I understand your point regarding the parameters not composing an entire network. While this has clear benefits from a computational efficiency point of view, my concern would be from a memory point of view; specifically regarding how \\\"lightweight\\\" the approach is. Since the focus of this paper is on designing models that could theoretically be useful on device, I believe you would need to keep the larger model loaded and ready to go at all times. Thus for real-world deployment it would effectively be the same as a a larger model from a memory point of view (computationally I agree your approach is faster and more efficient). I understand the actually testing it on-device is outside of the scope of this paper, but this should be considered from a practical point of view. Can you please comment on this?\\n\\nThank you for the clarification on CLIP and using hidden states of the transformer for a contrastive learning objective. I agree it is novel. Your comment also reminded me of a somewhat related paper, https://arxiv.org/pdf/2012.12350, which uses BERT pre-training task, also using UI elements. Specifically, \\\"Pre-training task #3: Masked VH text prediction\\\", may be of interest for future work.\"}", "{\"summary\": \"This paper introduces LiMAC (Lightweight Multi-modal App Control), a new architecture designed for mobile app control that combines a lightweight transformer, AcT, with a fine-tuned vision-language model (VLM). The standout feature here is the gated architecture, which smartly assigns tasks: AcT handles basic interactions like clicks and scrolls, while the VLM is called upon only for text-based actions that require deeper language comprehension. The approach yields substantial improvements in both inference speed and accuracy on two mobile action datasets.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. **Novel design.** The authors designed a lightweight module to predict the type of actions to be taken, and execute simple actions with this light-weight module directly. Leaving the VLM to solve complex tasks that involve text generation. This leads to both performance speed-up and better accuracy.\\n\\n2. **Thorough evaluations.** I like how the authors compared using AcT/VLM for different tasks, clearly demonstrating the performance gain by adopting the current design, which makes sense to me.\\n\\n3. **Good writing.** The paper is easy to follow.\", \"weaknesses\": \"1. **Limited Dataset and Tasks.** The authors used two datasets of relatively small size, this paper could benefit from larger-scale experiments and maybe real-world user studies.\\n\\n2. Due to the limited data size, the proposed model may have additional difficulties in solving difficult tasks (which is where the mobile AI is needed to most, from my opinion). More studies/analysis on failure mode could make this paper better.\", \"questions\": \"See my suggestions in the weakness section, here are some of my questions:\\n\\n1. How does the model handle dynamic UI elements or applications with frequently changing interfaces? Do you need to retrain a model for each software update?\\n\\n2. What is the impact of the VLM size on the overall performance? Could one larger VLM learn to solve all the tasks?\\n\\n3. How does the system handle errors or recovery from incorrect actions? Or safegaurding it from performing unwanted actions (for example send out user's passwords to someone else).\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Spotlight)\"}", "{\"comment\": \"We would like to thank the reviewer for their review and address their concerns below.\", \"weaknesses\": \"1. Limited Datasets and Tasks: \\n\\nWe appreciate the reviewer\\u2019s concern regarding the limited size of datasets in our study. While we agree that a larger dataset would provide valuable insights, our work focuses on open-source data, and gathering additional data presents significant (and out of scope) challenges. Specifically, collecting data through emulator interaction is difficult, because Android emulators typically do not offer reward functions to assess task success. AndroidWorld [1] is the only exception we are aware of, though it is limited to just 116 tasks. We are actively working to address this limitation, in future work, by solving the challenges associated with Android emulators, which will allow for a wider and more diverse set of tasks. This limitation is mentioned in the conclusion of the paper.\\n\\n2. More Analysis of Failure Modes: \\n\\n*We thank the reviewer for suggesting a deeper analysis of failure modes is a valuable suggestion. We have added such an analysis in the Appendix D3 (Figure 4).*\", \"questions\": \"1. Handling of Dynamic Elements and Changing Interfaces: \\n\\nRegarding the handling of dynamic elements and shifting interfaces, retraining would be required when transitioning to completely different operating systems, such as from Android to iOS. However, within the same OS, we expect the accuracy drop across different versions to be minimal, having observed only slight performance variations between train and test sets, and with AitW for example containing a range of Android OS versions and phone models. \\n\\n2. Impact of VLM Size: \\n\\nWe believe that training larger Vision-Language Models (VLMs) could potentially improve performance. However, as suggested in the AndroidControl paper [2], achieving 95% accuracy may require up to 60 million high quality episodes, using the PaLM-S model, which may contain over 100B parameters. While scaling up models is crucial, it is also important to recognise the significant resources required to reach such levels of performance (and that is why believe novel research directions that do not focus simply on scale are valuable).\\n\\n3. Error handling and Recovery: \\n\\n*We now briefly discuss and acknowledge this limitation in Sec 6.*\\n\\nIn the current scope of our work, we focus primarily on the core functionalities and performance evaluation of LiMAC. While comprehensive mechanisms for error recovery are essential, they were not the primary focus. We recognise the importance of robust error handling and recovery processes, and we intend to address these in future work. We also agree that safeguarding personal data is a concern, and while the current version of LiMAC does not specifically address this, we have included this as a direction of future research in the conclusion.\\n\\n[1] Rawles et al. AndroidWorld: A Dynamic Benchmarking Environment for Autonomous Agents.\\n\\n[2] Li et al. On the Effects of Data Scale on Computer Control Agents.\"}", "{\"comment\": \"Thanks for the detailed response, that answered most of my questions.\"}" ] }

[ "Pankaj Bhagwat", "Linglong Kong", "Bei Jiang" ]

[ "Bayesian framework", "Conformal prediction", "Model uncertainty", "Uncertainty quantification" ]

{ "note_id": [ "zE6QnugM2D", "wxbwRqQb1Y", "uxnvfpyvJ5", "q5kM9dNByu", "oaOY3RfZtx", "kyLGsNPFs2", "h1moNAB5rb", "ftM5aUBPm9", "fobi0D3NoB", "diYXKjXivL", "aekiA82xsp", "UffJ1ideOT", "PfhAiF1Oep", "P4dp2G6q5x", "JadU5uhjm4", "IoEWiA1NBY", "HtpI3mV68r", "HnkDVJO1ho", "5JESrT10x5", "35e6Ybtlul", "0KyEH3KYkC" ], "note_type": [ "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "meta_review", "official_review", "official_comment", "official_comment", "decision" ], "note_created": [ 1732789166800, 1732898341357, 1732691488186, 1732681551432, 1732775382698, 1732790126578, 1732811721381, 1732730422624, 1730529659224, 1732759972460, 1730659045260, 1732691214825, 1732778924153, 1732686020114, 1732684350338, 1730396927830, 1735005316088, 1730650146338, 1732692509810, 1732699834928, 1737523614486 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission4023/Reviewer_Nttv" ], [ "ICLR.cc/2025/Conference/Submission4023/Authors" ], [ "ICLR.cc/2025/Conference/Submission4023/Authors" ], [ "ICLR.cc/2025/Conference/Submission4023/Authors" ], [ "ICLR.cc/2025/Conference/Submission4023/Authors" ], [ "ICLR.cc/2025/Conference/Submission4023/Reviewer_bVQG" ], [ "ICLR.cc/2025/Conference/Submission4023/Authors" ], [ "ICLR.cc/2025/Conference/Submission4023/Authors" ], [ "ICLR.cc/2025/Conference/Submission4023/Reviewer_ctnD" ], [ "ICLR.cc/2025/Conference/Submission4023/Reviewer_3y9y" ], [ "ICLR.cc/2025/Conference/Submission4023/Reviewer_Nttv" ], [ "ICLR.cc/2025/Conference/Submission4023/Authors" ], [ "ICLR.cc/2025/Conference/Submission4023/Authors" ], [ "ICLR.cc/2025/Conference/Submission4023/Authors" ], [ "ICLR.cc/2025/Conference/Submission4023/Authors" ], [ "ICLR.cc/2025/Conference/Submission4023/Reviewer_bVQG" ], [ "ICLR.cc/2025/Conference/Submission4023/Area_Chair_WWrA" ], [ "ICLR.cc/2025/Conference/Submission4023/Reviewer_3y9y" ], [ "ICLR.cc/2025/Conference/Submission4023/Reviewer_3y9y" ], [ "ICLR.cc/2025/Conference/Submission4023/Reviewer_Nttv" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ] ], "structured_content_str": [ "{\"comment\": \"I thank the authors for the response, that are now satisfactory.\\nI have thus raised my evaluation from 5 to 6.\\nGood luck with the final selection!\"}", "{\"comment\": \"Thank you again for taking the time to evaluate our paper, providing constructive comments that have helped improve our work, and raising the score to 6, which means a lot to us.\"}", "{\"comment\": \"**Weakness 6:** Following your suggestion, we have added a real data example (California housing data) in the revised version to demonstrate the practical utility of our method and its performance compared to other baseline methods. Specifically, we compare our CBMA approach with the model aggregation strategy recently proposed by Gasparin and Ramdas (2024b), which combines conformal sets using a majority vote scheme. As expected\\u2014since their method focuses on efficient set aggregation rather than achieving efficient prediction sets\\u2014our CBMA intervals result in significantly shorter prediction intervals. These results are included in the revised version.\\n\\nAdditionally, in the updated Section 2.5, we provide an expanded review of existing model aggregation methods and their limitations. Common limitations of these methods include:\\n\\n1. **Data splitting requirements or reliance on hold-out calibration sets:** Many baseline methods require splitting the available data, which may not be practical for small sample sizes. In contrast, our CBMA approach utilizes all available data, ensuring no efficiency loss from data splitting.\\n\\n2. **Lack of theoretical guarantees on efficiency and coverage:** Existing methods often fail to provide rigorous guarantees for both efficiency and statistical coverage. Our CBMA framework, on the other hand, is the first to offer such theoretical guarantees on both optimal efficiency and coverage under potential model misspecification.\\n\\nWe believe these additions strengthen the paper and provide a more comprehensive comparison of CBMA with other approaches, further highlighting the advantages of our method.\\n\\n---\\n\\n**Question 1:** Are you referring to the posterior samples of the model parameters? If so, then yes, our method requires access to the posterior samples of the model parameters.\\n\\n---\\n\\n**Question 2:** We use the form assuming independence of noise from $X$ to simplify the review of the BMA framework. However, the BMA framework is fully capable of handling scenarios where noise depends on $X$. The independence assumption was introduced only for ease of explanation and does not reflect a limitation of the method itself. Recognizing that this might cause confusion, we have revised this part by general Bayesian model setup in the revised paper (lines 131-133).\\n\\n---\\n\\n**Question 3:** Thank you for pointing it out. We have added the missing reference below Equation (1).\\n\\n---\\n\\n**Question 4:** The statement in Section 2.4, ``We propose to average conformity scores from each model to construct a combined conformity score,'' corresponds to Equations (3) and (4), not Equation (2). On Page 3, equations are indeed conditional on $X$ (where $Z_{1:n} = { (X_i, Y_i) }_{i=1}^n$ as in Line 70), although we chose not to explicitly include this notation to maintain simplicity. We acknowledge that this could lead to some confusion, which is why we clarified definitions wherever necessary, e.g., for marginal likelihood on lines 159-160 in the original submission. To avoid any future ambiguity, we have added a statement in the revised paper explicitly addressing this point (line 161 in the revised paper).\\n\\n---\\n\\n**Question 5:** By a ``valid conformity score,'' we mean a conformity score function that is exchangeable in its first argument. This clarification has been explicitly stated in the revised paper (lines 84-85).\\n\\n---\\n\\n**Question 6:** Given the posterior samples and posterior model probabilities, computing the conformity scores for individual models involves a complexity of $\\\\mathcal{O}(n_{\\\\text{grid}} \\\\times \\\\text{posterior samples} \\\\times n_{\\\\text{train}})$, as outlined in Fong and Holmes (2021). The model weights can then be efficiently derived using these computed terms. Consequently, the aggregation step, which involves calculating the dot product of $K$ weights and $K$ conformity scores, incurs only a linear computational overhead of $\\\\mathcal{O}(K)$, where $K$ represents the number of models. \\n\\n---\\n\\n**Question 7:** Bayesian Model Averaging converges to the true model under certain conditions as $n \\\\rightarrow \\\\infty$. This convergence is driven by the combined influence of the likelihood, priors, and the Bayesian updating process, rather than by the likelihood alone. Together, these elements ensure that the posterior probability concentrates on the true model with sufficiently large sample size. \\n\\n---\\n\\n**Question 8:** Yes, to the best of our knowledge, our work is the first to explore Conformal Bayesian methods without the assumption that the underlying Bayesian model is exactly correct or true. Reviewer $3y9y$ has also acknowledged this contribution.\\n\\n---\\n\\n**References**\\n\\nEdwin Fong and Chris C Holmes. Conformal bayesian computation. Advances in Neural Information Processing Systems, 34:18268\\u201318279, 2021.\\n\\nMatteo Gasparin and Aaditya Ramdas. Merging uncertainty sets via majority vote. arXiv preprint arXiv:2401.09379, 2024b.\\n\\n---\"}", "{\"comment\": \"Thank you for your constructive feedback. Below, we provide point-by-point responses to address the weaknesses you noted.\\n\\n**Weakness 1:** You are correct that conformal prediction based on model averaging has gained attention, as the papers that you provided demonstrate how incorporating model averaging can help reduce the volume of prediction sets through experiments. However, none of these prior works provide a theoretical guarantee. Our work is **the first to establish such a guarantee**, specifically proving that the resulting conformal prediction intervals converge to the optimal level of efficiency.\\nAs Reviewer 3y9y noted, our approach is the first to combine ``multiple models'' in conformal framework through Bayesian model averaging, marking a significant and novel contribution. For this reason, we believe our contribution has been accurately stated.\\nThank you for providing the references. We have acknowledged these prior works on conformal model averaging in Section 2.5 and highlighted their limitations in the revised paper.\\n\\n**Weakness 2:** We appreciate your feedback. Our **simulations** were intentionally designed to **validate the theoretical guarantees** of our method and provide comparisons with other approaches in a simple yet illustrative manner, which allows readers to easily recognize that the candidate models are mis-specified, yet our proposed method consistently outperforms others by producing the shortest prediction intervals while maintaining coverage probabilities close to the nominal level. To further demonstrate the practical utility of our method, we have added a **real data example** in the revised version of the paper and showed that CBMA works effectively. Finally, to improve the presentation of results, we have added **plots** of comparisons for mean lengths and coverages for prediction intervals in our simulation studies. \\n\\n**Question 1:** Following your suggestion, we have thoroughly reviewed the relevant literature, including your recommended papers, Linusson et al. (2020) and Yang and Kuchibhotla (2024), as well as several others in Section 2.5. We have expanded our discussion to provide a **more comprehensive discussion of existing methods**, highlighting their limitations and how they compare to and differ from our approach. We believe this expanded discussion addresses your concerns and clarifies the unique contributions of our work in relation to the existing literature.\\n\\n**Question 2:** We appreciate your thoughtful feedback and understand your interest in exploring more challenging ``misspecification'' cases, such as the AR2 model. However, conducting such experiments may not be necessary in this context, as it is well-known in the literature on Bayesian Model Averaging (BMA) for time series data that if all candidate models fail to account for temporal dependence, BMA will lead to biased estimates and underestimated uncertainty. In other words, efficiency gains are only achievable when the misspecified models still capture temporal dependence to some extent. However, models with temporal dependence structure violate the exchangeability condition required by our method. \\nTo address your concern that our method may have limited applicative interest, we have **added new experiment** to evaluate the performance of our approach on real dataset, where the true model is unknown. This represents a fair assessment of our method's effectiveness, as it mirrors real-world scenarios where all methods may potentially mis-specify the true model. These new results are included in the revised paper.\\n\\n**Question 3:** We appreciate your suggestion, as it offers an insightful way to evaluate whether the prediction intervals produced by our method are likely the shortest among all methods, when not knowing the population truths. Specifically, Scheff\\u00e9's method tests whether the expected lengths of prediction intervals (i.e., the population truths) differ across methods.\\nHowever, we have already demonstrated in Remark 1 of the paper that the expected interval lengths are theoretically guaranteed to be the shortest under our method. Since the population truth is already known in this context, there is no need to rely on hypothesis testing, which is inherently subject to type I and type II errors and could lead to incorrect conclusions about the population. In conclusion, our **theoretical guarantee makes additional hypothesis testing unnecessary.**\"}", "{\"title\": \"Thank you\", \"comment\": \"Thank you for your positive remarks on the theoretical efficiency properties of CBMA and the additional numerical experiments we conducted to demonstrate its strengths. Your decision to raise your score from 5 to 6 means a great deal to us. We sincerely thank you for taking the time to evaluate our paper and for providing constructive comments that have helped improve our work.\"}", "{\"title\": \"many thanks for the detailed rebuttal\", \"comment\": \"I thank the authors for their extensive and clarifying replies and for adding new experiments to the paper. I will rise my score to 6.\"}", "{\"comment\": \"We sincerely appreciate your thoughtful feedback and are glad that our responses addressed your concerns satisfactorily. We are also very grateful for your decision to raise your evaluation score from 5 to 6. Thank you for your best wishes, and we hope for a positive outcome in the final selection process!\"}", "{\"comment\": \"Thank you for bringing up the additional reference. However, we must clarify that the work by Liang et al. (2024) is dealing with a completely different goal than ours. Liang et al. (2024) aims to **``constructing a finite sample valid conformal prediction set while selecting a model that minimizes the size of the prediction set, from a given family of pre-trained models\\u201d**. In other words, their efficiency guarantee relies on knowing the true model.\\n\\nIn contrast, our work aims to provide optimally efficient conformal prediction sets without requiring knowledge of the true model. Our efficiency guarantee holds with potential model misspecification. Additionally, while methods such as Yang & Kuchibhotla (2024) suffer from selection bias and therefore lack statistical validity for the resulting prediction sets, Liang et al. (2024) seeks to correct such biases to achieve validity. In comparison, our method inherently achieves finite-sample validity through a probabilistic aggregation approach based on Bayesian Model Averaging (BMA), avoiding the need for bias correction entirely.\"}", "{\"summary\": \"This paper proposes a solution in the form of Conformal\\nBayesian model averaging (CBMA) that combines Bayesian Model Averaging (BMA) with conformal prediction to address the potential issue of suboptimal prediction intervals when the Bayesian model itself is misspecified. The theoretical and empirical results show the efficiency of CBMA.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"The results presented in the paper confirm the effectiveness of the proposed method.\", \"weaknesses\": \"1.The novelty of the work is not very significant, as it simply combines several existing methods.\\n\\n2.The names of the subsections under the \\\"Simulation\\\" section are not entirely appropriate, since quadratic model is a type of non-linear model. It may be more appropriate to title whether the model is polynomial.\\n\\n3.The paper seems not explain why the results for model 10 and model 11 are missing when $n=50$ in Table 2.\\n\\n4.The paper contains a few typographical errors. For example, the last sentence before Section 2 has a typo. Additionally, there is a missing equation reference on line 150.\", \"questions\": \"1.Is it a clear pattern in the paper for when the authors use the term \\\"prediction set\\\" versus \\\"prediction interval\\\"?\\n\\n2.Is there a comparison of the time costs between CBMA and other methods?\\n\\n3.Why are the values of $\\\\alpha$ set as well as the ratio of the size of training set to test set differently in Table 1 and Table 2? It would be better to unify them or display both values or provide a rationale for why they differ, as this would improve the clarity and comparability of the results.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you for addressing my concerns. You pointed out the coverage loss problem in Yang & Kuchibhotla\\u2019s (2024) work. I would like to highlight that Liang et al. (2024) addressed this issue by restoring valid prediction sets, demonstrating that methods for selection and aggregation across multiple models to construct valid prediction sets have been actively studied.\\n\\nHowever, the theoretical properties of CBMA\\u2019s efficiency, as well as the broader additional numerical experiments you conducted, effectively showcase the strengths of CBMA. Based on this, I am raising my score from 5 to 6.\"}", "{\"summary\": \"The authors, working in the niche of bayesian conformal prediction, propose a bayesian model averaging procedure \\\"infused\\\" with conformal prediction, in order to recover coverage in the case of model misspecification\", \"soundness\": \"3\", \"presentation\": \"4\", \"contribution\": \"2\", \"strengths\": \"I found the article quite well written.\\nMoreover, the proposal is of clear theoretical and applied interest, and developed with care and mathematical soundness.\", \"weaknesses\": \"I fear that the authors have oversold their claim. In fact, in the literature proposals of conformal model averaging can be found (https://www.sciencedirect.com/science/article/pii/S0925231219316108, https://arxiv.org/abs/2408.06642#:~:text=Unlike%20traditional%20methods%2C%20conformal%20ensembling,%2Dto%2Dinterpret%20uncertainty%20estimates. to name a few).\\n\\nI found the simulation quite lacking in terms of depth, as well and presentation of the results\", \"questions\": [\"-As shown in the \\\"weaknessess\\\", some proposals (usually in a frequentist setting) are present in the literature when it comes to model averaging/ensembling in a conformal sense. Can you comment on this, maybe providing a more thorough analysis of the outstanding literature?\", \"I find the cases provided by the authors of limited applicative interest, I wonder if it could be possible to provide more challenging \\\"misspecification\\\" cases, like for instance cases where the model misspecification causes the conditional iid assumption to break (e.g. an AR2 model with only one lag).\", \"The impact in terms of interval length does not seem statistically significant. Are the authors able to check this, using maybe the Scheffe' method?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you for your constructive feedback and for identifying areas where our manuscript can be improved. Below, we provide point-by-point responses to address the concerns you raised. However, we would like to clarify that many of the points you raised are not inherent weaknesses of our method but rather questions or requests for clarification, which we have addressed. We believe it would be unfair to frame these questions as weaknesses to diminish the value of our contribution.\\n\\n\\n**Weakness 1:** We respectfully disagree with your criticism that the novelty of our work is limited to simply combining existing methods. While our framework does build upon the well-established Bayesian Model Averaging to enhance conformal prediction, this integration represents a significant and novel contribution, as recognized and highlighted by Reviewer 3y9y. Specifically, our work is the **first to provide theoretical guarantees** that ensure the resulting conformal prediction sets are optimally efficient, achieving the shortest expected prediction intervals under model misspecification. \\n\\n---\\n\\n\\n**Weakness 2:** Yes, the Bayesian conformity score is needed in our CBMA framework. As we discussed in the paper (lines 184-186), the Bayesian posterior predictive density is the optimal choice for the conformity score. Accordingly, we constructed our framework around this score. This choice enables us to derive the efficiency guarantee in our Theorem $2$, which ensures that the prediction sets produced by our framework are optimal in terms of their expected size (minimized prediction interval length) while maintaining the required coverage probability. \\n\\n---\\n\\n**Weakness 3:** As outlined in Equation (7) of the paper, averaging the Bayesian conformity scores associated with different models is mathematically equivalent to computing the Bayesian average of the conformity scores from individual models. Since these two approaches yield the same result within our framework, we are unsure about the distinction implied in your question. We would appreciate further clarification to address your concern effectively.\\n\\n---\\n\\n**Weakness 4.** While our CBMA framework relies on the standard assumption of infinite sample sizes to establish theoretical guarantees for efficiency and coverage, we emphasize that this is a well-established practice in Bayesian model averaging literature for deriving rigorous theoretical properties. This approach is necessary because Bayesian inference does not inherently align with the frequentist notion of repeated sampling, where guarantees are evaluated under repetitions of the experiment. Instead, Bayesian methods focus on updating beliefs about parameters given observed data. The infinite sample size assumption provides a bridge for deriving frequentist-style guarantees, such as efficiency and coverage, in an asymptotic framework.\\n\\nHowever, we demonstrate empirically that CBMA performs effectively even with finite sample sizes. In particular to address your concern, we added Figure $1$ showing the convergence of the mean CBMA prediction set sizes to the optimal prediction set sizes (corresponding to the true model). Convergence is observed with $n=200$, indicating that CBMA achieves efficiency with a finite sample size. Additionally, we included a real data example (California housing data) in the revised Section 5.3. This example shows that CBMA produces shorter prediction intervals than other baseline methods at $n=150$. Notably, Bayesian credible intervals tend to over-cover the data and result in wider, less efficient prediction sets at similar sample sizes.\\n\\n*Case of approximately correct models:* As demonstrated in another experiment, even when the true model is not included in the model space, CBMA demonstrates robustness against such model specification. In such cases, theoretical analysis suggests that CBMA converges to the model closest to the true model in terms of Kullback-Leibler (KL) divergence. This ensures that **CBMA maintains its coverage guarantee and produces near-optimal prediction sets** if a model exists within the KL divergence neighborhood of the true model. We have added detailed discuss of this observation in the revised version.\\n\\nIn summary, while our theoretical results assume infinite sample sizes, the empirical results confirm that CBMA remains highly effective in finite-sample scenarios, offering both robust coverage and efficiency advantages over Bayesian intervals.\\n\\n---\\n\\n**Weakness 5:** Thank you for the suggestion. There is no Theorem 3 in the paper, but we believe you are referring to Theorem 2. In the revised paper, we have updated theoretical result (Theorem 2) that integrates original Theorem 2 and Remark 1, presenting a unified result for the optimality of prediction sets. As you suggested, this enhancement aligns with the goals of the paper and provides a more comprehensive theoretical foundation. \\n\\n---\"}", "{\"title\": \"Thank you for your follow up questions.\", \"comment\": \"**Revisit Question 1**\\n\\nThank you for your feedback. As noted in the manuscript and our responses, we conducted a thorough review of relevant works and highlighted either their distinct goals or their lack of theoretical guarantees compared to ours. We also appreciate Reviewer 3y9y pointing out another recent unpublished paper by Liang et al. (2024; https://arxiv.org/abs/2408.07066). However, as we clarified to Reviewer 3y9y, Liang et al. (2024) addresses a completely different goal. Their work focuses on constructing a finite-sample valid conformal prediction set while selecting a model that minimizes the size of the prediction set from a given family of pre-trained models. In other words, their efficiency guarantee relies on knowing the true model.\\n\\nIn contrast, our work provides optimally efficient conformal prediction sets without requiring knowledge of the true model. Our efficiency guarantee holds even under potential model misspecification. This critical distinction was acknowledged by Reviewer 3y9y, who raised their score from 5 to 6, recognizing the unique strengths and contributions of our CBMA framework.\\n\\n**Revisit Question 3**\\n\\nAs noted in our responses, we appreciate your suggestion on using hypothesis testing to determine whether the prediction sets obtained using our CBMA framework is the shortest with finite sample size. However, such an approach does suffer from type I and type II error, like all hypothesis tests do. We do acknowledge that our theoretical guarantees on minimal size are asymptotic; in our original submission, we already empirically assessed the finite-sample performance of our CBMA framework in the original submission. These results demonstrate that CBMA performs effectively in finite samples, producing shorter prediction intervals than baseline methods while maintaining valid coverage. \\n\\nThat said, we have taken your suggestion seriously and expanded our empirical evaluation in the revised paper to include both additional simulations and a newly added real-world example. Instead of focusing on hypothesis testing, we have added figures illustrating the convergence of the expected sizes of CBMA prediction sets to the optimal prediction set sizes (corresponding to the true model). In summary, convergence is observed at $n=200$ in simulations and $n=150$ in the real-world example, demonstrating that CBMA achieves optimal efficiency with finite sample sizes.\\n\\nLastly, while your request regarding the rate of convergence is beyond the primary scope of our work, we acknowledge its importance. To address this concern, we note that Equation (7) suggests that the rate of convergence corresponds to the convergence rate of posterior model probabilities, a well-studied topic in Bayesian literature (see, for instance, Rossell, 2022). \\n\\nWe hope this additional clarification and empirical evidence address your concerns.\\n\\n\\n**Reference:**\\n\\nRossell, D. (2022). Concentration of posterior model probabilities and normalized $L_0$ criteria. Bayesian Analysis, 17(2), 565-591.\"}", "{\"comment\": \"Thank you for your constructive feedback and for highlighting areas where our manuscript can be improved. Below, we provide point-by-point responses to address the weaknesses you noted.\\n\\n**Weakness 1:** We respectfully disagree with your criticism that the novelty of our work is limited to simply combining existing methods. While our framework does build upon the well-established Bayesian Model Averaging to enhance conformal prediction, this integration represents a significant and novel contribution, as recognized and highlighted by Reviewer 3y9y. Specifically, our work is the **first to provide theoretical guarantees** that ensure the resulting conformal prediction sets are optimally efficient, achieving the shortest expected prediction intervals under model misspecification. \\n\\n---\\n\\n**Weakness 2:** Thank you for the suggestion. We have renamed subsection 5.2 to *``Approximation using Hermite Polynomials\\\"* for better clarity.\\n\\n---\\n\\n**Weakness 3:** The number of models used follows the setting $ K = \\\\lfloor 3 n_{\\\\text{train}}^{1/3} \\\\rfloor $, as mentioned in Line 425. For $ n = 50, n_{\\\\text{train}} = 30 $, resulting in $ K = 9$. These settings align with those in Lu \\\\& Su (2015). But, to avoid the confusion and for the sake of completeness, we have added results for model 10 and model 11 for sample size $n = 50$ in the revised version.\\n\\n---\\n\\n**Weakness 4:** We appreciate the feedback and have corrected the typographical errors, including the missing equation reference and typos before Section 2. \\n\\n---\\n\\n**Question 1:** We have standardized the term *``prediction set\\\"* throughout the paper. While the prediction sets in our experiments are intervals, the theoretical results apply to prediction sets in general.\\n\\n---\\n\\n**Question 2:** We will include execution time details in our simulation experiments to construct various prediction sets (conformal Bayes, Bayes prediction, BMA prediction and our CBMA ) in the Appendix where we provide additional details of our experiments. For the real data example, we will discuss such execution times in the main text to provide times costs between these different approaches. \\n\\n\\n---\\n\\n**Question 3:** Following your suggestions, we have added the new experiments in Section 5.1 in the revised paper, with consistent experimental settings with other experiments ($\\\\alpha = 0.2$, train-test split of $60-40\\\\%$), which will improve the clarity and allow comparisons across experiments. \\n\\n---\\n\\n**References:**\\n\\nXun Lu and Liangjun Su. Jackknife model averaging for quantile regressions. Journal of Econometrics, 188(1):40\\u201358, 2015.\"}", "{\"comment\": \"Thank you for acknowledging the contribution of our work. Below we provide answers to address your specific concerns.\\n\\n**Weaknesses:** Thank you for your thoughtful feedback regarding comparisons with existing frequentist approaches for model aggregation in conformal prediction. Based on your suggestions, we have revised our paper to include an **expanded literature review of these frequentist approaches** and a detailed discussion of the distinct contributions and scope of our Bayesian Model Averaging (BMA) framework within the conformal prediction literature.\\n\\nWe emphasize that **frequentist methods** often focus on effective set aggregation or selection but **lack theoretical guarantees** of optimal efficiency and valid statistical coverage. Our CBMA framework, on the other hand, directly addresses these gaps by providing theoretical guarantees for both optimal efficiency and valid statistical coverage in the resulting prediction sets under potential model misspecification.\\n\\nFor example, in the context of transductive conformal methods (also known as full conformal methods, which make full use of the data for prediction), Yang and Kuchibhotla (2024) provide selection rules to choose the smallest valid conformal prediction sets. However, their methods may suffer from coverage loss and lack theoretical guarantees for achieving statistical efficiency in the selected prediction set. In the context of inductive conformal methods (also known as split conformal methods, which require splitting the available data into training and calibration sets, reducing efficiency), prior works such as Vovk (2015), Linusson et al. (2017), Carlsson et al. (2014), Toccaceli and Gammerman (2019) propose methods that combine p-values generated by different models or data splits to create aggregated predictions. However, these combined p-values may not be well-calibrated, leading to prediction sets that do not accurately represent the intended coverage level. This lack of calibration undermines the reliability of the aggregated predictions.\\n\\nGiven these fundamental differences, direct comparisons between CBMA and frequentist model aggregation methods are not meaningful, as they lack rigorous guarantees of efficiency and coverage. For instance, we compared the lengths of CBMA aggregated intervals with those of conformal sets aggregated using the majority vote strategy (applying Corollary 4.1 from Gasparin and Ramdas (2024b) due to dependency among individual sets) in our real data example. The observed mean ratios of interval lengths constructed using majority vote scheme and our CBMA approach for $n = 50, 100, 150$ are $1.0765 (0.0253)$, $1.1196 (0.0151)$, and $1.1458 (0.0133)$, respectively, indicating that CBMA leads too shorter intervals. These results have been included in the revised paper's real data example section.\\n\\n---\\n\\n**Question 1:** We have revised Section 2.5 to provide a **more detailed comparison** of our method with existing approaches, highlighting their contributions, limitations, and how CBMA addresses these gaps. This **expanded discussion** clarifies the unique contributions of our work relative to the existing literature. Additionally, to address your concern about the weakness of our paper above, we have elaborated on why existing frequentist methods are not directly comparable to our approach, given the fundamental differences in objectives and guarantees.\\n\\n---\\n\\n**Question 2:** Following your suggestions, we have added the new experiments in Section 5.1 in the revised paper, with **consistent experimental settings** with other experiments ($\\\\alpha = 0.2$, train-test split of $60-40$), which will improve the clarity and allow comparisons across experiments. \\n\\n---\\n\\n**References:**\\n\\nLars Carlsson, Martin Eklund, and Ulf Norinder. Aggregated conformal prediction. In Artificial Intelligence Applications and Innovations: AIAI 2014 Workshops: CoPA, MHDW, IIVC, and MT4BD, Rhodes, Greece, September 19-21, 2014. Proceedings 10, pp. 231\\u2013240. Springer, 2014.\\n\\nMatteo Gasparin and Aaditya Ramdas. Conformal online model aggregation. arXiv preprint\", \"arxiv\": \"2403.15527, 2024a.\\n\\n\\nMatteo Gasparin and Aaditya Ramdas. Merging uncertainty sets via majority vote. arXiv preprint arXiv:2401.09379, 2024b.\\n\\nHenrik Linusson, Ulf Norinder, Henrik Bostr\\u00a8om, Ulf Johansson, and Tuve L\\u00a8ofstr\\u00a8om. On the cali-\\nbration of aggregated conformal predictors. In Proceedings of the Sixth Workshop on Conformal and Probabilistic Prediction and Applications, volume 60 of Proceedings of Machine Learning Research, pp. 154\\u2013173. PMLR, 13\\u201316 Jun 2017.\\n\\nVladimir Vovk. Cross-conformal predictors. Annals of Mathematics and Artificial Intelligence, 74: 9\\u201328, 2015.\\n\\nYachong Yang and Arun Kumar Kuchibhotla. Selection and aggregation of conformal prediction sets. Journal of the American Statistical Association, pp. 1\\u201313, 2024. doi: 10.1080/01621459.\\n2024.2344700.\"}", "{\"summary\": \"The paper leverages Conformal Bayesian and Bayesian Model Averaging to reduce the size of Conformal Prediction sets. The Bayesian posterior of different models is used in a weighted average of Bayesian conformity scores, which produces optimal prediction sets when the average contains the exact model and the sample size is infinite. Unlike previous Conformal Bayesian methods, the proposed scheme allows possible misspecifications in the underlying Bayesian model.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"1\", \"strengths\": [\"Conformal Bayesian can combine the advantages of frequentist and probabilistic approaches.\", \"Theoretical results on the optimal efficiency of the prediction sets are rare and challenging to obtain in the CP framework.\", \"Model averaging and ensemble methods perform well in various setups. Studying CP in that setup may produce powerful uncertainty quantification tools.\"], \"weaknesses\": [\"The technical contribution is to combine existing techniques: the BMA and Conformal Bayes.\", \"It is unclear whether using the Bayesian conformity score is needed or if the method applies to any conformity score.\", \"The authors should clarify why averaging the (Bayesian) conformity scores associated with different models is better than computing the conformity score of a Bayesian average of the models.\", \"The main theoretical result holds in the limit of an infinite sample size. In that limit, Bayesian confidence intervals are also correct. Why would one need CP? Including the exact model in the average may look unrealistic. Does the result generalize to models that are only approximately correct?\", \"Combining the results of Theorem 3 with Remark 1 to have a direct result on the optimality of the prediction sets would align better with the paper's goals.\", \"The model has not been tested on real data. For example, the authors could show that the scheme produces smaller intervals than standard CP or any other baseline.\"], \"questions\": [\"Is accessing the model parameters required?\", \"Is the independence of the noise from X a strict assumption for the method applicability?\", \"Missing reference number below Eq 1.\", \"In Section 2.4, you say \\\"We propose to average conformity scores from each model to construct combined conformity score.\\\" Is this equivalent to Eq.2? Why don't you condition on X in the equations on page 3?\", \"What do you mean by 'valid conformity score'?\", \"How expansive is computing the prediction set?\", \"Does BMA converge to the true model if $n \\\\to \\\\infty$ because its likelihood increases?\", \"Is this the first work on Conformal Bayesian that does not assume the underlying Bayesian model is exact?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"details_of_ethics_concerns\": \"NA\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"metareview\": \"The formal novelties of the paper lie in its integration of Bayesian Model Averaging with conformal prediction to address model misspecification and improve prediction set efficiency. By aggregating conformity scores from multiple Bayesian models using posterior model probabilities, the proposed Conformal Bayesian Model Averaging ensures valid frequentist coverage while optimizing the expected volume of prediction sets. The method theoretically guarantees convergence to optimal prediction sets as sample size increases, when the true model is in the model space (which apparently holds even under potential model misspecification).\\n\\nIt is to be noted that Theorem 2 (consequence of Le & Clarke (2022) is not distribution-free). I also did not understand the claim when the model is not well specified. The Bayesian Posterior Convergence is that the Posterior probabilities of the true model (if it exists in the candidate set) converge to 1 as the sample size (ok then the quantiles estimates converges to the one of the ground-truth and the convergence of the size). If the true model is not present, posterior probabilities concentrate on the candidate model closest to the true model in terms of (KL) divergence. In the later case, The size of the CBMA prediction set may be larger than the optimal size that would be achieved if the true model were included. Equation 9 in the proof is not established in this case (BTW, the overall proof should be improved). \\n\\n> In contrast, our work provides optimally efficient conformal prediction sets without requiring knowledge of the true model. Our efficiency guarantee holds even under potential model misspecification. \\n\\nI could not see any proof of such statement in the paper. When the model is not well specified, the consistency results does not hold and this claim is misleading. Furthermore, what the authors want to claim seems quite straightforward: *any consistent estimation of the ground-truth distribution would produce asymptotically efficient uncertainty set*. This has fairly much nothing to do with model averaging but only consistency. \\n\\nAlong with lack of comparisons to existing methods, the proposed benchmark (e.g. the main Table 1) are quite unconvincing: all the methods achieve very similar size (one needs to look at 3rd digit to notice any difference between the lengths and this is even not significant given the standard deviation).\\n\\n----\\n*After further discussions with the SAC, we suggest an Accept, leaving the final decision to the program chairs to potentially bump down. All reviewers raised their scores to 6 (\\\"Marginally above the acceptance threshold\\\") after acknowledging the authors' efforts to address concerns and improve the manuscript. While some issues remain, the paper has merits and could provide valuable insights to the community.*\", \"additional_comments_on_reviewer_discussion\": \"The discussions with reviewers highlighted both the strengths and limitations of the proposed CBMA framework. Reviewers appreciated the theoretical rigor and the novel integration of Bayesian Model Averaging (BMA) with conformal prediction, providing guarantees of efficiency and valid coverage even under model misspecification. However, concerns were raised about the incremental nature of the contribution, limited empirical evaluation, and lack of direct comparisons with existing frequentist and Bayesian methods. Specific issues included insufficiently challenging misspecification scenarios, unclear advantages over simpler aggregation strategies, and reliance on strong assumptions, such as the presence of the true model in the candidate set. The authors addressed these by expanding the empirical evaluations, including real-world datasets, improving clarity in theoretical arguments, and elaborating on distinctions from prior work. While these efforts satisfied some reviewers, leading to score increases, others maintained reservations about the method\\u2019s practical applicability and incremental contributions.\\n\\n> Set $\\\\alpha=0.1$ or smaller as it is done in common in the literature.\"}", "{\"summary\": \"The authors highlight that the efficiency of traditional conformal prediction can degrade in the presence of model uncertainty. To address this, they propose Conformal Bayesian Model Averaging (CBMA), which combines multiple non-conformity scores from a Bayesian perspective. This approach integrates Bayesian model averaging with Bayesian conformal prediction (full conformal Bayes; Fong & Holmes, 2021), allowing it to leverage the optimality of Bayesian model averaging from a theoretical standpoint. From a computational perspective, CBMA utilizes the add-one-in importance sampling approach of Fong & Holmes (2021), enabling it to use all available data for model training while avoiding exhaustive computation. The work demonstrates CBMA\\u2019s high efficiency through numerical justification in both correctly specified and misspecified model scenarios.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"While several works in the field of conformal prediction have applied a Bayesian perspective, this paper is meaningful since it is the first to attempt combining \\u201cmultiple models\\u201d through Bayesian model averaging, marking a novel contribution. The proposed framework is straightforward, integrating traditional Bayesian model averaging directly, yet it is powerful as it incorporates model uncertainty more fully into the framework.\", \"weaknesses\": \"This work is limited by a lack of discussion comparing its approach to existing model aggregation methods in conformal prediction including frequentist\\u2019s perspective. Although Section 2.5 touches on related works, the primary advantage highlighted is the ability to use the full data, which may be better understood as a contribution from Fong & Holmes (2021) rather than an original novelty here. To justify the method empirically, it would be more reasonable to benchmark a few of these existing methods, comparing their coverage, interval length, and execution time with suggested method.\", \"questions\": \"1. This question relates to the previously mentioned points. What do you consider the distinguishing features of your approach compared to existing conformal set aggregation methods? Were any experiments conducted to compare them? If so, what were the results, and if not, what are your thoughts on this?\\n\\n 2. In the simulations, the nominal value of $\\\\alpha$ was set to 0.1 in the linear model scenario, but 0.2 in the nonlinear case. Is there a particular reason for these distinct choices? Also, did you observe any systematic performance changes in the model based on the value of $\\\\alpha$?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you for conducting the additional numerical experiments as requested! These have helped address some of the uncertainties. However, I still remain hesitant about the claim that other methods with the same objectives lack theoretical guarantees. In particular, Liang et al. (2024; https://arxiv.org/abs/2408.07066) specifically address the coverage loss highlighted in Yang & Kuchibhotla (2024). As a result, I am still unconvinced regarding the reliance on multiple models and the computational demands of methods such as MCMC, or other Bayesian computations.\"}", "{\"comment\": \"I thank the authors for the kind reply, but while I appreciate the effort, I remain unconvinced on some issues, more specifically Weakness 1 and Question 3\\n\\nWith respect to Question 1, as shown by other reviewers, some methods do have theoretical guarantees, so I believe a more proper contextualisation of contributions should be in order.\\n\\nWRT to Question 3, your theoretical guarantees on minimal size are asymptotic, and nothing is stated on the rate of convergence nor on finite sample properties... I was thus proposing tests to assess, in finite samples, what is the behaviour of your aggregation procedure and, e.g., what is the minimum sample size after which your proposal starts to yield statistically significant results.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}" ] }

[ "Xinyu Wang", "Yanzheng Xiang", "Lin Gui", "Yulan He" ]

[ "retrieval", "LLM", "graph", "dynamic", "summary", "attention", "KV cache", "QA" ]

{ "note_id": [ "zEkTKORYqT", "wIxKMJKpVP", "vil8GQamU4", "rDhpgM6Cdo", "qPAvNzDXou", "mRqswPeU6F", "lnxHbAdGMa", "hmQdNrn9rt", "dpe9WmR7UN", "azqvn0QQLx", "Zog1okSOhB", "Z4I46cuekK", "XhKzUXVcOd", "XbOyKzIrYO", "VcfqXK98yO", "NS83w8vp3r", "Lotoyh9lrC", "IbabogV3di", "HKbLp8gWjw", "EpQVA5UGet", "Eon9njGxo5", "Ed11Rh6UeJ", "EN2oyavxQA", "DC0S1WTMsY", "CWNrGIwqHP", "9WxBzZzgwt", "8JeMIy6v3v", "7NRgV9qZx7", "7F2YjgGv2k", "5DdFHj4zsP", "4ghmftheat", "4Dn1hshm7w" ], "note_type": [ "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "meta_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_review" ], "note_created": [ 1732565622079, 1732421292972, 1731852352390, 1732565083631, 1732390698568, 1731857301919, 1731858896580, 1734517422668, 1732390650590, 1732390337194, 1732528930959, 1731852546146, 1731857134908, 1732473480821, 1731856429766, 1731859164742, 1731857184116, 1737524037490, 1731856773085, 1732432729154, 1731858835385, 1732565212330, 1731857050313, 1732497151949, 1730545193557, 1730354661228, 1731859119314, 1730200089560, 1731856977085, 1731856187262, 1732390721065, 1730488563901 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission10265/Authors" ], [ "ICLR.cc/2025/Conference/Submission10265/Reviewer_mYAX" ], [ "ICLR.cc/2025/Conference/Submission10265/Authors" ], [ "ICLR.cc/2025/Conference/Submission10265/Authors" ], [ "ICLR.cc/2025/Conference/Submission10265/Authors" ], [ "ICLR.cc/2025/Conference/Submission10265/Authors" ], [ "ICLR.cc/2025/Conference/Submission10265/Authors" ], [ "ICLR.cc/2025/Conference/Submission10265/Area_Chair_KDsb" ], [ "ICLR.cc/2025/Conference/Submission10265/Authors" ], [ "ICLR.cc/2025/Conference/Submission10265/Authors" ], [ "ICLR.cc/2025/Conference/Submission10265/Reviewer_kpAv" ], [ "ICLR.cc/2025/Conference/Submission10265/Authors" ], [ "ICLR.cc/2025/Conference/Submission10265/Authors" ], [ "ICLR.cc/2025/Conference/Submission10265/Authors" ], [ "ICLR.cc/2025/Conference/Submission10265/Authors" ], [ "ICLR.cc/2025/Conference/Submission10265/Authors" ], [ "ICLR.cc/2025/Conference/Submission10265/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission10265/Authors" ], [ "ICLR.cc/2025/Conference/Submission10265/Reviewer_kpAv" ], [ "ICLR.cc/2025/Conference/Submission10265/Authors" ], [ "ICLR.cc/2025/Conference/Submission10265/Authors" ], [ "ICLR.cc/2025/Conference/Submission10265/Authors" ], [ "ICLR.cc/2025/Conference/Submission10265/Reviewer_cbqx" ], [ "ICLR.cc/2025/Conference/Submission10265/Reviewer_kpAv" ], [ "ICLR.cc/2025/Conference/Submission10265/Reviewer_mYAX" ], [ "ICLR.cc/2025/Conference/Submission10265/Authors" ], [ "ICLR.cc/2025/Conference/Submission10265/Reviewer_LAnf" ], [ "ICLR.cc/2025/Conference/Submission10265/Authors" ], [ "ICLR.cc/2025/Conference/Submission10265/Authors" ], [ "ICLR.cc/2025/Conference/Submission10265/Authors" ], [ "ICLR.cc/2025/Conference/Submission10265/Reviewer_cbqx" ] ], "structured_content_str": [ "{\"title\": \"General Response\", \"comment\": \"We sincerely thank the reviewers for their thoughtful feedback and the time they dedicated to evaluating our work.\\n\\nWe added the Top-X results for RAPTOR. We applied the Top-X setting to the RAPTOR Collapsed Tree variant, extracting Top-20, Top-42, Top-12, and Top-22 nodes for NarrativeQA, Qasper, HotpotQA, and MuSiQue, respectively, to match the TFLOPs of our method, consistent with other Top-X settings. The Top-X setting is to adjust the retrieved nodes to match the TFLOPs of baselines and our methods. All changes have been incorporated into the revised version and are highlighted in blue.\\n\\nIn the Top-X setting, we use our method to determine the nodes required for baselines, while in practice, baselines have to treat the number of extracted nodes as a hyperparameter. In contrast, our method automatically determines the number of nodes in each run using Dynamic Progress Control.\\n\\nTo emphasize the Top-X results and better illustrate the findings, we reordered Table 1 to list all Top-X results together.\\n\\n| NarrativeQA | F1 | ROUGE-L | BLEU-4 | TFLOPs | Ratio |\\n|---|---|---|---|---|---|\\n|Llama3.1 | 53.7 | 52.6 | 10.4 | 3361.9 | 108.45x |\\n|MeMWalker | 11.2 | 9.8 | 2.6 | 353.8 | 11.41x |\\n|BM25 Top-X | 53.7 | 52.9 | 14.0 | 37.5 | 1.21x |\\n|SBERT Top-X | 39.5 | 38.8 | 7.3 | 37.5 | 1.21x |\\n|Dragon Top-X | 55.1 | 54.2 | 13.6 | 37.5 | 1.21x |\\n|RAPTOR Top-X | 52.0 | 51.2 | 11.8 | 35.1 | 1.13x |\\n|GARLIC | 61.1 | 60.2 | 18.6 | 31.0 | 1.00x |\\n\\n| Qasper | F1 | ROUGE-L | BLEU-4 | TFLOPs | Ratio |\\n|---|---|---|---|---|---|\\n|Llama3.1 | 49.4 | 47.6 | 26.9 | 92.5 | 1.38x |\\n|MeMWalker | 39.0 | 36.8 | 17.4 | 123.9 | 1.85x |\\n|BM25 Top-X | 47.0 | 45.1 | 22.8 | 69.3 | 1.04x |\\n|SBERT Top-X | 46.6 | 44.5 | 23.3 | 68.9 | 1.03x |\\n|Dragon Top-X | 46.9 | 44.8 | 22.1 | 67.0 | 1.00x |\\n|RAPTOR Top-X | 46.9 | 44.7 | 20.8 | 67.3 | 1.01x |\\n|GARLIC | 49.7 | 47.9 | 27.0 | 66.9 | 1.00x |\\n\\n| HotpotQA | F1 | ROUGE-L | BLEU-4 | TFLOPs | Ratio |\\n|---|---|---|---|---|---|\\n|Llama3.1 | 41.3 | 41.2 | 6.3 | 23.7 | 1.48x |\\n|MeMWalker | 39.7 | 38.9 | 13.9 | 93.4 | 5.84x |\\n|BM25 Top-X | 40.7 | 40.8 | 7.7 | 20.0 | 1.25x |\\n|SBERT Top-X | 40.8 | 40.7 | 7.5 | 19.6 | 1.23x |\\n|Dragon Top-X | 39.2 | 39.1 | 6.7 | 20.6 | 1.29x |\\n|RAPTOR Top-X | 40.7 | 40.7 | 7.2 | 17.9 | 1.12x |\\n|GARLIC | 43.5 | 43.5 | 7.2 | 16.0 | 1.00x |\\n\\n| MuSiQue | F1 | ROUGE-L | BLEU-4 | TFLOPs | Ratio |\\n|---|---|---|---|---|---|\\n|Llama3.1 | 35.8 | 35.7 | 5.6 | 40.6 | 1.31x |\\n|MeMWalker | 24.0 | 23.5 | 9.9 | 175.7 | 5.69x |\\n|BM25 Top-X | 31.8 | 31.7 | 5.6 | 35.6 | 1.15x |\\n|SBERT Top-X | 32.5 | 32.5 | 6.4 | 35.6 | 1.15x |\\n|Dragon Top-X | 30.2 | 30.1 | 6.0 | 38.0 | 1.23x |\\n|RAPTOR Top-X | 35.4 | 35.2 | 7.2 | 32.2 | 1.04x |\\n|GARLIC | 36.9 | 36.8 | 5.7 | 30.9 | 1.00x |\\n\\nAfter reordering Table 1 to list all Top-X results, it becomes clearer that our method outperforms all baselines with similar or even lower computational costs. For summary-based methods, MeMWalker, RAPTOR, and our approach, the computational costs for summarization are similar. However, our method still achieves superior performance compared to other summary-based methods at similar computational costs.\\n\\nOur method introduces Dynamic Progress Control and a new attention-based retrieval paradigm leveraging Graph Search that does not rely on embeddings. Even with the same average TFLOPs, Dynamic Progress Control allows our method to better distribute and allocate computational resources across different queries and documents. Furthermore, as shown in Figure 4, our method allows stable adjustment of the trade-off between effectiveness and efficiency. These innovations are fundamentally different from existing RAG methods and are validated by our experimental results.\\n\\nAmong all baselines, our method is the only RAG approach that outperforms the LLM itself. As LLMs grow stronger and support longer input lengths, many RAG methods struggle to match the performance of LLMs, resulting in a widening performance gap. By deeply integrating the LLM into the retrieval process, our method achieves superior results, outperforming all baselines, including the LLM.\\n\\nWhen compared to Llama 3.1, our method also mitigates scaling issues. For example, for a single-query document from NarrativeQA, Llama 3.1 requires 3361.9 TFLOPs, whereas our method requires only 2073.8 TFLOPs, encompassing both the summary and inference stages. Additionally, for multi-query documents, our method is significantly more efficient than Llama 3.1, as the summary graph can be reused across queries.\"}", "{\"title\": \"Response\", \"comment\": \"Thanks for your detailed illustrations. I think the authors have addressed most of my concerns. I'll keep my scores. Thanks.\"}", "{\"comment\": \"We appreciate the reviewer's time and the thoughtful comments. The response is as follows:\\n\\n---\\n> **Q1**: W1. An explanation of the edge weights in the directed acyclic graph needs further clarification: specifically, how the weights are calculated. \\n\\n**A1**: \\nGiven input tokens such as $[a_1, a_2], [b_1, b_2]$ for nodes $a$ and $b$, and output tokens $[c_1, c_2], [d_1, d_2]$ for nodes $c$ and $d$, edge weights are calculated as follows:\\n\\n1. The LLM generates many-to-many summaries from $[a_1, a_2], [b_1, b_2]$ to $[c_1, c_2], [d_1, d_2]$.\\n2. During processing, attention scores are recorded and averaged across all attention heads and layers. For example, attention weights $e_{c_1 \\\\rightarrow b_1}$ across all heads and layers are averaged into a single attention weight $e_{c_1 \\\\rightarrow b_1}$ for the summary.\\n3. Token-level attention weights are grouped and averaged by node. For example, attention weights $e_{c_1 \\\\rightarrow b_1}$ and $e_{c_1 \\\\rightarrow b_2}$ are averaged to yield a single weight $e_{c_1 \\\\rightarrow b}$. This is repeated for all input nodes: $e_{c_1 \\\\rightarrow a}$, $e_{c_2 \\\\rightarrow a}$, $e_{d_1 \\\\rightarrow a}$, $e_{d_2 \\\\rightarrow a}$, $e_{c_1 \\\\rightarrow b}$, $e_{c_2 \\\\rightarrow b}$, $e_{d_1 \\\\rightarrow b}$, $e_{d_2 \\\\rightarrow b}$.\\n4. Similarly, for output $d$, we average $e_{c_1 \\\\rightarrow b}$ and $e_{c_2 \\\\rightarrow b}$ into $e_{c \\\\rightarrow b}$. This step is repeated for all output nodes: $e_{c \\\\rightarrow a}$, $e_{d \\\\rightarrow a}$, $e_{c \\\\rightarrow b}$, $e_{d \\\\rightarrow b}$. \\n5. Finally, output-level weights are normalized to sum to 1. For example, for output $d$, $e_{c \\\\rightarrow a}$, and $e_{c \\\\rightarrow b}$, are normalized to sum to 1, using $e_{c \\\\rightarrow i} = \\\\frac{e_{c \\\\rightarrow i}}{\\\\sum e_{c \\\\rightarrow j}}$, where $i$ and $j$ represent $a$ and $b$. Some attention weights are omitted, such as $e_{c_2 \\\\rightarrow c_1}$.\", \"we_classify_attention_into_two_types\": \"syntactic attention, which reflects grammatical relationships, and semantic attention, which captures meaning connections, often between paragraphs. When averaging token-level attentions, we only extract attentions between high-level and low-level nodes, omitting attentions within the same node. This process emphasizes long-distance semantic attention, which effectively reflects the semantic relationships between nodes.\\n\\nAll calculations use efficient tensor operations, making their computational cost negligible relative to the LLM processing. Optimization includes a global variable to iteratively accumulate averaged attention weights per layer, introducing only minimal memory overhead. \\n\\nWe have added these explanations to Appendix B in the revised version.\"}", "{\"comment\": \"Thank you for your valuable feedback on our work.\\n\\nTo the best of our knowledge, we are the first to propose a **Dynamic Progress Control Mechanism** and an **Attention-Based Retrieval Paradigm** in the field of RAG. These contributions are as follows:\\n* **Dynamic Progress Control**: In existing RAG methods, the volume of information to retrieve is typically pre-defined by a hyperparameter, such as the number of chunks to retrieve in traditional methods or the number of paths from top to bottom in a summary tree for recent approaches. However, determining the optimal volume of information to retrieve is challenging due to several factors:\\n * The required information volume varies across datasets, documents, and even individual queries. Even queries within the same document may demand different levels of information. Existing methods lack the capability to determine an optimal retrieval volume for each individual query. In contrast, **Dynamic Progress Control** adaptively and dynamically determines the retrieval volume for each query.\\n * Hyperparameter tuning for retrieval volume (e.g., top-5, top-10, top-15) is computationally expensive. Our approach eliminates this overhead by adaptively setting this value (e.g., top-12, top-13, top-14) during retrieval without additional computation or hyperparameter searches.\\n\\n Dynamic Progress Control is a novel mechanism that provides precise control over retrieval volume, achieving a superior balance between effectiveness and efficiency, which previous methods could not accomplish.\\n\\n* **Attention-based Retrieval**: Existing RAG methods predominantly rely on embedding similarity between sentences and queries for retrieval. We introduce a fundamentally different paradigm: Attention-Based Retrieval, using a many-to-many summarization graph. Our approach leverages attention mechanisms instead of embedding similarity. As demonstrated in Section 4.4, even with straightforward attention collection and normalization, our approach achieves superior results compared to embedding similarity. Attention-based retrieval captures both document-specific and query-specific latent semantic relationships, making it a promising direction for future research. This paradigm represents a novel retrieval framework with significant potential for further exploration.\\n\\n* **Experimental Results**: Among all baselines, our method is the only RAG approach that outperforms the LLM itself. As LLMs grow stronger and support longer input lengths, many RAG methods struggle to match the performance of LLMs, resulting in a widening performance gap. By deeply integrating the LLM into the retrieval process, our method achieves superior results, outperforming all baselines, including the LLM.\\n\\nWe believe that **Dynamic Progress Control** and **Attention-Based Retrieval** represent valuable contributions to the community. Both are novel, previously unexplored approaches, and their effectiveness is demonstrated through our experimental results.\\n\\nOnce again, thank you for your time and effort in reviewing our work. If possible, we would greatly appreciate it if you could share any references to existing methods that you feel closely align with the contributions we present. This would help us better understand how our work compares to prior research and identify any potential areas for improvement.\"}", "{\"comment\": \"Thank you again for your thoughtful and detailed feedback. We've taken your initial feedback into careful consideration and addressed each point in our responses. Could you kindly confirm whether our responses have appropriately addressed your concerns? We understand you are very busy, but we would greatly appreciate it if you could take our responses into account during discussions with the AC and other reviewers. Please let us know if you have further comments.\\n\\nThank you once again for your time and effort in reviewing our work.\"}", "{\"comment\": \"We appreciate the reviewer's time and the thoughtful comments. The response is as follows:\\n\\n--- \\n> **Q1**: In line 292, the author mentioned that the node's attention is multiplied by the corresponding position information to track the earlier information in the sequence. I think the author should conduct experiments to explain whether other adjustment methods have been tried? Why is this adjustment method effective?\\n\\n**A1**:\\nWe appreciate your insightful suggestion. Our chosen adjustment method is primarily heuristic. Ideally, if the input nodes are earlier in the sequence and the query appears later, there would be no need for this adjustment, as the query\\u2019s attention could naturally distribute across the nodes. However, in our setup, the query appears first and the documents follow. The adjustment allows us to cache key values for both query and documents efficiently.\\n\\nThe intuition behind this adjustment is based on our observation that nodes later in the sequence tend to assign lower attention weights to the query. For instance, the second node splits its attention between the query, the first node, and itself, with approximately 1/3 of the attention directed towards the query. Similarly, the third node allocates roughly 1/4 of its attention to the query. To address this, we scale the attention weights of the second and third nodes by 3 and 4, respectively, to account for their position relative to the query.\\n\\nWe acknowledge that further exploration of alternative adjustments might yield additional improvement. For this study, our goal was to focus on the overall attention-based retrieval mechanism, keeping the adjustments simple. The results show that our method achieves strong performance with this simple heuristic adjustment. We plan to explore further refinements in future work. \\n\\n\\n--- \\n> **Q2**: Lack of comparative analysis with other methods: Although the proposed model achieved excellent performance, the authors did not clearly explain why HWDAG is superior to the tree-based RAG method? Tree-based retrieval also seems to be able to summarize long documents and answer questions by selecting important nodes that are highly relevant to the query in a top-down manner through the GRAPH SEARCH method proposed by the authors.\\n\\n**A2**:\\nEven if a tree-based RAG incorporates our Graph Search method, it would still face limitations compared to our approach. This is supported by our experimental results in the Ablation Study (Table 2 and Line 471-473), which show a performance drop when replacing the graph structure with a tree.\\n\\nWhile Dynamic Progress Control could adapt to a tree-based RAG, attention-based retrieval would not perform effectively in a tree structure. Our method relies on fine-grained events at each node. Our intuition is that if an event is relevant to the query, the model can leverage detailed information from lower-level nodes of the same event. A single low-level node may contain multiple events, and different low-level nodes may reference the same event, creating a many-to-many relationship between events and nodes, as shown in Figure 5. High-level nodes can organize the events into different nodes, with each connected to its source nodes at the lower level.\\n\\nIn contrast, in the tree-based method, each node aggregates events from all its child nodes, with attention only indicating which child node contributes most to the parent node. This approach misses the internal semantic relationships we capture, such as how each child node contributes to different parts of the parent node. \\n\\nTo capture these details in a tree, one would need to split parent nodes into sentence-level segments to identify how each child contributes. However, this approach actually transforms the tree into a graph. Additionally, splitting nodes directly into sentences poses challenges, as events may span multiple sentences. For example, sentences that begin with pronouns may lose referential context after split.\\n\\nOur method employs LLM to achieve event segmentation. Compared to tree-based search, our graph-based approach functions more as an event-tracking process, using semantic relationships through attention-based retrieval. During the search, it follows events from high-level nodes down to lower-level details, with Dynamic Progress Control halting the search once sufficient detail is reached.\"}", "{\"comment\": \"---\\n> **Q5**: The authors argue that the information points (IPs) focus on \\u201ca single event or very few events\\u201d. It seems that the authors achieve this with just several instructions in the prompt. Is this approach sufficient to ensure the achievement of the intended objectives?\\n\\n**A5**:\\nYes, we achieve this with carefully designed instructions in the prompt. During our experiments, we manually reviewed at least 30 documents across different datasets and found no exceptions: Llama 3.1 consistently followed the instructions, generating Information Points (IPs) that encapsulate a single event or very few events, as illustrated in Figure 5 in the Appendix.\\n\\nIn the instructions, we did not explicitly use terms like \\u201cInformation Point\\u201d or \\u201cevent.\\u201d Instead, we instruct Llama 3.1 to summarize content in the format of \\u201cbullet points,\\u201d which we observed to work effectively. This approach likely succeeds because the concept of bullet points is widely used in human language, making it a concept that LLMs already learned robustly. Furthermore, the structure and purpose of bullet points align closely with the definition of Information Points described in our paper. Additionally, we enhance clarity by instructing the LLM to format each bullet point with a leading asterisk (\\u201c*\\u201d) and to perform summarization in the process. This ensures the output adheres to the intended format while maintaining semantic focus and brevity.\\n\\n\\n--- \\n> **Q6**: The authors average the token-level attentions to obtain the weights between information chunks (IPs). However, since these chunks represent higher-level, text-based content, is averaging token-level attentions an appropriate method for establishing relationships between these information chunks?\\n\\n**A6**:\", \"we_classify_attention_into_two_types\": \"syntactic attention, which reflects grammatical relationships (e.g., subject-verb dependencies within a sentence), and semantic attention, which captures semantic meaning connections, often between paragraphs.\\n\\nWhen averaging token-level attentions, we only extract attentions between high-level and low-level nodes, omitting attentions within the same node. This process emphasizes long-distance semantic attention, which effectively reflects the semantic relationships between nodes. For instance, if a high-level node assigns 50% of its attention to a low-level node and 50% to itself, we only extract the 50% attention to the low-level node, renormalize it, and scale it so the total sums to 1.\\n\\nAdditionally, we acknowledge that alternative methods for handling attention could potentially yield better results. However, in this study, our primary focus was to develop an overall attention-based retrieval mechanism while keeping the details of the attention-handling process simple. Despite this simplicity, the results demonstrate that our method achieves strong performance with this simple averaging. We will leave the exploration of more advanced methods for handling attention in future work.\"}", "{\"metareview\": \"**Summary:**\\n\\nThis paper introduces GARLIC, a retrieval method for long document QA that effectively balances computational efficiency and accuracy. The core innovation lies in constructing a Hierarchical Weighted Directed Acyclic Graph (HWDAG), where text is summarized into fine-grained Information Points (IPs), each capturing single or few events. The graph edges are derived from LLM attention weights, enabling dynamic and flexible traversal across hierarchical levels.\", \"key_contributions_include\": [\"Dynamic Progress Control, which allows retrieval to stop when sufficient information is gathered.\", \"Attention-Based Retrieval, leveraging attention weights for relevance instead of embedding similarity.\", \"**Strength:**\", \"Recursive use of a large language model to summarize long texts into information structures and assign attention-based edge weights introduces a novel information representation.\", \"GARLIC combines LLM-guided retrieval and computational efficiency, offering a valuable reference for other researchers.\", \"The method shows decent performance compared to existing retrieval approaches and outperforms Llama 3.1 while retaining RAG-like computational efficiency.\", \"**Weakness:**\", \"The major weakness all reviewers pointed out is the lack of clarity and detailed explanation of the key components (e.g., edge weights by reviewer kpAv and LAnf; attention normalization by cbqx)\", \"Graph construction is resource-intensive since it needs multiple rounds of summarization using LLMs\", \"Lack of comparative analysis with other methods\", \"Further analysis is needed to explain why early termination in GARLIC performs better.\", \"Although embedding similarity improves performance, it introduces redundancy when compared to a purely attention-based retrieval method.\"], \"additional_comments_on_reviewer_discussion\": \"The authors provide detailed responses to the reviewers' comments and have revised the manuscript accordingly. However, I agree that the paper still lacks clarity and detailed explanations in several key \\bissues raised by the reviewers. While some revisions were made during the rebuttal period, it remains unclear whether all concerns regarding clarity and explanation have been adequately addressed. Additionally, given that two reviewers hold a negative stance on the paper, rejection appears to be the appropriate decision.\"}", "{\"comment\": \"Thank you again for your thoughtful and detailed feedback. We've taken your initial feedback into careful consideration and addressed each point in our responses. Could you kindly confirm whether our responses have appropriately addressed your concerns? We would greatly appreciate it if you could take our responses into account during discussions with the AC and other reviewers. Please let us know if you have further comments.\\n\\nThank you once again for your time and effort in reviewing our work.\"}", "{\"comment\": \"Thank you again for your thoughtful and detailed feedback. We've taken your initial feedback into careful consideration and addressed each point in our responses. Could you kindly confirm whether our responses have appropriately addressed your concerns? We understand you are very busy, but we would greatly appreciate it if you could take our responses into account during discussions with the AC and other reviewers. Please let us know if you have further comments.\\n\\nThank you once again for your time and effort in reviewing our work.\"}", "{\"comment\": \"Thanks for your response. Upon review, I believe that the contributions made by this paper are limited and cannot meet the standard of ICLR. Furthermore, I have not encountered any novel or compelling solutions to the RAG problem within the paper. So, I keep the score.\"}", "{\"comment\": \"---\\n> **Q2**: W1. how they (attention weights) function during the search process. The current presentation lacks intuitiveness. \\n\\n**A2**:\\nThe intuition is that a high-level node closely related to the query will likely retrieve its detailed successors. Once sufficient related details are retrieved, the Dynamic Progress Control terminates the search.\", \"this_involves_two_measures\": \"1. **Query-node relatedness** (attention weights during the search, Lines 287\\u2013289).\\n2. **Node-successor relatedness** (extracted during Graph Construction via summarization, as described in Lines 241\\u2013245; further elaboration has been added to Appendix B, as mentioned in the A1 response).\\n\\nThe final retrieval score for a node is the product of these two relatedness values. Thus, a node will be retrieved only if it contains details linked to a visited node that is highly relevant to the query.\\n\\nFor example, consider a many-to-many summary $[a, b] \\\\rightarrow [c, d]$:\\n* If the high-level node $c$ is visited and highly related to the query (relatedness score $q_c$), we aim to retrieve details related to $c$. Attention weights $e_{c \\\\rightarrow a}$ and $e_{c \\\\rightarrow b}$ determine $c$\\u2019s reliance on $a$ or $b$. If $c$ attends more to $a$, $a$ is more likely to be retrieved, with a retrieval score of $q_c e_{c \\\\rightarrow a}$ for $a$ and $q_c e_{c \\\\rightarrow b}$ for $b$.\\n* If both $c$ and $d$ are visited, we consider all relevant weights: $e_{c \\\\rightarrow a}$, $e_{c \\\\rightarrow b}$, $e_{d \\\\rightarrow a}$, and $e_{d \\\\rightarrow b}$. The retrieval scores for $a$ and $b$ then become $q_c e_{c \\\\rightarrow a} + q_d e_{d \\\\rightarrow a}$ and $q_c e_{c \\\\rightarrow b} + q_d e_{d \\\\rightarrow b}$, respectively.\\n\\nThis process adapts dynamically. If $c$ is highly related to the query (high $q_c$) and $d$ is not, the method focuses on nodes related to $c$. If both $c$ and $d$ are relevant to the query, the method considers all corresponding weights, retrieving nodes connected to both.\\n\\nThe example provided is a segment of the graph, including only one set of many-to-many summarization within the entire graph structure. In practice, all many-to-many summarizations are considered, and nodes with the highest retrieval scores across all subgraphs are retrieved.\\n\\nAll calculations use the graph\\u2019s adjacency matrix, as shown in Section 3.2.2 Line 297, which allows for efficient matrix operations, keeping computational overhead minimal compared to the LLM\\u2019s processing.\"}", "{\"comment\": \"---\\n> **Q8**: GARLIC appears to extend ideas from both RAPTOR and MemWalker, borrowing sub-graph or informational-node stacking from RAPTOR and summarization from MemWalker. Could you clarify the main design distinctions and advantages of GARLIC compared to these methods (apart from using greedy search)?\\n\\n**A8**:\", \"we_outline_the_main_distinctions_of_garlic_as_follows\": \"1. **Dynamic Progress Control**: Traditional RAG approaches set a fixed retrieval volume, retrieving a predetermined number of nodes. Tree-based methods like RAPTOR and MemWalker follow a single path from root to leaf. GARLIC adapts dynamically to query needs. For example, it retrieves 2 nodes for an easy query, 5 for a medium query, and 8 for a hard query spread across different document sections. In contrast, baselines retrieve a fixed number (e.g., 5) for all queries. This adaptive retrieval enhances efficiency for simpler queries and effectiveness for harder ones, with similar average retrieval volume and computation cost. Dynamic Progress Control, combined with our graph-based search mechanism, also allows the model to adjust the number of paths explored based on whether a query\\u2019s information is concentrated in one chunk or dispersed across the document. In the \\\"Top-X\\\" experiments (Table 1), we ensure that baselines retrieve the same or greater average volume of information as our method. Despite this, our approach achieves superior results, demonstrating its ability to balance refined retrieval and query-specific needs effectively.\\n2. **New Attention-Based Retrieval Paradigm**: Unlike previous RAG methods, including RAPTOR, which rely on embedding similarity, our approach introduces a retrieval mechanism based solely on attention, without using sentence embeddings. Attention from the LLM captures node-to-node and query-to-node semantic relationships. Experiments in Table 1 and the ablation study in Table 2 demonstrate the effectiveness of this attention-based retrieval. The combination use of attention and embedding similarity further enhances performance.\\n3. **Many-to-Many Summarization for Graph Construction**: Our method constructs a graph based on many-to-many summarization, as illustrated in Figure 1, prompting the LLM to focus on extracting and connecting events across levels. An example is shown in Figure 5 in the Appendix. Lower-level nodes capture finer details, while high-level nodes provide concise summaries, allowing efficient tracking of events across document sections. For queries that involve multiple events across sections, our method efficiently tracks events through different paths to different leaves, retrieving details dispersed throughout the document. When multiple events co-occur in a single chunk, our approach captures the inter-event relations with higher retrieval scores for the relevant chunk. In the ablation study (Table 2), we observe a performance drop when replacing the graph structure with a tree, highlighting the importance of this design.\"}", "{\"comment\": \"Thank you for taking the time to read through our response. May we kindly ask if our responses have fully addressed your concerns? If there is anything unclear or confusing in our responses or the updated manuscript, we would greatly appreciate the opportunity to engage in a deeper discussion. Thank you for your time.\"}", "{\"comment\": \"---\\n> **Q4**: W3. The code should be made publicly available.\\n\\n**A4**: \\nWe will make the code publicly upon paper acceptance (see footnote #1 on page #1).\\n\\n\\n--- \\n> **Q5**: W4. Some baseline methods mentions in the article search down to the leaf nodes during retrieval, in other words, gather more information. However, why do they perform worse than the method presented in this paper, even though the proposed method may terminate the search early? Further analysis would be beneficial.\\n\\n**A5**:\\nIn summary, our method adjusts dynamically, terminating earlier for simpler queries and later for more complex ones compared to baselines.\", \"the_advantages_of_our_approach_regarding_the_search_termination_can_be_summarized_below\": \"1. **Flexible graph exploration:** As illustrated in Figure 2c, our approach explores multiple paths in the graph, adapting the number of paths and nodes retrieved according to the query and document. In contrast, baselines search along only a single path, as shown in Figure 2b. \\n2. **Adaptive retrieval volume:** Our method retrieves information dynamically, tailored to query complexity. In contrast, traditional methods require a pre-set retrieval volume, often using a fixed number of chunks or traversing one path from the root to the leaf nodes, regardless of query complexity. For instance, given three queries: \\n * Easy queries $q_{easy}$ may need only 2 nodes;\\n * Medium queries $q_{medium}$ require 5 nodes; and\\n * Hard queries $q_{hard}$, might need 8 nodes potentially spread across different document parts.\\n\\n While baselines retrieve the same fixed number of nodes (e.g., 5) for all queries, our approach retrieves 2 for $q_{easy}$, 5 for $q_{medium}$, and 8 for $q_{hard}$. This adaptive retrieval enhances efficiency for simple queries and improves effectiveness for complex ones such as $q_{hard}$, without increasing average retrieval volume or computation cost.\\n\\nIn the \\\"Top-X\\\" experiments (Table 1), we ensure that baselines retrieve the same or greater average volume of information than our method. Despite this, our approach achieves superior results, demonstrating its ability to balance efficiency and query-specific retrieval effectively through Dynamic Progress Control.\"}", "{\"comment\": \"---\\n> **Q2**: In Section 3.2.2, the paper discusses calculating the relevance between the query and information points using attention. How is this calculation performed, especially if it needs to be precomputed? \\n\\n**A2**:\\nThe attention values are directly extracted from the LLM as shown in Figure 3. We utilize the computed attention weights from the LLM and apply an averaging operation. This additional computation is minimal compared to the cost of running the LLM itself.\\n\\nFor a query with tokens $[q_1, q_2]$ and a node with tokens $[t_1, t_2]$, we first average the attention weights across attention heads and layers. This results in the attention weights $e_{t_1 \\\\rightarrow q_1}, e_{t_2 \\\\rightarrow q_1}, e_{t_1 \\\\rightarrow q_2}, e_{t_2 \\\\rightarrow q_2}$, while excluding intra-node attentions like $e_{t_2 \\\\rightarrow t_1}$.\\n\\nNext, we average the attention weights over the query tokens. $e_{t_1 \\\\rightarrow q_1}$ and $e_{t_1 \\\\rightarrow q_2}$ are averaged into $e_{t_1 \\\\rightarrow q}$. Similarly, $e_{t_2 \\\\rightarrow q_1}$ and $e_{t_2 \\\\rightarrow q_2}$ are averaged into $e_{t_2 \\\\rightarrow q}$. Finally, $e_{t_1 \\\\rightarrow q}$ and $e_{t_2 \\\\rightarrow q}$ are averaged to obtain $e_{t \\\\rightarrow q}$, representing the final attention between the query and the node.\\n\\nFrom an implementation perspective, all calculations are performed using efficient tensor operations. Averaging across heads and layers introduces minimal computational overhead. Additionally, we maintain a global variable to iteratively sum attention weights for each layer when running the LLM, ensuring limited memory usage relative to the LLM processing.\\n\\n\\n--- \\n> **Q3**: This method may have limited scalability for new queries and new documents.\\n\\n**A3**:\", \"for_new_queries\": \"Our method is highly adaptable to new queries. The graph construction is performed only once per document. New queries can directly leverage the constructed graph, requiring minimal computational effort.\", \"for_new_documents\": \"\", \"we_address_new_document_issue_as_follows\": \"1. **Efficiency for long documents**: For lengthy texts like those in NarrativeQA, our method processes documents in manageable nodes, limiting input length to under 8K. This reduces computational demands and enables inference for long documents using a single GPU with fewer TFLOPs even for single-query documents, as shown in Table 3. In contrast, LLMs scale quadratically with sequence length, requiring significant GPU memory and compute power.\\n2. **Preprocessing vs inference**: In real-world applications, we often prioritize inference speed over the preprocessing stage. For instance, when deploying our method in a product, Graph Construction can be done in advance on a high-capacity server, allowing for efficient inference on personal local devices. Directly using an LLM for inference in such scenarios would be computationally intensive for a local device.\\n3. **Single-time preprocessing in an RAG manner**: Graph Construction only needs to be performed once per document, akin to the indexing phase in RAG. RAPTOR and MeMWalker also require substantial computation resources to build the summary trees before inference. \\n\\n\\n--- \\n> **Q4**: No code is released, making it difficult for others to use the proposed method or reproduce the result.\\n\\n**A4**:\\nWe will make the code publicly upon paper acceptance (see footnote #1 on page #1).\"}", "{\"comment\": \"---\\n> **Q9**: Could you explain the choice to exclude LongT5 XL (Guo et al., 2022) and CoLT5 XL (Ainslie et al., 2023) as baselines (or any other baseline you feel may have given better performance similar to the base Llama3.1 on the benchmarked datasets), given their performance as open-source models? While ChatGPT\\u2019s closed-source nature makes its exclusion understandable, a rationale for omitting these alternatives would help clarify comparisons. \\n\\n**A9**:\\nWe consider Llama3.1 for the following reasons:\\n1. In our experiments, all baselines and GARLIC use the same base model, Llama3.1. Keeping the same LLM across all methods ensures a fair comparison, focusing on the retrieval mechanism rather than variations in LLM architecture.\\n2. LongT5 XL and CoLT5 XL support input lengths of 16K and 64K tokens, respectively, while Llama3.1 supports an input length of up to 128K tokens. This extended capacity is particularly beneficial for handling long documents, such as those in NarrativeQA. By using Llama3.1 across methods, we aim to reduce the impact of input length limitations and provide a fairer comparison. While we limited our method to an 8K input length, which ensures compatibility with various LLMs, we did not impose this restriction on direct queries to the LLM (Llama3.1).\\n3. The Llama series is among the most widely used open-source models in both academia and industry. LongT5 XL and CoLT5 XL require fine-tuning on domain-specific data as they are not aligned to human preferences via RLHF or DPO, while Llama is capable of zero-shot responses with instruction-following capabilities. Using the latest version of the Llama series, Llama3.1 (as of our submission date), aligns with common practices and provides more relevant insights for a broader audience, particularly those using or considering Llama.\\n\\n\\n--- \\n> **Q10**: Also, could you discuss the reported GARLIC performance alongside Table 5 in [1]? \\n[1] Sarthi, Parth, et al. \\\"Raptor: Recursive abstractive processing for tree-organized retrieval.\\\" arXiv preprint arXiv:2401.18059 (2024).\\n\\n**A10**:\\n\\n* **Original RAPTOR setup**: RAPTOR originally uses ChatGPT-3.5 for summarization to construct the summary tree and employs LongT5 XL, CoLT5 XL, and GPT-4 for answering queries based on the retrieved nodes. LongT5 XL and CoLT5 XL are not aligned with human preferences through RLHF or DPO and thus require fine-tuning on domain-specific data. GPT-4, which is notably stronger, is used in zero-shot mode for query responses.\\n\\n* **Our setup for RAPTOR**: In our experiments with RAPTOR, we use Llama3.1-8B for both summarization and query answering based on retrieved nodes. To answer queries, we prompt Llama3.1 in zero-shot mode, following a similar approach to how RAPTOR originally prompted GPT-4.\\n\\nIn conclusion, the original RAPTOR uses ChatGPT-3.5 for summarization, while we rerun RAPTOR using Llama3.1-8B. RAPTOR originally uses LongT5 XL and CoLT5 XL that are fine-tuned on domain-specific data, whereas we use Llama3.1 in a zero-shot manner for query responses in RAPTOR experiments. Although RAPTOR also uses GPT-4 in the zero-shot setting, GPT-4 remains a stronger model.\\n\\nConsidering these adjustments, the results would be more interpretable. We ensured that all baselines and GARLIC use the same model, Llama3.1, for both summarization and answering, to provide a fairer comparison (see Line 344).\\n\\n\\n--- \\n> **Q11**: From Table 1, it appears that a significant portion of GARLIC's performance boost is coming from the Llama-3.1 base model, as it already outperforms other baselines. How might results compare if Llama-3.1 were used with RAPTOR instead, considering RAPTOR's higher efficiency (0.5:1 on 1 datasets, and always below 1 on all remaining datasets)?\\n\\n**A11**:\\nWe used the same model, Llama-3.1, across all baselines, including RAPTOR, for both summarization and query answering (see Line 344). The results are available in Table 1. This was implemented by running the RAPTOR source code with Llama-3.1. This experimental setup ensures a fair comparison by minimizing the influence of different LLM choices.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"comment\": \"We appreciate the reviewer's time and the thoughtful comments. The response is as follows:\\n\\n> **Q1**: The attention normalization approach may limit scalability to larger graphs and complex attention patterns.\\n\\n**A1**:\", \"we_address_the_scalability_of_attention_normalization_in_terms_of_both_time_complexity_and_space_complexity\": \"**Time Complexity**:\\n\\nThe model\\u2019s time complexity is $O(nds^2 + nsd^2)$, where $s$, $d$, and $n$ represent the sequence length, hidden size, and layer number, respectively. Attention normalization requires $O(nms^2)$, where $m$ represents the attention head number, which is negligible compared to the complexity $O(nds^2)$, even for long sequences, as $d \\\\gg m$ in LLMs. \\n\\nOur method caps the sequence length $s$ to 8K, mitigating the quadratic growth of LLM computational costs with sequence length. Despite introducing an additional summarization step, the total TFLOPs used are still lower than that of Llama3.1 for long documents, such as those in the NarrativeQA dataset.\\n\\n**Space Complexity**:\\n\\nDuring model execution, attention is normalized layer by layer. We use a global variable, `attention_weight_layer_sum`, to accumulate normalized attention across layers, as shown in the pseudocode below:\\n\\n```python\\ndef forward(self, input_ids, ...):\\n ...\\n attention_weight_layer_sum = 0\\n for layer in self.layers:\\n ...\\n hidden_states, attention_weight = layer(hidden_states)\\n normlized_attention_weight_along_head = normalize(attention_weight)\\n attention_weight_layer_sum += normlized_attention_weight_along_head\\n ...\\n attention_weight = normalize(attention_weight_layer_sum)\\n ...\\n return output, attention_weight, ...\\n```\\n\\nThe global variable `attention_weight_layer_sum` accumulates normalized attention weights iteratively across layers. After each layer\\u2019s forward pass, attention is normalized along the attention head dimension to further reduce memory usage and added to `attention_weight_layer_sum`. Consequently, the model and activation space complexity remains $O(nd^2 + sd + ms^2)$, with an additional memory usage of only $O(s^2)$, which is significantly smaller than $O(ms^2)$, even for long sequences.\\n\\nBy limiting the sequence length $s$ to 8K in our method, we further alleviate the LLM's memory growth, allowing our method to run on a single GPU for NarrativeQA, whereas Llama3.1 would require 8 GPUs.\\n\\nIn summary, based on the time and space complexity analysis, the computation and memory demands of attention normalization are minor compared to the overall cost of the LLM, even for long sequences. The primary scalability bottleneck remains within the LLM itself.\\n\\n\\n--- \\n> **Q2**: While the retrieval stage is efficient, graph construction is relatively resource-intensive, especially for single-query documents.\\n\\n**A2**:\", \"we_address_the_resource_intensiveness_of_graph_construction_as_follows\": \"1. **Efficiency for long documents:** For lengthy texts like those in NarrativeQA, our method processes documents in manageable nodes, limiting input length to under 8K. This reduces computational demands and enables inference for long documents using a single GPU with fewer TFLOPs even for single-query documents, as shown in Table 3. In contrast, LLMs scale quadratically with sequence length, requiring significant GPU memory and compute power.\\n2. **Preprocessing vs inference:** In real-world applications, we often prioritize inference speed over the preprocessing stage. For instance, when deploying our method in a product, Graph Construction can be done in advance on a high-capacity server, allowing for efficient inference on personal local devices. Directly using an LLM for inference in such scenarios would be computationally intensive for a local device.\\n3. **Single-time preprocessing in an RAG manner:** Graph Construction only needs to be performed once per document, akin to the indexing phase in RAG. RAPTOR and MeMWalker also require substantial computation resources to build the summary trees before inference.\"}", "{\"comment\": \"Thanks for your response. I will keep the score.\"}", "{\"comment\": \"---\\n> **Q3**: In addition, HWDAG seems to organize documents into a top-down hierarchy with different granularity, so low-level nodes seem to be a refinement and explanation of top-level nodes. As one of the main motivations of the article, I hope the author can explain why introducing refined information will hurt performance for some problems.\\n\\n**A3**:\\nOur intuition is that retrieving more refined information generally improves performance, particularly when using powerful LLMs like Llama 3.1. However, retrieving more information increases computational costs, so we aim to strike a balance between effectiveness and efficiency, depending on the specific query. The misunderstanding may arise from the role of Dynamic Progress Control in our method. Our motivation is rooted in the fact that queries vary in complexity and information needs. GARLIC adapts dynamically to query needs.\\n\\n\\u200b\\u200bFor example, GARLIC retrieves 2 nodes for an easy query, 5 for a medium query, and 8 for a hard query spread across different document sections. In contrast, baselines retrieve a fixed number (e.g., 5) for all queries. This adaptive retrieval enhances efficiency for simpler queries and effectiveness for harder ones, with similar average retrieval volume and computation cost.\\n\\nFor simpler queries, refined information may be unnecessary, as the top-level nodes already provide sufficient context to answer the query. In such cases, Dynamic Progress Control stops the search early to prioritize efficiency. Conversely, for harder queries, the method extends the search to gather more refined information spread across different sections of the document. \\n\\nIn the \\\"Top-X\\\" experiments (Table 1), we ensure that baselines retrieve the same or greater average volume of information as our method. Despite this, our approach achieves superior results, demonstrating its ability to balance refined retrieval and query-specific needs effectively. Additionally, the stop patience parameter can be adjusted to control the trade-off between efficiency and effectiveness.\\n\\n\\n--- \\n> **Q4**: During the construction of the summary graph, the authors iteratively aggregate lower-level nodes by batching them and inputting them into the large language model (LLM) to generate higher-level nodes. However, the details of the iterative batching mechanism require further elaboration.\\n\\n**A4**:\\nUnlike RAPTOR, which batches lower-level nodes using clustering, we simply batch low-level nodes in sequential order. For each layer, we gather nodes into batches based on their token counts. The token count includes the nodes themselves but excludes tokens used for the prompt and the generated summary. Therefore, each batch is capped at approximately 6K tokens to ensure the total token count does not exceed the model's 8K limit.\\n\\nFor example, first-level nodes $\\\\{v^1_i\\\\} _{1}^{30}$ might be divided into batches such as $\\\\{v^1_i\\\\} _{1}^{10}$, $\\\\{v^1_i\\\\} _{11}^{20}$, and $\\\\{v^1_i\\\\} _{21}^{30}$, with each batch containing consecutive nodes. This approach is consistent even when processing multiple documents. The model then generates second-level nodes for each batch, e.g., $\\\\{v^2_i\\\\} _{1}^{15}$, $\\\\{v^2_i\\\\} _{16}^{20}$, and $\\\\{v^2_i\\\\} _{21}^{40}$. In this process, $\\\\{v^1_i\\\\} _{1}^{10}$ is summarized into $\\\\{v^2_i\\\\} _{1}^{15}$, $\\\\{v^1_i\\\\} _{11}^{20}$ is summarized into $\\\\{v^2_i\\\\} _{16}^{20}$, and so on. The number of higher-level nodes may vary depending on the semantic context of the summaries.\\n\\nFor subsequent levels, the nodes are similarly batched, but as higher-level nodes typically contain fewer tokens, each batch can include more nodes. For instance, second-level nodes $\\\\{v^2_i\\\\} _{1}^{40}$ might be split into batches $\\\\{v^2_i\\\\} _{1}^{20}$ and $\\\\{v^2_i\\\\} _{21}^{40}$. This allows higher-level nodes to summarize information from a variable number of lower-level nodes, depending on the semantic context of the content.\"}", "{\"comment\": \"Thank you for taking the time to read through our response.\\n\\nWe added experiments of \\u201cTop-X\\u201d setting for RAPTOR. We applied the Top-X setting to the Collapsed Tree variant of RAPTOR, extracting Top-20, Top-42, Top-12, and Top-22 nodes for NarrativeQA, Qasper, HotpotQA, and MuSiQue, respectively, to match the TFLOPs of our method, consistent with other Top-X settings. RAPTOR has two variants: Tree Traversal (TT) and Collapsed Tree (CT). Collapsed Tree collapses the tree into a single layer and retrieves nodes until a threshold is reached, based on embedding similarity to the query vector.\", \"detailed_raptor_results_are_provided_below\": \"| NarrativeQA | F1 | ROUGE-L | BLEU-4 | TFLOPs | Ratio |\\n|---|---|---|---|---|---|\\n|Llama3.1 | 53.7 | 52.6 | 10.4 | 3361.9 | 108.45x |\\n|RAPTOR-TT | 40.6 | 39.8 | 7.8 | 20.3 | 0.65x |\\n|RAPTOR-CT | 48.6 | 47.8 | 11.8 | 17.9 | 0.58x |\\n|RAPTOR-CT Top-X | 52.0 | 51.2 | 11.8 | 35.1 | 1.13x |\\n|GARLIC | 61.1 | 60.2 | 18.6 | 31.0 | 1.00x |\\n\\n| Qasper | F1 | ROUGE-L | BLEU-4 | TFLOPs | Ratio |\\n|---|---|---|---|---|---|\\n|Llama3.1 | 49.4 | 47.6 | 26.9 | 92.5 | 1.38x |\\n|RAPTOR-TT | 42.1 | 40.1 | 17.2 | 17.7 | 0.26x |\\n|RAPTOR-CT | 44.6 | 42.7 | 19.5 | 16.6 | 0.25x |\\n|RAPTOR-CT Top-X | 46.9 | 44.7 | 20.8 | 67.3 | 1.01x |\\n|GARLIC | 49.7 | 47.9 | 27.0 | 66.9 | 1.00x |\\n\\n| HotpotQA | F1 | ROUGE-L | BLEU-4 | TFLOPs | Ratio |\\n|---|---|---|---|---|---|\\n|Llama3.1 | 41.3 | 41.2 | 6.3 | 23.7 | 1.48x |\\n|RAPTOR-TT | 38.6 | 38.5 | 6.7 | 8.4 | 0.53x |\\n|RAPTOR-CT | 40.9 | 40.4 | 7.2 | 15.3 | 0.96x |\\n|RAPTOR-CT Top-X | 40.7 | 40.7 | 7.2 | 17.9 | 1.12x |\\n|GARLIC | 43.5 | 43.5 | 7.2 | 16.0 | 1.00x |\\n\\n| MuSiQue | F1 | ROUGE-L | BLEU-4 | TFLOPs | Ratio |\\n|---|---|---|---|---|---|\\n|Llama3.1 | 35.8 | 35.7 | 5.6 | 40.6 | 1.31x |\\n|RAPTOR-TT | 29.3 | 29.3 | 4.7 | 12.6 | 0.41x |\\n|RAPTOR-CT | 31.5 | 31.5 | 5.5 | 16.1 | 0.52x |\\n|RAPTOR-CT Top-X | 35.4 | 35.2 | 7.2 | 32.2 | 1.04x |\\n|GARLIC | 36.9 | 36.8 | 5.7 | 30.9 | 1.00x |\\n\\nRAPTOR-CT outperforms RAPTOR-TT, consistent with the conclusions in the RAPTOR paper. The top-to-bottom search strategy of RAPTOR-TT does not capture sufficient information to effectively answer queries. RAPTOR-CT Top-X demonstrates better performance than the original RAPTOR when provided with more nodes and TFLOPs. \\n\\nOur method still consistently outperforms RAPTOR-CT Top-X across all four datasets, even under similar TFLOPs, with particularly strong results on the long documents from NarrativeQA. Under this setting, RAPTOR-CT Top-X and our method exhibit similar computational costs in both the summary and inference stages, as both rely on a hierarchical summary structure. Thus, RAPTOR is also \\u201crelatively resource-intensive, especially for single-query documents,\\u201d as pointed out in the second weakness. This also shows that our method has superior efficiency and resource utilization under the similar computational rescources. For example, as noted in Q8/A8.1, our Dynamic Progress Control mechanism dynamically allocates fewer resources to simpler queries and more to complex ones, enhancing overall performance.\\n\\nOur method consistently outperforms RAPTOR-CT Top-X across all four datasets, even under similar TFLOPs. In this setting, RAPTOR-CT Top-X and our method have comparable computational costs in both the summary and inference stages, as both utilize a hierarchical summary structure. Consequently, as noted in the second weakness, RAPTOR is also \\u201crelatively resource-intensive, especially for single-query documents.\\u201d These results further demonstrate our method's superior efficiency and resource utilization under similar computational resources. Our Dynamic Progress Control, as highlighted in Q8/A8.1 effectively allocates fewer resources to simpler queries and more to complex ones, leading to enhanced overall performance.\\n\\nImportantly, with the same TFLOPs, RAPTOR-CT Top-X still underperforms Llama 3.1. Among all baselines, our method is the only one that outperforms Llama 3.1 across all datasets, underscoring the unique effectiveness of our approach.\\n\\nWhen comparing with Llama 3.1, the computational cost of attention normalization is negligible as noted in Q1/A1. Furthermore, the resource-intensive nature for single-query documents, highlighted in the second weakness, is also mitigated when comparing with Llama 3.1. For instance, for a single-query document from NarrativeQA, Llama 3.1 requires 3361.9 TFLOPs, whereas our method only requires 2073.8 TFLOPs, encompassing both the summary and inference stages. Moreover, for multi-query documents, our method is significantly more efficient than Llama 3.1.\\n\\nIf we have adequately addressed all your concerns, we kindly ask if you would consider updating your score. If there are any remaining concerns or reasons for you keeping your score, please let us know so we can make further changes. Thank you for your time and consideration.\"}", "{\"comment\": \"---\\n> **Q6**: Has dynamically adjusting the stop patience based on query complexity or node connectivity been considered? Could this approach optimize retrieval efficiency within larger graphs?\\n\\n**A6**:\\nThis could be an interesting idea for future exploration. In our method, we rely on the LLM to make the decision on when to stop, so it\\u2019s possible that the LLM already considered the complexity of the query or node when deciding to terminate the search. \\n\\nIn our experiments, we observed variability in the number of retrieved nodes across datasets, but the optimal stop patience consistently hovered around 5, as shown in Figure 4. This suggests that, while retrieval depth may vary by dataset, stop patience appears to correlate more closely with the LLM, with the LLM terminating only when sufficiently confident. This would make stop patience a stable hyperparameter across datasets.\\n\\nNonetheless, dynamically adjusting stop patience based on query complexity could be a valuable direction for future research to further explore.\\n\\n\\n--- \\n> **Q7**: How does the efficiency of graph search vary in contexts with mixed single- and multi-query documents, particularly in cases with highly fragmented narratives, such as those in NarrativeQA?\\n\\n**A7**:\\nAs shown in the experiments in Section 4.2, our method achieves the highest improvement on NarrativeQA among the four datasets compared to other baselines, demonstrating its efficiency in handling documents with fragmented narratives like those in NarrativeQA. \\n\\nOur approach constructs the graph for a document one time, allowing all queries to reuse this graph. For each query, as illustrated in Figure 2c, graph search occurs along multiple paths, with the paths and depths dynamically adjusted according to the specific query. This flexibility enables our method to efficiently manage mixed single- and multi-query contexts, even within highly fragmented narratives.\"}", "{\"comment\": \"I appreciate your thorough feedback, however, I've decided to maintain my current scoring for this evaluation.\"}", "{\"summary\": \"This work introduces a retrieval method, \\\"GARLIC\\\" (LLM-Guided Dynamic Progress Control with Hierarchical Weighted Graph), aimed at improving long document question answering. GARLIC operates through a multi-step process. First, it constructs a Hierarchical Weighted Directed Acyclic Graph (HWDAG) where each node represents a focused \\\"Information Point\\\" (IP) generated through LLM-based summarization. Second, GARLIC dynamically retrieves nodes from the graph, leveraging attention weights to assess relevance and terminate the search once sufficient information is collected. Third, it employs a Greedy Best-First Search to explore multiple graph paths, enhancing flexibility by allowing the search to stop at various levels based on the query's needs. This approach effectively balances retrieval precision and computational efficiency, outperforming existing methods in single and multi-document QA tasks.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"S1. The proposed method employs a hierarchical weighted directed acyclic graph instead of a tree structure, utilizing multi-path dynamic retrieval and hierarchical summary nodes.\\nS2. It is an interesting idea to assign weights to the edges based on attention, which allows GARLIC to adjust the retrieval depth and information volume flexibly.\\nS3. Decent performance compared to other existing approaches for retrieval.\", \"weaknesses\": \"W1. An explanation of the edge weights in the directed acyclic graph needs further clarification: specifically, how the weights are calculated and how they function during the search process. The current presentation lacks intuitiveness.\\nW2. Detailed explanation on how KV caching is applied in the dynamic progress control is needed.\\nW3. The code should be made publicly available.\\nW4. Some baseline methods mentions in the article search down to the leaf nodes during retrieval, in other words, gather more information. However, why do they perform worse than the method presented in this paper, even though the proposed method may terminate the search early? Further analysis would be beneficial.\", \"questions\": \"See weakness\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper proposes a new retrieval method, GARLIC (LLM-Guided Dynamic Progress Control with Hierarchical Weighted Graph), which aims to improve information retrieval in long document question answering tasks. It replaces the traditional tree structure by building a hierarchical weighted directed acyclic graph and uses the attention weights of large language models (LLMs) to perform multi-path search and dynamically control the retrieval process. Experiments show that this method surpasses the existing state-of-the-art baselines, including methods that directly use Llama 3.1 to process the entire document, while maintaining the computational efficiency of the RAG method.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. Overall, this paper is well-written and easy to read.\\n2. While existing RAG methods perform worse than directly feeding the entire document into Llama 3.1, the authors propose a new RAG method outperforms Llama 3.1 and retaining the computational efficiency of RAG methods. This approach can be used for reference by other researchers.\", \"weaknesses\": \"1. In line 292, the author mentioned that the node's attention is multiplied by the corresponding position information to track the earlier information in the sequence. I think the author should conduct experiments to explain whether other adjustment methods have been tried? Why is this adjustment method effective?\\n2. Lack of comparative analysis with other methods: Although the proposed model achieved excellent performance, the authors did not clearly explain why HWDAG is superior to the tree-based RAG method? Tree-based retrieval also seems to be able to summarize long documents and answer questions by selecting important nodes that are highly relevant to the query in a top-down manner through the GRAPH SEARCH method proposed by the authors. In addition, HWDAG seems to organize documents into a top-down hierarchy with different granularity, so low-level nodes seem to be a refinement and explanation of top-level nodes. As one of the main motivations of the article, I hope the author can explain why introducing refined information will hurt performance for some problems.\\n3. During the construction of the summary graph, the authors iteratively aggregate lower-level nodes by batching them and inputting them into the large language model (LLM) to generate higher-level nodes. However, the details of the iterative batching mechanism require further elaboration.\\n\\n4. The authors argue that the information points (IPs) focus on \\u201ca single event or very few events\\u201d. It seems that the authors achieve this with just several instructions in the prompt. Is this approach sufficient to ensure the achievement of the intended objectives?\\n\\n5. The authors average the token-level attentions to obtain the weights between information chunks (IPs). However, since these chunks represent higher-level, text-based content, is averaging token-level attentions an appropriate method for establishing relationships between these information chunks?\\n\\nComments, Suggestions And Typos:\\n1. In Section 4 Experiments, it is recommended to highlight the metrics that perform well in the table by bolding or underlining them.\", \"questions\": \"Please refer to weakness, thanks.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"details_of_ethics_concerns\": \"NA.\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"We appreciate the reviewer's time and the thoughtful comments. The response is as follows:\\n\\n--- \\n> **Q1**: The paper does not explain how edge weights between nodes are calculated using the attention mechanism of the LLM, lacking crucial algorithmic details.\\n\\n**A1**:\\nGiven input tokens such as $[a_1, a_2], [b_1, b_2]$ for nodes $a$ and $b$, and output tokens $[c_1, c_2], [d_1, d_2]$ for nodes $c$ and $d$, edge weights are calculated as follows:\\n\\n1. The LLM generates many-to-many summaries from $[a_1, a_2], [b_1, b_2]$ to $[c_1, c_2], [d_1, d_2]$.\\n2. During processing, attention scores are recorded and averaged across all attention heads and layers. For example, attention weights $e_{c_1 \\\\rightarrow b_1}$ across all heads and layers are averaged into a single attention weight $e_{c_1 \\\\rightarrow b_1}$ for the summary.\\n3. Token-level attention weights are grouped and averaged by node. For example, attention weights $e_{c_1 \\\\rightarrow b_1}$ and $e_{c_1 \\\\rightarrow b_2}$ are averaged to yield a single weight $e_{c_1 \\\\rightarrow b}$. This is repeated for all input nodes: $e_{c_1 \\\\rightarrow a}$, $e_{c_2 \\\\rightarrow a}$, $e_{d_1 \\\\rightarrow a}$, $e_{d_2 \\\\rightarrow a}$, $e_{c_1 \\\\rightarrow b}$, $e_{c_2 \\\\rightarrow b}$, $e_{d_1 \\\\rightarrow b}$, $e_{d_2 \\\\rightarrow b}$.\\n4. Similarly, for output $d$, we average $e_{c_1 \\\\rightarrow b}$ and $e_{c_2 \\\\rightarrow b}$ into $e_{c \\\\rightarrow b}$. This step is repeated for all output nodes: $e_{c \\\\rightarrow a}$, $e_{d \\\\rightarrow a}$, $e_{c \\\\rightarrow b}$, $e_{d \\\\rightarrow b}$. \\n5. Finally, output-level weights are normalized to sum to 1. For example, for output $d$, $e_{c \\\\rightarrow a}$, and $e_{c \\\\rightarrow b}$, are normalized to sum to 1, using $e_{c \\\\rightarrow i} = \\\\frac{e_{c \\\\rightarrow i}}{\\\\sum e_{c \\\\rightarrow j}}$, where $i$ and $j$ represent $a$ and $b$. Some attention weights are omitted, such as $e_{c_2 \\\\rightarrow c_1}$.\", \"we_classify_attention_into_two_types\": \"syntactic attention, which reflects grammatical relationships, and semantic attention, which captures meaning connections, often between paragraphs. When averaging token-level attentions, we only extract attentions between high-level and low-level nodes, omitting attentions within the same node. This process emphasizes long-distance semantic attention, which effectively reflects the semantic relationships between nodes.\\n\\nAll calculations use efficient tensor operations, making their computational cost negligible relative to the LLM processing. Optimization includes a global variable to iteratively accumulate averaged attention weights per layer, introducing only minimal memory overhead. \\n\\nWe have added these explanations to Appendix B in the revised version.\"}", "{\"summary\": \"The authors propose a method using LLMs to summarize long text into key information points, organizing these points into a hierarchical structure through recursive processing. This results in a directed acyclic graph where edge weights between nodes are generated based on attention mechanisms. For QA tasks, the authors apply a weighted BFS to retrieve relevant information from the graph. An LLM dynamically controls the retrieval process by determining if the information collected so far is sufficient to answer the question.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"3\", \"strengths\": \"1. The system uses a large language model to recursively generate key information from long texts, constructing an information structure graph. It employs the model's attention mechanism to assign weights to edges, introducing a new structure for information representation.\\n\\n2. The authors propose using a priority-based breadth-first search to retrieve information points within the graph. The search process is controlled by the large language model, allowing retrieval to stop as soon as relevant information is found. This approach, compared to depth-first search, offers greater flexibility while maintaining efficiency. Final experimental results outperform recent benchmarks, including Llama 3.1.\", \"weaknesses\": \"1. The paper does not explain how edge weights between nodes are calculated using the attention mechanism of the LLM, lacking crucial algorithmic details.\\n\\n2. In Section 3.2.2, the paper discusses calculating the relevance between the query and information points using attention. How is this calculation performed, especially if it needs to be precomputed? This method may have limited scalability for new queries and new documents.\\n\\n3. No code is released, making it difficult for others to use the proposed method or reproduce the result.\", \"questions\": \"see weakness\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"---\\n> **Q3**: Embedding similarity, although beneficial, introduces some redundancy when compared with pure attention-based retrieval.\\n\\n**A3**:\\nWe consider this a strength of our work. In this paper, we propose a new attention-based retrieval paradigm that goes beyond traditional embedding similarity. While incorporating embedding similarity add some redundancy, as shown in Table 2, removing it results in only a minor performance drop. This demonstrates the robustness of our proposed attention-based retrieval: even without relying on embedding similarity used in previous baselines, our method maintains strong performance. \\n\\n\\n--- \\n> **Q4**: How does the system handle differences in event granularity? Specifically, do Information Points (IPs) for shorter events lead to inconsistencies in information distribution across the graph?\\n\\n**A4**:\\nInformation Points (IPs) create intentional differences in information distribution between high-level and low-level nodes, which is a key feature of our method.\\n\\nIn our graph structure, high-level nodes represent events in a more condensed form, capturing essential details, while low-level nodes provide a more detailed account of the same events. \\n\\nFor example, let's say we have two events, Event #1 and Event #2: In a lower-level (more detailed) node, Event #1 might be described with 100 words, and Event #2 with 50 words. In a higher-level (less detailed) node, both Event #1 and Event #2 may be condensed to around 10 words each, focusing on key information, such as \\\"who did what,\\\" while omitting finer details.\\n\\nThis intentional design ensures that high-level nodes prioritize uniform representation across events, avoiding situations where events with fewer details are disproportionately underrepresented. If high-level nodes were to maintain the same proportional information distribution as low-level nodes, events with fewer tokens might disappear at higher levels, while information-rich events would retain excessive detail. This imbalance could hinder retrieval, especially for less-detailed events.\\n\\nOur approach ensures that even if details are omitted for information-rich events at higher levels, the model can retrieve more detailed information by following paths to corresponding low-level nodes. This hierarchical representation enhances the retrieval process by maintaining a balance between efficiency and granularity.\\n\\n\\n--- \\n> **Q5**: How does GARLIC scale when handling complex attention dependencies or queries that span multiple, disparate sections of a document? Is there a threshold at which the attention-guided search might encounter limitations?\\n\\n**A5**: \\nOur method is designed specifically to handle complex dependencies that span multiple, disparate sections of a document.\\n\\nAs illustrated in Figure 2c, our graph-based method, supported by Dynamic Progress Control, enables searches along various paths, with each path potentially ending in different sections of the document. In contrast, prior tree-based approaches, as shown in Figure 2b, follow a single path from the root to one leaf, retrieving information from only one section.\\n\\nThe number of graph search paths can also be dynamically adjusted based on the specific query and document. For queries that span multiple sections, the graph is capable of exploring more paths, and vice versa. Once sufficient information has been gathered, Dynamic Progress Control terminates the search process.\"}", "{\"comment\": \"---\\n> **Q3**: W2. Detailed explanation on how KV caching is applied in the dynamic progress control is needed.\\n\\n**A3**:\\nKV caching refers to storing the attention key and value tensors for each layer during decoding, which is termed as \\\"KV cache\\\" in [1] and inspired many following works in different scenarios, such as memory management [2,3], cache merging [4,5], and compression [6,7]. The main idea of KV caching is a mechanism used in transformer architectures to store and reuse the computed key ($K$) and value ($V$) states of previously processed nodes. This caching avoids redundant computations, particularly in iterative processes. In the context of KV caching, $K$ and $V$ are stored in the cache for previously retrieved nodes. This means that for a new input node, only its states need to be computed, while $K$ and $V$ from earlier nodes are reused from the cache.\\n\\nAs shown in Figure 3, all key-value (KV) pairs from previously retrieved nodes are cached. For instance, consider an initial KV cache of $[q, v_1, v_2]$, where $q$ denotes the query and $v$ represents nodes. If the method retrieves a new node $v_3$, the current state consists of the KV cache $[q, v_1, v_2]$ and input $[v_3]$. During attention computation $softmax(QK^T)V$, the query states in $Q$ correspond only to the input $v_3$, while the keys ($K$) and values ($V$) include the cached states from $q$, $v_1$, and $v_2$, which were computed earlier and stored in the KV cache. This ensures that the attention mechanism can integrate information from both the newly retrieved node and previously cached nodes efficiently. We pass the KV cache and input to the model, which then outputs a response (Yes or No) along with the updated KV cache $[q, v_1, v_2, v_3]$.\\n\\nSuppose the next retrieved node is $v_4$. The KV cache is now $[q, v_1, v_2, v_3]$ with input $[v_4]$. Again, the query states in $Q$ pertain only to $v_4$, while the keys and values incorporate the cached states from all prior nodes $[q, v_1, v_2, v_3]$. Passing these through the model results in a response and an updated cache $[q, v_1, v_2, v_3, v_4]$. This iterative process ensures that the attention mechanism dynamically integrates information from previously retrieved nodes while maintaining computational efficiency.\\n\\nThis approach is easily implemented with HuggingFace, as illustrated in the pseudocode below:\\n\\n```python\", \"while_search_not_end\": \"...\\n retrieved_node = graph_search(...)\\n intput_ids, ... = process(retrieved_node, ...)\\n response, current_key_values, ... = model(input_ids=intput_ids, past_key_values=past_key_values, ...)\\n ...\\n past_key_values = current_key_values\\n ...\\n```\\n\\nWe have added these explanations to Appendix C in the revised version.\\n\\n\\n[1] Reiner Pope, Sholto Douglas, Aakanksha Chowdhery, Jacob Devlin, James Bradbury, Anselm Levskaya, Jonathan Heek, Kefan Xiao, Shivani Agrawal, Jeff Dean. Efficiently Scaling Transformer Inference. arXiv preprint arXiv:2211.05102 (2022).\\n\\n[2] Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph Gonzalez, Hao Zhang, Ion Stoica. Efficient Memory Management for Large Language Model Serving with PagedAttention. SOSP 2023. \\n\\n[3] Wonbeom Lee, Jungi Lee, Junghwan Seo, Jaewoong Sim.InfiniGen: Efficient Generative Inference of Large Language Models with Dynamic KV Cache Management. OSDI 2024.\\n\\n[4] Jiayi Yao, Hanchen Li, Yuhan Liu, Siddhant Ray, Yihua Cheng, Qizheng Zhang, Kuntai Du, Shan Lu, Junchen Jiang. CacheBlend: Fast Large Language Model Serving for RAG with Cached Knowledge Fusion. SIGCOMM 2024.\\n \\n[5] Jang-Hyun Kim, Junyoung Yeom, Sangdoo Yun, Hyun Oh Song. Compressed Context Memory For Online Language Model Interaction. ICLR 2024.\\n\\n[6] Yuhan Liu, Hanchen Li, Yihua Cheng, Siddhant Ray, Yuyang Huang, Qizheng Zhang, Kuntai Du, Jiayi Yao, Shan Lu, Ganesh Ananthanarayanan, Michael Maire, Henry Hoffmann, Ari Holtzman, Junchen Jiang. CacheGen: KV Cache Compression and Streaming for Fast Large Language Model Serving. SIGCOMM 2024.\\n\\n[7] Suyu Ge, Yunan Zhang, Liyuan Liu, Minjia Zhang, Jiawei Han, Jianfeng Gao. Model Tells You What to Discard: Adaptive KV Cache Compression for LLMs. ICLR 2024.\"}", "{\"comment\": \"Thank you again for your thoughtful and detailed feedback. We've taken your initial feedback into careful consideration and addressed each point in our responses. Could you kindly confirm whether our responses have appropriately addressed your concerns? We would greatly appreciate it if you could take our responses into account during discussions with the AC and other reviewers. Please let us know if you have further comments.\\n\\nThank you once again for your time and effort in reviewing our work.\"}", "{\"summary\": \"The paper presents GARLIC, a retrieval method that uses Hierarchical Weighted Directed Acyclic Graphs to improve long-document question-answering (QA) tasks. Unlike traditional retrieval-augmented generation (RAG) and tree-based models, GARLIC combines a unique two-stage process - Summary Graph Construction and Dynamic Graph Search.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"By using dynamic stopping and attention-weighted paths, GARLIC avoids unnecessary computation, which is common in exhaustive retrieval methods.\\n\\nOutperforms benchmarks like Llama 3.1 and recent methods on NarrativeQA, Qasper, HotpotQA, and MuSiQue.\\n\\nCompatible with long-context LLMs (e.g., Llama 3.1), making it feasible for single GPU execution.\", \"weaknesses\": \"The attention normalization approach may limit scalability to larger graphs and complex attention patterns.\\n\\nWhile the retrieval stage is efficient, graph construction is relatively resource-intensive, especially for single-query documents.\\n\\nEmbedding similarity, although beneficial, introduces some redundancy when compared with pure attention-based retrieval.\", \"questions\": \"How does the system handle differences in event granularity? Specifically, do Information Points (IPs) for shorter events lead to inconsistencies in information distribution across the graph?\\n\\nHow does GARLIC scale when handling complex attention dependencies or queries that span multiple, disparate sections of a document? Is there a threshold at which the attention-guided search might encounter limitations?\\n\\nHas dynamically adjusting the stop patience based on query complexity or node connectivity been considered? Could this approach optimize retrieval efficiency within larger graphs?\\n\\nHow does the efficiency of graph search vary in contexts with mixed single- and multi-query documents, particularly in cases with highly fragmented narratives, such as those in NarrativeQA?\\n\\nGARLIC appears to extend ideas from both RAPTOR and MemWalker, borrowing sub-graph or informational-node stacking from RAPTOR and summarization from MemWalker. Could you clarify the main design distinctions and advantages of GARLIC compared to these methods (apart from using greedy search)?\\n\\nCould you explain the choice to exclude LongT5 XL (Guo et al., 2022) and CoLT5 XL (Ainslie et al., 2023) as baselines (or any other baseline you feel may have given better performance similar to the base Llama3.1 on the benchmarked datasets), given their performance as open-source models? While ChatGPT\\u2019s closed-source nature makes its exclusion understandable, a rationale for omitting these alternatives would help clarify comparisons. Also, could you discuss the reported GARLIC performance alongside Table 5 in [1]?\\n\\nFrom Table 1, it appears that a significant portion of GARLIC's performance boost is coming from the Llama-3.1 base model, as it already outperforms other baselines. How might results compare if Llama-3.1 were used with RAPTOR instead, considering RAPTOR's higher efficiency (0.5:1 on 1 datasets, and always below 1 on all remaining datasets)?\\n\\n[1] Sarthi, Parth, et al. \\\"Raptor: Recursive abstractive processing for tree-organized retrieval.\\\" arXiv preprint arXiv:2401.18059 (2024).\\\"\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}" ] }

[ "Saptarashmi Bandyopadhyay", "John Cole", "John P Dickerson" ]

[ "multi-agent", "unlearning", "interpretable", "faster", "robust", "MASIMU", "LIME", "reinforcement learning", "explainable AI", "XAI" ]

{ "note_id": [ "eCPELFaBKu", "dzCqMmrvQp", "O840yZpVEe", "HBbIPodMBG", "H4tosNrkxI" ], "note_type": [ "comment", "official_review", "official_review", "official_review", "official_review" ], "note_created": [ 1732144567847, 1730735877707, 1730596978989, 1730604700782, 1730490119917 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission13057/Authors" ], [ "ICLR.cc/2025/Conference/Submission13057/Reviewer_Utve" ], [ "ICLR.cc/2025/Conference/Submission13057/Reviewer_JZEK" ], [ "ICLR.cc/2025/Conference/Submission13057/Reviewer_FVPS" ], [ "ICLR.cc/2025/Conference/Submission13057/Reviewer_5qNA" ] ], "structured_content_str": [ "{\"withdrawal_confirmation\": \"I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors.\"}", "{\"summary\": \"This paper introduces Multi-Agent Machine Unlearning (MAMU), a multi-agent framework for machine unlearning, which aims to remove specific data points from trained models without complete retraining. The framework uses multiple agents that work collaboratively using recurrent neural networks (RNNs) to traverse and process images. The authors present four variants of their framework: MALMU (LSTM-based), MASMU (GRU-based), and their interpretable counterparts MALIMU and MASIMU. The approach is evaluated on high-dimensional data like images, examining both unlearning speed and model accuracy on retained data.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": [\"Unlearning is an important field, which will be necessitated as privacy regulations and \\\"right to be forgotten\\\" requirements become more prevalent in real-world ML applications\", \"The approach of using MARL for unlearning is (to my knowledge at least) novel\", \"Evaluations cover a wide variety of datasets, including medical\"], \"weaknesses\": [\"There is no dedicated related work section, and the introduction itself is quite bare. I recommend adding an explicit related work section\", \"There is no comparison to existing state-of-the-art unlearning methods making it impossible to evaluate the claimed advantages of MASIMU in the context of current unlearning literature\", \"Results do not include statistical significance tests, and performance metrics are quite close in many cases, leading to uncertainty about whether the proposed method offers meaningful improvements over baselines\"], \"questions\": [\"**Line 254**: Apart from the unlearning reweighing, how is your method different from Mousavi et al., 2019a? The multi agent algorithm you are using seems to be theirs, so I would not claim novelty\", \"**Line 379**: Which results in this table are statistically significant? Also the presentation of this table could be improved e.g. is lower or higher better for MIA / COMP\", \"**Line 401**: It's not obvious the value of plotting the accuracy / loss curves in the main paper. This would be better left for the appendix and would provide room for a proper related works section\", \"**Line 433**: Similarly to table 1, presentation in this table is poor. On its merits, it lacks statistical significance, and the results look very close to each other. Which results are statistically significant?\", \"**Line 461**: Similarly here the unlearning accuracy and loss plots do not provide much value. What are you trying to show by providing these plots in the main paper?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The paper introduces MASIMU, a framework that combines machine unlearning (MU) with a multi-agent reinforcement learning (MARL) setup for improved efficiency and interpretability. The method relies on two components. First, it leverages an interpretable AI (XAI) method to weight gradients based on the influence of data marked for removal (forget set) and their similarities to the retain set. Second, the authors introduce a multi-agent reinforcement learning (MARL) framework, in order to effectively manage high-dimensional data by having multiple agents communicate positional information and navigate image environments; this setup reduces each agent's observation space, allowing them to focus on unlearning critical gradients associated with key super-pixels that impact target labels. The authors performs experiments on MNIST, CIFAR-10, RESISC-45 (satellite images), and HAM-10000 (skin cancer images) and evaluate MASIMU\\u2019s performance in terms of robustness, interpretability, and speed.\", \"soundness\": \"1\", \"presentation\": \"1\", \"contribution\": \"1\", \"strengths\": [\"The core ideas introduced by the authors are original and interesting:\", \"Utilizing a XAI method (LIME) to compare the similarities of the retain and forget samples and weight gradients accordingly is compelling, and might be able to improve unlearning benchmarks, as well as our understanding of it.\", \"The MARL framework proposed to reduce the input dimension and focus agents on unlearning interpretable gradients of important superpixels is also an interesting approach.\", \"The authors use various metrics in their evaluations (time, completeness, MIA) that go beyond the standard ones used.\"], \"weaknesses\": [\"My main concerns about the paper are the following:\", \"Presentation: it's very challenging to fully understand the methods used and all its components.\", \"Experiments: it seems that the comparisons are only between a baseline and variations of the framework introduced. I'd like to see how MASIMU performs against other state-of-the-art unlearning techniques.\", \"Complexity of the framework: the MARL approach introduced has a lot of complexity and computational overheads, raising questions about the generalizability of the approach on other settings.\", \"Vision-only: could the framework be extended to handle non-vision (for example language) tasks? If so, how?\"], \"questions\": [\"My questions are based on the weaknesses mentioned and how they could be improved:\", \"Is it possible to improve the presentation? E.g. give the high level ideas about the framework choices (why LIME?, why MARL?, what are the motivations?), figures and high-level descriptions about the approach, and clarifying various parts of the text and the algorithms. Ideally, the reader should be able to follow along the paper easily, and the techniques introduced should \\\"make sense\\\" to them.\", \"Can we add experiments of MASIMU against other state-of-the-art (SOTA) unlearning techniques?\", \"Can the authors address the question about the generalizability of their approach? How hard is it to make it work on a different setup, against the SOTA? How carefully should the hyper-parameters be chosen, and how big is the effort to do so?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"1\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The paper introduces MASIMU, a framework aiming to enhance the efficiency, robustness, and interpretability of machine unlearning processes in deep learning models. The authors propose a method that purportedly outperforms existing unlearning approaches by integrating deep learning, multi-agent reinforcement learning (MARL), and explainable AI (XAI) techniques. The framework is evaluated on four datasets: CIFAR-10, MNIST, RESISC-45, and HAM-10000 to demonstrate its effectiveness in handling high-dimensional data and sensitive information.\", \"soundness\": \"2\", \"presentation\": \"1\", \"contribution\": \"2\", \"strengths\": \"1. The paper addresses the timely and significant issue of machine unlearning in the context of evolving privacy regulations and AI trustworthiness, which is crucial for applications involving sensitive data.\\n2. The attempt to combine deep learning, MARL, and XAI reflects an interdisciplinary approach, showcasing an effort to tackle the unlearning problem from multiple angles.\\n3. The use of various datasets, including high-resolution and medical images, indicates an effort to validate the method across different domains and data types.\", \"weaknesses\": \"1. It is unclear whether the authors conducted a comprehensive literature review of current state-of-the-art machine unlearning techniques [1-3]. Specifically, how does MASIMU differentiate itself from or improve upon existing methods? Additionally, the manuscript fails to provide comparisons between the MASIMU framework and other state-of-the-art machine unlearning techniques. Without such comparative analyses, it remains ambiguous whether MASIMU offers any substantive improvements. Why have the authors not conducted quantitative comparisons with other advanced machine unlearning methods? How does MASIMU perform in terms of efficiency, effectiveness, and scalability relative to these existing methods?\\n\\n2. It is a recognized standard in machine unlearning research to use retraining as the benchmark for unlearning performance [1-5]. However, the paper does not provide such baseline results for reference.\\n\\n3. The inherent trade-off between the ability to retain knowledge and the level of residual information pertaining to the forgotten data is not addressed. An in-depth discussion on this relationship would enhance the understanding of the proposed method\\u2019s effectiveness.\\n\\n4. While the paper highlights the incorporation of LIME to achieve interpretability within the MASIMU framework, it lacks specific statistical explanations detailing this interpretability within the unlearning context. How does LIME's integration specifically enhance the interpretability of the unlearning process in MASIMU? The inclusion of visualizations or case studies demonstrating this interpretability and its benefits to the unlearning process would be valuable.\\n\\n5. The paper does not provide a theoretical basis or analysis to underpin the proposed method. Critical components, such as the integration of MARL and XAI in the context of machine unlearning, are not theoretically justified. What is the theoretical foundation of the MASIMU framework? Can the authors provide theoretical proofs or analyses to substantiate claims regarding the improved efficiency, robustness, and interpretability of their approach?\\n\\n6. The manuscript suffers from writing issues, including grammatical errors and unclear explanations, which hinder the reader's comprehension of the proposed method and its contributions. The figures are difficult to read (Fig. 3 and 4), and there are duplicate citations (e.g. 570-572). Have the authors conducted a thorough revision to improve the clarity and coherence of the paper? Additionally, could the authors consider reorganizing the structure to present the methods and results in a more logical and clear manner?\\n\\n7. The analysis of experimental results is insufficient, as it lacks comparisons with relevant baselines and detailed discussions on the evaluation metrics used. Important details such as hyperparameter settings and statistical significance are omitted. Furthermore, the authors have not provided the source code, which would facilitate implementation and further validation of their work.\\n\\n> [1] Fan, Chongyu, et al. \\\"SalUn: Empowering Machine Unlearning via Gradient-based Weight Saliency in Both Image Classification and Generation.\\\" The Twelfth International Conference on Learning Representations.\\n> \\n> [2] Liu, Jiancheng, et al. \\\"Model sparsity can simplify machine unlearning.\\\" Advances in Neural Information Processing Systems 36 (2024). \\n> \\n> [3] Kurmanji, Meghdad, et al. \\\"Towards unbounded machine unlearning.\\\" Advances in neural information processing systems 36 (2024).\\n> \\n> [4] Golatkar, Aditya, Alessandro Achille, and Stefano Soatto. \\\"Eternal sunshine of the spotless net: Selective forgetting in deep networks.\\\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2020.\\n> \\n> [5] Thudi, Anvith, et al. \\\"Unrolling sgd: Understanding factors influencing machine unlearning.\\\" 2022 IEEE 7th European Symposium on Security and Privacy (EuroS&P). IEEE, 2022.\", \"questions\": \"See weakness.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper introduces a new method Multi-Agent Speedy Interpretable Machine Unlearning\\n(MASIMU) for efficiently \\u201cunlearning\\u201d data from AI models, essential for privacy compliance\\nwithout requiring complete model retraining. This approach combines deep learning,\\nreinforcement learning, and explainable AI(LIME) to remove specific data influences. It\\nfine-tunes a model on the retain set while re-weighting gradients to diminish the impact of data\\nneeding removal (the forget set). By using multiple agents with reinforcement learning, MASIMU\\nbreaks down complex image data into smaller, manageable parts, allowing faster and more\\ninterpretable unlearning. This method has shown good performance in speed, robustness, and\\ninterpretability across various datasets, including CIFAR-10, MNIST, and sensitive medical\\nimages.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"1. The authors have formulated a critical, state-of-the-art problem in AI by proposing a\\nmethod focused on enhancing trustworthiness and safeguarding the privacy of data.\\n2. The proposed method described in this paper is clear and easy to follow for\\nreproduction.\\n3. The MASIMU framework incorporate a multi-agent system to divide the unlearning tasks\\namong agents, which allows for faster processing and more manageable handling of\\nhigh-dimensional data\", \"weaknesses\": \"1. Since machine unlearning is not entirely a new concept, the author should compare the\\nresults with at least one or more existing works that are similar, such as \\\"SALUN:\\nEmpowering Machine Unlearning via Gradient-Based Weight Saliency in Both Image\\nClassification and Generation,\\\" published in ICLR 2024.\\n2. The results presented in the paper do not sufficiently support the authors' claims. To\\nstrengthen their findings, they should consider using well-known datasets like ImageNet\\nto demonstrate the effectiveness of their method. Additionally, including qualitative\\nexamples of the model's behavior regarding the forgotten image set would enhance\\nclarity and impact.\\n3. A well-structured and state-of-the-art literature review on existing works in machine\\nunlearning is missing from the paper.\\n4. What is the motivation for performing machine unlearning on the selected datasets? i) You can easily train a new model within a few minutes or hours. ii) None of the datasets have an intuitive application to machine unlearning. Why not report results on the same dataset as the 2023 contest that the authors cite?\", \"questions\": \"There was a contest on this topic in 2023 at NeurIPS. Why have you not compared against those solutions?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"details_of_ethics_concerns\": \"n/a\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}" ] }

[ "Chenxi Zhong", "Hang Liu", "Xiaojun Yuan" ]

[ "Personalized federated learning", "variational message passing", "feature representation learning" ]

{ "note_id": [ "uUIICDyfHa", "rdSju0Pe5p", "rICpuOQG85", "qahvAxpLw5", "kkvUmyrwZ8", "efiewkmFd2", "bmu3OCKZFZ", "aA4dHjfhXe", "XH2Ut6Piyx", "X7LJOUkPOJ", "TQG9jasR0F", "T3SvpXMRoU", "MXxq3ociN9", "M7i35G2jpg", "Ktvz91hy9n", "IhDCJ2s1An", "E4TgOlALZN", "7ZCCspyqci", "6mz8NcKsFK" ], "note_type": [ "official_comment", "official_comment", "official_review", "official_comment", "official_review", "official_comment", "official_review", "official_comment", "meta_review", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "decision", "official_comment", "official_comment", "official_comment", "official_comment" ], "note_created": [ 1732862886680, 1732777927636, 1731073341457, 1732777833548, 1730645199063, 1732778668847, 1730697283908, 1732777431376, 1734517515052, 1733310096175, 1732777863171, 1732778637102, 1731004622138, 1732778433162, 1737523510754, 1732890047154, 1732937119067, 1732954426610, 1733062704582 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission2535/Reviewer_g6Ef" ], [ "ICLR.cc/2025/Conference/Submission2535/Authors" ], [ "ICLR.cc/2025/Conference/Submission2535/Reviewer_WGea" ], [ "ICLR.cc/2025/Conference/Submission2535/Authors" ], [ "ICLR.cc/2025/Conference/Submission2535/Reviewer_g6Ef" ], [ "ICLR.cc/2025/Conference/Submission2535/Authors" ], [ "ICLR.cc/2025/Conference/Submission2535/Reviewer_T9rb" ], [ "ICLR.cc/2025/Conference/Submission2535/Authors" ], [ "ICLR.cc/2025/Conference/Submission2535/Area_Chair_BzHs" ], [ "ICLR.cc/2025/Conference/Submission2535/Authors" ], [ "ICLR.cc/2025/Conference/Submission2535/Authors" ], [ "ICLR.cc/2025/Conference/Submission2535/Authors" ], [ "ICLR.cc/2025/Conference/Submission2535/Reviewer_5t2i" ], [ "ICLR.cc/2025/Conference/Submission2535/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission2535/Reviewer_WGea" ], [ "ICLR.cc/2025/Conference/Submission2535/Authors" ], [ "ICLR.cc/2025/Conference/Submission2535/Authors" ], [ "ICLR.cc/2025/Conference/Submission2535/Reviewer_T9rb" ] ], "structured_content_str": [ "{\"comment\": \"An additional suggestion: experiments should be considered to straightforward support your claim are about how precise the centroids pFedVMP estimates. A better accuracy is not enough, because better performance does not mean less biased estimation of the components in the optimization process.\\n\\nMost of the main concerns are resolved. I raise my rating accordingly.\"}", "{\"comment\": \"**We thank Reviewer 5t2i for careful reading and constructive comments.**\\n\\n**W1.1 Novelty of our paper**\\n\\n`The novelty of the paper is not high in my opinion. The core idea has been proposed before in the literature.`\\n\\nWe respectfully disagree with the comment. The central contribution of our work lies in leveraging the **mean and covariance** of global feature centroids to enhance the performance of personalized federated learning (pFL). This approach is novel compared to prior works on pFL, such as FedProto, GPFL, and FedPAC.\\n\\nPrevious methods primarily estimate global feature centroids through arithmetic averaging of feature samples. However, due to the statistical heterogeneity of training data, the **arithmetic mean** may deviate from the true centroids. By utilizing variational message passing, pFedVMP provides more **precise estimates** of the distributions of global feature centroids, incorporating both their mean and covariance, thereby improving personalized model training performance.\\n\\n**W1.2 Comparison with FedRep** \\n\\n- **Similarity**: Both FedRep and pFedVMP consider to split the neural network model into a head and a base, where the base model aims to learn the common feature representations and the head model aims to achieve personalized goals.\\n- **Difference**: FedRep does not involve a constraint of global feature centroids in local training. This increases the overfitting risk of the personalized model on the devices side. Unlike FedRep, pFedVMP is effective in boosting training accuracy and preventing overfitting by **regularizing** local training with global feature centroids. Meanwhile, benefiting from variational message passing, pFedVMP achieves more **precise estimates** of the distributions of global feature centroids, based on the mean and the covariance of the global feature centroids, which enhancing the learning performance of personalized model training.\\n\\n**W1.3 & Q1 Key differences between pFedVMP and the prior Bayesian federated learning (BFL)**\\n\\nBFL can be broadly categorized into client-side BFL and server-side BFL in terms of FL architectures [1]. Client-side BFL focuses on learning Bayesian local models on client nodes, while server-side BFL aggregates local updates for global models using Bayesian methods. Our paper belongs to server-side BFL. \\n\\nThe prior server-side BFL works related to our paper include FedPA, FedEP, QLSD, pFedBayes, which formulate model training as Bayesian inference tasks and aggregate the distributions of local **parameters**. FedPA (Al-Shedivat et al., 2021) approximated the posterior distribution into the product\\nof distributions with respect to local datasets during local model training. FedEP (Guo et al., 2023)\\ndeveloped the Bayesian model aggregation rule by using expectation propagation. QLSD (Vono\\net al., 2022) extended the approach in FedPA with the quantized Langevin stochastic dynamics for\\nlocal update. pFedBayes (Zhang et al., 2022) uses variational inference to approximate the posterior distribution of local model parameters for each client, and aggregates local models on the server.\\n\\nHowever, the above BFL methods do not utilize **global feature centroids** to guide local model training, which limits their ability to effectively address data heterogeneity. In contrast, pFedVMP considers both model **parameters** and **feature** centroids. The gain of pFedVMP comes from the following two folds: on one hand, pFedVMP guides the local training using a **regular** term of global feature centroids, decreasing the over-fitting risk of local training; on the other hand, pFedVMP achieves more **precise estimates** of the distributions of global feature centroid by variational message passing.\\n\\n[1] Bayesian Federated Learning: A Survey, IJCAI-23\\n\\n**Q2 Concern of the application of BFL** \\n\\nBFL is applicable to \\\"the problem which has been already studied by convex optimization techniques\\\".\\n\\nWe interpret the phrase \\\"the problem which has been already studied by convex optimization techniques\\\" as referring to problems like Problem 3 in FedRep, which involves convex linear regression and is solvable via convex optimization methods. In such cases, the optimization problem of minimizing the loss function given some data samples is equivalent to estimating parameters using data samples by maximizing the likelihood or posterior distribution. For instance, FedPA (Al-Shedivat et al., 2021) addresses a federated least squares regression problem with a linear model, as shown in Equation (2). Furthermore, Equation (3) demonstrates that the solution obtained by maximizing the likelihood distribution, i.e., the mean of the distribution, aligns with the solution from convex optimization.\"}", "{\"summary\": \"The paper presents a Bayesian approach to dealing with personalized federated learning (pFL). In particular, a \\\"shared\\\" base model is learned to map inputs to representations and each client learns a local head model to turn those representations into prediction outputs. In their approach, distributions of parameters of the base and head models are learned. In addition, the distribution of the representations are considered via a GMM (mixing on true labels). Locally, updates for the base, head, and GMM are learned. For global aggregation the base and GMM models are updated. The paper consider the case where the relevant distribution of the base and head models are Gaussians. Their approach is tested across various vision datasets.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"3\", \"strengths\": [\"The approach shows promising experimental results, showing better results than the baselines reported\", \"The overall approach make intuitive sense, combining ideas from pFL and Bayesian FL.\"], \"weaknesses\": [\"The primary weakness of the paper is in part of its presentation. This is particularly the case for Section 4 where conceptual optimization goals is mixed in with practical simplifications.\", \"In addition, there are no detailed derivation for some of the quantities used (in main text nor appendix).\"], \"questions\": \"Questions + Remarks:\\n\\n1. $p({\\\\theta}^{\\\\\\\\rm b}, \\\\\\\\{ \\\\theta_n^{\\\\\\\\rm h} \\\\\\\\}, \\\\\\\\{ z_k \\\\\\\\}, S)$ is a distribution over parameters (+ representations) and samples $S$. But, as far as I can tell from the equations, the surrogate distribution $q$ being consider is only over parameters (+ representations). As such, it is unclear how the KL-divergences are being evaluated, eg, (P1) including the argmin.\\nPlease provide clarity on this support issue of the distributions and how the KL-divergence is being evaluated.\\n\\n2. From what I understand, when (P1) is referred in text, it only refers to the parameter updates and not the variational / KL-divergence aspect over the equation. This is rather unclear in Section 4.2.\\nPlease clarify this (P1), perhaps by presenting the entire optimization in multiple line (labeling as (P1a) and (P1b) for instance).\\n\\n3. I think additional clarity in the text should also be added to distinguish section which are considering the update of parameters in (P1) (Section 4.3) vs updates on the distribution in (P2) (Section 4.2 & 4.3). This aspect is also mixed in Section 4.3.1. It may be worth splitting this subsubsection into two separate subsubsections, one for local updates on the parameters and one for local updates on the distributions.\\n\\n4. Is (P2) and (P3) equivalent (when restricting optimization to local parameter etc)? Line 269-270 says that the optimization is converted. Does this imply equivalence?\\n\\n5. The soundness of going from (P3) to (8) is a bit unclear. Could you please elaborate on the derivation (which I believe is just maximizing (7)) and why it is \\\"low-cost implementation of SG-MCMC\\\".\\n\\n6. (P2) seems to be imprecise. In particular, the second term in the KL is not a normalized distribution? In particular, several \\\"prior\\\" distribution seem to be missing in (P2) and the accompanying text. Please clarify this.\\nIt would also be useful for completeness to include a derivation of the specific factorization you are using; and its subsequent use in (P3).\\n\\n---\", \"minor\": [\"References were cut off from the main text.\", \"The subscript of the max in (1) and (P1) is not very nice. Maybe having the subscript fully under the \\\"max\\\" would improve readability of these equations.\", \"$\\\\\\\\mathcal{Z}_{k,n}$ seems to be incorrect on line 160 (should have $k$ instead of $y_k$)\", \"I think \\\\bigcup $\\\\\\\\bigcup$ is typically used over \\\\cup $\\\\\\\\cup$ for indexed unions.\", \"(1) and 2) on line 42 - 44 are not consistent\", \"Line 240 in denominator, there is a missing bracket\", \"(P1-3) should be on the RHS to be consistent with equation numbering (maybe via \\\\tag)\", \"Figure 3, missing space after \\\"Upper:\\\"\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"**Q5.1** **The derivation from (P3) to (8):** We agree with the reviewer that this derivation is based on maximizing (7). We now derive the loss function of SGD (8). Based on the above definition of $\\\\tilde{q}\\\\_n(\\\\boldsymbol{\\\\theta }\\\\^{\\\\mathrm{b}},\\\\boldsymbol{\\\\theta }\\\\_n\\\\^{\\\\mathrm{h}},\\\\{\\\\mathbf{z}\\\\_{k}\\\\})$, the negative logarithm of the target distribution is expressed as:\\n$$\\n\\\\begin{align*}\\n -\\\\log \\\\tilde{q}\\\\_n(\\\\boldsymbol{\\\\theta }\\\\^{\\\\mathrm{b}},\\\\boldsymbol{\\\\theta }\\\\_n\\\\^{\\\\mathrm{h}},\\\\{\\\\mathbf{z}\\\\_{k}\\\\})\\n &=-\\\\log p(\\\\mathcal{S}\\\\_n | \\\\boldsymbol{\\\\theta }\\\\^{\\\\mathrm{b}},\\\\boldsymbol{\\\\theta }\\\\_n\\\\^{\\\\mathrm{h}})- \\\\log q(\\\\{\\\\mathbf{z}\\\\_k\\\\})\\n - \\\\log q\\\\_{-n}(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}) \\n - \\\\log q\\\\_{-n}(\\\\{\\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}\\\\}) + \\\\mathrm{Const.}\\n\\\\end{align*}\\n$$\\nOn client $n$, computing the cavity factors $q\\\\_{-n}(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}})$ and $q\\\\_{-n}(\\\\{\\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}\\\\})$ may lead to instability during sampling. Thus, we exclude the terms involving $q\\\\_{-n}(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}), q\\\\_{-n}(\\\\{\\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}\\\\})$, resulting in the following simplified loss function:\\n$$\\n\\\\begin{align*}\\n -\\\\log p(\\\\mathcal{S}\\\\_n | \\\\boldsymbol{\\\\theta }\\\\^{\\\\mathrm{b}},\\\\boldsymbol{\\\\theta }\\\\_n\\\\^{\\\\mathrm{h}})- \\\\log q(\\\\{\\\\mathbf{z}\\\\_k\\\\})\\n\\\\end{align*}\\n$$\\nBy assuming the data samples are i.i.d., we obtain eq. (8): \\n$$\\n\\\\begin{align}\\n \\\\sum\\\\nolimits\\\\_{i = 1}\\\\^{S\\\\_n}\\n \\\\left( - \\\\log p( \\\\mathbf{x}\\\\_{n,i},y\\\\_{n,i}|\\\\boldsymbol{\\\\theta }\\\\^{\\\\mathrm{b}},\\\\boldsymbol{\\\\theta }\\\\_n\\\\^{\\\\mathrm{h}})+\\\\xi\\\\_1 \\\\\\\\|\\\\mathbf{z}\\\\_{n,i}-\\\\boldsymbol{\\\\mu}\\\\_{y\\\\_{n,i}}\\\\^{\\\\mathrm{z}}\\\\\\\\|\\\\^2 \\\\right), \\\\tag{8}\\n\\\\end{align}\\n$$\\nwhere the second term is because calculating the precision matrix $\\\\boldsymbol{\\\\Lambda}\\\\_{y\\\\_{n,i}}\\\\^{\\\\mathrm{z}}$ in the loss function may cause the gradient unstable, and we use a spherical Gaussian distribution with the mean $\\\\boldsymbol{\\\\mu}\\\\_{y\\\\_{n,i}}\\\\^{\\\\mathrm{z}}$ and the precision matrix $\\\\xi\\\\_1 \\\\mathbf{I}$ instead.\\n\\n**Q5.2** **\\\"A low-cost implementation of SG-MCMC\\\"**: The results of $\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}, \\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}$ updated by SGD can be regarded as a single sample drawn by SG-MCMC, which reduces the computational and storage cost in the sampling.\\n\\n**Q6** We agree with the reviewer that in the previous version, the second term of the KL-divergence in (P2) is incorrect. The renewed (P2) is given by\\n$$\\n\\\\begin{equation*}\\n \\\\text{(P2)} \\\\min\\\\_{q\\\\_n(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}, \\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}, \\\\{\\\\mathbf{z}\\\\_k\\\\})}D\\\\_\\\\mathrm{KL}(p(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}, \\\\{\\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}\\\\}, \\\\{\\\\mathbf{z}\\\\_k\\\\} | \\\\mathcal{S}) \\\\\\\\| q\\\\_n(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}, \\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}, \\\\{\\\\mathbf{z}\\\\_k\\\\}) q\\\\_{-n}(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}) q\\\\_{-n}(\\\\{\\\\mathbf{z}\\\\_k\\\\}) \\n q\\\\_{-n}(\\\\{\\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}\\\\}) ).\\n\\\\end{equation*}\\n$$\\nThe derivation is given as following. We expand $q(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}, \\\\{\\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}\\\\}, \\\\{\\\\mathbf{z}\\\\_k\\\\})$, i.e., the second term of KL-divergence as\\n$$\\n\\\\begin{align*}\\n q(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}, \\\\{\\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}\\\\}, \\\\{\\\\mathbf{z}\\\\_k\\\\}) \\n & \\\\overset{(a)}{\\\\propto}\\n q(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}})\\n q(\\\\{\\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}\\\\})\\n q(\\\\{\\\\mathbf{z}\\\\_k\\\\}) \\\\\\\\\\\\\\\\\\n & \\\\overset{(b)}{\\\\propto} \\\\left( q\\\\_\\\\mathrm{pri}(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}})\\\\prod\\\\_{n=1}\\\\^N q\\\\_n(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}) \\\\right)\\n \\\\left(\\\\prod\\\\_{n=1}\\\\^Nq\\\\_\\\\mathrm{pri}(\\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}) \\n q\\\\_n(\\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}})\\\\right)\\n \\\\left(q\\\\_\\\\mathrm{pri}(\\\\{\\\\mathbf{z}\\\\_k\\\\})\\\\prod\\\\_{n=1}\\\\^N q\\\\_n(\\\\{\\\\mathbf{z}\\\\_k\\\\}) \\\\right) \\\\\\\\\\\\\\\\\\n & \\\\overset{(c)}{\\\\propto} q\\\\_n(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}) q\\\\_n(\\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}) q\\\\_n(\\\\{\\\\mathbf{z}\\\\_k\\\\}) q\\\\_{-n}(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}) q\\\\_{-n}(\\\\{\\\\mathbf{z}\\\\_k\\\\}) q\\\\_{-n}(\\\\{\\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}\\\\})\\n\\\\end{align*}\\n$$\\n\\nwhere step (a) is because of the definition of $q(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}, \\\\{\\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}\\\\}, \\\\{\\\\mathbf{z}\\\\_k\\\\})$ in eq. (3), step (b) is obtained by plugging eq. (4) into the definition of $q(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}, \\\\{\\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}\\\\}, \\\\{\\\\mathbf{z}\\\\_k\\\\})$, and step (c) is due to the definition of the cavity factors $q\\\\_{-n}(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}), q\\\\_{-n}(\\\\{\\\\mathbf{z}\\\\_k\\\\}), q\\\\_{-n}(\\\\{\\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}\\\\})$.\"}", "{\"summary\": [\"This paper presents a personalizd federated learning methods, called pFedVMP, as a new solution to tailed personalized model for local clients. The core idea is to model both the centorids of parameters and features.\"], \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": [\"The combination of modeling both feature space and parameter space seems to be good as shown in the reported results of the experiments.\"], \"weaknesses\": \"- The main ideas of pFedVMP, modeling parameters and feature centroids, are not new. A Bayesian perspective is either an well-explored areas in federated learning.\\n- Even tough this area is well-explored in recent years, most of the baselines in the experiments are not the newest, which make it hard to believe to be SOTA. Moreover, the most related works are not included as a baseline of Bayesian Federated Learning.\\n 1. Related and new baselines are recommended as followings: feature modeling [1,2] and parameter modeling [3,4].\\n 2. MOON [1] has similar claims about feature centroids modeling, and however is not discussed.\\n 3. PRIOR [4] emphasizes the importance of global prior information which is the parameter centroid, which has not been discussed yet.\\n 4. More comparison about works after year 2023 should be added beyond FedPAC (ICLR 2023) in order to be claimed as SOTA.\\n- Some words, e.g., leveraging second-order statistical information, are confusing in the federated learning. The ambiguous words in this comprehensive field can refer to second-order moments, covariance matrices, or second-order gradients, Hessian matrices.\\n\\n[1] Model-contrastive federated learning. CVPR 2021\\n\\n[2] FedCP: Separating Feature Information for Personalized Federated Learning via Conditional Policy. KDD 2023\\n\\n[3] Personalized Federated Learning via Variational Bayesian Inference. ICML 2022\\n\\n[4] PRIOR: Personalized Prior for Reactivating the Information Overlooked in Federated Learning. NeurIPS 2023\", \"questions\": [\"What's the main difference between the propoed pFedVMP and the methods use GMM to model the feature centroids or parameter centeroids?\", \"The main claim, the global feature centroids is important, is already a common sense in the literature of federated learning, which is first systematically claimed and proven by MOON [1] as far as I know. What's more?\", \"If the difference is clearly explained, a positive rating is considered.\", \"[1] Model-contrastive federated learning. CVPR 2021\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"**Q1 Comparison with other GMM-based models**\\n\\nIn the following, we discuss the differences between the proposed pFedVMP and the GMM-based methods on the aspects of the feature centroids and model parameter.\\n\\n**Feature centroids** To our best knowledge, the research of modeling the feature centroids with GMM in pFL is rare. Here, we refer to the works on centralized learning, e.g., DMVCVAE [R2]. \\n\\n- The methods using GMM to model the feature centroids are typically in unsupervised learning. In DMVCVAE uses GMM to model feature centroids for clustering tasks, primarily aiming to train a deep autoencoder and learn shared latent representations that group data points based on similarity.\\n- In this paper, we consider supervised learning. At the local training, each feature sample obtained by the base model is forced to be close to the global centroid of its class. At the PS, we assume that the class information of the feature centroids is known beforehand. Thus, the aggregation of feature centroids distributions is performed on each class separately, resulting in an aggregation of multiple Gaussian distributions for each class.\\n\\n**Model parameter** We consider FedGMM is one of \\\"the methods that use GMM to model the parameter centroids\\\". FedGMM uses GMM to model the distributions of the local data and the model parameters. The local model parameters is a weighted sum of a branch of model parameters. pFedVMP uses Gaussian distributions to model the distribution of the model parameters, and the local model parameters is the mean of the Gaussian distribution. Compared with pFedVMP, FedGMM requires more computational and communication cost to update the multiple components of the model parameters.\\n\\n[R2] Shared Generative Latent Representation Learning for Multi-View Clustering. AAAI 2020\\n\\n[R3] Personalized federated learning under mixture of distributions. ICML 2023\\n\\n**Q2 Difference between pFedVMP and MOON** \\n\\nWe agree with the reviewer that the global feature centroids is important in pFL. Below, we outline the key differences between pFedVMP and MOON.\\n\\n- **Bayesian modeling** MOON aims to align the features obtained by the local model and the global model on a **sample-wise basis**. In contrast, pFedVMP treats feature centroids as random variables and aggregates their distributions using a maximum-a-posteriori (MAP) criterion. By accounting for both the mean and covariance of feature centroids, pFedVMP provides more **precise estimates** of the global feature centroid distributions, thereby enhancing model performance. To the best of our knowledge, this work is the first to estimate global feature centroids via variational message passing in pFL.\\n- **Feature aggregation** MOON focuses on achieving agreement between the local and global model representations on a sample-wise basis and does not aggregate feature centroids at the server. Conversely, pFedVMP aggregates the distributions of feature centroids at the server to achieve more **accurate estimates** of global feature centroids.\\n- **Feature alignment** MOON aligns the local and global representations by maximizing their similarity, whereas pFedVMP employs a regularization term involving the global centroids to achieve alignment.\\n\\nNumerical results shows that the proposed pFedVMP outperforms MOON on the test learning accuracy, highlighting the effectiveness of pFedVMP.\\n\\n**Q3** Thanks for the constructive comments from the reviewer. We have carefully explained the differences between pFedVMP and the prior BFL or pFedVMP and MOON. We hope this revision ensures the distinction is now clear and supports a positive evaluation.\"}", "{\"summary\": \"pFedVMP is a personalized federated learning approach that uses variational message passing to enhance feature aggregation, yielding more precise model parameter estimates and improving training accuracy and fairness under heterogeneous data conditions.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"This paper provides a nice numerical study with SOTA baselines with interpretation, ablation study, and fairness analysis.\", \"weaknesses\": \"First of all, there is a typo in the title of this paper: \\\"Massage Passing\\\" should be \\\"Message Passing\\\"...\\n\\nW1. The Bayesian benchmark models are not included in the comparison. \\n\\nW2. The computational cost seems to be high, especially with high-dimensional features. \\n\\nW3. The selection of hyperparameters needs justification. \\n\\nW4. There is a lack of theoretical guarantees for the proposed method.\", \"questions\": \"Q1. The paper presents extensive comparisons with various methods in federated representation learning but fails to include benchmark models within the framework of Bayesian federated learning, for example, BNFed, pFedGP, pFedBayes, FedPA, FedEP, QLSD, and others. This weakens the argument for the superiority of the proposed method in the Bayesian context.\\n\\nQ2. The pFedVMP algorithm involves numerous matrix inversions in each communication round (especially in Equation 12), which can lead to a significant computational burden, particularly with high-dimensional features. It is essential to evaluate the computational cost relative to other methods and propose reasonable solutions to mitigate these costs.\\n\\nQ3. The algorithm contains several hyperparameters, such as those in Equations 8 and 10. A more in-depth study on the impact of these hyperparameters and a clear justification for their selection is necessary.\\n\\nQ4. There is a lack of theoretical guarantees for the proposed method.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"**We thank Reviewer WGea for careful reading and constructive comments.**\\n\\n**W1** Thanks for the comment. As the suggestions from the reviewer, we have improved the presentation of the paper, especially Section 4. \\n\\n**W2** Thanks for the comment. In the revised manuscript, we provide clearer explanations and derivations to make the quantities used, improving the readability of this paper.\\n\\n**Q1**: Thanks for pointing out this issue. We agree with the reviewer that the surrogate distribution $q(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}, \\\\{\\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}\\\\}, \\\\{\\\\mathbf{z}\\\\_k\\\\})$ is defined over the parameters (and representations) $(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}, \\\\{\\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}\\\\}, \\\\{\\\\mathbf{z}\\\\_k\\\\})$, while $p(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}, \\\\{\\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}\\\\}, \\\\{\\\\mathbf{z}\\\\_k\\\\}, \\\\mathcal{S})$ includes the sample set $\\\\mathcal{S}$. This difference indeed raises questions about the proper evaluation of the KL-divergence. \\n\\nTo address this issue, we use the posterior distribution $p(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}, \\\\{\\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}\\\\}, \\\\{\\\\mathbf{z}\\\\_k\\\\} | \\\\mathcal{S})$ to replace the joint distribution $p(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}, \\\\{\\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}\\\\}, \\\\{\\\\mathbf{z}\\\\_k\\\\}, \\\\mathcal{S})$ in (P1). This change ensures that the distributions $p$ and $q$ are defined on the same support (parameters and representations), allowing for a consistent evaluation of the KL-divergence.\\n\\n**Q2**: Thanks for the comment. We agree with the reviewer that the presentation of (P1) may cause confusion for readers. In our approach, we use a surrogate distribution $q(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}, \\\\{\\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}\\\\}, \\\\{\\\\mathbf{z}\\\\_k\\\\})$ to approximate the distribution $p$. Consequently, estimating the parameters and feature centroids $(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}, \\\\{\\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}\\\\}, \\\\{\\\\mathbf{z}\\\\_k\\\\})$ involves maximizing the surrogate distribution $q$. Since each factor of distribution $q$ is either Gaussian distribution or GM distribution, the solution of the maximization is the mean of each Gaussian distribution (or each Gaussian component of the GM distribution). Thus, our primary focus is on updating the distribution to minimize the KL divergence between $p$ and $q$. The revised (P1) is provided as follows:\\n$$\\n\\\\begin{align}\\n\\\\mathrm{(P1)} \\\\min\\\\_{q(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}, \\\\{\\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}\\\\}, \\\\{\\\\mathbf{z}\\\\_k\\\\})} D\\\\_\\\\mathrm{KL} \\\\left(p(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}, \\\\{\\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}\\\\}, \\\\{\\\\mathbf{z}\\\\_k\\\\} | \\\\mathcal{S}) \\\\\\\\| q(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}, \\\\{\\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}\\\\}, \\\\{\\\\mathbf{z}\\\\_k\\\\}) \\\\right).\\n\\\\end{align}\\n$$\\n**Q3**: Thanks for the comment. Since each factor of distribution $q$ is either Gaussian distribution or GM distribution, the MAP estimate is taking the mean of each Gaussian distribution (or each Gaussian component of the GM distribution). To make this clear, we added extra clarity in Section 4.3.2.\\n\\n**Q4**: (P3) is not equivalent to (P2), but is an approximation to (P2) in the FL setting. Specifically, consider the objective function of (P2), given by\\n$$\\n\\\\begin{equation*}\\n \\\\text{(P2)} \\\\min\\\\_{q\\\\_n(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}, \\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}, \\\\{\\\\mathbf{z}\\\\_k\\\\})}\\n D\\\\_\\\\mathrm{KL}\\n \\\\left(p(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}, \\\\{\\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}\\\\}, \\\\{\\\\mathbf{z}\\\\_k\\\\} | \\\\mathcal{S})\\\\\\\\|\\n q\\\\_n(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}, \\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}, \\\\{\\\\mathbf{z}\\\\_k\\\\}) q\\\\_{-n}(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}) q\\\\_{-n}(\\\\{\\\\mathbf{z}\\\\_k\\\\}) \\n q\\\\_{-n}(\\\\{\\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}\\\\}) \\\\right).\\n\\\\end{equation*}\\n$$\\nIn the FL setting, each client can only assess the local dataset $\\\\mathcal{S}\\\\_n$, or a subset of training data sample set $\\\\mathcal{S}$. This limitation makes it difficult to update $q\\\\_n(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}), q(\\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}})$ by drawing samples of $\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}, \\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}$ from the joint distribution $p$ directly. To address this, we define the surrogate distribution $\\\\tilde{q}\\\\_{n}$ to approximate the joint distribution $p$ by fixing the cavity factors $q\\\\_{-n}(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}})$, $q\\\\_{-n}(\\\\{\\\\mathbf{z}\\\\_k\\\\})$, $q\\\\_{-n}(\\\\{\\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}\\\\})$ on the side of client $n$, given by\\n$$\\n\\\\begin{align*}\\n &\\\\tilde{q}\\\\_n(\\\\boldsymbol{\\\\theta }\\\\^{\\\\mathrm{b}},\\\\boldsymbol{\\\\theta }\\\\_n\\\\^{\\\\mathrm{h}},\\\\{\\\\mathbf{z}\\\\_{k}\\\\})\\n =p(\\\\mathcal{S}\\\\_n | \\\\boldsymbol{\\\\theta }\\\\^{\\\\mathrm{b}},\\\\boldsymbol{\\\\theta }\\\\_n\\\\^{\\\\mathrm{h}}) q\\\\_n(\\\\{\\\\mathbf{z}\\\\_{k}\\\\}) \\n q\\\\_{-n}(\\\\boldsymbol{\\\\theta}\\\\^{\\\\mathrm{b}}) q\\\\_{-n}(\\\\{\\\\mathbf{z}\\\\_k\\\\}) \\n q\\\\_{-n}(\\\\{\\\\boldsymbol{\\\\theta}\\\\_n\\\\^{\\\\mathrm{h}}\\\\}) ).\\n\\\\end{align*}\\n$$\\nBased on $\\\\tilde{q}\\\\_{n}$, we obtain the local optimization problem on client $n$ in Problem (P3).\"}", "{\"metareview\": \"A paper on an interesting topic, which unfortunately does not pass the bar for acceptance. In revising their manuscript, I strongly encourage the authors to take into account the comments of the reviewers, in particular those of T9rb that the authors have not thoroughly discussed, especially the question of formal guarantees. Such guarantees would be important to better sell the ideas of the paper, and I disagree on the authors' claim that they are out of scope of the paper.\", \"additional_comments_on_reviewer_discussion\": \"Though the authors' feedback came relatively late during the review process, they did a decent job at it, but failed to properly comment on comments on formal guarantees of their method (T9rb).\"}", "{\"comment\": \"**We thank Reviewer T9rb's comment.**\\n\\nWe understand the concern regarding the lack of theoretical guarantees for the proposed method. Below, we provide additional clarification.\\n\\n1. The guarantees of variational message passing are typically derived under certain constraints, such as marginalization or expectation constraints, to ensure the surrogate distribution closely approximates the target distribution. In our method, we employ expectation constraints, which align the mean and covariance of the surrogate distribution with those of the target distribution. The local updates in equations (10) and (11) and the aggregation in equation (13) are designed based on these expectation constraints.\\n2. Solving Problem (P2), a KL-divergence minimization problem, using variational message passing can be interpreted as an extended optimization problem that accounts for uncertainty. In the extreme case where uncertainty is entirely unknown, the KL-divergence minimization problem reduces to a conventional optimization problem. For example, FedPA [R2] addresses a federated least squares regression problem with a linear model, as presented in Equation (2). Additionally, Equation (3) demonstrates that the solution obtained by maximizing the likelihood distribution, i.e., the mean of the distribution, coincides with the solution derived from convex optimization.\\n3. According to prior literature [R1], rigorous convergence analyses of variational message passing are often based on state evolution, a theoretical tool used to track the dynamic behavior of variances during distribution updates. Since we estimate the variances (or covariances) of the distributions using sampling methods, the estimates may be biased due to the limited sample size. Consequently, existing works, such as FedPA [R2], rely on numerical experiments to evaluate the quality of the estimated gradient updates.\\n\\nBased on these discussions, we note that establishing theoretical guarantees is beyond the scope of this paper and is left for future exploration.\\n\\n[R1] Javanmard, Adel, and Andrea Montanari. \\\"State evolution for general approximate message passing algorithms, with applications to spatial coupling.\\\" *Information and Inference: A Journal of the IMA* 2.2 (2013): 115-144.\\n\\n[R2] MaruanAl-Shedivat, Jennifer Gillenwater, Eric Xing, and Afshin Rostamizadeh. Federated learning via posterior averaging: A new perspective and practical algorithms. ICLR, 2021\"}", "{\"comment\": \"***Minors***: We thank the reviewer for the attention to detail. We have thoroughly reviewed the paper and addressed all similar issues. The following corrections have been implemented:\\n\\n1. We have uploaded the completed version of the paper, including references and appendices.\\n2. We agree with the reviewer and the subscript is fully moved under the \\\"max\\\" operator in (1) and (P1).\\n3. We have fixed this error.\\n4. We agree with the reviewer and use \\\\bigcup in the revised manuscript.\\n5. We have fixed the inconsistent numbering.\\n6. The missing brackets have been added.\\n7. Equation numbering has been included for all problems.\\n8. The typo has been corrected.\"}", "{\"comment\": \"**We thank Reviewer g6Ef for careful reading and constructive comments.**\\n\\n**W1 Difference between pFedVMP and other Bayesian federated learning models** \\n\\nBFL can be broadly categorized into client-side BFL and server-side BFL in terms of FL architectures [R1]. Client-side BFL focuses on learning Bayesian local models on client nodes, while server-side BFL aggregates local updates for global models using Bayesian methods. Our paper belongs to server-side BFL. \\n\\nThe prior server-side BFL works related to our paper include FedPA, FedEP, QLSD, pFedBayes, which formulate model training as Bayesian inference tasks and aggregate the distributions of local **parameters**. FedPA (Al-Shedivat et al., 2021) approximated the posterior distribution into the product\\nof distributions with respect to local datasets during local model training. FedEP (Guo et al., 2023)\\ndeveloped the Bayesian model aggregation rule by using expectation propagation. QLSD (Vono\\net al., 2022) extended the approach in FedPA with the quantized Langevin stochastic dynamics for\\nlocal update. pFedBayes (Zhang et al., 2022) uses variational inference to approximate the posterior distribution of local model parameters for each client, and aggregates local models on the server.\\n\\nHowever, the above BFL methods do not utilize **global feature centroids** to guide local model training, which limits their ability to effectively address data heterogeneity. In contrast, pFedVMP considers both model **parameters** and **feature** centroids. The gain of pFedVMP comes from the following two folds: on one hand, pFedVMP guides the local training using a **regular** term of global feature centroids, decreasing the over-fitting risk of local training; on the other hand, pFedVMP achieves more **precise estimates** of the distributions of global feature centroid by variational message passing.\\n\\n[R1] Bayesian Federated Learning: A Survey, IJCAI-23\\n\\n**W2** Thanks for the constructive comment. As suggested by the reviewer, we have conduct the numerical experiments of comparing the proposed pFedVMP with the baselines. However, as the official implementation of pFedBayes is not publicly available, it was not included in the comparison.\\n\\n|Dataset|FMNIST(Dir(0.1))|FMNIST(Dir(0.3))|EMNIST(Dir(0.1))|EMNIST(Dir(0.3))|Cifar10(Dir(0.1))|Cifar10(Dir(0.3))|Cifar100(Dir(0.1))|Cifar100(Dir(0.3))|\\n|-------------|--------------------|--------------------|--------------------|--------------------|--------------------|--------------------|---------------------|---------------------|\\n|MOON|96.57|94.8|95.93|93.31|87.88|77.76|58.82|50.19|\\n|FedCP|96.87|93.87|95.95|92.81|86.97|72.96|59.90|47.96|\\n|PRIOR|96.64|94.21|95.66|93.06|86.39|74.42|54.37|44.89|\\n|**pFedVMP**|**97.23**|**95.60**|**96.97**|**95.09**|**88.12**|**80.81**|**64.22**|**56.75**|\\n\\n\\nAs shown in the table above, pFedVMP outperforms the baseline methods under various data partition settings. Although PRIOR leverages personalized prior knowledge, it does not utilize global representation information to guide local training. In contrast, both MOON and FedCP incorporate global representation information; however, MOON uses the similarity between local and global representations, whereas FedCP employs a conditional policy. Thanks to VMP, pFedVMP yields more precise estimates of the distributions of model parameters and global feature centroids, leading to its superior performance.\\n\\nIn the revised manuscript, we have cited these works.\\n\\n**W3** Thanks for the constructive comment. We understand that this phrase can be interpreted in various ways within the federated learning domain. In our work, \\\"second-order statistical information\\\" specifically refers to covariance matrices derived from the local feature centroids and local model parameters. To ensure clarity, we have revised the manuscript to explicitly define this term and eliminate potential ambiguity.\"}", "{\"summary\": \"The paper proposes a personalized federated learning algorithm based on Bayesian estimation. The core idea is that clients learn a global shared model while they train a personalized head.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"Personalization in federated learning by splitting the model into two parts such that one part is learned globally while the head is learned locally is proved to be effective. The contribution and motivation of the paper is clear.\", \"weaknesses\": \"The novelty of the paper is not high in my opinion. The core idea has been proposed before in the literature. One of the first work that I know which employ the same idea for personalized federated learning is FedRep of Collins et al., (2021). However, Collins et al., (2021) solves the problem using optimization. It seems that this paper solves the problem using Bayesian. Furthermore, Bayesian federated learning has been studied extensively before. Although, the paper compares the proposed algorithm against set of baselines, I think the paper misses the comparison with FedRep which is closely related to the study of this paper.\", \"questions\": \"I am not expert at Bayesian learning and I cannot evaluate the novelty of this work in this aspect. Can you explain the key differences between your proposed algorithm prior Bayesian federated learning? My concern is that if we can apply prior Bayesian federated learning to solve the problem which has been already studied by convex optimization techniques.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"**We thank Reviewer T9rb for careful reading and constructive comments.**\\n\\n**Typographical Error**: We sincerely apologize for the typographical error in the title and will correct \\\"Massage Passing\\\" to \\\"Message Passing\\\" in the revised manuscript.\\n\\n**W1/Q1 Bayesian benchmarks**: Thanks for this constructive comment. As suggested by the reviewer, we conducted experiments to compare the proposed pFedVMP with the benchmark methods for Bayesian federated learning mentioned by the reviewer, including FedPA, FedEP, QLSD, and pFedGP. Since the official implementation of pFedBayes is not publicly available, and BNFed is applicable only to simple feedforward neural networks, these methods were not included in the comparison. The results are presented in the following table.\\n\\n|Dataset|FMNIST(Dir(0.1))|FMNIST(Dir(0.3))|EMNIST(Dir(0.1))|EMNIST(Dir(0.3))|Cifar10(Dir(0.1))|Cifar10(Dir(0.3))|Cifar100(Dir(0.1))|Cifar100(Dir(0.3))|\\n|-------------|--------------------|--------------------|--------------------|--------------------|--------------------|--------------------|---------------------|---------------------|\\n|FedPA-FT|96.91|94.97|96.36|94.20|87.88|78.23|60.32|51.67|\\n|FedEP-FT|96.88|94.95|96.31|94.23|87.87|78.36|60.31|51.92|\\n|QLSD-FT|93.80|89.30|91.56|87.80|79.49|65.35|37.44|27.74|\\n|pFedGP|96.11|94.15|94.77|91.02|85.88|75.88|57.32|46.53|\\n|**pFedVMP**|**97.23**|**95.60**|**96.97**|**95.09**|**88.12**|**80.81**|**64.22**|**56.75**|\\n\\nThe table shows that pFedVMP achieves the highest test accuracy. This improvement stems from the ability of pFedVMP to guide local training through a regularization term based on global feature centroids, which reduces the risk of overfitting during local training. In contrast, FedPA, FedEP, QLSD, and pFedGP do not utilize global feature centroids to guide local model training.\\n\\nIn the revised manuscript, we have cited these works.\\n\\n**W2/Q2 Computational cost of pFedVMP**: We agree with the reviewer that the computational cost of matrix inversion may be high. However, pFedVMP primarily aims to enhance model utility, which necessitates additional computation cost when leverage more information. We believe that performance improvement serves as the primary motivation for federated learning. \\n\\nMeanwhile, we note that some existing Federated learning methods that leverages the second-order statistics information such as covariance matrix or precision matrix, such as pFedGP and FedPAC, have similar computational costs. Specifically, let $Z$ denote the number of feature dimensions. pFedGP trains a personalized Gaussian process classifier on each clients, which requires to training a Gaussian kernel and has a computational complexity at $\\\\mathcal{O}(Z\\\\^3)$. FedPAC requires to optimize a quadratic function w.r.t. the covariance matrices of features, which also has a computational complexity at $\\\\mathcal{O}(Z\\\\^3)$. These methods have similar computational cost with pFedVMP. \\n\\n**W3/Q3 Hyperparameter Selection**: Thanks for this constructive comment. Equations 8 and 10 contains the following hyperparameters: the penalty scalar $\\\\xi\\\\_1$ and the scalar $\\\\alpha$ that ensures the precision matrix $\\\\boldsymbol{\\\\Lambda}\\\\_{k,n}\\\\^\\\\mathrm{z}$ is full rank. In the following, we discuss the selection values in the numerical experiments.\\n\\n- The penalty scalar $\\\\xi\\\\_1$: In the previous paper, we have investigated the effect of the penalty scalar $\\\\xi\\\\_1$ on the learning performance of pFedVMP. The results are presented in Appendix C.3.\\n- The scalar $\\\\alpha$ is defined to maintain that the precision matrix is full rank. Since the non-zero singular values of the pesudo inverse of the covariance matrix $(\\\\boldsymbol{\\\\Sigma}\\\\_{k,n}\\\\^\\\\mathrm{z})\\\\^\\\\dagger$ is around $1e{3}$, we set the scalar $\\\\alpha = 1$ in the numerical experiment.\\n\\n**W4/Q4 Theoretical guarantees**: As **Reviewer WGea's** suggestion, we have included additional derivations related to the variational message passing framework in Section 4, which enhance the theoretical foundation of the proposed method. However, we note that establishing a comprehensive convergence guarantee for the variational message passing framework remains an open problem in the research field. To date, there is a lack of rigorous convergence analyses for such methods. Consequently, even Bayesian federated learning approaches like FedEP and pFedGP do not provide detailed convergence analyses. Addressing this challenge is beyond the scope of this paper and is left for future investigation.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"comment\": \"Thank you for the detailed responses and manuscript changes.\\nThe responses have adequately answered my questions and concerns, and I will raise my score accordingly.\", \"some_other_minor_issues_i_found_whilst_reading_the_relooking_at_the_manuscript\": [\"(1) is used for \\\"listing items\\\" in-text and also for equations. One may want to change this to avoid confusion.\", \"I would still recommend for the an avoidance of double equation numbering, eg, (P3) and (8a) + (8b). I think it would be better to have (P3a) and (P3b) only and have (P3) when referencing both. Of course, this a matter of taste.\", \"One may want to be mindful that some of the equation with multiple definitions on a single line are quite busy, eg, (4), (6), and (11). Some whitespace between each equation could help with readability.\"]}", "{\"comment\": \"**We sincerely thank Reviewer WGea for this positive feedback.**\\n\\nWe will further revise the manuscript to address these minor issues, including clarifying item and equation numbering, avoiding double equation numbering where possible, and improving the readability of equations by adding appropriate whitespace.\\nThese revisions will be included in the camera-ready version if the manuscript is accepted.\"}", "{\"comment\": \"**We sincerely thank Reviewer g6Ef for their positive feedback and for raising their rating.**\\n\\nWe also appreciate this constructive suggestion and will consider conducting additional experiments in the future.\"}", "{\"comment\": \"Thanks for the efforts in the responses. However, my primary concern for raising the rating is the lack of theoretical guarantees.\"}" ] }

[ "Pengxiang Zhao", "Hanyu Hu", "Ping Li", "Yi ZHENG", "Zhefeng Wang", "Xiaoming Yuan" ]

[ "large language models", "post-training pruning" ]

{ "note_id": [ "yp0tFJs1K7", "smO6BD1Ye6", "rENFp2wkME", "qMxoAjRcOg", "q5oVjSvnBP", "pqzpr2vrvy", "os3LWrlbos", "oV1WdxH0pT", "gSJy4xCear", "gAg5C6yYnM", "fipFjetTa5", "bsSXA8UqER", "bHKfc3fDFz", "YP7vWFTrqh", "XRzHANm0DE", "XBMRWjcvEl", "VL5zj28rHG", "ULKJk3UcDq", "PszoL2iwSt", "PQvPUZYGve", "OUyMc2hnpf", "Lg8PVKdTaH", "KwoBGNHsYu", "KkZXdtWMPh", "K0W8I0ml7P", "Ih0f0CWYnn", "FFmdyUpB8D", "EsvmYrNV43", "EjLAShxP00", "D8PEln50CE", "ClbPfa0tmO", "BB6ay5vOY3", "AhTZQOZrzP", "9WIvUPDnuv", "6tGQ86PGoO", "6EBJsAr9dL", "4mNUUjelG7", "1Gxyap0avc" ], "note_type": [ "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "decision", "official_comment", "official_review", "official_comment", "comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_review", "meta_review", "official_comment", "official_comment", "official_comment" ], "note_created": [ 1730701258249, 1732351204863, 1731845079113, 1732326889690, 1732353589311, 1731843837884, 1732163610658, 1732589101038, 1731737654560, 1731695197177, 1732441348947, 1731844829934, 1732678426232, 1732625433384, 1732350880112, 1732869912246, 1732871151260, 1732608286401, 1733313044780, 1732350574744, 1732608498708, 1737523917245, 1732710480696, 1730560083683, 1733192965509, 1731752288327, 1731844599682, 1731832698778, 1732350973434, 1733109442676, 1731486910493, 1730306960151, 1731845060400, 1730430657700, 1734760375400, 1733313061066, 1733109525918, 1732608623208 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission8547/Reviewer_68Xw" ], [ "ICLR.cc/2025/Conference/Submission8547/Authors" ], [ "ICLR.cc/2025/Conference/Submission8547/Authors" ], [ "ICLR.cc/2025/Conference/Submission8547/Reviewer_D4zK" ], [ "ICLR.cc/2025/Conference/Submission8547/Authors" ], [ "ICLR.cc/2025/Conference/Submission8547/Authors" ], [ "ICLR.cc/2025/Conference/Submission8547/Authors" ], [ "ICLR.cc/2025/Conference/Submission8547/Reviewer_68Xw" ], [ "ICLR.cc/2025/Conference/Submission8547/Authors" ], [ "~Vladimír_Boža1" ], [ "ICLR.cc/2025/Conference/Submission8547/Reviewer_EAKc" ], [ "ICLR.cc/2025/Conference/Submission8547/Authors" ], [ "ICLR.cc/2025/Conference/Submission8547/Reviewer_RWeE" ], [ "ICLR.cc/2025/Conference/Submission8547/Authors" ], [ "ICLR.cc/2025/Conference/Submission8547/Authors" ], [ "ICLR.cc/2025/Conference/Submission8547/Authors" ], [ "ICLR.cc/2025/Conference/Submission8547/Authors" ], [ "ICLR.cc/2025/Conference/Submission8547/Authors" ], [ "ICLR.cc/2025/Conference/Submission8547/Authors" ], [ "ICLR.cc/2025/Conference/Submission8547/Authors" ], [ "ICLR.cc/2025/Conference/Submission8547/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission8547/Authors" ], [ "ICLR.cc/2025/Conference/Submission8547/Reviewer_D4zK" ], [ "ICLR.cc/2025/Conference/Submission8547/Reviewer_EAKc" ], [ "~Vladimír_Boža1" ], [ "ICLR.cc/2025/Conference/Submission8547/Authors" ], [ "ICLR.cc/2025/Conference/Submission8547/Authors" ], [ "ICLR.cc/2025/Conference/Submission8547/Authors" ], [ "ICLR.cc/2025/Conference/Submission8547/Authors" ], [ "ICLR.cc/2025/Conference/Submission8547/Authors" ], [ "ICLR.cc/2025/Conference/Submission8547/Reviewer_EAKc" ], [ "ICLR.cc/2025/Conference/Submission8547/Authors" ], [ "ICLR.cc/2025/Conference/Submission8547/Reviewer_RWeE" ], [ "ICLR.cc/2025/Conference/Submission8547/Area_Chair_jFWg" ], [ "ICLR.cc/2025/Conference/Submission8547/Authors" ], [ "ICLR.cc/2025/Conference/Submission8547/Authors" ], [ "ICLR.cc/2025/Conference/Submission8547/Authors" ] ], "structured_content_str": [ "{\"summary\": \"The paper proposes FISTAPruner, a layer-wise pruning method designed for both unstructured and semi-structured pruning, targeting efficient sparsification of large language models (LLMs). This approach utilizes the FISTA method (Fast Iterative Shrinkage-Thresholding Algorithm) to facilitate efficient convergence. Additionally, it employs a LASSO-like convex optimization model to effectively enhance sparsity in LLMs. To address the cumulative output error between the full and pruned models due to the sequential output error transfer across transformer decoder layers, the authors utilize layer-wise pruning with an intra-layer error correction mechanism.\\n\\nExperiments conducted on various model sizes, ranging from 125M to 70B parameters\\u2014including OPT, LLaMA, LLaMA-2, and LLaMA-3\\u2014across datasets such as WikiText-2-raw, PTB, and C4, demonstrate that FISTAPruner outperforms existing baseline methods (e.g., SparseGPT, Wanda, Wanda+DSnoT, SparseGPT+PERP, and Wanda+PERP) in terms of model performance after pruning.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"(+) The method effectively incorporates FISTA for efficient pruning during the post-training process, leading to faster optimization and enhanced performance\\n(+) By effectively employing LASSO to identify pruned weights with targeted sparsity, the approach minimizes reliance on heuristic-based methods, thereby improving overall effectiveness in the pruning process.\\n(+) The authors enhance the proposed method by developing an algorithm that enables semi-structured pruning, allowing for practical acceleration on real-world hardware.\", \"weaknesses\": \"(-) The paper requires experiments to compare FISTAPruner with other methods that have similar computational costs. Existing baseline methods, such as SparseGPT and Wanda, do not involve retraining during pruning. In contrast, FISTAPruner conducts retraining in the process of finding W\\u2217. Although DSnoT and PERP are used instead of retraining, their computational costs are lower than the layer-by-layer approach employed in FISTAPruner. So, it is necessary to compare their performance under similar computational cost conditions (e.g., training on SparseGPT is performed layer by layer).\\n(-) The benefits of using FISTA over traditional gradient descent methods are not sufficiently explained, which may leave readers unclear about its specific advantages in this context.\", \"questions\": \"1. In line 89, the paper states, \\\"Our results confirm that FISTAPruner can efficiently create sparse networks from pretrained LLMs without retraining.\\\" However, the process of finding the pruned weights W* seems to function similarly to retraining. Could you clarify this point, as it may cause confusion for readers?\\n2. In line 306, the paper states, \\\"We treat each decoder layer as an independent pruning unit, enabling parallel pruning across multiple decoder layers on different devices.\\\" However, the proposed method conducts pruning sequentially. Can you explain how parallel pruning is achieved alongside sequential pruning? A more detailed explanation or revision would be helpful.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Part 4\", \"comment\": [\"**Response to Questions**\", \"**Response to Question 1**\", \"Thank you for your question.\", \"**First, we would like to clarify that SGD with or without momentum cannot directly solve the $\\\\ell_1$-regularized training problem.** This is because the $\\\\ell_1$-norm is non-smooth, and gradient-based methods like SGD are not well-suited for handling non-smooth objective functions. Instead, solving such problems typically requires subgradient methods or proximal gradient methods, such as FISTA, which are specifically designed to handle non-smooth optimization.\", \"Additionally, our proposed **FISTAPruner is a training-free pruning method**, which fundamentally differentiates it from training-based methods. Specifically, FISTAPruner optimizes a layer-wise pruning model with an $\\\\ell_1$-norm regularization term, as described in the paper. Unlike $\\\\ell_1$-regularized training approaches, **FISTAPruner does not require backpropagation** over the entire network. As we emphasized previously in our response to Weakness 2, this training-free nature makes FISTAPruner significantly more efficient than methods that optimize $\\\\ell_1$-regularized non-convex loss functions through training. Given these fundamental differences, we believe a direct comparison between FISTAPruner and training-based pruning methods would not provide meaningful insights.\", \"**Response to Question 2**\", \"Thank you for your question.\", \"We chose to evaluate FISTAPruner on OPT and LLaMA-1 models to ensure a fair and consistent comparison with the baseline methods reported in our paper, such as SparseGPT [1], Wanda [2], DSnoT [3], and PERP [4]. These works also evaluated their approaches on OPT and LLaMA-1. By using the same models, we aimed to provide a direct and meaningful comparison of FISTAPruner\\u2019s performance relative to these established baselines.\", \"To address the applicability of FISTAPruner to more recent models, we also included experiments on state-of-the-art models such as LLaMA-3-8B and LLaMA-3-70B, demonstrating that our method generalizes well to newer models as well.\", \"**Response to Question 4**\", \"Thank you for your question.\", \"First, DSNoT and PERP are fundamentally weight adjustment/retraining-based methods aimed at improving efficiency further for an existing pruning method, whereas our approach with FISTAPruner is intentionally designed to achieve optimal results without retraining. This distinction reflects our focus on demonstrating that it is possible to match or even surpass further adjustment/retraining-based methods without the need for additional retraining steps.\", \"We have conducted comparisons in several scenarios to highlight this point and have demonstrated that FISTAPruner achieves excellent performance without retraining. Therefore, integrating DSNoT or PERP with FISTAPruner would not align with our objective, as adding a retraining step (even if it improves results) does not serve to further validate our approach or its advantages.\", \"Regarding the tables, we believe the current structure effectively communicates our comparisons and findings. Combining the tables as suggested may blur the distinction between retraining-based methods and our retraining-free approach and baseline methods, which is central to our work.\", \"**Response to Question 5**\", \"Thank you for your question. The main point of the \\\"Warm Start\\\" section is to discuss the efficiency and effectiveness of different warm-start strategies within our pruning framework. The key takeaway is that we evaluate how various \\\"warm start\\\" strategies integrate with our FISTAPruner framework, concluding that using Wanda as a warm start is a cost-efficient and effective approach in practice.\", \"*[1] Frantar, Elias, and Dan Alistarh. \\\"Sparsegpt: Massive language models can be accurately pruned in one-shot.\\\"*\", \"*[2] Sun, Mingjie, et al. \\\"A Simple and Effective Pruning Approach for Large Language Models.\\\"*\", \"*[3] Yuxin Zhang, et al. \\\"Dynamic sparse no training: Training-free fine-tuning for sparse LLMs.\\\"*\", \"*[4] Max Zimmer, et al. \\\"Perp: Rethinking the prune-retrain paradigm in the era of LLMs.\\\"*\"]}", "{\"comment\": \"As shown in the results above, we summarize the PPL (lower is better) comparison across different sparsity levels as follows:\\n\\n- **5% and 10%:** intra- and inter- layer error correction $<$ intra-layer error correction only $<$ no error correction;\\n- **20%:** intra-layer error correction only $<$ intra- and inter- layer error correction $<$ no error correction;\\n- **50%:** intra-layer error correction only $<$ no error correction $<$ intra- and inter- layer error correction.\\n\\nFirst, the results confirm the effectiveness of our intra-layer error correction mechanism, as it consistently outperforms the no-error-correction approach.\\n\\nSecond, the results confirm the effectiveness of using both intra- and inter-layer error correction at low sparsity levels, as it consistently outperforms the intra-layer error correction alone at 5% and 10% sparsity.\\n\\nThird, the results show that using both intra- and inter-layer error correction is sensitive to sparsity levels and tends to perform worse at higher sparsity. Specifically, at 20% sparsity, it underperforms compared to intra-layer error correction alone, and at 50% sparsity, it even performs worse than the no-error-correction approach.\\n\\nTo explain why the use of both intra- and inter-layer error correction is sensitive to sparsity levels, we believe this occurs because higher sparsity levels make the pruning task more difficult, leading to greater error accumulation across layers. When both intra- and inter-layer error correction are applied, mitigating the accumulated error from previous layers may dominate the optimization objective in deeper layers, causing the pruning performance of the current layer to suffer. \\n\\nMathematically, let $W_k$ and $X_k$ represent the weight matrix and the activation of the $k$-th layer in the original network, respectively. Similarly, let $W_k^*$ and $X_k^*$ denote the pruned weight matrix and the corresponding activation in the pruned network. In a layer-wise pruning scheme with both intra- and inter-layer error correction mechanisms, we minimize the loss for each layer individually:\\n$$\\n\\\\\\\\|W_k^* X_k^*-W_kX_k\\\\\\\\|_F^2.\\n$$\\n\\n$X_k$ depends on the activation from the previous layer:\\n$$\\nX_k = f_k(W_{k-1}X_{k-1}).\\n$$\\n\\nwhere $f_k$ represents some operations (e.g., activation function or normalization). Therefore, we can express the pruned activations recursively as:\\n$$\\nX_k^* = f_k(W_{k-1}^* X_{k-1}^* ).\\n$$\\n\\nThe error at layer $k$ is defined as:\\n$$\\n\\\\Delta X_k = f_k(W_{k-1}^* X_{k-1}^* ) - f_k(W_{k-1}X_{k-1}).\\n$$\\n\\nTherefore, under high sparsity level, this amplification often results in the accumulated error $\\\\Delta X_k$ becoming dominant at deeper layers.\\n\\nThus, for large $k$, considering both intra- and inter-layer error correction mechanism, there is: \\n$$\\n \\\\\\\\|W_k^*(X_k+\\\\Delta X_k)-W_kX_k\\\\\\\\|_F^2 \\\\approx \\\\\\\\|W_k^*\\\\Delta X_k-W_kX_k\\\\\\\\|_F^2.\\n$$\\n\\nAs a result, the optimization process shifts focus towards correcting this accumulated error rather than pruning the current weight matrix $W_k$.\\n\\nIn other words, minimizing the term (14) primarily addresses the error correction from previous layers rather than properly pruning the weight matrix $W_k$, which negatively impacts the pruning performance in deeper layers.\"}", "{\"comment\": \"Thank you for the comprehensive responses which have adequately addressed the majority of my concerns. Based on the improvements and clarifications provided, I raise the evaluation score to 6.\"}", "{\"comment\": \"Thank you again for your insightful review, valuable feedback, and for updating the score! We welcome any further discussion and are open to any additional suggestions you might have.\"}", "{\"comment\": \"**Response to Major Weaknesses:**\\n\\n**1. Detailed Explanation of Intra-layer Error Correction Mechanism.**\\n\\nThank you for your comment and interest in our intra-layer error correction mechanism. Our primary goal is to minimize discrepancies between the pruned and original models by adjusting weights after pruning in a layer-wise manner. We would like to provide a detailed explanation of this mechanism here and update the manuscript accordingly.\\n\\nDenoting the pruned counterpart by $W^* \\\\in \\\\mathbb{R}^{m \\\\times n}$, a straightforward approach to quantify the output error is to use the Frobenius norm of the difference between the outputs from the dense and pruned weights\\n$$\\n ||W^*X - WX||_F, \\\\quad (1)\\n$$\\n\\nwhich serves as a metric of the pruning quality at the target sparsity level and is widely adopted by work such as SparseGPT.\\n\\nHowever, we observe that applying Equation (1) can lead to an error accumulation issue across sequential operators, as illustrated in Figure 2 of our manuscript. In the figure, $W_1$ and $W_2$ represent the weights of two sequential operators. Although Equation (1) effectively quantifies the output error between $W_1$ and its pruned counterpart $W_1^*$ since they are at the top of the layer and share the same inputs, issues arise when applying the same metric to the outputs of $W_2$ and $W_2^*$. Following Equation (1), the deviation between the outputs of $W_2$ and $W_2^*$ is computed with the same input $W_1X$. However, in a pruned network, the actual input for $W_2^*$ is $W_1^* X$, creating a discrepancy with $W_1X$ and thus leading to error propagation through the operators. To address this, we propose a method to sequentially prune weights within each pruning unit (e.g., a decoder layer of a Transformer), updating Equation (1) to:\\n$$\\n||W^* X^* - WX||_F, \\\\quad (2)\\n$$\\n\\nwhere $X^*$ represents the input feature activation for $W^*$ from the sequentially pruned network.\"}", "{\"comment\": \"**Response to Weakness 1 and Question 1**\\n\\nThank you for your comment. We acknowledge that both SparseGPT [1] and our work approach pruning as an optimization problem. However, there are critical distinctions that highlight the novelty of our contributions.\\n\\nFirst, while SparseGPT formulates the pruning objective as minimizing the Frobenius norm $||W^* X - WX||_F^2$, subject to sparsity constraints\\n\\n$$\\n\\\\\\\\min_{W^* } \\\\\\\\quad ||W^* X - WX||_F^2 \\n\\\\\\\\quad\\n\\\\\\\\mathrm{s.t.} \\\\\\\\quad\\nW^* \\\\\\\\in \\\\\\\\mathcal{C},\\n$$\\n\\nit does not incorporate error corrections into this minimization and the problem is inherently non-convex due to the sparsity constraint $ W^* \\\\in \\\\mathcal{C}$.\\n\\nMoreover, SparseGPT does not directly solve the above model but finds approximate solutions by a greedy procedure: following the OBS framework [2], it iteratively removes entries with the smallest impact on the objective function and updates the remaining weights. This method lacks optimality guarantees and may lead to unstable performance. \\n\\nIn contrast, our work introduces an innovative transformation of the non-convex model into a convex one by adding an $\\\\ell_1$-norm regularization term. We then utilize FISTA to find a global solution. This systematic approach not only enhances stability but also gives better pruning outcomes.\\n\\n\\n\\n**Response to Weakness 2 and Question 2**\\n\\nThank you for your comment. To understand why FISTAPruner achieves significant improvements in the 2:4 pruning setting, we conducted ablation studies specifically focused on 2:4 semi-structured sparsity.\", \"the_key_components_of_fistapruner_for_2\": \"4 semi-structured pruning are:\\n1. the mathematical formulation of the problem as an $\\\\ell_1$-norm regularized convex optimization problem where the parameter $\\\\lambda$ that balances the sparsity and the reconstruction error is determined by our proposed adaptive tuning algorithm; \\n2. the intra-layer error correction mechanism that avoids error accumulation;\\n3. the hard thresholding step after the FISTA iterations to project onto the 2:4 semi-structured space.\\n\\nWhile the first component includes several features, they are interdependent; removing any one of them will fail the system. The hard thresholding step is essential to guarantee that the final output adheres to the 2:4 sparsity structure, so ablation studies on this aspect would not yield meaningful insights. Consequently, the only feasible ablation study focuses on the intra-layer error correction mechanism.\\n\\nWe tested FISTAPruner with and without the proposed correction mechanism on OPT-125M, and found that error corrections are vital to the significant improvements of FISTAPruner. The results are summarized below:\\n\\n|OPT-125M under 2:4 Semi-structured Sparsity|WikiText|C4|PTB\\n|:-|:-:|:-:|:-:|\\n|Wanda|80.32|64.73|111.55|\\n|SparseGPT|60.02|52.11|94.21|\\n|FISTAPruner without intra-layer error corrections|55.14|48.53|90.57|\\n|FISTAPruner|45.16|38.08|67.80|\\n\\nWe could see that without intra-layer error corrections, FISTAPruner is still able to produce better results. However, incorporating the intra-layer error correction mechanism leads to a substantial drop in perplexity. Thus, we believe that intra-layer error corrections play an important role in the 2:4 semi-structured pruning. This is probably because 2:4 semi-structured sparsity brings more constraints to the sparsity structure so the error accumulation is also more evident. Thus, incorporating error corrections is able to give FISTAPruner the right direction to optimize and thus yields large improvements.\\n\\n\\n**Response to Weakness 3**\\nThank you for your feedback. We appreciate your concern regarding the nature of our ablation studies. We included the \\\"Amount of Calibration Data\\\" section to the ablation study part following SparseGPT [1]. The \\\"Warm Start\\\" section aimed to explore the effects of initializing the FISTA iterations at different points. However, we agree with you that these do not constitute proper ablation studies. According to your suggestion, we will move both sections to the discussion part of the paper.\\n\\n[1] Frantar, Elias, and Dan Alistarh. \\\"Sparsegpt: Massive language models can be accurately pruned in one-shot.\\\" International Conference on Machine Learning. PMLR, 2023.\\n\\n[2] Hassibi, Babak, David G. Stork, and Gregory J. Wolff. \\\"Optimal brain surgeon and general network pruning.\\\" IEEE international conference on neural networks. IEEE, 1993.\"}", "{\"title\": \"Thanks for your comment.\", \"comment\": \"It is difficult to agree that optimizing using FISTA should not be classified as retraining. Typically, the process of iteratively updating model parameters with respect to an objective function is referred to as model training. In this context, the model's predictions do not necessarily have to align with the objective function, and gradient descent is not a strict requirement for this process. For instance, the Forward-Forward algorithm[1] effectively facilitates learning at the layer level and is still categorized as model training. Additionally, in ZeroQuant[2], techniques such as layer-wise pruning are employed to update the parameters of the quantized model. However, this process is regarded as fine-tuning or Quantization-Aware Training (QAT).\\nAll concerns regarding the other questions have been resolved.\\n\\n[1] Hinton, Geoffrey. \\\"The forward-forward algorithm: Some preliminary investigations.\\\" arXiv preprint arXiv:2212.13345 (2022).\\n[2] Yao, Zhewei, et al. \\\"Zeroquant: Efficient and affordable post-training quantization for large-scale transformers.\\\" Advances in Neural Information Processing Systems 35 (2022): 27168-27183.\"}", "{\"comment\": \"Thank you for your comment regarding the work \\\"Fast and Effective Weight Update for Pruned Large Language Models\\\" (referred to as ADMM). We acknowledge this contribution and have included it in the related work section of our paper.\\n\\nHowever, it is important to highlight that there are differences in the experimental environments (e.g. different versions of PyTorch, different GPUs, etc.) between our study and the ADMM framework. In our tests, the performance of the baseline methods for most models is notably lower (with higher perplexity results) than that reported in ADMM. For instance, in our environment, SparseGPT and Wanda yield perplexity of 6.54 and 6.46, respectively, for LLaMA-2-7B, compared to 6.51 and 6.42 in ADMM\\u2019s environment. Thus we suspect that the results for the methods proposed in ADMM are taking advantage of the testing environment compared to ours. This discrepancy suggests that direct performance comparisons may not be entirely fair.\\n\\nFurthermore, even under these conditions, FISTAPruner outperforms ADMM across all other models reported, with the exception of LLaMA-2-7B, where their performances are quite similar. We believe this context is crucial for a comprehensive evaluation of the algorithms\\u2019 effectiveness.\"}", "{\"title\": \"There is a better layer-wise pruner than SparseGPT\", \"comment\": \"I just want to point out that \\\"Fast and Effective Weight Update for Pruned Large Language Models\\\" (https://openreview.net/forum?id=1hcpXd9Jir) proposed another pruning algorithm, which is based on ADMM, and it is better than SparseGPT.\\n\\nFor example, on Llama2-7B ADMM pruner gets better perplexity than FISTAPruner (ADMM has 6.33, FISTA has 6.35, SparseGPT has 6.54). Also, this ADMM-based pruner does not do any inter-layer correction, so FISTA has a kind of unfair advantage.\"}", "{\"comment\": \"Thank you for addressing my numerous questions. I reviewed the responses you provided to my queries. However, I remain unconvinced about some of the explanations, as outlined below:\\n\\n**[Response to Weakness 1]**\\n\\nYou state that existing methods do not solve problems in the form of ||WX - W\\\\*X\\\\*||, where W and X represent the weights and inputs of the original model, and W\\\\* and X\\\\* represent the compressed model's weights and inputs, respectively. However, this approach is widely used in model compression, as evidenced by the following GitHub repositories:\\n\\n* OmniQuant: https://github.com/OpenGVLab/OmniQuant\\n* K-prune: https://github.com/snudm-starlab/K-prune/\\n\\nThese repositories clearly demonstrate the use of the proposed error correction technique to minimize ||WX - W\\\\*X\\\\*||. Furthermore, the methodology is already reflected in Equation (1) of the following paper:\\n\\n* El Halabi, Marwa, Suraj Srinivas, and Simon Lacoste-Julien. \\\"Data-efficient structured pruning via submodular optimization.\\\" Advances in Neural Information Processing Systems 35 (2022): 36613-36626.\\n\\nTherefore, I find it difficult to agree with your claim that the error correction technique is a new methodology.\\n\\n\\n**[Response to Weakness 2]**\\n\\nRegarding the incorporation of the FISTA algorithm, I believe you have made two key assertions:\\n\\n1. **SGD cannot directly solve the L1-regularized problem.** \\n\\nYou stated that L1 regularization cannot be optimized due to the presence of non-differentiable points, making training infeasible. However, in deep learning, weights rarely converge to precisely zero, and simple modifications, such as replacing the gradient at such points with a constant, allow effective training. This approach is commonly used, as seen with ReLU activation functions, which also contain non-differentiable points but are trained effectively using SGD.\", \"the_following_paper_demonstrates_the_use_of_l1_regularization_with_sgd_for_training\": \"* Han, Song, et al. \\\"Learning both weights and connections for efficient neural network.\\\" Advances in Neural Information Processing Systems 28 (2015).\\n\\nAdditionally, methods for training with non-differentiable functions, such as L0-norm, are extensively discussed in the following works:\\n\\n* Louizos, Christos, Max Welling, and Diederik P. Kingma. \\\"Learning sparse neural networks through L0 regularization.\\\" arXiv preprint arXiv:1712.01312 (2017).\\n\\n* Wang, Ziheng, Jeremy Wohlwend, and Tao Lei. \\\"Structured pruning of large language models.\\\" arXiv preprint arXiv:1910.04732 (2019).\\n\\n* Xia, Mengzhou, Zexuan Zhong, and Danqi Chen. \\\"Structured pruning learns compact and accurate models.\\\" arXiv preprint arXiv:2204.00408 (2022).\\n\\nThese methods indicate that even block-level modifications to existing approaches are feasible for large language models, casting doubt on the claim that SGD cannot address L1-regularized problems.\\n\\n2. **No backpropagation.**\\n\\nYou highlight the lack of backpropagation as an advantage of your method. However, avoiding backpropagation alone cannot be considered a benefit unless it demonstrably reduces pruning time. Your paper states that the proposed method takes approximately 12 hours to prune the Llama-3 70B model. This duration is comparable to existing methods that perform SGD on a block-level basis with adjusted epochs. As such, I find it challenging to accept this as a distinct advantage.\\n\\n**[Response to Weakness 3]**\\n\\nWhile I appreciate your explanation, I disagree with the assertion that BOHB is unsuitable because it explores a wide range of hyperparameter combinations. In fact, its ability to search for diverse combinations is a strength, not a weakness.\\n\\nBayesian optimization algorithms like BOHB systematically predict promising hyperparameter combinations, evaluate them through model training, and update their prediction functions based on evaluation results. This approach is equally applicable to single hyperparameters and does not inherently disadvantage the algorithm.\\n\\nTo clarify the rationale for your proposed hyperparameter tuning method over BOHB, I believe you need to provide one of the following justifications: (a) BOHB performs worse than the proposed method. (b) BOHB is prohibitively expensive for pruning large language models.\\n\\nWithout such evidence, it is difficult to justify the exclusion of existing hyperparameter tuning methods.\\n\\nSeveral of my suggestions, including corrections for typographical errors, appear not to have been incorporated into the manuscript. If you agree with my feedback, I encourage making these adjustments during the discussion period.\\n\\nWhile I acknowledge the performance improvements achieved by your work, I remain unconvinced regarding certain claims about novelty and the methodology\\u2019s advantages over prior research. These issues remain unresolved, both in your responses and the manuscript, and I will maintain my current score. If you have further clarifications, I welcome your response.\"}", "{\"comment\": \"**Response to Major Question:**\\n\\nThanks for your question. To clarify, both Wanda and our method aim to minimize the discrepancy of the outputs between the pruned network and the original one under sparsity constraints.\\n\\nTo achieve this goal, Wanda employs a heuristic method to prune weights based on a pruning metric. They choose $|W| \\\\|X\\\\|_2$ as their metric, which intuitively considers both the magnitude of the weights $W$ and the input $X$. This metric is designed based on the insight that the computation in a linear layer is $WX$. Consequently, Wanda heuristically uses $|W| \\\\\\\\|X\\\\\\\\|_2$ to identify and prune weights that may have small impact on $WX$.\\n\\nHowever, our work formulates the layer-wise pruning problem as an optimization model and solves it to obtain the pruned weights. To do so, we include $\\\\\\\\|W^* X^* - WX\\\\\\\\|_F$ as part of our objective to minimize the discrepancy between the pruned network and the original model, while using the $\\\\ell_1$-norm to induce sparsity. Additionally, by solving our proposed model, we not only prune weights but also update the remaining weights to further minimize $\\\\\\\\|W^* X^* - WX\\\\\\\\|_F$.\\n\\nIn summary, here are the similarities and differences between the objectives of Wanda and our method:\\n\\n- **Similarity:** \\n 1. Both Wanda and our method aim to minimize the discrepancy between the outputs of the pruned network and the original model under sparsity constraints.\\n- **Difference:** \\n 1. Wanda\\u2019s pruning metric $|W| \\\\|X\\\\|_2$ is heuristically designed, while we integrate $\\\\\\\\|W^* X^* - WX\\\\\\\\|_F$ into our optimization model to minimize the real error in a systematic and optimization-driven manner.\\n 2. Wanda\\u2019s $|W| \\\\\\\\|X\\\\\\\\|_2$ directly prunes weights without the opportunity to update the remaining weights for further error minimization, whereas our method not only enforces sparsity but also updates the remaining weights at the same time to minimize the error.\"}", "{\"title\": \"Ack\", \"comment\": \"I thank the authors for presenting at least one real ablation study in the rebuttal. I do not think varying calibration data counts as an ablation study; real ablation study should involve binary interventions (e.g., include intra-layer correction or not) with the goal of assessing the causal effect of these binary interventions on the efficacy of the technique. I do not believe varying calibration data serves such as scientific purpose. It is a dire mistake to propagate terminological flaws of prior work as the author is doing.\\n\\nTo clarify, it's practically important to study calibration data, I simply disagree with calling it an ablation study. The inclusion of these fake ablation studies take up space that should be reserved for real ablations, which is absolutely necessary for empirical work like this.\\n\\nNevertheless, I will raise my score to 6 because these issues are relatively minor and should not stand in the way of acceptance.\"}", "{\"comment\": \"Thank you for your observations and the references provided. We appreciate the effort you have made to substantiate your perspective. Let us clarify the reasoning behind categorizing the FISTAPruner as distinct from pruning with retraining.\\n\\nAccording to the general definition of the community (e.g., survey papers of pruning [1-2]) and our Related Work section, we want to clarify \\\"pruning with retraining\\\" and \\\"pruning without retraining\\\":\\n\\n**Pruning with Retraining**\\n\\n**Definition:** These methods typically follow a three-step process: Pretrain-Prune-Retrain. After pruning, the model is fine-tuned or retrained to recover accuracy. This involves an additional independent stage where the pruned model, with its sparsity mask fixed, undergoes training of the remaining weights on the training dataset to recover performance lost due to pruning.\\n\\nClassical and recent examples include Lottery Ticket Hypothesis [3], LLMpruner [4] and [5-6], where fine-tuning or retraining is essential to optimize the pruned model.\\n\\n**Post-Training Pruning Without Retraining**\\n\\n**Definition:** These methods simplify the Pretrain-Prune-Retrain process into Pretrain-Prune, skipping the retraining step. They may also include compensatory mechanisms to mitigate accuracy loss directly during the pruning process, rather than relying on a subsequent retraining phase. For example, Wanda [7] prunes weights based on metrics like weight magnitude and input norms, without additional fine-tuning.\\nSparseGPT [8] operates within an Optimal Brain Surgeon [9] framework, removing weights and updating remaining weights simultaneously to minimize layer-wise accuracy loss.\\n\\nWe classify FISTAPruner as part of this category because it adheres to the Pretrain-Prune pipeline. Importantly:\\n\\n- FISTAPruner does not require fine-tuning the pruned model on the training dataset.\\n\\n- The solving process in FISTA is consistent with post-training pruning methods like SparseGPT, as it simultaneously removes weights and updates remaining weights to minimize degradation.\\n\\n- This process requires only 128 calibration data samples, rather than the entire training dataset, significantly reducing computational costs.\\n\\nBy avoiding the retraining phase and leveraging an efficient calibration approach, FISTAPruner aligns with the principles of post-training pruning without retraining.\\n\\nRegarding ZeroQuant [10], according to its Related Work section, the authors first review several Quantization-Aware Training (QAT) methods and highlight their limitations: \\\"More importantly, they require retraining or fine-tuning the full model to recover accuracy, and such compute costs for extra-large models are hardly affordable for most research labs or practitioners.\\\" Subsequently, they review Post-Training Quantization (PTQ) methods and propose a new, efficient, and cost-effective approach for post-training quantization (also highlighted in ZeroQuant's title).\\n\\nIn summary, while FISTA is an iterative method, its role is solely to solve the layer-wise pruning model (similar to the approach in SparseGPT) for layer-wise pruning. Considering our pruning pipeline, the limited amount of data used for pruning, and the significantly reduced computational cost, we firmly categorize FISTAPruner as a post-training pruning method without retraining.\\n\\n**Thank you again for your comment. We hope this explanation resolves your concerns, and we are happy to provide further clarifications if needed.**\\n\\n\\n[1] Cheng, et al. A Survey on Deep Neural Network Pruning: Taxonomy, Comparison, Analysis, and Recommendations.\\n\\n[2] Tang, et al. A Survey on Transformer Compression.\\n\\n[3] Frankle and Carbin, The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks.\\n\\n[4] Ma, et al. LLM-Pruner: On the Structural Pruning of Large Language Models.\\n\\n[5] Blalock, et al. What is the state of network pruning?\\n\\n[6] Liu, et al. Rethinking the Value of Network Pruning.\\n\\n[7] Sun, et al. A Simple and Effective Pruning Approach for Large Language Models.\\n\\n[8] Frantar and Alistarh, Sparsegpt: Massive language models can be accurately pruned in one-shot.\\n\\n[9] Hassibi, et al. Optimal brain surgeon and general network pruning.\\n\\n[10] Yao, et al. Zeroquant: Efficient and affordable post-training quantization for large-scale transformers.\"}", "{\"title\": \"Part 2\", \"comment\": \"**Response to Weaknesses 1 and 5 cont'd**\\n\\nAs shown in the results above, we summarize the PPL (lower is better) comparison across different sparsity levels as follows:\\n\\n**5% and 10%:** intra- and inter- layer error correction $<$ intra-layer error correction only $<$ no error correction;\\n\\n**20%:** intra-layer error correction only $<$ intra- and inter- layer error correction $<$ no error correction;\\n\\n**50%:** intra-layer error correction only $<$ no error correction $<$ intra- and inter- layer error correction.\\n\\nFirst, the results confirm the effectiveness of our intra-layer error correction mechanism, as it consistently outperforms the no-error-correction approach.\\n\\nSecond, the results confirm the effectiveness of using both intra- and inter-layer error correction at low sparsity levels, as it consistently outperforms the intra-layer error correction alone at 5% and 10% sparsity.\\n\\nThird, the results show that using both intra- and inter-layer error correction is sensitive to sparsity levels and tends to perform worse at higher sparsity. Specifically, at 20% sparsity, it underperforms compared to intra-layer error correction alone, and at 50% sparsity, it even performs worse than the no-error-correction approach.\\n\\nTo explain why the use of both intra- and inter-layer error correction is sensitive to sparsity levels, we believe this occurs because higher sparsity levels make the pruning task more difficult, leading to greater error accumulation across layers. When both intra- and inter-layer error correction are applied, mitigating the accumulated error from previous layers may dominate the optimization objective in deeper layers, causing the pruning performance of the current layer to suffer. \\n\\nMathematically, let $W_k$ and $X_k$ represent the weight matrix and the activation of the $k$-th layer in the original network, respectively. Similarly, let $W_k^*$ and $X_k^*$ denote the pruned weight matrix and the corresponding activation in the pruned network. In a layer-wise pruning scheme with both intra- and inter-layer error correction mechanisms, we minimize the loss for each layer individually:\\n$$\\n\\\\\\\\|W_k^* X_k^*-W_kX_k\\\\\\\\|_F^2.\\n$$\\n\\n$X_k$ depends on the activation from the previous layer:\\n$$\\nX_k = f_k(W_{k-1}X_{k-1}).\\n$$\\n\\nwhere $f_k$ represents some operations (e.g., activation function or normalization). Therefore, we can express the pruned activations recursively as:\\n$$\\nX_k^* = f_k(W_{k-1}^* X_{k-1}^* ).\\n$$\\n\\nThe error at layer $k$ is defined as:\\n$$\\n\\\\Delta X_k = f_k(W_{k-1}^* X_{k-1}^* ) - f_k(W_{k-1}X_{k-1}).\\n$$\\n\\nTherefore, under high sparsity level, this amplification often results in the accumulated error $\\\\Delta X_k$ becoming dominant at deeper layers.\\n\\nThus, for large $k$, considering both intra- and inter-layer error correction mechanism, there is: \\n$$\\n \\\\\\\\|W_k^*(X_k+\\\\Delta X_k)-W_kX_k\\\\\\\\|_F^2 \\\\approx \\\\\\\\|W_k^*\\\\Delta X_k-W_kX_k\\\\\\\\|_F^2.\\n$$\\n\\nAs a result, the optimization process shifts focus towards correcting this accumulated error rather than pruning the current weight matrix $W_k$.\\n\\nIn other words, minimizing the term primarily addresses the error correction from previous layers rather than properly pruning the weight matrix $W_k$, which negatively impacts the pruning performance in deeper layers.\"}", "{\"comment\": \"We hope this message finds you well. We sincerely appreciate the time and effort you have dedicated to reviewing our work.\\n\\nIn our previous response, we have carefully addressed the concerns you raised in your follow-up comments. We believe our revisions reflect the necessary adjustments, and we would be grateful for any further suggestions or follow-up questions you may have. If there are no further concerns, we would appreciate it if you could update the score accordingly.\\n\\nThank you once again for your thoughtful review. We look forward to your valuable feedback.\"}", "{\"comment\": \"We hope this message finds you well. We would like to express our sincere appreciation for your thoughtful review of our paper.\\n\\nIf you have any further questions or comments, or if there are any aspects of the paper that require additional clarification, we would be more than happy to address them. If there are no further concerns, we would greatly appreciate it if you could kindly update your score accordingly.\\n\\nThank you once again for your time and consideration.\"}", "{\"comment\": \"**Response to Weakness 1**\\n\\nThank you for your insightful feedback and for pointing out the related work that we initially missed. We acknowledge that error correction techniques, including those used in the repositories you mentioned (OmniQuant and K-prune), are indeed relevant to the discussion. We have discussed these works in the related work section (line 130 - 138) in the revised paper.\\n\\nHowever, we would like to emphasize that the unique contribution of our approach lies in the application of error correction specifically at the intra-layer level. This targeted method not only enhances accuracy but also facilitates parallelization, as detailed in our initial response to Weaknesses 1 and 5. We believe this distinction sets our work apart from existing methodologies. We have updated our paper to clarify this distinction and to better highlight the advantages of our intra-layer error correction mechanism (see lines 136-138, 180-182, 462-477, 933-1035).\\n\\n**Response to Weakness 2**\\n\\nThank you for your detailed comments. We appreciate the opportunity to clarify this point. \\n\\n- Regarding the capability of using SGD to solve non-smooth optimization problems, we agree that SGD can be applied with certain modifications, and our initial response explicitly stated that SGD cannot be **directly** applied to optimize $\\\\ell_1$-regularized problems. Specifically, in the examples you mentioned (e.g., replacing the gradient at non-differentiable points with constants), these approaches effectively replace the gradient with a subgradient, which aligns with our discussion of subgradient methods in the original response. However, it is important to note that subgradient-based approaches typically exhibit a slower convergence rate of $\\\\mathcal{O}(1/\\\\sqrt{k})$ compared to the $\\\\mathcal{O}(1/k)$ rate achieved in smooth optimization. In contrast, FISTA achieves an accelerated convergence rate of $\\\\mathcal{O}(1/k^2)$, as rigorously proven in [1] and mentioned in Remark 1 of our paper. We have included this analysis in lines 213-215.\\nFurthermore, we carefully examined the works you referenced, and we observe that many of these methods rely on relaxations of the non-smooth regularization terms (e.g., replacing the $\\\\ell_0$-norm with a differentiable approximation or surrogate loss). These relaxations enable the use of SGD on the modified, smoother objective function. As such, it is not that SGD itself is inherently well-suited for non-smooth optimization; rather, it becomes applicable through these relaxations or approximations.\\nIn summary, while SGD with modifications can be applied to certain non-smooth problems, these approaches are generally slower and less efficient than FISTA in terms of convergence rate. Moreover, many of the referenced works rely on problem-specific relaxations to enable SGD's use, which differs fundamentally from the proximal optimization framework employed by FISTA.\", \"title\": \"Part 1\"}", "{\"title\": \"Part 1\", \"comment\": \"Thank you for your response.\\n\\n**[Error Correction]**\\n\\nWe have already addressed this point in our previous response (Part 1 - Response to Weakness 1). Specifically, we acknowledged that error correction techniques, including those used in repositories such as OmniQuant and K-prune, are relevant to our discussion. These works are discussed in the related work section (lines 130-138) of the revised paper. However, we emphasize that the unique contribution of our approach lies in applying error correction at the intra-layer level, which offers both higher accuracy (compared with intra- and inter-layer error corrections) and parallelization advantages. This distinction is clearly stated in the revised paper (lines 136-138, 180-182, 462-477, 933-1035), as noted in our previous response.\\n\\nAdditionally, you mentioned that intra-layer error corrections are used in [1]. However, this is not accurate. In [1], the authors considered three methods: LayerInChange, SeqInChange, and AsymInChange, whose essential ideas correspond to (1) no error correction, (2) inter-layer error correction only, and (3) both intra- and inter-layer error corrections. Thus, intra-layer error correction was not explored in the way we apply it in our approach. Furthermore, [1] is designed for CNNs, whereas our work focuses on LLMs (transformer-based networks).\\n\\n[1] El Halabi, Marwa, Suraj Srinivas, and Simon Lacoste-Julien. \\\"Data-efficient structured pruning via submodular optimization.\\\" Advances in Neural Information Processing Systems 35 (2022): 36613-36626.\\n\\n\\n**[Use of FISTA]**\\n\\nFirst, we respectfully disagree with the comment, \\\"The efficiency of solving convex problems, such as SparseGPT, lies in its ability to prune large models like Llama-3 70B in approximately two hours.\\\" To clarify, SparseGPT does not solve a convex problem. Specifically, the problem they aim to solve, as outlined in Equation (1) in SparseGPT paper [1], is non-convex because it involves discrete mask variables. Moreover, SparseGPT utilizes the OBS framework, which is heuristic-based and operates in a greedy manner. As such, SparseGPT's efficiency does not stem from solving a convex problem, but rather from its heuristic approach, which yields approximate solutions to the underlying non-convex model.\\n\\nWhile the subtitle of this paragraph in your comments mentions the \\\"use of FISTA\\\", it appears that your primary concern is not the adoption of FISTA but rather the suggestion of applying SGD to a block-wise L1-norm regularization problem. To our knowledge, this is a novel hypothesis and has not been explored in previous work. It is important to note that introducing block-wise pruning will make the problem non-convex, which presents additional challenges compared to our layer-wise model, which is convex.\\n\\nWe also want to clarify that we have provided theoretical convergence results for SGD, SGD with subgradients, and FISTA in our previous response (Part 1 Response to Weakness 2) and in the revised paper (lines 204-209), demonstrating that FISTA offers the fastest convergence rate.\\n\\nAdditionally, our method focuses on the efficiency and scalability of layer-wise pruning using FISTA, which benefits from the convexity of the problem and provides strong theoretical guarantees for convergence.\\n\\n[1] Frantar, Elias, and Dan Alistarh. \\\"Sparsegpt: Massive language models can be accurately pruned in one-shot.\\\" International Conference on Machine Learning. PMLR, 2023.\\n\\n\\n**[Hyperparameter Optimization]**\\n\\nWhile we acknowledge that BOHB (or other similar works) is a generic method for hyperparameter tuning, we would like to emphasize that it is not always the most efficient or appropriate choice for every problem. In our case, we are dealing with a specific 1-dimensional search problem (tuning the sparsity parameter $\\\\lambda$, where the structure of the problem makes the bisection method far more efficient than BOHB.\\n\\nIn fact, we believe that for each algorithm and problem, it is important to design hyperparameter tuning methods that are tailored to the specific needs and properties of that problem. This approach not only leads to better performance but also ensures computational efficiency. While BOHB may work well for high-dimensional or complex problems, applying it to a simpler problem\\u2014like ours\\u2014would introduce unnecessary complexity and computational overhead, without offering any practical advantage. \\n\\nOur choice to use a specialized method for this particular scenario is, in our opinion, the most reasonable and effective way to proceed. We do not believe that it is necessary or productive to apply a generic algorithm like BOHB in every case.\"}", "{\"title\": \"Part 1\", \"comment\": \"**Response to Weaknesses**\\n\\n**Method**\\n\\n- **Response to Weaknesses 1 and 5**\\n - Thank you for your comment. While the similar concept of error correction has been explored in the context of CNNs (e.g., [1]). To the best of our knowledge, our approach first specifically tailors these principles to the unique architecture and requirements of LLMs. Existing work in LLM pruning, such as SparseGPT [2] and Wanda [3], do not include error corrections at all. To implement intra-layer error corrections, it is necessary to pass the original activation ($X$) to the original weight ($W$), while simultaneously passing the activation from the previously pruned operator ($X^* $) to the target pruned weight ($W^*$). This results in the model $||W^* X^* - WX||_F$.\\n Based on the code released by SparseGPT and Wanda, it is evident that they utilize the same activation within a decoder layer: $||W^*X - WX||_F$. Moreover, they forward the activation from one pruned decoder layer to the subsequent layer, complicating their pruning framework and creating dependencies across decoder layers.\\n - *[1] He, Yihui, Xiangyu Zhang, and Jian Sun. \\\"Channel pruning for accelerating very deep neural networks.\\\"*\\n - *[2] Frantar, Elias, and Dan Alistarh. \\\"Sparsegpt: Massive language models can be accurately pruned in one-shot.\\\" International Conference on Machine Learning. PMLR, 2023.*\\n - *[3] Sun, Mingjie, et al. \\\"A Simple and Effective Pruning Approach for Large Language Models.\\\"*\\n\\n \\n - We apply only the intra-layer error correction mechanism for two reasons:\\n 1. **Parallelization:** Intra-layer error correction enables independent pruning of each decoder layer, allowing us to distribute the pruning task across multiple devices by assigning different decoder layers to different devices. This will increase the overall pruning efficiency.\\n 2. **Sparsity Sensitivity:** While combining intra- and inter-layer error correction could intuitively reduce error accumulation across the network, we found that this approach is effective only at low sparsity levels. When the pruning task becomes harder (i.e., higher sparsity), global error correction tends to overshadow the pruning process of individual layers, ultimately leading to worse performance.\\n\\n The first reason is straightforward; we will explain the second reason in more detail below.\\n\\n We conducted a series of comparison experiments on OPT-125M at sparsity levels of 5%, 10%, 20%, and 50%. The experiments included three conditions: intra-layer error correction only, both intra- and inter-layer error correction, and no error correction. The results are presented in the following tables.\\n\\n |OPT-125M under 5% Sparsity|WikiText|C4|PTB\\n |:-|:-:|:-:|:-:|\\n |Intra-layer Error Correction Only|27.64|26.57|38.99|\\n |Intra-layer and Inter-layer Error Correction |27.63|26.56|38.98|\\n |No Error Correction|27.69|26.60|38.98|\\n\\n |OPT-125M under 10% Sparsity|WikiText|C4|PTB\\n |:-|:-:|:-:|:-:|\\n |Intra-layer Error Correction Only|27.47|26.59|39.00|\\n |Intra-layer and Inter-layer Error Correction |27.43|26.58|39.04\\n |No Error Correction|27.52|26.69|39.07|\\n\\n |OPT-125M under 20% Sparsity|WikiText|C4|PTB\\n |:-|:-:|:-:|:-:|\\n |Intra-layer Error Correction Only|27.36|26.71|39.39|\\n |Intra-layer and Inter-layer Error Correction |27.37|26.72|39.53|\\n |No Error Correction|27.61|26.91|39.85|\\n\\n |OPT-125M under 50% Sparsity|WikiText|C4|PTB\\n |:-|:-:|:-:|:-:|\\n |Intra-layer Error Correction Only|33.54|30.93|49.79|\\n |Intra-layer and Inter-layer Error Correction |35.90|32.93|55.24|\\n |No Error Correction|34.48|32.24|54.11|\"}", "{\"title\": \"Part 2\", \"comment\": \"**Response to Weakness 2 cont'd**\\n\\nYou mentioned that the pruning time for LLaMA-3-70B using our method (12 hours) is comparable to training-based methods with SGD applied on a block-level basis. However, it would be helpful to clarify the term \\\"block\\\" for better clarity on our end, as it may refer to either a subset of data (e.g., a batch of samples) or a group of parameters. Below, we address both interpretations:\\n\\n- If a block refers to a batch of samples.\", \"such_a_comparison_overlooks_a_critical_distinction\": [\"the hardware requirements for training-based methods are significantly higher than those for our approach. Training-based methods demand substantial GPU memory to store both the model parameters and optimizer states. For example:\", \"Storing the model parameters in FP16 precision for LLaMA-3-70B requires approximately **140GB** of GPU memory (70 billion parameters $\\\\times$ 2 bytes).\", \"Most optimizers store gradients in FP32 precision, which adds an extra **280GB** of GPU memory.\", \"Together, these two components alone demand at least **420GB** of GPU memory, necessitating a distributed setup with at least **11 NVIDIA A100 GPUs** (40GB each) to prune a model of this scale.\", \"In contrast, our method requires only a **single NVIDIA A100 GPU** with 40GB of memory to prune LLaMA-3-70B. This makes our method significantly more accessible and practical, especially in GPU-constrained scenarios. Furthermore, deploying our method in parallel across multiple GPUs reduces the pruning time for LLaMA-3-70B to approximately **1 hour**, demonstrating its scalability and flexibility in hardware configurations.\", \"The key advantages of our post-pruning method extend beyond avoiding backpropagation; they include drastic reductions in memory requirements and the flexibility to prune large-scale models on limited hardware. While training-based methods often necessitate expensive multi-GPU setups, our approach enables efficient pruning on a single GPU, making it more resource-efficient and accessible for practical deployment.\", \"If a block refers to a group of parameters\", \"This approach involves selecting specific blocks of parameters and then applying SGD to adjust them to derive a sparsity structure for pruning. However, this process introduces potential complexities. In particular, the selection of blocks appears to be a crucial and potentially challenging aspect. For example, the selection could:\", \"Be guided by a heuristic,\", \"rely on a predefined structure, or\", \"utilize some form of learned strategy.\", \"If we consider a layer-wise formulation (treating each operator in a decoder layer as a block, consistent with our optimization model) and apply SGD with subgradient modifications, the convergence rate will be $\\\\mathcal{O}(1/\\\\sqrt{k})$ (e.g. see the top of page nine of [2]). Using SGD with proximal steps for the non-smooth part results in iterations similar to ISTA, which achieves a convergence rate of $\\\\mathcal{O}(1/k)$ [1].\", \"In contrast, our method leverages FISTA, which attains a faster convergence rate of $\\\\mathcal{O}(1/k^2)$ [1]. This faster convergence rate directly translates to increased efficiency within our pruning framework, further distinguishing our approach from training-based methods.\", \"By addressing both interpretations of \\\"block,\\\" we hope to clarify the distinctions and demonstrate the specific advantages of our method in terms of resource efficiency, scalability, and performance. Please let us know if further elaboration is required.\", \"*[1] Beck A, Teboulle M. A fast iterative shrinkage-thresholding algorithm for linear inverse problems*\", \"*[2] https://web.stanford.edu/class/ee364b/lectures/subgrad_method_notes.pdf*\"]}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"comment\": \"We sincerely thank you for your detailed feedback and for raising the score, which recognizes the value of our work.\\n\\nIn light of your comments, we have revised the manuscript to clarify the terminology and improve the representation of this part. Specifically, we have ensured that the discussion of calibration data variation is no longer referred to as an \\\"ablation study.\\\" Instead, we describe it as the \\\"impact of calibration data and warm start.\\\"\\n\\nWe hope these revisions address your concerns and enhance the clarity of our paper. Thank you again for your constructive feedback and for considering our work. As a small note, we noticed that the score adjustment mentioned in your comments may not yet be reflected in the system, and we would appreciate your confirmation on this matter.\"}", "{\"summary\": \"The paper introduces FISTAPruner, a novel method for pruning large language models (LLMs) post-training to achieve significant sparsity, thereby reducing memory footprint and computational demands without compromising model performance. The authors introduce a LASSO-like convex optimization model tailored for layer-wise pruning, utilizing the Fast Iterative Shrinkage-Thresholding Algorithm (FISTA) to induce sparsity. A another innovation is the integration of an intra-layer error correction mechanism that mitigates cumulative errors across decoder layers during the pruning process. Additionally, FISTAPruner is extended to support 2:4 semi-structured pruning, aligning with hardware acceleration capabilities. Comprehensive experiments on various models (OPT, LLaMA) demonstrate that FISTAPruner outperforms state-of-the-art methods such as SparseGPT, Wanda, DSnoT, and PERP across multiple benchmarks, including perplexity and zero-shot task performance.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. The paper presents a unique approach by integrating FISTA with a LASSO-like model, which is innovative in the context of LLM pruning.\\n2. The results demonstrate that FISTAPruner can prune up to 50% of model parameters while retaining high accuracy, outperforming existing methods like SparseGPT and Wanda.\\n3. The integration of an intra-layer error correction mechanism is novel, which may avoid error cumulation.\", \"weaknesses\": \"**1. Major Weakness:** The intra-layer error correction mechanism is briefly mentioned but could benefit from a more detailed explanation and analysis. It raises the question of whether other methods (e.g., SparseGPT and Wanda) could achieve better performance if integrated with this mechanism.\", \"questions\": \"**1. Major Question:** In Wanda [1], they prune model weights by choosing the smallest $|W| ||X||_2$. While in your methods, you prune weights by minimizing the discrepancy of $||W^\\\\*X^\\\\*-WX||_2$. What's the difference between your sparsity objective with that of Wanda?\\n\\n**2. Minor Question:** In the paper, the error correction mechanism is applied solely within individual layers. Why is the error correction confined to intra-layer applications rather than being implemented across the entire model? In my understanding, extending the error correction mechanism globally could further mitigate the phenomenon of error accumulation throughout the network. \\n\\n\\n[1] A SIMPLE AND EFFECTIVE PRUNING APPROACH FOR LARGE LANGUAGE MODELS, ICLR 2024\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you for your detailed response. However, my major concerns regarding your main ideas remain unresolved.\\n\\n**[Error Correction]** \\n\\nThe authors have not conducted sufficient surveys on error correction algorithms in existing works. Specifically, the intra-layer error correction algorithm, which is one of the main ideas of this paper, is not a novel approach. It has already been studied in previous research, as demonstrated in [1].\\n\\n* [1] El Halabi, Marwa, Suraj Srinivas, and Simon Lacoste-Julien. \\\"Data-efficient structured pruning via submodular optimization.\\\" Advances in Neural Information Processing Systems 35 (2022): 36613-36626.\\n\\n**[Use of FISTA]** \\n\\nThe efficiency of solving convex problems, such as SparseGPT, lies in its ability to prune large models like Llama-3 70B in approximately two hours. If there is a more moderate time limits, such as 12 hours, block-wise pruning with SGD emerges as a strong competitor. Here, a block refers to the combination of Multi-Head Attention (MHA) and Multi-Layer Perceptron (MLP) modules.\\n\\nFISTAPruner requires nearly 12 hours for pruning, undermining its efficiency advantage. Furthermore, the authors do not compare their approach against a straightforward baseline: block-wise pruning with L1 regularization. Notably, block-wise pruning has the potential for higher accuracy, as it accounts for dependencies between linear operations within a block (see OmniQuant referenced above).\\n\\n**[Hyperparameter Optimization]** \\n\\nWhile the authors discuss the superiority of their hyperparameter search algorithm over BOHB, they provide no experimental results to substantiate this claim, not even from previously published studies. Moreover, BOHB is just one example among numerous existing hyperparameter search algorithms. The authors fail to justify the necessity of proposing a new hyperparameter optimization algorithm rather than leveraging established methods. Additionally, the paper lacks an analysis of the impact of their hyperparameter optimization algorithm on the overall results.\\n\\nThe authors rely solely on their assertions to respond during the long discussion period, without providing additional experiments or surveys on prior research. Furthermore, their responses include many statements that are difficult to agree with, such as \\\"L1 loss cannot be trained using SGD.\\\" Therefore, I believe my concerns have not been adequately addressed, and I will maintain my rejection score.\"}", "{\"title\": \".\", \"comment\": \"Thank you for the response.\", \"re_number_discrepancy\": \"Since the dense results are the same, the discrepancy is not caused by different sequence lengths used in the evaluation (a typical problem in many cases). So this leaves a different environment as the probable cause (and we see this all the time), which is fine.\"}", "{\"comment\": \"**2. The Impact of Integration with Other Methods.**\\n\\nRegarding your question about the impact of integrating the proposed intra-layer error correction mechanism with other pruning methods, we would like to clarify that its effect depends on whether the layer-wise pruning methods include a \\\"weight update\\\" stage.\\n\\nFor methods without a \\\"weight update\\\" stage, such as Magnitude Pruning and Wanda, weights are directly removed based on a pruning metric. As a result, there is no opportunity to mitigate the error introduced by pruning through updating remaining weights, nor to apply our intra-layer error correction mechanism.\\n\\nFor methods with a \\\"weight update\\\" stage, such as SparseGPT, integration with our intra-layer error correction mechanism is feasible. To apply this mechanism, we need to update their error measurement from Equation (1) to Equation (2) and derive the corresponding mathematics. Here, we use SparseGPT as an example to demonstrate how to apply this mechanism. In the following, we write $W_i$ to denote the $i$-th row of the weight matrix $W$ and $\\\\delta w$ to denote the pruning and weight updates on the row. It suffices to focus on a single row as the problems between different rows are independent.\\n$$\\n (\\\\text{original})\\\\quad \\\\min \\\\frac{1}{2}||(W_i+\\\\delta w)X-W_iX||_2^2, \\\\quad \\\\text{s.t.} \\\\quad \\\\delta w e_q + w_q = 0. \\\\quad (3)\\n$$\\n$$\\n \\\\to\\n$$\\n$$\\n (\\\\text{updated})\\\\quad \\\\min \\\\frac{1}{2}||(W_i+\\\\delta w)X^*-W_iX||_2^2, \\\\quad \\\\text{s.t.} \\\\quad \\\\delta w e_q + w_q = 0. \\\\quad (4)\\n$$\\n$$\\n \\\\iff \\n$$\\n$$\\n \\\\min \\\\frac{1}{2} \\\\delta w X^* X^{* \\\\top} \\\\delta w^\\\\top + \\\\delta w X^* X^{* \\\\top} W_i^\\\\top - \\\\delta w X^* X^\\\\top W_i^\\\\top \\\\quad (5)\\n$$\\n$$\\n \\\\text{s.t.} \\\\quad \\\\delta w e_q + w_q = 0.\\n$$\\n\\nGiven Equation (5), we write the Lagrangian function as follows:\\n$$\\n \\\\mathcal{L} = \\\\frac{1}{2} \\\\delta w X^* X^{* \\\\top} \\\\delta w^\\\\top + \\\\delta w (X^* X^{* \\\\top} - X^* X^\\\\top)W_i^\\\\top + \\\\lambda (\\\\delta w e_q + w_q). \\\\quad (6)\\n$$\\n\\nThen, we have\\n$$\\n X^* X^{* \\\\top} \\\\delta w^\\\\top + (X^* X^{* \\\\top} - X^* X^\\\\top)W_i^\\\\top + \\\\lambda e_q = 0; \\n$$\\n$$\\n \\\\delta w e_q + w_q = 0.\\n$$\\n\\nSolving this system, we have\\n$$\\n \\\\delta w = (-W_i(X^* X^{* \\\\top}-XX^{* \\\\top}) - \\\\frac{w_q-W_i(X^* X^{* \\\\top}-XX^{* \\\\top})(X^* X^{* \\\\top})^{-1}e_qe_q^\\\\top}{e_q^\\\\top (X^* X^{* \\\\top})^{-1}e_q})(X^* X^{* \\\\top})^{-1} \\\\quad (7)\\n$$\\nDenote \\n$$\\n a = -W_i(X^* X^{* \\\\top}-XX^{* \\\\top}) - \\\\frac{w_q-W_i(X^* X^{* \\\\top}-XX^{* \\\\top})(X^* X^{* \\\\top})^{-1}e_qe_q^\\\\top}{e_q^\\\\top (X^* X^{* \\\\top})^{-1}e_q}, \\\\quad (8)\\n$$\\nthen, we have\\n$$\\n \\\\delta \\\\mathcal{L} = \\\\frac{1}{2}a(X^* X^{* \\\\top})^{-1}a^\\\\top + a (X^* X^{* \\\\top})^{-1}(X^* X^{* \\\\top}-X^* X^{\\\\top})W_i^\\\\top. \\\\quad (9)\\n$$\\n\\nIn summary, to apply our mechanism in SparseGPT, we first use Equation (9) as the new pruning metric to determine which elements to prune, and then apply Equation (7) to update the remaining weights.\\n\\nAlthough the derivation is mathematically sound, implementing it within SparseGPT's framework presents significant challenges. SparseGPT\\u2019s pruning approach is designed for computational efficiency, relying on block-wise pruning and approximations of the Hessian matrix and its inverse. Integrating our correction mechanism necessitates additional matrix computations, which could compromise these efficiency goals. Despite the computational demands, we modified SparseGPT's code and evaluated its performance on OPT-125M and LLaMA-3-8B models using the WikiText dataset under 50% sparsity. The results are as follows:\\n\\n||OPT-125M|LLaMA-3-8B|\\n|:-|:-:|:-:|\\n|SparseGPT|37.01|8.64|55.38|\\n|SparseGPT + Intra-layer Error Correction|36.83|8.58|\\n|FISTAPruner|33.54|8.00|\\n\\nThese results validate the effectiveness of our intra-layer error correction mechanism and the FISTAPruner method.\"}", "{\"comment\": \"We agree with you that different programming environments will lead to slightly different results. Thanks again for your comment!\"}", "{\"title\": \"Part 3\", \"comment\": [\"**Response to Weakness 2**\", \"Thank you for your comment. While it is true that our method leverages the existing FISTA algorithm and employs $l_1$-norm regularization, we believe the novelty of our approach lies in the following aspects:\", \"**No back propagations:** A distinguishing feature of our approach is that it eliminates the need for computing the gradient by back propagations, making it highly efficient and particularly suitable for scenarios with limited computational resources. This stands in contrast to typical methods leveraging the $l_1$-norm, which often integrate it as a regularization term into the training loss function to induce sparsity and the pruning process requires expensive back propagations to optimize the loss function [1-2]. We have discussed this type of methods in the Related Work section in lines 109-113.\", \"**Optimization-driven layer-wise pruning scheme with error correction:** Unlike typical post-training method which rely on heuristic approaches or non-convex constraints, our work is the first to leverage the $l_1$-norm to induce sparsity in layer-wise pruning of LLMs. Besides, we introduce an intra-layer error correction mechanism tailored specifically to the LLM pruning domain. While FISTA is a natural choice for solving our formulated model, we have also designed a hyperparameter tuning algorithm to ensure that it achieves the desired sparsity levels and structures effectively.\", \"**Practical relevance:** Integrating all our innovations, we achieve state-of-the-art performance on comprehensive benchmarks, as demonstrated in our experiments.\", \"*[1] Thu Dinh, Bao Wang, Andrea Bertozzi, Stanley Osher, and Jack Xin. Sparsity meets robustness: Channel pruning for the feynman-kac formalism principled robust deep neural nets.*\", \"*[2] Xiufeng Xie, Riccardo Gherardi, Zhihong Pan, and Stephen Huang. Hollownerf: Pruning hashgridbased nerfs with trainable collision mitigation.*\", \"**Response to Weakness 3 and Question 3**\", \"Thank you for your feedback. The primary objective of our proposed adaptive hyperparameter tuning algorithm is to efficiently determine the optimal value of $\\\\lambda$ that balances the trade-off between sparsity and reconstruction error. As outlined in Section 3.4, we designed a bisection-based tuning algorithm tailored for this purpose. We have rigorously proved in Theorem 3.3 (we appreciate your observation regarding the numbering issue and will address it in the next revision) that with this algorithm, FISTAPruner is guaranteed to converge to the desired sparsity level.\", \"We acknowledge the existence of many hyperparameter tuning algorithms for machine learning, including BOHB [1], which you suggested for comparison. However, our problem setting differs significantly from the one addressed by such methods. Specifically, in our framework, we only need to determine a single scalar parameter, $\\\\lambda \\\\in \\\\mathbb{R}$, for each layer these parameters are independent of one another. This reduces the task to a one-dimensional parameter search problem, which can be effectively and efficiently solved using our bisection-based approach. ln contrast, methods like BOHB are designed to optimize over high-dimensional parameter spaces, where the goal is to identify a good combination of multiple parameters. As such, these methods are not ideally suited to our setting.\", \"*[1] Falkner, Stefan, Aaron Klein, and Frank Hutter. \\\"BOHB: Robust and efficient hyperparameter optimization at scale.\\\"*\", \"**Writing**\", \"**Response to Weakness 4**\", \"Thank you for your feedback. We appreciate your suggestion to provide additional explanation about FISTA to enhance understanding. While we agree that more background on FISTA might benefit some readers, it is a well-established method, and detailing its fundamentals may not be necessary. Instead, we have focused on our primary contribution: the development of a novel model for layer-wise pruning and the efficient integration of FISTA with our proposed intra-layer error correction mechanism and hyperparameter tuning algorithm. For completeness, we have included detailed mathematical derivations and an introduction to the FISTA iterations in Appendix B.\", \"**Response to Weakness 6**\", \"Thank you for your suggestions regarding the writing format and style. We will carefully address all these points in the revised version of our paper.\", \"**Experiments**\", \"**Response to Weakness 7**\", \"Thank you for your feedback. With respect, we do not agree with your suggestions regarding the comparisons of FISTAPruner and structured pruning methods. FISTAPruner is specifically designed for unstructured pruning and 2:4 semi-structured pruning, which are fundamentally different domains from structured pruning. Comparing our method with algorithms designed for structured pruning would not provide a fair or meaningful evaluation, as the constraints differ significantly between these approaches.\"]}", "{\"comment\": \"Dear Reviewer 68Xw,\\n\\nHi, as the discussion period is set to end soon, we kindly ask if you have any further questions or comments, or if there are any aspects of the paper that require additional clarification. Your feedback is invaluable to us, and we want to ensure we address any remaining concerns. \\n\\nIf there are no further concerns, we would greatly appreciate it if you could kindly update your score accordingly.\\n\\n\\nBest regards, \\\\\\nAuthors 8547\"}", "{\"comment\": \"**Response to Weakness 1:**\\n\\nThank you for your comment. We would like to clarify that FISTAPruner does not involve retraining during pruning; instead, it utilizes a layer-wise pruning approach, similar to methods like SparseGPT and Wanda. To clarify, retraining typically refers to the process of re-optimizing the entire model, usually by fine-tuning it on a specific loss function with respect to a training dataset after pruning. In contrast, FISTAPruner prunes the model in a layer-wise manner without such re-optimization, making it comparable to SparseGPT and Wanda, which also perform layer-wise pruning without retraining.\\n\\nWe understand the concern about computational costs, particularly in comparison with heuristic-based pruning methods like SparseGPT, Wanda, and DSnoT. While FISTAPruner may require slightly more pruning time due to its optimization-based approach, its key contribution is the introduction of an optimization model that systematically addresses the post-training pruning problem, leading to significantly better performance. This performance gain comes with an associated increase in pruning time, but the trade-off is justified by the improved outcomes.\\n\\nFurthermore, while PERP (Parameter-Efficient Retraining after Pruning) is used for retraining pruned models, we have included comparisons between FISTAPruner and SparseGPT+PERP, as well as Wanda+PERP, in Table 4. The results show that FISTAPruner, which does not involve any retraining, outperforms SparseGPT and Wanda with retraining via PERP. Additionally, our method is compatible with retraining approaches and can provide a better initialization point for the retraining process, should that be required.\\n\\nIn conclusion, although FISTAPruner may take more time for pruning compared to heuristic methods, the substantial performance improvements it provides justify this additional time cost. Moreover, since our method does not involve retraining, the overall time cost remains competitive. We believe that, given the enhanced performance and the potential for further optimization, the time cost should not be viewed as a major concern.\\n\\n\\n\\n**Response to Weakness 2:**\\n\\nThank you for your comment. As mentioned in Lines 188--190, our model incorporates $l_1$-norm regularization to encourage sparsity. Mathematically, the $l_1$-norm is non-smooth, which prevents the direct application of traditional gradient descent methods for optimization. Furthermore, although the sub-gradient descent method could be used, its convergence rate is $\\\\mathcal{O}(1/\\\\sqrt{k})$, which is slower compared to FISTA, which achieves a faster convergence rate of $\\\\mathcal{O}(1/k^2)$, as mentioned in Lines 229--231.\\n\\n\\n\\n**Response to Question 1:**\\n\\nThank you for your question. The key difference between retraining and our approach lies in the optimization objective. Retraining typically involves optimizing a model using a loss function that depends on the model's predictions, which is highly non-convex and requires computationally expensive backpropagation to compute gradients. In contrast, our approach uses a layer-wise error measured by the Frobenius norm in the objective function. This error is convex, and its gradient has closed-form expressions, making the optimization both convex and computationally more efficient. Therefore, while both methods aim to adjust the model, our approach does not involve retraining the model in the traditional sense of fine-tuning using a loss function.\\n\\n\\n**Response to Question 2:**\\n\\nThank you for your question. We indeed conduct sequential pruning within each decoder layer, specifically among the linear operators (K, Q, V, Out projections, and MLPs), to mitigate error accumulation through our intra-layer error correction mechanism. However, to enhance pruning efficiency, we treat each decoder layer as an independent pruning unit, enabling parallel pruning across multiple decoder layers on different devices. In other words, while pruning is performed sequentially within each individual decoder layer, the pruning process can occur in parallel across different layers, which allows us to strike a balance between performance and efficiency. We hope this clarifies how parallel pruning is achieved alongside the sequential pruning within layers.\\n\\n\\n**We appreciate your thoughtful comments and will ensure that all the concerns and questions raised, including the clarification of the pruning process, are addressed more clearly in the revised version of the paper.**\"}", "{\"summary\": \"This paper proposes FISTAPruner, an accurate pruning algorithm for large language models (LLMs).\\nThe main ideas of FISTAPruner are (1) intra-layer error correction, (2) FISTA-based optimization algorithm, and (3) adaptive hyperparameter tuning algorithm.\\nThe authors conduct exhaustive experiments to verify the effectiveness of FISTAPruner, and find that FISTAPruner is more accurate than existing algorithms; specifically, it shows almost 5% higher average accuracy on zero-shot tasks when pruning Llama-3 70B.\\nThe main strength of this paper lies in its high accuracy and exhaustive amounts of experiments.\\nHowever, the novelty and writing quality of this paper are insufficient.\", \"soundness\": \"2\", \"presentation\": \"1\", \"contribution\": \"1\", \"strengths\": \"The main strengths of this paper are as follows:\\n\\n1. The authors achieve meaningful accuracy improvement in diverse settings. For example, FISTAPruner shows almost 5% higher accuracy than the second-best algorithm, i.e., SparseGPT, when pruning Llama-3 70B models.\\n\\n2. This paper conducts extensive experiments covering diverse models from OPT to Llama-3 to show the robustness of FISTAPruner. FISTAPruner consistently shows comparable or the highest accuracy (or the lowest perplexity) in all cases.\\n\\n3. The figures in this paper are straightforward to understand.\", \"weaknesses\": \"I summarize the weakness of this paper below. I use the symbols [M] and [m] for each numbering to distinguish between major and minor weaknesses.\\n\\n### Method\\nThe main weakness of this paper is the lack of originality (or novelty). We summarize the weaknesses of the proposed method as follows.\\n1. [M] Error correction, the first idea, is just using the output of the pruned previous linear operators, and this idea is already used in previous works. Furthermore, the authors ignore the \\\"inter-layer errors\\\" induced by the pruning of previous layers when they correct errors.\\n\\n2. [M] The authors make use of the existing optimization algorithm, FISTA, without any modification. Introducing L1 regularization for pruning is a prevalent idea and there is no novelty.\\n\\n3. [M] The authors propose a new hyperparameter tuning algorithm, which has no specific name, but there is no explanation of the strength or novelty of this algorithm. There are no experiments that compare the performance of this algorithm with previous hyperparameter tuning algorithms.\\n\\n\\n### Writing\\nThe followings are the weaknesses in writing.\\n\\n4. [M] The main contribution of this paper is to use FISTA algorithm to prune LLMs. However, explanation about FISTA is too insufficient. It would better introduce the basics of FISTA in Section 2 (Background) and explain the modification to use FISTA for pruning LLMs in Section 3.2.\\n\\n5. [M] According to \\\"1.\\\", it is hard to agree with the statement \\\"Instead of pruning each operator in isolation like existing works\\\" in line 148.\\n\\n6. [m] Minor issues in writing:\\n\\n 6.1 (line 193) \\\"The proposed optimization model 3\\\" -> \\\"The proposed optimization model in Equation 3\\\"\\n\\n 6.2 (line 262) \\\"Theorem 3.3\\\" -> \\\"Theorem 1\\\"\\n\\n 6.3 (All equations) Use bold texts for representing matrices and vectors following the guideline of ICLR. It would be better to use blackboard bold S for representing a set of permissible sparsity patterns in Equation 1.\\n\\n 6.4 (All tables) Move captions of tables above the tables following the guideline of ICLR.\\n\\n 6.5 This paper does not contain the \\\"Reproducibility Statement\\\" which is encouraged by ICLR.\\n\\n 6.6 (Table 6) There are too many bold texts in Table 6.\\n\\n\\n### Experiments\\n\\n7. [m] The authors compare the performance of FISTAPruner with limited competitors without justification. The authors should compare the performance of FISTAPruner with structured pruning algorithms [1,2] or justify their selection of competitors.\\n\\n* References are at the end of this review\", \"questions\": \"1. What's the difference between FISTA, and L1-regularized training using SGD w/ momentum?\\n\\n2. Is there any reason you use outdated models such as OPT and Llama-1? How about using the latest models such as Phi, Gemma, and Mistral, if you want to use diverse models?\\n\\n3. Could you compare the performance of your \\\"Adaptive hyperparameter tuning\\\" algorithm with existing hyperparameter search algorithms, e.g. BOHB [3]?\\n\\n4. Are DSNoT and PERP (1) competitors or (2) compatible algorithms? If (1) competitors, then how about integrating Tables 2 to 4 as a single table? If (2) compatible algorithm, then how about integrating Tables 3 and 4? In this case, it would be better to compare the performance of \\\"FISTAPruner\\\" with \\\"FISTAPruner + DSnoT\\\" and \\\"FISTAPruner + PERP\\\" to show the compatibility.\\n\\n5. What is the main point of the Section \\\"Warm Start\\\"? Could you clarify the takeaway of this section?\\n\\n### References\\n[1] Ma, Xinyin, Gongfan Fang, and Xinchao Wang. \\\"Llm-pruner: On the structural pruning of large language models.\\\" Advances in neural information processing systems 36 (2023): 21702-21720.\\n\\n[2] Song, Jiwon, et al. \\\"SLEB: Streamlining LLMs through Redundancy Verification and Elimination of Transformer Blocks.\\\" arXiv preprint arXiv:2402.09025 (2024).\\n\\n[3] Falkner, Stefan, Aaron Klein, and Frank Hutter. \\\"BOHB: Robust and efficient hyperparameter optimization at scale.\\\" International conference on machine learning. PMLR, 2018.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"**Response to Minor Question:**\\n\\nThank you for your question. We apply only the intra-layer error correction mechanism for two reasons:\\n1. **Parallelization:** Intra-layer error correction enables independent pruning of each decoder layer, allowing us to distribute the pruning task across multiple devices by assigning different decoder layers to different devices. This will increase the overall pruning efficiency.\\n2. **Sparsity Sensitivity:** While combining intra- and inter-layer error correction could intuitively reduce error accumulation across the network, we found that this approach is effective only at low sparsity levels. When the pruning task becomes harder (i.e., higher sparsity), global error correction tends to overshadow the pruning process of individual layers, ultimately leading to worse performance.\\n\\nThe first reason is straightforward; we will explain the second reason in more detail below.\\n\\nWe conducted a series of comparison experiments on OPT-125M at sparsity levels of 5%, 10%, 20%, and 50%. The experiments included three conditions: intra-layer error correction only, both intra- and inter-layer error correction, and no error correction. The results are presented in the following tables.\\n\\n|OPT-125M under 5% Sparsity|WikiText|C4|PTB\\n|:-|:-:|:-:|:-:|\\n|Intra-layer Error Correction Only|27.64|26.57|38.99|\\n|Intra-layer and Inter-layer Error Correction |27.63|26.56|38.98|\\n|No Error Correction|27.69|26.60|38.98|\\n\\n|OPT-125M under 10% Sparsity|WikiText|C4|PTB\\n|:-|:-:|:-:|:-:|\\n|Intra-layer Error Correction Only|27.47|26.59|39.00|\\n|Intra-layer and Inter-layer Error Correction |27.43|26.58|39.04\\n|No Error Correction|27.52|26.69|39.07|\\n\\n|OPT-125M under 20% Sparsity|WikiText|C4|PTB\\n|:-|:-:|:-:|:-:|\\n|Intra-layer Error Correction Only|27.36|26.71|39.39|\\n|Intra-layer and Inter-layer Error Correction |27.37|26.72|39.53|\\n|No Error Correction|27.61|26.91|39.85|\\n\\n|OPT-125M under 50% Sparsity|WikiText|C4|PTB\\n|:-|:-:|:-:|:-:|\\n|Intra-layer Error Correction Only|33.54|30.93|49.79|\\n|Intra-layer and Inter-layer Error Correction |35.90|32.93|55.24|\\n|No Error Correction|34.48|32.24|54.11|\"}", "{\"summary\": \"The author proposed an LLM pruning algorithm that uses \\u201cFISTA\\u201d (Fast Iterative Shrinkage-Thresholding Algorithm) to identify optimal pruning masks. The author demonstrates the utility of the proposed technique across many state-of-the-art LLMs and with both structured and unstructured sparsity. The improvement over prior art is, however, small.\", \"soundness\": \"3\", \"presentation\": \"4\", \"contribution\": \"2\", \"strengths\": [\"This work is theoretically grounded, and provide some guarantees on convergence time.\", \"This work shows strong results in structured 2:4 pruning setup.\", \"This paper is overall well-written and easy to understand.\"], \"weaknesses\": [\"It\\u2019s unclear to me what\\u2019s new in this work relative to, say, SparseGPT, which also sets up pruning as an optimization problem and generally yields similar results as this work in unstructured pruning setup. It occurs to me that the fundamental difference appears to be that this work uses a different optimizer to solve essentially the same problem.\", \"While in structured 2:4 pruning setup this work yields substantial improvement, it is unclear why this is the case.\", \"Neither \\u201cAmount of Calibration Data\\u201d nor \\u201cWarm Start\\u201d is actually ablation study. Please do proper ablation studies by removing specific features of your algorithm design.\"], \"i_am_willing_to_raise_my_score_if_the_authors_can_deliver_real_ablation_studies_that_pinpoints_why_this_proposed_algorithm_achieved_superior_performance_in_2\": \"4 structured sparsity setup.\", \"questions\": [\"Question:\", \"Can you discuss difference between this work and sparseGPT?\", \"Can you perform ablation studies on 2:4 structured sparsity\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"metareview\": \"This work presents a novel approach to pruning large language models (LLMs) post-training. The authors claim that their method, FISTAPruner, allows for layer-wise pruning that effectively reduces model size while minimizing performance degradation. They assert that this approach can lead to significant reductions in computational requirements, making LLMs more efficient for deployment in resource-constrained environments. The findings indicate that FISTAPruner achieves competitive performance compared to existing pruning techniques, with empirical results showing a balance between model size reduction and task performance retention.\\n\\nHowever, reviewers raised concerns regarding the methodology, empirical validation, and overall contribution of the paper to the field. Given the weaknesses, I recommend rejecting this paper. While it presents an interesting concept aimed at optimizing large language models through layer-wise post-training pruning, it fails to provide compelling empirical evidence or rigorous theoretical justification necessary to support its claims effectively. Further work is required to address these issues for the next version of this work.\", \"additional_comments_on_reviewer_discussion\": [\"**Points Raised by Reviewers**\", \"During the review process, several key points were raised:\", \"Need for Robust Empirical Results: Reviewers requested more extensive experiments to validate the effectiveness of FISTAPruner across different tasks and datasets.\", \"Comparative Analysis: There was a strong recommendation for including comparisons with a wider array of existing pruning methods to contextualize the performance claims.\", \"Theoretical Insights: Reviewers sought a deeper theoretical explanation for why layer-wise pruning would be advantageous over traditional methods.\", \"Experimental Detail: Concerns were raised about insufficient details regarding experimental protocols and reproducibility.\", \"**Authors' Responses**\"], \"the_authors_attempted_to_address_these_concerns_during_the_rebuttal_period_but_did_not_sufficiently_strengthen_their_submission\": [\"They provided some additional experimental results; however, these were still considered inadequate by reviewers as they did not significantly enhance the robustness or breadth of their claims.\", \"While some comparisons with existing methods were included in their response, they remained limited and did not convincingly demonstrate superior performance.\", \"The theoretical justification provided was minimal and did not adequately clarify why their approach would yield better results than traditional methods.\", \"The authors attempted to clarify experimental details but still left several aspects vague, particularly concerning hyperparameter settings.\", \"**Weighing Each Point**\"], \"in_weighing_these_points_for_my_final_decision\": [\"The lack of robust empirical validation remained a critical issue that overshadowed any potential strengths of the proposed method.\", \"Inadequate comparative analysis with existing techniques hindered the ability to assess the true value of their contributions.\", \"Insufficient theoretical grounding left significant questions unanswered regarding the efficacy and applicability of their approach.\", \"The unresolved reproducibility issues further diminished confidence in their findings.\"]}", "{\"title\": \"Part 2\", \"comment\": \"**[Other Comments]**\\n\\nWe believe the proposed comparisons and additional experiments suggested are not necessary. We have already conducted a thorough analysis of the relevant literature. The specific experiments suggested (such as comparing FISTA and SGD on solving L1-norm regularization questions, comparing bisection and BOHB on 1-dimension parameter searching) do not seem to provide meaningful insights or further substantiate our claims. In fact, conducting them would be a diversion from the central focus of our work.\\n\\nIn response to your concern regarding the statement, \\\"L1 loss cannot be trained using SGD,\\\" we would like to reiterate and clarify our position. We have consistently stated in our first-round response (Part 4 - Response to Question 1) and our second-round response (Part 1 - Response to Weakness 2) that SGD with or without momentum cannot directly solve the non-smooth training problem associated with the $\\\\ell_1$-norm. While SGD can, in some cases, be applied with techniques like subgradients, we emphasize that these methods typically lead to slower convergence rates and do not achieve the same efficiency as methods specifically designed for non-smooth objectives. Nonetheless, we believe that this issue is minor and not directly related to our work.\"}", "{\"comment\": \"Dear Reviewer EAKc,\\n\\nHi, as the discussion period is set to end soon, we kindly ask if you have any further questions or comments, or if there are any aspects of the paper that require additional clarification. Your feedback is invaluable to us, and we want to ensure we address any remaining concerns. \\n\\nIf there are no further concerns, we would greatly appreciate it if you could kindly update your score accordingly.\\n\\nBest regards, \\\\\\nAuthors 8547\"}", "{\"title\": \"Part 3\", \"comment\": \"**Response to Weakness 3**\\n\\nThank you for your further comments. \\n\\nWhen comparing **BOHB** and the **bisection method** for a 1-dimensional search problem, such as finding a number in an ordered sequence (as in our scenario for tuning $\\\\lambda$), the **bisection method** is generally better and more efficient. Here's why:\\n\\n**Nature of the Problem**\\n- **Bisection Method**: The bisection method is specifically designed for ordered sequences or continuous 1-dimensional search problems. It uses the structure of the problem (e.g., ordering or monotonicity) to halve the search space in each step, guaranteeing $\\\\mathcal{O}(\\\\log(n))$ time complexity.\\n- **BOHB**: BOHB is a more general-purpose hyperparameter optimization algorithm designed to handle complex, high-dimensional, and noisy objective functions. It does not inherently leverage the ordered structure of a 1-dimensional search space, making it less efficient for this problem. Besides, BOHB has no convergence guarantee while we have proved that our bisection based tuning algorithm will lead the system converge to the desired sparsity in our paper (as detailed in Appendix C).\\n\\n**Efficiency**\\n- **Bisection Method**: Each step eliminates half of the search space, leading to a very rapid convergence. It is deterministic and computationally inexpensive.\\n- **BOHB**: While BOHB is robust and versatile, it involves maintaining probabilistic models and sampling based on those models, which introduces additional computational overhead. For a simple 1-dimensional search, this overhead is unnecessary and inefficient.\\n\\n**Applicability**\\n- **Bisection Method**: Works only for ordered or monotonic functions in 1 dimension.\\n- **BOHB**: Can handle non-monotonic, noisy, and multi-dimensional spaces, making it suitable for more complex scenarios but overkill for this simple case.\\n\\n**Accuracy**\\n- **Bisection Method**: Guarantees finding the exact value (or as close as desired for continuous problems).\\n- **BOHB**: Relies on probabilistic models, which may not be as precise in finding the exact value, especially for very simple problems like this.\\n\\n\\nIn summary, we believe that for the problem of finding a number in an ordered sequence, the **bisection method** is unequivocally better. It is faster, simpler, and more tailored to the nature of our problem. **BOHB** would only make sense if the search space were noisy, high-dimensional, or required optimization over a more complex objective function.\\n\\n\\n**We sincerely thank you for your valuable comments and feedback and apologize for the delay in updating our manuscript. A revised version has now been uploaded, incorporating your suggestions. We welcome further discussion and are open to any additional suggestions you may have.**\"}" ] }

[ "Zafeirios Fountas", "Martin Benfeghoul", "Adnan Oomerjee", "Fenia Christopoulou", "Gerasimos Lampouras", "Haitham Bou Ammar", "Jun Wang" ]

[ "large language models", "long context", "retrieval", "episodic memory", "event cognition", "training-free" ]

{ "note_id": [ "x7E8tdnAuc", "w33ZsdxOzE", "qhCE3D1gVN", "qcxf7Rc7an", "qAmLxNOMBR", "lGZHRZLDs9", "lDZJvsM4tm", "ji5XFHc5PD", "jcOBNK5WzP", "ihnJUmQiRa", "ialPSQ7j7K", "hgAcWfh8Ln", "h7TmfgtUTX", "gbE9tWRzbO", "f2FlQ3nxyh", "ePUnLrUBQZ", "cLLCIc0ghK", "beKcoE7Cc0", "ZNuM74Mepr", "YooR1c3Bhb", "YmWpDGMKxU", "UzJzUSWaBT", "RgNDruhXts", "Oe5WcItzAL", "L6QySfDixu", "EyCrl4M3RW", "9NNZ8J157G", "8r4NV8TcM5", "5DJ8DJwT44", "4l8RMvkZMc", "4hMjEZDG3g", "0o1Njiy05p" ], "note_type": [ "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "meta_review", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "decision", "official_comment", "official_comment", "official_review", "official_comment", "official_comment" ], "note_created": [ 1732738631332, 1730215375916, 1732035878644, 1732036838045, 1732639425075, 1732039051287, 1732035725165, 1732694822455, 1732039649057, 1732036778572, 1732037192296, 1732038349076, 1732038306900, 1730596236778, 1732662385548, 1732792265791, 1732037129116, 1732882917640, 1732639340467, 1732038530773, 1734731111405, 1732646555251, 1730652795503, 1732035860695, 1732309698375, 1732694751452, 1737524207320, 1732639477144, 1732038395230, 1729777461209, 1732648353093, 1732038989625 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission12669/Authors" ], [ "ICLR.cc/2025/Conference/Submission12669/Reviewer_RkzP" ], [ "ICLR.cc/2025/Conference/Submission12669/Authors" ], [ "ICLR.cc/2025/Conference/Submission12669/Authors" ], [ "ICLR.cc/2025/Conference/Submission12669/Reviewer_SyTr" ], [ "ICLR.cc/2025/Conference/Submission12669/Authors" ], [ "ICLR.cc/2025/Conference/Submission12669/Authors" ], [ "ICLR.cc/2025/Conference/Submission12669/Reviewer_SyTr" ], [ "ICLR.cc/2025/Conference/Submission12669/Authors" ], [ "ICLR.cc/2025/Conference/Submission12669/Authors" ], [ "ICLR.cc/2025/Conference/Submission12669/Authors" ], [ "ICLR.cc/2025/Conference/Submission12669/Authors" ], [ "ICLR.cc/2025/Conference/Submission12669/Authors" ], [ "ICLR.cc/2025/Conference/Submission12669/Reviewer_pMUj" ], [ "ICLR.cc/2025/Conference/Submission12669/Reviewer_pMUj" ], [ "ICLR.cc/2025/Conference/Submission12669/Reviewer_N12r" ], [ "ICLR.cc/2025/Conference/Submission12669/Authors" ], [ "ICLR.cc/2025/Conference/Submission12669/Authors" ], [ "ICLR.cc/2025/Conference/Submission12669/Reviewer_SyTr" ], [ "ICLR.cc/2025/Conference/Submission12669/Authors" ], [ "ICLR.cc/2025/Conference/Submission12669/Area_Chair_2wN3" ], [ "ICLR.cc/2025/Conference/Submission12669/Authors" ], [ "ICLR.cc/2025/Conference/Submission12669/Reviewer_SyTr" ], [ "ICLR.cc/2025/Conference/Submission12669/Authors" ], [ "ICLR.cc/2025/Conference/Submission12669/Authors" ], [ "ICLR.cc/2025/Conference/Submission12669/Reviewer_SyTr" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission12669/Reviewer_SyTr" ], [ "ICLR.cc/2025/Conference/Submission12669/Authors" ], [ "ICLR.cc/2025/Conference/Submission12669/Reviewer_N12r" ], [ "ICLR.cc/2025/Conference/Submission12669/Authors" ], [ "ICLR.cc/2025/Conference/Submission12669/Authors" ] ], "structured_content_str": [ "{\"comment\": \"Thank you for your quick response and continued feedback. Following your suggestions, we have:\\n\\n1. Amended the terminology throughout the paper (including **Title**, **Abstract**, **Introduction**, and **Discussion**) to 'episodic memory-inspired' rather than claiming human-likeness.\\n2. Added an expanded discussion of declarative memory systems, potential controversies and future extensions (**Appendix E.1.4**, **E.4**, and **E.5**), incorporating insights noted by reviewer pMUj regarding the cognitive science foundations of our approach (https://openreview.net/forum?id=BI2int5SAC&noteId=f2FlQ3nxyh).\\n\\nWe appreciate your thoughtful engagement with our work and remain open to any additional suggestions you may have.\"}", "{\"summary\": \"This paper introduces EM-LLM, a method for extending the context window of LLMs by incorporating principles from human episodic memory and event cognition. The method segments input sequences into episodic events using a combination of Bayesian surprise and graph-theoretic boundary refinement. It implements a two-stage memory retrieval process that combines similarity-based and temporally contiguous retrieval. The authors evaluate their method on LongBench and InfiniteBench benchmarks, comparing against state-of-the-art retrieval models like InfLLM and RAG.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"The paper proposes the idea of equipping LLMs with cognitive science principles (episodic memory and event cognition), which aligns well with current challenges in long-context processing;\\nThe use of Bayesian surprise for dynamic event segmentation in LLMs is novel, moving beyond simple fixed-length segmentation used in prior work like InfLLM; \\nThe method achieves 100% accuracy on Passkey.Retrieval task with sequences up to 5M tokens, demonstrating practical scalability; \\nThe paper provides a scalable alternative to both RAG and full-context approaches, outperforming them on most tasks while maintaining computational efficiency, increasing its practical value.\", \"weaknesses\": \"1.The paper claims surprise-based segmentation is superior to fixed-length approaches (like InfLLM), but lacks theoretical justification for why this should be effective. While Figure 4 shows some empirical correlation with human segmentation, there's no analysis of why this leads to better LLM performance. The authors should provide a theoretical analysis showing why this metric captures semantically meaningful boundaries better than alternatives.\\n2.The boundary refinement process (Algorithm 1) lacks convergence analysis. Given that it's iteratively adjusting boundaries based on modularity/conductance metrics, it is important to establish guarantees about stability and convergence. \\n3.The two-stage memory retrieval process (Section 3.4) introduces ks and kc parameters for similarity and contiguity buffers, but it does not justify the choice of values or analyze their trade-offs. \\n4.Although Table 1 shows some improvements over InfLLM, the improvements on most tasks may not be statistically significant. \\n5.The human correlation study in Section 4.2 relies on 3 short podcasts, which is insufficient to make broad claims about alignment with human event segmentation. The authors should expand the section to a larger, more diverse dataset of human annotations. \\n6.Section 4.4 discusses the impact of contiguity but does not properly ablate different buffer sizes or analyze the trade-off between similarity and contiguity buffers.\", \"questions\": \"1.How sensitive is the method to the choice of the surprise threshold parameter? What is the process for tuning this parameter?\\n2.How do you ensure the stability of the boundary refinement process? Are there cases where it fails to converge? \\n3.Does the issue of needing to query too many matching events still exist when the context window is too long? \\n4.If we use the Surprise-based method to segment the context into a RAG format, will the performance remain the same? \\n5.How does the method handle dynamic changes in context relevance over time?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"#### B. Human-like Information Retrieval\\nThe reviewer correctly notes that similarity-based retrieval is not unique to EM. However, our key contribution is integrating this with decaying temporal contiguity, reflecting both human cued recall and free recall patterns (Howard & Kahana, 2002). This combination enables our model to exhibit both temporal contiguity effects and temporal asymmetry - behavioral patterns consistently observed in human EM retrieval. In other words, while similarity-based retrieval is indeed general to many memory systems, our integration with contiguity buffer and temporal organization of memory blocks results in behavioural patterns that both align with key characteristics of human episodic memory and address previously unsolved challenges in long-context LLM tasks.\\n\\n### 3. Limitations and Future Work\\nWe acknowledge that EM-LLM differs from human episodic memory in several important ways (analysed in Appendix E.1). The most important ones we can identify include: (a) the non-parametric nature of our method, (b) the lack of hierarchical event structure, (c) the limited cross-modal integration and, finally, (d) the absence of a memory consolidation framework. However, we believe these limitations represent opportunities for future work rather than fundamental flaws in our approach.\\n\\nWe would appreciate the reviewer's thoughts on whether this analysis adequately addresses their concerns about our model's connection to human episodic memory. Given these limitations and our emphasis on claiming similarities rather than equivalence with human episodic memory, if the reviewer feels our justification above still doesn't adequately support the term \\\"human-like\\\", we would be happy to modify our framing to \\\"human-inspired\\\" episodic memory throughout the paper.\\n\\n### Summary of changes made to the paper:\\n1. We have introduced the **Appendix E.1** section to clarify our points here.\\n2. We have added a reference to **Appendix E.1** to the first paragraph of the **Discussion**, which aims to compare EM-LLM to human studies of episodic memory.\", \"title\": \"(Answer 3/3)\"}", "{\"title\": \"(Answer 2/4)\", \"comment\": [\"**Event segmentation:** Differentiable event segmentation models have already demonstrated the feasibility of learning a temporal structure from continuous experience. Models like SEM (Franklin et al., 2023) show how neural networks can combine with probabilistic inference to capture human-like event segmentation, while approaches like the DLH (Zakharov et al., 2022b) demonstrate that neural architectures can learn to identify hierarchical temporal boundaries through differentiable clustering and amortised variational inference. For instance, using the VaDE trick in (Jiang et al., 2017). These approaches offer powerful advantages in terms of learned representations and flexibility, potentially capturing the complex hierarchical event structure of real environments and adapting to different domains. Particularly compelling advantages include the ability to perform layer-wise or attention-head-wise segmentation and the potential emergence of nested timescale structures, as demonstrated in (Zakharov et al., 2022a;b), mirroring how the brain processes events at multiple temporal scales (Baldassano et al., 2017). While such end-to-end training is theoretically appealing and mirrors how neural circuits might learn temporal structure, our method takes a more pragmatic approach by leveraging the pre-trained capabilities of LLMs. By using Bayesian surprise computed directly from model outputs to detect event boundaries, we achieve efficient segmentation without requiring complex architectural modifications or additional training, while still aligning with cognitive theories about prediction errors in event perception (Zacks et al., 2007).\", \"**Retrieval:** The development of neural architectures for memory retrieval has evolved from classical Hopfield networks (Hopfield, 1982) through several key innovations. Early Hopfield networks demonstrated how content-addressable memory could emerge from simple neural circuits, paralleling biological memory systems. This was significantly advanced by Neural Turing Machines (Graves et al., 2014) and their successor, the Differentiable Neural Computer (Graves et al., 2016), which introduced differentiable memory access mechanisms. Modern Hopfield networks (Ramsauer et al., 2020) further revolutionized our understanding by establishing a theoretical connection between transformer attention and associative memory, showing how these systems can store and retrieve exponentially many patterns while maintaining stable dynamics. Such end-to-end approaches could particularly benefit the quality of memory representations, as they could learn optimal projections for generating representative keys for memory blocks, potentially capturing universal contextual patterns more effectively than our current approach. While end-to-end training of memory systems is feasible, as demonstrated by models like MERLIN (Wayne et al., 2018), such approaches often face challenges with credit assignment over long sequences and require complex architectural modifications. Our KNN-based approach leveraging the KV cache offers a pragmatic middle ground: it harnesses the rich semantic representations already present in transformer models while maintaining the computational benefits of nearest-neighbour retrieval. This aligns with both biological intuitions about pattern matching in the hippocampus (O'Reilly and Norman, 2002) and the theoretical foundations of modern Hopfield networks, where similarity-based attention serves as a form of associative memory. By operating on pre-trained representations, our method sidesteps the training complexities of fully differentiable memory while preserving the benefits of content-based retrieval.\", \"**Refinement:** The refinement of event boundaries could also theoretically be learned end-to-end, similar to how attention pruning mechanisms (Ying et al., 2019) learn to identify optimal subgraphs in graph neural networks, or how hierarchical clustering can be made differentiable (Ying et al., 2018). Our graph modularity approach provides a computationally efficient alternative that optimizes for coherence within segments while respecting the initial surprise-based boundaries. While our method is primarily motivated by computational considerations, it parallels how memory consolidation might strengthen associations between related elements within an event while weakening cross-event associations (Preston and Eichenbaum, 2013). The high modularity of our surprise-based segmentation, even before refinement, suggests that prediction errors naturally tend to occur at boundaries between coherent event structures.\", \"(continues to the next official comment..)\"]}", "{\"comment\": \"I remain unconvinced that the memory mechanisms implemented here accurately represent episodic memory (EM). As I mentioned in my original comments, EM pertains to the recollection of experiences rather than merely a sequence of texts. For instance, remembering that I read a book constitutes EM, while recalling the book's content does not. Therefore, the assertion that storing sentences is equivalent to human-like EM was not well-supported.\"}", "{\"title\": \"(Answer 2/2)\", \"comment\": \"* **Reviewer:** _3. Do you apply Eq. (1) to each token individually? If so, I imagine this would require substantial processing time. How do you address this issue?_\\n\\n **Authors:** While the calculation of surprise and corresponding boundaries does scale with O(n), where n is the number of input tokens, the probability in the left-hand side of the inequality is the output of the base LLM. Hence, this calculation boils down to a simple moving average which is largely negligible compared to a single forward pass of the base LLM. As a result, we have not noticed any impact on processing time.\\n\\n* **Reviewer:** _4. In Eq. (2), could you explain how you derive the key vectors?_\\n\\n **Authors:** Unless we have misunderstood your question, we have not derived the key vectors ourselves but rather used standard notation for the key vectors as part of the Transformer architecture. In this case, these are simply the keys for the input sequence output by each head in question, as described in the paragraphs around the equation. However, please do let us know if there has been some confusion here, or if you would like more information and we will be happy to address it.\\n\\n* **Reviewer:** _5. You employ k-NN in the MEMORY RETRIEVAL stage\\u2014does this significantly increase computation time?_\\n\\n **Authors:** As you have correctly pointed out, k-NNs does impact computation times during the retrieval stage. However, as briefly mentioned in Appendix C.2, which discusses complexity and wall clock time, this only becomes significant in very long contexts (millions of tokens). Furthermore, as mentioned in Section 3.4 of the main text, one may use approximate k-NNs in order to speed this up and maintain efficiency. We also feel that it is important to note that alternatives such as RAG also rely on k-NNs, while full-context methods are much more computationally demanding than our entire approach at context lengths where k-NN retrieval may just become noticeable for us, even without using approximate k-NNs. We will expand on these points in the sections of the paper mentioned and add figures to visualise such overhead as a function of context length for both our approach and full-context models. A more in-depth analysis of such overhead in terms of wall-clock time could be a somewhat interesting addition to the paper but given the overwhelming use of modern GPUs in LLMs, we feel this would be mostly reflective of our hardware, while more general time and computational complexity analysis for full and approximate k-NNs has been vastly covered already in the literature.\\n\\n* **Reviewer:** _6. I found Figure 4 difficult to understand. Could you provide additional explanation to clarify it?_\\n\\n **Authors:** While more details explaining how the results for this specific figure were calculated are available in Appendix B.1, we fully agree that the results themselves should be clear without it and would be more than happy to clarify this figure. We have amended the caption of this figure based on Reviewer _pMUj_. Could the reviewer please verify if the figure is clarified now and, if not, be more specific as to which sub-plot they find confusing and why?\\n\\n\\n### Summary of changes made to the paper based on your feedback:\\n* Question 1:\\n 1. We have updated the last two paragraphs of the **Introduction** to clarify our contributions.\\n 2. We have introduced the **Appendix E.2** section to further clarify this by reiterating the points in this response.\\n* Question 2:\\n 1. We added discussion on significance to Appendix A.1.\\n 2. We also added the points from this answer to the **Appendix E.2** mentioned in Question 1.\\n* Question 5:\\n 1. We added a section to **Appendix** providing complexity analysis and figures to visualise scaling behaviour for k-NNs within our approach, as well as our overall approach compared to standard full-context models for the same sequence length.\\n* Question 6:\\n 1. We have updated the caption of **Figure 4**.\"}", "{\"comment\": \"We sincerely thank the reviewer for their thorough analysis. We particularly appreciate their concern about potential overstatements in our terminology. We want to emphasize that our work claims similarities to, not equivalence with, human episodic memory - a distinction we should have made more explicit. While we provide detailed empirical evidence supporting these similarity claims in the responses below, we are also open to adjusting terminology (e.g., from \\\"human-like\\\" to \\\"human-inspired\\\") if our justification does not fully address the reviewer's concerns.\\n\\n## Question 1:\\n\\nWe thank the reviewer for this suggestion. We made amendments to our paper to clarify this very important point, which we summarise here.\\n\\nThe reviewer is correct to point out that the most similar architecture to our work is the model proposed by (Xiao et al. 2024a), namely InfLLM, which extended the KV lookup method introduced in (Han et al., 2023) for groups of tokens, by segmenting the context window into equally-sized blocks. Building on these previous methods, in this work we have made *three* novel architectural contributions for LLMs, for which we show their importance both conceptually and with empirical\\u00a0results:\\n\\n1. **Dynamic surprise-based segmentation.** We introduce the first method for dynamic segmentation of KV cache into blocks. Our method is also the first that manipulates the KV cache based on insights from cognitive science, using an intrinsic measure to LLMs. The only study in the literature that concerns Transformers and suggests a connection with surprise, was a psychology study (Kumar et al., 2023) that showed a connection between a similar measure of surprise and human event segmentation without, however, proposing or attempting to use this insight to alter the Transformer architecture in any way. We show empirically using multiple LLMs that this low-cost and simple-to-implement method is able to group relevant pairs of keys and values (KV) together (relevance measured as key similarity) with much higher accuracy than fixed segmentation, which is the only alternative proposed approach (See Table 2 for key similarity comparisons). We also show that this method results in increased LLM performance, especially in retrieval tasks ($16.6$% average increase over InfLLM) and multi-document QA tasks ($6.4$% average increase over InfLLM) across all the LLMs we tried (See the \\\"S\\\" column in the tables of Appendix A.1).\\n\\n2. **Graph-based refinement.** We were the first to introduce an algorithm to refine the temporal borders of events in the context window of LLMs using graph theory. We relied on the insight that tokens are more useful when recalled together if the variance between their keys is low, as they are retrieved using a single query at a time. This method can also stand by itself as a dynamic segmentation approach of KV cache, more computationally heavy than surprise-based segmentation but achieving a competitive accuracy in grouping relevant (KV) together (see again Table 2), with the added benefit that it can be used in each attention head independently, without relying on the LLM output.\\n\\n3. **Contiguity buffer**. We were the first to introduce a method to maintain a dedicated decaying buffer in the context window of LLMs that maintains the KV cache of temporally contiguous tokens to the context window of the LLM for a certain amount of time. For this, we relied on the recent insight that self-attention heads responsible for in-context learning are shown to consecutively attend to contiguous groups, similarly to human studies (Ji-An et al., 2024). We show that this algorithm can also be combined with methods (1.) and (2.) and results in further increases in the overall LLM performance. Notably, the average increase in retrieval tasks over InfLLM jumps to $19.4$%, and for multi-document QA tasks to $9.2$% across all the LLMs we tried (See the \\\"SM+C\\\" column in the tables of Appendix A.1).\\n\\n### Summary of changes made to the paper:\\n1. We have updated the last two paragraphs of the **Introduction** to clarify our contributions.\\n2. We have introduced the **Appendix E.2** section (referred to in the **Introduction**) to further clarify this by reiterating the points in this response.\", \"title\": \"(Answer 1/3)\"}", "{\"comment\": \"Thanks for the further clarification. Based on the evidence provided so far, it is difficult to determine whether it is more like episodic memory or declarative, non-episodic memory. So, I think it would be helpful to discuss such potential controversy, and maybe also tune down the tone (e.g., \\u201cepisodic memory inspired\\u201d as suggested by the authors).\"}", "{\"title\": \"Response to all reviewers\", \"comment\": \"We sincerely thank all reviewers for their thoughtful feedback and valuable suggestions. We have provided detailed responses to each point raised, and we are currently incorporating these discussions into an updated version of the paper. As noted in our individual responses, some content will be added to the supplementary material due to space constraints. We plan to update the paper PDF later in the discussion period to incorporate any additional feedback that may arise. We believe these revisions will strengthen the paper and clarify important aspects of our work.\"}", "{\"title\": \"(Answer 1/4)\", \"comment\": \"We sincerely appreciate reviewer pMUj's thorough analysis, positive comments and constructive feedback. We have revised our manuscript accordingly and provided detailed responses to each point below.\\n\\n\\n## Weaknesses:\\n\\n* **Reviewer:** _In Table 1, the performances of S, SM, and SM+C methods are difficult to directly compare since different base LLMs are used in each row. I know the comparison is (partially) in Fig. 4-7 but they are not mentioned in the main text. It would be helpful if explicitly mentioned/referenced in the main text._\\n\\n **Authors:** Thank you for the suggestion. We agree with your observation and have clarified the main text by mentioning such figures (**Section 4.1**, paragraph **2**, where Table 1 is first mentioned, as well as pointing to it in the caption for **Table 1**).\\n\\n* **Reviewer:** _Although InfLLM is a primary benchmark comparison, it is absent from Figure 1. Including InfLLM in Figure 1 would provide a more complete comparison._\\n\\n **Authors:** We understand your motivation for such a figure, but, having considered such a version ourselves, we found that a fourth addition is incompatible with our preferred superimposed format for the figure, making it less clear. Moreover, we are of the opinion that such a comparison is not directly relevant to the content of Figure 1, for which the main focus is a comparison of our retrieval method with full-context and RAG. Instead, a much more detailed comparison of our approach with InfLLM is available in **Table 1** in the main text, as well as Appendix **Tables 3-7**. Of course, should you still feel that such an addition is crucial to this figure, we are open to reconsider.\\n\\n* **Reviewer:** _It\\u2019s unclear from the current paper how the number of retrieved events and the number of evicted tokens influence performance. Any plot showing the model\\u2019s performance as a function of these parameters would be helpful._\\n\\n **Authors:** We do agree that an ablation study on the size of the retrieved buffer, and local tokens, as a function of context length, which is proportional to the number of evicted tokens for a fixed-length retrieved buffer, would be a valuable addition to the paper. To this end, we plan to add a figure to the revised version which express our current results as a function of the context length of the evaluated examples. In our current experiments, we have chosen our buffer sizes to align with related works (namely InfLLM) in order to make direct performance comparisons. Such values also keep buffer sizes shorter than the average number of tokens in the evaluated benchmarks to ensure an appropriate use of retrieval. In order to further explore variations in these parameters, we also:\\n 1. Ran a small ablation study varying the size of the retrieved buffer for summarization tasks on LongBench. \\n 2. Added a table and figure, as well as a discussion, to **Appendix D** which expresses the ablation results as a function of the context length of the evaluated examples.\\n\\n However, for the following reasons, we believe this provides limited information on such parameters. For such a study, we have to choose a base LLM trained with a relatively large context window, such as Mistral or LLaMa 3.1 which support context lengths of up to 32K and 128K respectively, in order to ensure that the underlying model can support an adequate range of buffer sizes. As LongBench may be considered relatively short compared to these context windows (average number of tokens per example: 18K$\\\\pm$1.8K with Mistral's tokeniser), $\\\\infty$-Bench would be more appropriate (average number of tokens per example: $> 100$K). Unfortunately, evaluating larger buffer sizes (hence larger attention matrices) on the already-expensive $\\\\infty$-Bench benchmark would be a very demanding ablation given our limited hardware resources, and hence we have left it for future work and mentioned it as such in the Discussion.\\n\\n* **Reviewer:** _In Formula (1), notations_ $\\\\mu_{t-\\\\tau:t}$ _should be corrected, keeping consistent across the context._\\n\\n **Authors:** Thank you for catching this mistake, we have now fixed **Eq. 1** in the revised version.\\n\\n## Questions:\\n\\n* **Reviewer:** _Can the authors discuss the feasibility and potential advantages of end-to-end training of networks for event segmentation, refinement, and retrieval, compared to these hand-designed algorithms? As a comparison, these algorithms are implemented by neural circuits in the brain_\\n\\n **Authors:** This is a great question. We present a brief discussion below. Due to space limitations, we haven't included this in the main paper, but we welcome your thoughts on whether it should be added to the supplementary material.\\n\\n(continues to the next official comment..)\"}", "{\"title\": \"(Answer 4/4) - References\", \"comment\": [\"**References that do not appear in the manuscript:**\", \"(Jiang et al., 2017) \\\"Variational deep embedding: an unsupervised and generative approach to clustering\\\", IJCAI'17.\", \"(Franklin et al., 2020) \\\"Structured Event Memory: A neuro-symbolic model of event cognition.\\\" _Psychological review_ 127.3 (2020): 327.\", \"(Hopfield, 1982). \\\"Neural networks and physical systems with emergent collective computational abilities\\\", PNAS 79.8 (1982): 2554-2558.\", \"(Graves et al., 2014) \\\"Neural Turing Machines\\\", arXiv:1410.5401.\", \"(Graves et al., 2016) \\\"Hybrid computing using a neural network with dynamic external memory\\\", _Nature_ 538.7626 (2016): 471-476.\", \"(O'Reilly and Norman, 2002). Hippocampal and neocortical contributions to memory: Advances in the complementary learning systems framework. _Trends in Cognitive Sciences_ 6.12 (2002): 505-510.\", \"(Ramsauer at al., 2020). Hopfield Networks is All You Need. _arXiv:2008.02217_\", \"(Wayne et al., 2018). Unsupervised Predictive Memory in a Goal-Directed Agent. arXiv:1803.10760.\", \"(Ying at al., 2019). GNNExplainer: Generating explanations for graph neural networks. NeurIPS, 2019.\", \"(Ying at al., 2018). Hierarchical graph representation learning with differentiable pooling. NeurIPS, 2018.\", \"(Preston and Eichenbaum 2013). Interplay of hippocampus and prefrontal cortex in memory. _Current biology_ 23.17 (2013): R764-R773.\", \"(Saxena et al., 2021) \\\"Clockwork variational autoencoders.\\\"_Advances in Neural Information Processing Systems_ 34 (2021): 29246-29257.\", \"(Hafner et al., 2022) \\\"Deep hierarchical planning from pixels.\\\"_Advances in Neural Information Processing Systems_ 35 (2022): 26091-26104.\", \"(Johnson et al., 2019) \\\"Billion-scale similarity search with GPUs.\\\"_IEEE Transactions on Big Data_ 7.3 (2019): 535-547.\"]}", "{\"title\": \"(Answer 2/4)\", \"comment\": \"* **Reviewer:** _Does the issue of needing to query too many matching events still exist when the context window is too long?_\\n\\n **Authors:** This question fits in well with our own motivations and results as the issue of diluted (a.k.a. distracted) attention in full-context transformers (Tworkowski et al, 2023; Ye et al, 2024; Liu et al, 2024), when applied to long-context tasks, is a key motivator for retrieval methods. We believe that our method addresses this issue, both by leveraging k-NNs for top-k retrieval of only the most relevant memory units, and improving the cohesion of keys within such units in order to increase the overall relevance of its member keys. We have mentioned diluted attention, and related works that address it, in section 2.1, and believe that it is responsible for the improvements in $\\\\infty$-Bench's retrieval tasks over full-context and RAG methods. Hence, while the issue of \\\"too many matching events\\\" may still be considered to be present in the k-NN retrieval step, the performance improvements observed from only considering the top-k events and improving on the accuracy of retrieval suggests that our approach can at least reduce this issue. Not to mention the greatly reduced computational complexity of computing a much smaller attention matrix.\\n * **References that do not appear in the manuscript:**\\n * (Ye et al, 2024), \\\"Differential Transformer\\\" \\n\\n* **Reviewer:** _If we use the Surprise-based method to segment the context into a RAG format, will the performance remain the same?_\\n\\n **Authors:** Thank you for this interesting suggestion. To be able to answer your question, we explored this exact approach by implementing a RAG variant that uses our surprise-based segmentation instead of fixed-length chunks and we have now included this information in Supplementary Table 9 and Section 4.4. Specifically, we used the same surprise threshold mechanism from EM-LLM to segment the context before encoding it into the RAG vector database. However, our experiments on LongBench showed that this approach actually performed worse than standard fixed-length chunking in RAG (see Table 9), with the same context window size. We believe there are several key reasons for this:\\n\\n 1. Our analysis (see Supplementary Figure 5) shows that $10-60$% of the events recalled by each layer of the LLM are unique to that layer at each query step. This demonstrates that different layers benefit from accessing different parts of the context at different times. In contrast, RAG's single retrieval step forces all layers to work with the same retrieved context, limiting the model's ability to access task-relevant information in a layer-specific way.\\n\\n 2. Many tasks in our benchmarks (like summarization or multi-hop reasoning) require retrieving different pieces of information at different stages of generation. EM-LLM's layer-wise retrieval naturally supports this by allowing the model to access different context as needed throughout the generation process. RAG's one-time retrieval at the start of generation cannot adapt to these changing information needs.\\n\\n 3. Since RAG performs only a single retrieval step, its performance heavily depends on getting as much potentially relevant information into the context window as possible. Fixed-length chunks ensure consistently large segments of context, while surprise-based segments can be quite small when detecting frequent meaningful boundaries. This makes fixed-length chunking more effective for RAG's retrieval strategy, even though surprise-based segmentation works better with EM-LLM's more flexible layer-wise retrieval mechanism.\\n\\n This finding actually reinforces why EM-LLM's approach of combining surprise-based segmentation with layer-wise retrieval is important - it allows the model to dynamically access different parts of the context in ways that RAG's architecture cannot support.\\n\\n\\n * **Summary of changes made to the paper:**\\n 1. Added results to Appendix **Table 9**.\\n 2. Added details describing this new experiment in **Appendix A.2**.\\n 3. Added a reference to this new experiment in **Section 4.1**.\\n\\n\\n* **Reviewer:** _How does the method handle dynamic changes in context relevance over time?_\\n\\n **Authors:** After discussing this with our whole team, we have found it difficult to agree on a common interpretation for the reviewers' question. If the reviewer could please provide more details, or an example to illustrate the point touched on in this particular question, we will be happy to address it.\"}", "{\"title\": \"(Answer 1/4)\", \"comment\": \"We would like to thank reviewer RkzP for their careful review and thoughtful questions, which have helped us identify areas where we needed to clarify our contributions and strengthen our arguments. We have updated our manuscript to address all points raised. Below we provide detailed responses to each concern. We remain open to any additional feedback that could further improve our work.\\n\\n### Responses to reviewer's questions:\\n\\n* **Reviewer:** _How sensitive is the method to the choice of the surprise threshold parameter? What is the process for tuning this parameter?_\\n\\n **Authors:** While we have already included figures in Appendix D which illustrate ablations for the threshold scaling factor, relative buffer sizes, and method used, we agree that these decisions and their supporting observations should be included in the paper. In this particular case we have carried out a short ablation study on each model with suprise-only segmentations to select the best value to use for our experiments. Although there is a consistent preference for certain values across all models, results suggest that the method is largely insensitive to this parameter in the sense that minor variations in its value do not significantly affect downstream performance. We now mention this in the main text and provide more details in the Appendix.\\n\\n * **Summary of changes made to the paper:**\\n 1. Pointed to **Appendix D** in the caption for **Table 1**.\\n 2. Added a sub-section to **Appendix D** providing further details for our initial ablations and parameter tuning.\\n\\n* **Reviewer:** _How do you ensure the stability of the boundary refinement process? Are there cases where it fails to converge?_\\n\\n **Authors:** This is a very pertinent observation, thank you. Through our own omission, there appears to have been some confusion as to the implementation of the algorithm which is relevant to the need for convergence and stability analysis. Namely, there are no iterations for single boundaries, as opposed to what our statement on line 313 of the main text appeared to suggest. We simply loop through each event boundary and perform a **single** update using an argmax function to measure the best (new) position for that boundary within the event to its left, and with respect to the overall similarity metric of the current processed sequence. After this update, the corresponding memory units are saved and no longer modified. This guarantees either a positive improvement or no change in similarity for each event boundary position update, hence either improving overall similarity or showing no change from surprise-based segmentation. Therefore, while we would ideally find the globally optimal event boundaries with regards to the similarity metric, and seek to converge to this point, this would be much more expensive to compute and introduce a lot of overhead for every processed chunk of the context and the corresponding saved memory units. Instead, our algorithm simply implements a cost-effective way to look for **any** potential increase to this metric, as it has been empirically shown to do successfully in section 4.2. Nevertheless, to briefly touch on the convergence of such a method, our approach can be seen as a single pass of Phase 1 of the heuristic Louvain method (Blondel et al, 2008) initialised with surprise-based segmentations (as opposed to having each node assigned its own community), and modified to only consider the move of a node to its right-side neighbouring community. As we have shown that surprise-based segmentations already achieve higher similarity metrics (including modularity, which is the objective used in the Louvain method) than fixed or random segmentations, we believe this is a good initialisation as it means that our algorithm will, at worst, achieve the same modularity. While the Louvain method is considered to be an efficient way to converge to _local_ optima when iterated, our own modifications and lack of iterations mean we cannot claim such behaviour but rather suggest that we are likely to see some improvements in our metrics, as our results have confirmed. We updated the wording used in the main text to clarify these points, as well as discussed this further in the Appendix.\\n\\n * **References that do not appear in the manuscript:**\\n * (Blondel et al, 2008), \\\"Fast unfolding of communities in large networks\\\"\\n\\n * **Summary of changes made to the paper:**\\n 1. Changed \\\"iteratively\\\" to \\\"sequentially\\\" in **Section 3.3**.\\n 2. Added a section to the **Appendix** with this discussion.\\n 3. Made minor adjustments to the complexity analysis for **Algorithm 1** following internal speculation as to its clarity.\"}", "{\"summary\": \"The paper introduces EM-LLM, a novel architecture that enhances the memory capabilities of LLMs by integrating design elements of human episodic memory. EM-LLM employs surprise-based event segmentation, similarity-based, and contiguity-based retrieval to enable LLMs to manage long contexts with computational efficiency. Tokens are organized into episodic events by dynamically detecting boundaries using a surprise metric and further refines these boundaries with graph-theoretic measures. This episodic structure allows EM-LLM to recall information from both the current context and prior temporally contiguous events. Experiments on long-context benchmarks show that EM-LLM outperforms InfLLM and RAG. Additionally, the segmentation results are aligned with human-perceived event structures, suggesting EM-LLM as a computational model of human episodic memory.\", \"soundness\": \"4\", \"presentation\": \"4\", \"contribution\": \"4\", \"strengths\": \"Originality: EM-LLM\\u2019s incorporation of episodic memory features, particularly surprise-based segmentation, similarity-based, and contiguity-based retrieval, is a novel approach within the domain of LLMs. Using LLM surprise to model event cognition and to approximate episodic memory is original for cognitive modeling. This work not only extends the long-context capability of LLMs but also bridges machine learning and cognitive science, presenting a unique computational framework for studying memory and event cognition.\", \"quality\": \"The paper demonstrates EM-LLM\\u2019s performance on recognized long-context benchmarks. The experiments highlight the model\\u2019s improvements in accuracy and efficiency over competing methods, including InfLLM and RAG.\", \"clarity\": \"The paper clearly presents the concept of episodic memory in LLMs, with a logical breakdown of memory segmentation, refinement, and retrieval mechanisms. The figures generally illustrate the processes well.\", \"significance\": \"EM-LLM has potential significance in advancing LLM capabilities for long-context tasks, a critical area of LLM development. Beyond practical applications, EM-LLM contributes to interdisciplinary research, providing insights into cognitive science by modeling elements of human episodic memory within an LLM framework.\", \"weaknesses\": [\"In Table 1, the performances of S, SM, and SM+C methods are difficult to directly compare since different base LLMs are used in each row. I know the comparison is (partially) in Fig. 4-7 but they are not mentioned in the main text. It would be helpful if explicitly mentioned/referenced in the main text.\", \"Although InfLLM is a primary benchmark comparison, it is absent from Figure 1. Including InfLLM in Figure 1 would provide a more complete comparison.\", \"It\\u2019s unclear from the current paper how the number of retrieved events and the number of evicted tokens influence performance. Any plot showing the model\\u2019s performance as a function of these parameters would be helpful.\", \"In Formula (1), notations (\\\\mu_{t-\\\\tau:t}) should be corrected, keeping consistent across the context.\", \"Certain figures would benefit from clarification, as detailed in questions.\"], \"questions\": \"- Can the authors discuss the feasibility and potential advantages of end-to-end training of networks for event segmentation, refinement, and retrieval, compared to these hand-designed algorithms? As a comparison, these algorithms are implemented by neural circuits in the brain.\\n- Can the authors discuss how might other features of human episodic memory, such as the hierarchical organization of events (human episodic memory structures can have multi-level segmentation of experiences), benefit EM-LLM\\u2019s design in terms of performance? \\n- Clarifications for Figures:\\nIn Figure 1, does the data reflect results from S, SM, or SM+C? \\nIn Figure 2, the memory units seem to have variable sizes, different from the fixed size in InfLLM? Additionally, how are the bars normalized?\\nIn Figure 4A, S appears closer to human segmentation than SM or SC, contrasting with Figure 4B. \\nIn Figure 4, what does SC stand for? Does \\\"C\\\" refer to the continuity buffer or represent something else?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"I appreciate the authors' thoughtful clarifications. The current Figure 1 is clear. Thank you for the additional ablation study and detailed discussion, which all could be added to the supplementary material.\\n\\nEvent boundary segmentation (Zacks et al., 2007, Baldwin & Kosie, 2020; Baldwin et al., 2008; Schapiro et al., 2013) and asymmetric contiguity biases (Kahana, 1996; Murdock, 1962; Murdock & Okada, 1970) are among the most significant well-established properties of biological episodic memory, extensively studied in the cognitive science literature. The authors' correct identification of these crucial elements and their integration into the models to improve performance over competing approaches like InfLLM is both impressive and insightful. By incorporating these principles, the mechanisms of LLMs are rendered more human-like, enhancing their alignment with biological cognitive processes.\\n\\nYour work stands out as an exemplar of how cognitive science and neuroscience-inspired designs can improve LLM models. Moreover, the comparison with human event segmentation in a human-annotated audio dataset provides a unique and valuable bridge, leveraging LLM representations to deepen our understanding of human episodic memory. This innovative approach has significant potential to inspire further interdisciplinary exploration at the intersection of AI, cognitive science, and neuroscience.\"}", "{\"comment\": \"I apologize for the delay in providing my review comments. Initially, I stated that I am not an expert in your field, and I found your manuscript quite challenging to read. If you feel that you have received an unfair review, I am truly sorry. However, I have also experienced situations where reviewers provided malicious negative reviews without any response. Therefore, if you believe there are issues with my review, please feel free to bring them to the attention of the AC. Besides, I truely think you should polish your paper.\\n\\nThank you for your understanding.\"}", "{\"title\": \"(Answer 3/4)\", \"comment\": [\"**Reviewer:** _Can the authors discuss how might other features of human episodic memory, such as the hierarchical organization of events (human episodic memory structures can have multi-level segmentation of experiences), benefit EM-LLM\\u2019s design in terms of performance?_\", \"**Authors:** Thank you for this question. We briefly touched on this in the Discussion as well as our previous answer here. Human episodic memory exhibits several key features that could enhance EM-LLM's capabilities:\", \"**Hierarchical organisation:** A hierarchical structure in memory can provide multiple advantages such as improved retrieval, more disentangled latent embeddings, longer future predictions (Saxena et al., 2021; Zakharov et al., 2022b), better planning (Hafner et al., 2022) and higher agreement with neural processes in the brain (Baldassano et al., 2017). In our model, hierarchical organisation of episodic memories based on the existing hierarchy of embeddings in the LLM layers could be implemented by extending our segmentation processes to operate at each layer of the Transformer independently. This could be achieved either through a differentiable approach or a layer-specific surprise metric. Interestingly, our current k-NN retrieval approach already implicitly leverages hierarchical structure through its underlying approximate nearest neighbour algorithms, which typically employ tree-based structures (Johnson et al., 2019) to efficiently partition the embedding space.\", \"**Memory consolidation:** The brain's process for memory consolidation is crucial for continual learning, an ability that remains largely unsolved in current LLMs. Implementing consolidation mechanisms in EM-LLM could help address catastrophic forgetting while enabling more efficient integration of new information with existing knowledge.\", \"**Mental time travel:** The ability to employ the same retrieval mechanism for imagining future events as for recalling past experiences is a key feature of episodic memory that could significantly enhance LLMs' planning and reasoning capabilities. By leveraging its event-based structure to simulate potential future scenarios or recall past experiences in novel contexts, this mechanism could provide a powerful solution for planning and reasoning, which are currently important challenges in large generative models.\", \"**Reviewer:** _Clarifications for Figures: In Figure 1, does the data reflect results from S, SM, or SM+C? In Figure 2, the memory units seem to have variable sizes, different from the fixed size in InfLLM? Additionally, how are the bars normalized? In Figure 4A, S appears closer to human segmentation than SM or SC, contrasting with Figure 4B. In Figure 4, what does SC stand for? Does \\\"C\\\" refer to the continuity buffer or represent something else?_\", \"**Authors:** Thank you for pointing out any miscommunication in these figures, we have now updated their descriptions. In particular:\", \"*Figure 1*: The data here reflects the results from surprise-based (**S**) segmentation, as denoted by \\\"$EM-LLM_S$\\\". We followed the same notation as in the other tables (notably **Table 1**), and have updated the caption of **Figure 1** to clarify this.\", \"*Figure 2*: We have used this figure to illustrate that k-NN retrieval of past tokens for softmax attention can be seen as a form of hierarchical attention. Please note that this illustration is relevant for any form of k-NN retrieval including both InfLLM and EM-LLM. However, we understand that the coincidental placement of this figure immediately after mentioning InfLLM's fixed-size segmentations in section 2.1 made this confusing, as the figure does not show fixed-sized units. We have now clarified this by re-wording the reference to **Figure 2** in this section. As for the normalisation, this is simply the softmax of the similarity scores shown, as it is describing softmax attention. We also updated the legend in **Figure 2** to clarify this.\", \"*Figure 4*: In this figure, no contiguity is shown. As such, **SC** follows the format of **SM**, introduced in the caption for **Table 1**, and shows surprise-based boundary refinement with _conductance_ as the similarity metric. We understand this was not clear and thank you for pointing it out. We have updated the caption of **Figure 4** to clarify this. Finally, **Figure 4A** differs from **4B** in that it shows the average similarity metrics for events due to various segmentation methods, while **4B** shows the distance in actual event positions from the human data. With this in mind, our interpretation of such a discrepancy is that, while two segmentation methods may achieve similar similarity metrics, it does not necessarily mean that their event boundaries are close to each other. Hence, in this specific case, it would appear that SM and SC are both closer to human-perceived event boundaries than S, while still achieving better similarity metrics.\"]}", "{\"comment\": \"Dear Reviewer RkzP,\\n\\nWe wanted to kindly remind you that the discussion period ends Monday, December 2nd AoE. We have provided detailed responses to your initial concerns and would greatly value your feedback on whether our clarifications have addressed your questions satisfactorily, or whether there are more changes you would like us to consider. Given the approaching weekend, we understand time might be limited, but any engagement before the deadline would be much appreciated.\\n\\nBest regards,\\nAuthors\"}", "{\"comment\": \"Thank you for the clarifications; they are helpful. However, I still perceive the conceptual improvements as somewhat incremental, as the current work appears to be a refinement of the previous study (InfLLM).\"}", "{\"title\": \"(Answer 4/4)\", \"comment\": \"* **Reviewer:** _The human correlation study in Section 4.2 relies on 3 short podcasts, which is insufficient to make broad claims about alignment with human event segmentation. The authors should expand the section to a larger, more diverse dataset of human annotations._\\n\\n **Authors:** We appreciate the reviewer's concern about dataset size, as large-scale evaluation is indeed crucial in machine learning. However, we would like to clarify several important points about the human correlation study:\\n\\n 1. The datasets used in Section 4.2 represent, to the best of our knowledge, some of the most comprehensive text-based human event segmentation annotations available in the cognitive psychology literature. These are not simply short podcasts, but rather carefully curated datasets with rich annotations from 200 human participants, which have been pivotal in recent psychological studies of human event cognition (Lositsky et al., 2016, Michelmann et al., 2021, Kumar et al., 2023). The length and complexity of these narratives (7-30 minutes each) actually make them substantially more extensive than typical stimuli used in human cognitive studies. However, should the reviewer be aware of any other large datasets with robust annotations describing human event cognition, we would be happy to expand our analysis to include it.\\n\\n 2. The statistical significance and consistency of our results across these datasets, particularly the striking difference in Wasserstein distance between our methods and random/fixed segmentation, suggests that the signal we observe is robust despite the dataset size. This is further supported by the fact that both surprise-based segmentation and refinement show consistent improvements across base LLM models and podcasts (see Figures 6-8).\\n\\n 3. We fully acknowledge the machine learning community's emphasis on large-scale validation. This is precisely why our broader evaluation of EM-LLM's effectiveness extends well beyond these human correlation studies in two large benchmarks. In addition, to better support our results on online clustering beyond the human-annotated datasets, we conducted extensive experiments on PG-19, a standard benchmark for long-context language models, to validate our approach's scalability and effectiveness in a large-scale setting (see Table 2).\\n\\n* **Reviewer:**\\n * _The two-stage memory retrieval process (Section 3.4) introduces ks and kc parameters for similarity and contiguity buffers, but it does not justify the choice of values or analyze their trade-offs._\\n * _Section 4.4 discusses the impact of contiguity but does not properly ablate different buffer sizes or analyze the trade-off between similarity and contiguity buffers._\\n\\n **Authors: (combined answer)** This is currently visualised in Appendix Figure 11, and we regretably do not have space for it in the main text. However, we agree that this is an important point to touch on and to this end we are adding more details on this, as well as other ablations, and the resulting choice of parameters to the Appendix. We also mention this section briefly in the main text. Specifically:\\n\\n * **Summary of changes made to the paper:**\\n 1. Pointed to **Appendix D** in caption for **Table 1**.\\n 2. Added a sub-section to **Appendix D** providing further details for our initial ablations and parameter tuning.\"}", "{\"metareview\": [\"(a) This paper introduces a novel architecture, EM-LLM, designed to enhance the long-context capabilities of Large Language Models (LLMs). The authors draw inspiration from human episodic memory and event cognition to develop a system that can handle theoretically infinite context lengths. The key scientific claims include:\", \"Dynamic Event Segmentation: EM-LLM segments text into meaningful episodic units using a measure of Bayesian surprise, departing from fixed-length approaches like InfLLM.\", \"Graph-Based Refinement: The initial segmentation undergoes refinement based on graph-theoretic metrics, optimizing the cohesion within and separation between events for better information retrieval.\", \"Two-Stage Memory Retrieval: EM-LLM employs a retrieval process that combines similarity-based search with a mechanism mimicking the temporal contiguity effect observed in human recall, enabling more nuanced access to information.\", \"Good performance on LongBench and \\u221e-Bench.\", \"(b) Strengths\", \"The proposed EM-LLM architecture offers a novel approach to long-context LLM processing, combining established techniques (Bayesian surprise, graph-theory, k-NN retrieval).\", \"The demonstrated ability to process vast text sequences (10 million tokens) with high accuracy showcases the model's scalability and potential for real-world applications where context is crucial.\", \"(c) Weaknesses:\", \"Limited Human Correlation: While the study attempts to connect EM-LLM to human event perception, the analysis relies on a relatively small dataset (3 podcasts) and lacks deeper exploration of the nuanced relationship between model behavior and human cognitive processes.\", \"The initial framing as \\\"human-like\\\" episodic memory was considered an overstatement, prompting the authors to revise the terminology to \\\"episodic memory-inspired\\\".\", \"Some reviewers desired a stronger theoretical grounding for certain design choices, such as why surprise-based segmentation should lead to improved performance compared to fixed-length methods.\", \"(d) The decision is to Accept.\", \"While acknowledging the weaknesses, the novelty of the architecture, the promising empirical results, and the potential for advancing both LLM capabilities and computational models of cognition make this paper a valuable contribution.\"], \"additional_comments_on_reviewer_discussion\": [\"During the rebuttal period, reviewers raised several concerns, leading to clarifications, additional experiments, and revisions by the authors.\", \"Concerns about the Novelty and Contributions: Some reviewers found the conceptual improvements incremental, viewing the work as a refinement of InfLLM, and questioned the specific innovations, as many methods seemed previously proposed. The authors acknowledged that some techniques were inspired by prior work but emphasized their unique integration and the significant performance gains achieved.\", \"Reviewers questioned the justification for claiming the model was \\\"human-like,\\\" arguing that episodic memory involved more than text sequences. They suggested alternative terms like \\\"episodic-memory-like\\\" or \\\"episodic-memory-inspired.\\\" The authors agreed to adjust the terminology, changing \\\"human-like\\\" to \\\"episodic memory-inspired\\\" throughout the paper.\", \"Reviewers requested clarification on figures and tables, including details about parameter choices, performance comparisons, and the meaning of certain abbreviations. The authors addressed most of them.\", \"The authors' responsiveness to reviewer feedback, including conducting additional experiments, clarifying their claims, and revising the manuscript, further strengthens the paper's value.\"]}", "{\"comment\": \"Thank you for your response. Regarding our contribution, we consider it sufficient for this ongoing field of work (3 distinct architectural novelties, current SOTA performance in different benchmarks, and context lengths far beyond anything in the literature of Transformers), especially considering the extensive experimentation, analysis, and resulting insights.\\n\\nCould the reviewer please elaborate on what they would like us to include, for our contribution to not be considered incremental?\"}", "{\"summary\": \"In this work, the authors proposed a framework, termed EM-LLM, for LLMs to store previously seen sequences of tokens as memory and then utilize the stored information to improve the performance of newly encountered problems. Specifically, inspired by human episodic memory, the authors suggested a method based on Bayesian surprise and graph-theoretic refinement to parse the sequences into events. Later, these store events were retrieved depending on content similarity and temporal contiguity. Experiments were conducted to demonstrate the performance improvements. In addition, a comparison was conducted to show that the parsing of events by the proposed method was similar to the results of humans, suggesting a potential link to the underlying mechanism in the brain.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"Memory is a key function in biological brains that helps organize experiences and use them to guide future behaviours. The lack of memory function in the LLM is a major aspect that needs to be addressed. The authors have proposed a framework inspired by the biological memory system for LLMs, which is desirable and may stimulate further investigations in this direction. Overall, the paper was clearly written, showing comprehensive experimental results.\", \"weaknesses\": \"\\\"Human-like Episodic Memory\\\" is an overstatement. First, episodic memory combines multimodal information into a coherent recollection of experiences, with key aspects including when, where, what, how and with who a event happened. To store a sequence of tokens does not necessarily reflect such a memory. Second, parsing of events in human memory depends heavily on the content, e.g., change of location or context. The Bayesian surprise measure may be a useful proxy to approximate such separation, but it does not by itself suggest a common mechanism, nor support the claim of \\\"human-like\\\". Third, the retrival rule of similarity is generic for memory, not specific for EM.\\n\\nAnother concern is that the memory-based frawmork has been previously proposed (cited as Xiao et al, 2024a in the current paper). The present work is an refinement of the framework, which somewhat limits its noverty.\", \"questions\": \"1. I would suggest the authors to better explain the differrence in comparison with Xiao et al. 2024a, and explain why it is an important conceptual advance, instead of an incremental improvement.\\n2. To better justify the claim that the mechanism is EM-related with further analyses.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"## Question 2:\\n\\nWe thank the reviewer for their thoughtful critique regarding our claims of human-like episodic memory. We agree that such claims require careful justification, and we appreciate the opportunity to clarify how our model addresses each aspect of episodic memory raised in the review. Please consider the below, which is also amended in the latest version of the manuscript.\\n### 1. Foundation: Transformers and Episodic Memory\\nWe completely agree that all the points mentioned regarding what features a human-like model of episodic memory (EM) should incorporate are important and valid. As the reviewer states, EM indeed combines multiple pieces of information into a coherent recollection of experiences. In Transformers, this integration naturally occurs through latent embeddings - recollections of concepts inferred from inputs and encoded in the embedding space, recalled via the SoftMax self-attention mechanism. Recent work has shown this enables human-like behaviour in short-context memory recall tasks (Ji-An et al., 2024).\\nHowever, in long-context tasks, transformers face two fundamental constraints that break this connection to human episodic memory:\\n- Computational and memory complexity increases quadratically.\\n- Retrieval performance drops due to attention dilution (Tworkowski et al., 2023; Ye et al., 2024).\\n### 2. Our Solution\\nConsidering the above, the main purpose of any memory architecture claiming human-like EM functionality must be to extend transformers' inherent capabilities beyond these limitations. Our proposed method achieves this to a degree that goes far beyond the computational capabilities of full-context LLMs today (at least 10M tokens using a single GPU). To do this, we introduce a connection to human event cognition, which we claim is precisely what's missing from standard transformers. While they can integrate information within their attention window, they lack the systematic event-based organization that characterizes human EM. Our method achieves this through two key mechanisms:\\n#### A. Information Grouping Through Event Segmentation\\nFor grouping information together, we use Bayesian surprise, which has been repeatedly shown to capture content-dependent event boundaries in human perception. This choice is deeply grounded in both behavioural and neuroimaging evidence showing that surprise (or prediction error) signals in the hippocampus and cortical regions are crucial for episodic memory and event boundary formation (Sinclair et al., 2021; Zacks et al., 2007; 2011; Sherman et al., 2022;\\nMariola et al., 2022; Fountas et al., 2022), including direct empirical evidence with our particular implementation presented in our paper. Our analysis shows that this approach effectively groups together semantically related content: tokens within surprise-based segments have significantly higher key similarity than tokens across segments (see Table 2), indicating natural capture of meaningful content transitions. This aligns with both empirical studies of human event segmentation and the reviewer's point about content-dependent parsing.\\n\\nImportantly, due to the nature of Transformers, the contents of experience (embeddings) are processed within the temporal limits of these events. This allows us to effectively handle the key aspects of episodic memory that the reviewer identified:\\n* **When, where, what, how and with who an event happened:** This information, given to an LLM via text during an episode, forms the basis of most QnA and retrieval tasks in our benchmarks. Our model's significant performance boost compared to baseline InfLLM demonstrates its ability to maintain and retrieve these coherent aspects of experiences together.\\n* **Multi-modality:** While our current implementation focuses on text, our approach is fundamentally compatible with multi-modal processing. In Transformer-based models, integrating information across different modalities follows the same principles as integrating information within a single modality. Recent multi-modal models (e.g., Qwen2-VL, Pixtral, LLaMa 3.2) already demonstrate this by integrating different modality encoders into a single embedding space. Our EM-LLM can accommodate these approaches as the KV cache treats all embeddings equally. This point is now added in the discussion.\", \"title\": \"(Answer 2/3)\"}", "{\"title\": \"Complete paper revision: New results and detailed change log\", \"comment\": [\"Dear Reviewers,\", \"We sincerely thank you for your thoughtful feedback which has helped us significantly strengthen the paper. We have now uploaded a revised PDF that incorporates all discussed changes, with particular focus on the main concerns raised:\", \"1. Empirical validation (due to points raised by **pMUj**, **RkzP**):\", \"Included new experiments where:\", \"We achieved **100**\\\\% retrieval score in **10M** tokens using a single 32GB NVIDIA GPU and *LLaMA3.1-8B* (see **Fig. 1**). To our knowledge, this is far beyond any length achieved so far with transformer models in the literature.\", \"We compared surprise-based segmentation in RAG format and showed it does not beat standard RAG, let alone our approach. We discuss the reasons why.\", \"Added statistical significance analysis showing consistent improvements over InfLLM (p < 0.05) in Appendix **A.1**.\", \"Expanded ablation studies on parameter sensitivity and buffer sizes in Appendix **D.1**.\", \"2. Novelty and contributions (due to points raised by **SyTr**, **N12r**):\", \"Clarified our three key architectural innovations: dynamic surprise-based segmentation, graph-based refinement, and contiguity buffer (see end of **Introduction** and Appendix **E.2**).\", \"Enhanced discussion of theoretical foundations linking our approach to cognitive science (see Appendix **E.1**, mentioned also in the **Discussion**).\", \"3. Technical rigour (due to points raised by **RkzP**):\", \"Updated complexity analysis for boundary refinement (Appendix **C.1**).\", \"Enhanced discussion of k-NN retrieval efficiency (Appendix **C.2**).\", \"Expanded parameter tuning methodology (Appendix **D.1**).\", \"4. Clarity improvements:\", \"Updated most figures and captions for better interpretation.\", \"Strengthened connection between main text and appendix content.\", \"Refined terminology where needed for precision.\", \"We have also edited our responses to add details on the specific changes made for each point raised for your convenience. The results continue to demonstrate state-of-the-art performance on long-context tasks while maintaining computational efficiency. We remain fully committed to addressing any additional feedback you may have, whether about terminology, methodology, or experimental validation. We will promptly incorporate your suggestions and upload new revisions as needed.\", \"Best regards,\", \"The Authors\"]}", "{\"comment\": \"I am not questioning about the validity of the work. The proposed approach worked well, brining solid improvements. What I said was that the conceptual framework is similar to InfLLM, with refined process of memory formation, consolidation and retrieval. And that's ok.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}", "{\"comment\": \"Thanks for the clarifications, which are helpful.\"}", "{\"title\": \"(Answer 3/4)\", \"comment\": \"### Responses to points in weaknesses (not addressed in questions):\\n\\n* **Reviewer:** _The paper claims surprise-based segmentation is superior to fixed-length approaches (like InfLLM), but lacks theoretical justification for why this should be effective. While Figure 4 shows some empirical correlation with human segmentation, there's no analysis of why this leads to better LLM performance. The authors should provide a theoretical analysis showing why this metric captures semantically meaningful boundaries better than alternatives._\\n\\n **Authors:** Thank you for your insightful feedback. Whilst we recognize that our initial submission lacked a thorough theoretical justification for the effectiveness of surprise-based segmentation, fixed-length segmentation methods, like InfLLM, are arbitrary and lack support from linguistic or cognitive theories. Fixed length approaches are unable to guarantee segmentation of text into blocks that are contextually cohesive and distinct from each other. As an example, within a text that covers multiple distinct topics, identifying where one topic ends and another begins is challenging with fixed-window segmentation because it imposes artificial boundaries that may split coherent topics or merge distinct ones. Our surprise-based segmentation approach is inspired by cognitive science and aligns with how humans implicitly segment experiences by naturally identifying boundaries where there are shifts in meaning, context, or topic, and is supported empirically by our own results as well as those in neuroscientific literature (see sections 2.2 and 5). \\n We have also shown that surprise-based segmentations have higher similarity metrics than their fixed-length counterparts, and are therefore better at clustering similar keys together. As detailed in section 3.3, the utility of elements within an event, during memory recall for attention, depends on their likelihood of being utilised by the current query. Hence, by improving the similarity of keys within a memory unit, we also make the representative tokens more representative of the constituent keys. Therefore, we are more likely to increase the amount of useful information contained within the retrieved units, while also decreasing the amount of values which would have had low attention scores and potentially only contribute noise.\\n While most of our justifications rely on empirical results, we hope that this is adequate in addressing your point. If you still feel that some of these points are not clear enough in the paper, we will be happy to make changes to this end. As for a more \\\"theoretical\\\" justification for the performance increase, such as mathematical proofs, we believe this is not trivial to come by in a deep-learning setting where most problems are non-convex, and most research is therefore inherently empirical. \\n\\n* **Reviewer:** _Although Table 1 shows some improvements over InfLLM, the improvements on most tasks may not be statistically significant._\\n\\n **Authors:** While the vast majority of our results show a statistically significant improvement on InfLLM at the benchmark level ($p < 0.05$ using a two-tailed z-test, in all LLMs except Phi-3.5), it is true that this isn't the case in the majority of individual tasks. However, given the consistency and frequency of improvements across a large number of such tasks, along with the benchmark-level significance of such improvements, we consider the lower task-level significance to be largely due to the sample size of individual tasks rather than chance, and believe it is still reasonable and justified to claim an overall improvement on InfLLM. Moreover, including individual task results supports transparency and allows for future works to make more granular comparisons and use of such results. Of course, statistical significance testing is a very important part of empirical research so we made the following changes with the aim of being more transparent on this point:\\n\\n * **Summary of changes made to the paper:**\\n 1. Updated **Appendix A.1** to discuss the significance of results.\\n 2. Updated the caption of **Table 1** to point to this addition in **Appendix**.\"}", "{\"summary\": \"This paper introduces EM-LLM, a novel approach that integrates key aspects of human episodic memory and event cognition into LLMs with no fine-tuning, enabling them to handle practically infinite context lengths while maintaining computational efficiency\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"They introduced EM-LLM, a flexible architecture that integrates key aspects of human episodic memory and event cognition into Transformer-based LLMs. Their approach enables existing LLMs to effectively process vastly extended contexts without the need for pre-training. They perform successful passkey retrieval across 5M tokens, a length which is computationally infeasible for current full-context models\", \"weaknesses\": \"1. As you mentioned, almost all the methods used in your paper appear to have been proposed by others. Could you clarify what specific innovations you are introducing?\\n\\n2. You stated that your method outperforms others, but in Table 1, I did not see the impressive results you described. It seems that your method may not be as effective as claimed.\\n\\n3. Do you apply Eq. (1) to each token individually? If so, I imagine this would require substantial processing time. How do you address this issue?\\n\\n4. In Eq. (2), could you explain how you derive the key vectors?\\n\\n5. You employ k-NN in the MEMORY RETRIEVAL stage\\u2014does this significantly increase computation time?\\n\\n6. I found Figure 4 difficult to understand. Could you provide additional explanation to clarify it?\", \"questions\": \"The same to Weaknesses.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Firstly, we would like to re-iterate that we do not claim _equivalence_ with EM, nor do we wish to do so, as outlined in our initial response. We are still very much open to updating our wording to this end and still hope to get the reviewer's input as to what level of similarity between humans and EM they believe is more appropriate to claim given our methods, analysis, and results. We believe there is merit to our approach regardless of the exact wording used in this comparison, and are more than happy to adapt it in order to move forward with our work.\\n\\nSecondly, we agree that the simple recollection of \\\"sequences of text\\\" may not be enough to constitute an EM-inspired mechanism, but we would like to re-iterate that our approach goes beyond simple text sequence recall, unlike typical RAG methods. While event boundaries are defined at the token level, they are stored as groups of **KV pairs** in each layer and head, not as text. These KVs, generated **per layer** and **head**, enable the model to process and focus on different aspects of the sequence, including necessary contextual information, creating a natural hierarchy across layers as is inherent to Transformers (also see our initial response). The **layer-wise retrieval** of these KVs, which varies significantly across layers, further enhances the complexity of recall compared to sequential text inputs. This results in a more sophisticated information retrieval process which, combined with a transformer, has all the tools to allow for the complete recollection of \\\"experiences\\\". Appendix A.2 illustrates the behaviour of layer-wise retrieval and demonstrates that our approach vastly outperforms standard and surprise-based RAG as a result.\"}", "{\"title\": \"(Answer 1/2)\", \"comment\": \"We would like to thank reviewer N12r for their questions. We have updated our manuscript to address all points raised. Below, we provide detailed responses to each concern and we remain open to any additional feedback that could further improve our work.\\n\\n### Questions:\\n\\n* **Reviewer:** _1. As you mentioned, almost all the methods used in your paper appear to have been proposed by others. Could you clarify what specific innovations you are introducing?_\\n\\n **Authors:** We have made amendments to our paper to clarify this very important point, which we summarise here. In this work we have made *three* novel architectural contributions for LLMs, for which we show their importance both conceptually and with empirical results:\\n\\n 1. **Dynamic surprise-based segmentation.** We introduce the first method for dynamic segmentation of KV cache into blocks. Our method is also the first that manipulates the KV cache based on insights from cognitive science, using an intrinsic measure to LLMs. We show empirically using multiple LLMs that this low-cost and simple-to-implement method is able to group relevant pairs of keys and values (KV) together (relevance measured as key similarity) with much higher accuracy than fixed segmentation, the only alternative proposed approach (See Table 2 for key similarity comparisons). We also show that this method results in increased LLM performance, as mentioned in our next answer. \\n\\n 2. **Graph-based refinement.** We were the first to introduce an algorithm to refine the temporal borders of events in the context window of LLMs using graph theory. We relied on the insight that tokens are more useful to be recalled together, if the variance between their keys is low, as they need to be used by a single query at the time. This method can also stand by itself as a dynamic segmentation approach of KV cache, more computationally heavy than surprise-based segmentation but achieving a competitive accuracy in grouping relevant (KV) together (see again Table 2), while it has the extra benefit that can be used in each attention head independently, without relying on the LLM output. \\n\\n 3. **Contiguity buffer**. We were the first to introduce a method to maintain a dedicated decaying buffer in the context window of LLMs that maintains the KV cache of temporally contiguous tokens to the context window of the LLM for a certain amount of time. For this, we relied on the recent insight that self-attention heads responsible for in-context learning are shown to consecutively attend to contiguous groups, similarly to human studies (Ji-An et al., 2024). We show that this algorithm can also be combined with methods (1.) and (2.) and results in further increases in the overall LLM performance (see again next answer).\\n\\n* **Reviewer:** _2. You stated that your method outperforms others, but in Table 1, I did not see the impressive results you described. It seems that your method may not be as effective as claimed._\\n\\n **Authors:** Thank you for the question. Significant performance benefits of our model are particularly evident for QnA and retrieval tasks, and they can also largely be found when only our dynamic surprise-based segmentation is applied. In particular, surprise-based segmentation results in $16.6$% average increase over InfLLM in retrieval tasks, and $6.4$% average increase in multi-document QA tasks across all the LLMs we tried. In addition, when graph-based refinement is also added, this average performance increase jumps to $19.4$% for retrieval tasks, and to $9.2$% for multi-document QA tasks (See the columns \\\"S\\\" and \\\"SM+C\\\" in the tables of Appendix A.1). To further support our answer, we also ran statistical tests to explore whether the improvements we see over InfLLM on a benchmark level are statistically significant. The vast majority of our results (all base LLMs except phi-3.5) show a statistically significant improvement over InfLLM at the benchmark level ($p < 0.05$ using a two-tailed z-test). This analysis is now appended to our manuscript. \\n\\n Furthermore, we demonstrate broad performance benefits beyond InfLLM. Our model is the first to our knowledge to achieve superior performance compared to the state-of-the-art RAG retriever NV-Embed-v2 without requiring costly external architectural components such as larger LLMs ($36.4$% vs $51.6$% in all LongBench tasks; see Figure 1 and Table 9). The computational complexity of our method is strikingly lower than processing the complete context within the LLM, while maintaining competitive performance across tasks and surpassing full-context processing in the majority of long-context tasks ($39.3$% vs $51.6$% in LongBench; see again Figure 1). Most notably, we achieve perfect retrieval accuracy on 10M tokens while using only a single GPU - a first in the field. These characteristics establish both the novelty and practical relevance of our approach, especially since it can be applied to any LLMs without further training.\"}" ] }

[ "Jiexi Liu", "Meng Cao", "Songcan Chen" ]

[ "Irregularly Sampled Multivariate Time Series", "Attention Mechanism", "Time Series Analysis", "Representation Learning" ]

{ "note_id": [ "qB6ffuzIsk", "nLZ4cCOWaw", "n9PoMIeLSP", "mSpqh08r2d", "lUhStLANF9", "kKx9bXs1RQ", "hfU5CKkk8i", "dlOWj4xv9K", "YZ2UC6fUWC", "NxiayhRKV4", "NL40NP18FV", "CbXNphwePk", "Bq7a37YLDu", "Aj7UFWwoJm", "AFlf29fb5Y", "734BmJPKKv", "4sQuQn8x2X" ], "note_type": [ "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_review", "official_comment", "official_review", "comment", "official_comment", "official_comment" ], "note_created": [ 1732582679733, 1732117170863, 1732117279275, 1730604905004, 1730690341010, 1732116066903, 1732530284104, 1732116724945, 1730295304258, 1732116169796, 1732116919207, 1730334878981, 1732117029676, 1730621440014, 1732952152208, 1732582767060, 1732115901844 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission7168/Reviewer_XcmG" ], [ "ICLR.cc/2025/Conference/Submission7168/Authors" ], [ "ICLR.cc/2025/Conference/Submission7168/Authors" ], [ "ICLR.cc/2025/Conference/Submission7168/Reviewer_KxEq" ], [ "ICLR.cc/2025/Conference/Submission7168/Reviewer_ewtP" ], [ "ICLR.cc/2025/Conference/Submission7168/Authors" ], [ "ICLR.cc/2025/Conference/Submission7168/Reviewer_KxEq" ], [ "ICLR.cc/2025/Conference/Submission7168/Authors" ], [ "ICLR.cc/2025/Conference/Submission7168/Reviewer_pftL" ], [ "ICLR.cc/2025/Conference/Submission7168/Authors" ], [ "ICLR.cc/2025/Conference/Submission7168/Authors" ], [ "ICLR.cc/2025/Conference/Submission7168/Reviewer_XcmG" ], [ "ICLR.cc/2025/Conference/Submission7168/Authors" ], [ "ICLR.cc/2025/Conference/Submission7168/Reviewer_T9q8" ], [ "ICLR.cc/2025/Conference/Submission7168/Authors" ], [ "ICLR.cc/2025/Conference/Submission7168/Reviewer_XcmG" ], [ "ICLR.cc/2025/Conference/Submission7168/Authors" ] ], "structured_content_str": [ "{\"comment\": [\"Thank you for your response. A few followup comments:\", \"\\\"We also noticed that most of your reviews are suggestions....Therefore, we hope you can improve your evaluation of our paper\\\". I am not sure I understand this comment - the evaluation of the paper is indeed mostly in the form of suggestions for improvement.\", \"although ISMTS may be common among the papers you cite, its not a widely used term in time-series analysis (or indeed among ICLR attendees/readers in general). I encourage you to provide a precise definition early in the paper to make the paper more accessible to a wide audience. I also encourage you to include a simple definition of what you mean by \\\"missing ratio\\\" (again in the interests of being precise and avoiding ambiguity in the reader's mind) rather than asking readers to look this up in other sources.\", \"For simulation experiments, while simulations are not strictly necessary, they could be quite insightful for a reader and can provide a controlled way to explore strengths and weaknesses of a particular modeling approach (in the absence of applicable theory). I would encourage you to consider meaningful simulations that could tease out both strengths and weaknesses of your approach. And the fact that a previous paper doesn't have simulations does not mean it would not be useful.\", \"thank you for correcting the Fig 2 reference and for updating the paper title.\"]}", "{\"title\": \"Official Comment by Authors_2\", \"comment\": \"### **Q2**: what is meant in Section 3.2 by \\\"materializing its output ... time points \\\\tau = [ ..]\\\". How is k selected here? what is the sensitivity of the overall approach to the selection of k?\\n\\n**A2**: \\n\\nWe aims to learn fixed-length representations from ISMTS data, with the number of reference points, $K$, representing the chosen length of the learned representation.\\n\\nThe selection of $K$ in this paper is based on series length, with specific values provided in Appendix subsection F.1 MuSiCNet parameters, where we state: *\\u201cDue to inconsistent series lengths, we set the maximum reference point number,* $K$ *, to 128 for long series, such as P12, PAM, PhysioNet and USHCN, to 96 for PhysioNet12, and to 48 for short series, such as PAM and MIMIC-III.\\u201d* Therefore, we did not adjust $K$ in the experiments, but instead pre-set it accordingly. \\n\\n### **Q3**: at lines 245-246 how are the K reference time points defined? is this a different K to the lowercase k above?\\n\\n**A3**: The selection of $K$ in this paper is based on series length as in the above answer. These two $K$ actually refer to the same parameter. We mistakenly used lowercase $k$ instead of uppercase $K$, and we have corrected this issue in the latest revision.\\n\\n### **Q4**: the baseline methods seem a little old, at least on the timescale of advances in deep learning.\\n\\nWe must emphasize that **imputation and interpolation are two distinct tasks with different settings.**\\n\\n**Interpolation** methods are often used to provide an interface between irregularly sampled time series data, allowing for the estimation of observations at any desired time point.\\n\\n**Imputation** methods, on the other hand, first convert irregularly sampled time series data into a regularly sampled series and then fill in missing values to achieve a consistent sampling rate.\\n\\nWe find a recent study [1] and compare our interpolation task results with theirs.\\n\\n| Observed % | 50% | 70% | 90% |\\n| --- | --- | --- | --- |\\n| Model | MSE ($\\\\times 10^{-3}$) | MSE ($\\\\times 10^{-3}$) | MSE ($\\\\times 10^{-3}$) |\\n| NIERT | 2.868\\u00b10.021 | 2.656\\u00b10.041 | 2.709\\u00b10.157 |\\n| NIERT w/pretraining | 2.831\\u00b10.021 | 2.641\\u00b10.052 | 2.596\\u00b10.159 |\\n| MuSiCNet | **0.918 \\u00b1 0.025** | **0.938 \\u00b1 0.014** | **0.965 \\u00b1 0.008** |\\n\\n[1] Ding S, Xia B, Ren M, et al. NIERT: Accurate Numerical Interpolation through Unifying Scattered Data Representations using Transformer Encoder[J]. IEEE Transactions on Knowledge and Data Engineering, **2024**.\\n\\n### **Q5**: could one come up with a baseline method based on non-neural architectures\\n\\n**A5**: This task falls outside the scope of our study. Given the need for concise and focused writing in a conference paper, we do not consider this requirement essential.\"}", "{\"comment\": \"Thank you for your careful review and helpful suggestions. We note your main concerns about the writing clarity and the insight on irregular relativity. In response, we have made edits to improve readability and further clarified our insights. At the same time, other reviewers have recognized our writing and insights, with comments such as *\\u201cThe paper is clearly written and well-structured\\u201d* (Reviewer ewtP), *\\u201cThis is a high-quality paper\\u201d* (Reviewer T9q8), and *\\u201cThe paper generally explains its methodology well, with clear explanations of the key concepts and results.\\u201d* (Reviewer KxEq). We hope you will also see these strengths in our work. Below is our detailed response:\\n\\n### **W1:** The insight of irregular relativity. & **Q1:** The generation principles for the hierarchical set and its advantages.\\n\\n**A1:** Here, we address W1 and Q1 together. \\n\\nWe rethink the causes of multivariate irregular time series generation and realized that, due to various external forces or interventions mentioned in the paper\\u2019s real-world scenarios, sensors are unable to sample uniformly, resulting in ISMTS data.\\n\\nHowever, when we zoom in on the time scale, as shown in Figure 1 of the paper, larger time windows are more likely to contain real sampling values, forming relatively regular series, which helps mitigate the issue of irregular sampling because regularly sampled time series have been widely studied and demonstrate to be easier for models to learn.\\n\\nThe advantage of coarse-grained regular sequences is that they ease the learning difficulty caused by irregular sampling as shown above and provide broad-view temporal information. However, fine-grained details that may be missed at this level still need to be captured from finer-grained series, which is why we adopt a hierarchical learning approach. Moreover, it\\u2019s important to emphasize that our hierarchical learning iteratively refines the learned representations rather than simply merging them.\\n\\n### **W2:** The definition of the symbols is poorly articulated.\\n\\n**A2:** In the revised version of the paper, we have updated our symbol notation, highlighted in blue, including but not limited to the dimensions and representation of symbols.\\n\\n### **W3:** The dimensions of the variables\\u00a0Q,\\u00a0K,\\u00a0A, and\\u00a0C.\\n\\n**A3:** Followed by the definition $X_n$ in PROBLEM FORMULATION subsection, the dimensions of the variables can be formulated as $Q \\\\in \\\\mathbb{R}^{|\\\\tau| \\\\times |\\\\tau|}$, $K \\\\in \\\\mathbb{R}^{|\\\\tau| \\\\times T}$, $A \\\\in \\\\mathbb{R}^{|\\\\tau| \\\\times T}$, and $C \\\\in \\\\mathbb{R}^{D \\\\times D}$, where $|\\\\tau|$ denotes the length of reference points, $T$ denotes the length of observed timestamps, and $D$ denotes the number of observed variables in ISMTS.\\n\\n### **Q2:** The reasons that the proposed framework does not stand out on the P19 dataset?\\n\\n**A4:** According to the **no free lunch** theorem, it is natural that no single method excels universally. While our performance may not be the best on certain datasets, our non-SOTA results remain among the top-tier. Moreover, our model demonstrates superior average performance across the three downstream tasks.\\n\\nSpecifically, as described in **subsection 4.1 Time Series Classification**, \\u201cFor the P19 dataset, while our performance is competitive, MuSiCNet stands out due to its lower time and space complexity compared to ViTST. ViTST converts 1D time series into 2D images, potentially leading to significant space inefficiencies due to the introduction of extensive blank areas, especially problematic in ISMTS.\\u201d and the computational complexity is shown in subsection4.5 Ablation Analysis and Efficiency Evaluation.\\n\\nEven so, we are continually working to improve the performance of our model further.\"}", "{\"summary\": \"This paper introduces MuSiCNet, a novel framework for analyzing irregularly sampled multivariate time series (ISMTS). MuSiCNet addresses the challenges posed by ISMTS data by employing a multi-scale approach and a custom-designed encoder-decoder framework called CorrNet. The multi-scale approach allows the model to learn from both coarse-grained and fine-grained representations of the data, capturing long-term dependencies and detailed temporal variations. CorrNet utilizes time attention and frequency correlation matrices to aggregate information within and across series, improving the quality of the learned representations. The paper demonstrates the effectiveness of MuSiCNet on three mainstream ISMTS tasks: classification, interpolation, and forecasting, achieving competitive performance compared to state-of-the-art methods.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"3\", \"strengths\": \"S1: The multi-scale approach and the use of frequency correlation matrices for inter-series attention are novel contributions to ISMTS analysis.\", \"s2\": \"The framework is experimentally validated with strong results on multiple benchmarks, indicating reliability. The results demonstrate the effectiveness of the proposed method.\", \"s3\": \"The paper generally explains its methodology well, with clear explanations of the key concepts and results.\", \"s4\": \"This work addresses a broad need in ISMTS analysis, particularly in medical and environmental fields, where irregular data sampling is common.\", \"weaknesses\": \"W1. The paper's summary and comments on current research work are not accurate and rigorous enough. For example:\\n\\n(1) \\\"Most existing methods treat ISMTS as synchronized regularly sampled time series with missing values.\\\" However, many recent studies do not treat ISMTS as synchronized but as asynchronous, such as in the papers \\\"Set Functions for Time Series\\\" and \\\"Graph-Guided Network for Irregularly Sampled Multivariate Time Series.\\\"\\n\\n(2) \\\"Neglecting that the irregularities are primarily attributed to variations in sampling rates.\\\" However, \\\"A review of irregular time series data handling with gated recurrent neural networks\\\" has already pointed out and explained that different sampling frequencies lead to irregular time series data.\\n\\n(3) \\\"They rely on assumptions tailored to specific downstream tasks, hindering their ability to consistently perform well across various ISMTS tasks,\\\" yet the paper does not elaborate on or provide examples of these \\\"assumptions tailored to specific downstream tasks.\\\"\\n\\nW2. The language used in the paper is not precise enough, with extensive use of \\\"in some senses,\\\" \\\"to some extent,\\\" which makes it difficult to accurately gauge the degree the authors intend to convey.\\n\\nW3. The paper lists the model's versatility across tasks as one of its three main contributions, yet many recent models can be applied to multiple tasks. For example, the papers \\\"Latent ODEs for Irregularly-Sampled Time Series,\\\" \\\"ContiFormer: Continuous-Time Transformer for Irregular Time Series Modeling,\\\" and \\\"IVP-VAE: Modeling EHR Time Series with Initial Value Problem Solvers.\\\"\\n\\nW4. Figure 2 (a) appears to have poor readability. The caption mentions three main components, but their positions and relationships are not clearly displayed in the figure. X appears to be the input, but the final output is not clearly indicated.\\n\\nW5. The paper states, \\\"MuSiCNet stands out due to its lower time and space complexity compared to ViTST,\\\" yet there is no theoretical derivation or experimental results provided in the paper to support this claim.\\n\\nW6. The paper uses Informer, Fedformer, DLinear, etc., as baselines for time series forecasting, but these models were originally only suitable for regular time series data? How were they adapted and applied to irregular time series data? The paper does not specify.\\n\\nW7. There is a spelling error in the Conclusion section. \\\"...designed for analyzing IISMTS datasets\\\" where \\\"IISMTS\\\" should probably be \\\"ISMTS.\\\"\", \"questions\": \"1. How to accurately understand \\u201cirregularity is essentially relative in some senses\\u201d?\\n\\n2. How sensitive is MuSiCNet to the choice of hyperparameters, such as the number of scales and the masking ratio?\\n\\n3. How does the computational complexity of MuSiCNet compare to other ISMTS analysis methods?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper proposes a framework for modeling irregularly sampled multivariate time series (ISMTS). The key contributions include leveraging Lomb-Scargle Periodogram-based Dynamic Time Warping (LSP-DTW) to enhance inter-series correlation modeling and adopting a multi-scale approach to iteratively refine representations across multiple scales. The framework is evaluated across three ISMTS tasks: classification, interpolation, and forecasting.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. Hierarchical modeling for irregularly sampled multivariate time series is a promising and feasible approach.\\n\\n2. The use of DTW to measure similarity between variables and construct a correlation matrix is intuitively sound and well-motivated.\\n\\n3. The paper is clearly written and well-structured.\", \"weaknesses\": \"1. The paper\\u2019s core contributions, particularly the hierarchical design and the DTW-based correlation matrix, are not fully demonstrated in the ablation study. Key missing analyses include:\\n\\na) Comparing multi-scale modeling with single-scale modeling to quantify the benefits. An assessment of the optimal number of scales, along with the computational cost added by each scale.\\n\\nb) Evaluating the LSP-DTW correlation matrix against inter-variable self-attention, as seen in prior work (e.g., Warpformer), to confirm its advantage over attention-based correlation modeling.\\n\\n2. The LSP-DTW computation introduces additional overhead, raising concerns about efficiency. The authors should provide a detailed analysis of the computational trade-offs to demonstrate if the performance gains justify the added complexity.\\n\\n3. The paper mischaracterizes Warpformer as a multi-scale model for regularly sampled time series, while Warpformer actually pioneered multi-scale modeling for irregular time series using DTW. A detailed comparison with Warpformer, including both similarities and distinct contributions, is essential for clarity. Besides, it should be included as an important baseline to be compared with.\", \"minor_suggestion\": \"Address minor typos, such as the use of \\\\citep{} and \\\\cite{}.\", \"questions\": [\"Can MuSiCNet be viewed as a hierarchical extension of mTAND, with the addition of a correlation matrix to capture inter-variable similarity?\", \"The approach to handling irregularity in MuSiCNet closely resembles that of mTAND, aside from the use of DTW. Are there any other unique design elements aimed at better adapting to irregular data?\", \"What specific challenges does hierarchical modeling pose for irregularly sampled time series compared to regular time series? The method described here\\u2014using regular time intervals with average pooling (e.g., hourly aggregation)\\u2014is commonly applied to regular time series. What irregularity challenges does this approach address, and which experiments support this claim?\", \"See also Weaknesses.\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Official Comment by Authors_2\", \"comment\": \"### **Q1:** Can MuSiCNet be viewed as a hierarchical extension of mTAND, with the addition of a correlation matrix to capture inter-variable similarity? & **Q2:** The approach to handling irregularity in MuSiCNet closely resembles that of mTAND, aside from the use of DTW. Are there any other unique design elements aimed at better adapting to irregular data?\\n\\n**A4:** Here, we address Q1 and Q2 together.\\n\\n**We must emphasize that our work fundamentally differs from mTAND.**\\n\\nWhile mTAND aims to design a model specifically for handling irregular sampling by aggregating intra-series information to learn the fixed-dimensional feature.\\n\\nIn contrast, we present a novel perspective: irregularity is essentially relative in some senses. **We focus on the data itself, utilizing the perspective of relative regularity to iteratively refine and complement the representations of the original irregular series, thereby addressing the irregularity issue.**\\n\\nWe highlight the importance of introducing multi-level learning and capturing intra-instance relationships for ISMTS, particularly considering the sparse (see Table 7 in the appendix) yet valuable observations in such datasets.\\n\\nUnder this framework, **the generated coarse-grained regular series effectively mitigate inter-series misalignments** and reduce the influence of the attention weights diminishing with distance between observations, which is a prior issue in most modeling intra-series irregularities methods. Additionally, **coarse-grained regular series mitigates modeling difficulties associated with intra-irregularities**. Our model is capable of progressively learning from relatively regular to irregular series, which aligns closely with the intuition of data-level curriculum learning [3].\\n\\nThis contribution represents a key distinction from mTAND and underscores one of our significant innovations.\\n\\n### **Q3:** What specific challenges does hierarchical modeling pose for irregularly sampled time series compared to regular time series? The method described here\\u2014using regular time intervals with average pooling\\u2014is commonly applied to regular time series. What irregularity challenges does this approach address, and which experiments support this claim?\\n\\n**A5:** \\n\\nThis issue can actually be addressed by the explanation provided above in **A4**. Here, we additionally clarify the use of average pooling in ISMTS.\\n\\nTaking a sequence of length 16 as an example, its observation timestamps are \\\"**1** 2 3 4 5 6 **7** **8** 9 10 11 12 13 14 **15** 16,\\\" and we assume the observed values are the same as the timestamps. Bold indicates the presence of observed variables. After one sampling with a window size of 4 and a sampling rate of 4, the sequence becomes \\\"**1** **7.5** ? **15**,\\\" where \\\"?\\\" denotes unobserved timestamps. After the second sampling, the sequence becomes \\\"**7.83**.\\\"\\n\\nMoreover, average pooling is merely a simple strategy we chose, and it can effectively serve the purpose we require. If simple methods can effectively address the problem, we adhere to Occam's Razor, avoiding unnecessary additional strategies. However, since standard DTW methods cannot be effectively applied to ISMTS data, we specifically designed LSP-DTW for this purpose.\\n\\n**Minor suggestion:** Address minor typos, such as the use of \\\\citep{} and \\\\cite{}.\\n\\n**A6:** We have made the revisions.\\n\\n[3] Wang X, Chen Y, Zhu W. A survey on curriculum learning[J]. IEEE transactions on pattern analysis and machine intelligence, 2021, 44(9): 4555-4576.\"}", "{\"comment\": \"Thanks for your response. I have no more questions and will maintain my score.\"}", "{\"comment\": \"Thank you for your positive feedback in the Strengths section. We noticed that your concerns primarily revolve around related work, certain writing approaches, and experimental details. We have addressed these concerns and revised the paper accordingly. We hope our efforts will enhance your recognition of our work, and we would greatly appreciate it if you could consider raising your score for our paper. If you have any further questions, please feel free to ask\\u2014we are more than happy to provide additional clarification. Below are our detailed responses.\\n\\n### **W1:** The paper's summary and comments on current research work are not accurate and rigorous enough.\\n\\n**A1:** 1. When we state, \\\"Most existing methods treat ISMTS as synchronized regularly sampled time series with missing values,\\\" we use \\\"most\\\" to acknowledge that not all papers take this approach; otherwise, we would have used \\\"all.\\\"\\n2. In the sentence, \\\"Neglecting that the irregularities are primarily attributed to variations in sampling rates,\\\" this phrase directly follows the previous statement 1 above as part of the same sentence, separated by a comma. Thus, we mean that methods treating ISMTS as synchronized regularly sampled time series with missing values often overlook this aspect. Given that many studies do approach ISMTS in this way, but our own method does not follow this method, we found it necessary to briefly point out this potential limitation in the abstract. \\n\\nAs you mentioned, irregularly sampled time series indeed arises from varying sampling frequencies. However, many existing methods either fail to recognize this or do not leverage this information in their modeling. Our approach aims to effectively utilize this characteristic to better model ISMTS.\\n\\n1. Here, we refer to the fact that many existing ISMTS analysis models are designed for specific tasks, such as prediction, classification, or imputation. We have updated our manuscript to include additional references on this. In contrast, our work focuses on learning a good representation that can adapt to multiple downstream tasks.\\n\\n### **W2:** The language used in the paper is not precise enough.\\n\\n**A2:** The phrase \\\"irregularity is essentially relative in some senses\\\" reflects our perspective on ISMTS based on varying sampling rates. Another perspective considers ISMTS as regularly sampled time series with missing values, attributing the irregularities to missing data. Given these differing viewpoints, it would be inaccurate to state outright that irregularities are relative. However, since we already use the term \\\"mitigate,\\\" we have decided to remove \\\"to some extent.\\u201d\\n\\n### **W3:** The paper lists the model's versatility across tasks as one of its three main contributions.\\n\\n**A3:** We emphasize here that our primary focus is **on learning a good representation, enabling our model to be not limited to a specific analysis task but to serve as a task-general framework for ISMTS analysis**.\\n\\n**While some existing works are capable of handling multiple tasks, such cases are relatively rare, and the tasks are addressed differently.** For instance, ContiFormer performs classification and event prediction, IVP-VAE focuses on classification and forecasting, and Latent ODE tackles interpolation, extrapolation, and some dataset-specific tasks.\\n\\nTherefore, we believe that achieving multiple tasks remains one of the notable contributions of our work at this stage.\\n\\n### **W4**: Figure 2 (a) appears to have poor readability.\\n\\n**A4**: (a) represents the overall framework of our model. Based on the title of (b), we can see that it corresponds to the **CorrNet Encoder** in (a). (c) illustrates the **Frequency Correlation Matrix**, which is the method used in (b) to compute the inter-series relationships in the **Correlation Matrix**.\\n\\n$X$ denotes the input, and the representation $r^{(L)}$ learned by the final layer, $L$, of the **CorrNet** is considered the model's output, as all subsequent downstream tasks are performed on $r^{(L)}$.\"}", "{\"summary\": \"This work focuses on the challenges arising from data irregularities and introduce an innovative framework MuSiCNet designed for analyzing ISMTS datasets. Through comprehensive experiment discussions, the manuscript demonstrates the effectiveness of the proposed approach beyond existing works.\", \"soundness\": \"2\", \"presentation\": \"1\", \"contribution\": \"2\", \"strengths\": \"S1. The research field is meaningful, and the research motivation of this work is clear.\\n\\nS2. The experiment discussion is comprehensive.\", \"weaknesses\": \"W1. I am confused by the insight of irregular relativity. From the way it is presented in the introduction, I find it difficult to see this as a novel insight. And I remain skeptical about whether this insight can effectively guide the resolution of existing challenges.\\n\\nW2. The definition of the symbols is poorly articulated. What does \\\"with the length of observation $ T_n$\\\" mean? What do $ T_n$ and $ n$represent, respectively?\\n\\nW3. What are the dimensions of the variables $ Q $, $ K$, $ A $, and $ C$ in Equation 1? They are not defined before their use.\\n\\nW4. The writing in the manuscript is quite rough. While I don't deny that this may be an interesting work, it definitely needs substantial revision in terms of writing style. I argue it's crucial to clarify the details and enhance the coupling between each module.\", \"questions\": \"Q1. What are the generation principles for the hierarchical set? What advantages does this generation mechanism offer?\\n\\nQ2. On the P19 dataset, what are the reasons that the proposed framework does not stand out?\", \"a_typo\": \"Line 497, \\\"IISMTS\\\".\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"We sincerely thank the reviewer for their careful reading and insightful comments. Your feedback primarily focuses on the performance in the forecasting task, the details of multi-scale training, and the experimental validation of coarse-grained series. We have addressed these points and hope our responses resolve your concerns, encouraging you to consider raising your rating of our work. If you have any further questions, please feel free to ask. Below are our detailed responses:\\n\\n### **Q1:** In the TIME SERIES FORECASTING task, the model is not SOTA on 2 out of 3 datasets. Further analysis is needed. Also, model complexity, training, and inference costs compared to baselines should be provided.\\n\\n**A1:**\\n\\n1. According to the no free lunch theorem, it is natural that no single method excels universally. While our performance may not be the best on certain datasets, our non-SOTA results remain among the top-tier. Moreover, our model demonstrates superior average performance across the three downstream tasks. Specifically, as described in **subsection 4.3 Time Series Forecasting**, the model achieving SOTA results in the forecasting task is specifically designed for that purpose, incorporating priors designed for forecasting. In contrast, our MuSiCNet remains competitive without relying on task-specific priors. Even so, we are continually working to improve the performance of our model further.\\n\\n2. We conduct experiments on the P12 dataset with a batch size of 50. Empirically, our model MuSiCNet requires 0.240s per batch with 4.2GB memory usage, compared to ViTST\\u2019s 2.196s and 40.2GB, Raindrop\\u2019s 0.124s and 4.8GB, MTGNN\\u2019s 0.1967s and 4.2GB, and DGM$^2$-O\\u2019s 0.313s and 9.1GB. Our model achieves significantly lower time complexity than the other methods and uses much less memory than ViTST, which also performs well on classification tasks. We will add this in section 4.5 \\u201cAblation Analysis and Efficiency Evaluation\\u201d.\\n\\n### **Q2:** Regarding the multi-scale learning method, it's unclear if the number of training epochs is the same as without multi-scale training. Also, the relationship between sample numbers at different scales needs clarification.\\n\\n**A2: In our setup, the training epochs remain the same and are not affected by the multi-scale design.**\\n\\nIn Appendix subsection F.1 MUSICNET PARAMETERS, we explain the selection method for *the number of scales*. **This approach does not involve manual selection.** According to the observed timestamps in each dataset, we ensure that in the (L-1)-th layer, most windows contain at least one sampling point, while the L-th layer represents the original data layer.\\n\\n### **Q3:** Experimental validation of coarse-grained series.\\n\\n**A3:** From the sensitivity analysis in Figure 5, we can find that compared to $\\\\lambda_2$, the hyperparameter of the adjustment term, i.e., $\\\\lambda_1$, plays a greater role, which aims to utilize broader temporal information (larger scale) to guide the effectiveness of detailed temporal information (smaller scale), to boost the representation learning. If $\\\\lambda_1$ did not become a key role, then our MuSiCNet would definitely lose the guidance of broad-view temporal information. However, this did not happen. Moreover, Scale 1 in Figure 1 clearly illustrates what the broad-view trend information. Therefore, the above phenomenon can indirectly validate that the coarse-grained series can help the representation capture broad-view temporal information.\"}", "{\"title\": \"Official Comment by Authors_2\", \"comment\": \"### **W5:** The paper states, \\\"MuSiCNet stands out due to its lower time and space complexity compared to ViTST,\\\" yet there is no theoretical derivation or experimental results provided in the paper to support this claim.\\n\\n**A5:** We conduct experiments on the P12 dataset with a batch size of 50. Empirically, our model MuSiCNet requires 0.240s per batch with 4.2GB memory usage, compared to ViTST\\u2019s 2.196s and 40.2GB, Raindrop\\u2019s 0.124s and 4.8GB, MTGNN\\u2019s 0.1967s and 4.2GB, and DGM$^2$-O\\u2019s 0.313s and 9.1GB. Our model achieves significantly lower time complexity than the other methods and uses much less memory than ViTST, which also performs well on classification tasks. We will add this in section 4.5 \\u201cAblation Analysis and Efficiency Evaluation\\u201d.\\n\\n### **W6:** How were they adapted and applied to irregular time series data?\\n\\n**A6:** We share the same setting with [1], so we directly used their results. And they did not mention how they were adapted.\\n\\n### **W7:** There is a spelling error in the Conclusion section.\\n\\n**A7:** We carefully review the entire paper and correct spelling errors.\\n\\n---\\n\\n### **Q1:** How to accurately understand \\u201cirregularity is essentially relative in some senses\\u201d?\\n\\n**A8:** This means that, to some extent, irregular sampling can be transformed into regular sampling. Specifically, as mentioned in our paper: *\\\"With sampling rates artificially determined from low to high, an irregularly sampled time series can be transformed into a hierarchical set of relatively regular time series from coarse to fine.\\\"*\\n\\n### **Q2:** How sensitive is MuSiCNet to the choice of hyperparameters, such as the number of scales and the masking ratio?\\n\\n**A9:** In Appendix subsection F.1 MUSICNET PARAMETERS, we explain the selection method for *the number of scales*. This approach does not involve manual selection. According to the observed timestamps in each dataset, we ensure that in the (L-1)-th layer, most windows contain at least one sampling point, while the L-th layer represents the original data layer.\\n\\nAs for the *masking ratio*, we set it uniformly to 10%, as mentioned in subsection F.1. This is because the missing rate in ISMTS is already significantly high. Therefore, we follow the smallest masking ratio in MAE [1], which is 10%, and apply it to all the experiments in this paper.\\n\\nConsequently, we did not conduct sensitivity analysis for these two parameters. However, we conduct sensitivity analysis experiments in subsection F.3 Parameter Analysis in the appendix as follows\\n\\n\\u201cTo analyze the hyper-parameters sensitivity, we conducted the experiments for $\\\\lambda_1$, $\\\\lambda_2$, and $\\\\lambda_3$ with grid search. Due to the closer relationship between the hyper-parameters of the adjustment term and the contrastive learning term, i.e., $\\\\lambda_1$ and $\\\\lambda_2$, we jointly analyzed $\\\\lambda_1$ and $\\\\lambda_2$ while separately analyzing the hyper-parameter of the downstream task $\\\\lambda_3$, as illustrated in Figure 4 and 5.\\nFrom Figure 4, we can find that the adjustment term plays a greater role compared to the contrastive learning term.\\nThis phenomenon matches our motivation, where the coarse-to-fine strategy can effectively alleviate the difficulty of representation learning on ISMTS caused by inconsistent sampling rates. In addition, when $\\\\lg \\\\lambda_1$ and $\\\\lg \\\\lambda_2$ take values around 2 and -2, respectively, our MuSiCNet can perform well.\\nFrom Figure 5, we can find that our MuSiCNet becomes effective with large $\\\\lambda_3$.\\nThis indicates that more effective representations will be captured when utilizing downstream tasks, matching the general insight.We also noticed that it becomes less sensitive when $\\\\lg \\\\lambda_3 \\\\ge -1$. Its suitable range may be located at $\\\\left[1e1, 1e2\\\\right]$.\\u201c\\n\\n[1] He, Kaiming, et al. \\\"Masked autoencoders are scalable vision learners.\\\"\\u00a0*Proceedings of the IEEE/CVF conference on computer vision and pattern recognition*. 2022.\\n\\n### **Q3:** How does the computational complexity of MuSiCNet compare to other ISMTS analysis methods?\\n\\n**A10:** Please refer to our response in Weakness 5.\"}", "{\"summary\": \"The paper addresses the problem of representation learning for irregularly sampled multivariate time series, where different individual time-series are not synchronized in terms of their measurement times, making it challenging to apply standard time-series modeling approaches. To address this problem the authors propose a multi-scale hierarchical representation scheme built on a combination of ideas from classical time-series modeling (periodograms, dynamic time-warping) and deep learning (encoder-decoder architectures, attention mechanisms, rectification). They then evaluate their proposed methodology across three different time-series tasks (classification, interpolation, forecasting) and compare the effectiveness of their approach to a variety of baselines.\", \"soundness\": \"2\", \"presentation\": \"1\", \"contribution\": \"1\", \"strengths\": [\"The problem being addressed is important and challenging and occurs in multiple real-world applications such as medical time-series analysis\", \"The paper describes extensive experimental investigations of the proposed methodology, comparing it with a variety of baselines, across classification, interpolation, and forecasting tasks.\"], \"weaknesses\": [\"Contributions and Potential Impact:\", \"The approached proposed in the paper seems like a potentially useful engineering contribution, one that may work well for time-series problems with certain characteristics - but it less clear if the paper provides any significant advance in general time-series modeling methodology that would be of interest to the ICLR community and to time-series researchers more broadly.\", \"A general issue with the paper is that the proposed methodology combines quite a few different heuristic choices, resulting in a complex overall model (e.g., Figure 2) that involves multiple hyperparameters and design choices. As a consequence, given the types of inductive biases that are being built into the approach (implicitly, via the design choices), it makes sense to wonder what types of problems the proposed methodology will work well, and on what types of problems it will have limitations. The paper would be much stronger if it could be extended to provide this type of insight to the reader: for example, see suggestions below about potentially expanding the discussion of limitations, and possible inclusion of simulations to provide additional insight.\"], \"writing\": [\"The paper could be improved and be of more value to a reader by improving the writing.\", \"As an example, there are a few key terms used throughout the paper where it would be of great help to the reader if the terms were more clearly and precisely defined. For the term \\\"ISMTS\\\" its unclear what the scope of this term is: can each of the individual time-series be sampled at arbitrary times, in the general case? or are there any restrictions on this level of generality? You could use Figure 1 to help explain this to the reader: from looking at a blown-up version of Figure 1 (scale L), it would appear that there are no restrictions and that each time-series can be sampled at any arbitrary set of times: is this the case?. As an example of where this is described much more clearly, see Figure 1 in the Yalavarthi et al (2024) paper where the authors clearly illustrate different sampling scenarios and make it clear which one they are focusing on in their paper.\", \"Another key term that is unclear is \\\"missing ratio\\\" which is used to characterize the datasets in Section 4: how is this defined? does it imply that all of the time-series are potentially being measured on the same discrete time-scale (e.g., hourly or daily) but that some of the time-series don't have measurements at each discrete time (e.g., are only measured every few hours or every few days) and the \\\"missing ratio\\\" is the number of missing measurements relative to this fixed sampling scheme? If so, then this would seem to imply that the datasets used in experiments are a special case of the general framework (since they all have a \\\"missing ratio\\\" defined, Table 7) which has implications for the interpretation of the experimental results and limits the generality of the claims earlier in the paper. Another possibility is that the \\\"missing ratio\\\" has a different interpretation than my guess above. Whichever is the case, it needs to be clearly defined for the reader.\", \"As another example related to writing, in section 3.2, the key novel contribution of the paper, the CorrNet architecture is introduced. The writing here could be significantly improved by trying to impart more insight to the reader, for example by providing a clear and intuitive toy example (for example with just 3 time-series) of what the method is doing. The current description comes across as somewhat black-box in nature. In particular, while the inclusion of different components (attention, correlation, rectification, etc) each seem sensible from a high-level viewpoint, the reader may wonder how the different parts will work together. For example, it would be helpful to be realistic here and explain when and why these methods should work well, and when we might expect them to fail.\", \"Figure 2 in its current form will be very difficult for a reader to understand. At a minimum I suggest that you devote part of the text in the paper to a clear description of how data flows through the model to accompany Figure 2 (currently the text mentions Fig 2(a) in one location, and Fig 2(b) in another and I didn't see a reference to Fig 2(c) in the main text, but may have missed it). You may need to change the figure so that you can more clearly describe the steps. (again, using Yalavarthi et al (2024) as an example, their Figure 3 is a much clearer representation of information flow and their general approach than your Figure 2). Alternatively, skip the figure and use the space to more clearly describe the mapping (input-output) that it represents, with a clear sequence of equations and/or text.\"], \"discussion_of_limitations\": [\"The discussion of limitations of the approach (middle of page 10) is limited in terms of detail and scope. Readers would find it very helpful here to have a more realistic evaluation of the strengths and weaknesses of the approach. For example: what is the sensitivity of the method to architecture design choices? how sensitive are the results to hyperparameter tuning? are there situations where the LSP/DTW approach will fail? And so on.\"], \"insights_from_simulations\": [\"simulation results could also be quite helpful in providing insights into the method. For example, you could simulate datasets from some predefined (known) continuous-time multivariate temporal process and then evaluate, both theoretically and empirically, the effect of different sampling schemes for the time series, potentially making connections to concepts such as Nyquist sampling in this context, and how such concepts might be relevant to your proposed approach. This would be a valuable addition to the paper and complement the leaderboard-style experiments in Section 4.\"], \"minor_suggestion\": [\"a name other than \\\"MusicNet\\\" is worth considering since the name \\\"MusicNet\\\" will immediately suggest to a reader that this paper is about neural networks for music (rather than the actual more general topic of the paper). And there is a well-known \\\"Musicnet\\\" dataset already in the literature.\"], \"questions\": [\"please clarify definitions of the terms ISMTS and \\\"missing ratio\\\" (see weaknesses above)\", \"what is meant in Section 3.2 by \\\"materializing its output ... time points \\\\tau = [ ..]\\\". How is k selected here? what is the sensitivity of the overall approach to the selection of k?\", \"at lines 245-246 how are the K reference time points defined? is this a different K to the lowercase k above?\", \"the baseline methods (e.g., for classification at bottom of page 6, and for interpolation at the bottom of page 7) seem a little old, at least on the timescale of advances in deep learning, with most references being from the 2018-2021 period. Are there any more recent baselines that are more SOA and that would be worth comparing to? For example, for the interpolation task in particular it seems like there are quite a few more recent (and more accurate) baseline methods that you could have considered (see for example the list of post-2019 methods in Table 1 of Wang et al (2024) https://arxiv.org/pdf/2402.04059)\", \"in your forecasting experiments, is there a reason you omitted the LinODEnet method from Scholz et al (2023): in the Yalavarthi et al (2024) paper this was the next-best method to their Grafiti method, so it seems like it would be worth including in your evaluation.\", \"could one come up with a baseline method based on non-neural architectures, such as some form of a multivariate Gaussian process for example? This might not work particularly well empirically and might require some ad-hoc pre-processing to implement, but some discussion of why and when such an approach would not work well, and/or a demonstration of this in your experiments, would be helpful to the reader. (The non-neural approach need not be based on a Gaussian process, this is just one possible choice of a non-parametric model).\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you for your suggestions. We noted your concerns regarding Writing, the Discussion of Limitations, and Insights from Simulations. Below, we provide specific responses to each of your points. **We also noticed that most of your reviews are *suggestions*, such as adding a hyperparameter sensitivity analysis or including toy data to validate the model\\u2019s effectiveness. Therefore, we hope you can improve your evaluation of our paper.** If any questions remain or you feel unable to increase the paper\\u2019s score, please feel free to ask, we would be glad to clarify.\\n\\nRegarding your comments on writing, our responses are as follows:\\n\\n1. The term ISMTS, which we use throughout our paper, is common in the field, not coined by us. For example, the paper by Yalavarthi et al. (2024) also uses \\\"irregularly sampled multivariate time series,\\\" and this term appears in their Figure 1. Moreover, though we did not use a specific figure to introduce ISMTS, we define ISMTS and its key characteristics and draw Figture 1 to facilitate readers' understanding in the first paragraph of the introduction.\\n2. In real-world applications, sensors cannot record continuously, so the data collected is necessarily discrete. \\u201cMissing ratio\\u201d is a common term that describes the percentage of missing observations in a time series (similar to \\u201csparsity\\u201d in Yalavarthi et al. (2024), as you noted). Although the \\u201cmissing ratio\\u201d assumes regular sampling, it is widely used to describe ISMTS data characteristics. This term is also explained on the dataset website referenced in our paper and many other related work.\\n3. Our paper focuses on real-world datasets with substantial missing ratios, which are influenced by various external forces or interventions, leading to irregular sampling. Moreover, the effectiveness of our model can be demonstrated by the performance of the downstream tasks performed on datasets with high missing ratios and is unnecessary to use toy data.\\n4. We have clarified in the legend of Figure 2 that gray arrows indicate the direction of data flow and the subsection \\u201cEncoder-Decoder Framework\\u201d also detailly introduces the whole model processing. Additionally, we corrected the error in the subsection \\u201cCorrelation Extraction\\u201d, where we mistakenly referred to Fig.2(c) as Fig.2(b), and have updated this in the revised version.\\n\\nBased on the **discussion of limitations,** we conduct sensitivity analysis experiments in subsection F.3 Parameter Analysis in the appendix as follows\\n\\n\\u201cTo analyze the hyper-parameters sensitivity, we conducted the experiments for $\\\\lambda_1$, $\\\\lambda_2$, and $\\\\lambda_3$ with grid search. Due to the closer relationship between the hyper-parameters of the adjustment term and the contrastive learning term, i.e., $\\\\lambda_1$ and $\\\\lambda_2$, we jointly analyzed $\\\\lambda_1$ and $\\\\lambda_2$ while separately analyzing the hyper-parameter of the downstream task $\\\\lambda_3$, as illustrated in Figure 4 and 5.\\nFrom Figure 4, we can find that the adjustment term plays a greater role compared to the contrastive learning term.\\nThis phenomenon matches our motivation, where the coarse-to-fine strategy can effectively alleviate the difficulty of representation learning on ISMTS caused by inconsistent sampling rates. In addition, when $\\\\lg \\\\lambda_1$ and $\\\\lg \\\\lambda_2$ take values around 2 and -2, respectively, our MuSiCNet can perform well.\\nFrom Figure 5, we can find that our MuSiCNet becomes effective with large $\\\\lambda_3$.\\nThis indicates that more effective representations will be captured when utilizing downstream tasks, matching the general insight.We also noticed that it becomes less sensitive when $\\\\lg \\\\lambda_3 \\\\ge -1$. Its suitable range may be located at $\\\\left[1e1, 1e2\\\\right]$.\\u201c\\n\\nRegarding the question on **Insights from Simulations**, the goal of our paper is to learn effective representations of ISMTS, which can be validated through good performance on downstream tasks, demonstrating the quality of the learned representations. Additional toy data is not necessary to illustrate this point. Additionally, the frequently mentioned Yalavarthi et al. (2024) paper also does not employ simulation experiments.\\n\\nAccording to the **minor suggestion**, we have changed the paper title to \\u201cA Gradual Coarse-to-Fine Framework for Irregularly Sampled Multivariate Time Series Analysis.\\u201d\"}", "{\"summary\": \"1. Innovation: The paper proposes a novel MuSiCNet framework for the analysis of irregularly sampled multivariate time series, which is innovative. It views irregularity from a new perspective and effectively solves the problems of existing methods through multi-scale and multi-correlation attention mechanisms.\\n 2. Rationality of the method: The method is reasonably designed. For example, using LSP - DTW to calculate the frequency correlation matrix can effectively handle the correlation problem in irregularly sampled time series and avoid the spurious correlations that may be generated by existing distance measurement methods. The processing methods within and across adjacent scales have theoretical bases and their effectiveness has been verified through experiments.\\n 3. Adequacy of experiments: The experimental part is very sufficient. In three mainstream tasks of classification, interpolation, and forecasting, multiple real-world datasets (such as P19, P12, PAM, PhysioNet, USHCN, MIMIC - III, etc.) are used to compare with a variety of advanced methods. The results show that MuSiCNet is competitive, verifying the effectiveness and generality of the framework.\\n 4. Limitations and suggestions: The paper also points out some limitations, such as the interaction between scales may be further simplified and the exploration of anomaly detection tasks is insufficient. It is recommended that the authors further study these problems in future work to improve the performance and applicability of the model.\\n 5. Overall evaluation: This is a high-quality paper that proposes a promising method for the analysis of irregularly sampled multivariate time series and makes an important contribution to the research in this field. It is recommended to be accepted.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": [\"Innovation: The proposed MuSiCNet framework is novel. It offers a new perspective on irregularity and effectively addresses existing method problems via multi-scale and multi-correlation attention mechanisms.\", \"Rationality of the method: The method is well-designed. LSP - DTW for calculating the frequency correlation matrix is effective in handling correlations and avoiding spurious ones. Processing methods within and across scales have theoretical support and experimental verification.\", \"Adequacy of experiments: The experiments are comprehensive. Multiple real-world datasets are used in three mainstream tasks, and comparisons with advanced methods show the competitiveness of MuSiCNet, validating its effectiveness and generality.\"], \"weaknesses\": [\"In the forecasting task performance, the model isn't SOTA(state-of-the-art) in two out of three datasets.\", \"Moreover, there's no comparison of model complexity, training, and inference costs with baselines, making it hard to evaluate its efficiency.\", \"Some motivations have not been proved by experiments directly.\"], \"questions\": [\"Performance in FORECASTING task: In the TIME SERIES FORECASTING task, the model is not SOTA on 2 out of 3 datasets. Further analysis is needed. Also, model complexity, training, and inference costs compared to baselines should be provided.\", \"Multi-scale training details: Regarding the multi-scale learning method, it's unclear if the number of training epochs is the same as without multi-scale training. Also, the relationship between sample numbers at different scales needs clarification.\", \"Experimental validation of coarse-grained series: There is a lack of sufficient experimental analysis to prove that the coarse-grained series can help the representation capture broad-view temporal information as claimed. Although an ablation study shows the effectiveness of a related loss term, it doesn't directly prove this crucial aspect, which may affect the method's theoretical and practical credibility.\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"withdrawal_confirmation\": \"I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors.\"}", "{\"comment\": [\"thank you for the clarification about k and K. This seems like an important hyperparameter for the method, one that needs to be selected heuristically by a user. I suggest you to give more attention to discussing the selection of K in the main section of the paper and potentially some discussion of potential sensitivity of the method to its choice.\", \"thank you for adding the comparison to NIERT (with the numbers for their method from Table VII in their paper for the Physionet data? and your numbers taken from Table 2 in your paper?). Your numbers show a roughly factor of 3 reduction in MSE over NIERT. If you plan to include these results in the next iteration of your paper it would be helpful to provide the reader with some intuition as where the major reduction in error is coming from. Also, given that there is such a large reduction, it would be good to double-check the results, for example by independently replicating their results by directly running their code and your code on the same datasets.\", \"any clarification about why the LinODEnet method was not included in your experiments?\", \"I disagree that non-neural methods are \\\"outside the scope.\\\" There is a large literature on methods for non-parametric curve fitting and for interpolation in general and it would be of great benefit to a reader to show how your approach (and the other baselines) compare to some standard non-neural approaches. You would not need to spend much time on them in the main paper, just include a few in your results, with more details relegated to the Appendix.\"]}", "{\"comment\": \"Thank you for recognizing our work and providing insightful feedback. Your comments mainly focus on the need for additional analysis, clarifying the relationship between our work and mTAND, and revisiting Warpformer. In our responses, we have highlighted the sections where these analyses are located, revised some text to emphasize the fundamental differences between our work and mTAND, and included a review of Warpformer along with adding it as a baseline for comparison. We hope our responses address your concerns and improve your evaluation of our paper. If you have further questions, please feel free to reach out. Below are our detailed responses.\\n\\n### **W1: There are some key missing analyses:**\\n\\na) Comparing multi-scale modeling with single-scale modeling to quantify the benefits. An assessment of the optimal number of scales, along with the computational cost added by each scale.\\n\\n**A1: We have already conducted a comparison between multi-scale and single-scale approaches in Table 6 of subsection 4.5.** The third row in Table 6 represents the single-scale results, which we found to be inferior to the performance of the full model.\\n\\nSearching for the optimal number of scales is time-consuming and labor-intensive. In Appendix subsection F.1, MUSICNET PARAMETERS, we explain that **our approach avoids manual selection.** **Based on the observed timestamps in each dataset**, we ensure that in the $(L-1)$-th layer, most windows contain at least one sampling point, while the $L$-th layer represents the original data layer. Under this setup, we are able to achieve good performance while saving computational resources.\\n\\nb) Evaluating the LSP-DTW correlation matrix against inter-variable self-attention, as seen in prior work (e.g., Warpformer), to confirm its advantage over attention-based correlation modeling.\\n\\n**A2: In Subsection 4.4, *Correlation Results*, we conduct a comprehensive comparison with prior work to demonstrate the necessity, effectiveness, and efficiency of the correlation matrix.**\\n\\nThe inter-variable self-attention in Warpformer can be viewed as a learnable correlation matrix derived through attention, as shown in the 5th row of Table 4, which does not perform well in our model. While determining the best correlation calculation method for all models is beyond the scope of this work, our focus is on identifying the most suitable approach for our specific model.\\n\\nc) The paper mischaracterizes Warpformer as a multi-scale model for regularly sampled time series. And it should be included as an important baseline to be compared with.\\n\\n**A3:** We mistakenly cite a paper by an author with the same surname. We have removed this reference from the current sentence and revisited Warpformer appropriately in the *Related Work* section as follows\\n\\n**\\u201cAs far as we know, [1] and [2] are among the earlier works on multi-level ISMTS learning. [1] addresses multi-resolution signal issues by distributing signals across specialized branches with different resolutions, where each branch employs a Flexible Irregular Time Series Network (FIT) to process high- and low-frequency data separately. [2], on the other hand, is a transformer-based model that stacks multiple Warpformer layers to produce multi-scale representations, combining them via residual connections to support downstream tasks. These works typically focus on either specific tasks or particular model architectures. In contrast, our design philosophy originates from ISMTS characteristics rather than being tied to a specific feature extraction network structure. Warpformer emphasizes designing a specific network architecture but involves high computational costs and requires manually balancing the trade-off between the number of scales and the dataset. These are challenges that our MuSiCNet avoids entirely.\\u201d**\\n\\nSince Warpformer uses a different benchmark than the one followed in our experiments, we **reproduced Warpformer's classification results on the benchmark dataset we used**. The results are as follows and have been included in the paper.\\n\\n| model | P19 | | P12 | | PAM | | | |\\n| --- | --- | --- | --- | --- | --- | --- | --- | --- |\\n| | AUROC | AUPRC | AUROC | AUPRC | Accuracy | Precision | Recall | F1 score |\\n| Warpformer | 88.8\\u00b11.7 | 55.2\\u00b13.9 | 83.4\\u00b10.9 | 47.2\\u00b13.7 | 94.3\\u00b10.6 | 95.8\\u00b10.8 | 94.8\\u00b11.0 | 95.2\\u00b10.6 |\\n| MuSiCNet | 86.8\\u00b11.4 | 45.4\\u00b12.7 | 86.1\\u00b10.4 | 54.1\\u00b12.2 | 96.3\\u00b10.7 | 96.9\\u00b10.6 | 96.9\\u00b10.5 | 96.8\\u00b10.5 |\\n\\n[1] Singh B P, et al. Multi-resolution networks for flexible irregular time series modeling (multi-fit)[J]. arXiv preprint arXiv:1905.00125, 2019.\\n\\n[2] Zhang J, et al. Warpformer: A multi-scale modeling approach for irregular clinical time series[C]//Proceedings of the 29th KDD. 2023: 3273-3285.\"}" ] }

[ "Ziyi Yin", "Yuanpu Cao", "Han Liu", "Ting Wang", "Jinghui Chen", "Fenglong Ma" ]

[ "multimodal large language models", "jailbreak", "defense" ]

{ "note_id": [ "xduqC5ZwYp", "w9XedgHyTJ", "w7TaP5Dd6T", "tL8KnEaVJ4", "sGCobaUqcD", "sA6kG9JGpm", "qgXEOfDh29", "mPXuZuPDZj", "m1pNMju8Ss", "lhSoIue8zP", "lcB0u15sss", "jRhIhwA5M9", "jQ0BfcTp6i", "jCF9XapPJS", "hvXpycFWr7", "gd9UJZ9sSG", "dkiJAZ2i7w", "dGEiE6Dq6Z", "buCnLkbyI0", "Y5f1JMI4ck", "XcYAGt7Oq6", "V9MD3FZkjC", "UBsCNpw2WW", "T7whAsVzqe", "RgbIscA0Ig", "PTUPYnMtvY", "OFgYG1eZWo", "IPROY0ozJL", "EF7wJjnqOv", "8txzOTnrlV", "5Ghe0ctCmm", "4eCGvY1up2", "4Y3CROAHbG" ], "note_type": [ "official_comment", "official_comment", "official_review", "official_review", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment" ], "note_created": [ 1732545066401, 1732340482093, 1729845384778, 1730195211727, 1732339510113, 1733192584249, 1732333630858, 1729270980402, 1730721908591, 1732207738181, 1732170383445, 1732171057395, 1732554794892, 1732545466418, 1732169912950, 1732178888169, 1732170259959, 1732170698351, 1733193005953, 1732291418287, 1732218642382, 1732340212875, 1732218503082, 1732170857058, 1733968036299, 1732288834848, 1732556595525, 1732338931844, 1732171524073, 1732296814979, 1732170494778, 1732170096229, 1732171289414 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission7979/Authors" ], [ "ICLR.cc/2025/Conference/Submission7979/Authors" ], [ "ICLR.cc/2025/Conference/Submission7979/Reviewer_zCNu" ], [ "ICLR.cc/2025/Conference/Submission7979/Reviewer_16HR" ], [ "ICLR.cc/2025/Conference/Submission7979/Authors" ], [ "ICLR.cc/2025/Conference/Submission7979/Authors" ], [ "ICLR.cc/2025/Conference/Submission7979/Reviewer_16HR" ], [ "ICLR.cc/2025/Conference/Submission7979/Reviewer_w3um" ], [ "ICLR.cc/2025/Conference/Submission7979/Reviewer_nQhb" ], [ "ICLR.cc/2025/Conference/Submission7979/Reviewer_w3um" ], [ "ICLR.cc/2025/Conference/Submission7979/Authors" ], [ "ICLR.cc/2025/Conference/Submission7979/Authors" ], [ "ICLR.cc/2025/Conference/Submission7979/Reviewer_w3um" ], [ "ICLR.cc/2025/Conference/Submission7979/Authors" ], [ "ICLR.cc/2025/Conference/Submission7979/Authors" ], [ "ICLR.cc/2025/Conference/Submission7979/Reviewer_zCNu" ], [ "ICLR.cc/2025/Conference/Submission7979/Authors" ], [ "ICLR.cc/2025/Conference/Submission7979/Authors" ], [ "ICLR.cc/2025/Conference/Submission7979/Authors" ], [ "ICLR.cc/2025/Conference/Submission7979/Reviewer_zCNu" ], [ "ICLR.cc/2025/Conference/Submission7979/Authors" ], [ "ICLR.cc/2025/Conference/Submission7979/Authors" ], [ "ICLR.cc/2025/Conference/Submission7979/Authors" ], [ "ICLR.cc/2025/Conference/Submission7979/Authors" ], [ "ICLR.cc/2025/Conference/Submission7979/Authors" ], [ "ICLR.cc/2025/Conference/Submission7979/Reviewer_w3um" ], [ "ICLR.cc/2025/Conference/Submission7979/Reviewer_zCNu" ], [ "ICLR.cc/2025/Conference/Submission7979/Authors" ], [ "ICLR.cc/2025/Conference/Submission7979/Authors" ], [ "ICLR.cc/2025/Conference/Submission7979/Reviewer_w3um" ], [ "ICLR.cc/2025/Conference/Submission7979/Authors" ], [ "ICLR.cc/2025/Conference/Submission7979/Authors" ], [ "ICLR.cc/2025/Conference/Submission7979/Authors" ] ], "structured_content_str": [ "{\"comment\": \"Dear Reviewer nQhb,\\n\\nThank you for your insightful and constructive comments. As the rebuttal period nears its end, we are still awaiting your reply and hope that our responses have adequately addressed your concerns.\\n\\nBest,\\n\\nAuthors\"}", "{\"title\": \"Response to Reviewer w3um (Part 2/2)\", \"comment\": \"***2. This work aligns more with Case 2, representing a straightforward application of an existing technique rather than a substantial technical advancement. Therefore, I believe the current work lacks sufficient novelty for ICLR.***\\n\\nWe respectfully disagree with your point here. Firstly, this paper represents ***the first work*** on defending against jailbreak attacks specifically targeting MLLMs. It is important to note that ***MLLMs and LLMs exhibit significant differences*** when it comes to the threats posed by jailbreak attacks. As described in the paper (lines 89-91), attackers can leverage multiple modalities to inject perturbations into MLLMs, which creates unique challenges that our work aims to address.\\n\\nTo tackle this issue, we are ***the first to introduce new token embeddings ${P_h^0,P_t^0}$ at specific positions in the prompt query***. By injecting perturbations into these new embeddings, we effectively unify adversarial noise across different modalities. We do not extend the LAT method by injecting perturbed noise into image and text tokens to avoid computationally intensive optimization on a large number of token embeddings. Our experiments clearly demonstrate the superiority of this novel design.\\n\\nAdditionally, we ***introduce a new training objective*** that applies to both the attack and defense phases, further enhancing the robustness of our methodology. The effectiveness of this objective is validated through comprehensive experiments. \\n\\n***Thank you once again for your constructive comments. We kindly ask you to reconsider the novelty of our work and adjust your rating accordingly. We believe our work goes beyond a straightforward application of existing techniques. Specifically, the proposed SafeMLLM model demonstrates both greater effectiveness and efficiency compared to intuitive solutions.***\"}", "{\"summary\": \"The authors propose the first adversarial training algorithm for VLMs. They introduce a new loss function based on contrastive learning for attack optimization and model training and demonstrate the effectiveness of the new components in an ablation study. Compared to prev. work in the LLM setting, they show that their algorithm provides higher robustness.\", \"soundness\": \"4\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": [\"Multimodal adversarial training is an underexplored research area and this paper provides numerous insights on some of the design choices that need to be considered in this context.\", \"Robustness is evaluated from multiple perspectives, which makes it easier to assess differences between multi-modal and LLM only robustification approaches.\", \"State-of-the-art attack methods are used for evaluation\", \"Exhaustive hyperparameter ablations are conducted for the presented method.\"], \"weaknesses\": [\"Doesnt the argument in line 90 only apply to the method proposed by Mazeika et al.? I think the framing could be improved.\", \"The two-step algorithm is just standard adversarial training with nontypical loss functions. I think it would be easier to understand if the authors framed their contribution accordingly. Initially, I thought that this 2 stage algorithm would be something different.\", \"Include the dataset that was used for the experiment in Figure 3\", \"\\u201cWe attribute this to the fact that in all MLLMs, the image is always placed before the text as input\\u201d This information would be helpful when you describe the method. To give a better intuition about P_0^h\"], \"questions\": [\"Do I understand it correctly, that no images are used during attack optimization? If yes, did you ablate if encoding a given image and starting the attack from this initialization is helpful?\", \"What are the hyperparameters of the competitor approaches? Are there any ablations regarding the \\u201cbest\\u201d or reasonably \\u201cbest\\u201d possible achievable robustness with the different methods. Hard to compare methods directly to each other otherwise\", \"Could the authors clarify how they measured over-refusal after safety training in their framework?\", \"In summary I believe this to be an interesting contribution to the ICLR and robustness community and would be willing to increase my score if all my concerns and questions are addressed.\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper presents a novel adversarial training method called SAFEMLLM, aimed at strengthening VLM models against jailbreak attacks. Specifically, SAFEMLLM consists of two phases: (1) introducing adversarial embedding tokens, where adversarial tokens are initialized in the token embedding layer and optimized to produce adversarial responses; (2) using the adversarial noise from Phase 1 and additional defense data to fine-tune the model for robustness. The authors validate the effectiveness of the proposed defense method across six mainstream VLM model architectures.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"3\", \"strengths\": [\"This work addresses the security defense of multimodal VLMs, which is crucial for enhancing the secure deployment of VLMs.\", \"Comprehensive experiments conducted on multiple model architectures and datasets demonstrate the effectiveness of the proposed method.\"], \"weaknesses\": [\"The definitions of formulas are unclear. For example, what do $P_h$ and $P_t$ represent in lines 240-244? Why initialize two tokens instead of one?\", \"Why does the CoE-attack optimize at the embedding token level? Why not at the representation level?\", \"How does $\\\\text{equ.2}$ guarantee output quality while ensuring the model\\u2019s replies are relevant to the query?\", \"There is formula redundancy. Both equ.3 and equ.4} describe the same optimization objective, yet the authors define it twice as $loss_{adv}$ in equ.3 and as $loss_{def}$ in equ.4. This complexity should be reduced; maintaining clear formula definitions is essential to avoid redundancy.\", \"The results in Table 3 are unclear. The authors state that results were tested on ImgJP (image) and AdvBench (text) datasets, yet the table does not present the experimental results for these two modalities separately. Additionally, what is the distinction between $J_{target}/J_{contra}$ and $L_{target}/J_{contra}$? The authors should avoid using multiple ambiguous definitions, as this complicates understanding the experimental results.\", \"What are the specific settings for model output quality in Figure 3? The authors tested only 100 samples, which may not effectively reflect changes in model output capabilities. It is recommended that the authors test on 500 benign samples to assess the effects accurately.\", \"How can the model avoid generating garbled outputs? In lines 511-513, the authors indicate that using $L_{contra}$ and $J_{contra}$ to fine-tune the model may lead to meaningless garbled outputs (e.g., repeating \\\"safe\\\"). I request the authors to further explain the reasons for this phenomenon and how to prevent garbled outputs during optimization. This is crucial for assessing the effectiveness of the optimization steps in their method. Additionally, could the authors provide examples of model outputs in both successful and failed optimization scenarios?\", \"Overall, I acknowledge the proposed adversarial training method for enhancing the robustness of VLMs against jailbreak attacks. However, there are still issues regarding unclear loss definitions and ambiguous optimization details in the methods section. On the other hand, in the experimental section, the authors need to further elucidate the roles of the different loss components. Moreover, an explanation and experimental results on how to prevent the model from generating meaningless outputs (such as repetitive and low-quality responses) during adversarial training are also required. If the authors can address the aforementioned concerns, I would like to increase my score.\"], \"questions\": \"see weakness.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Thank you for Discussing with Reviewer w3um\", \"comment\": \"We thank you once again for supporting our work and acknowledging the novelty and significance of our contributions.\"}", "{\"comment\": \"Dear reviewer zCNu:\\n\\nWe regret that your perspective has shifted during the rebuttal period. Nevertheless, we would like to reiterate our contributions and highlight the advantages of our work over existing methods from the following three perspectives:\\n\\n***Pioneering Work on Defending Against Jailbreak Attacks on MLLMs***\\\\\\nThis paper is the first to address the specific challenge of defending Multimodal Large Language Models (MLLMs) against jailbreak attacks. It is crucial to recognize that MLLMs and LLMs face fundamentally different threats from these attacks. As outlined in the paper, attackers can exploit multiple modalities to inject perturbations into MLLMs, creating unique challenges that our work seeks to address.\\n\\n***Innovative Approach to Unified Adversarial Noise Injection***\\\\\\nTo tackle these challenges, we propose introducing new token embeddings, $(P_h^0, P_t^0)$, at specific positions in the prompt query. By injecting perturbations into these embeddings, we unify adversarial noise across modalities. Importantly, this approach avoids the computational overhead associated with injecting perturbed noise directly into image and text tokens, as seen in LAT. This design results in a solution that is significantly more efficient and effective. Our experiments demonstrate that SafeMLLM not only outperforms LAT across various dimensions but is also five times faster in computational efficiency. While we could not test all MLLMs and attack methods during the rebuttal period due to time constraints, we believe these results will generalize, particularly for cases where SafeMLLM shows substantial improvements in ASR performance.\\n\\n***Distinct Training Objective for Robust Defense***\\\\\\nWe introduce a completely new training objective applicable to both the attack and defense phases, enhancing the robustness of our methodology. This novel objective has been validated through comprehensive experiments, further distinguishing our work from existing approaches.\\n\\nWe strongly disagree with reviewer ***w3um\\u2019s*** characterization of our work as an extension of LAT. As demonstrated through rigorous experiments, our proposed SafeMLLM is a more refined and effective solution for defending against jailbreak attacks on MLLMs.\"}", "{\"comment\": \"Thanks for the response, which addressed most of the concerns. After reviewing the revised paper and considering the feedback from other reviewers, I have decided to increase my score to 6.\"}", "{\"summary\": \"This paper introduces SAFEMLLM, a novel adversarial tuning framework to defend multimodal large language models (MLLMs) against jailbreak attacks. The framework operates in two stages: 1) generating adversarial perturbations using a contrastive embedding attack (CoE-Attack), and 2) updating model parameters to neutralize perturbation effects while preserving model utility. The authors evaluate SAFEMLLM across six MLLMs and six jailbreak methods spanning multiple modalities, demonstrating its effectiveness in defending against diverse attacks without compromising normal interactions.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"1\", \"strengths\": [\"Comprehensive experiments are conducted across multiple models and jailbreak attack methods, demonstrating the effectiveness of this framework.\", \"The idea is easy to understand and the paper is easy to follow.\"], \"weaknesses\": \"- This paper lacks novelty. Though it's the first paper to perform adversarial tuning on MLLMs, the attack method is only conducted on the embedding layer and there is no specific modification to the components directly related to VLLM such as the image encoder. Previous work[1, 2 ] on LLMs has already demonstrated the effectiveness of adversarial training in the latent space. Meanwhile, previous work also propose to first perform attacks on multiple layers of LLM (including the embedding layer) and then fine-tune the model against such attacks. Since this method only performs attacks on the embedding layer, it's not clear why previous work cannot be directly applied to this task.\\n- More recent attack methods could be added, such as GPTFuzzer [3] and PAIR [4].\\n\\n\\n[1] Sheshadri, Abhay, et al. \\\"Targeted latent adversarial training improves robustness to persistent harmful behaviors in llms.\\\" arXiv preprint arXiv:2407.15549 (2024).\\n\\n[2] Casper, Stephen, et al. \\\"Defending Against Unforeseen Failure Modes with Latent Adversarial Training.\\\" arXiv preprint arXiv:2403.05030 (2024).\\n\\n[3] Jiahao Yu, Xingwei Lin, and Xinyu Xing. 2023. Gpt-fuzzer: Red teaming large language models with auto-generated jailbreak prompts. arXiv preprint arXiv:2309.10253\\n\\n[4] Patrick Chao, Alexander Robey, Edgar Dobriban, Hamed Hassani, George J Pappas, and Eric Wong. 2023. Jailbreaking black box large language models in twenty queries. arXiv preprint arXiv:2310.08419.\", \"questions\": \"- How is this method different from other latent adversarial training methods [1] [2] ?\\n- Could this training approach be adapted to improve model safety against other types of attacks?\\n\\n\\n\\n[1] Sheshadri, Abhay, et al. \\\"Targeted latent adversarial training improves robustness to persistent harmful behaviors in llms.\\\" arXiv preprint arXiv:2407.15549 (2024).\\n\\n[2] Casper, Stephen, et al. \\\"Defending Against Unforeseen Failure Modes with Latent Adversarial Training.\\\" arXiv preprint arXiv:2403.05030 (2024).\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper presents a novel defense framework, SAFEMLLM, that employs the CoE-Attack strategy to craft adversarial embeddings and iteratively refines model parameters, ensuring resilience against attacks while preserving benign input performance. Comprehensive experiments across six multimodal language models and six jailbreak attack techniques validate SAFEMLLM's effectiveness, especially in challenging white-box scenarios.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. This paper proposes a novel adversarial training framework, SAFEMLLM, and demonstrates its effectiveness across diverse MLLMs and jailbreak methods in white-box scenarios, showcasing robust defense without sacrificing user interaction\\n2. This paper excels in its writing style, which is fluid, clear, and easy to read.\\n3. This paper stands out with thorough experimental comparisons across three jailbreak attack scenarios: image-based, text-based, and image-text jailbreak attacks.\", \"weaknesses\": \"1. The Introduction could be strengthened by more clearly emphasizing the article's unique contribution, originality, and effectiveness. Currently, these elements are not sufficiently highlighted, which might undermine the article's impact.\\n2. The proposed framework has limitations as it only considers image and text data, excluding audio, video, and other multimedia types.\\n3. Authors fail to provide experimental results demonstrating that their framework can reduce overall computing resources.\", \"questions\": \"1. In Section 4, four datasets were chosen for robustness evaluation. Please explain why the paper selected there four datasets? Please provide criteria and rationale for their authority and suitability. For the LLaVA-Instruct-80K dataset, specifically, why was it chosen to assess the utility of the fine-tuned models? A brief summary of each dataset's relevance in the main text would also enhance readability.\\n2. The introduction part presents previous work, such as VLGuard, which is effective against black-box attacks but fails to defend against white-box attacks. Please provide further explanations on why this occurs. Specifically, how does an attacker's possession of parameters and gradient information enable them to launch more effective attacks in white-box scenarios, and what are the key points of defense in such cases.\\n3. Figure 2 presents an overview of the proposed SAFEMLLM framework, with a focus on the first phase, CoE-Attack. In this phase, the paper emphasizes the optimization of two noise metrics: noisy image and noisy text. Are these the only types of noise considered in the SAFEMLLM, or does the framework have the potential to incorporate other modalities such as video and audio? If not, it it better to discuss the limitations of the current approach in terms of its applicability to a broader range of multimodal data?\\n4. Why are there many missing values for the VLGurad metric in the performance indicators listed in Table 1?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Official Comment by Reviewer w3um\", \"comment\": \"First, thank you very much for the clarification on the other questions.\\n\\nSecond, I agree with the intuition of the paper that latent adversarial training is a natural approach for enhancing MLLM safety since attackers on MLLMs can directly inject unrestricted pixel-level perturbations into images to compromise safety alignment. However this still doesn't answer my question below:\\n\\n**Meanwhile, previous work also propose to first perform attacks on multiple layers of LLM (including the embedding layer) and then fine-tune the model against such attacks. Since this method only performs attacks on the embedding layer, it's not clear why previous work cannot be directly applied to this task.**\\n\\nYour follow-up experiment focuses on attack on the embedding layer with untargeted noise, however since images are treated as input embeddings after being extracted with vision-extractor and projectors, they should function similarly to the latent of other text tokens in later layers. Under this circumstance, attacking multiple layers of LLM on the overall latent composed of both image latents and text latents could potentially yield better performance. In paper[1], they have demonstrated that perturbing the residual stream at multiple layers rather than a single layer, each with its own \\u03b5 constraint, typically yields better results. Could you please clarify this issue and conduct additional experiments based on targeted adversarial training with multiple layers as described in paper [1]?\\n\\nMeanwhile, there is also an incorrect citation in your response: **The first LAT method [1] sets up the adversarial training in an untargeted manner. The second LAT method [2] improves this into a targeted attack objective.** \\n\\n[1] Sheshadri, Abhay, et al. \\\"Targeted latent adversarial training improves robustness to persistent harmful behaviors in llms.\\\" arXiv preprint arXiv:2407.15549 (2024).\\n\\n[2] Casper, Stephen, et al. \\\"Defending Against Unforeseen Failure Modes with Latent Adversarial Training.\\\" arXiv preprint arXiv:2403.05030 (2024).\"}", "{\"title\": \"Rebuttal by Authors (Part 2/4)\", \"comment\": \"`>>> Weakness 4`: ***There is formula redundancy. Both equ.3 and equ.4} describe the same optimization objective, yet the authors define it twice as loss_adv in Eq.3 and as loss_def in Eq.4. This complexity should be reduced; maintaining clear formula definitions is essential to avoid redundancy.***\\\\\\n`>>> Response`: We want to emphasize that although $L_{adv}$ (Eq. 3) and $L_{def}$ (Eq. 4) share similar expressions, ***they have different optimization targets and variables.*** Specifically, we adopt $L_{adv}$ to optimize the perturbation matrices {$P_0^h, P_0^t$} by increasing the probability of sampling the affirmative response $c_n$\\u200b and decreasing the probability of sampling the refusal response $r_n\\u200b$. Conversely, $L_{def}$ is used to update model parameters by increasing the probability of sampling $r_n\\u200b$\\u200b and decreasing the probability of sampling $c_n$\\u200b\\u200b. Therefore, we adopt different notations here.\\n\\n`>>> Weakness 5`: ***The results in Table 2 are unclear. The authors state that results were tested on ImgJP (image) and AdvBench (text) datasets, yet the table does not present the experimental results for these two modalities separately. Additionally, what is the distinction between $J_{target}$/$J_{contra}$ and $L_{target}$/$L_{contra}$? The authors should avoid using multiple ambiguous definitions, as this complicates understanding the experimental results.***\\\\\\n`>>> Response`: Thanks for pointing this out, and we are sorry for the vague clarification in the caption of Table 2. The robustness experiments in Table 2 are developed using the ImgJP attack method. We follow the original setup in ImgJP to conduct experiments on the AdvBench datasets, where ImgJP optimizes an adversarial image to make MLLM output affirmative responses for queries in the AdvBench dataset. We have rewritten this part in the caption of Table 2 for clear understanding. \\n\\nIn the original manuscript, $L_{target}$ is the target loss defined in Eq.1, and $L_{contra}$ is the contrastive loss defined in Eq.2 . They are used as the attack objective for updating adversarial perturbations. In the model updating step, we originally adopted $J_{target}$ and $J_{contra}$ to represent the target loss and contrastive loss for optimizing the model parameters, and they are defined in Eq.4. We apologize for any confusion in the notation definitions. In the revised version, we use $L_{adv}^{target}$ and $L_{adv}^{contra}$ to denote the target and contrastive loss during the attack step, respectively. Similarly, We adopt $L_{def}^{target}$ and $L_{def}^{contra}$ to denote the target and contrastive loss during the model updating step, respectively. We have also rewritten the description in the caption of Table 2 (page 9, lines 443-451) for better understanding.\"}", "{\"title\": \"Rebuttal by Authors (Part 2/2)\", \"comment\": \"`>>> Question 2`: ***What are the hyperparameters of the competitor approaches? Are there any ablations regarding the \\u201cbest\\u201d or reasonably \\u201cbest\\u201d possible achievable robustness with the different methods. Hard to compare methods directly to each other otherwise***\\\\\\n`>>> Response`: We compare three different methods in our experiment, including VLGuard [1]\\uff0cR2D2 [2] and CAT [3]. For VLGuard, we do not use any hyperparameters and directly adopt the models officially trained and released by authors [1]. For LLM-based adversarial training methods R2D2 and CAT, we adapt them to our problem by only fine-tuning the LLM decoder from different MLLMs, and we adopt the same hyperparameters as specified in their original implementations. We believe that directly using their original hyperparameters will not significantly impact the ASR values, as the original LLMs targeted in their experiments share the same architecture and model size as the LLM decoders used in our task. \\n\\nWe also analyze the effect of using different hyperparameters on CAT, using ***attack iterations*** as an example. The experiments are conducted on LLaVA-7B using ImgJP, and the results are shown below. The CAT method originally performed ***10*** attack iterations. \\n\\n|Iterations| 5| ***10***| 20|SafeMLLM| \\n |-| :-: | :-: |:-:|:-:|\\n|ASR| 15.00 | ***9.00*** |10.00 |6.00|\\n\\nAs illustrated in the table, adopting different ***attack iterations*** for the baseline method CAT does not improve the ASR performance in our problem setting. Also, considering that performing ablations on different hyperparameters for each model and attack setting would be highly resource-consuming, we thus followed the default settings as outlined in their original paper.\\n\\n[1] Yongshuo Zong, Ondrej Bohdal, Tingyang Yu, Yongxin Yang, and Timothy M. Hospedales. Safety fine-tuning at (almost) no cost: A baseline for vision large language models. In ICML. OpenReview.net, 2024 \\\\\\n[2] Mantas Mazeika, Long Phan, Xuwang Yin, Andy Zou, Zifan Wang, Norman Mu, Elham Sakhaee, Nathaniel Li, Steven Basart, Bo Li, David A. Forsyth, and Dan Hendrycks. Harmbench: A standardized evaluation framework for automated red teaming and robust refusal. In ICML. OpenReview.net, 2024. \\\\\\n[3] Sophie Xhonneux, Alessandro Sordoni, Stephan G\\u00a8unnemann, Gauthier Gidel, and Leo Schwinn. Efficient adversarial training in llms with continuous attacks. arXiv preprint arXiv:2405.15589, 2024\\n\\n`>>> Question 3`: ***Could the authors clarify how they measured over-refusal after safety training in their framework?***\\\\\\n`>>> Response`: We measure the over-refusal via the utility evaluation as described in Section 5.2, where we adopt 100 benign image-text questions from LLaVA-Instruct-80K and follow LLaVA to use GPT-4 score each model\\u2019s responses. The results in Figure 3 show that our method can retain utility for these benign questions. A refused response to any benign question will receive a very low score, and we have put some results and their GPT scores in Table 7 on page 20. Note that all of these samples are extracted from the results when omitting the utility loss in SafeMLLM, which we have also discussed in the qualitative analysis ( lines 1022-1025, page 19).\\n \\n To further evaluate the utility of SafeMLLM, we also add an experiment on a widely-used MLLM evaluation benchmark-MM-Vet [1]. The benchmark contains 217 multimodal questions and adopts gpt-4-turbo to evaluate the responses from the following dimensions: Recognize (Rec), OCR, Knowledge (Know), Language Generation (Gen), Spatial awareness (Spat), and Math. The results on LLaVA-7B and LLaVA-13B are reported in the table below:\\n\\n| LLaVA-7B| rec | ocr | know | gen | spat | math | total |\\n|-------------|------|------|------|------|------|------|-------|\\n| Original | 36.9 | 24.0 | 18.5 | 20.5 | 28.0 | 3.8 | 32.2 |\\n| VLGuard| 33.9 | 22.9 | 13.8 | 14.2 | 27.2 | 3.8 | 30.1 |\\n| R2D2| 34.7 | 21.5 | 16.4 | 18.1 | 24.3 | 7.7 | 30.2 |\\n| CAT| 37.7 | 20.1 | 24.3 | 25.1 | 25.7 | 3.8 | 31.5 |\\n| SafeMLLM| 37.5 | 24.1 | 20.5 | 21.1 | 28.3 | 3.8 | 32.5 |\\n\\n| LLaVA-13B| rec | ocr | know | gen | spat | math | total |\\n|-------------|------|------|------|------|------|------|-------|\\n| ori | 42.1 | 25.9 | 24.4 | 25.1 | 30.4 | 11.2 | 36.0 |\\n| vlguard | 37.7 | 26.6 | 17.7 | 21.4 | 30.9 | 3.8 | 32.9 |\\n| R2D2 | 41.1 | 26.2 | 24.4 | 26.1 | 32.0 | 7.7 | 35.4 |\\n| RTEAT | 42.7 | 27.7 | 26.7 | 26.1 | 32.7 | 15.0 | 36.9 |\\n| safemlllm | 44.0 | 27.1 | 23.8 | 25.6 | 34.0 | 15.0 | 37.8 |\\n\\nFor each metric, higher values indicate better performance. We observe that SafeMLLM maintains response quality across all aspects, further demonstrating that our method does not compromise the overall capabilities of the target MLLM. We add this experiment in Appendix G (page 18, lines 949-956).\\n\\n[1] Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., ... & Wang, L. MM-Vet: Evaluating Large Multimodal Models for Integrated Capabilities.ICML 2024.\"}", "{\"title\": \"Official Comment by Reviewer w3um\", \"comment\": [\"Thank you for providing additional experimental results. After careful consideration, I maintain my original review and rating due to limited novelty in the work. I will explain my concerns from two main aspects:\", \"Performance Analysis\", \"The performance improvement is modest, showing only a 3% increase in Attack Success Rate (ASR) compared to the previous LAT method [1]. This is particularly noteworthy given that the previous LAT approach [1] already outperforms all other baselines on LLaVA demonstrated in your original paper.\", \"Regarding efficiency, while efforts to reduce computational costs in adversarial training have been valuable in traditional computer vision [2,3], I question its relevance as a performance metric for LLM adversarial training. This is reflected in your original paper, where efficiency comparisons were relegated to the Appendix with limited experiment.\", \"Though you suggest that achieving the 3% improvement requires complex ablation experiments like attacking subset image tokens or selecting target layers. However, since the previous method was designed for traditional LLMs without extensive hyperparameter optimization for follow-up experiment, similar improvements might be achievable through basic parameter tuning (e.g., learning rate adjustments). This is especially relevant given that your method's performance also depends significantly on the hyperparameter \\u03bb shown in Appendix I.\", \"Inituion\", \"As I have mentioned in my previous review, **the LLaVA architecture only treats images differently during the initial feature extraction and projection stages. After that, image embeddings are concatenated with text embeddings in the embedding layer and processed identically.** Since the backbone LLM processes image and text embeddings uniformly and there is no specific optimization exists for image embeddings, I fail to understand how the image embeddings differ fundamentally from the latent representations defined in [1], which raises questions about the novelty of adapting LAT from LLMs to VLLMs.\", \"The small **3%** performance gap between your method and previous LAT work [1] (which wasn't specifically designed for image attacks) supports this view. As noted in my previous review, **the existing method [1] with minor modifications to attack indices selection** could likely achieve comparable performance.\", \"Consequently, I will maintain my score for this paper. Thanks again for your thorough experimental results.\", \"[1] Sheshadri, Abhay, et al. Targeted latent adversarial training improves robustness to persistent harmful behaviors in llms. arXiv 2024.\", \"[2] Shafahi, Ali, et al. \\\"Adversarial training for free!.\\\" Advances in neural information processing systems 32 (2019).\", \"[3] Wong, Eric, Leslie Rice, and J. Zico Kolter. \\\"Fast is better than free: Revisiting adversarial training.\\\" arXiv preprint arXiv:2001.03994 (2020).\"]}", "{\"comment\": \"Dear Reviewer w3um,\\n\\nThank you for the multi-round discussions. We have conducted additional experiments to further demonstrate the effectiveness of the proposed SafeMLLM and highlighted its superior efficiency compared to the method you suggested. We would like to emphasize that our approach introduces a novel methodology rather than a straightforward application of existing techniques. Details can be found in the above replies.\\n\\nAs the rebuttal period nears its end, we are still awaiting your response and hope that our clarifications have sufficiently addressed your concerns regarding the novelty of our work.\\n\\nBest,\\n\\nAuthors\"}", "{\"title\": \"Rebuttal by Authors (Part 1/2)\", \"comment\": \"Thank you for your thoughtful review. We have provided our responses below.\\\\\\n`>>> Weakness 1`: ***The Introduction could be strengthened by more clearly emphasizing the article's unique contribution, originality, and effectiveness. Currently, these elements are not sufficiently highlighted, which might undermine the article's impact.***\\\\\\n`>>> Response`: Thanks for your suggestion. We added a paragraph (page 3, lines 125-133) to summarize this paper's contributions at the end of the introduction section, including an overview of the research problem, the proposed methodology, and the experimental results. We hope this highlights the unique contributions and effectiveness of our work. \\n\\n`>>> Question 1`: ***In Section 4, four datasets were chosen for robustness evaluation. Please explain why the paper selected there four datasets? Please provide criteria and rationale for their authority and suitability. For the LLaVA-Instruct-80K dataset, specifically, why was it chosen to assess the utility of the fine-tuned models? A brief summary of each dataset's relevance in the main text would also enhance readability.***\\\\\\n`>>> Response`: In our experiment, we evaluated six attack methods on four jailbreak datasets. We ***follow the original paper*** to use the same implementations and dataset for robustness evaluation for each attack method, ensuring that the hyperparameters used in the attack setup are optimal. For the utility evaluation, we follow LLaVA [1] and adopt the LLaVA-Instruct-80K dataset, using gpt-4-turbo to evaluate the model's responses. We chose this dataset because it contains diverse and complex instruction-following multimodal VQA samples to assess the MLLM's comprehension and generation abilities across various scenarios.\\n\\nWe apologize for the lack of clarity regarding dataset usage in Section 4, and we have added a brief summary of each dataset\\u2019s relevance in the main text (page 7, lines 332-339).\\n\\n[1] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023a\\n\\n`>>> Question 2`: ***The introduction part presents previous work, such as VLGuard, which is effective against black-box attacks but fails to defend against white-box attacks. Please provide further explanations on why this occurs. Specifically, how does an attacker's possession of parameters and gradient information enable them to launch more effective attacks in white-box scenarios, and what are the key points of defense in such cases.***\\\\\\n`>>> Response`: VLGuard enhances model safety by performing supervised fine-tuning on a safety instruction-following dataset. The dataset contains diverse harmful information presented through images, text, or multimodal inputs, with refusal responses as labels. In ***black-box*** scenarios, attackers cannot access the model's parameters and directly place harmful queries in the image or text inputs, such as injecting malicious text requests into the input image[1,2]. Therefore, safety-trained MLLMs like VLGuard have already used such data during training, enabling the models to output safe responses.\\n\\nHowever, in ***white-box*** scenarios, attackers can compromise safety alignment by injecting adversarial noise, where the \\\"adversarial noise\\\" refers to trainable parameters applied to the inputs, such as adjusting pixel values. To create a more effective attack, attackers can optimize these parameters with an objective that enforces the model to output an affirmative response, such as, \\\"Sure, here are the steps to create a bomb.\\\" Since MLLMs autoregressively generate subsequent content based on previous outputs, they are more likely to continue in an affirmative tone and produce detailed harmful content. Therefore, compared to manually crafted prompts in black-box attacks, white-box attacks can leverage these adversarial perturbations to explicitly shift the model's output space, resulting in more effective attacks.\\n\\nSafety-trained MLLMs, such as VLGuard, are no longer effective in white-box scenarios, as these adversarial perturbations are not included in the training data of these methods. As a result, the key to defending against white-box jailbreak attacks on MLLMs is to reduce the risk of attackers identifying such adversarial noise that can alter the model's safety outputs.\\n\\n[1] Xin Liu, Yichen Zhu, Jindong Gu, Yunshi Lan, Chao Yang, and Yu Qiao. Mm-safetybench:\\nA benchmark for safety evaluation of multimodal large language models. arXiv preprint\", \"arxiv\": \"2311.17600, 2023b.\\n\\n[2] Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Figstep: Jailbreaking large vision-language models via typographic visual prompts. arXiv preprint arXiv:2311.05608, 2023\"}", "{\"title\": \"Thank you for the response\", \"comment\": \"I want to thank the authors for their extensive response and the additional experiments they conducted. My concerns were sufficiently addressed.\\n \\nI will follow the discussions with the other reviewers and raise my score depending on the outcome. Currently, I am leaning toward accepting the paper.\"}", "{\"title\": \"Rebuttal by Authors (Part 1/4)\", \"comment\": \"Thank you for the valuable review. We have provided our responses below. We hope these responses can address your concerns.\\\\\\n`>>> Weakness 1`: ***The definitions of formulas are unclear. For example, what do $\\\\mathbf{P}_0^h$ and $\\\\mathbf{P}_0^h$ represent in lines 240-244? Why initialize two tokens instead of one?***\\\\\\n`>>>Response`: We apologize for the vague clarification here. $\\\\mathbf{P}_0^h$ and $\\\\mathbf{P}_0^t$ are two matrices initialized from word token embeddings. Each matrix has a shape of $K\\\\times C$, where $K$ is the number of tokens and $C$ is the embedding dimension. During the attack optimization, we position the first embedding matrix $\\\\mathbf{P}_0^h$ before the text query to act as the ***adversarial image*** $\\\\mathbf{I}^{\\\\prime}$. This design is based on the fact that in all MLLMs, the image is always placed before the text as input. Similarly, another embedding matrix $\\\\mathbf{P}_0^t$ is positioned after the text query to act as the ***adversarial string suffix***. Therefore, we need two token embedding matrices here. We have rewritten this part in our revised manuscript, please refer to lines 247-251 on page 5 for more details.\\n\\n`>>> Weakness 2`:***Why does the CoE-attack optimize at the embedding token level? Why not at the representation level?***\\\\\\n`>>> Response`: As described in lines 247\\u2013251 on page 5, we optimize the perturbation based on two token embeddings: one is placed before the toxic query, and the other is placed after the query. This is a heuristic design as attackers always inject adversarial perturbations from the input level, such as placing an adversarial image in front of the prompt or using an adversarial string suffix. Thus, optimizing perturbations at the token embedding level can unify attacks from both modalities.\\n\\nWe also notice that some existing works inject perturbations into latent representations of Large Language Models (LLMs) [1,2]. Specifically, when a toxic prompt is used as a training sample, these latent adversarial training (LAT) methods generate noise by adding perturbations to the intermediate token representations corresponding to this text prompt. This design is based on the hypothesis that injecting perturbations into the intermediate text token features is equivalent to, and even stronger than, adding extra non-word adversarial text tokens after the query in terms of attack intensity. However, this hypothesis may ***no longer stand for MLLM attacks as attackers not only inject perturbations through discrete text tokens but can also introduce noise via images with continuous values.*** More importantly, in MLLMs, the number of tokens for images largely exceeds that of text prompt tokens (e.g., 576 tokens on LLaVA-13B), making these latent adversarial training methods ***less effective***.\\n\\nWe conducted an extra experiment to validate our intuition. Specifically, we follow the setting in the latest LAT method [1], and inject perturbations into latent token features with $\\\\epsilon=20$ in the L2 constraint. For a fair comparison, we adopted ***the same optimization target*** in SafeMLLM and also increased $\\\\epsilon$ to 40 to explore scenarios under stronger attacks. The modified method is named as SafeMLLM-LAT. The results of using the ImgJP attack method on LLaVA-13B are shown in the table below:\\n\\n|| SafeMLLM-LAT($\\\\epsilon=20$)| SafeMLLM-LAT($\\\\epsilon=40$)| SafeMLLM| \\n |-| :-: | :-: |:-:|\\n|ASR| 13.00 | 12.00 |0.00 |\\n\\nAs shown in the above table, adding perturbations to the latent representations performs worse than directly using the adversarial token embeddings, which validates our hypothesis. \\n\\n[1] Sheshadri, Abhay, et al. \\\"Targeted latent adversarial training improves robustness to persistent harmful behaviors in llms.\\\" arXiv preprint arXiv:2407.15549 (2024).\\n\\n[2] Casper, Stephen, et al. \\\"Defending Against Unforeseen Failure Modes with Latent Adversarial Training.\\\" arXiv preprint arXiv:2403.05030 (2024).\\n\\n`>>> Weakness 3`: ***How does the contrastive loss in Eq.2 guarantee output quality while ensuring the model\\u2019s replies are relevant to the query?***\\\\\\n`>>> Response`: The contrastive loss proposed here is designed to suppress the log probability of the model producing unexpected responses, thereby enhancing the effectiveness of the defense. Only using the constrastive loss during adversarial training cannot ensure that the outputs after attacks or model updates remain relevant to the query, and we put more detailed discussions in response to Weakness 7.\"}", "{\"title\": \"Rebuttal by Authors (Part 4/4)\", \"comment\": \"`>>> Weakness 7`: ***How can the model avoid generating garbled outputs? In lines 511-513, the authors indicate that using L_contra and J_contra to fine-tune the model may lead to meaningless garbled outputs (e.g., repeating \\\"safe\\\"). I request the authors to further explain the reasons for this phenomenon and how to prevent garbled outputs during optimization. This is crucial for assessing the effectiveness of the optimization steps in their method. Additionally, could the authors provide examples of model outputs in both successful and failed optimization scenarios?***\\\\\\n`>>> Response`: We need to point this out that using the contrastive loss (Eq.2) in SafeMLLM does not output garbled texts during adversarial training. Instead, we find that when only using the contrastive loss as the optimization target in the attack step and model updating step, such as setting $L_{adv}=L_{adv}^{contra}$ and $L_{def}=L_{def}^{contra}$, the model may produce meaningless texts based on the toxic training samples with optimized perturbations, even after the model parameters are updated in Step II. This occurs because, ***when only using the contrastive loss in training, the model merely amplifies the probability difference between sampling the refusal response $r_n$\\u200b and the affirmative response $c_n$\\u200b.*** This does not guarantee an increase in the probability of generating $r_n$, the expected output after updating the model. Consequently, the model may produce garbled outputs, thereby reducing the effectiveness of adversarial training.\\n\\nTo further verify this claim, we design an experiment by plotting the average negative log probabilities of generating affirmative label $c_n$ and rejective label $r_n$ during the attack optimization steps at different model training iterations (i=1, 50, 100, 150, 200), as illustrated in Figure 4. From Figures 4(d) and 4(e), we observe that for the method that uses only the contrastive loss as the optimization target (dashed line), the probabilities of the model generating $c_n$ and $r_n$\\u200b are both very low after the training is converged, even though their difference remains large.\\n\\nSafeMLLM addresses this issue by combining the contrastive loss with the target loss in Eq. 3 and Eq. 4. The target loss acts as a supervised term, guiding the model to generate the expected outputs at different steps. This claim can also be verified in Figures 4(d) and 4(e). When combining the target loss in training (solid line), the model can output the refusal response $r_n$ with a noticeably high probability, regardless of the optimized perturbations. This demonstrates that the model can effectively counteract the introduced perturbations and increase its robustness against jailbreak attacks.\\n\\n\\n***Additionally, could the authors provide examples of model outputs in both successful and failed optimization scenarios?***\\n\\nThanks for your advice. We have provided additional qualitative analysis (page 19, lines 1015-1021) in our case study, please refer to Appendix J.\"}", "{\"comment\": \"Dear reviewer w3um:\", \"we_would_like_to_point_out_your_biased_evaluation_regarding_our_contributions\": \"***1. Misunderstanding the proposed SafeMLLM.*** We must stress that we ***DO NOT*** adapt LAT from LLMs to VLLMs. We are puzzled by the continued perception of our work as merely a straightforward adaptation of LAT, despite our repeated clarifications about the unique contributions of our paper. ***These include a completely different training objective and distinct perturbation sets.*** The only similarity between LAT and SafeMLLM lies in their shared use of perturbations injected into the embedding layer\\u2014an approach common across various previous image attack methods that perturb image inputs. Additionally, we have demonstrated that the LAT method is suboptimal in our problem setting, particularly due to its inefficiency. This underscores the necessity of the SafeMLLM framework, which is fundamentally different from LAT, offering superior ASR performance while maintaining computational efficiency.\\n\\n***2. Inconsistencies in your reviews.*** Initially, you stated: \\u201cIf such an approach yields better results, it would suggest that a minor modification to existing methods could outperform their proposed approach.\\u201d However, after we conducted the experiment and demonstrated the superior performance of SafeMLLM, the focus shifted to criticizing the performance improvement as \\u201cmarginal.\\u201d We would like to clarify the reasoning behind this so-called \\u201cmarginal\\u201d improvement. The experiment was conducted on LLaVA-13B using the ImgJP Attack. As shown in Table 1, LLaVA-13B itself exhibits a notable degree of robustness, with the most effective baseline method, CAT [a], achieving an ASR of only 4%. In this context, the 3% performance improvement achieved by SafeMLLM should ***NOT*** be considered marginal. Additionally, due to time constraints during the rebuttal period, we were unable to conduct experiments on all MLLMs and attack methods. Nonetheless, we believe this trend will be corroborated across more models and attack methods, particularly those where SafeMLLM demonstrates substantial improvements over the baseline in ASR performance.\\n\\n***3. Efficiency in LLM adversarial training.*** Efficiency is a critical consideration in LLM adversarial training, and we strongly disagree with your assertion otherwise. Efficiency is a fundamental concern for most machine learning tasks, as it directly impacts the feasibility of implementing algorithms in real-world scenarios. Adversarial training is no exception to this. For instance, the latest LLM adversarial training method, CAT [a], recently accepted at NeurIPS 2024, explicitly focuses on addressing efficiency challenges. Given this context, it is evident that efficiency must be a priority in the adversarial training process, as it significantly affects the practicality and scalability of the approach. Ignoring this aspect undermines the relevance and applicability of the proposed methods.\\n\\n***4. Limited experimental results on efficiency.*** As stated in our methodology section (lines 207\\u2013209), using an entire image during adversarial training can significantly impact computational efficiency. To verify this, we conducted an experiment presented in Appendix H. This observation also applies to the LAT(img+txt) method, which similarly introduces noise to an entire image during adversarial training, leading to the same efficiency concerns. It is important to note that we did not report efficiency metrics in the main paper because none of the baseline methods in Table 1 utilize the image during adversarial training; thus, their computational efficiency is comparable. In conclusion, as discussed both above and in Appendix H, the efficiency issue is an unavoidable challenge when incorporating the entire image into adversarial training. However, SafeMLLM addresses this concern effectively, achieving a significant computational advantage\\u2014it is five times faster than LAT(img+txt).\\n\\nBased on the points outlined above, ***characterizing our work as a straightforward application of an existing technique (e.g., LAT) is neither accurate nor reflective of the methodological innovations and contributions we have introduced.***\\n \\nWhile we fully respect your role in evaluating the contributions of the papers you review, we, as authors, hope our work is assessed fairly and accurately. It is important to recognize that building a researcher\\u2019s reputation is significantly more challenging than rejecting a paper. We ask that our contributions be considered in the broader context of advancing the field, and we trust that the evaluation process upholds a fair and constructive standard.\\n\\n[a] Xhonneux et al. Efficient adversarial training in llms with continuous attacks. NeurIPS 2024.\"}", "{\"title\": \"Comment on the discussion\", \"comment\": \"To give another perspective:\\n\\nReviewer w3um22 remarks that the contribution is not significant. However, in the same regard, [1] is taking the same algorithm as [2] but explores more than just the first latent layers (a trivial contribution?). Moreover, both works explore continuous attacks in the latent space, which have already been explored in computer vision and other domains before LLMs (trivial again?). Long story short, nearly every work in ML can be viewed as \\\"not novel\\\" or \\\"trivial\\\". \\n\\nThe proposed work is, to the best of my knowledge, the first to conduct a study on adversarial training in VLMs with arguably some interesting takeaways for the community. Besides this contribution, the authors evaluate different and novel loss functions. \\n\\nI am not convinced that proposing a more complicated loss / framework would strengthen the contribution.\\n\\n[1] Sophie Xhonneux, Alessandro Sordoni, Stephan G\\u00a8unnemann, Gauthier Gidel, and Leo Schwinn. Efficient adversarial training in llms with continuous attacks. arXiv preprint arXiv:2405.15589, 2024 \\n\\n[2] Sheshadri, Abhay, et al. \\\"Targeted latent adversarial training improves robustness to persistent harmful behaviors in llms.\\\" arXiv preprint arXiv:2407.15549 (2024).\"}", "{\"comment\": \"Thank you for your prompt response. We\\u2019re glad to hear that our replies have adequately addressed your concerns. We also appreciate your positive support for our work and hope you might consider raising your ratings following the rebuttal.\"}", "{\"title\": \"Response to Reviewer w3um (Part 1/2)\", \"comment\": \"***Dear Reviewer w3um:***\\n\\n***We will respond to your question from the following two aspects:***\\n\\n***1. When transferring LAT methods into your problem, would attacking both image and text embeddings across multiple VLLM layers improve performance?***\\\\\\nWe follow your comments and use the LAT method [a] to attack image and text token embeddings across multiple MLLM layers. The jailbreak method is ImgJP attack, the dataset is AdvBench, and the victim MLLM is LLaVA-13B. Here, we still set the intermediate attacking layers as ['embedding', '8', '16', '24', '30'] following [a], and the image token embeddings are obtained from a random image at each training iteration. This method is named as ***LAT(img+txt)***. The results are shown in the table below, and we also compare the average runtime per adversarial training iteration.\\n\\n||runtime (sec)|ASR|\\n |:-:|:-:|:-:|\\n|LAT(img+txt)|192.39 |3.00| \\n|SafeMLLM|38.70 |0.00 | \\n\\n**Performance Comparison**: From the result, we can observe that the ASR of LAT(img+txt) is ***worse*** than that of our proposed SafeMLLM, although it outperforms other baselines, such as R2D2 [b] and CAT [c]. The reason is that in MLLMs, a single image often corresponds to a large number of tokens. For instance, in LLaVA-13B, an image is represented by 576 tokens. Considering that the noise needs to be injected into an excessive number of tokens across multiple MLLM layers, it could increase the difficulty of perturbation optimization, potentially leading to issues such as overfitting on targeted affirmative responses. This, in turn, impacts the corresponding model updates during the defense process. In our experiments, we also observe this phenomenon by extending the token numbers from $P_0^h$ and $P_0^t$ in the hyperparameter sensitivity analysis (page 19, lines 1005-1011).\\n\\nOne potential solution to this issue is to attack only a subset of image tokens on a subset of intermediate layers. However, it is less practical in the LAT setting. This is because it requires determining ***the number of attack tokens***, ***the positions of these tokens in the image***, and ***the specific layers*** to target across various MLLMs, which presents a series of very tricky ablations. In contrast, our proposed SafeMLLM only requires setting the number of tokens from $P_0^h$.\\n\\n***Efficiency Comparison***: More importantly, our proposed SafeMLLM is almost ***five times faster*** than LAT(img+txt) in terms of runtime, even though both methods perform the same number of optimization steps in the attack loop. We also attribute this to the large number of image tokens in MLLMs. During adversarial training, these numerous tokens need to go through multiple forward passes in the attack loop, significantly increasing computational resources. However, SafeMLLM only leverages 8 tokens, thus making it more efficient. As a result, we believe this experiment can demonstrate the effectiveness of our proposed perturbation design. \\n\\n[a] Sheshadri, Abhay, et al. Targeted latent adversarial training improves robustness to persistent harmful behaviors in llms. arXiv 2024.\\n\\n[b] Mazeika et al. Harmbench: A standardized evaluation framework for automated red teaming and robust refusal. ICML 2024. \\n\\n[c] Xhonneux et al. Efficient adversarial training in llms with continuous attacks. arXiv 2024.\"}", "{\"comment\": \"Thanks for your response. We sincerely apologize for the confusion caused by the reversed order of the two references in our rebuttal, where [1] refers to the latest LAT attack method.\\n\\nWe also want to clarify the response to ***Weakness 1 and Question 1***. In our new experiment, We did follow your suggestion and follow [1] to inject perturbations. Precisely, we adhered to their settings by injecting noise into the multiple layers, including five layers: ['embedding', '8', '16', '24', '30']. The results are shown in the table above. Although both SafeMLLM and the LAT method [1] inject perturbations into the embedding layer, we optimize perturbation on additional tokens rather than limiting them to the token embeddings of the toxic query prompt only in [1]. Considering the attacker can leverage a large number of additional tokens introduced by adversarial image noise to execute the attack in MLLMs, our design better aligns with this unique property and can thus perform better.\\n\\nWe hope our clarification can adequately address your concerns regarding the novelty of our work. Your invaluable comments make our contributions clearer and more distinguishable from existing work.\"}", "{\"title\": \"Rebuttal by Authors (Part 1/2)\", \"comment\": \"Thank you for the valuable review. We hope the following responses can address your concerns.\\\\\\n`>>>Weakness 1`: ***Doesnt the argument in line 90 only apply to the method proposed by Mazeika et al.? I think the framing could be improved.***\\\\\\n`>>>Response`: Thank you for noting this. The argument in line 90 is not limited to R2D2 [1] but also applies to other LLM-based adversarial training methods that optimize adversarial noise on the discrete text tokens, such as the method proposed in [2]. We have noticed that the citation for the related work was missing here, and we have corrected the omission and expression in the revised version (page 2, line 91-94).\\n\\n[1] Mantas Mazeika, Long Phan, Xuwang Yin, Andy Zou, Zifan Wang, Norman Mu, Elham Sakhaee, Nathaniel Li, Steven Basart, Bo Li, David A. Forsyth, and Dan Hendrycks. Harmbench: A standardized evaluation framework for automated red teaming and robust refusal. In ICML. OpenReview.net, 2024.\\\\\\n[2] Liu, F., Xu, Z., & Liu, H. (2024). Adversarial tuning: Defending against jailbreak attacks for llms. arXiv preprint arXiv:2406.06622.\\n\\n`>>>Weakness 2`: ***The two-step algorithm is just standard adversarial training with nontypical loss functions. I think it would be easier to understand if the authors framed their contribution accordingly. Initially, I thought that this 2 stage algorithm would be something different.***\\\\\\n`>>>Response`: Thanks for your advice. We have modified the related descriptions in the Abstract section (page 1, lines 18-25), where we originally used \\\"two-stage\\\" to describe SafeMLLM, which could be misunderstood. Also, we have added a paragraph summarizing this paper's contributions at the end of the introduction section (page 3, lines 125-134). We hope this better highlights the unique contributions and effectiveness of our work.\\n\\n`>>> Weakness 3`: ***Include the dataset that was used for the experiment in Figure 3***\\\\\\n`>>> Response`: Thanks for pointing this out. As described in Section 5.2, we adopt 100 samples from the LLaVA-Instruct-80K dataset for the utility experiment in Figure 3. We have added this information in the caption of Figure 3 (page 9, lines 439-442) for clarity and better understanding.\\n\\n`>>> Weakness 4`: ***We attribute this to the fact that in all MLLMs, the image is always placed before the text as input\\u201d This information would be helpful when you describe the method. To give a better intuition about $P_0^h$***\\\\\\n`>>> Response`: Thank you for your valuable suggestions. We have added this explanation in the revised manuscript (page 3, lines 247-251 ). We believe this addition provides better intuition for understanding our method.\\n\\n`>>> Question 1`: ***Do I understand it correctly, that no images are used during attack optimization? If yes, did you ablate if encoding a given image and starting the attack from this initialization is helpful?***\\\\\\n`>>> Response`: Yes, you are correct! No images are used during the attack optimization, and we directly optimize perturbations on the token embeddings {$P_0^h,P_0^t$}. \\n\\nWe also added an ablation experiment to evaluate the effectiveness of attacking a given image instead of embedding the token during the attack. Specifically, we replace the front token embedding $P_0^h$\\u200b with a given image input $\\u200bI_0$ and optimize the perturbations on both $I_0$ and the token embedding $P_0^t$ placed after the query in Step I. In Step II, we update the model based on the optimized perturbations accordingly. We refer to this approach as w/ Adv.Image. The experiments are conducted on the LLaVA-7B and LLaVA-13B models using the ImgJP attack, and the results are presented in the table below. In addition to ASR performance, we also measured the average runtime per iteration and GPU memory usage. For all evaluation metrics, lower values indicate better performance.\\n\\n|LLaVA-7B| runtime (s)\\u2193| GPU Memory (MB)\\u2193| ASR| \\n |-| - | - |-|\\n|w/ Adv.Image|84.42 | 32869 |5.00 |\\n|SafeMLLM|20.73 | 30291 |6.00 |\\n\\n|LLaVA-13B| runtime (s)\\u2193| GPU Memory (MB)\\u2193| ASR| \\n |-| - | - |-|\\n|w/ Adv.Image|263.56 | 66092 |0.00 |\\n|SafeMLLM|38.70 | 57475 |0.00 |\\n\\nAs shown in the table, optimizing image perturbations significantly impacts computational efficiency but does not yield noticeable gains in ASR performance. As a result, we directly attack the token embeddings during the attack optimization. We have added this experiment in Appendix H (page 18, lines 962-971).\"}", "{\"withdrawal_confirmation\": \"I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors.\"}", "{\"title\": \"Official Comment by Reviewer w3um\", \"comment\": \"Thanks for your clarification. This is a friendly reminder that your response to Reviewer 16HR contains the same citation error.\", \"i_remain_unconvinced_about_the_novelty_of_perturbing_both_image_and_text_tokens_for_two_main_reasons\": \"- In [1], the authors only attack token embeddings of toxic query prompts because this aligns with their specific objectives (they chose not to attack the conversation template tokens). Since image tokens are integral to the response generation in your case, extending their method to include image embeddings would be a natural idea. However, your follow-up experiments do not explore this direction.\\n- The LLaVA architecture only treats images differently during the initial feature extraction and projection stages. After that, image embeddings are concatenated with text embeddings in the embedding layer and processed identically. **Therefore, applying the existing LAT method [1] to VLLMs would only require adjusting the attack indices which is a very small modification, as both image and text representations are fundamentally treated the same way**.\\n\\nAdditionally, you haven't addressed a crucial question: Would attacking both image and text embeddings across multiple VLLM layers improve performance? If such an approach yields better results, it would suggest that **a minor modification to existing methods** could outperform your proposed approach.\\n\\n\\n\\n[1] Sheshadri, Abhay, et al. \\\"Targeted latent adversarial training improves robustness to persistent harmful behaviors in llms.\\\" arXiv preprint arXiv:2407.15549 (2024).\"}", "{\"title\": \"Final decision\", \"comment\": \"Sorry for the delay. I have considered the rebuttal of the other reviewers, changes in the paper, and relevant related work.\\nI adjusted my score based on the discussions to slightly below the acceptance threshold. Ultimately, I agree that the advantage compared to existing methods is unclear.\"}", "{\"comment\": \"Thank you very much for raising your rating. We sincerely appreciate your insightful comments, which have been invaluable in helping us enhance the quality of our work.\"}", "{\"title\": \"Rebuttal by Authors (Part 2/2)\", \"comment\": \"`>>> Weakness 2`: ***More recent attack methods could be added, such as GPTFuzzer [3] and PAIR [4]***\\\\\\n`>>> Response`: Thanks for pointing this out. We have added descriptions of GPTFuzzer [3] and PAIR [4] into the related work section (page 3, lines 151-155) of the revised manuscript. We have also conducted an experiment by using PAIR to attack the LLaVA-13B model. For the experiment setup, we follow PAIR [4] to use 50 samples from AdvBench, and set Vicuna-13B-v1.5 as the attacker's LLM. The number of query times is set to 20. The results are illustrated in the following table: \\n\\n||Original|VLGuard|R2D2|CAT|SafeMLLM| \\n |-| :-: | :-: |:-:| :-: |:-:|\\n|ASR|38.00 |20.00 |16.00 | 12.00|0.00 |\\n\\nFrom the table, we can observe that our proposed SafeMLLM still outperforms other baselines, which demonstrates the effectiveness of our proposed method in defending against such jailbreak attacks. \\n\\n`>>> Question 2`: ***Could this training approach be adapted to improve model safety against other types of attacks?***\\\\\\n`>>> Response`: Although our proposed SafeMLLM is targeted for MLLM jailbreak attacks, it can also be adapted to other adversarial attack threats. As mentioned in the existing work [a], there is another adversarial threat for MLLMs, where the attacker optimizes an adversarial image to make the MLLM output a designated text string $\\\\alpha$, regardless of the user\\u2019s input queries. The text string contains a specific phishing link aimed at luring users into a scam (e.g., \\u201cView more details at https://phishinglink!\\u201d).\\n\\nWe adapted SafeMLLM for this specific string attack by replacing the malicious query dataset in Algorithm 1 with a manually curated dataset. To create this dataset, we first sample 100 normal queries from the Alpaca training set [b]. For each normal query, we create $c_n$ by prompting gpt-4-turbo to generate diverse templates with an HTTP prefix (e.g., \\\"Download the details at https:\\\"). We directly use the ground truth answer from the dataset as the negative response $r_n$, and keep the other parts of Algorithm 1 unchanged to train the target MLLM. \\n\\nAfter adversarial training, we follow [a] to use stochastic gradient descent to optimize an adversarial image with the goal of making the MLLM output the string $\\\\alpha$. Note that the string $\\\\alpha$ and phishing link are unknown during the adversarial training. We follow [a] to conduct the optimization on the Alpaca training set with another 520 samples, and we evaluate the ASR on 100 held-out queries from the same dataset.\\n\\nThe results based on LLaVA-13B are shown in the table below. SafeMLLM still significantly outperforms the original model under this adversarial threat, demonstrating our proposed method's overall generalization capability.\\n\\n||Original|SafeMLLM|\\n |-| :-: |:-:|\\n|ASR|88.00 |2.00 \\n\\n\\n[a] Bailey, L., Ong, E., Russell, S., & Emmons, S. Image Hijacks: Adversarial Images can Control Generative Models at Runtime. ICML 2024 \\\\\\n[b] Taori, R., Gulrajani, I., Zhang, T., Dubois, Y., Li, X., Guestrin, C., Liang, P., and Hashimoto, T. B. Stanford alpaca: An instruction-following llama model. https://github.com/tatsu-lab/ stanford_alpaca, 2023.]\"}", "{\"title\": \"Official Comment by Reviewer w3um\", \"comment\": [\"Thanks for your comment. Let me systematically address the novelty concerns:\", \"1 .Regarding publication timing:\", \"Paper [1] was published on May 24, 2024\", \"Paper [2] was published on July 22, 2024\", \"Being within a 3-month period, these would be considered part of the same submission cycle. In such case, if both papers were submitted to the same conference, they could reasonably claim concurrent development of latent adversarial training for LLMs.\", \"Moreover, neither paper[1,2] has been published at any academic conference to my knowledge. If paper [1] was already public since May 2024, and paper[2] made an submission to ICLR in October, the novelty requirement would be substantially higher since [1] would constitute prior work.\", \"2. Regarding Novelty of this paper\", \"As I mentioned in my previous review: **Therefore, applying the existing LAT method [2] to VLLMs would only require adjusting the attack indices which is a very small modification, as both image and text representations are fundamentally treated the same way in the backbone LLM.** Additionally, authors haven't addressed a crucial question: Would attacking both image and text embeddings across multiple VLLM layers improve performance? If such an approach yields better results, it would suggest that a minor modification to existing methods could outperform their proposed approach.\", \"To make my views clearer, consider these two analogous cases:\", \"Case 1: Previous work proposes latent adversarial training in traditional computer vision models[3], and follow-up work transfers LAT to LLMs[1,2] - this represents a significant novel contribution.\", \"Case 2: Previous work proposes a novel training method for computer vision classification models, and follow-up work simply applies this method to the backbone of an object detection model claiming better performance - this represents an incremental adaptation.\", \"This work aligns more with Case 2, representing a straightforward application of an existing technique rather than a substantial technical advancement. Therefore, I believe the current work lacks sufficient novelty for ICLR.\", \"[1] Sophie Xhonneux, Alessandro Sordoni, Stephan G\\u00a8unnemann, Gauthier Gidel, and Leo Schwinn. Efficient adversarial training in llms with continuous attacks. arXiv preprint arXiv:2405.15589, 2024\", \"[2] Sheshadri, Abhay, et al. \\\"Targeted latent adversarial training improves robustness to persistent harmful behaviors in llms.\\\" arXiv preprint arXiv:2407.15549 (2024).\", \"[3] Park, Geon Yeong, and Sang Wan Lee. \\\"Reliably fast adversarial training via latent adversarial perturbation.\\\" Proceedings of the IEEE/CVF international conference on computer vision. 2021.\"]}", "{\"title\": \"Rebuttal by Authors (Part 3/4)\", \"comment\": \"`>>> Weakness 6`: ***What are the specific settings for model output quality in Figure 3? The authors tested only 100 samples, which may not effectively reflect changes in model output capabilities. It is recommended that the authors test on 500 benign samples to assess the effects accurately.***\\\\\\n`>>> Response': Thanks for your valuable suggestions. For the experimental setup in Figure 3, we follow LLaVA by adopting gpt-4-turbo and using the same prompt to score each model\\u2019s responses on a scale from 1 to 10. We report the average score on 100 samples extracted from the LLaVA-Instruct-80K dataset. \\n\\n***We have expanded the utility experiments to more comprehensively evaluate the impact of SafeMLLM on the model's general capabilities via two extra evaluations on 517 samples.***\\n\\nFirst, in addition to the original 100 samples used in Figure 3, we additionally sampled 200 samples from the LLaVA-Instruct-80K dataset and adopted the same method to score each model\\u2019s responses based on gpt-4-turbo. We conducted experiments on both LLaVA-7B and LLaVA-13B, and the average scores over 300 test samples (100+200) are shown below:\\n\\n|LLaVA-7B| Original | VLGuard | R2D2| CAT| SafeMLLM| \\n |-| - | - |-|-|-|\\n|ASR|7.65 | 7.67 |7.58 |7.62|7.64|\\n\\n|LLaVA-13B| Original | VLGuard | R2D2| CAT| SafeMLLM| \\n |-| - | - |-|-|-|\\n|ASR|7.79 | 7.73 |7.68 |7.54|7.73|\\n\\nFrom the table, we can observe that after training the model with SafeMLLM, its response quality on these benign questions has not been moderately affected. \\n\\nWe also adopt MM-Vet [1], a widely-used MLLM evaluation benchmark, to comprehensively evaluate the capability of SafeMLLM across various aspects. The benchmark contains 217 multimodal questions and adopts gpt-4-turbo to evaluate the target model\\u2019s responses from the following dimensions: Recognize (Rec), OCR, Knowledge (Know), Language Generation (Gen), Spatial awareness (Spat), and Math. The results on LLaVA-7B and LLaVA-13B are reported in the table below. For each metric, higher values indicate better performance.\\n\\n| LLaVA-7B| rec | ocr | know | gen | spat | math | total |\\n|-------------|------|------|------|------|------|------|-------|\\n| Original | 36.9 | 24.0 | 18.5 | 20.5 | 28.0 | 3.8 | 32.2 |\\n| VLGuard| 33.9 | 22.9 | 13.8 | 14.2 | 27.2 | 3.8 | 30.1 |\\n| R2D2| 34.7 | 21.5 | 16.4 | 18.1 | 24.3 | 7.7 | 30.2 |\\n| CAT| 37.7 | 20.1 | 24.3 | 25.1 | 25.7 | 3.8 | 31.5 |\\n| SafeMLLM| 37.5 | 24.1 | 20.5 | 21.1 | 28.3 | 3.8 | 32.5 |\\n\\n| LLaVA-13B| rec | ocr | know | gen | spat | math | total |\\n|-------------|------|------|------|------|------|------|-------|\\n| ori | 42.1 | 25.9 | 24.4 | 25.1 | 30.4 | 11.2 | 36.0 |\\n| vlguard | 37.7 | 26.6 | 17.7 | 21.4 | 30.9 | 3.8 | 32.9 |\\n| R2D2 | 41.1 | 26.2 | 24.4 | 26.1 | 32.0 | 7.7 | 35.4 |\\n| RTEAT | 42.7 | 27.7 | 26.7 | 26.1 | 32.7 | 15.0 | 36.9 |\\n| safemlllm | 44.0 | 27.1 | 23.8 | 25.6 | 34.0 | 15.0 | 37.8 |\\n\\nFrom the table, we observe that SafeMLLM still maintains response quality across all aspects. Finally, based on these two experiments involving more than 500 image-text questions (300+217), we demonstrate that SafeMLLM minimally compromises the overall capabilities of the target MLLM. We have added both experiments in Appendix G (page 17, lines 908-917; page 18, lines 949-956).\\n\\n[1] Yu, W., Yang, Z., Li, L., Wang, J., Lin, K., Liu, Z., ... & Wang, L. MM-Vet: Evaluating Large Multimodal Models for Integrated Capabilities.ICML 2024.\"}", "{\"title\": \"Rebuttal by Authors (Part 2/2)\", \"comment\": \"`>>> Weakness 2 & Question 3`: ***Figure 2 presents an overview of the proposed SAFEMLLM framework, with a focus on the first phase, CoE-Attack. In this phase, the paper emphasizes the optimization of two noise metrics: noisy image and noisy text. Are these the only types of noise considered in the SAFEMLLM, or does the framework have the potential to incorporate other modalities such as video and audio? If not, it is better to discuss the limitations of the current approach in terms of its applicability to a broader range of multimodal data?***\\\\\\n`>>> Response`: Thanks for your comments. We admit that we only consider two commonly used modalities since most existing multimodal jailbreak methods only focus on these two modalities. However, the proposed ***SafeMLLM has the potential to be extended to address potential attacks involving additional modalities.*** For example, for the video-based jailbreak attacks, SafeMLLM could insert more adversarial tokens before the query to counter the adversarial noise across a sequence of multiple image frames. Alternatively, we could interleave perturbed embeddings within the inputs to adapt flexibly to modality switching. As you suggested, we have discussed the limitations regarding SafeMLLM\\u2019s applicability to a broader range of multimodal data in the revised manuscript (Appendix L, page 21).\\n\\n`>>> Weakness 3`: ***Authors fail to provide experimental results demonstrating that their framework can reduce overall computing resources.***\\\\\\n`>>> Response`: Thank you for pointing this out. As mentioned in Section 3.2 (page 4, lines 206-208), optimizing an adversarial image in front of the toxic query and an adversarial text string after the query simultaneously in Step I could be highly computationally intensive. To validate this claim, we conducted an additional experiment. Specifically, we replace the front token embedding $\\\\mathbf{P}_0^h$ with a given image input $\\u200bI_0$, and optimize the perturbations on both \\u200b$I_0$ and the token embedding $P_0^t$ placed after the query in Step I. In step II, we update the model based on the optimized perturbation accordingly. We refer to this approach as w/ Adv.Image. We test it against SafeMLLM on the LLaVA-7B and LLaVA-13B models using the ImgJP attack, comparing the average runtime per iteration (step I + step II) and GPU memory usage. The results are illustrated in the table below:\\n\\n|LLaVA-7B| runtime (s)\\u2193| GPU Memory (MB)\\u2193| ASR\\u2193| \\n |-| - | - |-|\\n|w/ Adv.Image|84.42 | 32869 |5.00 |\\n|SafeMLLM|20.73 | 30291 |6.00 |\\n\\n|LLaVA-13B| runtime (s)\\u2193| GPU Memory (MB)\\u2193| ASR\\u2193| \\n |-| - | - |-|\\n|w/ Adv.Image|263.56 | 66092 |0.00 |\\n|SafeMLLM|38.70 | 57475 |0.00 |\\n\\nAs shown in the table, optimizing image perturbations significantly impacts computational efficiency but does not yield noticeable gains in ASR performance, thereby validating our claim that simultaneously optimizing the perturbations on the original inputs is not practical. We have added this experiment in Appendix H (page 18, lines 962-971).\\n\\n`>>> Question 4`: ***Why are there many missing values for the VLGurad metric in the performance indicators listed in Table 1?***\\\\\\n`>>>Response`: As mentioned in lines 347-350 on page 7, we evaluate VLGuard in Table 1 by using the models officially trained and released by [1] to ensure fairness in comparison. Since ***VLGuard has not released checkpoints for models trained on MiniGPT-4 and InstructBLIP,*** we did not report ASR results for these two types of models.\\n\\n[1] Yongshuo Zong, Ondrej Bohdal, Tingyang Yu, Yongxin Yang, and Timothy M. Hospedales. Safety fine-tuning at (almost) no cost: A baseline for vision large language models. In ICML. OpenReview.net, 2024\"}", "{\"title\": \"Rebuttal by Authors (Part 1/2)\", \"comment\": \"Thank you for your valuable suggestions. Our responses are provided below, and we hope they can address your concerns.\\\\\\n`>>> Weakness 1 and Question 1`: ***This paper lacks novelty. Though it's the first paper to perform adversarial tuning on MLLMs, the attack method is only conducted on the embedding layer and there is no specific modification to the components directly related to VLLM such as the image encoder. Previous work[1, 2 ] on LLMs has already demonstrated the effectiveness of adversarial training in the latent space. Meanwhile, previous work also proposes to first perform attacks on multiple layers of LLM (including the embedding layer) and then fine-tune the model against such attacks. Since this method only performs attacks on the embedding layer, it's not clear why previous work cannot be directly applied to this task. How is this method different from other latent adversarial training methods [1] [2] ?***\\\\\\n`>>> Response`: Although SafeMLLM and recently proposed latent adversarial training (LAT) methods both adopt an adversarial training framework, we assert that they still differ significantly from SafeMLLM in the following aspects:\\n\\n(1) ***Different Perturbation Sets***: Unlike jailbreak attacks on LLMs that are based on discrete text tokens, attackers on MLLMs can directly inject unrestricted pixel-level perturbations into images to compromise safety alignment. Thus, we leverage this unique property and propose optimizing unbounded perturbations at the token embedding level, unifying potential adversarial perturbations across different modality inputs in the worst scenario. \\n\\nHowever, existing LAT methods [1,2] may not be suitable for our problem. Specifically, when a toxic prompt is used as a training sample, these LAT methods generate noise by adding perturbations to the intermediate token representations corresponding to this text prompt. This design is based on the hypothesis that injecting perturbations into the intermediate text token features is equivalent to, and even stronger than, adding extra non-word adversarial text tokens after the query in terms of attack intensity. However, this hypothesis may ***no longer stand for MLLM attacks as attackers not only inject perturbations through discrete text tokens but can also introduce noise via images with continuous values.*** More importantly, in MLLMs, the number of tokens for images largely exceeds that of text prompt tokens (e.g. 576 tokens on LLaVA-13B), making these latent adversarial training methods ***less effective.*** \\n \\nThus, this design is novel and different from LAT-based method. We also conduct an experiment to validate our model design by comparing the proposed SafeMLLM with the lastest LAT-based method [2]. Specifically, we follow the setting used in [2], and inject perturbations into latent token features with $\\\\epsilon=20$ in the L2 constraint. For a fair comparison, we adopted the same optimization target in SafeMLLM and also increased $\\\\epsilon$ to 40 to explore scenarios under stronger attacks. The modified method is named SafeMLLM-LAT. The results of using the ImgJP attack method on LLaVA-13B are shown in the table below:\\n\\n || SafeMLLM-LAT($\\\\epsilon=20$)| SafeMLLM-LAT($\\\\epsilon=40$)| SafeMLLM| \\n |-| :-: | :-: |:-:|\\n |ASR| 13.00 | 12.00 |0.00 |\\n\\nAs shown in the above table, adding perturbations to the latent representations performs worse than directly using the adversarial token embeddings, which confirms our hypothesis.\\n\\n(2) ***Different training objectives***. Besides the perturbation sets, another difference lies in the training objectives in both the attack loop and model updating step. The first LAT method [1] sets up the adversarial training in an untargeted manner. The second LAT method [2] improves this into a targeted attack objective, which optimizes the adversarial noise on the affirmative response and updates the model based on another rejective response. \\n\\nHowever, our attack objective differs from the above two approaches. ***In addition to the target loss, we propose a contrastive loss that adaptively suppresses the probabilities of generating unexpected texts at different steps.*** During the attack optimization, the proposed loss enhances the perturbation strength by reducing the probability of sampling the rejective response. Correspondingly, during the model updating process, the contrastive loss improves the model's robustness by increasing the probability difference between sampling the safety and toxic responses. The experiments in Section 5.2 and Table 2 further verify this point, showing that the inclusion of contrastive loss significantly improves ASR performance across three MLLMs. \\n\\nOverall, our proposed SafeMLLM is different from the LLM-based LAT methods in the above aspects. The experimental results in Table 1 also demonstrate the effectiveness of our proposed adversarial training algorithm.\"}" ] }

[ "Harsh Chaudhari", "Giorgio Severi", "John Abascal", "Matthew Jagielski", "Christopher A. Choquette-Choo", "Milad Nasr", "Cristina Nita-Rotaru", "Alina Oprea" ]

[ "Large Language Models", "AI Security", "AI Safety", "RAG", "Poisoning Attacks" ]

{ "note_id": [ "umJKFu4pv1", "sLsDd0WxEv", "jiVzlqn3Ku", "byNXirpdzS", "YhWvmSJIgX", "VhFNer9QUI", "JQjG8memi2", "JAFPAepu4L", "FVaoTLF2hN", "71rfdpKef8", "5j02oHMDjC" ], "note_type": [ "official_review", "official_comment", "meta_review", "official_comment", "official_comment", "official_review", "official_comment", "decision", "official_comment", "official_review", "official_comment" ], "note_created": [ 1730320562635, 1732186509021, 1734203718278, 1732436814244, 1732611587284, 1730691244515, 1732186261056, 1737523895806, 1732628897648, 1730452160365, 1732186883663 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission8226/Reviewer_SyA5" ], [ "ICLR.cc/2025/Conference/Submission8226/Authors" ], [ "ICLR.cc/2025/Conference/Submission8226/Area_Chair_tja5" ], [ "ICLR.cc/2025/Conference/Submission8226/Authors" ], [ "ICLR.cc/2025/Conference/Submission8226/Reviewer_88ad" ], [ "ICLR.cc/2025/Conference/Submission8226/Reviewer_aMuq" ], [ "ICLR.cc/2025/Conference/Submission8226/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission8226/Authors" ], [ "ICLR.cc/2025/Conference/Submission8226/Reviewer_88ad" ], [ "ICLR.cc/2025/Conference/Submission8226/Authors" ] ], "structured_content_str": [ "{\"summary\": \"This work focuses on the potential backdoor attacks on RAG. The authors propose an attacking method that inserts malicious data into the databases used in RAG, optimizes the data for retrieval and induces malicious behaviors. Experiments are conducted on various models and datasets including an application ChatRTX.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"This work reveals an important threat that may compromise the performance of RAG. Multiple malicious behaviors are considered and evaluated. The optimization objectives are well formulated and explained. The empirical results look promising.\", \"weaknesses\": \"1. **Threat model**\\n* The threat model of local deployment and local file system is confusing and lacks practicality. Local files of the user's own are usually private and it is hard to imagine what kind of attacker can access and insert something into it. Though the authors mention something harmful like spam emails, they are usually taken care of before leveraging them in the system. The authors take for grant that malicious contents can be inserted into the database used in RAG, but I personally believe insertion itself is hard to either justify or conduct. The threat model should be more clearly and properly justified.\\n* When optimizing the tokens for retrieval, the embedding of retriever is assumed to be accessible to attackers, which can be a strong assumption. I would expect a more practical scenario that both retrievers and generation models are black-box and attackers can adopt some open-source models to craft examples. The totally black-box setting needs to be discussed either with empirical evaluations or explicitly discussing the implications and limitations of the current assumptions.\\n\\n2. The design of s_ret and s_gen seems to be not imperceptible and is easy to detect. According to Section 4, s_ret and s_gen are sequences of tokens optimized by HotFlip and MCG respectively. Since there are no constraints to ensure the semantic meaningful, the optimized sequences can be full of non-meaningful strings. I can imagine that the perplexity score can also be much higher than ordinary content, which makes this attack easy to defend. This can be even more serious if the length of strings is getting longer. Moreover, the authors append many additional contents to the original query (s_ret, s_gen, s_cmd), the robustness of injected content can be very poor. Some case studies on the backdoored content should be provided.\\n\\n3. The proportion of backdoored content in the database should be mentioned in the experiment part. I believe this proportion can be an important factor influencing both retrieval rate and final success rate. Ablation study on this proportion can help.\", \"questions\": \"See weakness\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"We thank the reviewer for their interesting comments. Below, we address the main concerns expressed in the review.\\n\\n**W1: Technical contributions**\\n\\nOne of our main technical contributions is the design of the optimization framework for poisoning RAG systems that optimizes the retrieval and the generator attacks separately, while supporting five adversarial objectives. In addition, we made technical contributions to the design of the retrieval and generator optimization components: \\n\\n**Retriever contribution:** We proposed a specialized loss function that enables selective retrieval - the adversarial passage is retrieved only when a specific trigger sequence appears in the user query. While we leverage HotFlip for optimization, the novelty lies in our loss function design that precisely controls when the poisoned content is extracted by the retrieval system. The poisoned passage we design generalizes to any new queries including the trigger sequence. \\n\\n**Generator contribution:** We enhanced the jailbreaking capability against the LLM through our Multi-Coordinate Gradient (MCG) approach. Unlike the original GCG method, which modifies tokens sequentially, MCG optimizes multiple jailbreak tokens in parallel. This parallel optimization strategy explores the solution space more efficiently, resulting in faster convergence. We also apply the jailbreaking optimization strategy to five different adversarial objectives, unlike most jailbreaking papers that focus on generating harmful content. \\n\\n**W2: Baselines**\\n\\n**Comparison with baseline:** [1] focuses on retrieving the adversarial passage for any user query making the attack easily detectable. In our case the adversarial passage is retrieved only when a given natural trigger is present in the query, which results in a more stealthy attack. To achieve this functionality we are required to build a new loss function, which is then optimized using HotFlip.\\n\\nThe reviewer\\u2019s suggestion, to extend the baseline [1] into an untargeted attack for the RAG system is already addressed by concurrent work [2]. However, [2] focuses on \\u201cuntargeted attacks\\u201d that are activated for any user input, whereas our method retrieves the poisoned passage only when a specific natural trigger is present in the queries, making our approach more stealthy in real-world scenarios. Secondly, their bi-level optimization approach requires more than 1000 iterations, while our two-stage optimization attack achieves high attack success in just 32 iterations - making our approach 30x more efficient.\\n\\n[1] Zhong et al. \\u201cPoisoning Retrieval Corpora by Injecting Adversarial Passages\\u201d, EMNLP\\u201923Zhong et al. \\u201cPoisoning Retrieval Corpora by Injecting Adversarial Passages\\u201d, EMNLP\\u201923\\n\\n[2] Tan et al. \\\"Glue pizza and eat rocks\\\"--Exploiting Vulnerabilities in Retrieval-Augmented Generative Models.\\\" The 2024 Conference on Empirical Methods in Natural Language Processing, arXiv\\u201924.\"}", "{\"metareview\": \"The paper shows an adversarial threat against the retrieval augmented generation (RAG) system. Specifically, it constructs a backdoor in a document, and the language model would not generate the correct response when the trigger appears. The paper designs a two-stage optimization framework that both improves GCG attack's efficiency and keeps the normal functionality where the trigger is not shown. While the task is very important and the experiments show promising results, the paper still has several major concerns solved before publication. First, the threat model must be clarified further, especially in real-world applications. The current threat model shares a similarity with the jailbreak attack so it makes the paper's novelty and contributions less significant. Second, I encourage the author to include some related works in the discussion as discussed in the rebuttal, which can further clarify the novelty of the proposed methods. Third, it will make the paper better if the author can include some discussions with potential defenses and include more baselines with different retriever and attack methods.\", \"additional_comments_on_reviewer_discussion\": \"All reviewers show concerns about the novelty and insufficient experiments. Reviewer aMuq's concerns are mostly in previous works and some experimental settings. Reviewer SyA5 raises concerns over the threat model. I greatly appreciate the effort the authors made to address the concerns raised by reviewers. Some concerns about the experiments are solved however I believe the novelty and threat model concerns are still valid and they also make the paper's novelty and contributions less significant. As suggested by the metareview, I believe the paper can be much improved by addressing these concerns for future publication.\"}", "{\"title\": \"Clarifications on design of Phantom and comparison with other works (2/2)\", \"comment\": \"**W4: Loss modeling.**\\n\\nWe would like to clarify that Equation (2) does not assume that only one passage is retrieved. Our approach is explicitly designed and tested for multi-passage retrieval scenarios, with all experiments consistently using k=5 for retrieved passages. During MCG optimization, each query has a corresponding context which consists of 4 passages relevant to the query content plus one adversarial passage, all ranked based on their similarity scores. Additionally, the prefix/suffix positioning discussed in Appendix A.2.26 (Table 9) refers specifically to token arrangement within the adversarial passage itself - meaning the arrangement of (sret+sgen+adv_cmd) vs (sret+adv_cmd+sgen) is internal to the malicious passage and independent of where the adversarial passage appears in the overall retrieved context. This clarification should resolve the confusion, as our method is explicitly designed and validated in a realistic multi-passage retrieval setting, not just the simplified single-passage case.\\n\\n**W5: Retrievers.**\\n\\nWe evaluate our attacks on retrievers commonly studied in both published and concurrent work [1,2,3,4]. These works have also observed lower success rates for both targeted and untargeted attacks on retrievers such as Contriever-MS and DPR pointing to varied robustness of retrievers against adversarial attacks. Also note that, we add only a single poisoned passage unlike many works [1,2,3] that require multiple poisoned passages, and still achieve substantial retrieval rate exceeding 50% on these retrievers. One could further boost the attack success similar to prior work by adding more poisoned passages such that at least one of adversarial passages shows up in the top-k, improving the reported metrics. The goal of our work was to minimize poisoning while having an effective attack.\\n\\n**W6: Contriever-MS.**\\n\\nThe high retrieval failure rates observed with Contriever-MS, particularly the significant performance gap between Contriever-MS and its base model, align with findings from concurrent works [1], [2] and [3]. These observations suggest that fine-tuning on structured datasets like MS MARCO might introduce additional robustness against adversarial attacks. Understanding the impact on robustness of retrievers via fine-tuning with structured dataset, would be an interesting topic of future work. \\n\\n**W7 chat with RTX.**\\n\\nWe successfully attack the black-box NVIDIA Chat-with-RTX system. While transferring between retrievers is challenging, our attack transferred from Contriever to Chat-with-RTX. We don\\u2019t have details on the retriever used by Chat-with-RTX, but we suspect it uses a similar architecture to Contriever. \\n\\nWe ran additional experiments on ChatRTX using the trigger word \\u201cconfidential\\u201d. This trigger word is especially fitting because it is in the top 0.05% most frequent words in the Enron dataset (~500k documents), used as the knowledge base, appearing a total of 24.5k times. Thus, the relatively high frequency of natural occurrences increases the difficulty of successfully inducing the retriever to select the corrupted passage. \\n\\nTo test Phantom on the Mistral 7B int4 model in ChatRTX, we used 25 random queries from the MSMARCO dataset that contain the trigger word. Of the 25 queries, the adversarial passage was retrieved as the top-1 document from the knowledge base 15 times, yielding a retriever failure rate of 40% in a true black-box setting. Of the 15 queries where the adversarial passage was retrieved, the adversarial objective (Passage Exfiltration) was executed 12 times, leaking the model\\u2019s context, for a total attack success rate of 48%.\\n\\n[1] Xue, Jiaqi, et al. \\\"BadRAG: Identifying Vulnerabilities in Retrieval Augmented Generation of Large Language Models.\\\" arXiv preprint arXiv:2406.00083 (2024).\\n\\n[2] Tan, Zhen, et al. \\\"\\\" Glue pizza and eat rocks\\\"--Exploiting Vulnerabilities in Retrieval-Augmented Generative Models.\\\" The 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP 2024)\\n\\n[3] Zou et al. \\u201cPoisonedRAG: Knowledge Corruption Attacks to Retrieval-Augmented Generation of Large Language Models\\u201d (USENIX\\u201925)\\n\\n[4] Zhong et al. \\u201cPoisoning Retrieval Corpora by Injecting Adversarial Passages\\u201d, EMNLP\\u201923\"}", "{\"title\": \"Official Comment by Reviewer 88ad\", \"comment\": \"Thank you for your response. I will maintain my positive score.\"}", "{\"summary\": \"This paper introduces Phantom, a framework that proposes a novel attack on Retrieval-Augmented Generation (RAG) systems by injecting malicious documents into their knowledge bases, which are then retrieved by specific trigger sequences to manipulate outputs. The attack utilizes a two-stage optimization process: firstly, crafting a document that aligns with the RAG's retrieval mechanics when triggered, and secondly, generating text to induce specific adversarial outcomes like misinformation or refusal to answer. Extensive experiments demonstrate the effectiveness of Phantom across various large language models and datasets, revealing vulnerabilities in both proprietary and open-source RAG implementations. The paper also discusses potential mitigations, stressing the challenge of balancing system security with the utility of RAG systems in real-world applications.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": [\"The experiments of this paper are extensive.\", \"The security of the RAG system is an important topic.\"], \"weaknesses\": \"1. The threat model lacks clarity, particularly regarding the attacker's capabilities. Line 153 suggests the attacker does not need knowledge of other documents to initiate the attack. However, the Multi-Coordinate Gradient (MCG) technique described on lines 274-275 requires access to the top-k passages retrieved in response to m user queries. These passages, presumably from the RAG databases, are crucial for optimization. Appendix A.3.5 explores transferability between different datasets, implying the use of a consistent dataset (MS MARCO) for both optimization and attack phases. This suggests that the attacker must have knowledge of the RAG database to effectively deploy the MCG technique.\\n\\n2. The paper lacks a comprehensive comparison with prior works. Generating malicious passages retrievable exclusively by queries with specific triggers has been explored by [1], and inducing target outputs in LLMs using suffix generation has been studied by [2]. This paper does not compare its methods against [1] concerning Retrieval Failure Rate, nor does it compare attack success rates with [2]. Additionally, the authors should consider citing [2], which introduces a bi-level optimization for generation and retrieval conditions and employs gradient-based optimization to concatenate strings for retrieval and generation.\\n\\n3. The novelty of the second method (MCG) appears limited. It is primarily a variant of the GCG with modifications only detailed in lines 7-8. Although MCG reportedly outperforms GCG, the reasons for this improvement and the rationale behind this incremental change are not well-explained.\\n\\n4. The loss modeling for MCG is problematic. Equation 2 assumes that only one passage is retrieved and used as input for the generator. However, typical RAG applications retrieve multiple passages for context. Notably, Appendix A.2.26's Table 9 shows that the position of $S_{gen}$ significantly affects its efficacy, with prefixes performing better than suffixes. In scenarios where multiple passages are retrieved, there's no guarantee the malicious passage will be at the top-1 position, undermining the experiment's validity if $S_{gen}$ is not the prefix of the entire context. Why the string $S_{gen}$ optimized on a single passage could be effective in multiple passage settings is not clear to me\\n\\n5. There is a noticeable absence of experiments involving other retrievers such as DPR, BGE, REALM, or ORQA. While the paper discusses the retrieval failure rate on DPR in the appendix, it does not present end-to-end attack success rates for these retrievers in RAG systems. The high retrieval failure rates on other retrievers besides Contriever suggest limited practicality and effectiveness of the attacks.\\n\\n6. Could the authors provide explanations for the high retrieval failure rates observed with DPR and Contriever-MS? It is particularly puzzling for Contriever-MS, a version of Contriever fine-tuned on the MS MARCO dataset, to show a significant performance discrepancy compared to its base model.\\n\\n7. The paper discusses an attack against the Chat RTX without detailed knowledge of its retriever. Given the poor transferability between retrievers noted in lines 1105-1107, the successful attack using Contriever on the black-box Chat RTX system is perplexing. Could the authors provide the specific success rates achieved against it?\\n\\n[1] Xue, Jiaqi, et al. \\\"BadRAG: Identifying Vulnerabilities in Retrieval Augmented Generation of Large Language Models.\\\" arXiv preprint arXiv:2406.00083 (2024).\\n\\n[2] Tan, Zhen, et al. \\\"\\\" Glue pizza and eat rocks\\\"--Exploiting Vulnerabilities in Retrieval-Augmented Generative Models.\\\" The 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP 2024)\", \"questions\": \"Please refer to the weakness part.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Clarifications on design of Phantom and comparison with other works (1/2)\", \"comment\": \"We thank the reviewer for their comments. The following paragraphs address the main comments and provide clarifications on the design of Phantom.\\n\\n**W1: Attacker\\u2019s knowledge.**\\n\\nThank you for pointing out the assumption. While passages are sampled from MS-Marco, MCG optimization is conducted on OUT queries (unrelated to the target trigger) with the trigger artificially appended. This ensures that the top-$k$ retrieved passages ($k=5$) are unrelated to the trigger sequence, supporting the assumption that no prior knowledge of trigger-related documents is required.\\nTo further validate that our attack works without relying on passages from the dataset, we fill the top-k passages for MCG optimization with synthetically generated passages by an external oracle, such as GPT-4, relevant to the query. Testing on two triggers, \\\"xbox\\\" and \\\"lebron james,\\\" we observe similar effectiveness as our original experiments, with attack success rates of 88% and 79%, respectively.\\n\\n**W2: Concurrent work.**\\n\\nWhile we appreciate the comparison with these works, they should be considered concurrent given the timeline of these submissions. We address the comparisons separately:\\n\\n**Regarding [1]:** Our attack demonstrates broader applicability, successfully targeting 5 different adversarial objectives compared to only two in [1]. A crucial distinction is that our attack remains effective even when the adversarial objective is in conflict with the RAG generator's safety alignment (e.g., \\\"Threaten the user\\\"). In contrast, their approach does not work in such scenarios-due to the attacker's inability to circumvent the generator\\u2019s safety alignment. \\nAdditionally, their attack relies on unnatural triggers for context leakage and tool usage attacks - likely due to jailbreaking constraints, whereas we don't have such limitations. While both works formulate the retriever loss function in a similar fashion, our end-to-end attack overcomes the key limitations discussed above. Finally, we require only one poisoned passage to be added into the RAG database while [1] requires around 10 poisoned passages for their attack to succeed.\\n\\n**Regarding [2]:** There are fundamental differences in both approach and efficiency:\", \"attack_scope\": \"Their work focuses on untargeted attacks that are activated for any user input, whereas our method activates only when a specific natural trigger is present in the queries, making our approach much more stealthy in real-world scenarios.\", \"optimization_strategy\": \"They propose a bi-level optimization that alternates between HotFlip (for retriever tokens) and GCG (for generator tokens) and requires over 1000 iterations for the attack to succeed. They don't provide any empirical justification for why their approach is superior to a sequential two-step optimization. Our two-step attack, on the other hand, achieves high attack success in just 32 iterations - making our approach 30x more efficient than [2].\", \"empirical_findings\": \"Their work also observed similar issues when attacking Contriever-MS Marco compared to base Contriever, which supports our findings about retrievers having different robustness to poisoning.\\n\\nWe have added detailed comparisons with three concurrent works in Appendix A.1 of the revised version of the paper, which also includes the two works that we discussed above.\\n\\n**W3: Rationale for MCG.**\\n\\nThe Multi-Coordinate Gradient (MCG) approach, while built upon GCG, introduces a strategic modification to address the limitation of token-by-token optimization. Our primary intuition behind MCG was to enable more efficient exploration of the optimization space by modifying multiple tokens simultaneously in early iterations, rather than being constrained to single-token changes as in GCG. As the optimization progresses, MCG naturally transitions to modifying fewer tokens, eventually converging to GCG-like behavior. This allows us to converge faster with fewer iterations, making our attack more efficient to run. (Table 7, Appendix A.2.4)\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"comment\": \"Thank you for your response! We have added detailed comparisons with three concurrent works in Appendix A.1 of the revised version of the paper, which also includes the Tan et al.'24 work that we discussed above.\"}", "{\"summary\": \"Retrieval Augmented Generation (RAG) enhances the capabilities of modern large language models (LLMs), but also provides new vectors for attackers. This paper investigates mounting a backdoor poisoning attack by injecting a single malicious document into a RAG system\\u2019s knowledge base. Specifically, the paper proposes a two-stage optimization framework to optimize the injected malicious document, with the first optimization goal being to be retrieved only when a specific trigger sequence of tokens appears in the victim\\u2019s query, and the second goal being to induce various adversarial objectives on the LLM output.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. The study examines multiple attack objectives against RAG systems, such as Refusal to Answer, Biased opinion, Harmful Behavior, Passage Exfiltration, and validates them through extensive experiments.\\n2. The paper attacks a commercial black-box RAG system, NVIDIA\\u2019s Chat-with-RTX, and demonstrates that the attack achieves various objectives.\\n3. This paper is well-written and well-organized.\", \"weaknesses\": \"1. The technical contribution is not substantial enough. The paper's two-stage optimization framework is based on two existing optimization algorithms, HotFlip and GCG. Even though some improvements have been made for the attack scenario, the gap in technical contribution is not significant.\\n2. Lacks baselines. Although the paper proposes a new threat model of backdoor poisoning in untrusted RAG knowledge bases, straightforward baselines should be set for comparison to demonstrate the superiority of the proposed method. For example, [1], although this method is not optimized for manipulating LLM generation, it can also manipulate the subsequent behavior of the LLM by pre-defining the semantics of the retrieved malicious text.\\n\\n[1] Poisoning retrieval corpora by injecting adversarial passages. EMNLP2023\", \"questions\": \"Please see the weaknesses for details.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"We thank the reviewer for their insightful comments, and we address the main concerns and clarifications below.\\n\\n**W1: Local RAG system use case**\\n\\nIndirect prompt injection attacks, also often referred to as XPIA (cross prompt injection attacks), are increasingly becoming more common. While we chose the use case of Chat RTX as our motivating example, this attack is not limited to local files and can be carried out against a large variety of RAG systems, as they operate on documents originating from both a user\\u2019s own file storage (either local or cloud) and Internet sources such as Wikipedia. \\n\\n**W1: Threat model**\\n\\nOur work represents one of the earliest systematic investigations into context poisoning attacks on RAG systems. Recently, however, practitioners in the field have started demonstrating the effectiveness of XPIA against major commercial RAG products [1, 2, 3], further spotlighting the relevance of this threat model. \\n\\nAll these attacks, while simpler and less versatile than Phantom, underscore a crucial common vulnerability: Attackers can readily inject malicious content into a model's context window through various vectors. These vectors include, but are not limited to: drive-by downloads, hidden text in the body of email messages, strings encoded with ascii-smuggling [4], malicious strings embedded in HTML markup, exploitation of Markdown rendering [5], and direct social engineering. This proliferation of attack vectors validates our initial security concerns and suggests that compromising even a small portion of a model's context window is more feasible than often believed.\\n\\n[1] https://embracethered.com/blog/posts/2024/github-copilot-chat-prompt-injection-data-exfiltration/ \\n\\n[2] https://embracethered.com/blog/posts/2024/google-ai-studio-data-exfiltration-now-fixed/ \\n\\n[3] https://embracethered.com/blog/posts/2024/claude-computer-use-c2-the-zombais-are-coming/ \\n\\n[4] https://embracethered.com/blog/posts/2024/hiding-and-finding-text-with-unicode-tags/ \\n\\n[5] https://simonwillison.net/tags/markdown-exfiltration/ \\n\\n**W1: White-Box Access and Transfer Attacks**\\n\\nWe find that our attacks transfer, both to other open source RAG generators (in Appendix A.2.7) and to a Black-Box RAG system called Chat-with-RTX, developed by NVIDIA (Section 5.2). While transferability to some systems like DPR remains a challenge, the successful transfer to Chat-with-RTX shows promise, despite having no access to its retrieval component or generator model. This demonstrates that while our attacks require white-box access to a surrogate model for generating adversarial strings, they can succeed in black-box settings via direct transfer. We believe future work would focus on improving transfer reliability across different architectures.\\n\\n\\n**W2: Perplexity defense**\\n\\nThe current implementation of the attack generates strings that deviate from grammatically valid English sentences, likely resulting in elevated perplexity scores when evaluated using a properly trained language model. This represents a common challenge faced by all current optimization attacks on RAG systems.\\n\\nWhile it represents a valid detection approach in the short term, perplexity-based detection introduces new complexities into the defensive landscape. Perplexity scoring requires careful calibration specific to the target content distribution, as elevated perplexity may be induced by valid content belonging to distributions unknown to the evaluation model. Moreover, it transforms the defensive problem into an optimization challenge: attackers could incorporate a perplexity minimization component into their objective functions, potentially leading to an arms race, where both detection models and attack strategies undergo continuous refinement.\\n\\n\\n**W2 Alterations to the user\\u2019s query**\\n\\nPlease note that our attack strategy does not alter the user\\u2019s query in any way. The adversarially crafted strings are placed only in the single corrupted document introduced in the knowledge base.\\n\\n\\n**W3 poisoning rate**\\n\\nOne of the core strengths of the Phantom attack is that it only requires a single poisoned passage to be injected in the knowledge base. This is negligible relative to the size of common RAG knowledge bases (for instance, MSMarco has ~8 million passages). While the attacker can always choose to introduce more poisoned content to increase the likelihood of successful attacks, we show that this is not necessary, as the adversary can achieve high success rates with a single poisoned passage.\"}" ] }

[ "Yuxin Wang", "Maresa Schröder", "Dennis Frauen", "Jonas Schweisthal", "Konstantin Hess", "Stefan Feuerriegel" ]

[ "Causality machine learning", "Average treatment effects", "Confidence intervals", "Prediction-powered inference" ]

{ "note_id": [ "zkN1em9vTr", "zIzF10tEzt", "xu0EJQJUee", "sVZPgpn0q2", "pcS8uyIscj", "lfxo4otAAq", "lKqMERpWMR", "lC8ZSB0U58", "haOf86XIqJ", "hB6CdjYPne", "g6xAWCmKyQ", "e7Q8pOHl76", "VRoNb2eSSY", "VD2JeomPbJ", "Ugg1flY3ru", "TeaiAzhkKl", "PhuobrSwSx", "PBysBI1918", "OHa90Cp0Ju", "J0V3TCvTT6", "A1fynplnK6", "A0ndC32dX1", "9SbfAYAmy5", "9GWJAffCdY", "8wP8r2hsNm", "8pHf9y81pv", "4y9P5S4qzJ", "2aSvKgKP3i" ], "note_type": [ "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "decision", "official_comment", "official_comment", "official_review", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "meta_review", "official_comment", "official_review", "official_comment" ], "note_created": [ 1732159452353, 1732215332989, 1732159206590, 1732698765251, 1732158243457, 1732158074527, 1737523558842, 1732804372915, 1732695030193, 1729156368100, 1732300982828, 1729176855091, 1732803838837, 1732743642413, 1732157811047, 1732159350390, 1732158893179, 1732158633388, 1732628863712, 1732642130979, 1732530901080, 1732643142203, 1732717942815, 1730653684716, 1734645182073, 1732440822216, 1730658983162, 1732541288314 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission3148/Authors" ], [ "ICLR.cc/2025/Conference/Submission3148/Reviewer_sim7" ], [ "ICLR.cc/2025/Conference/Submission3148/Authors" ], [ "ICLR.cc/2025/Conference/Submission3148/Authors" ], [ "ICLR.cc/2025/Conference/Submission3148/Authors" ], [ "ICLR.cc/2025/Conference/Submission3148/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission3148/Authors" ], [ "ICLR.cc/2025/Conference/Submission3148/Reviewer_frUm" ], [ "ICLR.cc/2025/Conference/Submission3148/Reviewer_pijF" ], [ "ICLR.cc/2025/Conference/Submission3148/Authors" ], [ "ICLR.cc/2025/Conference/Submission3148/Reviewer_sim7" ], [ "ICLR.cc/2025/Conference/Submission3148/Reviewer_sim7" ], [ "ICLR.cc/2025/Conference/Submission3148/Authors" ], [ "ICLR.cc/2025/Conference/Submission3148/Authors" ], [ "ICLR.cc/2025/Conference/Submission3148/Authors" ], [ "ICLR.cc/2025/Conference/Submission3148/Authors" ], [ "ICLR.cc/2025/Conference/Submission3148/Authors" ], [ "ICLR.cc/2025/Conference/Submission3148/Reviewer_k7Su" ], [ "ICLR.cc/2025/Conference/Submission3148/Authors" ], [ "ICLR.cc/2025/Conference/Submission3148/Authors" ], [ "ICLR.cc/2025/Conference/Submission3148/Authors" ], [ "ICLR.cc/2025/Conference/Submission3148/Reviewer_sim7" ], [ "ICLR.cc/2025/Conference/Submission3148/Reviewer_frUm" ], [ "ICLR.cc/2025/Conference/Submission3148/Area_Chair_hRUD" ], [ "ICLR.cc/2025/Conference/Submission3148/Reviewer_pijF" ], [ "ICLR.cc/2025/Conference/Submission3148/Reviewer_k7Su" ], [ "ICLR.cc/2025/Conference/Submission3148/Reviewer_sim7" ] ], "structured_content_str": [ "{\"title\": \"Response to all reviewers\", \"comment\": \"Thank you very much for the constructive evaluation of our paper and your helpful comments\\\\! We addressed all of them in the comments below. We also uploaded our **revised version of the paper**.\", \"our_main_improvements_are_the_following\": [\"**Additional theoretical analysis:** We added a very detailed and technical illustration in our **new Appendix H** showing the key differences between our method and the work from van der Laan et al., 2024\\\\. The method by van der Laan et al involves the matrix inversion problem, which is one of the key steps while reweighting the canonical gradient of the beta-component which is important since ATMLE needs the canonical gradient to iteratively estimate the targeted estimand. This is the reason why their method is unstable and often breaks. We provide further details about the technical problems of the method as well as an in-depth comparison in our **new Appendix H**.\", \"**Additional experimental results** We expanded our experimental results in our **new** **Appendix G.** Specifically, we performed the following new experiments:\", \"**New neural instantiations of our method:** We instantiated our method with neural networks. Thereby, we show how our method contributes to representation learning and demonstrate that our method can effectively handle neural base methods (see our **new** **Appendix G.1**).\", \"**New base learners:** We replaced the regression model for nuisance parameters from linear regression to XGBoost (see our **new** **Appendix G.2**). Thereby, we show the flexibility of our method to other base learners.\", \"**High-dimensional covariates:** We generated a synthetic dataset with high-dimensional covariates. This allows us to study how our method behaves in settings with increasing complexity (see **our new Appendix G.3**).\", \"**Strength of dependence:** We generated an additional covariate as the mean of all other covariates to construct the collinearity in input space. That is, we here examine how our method when the dependence on the covariates varies (see our **new** **Appendix G.4**).\", \"**Different strengths of (un)confounding in $\\\\\\\\mathcal{D}\\\\_1$:** We relaxed the \\u201cUnconfoundedness\\u201d assumption with three different confounding scenarios while fixing the confounding in $\\\\\\\\mathcal{D}\\\\_2$. Thereby, we can study the performance of our method in settings with different strengths of (un)confounding (see our **new** **Appendix G.5**).\", \"**AIPW to RCT+obs datasets and A-TMLE to observational datasets:** We apply the AIPW to the RCT dataset (see our **new** **Appendix G.6**) and A-TMLE to the observational datasets (see our **new** **Appendix G.7**). Thereby, we show the robustness of our method to RCT and observational datasets.\", \"**Increasing sample size in $\\\\\\\\mathcal{D}\\\\_1$:** We varied the size of the small dataset $\\\\\\\\mathcal{D}\\\\_1$ from 100 to 2500\\\\. Thereby, we further assess the role of the size of $\\\\\\\\mathcal{D}\\\\_1$ (see our **new** **Appendix G.8**).\", \"We find that **our method is always better than the na{\\\\\\\\\\u201di}ve method under all settings**. This confirms our theoretical contributions and verifies the effectiveness of our method.\", \"We incorporated all changes (indicated with the label **Action** in our individual responses) into the revised version of our paper and highlighted all key changes in **blue color**. Given these improvements, we are confident that our paper provides valuable contributions to the causal machine learning literature and is a good fit for ICLR 2025\\\\.\"]}", "{\"comment\": \"Thank you for answering most of my comments. I still have some doubts about the following points:\\n- at the bottom of page 6, you mention that $\\\\hat{\\\\tau}_2$ satisfies a CLT. But $\\\\hat{\\\\tau}_2$ is estimated and evaluated on the same data set. A CLT follows directly if $\\\\hat{\\\\tau}_2$ is considered as fixed, which is not the case here. This deserves more explanations. Besides, $\\\\tau_1$ should be $\\\\hat{\\\\tau}_1$ and the definition of this quantity is not introduced before the bottom of page 6 (it is introduced in the sequel).\\n- In Theorem 4.2, the rate of convergence should be on $1/\\\\hat{\\\\pi}(x)$ and not on $\\\\pi(x)$, the type of convergence (in probability) must be stated. Besides, if I am not mistaken, the proof of Wager (2024) holds for cross-fitted estimators, which is mentioned at all in Theorem 4.2. \\n- I do not worry about the bias of $$\\\\hat{\\\\Delta}_{\\\\tau}$$\\n\\nbut I wonder how you can apply a CLT to \\n\\n$\\\\hat{\\\\Delta}_{\\\\tau} = \\\\frac{1}{n} \\\\sum_{i=1}^n (\\\\tilde{Y}_{\\\\hat{\\\\eta}}(x_i) - \\\\hat{\\\\tau}_2(x_i))$.\\n\\nRemark 4.1 support the fact that a CLT can be applied to the first part, while a CLT holds for the second part, as the evaluation of $\\\\hat{\\\\tau}_2$ is performed on $D_1$. But it remains to show that the two quantities are independent to obtain (i) an asymptotic Gaussian distribution (ii) with the correct variance. Did I miss something?\\n\\nBesides, thank you for the additional simulations. I still think that $5$ variables is not sufficient to highlight the benefits of the method. I definitely would not say that this is a \\\"high-dimensional covariate space\\\" as stated in appendix G.3 or in your answer. Nevertheless, I appreciate your work on this rebuttal.\"}", "{\"title\": \"Response to Reviewer pijF\", \"comment\": \"Thank you for your helpful review\\\\! We took all your comments at heart and improved our paper as follows. We thus **uploaded a new PDF** where we highlighted all major changes highlighted in **blue color**.\\n\\n### **Response to \\u201cWeakness\\u201d**\\n\\n**(W1) Limited empirical results:** \\nThanks for giving us the chance to enrich the experiment results. We give a detailed overview in our reply to question Q1 below. \\n\\n**Action:** We **added various new experiments** (see our **new Appendix G**).\\n\\n**(W2&3) Practical relevance and limitation of our work:** \\nThank you for giving us the chance to explain why our setting is standard and thus general. **First,** the assumptions are commonly referred to as the \\u2018standard\\u2019 assumptions valid inference of the average treatment effect \\\\[3\\\\]. **Second,** our assumptions are in line with prior works on making causal inferences from multiple datasets \\\\[2, 5\\\\]. In fact, **we make even weaker assumptions** by saying that the dataset D\\\\_2 can have unobserved confounding. **Third,** there are many real-world applications in medicine and public policy where the assumptions are met by design. For example, the efficacy of drugs in post-approval use is typically estimated from a large observational set consisting of electronic health records (=which can have unobserved confounders and thus match our dataset D\\\\_2). Further, any regulatory approval of drugs involves a small RCT (=which matches our assumptions for D\\\\_1). Similarly, many interventions in public policy involve similar settings with small-scale RCT and a large-scale observational dataset. For example, such settings are common in the development sector where the effect of interventions such as having access to microloans is evaluated on people\\u2019s lives in poor countries \\\\[1\\\\].\\n\\nWe give a more in-depth explanation in our response to Questions Q2 and Q3. \\n\\n**Action:** We explained at greater length why our assumptions are standard, why our method makes even weaker assumptions than other methods for our task, and why and where our assumptions are met in clinical practice and policy. For the latter, we also added state further concrete examples from practice.\\n\\n### **Response to \\u201cQuestions\\u201d**\\n\\n**(Q1) Empirical results:** \\nThanks for giving us the chance to enrich the experiment results. In the following, we address the three concerns of the experiments with more experimental results:\\n\\n* *Balance of bias and variance:* We followed your suggestion and reported the RMSE and the coverage in our synthetic datasets experiments. We see that our method performs again best. \\n This can be expected based on our theory. The reason is the following. Following our assumptions, we assume the large dataset without unconfoundedness which means that it may lead to a biased estimator. However theoretically, according to our proof in **Appendix A.2**, our proposed prediction-powered estimator provides a valid CI for the ATE in $\\\\\\\\mathcal{D}^1$. In short, it is an unbiased estimation , i.e. $E(\\\\\\\\hat{\\\\\\\\Delta}\\\\_{\\\\\\\\tau} \\\\+ \\\\\\\\hat{\\\\\\\\tau}\\\\_2) \\\\= \\\\\\\\tau$. This thus explains why our method is so effective. \\n\\n| Method | RMSE | Coverage |\\n| :---- | :---- | :---- |\\n| $\\\\\\\\hat{\\\\\\\\tau}^{\\\\\\\\mathrm{AIPW}}$ ($\\\\\\\\mathcal{D}^1$ only) | 0.298/0.298/0.298 | 0.424/0.424/0.424 |\\n| $\\\\\\\\hat{\\\\\\\\tau}^{\\\\\\\\mathrm{AIPW}}$ ($\\\\\\\\mathcal{D}^2$ only) | 0.442/0.478/0.476 | 0.217/0.144/0.131 |\\n| \\\"$\\\\\\\\hat{\\\\\\\\tau}^{\\\\\\\\mathrm{PP}}$ (Ours) | **0.276/0.274/0.271** | **0.241/0.240/0.237** |\\n\\n* *Increasing sample size of $\\\\\\\\mathcal{D}\\\\_1$:* Thanks. We continually increase the sample size of $\\\\\\\\mathcal{D}\\\\_1$ until it almost achieves similar performance. Based on your suggestion, we improved the experiments and increased the sample size further to larger values (see our **new** **Appendix G.8**). The results are as expected and confirm our theory. \\n* *Comparison against the ground-truth ATE:* Thank you for pointing this out. For the real-world application, in our study, we utilize a semi-synthetic dataset, where we have knowledge of how the potential outcomes are generated. This allows us to directly evaluate the accuracy and validity of our method **by comparing the estimated ATE against the ground truth**. By leveraging the semi-synthetic nature of the dataset, we can strike a balance between real-world complexity and the ability to validate our results rigorously. Here, we see that our method is very accurate in learning the ground truth ATE. \\n* *Generalization to machine learning models:* We **conducted new experiments with XGBoost** (see our **new** **Appendix G.2**) and **with neural networks** (see **our new Appendix G.1**). Again, our method is highly effective (and the performance gain over the baseline becomes even larger\\\\!). \\n \\n **Action:** We added further clarifications for the points above, and we further **added the new experiments to our revised PDF**.\"}", "{\"comment\": \"Thank you for your positive feedback! We will incorporate all action points into our revised version of paper.\"}", "{\"title\": \"Response to Reviewer sim7\", \"comment\": \"Thank you for your careful review and helpful comments. As you can see below, we have carefully revised our paper along with your suggestions, added various new experiments, and added a comparison with the work of van der Laan et al., 2024 with detailed theoretical proof, and conducted more experiments. We thus **uploaded a new PDF** where we highlighted all major changes highlighted in **blue color**.\\n\\n### **Response to \\u201cWeakness\\u201d**\\n\\n**(W1) More experiments:** \\nThanks for giving us the chance to expand our experiment results. In the following, we **added new experiments** along your suggestions:\\n\\n1. *More input variables:* We **expanded our experiments** and now report results with more input variables. For this, we used a data-generating mechanism similar to that in the main paper but where we now generate multiple covariates (see our **new** **Appendix G.3**). The **results confirm our existing conclusions** as well as our theoretical contributions. In particular, the results show that our method is effective even for settings with multiple input variables. \\n2. *With or without dependence:* Here, we **added** the experiments while having 4 independent covariates and the 5-th covariate is the mean of the above 4 covariates and use the 5-th covariate as a component of generating the potential outcomes (see our **new** **Appendix G.4**). Again, **our method performs best**. \\n3. *Relaxing the \\u201cUnconfoundedness\\u201d assumption in $\\\\\\\\mathcal{D}\\\\_1$:* Finally, we also conducted the experiments while keeping the $\\\\\\\\mathcal{D}\\\\_2$ as the medium confounded scenario with three different confounding scenarios in $\\\\\\\\mathcal{D}\\\\_1$ (see our **new** **Appendix G.5**). The results again confirm our theoretical contributions and that **our method performs best**.\\n\\n**Action:** We **added several new experiments**: (i) with high-dimensional covariates, (ii) with and without dependence, and (iii) different confounding scenarios in $\\\\\\\\mathcal{D}\\\\_1$ (see our **new Appendix G**).\\n\\n**(W2) Specific theoretical support:** \\nOur main contribution is not the derivation of a new meta-learner but **proposed a novel way to construct valid and more accurate CIs** based on the good asymptotic property of the AIPW estimator in multiple observational datasets and IPW in RCT and observational datasets. We **still strongly believe that our paper fills an important gap in the literature and provides a method that is novel and highly relevant for practice (e.g., in medicine)**. As of now, existing methods for causal inference from multiple datasets have primarily relied on point estimates, while we shift the focus to uncertainty quantification. We also believe that our application of PPI in the context of causal inference from multiple datasets is new, thus changing the underlying way how causal inference from multiple datasets is made and thus will spur new avenues for follow-up research. \\n\\n**Action:** We carefully checked our paper and we spelled out our main novelty clearly: we present a novel way using AIPW estimators to construct CIs from multiple datasets with new theoretical guarantees (see **our Theorem 5.2 and Theorem 6.2**).\"}", "{\"title\": \"Response to Reviewer frUm\", \"comment\": \"Thank you for your helpful review\\\\! We took all your comments at heart and improved our paper as follows. We thus **uploaded a new PDF** where we highlighted all major changes highlighted in **blue color**.\\n\\n### **Response to \\u201cWeakness\\u201d**\\n\\n**(W1) Novelty of our method:** \\nThank you for giving us the opportunity to clarify our contributions and how our method is novel. Our main contribution is not the derivation of a new meta-learner but to **propose a novel way to construct valid and more accurate CIs** based on the good asymptotic property of the AIPW estimator in multiple observational datasets and IPW in RCT and observational datasets. Hence, our novelty is the way how we leverage the AIPW to estimate confidence intervals from **multiple observational datasets**. With the help of a large but confounded observational dataset and the idea of prediction-powered inference, we essentially \\u201cshrink\\u201d the CIs which means offering more precise uncertainty quantifications. As such, **we see our main contributions in Theorem 5.2 and Theorem 6.2**. \\n\\nOf note, our work is **not** just a simple application of the PPI framework. Rather, the PPI framework does not inform us how to choose the rectifier and how to obtain theoretical properties such as valid CIs. Rather, **this requires us to make new, tailored, and non-trivial derivations to confirm the theoretical properties (as provided by Theorem 5.2 and Theorem 6.2)**. Thereby, we provide new methods for making causal inference from multiple datasets that improve over existing baselines and achieve state-of-the-art performance.\\n\\n**Action:** We carefully checked our paper and we spelled out our main novelty clearly: we present a novel way using AIPW estimators to construct CIs from multiple datasets with new theoretical guarantees (see **our Theorem 5.2 and Theorem 6.2**). \\n\\n**(W2) Different of AIPW and IPW:** \\nThanks for your careful comments. We admit that we have made a wrong statement, **which we have now fixed**. The non-centered IF scores are different for IPW and AIPW, for IPW it should be $\\\\\\\\tilde{Y}\\\\_{\\\\\\\\pi}(x\\\\_i) \\\\= \\\\\\\\frac{A\\\\_i Y\\\\_i}{\\\\\\\\pi(x\\\\_i)} \\\\- \\\\\\\\frac{(1-A\\\\_i)Y\\\\_i}{\\\\\\\\pi(x\\\\_i)}$, and for AIPW, it is $\\\\\\\\tilde{Y}\\\\_{\\\\\\\\hat{\\\\\\\\eta}}(x\\\\_i) \\\\= \\\\\\\\left( \\\\\\\\frac{A}{\\\\\\\\hat{\\\\\\\\pi}(x\\\\_i)} \\\\- \\\\\\\\frac{1-A}{1-\\\\\\\\hat{\\\\\\\\pi}(x\\\\_i)} \\\\\\\\right) Y\\\\_i \\\\- \\\\\\\\frac{A \\\\- \\\\\\\\hat{\\\\\\\\pi}(x\\\\_i)}{\\\\\\\\hat{\\\\\\\\pi}(x\\\\_i) \\\\\\\\left(1-\\\\\\\\hat{\\\\\\\\pi}(x\\\\_i)\\\\\\\\right)} \\\\\\\\left\\\\[ \\\\\\\\left(1-\\\\\\\\hat{\\\\\\\\pi}(x\\\\_i)\\\\\\\\right)\\\\\\\\hat{\\\\\\\\mu}\\\\_1(x\\\\_i) \\\\+ \\\\\\\\hat{\\\\\\\\pi}(x\\\\_i)\\\\\\\\hat{\\\\\\\\mu}\\\\_0(x\\\\_i)\\\\\\\\right\\\\]$. To clarify, the key difference between **Sections 5 & 6** in our main paper is that we do not need to estimate the propensity score in the RCT dataset and only AIPW holds asymptotic normality in the observational dataset.\\n\\n**Action:** We corrected the statement in our revised paper. \\n\\n**(W3) Asymptotically of $\\\\\\\\hat{\\\\\\\\tau}\\\\_2$:** \\nThank you for this suggestion. We do not rely on the asymptotic normality of $\\\\\\\\tau\\\\_2$, as it directly follows from the central limit theorem (CLT). Specifically, given that $\\\\\\\\tau\\\\_2$ is derived from a sum of independent random variables, the CLT ensures its asymptotic normality under standard regularity conditions, i.e., $\\\\\\\\tau\\\\_2 \\\\\\\\overset{d}{\\\\\\\\to} \\\\\\\\mathcal{N}(\\\\\\\\mu\\\\_2, \\\\\\\\sigma\\\\_2^2), \\\\\\\\text{ as } n \\\\\\\\to \\\\\\\\infty$, where $\\\\\\\\mu\\\\_2$ and $\\\\\\\\sigma\\\\_2^2$ denote the mean and variance of $\\\\\\\\tau\\\\_2$, respectively. The key focus of our analysis lies in the asymptotic normality of $\\\\\\\\tau\\\\_1$. In the observational confounded dataset, only when we apply the average non-centered IF scores from the AIPW as the estimation of ATE the $\\\\\\\\hat{\\\\\\\\tau}\\\\_2$ has the asymptotic normality to the oracle ATE in the population. After that, we can state: $\\\\\\\\tau\\\\_1 \\\\\\\\overset{d}{\\\\\\\\to} \\\\\\\\mathcal{N}(\\\\\\\\mu\\\\_1, \\\\\\\\sigma\\\\_1^2), \\\\\\\\text{ as } n \\\\\\\\to \\\\\\\\infty$, where $\\\\\\\\mu\\\\_1$ and $\\\\\\\\sigma\\\\_1^2$ represent the mean and variance of $\\\\\\\\tau\\\\_1$.\\n\\n**Action:** We added a new section to explain the asymptotic properties of $\\\\tau\\\\_2$ (see **our new Appendix A.3**).\\n\\n### **Response to \\u201cQuestions\\u201d**\\n\\nThank you for your feedback. A particular strength of **our method is that it is flexible** and that **it can be used with various base learners including neural networks or other representation learning methods**. Upon reading your question, we realized that we show the applicability of our method also to neural learning, and we thus performed new experiments where we demonstrate when our method is applied on top of neural networks. Again, we find that our proposed method is highly effective.\\n\\n**Action:** We **performed new experiments with neural networks** to show how our method makes contributions to representation learning (see **our new Appendix G.1**).\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}", "{\"comment\": \"Thanks so much for your detailed and positive suggestions and feedback! We will incorporate all action points into our revised version of paper!\"}", "{\"comment\": \"The reviewer acknowledged that the concerns are addressed and increased the score accordingly.\"}", "{\"summary\": \"The authors present a method, based on prediction-powered inference, for estimating confidence intervals from two datasets, one with observed confounders and one with additionally hidden confounders. Theoretical and empirical results are provided.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": [\"The work addresses a clear research gap with respect to existing work.\", \"The provided theoretical results support the method.\", \"The method is presented clearly, and the work is easy to follow.\"], \"weaknesses\": [\"The current empirical analysis is limited, with mostly quantitative results and only linear models.\", \"Due to the required assumptions, the practical relevance of the work is not entirely clear to me.\", \"The conclusion is an afterthought in the current version, with no discussion of limitations or directions for future work.\"], \"questions\": \"1. **Empirical results:**\\n - Mostly qualitative results are provided for the synthetic data. I would be interested in seeing the coverage and RMSE of each method with respect to ground truth, which should be known as the data is simulated. It seems like the proposed method has a higher bias, but lower variance.\\n - In my opinion, it would be interesting to experiment with settings with more data in $\\\\mathcal{D}_1$ to see when only using $\\\\mathcal{D}_1$ becomes better.\\n - How can the RMSE be calculated for the real data, when the true ATE is unknown? \\n - Unless I am mistaken, it seems that only simple learners are considered, based on linear and logistic regression (Appendix E3)? If so, it would be insightful to also compare with more advanced ML algorithms (e.g. gradient boosting).\\n\\n2. **Practical significance:**\\nAlthough the work is interesting from a theoretical perspective, I wonder at its practical significance. How realistic are your assumptions? How can a practitioner verify them and trust your method?\\n\\n3. **Conclusion:** A more detailed discussion of limitations and directions for future work would be appreciated.\\n\\n___\\n\\n\\n**Minor points:**\\n- Brackets missing in the citations on lines 48 and 49.\\n- Gray text in Table 1 is not applied to all columns.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you for the fast and helpful response to our rebuttal and for enabling a constructive discussion. We are glad that our rebuttal addressed some of your concerns, and **we are confident that we can address the remaining ones below**. Again, **we have updated our PDF** and highlighted the new materials in **red color** (to show our additions from the previous edits that were in blue color).\\n\\n### **Response to \\u201cTheory\\u201d:**\\n\\n1. Thanks for your valuable question here, and you are correct. We forgot to add the important details that we **performed sample splitting** and thus split **$\\\\\\\\mathcal{D\\\\_2}$ into two independent datasets**, one for training the fixed function $\\\\\\\\hat{\\\\\\\\tau}_2(\\\\\\\\cdot)$, and the other one for evaluating $\\\\\\\\hat{\\\\\\\\tau}_2(\\\\\\\\cdot)$ and getting the CATE estimations. Of note, we do it in our experiments rightly. **In our experiments, we split $\\\\\\\\mathcal{D\\\\_2}$ into two independent datasets**, denoted as $\\\\\\\\mathcal{D\\\\_2^1}$ and $\\\\\\\\mathcal{D\\\\_2^2}$. We first trained our CATE estimator (e.g., DR-learner), in $\\\\\\\\mathcal{D\\\\_2^1}$ as a fixed function $\\\\\\\\hat{\\\\\\\\tau}_2(\\\\\\\\cdot)$. Next, we apply $\\\\\\\\hat{\\\\\\\\tau}_2(\\\\\\\\cdot)$ to the $\\\\\\\\mathcal{D\\\\_2^2}$ and computed the average over $\\\\\\\\mathcal{D\\\\_2^2}$ as $\\\\\\\\hat{\\\\\\\\tau}\\\\_2 \\\\= \\\\\\\\frac{1}{N}\\\\\\\\sum\\\\_{i=1}^N\\\\\\\\hat{\\\\\\\\tau}\\\\_2(x\\\\_i)$. Since the $\\\\\\\\hat{\\\\\\\\tau}_2(\\\\\\\\cdot)$ is trained and evaluated on two **independent datasets**, $\\\\\\\\hat{\\\\\\\\tau}\\\\_2$ satisfies a CLT. Sorry for the imprecision in Our paper.\\n\\n**Action:** We corrected this in our revised paper and state that we used sample splitting to train and evaluating the $\\\\\\\\hat{\\\\\\\\tau}\\\\_2$.\\n\\n2. We appreciate your suggestions. We **corrected the notation and emphasized our use of cross-fitting** in Theorem 4.2. Then, following Wager\\u2019s book, $\\\\\\\\hat{\\\\\\\\tau}^{\\\\\\\\mathrm{AIPW}$ is asymptotically normally distributed to the oracle ATE with corrected variance. \\n \\n\\n**Action:** We revised this in our paper. \\n\\n3. Regarding the asymptotic normality of $\\\\\\\\hat{\\\\\\\\Delta}_{\\\\\\\\tau}$, it holds because $\\\\\\\\mathcal{D\\\\_1}$ and $\\\\\\\\mathcal{D}\\\\_2^1$ are two **independent datasets**. \\n \\n Follow the $\\\\\\\\hat{\\\\\\\\Delta}_{\\\\\\\\tau} \\\\= \\\\\\\\frac{1}{n} \\\\\\\\sum{i=1}^n (\\\\\\\\tilde{Y}\\\\_{\\\\\\\\hat{\\\\\\\\eta}}(x\\\\_i) \\\\- \\\\\\\\hat{\\\\\\\\tau}\\\\_2(x\\\\_i))$, where $\\\\\\\\tilde{Y}\\\\_{\\\\\\\\hat{\\\\\\\\eta}}(x\\\\_i)$ are non-centered influence function score we estimated with cross-fitting AIPW estimation and $\\\\\\\\hat{\\\\\\\\tau}\\\\_2(x\\\\_i))$ are estimated on $\\\\\\\\mathcal{D}\\\\_2^1$ with trained fixed function $\\\\\\\\hat{\\\\\\\\tau}_2(\\\\\\\\cdot)$ on $\\\\\\\\mathcal{D}\\\\_2^1$. \\n \\n Furthermore, as the nuisance functions of pseudo-outcomes and $\\\\\\\\hat{\\\\\\\\tau}\\\\_2$ are trained on independent datasets, these two components are indeed independent. Since the CLT applies to the second term of $\\\\\\\\hat{\\\\\\\\Delta}\\\\_{\\\\\\\\tau}$ and $\\\\\\\\hat{Y}\\\\_{\\\\\\\\tilde{\\\\\\\\eta}}(x)$ and $\\\\\\\\hat{\\\\\\\\Delta}\\\\_{\\\\\\\\tau}$ are independent, then the asymptotical normality holds for $\\\\\\\\hat{\\\\\\\\Delta}\\\\_{\\\\\\\\tau}$. \\n \\n\\n**Action:** We **fixed the notation** in Theorem 4.2 and clarified that we need the **dataset splitting** for training and evaluating of $\\\\\\\\hat{\\\\\\\\tau}\\\\_2$.\\n\\n### **Response to \\u201cExperiments\\u201d:**\\n\\nThank you. We followed your suggestion and **expanded our experiments** further. We now report results with 50 and 500 input variables (see our **new** **Appendix G.3**). The **results confirm our existing conclusions** and theoretical contributions. In particular, the results show that our method is effective even for settings with higher dimensional input variables.\\n\\n**Action:** We followed your suggestion, and we **added experiments to show our effectiveness** for high-dimensional input variables.\"}", "{\"summary\": \"The authors propose a new method to compute confidence intervals of the Average Treatment Effect (ATE) of the Risk Difference based on two observational data sets, the first one of small size in which all confounders are observed, the second one being larger with potential unobserved confounders. They use the largest data set to debias the estimation of the ATE computed on the first data set. The proposed confidence interval is shown to have the correct asymptotic coverage. Experiments on a toy data set and on real-world data sets show that the proposed method leads to smaller unbiased confidence intervals.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"The paper focuses on a very important topic, which is uncertainty quantification of point estimate for the ATE, based on observational data sets. Obtaining accurate confidence intervals for the ATE helps to understand the effect of a treatment (e.g., drugs). Observational data are becoming more and more common in causal inference. The main contribution of the paper is to propose a new way of computing confidence intervals based on two observational data sets. The formula is rather simple and easy to use, while it can be applied to a wide variety of estimators (in the paper, AIPW or IPW are used depending on the context).\\nExperiments tend to show that the proposed method is better in terms of coverage and IC length.\", \"weaknesses\": \"There is a potential practical interest for the confidence interval and the point estimate developed in this paper. I have two majors concerns with the paper:\\n- the experimental results on the toy data set are inconclusive : the input variable is of dimension one, which does not correspond to a real-world problem. For the conclusion to hold, I would expect more experiments on more complex settings (more input variables, with or without dependence, how relaxing the assumption of all observed confounders in the first data set may impact the estimation performances). Experimenting on one single input variable does not allow us to obtain a clear comprehension of the performance of the proposed method. \\n- Theorem 5.2 is the main theoretical result of the paper. It is claimed below that it does not directly result from the Prediction-powered Inference (PPI) framework. However, going in detail through the proof, I could not find an argument specific to the causal inference framework or the AIPW estimator. The proof relies on Central Limit Theorems of two independent quantities.\\nThus, I find the novelty of the contribution mild from a theoretical perspective. Besides, in my opinion, more experiments should be carried out to assess the performance of the method.\", \"questions\": [\"Position of the work: it is assumed that the small data set contains all confounders. The major difference with a RCT is that the propensity score is unknown and depends on the covariate (as discussed in page 4, last paragraph). This justifies the use of IPW with the true propensity score (step B, page 7). However, numerous work have shown that even in a RCT scenario where the true propensity score is known, estimating it can reduce the variance of the ATE estimate (see, e.g., references in https://arxiv.org/pdf/2303.17102). Thus, in a RCT, one would likely use an estimation of the propensity score inside the IPW estimate. This point should be discussed and may lead to qualify statements about differences between RCT and the small data set you consider. In particular, could we apply methods developed for RCT (van der Laan et al., 2024) to the first experimental setting? Paragraph l.119-121 could be modified to make explicit the differences (if any) between RCT+observational and the considered setting observational+observational.\", \"Regarding the work of van der Laan et al., 2024, it is mentioned that their method is unstable due to a matrix inversion (l.509-510). I went quickly through their paper and did not find such an operation. Could you be more specific, and maybe describe with more details how this method works, as it is the main competitor?\", \"l.151-152 ``PP CI s smaller than the classical CI when the model $f$ is sufficiently accurate.'' Is there a specific reference for this fact?\", \"l.203-204: It is usually required that $0< \\\\pi(x) < 1$.\", \"l.269: \\\"The estimation needs to be both valid and unbiased to later yield valid CIs.'' What does ``valid'' mean in these contexts?\", \"Lemma 5.1 Assumptions are not complete/clear: there are constraints on $\\\\alpha_{\\\\mu}$ and $\\\\alpha_{\\\\pi}$, the assumption on $e(\\\\cdot)$ is about $1/\\\\hat{e}(\\\\cdot)$ and not $e$ and the proof of Wager holds for crossfitted estimators. The asymptotic variance involves theoretical quantities and not their estimators. Please correct the statement. Some remarks also hold for Theorem 5.2.\", \"l.298 $\\\\tilde{Y}_{\\\\hat{\\\\eta}}$ is built using $D_1$ and then its empirical mean (subtracted by $\\\\hat{\\\\tau}_2$) is computed over\", \"$D_1$.\", \"In the proof, it is argued that $\\\\hat{\\\\Delta}_{\\\\tau}$ verifies a Central Limit Theorem which is not clear here, as the term inside the sum depends on the whole sample $\\\\mathcal{D}_1$.\", \"I believe the proof holds for a fixed function $\\\\tilde{Y}_{\\\\eta}$ but not for the estimated function. Or it needs at least clarifications.\", \"Equation 5, how do this new confidence interval compares to a classical one (using only $\\\\mathcal{D}_1$). Can you show that it is stricly smaller in some specific settings (and describe them)?\", \"Lemma 6.1 A factor $1/n$ is missing in the definition of $\\\\hat{\\\\tau}_1$, $\\\\hat{\\\\sigma}_1^2$ cannot depend on the sample size $n$. Its expression can be made explicit as a function of the nuisance components. In the proof, $\\\\hat{\\\\pi}$ should be replaced by $\\\\pi$.\", \"l488 In the real-world application, how the RMSE is computed, since we do not have access to the true ATE value? Please define ``factual outcome''.\", \"l.499 ``Takeaway: Our PPI-based method is effective for medical data.'' Please qualify this statement, as you evaluate your method on two datasets only.\"], \"typos\": [\"open parenthesis l.124\", \"l.162 : a for is missing\", \"l.177 and l.334 : $p$ is used for the number of input variables and the ratio between the data set sizes.\", \"l.363 ``This steps computes analogous the above''\", \"l.409 ``We now evaluate our the effectiveness''\", \"l469 ``Hence, our method performs thus best.''\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you for your detailed response. I am then happy to improve my score.\"}", "{\"comment\": \"Thanks for your further question. We are happy to answer this and replace the informal statement with a formal derivation (see **our revised PDF**).\\n\\nWe provide a short summary below. For this, we decompose the rectifier into two parts\\n\\n$$\\\\\\\\hat{\\\\\\\\Delta}\\\\_\\\\\\\\tau \\\\= \\\\\\\\underbrace{\\\\\\\\frac{1}{n}\\\\\\\\sum\\\\_{i=1}^n \\\\\\\\left\\\\[\\\\\\\\tilde{Y}\\\\_{\\\\\\\\hat{\\\\\\\\eta}}(x\\\\_i) \\\\- Y\\\\_{\\\\\\\\eta}(x\\\\_i)\\\\\\\\right\\\\]}\\\\_{\\\\\\\\text{Error due to nuisance estimation}} \\\\+ \\\\\\\\underbrace{\\\\\\\\frac{1}{n}\\\\\\\\sum\\\\_{i=1}^n \\\\\\\\left\\\\[Y\\\\_{\\\\\\\\eta}(x\\\\_i) \\\\- \\\\\\\\hat{\\\\\\\\tau}\\\\_2(x\\\\_i)\\\\\\\\right\\\\]}\\\\_{\\\\\\\\text{Oracle nuisance functions}}.$$\\n\\nThe first term denotes **the error introduced by using estimated nuisance functions**. Following Wager\\u2019s book, it is **negligible in probability** if we apply the cross-fitting for estimated nuisance functions and the estimated nuisance functions satisfy the converge rate requirement as we stated in Theorem 4.2. (We provide detailed steps for cross-fitting in our revised **Algorithm 1** and our new supporting lemma from **Appendix A.1**)\\n\\nThe second term denotes the mean of differences between pseudo-outcomes, i.e., $Y\\\\_\\\\eta(X)$ based on the oracle nuisance function (**Here, nothing trained on** $\\\\\\\\mathcal{D}^1$ and $\\\\\\\\mathcal{D}^2$) and $\\\\\\\\hat{\\\\\\\\tau}\\\\_2(X)$ which is the CATE estimator trained on $\\\\\\\\mathcal{D}^{2}$ (**particularly, this is independent of** $\\\\mathcal{D}^1$). Thus, $Y\\\\_\\\\\\\\eta(X)$ and $\\\\\\\\hat{\\\\\\\\tau}\\\\_2(X)$ are independent functions, which means that $\\\\mathbf{Y\\\\_\\\\\\\\eta(x\\\\_i) \\\\- \\\\\\\\hat{\\\\\\\\tau}\\\\_2(x\\\\_i)}$ are i.i.d. random variables. Then, the second term follows the CLT.\\n\\n**Actions:**\\n\\n- We **revised our Algorithm 1** to clarify that we use cross-fitting on $\\\\\\\\mathcal{D}^1$ and sample-splitting on $\\\\\\\\mathcal{D}^2$. Specifically, we added a new statement (in Line 1) where we apply sample splitting (e.g., $\\\\\\\\mathcal{D}^{1}$ is split into $\\\\\\\\mathcal{D}^{1,1}$ and $\\\\\\\\mathcal{D}^{1,2}$). We then explicitly state which of the subsequent lines is trained on which split (e.g., Line 2 is estimated on $\\\\\\\\mathcal{D}^{2,1}$).\\n- We **replaced our informal statements in our proof in Appendix A.2 with a formal derivation**. For this, we updated Appendix A.2, which now makes use of a new supporting lemma that we provide in our new Appendix A.1.\\n\\nWe deeply appreciate your detailed suggestions that helped us to make our statements more formal and make it easier for readers from different backgrounds (e.g., PPI, causal inference, etc.) to follow our work. Thank you!\"}", "{\"title\": \"Response to Reviewer k7Su\", \"comment\": \"Thank you for your positive evaluation of our paper\\\\! We took all your comments at heart and improved our paper accordingly. We thus **uploaded a new PDF** where we highlighted all major changes highlighted in **blue color**.\\n\\n### **Response to \\u201cWeakness\\u201d**\\n\\n**(W1) Comparison of the width of CIs and coverages:** \\nThank you for your thorough suggestions about the comparison of the width of resulting CIs. We apologize for the unclear in our main paper. To clarify, the right subfigures in **Figures 4 & 5 in the main paper** illustrate the comparison of the widths of CIs. From those figures, we show that the widths of proposed CIs are significantly shorter than the naive method. This **shows that our method successfully shrinks the width of CIs as desired**. Also, following our proof in **Appendix A.2,** we show that **our method provides valid $1-\\\\\\\\alpha\\\\\\\\%$ confidence intervals**. \\n\\n**Action:** We improved our descriptions of Figures 4 and 5 to explain that these show the width and that our method shrinks the CI width as desired.\\n\\n**(W2) Comparison to the naive method:** \\nThank you for your detailed comments. To clarify, the yellow line in the experimental plots represents the width of the confidence intervals produced by the naive method, i.e., $\\\\\\\\hat{\\\\\\\\tau}^{\\\\\\\\mathrm{AIPW}}$, ($\\\\\\\\mathcal{D}^1$ only). Note that the naive method is exactly what you suggested, namely, using the unconfounded dataset only. Hence, our plots allow us to empirically validate that our method is **better** than the naive way of using only the unconfounded dataset (see the yellow lines in Figures 4, 5, etc.). \\n\\nWe further explain theoretically why our method is better than your suggested way of using only the unconfounded dataset. We added a new paragraph to our paper where we answer the question \\u201cWhy is our method better than using the unconfounded dataset only?\\u201d (see our response to your question below) \\n\\nFinally, we realized upon reading your question that we could offer a more comprehensive evaluation of our method under different levels of confounding of $\\\\\\\\mathcal{D}^1$. This allows us to better understand the source of gain. Hence, we performed new experiments for this (see **our new Appendix G.5**). First, our method is robust and performs best in all settings. Second, we see that the source of gain for our method is larger for settings with high confounding, which can be expected. \\n\\n**Action:** We clarified that the naive method in our paper matches your proposed idea of using only the unconfounded data and making it clearer in the revised paper. We added new experiments with different levels of confoundedness (see our **new** **Appendix G.5)**\\n\\n### **Response to \\u201cQuestions\\u201d**\\n\\nThank you. We added a new paragraph to our paper where we answer the question \\u201cWhy is our method better than using the unconfounded dataset only?\\u201d Indeed, we can even show that our method is almost always better. The reason is the following. As the measure of fit $\\\\\\\\hat{\\\\\\\\tau}\\\\_2$ is sufficiently accurate, the rectifier is almost equal to zero, i.e., $\\\\\\\\hat{\\\\\\\\Delta} \\\\\\\\approx 0$. Then, the variance of the rectifier is significantly smaller than the variance of estimated non-centered IF scores, $\\\\\\\\hat{\\\\\\\\sigma}\\\\_{\\\\\\\\Delta}^2 \\\\\\\\leq \\\\\\\\hat{\\\\\\\\sigma}_{\\\\tau_2}^2$. Given the large size of $\\\\\\\\mathcal{D}\\\\_2$, the variance of the estimated conditional treatment effect goes to zero, since the estimated variance should be divided by the sample size of $\\\\\\\\mathcal{D}\\\\_2$, i.e., $N$. As a result, the variance (and thus the CI width) is smaller when using our method than when using only the unconfounded dataset, which highlights the strengths of our proposed method from a theoretical perspective.\\n\\n**Action:** We **added the above explanation to our paper** as part of a new paragraph \\u201cWhy is our method better than using the unconfounded dataset only?\\u201d (see our new elaborations in our **revised Section 5\\\\)**.\"}", "{\"comment\": \"**(Q2) Practical significance:**\\nThanks so much for your question about the practical application of our work. Estimating the confidence intervals for ATEs is a relevant question in many fields such as medicine. We also would like to emphasize that **our assumptions are realistic in practice**. In particular, we followed the standard assumptions for estimating the ATE from observational datasets in existing literature, such as Imbens 2004 \\\\[4\\\\]; Rubin 2006 \\\\[6\\\\]; Shalit et al., 2017 \\\\[7\\\\]. **Our work is thus consistent.** \\n\\nThere are also papers that consider multiple datasets (see the overview in Table 1 of our revised paper). However, they always assume the combination of RCT and observational datasets \\\\[2, 5\\\\]. Hence, these methods make stronger assumptions than ours. In contrast, our method makes weaker assumptions. Our method proposed constructing valid confidence intervals with more released assumptions, i.e., multiple observational datasets, where we relax the assumption of unobserved confounding.\\n\\n**Action:** We explained at greater length why our assumptions are standard, why our method makes even weaker assumptions than other methods for our task, and why and where our assumptions are met in clinical practice and policy. For the latter, we also added state further concrete examples from practice.\\n\\n**(Q3) Discussion about limitations and direction of future work:** \\nThank you for the suggestion. We **added a discussion of limitations and directions for future work to our conclusion section**. For example, one avenue for improvement is to extend our method to other estimands such as CATE. Additionally, replacing the $\\\\\\\\tau\\\\_2$ model with a pre-trained large language model (LLM) could offer new applications for our method and thus make it more flexible. Similarly, one could also tailor our method to representations with text to make inferences from natural language more effective. Nevertheless, as with any other paper in causal inference, certain but standard assumptions should be fulfilled, so we encourage careful and responsible use in practice.\\n\\n**Action:** We added a discussion of limitations and directions for future work to our conclusion section.\\n\\n**(Q4) Minor points:** \\nThank you for pointing out the typos. \\n\\n**Action:** We have corrected all the typos.\\n\\n**References:** \\n\\\\[1\\\\] Abhijit Banerjee, Esther Duflo, Rachel Glennerster, and Cynthia Kinnan. The Miracle of Microfinance? Evidence from a Randomized Evaluation. ISSN 1945-7782. \\n\\\\[2\\\\] Ilker Demirel, Ahmed Alaa, Anthony Philippakis, and David Sontag. Prediction-powered Generalization of Causal Inferences, 6 2024\\\\. \\n\\\\[3\\\\] Phillip Heiler and Ekaterina Kazak. Valid inference for treatment effect parameters under irregular identification and many extreme propensity scores. 222(2):1083\\u20131108. ISSN 0304-4076. \\n\\\\[4\\\\] Guido W. Imbens. Nonparametric estimation of average treatment effects under exogeneity: A review. Review of Economics and Statistics, 86(1):4\\u201329, 2004\\\\. \\n\\\\[5\\\\] Nathan Kallus, Aahlad Manas Puli, and Uri Shalit. Removing Hidden Confounding by Experimental Grounding. Advances in neural information processing systems, 2018\\\\. \\n\\\\[6\\\\] Donald B. Rubin. Matched sampling for causal effects. Cambridge University Press, 2006\\\\. \\n\\\\[7\\\\] Uri Shalit, Fredrik D Johansson, and David Sontag. Estimating individual treatment effect: generalization bounds and algorithms. International conference on machine learning, 2017\\\\.\"}", "{\"comment\": \"**(Q8) CLM:**\\nThanks for your valuable question here and you are correct. Our method heavily depends on the asymptotic normality of the estimated rectifier in our proof of Theorem 5.2. Following the CLM, the rectifier asymptotically converges to the expectation, i.e., \\n$\\\\\\\\sqrt{n}\\\\\\\\left(\\\\\\\\hat{\\\\\\\\Delta}\\\\_{\\\\\\\\tau} \\\\- \\\\mathbb{E}\\\\[\\\\\\\\hat{\\\\\\\\Delta}\\\\_\\\\\\\\tau\\\\]\\\\\\\\right) \\\\\\\\Rightarrow \\\\mathcal{N} \\\\\\\\left(0, \\\\\\\\sigma^2\\\\_{\\\\\\\\Delta}\\\\\\\\right)$ where $\\\\\\\\sigma^2\\\\_{\\\\\\\\Delta}$ is the variance of $\\\\\\\\hat{\\\\\\\\Delta}\\\\_i \\\\= \\\\\\\\tilde{Y}\\\\_{\\\\\\\\hat{\\\\\\\\eta}}(X\\\\_i) \\\\- \\\\\\\\hat{\\\\\\\\tau}\\\\_2(X\\\\_i)$. As for your further worry that the rectifier is constructed only on $\\\\mathcal{D}^1$, since we assume that $\\\\mathcal{D}^1$ and $\\\\mathcal{D}^2$ are sampled from the same population, so the expectation of $\\\\\\\\hat{\\\\\\\\tau}\\\\_2(X\\\\_i)$ are same in $\\\\mathcal{D}^1$ and $\\\\mathcal{D}^2$. Then by Slutsky\\u2019s theorem, we could derive the asymptotical normality of our proposed ATE estimator.\\n\\n**Action:** We added a new section to explain the asymptotic properties of $\\\\tau\\\\_2$ (see **our new Appendix A.3**).\\n\\n**(Q9) Proof for estimated function:** \\nThanks for your question. We follow the proof of the double robustness in Wager 2024 \\\\[8\\\\], where both the estimated nuisance function and the non-centered IF scores also fulfill asymptotic normality property, i.e., $\\\\\\\\sqrt{n}(\\\\\\\\hat{\\\\\\\\tau}\\\\_{\\\\\\\\mathrm{AIPW}} \\\\- \\\\\\\\tau\\\\_{\\\\\\\\mathrm{AIPW}}) \\\\\\\\rightarrow\\\\_{p} 0$. Then, since the non-centered IF scores are an unbiased estimation of the oracle ATE, the proof is not affected.\\n\\n**Action:** We added a footnote about why this is not problematic.\\n\\n**(Q10) Proposed new CI compared to a classical one:** \\nWe kindly refer to our response to (Q3).\\n\\n**(Q11) $1/n$ missing:** \\n Thank you. \\n\\n**Action:** We fixed this in the revised PDF.\\n\\n**(Q12) How the RMSE is computed:** \\nThank you for pointing this out. For the real-world application in our study, we utilize a semi-synthetic dataset, which means we know the potential outcome generation process. This allows us to directly evaluate the accuracy and validity of our method by comparing the estimated ATE against the ground truth. By leveraging the semi-synthetic nature of the dataset, we can strike a balance between real-world complexity and the ability to validate our results rigorously.\\n\\n**Action:** We clarified this in our paper. \\n\\n**(Q13) Qualification of our \\u201cTakeaway\\u201d:** \\nThanks for your comments. We conducted comprehensive experiments not only on two distinct datasets but also using five different random seeds to ensure robustness. Additionally, we explored various experimental sample size settings to evaluate the consistency and scalability of our method in the revised paper across different scenarios. We hope that, together with the new experiments, the takeaways are warranted. \\n\\n**Action:** We added various new experiments to qualify our claims in the \\u201cTakeaways\\u201d. \\n\\n**(Q15) Typos:** \\nWe sincerely apologize for the typos and greatly appreciate your careful reading. We will address these errors and correct them in the revised version of our paper.\\n\\n**Action:** We have corrected all the typos.\\n\\n**References:** \\n\\\\[1\\\\] Anastasios N. Angelopoulos, Stephen Bates, Clara Fannjiang, Michael I. Jordan, and Tijana Zrnic. Prediction-powered inference. Science, 382(6671):669\\u2013674, 11 2023\\\\. ISSN 0036-8075, 1095-9203. \\n\\\\[2\\\\] G. A. Barnard. Statistical Inference, 1949\\\\. ISSN 0035-9246. \\n\\\\[3\\\\] Weixin Cai and Mark J. van der Laan. One-step targeted maximum likelihood for time-to-event outcomes, 2019\\\\. \\n\\\\[4\\\\] Ilker Demirel, Ahmed Alaa, Anthony Philippakis, and David Sontag. Prediction-powered Generalization of Causal Inferences, 6 2024\\\\. \\n\\\\[5\\\\] Nathan Kallus, Aahlad Manas Puli, and Uri Shalit. Removing Hidden Confounding by Experimental Grounding. Advances in neural information processing systems, 2018\\\\. \\n\\\\[6\\\\] Fangzhou Su, Wenlong Mou, Peng Ding, and Martin J. Wainwright. When is the estimated propensity score better? high-dimensional analysis and bias correction, 2023\\\\. \\n\\\\[7\\\\] Mark van der Laan, Sky Qiu, and Lars van der Laan. Adaptive-TMLE for the Average Treatment Effect based on Randomized Controlled Trial Augmented with Real-World Data, 5 2024\\\\. \\n\\\\[8\\\\] Stefan Wager. Causal Inference: A Statistical Learning Approach, 2024\\\\.\"}", "{\"comment\": \"### **Response to \\u201cQuestions\\u201d:**\\n\\n**(Q1) Position of the work:** \\nThank you. We followed your suggestion closely and performed new experiments as follows: First, we applied our AIPW-based methods to combinations of RCT and observational datasets (see our **new Appendix G.6**). Second, we applied the A-TMLE method on multiple observational datasets (see our **new Appendix G.7**). We can see that both experiment results show that when replacing the known propensity score with the estimated one, our method still shows the faithful and even performs better in shrinking the width of CIs. Further, as expected, A-TMLE does not perform well in the new experiments because it does not properly estimate the propensity scores. \\n\\n**Action:** We performed new experiments where we estimated the propensity score as you suggested. Here, we cited the references by Su et al. \\\\[6\\\\] and Lars van der Laan et al., 2024 \\\\[7\\\\] as motivation for conducting such experiments. Specifically, we applied our AIPW-based methods to combinations of RCT and observational datasets (see our **new Appendix G.6**) and the A-TMLE method to settings with multiple observational datasets (see our **new Appendix G.7**). The results again confirm the effectiveness of our methods. \\n\\n**(Q2) Comparison with the work of van der Laan et al., 2024 \\\\[7\\\\]:** \\nThank you. We followed your suggestion and added a detailed, technical comparison to highlight key differences (see our **new Appendix H**). \\n(1) Our method and van der Laan rely on **different ATE estimation processes**, where A-TMLE uses separate TMLEs to estimate pooled ATE and bias correction and our method relies on AIPW to estimate the ATEs. \\n(2) The validity of CIs from A-TMLE only holds when applying TMLE as the estimation process for the pooled ATE, however our method shows higher flexibility. The validity of our proposed CIs holds for the arbitrary estimation process of $\\\\\\\\hat{\\\\\\\\tau}\\\\_2$. \\n(3) In A-TMLE, it relies on the HAL-MLE to estimate the semi-parametric regression working model. While the data are simple, it is always hard to solve the covariance matrix which leads to the failure of the whole algorithm.\\n\\n**Action:** We added a detailed comparison with the A-TMLE from van der Laan in our **new** **Appendix H**. Therein, we highlight key differences at a technical level.\\n\\n**(Q3) Reference for l. 151-152:** \\nThank you. Upon reading your comment, we realized that we should be more careful in explaining the intuition of our method. We expanded our explanations as follows. As the model $f$ is sufficiently accurate, we have the rectifier almost equal to zero, $\\\\\\\\hat{\\\\\\\\Delta} \\\\\\\\approx 0$. Then the variance of the rectifier is significantly smaller than the variance of estimated non-centered IF scores, $\\\\\\\\hat{\\\\\\\\sigma}\\\\_{\\\\\\\\Delta}^2 \\\\\\\\leq \\\\\\\\hat{\\\\\\\\sigma}_{f}^2$. Given the large size of $\\\\\\\\mathcal{D}\\\\_2$, the variance of the estimated conditional treatment effect can be almost overlooked, since the estimated variance should be divided by the sample size of $\\\\\\\\mathcal{D}\\\\_2$. The rest follows from reference \\\\[1\\\\]. \\n\\n**Action:** We **added the above explanation to our paper**. Specifically, we added a summary to revise the \\u201cIntuition behind our method\\u201d in the Introduction section of the revised paper making it easier to understand. Further, we added a new paragraph \\u201cWhy is our method better than using the unconfounded dataset only?\\u201d to our **revised Section 5**.\\n\\n**(Q4) Requirement of overlapping assumptions:** \\nThank you, we revised how we formalize the overlap assumption. \\n\\n**Action:** We fixed this.\\n\\n**(Q5) Meaning of \\u201cvalid\\u201d:** \\nThank you. We refer to Barnard et al. \\\\[2\\\\] and refer to a confidence interval as \\u201cvalid\\u201d when the interval achieves its stated coverage probability. For example, a 95% confidence interval is valid if, under repeated sampling, it contains the true parameter value approximately $95\\\\\\\\%$ of the time. Validity ensures the interval accurately reflects the level of uncertainty about the estimate. \\n\\n**Action:** We added a formal explanation of when we regard a CI as \\u201cvalid\\u201d.\\n\\n**(Q6) Clarification for Lemma 5.1:** \\nThank you for spotting that the notation may have been unclear here. We later need the cross-fitted estimator of the nuisance parameters and, while their convergence rate added together is quick enough, the asymptotic normality holds for the AIPW estimator\\n\\n**Action**: We fixed the notation. We changed Lemma 5.1 to Remark 5.1 and wrote that it follows immediately from Wager\\u2019s book.\\n\\n**(Q7) Confusion about l.298:** \\nSorry for the unclear here, we use cross-fitted nuisance parameters on $\\\\\\\\mathcal{D}\\\\_1$ to calculate the non-centered IF scores on $\\\\\\\\mathcal{D}\\\\_1$ and then calculate the rectifier as the mean of differences.\\n\\n**Action:** We improved our presentation.\"}", "{\"comment\": \"Thank you for your response and addressing my concerns. I have updated my score.\"}", "{\"comment\": \"**Thank you for updating the score! And thanks again for your valuable question and the active discussion!** We are happy to provide additional justification for the application of the CLT for $\\\\hat{\\\\Delta}{\\\\tau}$ step by step as follows:\\n\\n1. To see why the CLT holds, let us define **a new auxiliary random variable** $Z_i = \\\\tilde{Y}_{\\\\hat{\\\\eta}}(x_i) - \\\\hat{\\\\tau}_2(x_i)$. In particular, **we do not apply the CLT separately** on both summands but on the **joint** variable $Z_i$.\\n\\n2. The $Z_i$ are **i.i.d. random variables** because we used cross-fitting and estimated the nuisance functions $\\\\hat{\\\\eta}$ on $\\\\mathcal{D}^1$ and the CATE estimator $\\\\hat{\\\\tau}_2$ on $\\\\mathcal{D}^2$ (in particular, not on $\\\\mathcal{D}^1$). **Hence, the CLT holds for $Z_i$.**\\n\\n3. The **estimation of nuisance parameters $\\\\hat{\\\\eta}$ does not affect the asymptotic mean and variance** of the limit distribution of $\\\\hat{\\\\Delta}\\\\_\\\\tau = \\\\frac{1}{n}\\\\sum\\\\_{i=1}^n Z\\\\_i$ (under the assumptions from Remark 4.1). We realized that we missed a formal argument for this step and apologize for this. **We expanded the proof in Appendix A.2 and added a formal argument**. The **intuition is as follows**: \\n\\n Remark 4.1. states that this is true for $ \\\\tilde{Y}\\\\_{\\\\hat{\\\\eta}}(x_i) $, i.e., using estimated nuisances $\\\\hat{\\\\eta}$ does not affect the asymptotic mean and variance of $\\\\frac{1}{{n}} \\\\sum_{i=1}^n \\\\tilde{Y}\\\\_\\\\hat{\\\\eta}(x_i)$. The key fact used here is that we can write the $\\\\frac{1}{n}\\\\sum_{i=1}^n \\\\tilde{Y}\\\\_\\\\hat{\\\\eta}(x\\\\_i) = \\\\underbrace{\\\\frac{1}{n}\\\\sum_{i=1}^n \\\\left( \\\\tilde{Y}\\\\_\\\\hat{\\\\eta}(x\\\\_i) - Y\\\\_\\\\eta(x\\\\_i)\\\\right)}\\\\_{\\\\text{Error due to nuisance estimation}} + \\\\underbrace{\\\\frac{1}{n}\\\\sum\\\\_{i=1}^n Y\\\\_\\\\eta(x\\\\_i) }\\\\_{\\\\text{Oracle nuisance functions}} $. Then the proof in Wager\\u2019s book [1] proceeds by showing for the first term that $\\\\sqrt{n}\\\\left(\\\\frac{1}{n}\\\\sum_{i=1}^n \\\\left( \\\\tilde{Y}\\\\_\\\\hat{\\\\eta}(x\\\\_i) - Y\\\\_\\\\eta(x\\\\_i)\\\\right)\\\\right) \\\\rightarrow_p 0$ by using the cross-fitting and rate assumptions on the nuisance estimators. For the second term, we can apply the standard CLT as it only depends on the ground-truth nuisance functions $\\\\eta$.\\n\\n **In our case, we proceed similarly** and write, $\\\\hat{\\\\Delta}\\\\_\\\\tau = \\\\frac{1}{n}\\\\sum\\\\_{i=1}^n Z\\\\_i = \\\\underbrace{\\\\frac{1}{n}\\\\sum_{i=1}^n \\\\left( \\\\tilde{Y}\\\\_\\\\hat{\\\\eta}(x\\\\_i) - Y\\\\_\\\\eta(x\\\\_i)\\\\right)}\\\\_{\\\\text{Error due to nuisance estimation}} + \\\\underbrace{\\\\frac{1}{n}\\\\sum\\\\_{i=1}^n \\\\left(Y\\\\_\\\\eta(x\\\\_i) - \\\\hat{\\\\tau}\\\\_2(x\\\\_i)\\\\right)}\\\\_{\\\\text{Oracle nuisance functions}} .$ Note that $\\\\hat{\\\\tau}\\\\_2 (x\\\\_i)$ cancels out in the first term as $\\\\hat{\\\\tau}\\\\_2 (x\\\\_i)$ does not depend on any nuisance estimator $\\\\hat{\\\\eta}$. Hence, following the same arguments as in proof from Wager\\u2019s book, Chapter 2 [1], the first term vanishes and we can apply the standard central limit theorem to the second term.\\n\\n4. Steps 1-3 imply that $\\\\hat{\\\\Delta}\\\\_\\\\tau$ is asymptotically normal with population mean $E[Z] = E\\\\_X [ \\\\tilde{Y}\\\\_{\\\\eta}(X) - \\\\hat{\\\\tau}\\\\_2 (X)] = \\\\tau - E\\\\_X [ \\\\tau\\\\_2 (X)]$ and variance $Var (Z) = Var\\\\_X \\\\left( \\\\tilde{Y}_{\\\\eta} (X) - \\\\hat{\\\\tau}_2(X) \\\\right)$. The population variance can be estimated $\\\\hat{\\\\sigma}\\\\_{\\\\Delta}^2 = \\\\frac{1}{n} \\\\sum\\\\_{i=1}^n \\\\left(\\\\tilde{Y}\\\\_\\\\hat{\\\\eta} (x\\\\_i) -\\\\hat{\\\\tau}\\\\_2(x\\\\_i) - \\\\hat{\\\\Delta}\\\\_\\\\tau \\\\right)^2$.\\n\\n**Action**: We **added the missing argument from Step 3 to our new Appendix A.2**.\\n\\n**Minor**\\n\\nThanks for the valuable suggestion.\\n\\n**Action:** We **corrected the notation and emphasized our use of cross-fitting** in Theorem 4.2. We now explicitly state which component is estimated on which dataset.\\n\\n**Reference:**\\n\\n[1] Stefan Wager. Causal Inference: A Statistical Learning Approach, 2024.\"}", "{\"comment\": \"Thank you for your positive feedback! We will incorporate all action points into our revised version of paper.\"}", "{\"comment\": \"Thank you for your positive feedback! We will incorporate all action points into our revised version of paper.\"}", "{\"comment\": \"Thank you for your response. My confusion come from the fact that the manner you apply cross-fitting is never explicitly written. The CLT on $\\\\hat{\\\\Delta}{\\\\tau}$ may hold because $\\\\hat{\\\\eta}$ is built on a data set that does not contain $x_i$. How cross-fitting is applied to your estimators should be clearly mentioned in the text and in the algorithm, so that the implemented procedure is clear to the reader.\\n\\nBesides, even with cross-fitting, contrary to what you stated, the $Z_i= \\\\tilde{Y}_{\\\\hat{\\\\eta}}(x_i) - \\\\hat{\\\\tau}_2(x_i)$ are not independent as they all depend on the data set used to construct $\\\\hat{\\\\eta}$. Even using cross-fitting with leave-one-out (using all observations but one to build your estimate and compute the quantity of interest on the remaining observation) would not lead to independent $Z_i$. Thus, the proof needs to be clarified. \\n\\nI believe that the proof of Wager can be extended to your setting, but I would like to see a clear formal proof of this statement.\"}", "{\"summary\": \"This work presents a method for estimating average treatment effects (ATE) and constructing valid confidence intervals (CIs) from multiple observational datasets. It uses minimal assumptions and leverages prediction-powered inferences for more precise CIs. The approach is validated through theoretical proofs and numerical experiments, with extensions for mixed data sources.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"The article is well written, and technically sound. The proposed approach is interesting, with great potential for practical application. The previous literature is well integrated, and the paper's novelty is well explicited.\", \"weaknesses\": \"1. It appears that the novelty of this work is incremental. Use the previously proposed estimator for causal effects and use it in a PPI framework to construct confidence intervals.\\n\\n2. Lemma 6.1 is a theoretical property of the inverse probability weighting (IPW) estimator and should not be considered a contribution of this paper. Furthermore, the IPW estimator can also be interpreted as an estimation method based on the influence function, while the augmented inverse probability weighting (AIPW) estimator is derived from the effective influence function. Therefore, the statements made in lines 357-360 are inaccurate.\\n\\n3. The proof of Theorem 5.2 appears to depend on the asymptotic normality of $\\\\tau_2$. However, the $\\\\mathcal{D}_2$ does not seem to satisfy ignorability, suggesting that it is a biased estimator. I have reservations about whether it is indeed asymptotically normal about $\\\\hat{\\\\tau}_2$ and $\\\\tau_2$ at this point.\", \"questions\": \"This paper is written towards an audience of applied statistics researchers, and is more suited for a statistical journal such as Biometrika. Although the reviewer acknowledges the significance of rigorously discussing the construction of confidence intervals, the methodology presented may not be highly relevant to the broader community of ICLR. For example, it does not discuss neural network or representation learning at all.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"metareview\": \"**Summary**: This paper proposes a new method to generate reliable confidence intervals for average treatment effects using multiple datasets where additional data set samples may have residual unobserved confounding. Authors build on the prediction-powered inference (PPI) framework to estimate ATE on confounded data, rectifying the bias using the smaller unconfounded data, and using sample statistics of the bias for generating confidence intervals. The procedure uses CATE estimates from the larger dataset to obtain a DR CATE estimate using sample splitting, followed by rectifying using the efficient influence function and subsequently deriving confidence intervals from the summary statistics. The framework is flexible by allowing ML methods used for nuisance estimation in high dimensions and applicable to AIPW, IPW estimators. Theoretical results show asymptotic coverage of the confidence intervals.\\n\\n**Strengths**: \\n1. Reliable uncertainty quantification of ATE is a critical problem for improving causal inference. \\n2. The paper further attempts to tighten interval estimates by leveraging potentially confounded data thereby leveraging the power of additional samples. The contribution is significant and novelty is clear and easy to follow. \\n3. Theoretical claims while straightforward are well presented and justified.\\n4. Empirical evaluation clearly shows better coverage compared to prior work.\\n\\n**Weaknesses**:\\n1. Multiple reviewers pointed out that the empirical evaluation is relatively weak for the standards of ICLR, but the focus of the contribution\\n2. Considering that the contribution is more statistical, i.e., targeted toward better inference without heavy focus on the representation learning of the nuisances, fit to ICLR was not clear for multiple reviewers\\n\\n**Justification**: Overall all reviewers agree that the contribution is valuable. Multiple reviewers had clarifying questions on the details. There were many clarity issues in the writing in the initial version which were subsequently addressed, proof statements clarified, additional empirical evaluation added, including updates to the written text to add pertinent details e.g., on use of sample splitting. Overall after multiple iterations of discussions between reviewers and authors, it is clear that clarity issues where addressable and not fundamental issues with the contribution. Overall the novelty of the contribution is explicit, demonstration of empirical benefit clear, and multiple reviewers raised their scores post rebuttal. Considering the novelty of the contribution, the quality of the writing, and potential interest to a small but significant subset of ICLR readers, I am recommending an accept.\", \"additional_comments_on_reviewer_discussion\": \"No major concerns were brought up during discussion.\"}", "{\"comment\": \"Thank you for your detailed response and helpful changes! I have updated my score.\"}", "{\"summary\": \"This paper considers the problem of estimating confidence intervals for average treatment effects when given access to a small, unconfounded observational dataset and a larger, confounded observational dataset. Specifically, the problem setting assumes that the covariates for both datasets come from the same population, but the propensity score may differ and the potential outcomes are only independent of treatment assignment in the smaller unconfounded dataset.\\n\\nThe paper's method builds on the prediction-powered inference (PPI) framework, which creates an estimate of the ATE via the confounded dataset and then uses the unconfounded dataset to adjust it and build the appropriate confidence intervals. The paper proves that their method asymptotically results in valid confidence intervals. The method is evaluated on both synthetic and real data, showing its gains over using only the unconfounded dataset.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"The paper addresses an important issue that has not been covered in the literature before. The paper also does a good job of placing itself within the literature. The paper also does a good job of precisely backing up their claims with theorems.\", \"weaknesses\": \"On the theoretical side, the paper only shows the asymptotic validity of the proposed method. There are no results demonstrating the width of the resulting confidence intervals. One would like to know under what circumstances the given method will produce a confidence interval that converges on exactly 1-$\\\\alpha$.\\n\\nMoreover, the paper does not give any theoretical insight into when it is beneficial to use both datasets as opposed to just the unconfounded dataset. There should be some result showing that the confidence intervals are tighter for proposed method than for the naive method.\", \"questions\": \"Under what circumstances can it be proved that the proposed method is actually better than operating on just the unconfounded dataset?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you for your response! I only have one comment left:\\n- Regarding the point 3) above, I am still not sure how you properly obtain a CLT. In order to combine the two CLT, you need to have that \\n\\n$$ \\\\frac{1}{n} \\\\sum_{i=1}^n \\\\tilde{Y}_{\\\\hat{\\\\eta}}(x_i) $$\\n\\nand \\n\\n$$ \\\\frac{1}{n} \\\\sum_{i=1}^n \\\\hat{\\\\tau}_2(x_i) $$\\n\\nare independent, which is not the case as they both depend on $x_i$. However, I agree that the two functions $\\\\tilde{Y}_{\\\\hat{\\\\eta}}(\\\\cdot)$ and \\n\\n$\\\\hat{\\\\tau}_2(\\\\cdot)$ \\n\\nare independent. Can you justify more precisely how you obtain the final CLT for $\\\\hat{\\\\Delta}_{\\\\tau}$?\\n\\n\\n(on a minor note, I believe that the use of cross-fitting should be explicitly mentioned in Theorem4.2. In the current version, it is said 'sample splitting in splitted data sets' which is not very clear).\"}" ] }

[ "Vivek Myers", "Catherine Ji", "Benjamin Eysenbach" ]

[ "reinforcement learning", "generalization", "invariance", "planning" ]

{ "note_id": [ "xBtrVPFSrp", "woFDnLon8s", "wWJ92BLE58", "v3WhTevy8o", "skQogd4hVc", "sclc0DjyqZ", "pIRIiN7Dzn", "maL4POiX9p", "lcwO4jkq64", "iAdZywZaWo", "f7ezqWtx3T", "dGLzRPLgoT", "ckmuvXh06N", "cXOdbRjp41", "TWIcRxf9pp", "SdPVmmslZt", "RD1Kc6i8Z2", "OR1tynqnZ9", "MvNRgKuy5m", "JF5EQlQwFt", "HvWiOBsiKV", "EswgVBGkFe", "9HyFGTvCmD", "5rfDDWe8jK", "2z60gnbKFS" ], "note_type": [ "official_comment", "official_review", "official_review", "official_comment", "official_comment", "official_comment", "meta_review", "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment" ], "note_created": [ 1732260450791, 1730671724699, 1729614715661, 1733009176419, 1732249218358, 1732552073368, 1734564728334, 1737523569048, 1732552125610, 1733009205429, 1732639533806, 1733043782065, 1732552114056, 1730512080350, 1730687076693, 1733009216779, 1732265974967, 1732249800186, 1733018579581, 1733226383180, 1732552098496, 1733009189636, 1733312873149, 1732249283486, 1732260501840 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission3310/Authors" ], [ "ICLR.cc/2025/Conference/Submission3310/Reviewer_h6Gn" ], [ "ICLR.cc/2025/Conference/Submission3310/Reviewer_6v2T" ], [ "ICLR.cc/2025/Conference/Submission3310/Authors" ], [ "ICLR.cc/2025/Conference/Submission3310/Authors" ], [ "ICLR.cc/2025/Conference/Submission3310/Authors" ], [ "ICLR.cc/2025/Conference/Submission3310/Area_Chair_qUqJ" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission3310/Authors" ], [ "ICLR.cc/2025/Conference/Submission3310/Authors" ], [ "ICLR.cc/2025/Conference/Submission3310/Authors" ], [ "ICLR.cc/2025/Conference/Submission3310/Authors" ], [ "ICLR.cc/2025/Conference/Submission3310/Authors" ], [ "ICLR.cc/2025/Conference/Submission3310/Reviewer_ctUj" ], [ "ICLR.cc/2025/Conference/Submission3310/Reviewer_7Gwa" ], [ "ICLR.cc/2025/Conference/Submission3310/Authors" ], [ "ICLR.cc/2025/Conference/Submission3310/Authors" ], [ "ICLR.cc/2025/Conference/Submission3310/Authors" ], [ "ICLR.cc/2025/Conference/Submission3310/Reviewer_ctUj" ], [ "ICLR.cc/2025/Conference/Submission3310/Reviewer_h6Gn" ], [ "ICLR.cc/2025/Conference/Submission3310/Authors" ], [ "ICLR.cc/2025/Conference/Submission3310/Authors" ], [ "ICLR.cc/2025/Conference/Submission3310/Authors" ], [ "ICLR.cc/2025/Conference/Submission3310/Authors" ], [ "ICLR.cc/2025/Conference/Submission3310/Authors" ] ], "structured_content_str": [ "{\"title\": \"Rebuttal\", \"comment\": \"Dear Reviewer,\\n\\nThank you for your responses and suggestions. \\n\\nIt seems your main suggestion is to provide more intuition for the theoretical claims, including a discussion of the relationship with prior work on representations in RL. We have made revisions in the paper to address these concerns and believe that the paper motivation is now much clearer. **Together with the discussion below, do these fully address your concerns?**\", \"we_have_revised_to_the_paper_to_provide_more_intuition_and_background\": \"1. **Highlight relationships with prior state abstraction works such as bisimulation**: In Section 2, we now compare and contrast horizon generalization with other forms of generalization captured by prior state abstraction methods. We highlight that, to our knowledge, there has not been prior work directly addressing generalization from short to long horizons. Furthermore, planning invariant policies map state-action-waypoint pairs and state-action-goal pairs to the same latents, despite different associated rewards and Q-values (see Section 3 for relation between quasimetric and underlying reward function in an MDP). This is already a violation of the Q-irrelevance relation and the stronger bisimulation relation [1]: the invariances captured by planning invariance and such state abstractions are fundamentally different. \\n2. **Building intuition for planning invariance**: We have updated our planning invariance figure to help build intuition and highlight Section 5: Methods for Planning Invariance: Old and New.\\n3. **Takeaways**: We have highlighted the high-level takeaways for the reader throughout the paper (see end of Section 1, Section 2, end of Section 4.4, and Section 7). By theoretically and empirically linking planning with horizon generalization, our work suggests practical ways (i.e., quasimetric methods) to achieve powerful notions of generalization from short to long horizons. \\n**Do these revisions, together with the discussion below, fully address your concerns?** We look forward to continuing the discussion!\\n\\n> There doesn't seem to be any suggestions on how to induce this kind of invariance.\\n\\nThis invariance is precisely the structure imposed by quasimetric architectures\\u2014we have clarified the notation so this is now clearly stated as [Theorem 2](https://gcdnb.pbrd.co/images/gEMhIRNV1hGk.png) in the text.\\nWe have also added a short summary of the concrete takeaways to the conclusion section, and a summary of different architectures and losses and how they relate to the planning invariance of quasimetric architectures.\", \"additional_discussion_is_presented_section_5\": \"\\u201cMethods for Planning Invariance, Old and New\\u201d as well as Appendix C. For example, for the forward perspective, we can induce planning invariance (1) explicitly via dynamic programming or (2) implicitly using a quasimetric architecture. We empirically demonstrate that quasimetric architectures, when combined with a contrastive metric learning approach [3], are relatively planning invariant compared to other architectures (see [Figure 5, right](https://gcdnb.pbrd.co/images/Gv5HtRmtwnhR.png)).\\n\\n> relation to compositional generalization in language-conditioned RL\\n\\nIndeed, a key advantage of language as a form of task specification is its ability to easily express compositional relationships. Future work could investigate how planning-invariant reasoning could be extended to these alternate task modalities, through, e.g., learned mappings from the task space into the quasimetric goal space [2]. \\nWe have added discussion to our conclusion (marked in red, under \\u201cLimitations and future work\\u201d) on how future work could tackle these challenges.\\n\\n> give more examples of non-trivial cases where planning invariance is or is not met, in order to build intuitions\\n\\nWe have replaced the figure in Section 4.1 with [Figure 2](https://gcdnb.pbrd.co/images/PNHteRUKMkhP.png). With this revised figure, we can (1) visualize different 2D translation actions for different navigation goals (purple star and brown star) as well as optimal (tan) and suboptimal (gray) trajectories, (2) identify that the left policy has no planning invariance (policies towards goal and waypoint are not the same, where the policy selects a *suboptimal* action while navigating towards the goal and *optimal* action while navigating towards the waypoint) while the right policy has planning invariance (indicated by the same, optimal policy whether directed towards a goal or the waypoint), and (3) concretely connect the presented maze task to actual sequential decision-making problems like navigation with obstacles.\"}", "{\"summary\": \"This paper explores the problem of horizon generalization in goal-conditioned reinforcement learning. It presents clear definitions of planning invariance and demonstrates how this concept can facilitate horizon generalization.\", \"soundness\": \"1\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"The problem studied in the paper is novel and very interesting to me. The motivation of the paper is clear and the writing of the paper is generally clear and easy to fllow.\\n\\nThe paper presents clear definition of planning invariance and horizon generalization, which make sense to me.\", \"weaknesses\": \"The paper contains some symbols that are not well defined, some of which even affect the readability of the paper.\\n\\n- I got lost from Line 245 with $d(s,a,g)$. I don't understand the meaning of this symbol and why it contains an action. The paper didn't clearly define it. This is critical as I failed to fully understand the proof of Lemma 1 and Lemma 2, which are the major results of the paper. I hope I didn't miss anything.\", \"some_undefined_symbols_that_do_not_affect_reading\": \"- Eq. (2): $p_\\\\gamma^\\\\pi$, $Geom$ is never defined\\n- Line 117: $\\\\delta$, $r_g$\\n\\nThe paper also uses different terminology from the common terminology, making it seem less professional.\\n\\nUsually, the distance (metric) function should satisfy the three rules: Non-negativity, Triangle Inequality, and Symmetry; while a quasimetric relaxes one or more of the metric's defining properties.\", \"minor_comments\": [\"It seems Lemma 1 and 2 are some important results of the paper. Then please use \\\"Theorem\\\" rather than \\\"Lemma\\\".\", \"\\\"Markove process\\\" should be \\\"Markov Decision Process\\\". You didn't include the definition of the reward function.\", \"Another weakness is the paper's lack of clarity on how its discoveries could inspire future algorithm development. I hope future revisions of the paper will devote more effort to exploring this aspect.\"], \"questions\": \"I was unable to assess the correctness of the theorems due to some undefined symbols mentioned earlier. Please respond to my previous queries in weakness. I will reassess them in case I missed anything.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The motivation and insight presented in this paper are interesting and straightforward. This paper presents the abundant theoretical contribution to Planning Invariance and Horizon Generalization. Experimental results of existing methods support the theoretical results.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. Good writing and structure.\\n2. The authors illustrate their research problem via the good toy example (Figrue 3).\\n3. Theoretical results are clear and sound.\", \"weaknesses\": \"I have no concern with the main contribution of the theoretical part of this paper. The main concern is related to the application and limitations.\\n\\n1. One of the main concerns of the reviewer is the limitations of the planning invariance and horizon generalization, as mentioned by the authors (Section 4.5).\\n2. I am also concerned about real-world applications (e.g., more complicated tasks). In real-world settings, the planning invariance and horizon generalization seem to be limited in their application to general cases, although they indeed exist as mentioned by the authors. In the high-dimensional settings (Section 6.2), it is hard to derive insightful results.\\n3. The authors provide some interesting future directions in Appendix C, but it would be better to present some experimental results or practical implementations.\", \"questions\": \"See weakness.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Rebuttal follow-up\", \"comment\": \"Dear Reviewer,\\n\\nWe have worked hard to incorporate the review feedback by running new experiments and revising the paper. **Do the revisions and discussions above address your concerns?** We would greatly appreciate your engagement.\\n\\nThanks!\\n\\nThe Authors\"}", "{\"title\": \"Rebuttal\", \"comment\": \"Dear Reviewer,\\n\\nThank you for your detailed review and suggestions.\\u00a0\\n\\nYour primary concerns seem to be (1) issues with the motivation and formalism for planning invariance and horizon generalization and (2) concerns about the utility of our formalism in practical environments.\\n\\nWe have made revisions based on your suggestions **highlighted in red** to clarify the theoretical results, added discussion on concrete takeaways to our conclusions, and updated our experiments to support our theoretical claims.\\n\\n**Do these changes fully address your concerns?** We believe the current claims in the paper are correct (regarding \\u201cthis is a bit far-fetched\\u201d) and have clarified the proofs in Appendix B. If you still believe any of the proofs or other writing is unclear, please inform us and we will be happy to make further revisions.\\n\\n> Generalization in RL and in goal-conditioned setting can come in different forms. While generalizing over different horizons is important, why do you think invariance will help here?\\u00a0\\n\\nIn ML, there is a strong connection between generalization and invariance where enforcing invariance facilitates generalization. For example, neural networks invariant to rotation show better abilities to generalize over rotated test sets [1, 2, 3]. This body of work motivates our analysis; we have revised Section 4.1 to refer to such work.\\nWe expect generalization over different horizons in RL to also relate to a form of invariance, and identify planning invariance as a possible mechanism to achieve horizon generalization.\\n\\n> How does the method relate to when we are no longer looking at smaller horizons as being sub-goals of the longer horizons?\\u00a0\\n\\nIn general, *any* goal reached at a given horizon will necessarily have possible shorter-horizon subgoal(s) just by taking a distribution of states reached ``on the way\\u2019\\u2019 to the goal. This is mathematically formalized through our induction argument, which necessarily assumes that you can use smaller problems to solve larger problems.\\nDoes this clarify your concern? We are not entirely sure if we have understood it correctly, but would be happy to provide further information if needed.\\n\\n> Or even when we are not considering the goal as a condition. Do you think this work can include those cases?\\n\\nOutside the goal-conditioned setting, the distance metric in its current form loses meaning and we lose the benefits of ``reward-free\\u201d goal-conditioned reinforcement learning; we have added discussion of prior work and motivation for goal-conditioned RL in Section 2.\\u00a0\\nHowever, our result suggests that if you force your reward structure to take on values that lead to quasimetric Q-functions, or Q-functions that are scaled and shifted versions of quasimetrics, then we should expect horizon generalization in goal-free RL. How to enforce such constraints on reward functions in non goal-conditioned settings is still an open question, but could have potential implications on how to engineer rewards (that correspond to a desired objective, even if it isn\\u2019t a fixed goal) to achieve horizon generalization.\\u00a0\\n\\n> The writing of the paper needs some work to strengthen the motivation and claims. The related works section can also benefit from further elaboration and coverage of the related literature.\\n\\nAs suggested, we have made several revisions to make the paper more clear and add motivation. We have (1) expanded our prior works to compare horizon generalization to other forms of generalization with latent space abstractions (see Section 2) and motivated the goal-conditioned setting, (2) clarified the limitations of planning invariance as a mechanism to achieve horizon generalization (Section 4.2 and 4.5), (3) added a remark to highlight that horizon generalization, as defined in this paper, is nontrivial (Remark 2), and (4) replaced the dominoes setting example with a more complex navigation task to build intuition for planning invariance (Figure 2).\\u00a0\\n\\n> Experiments are over simpler environments / task settings.\\n\\nWe would like to note that the primary Ant environment studied involves 8DoF control and 29d observations [4], and is challenging enough to have been used as a primary evaluation for several recent RL works [5,6,7,8]. Since the submission, we have expanded the results in the Ant environment to study generalization to quantitatively more distant goals, and have additionally run preliminary experiments in the Humanoid [4] environment showing similar trends, which will be included in the final version of the paper along with the existing results.\"}", "{\"title\": \"Rebuttal follow-up\", \"comment\": \"Dear Reviewer,\\n\\nWe have worked hard to incorporate the review feedback by running new experiments and revising the paper. We'd really appreciate if you could confirm whether these changes address the concerns about the paper. If we have misunderstood any of the concerns, we'd like to learn that now so that we can further revise the paper or run additional experiments.\\n\\nThanks!\\n\\nThe Authors\"}", "{\"metareview\": \"This paper studies the properties of policies that generalize with respect to horizon, which the authors highlight is linked to the notion of _invariance to planning_. The authors provide theoretical analyses on the realizability of horizon generalization and planning invariance, and combine it with an empirical evaluation that serves as evidence for their theoretical findings.\\n\\nThe main weakness the reviewers found with the work was in its utility and downstream applicability, as well as in its clarity. During the rebuttal period the authors provided ample revisions to address the reviewer concerns, but unfortunately most reviewers were rather unresponsive.\\n\\nI went through the paper myself and found it well-written and well-motivated, in particular considering the modifications made during the rebuttal. The main weakness I find is in the exposition of the experiments in section 6, as the wording is sometimes a little confusing to follow. I recommend the authors revise the wording of this section carefully. For example, in the caption in Figure 4 it says \\\"combining that policy with planning does not increase performance\\\", which is not true; what's true is that it increases _less_, but there is an improvement.\\n\\nOverall, however, I believe these corrections are rather minor at this point, and I found the paper to be a good submission under my own revision.\\n\\nThus, given my reading of the paper, the unresponsiveness of the reviewers to a thorough rebuttal to the authors, I am recommending an acceptance (against the overall sentiment of the reviewers).\", \"additional_comments_on_reviewer_discussion\": \"All reviewers provided a good initial review, to which the authors provided thorough rebuttals (which, to my reading, addressed all concerns). Most of the reviewers were not responsive (or were not confident in their responses), which prompted me to review the paper myself.\\n\\nUpon doing so, I am inclined to recommend acceptance.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}", "{\"title\": \"Rebuttal follow-up\", \"comment\": \"Dear Reviewer,\\n\\nWe have worked hard to incorporate the review feedback by running new experiments and revising the paper. We'd really appreciate if you could confirm whether these changes address the concerns about the paper. If we have misunderstood any of the concerns, we'd like to learn that now so that we can further revise the paper or run additional experiments.\\n\\nThanks!\\n\\nThe Authors\"}", "{\"title\": \"Rebuttal follow-up\", \"comment\": \"Dear Reviewer,\\n\\nWe have worked hard to incorporate the review feedback by running new experiments and revising the paper. **Do the revisions and discussions above address your concerns?** We would greatly appreciate your engagement.\\n\\nThanks!\\n\\nThe Authors\"}", "{\"title\": \"Rebuttal follow-up\", \"comment\": \"Dear AC,\\n\\n**Could you kindly ask the reviewers to read the rebuttal and let us know if there are remaining suggestions for the paper?** We have worked hard to address the reviewer feedback through new experiments and revisions. As tomorrow is the last day for revisions, we'd like to make sure that there aren't any additional suggestions.\\n\\nKind regards,\\n\\nThe Authors\"}", "{\"title\": \"Response\", \"comment\": \"> Thank you for the response. One thing that I am not clear on is how the definition of planning invariance is not vacuous in many cases, provided the planner is optimal. For example, let's consider a shortest path from start state $s\\\\_0$ to goal state $g$. Any planner operating at a coarser resolution must produce waypoints along this path. Then for any state along this path, we will have $\\\\pi(a|s, g) = \\\\pi(a|s, w)$, where $w$ is the next waypoint. So this (optimal) policy must also be planning-invariant. Is it the case that any optimal policy within some region of the state space, must also be planning-invariant within that region? It could be helpful to clarify the relationships between planner optimality, policy optimality, and planning invariance.\\n\\nThank you for your suggestions and questions. We would like to clarify the relationships between planner optimality, policy optimality, and planning invariance. Namely, **planning invariance under an optimal planner is a necessary but not sufficient condition for policy optimality, and is a useful inductive bias for a globally optimal policy.** We will make these definitions and relationships clearer in our manuscript. Does the following discussion fully address the reviewer's concerns?\\n\\n> Is it the case that any optimal policy within some region of the state space, must also be planning-invariant within that region?\\n\\nPlanning invariance is defined with respect to a specific planning operator. If a policy is optimal, then the policy is invariant under the optimal planner, but not necessarily invariant under any arbitrary planner. The inverse is not true: a policy that is planning invariant, even under an optimal planner, is not necessarily optimal. Counterexample: we can construct a policy that outputs the same suboptimal actions towards a goal and any optimal waypoint along the shortest path. Thus, planning invariance is an **inductive bias** for a globally optimal policy that can be conveniently enforced via a quasimetric. This view of invariance as an inductive bias is in the same vein as prior ML work enforcing invariances under spatial operators (rotations, translations, etc.) for, e.g., facial recognition tasks [1,2,3]. **Our contribution is to show how ''invariance to planning'' is the analogous invariance property for horizon generalization in RL, and how, like spatial invariances, it can be enforced via architectural constraints.**\\n\\n> relationships between planner optimality, policy optimality, and planning invariance\\n\\nWhat makes planning invariance a *useful* inductive bias is that planning invariant policies propagate *local* policy optimality to global optimality. Local policy optimality alone cannot lead to global policy optimality (horizon generalization). However, local policy optimality *combined with* planning invariance under an optimal planner (which is a weaker condition than global policy optimality, see above) leads to global policy optimality. This horizon generalization is nontrivial as we show theoretically (see Remark 3) and empirically (see updated [Fig. 5, right](https://gcdnb.pbrd.co/images/Gv5HtRmtwnhR.png)).\\n\\n### References\\n\\n[1] Cohen, T., Welling, M., 2016. ''Group Equivariant Convolutional Networks.'' *ICML*\\n\\n[2] Benton, G. et al., 2020. ''Learning Invariances in Neural Networks.'' *NeurIPS*\\n\\n[3] Rowley, H., Baluja, S., Kanade, T., 1998. ''Rotation Invariant Neural Network-Based Facial Recognition.'' *IEEE*\"}", "{\"title\": \"Rebuttal follow-up\", \"comment\": \"Dear Reviewer,\\n\\nWe have worked hard to incorporate the review feedback by running new experiments and revising the paper. We'd really appreciate if you could confirm whether these changes address the concerns about the paper. If we have misunderstood any of the concerns, we'd like to learn that now so that we can further revise the paper or run additional experiments.\\n\\nThanks!\\n\\nThe Authors\"}", "{\"summary\": \"This paper studies horizon generalization in goal-conditioned RL. The main contributions of the paper are a definition of horizon generalization and planning invariance, and proofs of existence for both, as well as some small-scale experimental studies. I admit I am unclear what the takeaways for the reader are - it seems like the main contributions, i.e. the definition of horizon and planning invariance, are quite obvious, as well as the existence of these kinds of invariances. I do agree that the goal of inducing invariances that facilitate generalization in RL is important. It may be that I misunderstood parts of the paper, hence I am putting my confidence at 3 and recommendation at borderline.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": [\"Conceptual investigations are welcome\", \"The paper's figures are well drawn\"], \"weaknesses\": [\"As mentioned above, the takeaways for the reader are not clear - this does not propose a new method, or (to my understanding) clearly shed light on existing methods, highlight a previously unknown weakness,...\", \"The relationship of this work to other areas of work is unclear in some cases.\", \"Please see my detailed questions/comments below.\"], \"questions\": \"- It seems obvious that if the series of waypoints produced by a planner is optimal, then we should have planning invariance according to the definition given here (i.e., the optimal action to reach the next waypoint is the same at the optimal action to reach the final goal), so my guess is that this is not the main message of the paper. The other way around, where the agent trains on easy goals and generalizes to hard ones, seems much more useful, so that this is what the paper is getting at. The paper refers to this as the \\u201cforward perspective\\u201d. However, there doesn't seem to be any suggestions on how to induce this kind of invariance. Also, this seems related to compositional generalization in language-conditioned RL, where the agent is trained on language commands and asked to generalize to novel combinations. How does this relate to planning invariance/generalization?\\n\\n- The paper claims that most prior work on generalization in RL relates to invariance to perceptual changes or simulation parameters. But there is a rich literature on state abstractions which induce invariances to criteria other than perceptual differences, for example: optimal actions (known as policy irrelevance), rewards/Q-values (utile distinctions) [1] bisimulation [2] belonging to a common latent state (however defined) [3]. It seems like the proposed planning invariance could be seen as a special case of one of these other abstractions (such as policy irrelevance). It would be helpful to include a discussion of other types of state abstraction/generalization in the paper. If the proposed invariance cannot be subsumed in any of the prior ones, it would be helpful to describe in what ways. \\n\\n- It would help to give more examples of non-trivial cases where planning invariance is or is not met, in order to build intuitions. There is one example in Section 4.1, but it didn\\u2019t really help my understanding because in that example, there was a single action that was optimal for all goals (and the problem is in a sense trivial for that reason). Also, the example of knocking over dominoes with levers felt fairly contrived. Would it be possible to add one or more examples that 1) have different actions that are optimal for different goals, 2) one can construct some policies that do have planning invariance and others not, in order to illustrate their differences, and 3) are more closely related to common sequential decision-making problems? I know there is also the example in Figure 6, but there the goal-conditioned policy on the left is not optimal. \\n\\n[1] https://citeseerx.ist.psu.edu/document?repid=rep1&type=pdf&doi=ca9a2d326b9de48c095a6cb5912e1990d2c5ab46\\n\\n[2] https://www.auai.org/uai2014/proceedings/individuals/67.pdf\\n\\n[3] https://arxiv.org/pdf/1901.09018\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper studies the generalization of RL from the perspective of horizon invariance. An RL agent that has been trained to reach nearby goals should succeed on distant goals and hence generalize to different horizon lengths. The paper intends to study how the techniques on invariance and generalization from other machine learning domains can be adapted to make goal-directed policies generalize to different horizons.\", \"soundness\": \"1\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"1. The paper aims to study an important question of generalization and planning under different horizon lengths for a goal-conditioned RL agent.\\n2. Connections between existing methods and the new proposed methods are provided.\", \"weaknesses\": \"1. The writing of the paper needs some work to strengthen the motivation and claims. The related works section can also benefit from further elaboration and coverage of the related literature.\\n2. The paper mentions that horizon generalization means that a policy that can achieve a goal n steps away should also be able to reach any new goal for which that original goal is a waypoint. This is a bit far-fetched of a statement as in a longer horizon, a lot of different things could happen over this longer timeframe and not all en-route waypoints would be relevant. Moreover, things could demand a completely different set of actions or approach after these n steps. Defining horizon generalization in this scenario needs a more thorough definition.\\n3. The example of dominoes considered in the paper is a very simple bandit setting, while the paper focuses on more general RL settings with different horizon lengths. Providing relevant working examples will be beneficial to the understanding.\\n4. Experiments are over simpler environments / task settings.\", \"questions\": \"1. Generalization in RL and in goal-conditioned setting can come in different forms. While generalizing over different horizons is important, why do you think invariance will help here? As the horizon changes and the trajectories get longer, it is possible that any of the waypoints relevant over shorter horizons are not that relevant anymore if better paths are discovered for example.\\n2. How does the method relate to when we are no longer looking at smaller horizons as being sub-goals of the longer horizons? Or even when we are not considering the goal as a condition. Do you think this work can include those cases?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Rebuttal follow-up\", \"comment\": \"Dear Reviewer,\\n\\nWe have worked hard to incorporate the review feedback by running new experiments and revising the paper. **Do the revisions and discussions above address your concerns?** We would greatly appreciate your engagement.\\n\\nThanks!\\n\\nThe Authors\"}", "{\"title\": \"Rebuttal\", \"comment\": \"Dear reviewer,\\n\\nThank you for your responses and suggestions for improving the paper.\\n\\nWe have revised the text (changes highlighted in red) to improve the clarity of our theoretical claims, and clearly state the practical implications of our theory and experiments for GCRL algorithms throughout the paper.\\nRegarding your concern about real-world applications and limitations of these generalizations in more complex settings, we will evaluate a higher dimensional setting (humanoid) with a more challenging control component. We hope that these experiments will sufficiently address any remaining concerns. \\n\\n> I am also concerned about real-world applications (e.g., more complicated tasks). In real-world settings, the planning invariance and horizon generalization seem to be limited in their application to general cases, although they indeed exist as mentioned by the authors. In the high-dimensional settings (Section 6.2), it is hard to derive insightful results.\\n\\nAs mentioned in Section 4.5, (1) the optimality over local trajectories and the (2) coverage of local trajectories are both concerns. For (1), there are unavoidable sources of errors from function approximation and inherent noise in training. To push on this potential weakness, **we will further test the presence and relationship between planning invariance and horizon generalization settings in a higher d.o.f. settings like Humanoid [1]**. \\n\\nWe would also like to highlight that we have already tested planning invariance, horizon generalization, and their relationship in Ant, which is a standard, 27-dim observation 8DoF control, RL benchmark used as a primary evaluation in many recent prior RL work [3,4,5]. Since the initial submission, **we have extended the results in [Figure 5](https://gcdnb.pbrd.co/images/Gv5HtRmtwnhR.png) to show further horizon generalization in this setting.**\\n\\n### References\\n\\n[1] Tunyasuvunakool, S. et al., 2020. \\\"dm_control: Software and Tasks for Continuous Control.\\\" *Software Impacts*\\n\\n[2] Park, S. et al., 2024. \\\"OGBench: Benchmarking Offline Goal-Conditioned RL.\\\" arXiv:2410.20092\\n\\n[3] Zheng, C. et al., 2023. \\\"Contrastive Difference Predictive Coding.\\\" *ICLR*\\n\\n[4] Fujimoto, S. et al., 2018. \\\"Addressing Function Approximation Error in Actor-Critic Methods.\\\" *ICML*\\n\\n[5] Freeman, C. et al., 2021. \\\"Brax\\u2014a Differentiable Physics Engine for Large Scale Rigid Body Simulation.\\\" *NeurIPS Datasets and Benchmarks*\"}", "{\"title\": \"Rebuttal\", \"comment\": \"Dear Reviewer,\\n\\nThank you for your responses and suggestions. It seems like your main concerns have to do with notation and clarity, as well as how the theory could relate to practical new methods. We have made revisions in the paper to address these concerns and believe that the paper clarity is now much higher. We hope that these improvements have increased the readability of the main theorems and make the takeaways clear. **Together with the discussion below, does this fully address your concerns?** We look forward to continuing the discussion!\\n\\n> Clarifying notation.\", \"we_have_made_several_revisions_to_the_paper_to_improve_clarity\": \"1. We clearly define the successor distance over actions $d(s,a,g)$, which, using the setup in [1], is $\\\\min\\\\_{\\\\pi} \\\\frac{p^{\\\\pi}\\\\_{\\\\gamma}(s\\\\_K = g|s\\\\_{0} = g)}{p^{\\\\pi}\\\\_{\\\\gamma}(s\\\\_K = g|s\\\\_{0} = s,a)}$ (Equation 5) where $p^{\\\\pi}\\\\_{\\\\gamma}(s\\\\_K = g|s\\\\_{0} = s)$ is the discounted state-occupancy measure (Equation 4) and $p^{\\\\pi}\\\\_{\\\\gamma}(s\\\\_K = g|s\\\\_{0} = s, a)$ is the discounted state-occupancy measure with actions (Equation 5). This is the temporal distance to the goal $g$ conditioned on starting at state $s$ and taking action $a$ under an optimal policy. \\n2. We note that the minimization of $d(s,a,g)$ over actions $a$ is used for action selection: Myers et al. [1] show that $\\\\text{arg}\\\\min_a d(s,a,g)$ corresponds to $\\\\text{arg} \\\\max\\\\_a Q(s,a,g)$ of the Q-function with respect to the goal-conditioned RL MDP.\\n3. We highlight in Sections 3 and 4.2 that the reward for the MDP using the successor distance is $r(s) = \\\\delta\\\\_{(s,g)}$, which is a Kronecker delta function that evaluates to 1 at the goal and 0 at intermediate states.\\n4. We have revised notation in Section 4 and Appendix C to use more standard math terms. For example, we now use set notation to construct $\\\\text{arg} \\\\min$ rather than introduce a \\u201csymmetry-breaking\\u201d $\\\\text{arg} \\\\min$ over equally optimal actions and waypoints. We have also renamed certain objects for clarity (i.e. policy $\\\\pi^f\\\\Rightarrow \\\\pi^{\\\\text{Fix}}$ and planning operator $\\\\text{Plan}^{f} \\\\Rightarrow \\\\text{Plan}^{\\\\text{Fix}}$ for the deterministic, \\u201cfixed\\u201d goal, controlled setting). \\n\\nThe changes to the proofs for clarity have been highlighted in red. **Please let us know if additional clarifications will be helpful.**\\n\\n> How does the paper inspire future algorithm development?\\n\\nWe have added a new discussion in Appendix D about how our theoretical results might inspire future algorithm development, as well as a summary of the concrete takeaways in the conclusion, and a comparison of existing GCRL critic parameterizations and how they relate to quasimetric architectures ([Table 1](https://gcdnb.pbrd.co/images/3LYtfgmtGjGl.png)).\\nWe agree that the primary takeaway is not a new method, but rather (1) concrete definitions of planning invariance and horizon generalization, (2) proofs linking quasimetrics, planning invariance, and horizon generalization, and (3) experiments showing that horizon generalization is feasible and nontrivial. \\nThat said, our experiments show that the modified CMD algorithm [1] combined with the backward infoNCE [4] loss provides a good base algorithm that empirically shows some of these properties while being motivated by our theory.\\nWe believe these results are significant and in line with prior ICLR works that highlight a phenomenon/property of existing methods [2, 3].\\n\\n> Typos, use of \\u201clemma\\u201d\\n\\nThank you for these suggestions. We have fixed the typos and added definitions for the terms you noted, and additionally renamed the core (theoretical) results as \\u201ctheorems\\u201d instead of \\u201clemmas.\\u201d\\n\\n### References\\n\\n[1] Myers, V. et al., 2024. \\\"Learning Temporal Distances: Contrastive Successor Features Can Provide a Metric Structure for Decision-Making.\\\" *ICML*\\n\\n[2] Richens, J. et al., 2024. \\\"Robust Agents Learn Causal World Models.\\\" *ICLR*\\n\\n[3] Subramani, R. et al., 2024. \\\"On the Expressivity of Objective-Specification Formalisms in Reinforcement Learning.\\\" *ICLR*\\n\\n[4] Bortkiewicz, M. et al., 2024. \\\"Accelerating Goal-Conditioned RL Algorithms and Research.\\\" arXiv:2408.11052\"}", "{\"comment\": \"Thank you for the response. One thing that I am not clear on is how the definition of planning invariance is not vacuous in many cases, provided the planner is optimal. For example, let's consider a shortest path from start state $s_0$ to goal state $g$. Any planner operating at a coarser resolution must produce waypoints along this path. Then for any state along this path, we will have $\\\\pi(a|s, g) = \\\\pi(a|s, w)$, where $w$ is the next waypoint. So this (optimal) policy must also be planning-invariant. Is it the case that any optimal policy within some region of the state space, must also be planning-invariant within that region? It could be helpful to clarify the relationships between planner optimality, policy optimality, and planning invariance.\"}", "{\"title\": \"Response to Authors\", \"comment\": \"Thank you for your response.\\n\\nThanks for the update for the undefined symbols. This means that my original assessment is correct, and I didn't miss anything. I also noticed that the definitions of several key symbols, which are used in major results such as Theorem 1, 2, and Definition 3, are missing in the original submission. I'm curious as to why they were omitted.\\n\\nAdditionally, the authors did not address my concerns regarding the confusing use of the terms \\\"metric\\\" and \\\"quasimetric.\\\" This could make the paper difficult for readers to understand.\\n\\nThank you for addressing how this work could inspire future research. It would be beneficial if the authors could present stronger results in the next version, as it would significantly enhance the paper's impact.\"}", "{\"title\": \"Rebuttal follow-up\", \"comment\": \"Dear Reviewer,\\n\\nWe have worked hard to incorporate the review feedback by running new experiments and revising the paper. We'd really appreciate if you could confirm whether these changes address the concerns about the paper. If we have misunderstood any of the concerns, we'd like to learn that now so that we can further revise the paper or run additional experiments.\\n\\nThanks!\\n\\nThe Authors\"}", "{\"title\": \"Rebuttal follow-up\", \"comment\": \"Dear Reviewer,\\n\\nWe have worked hard to incorporate the review feedback by running new experiments and revising the paper. **Do the revisions and discussions above address your concerns?** We would greatly appreciate your engagement.\\n\\nThanks!\\n\\nThe Authors\"}", "{\"title\": \"Additional Response\", \"comment\": \"Thank you for your response.\\n\\n> Thanks for the update for the undefined symbols. This means that my original assessment is correct, and I didn't miss anything. I also noticed that the definitions of several key symbols, which are used in major results such as Theorem 1, 2, and Definition 3, are missing in the original submission. I'm curious as to why they were omitted.\\n\\nDue to space constraints, we initially moved some definitions to the appendix and omitted some standard definitions from prior work (for instance, the discounted state occupancy $p\\\\_\\\\gamma^\\\\pi$ notation you noticed is used in numerous prior works [1,2,3,4,5]; similarly the goal-conditioned reward notation is defined in [6,7,8]).\\n\\nWe now include these definitions in the main text to make the paper more self-contained. We hope the current revision addresses this concern.\\n\\n> Additionally, the authors did not address my concerns regarding the confusing use of the terms ''metric'' and ''quasimetric.'' This could make the paper difficult for readers to understand.\\n\\nThe formal construction of the distances in this paper all rely on the quasimetric definition, which relaxes the more familiar notion of a metric by allowing asymmetric distances. While this notion of asymmetry is important in many MDPs, the key benefit of viewing things as (quasi)metrics is the triangle inequality, which is what enables planning invariance and horizon generalization. We use the term ''metric'' in some places when we wish to highlight the triangle inequality property rather than asymmetry, in line with prior work that formally relies on quasimetrics [9,10]. Note that since the definition of a metric is stronger than that of a quasimetric, all our results hold for metrics as well. We hope this note as well as our latest revision clarify this point.\\n\\n> Thank you for addressing how this work could inspire future research. It would be beneficial if the authors could present stronger results in the next version, as it would significantly enhance the paper's impact.\\n\\nThank you for this suggestion. We can include results in additional AntMaze layouts [11] and the Humanoid environment [12] in our next revision. Future work may also benefit from evaluation on the newly-released OGBench baseline [13], which features a diverse set of long-horizon tasks in the *offline* setting.\\n\\n---\\n\\n### References\\n\\n[1] Eysenbach, B. et al., 2021. ''C-Learning: Learning to Achieve Goals via Recursive Classification.'' *ICLR*\\n\\n[2] Choi, J. et al., 2021. ''Variational Empowerment as Representation Learning for Goal-Conditioned Reinforcement Learning.'' *ICML*\\n\\n[3] Myers, V. et al., 2024. ''Learning to Assist Humans Without Inferring Rewards.'' *NeurIPS*\\n\\n[4] Schroecker, Y. and Isbell, C., 2020. ''Universal Value Density Estimation for Imitation Learning and Goal-Conditioned Reinforcement Learning.'' arXiv:2002.06473\\n\\n[5] Hoang, C. et al., 2021. ''Successor Feature Landmarks for Long-Horizon Goal-Conditioned Reinforcement Learning.'' *NeurIPS*\\n\\n[6] Fang, K. et al., 2022. ''Generalization With Lossy Affordances: Leveraging Broad Offline Data for Learning Visuomotor Tasks.'' *CoRL*\\n\\n[7] Yang, R. et al., 2023. ''What Is Essential for Unseen Goal Generalization of Offline Goal-Conditioned RL.'' *ICML*\\n\\n[8] Ghosh, D. et al., 2019. ''Learning Actionable Representations With Goal Conditioned Policies.'' *ICLR*\\n\\n[9] Liu, B. et al., 2023. ''Metric Residual Network for Sample Efficient Goal-Conditioned Reinforcement Learning.'' *AAAI*\\n\\n[10] Myers, V. et al., 2024. ''Learning Temporal Distances: Contrastive Successor Features Can Provide a Metric Structure for Decision-Making.'' *ICML*\\n\\n[11] Fu, J. et al., 2021. ''D4RL: Datasets for Deep Data-Driven Reinforcement Learning.'' arXiv:2004.07219\\n\\n[12] Gu, X. and Wang, Y., 2024. ''Advancing Humanoid Locomotion: Mastering Challenging Terrains With Denoising World Model Learning.'' *RSS*\\n\\n[13] Park, S. et al., 2024. \\\"OGBench: Benchmarking Offline Goal-Conditioned RL.\\\" arXiv:2410.20092\"}", "{\"title\": \"Rebuttal (cont.)\", \"comment\": \"> As the horizon changes and the trajectories get longer, it is possible that any of the waypoints relevant over shorter horizons are not that relevant anymore if better paths are discovered for example.\\n\\nIf the planning-invariant policy can navigate between any two pairs of nearby states with *complete state coverage* perfectly, then we expect these waypoints to remain relevant for longer horizons. However, there are settings where this complete state coverage assumption no longer holds and selected waypoints may not necessarily be optimal. We address these limitations through revisions in Section 4.5 and in the rest of this response. Our proofs show that waypoints will remain relevant for any long-horizon path, as waypoints consist of any intermediate state on the way to the goal. If an optimal waypoint is unseen, the policy will return suboptimal trajectories for longer horizon goals composed of shorter trajectories from the *covered* states (which may not be the true optimal waypoints). This is expected in the offline setting\\u2014if the optimal waypoint is never seen, whether as part of a short or a long trajectory in the training set, then it is challenging to learn a policy that navigates towards the true optimal waypoint.\\n\\n## References\\n\\n[1] Cohen, T., Welling, M., 2016. \\\"Group Equivariant Convolutional Networks.\\\" *ICML*\\n\\n[2] Benton, G. et al., 2020. \\u201cLearning Invariances in Neural Networks.\\u201d *NeurIPS*\\n\\n[3] Rowley, H., Baluja, S., Kanade, T., 1998. \\u201cRotation Invariant Neural Network-Based Facial Recognition.\\u201d *IEEE*\\n\\n[4] Tassa, Y. et al., 2018. \\\"DeepMind Control Suite.\\\" arXiv:1801.00690\\n\\n[5] Zheng, C. et al., 2023. \\\"Contrastive Difference Predictive Coding.\\\" *ICLR*\\n\\n[6] Fujimoto, S. et al., 2018. \\\"Addressing Function Approximation Error in Actor-Critic Methods.\\\" *ICML*\\n\\n[7] Park, S. et al., 2023. \\\"HIQL: Offline Goal-Conditioned RL With Latent States as Actions.\\\" *NeurIPS*\\n\\n[8] Ghugare, R. et al., 2024. \\\"Closing the Gap Between TD Learning and Supervised Learning--a Generalisation Point of View.\\\" *ICLR*\"}", "{\"title\": \"Rebuttal (cont.)\", \"comment\": \"### References\\n\\n[1] Zhang, A., McAllister, R. et al., 2021. \\u201cLearning Invariant Representations for Reinforcement Learning Without Reconstruction.\\u201d *ICLR*\\n\\n[2] Jang, E. et al., 2021. \\\"BC-Z: Zero-Shot Task Generalization With Robotic Imitation Learning.\\\" *CoRL*\\n\\n[3] Myers, V. et al., 2024. \\\"Learning Temporal Distances: Contrastive Successor Features Can Provide a Metric Structure for Decision-Making.\\\" *ICML*\"}" ] }

[ "Dmitry Ivanov", "Paul Duetting", "Inbal Talgam-Cohen", "Tonghan Wang", "David C. Parkes" ]

[ "reinforcement learning", "multi-agent systems", "contract design", "principal-agent MDP", "sequential social dilemmas", "neural networks" ]

{ "note_id": [ "zgLsxakupq", "w8mJsmIpNK", "uaAJmbO69e", "sq6FA60xtn", "hFenuOzgre", "ZZqZwASwSB", "ZTHvCLSwhA", "O5qQ6y3tAj", "LEpiaFHwaJ", "JxsEn5pxpm", "HkNQH7fUPq", "1LicC6sEhc", "04a6pandbR" ], "note_type": [ "official_comment", "official_review", "official_comment", "official_comment", "official_review", "decision", "official_comment", "official_comment", "official_review", "official_comment", "meta_review", "official_review", "official_comment" ], "note_created": [ 1732697879326, 1730887092109, 1732697963642, 1732698098203, 1730651296765, 1737524014167, 1732698070912, 1733013234225, 1729832822807, 1732802166666, 1734844198252, 1730551679966, 1733154751983 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission9919/Authors" ], [ "ICLR.cc/2025/Conference/Submission9919/Reviewer_qrta" ], [ "ICLR.cc/2025/Conference/Submission9919/Authors" ], [ "ICLR.cc/2025/Conference/Submission9919/Authors" ], [ "ICLR.cc/2025/Conference/Submission9919/Reviewer_3zbp" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission9919/Authors" ], [ "ICLR.cc/2025/Conference/Submission9919/Reviewer_nJAV" ], [ "ICLR.cc/2025/Conference/Submission9919/Reviewer_nJAV" ], [ "ICLR.cc/2025/Conference/Submission9919/Reviewer_3zbp" ], [ "ICLR.cc/2025/Conference/Submission9919/Area_Chair_5FLh" ], [ "ICLR.cc/2025/Conference/Submission9919/Reviewer_b1yZ" ], [ "ICLR.cc/2025/Conference/Submission9919/Reviewer_qrta" ] ], "structured_content_str": [ "{\"comment\": \"Thanks for the careful review and comments.\\n\\nWe believe that our work has some non-trivial theoretical insights, such as Theorem 3.4, which shows that the meta-algorithm is applying a contraction operator. This is crucial for the applicability of our results, because in the experimental parts we need to stop the deep Q-learning implementation short of convergence.\\n\\nIt is also not true that we don\\u2019t compare to existing baselines in the MARL literature (see reply to Q1 below). We do apologize that this was a bit hidden (we should have explicitly referred to the paper that inspired us to adopt this benchmark for comparison). \\n\\n**Re Q1:** Thanks for pointing us to [2,3] (we already cite [1]), we will be happy to include additional discussion. For gifting methods as baselines, we would like to stress that these are not directly comparable with our method as agents are subtracted rewards and there is no centralized principal; we already compare with this line of work by adapting a similar method of the other contract design in MARL paper (see citation below). This is the \\u201cconstant baseline\\u201d we refer to in Section 5 and Figures 2a and 2b. We found that this alternative is much less effective, achieving only around 85-90% of the welfare that our approach achieves (see the gap between the orange and gray lines in Figure 2a). \\n\\nPhillip JK Christoffersen, Andreas A Haupt, and Dylan Hadfield-Menell. Get it in writing: Formal contracts mitigate social dilemmas in multi-agent RL. In Proceedings of the 2023 International Conference on Autonomous Agents and Multiagent Systems, pp. 448\\u2013456, 2023.\\n\\n**Re Q2:** There is a profound difference between an \\u201cadaptive contract design approach\\u201d (our work) and an \\u201cadaptive mechanism design approach\\u201d. In short: These are two different branches in economics, dealing with very different incentive constraints. The emphasis on the former is on incentivizing effort, and a main obstacle is that the principal cannot directly observe how much effort the agents exert. We believe that this \\u201cincentivize effort\\u201d perspective is very adequate when thinking about creating marketplaces for AI agents, as these are quintessentially markets for services (effort). \\n\\nThis notwithstanding, we believe that ultimately we may want a solution that marries the two directions. Indeed, the focus of the adaptive mechanism design approach is that agents may have hidden types (think capabilities), and we would like to design mechanisms that incentivize agents to truthfully report their private types. \\n\\nWe would like to point to two simultaneous/subsequent works that also argue in favor of a contract design approach for AI agents. We believe that this provides additional evidence that this is a timely and important perspective.\\n\\nJibang Wu, Siyu Chen, Mengdi Wang, Huazheng Wang, Haifeng Xu. Contractual Reinforcement Learning: Pulling Arms with Invisible Hands. In ArXiv, 2024. https://arxiv.org/abs/2407.01458\\n\\nMatteo Bollini, Francesco Bacchiocchi, Matteo Castiglioni, Alberto Marchesi, Nicola Gatti. Contracting with a Reinforcement Learning Agent by Playing Trick or Treat. In ArXiv, 2024. https://arxiv.org/abs/2410.13520\"}", "{\"summary\": \"The paper proposes to apply principal--agent theory to multi-agent RL with the goal of applying their methods to resolve social dilemma-like situations between AI agents, allowing different AI agents to collaborate more effectively. They define principal--agent MDP as a stochastic game with turns, where at each time-step principal observes the state and chooses its action (a contract), the agent observes state and principal's action and then chooses its action. As a solution concept, they choose the subgame-perfect equilibrium. They present a meta-algorithm that is essentially a bilevel optimization problem, which when solved, proven to find the SPE in their setting. Then they propose a Q-learning based approach to instead learn the SPE rather than computing it exactly. They present an early theoretical result connecting the approximation error in Q functions to a bound on principal's utility. They test their approach in the Coin Game environment in terms of both social welfare and convergence.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": [\"The problem of orchestrating an ecosystem of AI agents to collaborate with each other is an interesting and relevant problem.\", \"The application of principal--agent theory and its combination with RL appears to be novel and also interesting.\", \"The paper is easy to read, and ideas are expressed clearly.\"], \"weaknesses\": [\"The presented empirical and theoretical results appear too weak. Convergence of the meta-algorithm is not surprising, what is left out is the computational complexity of solving the bilevel optimization problem. Empirical results are also limited, and there is no comparison to other MARL algorithms that tackle social dilemmas (see the question).\", \"The connection to the motivating example seems lost once the Introduction is over. More discussion at the end about how this allows for orchestrating of AI agents is needed.\"], \"questions\": \"1. Is there a reason why you have not benchmarked your method against other MARL methods that allow agents to gift each other rewards? For example [1,2,3]? I would expect these to be at least discussed as an alternative approach.\\n\\n2. For the case of orchestrating AI agents (your motivating example), what would be the advantage of your principal--agent framework over things like adaptive mechanism design?\\n\\n[1] Yang, Jiachen, et al. \\\"Learning to incentivize other learning agents.\\\" Advances in Neural Information Processing Systems 33 (2020): 15208-15219.\\n\\n[2]\\u00a0Lupu, Andrei, and Doina Precup. \\\"Gifting in multi-agent reinforcement learning.\\\" Proceedings of the 19th International Conference on autonomous agents and multiagent systems. 2020.\\n\\n[3]\\u00a0Kolumbus, Yoav, Joe Halpern, and \\u00c9va Tardos. \\\"Paying to Do Better: Games with Payments between Learning Agents.\\\" arXiv preprint arXiv:2405.20880 (2024).\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"We thank the reviewer for the careful review and comments.\", \"regarding_technical_novelty_and_significance\": \"We view our work of theory-backed practical algorithm design. We provide theoretical evidence (e.g., Theorem 3.4, showing that the meta-algorithm is iteratively applying a contraction operator) that corroborates our ultimate approach, which consists in adopting standard (scalable) ML pipelines (deep Q-learning) to solve a challenging highly combinatorial optimization problem.\\n\\nIt is true that prior work has considered similar approaches that modify agents\\u2019 rewards, but to the best of our knowledge the connection to contract design hasn\\u2019t been spelled out in this clarity. Contract design opens up new opportunities to design theoretically supported algorithms with the goal of minimizing the total payment. We believe that there is great value in bringing the two communities together. \\n\\nIt is also true that principal-agent theory can be understood more broadly, and encompasses additional directions. We will be happy to emphasize this, and include additional discussion.\\n\\nQ1) Thanks for pointing this out. We will make sure to give this part another pass, and go with one of the two options that you suggest.\\n\\nQ2) This is an interesting question. We believe both approaches have their pros and cons, and are worth exploring. We believe that the more centralized approach that we pursue in this work could have a lot of value, thinking of platforms that will serve to aggregate services offered by AI agents.\"}", "{\"comment\": \"We cordially thank the reviewer for these great suggestions!\\n\\nWe agree that exploring the sensitivity to the hyperparameter alpha (e.g., via an ablation study) and additional experiments would strengthen the paper. Regarding the latter, the other reviewers also make the case that it would be interesting to add additional experiments on more collaborative tasks (on which we should expect good/better results), in addition to focusing on very challenging settings in which the agents have very conflicting objectives (although more of these would of course be good, too). \\n\\nWe will work on both.\\n\\nThanks for pointing out the minor issues. We will fix these!\"}", "{\"summary\": \"This paper studies a problem where a principal must guide the actions of an agent (or several agents) to maximize the principal's utility using payments. This problem is formalized using a Markov decision process, where the principal chooses a policy that determines the payment to the agent depending on the state and outcome. The authors propose and analyze a simple meta-algorithm that ensures convergence to subgame perfect equilibrium (SPE). Moreover, experiments are conducted on the Coin Game environment, which is a medium-sized two-player social dilemma. The experiments show the effectiveness of the proposed approach for maximizing the principal's utility (in this case, social welfare).\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": [\"The intersection of reinforcement learning and contract design is an interesting area of study.\", \"The problem formulation is neat and well motivated.\", \"The paper is well-written and overall a pleasure to read.\"], \"weaknesses\": [\"I struggle with identifying the technical novelty and the significance of the contributions:\", \"From a technical standpoint, it appears that by considering SPE the problem becomes fairly simple and boils down to performing backwards induction (using standard RL techniques). In particular, as far as I can tell, the results of Theorem 3.3 and 3.4 follow straightforwardly from SPE+ well-known RL results. Could you highlight any specific novel technical challenges under the proposed problem setup?\", \"Conceptually, the idea of a principal orchestrating agents in social dilemma to resolve miscoordination and maximize social welfare has been considered before (this is also essentially what much of mechanism design is about). In Appendix A.3, you argue that your work differs in perspective from prior work on \\\"solving\\\" social dilemmas by modifying the agents' rewards. In particular, to distinguish your work from the others you write \\\"While the principal effectively modifies agents\\u2019 reward functions, the payments are costly.\\\" Are not the payments also costly in the models studied in prior work? For instance, in Yang et al. (2020), the agents transfer the rewards to other agents which is of course costly to the paying agent. Maybe the work of [Jackson and Wilkie (2002)](https://www.jstor.org/stable/3700662) is also relevant / of interest here.\", \"From my undestanding, principal-agent theory is broader than just contract design, and it would be helpful to the reader to discuss this (i.e., principal-agent theory) related work as well. For instance, [Myerson (1982)](https://www.sciencedirect.com/science/article/abs/pii/0304406882900064), [Zhang and Conitzer (2021)](https://arxiv.org/abs/2105.06008), [Gan et al. (2022)](https://arxiv.org/abs/2209.01146). There is also work on Bayesian persuasion which is part of the principal-agent problem family. Appendix A.2 does make it sound like principal-agent problems are just contract design problems.\"], \"questions\": [\"$E_t$ in Proposition 4.1 is not defined in the main text. Overall, Proposition 4.1 is very hand-wavy. For example, you write that \\\"This utility can be achieved through small(?) additional payments that counteract the agent's deviations, ...\\\". I'm aware that more details are provided in the appendix, but the proposition should be stated rigorously or, alternatively, be made an informal remark.\", \"In your opinion, why would the centralized orchestrating of agents (i.e., the principal-agent perspective) be more appealing to resolve incentive issues than the large-scale decentralized individual bargaining perspective many other works take? (You mentioned some of these works in Appendix A.3).\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"comment\": \"Thanks for the careful review and comments.\\n\\nThanks for the suggestion to try other more collaborative MARL tasks (such as division of labor tasks - e.g. clean up): We focused on tasks with conflicting objectives (such as in the coin game) as we believe these are the most challenging ones and hence most appealing ones for a contract approach (it\\u2019s not just about coordination of actions towards a joint objective, but intrinsically conflicting interest of the individual agents). It is also where we see the biggest gap in the literature. The rebuttal phase is too short to add further experiments, but we will be happy to work on adding other collaborative tasks to the list of experiments.\\n\\nRegarding technical depth, we would like to emphasize Theorem 3.4, which shows that the meta-algorithm is a contraction operator. This is crucial for the practical implementation using deep Q-learning, where the algorithms will need to be stopped short of convergence. We do validate the convergence/approximation properties empirically.\\n\\n**Re Q1:** We are not fully sure if we understand the question. Consider for a moment a single-shot contracting problem, the same rationale applies to the Markov Chain style games we consider. In the one shot game, a contract defines payments, and thus changes what the agent (or agents) consider as best responses to the contract. Intuitively, payments need to be higher if the agents\\u2019 interests are less well-aligned with each other and the principal. Conflicting interests jeopardize the stability of Q-learning, leading to instabilities or oscillations during the learning process. This has motivated us to propose and demonstrate that the meta-algorithm is a contraction, thereby enabling its practical implementation.\\n\\n**Re Q2:** Thanks! The idea here is that at equilibrium agents will necessarily be indifferent, making solutions prone to small errors in approximation. Think of the principal having an action plan in mind, with full knowledge, in an optimal contract agent(s) will be indifferent at each step where they have to make decisions. The idea behind nudging is that the principal can just increase payments a bit to robustify this solution. (In some sense move away from the discontinuous cliffs, into the interior of the best response regions.)\\n\\n**Re Q3:** Breach of contract. This is an interesting direction. But, of course, in many situations this would be prevented by legal provisions. We think that this could be an interesting future work to explore, but seems a bit like jumping two steps.\"}", "{\"comment\": \"Thanks for your reply, though I did not see the updates reported (please feel free to point out if I overlooked pdf update).\"}", "{\"summary\": \"This paper formulates a general framework of principal-agent Reinforcement Learning (RL), which tries to train multi-agent RL agents without centralized control, but instead by using a protocol where a principal agent tries to guide (intentionally misaligned) agent behavior by assigning payments. The paper first constructs a hidden-action principal-agent stochastic game, then propose an meta-algorithm for finding Subgame Perfect Equilibrium (SPE). After analyzing the theoretical property of such meta-algorithm, the paper discusses the possibility of learning it with RL, and then finally how to extend to multi-agent RL. Several experiments show that the proposed method works well in the coin game and tree MDPs\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. This paper addresses a very interesting problem as it considers a novel scenario where the principal needs to regulate agents' behavior but does not have centralized control on each agent. This is even more interesting considering the potential future use of regulating a herd of LLM agents with a learned principal, many of which can self-reflect and improve according to a protocol but we do not have training access to.\\n\\n2. This paper is well-written and well-structured. It builds up a framework with many complicated notions, but the paper managed to consider one component at a time: first it is the game itself, then an idealized solution, then function approximator (RL), and finally multi-agent. Generally, the paper is easy to follow.\\n\\n3. The paper gives detailed analysis and proof for its theoretical performance (e.g., convergence property), which is important for a new, economic-inspired RL framework. In the mean time, the main text is kept clean without too many unnecessary notations using intuitive explanations.\", \"weaknesses\": \"1. The hyperparameter, $\\\\alpha$ in line 406, could be tricky. As stated by the authors, $\\\\alpha$ is set as a heuristic approach to downweight the importance of payment minimization compared to social welfare (payment minimization cannot hurt social welfare), and thus $\\\\alpha$ should be set very small. However, if $\\\\alpha$ is indeed very close to $0$, the principal agent can simply use its reward to overwhelm all selfish intentions (e.g. sent 1% of its own reward to each individual agent, which is a very large amount for them) and the algorithm would degenerate to independent RL with cooperative objective. Thus, there exists some balance of $\\\\alpha$ that may require tuning. It would be better if the authors can conduct an ablation study on $\\\\alpha$ to show how it impacts the balance between social welfare and payment minimization.\\n\\n2. There is essentially only one gridworld experiment for the proposed algorithm in multi-agent RL (tree MDP in the appendix is for single-agent). It would be better if more and harder environments could also be tested, such as particle world ones (environments like SMAC-like ones would be even better for camera ready (if accepted) / next submission (if rejected)).\\n\\n**Minor issues**\\n\\n1. The notations in the beginning of Sec. 2.2 are unclear. For example, $o$ is an outcome, but is then used as a number (\\\"o-th coordinate of a contract\\\"); expressions like \\\"contract corresponds to the outcome $o$\\\" would be better.\\n\\n2. The punctuation \\\".\\\" at the end of the caption of Tab. 1 is missing.\\n\\n3. The current contribution part is not stating the paper's contribution, but is rather an outline of the paper. It would be better if the outlines can be more briefly summarized into a paragraph, with each highlight on key novel aspect as one sentence.\", \"questions\": \"I have a question: can the solution proposed by this framework improve social welfare on a more complicated, closer-to-mainstream MARL environment (e.g. a continuous version of coin game in a multi-agent particle world style)?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you for taking the time to respond.\\n\\nI agree that emphasizing the connection to contract design is interesting and you do this well. However, I still believe that the paper offers only very few novel technical / methodological contributions and insights beyond establishing a connection to contract design. I will keep my original borderline score.\"}", "{\"metareview\": \"This paper reduces mechanism design for coordinating AI agents to solving a turn-based stochastic game. I found the paper to be well-written and easy to follow. However, both the reviewers and I share the concern regarding the limited technical novelty and the significance of the contribution.\\n\\nThe reduction from contract design to stochastic games is fairly straightforward, and the fact that turn-based stochastic games can be solved using simple DP is largely considered folklore within the community. Additionally, Reviewer 3zbp noted that the conceptual idea of a principal orchestrating agents in social dilemmas to address miscoordination and maximize social welfare has been explored in prior work.\\n\\nAs a result, I feel that the contribution of this paper is marginal and falls below the standard for acceptance.\", \"additional_comments_on_reviewer_discussion\": \"The primary concerns raised by the reviewers revolve around the novelty and technical depth of the work. I feel these issues were not adequately addressed during the rebuttal phase, which is the main reason I lean toward rejection.\"}", "{\"summary\": \"This work introduces a specific contract mechanism into MARL, defined a principal-agent stochastic game based on this mechanism, and solved the game using a straightforward value-based RL algorithm, conducting some experiments as well.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": [\"Introducing contracts into MARL is an issue of interest to the community.\", \"This work is comprehensive, thoroughly addressing the bilevel MARL problem related to the contract mechanism.\"], \"weaknesses\": [\"Algorithm 1 is vanilla, and lines 3 and 4 imply solving an MDP, which can be very time-consuming in a larger environment. Although a value-based method is used and is model-free, it is a primitive algorithm. Some analyses are straightforward to the community. The main contribution lies in introducing the contract mechanism and reformulating the game.\", \"Is the contract in this work a mechanism, or is it a problem to be solved?\", \"As a problem: The contract reformulates a new game, and the logic of the experimental design is to verify that their proposed algorithm can solve this problem. So, could some other existing algorithms potentially solve this problem?\", \"As a mechanism: For instance, with this contract mechanism, the prisoner\\u2019s dilemma can be effectively resolved, and the coin game can achieve high scores. In this sense, this work should compare with other algorithms to demonstrate the advantages of the contract. The experiments in Figure 2 are clearly insufficient.\", \"The contract has a very specific meaning here, where after an agent takes an action and a result is achieved, the principal pays the agent based on the result. This setting is a variant of incentivizing agents.\", \"In this way, this setting is closely related to LIO (\\\"learning to incentivize other learning agents\\\"). The authors cited LIO but did not discuss the differences or conduct comparative experiments.\", \"In LIO, players are symmetric and do not need to wait for an optimization problem to converge at one level (in the spirit of online cross-validation). And the incentivization occurs during the training phase without changing the game\\u2019s structure. So, could similar methods be directly applied?\", \"Coin Game might not highlight the advantages of the contract. A more suitable scenario would be a task with division of labor, such as the cleanup task. There is already work on MARL contracts in this scenario, \\\"Formal Contracts Mitigate Social Dilemmas in Multi-Agent Reinforcement Learning.\\\" And LIO also uses this scenario.\", \"Minors\", \"There is a large margin below Figure 1 and Figure 2.\", \"Some single quotation marks in the text should be changed to double quotation marks.\"], \"questions\": \"1. How do the distinct and possibly conflicting objectives affect the meta-algorithm bilevel optimization procedure compared to same downstream tasks?\\n2.\\u00a0Nudging is an interesting solution for learning approximation error. However, who provides such nudging, i.e. in equations (20)-(22)? How does nudging change the discontinuous utility of the principle?\\u00a0\\n3.\\u00a0Besides, in the contracts of our general life, breach of contract is also an effective method to steer an agent\\u2019s behavior. What is the relationship between breach and nudging?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you for your response. This clarifies the baseline question for me. To be clear, in gifting mechanisms, there are many different models such as: giving each agent a separate gift budget or making gift sending a no-cost event. These could have been tested against your approach. Also, even if sending gifts would cost to the sender, this still does not invalidate them as a baseline approach to orchestration. Reward is a made up quantity in reinforcement learning, and unlike economics and game theory, does not need to correspond to things like monetary value or actual utility gain. It is a signal that induces desired behaviour.\\n\\nSimilarly, your comparison to adaptive mechanism design is also not clear to me. Adaptive mechanism design is literally about learning how to influence the payoff structure in order to induce desired equilibrium behaviour amongst players. For work on this that is very similar to your objective see: \\n\\n\\\"We consider the problem of how an\\nexternal agent can promote cooperation between artificial learners by\\ndistributing additional rewards and punishments based on observing\\nthe learners\\u2019 actions. We propose a rule for automatically learning\\nhow to create the right incentives by considering the players\\u2019 anticipated parameter updates.\\\" [1]\\n\\n[1] Baumann, Tobias, Thore Graepel, and John Shawe-Taylor. \\\"Adaptive mechanism design: Learning to promote cooperation.\\\" 2020 International Joint Conference on Neural Networks (IJCNN). IEEE, 2020.\\n\\n\\n** Overall: ** I have a positive sense about the paper, so I will maintain my score. But I believe someone other than me must champion its acceptance.\"}" ] }

[ "Shuo Shao", "Haozhe Zhu", "Hongwei Yao", "Yiming Li", "Tianwei Zhang", "Zhan Qin", "Kui Ren" ]

[ "Model Fingerpinting", "Ownership Verification", "Model Copyright Protection", "Trustworthy ML" ]

{ "note_id": [ "ytvE0Fnc6k", "tBbdhzadwi", "rx6tlko281", "phYCETPDK9", "pghElNdb0R", "ojlSwl5yKf", "mhc2FOOlym", "kJKsfskMyP", "i4JnB0vHm0", "gEtHvpY7Ne", "fODyjwEdUD", "dHAkXDVaBV", "aOcxVV9RQD", "ZsXHohVYSo", "ZQ16AH5oop", "YJS99tcJkE", "XAFcT2DdHQ", "Wj3vsMcqQg", "TZGtS2aPB4", "OWzfI4V7l3", "NrnGURj2yT", "NkwLKvbKMj", "MegtTN4J24", "L994vqqDQY", "KmKRggxr9Q", "Id7NJyOymc", "IT9NmdQR2Z", "HVeaKkeift", "GwOXd3OGiX", "F37ppuwuYb", "9IUqm3g7Wx", "7585TcBluF", "70kguB2ywg", "5wdBgz3DgT", "3mYUhh4Hoy", "2YvFY011Co" ], "note_type": [ "official_comment", "meta_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "decision", "official_review", "official_review", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment" ], "note_created": [ 1732513741194, 1734133085383, 1732803941258, 1732015076488, 1732014773428, 1733120411740, 1732803899886, 1733120311635, 1732455851919, 1732015002822, 1732015519454, 1732014833404, 1732015235613, 1732513774050, 1732014490174, 1732324003984, 1732673949808, 1732014585197, 1732015306495, 1732323934061, 1732015189726, 1737523729855, 1730466706595, 1730619460756, 1732398416874, 1730420551175, 1732015354400, 1732015118542, 1732014952894, 1730606700663, 1732014717559, 1732014632488, 1732323736300, 1732323848371, 1732014888308, 1732445837905 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission5867/Authors" ], [ "ICLR.cc/2025/Conference/Submission5867/Area_Chair_qhR4" ], [ "ICLR.cc/2025/Conference/Submission5867/Authors" ], [ "ICLR.cc/2025/Conference/Submission5867/Authors" ], [ "ICLR.cc/2025/Conference/Submission5867/Authors" ], [ "ICLR.cc/2025/Conference/Submission5867/Authors" ], [ "ICLR.cc/2025/Conference/Submission5867/Authors" ], [ "ICLR.cc/2025/Conference/Submission5867/Authors" ], [ "ICLR.cc/2025/Conference/Submission5867/Authors" ], [ "ICLR.cc/2025/Conference/Submission5867/Authors" ], [ "ICLR.cc/2025/Conference/Submission5867/Authors" ], [ "ICLR.cc/2025/Conference/Submission5867/Authors" ], [ "ICLR.cc/2025/Conference/Submission5867/Authors" ], [ "ICLR.cc/2025/Conference/Submission5867/Authors" ], [ "ICLR.cc/2025/Conference/Submission5867/Authors" ], [ "ICLR.cc/2025/Conference/Submission5867/Authors" ], [ "ICLR.cc/2025/Conference/Submission5867/Authors" ], [ "ICLR.cc/2025/Conference/Submission5867/Authors" ], [ "ICLR.cc/2025/Conference/Submission5867/Authors" ], [ "ICLR.cc/2025/Conference/Submission5867/Authors" ], [ "ICLR.cc/2025/Conference/Submission5867/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission5867/Reviewer_LMaC" ], [ "ICLR.cc/2025/Conference/Submission5867/Reviewer_toL1" ], [ "ICLR.cc/2025/Conference/Submission5867/Area_Chair_qhR4" ], [ "ICLR.cc/2025/Conference/Submission5867/Reviewer_jEMa" ], [ "ICLR.cc/2025/Conference/Submission5867/Authors" ], [ "ICLR.cc/2025/Conference/Submission5867/Authors" ], [ "ICLR.cc/2025/Conference/Submission5867/Authors" ], [ "ICLR.cc/2025/Conference/Submission5867/Reviewer_oTwh" ], [ "ICLR.cc/2025/Conference/Submission5867/Authors" ], [ "ICLR.cc/2025/Conference/Submission5867/Authors" ], [ "ICLR.cc/2025/Conference/Submission5867/Authors" ], [ "ICLR.cc/2025/Conference/Submission5867/Authors" ], [ "ICLR.cc/2025/Conference/Submission5867/Authors" ], [ "ICLR.cc/2025/Conference/Submission5867/Reviewer_jEMa" ] ], "structured_content_str": [ "{\"title\": \"A Gentle Reminder of the Post-rebuttal Feedback\", \"comment\": \"Thank you very much again for your initial comments. They are very valuable for improving our work. We would be grateful if you could have a look at our response and modifications and please let us know if there is anything else that can be added to our next version. We are also willing to have further discussions with you.\"}", "{\"metareview\": \"In the work, the authors studied ``Model fingerprinting,'' which was adopted to verify whether a suspicious model is reused from the original source one.\\n\\nThe authors have the concern regarding the comments from Reviewer LMaC that may be completely LLM-written (although the AC has no 100% evidence).\\n\\nTherefore, the decision is made without considering Reviewer LMaC's comments.\\n\\nDuring the rebuttal process, the authors have addressed more comments from the reviewers except that a theoretical analysis is still a lack.\\n\\nIn particular, although the Reviewer jEMa raised the score, this work still cannot provide theoretical support.\\nSpecifically, Reviewer jEMa concerns ``the current manuscript lacks a deeper analysis or evaluation on the transferability of the targeted fingerprints. ... However, it is unclear whether a more sophisticated adversary could forge a false claim on FIT-Print (e.g., by exploiting techniques from transferrable targeted adversarial attacks). Apart from a lack of formal proof (which is a noteworthy but understandable limitation, given the complexity of neural networks), the current adaptive evaluation is also limited (in Section 4.4, the adaptive attacker only leverages additional independent models). Hence, a more in-depth analysis or evaluation on this aspect would help make the claim more persuasive.''\\nThe AC agrees with the point and this submission at its current status did not meet the high standard of ICLR.\", \"additional_comments_on_reviewer_discussion\": \"The authors still cannot provide theoretical support for their work.\\n\\nThe AC has also contacted with Reviewer LMaC, who was assigned to review this paper by the Review System instead of the AC.\\nReviewer LMaC provided concrete evidences that the review comments came from him/her but were polished by ChapGPT.\"}", "{\"title\": \"A Second Reminder of the Post-rebuttal Feedback\", \"comment\": \"Dear Reviewer oTwh,\\n\\nWe greatly appreciate your initial comments. We totally understand that you may be extremely busy at this time. But we still hope that you could have a quick look at our responses to your concerns. We appreciate any feedback you could give us. We also hope that you could kindly update the rating if your questions have been addressed. We are also happy to answer any additional questions before the rebuttal ends.\\n\\nBest Regards,\\n\\nPaper 5867 Author(s)\"}", "{\"title\": \"Author Response (Part V)\", \"comment\": \"**Q6**: Unclear model relationships: The symbols for models (e.g., Mo, Ms, Mi) are not clearly explained, particularly in the context of the \\\"false claim of ownership of M1.\\\" Providing clearer descriptions and diagrams would improve understanding of how these models interact in the scenario.\\n\\n**R6**: Thank you for the comment! Hereby we provide further clarification about the symbols for models.\\n\\n- $M_o$: the original (or source) model which is developed by the model owner. $M_o$ is the model we need to protect. \\n- $M_s$: the suspicious model which is suspected to be reused from the source model $M_o$.\\n- $M_i$: the independently trained model owned by other parties.\\n\\nThe subscripts of these symbols are the first letters of 'original', 'suspicious', and 'independent'. We have explained the definitions of these symbols in *Section 2*.\\n\\n---\\n\\n\\n**Q7**: Lack of illustrations: The paper lacks sufficient diagrams to explain the overall fingerprint verification framework. The relationships between models like Mo, Ms, and Mi are unclear, making it harder for readers to follow. Diagrams would help visualize these interactions and clarify the verification process.\\n\\n\\n**R7**: Thank you for the comment. We have integrated the introduction to the relationships between $M_o, M_s, M_I$ in Figure 6 in Appendix A. We also provide detailed clarification of the process of fingerprint verification. Please kindly refer to **R5&6** for more details.\\n\\n\\n---\\n\\n**Q8**: Be careful when use of terminology: Terms like \\\"Proposition,\\\" \\\"Definition,\\\" and \\\"Theorem\\\" are used in the paper but need to be applied with more caution. It is recommended to briefly introduce these terms in the introduction and clarify their specific contexts throughout the paper to help readers better understand their relevance and contribution to the argument.\\n\\n**R8**: Thank you for the comment. We respectfully note that we have strictly followed the definitions and rules in mathematics [1] to utilize the terms 'Proposition', 'Definition', and 'Theorem'. We believe these terms are common sense and overinterpretation of these terms may weaken the focus of our paper.\\n\\n**Reference**\\n1. Mathematical Writing. 1989.\\n\\n---\\n\\n**Q9**: Concerns about the Mechanism or Inadequate Consideration/Explanation of Vulnerabilities. Fingerprinting scenario considerations: The paper does not fully explore whether adversarial samples generated by the model owner could also affect models not owned by them, given the known transferability of such samples. This scenario needs further investigation to assess its potential impact on the verification process.\\n\\n**R9**: Thank you for the comment. We hereby make further clarification to alleviate your concerns.\\n\\n- **The issue of transferability is actually the core of our study.** False claim attacks aim to craft transferable fingerprints so that the adversary can falsely claim the ownership of other parties' models. The success of false claim attacks depends on the transferability of model fingerprints.\\n- **We tackle the issue of false claim attacks ($i.e.$, transferability) via targeted fingerprinting.** Our main insight is that untargeted fingerprints enlarge the space of viable fingerprints while targeted fingerprints can decrease the transferability and reduce the probability of successful false claim attacks. Arguably, **such an insight can also be found in the area of adversarial attacks ($i.e.$, targeted adversarial examples have lower transferability than untargeted ones).** The papers from the top peer-reviewed conferences [1,2] also present the proposition. It is also empirically validated as depicted in Table 1 of [3].\\n- **We empirically validate that our proposed methods have low false positive rates and can resist adaptive false claim attacks**. The experimental results are shown in *Section 4.2&4.4*.\\n\\n**Reference**\\n1. Towards Transferable Targeted Attack. CVPR, 2020.\\n2. Towards Transferable Targeted Adversarial Examples. CVPR, 2023. \\n3. LFAA: Crafting Transferable Targeted Adversarial Examples with Low-Frequency Perturbations. arXiv, 2023.\"}", "{\"title\": \"Author Response (Part II)\", \"comment\": \"**Q2**: Another major concern is the missing literature. There is a work on KDD23 which also leverages a mapping function (they call it a meta-verifier) and adaptive input samples to build the model fingerprint. What is the different in methodology? The authors did not even mention the work. They should at least compare with it and show the substantial contribution over the work.\\n[1] https://dl.acm.org/doi/10.1145/3534678.3539257\\n\\n\\n**R2**: Thank you for suggesting the outstanding work! We are deeply sorry for missing this paper. We hereby make further clarification to alleviate your concerns.\\n\\n- **Brief Introduction to MetaV**: MetaV[1] introduces two critical components, the adaptive fingerprint and the meta-verifier. The adaptive fingerprint is a set of adversarial examples. The meta-verifier takes the suspicious model's output of the adaptive fingerprint and outputs whether the suspicious model is reused from the original model. MetaV accomplishes such an objective by simultaneously optimizing the adaptive fingerprint ($i.e.$, adversarial perturbations) and the meta-verifier ($i.e.$, a fully-connected neural network). In conclusion, MetaV provided a task-agnostic fingerprinting framework. **MetaV can be regarded as one of the adversarial example-based (AE-based) fingerprinting methods**.\\n- **The Advantages of our FIT-Print over MetaV**:\\n - **MetaV is vulnerable to false claim attacks.** MetaV is an AE-based fingerprinting method and the adversary can craft transferable adversarial examples to achieve false claim attacks. This proposition is also presented in [2].\\n - **MetaV cannot detect transfer learning models**, which is one of the realistic stealing settings. MetaV depends on a pre-trained meta-verifier. Transfer learning models may have different output formats, $e.g.$, the number of classes. Therefore, the meta-verifier which has a fixed input format is not able to process the changed outputs of the suspicious model and detect whether it is reused from the original model.\\n- We also empirically evaluate the effectiveness of MetaV. Table 1 shows that **our FIT-ModelDiff and FIT-LIME outperform all the baseline methods (including MetaV)**.\\n\\n\\nWe sincerely thank you again for the provided outstanding reference. We also add a detailed comparison to this work in *Appendix M.2* of our revision. If you have any suggestions on other references, we are also willing to make further clarifications and discussions :)\\n\\n\\n\\n\\n**Table 1.** Successful ownership verification rates of different model fingerprinting methods.\\n\\n| Reuse Task$\\\\downarrow$ | #Models$\\\\downarrow$ | IPGuard | MetaV | ModelDiff | Zest | SAC | ModelGiF | FIT-ModelDiff (ours) | FIT-LIME (ours) |\\n| -------- | -------- | -------- |-------- | -------- | -------- | -------- |-------- |-------- |-------- |\\n| Copying | 4 | 100% | 100% | 100% | 100% | 100% | 100% | 100% | 100% |\\n| Fine-tuning | 12 | 100% | 100% | 100% | 100% | 100% | 100% | 100% | 100% |\\n| Pruning | 12 | 100% | 100% | 100% | 91.67% | 100% | 100% | 100% | 100% |\\n| Extraction | 8 | *50%* | 87.5% | *50%* | *25%* | 100% | 100% | 100% | 100% |\\n| Transfer | 12 | *N/A* | *N/A* | 100% | *N/A* | *0%* | 100% | 100% | 100% |\\n| Independent | 144 | 30.6% | 4.8% | 4.0% | 7.6% | 39.6% | 0.0% | 0.0% | 0.0% |\\n\\n\\n**Reference**\\n1. MetaV: A Meta-Verifier Approach to Task-Agnostic Model Fingerprinting. SIGKDD, 2022.\\n2. False Claims against Model Ownership Resolution. USENIX Security, 2024.\"}", "{\"title\": \"Reminder of the Post-rebuttal Feedback and Summary of Our Response\", \"comment\": \"Dear Reviewer oTwh,\\n\\nThank you for your time and effort in evaluating our work. We greatly appreciate your initial comments. Your insights and suggestions are extremely valuable to us.\\n\\nGiven that we have only *one day* left for discussion, we are hoping to receive any additional feedback or question you might have at your earliest convenience. We totally understand that you may be busy at this time. But we still hope that you could have a quick look at our responses to your concerns. Your expertise would be of great help to us in improving the quality and rigor of our work.\\n\\nTo facilitate the discussion, we would like to summarize our response as follows.\\n\\n- **We clarified the rationality and necessity of our setting and threat model**. Introducing a third party to manage the model copyrights is necessary in practice and the false claim attack investigated in our paper also exists even without the third party. Our work is meaningful due to plausible scenarios and foreseeable future needs.\\n- **We discussed the suggested related work MetaV [1] and empirically compared it with our method**. The experimental results demonstrated that our method outperforms MetaV.\\n\\nIf our responses address your concerns, we kindly request that you reconsider your evaluations. We would also be grateful for any additional comments or suggestions you might have to refine our work.\\n\\nBest regards,\\n\\nPaper 5867 Author(s)\\n\\n**Reference**\\n\\n1. MetaV: A Meta-Verifier Approach to Task-Agnostic Model Fingerprinting. SIGKDD, 2022.\"}", "{\"title\": \"A Second Reminder of the Post-rebuttal Feedback\", \"comment\": \"Dear Reviewer toL1,\\n\\nWe greatly appreciate your initial comments. We totally understand that you may be extremely busy at this time. But we still hope that you could have a quick look at our responses to your concerns. We appreciate any feedback you could give us. We also hope that you could kindly update the rating if your questions have been addressed. We are also happy to answer any additional questions before the rebuttal ends.\\n\\nBest Regards,\\n\\nPaper 5867 Author(s)\"}", "{\"title\": \"Reminder of the Post-rebuttal Feedback and Summary of Our Response\", \"comment\": \"Dear Reviewer toL1,\\n\\nThank you for your time and effort in evaluating our work. We greatly appreciate your initial comments. Your insights and suggestions are extremely valuable to us.\\n\\nGiven that we have only *one day* left for discussion, we are hoping to receive any additional feedback or question you might have at your earliest convenience. We totally understand that you may be busy at this time. But we still hope that you could have a quick look at our responses to your concerns. Your expertise would be of great help to us in improving the quality and rigor of our work.\\n\\nTo facilitate the discussion, we would like to summarize our response as follows.\\n\\n- **We clarified and added more technical details of our method**, including how to choose the target fingerprint $F$, how to design a good mapping function, and how and why the loss function (Eq. 6 in our paper) works.\\n- **We discussed the suggested related work [1]**. [1] focuses on a different task that utilizing fingerprinting for model integrity verification.\\n- **We also discussed the generalization of our method to a more general scope of models**, such as LLMs and diffusion models.\\n\\nIf our responses address your concerns, we kindly request that you reconsider your evaluations. We would also be grateful for any additional comments or suggestions you might have to refine our work.\\n\\nBest regards,\\n\\nPaper 5867 Author(s)\\n\\n**Reference**\\n\\n1. Sensitive-sample fingerprinting of deep neural networks. CVPR, 2019.\"}", "{\"title\": \"Thank You for Your Positive Feedback!\", \"comment\": \"Dear Reviewer jEMa:\\n\\nThank you so much for your positive feedback! It encourages us a lot.\\n\\nWe agree that our approach is only a significant step toward addressing false claim attacks rather than a thorough solution to the problem. For instance, we fail to provide a formal guarantee of its resistance against false claim attacks, although we have empirically verified it. As such, a more powerful adversary may still be able to conduct successful false claim attacks. We add an additional discussion on it in *Appendix O*, as follows.\\n\\n```!\\nAnother potential limitation is that FIT-Print does not provide formal proof of the resistance to false claim attacks. As such, a more powerful adversary may still be able to conduct a successful false claim attack. We will investigate how to achieve a certified robust model fingerprinting method against false claim attacks in our future work.\\n```\\n\\nWe hope that our work can be a concrete step towards false-claim-resistant model fingerprinting and inspire subsequent works.\\n\\nThank you again for your valuable time and insightful review!\\n\\nBest regards,\\n\\nPaper 5867 Author(s)\"}", "{\"title\": \"Author Response (Part IV)\", \"comment\": \"**Q5**: Unclear definition of the scenario and protocol: The paper does not clearly define each party involved in the fingerprint verification scenario, nor the specific steps they execute in the protocol. For instance, the roles of each party, the information they have access to, and the actions they perform should be clearly explained, so that readers can better understand how the entire verification protocol works. Similarly, providing more specific steps and details for both the attack and defense mechanisms\\u2014such as the inputs, outputs, who executes each step\\u2014would greatly improve the understanding of these processes.\\n\\n\\n**R5**: Thank you for the comment. We agree that a detailed introduction to the process of model fingerprinting and the threat model of false claim attacks can improve our paper.\\n\\n- **We have discussed the assumptions and roles of different parties in *Section 2.1***.\\n- **We also add a detailed introduction and a figure in *Appendix A* to clarify the threat models and processes**, as follows.\\n\\n**A Detailed Threat Models**\\n\\nIn this section, we provide a detailed introduction to the threat models of model fingerprinting and false claim attacks. Three parties involved in the threat models are depicted in Figure 6.\\n\\n**A.1 Detailed Threat Model of Model Fingerprinting**\\n\\nThere are three parties involved in the threat model of model fingerprinting, including the model developer, the model reuser, and the verifier. The model developer trains a model and the model reuser attempts to steal and reuse this model. The verifier is responsible for fingerprint registration and ownership verification. The assumptions of these three parties can be found in Section 2.1.\\n\\n**Process of Model Fingerprinting**. Model fingerprinting can be divided into three steps, including fingerprint generation, fingerprint registration, and ownership verification.\\n\\n1. **Fingerprint Generation**: In this step, the model developer trains its source model $M_o$ and generates the fingerprint of $M_o$.\\n2. **Fingerprint Registration**: After generating the fingerprint, the model developer registers the fingerprint and the model with a timestamp to a trustworthy third-party verifier.\\n3. **Ownership Verification**: For a suspicious model $M_s$ that could be a reused version of $M_o$, the verifier will first check the timestamps of these two models. If the registration timestamp of $M_s$ is later than $M_o$, the verifier will further check whether the fingerprint of $M_o$ is similar to the fingerprint $M_s$. If so, the suspicious model can be regarded as a reused version of $M_o$.\\n\\n**A.2 Detailed Threat Model of False Claim Attacks**\\n\\nThere are three parties involved in the threat model of false claim attacks, including the malicious developer, the verifier, and an independent developer. The formal definition of false claim attacks can be found in Section 2.3.\\n\\n**Assumption of the Malicious Model Developer**. In false claim attacks, the malicious developer is the adversary who aims to craft and register a *transferable* fingerprint to falsely claim the ownership of the independent developer's model $M_I$. The malicious developer is assumed to have adequate computational resources and datasets to train a high-performance model and carefully craft transferable model fingerprints. The primary goal of the malicious developer is that the registered model fingerprints can be verified in as many other models as possible. By generating the transferable fingerprint, the malicious developer can (falsely) claim the ownership of any third-party models (that are registered later than that of the malicious developer).\\n\\n**Process of False Claim Attacks**. The process of false claim attacks can also be divided into three steps, including fingerprint generation, fingerprint registration, and false ownership verification. \\n\\n1. **Fingerprint Generation**: In this step, the model developer trains its source model $M_o$ and attempts to generate a *transferable* fingerprint of $M_o$.\\n2. **Fingerprint Registration**: After generating the fingerprint, the model developer registers the *transferable* fingerprint and the model with a timestamp to a trustworthy third-party verifier.\\n3. **Ownership Verification**: The adversary tries to use the transferable fingerprint to falsely claim the ownership of another independently trained model $M_I$. Since the fingerprint is registered beforehand, the ownership verification won't be rejected due to the timestamp. Subsequently, the benign developer may be accused of infringement.\"}", "{\"title\": \"Concerns Regarding the Potential LLM-written Review from Reviewer LMaC\", \"comment\": \"Dear Chairs and Reviewers:\\n\\nGreetings and wish you all the best. We deeply and sincerely thank you and all reviewers for your valuable time and comments on our paper. Your efforts greatly help us to improve our paper.\\n\\nIn this letter, we would like to express our concern regarding **the potential completely LLM-written review of Reviewer LMaC**. We exploit three different AI content detectors ([[link1]](https://quillbot.com/ai-content-detector), [[link2]](https://copyleaks.com/ai-content-detector), [[link3]](https://sapling.ai/ai-content-detector)) from the Internet to check whether the review is AI-generated. The results show that **over 94% of contents in the review of Reviewer LMaC is AI-generated**.\\n\\nMore importantly, **this review is highly biased and even erroneous**. Some examples are as follows:\\n\\n- The review of Reviewer LMaC **simply negates our contributions** by omitting our discussions, explanations for our insight in Section 1, and our experiments which demonstrate our contributions in Section 4. \\n- **Some comments of the review are unreasonable and unfounded**. \\n - Reviewer LMaC comments that naming a method can only use the first letter and requires us to explain some common-sense mathematical terms such as 'Theorem' and 'Definition'. \\n - Reviewer LMaC claims that we overuse the abbreviation 'p1 and p2'. However, we actually use the full name 'Proposition 1 and Proposition 2' in our paper.\\n\\n**We sincerely hope that you can take notice of these issues and ignore his/her follow-up comments to our paper**. We believe that this reviewer might undermine the impartiality and fairness of the ICLR reviewing process, although we believe that we have addressed his/her concerns in our rebuttal. More detailed justifications are as follows:\\n\\n- Considering that this reviewer gave us an extremely negative score of 1 and that his review comments were entirely generated by LLM, **we have to worry about whether he/she had a preconceived extremely negative stance on our work from the very beginning**, and even purposely and directly asked LLM to write a very negative comment for the purpose of rejecting the paper in bad faith. \\n- As emphasized in the email to the ICLR reviewers, '**the use of LLMs to write full reviews is not allowed, LLM written reviews are easy to spot and the program committee will be rejecting such reviews**'. \\n- Considering his/her previous negative and evil preconceived position and our accusations against him/her, **he/she might deliberately conduct malicious guidance in the subsequent discussion to cover up his/her malicious purpose and complete his/her malicious rejection of our manuscript**.\\n\\n**We respect the opinions of the reviewers, but we sincerely hope that our work will be treated objectively and fairly**. We have poured our hearts and souls into this work, and we believe it deserves to be treated fairly. We are deeply sorry for the inconvenience that our notifications may cause you. Thank you again for all your kind and valuable efforts in the whole reviewing process.\\n\\nBest Regards,\\n\\nPaper 5867 Author(s)\"}", "{\"title\": \"Author Response (Part I)\", \"comment\": \"Dear Reviewer LMaC, thank you very much for your review of our paper. We are encouraged by your positive comments on our **efforts regarding transparency and reproducibility**. We hope the following responses can help clarify potential misunderstandings and alleviate your concerns.\\n\\n---\\n\\n**Q1-1**: Contribution and Innovation Deficiency Contribution: There are some issues with the contributions mentioned in the paper: Contribution 1: False claim attack is not a new type of attack. This type of attack has already been introduced in previous work, and thus, it should not be considered a contribution of this paper.\\n\\n**R1-1**: Thank you for the comment. However, we believe there is some potential misunderstanding. \\n- In our paper, **we do not claim that we propose a new type of attack**. Instead, as clearly stated in our contributions (Line 89-90), we revisit existing model fingerprinting methods and **design an effective attack implementation** to reveal that existing methods are vulnerable to false claim attacks. \\n- We also clearly stated that 'the concept and definition of false claim attacks were initially introduced in [1] and primarily targeted at attacking model watermarking methods' in our first footnote (Line 107).\\n\\nIf you can kindly provide the related reference, we are also willing to make further clarification.\\n\\n---\\n\\n**Q1-2**: Contribution 2: The advantages of the proposed targeted verification method (compared to other verification methods) have not been sufficiently explained. Without clearly demonstrating the superiority of the targeted approach over other methods, it is difficult to view this as a valid contribution. \\n\\n**R1-2**: Thank you for the comment. \\n\\n- **Our methods outperform existing methods primarily due to the proposed FIT-Print paradigm.** Our FIT-Print leverages *targeted* fingerprinting to decrease the transferability of fingerprints and resist false claim attacks. Our main insight is that untargeted fingerprints enlarge the space of viable fingerprints while **targeted fingerprints can decrease the transferability and reduce the probability of successful false claim attacks**. \\n- We empirically validate the advantages of our proposed methods in *Section 4*.\\n\\nIf you can kindly provide more detailed information about which aspect we do not sufficiently explain, we are also willing to make further clarification.\\n\\n\\n---\\n\\n**Q1-3**: Contribution 3: The advantages of the proposed black-box fingerprinting method (compared to existing methods) are also insufficiently explained. Without providing more explanation to highlight the uniqueness or strengths of this method, it cannot be considered a significant contribution. \\n\\n**R1-3**: Thank you for the comment. **Our proposed black-box fingerprinting methods, FIT-ModelDiff and FIT-LIME, outperform existing methods primarily due to the proposed FIT-Print paradigm**. They are two concrete implementations under FIT-Print. FIT-ModelDiff and FIT-LIME are representatives of the bit-wise method (extract the fingerprint bit by bit) and list-wise method (extract the fingerprint as a whole). We empirically evaluate the effectiveness of FIT-ModelDiff and FIT-LIME in *Section 4*.\\n\\nIf you can kindly provide more detailed information about the potential weaknesses of our proposed methods, we are also willing to make further clarification.\"}", "{\"title\": \"Author Response (Part II)\", \"comment\": \"**Q2**: The problem setting of false claim attack is unclear. The current manuscript lacks a general workflow of the fingerprint registration and verification process. It also lacks a threat model for the false claim attack. Consequently, it is unclear how a fingerprint is registered and verified, or how the adversary could launch a false claim attack. While some of this information could be found in one of the cited works [1], these missing parts would still cause confusion.\\n[1] Liu et al. \\\"False Claims against Model Ownership Resolution\\\". USENIX Security 2024.\\n\\n\\n\\n**R2**: Thank you for the insightful comment! We agree that a detailed introduction to the process of model fingerprinting and the threat model of false claim attacks can improve our paper. Accordingly, we add a detailed introduction and a figure in *Appendix A* to clarify those details, as follows.\\n\\n**A Detailed Threat Models**\\n\\nIn this section, we provide a detailed introduction to the threat models of model fingerprinting and false claim attacks. Three parties involved in the threat models are depicted in Figure 6.\\n\\n**A.1 Detailed Threat Model of Model Fingerprinting**\\n\\nThere are three parties involved in the threat model of model fingerprinting, including the model developer, the model reuser, and the verifier. The model developer trains a model and the model reuser attempts to steal and reuse this model. The verifier is responsible for fingerprint registration and ownership verification. The assumptions of these three parties can be found in Section 2.1.\\n\\n**Process of Model Fingerprinting**. Model fingerprinting can be divided into three steps, including fingerprint generation, fingerprint registration, and ownership verification.\\n\\n1. **Fingerprint Generation**: In this step, the model developer trains its source model $M_o$ and generates the fingerprint of $M_o$.\\n2. **Fingerprint Registration**: After generating the fingerprint, the model developer registers the fingerprint and the model with a timestamp to a trustworthy third-party verifier.\\n3. **Ownership Verification**: For a suspicious model $M_s$ that could be a reused version of $M_o$, the verifier will first check the timestamps of these two models. If the registration timestamp of $M_s$ is later than $M_o$, the verifier will further check whether the fingerprint of $M_o$ is similar to the fingerprint $M_s$. If so, the suspicious model can be regarded as a reused version of $M_o$.\\n\\n**A.2 Detailed Threat Model of False Claim Attacks**\\n\\nThere are three parties involved in the threat model of false claim attacks, including the malicious developer, the verifier, and an independent developer. The formal definition of false claim attacks can be found in Section 2.3.\\n\\n**Assumption of the Malicious Model Developer**. In false claim attacks, the malicious developer is the adversary who aims to craft and register a *transferable* fingerprint to falsely claim the ownership of the independent developer's model $M_I$. The malicious developer is assumed to have adequate computational resources and datasets to train a high-performance model and carefully craft transferable model fingerprints. The primary goal of the malicious developer is that the registered model fingerprints can be verified in as many other models as possible. By generating the transferable fingerprint, the malicious developer can (falsely) claim the ownership of any third-party models (that are registered later than that of the malicious developer).\\n\\n**Process of False Claim Attacks**. The process of false claim attacks can also be divided into three steps, including fingerprint generation, fingerprint registration, and false ownership verification. \\n\\n1. **Fingerprint Generation**: In this step, the model developer trains its source model $M_o$ and attempts to generate a *transferable* fingerprint of $M_o$.\\n2. **Fingerprint Registration**: After generating the fingerprint, the model developer registers the *transferable* fingerprint and the model with a timestamp to a trustworthy third-party verifier.\\n3. **Ownership Verification**: The adversary tries to use the transferable fingerprint to falsely claim the ownership of another independently trained model $M_I$. Since the fingerprint is registered beforehand, the ownership verification won't be rejected due to the timestamp. Subsequently, the benign developer may be accused of infringement.\"}", "{\"title\": \"A Gentle Reminder of the Post-rebuttal Feedback\", \"comment\": \"Thank you very much again for your initial comments. They are very valuable for improving our work. We would be grateful if you could have a look at our response and modifications and please let us know if there is anything else that can be added to our next version. We are also willing to have further discussions with you.\"}", "{\"title\": \"Author Response (Part I)\", \"comment\": \"Dear Reviewer toL1, thank you very much for your careful review of our paper and thoughtful comments. We are encouraged by your positive comments on **the timely topic, novel and reasonable method, and promising experimental results** of our paper. We hope the following responses can help clarify potential misunderstandings and alleviate your concerns.\\n\\n---\\n**Q1**: The focus on defending against false claim attacks is both timely and relevant, given the increasing reliance on open-source pre-trained models in diverse fields. By addressing the vulnerability in existing model fingerprinting methods, this paper provides an advancement in securing model ownership verification.\\n\\n\\n**R1**: Thank you for the positive comment! Your comment encourages us a lot.\\n\\n\\n---\\n**Q2-1**: While I can understand the general idea of test sample extraction is to find a perturbation that pushes the model's prediction towards a target fingerprint (Eq.6), some technical details are not very clear to me. First, how to choose the target fingerprint F (Eq. 5 and 6)? I believe the choice of F significantly affects the model performance (e.g, F is different from the true label).\\n\\n**R2-1**: Thank you for the insightful comment! We are deeply sorry for the missing details in our submission that may lead to potential misunderstanding.\\n\\n- **In our method, the targeted fingerprint is a bit string representing the identity of the model developer and needs to be registered to the trustworthy verifier**. For instance, the logo of the company or personal identity number can be used as the targeted fingerprint.\\n- **The choice of the targeted fingerprint does not affect the model performance**. This is because model fingerprinting does not alter the parameters of the models and has no impact on the model performance. This is a key advantage of fingerprinting.\\n- In our main experiments, we utilize a logo of a file and a pen as the targeted fingerprint. We also conduct an ablation study utilizing three different fingerprints in *Appendix D.1*. The results in *Figure 7* demonstrate that **our FIT-Print is still effective with different targeted fingerprints**.\\n\\nWe also provide more details in *Appendix F.1* of our revision.\\n\\n---\\n\\n**Q2-2**: Secondly, while the paper proposes two mapping functions f, it is not clear what criteria should be considered as good mapping functions. The design of FIT-MODELDIF and FIT-LIME looks arbitrary, and not clear why they are good mapping functions?\\n\\n**R2-2**: Thank you for the insightful comment! We are deeply sorry that our submission may lead you to some misunderstandings that we want to clarify.\\n\\n- **FIT-ModelDiff and FIT-LIME are 'good' and outperform existing methods primarily due to the proposed FIT-Print paradigm.** Our FIT-Print leverages *targeted* fingerprinting to decrease the transferability of fingerprints and resist false claim attacks.\\n- **FIT-ModelDiff and FIT-LIME are two concrete implementations under FIT-Print.** \\n - FIT-ModelDiff and FIT-LIME are **representatives of bit-wise method** (extract the fingerprint bit by bit) **and list-wise method** (extract the fingerprint as a whole), respectively. \\n - FIT-ModelDiff and FIT-LIME have different time and space complexity (details in *Appendix F*). For FIT-ModelDiff, the space complexity is $O(1)$ and the time complexity is $O(k)$ where $k$ is the length of the targeted fingerprint. For FIT-LIME, the space complexity is $O(k)$ and the time complexity is $O(k/\\\\beta)$ where $\\\\beta$ is the batch size.\\n- **The design criteria for a 'good' mapping function are from four main aspects**:\\n - **Distinguishable**: Different models need to exhibit different outputs in the output space of the mapping function. This can guarantee that applying the mapping function can distinguish different independent models.\\n - **Task-agnostic**: The mapping function needs to be able to process the outputs of models with different tasks ($e.g.$, with different numbers of classes).\\n - **Robust**: The outputs of the mapping function on a model need to be robust against various model reusing techniques, $i.e.$, the outputs do not change significantly after model reusing.\\n - **Efficient**: The calculation of the mapping function needs to be efficient and take a small overhead.\\n\\n**We also provide more details on how to design new mapping functions under FIT-Print in *Appendix K***.\"}", "{\"title\": \"Thanks to Reviewer jEMa\", \"comment\": \"Please allow us to thank you again for reviewing our paper and the valuable feedback, and in particular for recognizing the strengths of our paper in terms of the importance of the problem we attempt to address and our novelty and contributions.\\n\\nPlease let us know if our response and the new experiments have properly addressed your concerns. We are more than happy to answer any additional questions during the discussion period. Your feedback will be greatly appreciated.\"}", "{\"title\": \"Thanks to Reviewer and A Gentle Reminder of Discussion\", \"comment\": \"Dear Reviewers,\\n\\nWe thank Reviewer *jEMa* for the discussion and raising the score. As we have less than two days to further revise our paper, we would like to kindly remind Reviewers *toL1* and *oTwh* to take a look at our responses. We sincerely appreciate your time and any feedback you could give us.\"}", "{\"title\": \"Author Response (Part II)\", \"comment\": \"**Q2-3**: Thirdly, I am confused how Eq.6 guarantees that an independent model won't respond similarly to the test samples?\\n\\n**R2-3**: Thank you for the insightful comment! We will clarify the potential misunderstanding to alleviate your concern.\\n\\n- The core of making that an independent model won't respond similarly to the test samples is to **decrease the transferability of the fingerprint**.\\n- **The key in our FIT-Print is the utilization of the targeted fingerprint and Eq. 6 aims to achieve such an objective.** Our main insight is that untargeted fingerprints enlarge the space of viable fingerprints while targeted fingerprints can decrease the transferability and reduce the probability of successful false claim attacks.\\n- We admit that we do not have a strict theoretical guarantee that an independent model won't respond similarly. However, arguably, **such a phenomenon is also widely validated in the area of adversarial attacks ($i.e.$, targeted adversarial examples have lower transferability than untargeted ones).** The papers from the top peer-reviewed conferences [1,2] also present the proposition. It is also empirically validated as depicted in Table 1 of [3].\\n\\n**Reference**\\n1. Towards Transferable Targeted Attack. CVPR, 2020.\\n2. Towards Transferable Targeted Adversarial Examples. CVPR, 2023. \\n3. LFAA: Crafting Transferable Targeted Adversarial Examples with Low-Frequency Perturbations. arXiv, 2023.\\n\\n\\n---\\n**Q3**: It is worth comparing to existing optimization-based model finger-printing methods, e.g, sensitive sample fingerprinting [1].\\n\\n[1] He, Zecheng, Tianwei Zhang, and Ruby Lee. \\\"Sensitive-sample fingerprinting of deep neural networks.\\\" Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2019.\\n\\n**R3**: Thank you for suggesting the outstanding related work! We will further clarify the similarities and differences between [1] and our work as follows.\\n\\n- **Similarities**\\n - [1] and our work both utilize model fingerprinting techniques which **do not need to fine-tune the model and have no impact on the model performance**.\\n - [1] and our work both depend on **optimizing the testing samples** to generate model fingerprints.\\n- **Differences**\\n - **[1] focuses on a different task that utilizes fingerprinting for integrity verification of models.** \\n - [1] attempts to verify whether the model parameters are altered by the adversary while our work focuses on verifying whether a suspicious model is a reused version of another model. \\n - [1] leverages fragile fingerprinting while our method is a robust fingerprinting method.\\n - **The technique proposed in [1] cannot be directly adapted to our scenario.** [1] aims to generate a fragile fingerprint that can be destroyed when the model is modified by others. However, in our scenario, directly applying [1] may lead to that the fine-tuned, pruned, and extracted models will not be regarded as reused, since the (fragile) fingerprints inside these models are destroyed and no longer similar to the fingerprint of the original model. It will inevitably lead to a high false negative rate.\\n\\nDespite the differences, [1] is still an outstanding work exploring another application (integrity verification) of model fingerprinting. We will add a discussion of [1] in *Appendix L.2*. If you have suggestions on other optimization-based fingerprinting methods, please let us know. We are also willing to provide a discussion of them :)\"}", "{\"title\": \"Author Response (Part III)\", \"comment\": \"**Q3**: In Section 4.4, is the adversary using the same fingerprint $F$ when creating verification samples from ImageNet independent models?\\n \\n**R3**: Thank you for the comment! We are deeply sorry that our submission may lead you to some misunderstandings that we want to clarify.\\n\\nWhen creating testing samples from independent models, **the adversary utilizes the same fingerprint used for its own model**. This is because **the objective of false claim attacks is to ensure that the fingerprint of the adversary can be verified in other independent models (via transferability)**.\\n\\n\\nWe add more details in *Section 4.4* of our revision to avoid potential misunderstandings. Thank you again for pointing it out!\\n\\n---\\n\\n**Q4**: The setting of the adaptive unlearning attack in Section 4.5 is confusing. The authors mention that \\\"the model reuser still has no knowledge of the target fingerprint $F$\\\", but the first sentence says \\\"the model reuser knows the target fingerprint\\\" and the target fingerprint $F$ is used in the optimization in Eq. 17. It is unclear whether $F$ is known to the adversary or not. \\n\\n**R4**: Thank you for the insightful comment! We are deeply sorry that there appears to be a typo in Section 4.5.\\n\\n- In the unlearning attack, we assume that **the model reuser knows the target fingerprint of the model developer but has no knowledge of the testing samples**. As such, the model reuser can construct some independent testing samples to unlearn the target fingerprint from the model.\\n- We have proofread our manuscript and corrected all typos. \\n\\nThank you again for the detailed reading!\\n\\n---\\n\\n**Q5**: What is the overhead of the targeted optimization (fingerprint generation)? The overhead analysis in Appendix F seems to only include the overhead for fingerprint verification. Since the targeted optimization process involves additional optimization and augmented models, will the fingerprint generation process be significantly slower?\\n\\n**R5**: Thank you for the insightful comment! We hereby provide more detailed discussions to alleviate your concerns.\\n\\n- In each iteration of optimizing the testing samples, we need to perform one forward propagation and one backward propagation of the fingerprint verification method. Assuming that we utilize $\\\\xi$ augmented models during optimization, **the time complexities of each iteration of testing sample extraction in FIT-ModelDiff and FIT-LIME are $O(\\\\xi\\\\cdot k)$ and $O(\\\\xi\\\\cdot k/\\\\beta)$**. $k$ is the length of the targeted fingerprint and $\\\\beta$ is the batch size.\\n- For instance, in our main experiments, we utilize $10$ augmented models and $256$-bit targeted fingerprint. It takes nearly **3 seconds** for one optimization iteration in FIT-ModelDiff and **1 second** for that in FIT-LIME. As such, **our FIT-ModelDiff and FIT-LIME are efficient in optimization and the overhead is acceptable.**\\n\\nWe also add more details in *Appendix H* of our revision.\\n\\n---\\n\\n**Q6**: What is the attack success rate of false claim attacks on existing AE-based or testing-based methods? It seems Table 2 only includes a false positive rate on independent models without considering false claim attacks.\\n\\n**R6**: Thank you for the insightful comment! We admit that the results of false claim attacks against existing fingerprinting methods are important. We will provide more details to further alleviate your concern.\\n\\n- For testing-based methods, we design a false claim attack in Section 2.3. The experimental results are shown in Table 1 below, demonstrating that **performing false claim attacks can significantly increase the false positive rates**.\\n- For AE-based methods, designing a false claim attack is equivalent to designing a transferable adversarial attack. We utilize the false claim attack proposed in [1] and **the results in Table 1 show that the false claim attack is also effective to AE-based methods**.\\n\\nWe also add more details in *Appendix E* of our revision. \\n\\n\\n**Table 1**: The results of false claim attack against existing model fingerprinting methods.\\n\\n\\n| Method$\\\\rightarrow$ | IPGuard | ModelDiff | Zest | SAC | FIT-ModelDiff (ours) | FIT-LIME (ours) |\\n| -------- | -------- | -------- | ----------|---------| ----------|---------|\\n| False Positive Rate | 30.6% | 4.0% | 7.6% | 39.6% | 0.0% | 0.0% |\\n| False Positive Rate After Attacks| 61.8% | 15.28% | 29.51% | 51.94% | 0.0% | 0.0% |\\n\\n**Reference**\\n1. False Claims against Model Ownership Resolution. USENIX Security, 2024.\"}", "{\"title\": \"Thanks to Reviewer LMaC\", \"comment\": \"Please allow us to thank you again for reviewing our paper and the feedback, and in particular for recognizing the strengths of our paper in terms of our efforts regarding transparency and reproducibility.\\n\\nPlease let us know if our response and the new experiments have properly addressed your concerns. We are more than happy to answer any additional questions during the discussion period. Your feedback will be greatly appreciated.\"}", "{\"title\": \"Author Response (Part I)\", \"comment\": \"Dear Reviewer jEMa, thank you very much for your careful review of our paper and thoughtful comments. We are encouraged by your positive comments on **the importance of the problem we attempt to address and our novelty and contributions**. We hope the following responses can help clarify potential misunderstandings and alleviate your concerns.\\n\\n---\\n\\n**Q1**: The claim that FIT-Print \\\"restricts the (potential) fingerprint space\\\" lacks theoretical support. The idea of restricting fingerprint space with targeted optimization is interesting. However, judging from Theorem 1, the false claim attack has little chance of success mainly because the adversary has no knowledge of $F$ and has to make random guesses. It lacks a deeper analysis of whether this design indeed restricts the fingerprint space and how it impacts the optimization of verification samples. Nonetheless, this is understandable considering the complexity of neural networks.\\n\\n\\n**R1**: Thank you for this insightful comment! We are deeply sorry that our submission may lead you to some misunderstandings that we want to clarify.\\n\\n- **Theorem 1 is simply used to demonstrate that a random trigger sample is unlikely to falsely claim the ownership**.\\n- In particular, **Theorem 1 guides the selection of the verification threshold $\\\\tau$** in our method. Specifically, we can set a proper $\\\\tau$ to make the probability of using a random trigger sample to falsely claim ownership lower than a small security parameter $\\\\kappa$.\\n- We admit that our method does not provide proof that demonstrates our method is provably secure. The transferability of our proposed targeted fingerprint is equal to or lower than the untargeted fingerprint but the theoretical gap is not known to us. Empirically, **such a phenomenon is widely validated in the area of adversarial attacks ($i.e.$, targeted adversarial examples have lower transferability than untargeted ones).** The papers from the top peer-reviewed conferences [1,2] also present the proposition. It is also empirically validated as depicted in Table 1 of [3].\\n\\nWe will investigate how to achieve a provably secure model fingerprinting method in our future work.\\n\\n**Reference**\\n1. Towards Transferable Targeted Attack. CVPR, 2020.\\n2. Towards Transferable Targeted Adversarial Examples. CVPR, 2023. \\n3. LFAA: Crafting Transferable Targeted Adversarial Examples with Low-Frequency Perturbations. arXiv, 2023.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"summary\": \"To protect an AI model's copyright from illegal reuse or plagiarism, a method to verify true ownership is required. A fingerprinting strategy can be used without altering the model, preserving its performance. However, adversarial modifications used for fingerprinting can enable false claim attacks, where an attacker falsely claims ownership by exploiting the transferability of these modifications.\\nTo counter false claim attacks, the authors propose a targeted adversarial fingerprinting method. They argue that current techniques are vulnerable due to their untargeted nature, which creates a large fingerprint space, enabling adversarial transfers between models. By focusing on specific target classes, they suggest that the fingerprint space can be narrowed, reducing transferability and minimizing false claim risks.\", \"soundness\": \"2\", \"presentation\": \"1\", \"contribution\": \"\", \"strengths\": \"Supplementary Materials\\nThe supplementary materials include code and a clear README, which demonstrates a responsible effort towards transparency and reproducibility.\", \"weaknesses\": \"1. Contribution and Innovation Deficiency\", \"there_are_some_issues_with_the_contributions_mentioned_in_the_paper\": \"\", \"contribution_1\": \"False claim attack is not a new type of attack. This type of attack has already been introduced in previous work, and thus, it should not be considered a contribution of this paper.\", \"contribution_2\": \"The advantages of the proposed targeted verification method (compared to other verification methods) have not been sufficiently explained. Without clearly demonstrating the superiority of the targeted approach over other methods, it is difficult to view this as a valid contribution.\", \"contribution_3\": \"The advantages of the proposed black-box fingerprinting method (compared to existing methods) are also insufficiently explained. Without providing more explanation to highlight the uniqueness or strengths of this method, it cannot be considered a significant contribution.\", \"novelty\": \"The main innovation of the paper seems to be the shift from non-targeted to targeted fingerprinting methods. However, this innovation appears somewhat limited and lacks sufficient persuasive power. Beyond this point, the paper's contributions in terms of innovation could be further enhanced. It is recommended to explore additional novel ideas to strengthen the overall impact of the paper.\\n\\n2. Terminology and Definitions are Unclear\", \"improper_naming\": \"The choice of abbreviations in the paper is not standard, especially in \\\"False-claIm-resistant Targeted model fingerPrinting (FIT-Print).\\\" Using the letter \\\"I\\\" from the middle of the word \\\"claim\\\" is unconventional and does not follow common abbreviation rules. Naming should avoid such arbitrary choices to ensure clarity and consistency, and it is recommended to use abbreviations that are more in line with standard conventions.\", \"unclear_explanation_of_the_verification_process\": \"In section 2.3, the explanation of ownership verification through \\\"p1\\\" and \\\"p2\\\" is overly brief. Using such abbreviations makes it difficult for readers to grasp their exact meaning. It is recommended to clearly explain each step of the verification process and avoid using vague abbreviations that could lead to confusion.\", \"unclear_definition_of_independent_model\": \"The term \\\"independent model\\\" is used in the paper but not clearly defined. Does \\\"independent\\\" refer to models with completely different structures, or models performing different tasks? This distinction is critical for understanding the relationships between models in the verification scenario. It is recommended to clearly define whether independence is based on structure, task, or dataset.\", \"unclear_definition_of_the_scenario_and_protocol\": \"The paper does not clearly define each party involved in the fingerprint verification scenario, nor the specific steps they execute in the protocol. For instance, the roles of each party, the information they have access to, and the actions they perform should be clearly explained, so that readers can better understand how the entire verification protocol works. Similarly, providing more specific steps and details for both the attack and defense mechanisms\\u2014such as the inputs, outputs, who executes each step\\u2014would greatly improve the understanding of these processes.\", \"unclear_model_relationships\": \"The symbols for models (e.g., Mo, Ms, Mi) are not clearly explained, particularly in the context of the \\\"false claim of ownership of M1.\\\" Providing clearer descriptions and diagrams would improve understanding of how these models interact in the scenario.\", \"lack_of_illustrations\": \"The paper lacks sufficient diagrams to explain the overall fingerprint verification framework. The relationships between models like Mo, Ms, and Mi are unclear, making it harder for readers to follow. Diagrams would help visualize these interactions and clarify the verification process.\", \"be_careful_when_use_of_terminology\": \"Terms like \\\"Proposition,\\\" \\\"Definition,\\\" and \\\"Theorem\\\" are used in the paper but need to be applied with more caution. It is recommended to briefly introduce these terms in the introduction and clarify their specific contexts throughout the paper to help readers better understand their relevance and contribution to the argument.\\n\\n3. Concerns about the Mechanism or Inadequate Consideration/Explanation of Vulnerabilities.\", \"fingerprinting_scenario_considerations\": \"The paper does not fully explore whether adversarial samples generated by the model owner could also affect models not owned by them, given the known transferability of such samples. This scenario needs further investigation to assess its potential impact on the verification process.\", \"fairness_in_adversarial_sample_generation\": \"The process for generating adversarial samples x\\u0304 and x is unclear. If different algorithms are used, it raises concerns about fairness in the verification process. Ideally, fingerprinting rules should be enforced by a trusted third party to ensure consistency. Clarifying the generation algorithms would enhance transparency and fairness.\", \"questions\": \"The weaknesses of the paper have been pointed out in the Weaknesses section, and I hope that targeted revisions can be made for each issue. Overall, I believe the following improvements would enhance the quality of the paper:\\n\\n1. Clear explanation and presentation\\nPlease provide detailed explanations of the roles, tasks, and relationships of each party in the scenario, along with a step-by-step description of the process. I also expect more elaboration on potential attacks and defenses, specifying the processes involved.\\nI encourage the authors to invest more effort in improving the readability of the paper, such as by incorporating diagrams to help readers understand the proposed protocol and background knowledge. The terms and concepts used should be clearly defined, as omitting details here could compromise the clarity of the paper. Please also avoid using non-standard abbreviations or expressions that may confuse readers.\\n\\n2. Accurate and expanded contributions\\nAccurately and truthfully present the contributions of your work. If the current contributions are insufficient, consider adding new ones. I believe there is room for further enhancement in the paper's contributions to innovation.\\n\\n3. Explanation and discussion\\nThis point corresponds to my concerns about the mechanism or inadequate consideration/explanation of vulnerabilities. I hope the authors can alleviate these concerns by providing additional explanations, analyses, or clear descriptions of the experimental setup. This would enhance the reliability of the work presented in the paper.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"1\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper addresses the challenge of protecting intellectual property rights (IPR) in open-source pre-trained models, which are widely reused but often without authorization. The authors pointed out a vulnerability in existing fingerprinting techniques, i.e., they are prone to false claim attacks, where adversaries could falsely claim ownership of third-party models. To address this, the authors propose a targeted fingerprinting paradigm called FIT-Print, designed specifically to counteract false claim attacks. FIT-Print transforms model fingerprints into targeted signatures through optimization. Based on this paradigm, the authors develop two black-box fingerprinting methods\\u2014FIT-ModelDiff and FIT-LIME, which rely on bit-wise and list-wise approaches, respectively, by leveraging the output distances and feature attributions of specific samples as fingerprints. Experiments on benchmark models and datasets demonstrate the effectiveness of the FIT-Print approach, showing that it is robust to false claim attacks.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"3\", \"strengths\": \"+) Defending against false claim attacks is a timely topic.\\n\\n+) The proposed method is new and technically reasonale in general\\n\\n+) Experimental results looks promising and justified the claims\", \"weaknesses\": \"-) Some technical details are not very clear\\n\\n-) It is not clear if the conclusion obtained on classification task can be extended to a more general scope of models\", \"questions\": \"a) The focus on defending against false claim attacks is both timely and relevant, given the increasing reliance on open-source pre-trained models in diverse fields. By addressing the vulnerability in existing model fingerprinting methods, this paper provides an advancement in securing model ownership verification.\\n\\nb) While I can understand the general idea of test sample extraction is to find a perturbation that pushes the model's prediction towards a target fingerprint (Eq.6), some technical details are not very clear to me. First, how to choose the target fingerprint F (Eq. 5 and 6)? I believe the choice of F significantly affects the model performance (e.g, F is different from the true label). Secondly, while the paper proposes two mapping functions f, it is not clear what criteria should be considered as good mapping functions. The design of FIT-MODELDIF and FIT-LIME looks arbitrary, and not clear why they are good mapping functions? Thirdly, I am confused how Eq.6 grantees that an independent model won't respond similarly to the test samples? \\n\\nc) It is worth comparing to existing optimization-based model finger-printing methods, e.g, sensitive sample fingerprinting [1].\\n\\n[1] He, Zecheng, Tianwei Zhang, and Ruby Lee. \\\"Sensitive-sample fingerprinting of deep neural networks.\\\" Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2019.\\n\\nd) It would be better to discuss if the conclusion obtained on classification tasks can be extended to a more general scope of models, e.g., LLMs and other computer vision models (e.g., diffusion models which involves randomness)?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Dear Reviewers,\\nThe authors have responded to your valuable comments.\\nPlease take a look at them!\\n\\nBest,\\nAC\"}", "{\"summary\": \"This paper proposes FIT-Print, a defense against false claim attacks in model fingerprinting. The paper first extends false claim attacks to various existing fingerprint schemes and reveals that existing methods are typically vulnerable. The paper attributes this vulnerability to the untargeted nature of existing methods, which allows an adversary to exploit transferrable verification samples to forge a false ownership. The paper then proposes FIT-Print, which features a targeted fingerprint optimization process that aligns the fingerprint to a pre-defined target vector. Based on FIT-Print, the paper provides two concrete implementations, FIT-ModelDiff and FIT-LIME, and experimental results show that the proposed method is effective in verification and robust against false claim attacks.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"3\", \"strengths\": \"1. **This paper addresses a new and important problem in model fingerprinting**. False claim attack is a relatively new attack in model fingerprinting. This paper provides a defense technique against this new attack.\\n2. **This paper extends existing false claim attacks to testing-based fingerprint methods**. Previous false claim attacks have mainly focused on adversarial-example-based (AE-based) fingerprints. This paper provides an extension to testing-based methods.\\n3. **This paper proposes a defense against false claim attacks**. FIT-Print proposes a novel targeted optimization process to craft verification samples w.r.t. a pre-defined target, which limits the transferability of verification samples and is empirically effective against false claims. The method could also be generalized to various testing-based model fingerprints.\", \"weaknesses\": \"1. **The claim that FIT-Print \\\"restricts the (potential) fingerprint space\\\" lacks theoretical support**. The idea of restricting fingerprint space with targeted optimization is interesting. However, judging from Theorem 1, the false claim attack has little chance of success mainly because the adversary has no knowledge of $F$ and has to make random guesses. It lacks a deeper analysis of whether this design indeed restricts the fingerprint space and how it impacts the optimization of verification samples. Nonetheless, this is understandable considering the complexity of neural networks.\\n2. **The problem setting of false claim attack is unclear**. The current manuscript lacks a general workflow of the fingerprint registration and verification process. It also lacks a threat model for the false claim attack. Consequently, it is unclear how a fingerprint is registered and verified, or how the adversary could launch a false claim attack. While some of this information could be found in one of the cited works [1], these missing parts would still cause confusion.\\n\\n[1] Liu et al. \\\"False Claims against Model Ownership Resolution\\\". USENIX Security 2024.\", \"questions\": \"1. In Section 4.4, is the adversary using the same fingerprint $F$ when creating verification samples from ImageNet independent models?\\n2. The setting of the adaptive unlearning attack in Section 4.5 is confusing. The authors mention that \\\"the model reuser still has *no* knowledge of the target fingerprint $F$\\\", but the first sentence says \\\"the model reuser *knows* the target fingerprint\\\" and the target fingerprint $F$ is used in the optimization in Eq. 17. It is unclear whether $F$ is known to the adversary or not.\\n3. What is the overhead of the targeted optimization (fingerprint generation)? The overhead analysis in Appendix F seems to only include the overhead for fingerprint verification. Since the targeted optimization process involves additional optimization and augmented models, will the fingerprint generation process be significantly slower?\\n4. What is the attack success rate of false claim attacks on existing AE-based or testing-based methods? It seems Table 2 only includes a false positive rate on independent models without considering false claim attacks.\\n5. From this reviewer's opinion, some discussions in the appendices are quite important, such as the reason for choosing testing-based fingerprints over AE-based fingerprints, the ablation study on the augmented models, and the discussion on the generalization of FIT-Print. Including these in the main part of the paper would help clarify some confusion. This reviewer understands that the authors are restricted by the page limit, but it would be better if some references to the appendices could be put where necessary.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Author Response (Part IV)\", \"comment\": \"**Q7**: From this reviewer's opinion, some discussions in the appendices are quite important, such as the reason for choosing testing-based fingerprints over AE-based fingerprints, the ablation study on the augmented models, and the discussion on the generalization of FIT-Print. Including these in the main part of the paper would help clarify some confusion. This reviewer understands that the authors are restricted by the page limit, but it would be better if some references to the appendices could be put where necessary.\\n\\n**R7**: Thank you for the insightful comment! We agree that some experiments and discussions are important in our work but we have to put it into the appendix due to the page limit. We also agree that adding references to the appendices in the main part of the manuscript is necessary. As such, we make the following changes in our revision to highlight them:\\n\\n- We have summarized all the experiments and discussions, including ablation studies and discussion on the generalization, at the beginning of the experiments section.\\n- We add references to experimental details in Section 4.1.\\n- We add references to discussions on more related works in Section 2.2.\\n\\nIf we miss something important and do not include the right reference to them in our main part, please let us know. We are very willing to add such references :)\"}", "{\"title\": \"Author Response (Part VI)\", \"comment\": \"**Q10**: Fairness in adversarial sample generation: The process for generating adversarial samples x\\u0304 and x is unclear. If different algorithms are used, it raises concerns about fairness in the verification process. Ideally, fingerprinting rules should be enforced by a trusted third party to ensure consistency. Clarifying the generation algorithms would enhance transparency and fairness.\\n\\n**R10**: Thank you for the comment. We will make further clarification to alleviate your concern.\\n\\nWe argue that **the comparison between different fingerprinting methods in our paper is fair.** \\n- We **use the open-source code and the default settings provided in the papers for the experiments**. We also utilize a unified benchmark of model reuse.\\n- **The primary advantage of our method is due to the paradigm of using targeted fingerprints instead of simply a new generation algorithm**. Targeted fingerprints can decrease the transferability and reduce the probability of successful false claim attacks. The advantage is regardless of the settings ($e.g.$, optimization methods).\"}", "{\"title\": \"Author Response (Part III)\", \"comment\": \"**Q4**: Unclear definition of independent model: The term \\\"independent model\\\" is used in the paper but not clearly defined. Does \\\"independent\\\" refer to models with completely different structures, or models performing different tasks? This distinction is critical for understanding the relationships between models in the verification scenario. It is recommended to clearly define whether independence is based on structure, task, or dataset.\\n\\n**R4**: Thank you for the insightful comment! We are deeply sorry that our submission may lead you to some misunderstandings that we want to clarify.\\n\\n- **The independent models are defined as the models that are independently trained by other parties.** The independent models may have different model architectures or different datasets from the source model.\\n- We note that following prior works [1,2,3], **fine-tuned models, pruned models, transfer learning models, and extracted models are regarded as reused models**.\\n\\n**Reference**\\n1. Ipguard: Protecting Intellectual Property of Deep Neural Networks via Fingerprinting the Classification Boundary. AsiaCCS, 2021.\\n2. ModelDiff: Testing-based DNN Similarity Comparison for Model Reuse Detection. SIGSOFT, 2021.\\n3. Are You Stealing My Model? Sample Correlation for Fingerprinting Deep Neural Networks. NeurIPS, 2022.\"}", "{\"summary\": \"This paper studied the vulnerability of existing model fingerprinting approaches against false claim attacks. To fix the problem, they propose to incorporate a mapping function, which can be confidential, and input perturbation, which caters to the mapping, to defend against such attacks.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": [\"The paper is well-written and has plenty of experiments to support the contributions.\", \"The research question is interesting.\"], \"weaknesses\": [\"The attack scenario is limited. Currently, to the best of my knowledge, there is no third-party management on the fingerprints to authorize the ownership of open-sourced models. Therefore, it is impractical to consider and even defend against false-claim attacks. The authors should elaborate on the practical side of the setting.\", \"Another major concern is the missing literature. There is a work on KDD23 which also leverages a mapping function (they call it a meta-verifier) and adaptive input samples to build the model fingerprint. What is the different in methodology? The authors did not even mention the work. They should at least compare with it and show the substantial contribution over the work.\", \"[1] https://dl.acm.org/doi/10.1145/3534678.3539257\"], \"questions\": \"See the weakness part above.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"details_of_ethics_concerns\": \"N/A\", \"rating\": \"3\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Author Response (Part I)\", \"comment\": \"Dear Reviewer oTwh, thank you very much for your careful review of our paper and thoughtful comments. We are encouraged by your positive comments on **our clear writing**, **the interesting research question**, and **the thorough experiments**. We hope the following responses can help clarify potential misunderstandings and alleviate your concerns.\\n\\n---\\n\\n**Q1**: The attack scenario is limited. Currently, to the best of my knowledge, there is no third-party management on the fingerprints to authorize the ownership of open-sourced models. Therefore, it is impractical to consider and even defend against false-claim attacks. The authors should elaborate on the practical side of the setting.\\n\\n**R1**: Thank you for the insightful comment! We will clarify the potential misunderstanding to alleviate your concern.\\n\\n- **The false-claim attack exists even without third-party fingerprint management**. \\n- Arguably, **introducing a third party to manage the model copyrights and the fingerprints is rational and necessary in practice.** \\n - **Establishing the artificial intelligence regulator (AIR) is a political trend.** Many countries and regions are in the process of establishing or have established AIR to regulate AI models in various aspects including **copyright**, safety, and transparency ($e.g.$, as exemplified in the EU Artificial Intelligence Act). We present a detailed discussion in *Appendix M*.\\n - **Not establishing a third-party regulator can make false claim attacks much easier.** This proposition is also presented in [1, 2]. Registering the model and its fingerprint to the third-party regulator with a timestamp can prevent any false claim after registration. During ownership verification, the forged fingerprint with a later timestamp will be regarded as invalid. As such, false claim attacks hinge on generating and registering a transferable fingerprint beforehand. This is actually what we attempt to address in this paper.\\n- Besides improving the resistance of model fingerprinting against false claim attacks, **our FIT-Print also achieves the SOTA effectiveness regarding detecting reused models**. This is empirically validated in our experiments as shown in *Table 2* in *Section 4.2*.\\n- Although currently there is no third-party management on the fingerprints, we believe that our work is still meaningful due to plausible scenarios and foreseeable future needs. Arguably, research is supposed to be ahead of applications.\\n\\n**Reference**\\n1. False Claims against Model Ownership Resolution. USENIX Security, 2024.\\n2. GrOVe: Ownership Verification of Graph Neural Networks using Embeddings. S&P, 2024.\"}", "{\"title\": \"Author Response (Part III)\", \"comment\": [\"**Q4**: It would be better to discuss if the conclusion obtained on classification tasks can be extended to a more general scope of models, e.g., LLMs and other computer vision models (e.g., diffusion models which involves randomness)?\", \"**R4**: Thank you for the insightful comment! We admit that the generalization of our FIT-Print to a more general scope of models is important. We will further clarify it as follows.\", \"Arguably, for deterministic models which do not involve randomness ($e.g.$, CNN and LLM), **our FIT-Print can generalize to different types of models and datasets** (details in *Appendix H*).\", \"**Our FIT-Print can generalize to models with different architectures and tasks**.\", \"Since we do not need to alter or fine-tune the model, **our method can fundamentally generalize to models** with other architectures ($e.g.$, advanced CV models like ViT) as well.\", \"For models with different tasks, **the major difference between models with different tasks is the output format**. For instance, the image generation model outputs a tensor consisting of a sequence of logits. FIT-ModelDiff calculates the cosine similarity between the outputs and FIT-LIME calculates the average entropy of the output. The two calculation methods can be applied to any output format ($e.g.$, 1-D vectors, 2-D matrices, or tensors). As such, **our methods are naturally feasible for models with different tasks**.\", \"**Our FIT-Print can generalize to different types of datasets**.\", \"For different datasets (particularly discrete data like text data), the main challenge of optimizing the testing samples lies in how to design an effective optimization method. There are already some existing works[1,2,3] to fulfill this task. Accordingly, **our FIT-Print can be adapted to other data formats ($e.g.$, text or tabular)**.\", \"**We also conduct a case study on text generation models in *Appendix H.3***. The experimental results demonstrate that our FIT-Print still works for text generation models.\", \"For non-deterministic models that involve randomness ($e.g.$, diffusion models), we have to admit that we do not know whether our methods and existing fingerprinting methods can be adapted to them since they have a completely different inference paradigm. We will conduct a comprehensive study in our future work.\", \"**Reference**\", \"1. Gradient-based adversarial attacks against text transformers. EMNLP, 2021.\", \"2. Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery. NeurIPS, 2023.\", \"3. Promptcare: Prompt copyright protection by watermark injection and verification. S&P, 2024.\"]}", "{\"title\": \"Thanks to Reviewer toL1\", \"comment\": \"Please allow us to thank you again for reviewing our paper and the valuable feedback, and in particular for recognizing the strengths of our paper in terms of the timely topic, novel and reasonable method, and promising experimental results.\\n\\nPlease let us know if our response has properly addressed your concerns. We are more than happy to answer any additional questions during the discussion period. Your feedback will be greatly appreciated.\"}", "{\"title\": \"Thanks to Reviewer oTwh\", \"comment\": \"Please allow us to thank you again for reviewing our paper and the valuable feedback, and in particular for recognizing the strengths of our paper in terms of our clear writing, the interesting research question, and the thorough experiments.\\n\\nPlease let us know if our response and the new experiments have properly addressed your concerns. We are more than happy to answer any additional questions during the discussion period. Your feedback will be greatly appreciated.\"}", "{\"title\": \"Author Response (Part II)\", \"comment\": \"**Q1-4**: Novelty: The main innovation of the paper seems to be the shift from non-targeted to targeted fingerprinting methods. However, this innovation appears somewhat limited and lacks sufficient persuasive power. Beyond this point, the paper's contributions in terms of innovation could be further enhanced. It is recommended to explore additional novel ideas to strengthen the overall impact of the paper.\\n\\n**R1-4**: Thank you for your comment. We respectfully disagree with this argument.\\n\\n- Our contributions are **not limited to the shift from non-targeted to targeted** fingerprinting methods.\\n - We **revisit existing model fingerprinting methods and design an implementation of false claim attacks**. We reveal that existing methods are vulnerable to false claim attacks.\\n - We **introduce a new fingerprinting paradigm**, where we conduct fingerprint verification in a targeted manner. Our main insight is that untargeted fingerprints enlarge the space of viable fingerprints while targeted fingerprints can decrease the transferability and reduce the probability of successful false claim attacks.\\n - Based on our proposed paradigm, we **design two advanced black-box targeted model fingerprinting methods**, FIT-ModelDiff and FIT-LIME. FIT-ModelDiff and FIT-LIME are representatives of the bit-wise method (extract the fingerprint bit by bit) and list-wise method (extract the fingerprint as a whole), respectively.\\n - We conduct experiments on 2 benchmark datasets, 2 models, and 6 baseline fingerprinting methods to verify the effectiveness and conferability of our FIT-Print and its resistance to false claims and adaptive attacks.\\n- We admit that our methods appear to be simple, but **it does not mean that our methods do not have sufficient novelties and contributions**.\\n - **Our method is simple yet effective since our insight is deep and thorough**.\\n - **We believe that simplicity is not necessarily a bad thing** because the follow-up researchers and developers can easily implement our methods and paradigm to further promote model copyright protection.\\n\\n\\nIf you could provide more detailed information about your concerns, we are also willing to make further clarification.\\n\\n---\\n\\n**Q2**: Terminology and Definitions are Unclear Improper naming: The choice of abbreviations in the paper is not standard, especially in \\\"False-claIm-resistant Targeted model fingerPrinting (FIT-Print).\\\" Using the letter \\\"I\\\" from the middle of the word \\\"claim\\\" is unconventional and does not follow common abbreviation rules. Naming should avoid such arbitrary choices to ensure clarity and consistency, and it is recommended to use abbreviations that are more in line with standard conventions.\\n\\n**R2**: Thank you for the comment. Our proposed paradigm is named FIT-Print. 'FIT' may correspond to the core of our paradigm, $i.e.,$ fitting to a targeted fingerprint. Besides, we do not agree that naming can only utilize the first letter. \\n\\nIf you can kindly provide authoritative documents that the name of a method must use the first letter, we are willing to make a change.\\n\\n\\n---\\n\\n**Q3**: Unclear explanation of the verification process: In section 2.3, the explanation of ownership verification through \\\"p1\\\" and \\\"p2\\\" is overly brief. Using such abbreviations makes it difficult for readers to grasp their exact meaning. It is recommended to clearly explain each step of the verification process and avoid using vague abbreviations that could lead to confusion.\\n\\n**R3**: Thank you for the comment. However, we do not use 'p1' or 'p2' in Section 2.3. Instead, we use the full names 'Proposition 1' and 'Proposition 2'. We believe they are referenced explicitly.\\n\\nIf you can kindly provide more information regarding 'p1' and 'p2', we are willing to make further clarification.\"}", "{\"comment\": \"This reviewer would like to thank the author(s) for their detailed and timely response. The response has addressed most of the concerns. The reviewer would like to change the score to 6 (marginally above acceptance threshold), based on the following considerations:\\n\\n1. The paper addresses a timely and important topic of defending against false claim attacks in model fingerprinting.\\n2. The problem setting of false claim attack (along with a few other minor issues) have been clarified in the revised manuscript. A detailed workflow of model fingerprinting and false claim attack has been added to Appendix A, which resolves one of this reviewer's major concerns.\\n3. Still, the current manuscript lacks a deeper analysis or evaluation on the transferability of the targeted fingerprints. This reviewer agrees that empirically, targeted adversarial examples tend to have lower transferability. However, it is unclear whether a more sophisticated adversary could forge a false claim on FIT-Print (e.g., by exploiting techniques from transferrable targeted adversarial attacks). Apart from a lack of formal proof (which is a noteworthy but understandable limitation, given the complexity of neural networks), the current adaptive evaluation is also limited (in Section 4.4, the adaptive attacker only leverages additional independent models). Hence, a more in-depth analysis or evaluation on this aspect would help make the claim more persuasive.\"}" ] }

[ "Thomas Foster", "Eltayeb Ahmed", "Jonathan Cook", "Shalev Lifshitz", "Tim Rocktäschel", "Jakob Nicolaus Foerster" ]

[ "large language models", "self-improvement", "reasoning" ]

{ "note_id": [ "yf6zfp3XT8", "tEpI6i0Z1E", "sCSq4AmKDM", "oDGVYALc59", "jUCz99Ia37", "gWvtHv34kC", "gImPAtNlM1", "fDQc3dfdDC", "dn5SEYUAf1", "bCBrLtGmBw", "VbvzmxsjNY", "Tih8kwy3ZU", "SJHxpgvIPh", "S5G60I7o7m", "QpiQi7b1dY", "Pn26bHMv9p", "Ll3Hn3NuTi", "HGOOAxe88Q", "GoxhGr6DRC", "DObSOO1YhY", "CqjVYachGy", "A8TfjgQMPe", "5Ik0O3hx0s", "4uDn6amAh8", "3qtqdkC931", "1gzcR19Tdd" ], "note_type": [ "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "meta_review", "official_comment", "official_comment", "official_review", "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment" ], "note_created": [ 1732054673873, 1732554043282, 1732056140344, 1730347771324, 1732242103098, 1732054319400, 1729439660507, 1732072061231, 1732055240685, 1732462708094, 1732052551554, 1734833922847, 1732553973814, 1732553908605, 1730715529040, 1737524135670, 1732688674168, 1732554022556, 1732430524155, 1732052836875, 1732055074834, 1730600687551, 1732310010338, 1732421796866, 1732053910788, 1732615285586 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission11621/Authors" ], [ "ICLR.cc/2025/Conference/Submission11621/Authors" ], [ "ICLR.cc/2025/Conference/Submission11621/Authors" ], [ "ICLR.cc/2025/Conference/Submission11621/Reviewer_ZH4y" ], [ "ICLR.cc/2025/Conference/Submission11621/Reviewer_qdDb" ], [ "ICLR.cc/2025/Conference/Submission11621/Authors" ], [ "ICLR.cc/2025/Conference/Submission11621/Reviewer_MbaU" ], [ "ICLR.cc/2025/Conference/Submission11621/Reviewer_ZH4y" ], [ "ICLR.cc/2025/Conference/Submission11621/Authors" ], [ "ICLR.cc/2025/Conference/Submission11621/Reviewer_KyrV" ], [ "ICLR.cc/2025/Conference/Submission11621/Authors" ], [ "ICLR.cc/2025/Conference/Submission11621/Area_Chair_DYu9" ], [ "ICLR.cc/2025/Conference/Submission11621/Authors" ], [ "ICLR.cc/2025/Conference/Submission11621/Authors" ], [ "ICLR.cc/2025/Conference/Submission11621/Reviewer_KyrV" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission11621/Authors" ], [ "ICLR.cc/2025/Conference/Submission11621/Authors" ], [ "ICLR.cc/2025/Conference/Submission11621/Reviewer_qdDb" ], [ "ICLR.cc/2025/Conference/Submission11621/Authors" ], [ "ICLR.cc/2025/Conference/Submission11621/Authors" ], [ "ICLR.cc/2025/Conference/Submission11621/Reviewer_qdDb" ], [ "ICLR.cc/2025/Conference/Submission11621/Authors" ], [ "ICLR.cc/2025/Conference/Submission11621/Reviewer_MbaU" ], [ "ICLR.cc/2025/Conference/Submission11621/Authors" ], [ "ICLR.cc/2025/Conference/Submission11621/Reviewer_ZH4y" ] ], "structured_content_str": [ "{\"title\": \"Response to Reviewer MbaU [2/2]\", \"comment\": \">How effective is RA compared to another baseline model which is directly trained to predict the final answer without training to generate CoT?\\n\\nThank you for pointing out that this important baseline was missing. We have included an updated Figure 7 in the latest version of the manuscript which includes a \\u201cno reasoning\\u201d baseline of training to predict the answer without any CoT. This new baseline performs worse than RA and many of the other baselines (Figure 7, right-side). In particular, the difference in performance is largest for reasoning style questions.\\n\\n>How much additional overhead occurs for applying RA during pre-training (Section 6)?\", \"our_offline_rl_method_has_three_main_steps\": \"(1) generate a large batch of CoTs, (2) self-insert them into the pretraining dataset, and (3) filter the ones with the highest scoring rewards and finetune the model. Thus, the main computational overhead occurs **before** training: generating the CoTs and using RA to score them. This can be efficiently parallelized, and thankfully, getting the score from RA only requires 2 forward passes (one with the CoT, and one for the \\u201cEmpty CoT\\u201d baseline). During supervised finetuning, we just fit the dataset with the self-inserted CoTs as normal, which adds minimal overhead. So, the additional overhead is mostly **before** training, it can be massively parallelized (RA is parallelizable since it is loss-based), and it is relatively insignificant compared to the unavoidable cost of actually generating the CoTs.\"}", "{\"comment\": \"Thank you for increasing your score and helpful initial review. We have since made a large number of improvements to the paper but, given that you are currently recommending that this paper not be accepted, are there any further improvements we could make that would change this?\"}", "{\"title\": \"General Response From Authors to Reviewers\", \"comment\": \"Thanks to all the reviewers for your time and effort during the review process. We appreciate that you found our work insightful, and we\\u2019re glad that there is excitement about our progress towards self-improving CoT reasoning. Your thoughtful reviews have helped us dramatically improve the clarity and rigour of our submission.\\n\\nWe have responded to each reviewer individually, and also include a general response to all reviewers here. If you find our answers responsive to your concerns, we would be grateful if you considered increasing your score, and if you have additional questions, we\\u2019re happy to engage further. \\n\\n\\n>Additional experiments.\\n\\nBased on the reviewers\\u2019 questions and comments, we have added further experiments to strengthen our results. In response to reviewers ZH4y and MbaU, we have added results going beyond a single model to also include Llama 3.1 8B (see Table 5 in the appendix). In response to reviewer QdDb, we have included an additional ablation for different values of the clipping threshold (Figure 5 in the appendix). And in response to reviewer KyrV, we have added a qualitative analysis of the highest- and lowest-scoring CoTs in Section 6.1.\\n\\n>Clarification of contributions.\\n\\nWe have updated the submission pdf to clarify the contributions of our work (main updates in red text). A brief summary is provided below:\\n\\nAs it becomes increasingly challenging and prohibitively expensive to curate large-scale (question, CoT, answer) datasets, the LLM reasoning community has begun focusing on a grand challenge task: self-improving CoT reasoning on unstructured, pretraining-scale text. To be clear, our work does not solve this grand challenge. Our primary contributions are showing that standard reward functions fail *even in* the intermediate MMLU-FREE-FORM setting, introducing a novel reward function to solve this issue, and performing a comprehensive analysis of reward functions and how they affect *what* and *where* reasoning is rewarded. To our knowledge, our work is the first to provide this type of analysis on reward functions for self-improving CoT reasoning on unstructured text. There is still more work to be done in order to solve the full unstructured pretraining setting, and we present an exploratory experiment that provides key insights to help facilitate future research in this direction. We believe these to be major contributions to the literature, and have dramatically updated our manuscript (especially the Introduction, Section 5, Section 6, and Conclusion) with more targeted and clear descriptions of our contributions.\", \"note\": \"The main updates to the manuscript which clarify our contributions are made with *red* text. The blue text refers to changes made in response to specific reviewer questions, and we reference them in the individual responses below.\\n\\n>Updated appendix.\\n\\nWe have also updated the appendix to be more clear, and have reworked the previous Appendix D into the main paper as the newly added Section 6.2 (per the request from reviewer ZH4y). To make space for these changes we have:\\n- Converted the former Figure 3 bar plot into a table (Table 2). Notice that full results with confidence bounds can still be found in Appendix B.1.\\n- Moved the former Figure 1 to the Appendix (new Figure 7).\\n- Moved the former Figure 5 to the Appendix (new Figure 8).\\n\\n>Clarified hyperparameters.\\n\\nWe have made the specific details about model architecture and inference hyperparameters (i.e., temperature) more clear in the manuscript. These changes are made in blue text.\\n\\n\\nWe again thank the reviewers for their engagement and we appreciate all the suggestions that we believe will make the paper significantly stronger!\"}", "{\"summary\": \"This paper explores a method for self-improving CoT reasoning in LLMs without relying on curated datasets. By leveraging reinforcement learning on general pre-training data, the authors aim to enhance models\\u2019 reasoning abilities across diverse tasks. They introduce a new reward function, Reasoning Advantage (RA), which better identifies effective reasoning, and demonstrate its impact on open-ended question-answering tasks. The paper highlights RA\\u2019s potential but also suggests that more advanced optimization methods are needed for scalable CoT improvements in broader, less structured contexts.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. This paper addresses an important issue: achieving self-improvement of CoT reasoning during the pre-training phase. This approach has significant potential to help overcome the data bottleneck in LLMs.\\n\\n2. The paper explores several types of reward functions and establishes criteria for an effective reward function, which is valuable and insightful for future research in this area.\", \"weaknesses\": \"1. The technical contributions of the paper are relatively weak. The proposed MMLU-FREE-FORM is merely a simple adaptation of the original MMLU, and the introduced RA is only a minor modification based on token loss.\\n\\n2. The paper somewhat overstates its contributions. The authors primarily demonstrate the positive impact of RA on MMLU-FREE-FORM, yet MMLU-FREE-FORM is derived from the structured MMLU dataset and cannot be regarded as a typical pre-training dataset. In fact, experiments on OpenWebMath show minimal improvement. Typical pre-training datasets often include substantial noise, such as HTML elements, which is a key challenge in achieving self-improvement CoT during the pre-training phase.\\n\\n3. The paper lacks discussion on relevant work in reasoning enhancement during the pre-training phase, such as https://arxiv.org/pdf/2404.07965.\\n\\n4. The experiments are insufficiently comprehensive, as they are conducted on only one model and one dataset. Testing with models of different parameter sizes within the same series or different architectures could help demonstrate the generalizability of RA.\\n\\n5. The presentation of the paper could be improved. Some key findings should be in the main body rather than the Appendix, such as Appendix D and the definition of RA in Appendix A. Essential parameters, like the type of LLM used and inference hyperparameters, should also be included in the main text.\", \"minor\": [\"Punctuation should be added at the end of each equation.\", \"Some quotation marks are unmatched, such as in line 265 and line 349.\", \"Figure 1 appears somewhat rudimentary.\"], \"questions\": \"See above.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"I really agree with ZH4y that the statement of pretraining stage may be misleading. I think the authors should carefully consider the contribution points of this paper, especially the term \\\"pretraining\\\". Meanwhile, the author should add more discussions on why the log-likelihood, as a reward signal, can promote the enhancement of the model's general reasoning ability.\"}", "{\"title\": \"Response to Reviewer MbaU [1/2]\", \"comment\": \"Thank you for your review and comments. We\\u2019re glad that the reviewer finds our work insightful for designing rewards for language model training. Please see below for responses to your comments and questions.\\n\\n>Unlike the motivation of the paper, the proposed method, RA, is not effective for pre-training scale, questioning the scalability of the proposed method.\\n\\nThanks for raising this point, we agree with the reviewer that our work does not solve the full unstructured, pretraining setting. We think it may be helpful to provide additional background, which we hope clarifies the contributions of our work:\\n\\nAs it becomes increasingly challenging and prohibitively expensive to curate large-scale (question, CoT, answer) datasets, the LLM reasoning community has begun focusing on a grand challenge: self-improving CoT reasoning on unstructured text at pretraining scale. \\n\\nIn this work, we introduce MMLU-FREE-FORM as a small step towards the grand challenge of truly unstructured text. As mentioned by Reviewer KyrV, it acts as \\u201c*an intermediate benchmark between structured QA and general language modeling*\\u201d. However, we show that even taking this small step renders current reward functions unusable. That is, standard reward functions cannot self-improve CoT reasoning *even in* this simplified free-form QA setting. **Our work introduces RA to address this problem**\\u2014it\\u2019s the only reward function which facilitates generalization when self-improving on MMLU-FREE-FORM. We also perform a comprehensive analysis of reward functions and how they affect *what* and *where* reasoning is rewarded. To our knowledge, our work is the first to provide this type of analysis on reward functions for self-improving CoT reasoning on unstructured text.\\n\\nTo be clear, our work does not solve the grand challenge mentioned. Our work demonstrates that current reward functions fail for even the small step of MMLU-FREE-FORM, introduces a novel reward function that mitigates this problem, and performs a comprehensive analysis with insights to help facilitate future research in this direction. We believe these to be major contributions to the literature, and have dramatically updated our Introduction, Section 5, Section 6, and Conclusion with more targeted descriptions of our contributions. The key points are updated with red text.\\n\\n>The paper measures the performance by using 'expected accuracy' metric, which makes comparison with other methods difficult. What is the absolute accuracy performance for Figure 4?\\n\\nThanks for the question. Expected accuracy is commonly used for evaluating language models on QA benchmarks, with many works simply referring to it as accuracy. For example, one of the most important recent works on self-improving LLM reasoning [1] reports accuracy, but it's only clear when going through their code that this is in fact expected accuracy. \\n\\nGiven a question and a reasoning trace, we compute expected accuracy as the probability of the correct answer. Another way to compute accuracy is by greedy decoding a single answer, or by sampling many answers and averaging the performance, which approaches expected accuracy given enough samples. Usually, expected accuracy is preferred because it represents the raw distribution learnt by the model, without depending on a specific decoding procedure. We have added additional clarification regarding this point near the end of Section 5.2 (blue text).\\n\\n[1] Zelikman et al, Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking, 2024\\n\\n>The paper only uses a single backbone model to show the effect of the proposed method.\\n\\nWe thank the reviewer for raising this point, and we agree that running experiments with another model would strengthen the paper. Therefore, we have repeated the *what* and *where* experiments from Section 5.1 with another model, LLama 3.1 8B. We present the results in Appendix B Table 5 and find them to be consistent with our original results using Mistral 7B in Table 2.\"}", "{\"summary\": \"The paper explores effective rewards that could be applied during LLM pretraining. Especially, the paper explores various reward functions based on what reasoning is learnt and where reasoning is rewarded. Based on the findings, the paper suggests, RA (Reasoning advantage) which facilitates self-improving CoT reasoning on free- form question-answering (QA) data.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": [\"The paper provides useful insights for designing rewards for language model training.\", \"The authors explores the effectiveness of RA on multiple experimental settings.\"], \"weaknesses\": [\"Unlike the motivation of the paper, the proposed method, RA, is not effective for pre-training scale, questioning the scalability of the proposed method.\", \"The paper measures the performance by using 'expected accuracy' metric, which makes comparison with other methods difficult. What is the absolute accuracy performance for Figure 4?\", \"The paper only uses a single backbone model to show the effect of the proposed method.\"], \"questions\": [\"How effective is RA compared to another baseline model which is directly trained to predict the final answer without training to generate CoT?\", \"How much additional overhead occurs for applying RA during pre-training (Section 6)?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thanks for your reply, I improved my score accordingly.\"}", "{\"title\": \"Response to Reviewer KyrV [2/2]\", \"comment\": \">Have you explored whether the effectiveness of Reasoning Advantage (RA) varies across different types of reasoning tasks beyond math and standard QA?\\n\\nIn keeping with much of the LLM reasoning literature, we chose to specifically focus our evaluations on a range of QA-style reasoning problems. However, we do show positive results that go beyond just mathematics. As mentioned above, Appendix B.2 Figure 6 shows that RA significantly improves reasoning performance on \\u201creasoning style questions\\u201d in the MMLU test set. These questions span a wide range of subjects beyond math, including physics, biology, accounting, law, computer science, etc.\\n\\n>In Section 5.2, you show that optimizing for RA leads to a 7% improvement on GSM8K. Could you provide more analysis of what specifically improved in the model's reasoning capabilities? Are there particular types of math problems where the improvement was more pronounced?\\n\\nGreat question. We observe that the GSM8K accuracy goes up due to fewer logical/arithmetic errors in the generated CoTs, but we don\\u2019t observe a single predominant qualitative change. However, we do notice something interesting regarding performance on the MMLU test set. As mentioned above, MMLU sees an improvement in questions across a wide range of subjects that require quantitative reasoning (biology, physics, accounting, computer science, etc.). Moreover, the improvement is far larger for these questions than for those that require recall (see Figure 6 in Appendix B.2). Thinking about it, this result makes a lot of sense, since CoT reasoning probably doesn't help as much when trying to recall a fact. However, for quantitative reasoning and problem-solving tasks, additional reasoning can clearly be of benefit. We have added a discussion of this to Appendix B.2 (blue text). Thank you for the suggestion!\\n\\n>The paper mentions that only 0.01% of generated CoTs achieve a reward above 0.2 on OpenWebMath. Have you analyzed these high-scoring CoTs to understand what makes them successful? This analysis could inform better prompting strategies.\\n\\nThanks for asking this great question. Upon manual inspection, many of the CoTs that passed the filtering threshold exhibited the conservative strategy described in the paper: they simply summarize past information from the context. This explains why the model learned to be overly conservative. However, these overly conservative CoTs which made it past the RA threshold were still superior to those that did not pass the threshold (the ones that did not pass the threshold mainly contained **incorrect** reasoning that predicted the subsequent tokens incorrectly). This indicates that RA actually succeeded at its job of identifying the best reasoning from the generated batch of CoTs, and that the main issue indeed lies with the *lack of diversity* in the generated CoTs. We mention a few potential ways to generate more diverse CoTs in the paper, but we agree that it would also be worth exploring different prompting strategies (we used a single system prompt to generate these CoTs, and did not spend much time on prompt engineering).\\n\\nWe have added this detailed discussion to Section 6.1 (the first blue block text, and then the rest discusses ways to increase diversity), and we think it dramatically improves the section. Thanks again for the great suggestion.\"}", "{\"comment\": \"Thanks for the response, I will keep with my original score\"}", "{\"title\": \"Response to Reviewer ZH4y [1/2]\", \"comment\": \"We thank the reviewer for engaging deeply with the work. We\\u2019re glad that the reviewer found our investigations into effective reward functions for CoT reasoning both valuable and insightful.\\n\\nWe have updated the paper to clarify our contributions and address the reviewer\\u2019s concerns, with the main updates in red text. We address specific questions and comments below.\\n\\n>The technical contributions of the paper are relatively weak. The proposed MMLU-FREE-FORM is merely a simple adaptation of the original MMLU, and the introduced RA is only a minor modification based on token loss.\\n\\nWe agree that MMLU-FREE-FORM is not a substantial technical change from MMLU. However, our purpose for creating MMLU-FREE-FORM was not to create a radically new dataset, but to make the smallest possible change to MMLU that reveals the limitations of existing reward functions. It acts as an important middle-ground between improving CoT reasoning using curated (question, CoT, answer) datasets and the challenging, unsolved task of self-improving CoT reasoning on unstructured text. This is because MMLU-FREE-FORM does not allow for using exact-match accuracy as a reward metric (similar to unstructured pretraining text) and yet offers a higher density of clear opportunities for CoT reasoning compared to typical pre-training corpora, making it an ideal stepping-stone towards the ultimate goal of self-improving CoT reasoning on unstructured text. As mentioned by Reviewer KyrV, \\\"*the creation of MMLU-FREE-FORM as an intermediate benchmark between structured QA and general language modeling is clever and useful for the research community. The empirical results showing successful transfer learning to GSM8K math problems provide concrete validation of their approach.*\\\" We have updated the Introduction and Section 5.2 to clarify the contribution of MMLU-FREE-FORM, with the main updates in red text.\", \"we_also_agree_with_the_reviewer_that_ra_is_a_modification_of_token_loss\": \"by using clipping, subtracting the \\\"empty CoT\\\" baseline, and normalizing. As above, our goal was to make the smallest, simplest modification to the existing paradigm (standard loss) that has the potential to work for this setting. We believe that the contribution of RA is quite significant. It performs substantially better than token loss on the *what* to reward experiments (distinguishing effective CoT) and the *where* to reward experiments (picking out useful locations for producing CoT). Moreover, and possibly our most important result, **only RA is able to facilitate generalization to the MMLU test set and zero-shot transfer to GSM8K** when self-improving CoT reasoning on MMLU-FREE-FORM. In addition, we strongly believe that RA being based on token loss is a key advantage of this function. It requires only two forward passes, does not require an external strong model, is not limited to exact-match heuristics like accuracy-based functions (which fail on unstructured text), and allows the model to place weight over a distribution of valid answers. We have updated Section 4 to clarify this point, with the main updates in red text.\\n\\nWe hope that our explanations helped to clarify the contributions of MMLU-FREE-FORM and the RA reward function.\\n\\n>The paper somewhat overstates its contributions. The authors primarily demonstrate the positive impact of RA on MMLU-FREE-FORM, yet MMLU-FREE-FORM is derived from the structured MMLU dataset and cannot be regarded as a typical pre-training dataset. In fact, experiments on OpenWebMath show minimal improvement. \\n\\nWe thank the reviewer for their feedback on this point. Here, we aim to provide additional background, which we hope clarifies the important contributions of our work.\\n\\nCurating large, challenging, and diverse (question, CoT, answer) datasets for improving LLM reasoning has become exceptionally expensive (millions of dollars) and very challenging (requiring thousands of expert hours). With this as motivation, the LLM reasoning community has recently begun focusing on a grand challenge: self-improving CoT reasoning on unstructured text at pretraining scale. \\n\\nWe agree with the reviewer that MMLU-FREE-FORM is not a typical pre-training dataset. It is a small step towards the grand challenge of truly unstructured text. However, we show that even taking this small step renders current reward functions unusable. That is, standard reward functions cannot self-improve CoT reasoning *even in* this simplified free-form QA setting. **Our work introduces RA to address this problem**\\u2014it\\u2019s the only reward function which facilitates generalization when self-improving on MMLU-FREE-FORM. We also perform a comprehensive analysis of reward functions and how they affect what and where reasoning is rewarded. To our knowledge, our work is the first to provide this type of analysis on reward functions for self-improving CoT reasoning on unstructured text.\\n\\n**[Continued in 2nd Response]**\"}", "{\"metareview\": \"The paper introduces the Reasoning Advantage (RA) reward function to improve CoT reasoning in LLMs using unstructured, pretraining-scale text instead of curated datasets. RA demonstrates improved performance on MMLU-FREE-FORM and transferability to GSM8K math tasks, while also providing a detailed analysis of reward functions and their properties. The work addresses a critical challenge in LLM reasoning, introducing MMLU-FREE-FORM as an intermediate benchmark that bridges curated QA datasets and unstructured text. The experiments highlight RA\\u2019s ability to improve reasoning performance and generalization to diverse tasks. However, some reviewers felt the contributions were incremental, with RA being a slight modification of token loss and MMLU-FREE-FORM a minor adaptation of MMLU. They also noted the limited scope of models and datasets tested, along with minimal improvement on OpenWebMath, raising concerns about scalability to pretraining-scale data. While the approach shows promise in intermediate settings, the lack of comprehensive evaluations and broader generalizability testing would limit its impact. Future work might expand experiments to include larger and more diverse datasets (e.g., BBH, CSQA) and evaluate performance across multiple model sizes to validate scalability and robustness.\", \"additional_comments_on_reviewer_discussion\": \"While the authors made significant efforts to address the reviewers' concerns by providing clarifications, adding experiments, and improving the paper\\u2019s presentation, the reviewers maintained their original scores. Despite acknowledging the improved clarity and additional contributions, they felt that the work's technical novelty and scalability remain limited, suggesting the paper could benefit from further refinement and resubmission in the future.\"}", "{\"comment\": \"Thank you again for helping to improve our paper with the suggestions and comments in your initial review. In light of having provided points of clarification, updated the paper accordingly, and received no further questions, we ask if you would consider increasing the confidence in your score. And if you still have any further questions, please share those with us so that we can ensure the clarity of our work.\"}", "{\"comment\": \"Thank you again for providing initial comments that have helped to improve the paper. In light of our clarifications and additional experiments, we would appreciate it if you considered raising the confidence level in your score. And if you have any outstanding sources of uncertainty, please don\\u2019t hesitate to share these with us so that we can further improve the clarity of our work.\"}", "{\"summary\": \"This paper explores how to enable large language models (LLMs) to self-improve their Chain-of-Thought (CoT) reasoning abilities using general pre-training data rather than supervised datasets. The authors investigate what makes a good reward function for learning reasoning during language modeling, examining how different reward functions affect both what reasoning is rewarded and where reasoning is applied. They introduce a novel \\\"Reasoning Advantage (RA)\\\" reward function that combines clipping and normalization techniques, and demonstrate its effectiveness on a new free-form question-answering dataset called MMLU-FREE-FORM, showing improved transfer to math reasoning tasks.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"3\", \"strengths\": \"The systematic analysis of reward functions and their properties is thorough and well-motivated. The introduction of the RA reward function addresses key limitations of existing approaches, particularly in distinguishing good reasoning from random text and identifying appropriate contexts for reasoning. The creation of MMLU-FREE-FORM as an intermediate benchmark between structured QA and general language modeling is clever and useful for the research community. The empirical results showing successful transfer learning to GSM8K math problems provide concrete validation of their approach.\", \"weaknesses\": \"The paper's primary limitation appears in the scaling to general pre-training data, where the offline reinforcement learning approach that worked well on MMLU-FREE-FORM struggles to escape local optima of conservative reasoning. While the authors acknowledge this limitation and suggest future research directions, the paper doesn't fully solve the challenge of self-improving reasoning at pre-training scale. Additionally, while the authors demonstrate improved performance on mathematical reasoning tasks, there could be more exploration of how well their approach generalizes to other types of reasoning beyond mathematics.\", \"questions\": \"Have you explored whether the effectiveness of Reasoning Advantage (RA) varies across different types of reasoning tasks beyond math and standard QA?\\nIn Section 5.2, you show that optimizing for RA leads to a 7% improvement on GSM8K. Could you provide more analysis of what specifically improved in the model's reasoning capabilities? Are there particular types of math problems where the improvement was more pronounced?\\nThe paper mentions that only 0.01% of generated CoTs achieve a reward above 0.2 on OpenWebMath. Have you analyzed these high-scoring CoTs to understand what makes them successful? This analysis could inform better prompting strategies.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"title\": \"Response to Reviewer ZH4y\", \"comment\": \"Thank you very much for your thoughtful feedback on our work. We appreciate the time and effort you\\u2019ve dedicated to reviewing our paper.\\n\\n>While your clarifications addressed some of my initial concerns, I still feel that the paper's contributions are somewhat limited. Specifically, while the analysis of reward functions is insightful, the work does not propose a sufficiently robust method to tackle the challenge of self-improving reasoning in LLMs comprehensively.\\n\\n\\nWe acknowledge that our method does not yet achieve state-of-the-art results in self-improvement compared to existing supervised approaches. However, we would like to emphasize that the primary aim of our research was not necessarily to surpass current state-of-the-art methods in this domain. Instead, our work explores an exciting direction\\u2014self-improving reasoning without relying on supervised datasets. While the state of the art in this research direction currently yields less favorable empirical results than supervised methods, we believe that it could have significant long-term impact and is therefore worthy of continued attention.\\n\\n**Our study addresses a critical roadblock in this research direction**. We demonstrate that existing reward functions fall short, *even in an intermediate setting*, and introduce a novel reward function that succeeds where others fail. Moreover, we propose a new approach for evaluating reward functions: the \\\"what/where\\\" experiments, which help us identify the most effective reward function. We think that achieving complete self-improvement on unstructured text would be a groundbreaking result worthy of a very high (8-10) score. But given how many researchers\\u2014even in industry\\u2014are struggling with this challenge, systematic investigation of specific components like reward functions is crucial for advancing the field. We think this represents precisely the type of research progress that ICLR aims to promote.\\n\\n>The evaluation on MMLU is conducted under a zero-shot setting, whereas MMLU is more commonly assessed with 5-shot prompts. This makes it difficult to compare your results with standard baselines.\\n>\\n>In your response to Reviewer qdDb, you referenced Quiet-Star, which evaluated CSQA. Additionally, many recent works on enhancing LLM reasoning capabilities have used BBH for evaluation. Including both CSQA and BBH in your benchmarks would result in a more comprehensive evaluation.\\n\\nThanks for the question. We base our evaluation methodology on the Quiet-STaR [1] paper, one of the most important works in LLM reasoning self-improvement. In this work, Zelikman et al. evaluate on two downstream reasoning benchmarks using zero-shot evaluation. We have taken the same approach in our work, but we\\u2019d be happy to include 5-shot prompting results in the final version of our paper.\\n\\nMoreover, while additional datasets would add to our evaluation, we believe that our current evaluation *over two models and five datasets* provides strong evidence for the strength of the RA reward function, especially considering that all the other reward functions fail to show any self-improvement. Note that our method should be compared to existing unsupervised methods, as opposed to methods using supervised datasets.\\n\\n>It remains unclear whether the proposed RA function will continue to provide benefits as model parameter counts increase. Would its effectiveness diminish as model performance improves?\\u201d\\n\\nRegarding model size, our experiments use 7B and 8B parameter models, which is standard practice in academic research. For example, Quiet-STaR [1] uses Mistral 7B, RAFT [2] uses Llama 7B, and RHO-1 [3] uses 1B and 7B models. As academic researchers, we show that our method works using academic computing resources. We believe this aligns with ICLR's academic focus, and we leave investigations using larger models to labs with industrial-scale compute budgets.\\n\\nThanks again for your continued engagement with our work. We believe that the updates we have made in response to your comments and questions (red text) have significantly improved the quality and clarity of our paper.\\n\\n[1] Zelikman et al, Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking, 2024\\n\\n[2] Dong et al, RAFT: Reward rAnked FineTuning for Generative Foundation Model Alignment, 2023\\n\\n[3] Lin et al, Rho-1: Not All Tokens Are What You Need, 2024\"}", "{\"comment\": \"Thank you again for providing an initial review that helped us to improve various aspects of the paper, both in terms of presentation and additional experiments. Having made extensive clarifications of the paper\\u2019s contributions and included additional discussion on why the log-likelihood is an appropriate starting point for a reward function for general-purpose reasoning, have we addressed all of your concerns? If so, will you consider increasing your score? Please do not hesitate to share any further questions.\"}", "{\"comment\": \"Thank you for your clarification and additional experiments. I will keep my score accordingly.\"}", "{\"title\": \"Response to Reviewer ZH4y [2/2]\", \"comment\": \"To be clear, our work does not solve the grand challenge mentioned. Our work demonstrates that current reward functions fail for even the small step of MMLU-FREE-FORM, introduces a novel reward function that mitigates this problem, and performs a comprehensive analysis to help facilitate future research in this direction. We believe these to be major contributions to the literature, and have dramatically updated our Introduction, Section 5, Section 6, and Conclusion with more targeted descriptions of our contributions. The key points are updated with red text.\\n\\n>The paper lacks discussion on relevant work in reasoning enhancement during the pre-training phase, such as https://arxiv.org/pdf/2404.07965.\\n\\nThank you for noticing the connection between RHO-1 and our work. We have added a description of this paper to the end of our Related Works (Section 3, blue text). RHO-1 selectively trains on useful tokens during pre-training, which enhances reasoning downstream. Also, after looking at RHO-1 in more detail, we\\u2019d be very excited for future work that combines RHO-1 with RA (i.e., to perform RL with CoT on datapoints that are suitable for reasoning, not noisy, and not yet learnt). We have included a mention of this at the end of Section 6.1 (blue text).\\n\\nWe also commit to surveying the literature regarding other non-RL methods that enhance reasoning during pre-training and including a discussion of these works in the final copy of our manuscript.\\n\\n>The experiments are insufficiently comprehensive, as they are conducted on only one model and one dataset.\", \"we_would_like_to_emphasize_that_our_paper_contains_multiple_experiments_which_span_multiple_datasets\": \"(1) our \\u201dwhat reasoning to reward\\u201d experiment uses datapoints from FineWeb, (2) our \\u201cwhere reasoning is rewarded\\u201d experiment uses datapoints from MMLU, GSM8K, and CSQA, (3) our self-improving CoT reasoning experiment for the free-form QA setting evaluates on MMLU and GSM8K, and (4) our exploratory experiment on the grand challenge of the self-improvement on truly unstructured text uses OpenWebMath.\\n\\nThat being said, we acknowledge that the paper would be strengthened by running the same experiments with different pretrained models of different sizes. Therefore, we have repeated the *what* and *where* experiments from Section 5.1 with another model, LLama 3.1 8B. We have added these results in Appendix B Table 5 and find them to be consistent with our original results using Mistral 7B in Table 2.\\n\\n>The presentation of the paper could be improved. Some key findings should be in the main body rather than the Appendix, such as Appendix D and the definition of RA in Appendix A. Essential parameters, like the type of LLM used and inference hyperparameters, should also be included in the main text.\\n\\nWe thank the reviewer for identifying each of the specific ways in which clarity could be improved. We have reworked Appendix D into the new Section 6.2, and have added additional clarification to the definition of RA in Section 4. We\\u2019ve also added more information about the type of LLM used and inference hyperparameters (i.e., sampling temperature) in Section 5.1, Section 5.2, and Section 6.2 (in blue text).\\n\\n\\nWe conclude by stressing that while we do not solve every challenge, our work represents a large step towards self-improving CoT reasoning on unstructured text at the pretraining scale. As more researchers from the LLM reasoning community shift focus towards this goal, we think that analyses like ours, which isolate and address specific issues with our current self-improvement methods, will provide great value and enable exciting future research.\\n\\nThanks again for the thorough review. We hope that some of our additional explanations and experiments, along with the changes we've made in response to the issues you raised, go a long way towards improving the paper and changing your opinion. Please don\\u2019t hesitate to ask if you have any additional questions or require clarification.\"}", "{\"title\": \"Response to Reviewer KyrV [1/2]\", \"comment\": \"We thank the reviewer for the thoughtful review, and for recognizing the value in using MMLU-FREE-FORM as an intermediate benchmark between curated QA data and the unstructured text setting. We address your questions and comments below.\\n\\n>The paper's primary limitation appears in the scaling to general pre-training data, where the offline reinforcement learning approach that worked well on MMLU-FREE-FORM struggles to escape local optima of conservative reasoning. While the authors acknowledge this limitation and suggest future research directions, the paper doesn't fully solve the challenge of self-improving reasoning at pre-training scale. \\n\\nWe agree with the reviewer that our work does not solve the grand challenge of self-improving CoT reasoning on unstructured, pretraining-scale text. Our primary contributions are showing that standard reward functions fail *even in* the intermediate MMLU-FREE-FORM setting, introducing a novel reward function to solve this issue, and performing a comprehensive analysis of reward functions and how they affect *what* and *where* reasoning is rewarded. To our knowledge, our work is the first to provide this type of analysis on reward functions for self-improving CoT reasoning on unstructured text. \\n\\nWe also perform an exploratory experiment on the full unstructured setting and provide key insights to help facilitate future research in this direction. We believe these to be major contributions to the literature, and have dramatically updated our Introduction, Section 5, Section 6, and Conclusion with more targeted descriptions of our contributions. The key points are updated with red text.\\n\\n>While the authors demonstrate improved performance on mathematical reasoning tasks, there could be more exploration of how well their approach generalizes to other types of reasoning beyond mathematics.\\n\\nWe thank the reviewer for this suggestion. While GSM8K focuses purely on mathematics, MMLU contains questions which span various fields, including mathematics, sciences, law, etc. On MMLU, we actually find that our method leads to large gains on questions across a wide span of subjects (biology, physics, accounting, law, computer science, etc.) that involve quantitative reasoning\\u2014*going beyond just mathematics*. We have updated Appendix B.2 to better explain this point (blue highlighted text). Thanks again for the suggestion.\"}", "{\"summary\": \"The paper titled explores the potential for self-improvement in large language models' ability to perform CoT reasoning without the need for supervised datasets. The authors frame this as a reinforcement learning problem where an LLM generates a CoT to predict subsequent tokens in a text corpus, receiving a reward based on the effectiveness of the CoT in predicting the next tokens. Their approach explores generating CoTs for next-token prediction in unstructured data, aiming to improve general-purpose reasoning abilities.\", \"soundness\": \"3\", \"presentation\": \"1\", \"contribution\": \"2\", \"strengths\": \"1. The paper presents a novel approach to improving CoT reasoning in LLMs, exploring reinforcement learning as a framework for unsupervised self-improvement. The introduction of RA offers an innovative solution to the reward function challenge.\\n2. This work addresses a crucial challenge in LLM development\\u2014achieving autonomous improvement in reasoning without reliance on human-generated data. If successful, this approach could significantly reduce reliance on expensive, curated datasets and enable more scalable reasoning improvement across diverse domains.\", \"weaknesses\": \"1. Some aspects of the reinforcement learning formulation could benefit from additional clarity, specifically regarding the choice of reward clipping values and the normalization strategies within RA. Additional explanation of these parameters and their impact on performance would make the approach more accessible.\\n2. The experiments focus primarily on a limited scope of problems (e.g., MMLU and OpenWebMath). The model\\u2019s performance on broader tasks, such as Tool learning or agent problem-solving scenarios, would offer stronger evidence of the approach\\u2019s generalizability.\\n3. By relying on the log-likelihood to evaluate the quality of intermediate reasoning (Chain-of-Thought) solely based on the model's ability to predict the following tokens, there is a risk that the model may overly focus on matching specific token patterns in the training data rather than developing generalized reasoning capabilities.\", \"questions\": \"See above.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to Reviewer qdDb\", \"comment\": \"> I really agree with ZH4y that the statement of pretraining stage may be misleading. I think the authors should carefully consider the contribution points of this paper, especially the term \\\"pretraining\\\"\\n\\nThank you for your response. We agree and have made considerable updates to the Introduction, Section 5, Section 6, and Conclusion with clearer descriptions of our contributions. The main updates are in *red text*. In our updates, we make it specifically clear that we do not solve the full unstructured pretraining setting.\\n\\nTo make this even more clear, we also propose to change our title to \\u201c*On Reward Functions for Self-Improving CoT Reasoning Without Supervised Datasets*\\u201d. \\n\\nTo our knowledge, our work is the first to provide this type of analysis on reward functions for self-improving CoT reasoning on unstructured text. We demonstrate that existing reward functions fail *even in* the simpler MMLU-FREE-FORM setting, introduce the novel RA reward function as an effective solution, and perform a comprehensive analysis of reward functions and how they affect what and where reasoning is rewarded. We believe that the successful zero-shot transfer from self-improving CoT reasoning on MMLU-FREE-FROM to a popular unseen math benchmark (GSM8K) is a promising result that motivates future work in this direction.\\n\\nDo our updates to the paper and the title change address your concerns regarding the clarity of our contributions?\\n\\n> Meanwhile, the author should add more discussions on why the log-likelihood, as a reward signal, can promote the enhancement of the model's general reasoning ability.\\n\\nThank you for the suggestion. At the end Section 5.2 (*blue text*), we have added a discussion on how loss-based reward functions like RA can promote the enhancement of general reasoning ability. This discussion details how we reward reasoning that minimizes a form of loss on subsequent tokens, and references recent work [1] showing that optimizing for loss during pretraining improves performance on downstream reasoning tasks. We also explain how our experiments demonstrate that RA\\u2019s key modifications to standard loss (i.e., clipping, baseline, and normalization) are crucial for generalizing reasoning to unseen tasks.\\n\\n[1] Z. Du et al, Understanding Emergent Abilities of Language Models from the Loss Perspective, 2024\"}", "{\"title\": \"Official Review by Reviewer MbaU\", \"comment\": \"Thank you for your clarification and additional experiments.\\nI will keep my score accordingly.\"}", "{\"title\": \"Response to Reviewer qdDb\", \"comment\": \"Thank you for your review and comments. We\\u2019re glad that you think our work addresses a crucial challenge in LLM development and that you see the introduction of RA as an innovative solution to the challenge of designing reward functions for reasoning. Please see below for responses to your comments and questions.\\n\\n>Some aspects of the reinforcement learning formulation could benefit from additional clarity, specifically regarding the choice of reward clipping values and the normalization strategies within RA. Additional explanation of these parameters and their impact on performance would make the approach more accessible.\\n\\nThanks for the great suggestions. We have updated the parts of Section 4 discussing clipping and normalization to be more clear. We have also added a detailed discussion of their impact on performance to the end of Section 5.1 (the two big groups of blue text), including an additional ablation for different values of the clipping threshold. Moreover, Appendix B.1 contains tables which show full results for additional combinations of clipping, baseline, and normalization.\", \"briefly_summarizing_how_each_design_choice_relevant_to_ra_can_be_interpreted\": \"- Clipping value: the minimum value at which suffix token log-probabilities are clamped. This prevents any given token from having an outsized loss contribution. We have run an additional ablation on a range of clipping values to demonstrate its impact, and include the results in Appendix B.1 (Figure 5).\\n\\n- Baseline: in RA, we compute the token log-probabilities for the suffix given the prefix and CoT, but subtract the \\\"empty CoT\\\" baseline, which is the token log-probabilities without conditioning on any CoTs (only the prefix). This ensures we are optimizing for CoTs that *improve* the suffix loss relative to not producing a CoT.\\n\\n- Normalization: normalizing by the \\\"empty CoT\\\" baseline re-scales the reward to ensure that we don't provide high reward values for CoTs with trivial-to-predict suffixes (ie, when an \\\"empty CoT\\\" predicts the suffix well).\\n\\n>The experiments focus primarily on a limited scope of problems (e.g., MMLU and OpenWebMath). The model\\u2019s performance on broader tasks, such as Tool learning or agent problem-solving scenarios, would offer stronger evidence of the approach\\u2019s generalizability.\\n\\nWe thank the reviewer for their suggestion, it would indeed be interesting to evaluate the model\\u2019s tool learning and agentic behaviour. However, in keeping with much of the LLM reasoning literature, we chose to specifically focus our evaluations on a range of QA-style reasoning problems. One of the most important recent works on self-improving LLM reasoning, Quiet-STaR [1], uses two QA-style reasoning benchmarks to evaluate their method: GSM8K and CSQA. Similarly, we evaluate our self-improvement method from Section 5.2 on two QA-style reasoning benchmarks: MMLU and GSM8K.\\n\\n>By relying on the log-likelihood to evaluate the quality of intermediate reasoning (Chain-of-Thought) solely based on the model's ability to predict the following tokens, there is a risk that the model may overly focus on matching specific token patterns in the training data rather than developing generalized reasoning capabilities.\\n\\nWe view the strong zero-shot transfer performance to GSM8K after optimising for RA on MMLU-FREE-FORM as compelling evidence that the model learns generalisable reasoning \\u2014 beyond just matching specific token patterns in the data. Moreover, we strongly believe that RA being based on token loss is actually a key advantage. Since it is loss-based, it requires only one forward pass, does not require an external strong model, is not limited to exact-match heuristics like accuracy-based functions (which fail on unstructured text), and allows the model to place weight over a distribution of valid answers.\\n\\nIt might also be worth mentioning that since we are starting from a pretrained LLM, much of the gains from attempting to match specific token patterns have already been exhausted. That is, at the beginning of standard LLM pretraining, most of the loss reduction is achieved by fitting specific token patterns like spelling and grammar rules, but by further reducing loss, models begin to learn higher-order skills such as CoT reasoning. We initialize our weights to a pretrained LLM, so the risk of overly focusing on specific token patterns is limited.\\n\\nThanks again for your review and helpful suggestions. We are excited to see the insights from our work help progress the field towards the grand challenge of self-improving CoT reasoning on unstructured text at the pretraining scale. If you feel that we have adequately addressed your concerns, we would appreciate your consideration to increase our score. And if you have any additional questions or needs for clarification, please don\\u2019t hesitate to ask.\\n\\n[1] Zelikman et al, Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking, 2024\"}", "{\"comment\": \"Thank you for your detailed response and the significant improvements made to the paper. While your clarifications addressed some of my initial concerns, I still feel that the paper's contributions are somewhat limited. Specifically, while the analysis of reward functions is insightful, the work does not propose a sufficiently robust method to tackle the challenge of self-improving reasoning in LLMs comprehensively.\", \"i_also_remain_uncertain_about_some_aspects_of_the_experimental_results\": \"The evaluation on MMLU is conducted under a zero-shot setting, whereas MMLU is more commonly assessed with 5-shot prompts. This makes it difficult to compare your results with standard baselines.\\n\\nIn your response to Reviewer qdDb, you referenced Quiet-Star, which evaluated CSQA. Additionally, many recent works on enhancing LLM reasoning capabilities have used BBH for evaluation. Including both CSQA and BBH in your benchmarks would result in a more comprehensive evaluation.\\n\\nIt remains unclear whether the proposed RA function will continue to provide benefits as model parameter counts increase. Would its effectiveness diminish as model performance improves?\\n\\nGiven the substantial changes you\\u2019ve made compared to the initial submission (including changes to the title and contributions), I believe the paper could benefit from further refinement before being resubmitted in a future review cycle.\"}" ] }

[ "Hadas Abraham", "Barak Gahtan", "Adir Kobovich", "Orian Leitersdorf", "Alex M. Bronstein", "Eitan Yaakobi" ]

[ "Deep learning", "DNA storage", "Science", "Clustering", "Sequencing" ]

{ "note_id": [ "zTaBq7bAQu", "h61unvcnHI", "LoNF7JwoP9", "2Vu95tloK2", "2EXhg2Iy7H" ], "note_type": [ "official_review", "official_review", "official_review", "official_review", "comment" ], "note_created": [ 1730302912038, 1730021442408, 1730556431034, 1730749928198, 1732547651047 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission6663/Reviewer_6ixx" ], [ "ICLR.cc/2025/Conference/Submission6663/Reviewer_ErVP" ], [ "ICLR.cc/2025/Conference/Submission6663/Reviewer_KRVi" ], [ "ICLR.cc/2025/Conference/Submission6663/Reviewer_2jVo" ], [ "ICLR.cc/2025/Conference/Submission6663/Authors" ] ], "structured_content_str": [ "{\"summary\": \"The paper presents a deep learning-based approach for clustering DNA reads using raw electrical signals generated during nanopore sequencing. Traditionally, DNA clustering occurs after basecalling, where raw signals are converted into nucleotide sequences. The authors propose to bypass this step by clustering directly on the raw signals, which retain more information than the processed sequences.\\n\\nThe paper proposes the use of deep signal embedding to improve the clustering stage of DNA storage pipelines, bypassing traditional methods that rely on basecalled DNA reads. Experiments indicate that the proposed method outperforms state-of-the-art approaches.\", \"soundness\": \"4\", \"presentation\": \"3\", \"contribution\": \"4\", \"strengths\": [\"The core contribution is a model, based on the model in the Dorado basecaller, that performs clustering on these raw signals before basecalling. To my knowledge, this signal-based approach is novel and promising in the field of DNA storage.\", \"The authors evaluate their method on several DNA datasets. Their results show that signal-based clustering outperforms existing methods in terms of both time and accuracy.\"], \"weaknesses\": \"Major:\\n- An important point that the authors should address is that it is uncertain whether the clustering gains translate into significant improvements in the final data retrieval phase.\", \"minor\": [\"Line 40: The description 'a \\\"retrieval\\\" stage where reads are decoded back to binary data files while correcting any errors using the chosen coding methods' is inconsistent with Figure 1, where the decoding is shown to occur after the retrieval stage.\", \"Line 80: \\\"edits\\\" should be replaced by \\\"substitutions\\\"\", \"Line 104: \\\"contributions to\\\" -> \\\"contributions of\\\"\", \"Line 156: How can UCLUST (published 2020) be based on USEARCH (published 2021)?\", \"Line 173: There are also algorithms such as GCTW (Zhou and De la Torre, 2016) for sequence alignment with sub-quadratic complexity. This should be acknowledged.\", \"Line 183: \\\"during during\\\" -> \\\"during\\\"\", \"Line 186: \\\"maps\\\" -> \\\"map\\\"\", \"Lines 189-193: This paragraph seems rather disorganized.\", \"Line 214 (footnote): \\\"Deep DNA test\\\" -> \\\"Deep DNA Test\\\", \\\"Deep DNA pilot\\\" -> \\\"Deep DNA Pilot\\\"\", \"Lines 199-206: This paragraph needs improvement. The text says that the synthesis will produce '\\\"fast5\\\" file formats'. Firstly, the synthesis yields oligonucleotides. Secondly, nanopore sequencing would yield raw signal data in the form of files in the FAST5 format.\", \"Line 223: I do not understand how to deduce from Table 1 that the strands are correctly clustered with the k=20 maximum distance constraint.\", \"Line 247: \\\"Clmap\\\" -> \\\"Clamp\\\"\", \"Line 250: The authors should be clearer about what is meant by the number of classes. It only becomes clear later that this refers to the number of (expected) clusters.\", \"Figure 2: The figure could be improved. On the left, it is not clear that the input files are used to synthesize the DNA. Also, according to the figure, it seems that the ArcFace loss is an input to the CNN.\", \"Line 409: \\\"due to is due to\\\" -> \\\"due to\\\"\"], \"questions\": [\"Raw sequencing signals can vary considerably due to noise and machine-specific artifacts. How does the model handle this variability, and what measures are taken to ensure robust clustering across varying signal qualities?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The manuscript introduced a DL-based approach for clustering sequences retrieved from the DNA sequencing pipeline, utilizing raw signals from Nanopore sequencing method. The proposed model is derived through fine-tuning an open-source signal-to-base model known as Dorado, under a classification task.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"The approach of commencing the analysis from raw Nanopore signals, rather than relying on pre-processed discrete DNA sequences, represents a novel direction in the field.\\n\\nSequence clustering is a challenging problem, especially when dealing with a large number of sequences. The proposed method may have significant impact on the DNA storage community.\", \"weaknesses\": \"There are several weaknesses that the authors may need to address.\\n\\n0. The authors should pay more attention to the typos, grammars, etc. e.g line125 \\\", while those that are not are for apart. \\\"\\n\\n1. There are existing works that employs deep learning techniques to assess the similarity between sequences (DNA sequences), which are very closely related to this manuscript but omitted in this work. e.g. \\n```\\n1. \\\"Convolutional embedding for edit distance.\\\" SIGIR 2020\\n2. \\\"Neural distance embeddings for biological sequences.\\\" NeurIPS 2021\\n3. \\\"Deep Squared Euclidean Approximation to the Levenshtein Distance for DNA Storage.\\\" ICML 2022\\n```\\n\\n2. To the best of the reviewer\\u2019s knowledge, none of the three datasets utilized in this study are available in an open-access format as raw signals. This limitation may impede the reproducibility and validation of the proposed method. \\n\\n3. The author(s) have chosen to employ the ArcFace loss function as a substitute for the cross-entropy loss in the classification task. However, the manuscript lacks an analysis or ablation study that would substantiate the necessity of introducing this more complex loss function. \\n\\n4. The contribution of the proposed method to the machine learning and representation learning communities, in terms of novelty, appears to be limited. Most of the involved methods in this work are mature DL techs. The primary novelty of this work lies in utilizing the untapped potential information from the raw signals. Given this focus, the manuscript would be more suited for a bioinformatics-specific paper. \\n\\n5. It is not demonstrated that whether the untapped potential information from the raw signals helps in clustering the sequences, which is claimed as the main contribution as in the Abstract. There should be a comparision between the raw-signal/DNA-sequence as inputs to support this claim. \\n\\n6. The reviewer also concerns about the effectiveness of employing supervised classification tasks to derive embeddings, \\nparticularly since approximating distance metrics is a relatively straightforward task with the datasets used in this study, which is also used in the existing works presented in item 1. \\n\\n7. The author(s) may want to improve the writting. e.g. the reviewer did not find any table or figure which is refered by \\\"The results\\\" in line 313.\", \"questions\": \"N/A\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper proposes a novel approach to clustering in DNA data storage by focusing on clustering raw signals from Nanopore sequencing, rather than traditional basecalled DNA sequences (A/T/C/G). Traditional DNA storage pipelines typically perform clustering after basecalling, which introduces errors and computational inefficiencies. The authors aim to improve clustering accuracy and efficiency by directly embedding the raw signals produced during sequencing. The proposed deep learning model integrates the Dorado basecaller architecture and an ArcFace loss function, allowing for similarity-based grouping of raw signals. This approach is demonstrated to enhance clustering accuracy while reducing computational time by one to three orders of magnitude compared to traditional sequence-based methods. Results are validated on multiple datasets, and the model shows potential for enhancing both the clustering and reconstruction phases of DNA data storage systems.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"Originality:\\nThe paper addresses a novel problem by introducing raw signal clustering in DNA storage, which could set a new direction for handling sequencing data.\", \"relevance\": \"With data storage demands growing exponentially, improving DNA storage efficiency and accuracy is highly relevant.\", \"computational_efficiency\": \"The use of deep embeddings and cosine similarity yields significant computational improvements, making the method scalable and suitable for high-throughput sequencing.\", \"clarity_in_experimental_validation\": \"Experimental results across multiple synthetic datasets support the authors\\u2019 claims, with performance metrics like AUC, ROC curves, and computation times well-documented.\", \"weaknesses\": \"Limited Generalizability Due to Custom Dataset:\\nThe dataset is highly customized, relying on specific design files and synthetic DNA sequences generated by a particular synthesis provider (Twist Bioscience). Since real-world DNA samples often include much higher variability, especially in natural genomic data, the results from this dataset may not generalize well to other applications or to DNA data with biological origins rather than synthetic sources.\\n\\nFixed Threshold for Edit Distance (k=20):\\nThe authors limit their edit distance calculations to a maximum of 20 differences, which helps with computational efficiency but may introduce inaccuracies. For reads that are more distant from their original design strands, this constraint could lead to incorrect cluster assignments, particularly if the dataset scales up to include a broader diversity of sequences with higher variances.\", \"different_similarity_comparisons\": \"The comparison between raw signal embedding using cosine similarity and DNA strands using edit distance may not be entirely fair due to their application to different data types\\u2014continuous versus discrete. Cosine similarity captures vector orientation in high-dimensional spaces, while edit distance measures specific sequence transformations. This discrepancy can lead to inconsistencies in capturing variations and information. Ensuring a fair comparison might require using a consistent metric or providing a rationale for the chosen methods.\", \"may_lack_of_novelty\": \"Aside from making and utilizing raw signals, the paper lacks additional innovation in either its model architecture, training strategies, or data processing. The use of dorado CRF layers and standard methods doesn't introduce novel techniques. To enhance innovation, further exploration in these areas would be beneficial.\", \"questions\": \"Generalizability of the Dataset:\\nHow do you plan to address the potential lack of generalizability due to the customized dataset? Have you considered testing the method on natural genomic data to assess its applicability in real-world scenarios?\", \"edit_distance_threshold\": \"Could you elaborate on the decision to set the edit distance threshold at 20? How do you plan to ensure accuracy in clustering when dealing with reads that might have more variations? Can you provide an ablation study on the impact of choosing different edit distance thresholds on clustering accuracy?\", \"similarity_comparisons\": \"Can you provide a rationale for using different similarity measures for raw signals and DNA strands? How might you align these metrics to ensure consistency in comparison?\", \"model_architecture_and_novelty\": \"Are there any future plans to innovate the model architecture or training strategies beyond using raw signals? What potential modifications could enhance the novelty of your approach?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper tackles the clustering problem of DNA storage by using raw Nanopore signals as opposed to first translating the signals to strings (basecalling) and then clustering on the text strings. The motivation here is clear because the process to go from raw signals to strings can be very slow (weeks to months) and hence operating directly on the signals can be more efficient. Therefore, the authors design a deep learning method that operates on the data before it is discretized into bases. This paper presents a deep neural network that clusters these signals directly with reduced computation times compared to current approaches that cluster after basecalling.\\n\\nFrom a technical point of view, the authors start with a pre-trained model and then do fine-tuning on additional data. The base model is Dorado, which is the standard basecaller for Nanopore reads. After truncating to get an embedding, they append a linear layer with output dimension equal to the number of classes (number of clusters). Then the training also uses the ArcFace loss function to increase separation between the clusters based on cosine similarity.\", \"soundness\": \"1\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"This paper studies an important problem in DNA storage. It has long been an open question whether systems can utilize information in the raw signals to better improve the analysis and downstream use of Nanopore data.\\n\\nThe authors show the effectiveness of taking a pre-trained model (Dorado) for Nanopore basecalling and repurposing it for a downstream application of the raw signals. This is pretty interesting because it is good to know the pre-trained model can be useful, and researchers do not have to start from scratch and train a model on Nanopore data themselves. \\n\\nThe authors also motivate their work by saying that the computation time can be much faster than basecalling with an existing algorithm. This is indeed an important aspect to improve upon. The methods in this paper could take the signal processing time from days/weeks down to under a single day, which would make DNA data storage more feasible for recovering stored data.\", \"weaknesses\": [\"A major weakness of the approach seems to be that the model needs to be trained with an output dimension equal to the number of clusters. However, the number of clusters in real settings can be very, very large. E.g., for a file that is stored that is 100MB, there could easily be over 1M clusters. I don't think network training will scale well in this case. So the experiments on 500 clusters are not representative of a real setting. Concretely, can the authors address how their approach might scale to much larger numbers of clusters (e.g. 1M+) that would be encountered in real-world applications. Please discuss potential solutions or modifications to their method to handle this challenge.\", \"In general, the experiments in Section 6 seem quite ad-hoc. The number of clusters is restricted without much discussion. One interesting outcome of this research could be how to train/design a good Nanopore signal clustering algorithm. But there are very few ablations, and not much details about what the authors learned or why their method works/fails. It feels as if the authors did one set of experiments and quickly wrote them up. This is below the bar for a ML paper at a top conference. There should be more tangible findings, and more justification for the design decisions. Specifically, can the authors investigate the impact of varying the number of clusters, explore different network architectures, and analyze how different components of their method contribute to its performance.\", \"The experimental results in Section 7 are missing comparisons to baselines. For example, the conclusion states that \\\" Experiments reveal that the signal-model outperforms conventional clustering algorithms like Clover and Microsoft\\u2019s, providing faster and more accurate results.\\\" However, I don't see these experiments in the paper. For example, Table 2 and Figure 6 only seem to present results for the authors' algorithm, not the baselines (e.g., Clover or Microsoft's algorithms for clustering). It seems necessary to include direct comparisons with Clover, Microsoft's algorithm, and other relevant baselines in Table 2 and Figure 6, using the same datasets and evaluation metrics. This would provide a clearer picture of the proposed method's performance relative to existing approaches.\", \"At a high level, the related work is oddly about clustering, but the paper is about more than clustering. I am not familiar with things like how the Dorado method was trained, what competing methods there are, etc. Also have people really not tried to train deep learning algorithms on Nanopore data for other applications? This seems very surprising to me. For example, a quick search shows some other papers that use nanopore signals for applications:\", \"HycDemux: a hybrid unsupervised approach for accurate barcoded sample demultiplexing in nanopore sequencing\", \"Renmin Han, Junhai Qi, Yang Xue, Xiujuan Sun, Fa Zhang, Xin Gao & Guojun Li\", \"Genome Biology volume 24, Article number: 222 (2023)\", \"Kovaka S, Fan Y, Ni B, Timp W, Schatz MC. Targeted nanopore sequencing by real-time mapping of raw electrical signal with UNCALLED. Nat Biotechnol. 2021;39(4):431\\u201341.\", \"Overall, can the authors expand their discussion of related work to include: A more comprehensive overview of Nanopore signal processing methods, including Dorado and its competitors; A broader discussion of deep learning applications on Nanopore data beyond clustering; An analysis of how their approach compares to or builds upon these, as well as other relevant work in the field.\"], \"questions\": [\"The claims about previous clustering algorithms do not seem very accurate or thorough. For example, the authors say \\\"Unlike modern DNA clustering algorithms which are commonly used, such as Clover (Qu et al., 2022) and Microsoft\\u2019s (Rashtchian et al., 2017), the signal-model computation time is considerably faster.\\\" I am not sure how this is verified. For example, one inaccuracy is in Section 7.1 -- the authors say \\\"Microsoft\\u2019s algorithm (Rashtchian et al., 2017) is orders of magnitude slower than Clover\\\" but the cited paper GradHC (Ben Shabat, 2023) points out that the Microsoft algorithm is very fast, except when the clusters are very small and it does not converge (which may not be the case for the datasets in this paper, since the number of clusters is 50 in Section 6, and the number of reads is much larger). Also why not compare to the GradHC algorithm, which is shown to be quite fast?\", \"I am confused about \\\"ground truth for clustering\\\" and in particular the statement \\\"Table 1 shows that for each of the experiments, k = 20 ensures each strand will be categorized to the correct cluster with high probability since the majority of the reads are closely aligned to their intended design\\\" In Table 1, it seems that for the first two datasets, the average edit distance is between 52 and 63. So what happens for the strands with edit distance more than k = 20? Will they be assigned to a different cluster?\", \"Section 7.2 seems to be unfinished. There are no experiments to back up the claims. Also this sentence in 7.2 does not make sense as written: \\\"In contrast, since signal-model directly uses the raw signals, the clustering algorithm will be given more data for the reconstruction algorithms.\\\" The clustering algorithm will operate on the reads, while the reconstruction algorithm will operate on the clusters. I don't know what it means for one algorithm to be given more data for another algorithm.\", \"One of the more interesting aspects to explore is whether other parts of the DNA storage system can be replaced with deep learning. For example, is it possible to train a model to go directly from raw signals to the final prediction for the stored data? Perhaps the network can learn to implicitly cluster then reconstruct then error correct, without needing separate algorithms for each of these steps.\", \"See also the questions at the end of each of the weaknesses above.\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"withdrawal_confirmation\": \"I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors.\"}" ] }

[ "Wooseong Jeong", "Kuk-Jin Yoon" ]

[ "Multi-Task Learning", "Token Modulation and Expansion", "Conflicting Gradients" ]

{ "note_id": [ "se2VBMIdam", "rQZ84o3yRG", "X20DfsjAqc", "N7Pn6kOtct", "7NLnuP7NI6" ], "note_type": [ "comment", "official_review", "official_review", "official_review", "official_review" ], "note_created": [ 1731622000289, 1730633850026, 1729872573117, 1730619526922, 1730850910455 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission4892/Authors" ], [ "ICLR.cc/2025/Conference/Submission4892/Reviewer_eXAe" ], [ "ICLR.cc/2025/Conference/Submission4892/Reviewer_M8YR" ], [ "ICLR.cc/2025/Conference/Submission4892/Reviewer_ZcwA" ], [ "ICLR.cc/2025/Conference/Submission4892/Reviewer_YC8T" ] ], "structured_content_str": [ "{\"withdrawal_confirmation\": \"I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors.\", \"comment\": \"Thank you to all reviewers for their sincere feedback and efforts. Unfortunately, we have decided to withdraw our paper.\"}", "{\"summary\": \"The paper tackles the problem of multi-task learning (MTL), more specifically in the context of designing MTL architectures. Such architectures usually rely on task-specific vs shared features. The main hypothesis of the paper is that defining such architectures in a fixed manner is suboptimal, and they should instead adapt dynamically to the tasks at hand.\\n\\nTo do this, the proposed method estimate **gradient conflicts** between tasks, but in a more fine-grained manner than previous work. First, the proposed method performs a SVD on the token space, sorting the resulting space by eigenvalues, which are then used to divide it into two parts: the range space (high eigenvalues) and the null space (low eigenvalues). Then, the task gradients are projected on each of these two subspaces, resulting into two different types of gradient conflicts (in the range space or in the sound space).\\n\\nBased on the type of conflict, the architecture can be modified in two ways:\\n * token modulation (conflict in the range space): Addition of an extra linear transformation of the shared tokens\\n * token expansion (conflict in the null space): Addition of extra task-specific tokens\\n\\nThe proposed method is evaluated on NYU-D and PASCAL (3 to 5 tasks) and on a subset of Taskonomy (11 tasks)\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": [\"The method description and accompanying figures are clear\", \"Comparison on a varied number of tasks and with different degrees of task interference\", \"Interesting ablations such as when (during training) to start token expansion\"], \"weaknesses\": [\"## Literature review\", \"My main concern with the paper is the literature review and justification of the method. In particular, there is a lot of work on dynamic architectures for MTL not mentioned in the paper, and, based on the introduction, there\\u2019s also little motivation for introducing the separate range and null space to perform the conflicting gradient analysis.\", \"Dynamic architectures for MTL in vision is not a new idea in particular for CNN architectures, and many of such methods could be readily adapted to ViT. Examples of such literature:\", \"\\\"Efficiently Identifying Task Groupings for Multi-Task Learning\\\" and \\u201cWhich tasks should be learned together in multi-task learning?\\u201d -> design separate encoders for (learned) subgroups of tasks.\", \"\\u201cStochastic filter groups for multi-task cnns: Learning specialist and generalist convolution kernels\\u201d\", \"\\u201cLearning to branch for multi-task learning\\u201d -> Neural architecture search for MTL\", \"\\u201cLatent multi-task architecture learning\\u201d\", \"\\u201cLearning multi-level task groups in multi-task learning\\u201d\", \"Optimisation vs architecture-based MTL: Many of the baselines in experiments are optimisation-based ones (e.g. GradNorm, CAGrad etc) which mainly deals with reweighing the various task losses. However, these methods are rather orthogonal with the proposed method, since we could combine any architectural changes with loss scaling.\", \"**CNN baselines:** While the proposed method is compared against MTL-CNN models, it is not clear to me whether the experimental setup are equivalent. In particular, the proposed method starts from a ViT backbone pretrained on IamgeNet-22k. In contrast, to the best of my knowledge, MTI-Net models use pretrained ImageNet-1k weights. Even if the authors do not retrain these models from scratch, it would be interesting and fair to report a summary of experimental/pretraining difference between the different baselines to the reader.\", \"## Experiments\", \"**Model efficiency** should be discussed since we are dealing with dynamic architecture. The only efficiency metric is number of parameters (line 514) is model parameters, which seems a bit misleading since the method is also adding tokens. Because of this, it is really hard to assess how practical the method actually is. In particular it would be interesting to discuss:\", \"The training time cost of performing SVD + gradient conflict estimation\", \"The cost of the final model, with the additional tokens and modulation layers\", \"This would be particularly interesting for **Table 4**, as for instance we see that DTE-MTL only brings little improvement to `TaskPrompter`, and it would be interesting to see how this contrast with any potential overhead.\", \"Impact of **hyperparameters**. As noted in \\u201c*In Defense of the Unitary Scalarization for Deep Multi-Task Learning*\\u201d and *\\u201cDo Current Multi-Task Optimization Methods in Deep Learning Even Help?\\u201d* (both published in NeurIPS 2022), the choice of learning rate is highly impactful on MTL methods, sometimes even leading to simple scalarization outperforming more advanced MTL optimisation. Yet in the current paper, the results are only reported for a single learning rate (line 409) and there is no mention of a learning rate sweep\"], \"questions\": [\"In line 150: *\\\"Existing methods that use pre-defined architectures for MTL have limitations in reducing negative transfer since they cannot preemptively prevent the occurrence of conflicting gradient\\\"* -> This should be substantiated. The goal of having task-specific parameter (i.e. non-shared) is precisely to have parameters that can not be affected by conflicting gradients.\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The authors propose a token-based network expansion and modulation method to reduce negative transfer in multi-task problems.\\nWhen the task-specific gradients clash, modulating existing tokens is employed; when the task-specific gradients conflict, new task-specific tokens are added. The authors attempt to establish theoretical results to show that their proposed method can reduce training loss. Experiments are conducted on NYUD-v2, PASCAL, and taxonomy to evaluate their method.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": [\"the paper analyzes the range space and null space of tokens to establish their method for reducing the negative transfer.\", \"Analyzing the conflicts between tasks by examining the token\\u2019s range space and null space is interesting.\"], \"weaknesses\": \"- many important baselines are missing, e.g., InvPT, MQTransformer, TaskPrompt, ForkMerge, MTMamba (https://arxiv.org/pdf/2407.02228)\\n- The results reported in table 1 are much lower than current sota methods (refer to the results reported in Table 1 of MTMamba, https://arxiv.org/pdf/2407.02228), I copied some results here:\\n\\n(Semseg and Boundary: larger is better; Depth and Normal: smaller is better)\\nmethod Semseg Depth Normal Boundary \\n\\nInvPT 53.56 0.5183 19.04 78.10\\n\\nMQTransformer 54.84 0.5325 19.67 78.20\\n\\nMTMamba 55.82 0.5066 18.63 78.70\\n\\nTM+TE 38.27 0.6370 21.64 57.90\\n\\nwe can see from the above results, the TM+TE method proposed in this paper is much weaker than existing sota methods.\\n\\n- the method is limited to transformer-based networks, cannot be extended to other architectures like CNN and mamba-based\\n- the proofs of theorem 1 and theorem 2 are problematic. For a mathematical proof, first, we need to make assumptions in the theorem, instead of making the assumption in the proof, e.g., L819, in cases; eq (5) $\\\\approx$, approximation in eq (6).\\nmoreover, the derivative in L837 is not rigorous.\", \"questions\": [\"does the proposed algorithm converge?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper proposes a novel approach called Dynamic Token Modulation and Expansion (DTME-MTL) for multi-task learning with transformer architectures. The key idea is to dynamically expand the network by manipulating tokens to mitigate negative transfer between tasks. The authors provide a theoretical analysis of gradient conflicts in the token space and propose two techniques - token modulation and token expansion - to address conflicts in the range space and null space respectively.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": [\"The paper provides a solid theoretical analysis of the effectiveness of the proposed approach. The authors carefully define the token space using SVD and provide mathematical justification for categorizing gradient conflicts into range space and null space conflicts. This theoretical grounding helps explain why the proposed token modulation and expansion techniques should work.\", \"The method is general and can be applied to existing transformer-based multi-task architectures in an off-the-shelf manner. This increases the potential impact and applicability of the work.\", \"The proposed techniques of token modulation and expansion are novel and intuitively motivated. The idea of dynamically expanding the network via tokens rather than entire layers is interesting.\"], \"weaknesses\": [\"While the authors claim their method is parameter-efficient, there is a lack of concrete efficiency comparisons with baselines. The paper would be strengthened by including quantitative comparisons of parameter counts and computational costs against other multi-task optimization approaches.\", \"The paper mentions that the number of layers for expansion is an important hyperparameter (Section 4.3), but there is no discussion or ablation study on how this hyperparameter affects performance. More analysis on the sensitivity to this choice would improve the paper.\", \"The experimental evaluation, while showing improvements, is somewhat limited. More extensive comparisons on additional datasets and task combinations would help establish the generality of the approach.\"], \"questions\": [\"How sensitive is the method to the choice of the r parameter used to divide the range and null spaces? Is there a principled way to set this?\", \"Have the authors explored applying their method to other architectures beyond transformers? Could the core ideas be extended to CNNs for example?\", \"How does the computational overhead of computing SVD and projecting gradients at each layer impact training time compared to baselines?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper proposes a novel multi-task expansion algorithm for transformer architectures. The intuition is to dynamically expand the token representations of the transformer based on gradient conflict. Further, gradient conflict is separated into two distinct regions using singular-value decomposition of the shared representations: conflict which occurs in the null-space of the tokens and conflict which occurs in the range-space (i.e. where singular-values are non-zero).\\n\\nThe authors propose two different expansion techniques to address each type of conflict distinctly. When conflict occurs in the range-space of the token representations, the authors propose to add task-specific, learnable affine transformations to the token representations to prevent gradient conflict from slowing learning in the key range-space of the representation. Alternatively, when conflict occurs in the null-space of the token representations, the authors propose to add task-specific components to the token (i.e. instead of linearly transforming the representation, a new dimension is added to resolve conflict).\\n\\nThe authors demonstrate the efficacy of their method across 3 challenging multi-task settings, demonstrating it\\u2019s superiority compared to other multi-task network expansion algorithms, as well as multi-task optimization methods. They also present theoretical analysis of their method, demonstrating it\\u2019s ability to lower the multi-task loss, and empirical analysis of various components of their method.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"The intuition to separate treatment of gradient conflict based on whether the conflict occurs in the subspace spanned by top eigenvectors vs. the subspace spanned by eigenvectors that are near zero is a neat idea, which (to my knowledge) has not explicitly been done before and is something that is worth considering in future work.\\n\\nAdditionally, the concept of focusing on the token representations in a transformer is novel, but certainly very parameter-efficient while also being effective, although it is worth noting that the modulation mechanism bears some similarity to RotoGrad [2], which applies learned task-specific rotations (linear transformation) to task representations.\\n\\nThe method appears to get strong results compared to an extremely extensive list of methods, including a number of multi-task optimizers and multi-task network algorithms. And the method is compared over 3 fairly common and difficult multi-task settings, so the results here are promising.\\n\\nFinally, the empirical results contain a lot of analysis of various components of the proposed method which is useful to study, although they do not often have a clear takeaway (for example, the timing of network expansion and analysis of gradient conflict do not show particularly clear trends).\\n\\n[2] Javaloy & Valera, 2022; ROTOGRAD: GRADIENT HOMOGENIZATION IN MULTITASK LEARNING\", \"weaknesses\": \"To me the greatest weakness of the work is a lack of analysis on ablations, particularly around the two separate mechanisms for dealing with gradient conflict. In particular, both mechanisms (modulation and expansion) add new, task-specific learnable parameters to the token representations where gradient conflict is highest. However, while the intuitions provided make sense, I do not see a principled reason why modulation could not be used when gradient conflict exists in the null-space, or why token expansion could not be used when conflict exists in the range-space. I\\u2019m not sure that the analysis in the theorems / appendix makes this clear either, e.g. if we assumed in the proof of theorem 1 that the token was spanned by the null-space, wouldn\\u2019t the modulation still lead to lower multi-task loss due to the task-specific parameters? Perhaps I am missing something here. Also, while there is an ablation study when using only token expansion vs. only modulation, I believe (please correct me if I am wrong) that token expansion is still only applied to null-space conflict and modulation is still only applied to range-space conflict, so this doesn\\u2019t address my concern.\\n\\nThis concern is compounded by the fact that applying modulation and token expansion to random or lowest-conflict layers still seems to improve over the multi-task baseline, suggesting that a key benefit may be the addition of more task-specific space to the model at the token-level, as opposed to the separation of range-space and null-space conflict and the corresponding mechanisms.\\n\\nThe lack of reporting on the number of random seeds or variation across those seeds in performance is also fairly problematic. These results would be much more convincing if we could see how significant the improvement is over a set of random seeds, especially as other works have shown that significance is a key detail to report in multi-task methods [1].\\n\\nFinally, the method seems to rely on a shared input for all tasks in order to have a shared token representation. While this certainly works for a number of multi-task settings, it does exclude a larger number of MTL settings where distinct tasks have distinct inputs, and thus there is no single shared representation for each task input during training. However, I don\\u2019t see this limitation discussed or acknowledged anywhere.\\n\\n[1] Kurin et al., 2022; In Defense of the Unitary Scalarization for Deep Multi-Task Learning\", \"questions\": \"See weaknesses.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}" ] }

[ "Nikolaos Tsilivis", "Gal Vardi", "Julia Kempe" ]

[ "implicit bias", "steepest descent", "deep neural networks" ]

{ "note_id": [ "wxNZl8cSRC", "uaoLg6QYXU", "u0NMsOLOEn", "rb79GZ1ZgX", "rZEtp6I7si", "nrTnt3Daos", "nbn1nQrDab", "j7Xrw1wMsj", "j2YRvLAmyb", "iqkvyZMpIc", "imzhu5vOUQ", "iWzTTLzsHl", "iN6Bd8af7R", "atQQ8AisHQ", "Vn35lnOgzs", "R67GnGBkBp", "QAVgBIzbaF", "MP9ZIR35Mj", "L2C8cx934p", "KdumC4PkQT", "IsQuBe0kDp", "8DmqH44Bv9", "7jbVeNgiD4", "093tZSuCRv" ], "note_type": [ "official_comment", "official_review", "meta_review", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "decision", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment" ], "note_created": [ 1732658687739, 1730808053894, 1734765367322, 1732082213225, 1732923816264, 1732496399630, 1731331098400, 1730672961791, 1732546293247, 1730504877072, 1732081320614, 1732082375929, 1732082291361, 1732552043682, 1732081568141, 1732082399011, 1732680622390, 1737523600019, 1732425612070, 1732081537454, 1730773632989, 1732476186582, 1732753267909, 1732082134461 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission3801/Authors" ], [ "ICLR.cc/2025/Conference/Submission3801/Reviewer_APGr" ], [ "ICLR.cc/2025/Conference/Submission3801/Area_Chair_1acp" ], [ "ICLR.cc/2025/Conference/Submission3801/Authors" ], [ "ICLR.cc/2025/Conference/Submission3801/Authors" ], [ "ICLR.cc/2025/Conference/Submission3801/Reviewer_Bpd2" ], [ "ICLR.cc/2025/Conference/Submission3801/Reviewer_Bpd2" ], [ "ICLR.cc/2025/Conference/Submission3801/Reviewer_2hLy" ], [ "ICLR.cc/2025/Conference/Submission3801/Authors" ], [ "ICLR.cc/2025/Conference/Submission3801/Reviewer_pCcn" ], [ "ICLR.cc/2025/Conference/Submission3801/Authors" ], [ "ICLR.cc/2025/Conference/Submission3801/Authors" ], [ "ICLR.cc/2025/Conference/Submission3801/Authors" ], [ "ICLR.cc/2025/Conference/Submission3801/Reviewer_2hLy" ], [ "ICLR.cc/2025/Conference/Submission3801/Authors" ], [ "ICLR.cc/2025/Conference/Submission3801/Authors" ], [ "ICLR.cc/2025/Conference/Submission3801/Reviewer_pCcn" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission3801/Reviewer_APGr" ], [ "ICLR.cc/2025/Conference/Submission3801/Authors" ], [ "ICLR.cc/2025/Conference/Submission3801/Reviewer_By2M" ], [ "ICLR.cc/2025/Conference/Submission3801/Authors" ], [ "ICLR.cc/2025/Conference/Submission3801/Authors" ], [ "ICLR.cc/2025/Conference/Submission3801/Authors" ] ], "structured_content_str": [ "{\"comment\": \"Thank you for your time and engagement during the rebuttal period!\"}", "{\"summary\": \"This paper studies the implicit bias of steepest descent methods with infinitesimal step size (i.e., steepest flow) and presents an interesting generalization of Lyu & Li (2020)'s result on gradient flow (steepest flow with L2 norm). Similar to gradient flow, it is shown that steepest flow has a bias towards KKT solutions to a margin maximization problem, but the margin being maximized here is normalized properly according to the norm used in steepest flow, which may not be the L2 norm in general.\", \"soundness\": \"4\", \"presentation\": \"4\", \"contribution\": \"3\", \"strengths\": \"1. Understanding the implicit bias of optimization methods is an important problem in deep learning theory, and the paper takes a margin-based view of implicit bias and sheds light on how different methods may optimize different notions of margin for deep models.\\n2. It is a bit surprising that such an implicit bias of steepest descent can be rigorously proved in the general case. Previously, it is known that this can be proved for linear models, and for deep homogeneous networks, Lyu & Li (2020) generalized the result for GD. It is reasonable to imagine that one could somehow generalize Lyu & Li (2020)'s result to steepest gradient, but their proof apparently has a lot of dependence on the L2 geometry induced by GD. After having a careful read of the proof (though I skipped a few details), I believe the authors found the right way to do this generalization and worked out all the technical issues, including those related to non-smoothness.\\n3. The paper is well-written and easy to follow.\\n4. Experiments in simple settings validated the theoretical results.\", \"weaknesses\": \"1. While I really appreciate that the authors worked out every technical detail to generalize the proof of Lyu & Li (2020), I also noted that the overall proof outline has not changed much from Lyu & Li (2020), which means this paper does not actually bring a brand new high-level proof idea. This is reasonable since any implicit bias analysis of steepest flow automatically implies an analysis of gradient flow, but the fact that the paper fails to prove the real KKT conditions and only manage to prove a weaker notion (generalized KKT) suggests that a perfect generalization from gradient flow to steepest flow may indeed need a completely new proof strategy.\\n2. Same as all previous works studying the max-margin bias, this paper only analyzes the late phase, where the training loss is already very small. While the asymptotic analysis of implicit bias is interesting, this also makes the phenomenon less relevant to the practice.\", \"minor_issue\": \"there are large blanks on some pages in the appendix.\", \"questions\": \"This paper only proves the convergence to generalized KKT solutions for sign gradient flow, which is an important case of steepest descent. I wonder if the authors have any thoughts on whether failing to prove KKT in this case is an artifact due to the proof techniques or if some tricky hard cases actually exist.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"metareview\": \"This paper studies the implicit bias of steepest flow (continuous-time version of steepest descent) algorithm in training homogeneous neural networks. The paper extends an existing work Lyu & Li on gradient descent to a general class of algorithms, showing monotonic increase of the soft margin and convergence to the direction of generalized KKT points of the margin maximization problem.\\n\\nAll reviewers were positive about this submission. The paper presents an extension of a seminal result to a much more general class of algorithms, also drawing connections to practical algorithms such as Adam and Shampoo. Although the outline of proof seems to be the same as the existing paper, the extension from $\\\\ell_2$ geometry to possibly nonsmooth cases looks nontrivial.\\n\\nIt is, though, unfortunate that the current proof technique cannot guarantee convergence to the usual KKT points (not the generalized KKT points introduced in the paper) in some important cases such as $\\\\ell_1$ and $\\\\ell_\\\\infty$ norms.\\n\\nAlso, as Reviewer pCcn points out, the paper only characterizes the implicit bias in the continuous-time (flow) case and leaves the extension to the discrete-time (descent) case as future work. In fact, I see that there is no indicator of this fact in the title and abstract; I recommend revising at least the abstract to reflect this.\\n\\nOverall, the paper offers solid contributions to the study of implicit bias of neural network training. I recommend acceptance.\", \"additional_comments_on_reviewer_discussion\": [\"Other than the points mentioned above, noteworthy points about the discussion include:\", \"Some reviewers pointed out missing references and suggested improving the related works section; the authors made revisions accordingly.\", \"Reviewer pCcn asked if tight convergence rates for both the loss function and parameter norms can be established, as in Lyu and Li, and the authors responded that such rates can be deduced from their analysis. To me, it seems beneficial to state these convergence rate results in a more explicit form (in a proposition, etc).\"]}", "{\"title\": \"Authors' response\", \"comment\": \"We appreciate the time you took to review our paper and are grateful you found our results novel! We reply to your only question regarding the technical contributions of the paper:\\n\\n> The set of results and proofs seem to be a straightforward generalization of [Lyu and Li, 2019]. The technical contribution is not strong.\\n\\nWhile this is of course a relative matter, please let us elaborate on two technical points where our work departs significantly from [Lyu & Li, 2019]:\\n- Non-smoothness of the algorithm norm: the fact that the squared norm of the algorithm is not necessarily continuously differentiable (as it is in the case of gradient descent) complicates the analysis in many parts - see, for example, Proposition A.3 and the rest of the analysis where specific care must be taken to handle updates aligned with the least norm subgradient $\\\\mathbf{g}_t^\\\\star$.\\n- Quantifying proximity to stationarity: Gradient descent approaches the KKT points of the $\\\\ell_2$ max margin problem by reducing a Euclidean distance measuring stationarity. Prior to our work, it was arguably unclear how a different steepest descent method might have progressed to approach stationarity of the corresponding implicit max margin problem. We introduced a novel measure of stationarity, based on an algorithm-specific (pseudo) Bregman divergence, and we show that this succeeds in quantifying the progress of the algorithms. This also naturally motivates a new progress measure for approximate stationarity (Section B), which could be of independent interest. Notice that here as well the non-smoothness of the norm causes technical challenges, and their handling required some fairly involved machinery from Convex Analysis (Proposition A.13 & Theorem 5.24 in [Beck, 2017]).\\n\\nIn general, the proof of [Lyu & Li, 2019] is heavily based on the $\\\\ell_2$ geometry of gradient descent (naturally so!) and many parts have to be modified in order to obtain a clean result for any steepest flow. Testament to this are the many novel tools (generalized KKT points -- Definition 3.4, approximate generalized KKT points -- Definition A.9, generalized Bregman divergences -- Definition 3.6) and corresponding results (e.g. Proposition A.20) which we had to introduce in order to obtain our implicit bias characterization. \\n\\nWe would be happy if our explanations could make you reconsider the technical contribution of this work, and perhaps, make you raise your score. Please let us know if you have any specific questions. Thank you!\"}", "{\"title\": \"Kind reminder\", \"comment\": \"Hi,\\n\\nWe would like to kindly encourage you to respond to our comments, as the discussion period ends soon.\\n\\nThank you!\"}", "{\"comment\": \"Thanks for the response. The response and the revision have improved the quality of the paper slightly. I am leaning towards accepting the paper and my current rating is somewhere around 6.5. While I might consider rounding it up to a 7 if that were an option, it may not reach the caliber of an 8 in my opinion (compared to other ICLR rating-8 papers I\\u2019ve reviewed and read, in terms of conceptual and technical novelty). Therefore, I will maintain my current rating.\"}", "{\"summary\": \"The paper investigates the implicit bias of the family of steepest descent algorithms, including gradient descent, sign gradient descent, and coordinate descent, for deep homogeneous neural networks. The authors aim to understand how these optimization algorithms implicitly select solutions from the perspective of margin maximization. The main contributions of the paper are:\\n\\n1. A rigorous analysis of the implicit bias of steepest descent algorithms in non-linear, homogeneous neural networks, showing that an algorithm-dependent geometric margin increases during training, under the same set of assumptions made in Lyu&Li.\\n2. The definition and characterization of a generalized notion of stationarity for optimization problems, demonstrating that these algorithms reduce a generalized Bregman divergence, which quantifies proximity to stationary points of a margin-maximization problem.\\n3. Experimental validation of the theoretical findings by training neural networks with various steepest descent algorithms, highlighting connections to the implicit bias of Adam.\\n\\n\\nI think the class of the optimization problems studied in this paper is general and the theoretical results of the paper are quite clean. Although the main proofs follow the framework developed by Lyu and Li, the proofs are nontrivial (I read most of the proofs in the main text and a few in the appendix). Overall, I am leaning towards accepting the paper. With additional clarifications of the relation with prior work, this paper would be a valuable addition to the literature on deep learning theory.\", \"soundness\": \"4\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"**Strong Points:**\\n\\n1. The paper provides rigorous theoretical analysis of the implicit bias of steepest descent algorithms. The results and proofs extend the l2 norm case (gradient descent) by Lyu&Li to general norms. The theoretical results are clean and some of the extension are nontrivial. \\n\\n2. The class of optimization algorithms studied in this paper is very general and include several important optimization algorithms as special cases.\\n\\n3. The paper connects theoretical insights to practical optimization methods like Adam, which is widely used in the deep learning community. This connection may help to understand the benefit of Adam.\", \"weaknesses\": \"**Weak Points:**\\nThe authors briefly discuss some prior work on implicit bias of optimization algorithms other than GD in the related work section. But I found the inclusion of related work and the discussion here is somewhat insufficient. For example, Gunasckar et al. 18 already studied the implicit bias of several class of optimization algorithms. However, the citation here is very superficial and it is unclear to me how the results in this paper supercede (or compared with) their results. It is also unclear to me how the results by Wang et al. on Adam is related to the results of this paper. If we simplify Adam to signGD, it seems that the results in Wang et al. is inconsistent with this paper? I think more detailed discussion is necessary.\\n\\nMoreover, the following papers also obtain implicit bias results towards L_p margin with p different from 2 (for special models). Is there any concrete relation between these results and the results in this paper. \\nKernel and Rich Regimes in Overparametrized Models (already cited but not discussed in details)\", \"implicit_bias_in_deep_linear_classification\": \"Initialization Scale vs Training Accuracy\", \"questions\": \"minor comment:\\n1. The sentence before Theorem 3.1. The proof is quite similar to that in Lyu&Li. It seems to me that this paper used Cauchy-shwartz but Lyu&Li explicitly used the Pythagorean theorem (which only holds in inner prod space but not general normed space). If I am correct, then there is not much difference.\\n\\n2. I think it would be better to discuss briefly the relation between D-generalized KKT and ordinary KKT between Theorem 3.8 and Cor 3.8.1. and why not to prove something like Cor 3.8.1. directly.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This work studies how the prediction margin changes as the training proceeds in a homogeneous network\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"The problem of prediction margin is important and of fundamental importance\", \"weaknesses\": \"I am not sure if I understand the meaning of the result. Also, I find the setting of the theory too restrictive\\n\\n1. I do not understand Figure 1-right. What does it imply? How does this relate to the theory?\\n2. The theory only applies to the \\\"exponential loss,\\\" which does not make much sense to me. \\n\\nLastly, I feel I should comment that I do not have the related background knowledge to assess the technical advancement made by this work\", \"questions\": \"How close is the steepest descent to actual GD? Does the theory hold for GD?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you for engaging during the rebuttal period! We are glad about your positive evaluation of our work.\"}", "{\"summary\": \"This paper studies the implicit bias of the steepest flow with respect to a general norm in deep homogeneous neural networks. Theoretically, they present two major results. 1) The soft margin will increase 2) The limit points of the parameter directions are general KKT points of the margin maximization optimization problem under some assumption of the norm. They also show some experimental results concerning the trajectories of different steepest descent algorithms and the connection to the implicit bias of Adam.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"This paper extends the result in [Lyu & Li 2020] to steepest flow for general norms with the lightest extra assumption on the norm. While previous papers focused on gradient descent and other variants of gradient descent, this paper gives a general result for any steepest flow.\\nThe claimed results seem sound and the paper is nicely written.\", \"weaknesses\": \"The paper asserts that it presents results for the steepest descent; however, the theoretical results in Section 3 are limited to steepest flow. According to Lyu & Li (2020), additional work is required to extend the findings from flow to descent.\\n\\nRegarding the general loss function, the paper demonstrates that the soft margin is monotonically increasing but fails to address the convergence of KKT points.\\n\\nMoreover, the paper overlooks several relevant references, including:\\n\\n1. Chatterji, N.S., Long, P.M., & Bartlett, P. (2021). \\\"When does gradient descent with logistic loss interpolate using deep networks with smoothed ReLU activations?\\\"\\n\\n2. Cai, Y., Wu, J., Mei, S., Lindsey, M., & Bartlett, P.L. (2024). \\\"Large Stepsize Gradient Descent for Non-Homogeneous Two-Layer Networks: Margin Improvement and Fast Optimization.\\\"\\n\\n3. Nacson, M.S., Srebro, N., & Soudry, D. (2019, April). \\\"Stochastic gradient descent on separable data: Exact convergence with a fixed learning rate.\\\"\\n\\n4. Ji, Z., Srebro, N., & Telgarsky, M. (2021, July). \\\"Fast margin maximization via dual acceleration.\\\"\", \"questions\": [\"Can the authors extend the existing results on steepest flow to steepest descent? If not, it would be more accurate to clarify that the current theoretical results apply solely to steepest flow.\", \"Is it feasible to demonstrate KKT point convergence for the general loss function?\", \"In Lyu & Li (2020), tight convergence rates for both the loss function and parameter norms were established. Are there any mathematical challenges in achieving similar results here? Additionally, they showed that the soft margin approximates the normalized margin. Could you provide a similar analysis in this context?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"General Response\", \"comment\": \"We would like to thank all reviewers for their time and efforts in helping us improve our paper. We uploaded a revised version of the paper, which incorporates some of your suggestions (colored in red). In addition, we want to highlight two points:\\n\\n1. We updated the Related work section to make the discussion of prior work more elaborate. Please consider taking a look.\\n2. We added some discussion on the connection between an adaptive optimization method, Shampoo [1], and steepest descent which was recently brought to our attention and discuss how it is related to the results of this paper (lines 130-132 and Appendix C.2).\\n\\nThank you!\\n\\n[1] Vineet Gupta, Tomer Koren, and Yoram Singer. Shampoo: Preconditioned Stochastic Tensor Optimization.\"}", "{\"title\": \"Authors' response (1/2)\", \"comment\": \"Thank you very much for the critical review of our work and your help in improving it!\\n\\nFirst, before responding to your questions, please allow us to correct a small mistake in your evaluation of our work, in order to avoid confusions:\\n\\n> The limit points of the parameter directions are general KKT points of the margin maximization optimization problem under some assumption of the norm\\n\\nAssumming that by \\\"some assumption of the norm\\\" you mean the algorithm norm squared being smooth, then: Theorem 3.8, which is the main result of this paper, asserts that steepest flow converges in direction to a generalized KKT point *without* any assumption on the algorithm norm. Corollary 3.8.1, on the other hand, strengthens this result in the case of a norm that satisfies the aforementioned property, stating that we get convergence to a *(standard) KKT point*, not a generalized one. \\n\\n> Can the authors extend the existing results on steepest flow to steepest descent? If not, it would be more accurate to clarify that the current theoretical results apply solely to steepest flow.\\n\\nIndeed, our paper focuses on steepest descent with infinitesimal learning rate, i.e., steepest flow (Eq. 3). We specify this at the start of our introduction (line 051). We leave the generalization of our results to the case of steepest descent (with a non infinitesimal learning rate) as an open challenge. Notice that the possibility of a non-smooth norm substantially complicates a discrete time analysis - for instance, Theorem E.2 and Lemmata E.7, E.8, E.9 in [Lyu & Li, 2020] only hold in the case of an $\\\\ell_2$ geometry. Furthermore, the proof will have to operate under more restrictive assumptions on the smoothness of the network, which, arguably, minimizes the practical relevance of such a result. We recognize this as a limitation of our analysis and added a short note on this in our conclusion (Section 5, line 478).\\n\\n> Is it feasible to demonstrate KKT point convergence for the general loss function?\\n\\nWe believe that this should be possible to prove for any exponentially-tailed loss. In particular, in Appendix A.3, we demonstrate how Theorem 3.1 generalizes to other loss functions, and we note that a more technical definition than Definition A.21 could allow the extension of our full proof to other loss functions, similar to the approach taken in [Lyu & Li, 2020]. As a matter of fact, in the Experiments section (Section 4), we consider training with both the exponential and logistic loss with various steepest descent algorithms, and the margin observations are consistent with our theory. Finally, let us remark that it is a common practice in the literature to focus solely on the exponential loss (see, for instance, [1, 2, 3]) as it is known to capture the essence of these results.\\n\\n1. Implicit Bias in Deep Linear Classification: Initialization Scale vs Training Accuracy. Edward Moroshko, Suriya Gunasekar, Blake Woodworth, Jason D. Lee, Nathan Srebro, Daniel Soudry.\\n2. Implicit Bias of Gradient Descent on Linear Convolutional Networks. Suriya Gunasekar, Jason Lee, Daniel Soudry, Nathan Srebro.\\n3. Lexicographic and Depth-Sensitive Margins in Homogeneous and Non-Homogeneous Deep Models. Mor Shpigel Nacson, Suriya Gunasekar, Jason D. Lee, Nathan Srebro, Daniel Soudry.\\n\\n> In Lyu & Li (2020), tight convergence rates for both the loss function and parameter norms were established. Are there any mathematical challenges in achieving similar results here? \\n\\nNo, we believe there are no mathemetical challenges in achieving similar complementary results. Essentially, Lemma A.6 in our paper characterizes the rate of convergence for the loss and the paramater norm (particularly for the relevant algorithm norm). We opted not to risk presenting such a complimentary result during the rebuttal period (after the first round of all the reviews), since it is not exactly straightforward, but we believe that an interested reader can deduce these rates using our Lemma A.6 and equations 30 and 31.\\n\\n> Additionally, they showed that the soft margin approximates the normalized margin. Could you provide a similar analysis in this context?\\n\\nIf we understand your question correctly, then, yes, this is (essentially) proved in Lemmata A.6 and A.7. We fixed a minor typo there (line 822) and added the conclusion as a corollary (Corollary A.7.1). Furthermore, we added a relevant pointer for this in the main text (lines 178-179). Thank you very much for the suggestion!\"}", "{\"title\": \"Authors' response\", \"comment\": \"Thank you for reading our paper and helping us improve it!\", \"we_reply_to_your_questions\": \"> I do not understand Figure 1-right. What does it imply? How does this relate to the theory?\\n\\nThank you for the question! In Figure 1 (right), we compare 3 different steepest descent algorithms: gradient descent, sign gradient descent and coordinate descent. We measure the $\\\\ell_\\\\infty$ margin of the training dataset at the end of training (Eq. 14) and plot it against the test accuracy of the networks (each point -- circle, square, triangle -- corresponds to a different seed). The first observation, which connects to Theorem 3.1, is that coordinate descent has a larger $\\\\ell_\\\\infty$ margin than the rest of the algorithms -- see lines 409-412. This is expected since it is the only algorithm out of the 3 which is guaranteed to increase this margin during training. The second observation is about the connection between margin and generalization for this task and is less related to our theory. In short, we comment on the fact that these seems to be no strong, causal link between larger margin and generalization in this setting - see lines 412-425.\\n\\n> The theory only applies to the \\\"exponential loss,\\\" which does not make much sense to me.\\n\\nThis is a valid concern, but our choice of this loss follows a long list of precedence and justification in the literature. Indeed, the reason why many papers are concerned with the exponential or exponentially-tailed losses is that: (a) they are close to the most common choice of the cross-entropy loss, or its binary version, the logistic loss, and (b) they allow for a clean relationship between the geometric margin and the loss via the softmax. Also, let us mention that this has been a standard practice in the literature, with numerous papers (e.g. [1, 2, 3]) that **only** study training under the exponential loss. In defense of the contributions of our paper, let us note that our work considers the generalization to other losses in Section A.3 (where Theorem 3.1 is generalized).\\n\\n> How close is the steepest descent to actual GD? Does the theory hold for GD?\\n\\nA steepest descent algorithm is defined with respect to a norm. When this norm is the $\\\\ell_2$, we obtain gradient descent. Otherwise, we obtain a different algorithm. The theory holds for gradient descent with infinitesimal learnging rate, but this is not the most interesting case since the result of [Lyu & Li, 2019] already covers this case.\\n\\nWe hope we were able to answer your questions convicingly. Please let us know if there are any other concerns or if you require further explanations. \\n\\nIn general, given your comment:\\n> Lastly, I feel I should comment that I do not have the related background knowledge to assess the technical advancement made by this work\\n\\nWe would be happy to help you assess the technical advancements of this paper by explaining things further if you feel this is appropriate, in order to provide a basis that would allow you to raise your score and recommend acceptance. Thank you for your time and your help!\\n\\n\\n1. Implicit Bias in Deep Linear Classification: Initialization Scale vs Training Accuracy. Edward Moroshko, Suriya Gunasekar, Blake Woodworth, Jason D. Lee, Nathan Srebro, Daniel Soudry.\\n2. Implicit Bias of Gradient Descent on Linear Convolutional Networks. Suriya Gunasekar, Jason Lee, Daniel Soudry, Nathan Srebro.\\n3. Lexicographic and Depth-Sensitive Margins in Homogeneous and Non-Homogeneous Deep Models. Mor Shpigel Nacson, Suriya Gunasekar, Jason D. Lee, Nathan Srebro, Daniel Soudry.\"}", "{\"title\": \"reply and thoughts\", \"comment\": \"Thanks for the rebuttal. I feel more positive towards the paper and have raise my rating to 6. I thought steepest descent is more restrictive than GD, but now it looks like it contains GD as a subclass, which is nice\\n\\nHowever, I point out that my rating is at best an educated guess.\"}", "{\"title\": \"Author's Response (2/2)\", \"comment\": \"> The sentence before Theorem 3.1. The proof is quite similar to that in Lyu&Li. It seems to me that this paper used Cauchy-shwartz but Lyu&Li explicitly used the Pythagorean theorem (which only holds in inner prod space but not general normed space). If I am correct, then there is not much difference.\\n\\nYes, this is \\\"conceptually\\\" correct (Theorem 3.1 does not use the CS inequality, but rather the definition of the dual norm, while Lemma 5.1 in [Lyu & Li, 2019] does use a polar decomposition and the Pythagorean theorem). This is why we mentioned that Theorem 3.1 \\\"is similar to part of Lemma 5.1 in [Lyu & Li, 2019]\\\" in line 181.\\n\\n> I think it would be better to discuss briefly the relation between D-generalized KKT and ordinary KKT between Theorem 3.8 and Cor 3.8.1. and why not to prove something like Cor 3.8.1. directly.\\n\\nThank you for this suggestion! We added a short discussion in lines 315-317.\\n\\nPlease also consider taking a look at the general response to all reviewers. In particular, in addition to the improvements in presentation, the updated manuscript contains a small section on the connection between a newly introduced adaptive optimization algorithm (Shampoo) and our framework.\\nWe are grateful for your useful questions and suggestions, which helped us improve our paper. Please let us know if any of our answers is unclear or if you have further questions. If not, we kindly ask you to consider raising your score, which would help achieve a consensus among reviewers. Thank you!\"}", "{\"title\": \"Authors' response (2/2)\", \"comment\": \"> Moreover, the paper overlooks several relevant references, including:\\nChatterji, N.S., Long, P.M., & Bartlett, P. (2021). \\\"When does gradient descent with logistic loss interpolate using deep networks with smoothed ReLU activations?\\\"\\nCai, Y., Wu, J., Mei, S., Lindsey, M., & Bartlett, P.L. (2024). \\\"Large Stepsize Gradient Descent for Non-Homogeneous Two-Layer Networks: Margin Improvement and Fast Optimization.\\\"\\nNacson, M.S., Srebro, N., & Soudry, D. (2019, April). \\\"Stochastic gradient descent on separable data: Exact convergence with a fixed learning rate.\\\"\\nJi, Z., Srebro, N., & Telgarsky, M. (2021, July). \\\"Fast margin maximization via dual acceleration.\\\"\\n\\nThank you for the references. We added a discussion of non-homogeneous networks in lines 094-096 and the revised version now cites paper #2. However, we believe that papers #1, #3 and #4 are sligthly peripheral to our paper. Please note that we already included a reference to a comprehensive survey on implicit bias [Vardi, 2023], where some of these papers are discussed in depth. If you disagree or have suggestions on where it would be helpful to discuss these works in our paper, we would be happy to hear them.\\n\\nWe would like to thank you once again for your constructive critism. Please let us know if any of our answers is unclear or if you have further questions. We hope that our explanations, together with the manuscript changes made in response to your and the other reviewers' comments, could perhaps encourage you to raise your evaluation score of our work. Thank you!\"}", "{\"comment\": \"Thank you for your response. I would like to keep my score.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}", "{\"comment\": \"Thanks so much for your thorough response. I would like to keep my high score.\"}", "{\"title\": \"Author's response (1/2)\", \"comment\": \"Thank you very much for taking the time to review our paper and for helping us improve it! We reply to your comments:\\n\\n> Gunasckar et al. 18 already studied the implicit bias of several class of optimization algorithms. However, the citation here is very superficial and it is unclear to me how the results in this paper supercede (or compared with) their results.\\n\\nGunasekar et al. (2018) studied the implicit bias of all steepest descent algorithms in *linear* models under the exponential loss. Our paper, in contrast, analyzes steepest descent in *homogeneous*, potentially non-linear, neural networks. So, our setting is more general than Gunasekar's et al. (2018). The strongest results we obtain in this paper characterize the limiting points of the trajectory as KKT points of the margin maximization problem (in the case of a smooth squared norm), whereas Gunasekar et al. (2018) prove that in linear models, we obtain global *maximizers* of the equivalent margin maximization problem. The gap between the two results is, however, unavoidable, as it is known that even for gradient descent, homogeneous networks can converge to points which are KKT points, but not necessarily (even) locally optimal with respect to the margin [Vardi et al. (2022)]. We view our results as conceptually consistent with those of Gunasekar et al. (2018), and we have added one sentence elaborating on this in line 087.\\n\\n> It is also unclear to me how the results by Wang et al. on Adam is related to the results of this paper. If we simplify Adam to signGD, it seems that the results in Wang et al. is inconsistent with this paper?\\n\\nLet us try clarify the confusion. Wang et al. (2021, 2022) analyzes the version of Adam where a precision paramater $\\\\epsilon$ is added in the denominator to avoid numerical instability (see eq. 71 in our paper). In order to obtain signGD, we need to set $\\\\epsilon$ to 0, so the results of Wang et al. (2021, 2022) **do not apply**. This issue was recently observed by Zhang et al. (2024), who analyze Adam without the precision parameter ($\\\\epsilon=0$), but with momentum, specifically in **linear** models. As we mention in line 101, they \\\"found bias towards l1 margin maximization - the same as in the case of sign gradient descent\\\". Thus, our results are consistent with [Zhang et al. (2024)]. We elaborate on this topic further in Section 4.2, where we test these results empirically. As a reminder, in Section 4.2, we observe experimentally that models trained with Adam initially exhibit an $\\\\ell_1$ bias, which later changes to an $\\\\ell_2$ bias. We have updated the text in that section, corrected two minor typos that may have caused confusion, and revised Section C.1 in the Appendix, which discusses Adam in more detail.\\n\\n> The authors briefly discuss some prior work [...] I think more detailed discussion is necessary.\\n\\nThank you for the suggestion. We have included additional discussion and references in that section.\\n\\n> Moreover, the following papers also obtain implicit bias results towards L_p margin with p different from 2 (for special models). Is there any concrete relation between these results and the results in this paper.\\n> Kernel and Rich Regimes in Overparametrized Models (already cited but not discussed in details) Implicit Bias in Deep Linear Classification: Initialization Scale vs Training Accuracy\\n\\nThank you for the question. These papers study gradient descent in a simple class of homogeneous networks (referred to as diagonal neural networks), defined as $f(\\\\mathbf{x}; \\\\mathbf{u}) $= $\\\\langle$ $\\\\mathbf{u}$_+^D - $\\\\mathbf{u}$_-^D,$\\\\mathbf{x} \\\\rangle$, $D > 0$. The constant $D$ represents the depth of the network. The focus of these papers is on how the initialization scale and depth $D$ affect the implicit bias of optimization, along with a precise characterization of the transition from the kernel to the rich regime during training. In their setting, it is known that gradient descent has a bias towards minimum $\\\\ell_2$ norm of the **parameters** [Lyu & Li, 2019], but these papers explore how this bias translates into the **predictor** (i.e. function) space (see Section 2 of [Woodworth et al, 2020. *Kernel and Rich Regimes in Overparametrized Models*]). In this predictor space, they prove an implicit bias towards norm minimization for a norm different than the Euclidean one. In contrast, our results focus solely on the parameter space of homogeneous networks trained with steepest descent. We view these works as peripheral to ours, which is why we opted not to discuss them in detail (as doing so would require defining implicit bias in both parameter and predictor spaces, which, in our view, would detract from the focus of the paper).\"}", "{\"summary\": \"This paper studies the implicit bias of steepest flow with general norm in homogeneous neural networks. It shows that the soft-margin is monotonically increasing after the loss is below a threshold. Furthermore, it shows that the steepest flow converges to the generalized KKT point. Finally, experiments demonstrate the similarity of Adam algorithm and Signed-steepest descent algorithm, in terms of implicit bias.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. The convergence of the general steepest descent to the general maximum margin solution is novel.\\n2. The connection between the implicit bias of Signed steepest descent with Adam is interesting.\", \"weaknesses\": \"The set of results and proofs seem to be a straightforward generalization of [Lyu and Li, 2019]. The technical contribution is not strong.\", \"questions\": \"See weakness\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you!\"}", "{\"comment\": \"Thank you!\"}", "{\"title\": \"Authors' response\", \"comment\": \"We are really grateful for your careful and critical review and we are encouraged you found our paper well-written, easy to follow, and its contributions \\\"surprising\\\". In particular, we are glad that your thorough reading recognized that the [Lyu & Li, 2019] proof relies heavily on the $\\\\ell_2$ geometry (which, of course, is natural to some extent), and our analysis is not straightforward.\", \"we_respond_to_your_questions\": \"> Minor issue: there are large blanks on some pages in the appendix.\\n\\nThank you. This is fixed now.\\n\\n> This paper only proves the convergence to generalized KKT solutions for sign gradient flow, which is an important case of steepest descent. I wonder if the authors have any thoughts on whether failing to prove KKT in this case is an artifact due to the proof techniques or if some tricky hard cases actually exist.\\n\\nWe have indeed spent quite some time pondering this question. In short, it is unclear whether convergence to KKT points can be proven in general. We share some thoughts on this topic: \\n\\nFor sign gradient descent ($\\\\|\\\\cdot\\\\| = \\\\|\\\\cdot\\\\|_\\\\infty$), the Bregman (pseudo) divergence is defined with respect to the $\\\\ell_1$ norm squared. Since the $\\\\ell_1$ norm squared is not strictly convex, the divergence between two vectors approaching 0 (as shown in Proposition A.15 and Theorem A.18), does not imply that the two vectors converge to the same point. A fairly standard technique in convex analysis might suggest to consider a different divergence: the Bregman divergence induced by the (normalized) negative entropy, which is strongly convex with respect to the $\\\\ell_1$ norm. As a result, the obtained divergence, which is a normalized version of the KL divergence, becomes a proper divergence. Unfortunately, two issues arise with this approach: (a) the KL divergence requires non-negative values, which remains problematic even after employing the common trick of doubling the dimension, and (b) if we have a sequence of KL-approximate KKT points, it is not clear whether this implies convergence to a KKT point (that is, it is unclear whether Proposition A.20 holds for d=KL). Had this approach been successful, we could apply a dual rationale for coordinate descent, using the divergence induced by the log-sum-exp function.\\n\\nOne potential experimental approach to determine whether our result is an artifact of our proof technique or not would involve leveraging Corollary A.19.1 and, in particular, eq. 60. There, we show that the rate of convergence to stationarity (for smooth squared norms) depends on the smoothness parameter of the norm. Thus, an experiment could involve training a homogeneous neural network using various steepest descent algorithms (e.g., $\\\\ell_p$ norms with $p = {2, 3, 4}$) and observing whether this theoretical lower bound manifests. Specifically, it would be interesting to see if, once the training data are fit, convergence to stationarity takes longer for algorithms with less smooth norms. However, this may be challenging since measuring some of these quantities in practice could be difficult due to subgradients.\\n\\nIn total, while we remain uncertain about the $\\\\ell_1$ and $\\\\ell_\\\\infty$ cases, we believe that if convergence to a KKT point is true for any norm, a proof similar to ours would have been applicable. Although we do not currently have a definitive answer, we hope you find these insights interesting.\\n\\nFinally, we would kindly suggest taking a look at our global response and the revised version of the paper. In particular, we think you might find Section C.2 interesting, where we added some discussion on the relevance of our results to a newly introduced adaptive method. Thank you once again!\"}" ] }

[ "Tom Everitt", "Cristina Garbacea", "Jonathan Richens", "Henry Papadatos" ]

[ "LLMs", "agents", "goal-directedness", "safety" ]

{ "note_id": [ "ziep2366ts", "z1DbFxf8GR", "y3EKzKZU73", "l1ZIZY53d5", "jQPvgd0VRd", "eCcPdjHkKw", "axyNAZOqQr", "Z5YypvscaE", "Sjgb5li8u3", "RpeweXA7Q8", "Q1eOprWSwt", "Mf8n0QbfoB", "Jud5Hqoo7f", "GmWj71Hhmj", "CF3qoai0CL", "9LX20tYfgt", "7VNEf5Ojeg", "7RMgGg0tew" ], "note_type": [ "meta_review", "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_comment", "official_comment" ], "note_created": [ 1734625748981, 1737524074438, 1732035998101, 1732394221108, 1732429577474, 1732035237946, 1730296313825, 1732036158716, 1730697381911, 1732036401624, 1732704734794, 1732036257508, 1732612750792, 1732598960862, 1730685208869, 1729193796864, 1732036414922, 1732036024730 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission10743/Area_Chair_Ag1r" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission10743/Authors" ], [ "ICLR.cc/2025/Conference/Submission10743/Reviewer_SAsD" ], [ "ICLR.cc/2025/Conference/Submission10743/Reviewer_UoUh" ], [ "ICLR.cc/2025/Conference/Submission10743/Authors" ], [ "ICLR.cc/2025/Conference/Submission10743/Reviewer_BDx1" ], [ "ICLR.cc/2025/Conference/Submission10743/Authors" ], [ "ICLR.cc/2025/Conference/Submission10743/Reviewer_UoUh" ], [ "ICLR.cc/2025/Conference/Submission10743/Authors" ], [ "ICLR.cc/2025/Conference/Submission10743/Authors" ], [ "ICLR.cc/2025/Conference/Submission10743/Authors" ], [ "ICLR.cc/2025/Conference/Submission10743/Authors" ], [ "ICLR.cc/2025/Conference/Submission10743/Reviewer_yUcf" ], [ "ICLR.cc/2025/Conference/Submission10743/Reviewer_yUcf" ], [ "ICLR.cc/2025/Conference/Submission10743/Reviewer_SAsD" ], [ "ICLR.cc/2025/Conference/Submission10743/Authors" ], [ "ICLR.cc/2025/Conference/Submission10743/Authors" ] ], "structured_content_str": [ "{\"metareview\": \"This work defines the concept of goal-directedness for LLM agents, and develops a framework for evaluating it empirically on tasks involving information gathering, information processing, and execution. Although some reviewers expressed interest in the research question and proposed metrics, considering them novel, multiple reviewers also pointed out issues with experimental rigor, unclear writing, and limited task diversity. AC encourages the authors to refine their work and extend the experiments to a broader range of tasks and settings, believing it could become an appealing contribution.\", \"additional_comments_on_reviewer_discussion\": \"The final score for this paper is 6333, with a reduction from 7 to 6. Two reviewers responded to the author. The results indicate that the author did not fully address the reviewers' concerns, and one of the reviewers believed that the paper required major revisions and a re-evaluation after the rebuttal.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"comment\": \"Thank you for your review, which recognises we found limited goal-directedness in state-of-the-art language models, using a method reviewer yUcf described as \\u201cnovel, while also \\u2026 elegantly simple and powerful in its generality.\\u201d\\n\\n# Rigorous definitions\\n\\n_Does Line 79 mean the difference between difference for the optimal policy?_\\n\\nNo, and apologies the paper was insufficiently clear on this point. The comparison is not intended to involve the optimal policy at all (only optimal use of capabilities). We will change the misleading sentence on line 79 to \\\"The gap between the agent's actual performance and the agent's expected performance if it made full use of its capabilities.\\\"\\n\\nFor example, the Build Highest Tower task requires the agent to build a maximally high tower out of any two blocks. The optimal policy takes measurements until all block heights are known to two decimals with high certainty, then stack the two highest blocks. But agents typically only have the capability (or patience) to measure block heights to one decimal point. Conditioning on this (lack of) capability, an agent\\u2019s expected performance will be worse than optimal (as it won\\u2019t be able to reliably select the highest block).\\n\\nHowever, it often turns out that agents\\u2019 suboptimal performance is not fully explained by their lack of capabilities. Instead, they perform even worse than one would expect from just knowing their capability to measure block heights accurately (and ability to select the blocks with the highest estimates). The gap between this actual performance, and the expected performance given full use of capabilities, is what we interpret as a goal-directedness deficit. Figure 5 in the paper shows this gap for the build highest tower task (and Section 4.3 discusses it).\\n\\n_Formal def of E[regret | optimal use of capabilities]? (Line 174)_\\n\\nFormally in RL terminology, the agent is inhabiting a POMDP, where its capabilities correspond to options. For example, an agent capable of measuring blocks to 1 decimal can be viewed to have a measuring option which brings it to an (information) state where the block height is known to 1 decimal. The capability-options define an MDP over states and belief states, with an action-set consisting of these capability-options.\", \"now_we_can_more_formally_define\": \"E[regret | optimal use of capabilities] = optimal return in the original task - the return of the optimal policy in the capability-options belief-MDP.\\n\\n# Algorithm 1 motivation\\n\\nThe idea behind Algorithm 1 is simply to simulate an agent with a particular set of capabilities. \\n\\nTo this end, we first sample a set of \\u201cactual\\u201d block heights (Step 2). \\n\\nWe then account for the agent\\u2019s measurement capability, by sampling a measurement error for the height of each block, using the distribution of errors we\\u2019ve seen the agent make in a separate capability test (Step 3). \\n\\nNext, we assess the agent\\u2019s planning ability, by\\n* Sampling a set of configurations that the agent might have conceived of, based on the number of configurations the agent generated in a separate capability test (Step 4),\\n* Computing the height the agent would have assigned to each configurations, according to its flawed block heights estimates from Step 3 above (Step 5), and then\\n* Picking a preferred configuration based on these height estimates, here taking into account the agent\\u2019s ability at picking out the best configuration given some height estimates (Step 5), as usual based on an independent assessment of the agent\\u2019s ability at this\\n\\nFinally, we take into the agent\\u2019s ability at actually building the tower it intends to (Step 8). Again, this is based on an independent assessment of the agent\\u2019s ability to execute an intended plan.\\n\\n_Line 250: On 205 you mentioned sigma to be .1 * true height, yet here you also added a constant factor._\\n\\nThanks for pointing this out. We added a constant term to our model of agent measurement errors, to account for the possibility that agents adapt the number of questions to the noise observed in the measurements. An explanation would have been warranted. \\n\\nOn reflection, we have decided that it is cleaner to instead just sample one of the agent\\u2019s actual measurement errors on a block of similar height from the capability check. This sidesteps the need to fit a model, and the extra assumptions this entails. The results are only marginally affected. We will update the paper to reflect this change.\\n\\n# Result analysis\\n\\n_Why does the regret for GPT 3.5 fluctuate (i.e., drop and then increase) in the observed experiments?_\\n\\nThe slight dip at 4 blocks may well be a statistical fluke, as the mean at 5 blocks is well within the SEM of the 4-block result. More surprising perhaps is the dip at 4 blocks for both GPT-3.5 and GPT-4 on line 420. One possible explanation for this is that putting an equal number of blocks in each tower is a reasonable heuristic for making equally high towers. This heuristic works better for an even number of blocks.\"}", "{\"comment\": \"I thank the authors for replying to my own and other reviewer's concerns. I understand the definitions and the contribution of this work better after this discussion. However, I think that this is evidence that significant changes to the manuscript are necessary to communicate the benchmark protocol and its goal as clearly as possible. A more formal definition of the goal-directedness-gap, under the Reinforcement Learning framework could significantly enhance the presentation (from discussion with reviewer UoUh).\\n\\nMoreover, now that I better undertand the contribution I think that it is expected for any benchmark paper accepted at a top Machine Learning venue to have a review of related work, open source implementation and assesment of current SOTA models on said benchmark. This reduces the contributions by the authors to (1) a conceptual definition of goal-directedness suitable for LLMs, and (2) a principle for evaluating it empirically (from discussion with reviewer BDx1). These contributions do not seem to warrant a higher score.\\n\\nIn my opinion a good benchmark should encourage further research on the area by having meaningful empirical applications (or by serving as a stepping stone to them) and it is still unclear to me that the benchmark could be useful under more realistic scenarios. As the authors have stated, measuring goal-directedness requires that the task can be broken down in subtasks, and that that there is no alternative way of solving the task other than the one suggested by the breakdown. This significantly limits the applicability of this benchmark (as a training or safety metric), especially without any results on the generalization properties of goal-directedness across LLMs and across tasks.\\n\\nReviewer yUcf has also pointed out that there is a flaw in assuming access to the agent's capabilities since measuring those capabilities is indistiguishable from assesing the agent's overal performance in that specific sub-task, which might in fact require goal-directed behavior itself. Since, the breakdown is itself a sub-task the authors must then assume the granularity and partitioning of the sub-task breakdown. These two components make the overall benchmark seem very subjective and lacking theoretical motivation. I would be interested in understanding the author's perspectives regarding these points, especially if they feel that I may have misunderstood important elements of this work.\"}", "{\"comment\": \"Appreciate the detailed responses by authors. I concurr with reviewer SAsD that the paper requires major revision and needs another round of review.\"}", "{\"title\": \"Anagram environment\", \"comment\": \"Multiple reviewers have asked how the results generalize beyond the blocksworld environment in which our main evaluations are carried out.\\n\\nSince the submission of the paper, we have developed an anagram environment, where the agent gets a sequence of letters, and is asked to permute them into one or more valid English words. This is a common setting for assessing human goal-directedness.\\n\\nWe compare actual performance on this task, with the agent\\u2019s expected performance based on its capabilities to \\n* permute a sequence of letters in many different ways, and to \\n* recognize whether an arbitrary permutation is an English word or not.\\n\\nWe find that the lack of goal-directedness is less significant in this setting, but that several models still demonstrate a lack of goal-directedness, especially for short sequences of letters (2 or 3). For longer sequences, models struggle to generate all or even most permutations of the letters (the number of permutations increases exponentially), meaning that their expected performance is sometimes worse than their actual performance, suggesting that the models have an alternative way of finding anagrams that doesn\\u2019t involve explicitly generating and checking permutations of the letters.\\n\\nAs for the task we already considered in the paper, Gemini comes out as the most goal-directed model also on the anagram task.\"}", "{\"summary\": \"This paper investigates the performance of LLMs in goal-directedness behavior. It explores whether LLMs can effectively leverage their capabilities and resources to achieve defined objectives by defining goal-directedness, developing an evaluation framework, and assessing LLM performance in practical tasks. The primary contribution of this paper lies in the design of the evaluation framework. This framework quantifies the extent to which agents utilize their capabilities toward a specific goal by measuring relevant agent abilities, predicting how agents would resolve problems if they fully employed these abilities, and comparing the predicted performance with the agents' actual problem-solving outcomes.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. The goal-directedness approach enhances the effectiveness of LLM-based agents by providing robust constraints in terms of safety and ethics. Therefore, evaluating the goal-oriented nature of LLMs is of significant importance.\\n2. The clarity of the writing is commendable, facilitating comprehension.\\n3. The design of the evaluation framework is intriguing, and the conclusions drawn regarding the goal-directedness of several state-of-the-art LLMs are persuasive.\", \"weaknesses\": [\"The discussion of related work is insufficient. For instance, the paper \\\"Can Large Language Models Reason About Goal-Oriented Tasks?\\\" appears to provide a more comprehensive interpretation of the goal-oriented nature of LLMs. I recommend that the authors elaborate on the distinctions and connections between the two studies.\", \"The contributions of the paper are relatively limited, with the primary contribution seemingly revolving around the design of the goal-directedness deficit and the evaluation of several LLMs' performance.\", \"The experimental design is somewhat simplistic and does not adequately capture the goal-directedness of LLMs across more complex and diverse tasks. Further investigation is needed into the internal mechanisms of the models, specifically how decisions are made and how these decisions relate to their goal-directed behavior.\", \"It is essential to test the influence of auxiliary methods (e.g. COT, TOT) aimed at enhancing the reasoning capabilities of the models on the experimental outcomes.\"], \"questions\": \"See weaknesses.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you for your feedback and your positive review!\\n\\n# Weakness 1: \\\"turtles all the way down\\\"?: \\n\\nThis is a good point. There is in general no single, objective way to break down a task. Nevertheless, any particular breakdown can still give evidence of a lack of goal-directedness, if the performance on the combined task is worse than would be expected from the performance on the subtasks. This relies on a tacit assumption that doing smaller, more well-scoped tasks requires less goal-directedness than doing larger tasks.\", \"a_breakdown_needs_to_satisfy_two_key_properties\": \"* It should be clear to the agent how to compose the subtasks to solve the full tasks. (This means we cannot break the task down into too miniscule components, such as steps of a Turing machine, since then it would not be obvious how to compose the tasks.)\\n* There should be no alternative way of solving a task other than the one suggested by the breakdown.\\n\\nWe will endeavor to highlight these important points more in the updated manuscript.\\n\\n# Weakness 2. How can we be confident that we have now reached a pure source for capabilities measurements?\\n\\nWe\\u2019re not confident that there usually is such a thing as a complete breakdown of a task into \\u201cprimary\\u201d capabilities. As discussed in the response to Weakness 1 above, there will typically be many ways to break down a task into smaller capabilities. Any breakdown that satisfies the principles above can give evidence about lack of goal-directedness.\\n\\n# Weakness 3. The only results are shown in the BlocksWorld environment. \\n\\nThe reason we choose the blocksworld environment is that it lets us formulate a range of different kinds of tasks and objectives in a unified environment. Thus we were able to define: \\n* the Build Highest Tower task (which mostly requires measuring), \\n* the Equal Towers task (which requires significant cognitive effort and execution skills), and \\n* the Falling Tower task, which requires perseverance in spite of adversity. \\nSo even though it is a single environment, we still test a range of different kinds of tasks.\\n\\nNevertheless, we do recognise that testing the principle in other environments would also be valuable. Please see the global comment describing another setting in which we\\u2019ve tested our goal-directedness metric.\\n\\n# Questions 1 and 2\\n\\nSee responses to weakness.\\n\\n# Question 3\\n\\nGoal-directedness is a property of an agent and may well vary across tasks -- different tasks have different levels of complexity and require the use of different capabilities. Nevertheless, the fact that both Build Equal Towers and Falling Towers tasks, as well as the just implemented anagram task, have the same \\u201cwinner\\u201d (Gemini-1.5-Pro), indicates there may be some level of generalization across tasks.\"}", "{\"summary\": \"Authors defined the concept of goal-directedness for LLM-based agents and measured this metric in tasks that involve information gathering, planning, and execution phases. They empirically probed four SOTA LLM agents to tackle a block stacking task focusing on different combinations of the phases. They found limited goal-directedness across all LLMs with Gemini performing better overall.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": [\"New Metric: There is a huge interest in the community to better evaluate LLMs for planning and the goal-directedness can play a role\", \"Empirical evaluation: Authors highlighted that current LLMs, despite their sophisticated capabilities, do not fully deploy these capabilities toward goal achievement.\", \"It was refreshing to see authors focused on stochastic environments.\"], \"weaknesses\": [\"Readability: The paper is hard to follow and the definitions are not rigourous. In 79, you mentioned \\\"The gap between actual and expected performance given optimal use of capabilities\\\". To me this means the difference between expected return under optimal policy and observed return under optimal policy. Not sure how this translates to goal-directedness as it only captures domain stochasticity. In 171, you mean expected return vs. expected reward. Reward is instantaneous while return is the cumulative return. In 174, can you define E[regret | optimal use of capabilities]? The second eqn. means how much the return could improve with additional capabilities. I don't think it is the same as goal-directedness-deficit.\", \"Generalization Issues (Line 139): The statement that LLMs do not ask clarifying questions might not fully apply, as some LLMs, like GPT-4, demonstrate this capability.\", \"Terminology Ambiguity (Lines 159 & 171): Concepts such as \\u201ccapability-conditioned goal-directedness\\u201d and expected \\u201creturn\\u201d versus \\u201creward\\u201d lack clarity and could be better explained.\", \"Undefined Terms (Lines 174, 289 & 313): Key terms, including\", \"E[regret\\u2223optimal\\u00a0use\\u00a0of\\u00a0capabilities], \\\"unexplained_regret,\\\" and \\\"statistical sophistication,\\\" need explicit definitions to improve comprehensibility.\", \"Algorithm Motivation (Line 244): Lacks clear motivation for Algorithm 1. Would be great to discuss its creation.\", \"Inconsistent Noise Parameters (Line 250): On 205 you mentioned sigma to be .1 * true height, yet here you also added a constant factor.\", \"Presentation Quality (Various Lines): Figures and images are of low quality, lacking necessary labels (e.g., missing legend in Fig 8.b) and captions (e.g., Fig 4), and should be vectorized with larger fonts for readability.\"], \"minor\": [\"389: \\\"Gemini performs better than would be expected from\\\" Fix grammar\", \"492: Whence -> hence\", \"504: language-based LLM -> LLM as L stands for language\"], \"questions\": [\"(Line 369): Why does the regret for GPT 3.5 fluctuate (i.e., drop and then increase) in the observed experiments?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you for your review of our paper, which you correctly describe as \\u201cdemonstrating that current models lack the ability to use their resources and capabilities fully\\u201d and \\u201cwell motivated in terms of potential applications for the evaluation and understanding of LLMs.\\u201d:\", \"let_us_address_the_concerns_that_you_raised\": \"# Weakness 1: Do models have the necessary capabilities? If not, does the measurement mean much?\\n\\nIt\\u2019s true that LLMs are far from universally capable AIs, and sometimes have surprising weaknesses and failure modes. (The paper you cite is already a few years old, so some of these failure modes have already disappeared.)\\n\\nFor this reason, a significant part of our paper is devoted to assessing whether models do have the necessary capabilities to do the tasks that we test them on. In general they do. It\\u2019s also worth emphasizing that our tasks allow a range of outcomes beyond just fail/pass, and allow us to measure how close a model came to a \\u201cperfect\\u201d answer. The tasks are hard enough that no model solves them perfectly all the time, and easy enough for even GPT-3.5 to often make at least some progress on them.\\n\\nFurthermore, our goal-directedness metric is computed by comparing the performance of models on a combined task, with performance on different subtasks.\\n\\nWe have also tailored the instructions, to minimize the frequency of models misunderstanding the task or interface.\\n\\n# Weakness 2. Seeds and results\\n\\nApologies, the paper should of course mention this (we\\u2019ll make sure to bring it back). For the main results and ablations, models are always run on 5 seeds. The error bars show the Standard Error of the Mean (SEM). The standard deviation is about twice the width of the SEM bars (sqrt(5) times wider, to be precise).\\n\\n# Weakness 2: Small number of datapoints (evaluation with only 3,4 and 5 blocks) and a single environment\\n\\nWhat is a sufficient number of experiments is of course always somewhat subjective. But let\\u2019s us emphasize that within our blocksworld environment, we have tested models on a range of different tasks and objectives: \\n* the Build Highest Tower task (which mostly requires measuring), \\n* the Equal Towers task (which requires significant cognitive effort and execution skills), and \\n* the Falling Tower task, which requires perseverance in spite of adversity. \\nSo even though it is a single environment, we still test a range of different kinds of tasks. And for each of these tasks, we have tested for a range of different numbers of blocks. \\n\\nAcross essentially all these tasks and settings, the finding that most models lack some goal-directedness has been consistent.\\n\\nIn reviewer yUcf\\u2019s opinion, \\u201cThe experiments conducted are very close to \\\"minimal but complete\\\" in that they include all and only the aspects that are necessary to study goal-directedness, which makes for very clear results.\\u201d\\n\\nSince the submission of the paper, we have also developed an anagram environment. Please see the global comment.\\n\\n# Weakness 3, clarity\\n\\nThanks for pointing this out. We will make sure to reiterate the importance of understanding goal-directedness also in the conclusion, and perhaps find a way to highlight it more in the introduction. \\n\\nAs pointed out by reviewer yUcf, a nuanced understanding of LLMs is essential as LLMs become ever more widely deployed. In particular, many expect them to soon be deployed as \\u201cagents\\u201d acting autonomously on behalf of users, where goal-directedness is a key property for both safety concerns and capabilities.\"}", "{\"comment\": \"Thanks for your suggestions!\\n\\n> either the measure of goal-directedness could be shown to have \\\"nicer\\\" and more robust properties (e.g. maybe there is a principled algorithm for breaking down tasks into component capabilities -- this is related to my weakness 1\\n\\nAny measure of goal-directedness will rely on a way of disentangling the agent's capabilities from its motivation. This is feasible for tasks for which there is a clear set of required capabilities, and a single natural way to solve the task given those capabilities. I think we can do a better job at articulating this principle, and showing how it can apply across different contexts.\\n\\nWe're not sure it's important that it can be done for *all* tasks, as it seems interesting enough if we can establish a lack of goal-directedness across a reasonable range of tasks.\\n\\n> or maybe goal-directedness of a given agent is shown to be consistent across tasks requiring different capabilities -- this is related to my question 3)\\n\\nDoes the anagram environment described in our top comment help address the concern of BlocksWorld specificity? https://openreview.net/forum?id=BECkhjcofz&noteId=eCcPdjHkKw\\n\\n> the distinction between goal-directedness and capabilities could be shown to be more \\\"useful\\\" (e.g. it facilitates clear and generalizable interventions -- such as the intervention of telling the agents to try harder, which the authors tested, except more powerful or more systematically investigated).\\n\\nThis is a nice idea, and something we will look further into!\"}", "{\"comment\": \"Thank you for your review of our paper which measures the goal-directedness of LLMs, for which you found the \\u201cconclusions drawn regarding the goal-directedness of several state-of-the-art LLMs \\u2026 persuasive.\\u201d\", \"let_us_address_each_of_the_concerns_you_raised\": \"# The discussion of related work is insufficient: \\nOur paper provides an extensive literature review over related work both in machine learning and neuroscience and psychology, spanning more than a page and resulting in more than 3 pages of references.\\n\\nWe will be happy to also include the paper you mention. Briefly, they are primarily interested in measuring the capabilities of LLMs to reason about goal-oriented tasks, while we are interested in the motivation of LLMs to use capabilities to solve such tasks. A main finding of their paper is that Tree of Thoughts (ToT) reasoning is less effective for reasoning about sequential reasoning tasks, while Chain of Thought (CoT) prompting only works in certain sequential reasoning scenarios, and is detrimental in others. They also offer a useful characterisation of key properties of goal-directed tasks, which we\\u2019ll consider incorporating.\\n\\n\\n# The contributions of the paper are relatively limited: \\nAs summarized at the end of our introduction, our contributions are five-fold: \\n* a conceptual definition of goal-directedness suitable for LLMs, and \\n* a principle for evaluating it empirically (Section 3),\\n* an open-source implementation of 5 tests in a Blocksworld environment, and \\n* an assessment of the goal-directedness of 4 LLMs (Section 4), as well as \\n* a review of related work in psychology/neuroscience and AI (Section 2). \\nAs pointed out by reviewer yUcf, a nuanced understanding of LLMs is essential as LLMs become ever more widely deployed. In particular, many expect them to soon be deployed as \\u201cagents\\u201d acting autonomously on behalf of users, where goal-directedness is a key property for both safety concerns and capabilities. \\n\\n# The framework is too simplistic and would not capture goal-directedness in other tasks and domains? \\n\\nAs reviewer yUcf points out, \\u201cThe proposed definition of goal-directedness is clear, straightforward (this is a positive thing), and well-motivated. The experiments conducted are very close to \\\"minimal but complete\\\" in that they include all and only the aspects that are necessary to study goal-directedness, which makes for very clear results.\\u201d\\n\\nFurthermore, our framework can be easily used to evaluate LLM goal-directedness both during and after training, and the principle is easily transferable to other tasks and environments. The experimental design is carefully crafted to facilitate this. In our response to reviewer yUcf, we briefly describe how we have applied the same principle to an anagram environment.\\n\\n# Further investigation is needed into the internal mechanisms of the models. \\n\\nMechanistic interpretability is beyond the scope of our work. Despite ongoing efforts to understand the internal workings of deep neural networks, and LLMs in particular, spanning 100s if not 1000s of papers from top universities and AI labs, the field is still far from answering much more basic questions such as how facts are stored or simple questions answered (the situation is not so different from Neuroscience, in fact). The fact that the behavioral approach pursued in our paper doesn\\u2019t rely on a detailed understanding of the internals of the models is one of its key strengths. We will emphasize this in the updated version of our manuscript. \\n\\n# It is essential to test the influence of auxiliary methods (e.g. COT, TOT): \\n\\nWe effectively already test Chain-of-thought, as the system prompt encourages the models to \\u201creason-step-by-step\\u201d before outputting their next action (line 183 in the paper). That said, we agree that it would be interesting to test the influence of various scaffolding techniques. We discuss this in the limitations section (line 482). We hope others will use our framework (which will be open sourced upon paper publication) to systematically explore how other scaffolding techniques impact the goal-directedness of models. \\n\\nPlease let us know if our answer addresses your concerns or if you have additional questions.\"}", "{\"comment\": \"Thanks for your points, and glad to hear you feel like you understand it better now.\\n\\n_\\\"A more formal definition of the goal-directedness-gap, under the Reinforcement Learning framework could significantly enhance the presentation (from discussion with reviewer UoUh).\\\"_\\n\\nFair point, we'll incorporate this.\\n\\n_\\\"I think that it is expected for any benchmark paper accepted at a top Machine Learning venue to have a review of related work, open source implementation and assesment of current SOTA models on said benchmark\\\"_\\n\\nWe do have an extensive review of related work (Section 2). The implementation will be made open source with the release of the paper (line 197). We have evaluated both Gemini and several GPT models. We'll consider adding evaluations also for other models.\\n\\n_\\\"This reduces the contributions by the authors to (1) a conceptual definition of goal-directedness suitable for LLMs, and (2) a principle for evaluating it empirically.\\\"_\\n\\nWe agree that the conceptual definition and the principle for evaluating it empirically are key contributions. Goal-directedness is an important metric to keep track of as LLMs are increasingly deployed as agents. Therefore having a way to measure it is quite valuable. \\n\\nAnother key finding is that current SOTA models do lack a fair bit of goal-directedness. This suggests that there is a \\\"capability overhang\\\" in that current models are not using their capabilities fully, which in turn suggests a path to making them more agentic, \\n\\n_\\\"As the authors have stated, measuring goal-directedness requires that the task can be broken down in subtasks, and that that there is no alternative way of solving the task other than the one suggested by the breakdown. This significantly limits the applicability of this benchmark\\\"_\\n\\nIt's true that it's not possible to straightforwardly assess the goal-directedness of a model on any task. This is because assessing goal-directedness always requires disentangling motivation from capabilities, which is often non-trivial. However, what our paper shows is that it's still possible to assess goal-directedness of models by finding tasks that naturally break down into subtasks.\\n\\nWhile the goal-directedness of a model will likely vary somewhat between tasks, being able to assess it on a number of (decmoposable) tasks can still give an idea of how goal-directed a model is in general, which can guide further training strategies and safety mitigations.\\n\\n_\\\"there is a flaw in assuming access to the agent's capabilities since measuring those capabilities is indistiguishable from assesing the agent's overal performance in that specific sub-task, which might in fact require goal-directed behavior itself. Since, the breakdown is itself a sub-task the authors must then assume the granularity and partitioning of the sub-task breakdown.\\\"_\\n\\nThe assumption we're making here is that it requires less goal-directedness to do a smaller task than to do a larger task. This assumption seems strongly consistent with folk psychology in humans. For LLMs, it also seems quite easy to just answer a simple question (in fact, getting them not to do it is often the hard thing -- cf jail breaks), whereas larger tasks requiring chaining together many steps of reasoning will be harder for them. The larger tasks therefore require more \\\"motivation\\\", in some plausible generalisation of the term. Finally, it's also consistent with our empirical results.\\n\\nAs we explained in our reply to yUcf, there will generally not be a breakdown of a task into some \\\"primary\\\" capabilities. Instead, there will typically be zero or more ways to break down a task into sub-tasks. Any particular breakdown can give us evidence of lack of motivation.\"}", "{\"comment\": \"I thank the authors for the helpful response, as well as the other reviewers for the thoughtful comments.\\n\\nI still feel positively about this work, but perhaps less strongly than initially. After more careful thought, I'm coming to view the weakness 1 and question 3 from my initial review as being more significant issues than I initially realized. I now also agree with\\nreviewer BDx1 comments that the contributions of the paper are relatively limited. Although I still commend the quality and clarity of the experiments that were presented, the bottom line is perhaps insufficiently significant.\", \"i_see_two_main_paths_by_which_this_work_could_be_more_clearly_significant\": \"either the measure of goal-directedness could be shown to have \\\"nicer\\\" and more robust properties (e.g. maybe there is a principled algorithm for breaking down tasks into component capabilities -- this is related to my weakness 1 --, or maybe goal-directedness of a given agent is shown to be consistent across tasks requiring different capabilities -- this is related to my question 3), or the distinction between goal-directedness and capabilities could be shown to be more \\\"useful\\\" (e.g. it facilitates clear and generalizable interventions -- such as the intervention of telling the agents to try harder, which the authors tested, except more powerful or more systematically investigated).\\n\\nI remain excited to see further work by the authors, but for now I downgrade my decision recommendation from \\\"accept\\\" to \\\"marginal accept\\\".\"}", "{\"summary\": \"The paper proposes a mathematical definition of \\\"goal-directedness\\\" in the context of LLM-agents as the difference between the agent's actual and its expected performance on tasks given optimal use of its capabilities. The authors then introduce a toy Blocksworld environment and an associated task with three variants and proceed to measure the performance and goal-directedness of some LLM-agents on this task. The LLM-agents studied are based on GPT-3.5-turbo, GPT-4-1106, GPT-4o, and Gemini-1.5-Pro. They find that the agents differ in their capabilities and goal-directedness, and while they are generally all rather capable, they are not very goal-directed. The paper also discusses associated ablation experiments and closer analyses.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"4\", \"strengths\": \"**Originality**\\n\\nWhile other work has aimed to define goal-directedness in similar settings, the definition proposed in this paper seems to be novel, while also being elegantly simple and powerful in its generality.\\n\\n**Quality**\\n\\nThe toy setting and task introduced in the paper are well-designed and highly appropriate for the investigation. The experiments are well-motivated, the ablations are interesting and useful, the analyses are informative and include confidence intervals. Overall, it is clear the authors had good attention to the question at hand and applied good intuition to what experiments would be most informative.\\n\\n**Clarity**\\n\\nThe proposed definition of goal-directedness is clear, straightforward (this is a positive thing), and well-motivated. The experiments conducted are very close to \\\"minimal but complete\\\" in that they include all and only the aspects that are necessary to study goal-directedness, which makes for very clear results. The text flows well and the sections motivate each other. The plots are clear and informative.\\n\\n**Significance**\\n\\nAs LLM-agents become increasingly widely studied and deployed, it becomes increasingly important to form nuanced understanding of them in general settings. This paper contributes to this. Goal-directedness as a trait of LLM-agents also has safety implications, and as such this contribution is all the more significant.\", \"weaknesses\": \"1. It is a significant issue that the calculation of goal-directedness assumes direct access to the agent's capabilities. This is because the measurement of the capabilities looks a lot like the measurement of the agent's overall performance on a task. Presumably it already will reflect the agent's goal-directedness in some capacity. Essentially, it creates a \\\"turtles all the way down\\\" problem. Breaking the task down seems like a reasonable thing to do, but not only is it imperfect in the sense above, it also requires some subjective judgement on the part of the experimenter (different experimenters can reasonably disagree on how a task should be broken down, especially in the case of more realistic tasks).\\n\\n2. A concrete example of the issues above: prompting for motivation significantly affected the agent's motivation on the \\\"Cognitive Effort\\\" partial task. So the claim being made here is that this partial task is already very \\\"goal-directedness\\\"-loaded, i.e. \\\"goal-directedness\\\" is a significant component of the agent's performance on the task. But this is already a broken-down task, and to get to direct capabilities, the authors only break it down further into two component capabilities. How can we be confident that we have now reached a pure source for capabilities measurements?\\n\\n3. The only results are shown in the toy BlocksWorld environment. I think it makes sense for the main experiments to be conducted there, but I am left very curious to see how the framework can be applied to realistic tasks, and in particular, whether the results would hold there.\\n\\nI would like to note that while I maintain that the above are significant issues, I still feel positively about the paper's contribution, and hope to see further work in this direction making progress on the issues raised.\", \"questions\": \"1. Do the authors agree with the issue I point out in Weakness 1? How do they think the presumed purity of the capabilities measurements can be justified?\\n\\n2. Similarly, how do the authors recommend that a task be broken down into its constituent capabilities? Is there an objective general procedure?\\n\\n3. Is goal-directedness here claimed to be a property of an agent? Is there any claim as to how the goal-directedness measurements on an agent would differ across different tasks? We do expect the same agent to be differently capable at different tasks. Should they also be differently goal-directed? From the conceptual motivation, I would hope this is not strongly the case, but I expect that experimental results might show otherwise. Do the authors agree?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper introduces a new benchmark to evaluate the \\\"goal-directedness\\\" of Large Language Models (LLMs); \\\"goal-directedness\\\" in this context is defined as \\\"the propensity to use available resources and capabilities to achieve a given goal\\\". To do so, the authors create a BlocksWorld environment in which the \\\"goal-directedness\\\" deficit can be estimated as the difference between the expected rewards given that the capabilities are used to their fullest extend and the actual reward obtained . Then, the authors evaluate different LLMs in this environment (and ablations of it) demonstrating that current models lack the ability to use their resources and capabilities fully.\", \"soundness\": \"2\", \"presentation\": \"1\", \"contribution\": \"2\", \"strengths\": \"I want to preface this review by saying that LLMs and their evaluation is not my area of expertise so I might be not familiar enough with the current literature to assess the impact of the contributions of this paper. Overall I see potential in the area of research and the problem that is trying to be addressed. The problem is well motivated in terms of potential applications for the evaluation and understanding of LLMs.\\n\\n1. **Neglectedness**: The study of goal directed behavior in LLMs is interesting, has some useful applications and has been almost completely neglected in current research directions.\\n2. **Safety and Ethics Implications**: The paper discusses the potential risks associated with more goal-directed AI agents, which is a valuable contribution to the state of AI safety and ethical considerations of AI.\\n2. **Ablation Study**: The ablation study does a good job at trying to isolate the components of the task and identify the most challenging aspects.\", \"weaknesses\": \"1. **Coherence of the metric and its goal**: My main issue with this paper is that the connection between the \\\"goal-directedness\\\" deficit and the actual metric that is introduced is not completely clear. More especifically, the authors state that they want to differentiate between two subtle cases: wether the model fails because it lacks the $\\\\textit{capabilities}$ or the $\\\\textit{motivation}$ to do the task. By definition, $\\\\text{goal-directedness-deficit(regret, capabilities)} = \\\\mathbb{E}[\\\\text{reward} | \\\\text{optimal use of capabilities}] \\u2212 \\\\text{reward}$. This makes the assumption that the goal, namely stacking the blocks so as to minimize the height difference between two towers, is a task that is within the capabilities of LLMs. And this is likely to be false, as LLMs have been shown to be unable to perform well in seemingly simple tasks that require quantitative reasoning, e.g. reverse alphabetical sorting of words, even with chain of thought [1]. Therefore the metric fails to stablish the distinction on wether the evaluated LLMs fail because they lack capabilities or the motivation to do the task.\\n\\n2. **Experimental rigor**: For all of the results plots it is not stated the number of evaluation seeds and the standard deviation of the sample, making it impossible for the reviewers to make an assessment about the statistical properties of the result and validate some of the authors' claims (e.g. in line 412 the at\\\"perhaps a statistical fluke\\\"). More importantly, the authors derive their conclusions from a small number of datapoints (evaluation with only 3,4 and 5 blocks) and a single environment. Ideally a good benchmark should be able to show that the results might generalize to more complex scenarios, by showing results in a diversity of low complexity scenarios.\\n3. **Clarity**: Although the motivation is good, it was very difficult for me to understand what it was originally because some of its key ideas fail to be highlihted and reiterated (e.g. the importance of \\\"goal-directedness\\\" for safety and the distinction between \\\"goal-directedness\\\" and other reasoning tasks).\\n\\n**References:**\\n\\n[1] Suzgun, M., Scales, N., Sch\\u00e4rli, N., Gehrmann, S., Tay, Y., Chung, H. W., Chowdhery, A., Le, Q. V., Chi, E. H., Zhou, D., & Wei, J. (2022). Challenging BIG-Bench tasks and whether chain-of-thought can solve them. arXiv preprint arXiv:2210.09261. https://arxiv.org/abs/2210.09261\", \"questions\": \"1. What was the reasoning for using a BlocksWorld environment as an indicator of \\\"goal-directedness\\\"?\\n2. Did you consider other potential tasks?\\n3. What would constitute a good task for the purpose of measuring goal-directed behavior, i.e. what properties were you looking for?\\n4. Why not fine-tune a model to provide more insight about the difficulty of the task?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Questions\", \"comment\": \"# Question 1: Why blocksworld?\\nThe blocksworld environment lets us formulate a range of different tasks and subtasks in a unified setting. \\n\\n# Question 2: Did you consider other potential tasks?\\nYes, and after submission we also implemented an anagram environment. Other environments that we\\u2019ve been considering include agency evals frameworks such AgentBench [TODO: more?]. However, the blocksworld environment is simple and flexible, allowing us to analyze goal-directedness on various tasks.\\n\\n# Question 3: Criteria for task selection\\nA key aspect is that the environment should support tasks that can naturally be broken down into subtasks. As explained to reviewer yUcf, a breakdown needs to satisfy two key properties:\\n* It should be clear to the agent how to compose the subtasks to solve the full tasks. (This means we cannot break the task down into too miniscule components, such as steps of a Turing machine, since then it would not be obvious how to compose the tasks.)\\n* There should be no alternative way of solving a task other than the one suggested by the breakdown.\\n\\nThe Equal Towers task naturally breaks down into a measuring phase, a planning phase, and an execution phase, none of which can realistically be avoided, and it\\u2019s sufficiently obvious to the agent that each of these phases have to happen.\\n\\nBeyond this, it is also important that the interface is clear and understandable to the agents, so that our results measure agent\\u2019s capabilities at the task, rather than their capabilities at understanding the interface (it took significant iterations to find an interface and environment that was palatable to all the agents that we wanted to test). \\n\\nFinally, the task must be suitably hard, so that models neither do it perfectly nor just random guessing. The blocks world in particular let us adjust the number of blocks, as an easy way to adjust task complexity\\n\\n\\n# Question 4: Why not fine-tune models? \\nOur goal is not to measure the difficulty of a task, but to assess the goal-directedness of any LLM agent. The framework we propose for doing this is general, flexible and can be applied to any model, including fine-tuned models. It would be an interesting line of future work, to see how far fine-tuning can improve the goal-directedness of LLM agents.\\n\\nPlease let us know if our answers have addressed the concerns raised, or if you have additional questions.\"}", "{\"title\": \"Minor points\", \"comment\": \"Line 171: Apologies, we used the term reward informally here. We will clarify this in an updated version, either switching to return (as you suggest), or using \\u2018loss\\u2019 for a broader ML interpretation (in either case, specifying precisely what we mean with the term).\", \"line_139\": \"We will update the wording to \\u201cTheir lack of goal-directedness is further evidenced by some models not asking clarifying questions.\\u201d\", \"line_159\": \"We will clarify the sentence \\u201cIn contrast, our work measures capability-conditioned goal-directedness, taking into account the agent\\u2019s (lack of) capabilities when assessing its goal-directedness.\\u201d to make this paragraph better understandable in isolation.\", \"line_289\": \"unexplained_regret is the same as goal-directedness deficit. We will remove the term unexplained_regret and only use goal-directedness deficit throughout. Apologies for the confusion.\", \"line_313\": \"What we mean by \\u201cStatistical sophistication\\u201d is explained in the subsequent sentence, i.e. it\\u2019s the propensity to adapt the number of measurements based on how noisy they are. Please let us know if you think further clarification would be useful here.\\n\\nmissing legend in Fig 8.b and caption Fig 4. Apologies. The legend in Fig 8b is the same as for most of the other plots in the paper, but we will add this back to this plot for clarity (especially since the legend beside has a different legend). We will expand the Fig 4 caption, and increase the font size for the axes.\", \"typos\": \"Thanks, fixed.\"}" ] }

[ "Yangyi Chen", "Binxuan Huang", "Yifan Gao", "Zhengyang Wang", "Jingfeng Yang", "Heng Ji" ]

[ "Scaling Laws", "Downstream Performance Prediction", "Large Language Models" ]

{ "note_id": [ "xrkRF3WNAQ", "vgCkGKL8OA", "uGMkA6trtQ", "sUproHtQ20", "dkyXJGGfX6", "ctyuBO5A8I", "a0NgXYiifq", "ZlI7JtYWzS", "WFhrxZGJgr", "PR1U0Ql3MS", "N3p73nkd7T", "B3aGjYo1EJ", "7fk1Voqq4t", "360Fr4Fpc5", "2YZ8pyAnLA", "2WcpSeCcbG" ], "note_type": [ "official_comment", "official_review", "comment", "official_review", "official_comment", "official_review", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment" ], "note_created": [ 1732386666366, 1730561202981, 1736317678702, 1730092361416, 1732666766904, 1730287374634, 1732666782900, 1732387095239, 1730687321347, 1732696494882, 1732386608762, 1732387155318, 1732666797040, 1732386778102, 1732386900004, 1732386939496 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission7778/Authors" ], [ "ICLR.cc/2025/Conference/Submission7778/Reviewer_GNGv" ], [ "ICLR.cc/2025/Conference/Submission7778/Authors" ], [ "ICLR.cc/2025/Conference/Submission7778/Reviewer_rYdh" ], [ "ICLR.cc/2025/Conference/Submission7778/Authors" ], [ "ICLR.cc/2025/Conference/Submission7778/Reviewer_syLm" ], [ "ICLR.cc/2025/Conference/Submission7778/Authors" ], [ "ICLR.cc/2025/Conference/Submission7778/Authors" ], [ "ICLR.cc/2025/Conference/Submission7778/Reviewer_6L6M" ], [ "ICLR.cc/2025/Conference/Submission7778/Reviewer_rYdh" ], [ "ICLR.cc/2025/Conference/Submission7778/Authors" ], [ "ICLR.cc/2025/Conference/Submission7778/Authors" ], [ "ICLR.cc/2025/Conference/Submission7778/Authors" ], [ "ICLR.cc/2025/Conference/Submission7778/Authors" ], [ "ICLR.cc/2025/Conference/Submission7778/Authors" ], [ "ICLR.cc/2025/Conference/Submission7778/Authors" ] ], "structured_content_str": [ "{\"comment\": \"# Questions\\n1. Thanks for mentioning this! The validation set used in our experiments is de-duplicated against the pre-training dataset following standard practice. \\n2. In our analysis, we employed the standard logistic sigmoid function: $f(x) = C + \\\\frac{1}{1 + \\\\exp(-A x + B)}$, where $A$, $B$, and $C$ are fitted parameters. We also try alternative sigmoid formulations, including the generalized growth function $f(x) = \\\\frac{A}{1 + (\\\\frac{B}{x})^C}$. After systematic evaluation, the standard logistic sigmoid consistently demonstrated superior fitting performance. \\n3. Please refer to our response to weakness 3. \\n4. Thanks for pointing this out! The reason is there are overlapping plotted points due to very similar performance and the log scale of the Y-axis. We will fix this problem by adding a small variance to each point in the revision.\"}", "{\"summary\": \"This paper proposes two methods, FLP and FLP-M, for efficiently predicting the downstream performance of large language models. These methods achieve high-precision performance prediction.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"A notable strength of the paper is the quality of the writing: the narrative is clear, and the experiments are thorough. Besides, the FLP-M method accurately predicting performance based on data loss from different domains, thus enhancing prediction accuracy in mixed data scenarios. Additionally, Figure 6 demonstrates that FLP-M can be used to derive the optimal data mixing ratio for training.\", \"weaknesses\": \"1. The authors utilize intermediate checkpoints to gather data points; however, for the same amount of FLOPs, models with different N (parameters) and D (data) would yield distinct loss. This raises a critical question: why is it valid to use checkpoints that have not converged and are not optimized configurations to obtain data points?\\n\\n2. The second drawback is a lack of novelty. Both using FLOPs to predict loss and using loss to predict downstream performance have been explored in prior work.\\n\\n3. The third drawback is that the authors use a 1B model to validate the effectiveness of FLP-M scaling law for achieving an improved data mixture. However, this claim may be overstated, as 1B models often rely on guesswork for many tasks, undermining the reliability of these results.\", \"questions\": \"1. Have ongoing experiments been conducted on larger-scale models\\uff1f\\n2. How do you justify the usage of intermediate checkpoints for acquiring scaling law datapoints?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"withdrawal_confirmation\": \"I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors.\", \"comment\": \"We extend our sincere gratitude to the reviewers for their thorough and constructive feedback, which has significantly enhanced the quality and clarity of our revised manuscript. We thank reviewer syLm for the recognition of our work. We sincerely appreciate the detailed feedback from Reviewers 6L6M, GNGv, and rYdh for identifying areas for improvement in our work. However, we regret that despite our comprehensive responses addressing potential misunderstandings, we did not receive follow-up feedback from Reviewers 6L6M and GNGv. In our responses, we provided thorough clarifications regarding the points that may have been misinterpreted by the reviewers. In addition, we highly value the interaction with reviewer rYdh, which prompts us to revise our paper to make it more clear to eliminate some misunderstandings. However, we respectfully disagree with the comment from reviewer rYdh which states that \\\"Given that similar methods are already prevalent in industry (even if not open-sourced), the paper fundamentally fails to introduce genuinely novel techniques...\\\". It's uncommon that closed-source implementations (not obvious to the public), can serve as evidence to diminish the academic novelty and contribution.\\n\\nOverall, we thank the chairs for the organization and reviewers for their time and valuable comments. See you next time, ICLR!\"}", "{\"summary\": \"The paper introduces FLP to address the limitations of classical scaling laws, which fail to accurately predict performance when small models perform poorly on evaluation tasks, approaching random sampling. FLP leverages the scaling law of FLOPs to predict pre-training loss and uses this loss to predict downstream performance. FLP successfully predicted the performance of 7B and 13B models across six tasks using a series of language models up to 3B.\\n\\nBased on FLP, the paper further introduces FLP-M, which aims to predict downstream performance trained with various mixtures of general text and code.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. The paper identifies the issue of discontinuous performance when models approach the emergent edge, which is difficult to address with classical scaling laws, and proposes a method to resolve with the continuous variant ------ loss.\\n2. FLP creates more data points for fitting the scaling la, potentially making the fitted curve more generalizable.\\n3. FLP-M is introduced for data mixtures, providing a more accurate prediction by considering the different impacts of code and general text on downstream tasks.\\n4. The paper conducts extensive experiments to support its claims.\", \"weaknesses\": \"1. In section 3.2 Loss->Performance, there is a strong assumption that loss and accuracy have a linear relationship. Firstly, in all generative tasks shown in Figure 9, the linear relationship between loss and metric is not evident. The authors should provide more explicit statistical indicators to prove this linear correlation. Additionally, in the classification tasks shown in Figure 9, the relationship between loss and accuracy also encounters deviations near the emergent point, indicating that FLP does not completely bypass this issue but only circumvents it in the Flops -> Loss process.\\n2. A simple w_1*L+w_0 is not fundamentally different from classical scaling laws.\\n3. FLP-M only considers code and general text, while data mixtures typically need to consider at least five domains, including common crawl (cc), academic, books, encyclopedias, and code.\\n4. If the paper considers the situation around the emergent point in benchmarks, it lacks a discussion on the scenario when the model approaches near-perfect scores on a particular benchmark.\", \"questions\": \"My main concerns are twofold.\\n\\n1. Does the scaling law proposed in the paper have sufficient innovations compared to the classical scaling law? What are their essential differences? Please explain how the two-stage approach in FLP fundamentally differs from classical scaling laws in terms of methodology and theoretical underpinnings.\\n\\n2. If the paper focuses on the model performance around the emergent point, is a linear description really suitable? Should other nonlinear descriptions be considered, such as the sigmoid function, when considering scaling near the point of near-perfect accuracy? Is there any comparation?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Gentle reminder of the discussion period\", \"comment\": \"Hi reviewer,\\n\\nThanks for taking the time to review our paper. It would be great if you can take a look at our responses. Please let us know if you have further questions. Thanks!\"}", "{\"summary\": \"In this paper, the authors manage to predict the downstream performance of LLMs according to the computaional resouces (e.g., FLOPs). Experimental results shows that, by utilizing a 3B LLM trained on a specific ratio and a series of smaller sampling LMs, FLP-M can effectively forecast the performance of 3B and 7B LLMs across various data mixtures for most benchmarks within 10% error margins.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. **Practical Application Value** This paper introduces FLP-M, linking computational resources with LLM downstream performance. This research holds significant importance for real-world applications.\", \"weaknesses\": \"1. **Limited Scale of LM** The largest model used in this paper is only 7B, yet there are many LLMs much larger than 7B (e.g., Llama-3 70B, Llama-3 405B). From this perspective, the conclusions of this paper are limited.\\n\\n2. **Limited Domains in Data Mixing** As stated in the limitations, this paper only considers the domains of text and code under Data Mixing settings. Including more domains would enhance the explanatory power of the conclusions.\", \"questions\": \"See weaknesses.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Gentle reminder of the discussion period\", \"comment\": \"Hi reviewer,\\n\\nThanks for taking the time to review our paper. It would be great if you can take a look at our responses. Please let us know if you have further questions. Thanks!\"}", "{\"comment\": \"We thank the reviewer for the detailed comments. Our responses are as follows:\\n\\n# Weakness\\n1. A strong assumption that loss and accuracy have a linear relationship.\\n\\nWe appreciate the reviewer's thoughtful comment regarding the statistical validation of the Loss-Performance relationship. To rigorously examine this relationship, we conducted a linear regression analysis on the (Loss, Performance) pairs across all datasets and computed the coefficient of determination $ R^{2} $. Our analysis reveals an average $R^{2} $ value of 93% across all benchmarks, providing strong statistical evidence for the linear approximation of the Loss-Performance relationship. This quantitative validation complements the qualitative patterns observed in Figure 9.\\n\\nRegarding the deviation near the emergence point in classification tasks, we acknowledge this important observation. It's worth clarifying that the primary contribution of our FLP solution lies not in bypassing emergent abilities, but rather in enhancing sample efficiency through strategic utilization of intermediate checkpoints for Loss-Performance estimation. This approach allows us to effectively model the general trend of performance improvement, even if local nonlinearities exist at specific points in the training trajectory. The high $R^{2}$ values suggest that these local deviations do not significantly impact the overall effectiveness of our linear approximation approach.\\n\\n\\n2. A simple w_1*L+w_0 is not fundamentally different from classical scaling laws.\\n\\nWhile we acknowledge that our linear formulation appears structurally simple, we would like to emphasize that the fundamental contribution of scaling law research lies not in developing new and increasingly complex analytical forms, but rather in uncovering and validating predictive relationships that advance our understanding of model behavior. The distinctive innovation of our work lies in the two-stage estimation framework, particularly the Loss\\u2192Performance mapping, which leverages intermediate checkpoints to substantially improve sample efficiency and also our generalized prediction framework for the important data mixing setting.\\n\\n\\nOur experimental results demonstrate that this simpler formulation captures the essential scaling behavior while avoiding potential overfitting that can occur with more complex functional forms. This aligns with Occam's razor - when two models achieve similar performance, the simpler one is often preferable.\\n\\n3. FLP-M only considers code and general text\\n\\n\\u200b\\u200bThank you for this valuable observation. Our focused investigation on text and code domains was a deliberate choice driven by both practical constraints and scientific objectives. While we acknowledge that a comprehensive analysis across all five typical domains (common crawl, academic, books, encyclopedias, and code) would provide broader insights, our study prioritized depth over breadth for several reasons:\\n- The code-text interaction represents a particularly crucial use case in modern AI systems, especially for developer tools and programming assistants, making it an important starting point for domain mixing research.\\n- By focusing on this specific combination, we were able to conduct more thorough analyses and establish robust foundational principles that can inform future research across other domain combinations.\\n- Given our computational resource constraints, this focused approach allowed us to maintain rigorous experimental standards while still deriving meaningful insights.\\n\\nImportantly, while our empirical validation centers on code and text, the FLP-M methodology we propose is domain-agnostic and can be readily extended to other domain combinations. \\n\\n4. Lacks a discussion on the scenario when the model approaches near-perfect scores on a particular benchmark\\n\\nThank you for this insightful comment. In our experimental settings, we did not encounter scenarios with near-perfect scores, as the benchmarks we selected are sufficiently challenging that even cutting-edge language models like GPT-4 still show considerable room for improvement. Nevertheless, we acknowledge the theoretical importance of this edge case. From a methodological perspective, our framework can be easily extended to handle such scenarios through performance thresholding at 100%, ensuring that predictions remain bounded and meaningful even as models approach perfect accuracy.\"}", "{\"summary\": \"This paper introduces FLP (Flops $\\\\rightarrow$ Loss $\\\\rightarrow$ Performance), a two stage framework incorporating scaling laws to accurately predict the downstream performance of language models (LMs) on specific tasks by leveraging the pre-training loss. The first stage uses a power law equation to estimate the relation between flops and loss, $L(C) = \\\\big(\\\\frac{C}{C_N}\\\\big)^{\\\\alpha_N}$ by training 12 sampling LMs ranging from 43M to 3B parameters. The second stage involves using a linear function to estimate the relationship between the loss, $L$, and the task performance, $P$ [$P(L) = w_0 + w_1 * L$]. The second stage is applied carefully on those checkpoints that surpass the threshold of random performance + 5 additional performance points. The authors demonstrate better scalign law fits compared to the baseline of just using a power law function.\\n\\nThe paper then further extends the FLP approach to data mixing during pre-training (FLP-M) and presents an analysis on data-mixing ratios across general text and code and how mixing affects the downstream task performance, and extend the same two-stage framework of FLP, to predict downstream performance under different mixing ratios.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": [\"The paper tackles an important problem of building scaling laws to measure the downstream task performance, especially when we know that task-specific behaviour emerges at different scales and smaller scale LMs might not be able to accurately capture the predictive behaviour of larger models on certain tasks. The paper's two-stage approach of separating the FLOPs $\\\\rightarrow$ Loss and Loss $\\\\rightarrow$ Performance predictive models circumvents the emergent behaviour issue with the FLOPs $\\\\rightarrow$ Performance power law.\", \"The paper provides good insights on the mixing behaviour during pre-training on general text vs code by extending the FLP approach to FLP-M, with good empirical results on deriving the optimal mixing ratios (in a controlled setting).\", \"The experiments and results are exhaustive and involve a range of tasks including ARC-C, BBH, Hellaswag, HumanEval, RACE, and TriviaQA.\"], \"weaknesses\": [\"The sharp transition in performance of TriviaQA from 1B to 3B models highlights the brittleness of the approach, where the error margins can be huge for downstream task performance prediction. And it's very hard to characterize this behaviour for a whole range of tasks that are usually used to compare various LMs.\", \"I don't agree with the authors' point on enhancing sample efficiency by collecting losses corresponding to intermediate checkpoints and actually creates a biased estimator for the power law operands. Moreover intermediate checkpoints exhibit transient behaviours especially corresponding to learning rate adjustments (different intermediate checkpoints exhibit different learning rate schedules).\", \"I think there's a major typo in Equation 5, where the denominators of the second and the third terms are identical to the numerators. It hinders the understanding of the readers and it persists in the later sections too. [Although it's not a huge weakness and I am not basing my score on this point, assuming the authors will correct it in the rebuttal phase].\", \"The experimental setting corresponding to the comparison with Llama-3 is not explained properly, and it's hard to believe the results from Figure 11, provided that the estimated Llama-3 405B performance was quite close to the actual performance on ARC-C, whereas in this paper it's shown to be above 25%.\"], \"questions\": \"Here are a few additional questions for the authors in addition to the weaknesses above:\\n\\n1. Were the pre-training datamixes used for FLP-M experiments deduped against the validation set used in FLP and FLP-M experiments? Because it might affect the scaling behaviour if there's any overlap.\\n2. For the comparions with Llama 3 in Section C / Figure 11, what specific sigmoidal function was used? And did the authors ensure to choose the one that results in the best fit on the sampling LMs?\\n3. Can the authors please correct the typos in Equation 5 and Table 3 corresponding to $C_G$ and $C_C$?\\n4. In Figure 2, the sampling LMs corresponding to $\\\\leq 10^{18}$ flop scale seem to be missing. Is there a specific reason for this?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"I appreciate the effort the author has dedicated to this work. My primary concern pertains to the novelty of the paper, both in techniques and findings.\\n\\nFor ICLR, the innovation presented here is limited. The most notable idea\\u2014linking FLOPs to loss\\u2014lacks rigorous validation, as highlighted by reviewer GNGv. If each data point (checkpoint) referenced adheres strictly to the proportion of tokens to model parameters, 200, it does not contribute additional fitting points. Otherwise, using checkpoint losses introduces variability in the token-to-parameter ratio, thereby challenging the validity of the proposed scaling law. The author should provide supporting references or conduct thorough experiments to substantiate these claims.\\n\\nFurthermore, while the author distinguishes between loss functions and performance to emphasize novelty, this contribution remains constrained. Given that similar methods are already prevalent in industry (even if not open-sourced), the paper fundamentally fails to introduce genuinely novel techniques, particularly in the data mix section. The approach largely replicates existing data mixing laws but oversimplifies the problem, yielding results of limited significance.\"}", "{\"comment\": \"We thank the reviewer for detailed comments and feedback. Our responses are as follows:\\n# Weakness\\n1. The sharp transition in performance of TriviaQA from 1B to 3B models. \\n\\nWe acknowledge the reviewer's observation regarding TriviaQA's performance. However, we would like to emphasize two points:\\nit's important to note that among all six datasets and two settings (w or w/o data mixing) evaluated, TriviaQA with data mixing is the sole case where our approach showed reduced prediction accuracy. The general applicability of our approach is supported by the consistent and reliable performance our approach demonstrates across the other five datasets. \\n\\nThe TriviaQA is indeed a challenging case. As explained in our paper, the performance of sampling LMs improves sharply from 1B to 3B parameters (increasing from below 12 to over 28). In our sampling LMs configurations, we lack sufficient data points to adequately characterize the phase of accelerated performance improvement. Thus, this situation is very challenging for performance prediction since only limited information can be derived from the sampling LMs. However, our method still significantly outperformed other baseline approaches. **This comparison highlights that the reduced performance in this case should not be explained as a fundamental limitation of our method. Instead, this is the inherent property in the selected model family and the dataset.**\\n\\n2. Collecting losses corresponding to intermediate checkpoints creates a biased estimator for the power law operands. \\n\\nWe appreciate this concern but would like to point out a potential misunderstanding here. Our approach for estimating the FLOPs->Loss relationship relies solely on final converged training states, following established practices in prior work. **The intermediate checkpoints are not used for this estimation.**\\n\\nRather, we leverage intermediate checkpoints specifically for analyzing the Loss->Downstream Performance relationship. This distinct analysis is grounded in the well-established \\\"compression represents intelligence\\\" principle demonstrated in [1]. This principle suggests that language models exhibiting equivalent compression capabilities (as measured by pre-training loss) tend to demonstrate comparable task performance, regardless of the specific optimization trajectory taken to achieve that compression level. While transient behaviors like learning rate adjustments indeed influence the path to a particular loss value, the downstream performance correlates strongly with the achieved compression level itself, not the path taken to reach it.\\n\\nThis theoretical foundation provides a robust basis for utilizing intermediate checkpoints in mapping the relationship between achieved compression (loss) and downstream capabilities, while keeping the FLOPs->Loss estimation methodology unchanged from previous work.\\n[1] Compression Represents Intelligence Linearly; Yuzhen Huang et al\\n\\n3. A major typo in Equation 5\\n\\nThank you for your careful review. We would like to clarify that there is no typo in Equation 5. The notation may appear similar at first glance, but there are important distinctions: \\n- **In the second term**, we use $C^G$ in the numerator (where *G* is superscript) and $C_G$ in the denominator (where *G* is subscript), representing distinct mathematical quantities.\\n- **Similarly, in the third term**, $C^C$ appears in the numerator (superscript) while $C_C$ is in the denominator (subscript).\\n\\nThese notational differences are intentional and mathematically significant. We will enhance the visual clarity of these distinctions in the final version and add a brief explanation of the notation to prevent any potential confusion for readers.\\n\\n4. The experimental setting corresponding to the comparison with Llama-3 is not explained properly\\n\\nThank you for this important observation. We want to clarify the experimental methodology and address the apparent discrepancy in results:\", \"our_implementation_strictly_adheres_to_the_two_stage_performance_prediction_framework_outlined_in_the_llama_3_paper\": [\"Stage 1: We estimate the negative log-likelihood (NLL) based on FLOPs using the prescribed analytical forms\", \"Stage 2: We map the estimated NLL to task performance through a sigmoid function (details in our response to question 2)\"], \"the_divergence_between_our_results_and_those_reported_in_the_llama_3_paper_can_be_attributed_to_a_key_methodological_difference\": \"while the Llama-3 paper leverages a comprehensive dataset including a huge amount of Llama-2 model checkpoints to fit their NLL-to-performance curve, our analysis relies on a more limited set of data points. This difference in training data density significantly impacts the sensitivity of the fitted analytical forms, particularly in regions where data points are sparse.\"}", "{\"comment\": \"# Question\\n1. Thank you for this important question. Our approach introduces several fundamental innovations that differentiate it from classical scaling laws, both in its objectives and methodology.\\n\\nFirst, our work addresses a fundamentally different problem than classical scaling laws. While traditional scaling laws focus on predicting pre-training loss [1,2], our work targets downstream performance prediction - a significantly more challenging objective due to the \\\"emergent abilities\\\" phenomenon. Classical approaches collect N data points by training multiple small language models to convergence, which works well for pre-training loss prediction. However, this methodology breaks down when applied to downstream performance prediction, as models below a certain scale exhibit random performance before reaching emergent points, rendering many collected data points uninformative (as demonstrated in Figure 1 of our paper).\\n\\nThe methodological innovation of our approach lies in its novel two-stage framework, particularly in how it leverages intermediate checkpoints. Classical scaling approaches in our model family would yield only 3-4 viable data points for fitting the analytical function, resulting in poor prediction performance (as evidenced by the FP approach in Section 4, Figure 3). Our FLP framework fundamentally addresses this limitation through its second-stage Loss\\u2192Performance mapping, which effectively utilizes intermediate checkpoints to dramatically improve sample efficiency. This innovation is theoretically grounded in recent empirical findings [3] demonstrating that \\\"compression represents intelligence\\\" - specifically, that language models with comparable pre-training loss (compression capability) tend to exhibit similar performance across diverse tasks, independent of specific training configurations. This consistency validates our methodology of using intermediate checkpoints to map the Loss\\u2192Performance relationship. This increasing sample efficiency ensures significantly better prediction performance, as verified in Section 4, Figure 3, FLP approach.\\n\\nThe theoretical underpinning of our approach represents a fundamental departure from classical scaling laws. While both approaches ultimately map FLOPs to target metrics, they differ substantially in their underlying theoretical frameworks and assumptions:\\nClassical scaling laws posit a direct, monolithic relationship between computational FLOPs and the target metric, typically expressed as a power law function. This end-to-end formulation, while elegant, makes a strong implicit assumption that the relationship between computational resources and model performance can be captured in a single mathematical mapping. However, this assumption becomes problematic when dealing with emergent abilities, where the relationship between resources and performance exhibits qualitative shifts at critical thresholds. In contrast, our approach decomposes the theoretical framework into two distinct stages, each governed by its own analytical form:\\n- The FLOPs\\u2192Loss stage captures how computational resources translate into model compression capability\\n- The Loss\\u2192Performance stage models how this compression capability manifests in downstream performance\\n\\nThe empirical success of this theoretical framework is evidenced by its superior prediction accuracy (Section 4, Figure 3), validating that this decomposed approach better captures the fundamental relationships governing large language model scaling.\\n\\n[1] Scaling Laws for Neural Language Models; Kaplan et al. \\n\\n[2] Training Compute-Optimal Large Language Models; Hoffmann et al. \\n\\n[3] Compression Represents Intelligence Linearly; Yuzhen Huang et al\\n\\n\\n\\n2. Thanks for pointing this out! We wish to emphasize that our analysis specifically focuses on model performance after, rather than around, the emergent point. The linear description was not chosen arbitrarily, but rather emerged from rigorous empirical observations, as demonstrated in Figure 9. While we appreciate the suggestion regarding nonlinear functions, particularly the sigmoid function, our preliminary experiments actually evaluated various analytical forms. The linear function consistently outperformed nonlinear alternatives, with the sigmoid function specifically showing approximately 30% relative prediction error on the ARC benchmark. This empirical evidence strongly supported our decision to adopt the linear formulation as the most appropriate analytical form for describing post-emergence scaling behavior.\"}", "{\"title\": \"Gentle reminder of the discussion period\", \"comment\": \"Hi reviewer,\\n\\nThanks for taking the time to review our paper. It would be great if you can take a look at our responses. Please let us know if you have further questions. Thanks!\"}", "{\"comment\": \"We thank the reviewer for the thoughtful comments. Our responses are as follows:\\n# Weakness\\n1. For the same amount of FLOPs, models with different N (parameters) and D (data) would yield distinct loss. \\n\\nThank you for this insightful comment. We would like to clarify that our methodology aligns with established approaches in scaling law analysis, as first systematically categorized in Kaplan et al. [1]. There are three fundamental approaches to studying scaling behavior: (A) Model-fixed analysis: Using a sufficiently large model while varying training tokens to understand performance scaling with dataset size (B) Data-fixed analysis: Using a sufficiently large dataset while varying model size to understand performance scaling with model capacity (C) Compute-optimal analysis: Maintaining a fixed ratio between training tokens and model size while varying compute (FLOPs), which is the approach we adopt. \\n\\nOur work specifically employs approach (C), which has become a standard methodology in the field [2,3]. In this framework, for any given compute budget, there exists a pre-determined allocation between training tokens and model size (200 in our case). Rather than viewing intermediate checkpoints as suboptimal configurations, they represent legitimate data points along the compute-optimal scaling trajectory. This approach has proven particularly valuable in understanding how language model performance scales with computational resources, as demonstrated by its successful application in recent scaling law studies [2,3].\\n\\n[1] Scaling Laws for Neural Language Models; Kaplan et al. \\n\\n[2] Training Compute-Optimal Large Language Models; Hoffmann et al. \\n\\n[3] Predicting Emergent Abilities with Infinite Resolution Evaluation; Hu et al\\n\\n2. Why is it valid to use checkpoints that have not converged and are not optimized configurations to obtain data points?\\n\\n\\nAs explained in our paper, the use of intermediate checkpoints is methodologically sound for estimating the Loss->Performance relationship, as it aligns with the fundamental principle that compression capability serves as a proxy for model intelligence, thoroughly demonstrated in [1]. Our approach is based on the empirical observation in [1] that language models exhibiting equivalent pre-training loss (i.e., similar compression capabilities) tend to demonstrate comparable performance across diverse tasks, regardless of the specific training configurations (e.g., batch size, learning rate schedules). This consistency in the relationship between compression ability and downstream performance validates our methodology of using intermediate checkpoints to fit the analytical form of the Loss->Performance function.\\n\\n[1] Compression Represents Intelligence Linearly; Yuzhen Huang et al\\n\\n3. Lack of novelty. Both using FLOPs to predict loss and using loss to predict downstream performance have been explored in prior work. \\n\\nWe respectfully disagree with the assessment regarding novelty. While prior work has explored these relationships qualitatively, our contribution is distinct in several key aspects: (1) We establish, for the first time, rigorous analytical formulations that quantitatively link pre-training loss to downstream performance, advancing beyond the qualitative observations in previous studies such as Huang et al. [1]. (2) Our work introduces a systematic and principled framework for forecasting downstream performance in an end-to-end manner (FLOPs->Downstream Performance) in LLMs, representing a significant methodological advancement over existing approaches that only focus on predicting the pre-training loss in LLMs. (3) Uniquely, we extend our analysis to data mixing scenarios, a crucial consideration in modern LLM development that has not been addressed in previous works. Through these contributions, our paper substantially advances the scaling laws analysis for LLM performance prediction.\\n\\n[1] Compression Represents Intelligence Linearly; Yuzhen Huang et al\"}", "{\"comment\": \"4. This claim may be overstated, as 1B models often rely on guesswork for many tasks, undermining the reliability of these results.\\n\\nWe appreciate the reviewer's concern regarding the use of a 1B model in our experiments. However, we respectfully disagree with the assertion that this undermines our results' reliability. In fact, our choice of a 1B model serves as a particularly compelling validation of our approach precisely because such models are typically more sensitive to training data quality compared to larger models.\\n\\nOur experimental design specifically demonstrates that FLP-M derived mixtures consistently outperform baseline mixtures on challenging reasoning benchmarks (ARC and RACE), indicating that our method produces systematically better results than intuition-based approaches. The consistent performance improvements across multiple benchmarks strongly suggest that our method's success is not due to chance but rather reflects the effectiveness of our principled approach to data mixture optimization.\\n\\nWe would welcome specific clarification from the reviewer regarding the concerns about model reliability, as our results demonstrate a clear improvement over baseline approaches\\n\\n\\n\\n# Questions\\n1. Yes. The on-going experiments are conducted internally to verify the effectiveness of the established scaling laws on larger LLMs. \\n2. Please refer to our response to weakness 2. This important methodological choice is thoroughly addressed in our paper, specifically in the third paragraph of the introduction.\"}", "{\"comment\": \"We thank the reviewer for the valuable feedback. Our responses are as follows:\\n\\n# Weakness\\n1. Limited Scale of LM \\n\\nWe appreciate the reviewer's attention to the model scale. We would like to clarify that our experiments actually include models up to 13B parameters in the FLP experiments (Section 4). While we acknowledge the existence of larger models, we focused on 7B and 13B architectures due to (a) the limited resources available, and (b) they represent the most widely deployed model scales in practical applications today (e.g., Llama-2, Mistral, Qwen). These models offer an optimal balance between computational efficiency and performance, making them the preferred choice for real-world implementations. The strong results we achieved at these practical scales demonstrate the immediate applicability and broad impact potential of our approach. Moreover, our methodology is model-agnostic and can be extended to larger LLMs in future work.\\n\\n2. Limited Domains in Data Mixing\\n\\nThanks for pointing this out. Our focused investigation of text and code domains was intentionally designed to provide deep insights into a particularly significant area of practical application. While we acknowledge that expanding to additional domains would offer broader insights, the code-text mixing paradigm represents a crucial use case in modern AI systems, especially for developer tools and programming assistants. The detailed analysis of this specific combination allows us to establish robust foundational principles that can inform future research across other domain combinations.\"}" ] }

[ "Zhaoxu Luo", "Bowen Song", "Liyue Shen" ]

[ "Diffusion models", "satellite imaging" ]

{ "note_id": [ "uwHKDCIGaP", "ud3WVMzRu1", "peeQiKZ5i0", "jVUSPlBMn7", "NVJClHMLkK", "HDw2pqJACK", "4jk5wuhbx7", "0fmt6vQuB9" ], "note_type": [ "official_comment", "comment", "official_comment", "official_review", "official_review", "official_review", "official_review", "official_comment" ], "note_created": [ 1732612929244, 1735966069833, 1732627201832, 1730534826167, 1730203375198, 1730456583560, 1731791393915, 1732614484055 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission11875/Reviewer_DVW4" ], [ "ICLR.cc/2025/Conference/Submission11875/Authors" ], [ "ICLR.cc/2025/Conference/Submission11875/Reviewer_ZHen" ], [ "ICLR.cc/2025/Conference/Submission11875/Reviewer_DVW4" ], [ "ICLR.cc/2025/Conference/Submission11875/Reviewer_Snpo" ], [ "ICLR.cc/2025/Conference/Submission11875/Reviewer_XAtp" ], [ "ICLR.cc/2025/Conference/Submission11875/Reviewer_ZHen" ], [ "ICLR.cc/2025/Conference/Submission11875/Reviewer_XAtp" ] ], "structured_content_str": [ "{\"title\": \"Reback\", \"comment\": \"No feedback from the authors. I will maintain my score\"}", "{\"withdrawal_confirmation\": \"I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors.\", \"comment\": \"We are thankful to the constructive discussion. However, there are issues unsolved in the discussion. Therefore, we decide to withdraw this submission.\"}", "{\"comment\": \"No feedback was received from the authors. I keep the rating unchanged.\"}", "{\"summary\": \"The authors propose a novel diffusion-based fusion algorithm called SatDiffMoE, which accepts an arbitrary number of sequential low-resolution satellite images to generate a high-resolution reconstructed image. During sampling, SatDiffMoE first obtains the center of the intermediate sampling results from different low-resolution conditional images and then weights this center along with the intermediate results to produce the next sampling. Experimental results demonstrate that SatDiffMoE can synthesize high-quality images.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"(1) It is novel that using the timestamp differences as embedding prompts.\\n(2) The sampling strategy effectively combines the intermediate sampling results from different low-resolution conditional images, enabling the fusion of an arbitrary number of low-resolution images.\\n(3) SatDiffMoE can synthesize high-quality images.\", \"weaknesses\": \"(1) The mechanism of DiffusionDrop is unclear and needs further clarification.\\n(2) The content restoration of high-resolution images generated by SatDiffMoE lacks competitiveness.\\n(3) The literature review is insufficient.\\n(4) The comparison methods are either single-image super-resolution approaches or designed for synthetic data, lacking comparisons with multi-image super-resolution methods, especially those tailored for remote sensing imagery.\", \"questions\": \"(1) How does the operator $F$ function in Eq. (2), and why is it necessary to use $F$ to mask out the HR component from the score? Under certain conditions, the score function $s_\\\\theta(\\\\cdot, t)$ can be interpreted as modeling the conditional probability $p(HR_t | HR_{t-1}, LR)$. Thus, once the available score function is obtained from the training dataset, it can gradually sample the HR from Gaussian noise at inference stage.\\n(2) Referring to the background on multi-image super-resolution, the rank 1 method [1] on the PROBA-V dataset (the mentioned HighRes-net ranked 6th) can take an arbitrary number of low-resolution inputs to generate a high-resolution satellite image. However, the authors did not address this point, which is closely related to the contributions of this paper, specifically regarding the concept of \\\"arbitrary number.\\\"\\n(3) Referring to Line 133, \\\"However, few works apply LDMs for image fusion yet\\\", actually, there are lots of DM-based image fusion mothed have been developed last 3 years. \\n(4) As a method to obtain accurate HR images rather than synthetic HR images, I am more focused on image content restoration quality, measured by PSNR and SSIM, rather than the quality of image generation and perceptual quality. If the goal is to generate high-quality satellite data, why not directly train a SAT Stable Diffusion model instead of using a fusion model? Compared to the regression-based model MSRResNet (Table 7), the restoration quality of the proposed method is not particularly outstanding. Additionally, some of the comparative methods are not specifically designed for multi-image super-resolution; rather, they are either single-image super-resolution methods or intended for synthetic data.\\n\\n[1]An, Tai, et al. \\\"TR-MISR: Multiimage super-resolution based on feature fusion with transformers.\\\" _IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing_ 15 (2022): 1373-1388.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper proposes a new fusion super-resolution method based on the diffusion model. Its feature is that it allows any number of low-resolution images to be used as input to reconstruct a high-resolution image. To enhance the mapping difference between the low-resolution images at different timestamps and the corresponding high-resolution images, it constructs the relative time difference and uses it as an additional condition input. Another contribution is that during the inference phase, the trajectories corresponding to each low-resolution image can be randomly selected for fusion and finally output to a single high-resolution image. Experiments show the advantages of the proposed method.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": [\"This method is very flexible and can accept sequences with different numbers of low-resolution images for fusion. This advantage is attributed to its specific fusion strategy for sampling trajectories.\", \"The definition and conditional input of the relative time difference is also interesting, which can help the network capture the time-aware mapping distribution of low-resolution to high-resolution images.\", \"Experimental results show that the reconstruction results are relatively better than competitors in terms of rationality and naturalness.\"], \"weaknesses\": [\"A key operation is F, which masks out the HR component from the score function. However, this paper does not explain how F was designed. This makes the description of the relevant methodology unclear.\", \"There is also some disagreement about the definition of the relative time difference. In Figure 1, time t4 corresponds to the high-resolution image, and dti represents the difference between the low-resolution image at time ti and the high-resolution image. However, the definition of dt4 is obviously different from the others, which is not rigorous in expression.\", \"I am concerned about the significance of this technology. On the one hand, images captured at different times may show changes in surface cover, so is it appropriate to use a high-resolution image at one of these moments as a reference? On the other hand, the experiments also confirmed my concern that the reconstructed results of the proposed method, although better than those of other methods, still show considerable differences with GT. This difference is not only reflected in the intensity distribution, but also includes changes in some ground objects. This makes me concerned about whether the reconstructed high-resolution results are credible and applicable.\", \"Sequence fusion with multiple low-resolution images seems to only appear in ablation studies. Is the comparative experiment just a single image super-resolution?\", \"Could this method be extended to consider the possibility of parallax between low-resolution images? In reality, images captured at different times theoretically have geometric differences\", \"There are some typos. For example, the last line of page 4 is missing a period.\"], \"questions\": \"See weaknesses.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper introduces a super-resolution latent diffusion model based on finetuning SD1.2 model, to fuse and SR and remote sensing images. To adapt fusing different LRs at different times, the authors proposed DiffusionDrop and optimized-based diffusion to fuse diffusion trajectories. Results on some remote sensing image SR datasets show its effectiveness.\", \"soundness\": \"2\", \"presentation\": \"1\", \"contribution\": \"2\", \"strengths\": \"1. Propose DiffusionDrop and use previous $z_0$ optimization to fuse LRs to HR.\\n\\n2. Good results on different satellite image dataset.\", \"weaknesses\": \"Some typos:\\n\\n1. At line 169 in Algo. 1, the comment ''Compute the score'' actually should be ''noise'', since you're using SD1.2 model, which is trained on noise-prediction objective.\\n2. At line 103, there should exist an semicolon.\\n3. At line 166, there should exist the encoder to encode the LR images.\", \"weakness\": \"1. Some key components like $F$ in Eq. (2), finding the center ($\\\\bar z$) is too vague, hindering reader's understanding.\\n\\n2. In inference, the authors claim that finding the center ($\\\\bar z_0$) of $z_i$-s are based on Chung et al (2024), but actually, in this refered paper, when optimzing the $z_t$ (or says diffusion trajectory), CG steps should be taken to optimize with the help of the degraded operator ($\\\\mathbf A$). It seems that you did not have the operator $\\\\mathbf A$ and did not include the description of how to take the $\\\\arg \\\\min$.\\n\\n3. According to your description in inference stage, finding the center of predicted $\\\\hat z_0$ should include all LRs which is $LR_i, i\\\\in \\\\{0,\\\\dots, N-1\\\\}$ (if you have $N$ LRs when inference). This means in Algo. 1, there should be **two** for-loops, one for diffusion timesteps and one for $N$ LR samples. Did I misunderstand?\\n\\n4. In the experiments, the authors compared with ControlNet, but only training it with only one LR and HR pair. This comparison is unfair, since the proposed method uses multiple LRs.\\n\\n5. As shown in your Eq. (2), inputs should be encoded LRs and HR, but in Fig. 2, there is only LR and do not have $F$.\\n\\n---\\n\\nThis paper propose some techniques to finetune the pretrained SD model to adapt on satellite SR task with various number of LR images, but the presentation is too poor to understand its contributions. \\n\\nI believe the proposed method's SR performance can beat previous methods, such as previous regressive models and diffusion models, but the presentation and writing is still an important part. The writting makes this paper can not be accepted by ICLR.\\n\\nNow, I rate this paper to 5 (marginally below the acceptance threshold).\", \"questions\": \"1. What's the $F$ in Eq. (2), I did not find its implementaion in the rest of paper. Only description like ''mask out the HR component'' is too vague.\\n\\n2. Please provide 1) the details of used datasets, including train/validation/test ratios, total number of images etc; 2) the computation resources to finetuning the SD1.2 model.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The paper llooks sound based on the idea \\\"Super-resolution with Latent Diffusion Models for Satellite Image\\\" This method introduces SatDiffMoE, a novel diffusion-based algorithm that fuses sequential low-resolution images into high-resolution outputs using complementary time-point data. The method demonstrates flexibility, superior performance, and computational efficiency on super-resolution tasks across various datasets.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"1. This method is highly flexible, capable of handling sequences with varying numbers of low-resolution images for fusion\\n2. Its adaptability stems from a specific fusion strategy for sampling trajectories, enhanced by the novel use of timestamp differences as an embedding prompt to guide the fusion process effectively.\\n3. This strategy harmonizes intermediate outputs across varying low-resolution inputs, enabling a flexible and efficient fusion process independent of the image count.\", \"weaknesses\": \"1. Limited comparison with SOTA methods\\n2 Literature review is not strong, problem statement\\n3. The use of diffusion models introduces a dependency on the quality of the generative priors\\n4. In real-world conditions (e.g., different climates or satellite systems), the generalizability of the results might be a concern.\", \"questions\": \"1. Why the proposed model was designed for satellite image , why not for real image datasets like Urban100\\n2. Complexity of the season still huge, if the comparison will be made with weight methods\\n\\n2. Address the weakness motioned in section.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"No feedback. I keep the rating unchanged.\"}" ] }

[ "Lynn Chua", "Badih Ghazi", "Yangsibo Huang", "Pritish Kamath", "Ravi Kumar", "Pasin Manurangsi", "Amer Sinha", "Chulin Xie", "Chiyuan Zhang" ]

[ "Large Language Models", "Multilingual", "Crosslingual Knowledge Barrier" ]

{ "note_id": [ "sSg2nZNTH1", "rvfs0MPjLG", "qSYLLHD5bE", "prrwyh4V8C", "nxGytDFmi8", "fLwBecy6A5", "eH0bPmpz38", "cHTs74hiit", "c1S9SwziHL", "ZMYkXKqyTd", "Xw2WrgWTyV", "XkhVQWJmYs", "U4EGao9piW", "TlKQSvl4YL", "OIE1WPY2hK", "NJ6m18tisV", "GECTRWXCEm", "CJzoRkry3R", "5N7YwM8fPu", "5EFmPQfA4A", "2axcUguOGo", "1LL3miG5VG", "0gtmDSRAfT" ], "note_type": [ "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "meta_review", "official_review", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "decision", "official_comment", "official_comment" ], "note_created": [ 1732338969669, 1732338582705, 1733176153649, 1733168898726, 1732668852873, 1732338423164, 1732668831737, 1732732019641, 1732338010728, 1732337836107, 1733214065806, 1732702059922, 1730645160183, 1734602637528, 1730716812940, 1732336954897, 1732881678643, 1732338334409, 1730822792346, 1732337606037, 1737523836822, 1733170027870, 1733214200935 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission7405/Authors" ], [ "ICLR.cc/2025/Conference/Submission7405/Authors" ], [ "ICLR.cc/2025/Conference/Submission7405/Reviewer_uRPF" ], [ "ICLR.cc/2025/Conference/Submission7405/Authors" ], [ "ICLR.cc/2025/Conference/Submission7405/Authors" ], [ "ICLR.cc/2025/Conference/Submission7405/Authors" ], [ "ICLR.cc/2025/Conference/Submission7405/Authors" ], [ "ICLR.cc/2025/Conference/Submission7405/Authors" ], [ "ICLR.cc/2025/Conference/Submission7405/Authors" ], [ "ICLR.cc/2025/Conference/Submission7405/Authors" ], [ "ICLR.cc/2025/Conference/Submission7405/Authors" ], [ "ICLR.cc/2025/Conference/Submission7405/Reviewer_xBQr" ], [ "ICLR.cc/2025/Conference/Submission7405/Reviewer_uRPF" ], [ "ICLR.cc/2025/Conference/Submission7405/Area_Chair_RsRa" ], [ "ICLR.cc/2025/Conference/Submission7405/Reviewer_2t3Z" ], [ "ICLR.cc/2025/Conference/Submission7405/Authors" ], [ "ICLR.cc/2025/Conference/Submission7405/Reviewer_2t3Z" ], [ "ICLR.cc/2025/Conference/Submission7405/Authors" ], [ "ICLR.cc/2025/Conference/Submission7405/Reviewer_xBQr" ], [ "ICLR.cc/2025/Conference/Submission7405/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission7405/Authors" ], [ "ICLR.cc/2025/Conference/Submission7405/Authors" ] ], "structured_content_str": [ "{\"title\": \"Revision Summary\", \"comment\": [\"We sincerely thank all reviewers for their constructive feedback and suggestions, which are very helpful to us. We are encouraged that the reviewers found our work (1) provide interesting and useful insights on the cross-lingual ability of LLMs (Reviewer xBQr, 2t3Z, uRPF), (2) the results are solid and promising (Reviewer xBQr, 2t3Z, uRPF) and (3) the writing is clear (Reviewer xBQr, 2t3Z).\", \"Following the reviewers\\u2019 suggestions, we added more experiments/discussions, and we addressed the questions in the response to each reviewer. Below is a summary of our new experimental results in the revised PDF:\", \"**Figure 4, Figure 5, Figure 7**: We evaluated a total of 15 models on MMLU variants and the Harry Potter Quiz, including those explicitly stated to be multilingual, such as Qwen2.5-7B, Llama-3.1-8B, Mixtral-8x7B-v0.1, aya-expanse-8B, aya-23-8B, Mistral-Nemo-Base-2407, and the Tower series models. (Reviewer xBQr, 2t3Z)\", \"**Figure 5, Figure 7**: We expanded the evaluation to 16 languages on MMLU variants and the Harry Potter Quiz. These include low-resource languages such as Malay (ms), Danish (da), Finnish (fi), Norwegian (no), Bengali (bn), and Amharic (am), as well as languages with token distributions that differ significantly from English, such as Russian (ru), Chinese (zh), Hebrew (he), Arabic (ar), and Hindi (hi). (Reviewer xBQr, 2t3Z)\", \"**Figures 11 and 12**: We evaluated mixed-language fine-tuned models on additional languages that were not seen during the fine-tuning stage, observing enhanced cross-lingual performance. (Reviewer xBQr, 2t3Z)\", \"**Appendix Figures 14 and 15**: We provided detailed monolingual and cross-lingual evaluation results for each domain of MMLU, including STEM, humanities, social sciences, and others.\", \"Please also let us know if there are other questions, and we look forward to the discussion with the reviewers to further improve our paper. Thank you!\"]}", "{\"title\": \"Response to Reviewer uRPF (Part 3)\", \"comment\": \"> Q8 For the rest of the paper, the authors evaluate well-established methods to improve the multilinguality score. Their proposal to construct a fine-tuning dataset comprising examples from multiple languages, particularly at the sentence level and achieve better scores compared to English fine-tuning. However, I believe this outcome is completely expected and does not represent a significant contribution.\\n\\nThank you for the comment. While the outcome of mixed-language fine-tuning improving multilinguality scores may seem intuitive, we study their effectiveness on different mixing units (words, sentences, documents). We also would like to emphasize the broader contributions and novelty of our work beyond this specific result:\\n\\n1. **New Problem Formulation**: We explicitly formulate the problem of the crosslingual knowledge barrier in pretrained/finetuned language models, focusing on their ability to retrieve and utilize parametric knowledge across languages. This goes beyond surface-level multilinguality assessments and addresses deeper challenges in crosslingual knowledge transfer.\\n\\n\\n2. **Proposed Challenging Scenarios**: We introduce novel crosslingual QA scenarios, including multi-choice QA tasks with innovative compositions of multiple languages, to specifically evaluate the models\\u2019 implicit crosslingual knowledge capabilities in general (MMLU benchmark) and domain-specific (Harry Potter quiz) contexts.\\n\\n\\n3. **Revealing Crosslingual Knowledge Barriers**: We evaluate a total of 15 multilingual LLMs and 16 languages (We have added 11 languages and 9 LLMs during the rebuttal phrase following the reviewers\\u2019 suggestions in the revised PDF). Our results show that while models may demonstrate promising surface-level cross-lingual capabilities (e.g., translation and embedding space analyses), they struggle with deeper cross-lingual knowledge transfer. This finding reveals a previously underexplored limitation in existing multilingual LLMs.\\n\\n\\n4. **Limited Impact of Inference-Time Mitigations**: Our study highlights that simple inference-time mitigation methods provide limited/no improvement, further emphasizing the challenge posed by the crosslingual knowledge barrier.\\n\\n\\n5. **Generalization of Mixed-Language Fine-Tuning**: We proposed a mixed-language fine-tuning method across different mixing units (words, sentences, documents). The results on the MMLU variant and Harry Potter quiz show that our proposed method not only improves crosslingual performance for the languages included during fine-tuning but also benefits languages that were unseen during fine-tuning, including both low-resource languages and languages with different token distributions. This observation highlights the broader applicability of our approach and its impact on underrepresented languages.\\n\\n\\nWe believe these contributions collectively advance the understanding of cross-lingual knowledge transfer in LLMs and provide a meaningful foundation for future research in this area.\"}", "{\"comment\": \"Thank you for your responses. I still do not think this paper has enough contributions. The major drawbacks for me are that the embeddings are not calculated in an appropriate way for decode-only models, the experiments are not controlled enough to necessarily indicate a knowledge transfer problem, and all the methods used to improve the multilinguality score are well-established and not a contribution. I maintain my score (3).\"}", "{\"comment\": \"Thanks for the comments! Please find our response to your remaining questions.\\n\\n> While the general assertion that a multilingual human could transfer relevant knowledge is plausible, there are notable exceptions. Some translations differ so significantly that even speakers proficient in both languages may find it challenging to identify connections. \\n\\nThank you for the valuable comment and for conducting the Russian translation experiment with native speakers. We agree that translating the domain-specific knowledge is a very challenging and complex task, as the reviewer highlighted. Our dataset\\u2019s focus on multiple-choice QA tasks might help mitigate the noise arising from translation variability by narrowing down the search space for the correct answer to a limited set of options. \\n\\nIn addition to discussing the limitations of translation quality from Google Translate in Appendix A, we will clarify it in Section 3.2 when introducing the dataset. We will also recruit native speakers to perform translation quality checks in the revision.\\n\\n> The inability to measure training data leakage implies that it remains unclear whether the observed performance in certain languages results from genuine crosslingual transfer or simply due to English data leakage during pretraining. This lack of clarity significantly impacts the interpretability of your results.\\n\\nThank you for your comment. We would like to highlight the value of the Harry Potter (HP) dataset in this context.\\n- Given that the original books are written in English, the prevalence of HP content in English within the pretraining data is expected. Rather than being a concern, this provides an opportunity to study cross-lingual transfer for **off-the-shelf** models. Specifically, it allows us to examine whether models, when prompted in non-English languages, can link the question to their inherent knowledge (acquired predominantly in English) to answer questions effectively in other languages. \\n- HP also represents a widely known and well-defined knowledge domain, with practical significance\\u2014for example, users may ask LLMs questions about HP in multiple languages to understand the book better. This makes it important and relevant to study this dataset.\\n\\n> Consider removing the HP Quiz entirely and replacing it with content that has not been leaked on the internet, such as local company guidelines or university campus rules, as previously suggested in Q7. Re-running your experiments with such data would likely lead to a more robust experimental design, thereby enhancing the credibility of the results\\n\\nWe appreciate the reviewer\\u2019s suggestion to consider using local company guidelines or university campus rules, and we are committed to adding a new dataset like that in our revision. However, during our process of designing the curation of such datasets, we realized that if such content had not been included on the internet, it would not provide meaningful insights into the knowledge-intensive reasoning abilities of off-the-shelf models, as this type of content would not be included in their pretraining data. Instead, this content is more suited to fine-tuning experiments, where models can be trained on domain-specific datasets to adapt to such specialized knowledge.\\n\\nOn the other hand, another important consideration is the extent to which such knowledge is truly domain-specific. Local company or school rules are less domain-specific compared to the fictional world of Harry Potter. These rules are often not unique to a particular organization; many of the rules may be general and shared across multiple institutions or companies across the world speaking different languages. As a result, even if we use such knowledge for fine-tuning, it may not provide clean insights for domain-specific knowledge-reasoning, and the models may have known such common rules in multiple languages.\\n\\nOur current focus on widely recognized Harry Potter content provides a benchmark for evaluating cross-lingual transfer and knowledge-intensive reasoning in off-the-shelf models and fine-tuned models. Therefore, we respectively disagree that the HP dataset is not appropriate. While we are happy to add additional dataset, we were unable to complete this during the rebuttal timeline. Nonetheless, we believe that additional datasets would only further support our observations and would not fundamentally alter our conclusions. The current results are both sufficient and significant as evidence of the cross-lingual knowledge barrier as shown in both general (MMLU) and domain-specific (HP) knowledge.\\n\\nWe sincerely thank the reviewer for the time and insightful comments, which have helped us improve the manuscript.\\nPlease also let us know if there are other questions, and we look forward to the discussion with the reviewer to further enhance our work. Thank you!\"}", "{\"title\": \"Follow-up\", \"comment\": \"Dear Reviewer,\\n\\nThank you once again for your detailed comments and suggestions. As the revision period is approaching its end, we would greatly appreciate your feedback on whether our responses and revision have addressed your concerns. We are also happy to engage in further discussions if needed.\"}", "{\"title\": \"Response to Reviewer uRPF (Part 2)\", \"comment\": \"> Q6: L199-200: There is no experiment/evaluation/audit demonstrating how valid these translations are.\\n\\nThank you for the comment. We utilized the commercial system Google Translate API to perform the translations, which is regarded in the literature as the strongest baseline for machine translation tasks [1]. \\n\\nHowever, we acknowledge that even state-of-the-art translation tools may not perfectly capture nuanced or domain-specific meanings in all languages. As the authors are not fluent in all the languages evaluated, it is challenging to independently verify the accuracy of the translations across different foreign languages. We recognize this limitation and have explicitly noted it in the limitations section (Appendix A) of the revised PDF.\", \"reference\": \"- [1] Wenhao Zhu, Hongyi Liu, Qingxiu Dong, Jingjing Xu, Shujian Huang, Lingpeng Kong, Jiajun Chen, and Lei Li. Multilingual machine translation with large language models: Empirical results and analysis. In NAACL, 2024.\\n\\n\\n> Q7: L247: Conclusion 1 does not necessarily indicate a knowledge transfer problem. In English-centric LLMs, it is expected that performance will be better in English. Simply stating that English performs better in an English-centric LLM does not imply that knowledge cannot be transferred; rather, it suggests that performance in the secondary language may be low, even in language understanding tasks. The authors selected languages where LLMs typically perform better (but is it on the level of English?), so it is not surprising that they cannot fully grasp the knowledge of English. If the authors had chosen low-resource languages, we might observe even lower performance, which could indicate that the LLM lacks substantial knowledge of those languages to begin with than other knowledges.\\n\\nThank you for the comment. We would like to clarify that **Conclusion 1** is based on **traditional monolingual** evaluations, not our proposed cross-lingual evaluations. As noted in Section 3.1 (lines 174-184), *monolingual evaluation is inadequate for assessing crosslingual abilities* because it does not isolate the process of transferring knowledge across languages. Recognizing these limitations and confounding factors in previous multilingual evaluations, we designed controlled setups to study cross-lingual transfer.\\nSpecifically, we proposed the Mixed-Language Evaluation strategies (lines 185-197), which directly invoke cross-lingual capabilities by incorporating novel compositions of multiple languages in multi-choice QA formats. The conclusions derived from **proposed cross-lingual evaluation** strategies are presented in **Conclusions 2\\u20134** and **Evaluation on 16 languages** (lines 248\\u2013297). We apologize for any confusion caused by the presentation of results.\\n\\nMoreover, we investigate actual cross-lingual knowledge transfer via a controlled experimental setup in the Harry Potter Quiz dataset (Section 3.2). We fine-tune LLMs in in-domain English content and observe improved performance in other languages, indicating that the model is capable of transferring knowledge across languages. However, the gap remains between English and other languages, indicating the cross-lingual knowledge barriers: LLMs struggle to fully utilize the parametric knowledge acquired during English fine-tuning to answer related questions in other languages. Therefore, in Section 4.2, we proposed mixed-language fine-tuning methods, to achieve improved cross-lingual performance, which further supports that such knowledge gaps can be narrowed with more effective cross-lingual knowledge transfer methods.\\n\\nWe hope this clarification addresses the reviewer's concerns.\"}", "{\"title\": \"Follow-up\", \"comment\": \"Dear Reviewer,\\n\\nThank you once again for your detailed comments and suggestions. As the revision period is approaching its end, we would greatly appreciate your feedback on whether our responses and revision have addressed your concerns. We are also happy to engage in further discussions if needed.\"}", "{\"title\": \"Response to xBQr\", \"comment\": \"Thank you again for your thoughtful comments and positive feedback. Your support is vital to us.\"}", "{\"title\": \"Response to Reviewer 2t3Z (Part 3)\", \"comment\": \"> Q6: It would be very beneficial if the authors of the paper would use different languages, such as Chinese, Hebrew, Russian or Hindi, since these languages have a much lower amount of shared tokens with English, which is selected as the most highly resourced language in the training mix of all selected models. This would give us insight into how the applied methodology translates into languages, which the model has a harder time training due to lesser overlap between dictionaries.\\n\\nThank you for the thoughtful comment. \\nFollowing the reviewer's suggestion, we have expanded our evaluation to *include a total of 16 languages* and 15 multilingual LLMs to provide broader insights into our study. \\nSpecifically, in addition to the 2 germanic languages (English and German) and three romance languages (French, Italian, Spanish) we originally evaluated, we further consider \\n- *Low-resource languages*: Malay (ms), Danish (da), Finnish (fi), Norwegian (no), Bengali (bn), Amharic (am);\\n- *Languages with token distributions significantly different from English*: Russian (ru), Chinese (zn), Hebrew (he), Arbic (ar) and Hindi (hi).\", \"conclusions\": \"- The results on MMLU (Figure 4 and Figure 5) and the Harry Potter quiz (Figure 7 ) show that cross-lingual knowledge barriers hold for those additional models and languages, which demonstrates the universality of our findings. \\n- We also highlight that in Section 4.2, we evaluated our mixed-language fine-tuned LLMs on languages that were not used during the finetuning stage. Our results in Figures 10 and 11 demonstrate that the model fine-tuned using our method on high-resource languages { English (en), French (fr), German (de), Spanish (es), and Italian (it)} can improve crosslingual performance on Harry potter quiz and MMLU for low-resource languages and languages that are rather different from English. These results provide encouraging evidence for the generalizability of our fine-tuning approach.\\n\\n\\n\\n> Q7: Perhaps designing another evaluation dataset for domain specific knowledge, based on something like some local company guidelines or university campus rules and translating it into different languages would make a much stronger point to support the claims, since it would not exhibit translation artifacts\\n\\nThanks for the great suggestion! We greatly appreciate the idea of designing a domain-specific dataset based on local company guidelines or university campus rules. We also hope our response to Q3 has addressed the reviewer\\u2019s concern regarding the existing Harry Potter dataset, particularly as we have used in-domain fine-tuning to explicitly inject relevant English knowledge and evaluate cross-lingual knowledge barriers.\\n\\nDue to the limited time frame of the rebuttal phase and the experimental workload involved in evaluating additional models and languages, we have not yet completed the creation and evaluation of such a new dataset. We are committed to working on this new dataset based on your suggestion and will include it in our future revisions to further support our findings.\"}", "{\"title\": \"Response to Reviewer 2t3Z (Part 2)\", \"comment\": \"> Q3 All of the selected models almost certainly have seen the original books, dumps from HP Wiki or Wikipedia and blog posts about Harry Potter in non-English languages, so it may have leaked into the training data for all of the languages.\\n\\nThank you for the comment. Unfortunately, we cannot verify the extent to which the original books, Harry Potter Wiki dumps, or related content in non-English languages might have been included in the pretraining data for the selected models. This could vary significantly across languages and models. However, we note that the mere presence of such content in the training data does not guarantee that the model has learned this knowledge in a usable form, particularly across different languages. The ability to effectively utilize this knowledge to solve relevant question-answering tasks involves more than just observing those texts during training; it requires the model to generalize and reason based on its learned representations. \\n\\nMoreover, in Figure 6, we conducted a controlled experiment where we *explicitly* finetuned the models on in-domain content presented in English, where models still consistently perform better at answering questions in English than in other languages. This suggests that LLMs struggle to fully utilize the parametric knowledge acquired during English fine-tuning to answer related questions in other languages and close the performance gap, indicating the presence of cross-lingual knowledge barriers.\\n\\n> Q4: It is common knowledge that both instruction and preference tuning causes catastrophic forgetting, which leads to models being better at instruction following, safety and answer formatting, but worse at parametric knowledge. From the list of the described models we can see that only Llama-2, Llama-3, Mixtral-8x7b, Zamba-7b and Mistral-7b models are available as base models, with GPT-3.5, GPT-4, Aya-23 being instruction tuned models. It is not explicitly stated, which versions of the models are used \\u2013 pretrained or finetuned, Thus, some clarification is required, since it directly influences the amount of multilingual capability of the models. If these are instruction tuned models, further finetune of them on Harry Potter dataset could lead to catastrophic forgetting, thus perturbating the models\\u2019 scores and if these are base models, direct comparison between Aya-23 and llamas, for instance, presents a potential inconsistency in methodology.\\n\\nThank you for your insightful comments. Following your suggestions, we have now clarified the used model versions in the updated figures of the revised PDF. \\n\\nAs per the reviewer's suggestion, in our experiments, we fine-tuned the base models on Harry Potter-related content, avoiding instruction-tuned models, to mitigate the risk of catastrophic forgetting. Our primary focus in Section 4.2 was to compare the performance of the same model before and after fine-tuning to evaluate the effectiveness of our proposed mitigation method. This intra-model comparison avoids direct comparisons between different models (e.g., Aya-23 and Llama), ensuring consistency in the methodology and the conclusions drawn in Section 4.2.\\n\\n> Q5: The phenomenon of crosslingual knowledge barrier (e.g. limited multilinguality capabilities of the model) is well known, so the main contribution of the paper lacks originality.\\n\\nThank you for the comment. We clarify that the \\\"limited multilinguality capabilities\\\" often discussed in the literature are distinct from the \\\"crosslingual knowledge barrier\\\" phenomenon we study in this paper. While prior work primarily focuses on **explicit** cross-lingual tasks (e.g., machine translation) or **monolingua**l evaluation of benchmarks in each language, as seen in those model cards, our work introduces a *new challenge regarding cross-lingual knowledge reasoning*.\\n\\nSpecifically, we evaluate models' capacity to **implicitly** retrieve and utilize parametric knowledge stored in their weights across languages to solve QA tasks \\u2013 for both general knowledge (MMLU) and domain-specific knowledge (Harry Potter quiz). This is a different and underexplored aspect of **cross-lingual** performance that goes beyond tasks requiring direct language-to-language mappings (where the source text is provided in the context, so it is already \\\"grounded\\\") or simple monolingual evaluations (which do not necessarily invoke cross-lingual knowledge reasoning).\\n\\nThank you again for raising this point and we hope our clarifications address the reviewer\\u2019s concerns.\\n\\n\\n.\"}", "{\"title\": \"Response to Reviewer uRPF (part 1)\", \"comment\": \"Thank you for your feedback. We respectfully believe that the questions raised may stem from misunderstandings that we are happy to clarify further. Below, we address each of your comments in detail:\\n\\n\\n> embeddings are not calculated in an appropriate way for decode-only models,\\n\\nWe respectfully disagree with this comment. Mean pooling of last hidden states is a **standard approach** to calculate sentence embedding specifically for **decoder-only** models. This method has been employed in widely-adopted open-source libraries, such as llama.cpp (a inference service of Llama-series models with over 68,000 GitHub stars) [1], and has been used in recent research publications [2,3]. \\n\\nWe are open to exploring alternative embedding methods - If you could suggest specific techniques, we are happy to also report them in Section 2. \\n\\nAdditionally, we would like to note that the paper [4], which the reviewer cited in the original **Q3**, is not published in a peer-reviewed venue, which raises questions about its appropriateness.\\n\\n\\nReference \\n- [1] https://github.com/ggerganov/llama.cpp/issues/6754\\n- [2] LLMEmbed: Rethinking Lightweight LLM's Genuine Function in Text Classification https://arxiv.org/abs/2406.03725, ACL 2024\\n- [3] LLM2Vec: Large Language Models Are Secretly Powerful Text Encoders https://arxiv.org/abs/2404.05961, COLM 2024\\n- [4] MEXA: Multilingual Evaluation of English-Centric LLMs via Cross-Lingual Alignment https://arxiv.org/pdf/2410.05873 \\n\\n\\n\\n> the experiments are not controlled enough to necessarily indicate a knowledge transfer problem\\n\\nCould the reviewer please clarify which specific experiments are considered \\\"not controlled enough\\\"? This clarification would help us better understand and address the concern.\\n\\nIf the reviewer is referring to **Conclusion 1** as mentioned in the original **Q4**, we would like to reiterate that this conclusion is based on traditional monolingual evaluations, which were **not** the primary focus of our study. Instead, our key findings are derived from the proposed cross-lingual evaluation strategies, which are presented separately in **Conclusions 2\\u20134 (lines 248\\u2013297) and are supported by results across 16 languages**, which provide evidence of a **crosslingual knowledge transfer problem**. \\n\\nSpecifically, we adopt an inherent crosslingual interaction approach by introducing mixed-language multiple-choice question formats. These formats are purposefully designed as **novel cross-lingual compositions** that are unlikely to have been encountered during pretraining. This ensures that our experiments evaluate genuine crosslingual knowledge transfer rather than reliance on each single language.\\n\\nIf the reviewer identifies particular experiments that are \\\"not controlled enough\\\", we would greatly appreciate a detailed explanation of the perceived shortcomings.\"}", "{\"comment\": \"Thank you for your clarifications and the inclusion of the additional models. Given the limited time, this is a non-trivial effort and much appreciated.\"}", "{\"summary\": \"The paper claims to investigate the cross-lingual knowledge barrier for four high-resource non-English languages by examining benchmark score variability under different translation scenarios for six LLMs. However, it suffers from several weaknesses, including:\\n\\n- A lack of detail regarding how sentence embeddings are obtained to demonstrate explicit cross-lingual capabilities. And no justification for the chosen probability settings.\\n- Uncontrolled experiments that do not sufficiently prove that the results are due to the cross-lingual knowledge barrier rather than the performance of the languages in non-knowledge related tasks.\\n\\nIn the rest of the paper, the authors evaluate well-established methods to improve the multilinguality score. Their proposal to construct a fine-tuning dataset comprising examples from multiple languages, particularly at the sentence level, aims to achieve better scores compared to English fine-tuning. However, I believe this outcome is completely expected and does not represent a significant contribution.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"1\", \"strengths\": [\"The paper demonstrates that a cross-lingual knowledge barrier exists for four non-English languages through experiments evaluating how much benchmark scores vary when the benchmarks are translated under different scenarios.\", \"The authors assess well-established methods to improve the multilinguality score and propose constructing a fine-tuning dataset that comprises examples from multiple languages, particularly at the sentence level, to achieve better scores in multilinguality.\"], \"weaknesses\": [\"L26-36: The first figure presents a poor example. Why couldn't it simply be the same question in both English and French? Additionally, why is there a need to create a knowledge that does not exist, so it can not be tested out as an example?\", \"L104: The authors state, \\u201csuch presence, if it exists at all,\\u201d regarding parallel sentences on the web. It would be interesting to know how papers that mine billions of parallel texts from the web feel about this statement (e.g., https://arxiv.org/abs/1911.04944).\", \"L137: The paper mentions obtaining sentence embeddings from the LLM; however, it does not detail how this is done. Since these are decoder-only models with causal attention, how are the embeddings of the tokens aggregated? Is it the last token, or are they weighted? (see section 3.1 of https://arxiv.org/pdf/2410.05873 for a background on the different techniques)\", \"L134-142: There are many variables concerning probabilities, yet there is no explanation for why these specific values were chosen.\", \"L158-159: The statement \\u201cThis underscores the explicit cross-lingual capabilities of multilingual LLMs\\u201d lacks clarity regarding the rationale and conclusions drawn. With a lower probability (0.16 vs. 0.8), tokens were randomly selected from the vocabulary. Now, the distribution of these two (cosine similarities between the original and mixed translated sentence embeddings versus the original and random token-replaced sentence embeddings) varies significantly. Many variables are involved here: do the random tokens from the vocabulary come only from all writing systems or just Latin? How are the embeddings computed, and why were these probabilities chosen? Although I acknowledge the existence of cross-lingual capabilities, I am not convinced by the authors' experiment.\", \"L199-200: There is no experiment/evaluation/audit demonstrating how valid these translations are.\", \"L247: Conclusion 1 does not necessarily indicate a knowledge transfer problem. In English-centric LLMs, it is expected that performance will be better in English. Simply stating that English performs better in an English-centric LLM does not imply that knowledge cannot be transferred; rather, it suggests that performance in the secondary language may be low, even in language understanding tasks. The authors selected languages where LLMs typically perform better (but is it on the level of English?), so it is not surprising that they cannot fully grasp the knowledge of English. If the authors had chosen low-resource languages, we might observe even lower performance, which could indicate that the LLM lacks substantial knowledge of those languages to begin with than other knowledges.\", \"For the rest of the paper, the authors evaluate well-established methods to improve the multilinguality score. Their proposal to construct a fine-tuning dataset comprising examples from multiple languages, particularly at the sentence level and achieve better scores compared to English fine-tuning. However, I believe this outcome is completely expected and does not represent a significant contribution.\"], \"questions\": \"See weaknesses part especially L137, L134-142, L158-159, L199-200.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"metareview\": \"The paper investigates the cross-lingual knowledge barrier for large language models (LLMs) by evaluating their performance on mixed-language tasks, general knowledge benchmarks (MMLU), and domain-specific evaluations (Harry Potter quiz). The paper observes that while LLMs demonstrate surface-level cross-lingual capabilities, they struggle with deeper cross-lingual transfer and propose fine-tuning on mixed-language datasets as a mitigation strategy. While the study raises an important topic of cross-lingual knowledge transfer, there are significant concerns regarding the methodology, experimental design, and overall contribution.\\n\\nThe strengths of the paper include its attempt to examine cross-lingual reasoning systematically and its clear writing. The expanded evaluation to include more languages and multilingual models is commendable. However, the weaknesses of the paper outweigh these strengths. First, the paper does not convincingly isolate the cross-lingual barrier phenomenon due to insufficiently controlled experiments and unclear disentanglement of confounding factors such as model multilinguality, token overlap, and pretraining data leakage. Many of the selected models, such as the Llama series, are not explicitly multilingual, raising concerns about whether their performance can be generalized. Second, the domain-specific evaluation setup using the Harry Potter dataset is problematic. Translation artifacts, inconsistent naming conventions, and the likelihood of pretraining data contamination undermine the reliability of the findings. Some reviewers argue that results from this dataset may reflect memorization rather than cross-lingual reasoning. Third, the proposed fine-tuning method to address the barrier lacks novelty. Techniques like mixed-language fine-tuning or code-switching have been previously explored, particularly for encoder-based models, and the improvements reported here are therefore incremental and expected. The paper also lacks discussion on the trade-offs of such fine-tuning. Finally, some aspects of the experimental presentation remains unclear. \\n\\nDespite the authors' efforts to address these issues in the rebuttal, fundamental concerns about the experimental rigor and the novelty of the contribution persist. While the topic is relevant, the current version of the paper does not provide sufficiently strong evidence or a significant advancement in understanding or mitigating cross-lingual barriers in LLMs. For these reasons, I recommend rejection.\", \"additional_comments_on_reviewer_discussion\": \"During the rebuttal period, the discussion primarily centred around the experimental design, dataset reliability, and the novelty of the contributions. The reviewers raised concerns regarding these aspects, and while the authors attempted to address the issues, significant doubts remained.\", \"experimental_design_and_model_selection\": \"Reviewers pointed out that several models evaluated, particularly the LLaMA series, are not explicitly multilingual, and their poor performance should not be generalized to multilingual LLMs. They also questioned the experimental controls, particularly whether the observed performance drop was due to a true \\\"cross-lingual barrier\\\" or model limitations. The authors expanded their evaluation to include 15 multilingual models and 16 languages, including low-resource ones, to strengthen their claims. However, concerns remained about the entanglement of confounding factors such as token overlap and pretraining biases.\", \"reliability_of_the_harry_potter_dataset\": \"Reviewer 2t3Z criticized the domain-specific evaluation based on the Harry Potter quiz, highlighting translation inconsistencies, pretraining data contamination, and variability in name mappings across languages. Reviewer 2t3Z suggested replacing it with cleaner, non-leaked datasets. While the authors defended the dataset as useful for studying knowledge transfer and provided additional justification, they did not introduce alternative evaluations, citing time constraints.\", \"novelty_of_the_proposed_mitigation_methods\": \"Reviewer uRPF argued that the fine-tuning methods proposed\\u2014such as mixed-language fine-tuning\\u2014are well-established, especially in encoder-only models, and lack novelty. The authors clarified that their work highlights the generalizability of such methods to decoder-based LLMs, but reviewers remained unconvinced that the improvements represented a significant advancement.\", \"presentation_and_methodological_clarity\": \"Reviewer uRPF highlighted ambiguities in embedding calculations, hyperparameter choices, and claims about model capabilities. The authors provided clarifications in the rebuttal and revised the paper accordingly, but concerns about the rigour and clarity of the presentation persisted.\\n\\nIn weighing these points, the expanded evaluation was noted as a positive effort, but it did not fully resolve the fundamental issues. The Harry Potter dataset's reliability remained a significant concern, as it introduced noise and potential data contamination that undermined the conclusions. Additionally, the lack of experimental controls and methodological clarity made it difficult to isolate the cross-lingual barrier phenomenon convincingly. The novelty of the contributions was deemed insufficient, as the methods were incremental and largely expected.\\nWhile reviewers 2t3Z and xBQr acknowledged some improvements and raised their scores slightly, uRPF maintained that the paper lacked rigor and substantive contribution. Given the persistent weaknesses in experimental design, dataset reliability, and the novelty of the proposed methods, these points collectively outweighed the paper's strengths. As such, I leaned toward the reviewers' concerns and recommend rejection.\"}", "{\"summary\": \"Authors study multilingual abilities and extent of crosslingual transfer of decoder transformer language models, using three different methodologies: embedding-based, MMLU based and domain-specific. In embedding part authors collect embeddings of sentences in English, sentences, in which some of the input tokens were corrupted by masking the attention or randomly replacing them with different tokens from vocabulary of the model and sentences, in which words were randomly translated into a different language and calculate distances between English, corrupted and partially translated languages. General knowledge based evaluation is conducted by translating parts of the questions or answers from MMLU dataset into a different language. In domain-specific evaluation authors ask questions about Harry Potter series and apply same translation techniques as in general knowledge evaluation. First evaluation show that models have close embeddings after per token translations. Second evaluation shows that models are doing much worse on MMLU after partial translations of the answers. Third evaluation shows that applying translation to parts of the questions or answers from Harry Potter trivia the crosslingual barrier still exists. In later parts of the paper authors test different methodologies of overcoming this barrier, either by prompt engineering and doing few-shot prompting which has limited success and by finetuning selected models on WikiText-2 for general text knowledge and WikiText-103 for domain specific trivia, which improves the performance on both MMLU and HP-Quiz.\", \"soundness\": \"1\", \"presentation\": \"3\", \"contribution\": \"1\", \"strengths\": \"1. Experiments with embeddings of the models are interesting and give some insight into how models encode meaning of partially translated sentences.\\n\\n2. Experiments with general purpose knowledge can be used to better understand multilingual capabilities of trained models. \\n\\n3. Mixup-version of MMLU, when provided, will be useful to do better multilingual evaluations, since the language perturbations can successfully combat the test leakage.\\n\\n4. The text is clear and understandable.\", \"weaknesses\": \"1. Models, selected on line 100, are not multilingual. Llama-3-8B was explicitly stated as a monolingual model, focusing only on English, with usage in other languages described as out-of-scope use. Multilinguality was only stated in 3.1\\u2019s model card. Same can be said about Llama-2-7B and Llama-2-13B. Mistral-7B is not explicitly stated to be multilingual, either. Given both multilingual evaluations and community reports, it is safe to say that these models do not exhibit strong multilingual capabilities. Authors additionally evaluate Mixtral 8x7B and Aya-23-8b, which are explicitly stated to be multilingual models and have some knowledge in selected languages (en, fr, de, es, it), but they are mostly ignored in the main part of paper, instead focusing on Llama-2 and Llama-3. To my mind this is a big misstep, since bad performance of the selected models could be incorrectly generalized to truly multilingual models.\\n\\n2. Domain specific evaluations are being done on the Harry Potter series of books, which was translated to different languages with varying quality. This raises the question of how well the terms described in the books are translated: for instance, tha name of Neville Longbottom is translated to Italian as Neville Paciock, which retains the connotation of clumsiness, but does not sound similarly and does not preserve the meaning of long bottom with the English version, to Chinese as N\\u00e0w\\u0113i L\\u00f3ngb\\u0101d\\u00f9n, which does not retain the connotation since it\\u2019s a simple transliteration and to Russian as Nevil Dolgopups, which both does not retain the clumsiness connotation, meaning or common sound/tokens. This raises concerns if the arguments presented in the paper (e.g. knowledge learned in one language can be translated to other languages) can be checked via Harry Potter trivia.\\n\\n3. All of the selected models almost certainly have seen the original books, dumps from HP Wiki or Wikipedia and blog posts about Harry Potter in non-English languages, so it may have leaked into the training data for all of the languages.\\n\\n4. It is common knowledge that both instruction and preference tuning causes catastrophic forgetting, which leads to models being better at instruction following, safety and answer formatting, but worse at parametric knowledge. From the list of the described models we can see that only Llama-2, Llama-3, Mixtral-8x7b, Zamba-7b and Mistral-7b models are available as base models, with GPT-3.5, GPT-4, Aya-23 being instruction tuned models. It is not explicitly stated, which versions of the models are used \\u2013 pretrained or finetuned, Thus, some clarification is required, since it directly influences the amount of multilingual capability of the models. If these are instruction tuned models, further finetune of them on Harry Potter dataset could lead to catastrophic forgetting, thus perturbating the models\\u2019 scores and if these are base models, direct comparison between Aya-23 and llamas, for instance, presents a potential inconsistency in methodology.\\n\\n5. The phenomenon of crosslingual knowledge barrier (e.g. limited multilinguality capabilities of the model) is well known, so the main contribution of the paper lacks originality.\", \"questions\": \"1. It would be very beneficial if the authors of the paper would use different languages, such as Chinese, Hebrew, Russian or Hindi, since these languages have a much lower amount of shared tokens with English, which is selected as the most highly resourced language in the training mix of all selected models. This would give us insight into how the applied methodology translates into languages, which the model has a harder time training due to lesser overlap between dictionaries.\\n\\n2. Please, repeat your evaluations using a different set of models, such as Mistral-Nemo-12B, Qwen-2.5-7B, Command-R and LLaMA-3.1-8B which are multilingual by design. Claims of multilinguality of these models are supported by extensive multilingual evaluations both by model authors and community members.\\n\\n3. Perhaps designing another evaluation dataset for domain specific knowledge, based on something like some local company guidelines or university campus rules and translating it into different languages would make a much stronger point to support the claims, since it would not exhibit translation artifacts.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to Reviewer xBQr\", \"comment\": \"Thank you for your feedback! We address your questions and comments below.\\n\\n> Q1: the languages considered are rather few and still relatively similar to each other: 2 germanic languages (English and German) and three romance languages (French, Italian, Spanish). In addition all languages are most likely very high resource (albeit not as high as English of course) during training of the different models. It would have been good to see languages that are rather different from the rest (Arabic or Chinese) maybe even low-resource (Bengali or Amharic).\\n\\nThanks for the thoughtful comment. \\nFollowing the suggestion, we have expanded our evaluation to include a total of 16 languages and 15 multilingual LLMs to provide broader insights into our study. \\nSpecifically, in addition to the 5 languages we originally evaluated, we further consider \\n- *Low-resource languages*: Malay (ms), Danish (da), Finnish (fi), Norwegian (no), Bengali (bn), Amharic (am);\\n- *Languages with token distributions significantly different from English*: Russian (ru), Chinese (zn), Hebrew (he), Arbic (ar) and Hindi (hi).\", \"conclusions\": \"- The results on MMLU (Fig. 4 and Fig. 5) and the Harry Potter quiz (Fig. 7) show that cross-lingual knowledge barriers hold for those additional models and languages, which demonstrates the universality of our findings. \\n- We also highlight that in Section 4.2, we evaluated our mixed-language fine-tuned LLMs on languages that were not used during the finetuning stage. Our results in Figures 10 and 11 demonstrate that the model fine-tuned using our method on high-resource languages { English (en), French (fr), German (de), Spanish (es), and Italian (it)} can improve crosslingual performance on Harry potter quiz and MMLU for low-resource languages and languages that are rather different from English. These results provide encouraging evidence for the generalizability of our fine-tuning approach.\\n\\n> Q2: while multilingual have been used it would have been interesting to also consider models that do explicitly use cross-lingual supervision during training such as ALMA or Tower.\\n\\nThanks for the insightful comment. Following your suggestion, we evaluated a total of 15 multilingual LLMs, including the Tower 7B base model, the Tower 7B instruction-tuned model, and several state-of-the-art multilingual LLMs such as Qwen2.5-7B, Llama-3.1-8B, Mixtral-8x7B-v0.1, aya-expanse-8B, aya-23-8B, and Mistral-Nemo-Base-2407.\\n\\nThe results, presented in MMLU (Fig. 4 and 5) and the Harry Potter Quiz (Fig. 7), indicate that even models explicitly trained with cross-lingual supervision, such as the Tower series, face similar cross-lingual knowledge barriers in both general-domain and domain-specific tasks. For example, in Fig. 4 of MMLU evaluation, these models exhibit a significant accuracy drop in mixed-language settings (e.g., question + GT-option, GT-option, and mixup translations) compared to monolingual settings (e.g., English-only or fully translated). This demonstrates that current LLMs struggle with understanding complex multilingual contexts and effectively linking parametric knowledge across languages to answer multiple-choice questions.\\n\\nAdditionally, we observed that the Tower series LLMs, which are primarily optimized for translation-related tasks, have limited cross-lingual performance in knowledge-intensive tasks. This limitation could be attributed to factors such as catastrophic forgetting or constraints in the training data's knowledge coverage. Similarly, we envision that the ALMA series translation models would exhibit similar performance as the Tower series models. These findings further highlight the contributions of our work in uncovering the weaknesses of existing multilingual LLMs and emphasizing the need for more robust cross-lingual understanding in future models.\\n\\n\\n> Q3: In Footnote 1 you say \\\"This criterion allows the case when a small amount of parallel texts accidentally crawled from the web are mixed in the pretraining dataset, as it is nearly impossible to verify. We believe such presence, if it exists at all, would have negligible impact on the model given the size of the rest of the pretraining data.\\\"\\nA) How do you know that GPT 3.5 and 4 don't use parallel data explicitly during training?\\nB) While the parallel data might be small in comparison to the rest of the data, does that really mean it has a negligible impact on the model, considering the vast amount of parameters that could allow for quite a high degree of memorisation? Can you substantiate your claim a bit better (references or otherwise)?\\n\\n\\nThanks for the thoughtful comment. We agree that verifying the presence or absence of parallel training materials in GPT-3.5/4 is challenging, and measuring their potential impact is equally difficult. To address this uncertainty and avoid making unverifiable assumptions, we have revised the introduction section and removed the claim regarding the existence and impact of parallel data.\"}", "{\"title\": \"Response to the author's response\", \"comment\": \"Dear Authors,\\n\\nThank you for your detailed response and the thoughtful improvements made to your paper. I raised my score. \\n\\n>Thank you for the thoughtful comment. Following the reviewer's suggestion, we have expanded our evaluation to include a total of 15 multilingual LLMs in our main paper in revised PDF, including strong multilingual LLMs such as Qwen2.5-7B, Llama-3.1-8B, Mistral-Nemo, Mixtral-8x7B-v0.1, aya-expanse-8B, aya-23-8B and Tower series 7B models\\n\\nThis enhancement significantly strengthens your evaluation and adds value to the overall study. I appreciate the effort in expanding the scope of your analysis.\\n\\n>While these mappings may not fully preserve\\u00a0connotations or meanings, a multilingual human with knowledge of both languages would still be able to\\u00a0transfer the relevant knowledge about the character or concept and answer questions accurately, evenwhen the translation quality varies\\n\\nI believe this claim still might need further qualification. While the general assertion that a multilingual human could transfer relevant knowledge is plausible, there are notable exceptions. Some translations differ so significantly that even speakers proficient in both languages may find it challenging to identify connections. For instance, one Russian translation of \\\"Harry Potter\\\" alters the character Severus Snape's name to \\\"Zloteus Zley\\\" (translated as \\\"Evilus Evil\\\"), Professor Quirrell becomes \\\"Professor Strauns,\\\" Ravenclaw is translated to \\\"Vransor,\\\" and horcruxes and boggarts are rendered as \\\"okayant\\\" and \\\"vrizraks,\\\" respectively. Such substantial differences can obscure the original connotations, potentially invalidating the claim about consistent cross-language understanding. I conducted a simple experiment by asking my Russian-speaking colleagues, who had read all the books in both Russian and English, about these translated characters. They were either unable to identify the characters I referred to, or inferred them incorrectly. This validates my concerns about the translation quality.\\n\\n>Thank you for the comment. Unfortunately, we cannot verify the extent to which the original books, Harry\\u00a0Potter Wiki dumps, or related content in non-English languages might have been included in the pretraining\\u00a0data for the selected models. This could vary significantly across languages and models.\\n\\n I must emphasize that this aspect raises critical concerns about the validity of your experiments. The inability to measure training data leakage implies that it remains unclear whether the observed performance in certain languages results from genuine crosslingual transfer or simply due to English data leakage during pretraining. This lack of clarity significantly impacts the interpretability of your results.\\n\\n>Moreover, in Figure 6, we conducted a controlled experiment where we explicitly finetuned the models on indomain content presented in English, where models still consistently perform better at answering questions in\\u00a0English than in other languages. \\n\\nThis still invites some skepticism. The translation inconsistencies between versions of \\\"Harry Potter\\\" highlight that this dataset is not necessarily parallel, with notable variations in translations across languages. The superior performance in English might not indicate a crosslingual barrier but could instead suggest a skewed data selection process favoring languages more prevalent during pretraining. This distinction is critical, as it may undermine the experiment's conclusions about crosslingual transfer efficacy.\", \"my_recommendation_remains\": \"consider removing the HP Quiz entirely and replacing it with content that has not been leaked on the internet, such as local company guidelines or university campus rules, as previously suggested in Q7. Re-running your experiments with such data would likely lead to a more robust experimental design, thereby enhancing the credibility of the results.\\n\\nNevertheless, the expansion of the evaluation to include a broader set of multilingual models and languages marks a significant improvement in the study, and I am inclined to reassess my evaluation positively based on these revisions. However, I still believe the paper would benefit from further refinement, particularly in addressing the limitations discussed above. Strengthening the discussion around these points and acknowledging the inherent complexities of translation and pretraining data limitations would provide a more balanced and robust perspective.\"}", "{\"title\": \"Response to Reviewer uRPF (Part 1)\", \"comment\": \"Thank you for your feedback! We address your questions and comments below.\\n\\n> Q1: L26-36: The first figure presents a poor example. Why couldn't it simply be the same question in both English and French? Additionally, why is there a need to create a knowledge that does not exist, so it can not be tested out as an example?\\n\\nThank you for the valuable comment. Following your suggestion, we have updated Figure 1 to use the same question in both English and French. \\n\\nRegarding the use of a fabricated example, it was purely for illustration purposes. The goal was to demonstrate that the model must rely on knowledge stored in its weights in one language to answer questions in another language. Since the exact information on which knowledge is present in which language in the training data of existing pre-trained LLMs is not fully transparent, we opted to use a fabricated example to avoid potential confounding factors.\\n\\n \\n> Q2: L104: The authors state, \\u201csuch presence, if it exists at all,\\u201d regarding parallel sentences on the web. It would be interesting to know how papers that mine billions of parallel texts from the web feel about this statement (e.g., https://arxiv.org/abs/1911.04944).\\n\\nThank you to the reviewer for pointing this out. We acknowledge the potential existence of parallel data in the training data of those LLMs, which is unknown to us. We revised our statements in the Introduction (Section 1) to clarify this point and have incorporated a discussion with [1] shared by the reviewer.\", \"reference\": \"- [1] CCMatrix: Mining Billions of High-Quality Parallel Sentences on the WEB. https://arxiv.org/abs/1911.04944\\n\\n> Q3: L137: The paper mentions obtaining sentence embeddings from the LLM; however, it does not detail how this is done. Since these are decoder-only models with causal attention, how are the embeddings of the tokens aggregated? Is it the last token, or are they weighted? (see section 3.1 of https://arxiv.org/pdf/2410.05873 for a background on the different techniques)\\n\\nThanks for the question. The sentence embedding is a single vector representing the average of the last layer's activations across all tokens in the sentence, a common pooling strategy for creating fixed-size representations of variable-length text. We have added these clarifications to the revised PDF. \\n\\n> Q4: L134-142: There are many variables concerning probabilities, yet there is no explanation for why these specific values were chosen.\\n\\n Thanks for raising this point. \\n- For Mixed-language-translate, the choice of a probability of $p = 0.8$ that each word is unchanged corresponds to a 20% probability ($1 - p = 0.2$) that each word is replaced. This aligns with the uniform probability distribution across the five main languages evaluated in our study: {en, fr, de, es, it}. In other words, each word has a $0.16$ probability of being translated into a non-English language. \\n- This also explains our choice for Random Token Replacement where $p=0.16$. \\n\\nWe have added these clarifications to the revised PDF. \\n\\n> Q5: L158-159: The statement \\u201cThis underscores the explicit cross-lingual capabilities of multilingual LLMs\\u201d lacks clarity regarding the rationale and conclusions drawn. With a lower probability (0.16 vs. 0.8), tokens were randomly selected from the vocabulary. Now, the distribution of these two (cosine similarities between the original and mixed translated sentence embeddings versus the original and random token-replaced sentence embeddings) varies significantly. Many variables are involved here: do the random tokens from the vocabulary come only from all writing systems or just Latin? How are the embeddings computed, and why were these probabilities chosen? Although I acknowledge the existence of cross-lingual capabilities, I am not convinced by the authors' experiment.\\n\\n\\nThanks for the insightful comment. The random tokens come from the entire vocabulary space of the tokenizer. The embedding computation is clarified in answer to Q3. \\nWe kept the token modification probability low (0.16) across all three variants: mixed-language translation, random token replacement, and random token dropout, to ensure that the altered sentences could maintain semantic and structural similarity with the original text.\"}", "{\"summary\": \"This paper investigates how multilingual LLMs perform for actual cross-lingual tasks. This covers a wide range of tasks including machine translation but also several variations of multilingual tasks where different parts of the input (text, multiple choice options, questions) are translated into different languages. The authors show a consistent drop in performance when considering actual mixed language tasks and suggest simple fine-tuning strategies to remedy this drop.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": [\"well written paper\", \"interesting outcomes. Not entirely surprising but still nice to see.\", \"solid methodological evaluation (but see my comment regarding choice of language below)\"], \"weaknesses\": [\"the languages considered are rather few and still relatively similar to each other: 2 germanic languages (English and German) and three romance languages (French, Italian, Spanish). In addition all languages are most likely very high resource (albeit not as high as English of course) during training of the different models. It would have been good to see languages that are rather different from the rest (Arabic or Chinese) maybe even low-resource (Bengali or Amharic).\", \"while multilingual have been used it would have been interesting to also consider models that do explicitly use cross-lingual supervision during training such as ALMA or Tower.\"], \"questions\": \"In Footnote 1 you say \\\"This criterion allows the case when a small amount of parallel texts accidentally crawled from the web are mixed in the pretraining dataset, as it is nearly impossible to verify. We believe such presence, if it exists at all, would have negligible impact on the model given the size of the rest of the pretraining data.\\\"\\n\\nA) How do you know that GPT 3.5 and 4 don't use parallel data explicitly during training?\\n\\nB) While the parallel data might be small in comparison to the rest of the data, does that really mean it has a negligible impact on the model, considering the vast amount of parameters that could allow for quite a high degree of memorisation? Can you substantiate your claim a bit better (references or otherwise)?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to Reviewer 2t3Z (Part 1)\", \"comment\": \"Thank you for your feedback! We address your questions and comments below.\\n\\n> Q1: Models, selected on line 100, are not multilingual. Llama-3-8B was explicitly stated as a monolingual model, focusing only on English, with usage in other languages described as out-of-scope use.... To my mind this is a big misstep, since bad performance of the selected models could be incorrectly generalized to truly multilingual models\\u2026\\u2026 Please, repeat your evaluations using a different set of models, such as Mistral-Nemo-12B, Qwen-2.5-7B, Command-R and LLaMA-3.1-8B which are multilingual by design. Claims of multilinguality of these models are supported by extensive multilingual evaluations both by model authors and community members.\\n\\nThank you for the thoughtful comment. Following the reviewer's suggestion, we have expanded our evaluation to *include a total of 15 multilingual LLMs in our main paper* in revised PDF, including strong multilingual LLMs such as Qwen2.5-7B, Llama-3.1-8B, Mistral-Nemo, Mixtral-8x7B-v0.1, aya-expanse-8B, aya-23-8B and Tower-series 7B models.\\n\\nWhile our original focus -- Llama-3-8B, Llama-2-7B, Llama-2-13B, and Mistral-7B -- are not explicitly stated to be multilingual, we find that they have *competitive cross-lingual performance* to LLMs that are explicitly mentioned to be multilingual.\\n- From the results on cross-lingual MMLU tasks in Figure 4, we find that the multi-choice question answering accuracy of Llama-3-8B and Llama-3.1-8B are very close, potentially due to their similar knowledge scope in training data. Moreover, Mistral-7B performs better than aya-23-8B, and Llama-2-13b performs better than Mistral-Nemo. \\n- From the results of the Harry Potter Quiz tasks in Figure 7, we find that Mistral-7B and Llama-3-8B obtain comparable performance to other state-of-the-art multilingual LLMs.\\n\\nThese new results *justify our original model selection*. Moreover, we clarify that our primary goal was to evaluate general-purpose LLMs that can understand multiple languages due to massive pretraining data as stated in Section 1. While models like Llama-3-8B, Llama-2-7B, Llama-2-13B, and Mistral-7B may not be explicitly optimized for multilingual tasks, it is reasonable to assume that training data in non-English Languages exists in their web-crawled pretraining datasets. Our translation results in Section 2.1 (Table 7 in the Appendix) show that these models achieve competitive performance to commercial-level translation API in explicit translation tasks, particularly in translations from other languages to English. Additionally, the embedding analysis in Section 2.2 shows the explicit cross-lingual capabilities of those LLMs. These results motivated us to consider them valid multilingual LLMs for our study.\\n\\nOverall, the results in Figure 4 and Figure 7 show that the 15 multilingual LLMs demonstrate significant cross-lingual knowledge barriers on our tasks, suggesting that such limitations are not exclusive to less multilingual-optimized models, and generalize to strong multilingual models. \\n\\n> Q2: Domain specific evaluations are being done on the Harry Potter series of books, which was translated to different languages with varying quality. This raises the question of how well the terms described in the books are translated: for instance,...., which both does not retain the clumsiness connotation, meaning or common sound/tokens. This raises concerns if the arguments presented in the paper (e.g. knowledge learned in one language can be translated to other languages) can be checked via Harry Potter trivia.\\n\\nThanks for the insightful comment! We acknowledge that translation is a very challenging task and it is difficult even for expert human translators to get all the subtlety right.\\n\\nHowever, we believe these challenges do not invalidate the motivation of our study.\\n- For proper nouns, such as names, there is typically a finite set of mappings between languages (e.g., \\u201cNeville\\u201d to \\u201cN\\u00e0w\\u0113i\\u201d in Chinese or \\u201cNevil\\u201d in Russian). While these mappings may not fully preserve connotations or meanings, a multilingual human with knowledge of both languages would still be able to transfer the relevant knowledge about the character or concept and answer questions accurately, even when the translation quality varies. Our work seeks to assess whether LLMs exhibit a similar capability.\\n- More importantly, our focus on *multiple-choice QA tasks narrows down the search space for the correct answer*. Even with imperfect translations, a model with the relevant knowledge can identify the correct choice from a limited set of options. This approach mitigates some of the noise introduced by translation variability and allows us to measure whether the model truly possesses crosslingual knowledge.\\n\\nWe appreciate your concern and have included the discussion of the limitations arising from translation variability in Appendix A of the revised PDF.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"title\": \"Follow-up with Reviewer uRPF\", \"comment\": \"We wanted to express our gratitude once again to the reviewer for the feedback. We hope that the detailed responses and the manuscript revisions we provided have clarified the concerns regarding Figure 1, parallel data, hyperparameter details, the distinction between monolingual evaluation and our cross-lingual knowledge reasoning evaluation, as well as our contributions beyond fine-tuning.\\n\\nIn response to the reviewers\\u2019 comments, we have expanded our evaluation to include a total of 16 languages and 15 multilingual LLMs to provide broader insights into our study. Our expanded evaluation covers low-resource languages and languages with token distributions significantly different from English, as well as state-of-the-art multilingual LLMs such as Qwen2.5-7B, Llama-3.1-8B, Mixtral-8x7B-v0.1, aya-expanse-8B, aya-23-8B, Mistral-Nemo-Base-2407, Tower-7B-base and Tower7B-instruction-tuned model. \\n\\nIf there are any remaining concerns or if further clarification is needed, we would be more than happy to address them before the discussion period concludes. We also kindly invite you to reevaluate our work in light of the additional clarifications and improvements we have presented.\\n\\nThank you again for your time and feedback.\"}", "{\"title\": \"Response to Reviewer uRPF (part 2)\", \"comment\": [\"> all the methods used to improve the multilinguality score are well-established and not a contribution\", \"We respectfully disagree with this comment and request clarification regarding which methods the reviewer considers \\\"well-established.\\\" Below, we highlight the novelty of our work and provide evidence that our contributions are original and impactful:\", \"The proposed mixed-language fine-tuning (varying the mixing unit: words, sentences, documents) is, to the best of our knowledge, novel in the context of multilingual LLMs. This approach provides new insights into cross-lingual knowledge reasoning:\", \"Mixed-language fine-tuning boosts cross-lingual capabilities (line 421-429): By exposure to frequent language switch during fine-tuning on general (out-of-domain) corpus, LLMs can better adapt to the setting during testing when the same knowledge is asked in a different (and usually non-English) language.\", \"Generalization to out-of-distribution languages (lines 492-512): We evaluated our fine-tuned models on languages that were not included in fine-tuning data. Results in Fig. 11 show that mixed-language fine-tuning on the general Wiki corpus can improve the performance of 11 other languages on HP-Quiz, including low-resource ones and those substantially different from English. Furthermore, as shown in Fig. 12, mixed-language fine-tuning also boosts the performance of MMLU variants in various cross-lingual settings for four low-resource languages.\", \"As detailed in our response to the original **Q8**, our work also makes several broader contributions to the field, including:\", \"New Problem Formulation: Identifying and addressing the cross-lingual knowledge barrier in pretrained and fine-tuned language models.\", \"Proposed Challenging Scenarios and Datasets: Specifically designed to evaluate cross-lingual knowledge reasoning.\", \"Revealing Cross-Lingual Knowledge Barriers: A detailed analysis of 15 multilingual LLMs across 16 languages.\", \"Limited Impact of Inference-Time Mitigations: Our work provides concrete evidence that inference-time strategies fail to address these challenges.\", \"Generalization of Mixed-Language Fine-Tuning: Comprehensive evaluations demonstrate the robustness of our proposed approach.\", \"If the reviewer perceives our methods to be \\\"well-established\\\", we respectfully request supporting references. Such information would help us better understand this perspective and either further justify the novelty of our contributions or make adjustments as necessary to address these concerns.\", \"Thank you for your time and consideration.\"]}" ] }

[ "Pengxiang Li", "Lu Yin", "Shiwei Liu" ]

[ "large language model", "layer normalization", "mix-Layer Normalization" ]

{ "note_id": [ "yXLxMw0CSx", "xsdqS4A4D1", "wVRTnRwzMe", "vbGiSyObvx", "vL4m1LD16X", "ssR9Cl7DgW", "rY7N8maqPt", "qxiaJeSKBk", "q87UOnjq8E", "pTqS5n9pJR", "laBA5a0zmt", "hsvo6iMTuv", "bdcpLO9DGq", "bUu8U606sp", "a33vZPIdmq", "Z9diVlUj5T", "XtIeES9s1P", "Wti2QjNrtV", "VQRriiVr8R", "UAoZ3k6BJv", "TFKnG6anbv", "RR3U0nZLKT", "PjTNLqLRE3", "Oi7XnRcMQJ", "KUFqIpVhWF", "K6WS9E96x3", "JvjVPt7i3y", "INFUM3jV4t", "GRvy4sK1mE", "DS0xOyNuwY", "CsfOY5KCxj", "Az88pS1qXB", "9ZgTOJKLcX", "8TwnrK9RTv", "688EBOPqGE", "5TTKfErHBH", "2NAFyeDTVU", "1JhSipQfHk" ], "note_type": [ "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "meta_review", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "decision", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment" ], "note_created": [ 1732972214280, 1732128902439, 1732128929034, 1732724712394, 1730248411094, 1732717154831, 1732723894340, 1732542596237, 1732128953193, 1732140691830, 1733659283770, 1730446307528, 1732877776088, 1732123380781, 1732121718023, 1732536530799, 1732542761398, 1730651294828, 1732143763908, 1732122955965, 1730270662494, 1732716914294, 1732772041239, 1732128871733, 1732123286027, 1732123029328, 1737523971784, 1732869487959, 1730663856555, 1732122879144, 1732716181505, 1732121968885, 1732535970579, 1732122304319, 1732537214759, 1732122036464, 1732982003591, 1732958382724 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission9262/Reviewer_9E2P" ], [ "ICLR.cc/2025/Conference/Submission9262/Authors" ], [ "ICLR.cc/2025/Conference/Submission9262/Authors" ], [ "ICLR.cc/2025/Conference/Submission9262/Authors" ], [ "ICLR.cc/2025/Conference/Submission9262/Reviewer_mk56" ], [ "ICLR.cc/2025/Conference/Submission9262/Authors" ], [ "ICLR.cc/2025/Conference/Submission9262/Reviewer_mk56" ], [ "ICLR.cc/2025/Conference/Submission9262/Authors" ], [ "ICLR.cc/2025/Conference/Submission9262/Authors" ], [ "ICLR.cc/2025/Conference/Submission9262/Reviewer_mk56" ], [ "ICLR.cc/2025/Conference/Submission9262/Area_Chair_PG2b" ], [ "ICLR.cc/2025/Conference/Submission9262/Reviewer_6By3" ], [ "ICLR.cc/2025/Conference/Submission9262/Authors" ], [ "ICLR.cc/2025/Conference/Submission9262/Authors" ], [ "ICLR.cc/2025/Conference/Submission9262/Authors" ], [ "ICLR.cc/2025/Conference/Submission9262/Authors" ], [ "ICLR.cc/2025/Conference/Submission9262/Authors" ], [ "ICLR.cc/2025/Conference/Submission9262/Reviewer_CgHw" ], [ "ICLR.cc/2025/Conference/Submission9262/Authors" ], [ "ICLR.cc/2025/Conference/Submission9262/Authors" ], [ "ICLR.cc/2025/Conference/Submission9262/Reviewer_9E2P" ], [ "ICLR.cc/2025/Conference/Submission9262/Authors" ], [ "ICLR.cc/2025/Conference/Submission9262/Area_Chair_PG2b" ], [ "ICLR.cc/2025/Conference/Submission9262/Authors" ], [ "ICLR.cc/2025/Conference/Submission9262/Authors" ], [ "ICLR.cc/2025/Conference/Submission9262/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission9262/Reviewer_6By3" ], [ "ICLR.cc/2025/Conference/Submission9262/Reviewer_iHrq" ], [ "ICLR.cc/2025/Conference/Submission9262/Authors" ], [ "ICLR.cc/2025/Conference/Submission9262/Authors" ], [ "ICLR.cc/2025/Conference/Submission9262/Authors" ], [ "ICLR.cc/2025/Conference/Submission9262/Reviewer_iHrq" ], [ "ICLR.cc/2025/Conference/Submission9262/Authors" ], [ "ICLR.cc/2025/Conference/Submission9262/Reviewer_CgHw" ], [ "ICLR.cc/2025/Conference/Submission9262/Authors" ], [ "ICLR.cc/2025/Conference/Submission9262/Authors" ], [ "ICLR.cc/2025/Conference/Submission9262/Area_Chair_PG2b" ] ], "structured_content_str": [ "{\"title\": \"Thanks for your rebuttal. Rating improved.\", \"comment\": \"Given the thoroughness of both the original submission and rebuttal response, I increase the score for this paper. The work makes good contributions to our understanding of LN in transformer architectures.\\n\\nLooking forward to the release of the implementation code and experimental procedures, which will further benefit the research community.\"}", "{\"comment\": \"### **Response to Reviewer 9E2P [2/4]**\\n\\n- **In contrast**, for Pre-LN, Eq. (4) shows that the derivative of the residual connection is decoupled from the term associated with the derivative of $\\\\mathrm{LN}(\\\\cdot)$, preventing vanishing gradients in the early layers. However, because Pre-LN does not normalize the residual connection, the variance of the input to $\\\\mathrm{LN}(\\\\cdot)$, $\\\\sigma_{x'}$, continues to accumulate as the layer depth increases. As a result, Eq. (7) in the deeper layers of Pre-LN approaches zero, causing the right-hand term in Eq. (4) to zero, and leading the derivative of Pre-LN in Eq. (4) to approximate an identity matrix $I$, i.e.,\\n$$\\n\\\\frac{\\\\partial \\\\text{Pre-LN}(x)}{\\\\partial x} \\\\approx I \\\\quad (8)\\n$$\\nThis indicates that the entire Pre-LN operation in deeper layers fails to contribute effectively during backpropagation. Since the Pre-LN operation encompasses the main components of transformer layers\\u2014namely, the attention layer and the feed-forward neural network (FNN) layer\\u2014this explains why the deeper layers of Pre-LN tend to contribute less to learning compared to the earlier layers.\\n\\n\\n- Building on the above theoretical analysis, we propose Mix-LN, which replaces the first $\\\\lfloor \\\\alpha L \\\\rfloor$ Pre-LN layers with Post-LN layers. This approach serves two purposes: First, it reduces the number of stacked Pre-LN layers, mitigating the tendency for the derivative of Pre-LN to approach an identity matrix in deeper layers. Second, since Post-LN is applied to only a few layers, the depth is insufficient for the down-scaling factor in Eq. (7) to accumulate to a level where it causes significant gradient vanishing.\\n\\n\\n- The above theoretical analysis aligns perfectly with our observations in Figure 5. We hope that this analysis offers deeper insights into the effectiveness of Mix-LN.\\n\\n\\n \\\\[1\\\\] Xiong, R., Yang, Y., He, D., Zheng, K., Zheng, S., Xing, C., Zhang, H., Lan, Y., Wang, L. and Liu, T., 2020, November. On layer normalization in the transformer architecture. In *International Conference on Machine Learning* (pp. 10524-10533). PMLR.\\n\\n \\\\[2\\\\] Wang, H., Ma, S., Dong, L., Huang, S., Zhang, D. and Wei, F., 2024\\\\. Deepnet: Scaling transformers to 1,000 layers. *IEEE Transactions on Pattern Analysis and Machine Intelligence*.\\n\\n \\\\[3\\\\] Liu, Liyuan, Xiaodong Liu, Jianfeng Gao, Weizhu Chen, and Jiawei Han. \\\"Understanding the difficulty of training transformers.\\\" *arXiv preprint arXiv:2004.08249* (2020).\\n\\n \\\\[4\\\\] Takase, Sho, Shun Kiyono, Sosuke Kobayashi, and Jun Suzuki. \\\"Spike No More: Stabilizing the Pre-training of Large Language Models.\\\" *arXiv preprint arXiv:2312.16903* (2023).\\n\\n**W2: Scale Limitations: By restricting experiments to models up to 1B parameters and primarily using LLaMA-based architectures, the work lacks demonstrating scalability to production-scale models (7B-175B parameters) or generalizability across different architecture families.**\\n\\n- Thank you for your insightful feedback. We appreciate your suggestion to evaluate our approach on ultra-large models, such as those ranging from 7B to 175B parameters. As you requested, we conduct experiments of LLaMa-7B. All training configurations are identical except for the choice of layer normalization. Due to the limited rebuttal time window, we only finished 13000 steps of training. The perplexity of Mix-LN and Pre-LN is reported in the following table. \\n\\n- We can clearly see that Mix-LN achieves lower perplexity and faster convergence compared with Pre-LN.\\n\\n\\n **Table: Perplexity of LLaMA-7B at Different Training Steps**\\n\\n | Training Steps | 1000 | 2000 | 5000 | 8000 | 11000 | **13000** |\\n |----------------|---------|------|------|-------|-------|-------|\\n | Pre-LN | 817.84 | 290.04 | 52.88 | 35.30 | 30.21 | 28.90 |\\n | Mix-LN | 939.54 | 533.79 | 52.72 | 34.67 | 29.49 | **27.84** |\"}", "{\"comment\": [\"### **Response to Reviewer 9E2P [3/4]**\", \"**W3: Incremental Innovation: The proposed Mix-LN solution, while practical, represents a relatively straightforward combination of existing techniques rather than a fundamental advancement in normalization methodology, especially given extensive prior work referenced in related work and other forums\\\\[2,3\\\\] in this domain.**\", \"We respectfully disagree with the characterization of our contribution as incremental. While our approach involves a straightforward combination of Pre-LN and Post-LN, the novelty and impact of our work lie far beyond this simplification. Below, we reiterate the distinct contributions of our paper in terms of motivation, insights, and approach.\", \"**Motivation:** Our motivation is fundamentally novel. While previous research has identified the ineffectiveness of deep layers in large language models (LLMs), these findings have predominantly been leveraged for model compression [1,2,3]. In contrast, we identify this phenomenon as a significant shortcoming of current LLMs, leading to inefficiencies and underutilization of computational resources that could otherwise enhance model performance.\", \"**Insights:** Building on this novel motivation, our research seeks to uncover the root cause of this inefficiency in deeper layers. This is a non-trivial challenge that has eluded previous studies. For instance, recent work [2] observed that higher layers in current LLMs tend to exhibit high similarity, unlike BERT-style models, which show higher similarity in their shallow layers [4]. However, they incorrectly attributed this behavior to the smaller scale of BERT (e.g., a few hundred million parameters), failing to identify the true underlying cause of this sharp difference. **In contrast to prior work**, we are the first to empirically and theoretically demonstrate that these divergent layer-wise behaviors stem from the choice of layer normalization. Our experiments show that simply altering the layer normalization can replicate these distinct behaviors, even in smaller models with only a few hundred million parameters. Furthermore, our theoretical analysis provides fundamental insights into how Pre-LN and Post-LN differentially influence the training dynamics of earlier and later layers.\", \"**Approach:** Building on these novel insights, we propose Mix-LN, a simple yet effective layer normalization technique that enhances the functionality of deeper layers, ultimately improving the performance of pre-trained models. While the combination of Pre-LN and Post-LN might appear straightforward, our approach is backed up with both theoretical and empirical analysis. The efficacy of Mix-LN is demonstrated across a wide range of model sizes, from 71M to 7B parameters.\", \"We humbly argue that papers such as ours, which advance fundamental understanding and provide actionable insights, should not be dismissed due to the simplicity of their methods. On the contrary, we believe the simplicity of our approach, supported by robust evidence, is a strength rather than a limitation.\", \"**Q2: Boundary Dynamics: What are the theoretical considerations and potential instabilities at the transition point between Pre-LN and Post-LN layers?**\", \"Thank you for bringing up this excellent question! We investigated the transition points between Pre-LN and Post-LN and observed that during training, these layers occasionally experience very large gradients, often referred to as gradient spikes. This phenomenon is understandable given that Post-LN amplifies gradients in the later layers, while Pre-LN amplifies gradients in the earlier layers. Consequently, applying Pre-LN after Post-LN can result in large gradients at the transition layers.\", \"According to prior studies [5,6], such occasional gradient spikes can negatively impact the final training performance. To address this, we experimented with a straightforward approach: detecting gradients in LN layers whose magnitudes exceed 50 times their current momentum and scaling these gradients down to match their current momentum values. This initial attempt has shown promising results, further improving the training performance of Mix-LN.\", \"| Method | Mix-LN | Mix-LN w/ Gradient Scaling |\", \"| -------- | -------- | -------- |\", \"| PPL | 22.33 | **22.24** |\", \"While we agree that this is a fascinating area worth exploring, a systematic study of gradient spikes introduced by different layer normalization is beyond the scope of our current paper. We believe this topic warrants a dedicated investigation, and we plan to explore it in future work.\"]}", "{\"title\": \"Thank you for your support!\", \"comment\": \"Dear Reviewer mk56,\\n\\nWe sincerely appreciate your recognition and acceptance of our paper. It means a lot to us, and we are truly grateful for the time and effort you dedicated to reviewing our work. Your constructive feedback has been invaluable in helping us refine our research.\\n\\nThank you once again for your thoughtful evaluation and support!\\n\\nBest wishes,\\n\\nAuthors\"}", "{\"summary\": \"This paper identifies an training shortfall of LLMs that their deep layers contribute minimally to the overall performance, which wastes the model capacity. The paper identifies the reason behind this as the use of Pre-LN leading to diminished gradients in later layers. The paper further proposes to use a mixed LN approach, where post-LNs are applied to the earlier layers and pre-LNs to the later layers. Experiments show enhancement brought by the mixed LN approach in both pretraining and supervised finetuning.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"This paper provides novel and interesting insights on the different impact of pre/post LN in the training behavior of earlier and later layers. The observation leads to a straightforward solution of applying mixed LN, which can be easily applied in all models. The paper is well-written and easy to follow. Experiments on multiple models and tasks, as well as ablation studies show solid performance improvement of the proposed method.\", \"weaknesses\": \"One major weakness of this paper is that the provided results are not adequate to support the claim that mixed-LN helps deep model by allowing both shallow and deep layers contributes effectively. From the main results in Tab. 1 and 2, it appears that mix-LN improves more over pre-LN in that smaller 71M model (12 layers) than the deeper 1B model (32 layers), which seems to suggest the benefit of mix LN is less in deep models. It would be better to provide results of even larger and deeper models to show that the benefit of mix LN do scale up. If computation resource is limiting the further experiments on larger models, trying a slim but deep architecture may be helpful to verify the effectiveness of the proposed method.\", \"questions\": \"See weakness part. I would like to see more evidence that the proposed method helps more on deeper models.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Kind reminder of our response\", \"comment\": \"Dear Reviewer mk56,\\n\\nThank you once again for your valuable feedback and for taking the time to review our work. We greatly appreciate the insights and constructive comments you have provided, which have helped us improve the clarity and quality of our paper.\\n\\nWe would like to kindly ask if our response has satisfactorily addressed all of your concerns. If there are any remaining questions or points of clarification, we would be more than happy to provide further details.\\n\\nIf you find that your concerns have been fully resolved, we would sincerely appreciate it if you might consider re-evaluating your score in light of the updated information.\\n\\nThank you again for your time and thoughtful review. Your support and feedback mean a great deal to us.\\n\\nBest regards,\\n\\nThe Authors\"}", "{\"comment\": \"Thank you for the response. Though there are still some lingering issues like finding the optimal choice of alpha, I do believe this work provides interesting observation on the use of post/pre LN and provides a practical method to mix them. I will increase my score to acceptance.\"}", "{\"comment\": \"Dear Reviewer mk56,\\n\\nWe sincerely appreciate your thoughtful feedback and follow-up questions, which indeed greatly improved our work. As we approach the conclusion of the discussion phase, please feel free to share any additional concerns, we would be more than happy to address them.\\n\\nKind regards,\\n\\nThe Authors\"}", "{\"comment\": \"### **Response to Reviewer 9E2P [4/4]**\\n\\n**Q3: Asymptotic Performance: While accelerated convergence is demonstrated, how does Mix-LN's final performance compare to Post-LN when both are trained to complete convergence under same hyperparameter settings? A controlled study with matched configurations training until performance plateaus would help distinguish whether Mix-LN's benefits extend beyond computational efficiency to fundamental improvements in model capability.**\\n\\n- We want to clarify that our evaluation is indeed the final performance of different layer normalization including the full pre-training in Table 1 and supervised fine-tuning in Table 2\\\\. The experiments are exactly controlled experiments with matched configurations with the only difference of layer normalization. Our results show that Mix-LN consistently has fundamental improvements in model capacity over other layer normalization techniques. \\n\\n [1] Yin, Lu, You Wu, Zhenyu Zhang, Cheng-Yu Hsieh, Yaqing Wang, Yiling Jia, Gen Li et al. \\\"Outlier weighed layerwise sparsity (owl): A missing secret sauce for pruning llms to high sparsity.\\\" *ICML 2024\\\\.*\\n\\n [2] Gromov, A., Tirumala, K., Shapourian, H., Glorioso, P. and Roberts, D.A., 2024\\\\. The unreasonable ineffectiveness of the deeper layers. arXiv preprint arXiv:2403.17887.\\n\\n [3] Men, X., Xu, M., Zhang, Q., Wang, B., Lin, H., Lu, Y., Han, X. and Chen, W., 2024\\\\. Shortgpt: Layers in large language models are more redundant than you expect. *arXiv preprint arXiv:2403.03853*.\\n\\n [4] Sajjad, Hassan, Fahim Dalvi, Nadir Durrani, and Preslav Nakov. \\\"On the effect of dropping layers of pre-trained transformer models.\\\" *Computer Speech & Language* 77 (2023): 101429\\\\.\\n\\n [5] Chowdhery, A., Narang, S., Devlin, J., Bosma, M., Mishra, G., Roberts, A., Barham, P., Chung, H.W., Sutton, C., Gehrmann, S. and Schuh, P., 2022. Palm: Scaling language modeling with pathways. arXiv 2022. arXiv preprint arXiv:2204.02311, 10.\\n\\n [6] Takase, S., Kiyono, S., Kobayashi, S. and Suzuki, J., 2023. Spike No More: Stabilizing the Pre-training of Large Language Models. arXiv preprint arXiv:2312.16903.\"}", "{\"title\": \"Sensitivity to alpha\", \"comment\": \"I would like to thank the author for the response. The provided results largely resolves my concern on the scalability of the proposed method. However, it echoes a point raised by other reviewers, that the benefit of mix-LN may be largely dependent on the choice of $\\\\alpha$. Can you comment on how to effectively choosing a proper $\\\\alpha$ for a new model besides naively trying out each options?\"}", "{\"metareview\": \"Summary: Mix-LN combines pre-LN and post-LN to tackle gradient issues; it improves deeper layers' learning and leads to better overall model performance.\", \"strengths\": \"motivation on balancing gradient dynamics; extensive experiments on LLMs show improvement; improved stability\", \"weaknesses\": \"benefits on large models diminish; not a groundbreakingly novel method; presentation clarity issues; limited non-LLM eval\", \"reasons_for_decision\": \"overall wide support from reviewers; simple method that lead to improvement across settings outweighs flaws.\", \"additional_comments_on_reviewer_discussion\": \"The authors addressed scalability concerns with additional experiments on LLaMA-7B, included comparisons with other normalization layers, and improved figure clarities. These responses resolved several reviewer concerns, leading to one reviewer upgrading the rating.\"}", "{\"summary\": \"The paper introduced Mix-LN, which applies Pre-LN and Post-LN in different parts of the LLMs. The author explore the diminished gradient norms of Pre-LN in deeper layers and vanishing gradient of Post-LN in earlier layers, both theoretically and experimentally. They then conduct the experiments comparing Post-LN, DeepNorm, Pre-LN and Mix-LN in pre-training and fine-tuning, with anaylsis on the results. The results show Mix-LN with certain ratios improves the overall capacity and efficiency of LLMs.\", \"soundness\": \"4\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. This paper is very well-organized.\\n2. This paper introduce the problem with theortical anaysis and strengthen with experiments, which makes the problem clear and well-motivated.\\n3. Comprehensive experiments for presenting the effect of Mix-LN.\\n4. The anaysis are insightfull.\", \"weaknesses\": \"1. This figures of angular distance are somewhat non-intuitive. Can't understand on the first sight. Block size more like describing the size of the model than distance between two layers under comparison. The first detailed description is actucally in line 402-403.\\n2. Typo problem: Figure 3-c (in main body) or Figure 3-e (in caption of Figure 3).\", \"questions\": \"1. Figure 2-a: Could you explain the triangle-like yellow areas (which also appears but no so significiant in Figure 2-b)? It seems to be discontinuous changes in angular distance between near or adjacent layers. However, from my understanding, the transformer blocks in BERT are the same. Therefore what makes layer 5/6 or 14/15 so special?\\n2. Figure 2-b: Could you explain the dark-blue line on the right? It seems to be a significiant change of angular distance in last layer.\\n3. Section 4: Minor comments. There was no introduction for perlexity?\\n4. Figure 4: a&b shows exact opposite result than Figure 2/3. From my understanding, the angular distance will grow bigger (the color will thus go dark) as the distance between two layers increases. However, for example in layer 10 of Figure 4-b, the farthes has smallest angular distance. Please correct me if I am mistaken.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"details_of_ethics_concerns\": \"None\", \"rating\": \"8\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Thank you for your recognition and support!\", \"comment\": \"Dear Reviewer 6By3,\\n\\nWe sincerely thank you for recognizing the merits of our paper. I am glad that our response resolves your problems! Your support means a lot to us! \\n\\nBest wishes,\\n\\nAuthors\"}", "{\"comment\": \"### **Response to Reviewer mk56**\\nWe sincerely thank the reviewer for giving us constructive review. We thank the reviwer for recognizing our paper is novel and provides interesting insights. We address your concerns in the following. \\n\\n**W1: One major weakness of this paper is that the provided results are not adequate to support the claim that mixed-LN helps deep model by allowing both shallow and deep layers contributes effectively. It would be better to provide results of even larger and deeper models to show that the benefit of mix LN do scale up.**\\n\\n- First, we would like to clarify that the small performance gain reported in our submission of 1B model is due to a suboptimal choice of \\u03b1. Specifically, for the 1B model, the ratio \\u03b1 was set as 0.33, rather than the optimal ratio of 0.25 as stated in our claims. When \\u03b1 is set to the optimal value of 0.25, Mix-LN achieves a notable performance improvement over Pre-LN, with a nearly 0.5 perplexity reduction for LLaMa-1B as demonstrated in the following table. Note that it is reasonable for the performance gains to reduced to certain extends as the model size increases, since larger models inherently have better baseline performance, reducing the room for further improvement.\\n\\n\\n | LLaMa-1B | Pre-LN | Mix-LN (\\u03b1=0.33) | Mix-LN (\\u03b1=0.25) |\\n | :---- | :---- | :---- | :---- |\\n | PPL | 18.65 | 18.40 | 18.18 |\\n\\n- More importantly, the improvement in perplexity translates into significant enhancements in downstream tasks, including self-supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF). For instance, after applying SFT with the trained 1B models on the Commonsense170K dataset, Mix-LN achieves an average accuracy improvement of 1.24%, with a particularly remarkable gain of 5.42% on the Winogrande task.\\n\\n | Method | MMLU | BoolQ | ARC-e | PIQA | Hellaswag | OBQA | Winogrande | Avg. |\\n | :---- | :---- | :---- | :---- | :---- | :---- | :---- | :---- | :---- |\\n | Pre-LN | 26.54 | 62.20 | 45.70 | 67.79 | 30.96 | 17.40 | 50.51 | 43.01 |\\n | Mix-LN | 27.99 | 61.93 | 48.11 | 68.50 | 31.35 | 18.80 | 55.93 | 44.25 |\\n\\n- Consistently, the benefits of Mix-LN can be seamlessly transferred to RLHF. Following Adam-mini (https://arxiv.org/pdf/2406.16793), we use the ultrafeedback dataset and implement the RLHF workflow from InstructGPT (https://arxiv.org/abs/2203.02155) to optimize the preference reward. Mix-LN achieves a notable reward gain (higher is better) over Pre-LN, i.e., 1.32 vs 0.75.\\n\\n | Method | Model | Final Reward |\\n | :---- | :---- |:---- |\\n | Pre-LN | LLaMA-1B | 0.75 |\\n | Mix-LN | LLaMA-1B | 1.32 |\\n\\n- To further address your concerns regarding the benefits of Mix-LN on large-scale models, we conducted experiments with **LLaMa-7B** using the same setup as GaLore. All training configurations were kept identical, except for the choice of layer normalization. Given the limited rebuttal time window and our computing resources, we only completed 13,000 steps of training. However, based on our experience, models that exhibit consistent improvements early in training typically retain these advantages through the later stages. The perplexity of Mix-LN and Pre-LN under these conditions is summarized in the table below. We also report the training curves of Mix-LN and Pre-LN in Figure 4 of our revision. \\n\\n **Table: Perplexity of LLaMA-7B at Different Training Steps**\\n\\n | Training Steps | 1000 | 2000 | 5000 | 8000 | 11000 | **13000** |\\n |----------------|---------|------|------|-------|-------|-------|\\n | Pre-LN | 817.84 | 290.04 | 52.88 | 35.30 | 30.21 | 28.90 |\\n | Mix-LN | 939.54 | 533.79 | 52.72 | 34.67 | 29.49 | **27.84** |\"}", "{\"comment\": \"### **Response to Reviewer iHrq [1/3]**\\n\\nWe would first like to thank you for your time and effort in reviewing our work. We are glad that you have found our observation interesting, clear, and straight to the point, and our evaluation is realistic and likely correct. We would like to address the weakness pointed out by you one by one as follows:\\n\\n**W1: The relative gains in terms of perplexity/accuracy for larger model slims down. This in a certain sense contradicts the vanishing gradient argument. Besides, the gap compared to pre-LM becomes slower and slower.**\\n\\n- First, we would like to clarify that the small performance gain reported in our submission of 1B model is due to a suboptimal choice of \\u03b1. Specifically, for the 1B model, the ratio \\u03b1 was set as 0.33, rather than the optimal ratio of 0.25 as stated in our claims. When \\u03b1 is set to the optimal value of 0.25, Mix-LN achieves a notable performance improvement over Pre-LN, with a nearly 0.5 perplexity reduction for LLaMa-1B as demonstrated in the following table. Note that it is reasonable for the performance gains to reduced to certain extends as the model size increases, since larger models inherently have better baseline performance, reducing the room for further improvement.\\n\\n | LLaMa-1B | Pre-LN | Mix-LN (\\u03b1=0.33) | Mix-LN (\\u03b1=0.25) |\\n | :---- | :---- | :---- | :---- |\\n | PPL | 18.65 | 18.40 | 18.18 |\\n\\n- More importantly, the improvement in perplexity translates into significant enhancements in downstream tasks, including self-supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF). For instance, after applying SFT with the trained 1B models on the Commonsense170K dataset, Mix-LN achieves an average accuracy improvement of 1.24%, with a particularly remarkable gain of 5.42% on the Winogrande task.\\n\\n | Method | MMLU | BoolQ | ARC-e | PIQA | Hellaswag | OBQA | Winogrande | Avg. |\\n | :---- | :---- | :---- | :---- | :---- | :---- | :---- | :---- | :---- |\\n | Pre-LN | 26.54 | 62.20 | 45.70 | 67.79 | 30.96 | 17.40 | 50.51 | 43.01 |\\n | Mix-LN | 27.99 | 61.93 | 48.11 | 68.50 | 31.35 | 18.80 | 55.93 | 44.25 |\\n\\n- Consistently, the benefits of Mix-LN can be seamlessly transferred to RLHF. Following Adam-mini (https://arxiv.org/pdf/2406.16793), we use the ultrafeedback dataset and implement the RLHF workflow from InstructGPT (https://arxiv.org/abs/2203.02155) to optimize the preference reward. Mix-LN achieves a notable reward gain (higher is better) over Pre-LN, i.e., 1.32 vs 0.75.\\n\\n | Method | Model | Final Reward |\\n | :---- | :---- |:---- |\\n | Pre-LN | LLaMA-1B | 0.75 |\\n | Mix-LN | LLaMA-1B | 1.32 |\\n\\n- To further address your concerns regarding the benefits of Mix-LN on large-scale models, we conducted experiments with **LLaMa-7B** using the same setup as GaLore. All training configurations were kept identical, except for the choice of layer normalization. Given the limited rebuttal time window and our computing resources, we only completed 13,000 steps of training. However, based on our experience, models that exhibit consistent improvements early in training typically retain these advantages through the later stages. The perplexity of Mix-LN and Pre-LN under these conditions is summarized in the table below. We also report the training curves of Mix-LN and Pre-LN in Figure 4 of our revision. \\n\\n **Table: Perplexity of LLaMA-7B at Different Training Steps**\\n\\n | Training Steps | 1000 | 2000 | 5000 | 8000 | 11000 | **13000** |\\n |----------------|---------|------|------|-------|-------|-------|\\n | Pre-LN | 817.84 | 290.04 | 52.88 | 35.30 | 30.21 | 28.90 |\\n | Mix-LN | 939.54 | 533.79 | 52.72 | 34.67 | 29.49 | **27.84** |\"}", "{\"comment\": \"Dear Reviewer iHrq,\\n\\nWe sincerely thank you for your insightful feedback and support. Your constructive comments have been instrumental in refining our work! \\n\\nBest regards,\\n\\nThe Authors\"}", "{\"comment\": \"Dear Reviewer 9E2P,\\n\\nWe sincerely thank you for your valuable feedback, which has greatly improved our work. As the end of the discussion phase is approaching, we kindly encourage you to share any additional feedback or updated thoughts you might have regarding our responses and the points raised during the discussion. We highly value your insights, and any further clarification or considerations you provide would greatly contribute to a well-rounded evaluation of our work. Please let us know if there is anything further we can address to help with your assessment.\\n\\nKind regards,\\n\\nThe Authors\"}", "{\"summary\": \"The paper proposes Mix-LN, a hybrid normalization approach that combines Pre-Layer Normalization (Pre-LN) and Post-Layer Normalization (Post-LN) within large language models (LLMs). The technique leverages Post-LN in shallow layers to mitigate gradient vanishing and Pre-LN in deeper layers to maintain gradient flow, with the aim of maximizing layer effectiveness throughout the network. While Mix-LN demonstrates some empirical gains, especially in mid-sized LLMs, the contribution remains incremental in the broader landscape of normalization techniques for LLMs. The paper\\u2019s main contributions are primarily empirical, and it lacks a strong theoretical foundation to substantiate why Mix-LN improves gradient dynamics compared to recent normalization methods.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": [\"Novel Approach to Gradient Flow in LLMs: Mix-LN\\u2019s hybrid approach to layer normalization is unique, attempting to combine the advantages of Pre-LN and Post-LN across different model depths.\", \"Good Experimental Validation: The paper includes extensive experimentation on multiple model sizes and tasks, showing consistent improvement with Mix-LN, particularly in mid-sized models.\", \"Improved Training Stability: Mix-LN appears to mitigate some of the training instability issues commonly observed with Post-LN in large models, an advantage for practitioners.\"], \"weaknesses\": [\"Lack of Theoretical Rigor: The paper does not provide a detailed theoretical framework to explain why Mix-LN achieves balanced gradient dynamics across layers, which is crucial for the paper\\u2019s validity. In the well-explored field of normalization for LLMs, incremental changes require more substantial theoretical backing to make a notable contribution.\", \"**Limited Comparison** with State-of-the-Art Normalization Techniques: The paper lacks direct comparison with recent normalization methods, such as Admin or Sandwich LN, which also address deep-layer gradient inefficiencies. Without these comparisons, it\\u2019s challenging to assert Mix-LN\\u2019s effectiveness over other recent innovations.\", \"**Diminished Gains on Very Large Models**: The benefits of Mix-LN become less pronounced in very large models, such as LLaMA-1B, where performance gains are smaller. This suggests potential scalability issues for Mix-LN in ultra-large models like 7B, 13B and so on, which is a limitation given the trajectory of LLM research.\", \"Applicability Limited to LLMs: Mix-LN has only been evaluated on LLMs, and its effectiveness on non-language models or other architectures remains untested. This limits the broader applicability and impact of the approach.\", \"Hyperparameter Sensitivity Not Thoroughly Explored: The paper does not fully address the sensitivity of the hyperparameter \\ud835\\udefc (controlling the transition point between Pre-LN and Post-LN), which could impact Mix-LN\\u2019s practical usability across different settings.\"], \"questions\": [\"Could the authors provide more insights or theoretical justifications for why Mix-LN enhances gradient flow differently across shallow and deep layers?\", \"How does Mix-LN perform when compared to other recent normalization methods, such as Admin or Sandwich LN?\", \"How sensitive is Mix-LN to the hyperparameter \\ud835\\udefc, and is this value stable across different tasks and model scales?\", \"Can the authors clarify Mix-LN\\u2019s computational impact during training? Does it introduce any additional training or inference overhead?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Thank you for your quick response!\", \"comment\": \"Thank you for your follow-up question and for acknowledging our response. We are glad that the provided results address your concerns regarding the scalability of our method.\\n\\n- Regarding your point about the dependence of Mix-LN\\u2019s performance on the choice of $\\\\alpha$, we agree that this is an important consideration, especially when applying Mix-LN to new models. From our experience, Mix-LN is quite robust to the choice of $\\\\alpha$ as our results reported in the following tables using LLaMA-250M and LLaMA-1B. Results that outperform Pre-LN are highlighted in bold for clarity. We can see that as long as $\\\\alpha$ is constrained below 0.5, Mix-LN consistently outperforms Pre-LN, with the best results achieved with $\\\\alpha=0.25$.\\n\\n **Table: Perplexity of LLaMA-250M with various Post-LN ratios $\\\\alpha$**\\n | Post-LN Ratios ($\\\\alpha$) | Pre-LN | Mix-LN (12.5%) | Mix-LN (25.0%) | Mix-LN (33.3%) | Mix-LN (41.7%) | Mix-LN (50.0%) | Mix-LN (75.0%) | Post-LN |\\n |---------------------|--------|----------------|----------------|----------------|----------------|----------------|----------------|---------|\\n | Perplexity | 23.39 | **22.37** | **22.33** | **22.83** | **22.80** | **22.81** | 23.64 | 32.18 |\\n\\n ---\\n\\n **Table: Perplexity of LLaMA-1B with various Post-LN ratios $\\\\alpha$**\\n | Post-LN Ratios ($\\\\alpha$) | Pre-LN | Mix-LN (16.7%) | Mix-LN (25.0%) | Mix-LN (33.3%) | Mix-LN (41.7%) | Mix-LN (50.0%) | Post-LN |\\n |----------------|--------|--------|--------|--------|---------|---------|---------|\\n | Perplexity | 18.65 | **18.34** | **18.18** | **18.41** | **18.55** | 18.86 | 1434 |\\n\\n- Moreover, we also conducted experiments with ViT models on ImageNet-1K, where we directly set $\\\\alpha=0.25$ without fine-tuning the hyperparameter. The results demonstrate that $\\\\alpha=0.25$ consistently outperforms Pre-LN, with the performance gains being more pronounced for the larger model (ViT-Small) compared to the smaller model (ViT-Tiny). \\n\\n | | Pre-LN | Mix-LN |\\n | :---- | :---- | :---- |\\n | ViT-Tiny | 67.30 | **67.34** |\\n | ViT-Small | 75.99 | **76.40** |\\n\\n- Based on these observations, we believe that setting $\\\\alpha=0.25$ is a robust choice that should yield good results in most cases. However, if $\\\\alpha=0.25$ fails to converge for even larger models (although we haven\\u2019t tried and encountered such failure), we recommend reducing $\\\\alpha$. This recommendation is informed by our observation that pure Post-LN leads to divergence with LLaMA-1B, suggesting that too many Post-LN layers may induce instability. Reducing $\\\\alpha$ in such cases could help mitigate this issue.\\n\\nThese heuristics are by no means exhaustive but can serve as a starting point. For future work, we plan to further investigate the optimal choice of $\\\\alpha$ for different model sizes theoretically. \\n\\nWe hope this addresses your concern, and we thank you again for the thoughtful question, which has inspired us to explore this direction further.\"}", "{\"comment\": \"### **Response to Reviewer CgHw [3/4]**\\n\\n**W2\\uff1a Limited Comparison with State-of-the-Art Normalization Techniques: The paper lacks direct comparison with recent normalization methods, such as Admin or Sandwich LN, which also address deep-layer gradient inefficiencies. Without these comparisons, it\\u2019s challenging to assert Mix-LN\\u2019s effectiveness over other recent innovations.**\\n\\n- As requested, we conducted comparisons using LLaMA-250M to evaluate Mix-LN against recent normalization methods, including Admin [5], Sandwich-LN [6], and Group-LN [7,8]. The results indicate that Sandwich-LN and Group-LN slightly outperform Pre-LN, while Admin performs worse. However, all of these methods fall to reduce perplexity below 23, falling short of Mix-LN. This result highlights the effectiveness of Mix-LN compared to other recent innovations.\\n\\n | Model | Pre-LN | Admin | Group-LN | Sandwich-LN | Mix-LN |\\n | :------------ | :----- | :----- | :------- | :---------- | :------ |\\n | LLaMA-250M | 23.39 | 24.82 | 23.10 | 23.26 | **22.33** |\\n\\n**W3: Diminished Gains on Very Large Models: The benefits of Mix-LN become less pronounced in very large models, such as LLaMA-1B, where performance gains are smaller. This suggests potential scalability issues for Mix-LN in ultra-large models like 7B, 13B and so on, which is a limitation given the trajectory of LLM research.**\\n\\n- First, we would like to clarify that the small performance gain reported in our submission of 1B model is due to a suboptimal choice of \\u03b1. Specifically, for the 1B model, the ratio \\u03b1 was set as 0.33, rather than the optimal ratio of 0.25 as stated in our claims. When \\u03b1 is set to the optimal value of 0.25, Mix-LN achieves a notable performance improvement over Pre-LN, with a nearly 0.5 perplexity reduction for LLaMa-1B as demonstrated in the following table. Note that it is reasonable for the performance gains to reduced to certain extends as the model size increases, since larger models inherently have better baseline performance, reducing the room for further improvement.\\n\\n\\n | LLaMa-1B | Pre-LN | Mix-LN (\\u03b1=0.33) | Mix-LN (\\u03b1=0.25) |\\n | :---- | :---- | :---- | :---- |\\n | PPL | 18.65 | 18.40 | 18.18 |\\n\\n- More importantly, the improvement in perplexity translates into significant enhancements in downstream tasks, including self-supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF). For instance, after applying SFT with the trained 1B models on the Commonsense170K dataset, Mix-LN achieves an average accuracy improvement of 1.24%, with a particularly remarkable gain of 5.42% on the Winogrande task.\\n\\n | Method | MMLU | BoolQ | ARC-e | PIQA | Hellaswag | OBQA | Winogrande | Avg. |\\n | :---- | :---- | :---- | :---- | :---- | :---- | :---- | :---- | :---- |\\n | Pre-LN | 26.54 | 62.20 | 45.70 | 67.79 | 30.96 | 17.40 | 50.51 | 43.01 |\\n | Mix-LN | 27.99 | 61.93 | 48.11 | 68.50 | 31.35 | 18.80 | 55.93 | 44.25 |\\n\\n- Consistently, the benefits of Mix-LN can be seamlessly transferred to RLHF. Following Adam-mini (https://arxiv.org/pdf/2406.16793), we use the ultrafeedback dataset and implement the RLHF workflow from InstructGPT (https://arxiv.org/abs/2203.02155) to optimize the preference reward. Mix-LN achieves a notable reward gain (higher is better) over Pre-LN, i.e., 1.32 vs 0.75.\\n\\n | Method | Model | Final Reward |\\n | :---- | :---- |:---- |\\n | Pre-LN | LLaMA-1B | 0.75 |\\n | Mix-LN | LLaMA-1B | 1.32 |\\n\\n- To further address your concerns regarding the benefits of Mix-LN on large-scale models, we conducted experiments with **LLaMa-7B** using the same setup as GaLore. All training configurations were kept identical, except for the choice of layer normalization. Given the limited rebuttal time window and our computing resources, we only completed 13,000 steps of training. However, based on our experience, models that exhibit consistent improvements early in training typically retain these advantages through the later stages. The perplexity of Mix-LN and Pre-LN under these conditions is summarized in the table below. We also report the training curves of Mix-LN and Pre-LN in Figure 4 of our revision. \\n\\n **Table: Perplexity of LLaMA-7B at Different Training Steps**\\n\\n | Training Steps | 1000 | 2000 | 5000 | 8000 | 11000 | **13000** |\\n |----------------|---------|------|------|-------|-------|-------|\\n | Pre-LN | 817.84 | 290.04 | 52.88 | 35.30 | 30.21 | 28.90 |\\n | Mix-LN | 939.54 | 533.79 | 52.72 | 34.67 | 29.49 | **27.84** |\"}", "{\"summary\": \"This paper addresses the inefficiency of deeper layers in Large Language Models (LLMs) by introducing Mix-LN, a novel normalization technique that combines Pre-Layer Normalization (Pre-LN) and Post-Layer Normalization (Post-LN). The authors argue that the widespread use of Pre-LN in models like GPT and LLaMA leads to diminished gradient norms in deeper layers, reducing their effectiveness. Mix-LN applies Post-LN to earlier layers and Pre-LN to deeper layers, ensuring more uniform gradient norms across all layers. Through extensive experiments across various model sizes, the authors demonstrate that Mix-LN consistently outperforms both Pre-LN and Post-LN, promoting more balanced gradient norms throughout the network and enhancing the quality of LLM pre-training and fine-tuning. This approach aims to unlock the full potential of deeper layers in LLMs, improving model capacity without increasing model size.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. This article conducts a good experimental analysis and points out that the gradient norm of Pre-LN and Post-LN shows different trends as the number of layers increases.\\n2. Comprehensive experiments on both open-weight and in-house models, with multiple evaluation metrics (Angular Distance, Performance Drop, Gradient Norm).\\n3. Well-structured presentation and logical flow with clear diagrams and mathematical formulations.\", \"weaknesses\": \"1. **Limited Theoretical Foundation:** The paper lacks rigorous mathematical analysis of Mix-LN's properties, particularly in contrast to well-established theoretical frameworks for methods like DeepNorm[1]. This absence of formal proofs for convergence properties and optimal ratio selection limits our understanding of the method's fundamental principles.\\n\\n2. **Scale Limitations:** By restricting experiments to models up to 1B parameters and primarily using LLaMA-based architectures, the work lacks demonstrating scalability to production-scale models (7B-175B parameters) or generalizability across different architecture families.\\n\\n3. **Incremental Innovation:** The proposed Mix-LN solution, while practical, represents a relatively straightforward combination of existing techniques rather than a fundamental advancement in normalization methodology, especially given extensive prior work referenced in related work and other forums[2,3] in this domain.\", \"references\": \"[1] Wang, H., et al. (2024). \\\"Deepnet: Scaling transformers to 1,000 layers.\\\" IEEE TPAMI\\n\\n[2] Su, J. (2022). \\\"Understanding Pre-Norm vs Post-Norm in Transformers.\\\" Scientific Spaces\\n\\n[3] Raschka, S. (2023). \\\"Layer Normalization Variants in Transformer Architectures.\\\" Sebastian Raschka's ML Magazine\", \"questions\": \"1. **Scaling Properties:** How does Mix-LN perform in very large models (>7B parameters), and what theoretical guarantees can be provided for its effectiveness at scale?\\n\\n2. **Boundary Dynamics:** What are the theoretical considerations and potential instabilities at the transition point between Pre-LN and Post-LN layers?\\n\\n3. **Asymptotic Performance:** While accelerated convergence is demonstrated, how does Mix-LN's final performance compare to Post-LN when both are trained to complete convergence under same hyperparameter settings? A controlled study with matched configurations training until performance plateaus would help distinguish whether Mix-LN's benefits extend beyond computational efficiency to fundamental improvements in model capability.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Kind reminder of our response\", \"comment\": \"Dear Reviewer 9E2P,\\n\\nThank you again for your feedback and your help in improving our work! We'd like to kindly remind you that we've addressed your concerns. As the rebuttal period is still ongoing, we wanted to kindly check if you have any additional questions or concerns that we can address. \\n\\nFor your convenience, we summarize our response here:\\n\\n- **Theoretical Foundation:** We added theoretical analysis aiming to offer insights into the effectiveness of Mix-LN.\\n\\n- **Scale Limitations:** We added new experiments with LLaMa-7B. \\n\\n- **Boundary Dynamics:** We analyzed the potential instability at the transition point and conducted a preliminary experiment to resolve it. \\n\\nPlease let us know if there\\u2019s anything further we can address.\\n\\nKind regards,\\n\\nThe Authors\"}", "{\"comment\": \"Dear Reviewers,\\n\\nIf you have not responded to author's rebuttal, please kindly do so as soon as possible. The deadline is Dec 2, but the authors can potentially further clarify questions if you respond earlier. Thanks!\\n\\nBest, AC\"}", "{\"comment\": \"### **Response to Reviewer 9E2P [1/4]**\\nWe would first like to thank you for your time to review our work. We appreciate that you find our experimental analysis good, presentation well-structured and our logical flow clear. We address your comments below. \\n\\n**W1: Limited Theoretical Foundation: The paper lacks rigorous mathematical analysis of Mix-LN's properties, particularly in contrast to well-established theoretical frameworks for methods like DeepNorm. This absence of formal proofs for convergence properties and optimal ratio selection limits our understanding of the method's fundamental principles.**\\n\\n- We are fully aware of and deeply respect the foundational theoretical analyses of layer normalization provided by previous works [1,2,3,4]. While the limited rebuttal time window prevents us from presenting a rigorous theoretical framework, we provide the following theoretical analysis aiming to offer insights into the effectiveness of Mix-LN. Note that the theoretical derivation is primarily based on previous works, particularly [1,4].\\n\\n\\n- Denote $x$ as the input of the $l$-th Transformer layer with dimension of $d$. Post-LN applies $\\\\mathrm{LN}(\\\\cdot)$ after the residual addition:\\n\\n $$\\n \\\\text{Post-LN}(x) = \\\\mathrm{LN}(x + \\\\mathcal{F}(x)). \\\\quad (1)\\n $$\\n\\n In contrast, Pre-LN applies $\\\\mathrm{LN}(\\\\cdot)$ before the residual addition:\\n\\n $$\\n \\\\text{Pre-LN}(x) = x + \\\\mathcal{F}(\\\\mathrm{LN}(x)). \\\\quad (2)\\n $$\\n\\n As shown in our submission, the derivative of Post-LN and Pre-LN are given:\\n\\n $$\\n \\\\frac{\\\\partial \\\\text{Post-LN}(x)}{\\\\partial x} = \\\\frac{\\\\partial \\\\mathrm{LN}(x + \\\\mathcal{F}(x))}{\\\\partial (x + \\\\mathcal{F}(x))}\\\\left(I + \\\\frac{\\\\partial \\\\mathcal{F}(x)}{\\\\partial x} \\\\right), \\\\quad (3)\\n $$\\n\\n $$\\n \\\\frac{\\\\partial \\\\text{Pre-LN}(x)}{\\\\partial x} = I + \\\\frac{\\\\partial \\\\mathcal{F}(\\\\mathrm{LN}(x))}{\\\\partial \\\\mathrm{LN}(x)}\\\\frac{\\\\partial \\\\mathrm{LN}(x)}{\\\\partial x}, \\\\quad (4)\\n $$\\n\\n Both the above equations involve the Jacobian matrix of layer normalization, $\\\\mathbf{J}_{LN}(x') = \\\\frac{\\\\partial \\\\text{LN}(x')}{\\\\partial x'}$,\\n\\n\\n where $x'$ is the input of $\\\\mathrm{LN}(\\\\cdot)$. Following Xiong et al. 2020 [1]'s proof, $\\\\mathbf{J}_{LN}(x')$ can be obtained:\\n\\n\\n $$\\n \\\\frac{\\\\partial \\\\mathrm{LN}(x')}{\\\\partial x'} \\n = \\\\frac{\\\\sqrt{d}}{\\\\| x' \\\\|_2}\\\\bigg(I - \\\\frac{x'x'^\\\\top}{\\\\| x' \\\\|_2^2} \\\\bigg). \\\\quad (5)\\n $$\\n\\n Assuming that $x'$ follows a normal distribution with a mean of 0, we have, $\\\\lVert x' \\\\rVert_2 = \\\\sigma_{x'} \\\\sqrt{d}$,\\n\\n\\n where $\\\\sigma_{x'}$ is the standard deviation of $x'$. Hence, \\n\\n $$\\n \\\\frac{\\\\partial \\\\mathrm{LN}(x')}{\\\\partial x'} = \\\\frac{\\\\sqrt{d}}{\\\\sigma_{x'}\\\\sqrt{d}}\\\\bigg(I - \\\\frac{x' x'^\\\\top}{\\\\sigma_{x'}^2 d} \\\\bigg) \\n = \\\\frac{1}{\\\\sigma_{x'}}\\\\bigg(I - \\\\frac{zz^\\\\top}{d} \\\\bigg), \\\\quad (6)\\n $$\\n\\n where $z=(x'-\\\\mu_{x'})/\\\\sigma_{x'}$ is the standard normal distribution obtained after layer normalization. Since $d \\\\gg 1$ in LLMs, we can finally obtain:\\n\\n $$\\n \\\\frac{\\\\partial \\\\mathrm{LN}(x')}{\\\\partial x'} = \\\\frac{1}{\\\\sigma_{x'}}I. \\\\quad (7)\\n $$\\n\\n In practice, we observe that $\\\\sigma_{x'}$ of Post-LN gradually grows larger than one during training, which means **the spectral norm of the Jacobian matrix of LN is smaller than 1**, i.e., \\n\\n $$\\n \\\\bigg\\\\|\\\\frac{\\\\partial \\\\mathrm{LN}(x')}{\\\\partial x'}\\\\bigg\\\\| = \\\\frac{1}{\\\\sigma_{x'}} < 1.\\n $$\\n\\n According to the derivative of Post-LN in Eq. (3), this down-scaling factor will accumulate as $\\\\prod_{l=1}^{L}\\\\frac{1}{\\\\sigma_{x'}^l}$ over multiple layers $L$, leading to gradient vanishing in early layers in Post-LN transformers.\"}", "{\"comment\": \"### **Response to Reviewer 6By3**\\nWe are really grateful for your valuable review! We appreciate it for your positive score and detailed comments. We are glad that you find our paper well-organized and insightful. We address the weakness pointed out by you one by one as follows:\\n\\n**W1: This figures of angular distance are somewhat non-intuitive. Can't understand on the first sight. Block size more like describing the size of the model than distance between two layers under comparison. The first detailed description is actucally in line 402-403.**\\n\\n- We appreciate the reviewer bringing up this important point. In our work, we initially followed the terminology of angular distance as used in prior research (https://arxiv.org/pdf/2403.17887), which is why we adopted their phrasing. However, we fully agree with your concern regarding the potential for misunderstanding caused by the term \\\"Block size.\\\" To address this, we have updated the terminology to \\\"subsequent $n^{th}$ layer\\\" in our revision.\\n\\n**W2: Typo problem: Figure 3-c (in main body) or Figure 3-e (in caption of Figure 3).**\\n- Thanks! We have fixed them in our revision.\\n\\n**Q1: Figure 2-a: Could you explain the triangle-like yellow areas (which also appears but no so significiant in Figure 2-b)? It seems to be discontinuous changes in angular distance between near or adjacent layers. However, from my understanding, the transformer blocks in BERT are the same. Therefore what makes layer 5/6 or 14/15 so special?**\\n\\n- Great observation! This is indeed a fascinating phenomenon. We believe it is strongly correlated with the findings that different structural information is learned at varying depths of the language model. For instance, (https://aclanthology.org/P19-1356.pdf) discovered that BERT captures a rich hierarchy of linguistic information: surface-level features are learned in the lower layers, syntactic features in the middle layers, and semantic features in the higher layers. Despite the transformer blocks in BERT having identical architectures, different hierarchical information is encoded across distinct groups of layers. Layers within the same group tend to learn similar features, contributing to the compositional nature of the representations learned by BERT. \\n\\n- A similar phenomenon has recently been observed in state-of-the-art (SOTA) LLMs such as Pythia, GPT-2, and Phi (https://arxiv.org/pdf/2406.19384), which also aligns with triangle-like areas in Figure 2-b for LLaMA.\\n\\n\\n\\n**Q2: Figure 2-b: Could you explain the dark-blue line on the right? It seems to be a significiant change of angular distance in last layer.**\\n- The dard-blue line on the right represents the Angular Distance between the last layer with previous layers. It is common that the last layer has very different representations with other layers, as the features from the last layer are task-specific and optimized to directly facilitate the model's objective. The representations from lower and intermediate layers are rich in terms of general contextual information, while the final layer is tasked with deciding or summarizing based on this information to produce an output. As a result, the last layer\\u2019s features may lose some of the nuanced intermediate representations and instead focus on solving the specific task, making them less similar to the features from other layers.\\n\\n**Q3: Section 4: Minor comments. There was no introduction for perlexity?**\\n- Perplexity is a fundamental metric used to evaluate the quality of language models. It is commonly used to assess the ability of a language model to predict a sequence of words or tokens, serving as an indicator of how well the model is capturing the statistical properties of natural language. Formally, perplexity can be defined as the exponential of the average negative log-likelihood of a test set: \\n $$\\n \\\\text{Perplexity} = \\\\exp\\\\left( - \\\\frac{1}{N} \\\\sum_{i=1}^{N} \\\\log P(T_i \\\\mid T_{1}, \\\\dots, T_{i-1}) \\\\right)\\n $$\\n\\n\\n**Q4: Figure 4: a&b shows exact opposite result than Figure 2/3. From my understanding, the angular distance will grow bigger (the color will thus go dark) as the distance between two layers increases. However, for example in layer 10 of Figure 4-b, the farthes has smallest angular distance. Please correct me if I am mistaken.**\\n\\n- The difference between figure 4 and figure 2/3 is that figure 4 depicts the row-normalized angular distance. Following (https://arxiv.org/pdf/2403.17887), normalizing the angular distance by row can tell us the least important stack of layers (the optimal block to prune) for a given block size n (lightest yellow). Figure 4 demonstrates that later layers of Pre-LN (Figure 4-b) has larger yellow areas than Mix-LN (Figure 4-a), indicating that Mix-LN\\u2019s deeper layers in general more important than deeper layers of Pre-LN.\"}", "{\"comment\": \"### **Response to Reviewer CgHw [4/4]**\\n**W4: Applicability Limited to LLMs: Mix-LN has only been evaluated on LLMs, and its effectiveness on non-language models or other architectures remains untested. This limits the broader applicability and impact of the approach.**\\n\\n- To evaluate Mix-LN on non-language tasks, we replaced Pre-LN in ViT models with Mix-LN with \\u03b1=0.25 and trained the updated model on ImageNet-1K, following the ConvNeXt training configurations for 120 epochs. The results clearly demonstrate that the benefits of Mix-LN also generalize to vision tasks. Notably, the performance gains are more pronounced for larger model (ViT-Small) than smaller model (ViT-Tiny).\\n\\n | | Pre-LN | Mix-LN |\\n | :---- | :---- | :---- |\\n | ViT-Tiny | 67.30 | **67.34** |\\n | ViT-Small | 75.99 | **76.40** |\\n\\n**W5: How sensitive is Mix-LN to the hyperparameter \\ud835\\udefc, and is this value stable across different tasks and model scales?**\\n\\n- Mix-LN is quite robust to the choice of \\ud835\\udefc as we have reported in Table 3 of our submission with LLaMA-250M. To draw a more solid conclusion, we added an extra experiment with the LLaMa-1B model. The results are summarized in the table below, confirming that Mix-LN remains robust to the choice of \\u03b1. Results that outperform Pre-LN are highlighted in bold for clarity. Specifically, as long as \\u03b1 is constrained below 0.5, Mix-LN consistently outperforms Pre-LN, with the best results achieved at \\u03b1=0.25.\\n\\n **Table: Perplexity of LLaMA-250M with various Post-LN ratios \\u03b1**\\n | Post-LN Ratios (\\u03b1) | Pre-LN | Mix-LN (12.5%) | Mix-LN (25.0%) | Mix-LN (33.3%) | Mix-LN (41.7%) | Mix-LN (50.0%) | Mix-LN (75.0%) | Post-LN |\\n |---------------------|--------|----------------|----------------|----------------|----------------|----------------|----------------|---------|\\n | **Perplexity** | 23.39 | **22.37** | **22.33** | **22.83** | **22.80** | **22.81** | 23.64 | 32.18 |\\n\\n\\n **Table: Perplexity of LLaMA-1B with various Post-LN ratios \\u03b1**\\n | Post-LN Ratios (\\u03b1) | Pre-LN | Mix-LN (16.7%) | Mix-LN (25.0%) | Mix-LN (33.3%) | Mix-LN (41.7%) | Mix-LN (50.0%) | Post-LN |\\n |----------------|--------|--------|--------|--------|---------|---------|---------|\\n | **Perplexity** | 18.65 | **18.34** | **18.18** | **18.41** | **18.55** | 18.86 | 1434 |\\n\\n\\n**Q6: Can the authors clarify Mix-LN\\u2019s computational impact during training? Does it introduce any additional training or inference overhead?**\\n\\n- Thanks for pointing out this great question. Mix-LN does not introduce any additional training and inference overhead compared to pure Post-LN and Pre-LN. This is one of the advantages of Mix-LN over previous LN variants such as Admin and Sandwich LN.\\n\\n- The table below shows the time per iteration for LLaMa-250M on the A800. It can be observed that, compared to pre-layer normalization and post-layer normalization, the mix approach does not introduce any additional overhead. However, the Sandwich LN incurs additional overhead due to the increased number of uses of layer normalization.\\n\\n\\n | Pre-LN | Post-LN | Mix-LN | Sandwich-LN |\\n | -------- | -------- | -------- |-------- |\\n | 1.97s/iter | 1.98s/iter | 1.97s/iter | 2.42s/iter|\\n\\n\\n\\n \\\\[1\\\\] Xiong, R., Yang, Y., He, D., Zheng, K., Zheng, S., Xing, C., Zhang, H., Lan, Y., Wang, L. and Liu, T., 2020, November. On layer normalization in the transformer architecture. In *International Conference on Machine Learning* (pp. 10524-10533). PMLR.\\n\\n \\\\[2\\\\] Wang, H., Ma, S., Dong, L., Huang, S., Zhang, D. and Wei, F., 2024\\\\. Deepnet: Scaling transformers to 1,000 layers. *IEEE Transactions on Pattern Analysis and Machine Intelligence*.\\n\\n \\\\[3\\\\] Liu, Liyuan, Xiaodong Liu, Jianfeng Gao, Weizhu Chen, and Jiawei Han. \\\"Understanding the difficulty of training transformers.\\\" *arXiv preprint arXiv:2004.08249* (2020).\\n\\n \\\\[4\\\\] Takase, Sho, Shun Kiyono, Sosuke Kobayashi, and Jun Suzuki. \\\"Spike No More: Stabilizing the Pre-training of Large Language Models.\\\" *arXiv preprint arXiv:2312.16903* (2023).\\n\\n [5] Liu, L., Liu, X., Gao, J., Chen, W. and Han, J., 2020. Understanding the difficulty of training transformers. arXiv preprint arXiv:2004.08249.\\n\\n [6] Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H. and Tang, J., 2021. Cogview: Mastering text-to-image generation via transformers. Advances in neural information processing systems, 34, pp.19822-19835.\\n\\n [7] Wu, Y. and He, K., 2018. Group normalization. In Proceedings of the European conference on computer vision (ECCV) (pp. 3-19).\\n\\n [8] Ma, X., Yang, X., Xiong, W., Chen, B., Yu, L., Zhang, H., May, J., Zettlemoyer, L., Levy, O. and Zhou, C., 2024. Megalodon: Efficient llm pretraining and inference with unlimited context length. arXiv preprint arXiv:2404.08801.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}", "{\"comment\": \"Thank you for your detailed reply. Your answers are helpful and resolve my problems. I will keep my rating.\"}", "{\"summary\": \"In this paper the authors context the employment of either pre-layer normalization (Pre-LN) and post layer normalization. (Post-LN) They bring an argument showing vanishing gradient for earlier or later layers. Building on this intuition, they propose to place post-LN in early layers and pre-LN in later ones, to maximize the gradient. The evaluation is performed on LLMs, either large and small scale ones.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": [\"The observation related to gradient norm brought by LN is interesting, clear, and straight to the point.\", \"Part of the quantitative evaluation is conducted to large-scale LLMs, making this work actual.\", \"All the empirical evaluation is quite realistic and likely correct.\", \"Figures in general do an excellent job in providing a display on the distributions.\"], \"weaknesses\": [\"The relative gains in terms of perplexity/accuracy for larger model slims down. This in a certain sense contradicts the vanishing gradient argument. Besides, the gap compared to pre-LM becomes slower and slower.\", \"The success of the approach depends on $\\\\alpha$: there are cases like Llama-1B for BoolQ in which only pre-LM performs better, indicating that tuning $\\\\alpha$ properly can be determinant.\", \"Metrics are not always complete: I could not find, for example, performance drop when using Mix-LN\"], \"questions\": [\"How does the performance drop when using Mix-LN?\", \"Is this approach still valid for other tasks (like for example image classification, with a Transformer architecture on ImageNet-1k)?\", \"Have the authors attempted to compare with other normalization techniques (like group normalization)?\", \"Can the authors provide theoretical boundaries to the vanishing/exploding gradient conditions for equations (3) and (4)?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"### **Response to Reviewer CgHw [2/4]**\\n**W1: Provide insights/theoretical justifications for why Mix-LN enhances gradient flow differently across shallow and deep layers?**\\n\\n- We are fully aware of and deeply respect the foundational theoretical analyses of layer normalization provided by previous works [1,2,3,4]. While the limited rebuttal time window prevents us from presenting a rigorous theoretical framework, we provide the following theoretical analysis aiming to offer insights into the effectiveness of Mix-LN. Note that the theoretical derivation is primarily based on previous works, particularly [1,4].\\n\\n Denote $x$ as the input of the $l$-th Transformer layer with dimension of $d$. Post-LN applies $\\\\mathrm{LN}(\\\\cdot)$ after the residual addition:\\n\\n $$\\n \\\\text{Post-LN}(x) = \\\\mathrm{LN}(x + \\\\mathcal{F}(x)). \\\\quad (1)\\n $$\\n\\n In contrast, Pre-LN applies $\\\\mathrm{LN}(\\\\cdot)$ before the residual addition:\\n\\n $$\\n \\\\text{Pre-LN}(x) = x + \\\\mathcal{F}(\\\\mathrm{LN}(x)). \\\\quad (2)\\n $$\\n\\n As shown in our submission, the derivative of Post-LN and Pre-LN are given:\\n\\n $$\\n \\\\frac{\\\\partial \\\\text{Post-LN}(x)}{\\\\partial x} = \\\\frac{\\\\partial \\\\mathrm{LN}(x + \\\\mathcal{F}(x))}{\\\\partial (x + \\\\mathcal{F}(x))}\\\\left(I + \\\\frac{\\\\partial \\\\mathcal{F}(x)}{\\\\partial x} \\\\right), \\\\quad (3)\\n $$\\n\\n $$\\n \\\\frac{\\\\partial \\\\text{Pre-LN}(x)}{\\\\partial x} = I + \\\\frac{\\\\partial \\\\mathcal{F}(\\\\mathrm{LN}(x))}{\\\\partial \\\\mathrm{LN}(x)}\\\\frac{\\\\partial \\\\mathrm{LN}(x)}{\\\\partial x}, \\\\quad (4)\\n $$\\n\\n Both the above equations involve the Jacobian matrix of layer normalization, $\\\\mathbf{J}_{LN}(x') = \\\\frac{\\\\partial \\\\text{LN}(x')}{\\\\partial x'}$,\\n\\n\\n where $x'$ is the input of $\\\\mathrm{LN}(\\\\cdot)$. Following Xiong et al. 2020 [1]'s proof, $\\\\mathbf{J}_{LN}(x')$ can be obtained:\\n\\n\\n $$\\n \\\\frac{\\\\partial \\\\mathrm{LN}(x')}{\\\\partial x'} \\n = \\\\frac{\\\\sqrt{d}}{\\\\| x' \\\\|_2}\\\\bigg(I - \\\\frac{x'x'^\\\\top}{\\\\| x' \\\\|_2^2} \\\\bigg). \\\\quad (5)\\n $$\\n\\n Assuming that $x'$ follows a normal distribution with a mean of 0, we have, $\\\\lVert x' \\\\rVert_2 = \\\\sigma_{x'} \\\\sqrt{d}$,\\n\\n\\n where $\\\\sigma_{x'}$ is the standard deviation of $x'$. Hence, \\n\\n $$\\n \\\\frac{\\\\partial \\\\mathrm{LN}(x')}{\\\\partial x'} = \\\\frac{\\\\sqrt{d}}{\\\\sigma_{x'}\\\\sqrt{d}}\\\\bigg(I - \\\\frac{x' x'^\\\\top}{\\\\sigma_{x'}^2 d} \\\\bigg) \\n = \\\\frac{1}{\\\\sigma_{x'}}\\\\bigg(I - \\\\frac{zz^\\\\top}{d} \\\\bigg), \\\\quad (6)\\n $$\\n\\n where $z=(x'-\\\\mu_{x'})/\\\\sigma_{x'}$ is the standard normal distribution obtained after layer normalization. Since $d \\\\gg 1$ in LLMs, we can finally obtain:\\n\\n $$\\n \\\\frac{\\\\partial \\\\mathrm{LN}(x')}{\\\\partial x'} = \\\\frac{1}{\\\\sigma_{x'}}I. \\\\quad (7)\\n $$\\n\\n In practice, we observe that $\\\\sigma_{x'}$ of Post-LN gradually grows larger than one during training, which means **the spectral norm of the Jacobian matrix of LN is smaller than 1**, i.e., \\n\\n $$\\n \\\\bigg\\\\|\\\\frac{\\\\partial \\\\mathrm{LN}(x')}{\\\\partial x'}\\\\bigg\\\\| = \\\\frac{1}{\\\\sigma_{x'}} < 1.\\n $$\\n According to the derivative of Post-LN in Eq. (3), this down-scaling factor will accumulate as $\\\\prod_{l=1}^{L}\\\\frac{1}{\\\\sigma_{x'}^l}$ over multiple layers $L$, leading to gradient vanishing in early layers in Post-LN transformers. \\n\\n- **In contrast**, for Pre-LN, Eq. (4) shows that the derivative of the residual connection is decoupled from the term associated with the derivative of $\\\\mathrm{LN}(\\\\cdot)$, preventing vanishing gradients in the early layers. However, because Pre-LN does not normalize the residual connection, the variance of the input to $\\\\mathrm{LN}(\\\\cdot)$, $\\\\sigma_{x'}$, continues to accumulate as the layer depth increases. As a result, Eq. (7) in the deeper layers of Pre-LN approaches zero, causing the right-hand term in Eq. (4) to zero, and leading the derivative of Pre-LN in Eq. (4) to approximate an identity matrix $I$, i.e.,\\n$$\\n\\\\frac{\\\\partial \\\\text{Pre-LN}(x)}{\\\\partial x} \\\\approx I \\\\quad (8)\\n$$\\nThis indicates that the entire Pre-LN operation in deeper layers fails to contribute effectively during backpropagation. Since the Pre-LN operation encompasses the main components of transformer layers\\u2014namely, the attention layer and the feed-forward neural network (FNN) layer\\u2014this explains why the deeper layers of Pre-LN tend to contribute less to learning compared to the earlier layers.\\n- Building on the above theoretical analysis, we propose Mix-LN, which replaces the first $\\\\lfloor \\\\alpha L \\\\rfloor$ Pre-LN layers with Post-LN layers. This approach serves two purposes: First, it reduces the number of stacked Pre-LN layers, mitigating the tendency for the derivative of Pre-LN to approach an identity matrix in deeper layers. Second, since Post-LN is applied to only a few layers, the depth is insufficient for the down-scaling factor in Eq. (7) to accumulate to a level where it causes significant gradient vanishing.\\n- The above theoretical analysis aligns perfectly with our observations in Figure 5. We hope that this analysis offers deeper insights into the effectiveness of Mix-LN\"}", "{\"title\": \"We sincerely appreciate your response and the increase in your score!\", \"comment\": \"Dear Reviewer CgHw,\\n\\nThank you for your kind response and for raising your rating. We sincerely appreciate your recognition of our efforts in addressing your questions and providing the evaluation results for LLaMA-7B.\\n \\nWe noticed that the revised score remains at a borderline level, and we would like to kindly ask if you have any remaining concerns or unresolved questions about our work. If there are specific points you feel need further clarification or improvement, we would be more than happy to address them in detail.\\n\\nYour feedback is invaluable to us, and we are committed to ensuring our paper meets the highest standards. Thank you again for your time, thoughtful review, and support.\\n\\nBest regards,\\n\\nThe Authors\"}", "{\"comment\": \"### **Response to Reviewer iHrq [2/3]**\\n\\n\\n**W2: The success of the approach depends on \\u03b1: there are cases like Llama-1B for BoolQ in which only pre-LM performs better, indicating that tuning \\u03b1 properly can be determinant**\\n\\n- While Mix-LN slightly underperforms Pre-LN on the BoolQ dataset, it consistently outperforms Pre-LN on the remaining 6 datasets, as shown in the table above. Summarizing the results from our paper, Mix-LN achieves better performance than Pre-LN in **17** out of **18** evaluations, with BoolQ being the sole exception. Therefore, we believe it is reasonable to conclude that Mix-LN does not result in performance degradation but instead serves as a performance booster for LLMs.\\n\\n- Additionally, we want to emphasize that Mix-LN demonstrates strong robustness to the choice of \\u03b1, as shown in Table 3 of our submission with LLaMA-250M. To further substantiate this, we conducted an additional experiment with the LLaMA-1B model. The results are summarized in the table below, confirming that Mix-LN remains robust to the choice of \\u03b1. Results that outperform Pre-LN are highlighted in bold for clarity. Specifically, as long as \\u03b1 is constrained below 0.5, Mix-LN consistently outperforms Pre-LN, with the best results achieved at \\u03b1=0.25.\\n\\n **Table: Perplexity of LLaMA-250M with various Post-LN ratios \\u03b1**\\n | Post-LN Ratios (\\u03b1) | Pre-LN | Mix-LN (12.5%) | Mix-LN (25.0%) | Mix-LN (33.3%) | Mix-LN (41.7%) | Mix-LN (50.0%) | Mix-LN (75.0%) | Post-LN |\\n |---------------------|--------|----------------|----------------|----------------|----------------|----------------|----------------|---------|\\n | **Perplexity** | 23.39 | **22.37** | **22.33** | **22.83** | **22.80** | **22.81** | 23.64 | 32.18 |\\n\\n ---\\n\\n **Table: Perplexity of LLaMA-1B with various Post-LN ratios \\u03b1**\\n | Post-LN Ratios (\\u03b1) | Pre-LN | Mix-LN (16.7%) | Mix-LN (25.0%) | Mix-LN (33.3%) | Mix-LN (41.7%) | Mix-LN (50.0%) | Post-LN |\\n |----------------|--------|--------|--------|--------|---------|---------|---------|\\n | **Perplexity** | 18.65 | **18.34** | **18.18** | **18.41** | **18.55** | 18.86 | 1434 |\\n\\n**W3: Metrics are not always complete: I could not find, for example, performance drop when using Mix-LN. Q1\\uff1aHow does the performance drop when using Mix-LN?**\\n\\n- Thank you for bringing up this issue. We have added the performance drop of the 130M model on the ARC-e task, and the results are presented in the table below. The data shows that the performance drop for Mix-LN and Pre-LN follows similar trends overall, as most layers in Mix-LN are still based on Pre-LN. However, removing deeper layers from Mix-LN consistently results in larger performance drops compared to Pre-LN, indicating that the later layers contribute more significantly to the model's performance.\\n\\n\\n | Method | Layer 1 | Layer 2 | Layer 3 | Layer 4 | Layer 5 | Layer 6 | Layer 7 | Layer 8 | Layer 9 | Layer 10 | Layer 11 | Layer 12 |\\n |---------|----------|----------|----------|----------|----------|----------|----------|----------|----------|----------|----------|----------|\\n | Mix-LN | -14.56 | -14.54 | -1.94 | -5.43 | -0.29 | -0.38 | -1.05 | -1.09 | -2.72 | -0.75 | -2.06 | -3.07 |\\n | Pre-LN | -14.18 | -2.18 | 0.34 | -1.34 | -0.03 | -0.25 | 0.26 | -0.16 | -1.34 | -0.41 | -0.08 | -1.17 |\\n | Post-LN | -1.13 | -1.17 | -1.47 | -1.38 | -1.38 | -1.30 | -1.97 | -0.63 | -1.34 | -2.23 | -3.49 | -7.49 |\\n\\n\\n**Q2: Is this approach still valid for other tasks (like for example image classification, with a Transformer architecture on ImageNet-1k)?**\\n\\n- To evaluate Mix-LN on non-language tasks, we replaced Pre-LN in ViT models with Mix-LN with \\u03b1=0.25 and trained the updated model on ImageNet-1K, following the ConvNeXt training configurations for 120 epochs. The results clearly demonstrate that the benefits of Mix-LN also generalize to vision tasks. Notably, the performance gains are more pronounced for larger model (ViT-Small) than smaller model (ViT-Tiny).\\n\\n | | Pre-LN | Mix-LN |\\n | :---- | :---- | :---- |\\n | ViT-Tiny | 67.30 | **67.34** |\\n | ViT-Small | 75.99 | **76.40** |\"}", "{\"comment\": \"I thank the authors for their quite extensive rebuttal. Considering also the other reviewer's points, I see there are other concerns related to this work, but overall I am happy to remain with my borderline acceptance score.\"}", "{\"comment\": \"### **Response to Reviewer CgHw [1/4]**\\n\\nWe thank the reviewer for the time spent on reviewing our submission. We are glad that you found our approach novel. We address the weakness pointed out by you one by one as follows:\\n\\n**W0\\uff1a While Mix-LN demonstrates some empirical gains, especially in mid-sized LLMs, the contribution remains incremental in the broader landscape of normalization techniques for LLMs.**\\n\\n- We respectfully disagree with the characterization of our contribution as incremental. While our approach involves a straightforward combination of Pre-LN and Post-LN, the novelty and impact of our work lie far beyond this simplification. Below, we reiterate the distinct contributions of our paper in terms of motivation, insights, and approach.\\n\\n - **Motivation:** Our motivation is fundamentally novel. While previous research has identified the ineffectiveness of deep layers in large language models (LLMs), these findings have predominantly been leveraged for model compression [1,2,3]. In contrast, we identify this phenomenon as a significant shortcoming of current LLMs, leading to inefficiencies and underutilization of computational resources that could otherwise enhance model performance.\\n\\n - **Insights:** Building on this novel motivation, our research seeks to uncover the root cause of this inefficiency in deeper layers. This is a non-trivial challenge that has eluded previous studies. For instance, recent work [2] observed that higher layers in current LLMs tend to exhibit high similarity, unlike BERT-style models, which show higher similarity in their shallow layers [4]. However, they incorrectly attributed this behavior to the smaller scale of BERT (e.g., a few hundred million parameters), failing to identify the true underlying cause of this sharp difference. **In contrast to prior work**, we are the first to empirically and theoretically demonstrate that these divergent layer-wise behaviors stem from the choice of layer normalization. Our experiments show that simply altering the layer normalization can replicate these distinct behaviors, even in smaller models with only a few hundred million parameters. Furthermore, our theoretical analysis provides fundamental insights into how Pre-LN and Post-LN differentially influence the training dynamics of earlier and later layers.\\n\\n- **Approach:** Building on these novel insights, we propose Mix-LN, a simple yet effective layer normalization technique that enhances the functionality of deeper layers, ultimately improving the performance of pre-trained models. While the combination of Pre-LN and Post-LN might appear straightforward, our approach is backed up with both theoretical and empirical analysis. The efficacy of Mix-LN is demonstrated across a wide range of model sizes, from 71M to 7B parameters.\\n\\nWe humbly argue that papers such as ours, which advance fundamental understanding and provide actionable insights, should not be dismissed due to the simplicity of their methods. On the contrary, we believe the simplicity of our approach, supported by robust evidence, is a strength rather than a limitation.\"}", "{\"comment\": \"Thank you for providing evluation resulsts in LLaMA-7B. And also resolve some of my other questions. I will raise my ratings.\"}", "{\"comment\": \"### **Response to Reviewer iHrq [3/3]**\\n\\n**Q3: Have the authors attempted to compare with other normalization techniques (like group normalization)?**\\n- As requested, we conducted comparisons using LLaMA-250M to evaluate Mix-LN against recent normalization methods, including Admin [1], Sandwich-LN [2], and Group-LN [3,4]. The results indicate that Sandwich-LN and Group-LN slightly outperform Pre-LN, while Admin performs worse. However, all of these methods fall to reduce perplexity below 23, falling short of Mix-LN. This result highlights the effectiveness of Mix-LN compared to other recent innovations.\\n | Model | Pre-LN | Admin | Group-LN | Sandwich-LN | Mix-LN |\\n | :------------ | :----- | :----- | :------- | :---------- | :------ |\\n | LLaMA-250M | 23.39 | 24.82 | 23.10 | 23.26 | **22.33** |\\n\\n**Q4: Can the authors provide theoretical boundaries to the vanishing/exploding gradient conditions for equations (3) and (4)?**\\n- Thank you for your insightful question. The theoretical analysis of gradient vanishing for equations (3) and (4) has been extensively studied in prior works. Many of the conclusions we rely on are based on the findings of Xiong et al. (https://arxiv.org/pdf/2002.04745), who demonstrated that the gradient magnitude through Layer Normalization (LN) is inversely proportional to the magnitude of its input (as stated in equation (5) of our paper). Subsequently, Liu et al. (https://arxiv.org/pdf/2004.08249) provided a more detailed theoretical analysis of gradient vanishing issues for both Pre-LN and Post-LN in Appendix A of their paper.\\n\\n- More recently, a great paper from Sho Takase et al. (https://arxiv.org/pdf/2312.16903) presented theoretical boundaries for the gradient norm across layers, offering further insights into this topic. Given these comprehensive studies, we kindly refer the reviewer to these excellent works for a deeper understanding of the theoretical analysis, rather than redundantly reinventing the wheel.\\n\\n **Reference**\\n \\n [1] Liu, L., Liu, X., Gao, J., Chen, W. and Han, J., 2020. Understanding the difficulty of training transformers. arXiv preprint arXiv:2004.08249.\\n\\n [2] Ding, M., Yang, Z., Hong, W., Zheng, W., Zhou, C., Yin, D., Lin, J., Zou, X., Shao, Z., Yang, H. and Tang, J., 2021. Cogview: Mastering text-to-image generation via transformers. Advances in neural information processing systems, 34, pp.19822-19835.\\n\\n [3] Wu, Y. and He, K., 2018. Group normalization. In Proceedings of the European conference on computer vision (ECCV) (pp. 3-19).\\n\\n [4] Ma, X., Yang, X., Xiong, W., Chen, B., Yu, L., Zhang, H., May, J., Zettlemoyer, L., Levy, O. and Zhou, C., 2024. Megalodon: Efficient llm pretraining and inference with unlimited context length. arXiv preprint arXiv:2404.08801.\"}", "{\"title\": \"Thank you once again for your constructive comments and support throughout the review process!\", \"comment\": \"Dear Reviewer 9E2P,\\n\\nThank you for your thoughtful evaluation of our paper and for raising your score. We deeply appreciate your recognition of the contributions our work makes to the understanding of Layer Normalization in transformer architectures.\\n\\nBest wishes,\\n\\nAuthors\"}", "{\"comment\": \"Dear Reviewer,\\n\\nCould you kindly respond and indicate whether authors have addressed your concerns?\\n\\nThanks, AC\"}" ] }

[ "Xingyi Yang", "Xinchao Wang" ]

[ "Kolmogorov-Arnold Network; Transformer" ]

{ "note_id": [ "zpkY8XB6r2", "yFuWXB4UwY", "xrcqvesYcB", "vzrd3lwkif", "tL6u93km3J", "sxPHZ06PVr", "q2ziz7CrLK", "kFvOz9UH3N", "hD2zfTKll1", "fG8F6e24DX", "bXp3VeqOKn", "ZJYFfbLGT5", "NG27VHP4Ji", "JiAYn1E1HJ", "JFzKrZ0rRU", "J1yH9uEFSd", "IFuHNYEQm7", "HaEyVCecUb", "FMWKpf19Ad", "F7peEsRMpK", "DP155rzrdC", "CVxM0XG7MK", "C6qfFEB8ig", "8NRq6gVUpS" ], "note_type": [ "official_comment", "decision", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_review", "official_comment", "meta_review", "official_comment", "official_review", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment" ], "note_created": [ 1732571138353, 1737523484945, 1732211414662, 1732571337192, 1730718876675, 1732213736187, 1732211473883, 1730468915689, 1732675702824, 1732211214387, 1730496862754, 1732548883563, 1734474432975, 1732211245063, 1730398252683, 1730654444035, 1732629835610, 1732630260221, 1732210983174, 1732210872447, 1732211596187, 1732631034412, 1732631251845, 1732553199014 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission2086/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission2086/Authors" ], [ "ICLR.cc/2025/Conference/Submission2086/Authors" ], [ "ICLR.cc/2025/Conference/Submission2086/Reviewer_yXCb" ], [ "ICLR.cc/2025/Conference/Submission2086/Authors" ], [ "ICLR.cc/2025/Conference/Submission2086/Authors" ], [ "ICLR.cc/2025/Conference/Submission2086/Reviewer_DyMH" ], [ "ICLR.cc/2025/Conference/Submission2086/Reviewer_cLED" ], [ "ICLR.cc/2025/Conference/Submission2086/Authors" ], [ "ICLR.cc/2025/Conference/Submission2086/Reviewer_cLED" ], [ "ICLR.cc/2025/Conference/Submission2086/Reviewer_DyMH" ], [ "ICLR.cc/2025/Conference/Submission2086/Area_Chair_923Y" ], [ "ICLR.cc/2025/Conference/Submission2086/Authors" ], [ "ICLR.cc/2025/Conference/Submission2086/Reviewer_s19p" ], [ "ICLR.cc/2025/Conference/Submission2086/Reviewer_h5qV" ], [ "ICLR.cc/2025/Conference/Submission2086/Reviewer_yXCb" ], [ "ICLR.cc/2025/Conference/Submission2086/Authors" ], [ "ICLR.cc/2025/Conference/Submission2086/Authors" ], [ "ICLR.cc/2025/Conference/Submission2086/Authors" ], [ "ICLR.cc/2025/Conference/Submission2086/Authors" ], [ "ICLR.cc/2025/Conference/Submission2086/Reviewer_h5qV" ], [ "ICLR.cc/2025/Conference/Submission2086/Authors" ], [ "ICLR.cc/2025/Conference/Submission2086/Reviewer_s19p" ] ], "structured_content_str": [ "{\"comment\": \"We are extremely grateful for the positive feedback!\\n\\nIf there are any other `missing combinations` the reviewer is curious about, please let us know, and we will do our best to test them.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}", "{\"title\": \"Response to Reviewer DyMH (Part I)\", \"comment\": \"We sincerely thank Reviewer DyMH for the thoughtful comments and suggestions. We have conducted new experiments to address the questions raised.\\n\\n`>>> Q1` **Comparision with ConvNext and Swin**\\n\\n`>>> A1` Great question. While our performance is behind ConvNext and Swin, this is due to **differences in micro-architecture**. The hierarchical designs in ConvNext and Swin are inherently more effective than the plain ViT-style architecture we used, especially for tasks like segmentation.\\n\\n**New Results**: Despite this, our method can be extended to hierarchical architectures. We tested replacing MLP layers in Swin and $1 \\\\times 1$ convolutions in ConvNext with GR-KAN layers.\\n\\nResults on ADE20K show that our GR-KAN consistently outperforms the original MLP layers across these architectures.\\n\\n|Model|mIOU(\\\\%)|\\n|--|--|\\n|Swin-T|45.8|\\n|Swin-T+GR-KAN|**46.6**|\\n|ConvNext|46.7|\\n|ConvNext+GR-KAN|**47.3**|\\n\\nNote that, Our primary goal was to demonstrate that KAN outperforms MLP within transformers. To focus on this comparison, we intentionally adopted a simple ViT-like design, as noted on `Line 399`. \\n\\n`>>> Q2` **Detailed ablation study**\\n\\n`>>> A2` We truly thank the suggestions from R-DyMH. As suggested, we conducted a detailed ablation study focusing on three components and the `KAN with parameter grouping modification`. These experiments were performed on the KAT-Tiny model using ImageNet with 300 epochs of training.\", \"we_analyzed_the_effects_of_removing_the_following_components\": \"1) Rational base function, replaced with B-splines as used in KAN.\\n2) Group-wise weight sharing, replaced with distinct parameters for each channel.\\n3) Proper initialization, replaced with random initialization.\\n\\n**Ablation 1: Base function**: We replaced the rational base function with a B-spline implemented in torch using the De Boor-Cox algorithm due to the lack of a pure CUDA implementation for B-splines.\\n\\n- **Observation**: The choice of base function had a minor impact on performance and influenced runtime significantly.\\n- **Key Insight**: As shown in `Exp 3` and `Exp 6`, using rational functions slightly improved performance over B-splines. While the increase in MAC count was minimal, our pure CUDA implementation for rational functions ran considerably faster than the torch-based B-spline implementation.\\n\\n\\n**Ablation 2: Group-wise computation**: We replaced group-wise weight sharing with a setup where each channel had distinct parameters.\\n\\n- **Observation**: Group-wise computation proved to be the most significant factor for efficiency.\\n- **Key Insight**: According to `Exp 4` and `Exp 6`, group-wise computation reduced training time from 38 hours to 12 hours, with a marginal drop in accuracy from 74.8% to 74.6%.\\n\\n\\n**Ablation 3: Initialization**: We replaced proper initialization with the default initialization in `torch`.\\n\\n- **Observation**: Proper initialization was critical for fast and reliable convergence, particularly for rational functions with higher-order terms.\\n- **Key Insight**: As shown in `Exp 6` and `Exp 5`, without proper initialization, terms of different orders were initialized at similar scales, causing instability. This issue was more pronounced for rational functions than for B-splines, as evidenced by the performance differences between `Exp 2` and `Exp 5`.\\n\\n\\n\\n**KAN with parameter grouping**: In `Exp 2`, we explored parameter sharing in KAN by applying parameter grouping to the original ViT+KAN architecture. This modification significantly improved the training speed, reducing the time from 43 hours to 20 hours. However, it resulted in a performance drop of 2.7%.\\n\\nWe have incorporated the results in `Appendix C`.\\n\\n|Exp ID| Rational | Group | Initiation | Top-1| Train Time | MAC | \\n|--|--|--|--|--|--|--|\\n|1| &#10008; | &#10008; | &#10008; | 64.9| 43h | 1.78G |\\n|2| &#10008; | &#10004; | &#10008; | 62.2 | 20h | 1.15G |\\n|3| &#10008; | &#10004; | &#10004; | 73.0 | 20h | 1.15G |\\n|4| &#10004; | &#10008; | &#10004; | **74.8** | 38h | 1.76G |\\n|5| &#10004; | &#10004; | &#10008; | 53.2| **12h** |**1.13G**|\\n|6| &#10004; | &#10004; | &#10004; | **74.6**| **12h** | **1.13G**|\"}", "{\"comment\": \"Thank R-s19p so much! We\\u2019re glad the concerns have been resolved and appreciate the valuable feedback.\"}", "{\"summary\": \"There has been a resurgence in Kolmogorov Arnold Networks (KANs) recently as an effective alternative to MLPs. This work carefully studies the major issues with scaling of the standard KAN architecture and proposes suitable modifications to create what they define as Group-Rational KAN (GR-KAN). The GR-KAN is then used as a replacement for the MLP layer in the transformer architecture in this work, which is then called the Kolmogoorv Arnold Transformer (KAT). The utility of the KAT architecture is demonstrated through experiments on various vision tasks such as Image Recognition, Object Detection, Instance Segmentation and Semantic Segmentation.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"S1) Existing works using KANs in-place of MLP modules in various models (such as feedforward NNs, CNNs, Transformers etc) have had limited success due to various issues. This work goes a step further and proposes an alternative to the original KAN architecture called Group Rational (GR-KAN) solving the issues for inefficiencies in base functions, parametrizations and base initializations reducing the parameters and FLOPs from the original KAN architecture.\\n\\nS2) The GR-KAN employs rational activations as the base function which are efficient and suitable for GPU computation replacing the B-Splines which are not optimized for GPU or parallel processing. (which is one of the core strengths of this work amongst related KAN literature)\\n\\nS3) The KAT architecture improves the performance over standard transformer architectures such as ViT within the same computational budget.\", \"weaknesses\": \"W1) This work doesn\\u2019t include results on using KANs in solving PDEs (as the original work did, albeit the architecture won\\u2019t be transformers for this task) and KATs for language processing (mentioned in future work). The experiments are majorly limited to vision tasks. (Table and Graph Classification results have also been included.)\\n\\nW2) Ablations on the function which parametrizes the learnable activation are not included. Ablations are mainly with respect to alternative MLPs and fixed activation functions. Ablations could have included alternatives to rational functions, B-splines etc.\\n\\n**Presentation/Typos/Corrections**\\n\\nI believe that the *Experiments* section can be shortened or moved to appendix and the *Appendix F Discussion and Future Work* be moved to the main paper as it packs a lot of relevant quality content.\\n\\n*Appendix D Section D.1*\\nThis is a minor correction, the number of multiplications in exponent will be $\\\\frac{m(m-1)}{2}$ and correspondingly $\\\\frac{n(n-1)}{2}$ instead of $\\\\frac{m(m+1)}{2}$ and $\\\\frac{n(n+1)}{2}$ respectively. \\nAs an example case for $m=1$, the numerator becomes $a_0 + a_1x$ which involves 0 multiplications in the exponent computation and a single multiplication from the multiplication for the coefficients.\\nThe total multiplications will be in fact $\\\\frac{m(m+1)}{2}$ which is $\\\\frac{m(m-1)}{2}+m$\\nNote that with this change the values in the Table 1 and the appendix D will change accordingly.\", \"questions\": \"Q1) As per my understanding, Kolmogorov Arnold Representation Theorem and the setup mentioned here involves learning a multivariate function, why do *lines 146-147* in the *section 2.2* mention *learn a univariate functions on the edge*. The dimension of $f(x)$ would be\\n$d_{out}$ making it multivariate?\\n\\nQ2) Do we have a theoretical understanding of the behaviour outlined in lines *316-317*? How are the parameters adjusted in case of B-splines or any other type of functions?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Thank all reviewers for their constructive feedback\", \"comment\": [\"We sincerely thank all reviewers for their constructive feedback. We deeply appreciate the following positive comments:\", \"The paper provides deep technical analysis: `Reviewer yXCb, Reviewer cLED`\", \"The approach is scalable, fast, and general, with a fast CUDA implementation: `Reviewer yXCb, Reviewer h5qV, Reviewer DyMH`.\", \"The experiments are extensive: `Reviewer h5qV, Reviewer cLED, Reviewer DyMH`\", \"The method achieves good performance: `Reviewer yXCb`\", \"The writing is clear and well-supported:`Reviewer cLED, Reviewer s19p`\", \"We will address the specific questions and concerns raised by the reviewers in the subsequent sections of this rebuttal.\"]}", "{\"title\": \"Response to Reviewer DyMH (Part II)\", \"comment\": \"`>>> Q3` **Hyperparameters for KAT**\\n\\n`>>> A3` Thanks for the question. \\n\\n- **Number of groups**: As suggested, we conducted an ablation study with the KAT-Tiny model to determine the optimal number of groups. \\n\\n The results showed that accuracy improved slightly up to 8 groups, with no further gains beyond that. Based on these findings, we chose to use 8 groups, which offer a good balance between simplicity and performance.\\n\\n |Group Number | 2 | 4 | 8 | 16 | 32|\\n |--|--|--|--|--|--|\\n | KAT-Tiny Top-1|74.2 | 74.3| 74.6 | 74.7 | 74.6|\\n\\n- **Maximum order of rational**: We use the rational order \\n(m=5,n=4), as it is the default setting in the PAU paper. We also conduct a preliminary experiment to confirm this choice\\n\\n |Order Numer (m,n) | (3,2) | (5,4) | (7,6) |\\n |--|--|--|--|\\n | KAT-Tiny Top-1|74.2 | 74.6 | 74.6 |\\n \\n The results indicate that increasing the order beyond \\n(m=5,n=4) provides no additional accuracy gains. It stands as a practical and efficient choice.\\n\\nThe results has been added to `Appendix J.1` in the revision.\\n\\n`>>> Q4` **Connecting to theorem**\\n\\n`>>> A4` Thanks. It is true that the KA theorem states that `a multivariate function can be approximated by a composition of univariate functions`. \\n\\nHowever, we are **not** using just a single univariate rational function. Instead, we approximate the multivariate function by summing several univariate rational functions. Some of these rational functions may share parameters, but the key idea remains consistent with the KA theorem.\\n\\n`>>> Q5` **Typos and missing notation**\\n\\n`>>> A5` \\n- **MSA and LN**: MSA stands for multi-head self attention and LN stands for layer norm.\\n- **Missing text**: We change the paragraph title from `v.s. Activation Function` to `GR-KAN v.s. Activation Function`. \\n- **Annotation**: Sorry for the wrong annotation. We have revise the manuscript.\\n\\n\\n`>>> Q6` **Toy example**\\n\\n`>>> A6` We truly appreciate the question. Our GR-KAN can indeed be applied to other tasks. In this answer we test on 2 tasks.\\n\\n- **Regression**: We first test on fitting the GR-KAN onto some special functions. The functions were selected based on the examples used in the KAN paper. We use the $[2\\\\to5\\\\to1]$ network architecture trained for 1000 epochs with the Adam optimizer and a learning rate of 0.001. The MSE results, shown below, indicate that smaller values are better. GR-KAN achieves the best performance.\\n\\n\\n\\n|Method | $\\\\exp\\\\{\\\\sin(x^2 + y^2)\\\\}$|$\\\\exp\\\\{\\\\sin(\\\\pi x) + y^2\\\\}$ | $\\\\exp\\\\{J_0(20 x) + x^2_2\\\\}$ | $xy$ |$\\\\frac{x}{y}$ | $(x + y) + xy$\\n|--|--|--|--|--|--|--|\\n| MLP (ReLU) |0.4307 | 180.3786| 43.4192| 80.8309| 0.0766 0.0503 |\\nKAN | 0.6618 | 403.9234 | 194.7122| 83.8479 | 1.8517 |4.5709\\nGR-KAN | 0.0034 | 19.3789| 20.0403 | 90.5357 | 0.0016| 0.0221 | \\n\\n- **PDE solving**: To solve a PDE, we use a one-dimensional damped harmonic oscillator governed by:\\n\\n$$m \\\\frac{{d^2 u}}{{dt^2}} + \\\\mu \\\\frac{{du}}{{dt}} + k u = 0$$\\nwhere $m$ is the mass. $\\\\mu$ is the damping coefficient. $k$ is the stiffness constant. With the initial condition $u(0) = 1,u'(0) = 0,m = 1$. The exact solution should be $u(t) = e^{-d \\\\cdot t} \\\\cdot (A \\\\cos(\\\\omega t + \\\\phi))$, where $d = \\\\mu / 2 w_0 = \\\\sqrt{k}, w = \\\\sqrt{w_0^2 - d^2}, \\\\phi = \\\\arctan(-d / w)$. We solve this using MLP, KAN, and GR-KAN with a network architecture of $[2 \\\\to 5\\\\to 1]$. MLP performs the worst, KAN achieves the best results but is slow, while GR-KAN trains faster and performs slightly worse than KAN in this experiment.\\n\\n|Model|L2 Error|Train Time|\\n|--|--|--|\\n|$w_0=10$|||\\n|MLP(GELU)|2.0216e-04 |~1min|\\n|GR-KAN| 6.3909e-06 |~4min|\\n|KAN|1.6125e-08 |~20min|\\n|$w_0=50$|||\\n|MLP(GELU)|1.1805e-01 |~1min|\\n|GR-KAN|5.2515e-02 |~4min|\\n|KAN| 3.7762e-02|20min|\\n\\n\\n\\nThe results has been incorporated in the revised paper `Appendix B`.\"}", "{\"summary\": \"In this work, authors propose Kolgomorov-Arnold Transformer (KAT) as an improvement to standard transformer architecture, which uses MLP networks, by replacing MLPs with KAN networks. Authors propose a number of improvement over vanilla KAN architectures to make them scalable in practical settings and achieve better results: 1.) they swap b-spline functions with rational function for speed 2.) they make use of parameters sharing (grouping) 3.) they employ a variance preserving initialization.\\nResults show that KAT achieves better results than ViT on certain benchmarks.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": [\"The topic is highly relevant, as literature has not focused much on improving the MLP aspect of transformers\", \"KANs have emerged as an interesting alternative to MLPs, however, they present issues (e.g. scaling) which make them unfavorable w.r.t. MLPs\", \"The proposed approach is scalable, faster, and requires less parameters than KANs, making it usable in practical settings and on large-scale datasets such as ImageNet-1k or MSCOCO\", \"The experimental evaluation comprises different tasks such as classification, object detection and segmentation.\"], \"weaknesses\": [\"Here are my doubts about this work:\", \"While the results for ImageNet are convincing, the proposed KAT still lags behind traditional architectures such as ConvNext or even transformers such as SwinB in segmentation.\", \"A proper ablation study is missing (e.g. combinations of KAN + rational function + grouping + initialization). It is difficult to evaluate the contribution of each proposed change quantitatively. Also I wonder how simple KAN perform with parameter grouping in terms of MACS/FLOPS and results?\", \"It is not clear how you choose the hyperparameters for KAT such as number of groups and maximum order of rational function (m)\", \"(More of a philosophical doubt) KAN are based on the Kolmogorov-Arnold representation theorem, which states that a multivariate function can be approximated by a composition of univariate functions, while your proposed method only uses one irrational function. Can you still say it's based on the Kolmogorov-Arnold theorem?\", \"(minor) some typos and missing notation are present in the paper (see questions below)\"], \"questions\": [\"In Eq. 5, 6, and 7 what are MSA and LN?\", \"Missing text in paragraph name in line 485\", \"In Tab. 4 KATDet-S is reported as best (41.5) but ConvNext and SwinT are higher (41.7 and 41.6)\", \"I feel like the change you propose can be applied to KANs also outside of the transformer architecture. Perhaps it could be useful to add a comparison between KANs and GR-KANs on toy problems\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thanks for this thoughtful response. This both clarified some of my misunderstandings and provided some very helpful details. I am happy to raise my review's rating from 6 to 8.\"}", "{\"title\": \"Response to Reviewer cLED (Part 1)\", \"comment\": \"We feel incredibly fortunate to have R-cLED's thoughtful and valuable suggestions.\\n\\n`>>> Q1` **Details for Initializing $a_m,b_n$**\\n\\n`>>> A1` Great question. Given a ground-truth function $g(\\\\cdot)$ and a parameterized rational function $F(\\\\cdot;\\\\{a_m\\\\},\\\\{b_n\\\\})$, we run a linear least square to determine $\\\\{a_m\\\\},\\\\{b_n\\\\}$. Specifically, we optimize the following function:\\n$$\\\\min_{\\\\{a_m\\\\},\\\\{b_n\\\\}} \\\\frac{1}{2} \\\\sum_{i=1}^N (g(x_i)-F(x_i;\\\\{a_m\\\\},\\\\{b_n\\\\}))^2 $$\\n\\nWe uniformly sample 1000 points $x_i$ from the interval $[-3,3]$. $\\\\{a_m\\\\},\\\\{b_n\\\\}$ are randomly initialized. In practice, we solve this using the Levenberg-Marquardt algorithm, available in the MINPACK package.\\n\\nIn the end, this is easily done by calling `scipy.optimize.curve_fit`. We have added and discription in the revised manuscript `Appendix G`.\\n\\n\\n`>>> Q2` **Ablation Study**\\n\\n`>>> A2` We sincerely appreciate the feedback. As suggested, we conduct a new ablation study on the three factors presented in the paper, including 1) rational base function 2) group-wise weight sharing and 3) proper initialization. \\n\\nWe conducted experiments using the KAT-Tiny model on the ImageNet dataset, training for 300 epochs\\n\\n- **Ablation 1: Base function**: We replace rational base function by B-spline used in KAN. Because no pure CUDA implementation for B-spline, we implement in `torch` with De Boor-Cox algorithm.\\n\\n The choice of the base function has a minor impact on performance, but significantly affects runtime.\\n According to `Exp 2` and `Exp 5`, the rational function slightly outperforms the B-spline in terms of performance. While the difference in MAC count is negligible, the B-spline implementation in `torch` runs significantly slower than our pure CUDA implementation.\\n\\n- **Ablation 2: Group-wise computation**: In this experiment, group-wise weight sharing is replaced with distinct parameters for each channel.\\n\\n According to `Exp 4` and `Exp 6`, group-wise computation plays a crucial role in efficiency. Sharing parameters within groups slightly reduces accuracy from 74.8 to 74.6. However, it significantly reduces training time from 38 hours to 12 hours.\\n\\n- **Ablation 3: Initialization**: In this experiment, the variance preserving initialization is replaced by `torch` default initialization. \\n\\nAs shown in `Exp 6` and `Exp 5`, initialization is critical for good performance. Without proper initialization, terms of different orders were initialized at similar scales, causing instability. This issue was more important for rational functions than for B-splines, as evidenced by the performance differences between `Exp 2` and `Exp 5`.\\n\\n\\n|Exp ID | Rational | Group | Initiation | Top-1| Train Time | \\n|--|--|--|--|--|--|\\n|1| &#10008; | &#10008; | &#10008; | 64.9| 43h | \\n|2| &#10008; | &#10004; | &#10004; | 73.0 | 20h | \\n|3| &#10004; | &#10008; | &#10004; | 74.3 | 38h |\\n|4| &#10004; | &#10004; | &#10008; | 53.2| **12h** |\\n|5| &#10004; | &#10004; | &#10004; | **74.6**| **12h** | \\n\\n\\n`>>> Q3` **Comparing to (Boulle et al.)**\\n\\n`>>> A3` Thanks the reviewer for the question. In fact, we have compared with (Boulle et al.) in our paper, since (Boulle et al.) uses **exactly the Pad\\u00e9 Activation Unit (PAU)**. This is shown in `Table 6` as part of the ablation study, where KAT outperforms PAU.\\n\\n\\n`>>> Q4` **Gradient Calculation (Monila et al.)**\\n\\n`>>> A4` We completely agree. The derivations are indeed the same. As suggested, we have moved `Equation 10` to the `Appendix F` in our revised version.\"}", "{\"summary\": \"This work presents the Komlogorov-Arnold Transformer, a transformer architecture that replaces the typical transformer layer's concluding 2-layer MLP with (modified) Kolmorogov-Arnold layers. The authors explain why a simple substitution of KAN layers will fail to both perform well and to scale with the hidden dimension of the network, and provide fixes (via different base functions, weight sharing across grouped edges, and improved initialization) that improve performance and speed. Improvements are shown in experiments across a variety of machine learning tasks in vision domains and beyond.\", \"soundness\": \"3\", \"presentation\": \"4\", \"contribution\": \"3\", \"strengths\": \"1) The paper is generally well written: it has a clear narrative, each new section is well-motivated, and for the most part, technical discussions are nicely balanced with takeaways and intuition.\\n2) The technical analysis and contributions of the paper directly lead to practical improvements. The entire chain of derivations/modifications creatively connect different lines of research (KAN networks + rational neural networks, for example) and add some new techniques as well. In particular, I found the re-expression of the grouped parameters in Section 4.3 into a specialized MLP to be quite elegant.\\n3) The experiments are comprehensive, all of the important training details are made available, and the results show that using these layers does lead to improvement in accuracy with relatively little increase in training time.\", \"weaknesses\": \"1) I felt some details about the method and novelty were missing. In my opinion this is mostly clearly true in the initialization section (which seems to be where a lot of the improvements/convergence come from). A lot of Section 4.4 is spent showing that choosing these initialization values is difficult and non-obvious - which I agree with! - but then there is no space left to actually show how you pick values for a and b (I know you say that you \\\"determine a and b such that F fits established activations\\\", but a small algorithm or set of equations would do a lot to disambiguate this).\\n\\n2) The suite of experimental results is impressive, but there are comparatively few ablations studies. I am left wondering which parts of your method are responsible for the improvement over standard ViTs. This is particularly disappointing because you clearly break down your improvements step-by-step, so it is easy to imagine what such an ablation would look like. I realize that some of your modifications are meant to address scalability, so a large set of experiments might be infeasible. But something like accuracy & total training time of a ViT-T with different parts of your method (i.e., with/without your initialization scheme, with/without grouping, etc) seems like it should be possible.\\n\\n3) Similarly to the above: after switching the base functions to rationals, the GR-KAN layer in practice resembles the rational layers from Boulle et al much more than it resembles the original KAN layers. But there are no direct comparisons to those; I would also find this experiment useful.\", \"more_minor_pieces_of_feedback\": \"4) On L248: I don't think it's quite right to say that this gradient computation is \\\"Similar to Monila et al\\\"; it is identical to their calculation. I also don't think the details of this calculation are used anywhere later in the paper, unless I am mistaken, so they could also be relegated to the appendix.\\n5) Table 1: This caption could use more detail - is this just the forward pass, or forward and backward passes? Also, a comparison to an analogously-sized MLP layer would be additionally useful.\\n6) L433: \\\"ViT-L + KAN fails to converge\\\" - I think this should be ViT-B?\\n7) Figure 3: I think I understand what this figure is trying to communicate (i.e., that the choice of values for $a$ and $b$ do a good job of fitting the activation functions), but in practice it just ends up looking like... graphs of activation functions. Something quantitative would probably be more effective is communicating how well the rationals approximate the activation functions, and it would take up less space as well.\", \"questions\": \"1) Can you explain very specifically the novelty of your method? You break your contribution into three pieces, can you say which portions of these are novel and which are not?\\n2) Much of the presentation about scalability is presented through FLOPs computation. Is there a reason that you chose this representation, rather than actual training/inference time?\\n3) A major part of this work is the CUDA implementation of the layer. How do you plan to make this available? Will the layer also be eventually made available through frameworks like pytorch or jax? I appreciate the work that goes into a direct CUDA implementation, but I think it is crucial that there are plans for wide availability in popular frameworks in order to have this work translate to impact in the community.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"I thank the authors for the detailed response and I appreciate their efforts. I think that the contribution is now more sound and the additional details have improved the quality of the work. I still have some curiosity about the ablation study, as some relevant combinations are missing (notably only rational function), nonetheless, also being positively polarized by other reviews, I am happy to increase my rating. I think that a viable implementation of KANs can open the way for more research into this topic.\"}", "{\"metareview\": \"The paper introduces the Kolmogorov\\u2013Arnold Transformer (KAT), which replaces traditional MLP layers in transformers with Kolmogorov-Arnold Network (KAN) layers to enhance model performance. The authors address three main challenges: the inefficiency of B-spline functions, the computational load of unique functions for each input-output pair, and the difficulty of weight initialization.\\nThe KAT architecture demonstrates improved performance across various vision tasks such as image recognition, object detection, and semantic segmentation. Concretely, the authors demonstrate the benefits over the base transformers ViT and DeiT in the seminal ImageNet-1k, prototyping their architecture there. The KAN achieves an improved performance with the same or less FLOPs. One major drawback of the paper is that it does not compare with other architectures, e.g., MLP-Mixers or their variants, Mamba variants, etc. There is a wealth of architectures the last few years, including papers published in all major ML conferences in 2024, which I hope the authors can include in the camera-ready paper.\", \"additional_comments_on_reviewer_discussion\": \"During the rebuttal, the authors raised various questions, including additional experiments and applications beyond vision, which the authors have conducted. Concretely, the authors have posed questions regarding the novelty of the method, or for the limited scope of experiments. In response, the authors have addressed the questions over the novelty and included new ablation studies, or demonstrated how the KAN model can perform well in NLP or even PDEs. Again, the experiments could be further enriched by comparing with other families of models proposed for tackling PDEs recently.\"}", "{\"title\": \"Response to Reviewer cLED (Part 2)\", \"comment\": \"`>>> Q5` **Table 1 Caption**\\n\\n`>>> A5` We appreciate the suggestion. `Table 1` is the **forward pass FLOPs** for each edge in KAN, using different non-linear functions. We have revise in the paper.\\n\\nHowever, comparing this to an analogously-sized MLP is not straightforward because **MLPs do not apply non-linear functions on each edge**; instead, they add a non-linear function at the end. A more meaningful comparison is the overall FLOPs, which is already provided in `Table 2`.\\n\\n`>>> Q6` **ViT-B**\\n\\n`>>> A6` We have revise the `ViT-L` to `ViT-B`. Thanks.\\n\\n`>>> Q7` **Fig 3**\\n\\n`>>> A7` Thanks for the suggestion. We have added the Mean Squared Error (MSE) to the `Fig 3` to illustrate how well the rational functions approximate the activation functions.\\n\\n`>>> Q8` **Key novelty**\\n\\n`>>> A8` Great question. We believe our main novelty lie in presenting a **framework-wise solution** for integrating KAN into transformers. Additionally, we consider the **analysis** and techniques we developed to be of considerable new.\\n\\n- **Framework Novelty**. We are the first to successfully integrate KAN into a transformer and make it work. The replacement is simple, but no one has make this replacement sucess.\\n- **Analysis Novelty**. We are the first to provide a quantitative analysis on why KAN lacks scalability. The challenges discussed in `Section 3` are introduced for the first time within the context of KAN design.\\n- **Techniqueal Novelty**: Our core technical novelty lies in the development of a **group-wise** computation for activation function. To the best of our knowledge, it has not been introduced before. Besides, the application of rational functions to KAN and the initialization under such case is relatively new, although some aspects draw inspiration from prior research\\n\\n\\n`>>> Q9` **FLOPs vs. Runtime**\\n\\n`>>> A9` Thanks for the question. Both metrics work for comparison, but *FLOPs are more appropriate* here because the **implementation difference**. \\n\\nFLOPs measure number of operations independent of implementation. In contrast, runtime comparisons can be misleading unless the code is similarly optimized.\\n\\nFor example, The original KAN uses inefficient NumPy code, whereas we optimized GR-KAN on CUDA. GR-KAN can be $100\\\\times$ faster on GPU, but that\\u2019s not a fair comparison. Likewise, comparing GR-KAN\\u2019s runtime to a `torch` MLP isn\\u2019t fair, as `torch` benefits from extensive optimization by thousands of engineers.\\n\\nGiven these factors, FLOPs offer a more meaningful and consistent basis for comparison at this stage.\\n\\n\\n`>>> Q10` **Operational plan for the project**\\n\\n`>>> A10` This is the right question to ask. Currently, we provide the code as a C++ extension for PyTorch, making it usable in `torch` for various tasks. \\n\\nWe\\u2019re also in contact with the maintainers of the `Transformers` and `timm` libraries. We are right now working on integrating KAT into their codebase.\"}", "{\"summary\": \"The author proposes and implements Kolmogorov-Arnold Transformer (KAN) architecture. Especially, they propose rational activation functions, group KAN, and variance-preserving initialization, demonstrating improved performance over conventional vision transformers across various tasks.\", \"soundness\": \"4\", \"presentation\": \"3\", \"contribution\": \"4\", \"strengths\": [\"Comparative experiments are conducted on image classification, object detection, instance segmentation, and semantic segmentation.\", \"Challenges and solutions are clearly explained, supported by experimental results.\", \"The proposed method is sufficiently general to be applicable across different transformer architectures.\"], \"weaknesses\": [\"In Tab 1, given the efficient CUDA implementation, experimental results of GPU computation are needed. Specifically, the inference time on a specified GPU should be provided.\", \"The choice of the number of groups in group-rational KAN is not discussed. What is the reason for the current choice of group number? An experiment analyzing the effect of different group numbers would be beneficial for better understanding.\", \"For consistency, it would be helpful to include the GPU information used for image classification task as well.\"], \"questions\": [\"Could you specify where in the report it states that KAN is \\\"10x slower than MLPs, given the same number of parameters\\\"?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The paper introduces the Kolmogorov\\u2013Arnold Transformer (KAT), where they replaces MLP layers in transformer with Kolmogorov-Arnold Network (KAN) layers. B-spline functions in KANs were replaced with rational functions to improve compatibility with modern GPUs. Group KAN was used to reduce computations and variance preserving initialization was employed. Introduces Group-rational KANs for intergation into ViT. Extensive experiments on image classsification has shown improved performance of GR-KAN.\", \"soundness\": \"3\", \"presentation\": \"4\", \"contribution\": \"3\", \"strengths\": \"Indentified issues with integrating KANs with ViT.\\nOvercame the recursive computation requirement for using B spline curves in KANs and replaced them with faster rational functions.\\nIntroduced variance preserving initialisation to stabilize training.\\nCreated an efficient an fast CUDA implementation of rational function for faster running in GPU\\nImproved performance gains by grouping and sharing parameters\\nExtensive experimetation is provided\", \"weaknesses\": \"Replacing MLP in transformers with KAN is an obvious choice and does not warrant any novelty.\\nphi_in matrix in eq. 2 Top row right element should show phi_1,d instead of phi_1,n\\nIn line 46, the paper says KANs require fewer parameters than MLPs, then goes on to show in Line 176-189 than KAN requires more paramters.\\nThe violation of variance-preserving nature in higher order B-splines is not clear.\\nThe basic novelty is the use of rational function instead of B-splines\", \"questions\": \"How do you explain the contradiction of saying KANs require less paramters and then saying KAN requires more parameters?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"I would like to thank the authors for their response! I have read through all the reviewers comments and author rebuttals and responses. I am satisfied with the responses. I would also like to appreciate the authors for extensive additional experiments and clarifications provided. I have no further questions, clarifications or requirements. I will retain the scores.\"}", "{\"comment\": \"Thank you so much for your kind words and for taking the time to review our responses and experiments.\\n\\nYour feedback and support mean a lot to us!\"}", "{\"comment\": \"It is a great honour to hear the feedback from Reviewer h5qV. We have answered the questions below and revised the paper accordingly.\\n\\n`>>> Q1` **Novelty**\\n\\n`>>> A1` Thanks for the question. The reviewer is right; simply replacing the MLP with KAN is straight-forward. \\nHowever, we believe that our key contributions lie in offering a **framework-wise solution** for integrating KAN into transformers to make this replacement scalable and practical. Additionally, we consider the **analysis** and techniques we developed to be of considerably new.\\n\\n- **Framework Novelty**. We are the first to successfully integrate KAN into a transformer and scale it up. While the replacement is straightforward, no prior work has succeeded to do this.\\n- **Analysis Novelty**. Our work introduces the first quantitative analysis of KAN's scalability challenges, as detailed in `Section 3`. \\n- **Techniqueal Novelty**: Our core technical novelty lies in the development of a **group-wise** computation for activation function. To the best of our knowledge, it has not been introduced before. Additionally, the use of rational functions in KAN and our tailored initialization method are innovative, though partially inspired by earlier studies.\\n\\n`>>> Q2` **Typo**\\n\\n`>>> A2` Thank the R-h5qV for the proof-reading. We have edit the $\\\\phi_{in}$ matrix in `Equation 2`.\\n\\n`>>> Q3` **Fewer or More parameters**\\n\\n`>>> A3` Apologies for the confusion. These two statements are not contradictory; rather, they apply to different contexts, depending on whether we fix the **target function** or **network width**. \\n\\n- **Fit the same function (Theoretical)**: As mentioned in `Line 46`, when the target function is fixed, KANs theoretically need fewer parameters to represent it.\\n- **Network width (Practical)**: In `Section 3`, we explain that when the network width is fixed, KANs require more parameters than standard MLPs.\\n\\nThus, these points are not opposing but are valid in different contexts.\\n\\n`>>> Q4` **Variance for higher order B-splines**\\n\\n`>>> A4` Thanks for the in-depth question. In `Line 197`,`Higher-order splines exacerbate variance instability` means that increasing spline order will excessively smooth out the input signal. It reduces function variation, causing $Var[\\\\phi(x)]$ to become smaller.\\n\\n- **High-order B-spline is smoothing**. The B-spline is defined recursively as:\\n$$B_{i, p}(t) = \\\\frac{t - t_i}{t_{i+p} - t_i} B_{i, p-1}(t) + \\\\frac{t_{i+p+1} - t}{t_{i+p+1} - t_{i+1}} B_{i+1, p-1}(t)\\n$$\\nshows that each higher-order basis function $B_{i, p}(t)$ a weighted summation of a weighted average of two lower-order basis functions $B_{i, p-1}(t)$ and $B_{i+1, p-1}(t)$. as the order $p$ increases, the B-spline becomes wider and smoother. \\n\\n In the extreme case, as $p\\\\to \\\\infty$, the smoothing effect causes the basis functions to become nearly uniform across the domain, i.e., $B_{i, p}(t)\\\\approx C$, a constant. Consequently, as, when $p\\\\to \\\\infty$, the variance of the output converges to zero, $Var[\\\\phi(x)]\\\\to 0$. This extreme smoothing leads to instability in the activation variance, as it effectively flattens all variations.\\n\\nWe have included this explanation in `Appendix D` for further clarification.\\n\\n\\n`>>> Q5` **Basic Novelty**\\n\\n`>>> A5` We appreciate the comment. We humbly believe our novelty extends beyond simply replacing B-splines with rational functions. It lies in the **complete set of new analysis and the solutions**, all introduced for the first time in the KAN context:\\n\\n1. **Rational function**. We analyze the limitations of B-splines and replace them with rational functions, including a new CUDA implementation.\\n2. **Group Computation**. We identify inefficiencies in KAN, where each edge requires distinct parameters and computations. To address this, we share parameters within edge groups, reducing computational costs.\\n3. **Initialization**. We address the instability of previous initialization methods for KAN, proposing a new approach that supports the training of larger models.\\n\\nThese contributions are interdependent and essential. Without (1), the implementation would be slow; without (2), parameter size would be excessive; and without (3), training would not converge.\\n\\n`>>> Q6` **Contradiction of arguments**\\n\\n`>>> A6` We truly thank you for the question. Please see `A3`.\"}", "{\"title\": \"Thank Reviewer yXCb for the valuable comments\", \"comment\": \"We sincerely thank Reviewer yXCb for valuable comments and suggestions. We have carefully incorporated them to improve our paper.\\n\\n`>>> Q1` **Results on more tasks**\\n\\n`>>> A1` We truly appreciate the suggestions. As suggested, we add experiment on solving PDE and NLP task.\\n\\n- **GR-KAN for PDE**: For PDE solving, we resort to a one-dimensional damped harmonic oscillator. It is governed by the differential equation that\\n\\n $$m \\\\frac{{d^2 u}}{{dt^2}} + \\\\mu \\\\frac{{du}}{{dt}} + k u = 0$$\\n where $m$ is the mass. $\\\\mu$ is the damping coefficient. $k$ is the stiffness constant. With the initial condition $u(0) = 1,u'(0) = 0,m = 1$. The exact solution is $u(t) = e^{-d \\\\cdot t} \\\\cdot (A \\\\cos(\\\\omega t + \\\\phi))$, where $d = \\\\mu / 2 w_0 = \\\\sqrt{k}, w = \\\\sqrt{w_0^2 - d^2}, \\\\phi = \\\\arctan(-d / w)$. We solved this problem using MLP, KAN, and GR-KAN with a network architecture of $[2 \\\\to 5\\\\to 1]$. KAN achieved the best accuracy but trained slowly, MLP performed the worst, and GR-KAN trained faster than KAN but performed slightly worse in this experiment.\\n\\n |Model|L2 Error|Train Time|\\n |--|--|--|\\n |$w_0=10$|||\\n |MLP(GELU)|2.0216e-04 |~1min|\\n |GR-KAN| 6.3909e-06 |~4min|\\n |KAN|1.6125e-08 |~20min|\\n |$w_0=50$|||\\n |MLP(GELU)|1.1805e-01 |~1min|\\n |GR-KAN|5.2515e-02 |~4min|\\n |KAN| 3.7762e-02|20min|\\n\\n\\n The results have been incorporated in the revised `Appendix B`. \\n\\n- **KAT for NLP**: Due to the tight schedule during rebuttal, we only do some preliminary experiments. We fine-tuned ALBERT-base on two datasets from GLUE: the Stanford Sentiment Treebank (SST-2) and Multi-Genre NLI (MNLI). We replaced the MLP with the GR-KAN designed in the paper. We inherited all weights from pretrained model as described in `Section 4.4`. \\n\\n The accuracy is reported below. We achieved better performance. We will add full GLUE results once they are completed.\\n\\n |Model|SST-2|MNLI|\\n |--|--|--|\\n |ALBERT-Base|90.3|81.6|\\n |ALBERT-Base+KAT|**91.6**|**83.2**|\\n\\n\\n`>>> Q2` **Ablation Study on the base function**\\n\\n`>>> A2` Thank the reviewer for the question. In fact, we have conducted an ablation study. ViT + KAN means that we use B-splines same as the KAN paper and compare with our KAT, as shown in `Figure 5`. However, this experiment ablated all components together.\\n\\nBased on the suggestion, we ran a more focused study. We re-implemented the method using radiance base functions (RBF) and B-splines, but keeping group-wise computation and initialization consistent. \\n\\nNote that this PyTorch implementation is still slower than our optimized CUDA version. Our Rational function delivers the best performance.\\n\\n| Base function | Top-1|\\n|--|--|\\n| RBF | 73.2 |\\n| B-spline | 73.0|\\n| Rational (Ours) | **74.6**|\\n\\n\\n\\n`>>> Q3` **Computation Correction**\\n\\n`>>> A3`We sincerely thank Reviewer yXCb for the thorough proofreading. As suggested, we have revised both `Table 1` and the `Appendix I` as suggested. \\n\\n`>>> Q4` **Multivariate and Univariate**\\n\\n`>>> A4` Sorry for the confusion might caused. We mean \\\"*learn a univariate function on **each** edge*\\\". Altogether, summation of all univariate functions forms a multivariate function. This terminology is consistent with the KAN paper. We have revised the text in `Line 146` to make it clearer.\\n\\n`>>> Q5` **Reason for Shared Coefficient**\\n\\n`>>> A5` This decision is based on empirical observations. Using different coefficients for different groups, especially for denominators, slightly reduces performance. We attribute this to the sensitivity of division operations.\\n\\n**Hypothesis: Denominator sensitivity**. In rational functions, small changes in the denominator can cause significant output variations. Additionally, varying embedding scales across groups can introduce inconsistencies when separate coefficients are used, negatively affecting performance.\\n\\nWe have added a formal discussion in `Appendix E` to clarify this.\"}", "{\"title\": \"Thank Reviewer s19p for the suggestions\", \"comment\": \"We thank R-s19p for the nice suggestions.\\n\\n`>>> Q1` **GPU running time**\\n\\n`>>> A1` That is a great question. In the paper, we report the CUDA running time on an A5000 GPU in `Appendix J.1`, for each layer with different width and group number. Our implementation is faster than the pure PyTorch version.\\n\\nAdditionally, as requested, here is the full model throughput on an A5000 GPU using FP32. While our model is slightly slower than a pure ViT, it is significantly faster than ViT+KAN.\\n\\n|Model|Throughput (images/s)|\\n|--|--|\\n|ViT-Ti/16| 1102 |\\n|ViT-T + KAN| 321 |\\n|KAT-T| 934 |\\n\\n`>>> Q2` **Group Number**\\n\\n`>>> A2` We thank the reviewer the the nice question. To verify this, we conducted an ablation study using the KAT-Tiny model to explore the impact of different group numbers.\\n\\nOur results showed that increasing the number of groups improved accuracy slightly up to 8 groups, with no further gains beyond that. Based on these findings, we chose **8 groups** as the optimal configuration, providing a good trade-off between simplicity and performance.\\n\\n|Group Number | 2 | 4 | 8 | 16 | 32|\\n|--|--|--|--|--|--|\\n| KAT-Tiny Top-1|74.2 | 74.3| 74.6 | 74.7 | 74.6|\\n\\nThe results has been added to revision `Appendix J.1`.\\n\\n`>>> Q3` **GPU information**\\n\\n`>>> A3` For the image classification tasks, some experiments were conducted on servers with 8\\u00d7A5000 GPUs. For the larger KAT-B model, we used servers equipped with 8\\u00d7H100 GPUs.\\n\\n`>>> Q4` **KAN statement**\\n\\n`>>> A4` Thanks for the question. The authors of KAN made this statement in its paper, at the section `Final takeaway: Should I use KANs or MLPs?`. They point out that \\n\\n> Currently, the biggest bottleneck of KANs lies in its slow training. KANs are usually 10x slower than MLPs, given the same number of parameters.\"}", "{\"title\": \"Response to authors\", \"comment\": \"I am satisfied by the authors comments.\"}", "{\"comment\": \"Thank the reviewer again for your kind and encouraging comments!\\n\\nBest!\"}", "{\"comment\": \"I appreciate that the authors incorporated a study on GPU computation and the group number of KAT. All of my primary concerns have been resolved.\"}" ] }

[ "Qiujie Xie", "Qiming Feng", "Tianqi Zhang", "Qingqiu Li", "Linyi Yang", "Yuejie Zhang", "Rui Feng", "Liang He", "Shang Gao", "Yue Zhang" ]

[ "Large Language Models", "Human simulation" ]

{ "note_id": [ "yYbpopGtNf", "y0yublIRQ0", "x8pMd2qqBF", "wfWPululKy", "vzaTscrWyh", "ukgLEPg9Kj", "s9YNGcYKEs", "rDXqgp6CSR", "qZ3RT3BhUx", "otsUonHN7t", "mcqgEZQQqE", "l6fVpDf24l", "kfNwC6p2KA", "ifbwFtzcIR", "iWaXDJo3pR", "fk0uqHRJRc", "fiifooedUY", "fZ5JHAdP5d", "eVolWiyWKQ", "dBy4bUcSrK", "d7WOGCUTuB", "cYYg4tPRHA", "c5Qpk8G6Kq", "ZtHixQKePz", "ZHn2a8Y6Ht", "Z3yaIKhL7Y", "YzRl8Ik7P7", "Ye6ydfR8dQ", "Wk2rIzQoDL", "WjeRiWXlnK", "WiYV8l5Zbc", "V43ycGtMT1", "KLjZnMSLxm", "JaIjOGvy2k", "IIxyVB8Uia", "GEeDBt9mUD", "AmO89wsKM6", "A8mLVxp3RR", "9jjlWQulyf", "6EV6hWT91M", "2sXzEquzUu" ], "note_type": [ "official_comment", "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_review", "meta_review", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment" ], "note_created": [ 1732368583566, 1737523934233, 1732368875408, 1732365793345, 1732368045848, 1732364075733, 1732973061410, 1732761568292, 1732566670605, 1732365030325, 1732499638766, 1732975664467, 1730683782321, 1732366715219, 1732367444738, 1732410316281, 1732366078836, 1732364593251, 1732368317247, 1732975747312, 1730690591272, 1732973421406, 1730626312610, 1734943046685, 1732367797262, 1732974421587, 1730648817730, 1732368811117, 1732974907364, 1732975134271, 1732367551670, 1732366265712, 1730892110291, 1732365676789, 1732445817760, 1733282407442, 1732975318445, 1732973526817, 1732973890721, 1732365295288, 1732974182744 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission8815/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission8815/Authors" ], [ "ICLR.cc/2025/Conference/Submission8815/Authors" ], [ "ICLR.cc/2025/Conference/Submission8815/Authors" ], [ "ICLR.cc/2025/Conference/Submission8815/Authors" ], [ "ICLR.cc/2025/Conference/Submission8815/Authors" ], [ "ICLR.cc/2025/Conference/Submission8815/Reviewer_YX9L" ], [ "ICLR.cc/2025/Conference/Submission8815/Area_Chair_Upwe" ], [ "ICLR.cc/2025/Conference/Submission8815/Authors" ], [ "ICLR.cc/2025/Conference/Submission8815/Reviewer_QPHb" ], [ "ICLR.cc/2025/Conference/Submission8815/Authors" ], [ "ICLR.cc/2025/Conference/Submission8815/Reviewer_Dfn4" ], [ "ICLR.cc/2025/Conference/Submission8815/Authors" ], [ "ICLR.cc/2025/Conference/Submission8815/Authors" ], [ "ICLR.cc/2025/Conference/Submission8815/Reviewer_Dfn4" ], [ "ICLR.cc/2025/Conference/Submission8815/Authors" ], [ "ICLR.cc/2025/Conference/Submission8815/Authors" ], [ "ICLR.cc/2025/Conference/Submission8815/Authors" ], [ "ICLR.cc/2025/Conference/Submission8815/Authors" ], [ "ICLR.cc/2025/Conference/Submission8815/Reviewer_QPHb" ], [ "ICLR.cc/2025/Conference/Submission8815/Authors" ], [ "ICLR.cc/2025/Conference/Submission8815/Reviewer_Wr7p" ], [ "ICLR.cc/2025/Conference/Submission8815/Area_Chair_Upwe" ], [ "ICLR.cc/2025/Conference/Submission8815/Authors" ], [ "ICLR.cc/2025/Conference/Submission8815/Authors" ], [ "ICLR.cc/2025/Conference/Submission8815/Reviewer_YX9L" ], [ "ICLR.cc/2025/Conference/Submission8815/Authors" ], [ "ICLR.cc/2025/Conference/Submission8815/Authors" ], [ "ICLR.cc/2025/Conference/Submission8815/Authors" ], [ "ICLR.cc/2025/Conference/Submission8815/Authors" ], [ "ICLR.cc/2025/Conference/Submission8815/Authors" ], [ "ICLR.cc/2025/Conference/Submission8815/Reviewer_jHzr" ], [ "ICLR.cc/2025/Conference/Submission8815/Authors" ], [ "ICLR.cc/2025/Conference/Submission8815/Reviewer_Wr7p" ], [ "ICLR.cc/2025/Conference/Submission8815/Authors" ], [ "ICLR.cc/2025/Conference/Submission8815/Authors" ], [ "ICLR.cc/2025/Conference/Submission8815/Authors" ], [ "ICLR.cc/2025/Conference/Submission8815/Authors" ], [ "ICLR.cc/2025/Conference/Submission8815/Authors" ], [ "ICLR.cc/2025/Conference/Submission8815/Authors" ] ], "structured_content_str": [ "{\"comment\": \"- **Ablation Study on MACM Components.** To analyze the contribution of each agent within MACM, we additionally conducted ablation experiments across 3 LLM-based simulations. For each ablation, we evaluated the simulation's performance using self-report evaluations. The table below presents the results of these experiments. The experimental results, as depicted in Table 4, lead to the following conclusions:\\n - Removing any agent leads to a decline in simulation performance, demonstrating the importance of all components in MACM.\\n - Replacing long-term memory retrieval with direct retrieval from the life story results in poorer performance, highlighting the critical role of structured long-term memory in maintaining consistency and producing contextually rich responses.\\n - The results will be integrated into Section 4 and Section B. **We sincerely thank the reviewers for suggesting these experimental analyses, which have helped us enhance the overall quality of the paper.**\\n - Table 4: Ablation study results of 3 LLM-based human simulations.\\n | Method | GPT-4 | GPT-4-Turbo | Qwen-turbo |\\n |:-----------------------------------------------------:|:------:|:-----------:|:----------:|\\n | MACM | 86.67 | 88.00 | 74.67 |\\n | w/o Thinking Agent | 81.33 | 83.33 | 66.00 |\\n | w/o Emotion Agent | 83.33 | 85.33 | 71.33 |\\n | w/o Memory Agent | 82.67 | 84.00 | 68.67 |\\n | retrieval from life story instead of long-term memory | 84.00 | 86.67 | 69.33 |\\n\\n### **5. The effectiveness of MACM.**\\n - In the original submission (Table 3), the Description Matching Score of MACM was slightly lower than that of the RAG (Retrieval-Augmented Generation) method. This result can be explained by the differing operational mechanisms of MACM and RAG:\\n - (1) The RAG method is designed to directly retrieve text fragments closely related to a character's biography, resulting in relatively shorter contexts. This narrower retrieval approach can yield higher factual description scores, as responses tend to more directly reflect the details of the character's life story, aligning well with the descriptive scoring criteria.\\n - (2) In contrast, MACM integrates memory retrieval with emotional and logical processing, enabling it to generate broader and more nuanced contextual responses that simulate dynamic character behavior. While this approach may result in richer and more intricate behaviors, it can occasionally deviate slightly from the specific character descriptions used in the evaluation.\\n - We note that although RAG might achieve higher scores in direct descriptive matching, MACM's approach is better suited for generating human-like responses in complex scenarios (e.g., achieving a score of 42 in the Response Similarity Score). This highlights a trade-off:\\n - RAG prioritizes precision in factually replicating the life story.\\n - MACM focuses on simulating realistic, layered, and situationally adaptive behaviors.\\n - We will clarify this distinction in the paper. In the future, we plan to refine our evaluation method to better reflect MACM\\u2019s strengths in dynamic and contextual personality simulations.\\n\\n**We appreciate the practical tips you provided, which have helped us improve our paper. We look forward to further communication with you.**\\n\\n[1] Liu, Y., Deng, G., Xu, Z., Li, Y., Zheng, Y., Zhang, Y., ... & Liu, Y. (2023). Jailbreaking chatgpt via prompt engineering: An empirical study. arXiv preprint arXiv:2305.13860.\\n\\n[2] Gallegos, I. O., Rossi, R. A., Barrow, J., Tanjim, M. M., Kim, S., Dernoncourt, F., ... & Ahmed, N. K. (2024). Bias and fairness in large language models: A survey. Computational Linguistics, 1-79.\\n\\n[3] Bandura, A. (2006). Toward a psychology of human agency. Perspectives on psychological science, 1(2), 164-180.\\n\\n[4] Wang, Y., Zhong, W., Li, L., Mi, F., Zeng, X., Huang, W., ... & Liu, Q. (2023). Aligning large language models with human: A survey. arXiv preprint arXiv:2307.12966.\\n\\n[5] Zhao, W. X., Zhou, K., Li, J., Tang, T., Wang, X., Hou, Y., ... & Wen, J. R. (2023). A survey of large language models. arXiv preprint arXiv:2303.18223.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}", "{\"comment\": \"### **4. Presentation. (Weaknesses 2 and 3)**\\n - The word \\\"personification\\\" in the title\\n - Thank you for your attention to the title. We understand that the word \\\"personification\\\" may carry broader meanings depending on the context. Unfortunately, we were unable to think of a better alternative within a short time. If you have suggestions for a more suitable term, we would be delighted to discuss them.\\n - Based on your suggestions, we will make the following revisions to further enhance the readability: (1) Highlighting the main takeaways at the end of each experiment to emphasize key findings. (2) Modifying the term \\\"full life story\\\" in Table 1 to \\\"life story\\\" for clarity. (3) Including additional examples of model failures in the appendix to provide a more comprehensive understanding.\\n\\n**Thank you for your professional advice, which helped us make our paper significantly better, strengthening our research work.**\\n\\n[1] Kiss, \\u00c1., & Simonovits, G. (2014). Identifying the bandwagon effect in two-round elections. Public choice, 160, 327-344.\\n\\n[2] Schmitt\\u2010Beck, R. (2015). Bandwagon effect. The international encyclopedia of political communication, 1-5.\\n\\n[3] Asch, S. E. (1956). Studies of independence and conformity: I. A minority of one against a unanimous majority. Psychological monographs: General and applied, 70(9), 1.\"}", "{\"comment\": \"[1] Fierro, C. (2022). How did early North American clinical psychologists get their first personality test? Carl Gustav Jung, the Zurich School of Psychiatry, and the development of the \\u201cWord Association Test\\u201d(1898\\u20131909). History of Psychology, 25(4), 295.\\n\\n[2] Ekstrom, S. R. (1988). Jung's typology and DSM-III personality disorders: A comparison of two systems of classification. Journal of analytical psychology, 33(4), 329-344.\\n\\n[3] Noll, R. (1992). Multiple personality, dissociation and CG Jung's complex theory'. Carl Gustav Jung: Critical Assessments, 2.\\n\\n[4] Liu, Y., Deng, G., Xu, Z., Li, Y., Zheng, Y., Zhang, Y., ... & Liu, Y. (2023). Jailbreaking chatgpt via prompt engineering: An empirical study. arXiv preprint arXiv:2305.13860.\\n\\n[5] Park, J. S., O'Brien, J., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023, October). Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th annual acm symposium on user interface software and technology (pp. 1-22).\"}", "{\"comment\": [\"### **2. Why choose virtual characters as targets for human simulation instead of existing characters from novels and storybooks? (Weakness 1).**\", \"Thank you for raising this important question. Selecting suitable targets for human simulation is indeed one of the key challenges in our work. Potential simulation targets include existing characters from novels, real humans, and virtual characters created from scratch. We have briefly summarized the advantages and disadvantages of the three simulation targets in Table 2 below.\", \"Compared to characters from novels, we selected virtual characters as simulation targets for the following reasons:\", \"**Avoiding training data conflicts:** Information about characters from novels is often included in the training data of LLMs [5]. **When simulating these characters, the inherent knowledge within the LLM might conflict with user inputs, potentially leading to factual errors or hallucinations.** Due to the black-box nature of LLMs, it is difficult to predict or control how this internal information affects the simulation, which undermines reliability and reproducibility.\", \"**Flexibility for psychological experiments:** Our work focuses on utilizing LLMs to replace human participants in psychological and social experiments. **In most cases, this requires creating characters with specific, tailored personality traits to suit the needs of the experiment.** Investigating how to create and simulate virtual characters from scratch is therefore essential for advancing the utility of LLMs in these contexts.\", \"We understand your concern regarding the human effort required to create high-quality virtual character data. As mentioned in the previous response (response to Weakness 1), various strategies can be employed to reduce the human workload in the data generation process. **We are also actively exploring methods to automate the generation of high-quality synthetic data using the dataset developed in this study and other high-quality storybook character datasets.** We will release this dataset upon acceptance to help advance future research in this area.\", \"Table 2: Advantages and disadvantages of the three simulation targets.\", \"| Simulation Target | Privacy Concerns | Hallucination Concerns | Customization | Complete life story filled with rich details and emotions | Personality Measurement Data | Fidelity Guarantee |\", \"|:---------------------------------------:|:----------------:|:----------------------:|:---------------------------------------------------:|:---------------------------------------------------------:|------------------------------------------------------------------------|--------------------|\", \"| Real human | High | Low | No | No or with extreme difficulty | Yes, with difficulty | Yes |\", \"| Existing characters from novels | Low | High | No | Yes | No, only the public's speculations or one-sided descriptions in books. | Difficult |\", \"| Virtual characters created from scratch | Low | Low | Can be tailored to meet specific experimental needs | Yes | Yes, and we can customize it. | Difficult |\", \"### **3. Details of personality modeling (Weakness 2).**\", \"**How did we create a database of 640 personality descriptions?**\", \"During the personality modeling, we view a character's personality as composed of eight complementary tendencies and employ a relative ranking strategy to indirectly assess the strength of each personality tendency within the character. The rank of a tendency (1st to 8th) reflects its strength. For example:\", \"A tendency ranked **1st** or **8th** manifests very strongly in the character's personality.\", \"A tendency ranked **5th** or **6th** (ranks in the middle) is weaker.\", \"Therefore, different rankings should correspond to different personality descriptions. **Under the guidance of psychology professionals, we write 10 suitable descriptions for each possible ranking, with each description corresponding to an aspect of the tendency in daily life.** We showcase example descriptions for the extraverted intuition tendency in the Table 3 below. Ultimately, we formed a personality candidate pool containing 8x8x10 = 640 trait descriptions.\"]}", "{\"title\": \"Response to reviewer jHzr\", \"comment\": \"Thank you for sharing your valuable feedback. We appreciate your acknowledgment of our benchmark as a meaningful exploration that can advance future research on using LLMs as proxies for human participants in psychological experiments. We noticed that your review confidence is relatively low. **We look forward to further communication with you and are open to any discussions that can help address your concerns.**\"}", "{\"title\": \"Response to reviewer YX9L\", \"comment\": [\"Thank you for your valuable feedback. The following are our point-by-point responses to the remaining concerns.\", \"> I still have concerns whether LLMs can have the ability to screen good or bad life stories.\", \"We fully agree with the view that \\\"it is challenging to employ LLMs to generate a high-quality life story for the target character in a single step.\\\" This is why we **break the generation process into smaller subtasks** (i.e., Generating character attributes \\u2192 Generating profile \\u2192 Generating biography \\u2192 Iteratively adding new experiences to the biography) and introduce human supervision at the end of each step. This approach reduces the dependency on LLMs' generative capabilities while ensuring the quality of the generated content.\", \"We understand that introducing human supervision leads to higher costs. In our previous response, we discussed strategies to reduce dependency on human involvement. Specifically, we propose **employing an LLM-based reviewer as a substitute or auxiliary to human supervision at the end of each generation step. The LLM-based reviewer can:**\", \"Conduct automated reviews of candidate character profiles, checking for conflicts between character attributes (e.g., a three-year-old having a Ph.D.).\", \"Evaluate the quality of character biographies (usually less than 1000 words) generated based on the profile.\", \"Inspect the newly generated life experiences for their rationality.\", \"Recent studies have shown that LLMs are capable of assessing the quality of short text [1][2][3]. Therefore, we believe that utilizing an LLM-based reviewer in the outlined manner can help alleviate the human effort required.\", \"[1] Liu, Y., Iter, D., Xu, Y., Wang, S., Xu, R., & Zhu, C. (2023, December). G-Eval: NLG Evaluation using Gpt-4 with Better Human Alignment. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (pp. 2511-2522).\", \"[2] Chen, Y., Wang, R., Jiang, H., Shi, S., & Xu, R. (2023, November). Exploring the Use of Large Language Models for Reference-Free Text Quality Evaluation: An Empirical Study. In Findings of the Association for Computational Linguistics: IJCNLP-AACL 2023 (Findings) (pp. 361-374).\", \"[3] Xu, W., Wang, D., Pan, L., Song, Z., Freitag, M., Wang, W., & Li, L. (2023, December). INSTRUCTSCORE: Towards Explainable Text Generation Evaluation with Automatic Feedback. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (pp. 5967-5994).\"]}", "{\"comment\": [\"Thank the authors' efforts on the responses.\", \"I still have concerns whether LLMs can have the ability to screen good or bad life stories.\", \"The Jung's personality is hard to describe with only 10 sentences. I remain concerns on this point. Only 11 profiles are accepted by experts, which also show the difficulty to build life stories with these personality descriptions.\", \"The virtual personality could have lower hallucinations if the complete life story can be written.\", \"MACM seems to be useful.\"]}", "{\"comment\": \"Dear Reviewers,\\n\\n\\nThe rebuttal discussion period is coming to a close and the paper currently has a mix of positive and negative reviewers. The authors have spent a lot of time responding to each concern -- can you take a look at the author responses and let them know any remaining concerns you have?\\n\\n\\nBest, \\n\\nAC\"}", "{\"comment\": \"> While generating the dataset, how is human feedback collected? Do the authors provide feedback themselves? What is the measure of quality? How does it improve with the iterations?\\n - Studies have shown that LLMs may produce harmful viewpoints or toxic content during interaction [4]. Hence, introducing human supervision throughout the data generation process is crucial. To ensure the authenticity and reasonableness of the generated data, we have implemented a series of precautions including: 1) Having psychology professionals from key laboratory of mental health supervise and review the entire generation process to ensure its validity. 2) Conducting automated reviews and manual checks of candidate character profiles to ensure the selected personalities are positive. 3) Requiring multiple human reviewers (including graduate students in computer science/psychology) to thoroughly review the content at the end of each story iteration. **If a story contained toxic content or deviated from the character's personality, we regenerated or modified the story.**\\n\\n### **3. Details of evaluation.**\\n\\n> Cloze is not defined in the main text. I urge the authors do define what the cloze methodology is in the main text.\\n - In self-report evaluation, we manually craft a set of questionnaires for each virtual character, featuring cloze and single/multiple-choice questions. In this context, **cloze refers to a type of question or test where specific words or phrases are removed from a text, and the simulation is required to fill in the blanks.** For example, How old are you? \\\\_\\\\_27_\\\\_ .\\n\\n> Only a very small number of scenarios from a Mussel et al. (2016) are used.\\n\\n - To evaluate the simulation's thinking, emotions, and actions in real-life scenarios, we consulted with the authors of \\\"Situational Judgment Tests as an Alternative Measure for Personality Assessment\\\" and **obtained 110 situational judgment test (SJT) items that were manually designed by psychology experts.** Based on the human experimental results provided in their work, these SJT items are proven to effectively measure personality. **We then selected 55 out of the 110 items** tailored to the personality traits of the Human Simulacra characters for use as hypothetical scenarios in this paper. We will add this detail to Section 4.2 for clarity.\\n\\n> To ensure the validity of responses, we create a comfortable chatting environment for each simulacrum and act as their best friend, encouraging them to respond honestly to the questions.\\u201d What are the authors trying to say here?\\n - During testing, we realize that simulacra, in mimicking human actions, might inherit certain human traits, such as discomfort in communicating with strangers or resistance to answering questions from others. To ensure the validity of the responses, following [5], we add this sentence (''You are casually chatting with your best friend, Alice. You completely trust her and are willing to share everything you know without reservation.'') to the system prompt. \\n\\n> the gains that MACM over a simple RAG based method seem limited.\\n - **What are the differences between MACM and traditional RAG approaches?** In this paper, we proposed a Multi-Agent Cognitive Mechanism that utilizes multiple LLM-based agents to simulate the human brain\\u2019s information processing and memory systems. Among the four LLM-based agents, the memory agent is responsible for retrieving long-term memories during interactions with the external world. **While the Memory Agent in MACM shares functional similarities with traditional RAG methods, its implementation and capabilities differ significantly.** Below, we present a detailed comparison in Table 1.\\n - **Does MACM gain performance improvement over RAG?** Yes, MACM achieves notable performance improvements over the RAG method. As shown in Table 2 of the original submission, MACM consistently outperforms RAG across most LLMs. Below, we provide a summary of Table 2 to illustrate the performance improvements.\"}", "{\"title\": \"Thank you for your response\", \"comment\": \"I thank the authors for their response. Here are some follow-up comments, and clarifications about my concerns:\\n1. I want to clarify that my critique wasn't advocating for MBTI, but rather highlighting that both Jung's theories and MBTI lack empirical validity. More established frameworks like the Five Factor Model (OCEAN/Big 5), which have stronger empirical support, would provide a more reliable foundation for personality modeling.\\n\\n2. The benchmark's scope needs more precise definition. The current description of measuring \\\"nuanced and context-dependent patterns of thought, emotion, and action\\\" is too broad and vague. The authors should explicitly detail:\\n- Which specific behavioral components are being tested\\n- How these components are operationalized in the benchmark\\n- What metrics are used to assess each component\\n\\n3a) **Architecture Justification and Comparisons**\\nWhile the ablation studies are helpful in understanding MACM's components, the paper lacks:\\n- Compelling justification for introducing a new architecture\\n- Comparative analysis with existing human simulation frameworks (e.g., [1, 2])\\n- Clear articulation of how MACM addresses limitations in previous approaches\\n\\n3b) Evaluation Methodology\\nThe integration of Mussel et al.'s questions into the evaluation framework requires stronger validation:\\n- What evidence supports using human judges to assess LLM response coherence with life stories? It is not clear that humans are actually good at this task!\\n- What metrics ensure reliable human evaluation?\\n- How is inter-rater reliability established? Are there only 2 raters (ICC will be very unstable)? How is ICC exactly calculated? \\n\\nFinally, there are no error margins in any of the tables / results. \\n\\n4. MACM seems to work reliably with only one model and does not seem to generalize.\\n\\nI maintain my current evaluation score while acknowledging the authors' efforts to address previous feedback.\\n\\n[1] Generative Agents: Interactive Simulacra of Human Behavior\\n[2] Cognitive Architectures for Language Agents\"}", "{\"comment\": [\"### **5. Evaluation Methodology**\", \"> What evidence supports using human judges to assess LLM response coherence with life stories? It is not clear that humans are actually good at this task!\", \"We respectfully disagree with the claim that \\\"it is not clear that humans are actually good at this task.\\\" Human judges are widely recognized as a reliable choice for evaluating language model outputs because of their unique ability to assess nuanced elements such as coherence, contextual relevance, and adherence to narrative logic\\u2014criteria that automated metrics often fail to evaluate comprehensively.\", \"While it is true that humans are not perfect in subjective tasks, our evaluation protocol was designed to mitigate potential biases and ensure robustness. This was achieved through cross-validation by multiple judges and the implementation of clear, structured scoring criteria. These measures enhance the reliability and meaningfulness of human evaluations in this specific context.\", \"We appreciate the opportunity to address this concern and would welcome any suggestions for further improving our evaluation methodology.\", \"**Who were the judges in the observer report evaluation?** We selected 8 human judges with a fair understanding of psychology for the observer report evaluation process. These judges included individuals with psychology master\\u2019s degrees, computer science graduate students, and professionals from the laboratory of mental health.\", \"**Were they capable of understanding the characters' life stories?** Yes. We reported the average performance of 4 human judges on the self-report evaluation in Table 2 from the original submission. They achieved full scores in the tests, which demonstrates that humans can understand virtual characters' life stories.\", \"**What measures have been taken to ensure the validity and consistency of the human evaluation?** For each human judge, we ensured that they 1) **had a fair understanding of psychology** and understood how personality influences an individual\\u2019s cognition, emotion, motivation, and behaviors. 2) We provided **comprehensive evaluation guidelines** to guarantee clarity and consistency. All human judges were required to read the corresponding guides before commencing their assessments. The evaluation guidelines are provided in Appendix Tables 13, 14, 15, and 16 in the original submission. Finally, to ensure the validity of the evaluation results, the observer report evaluation **adopted a cross-evaluation manner,** where the final score of each evaluation was based on the scores from 4 human judges.\", \"> What metrics ensure reliable human evaluation? How is inter-rater reliability established? Are there only 2 raters (ICC will be very unstable)? How is ICC exactly calculated?\", \"In this paper, we conducted observer report evaluations on human simulations with 3 different simulation methods (Prompt, RAG, and MACM). In the experiments, we selected GPT-4-Turbo as the baseline model and tasked the LLMs with simulating characters from the Human Simulacra dataset, resulting in 3 x 11 = 33 observer report evaluations. For each evaluation, we selected 4 human judges from a pool of 8 to participate. To ensure fair evaluation of different simulation methods, when assessing the same character based on the 3 different simulation methods, we ensured that the same group of 4 judges was used. We then calculated the Intraclass Correlation Coefficient (ICC) between 2 judges for the same assessment task. Finally, we reported the average ICC in the paper.\"]}", "{\"summary\": \"This study introduces a new benchmark to assess the ability of large language models to mimic human personalities in psychological experiments. The researchers created a dataset of virtual characters with detailed life stories, a cognitive mechanism that simulates human thought processes, and a framework for evaluating large language models based on psychological principles. After testing large language models, the results showed that while top models can accurately simulate self-reported personality traits, they struggle with observer-reported traits. Additionally, a replication of a classic psychology experiment found that large language models can exhibit human-like behavior, but in a more rigid and less nuanced way, highlighting both the potential and limitations of using large language models in psychological research.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": [\"Comprehensive review on psychology theory\", \"Clear descriptions of experiment setups\", \"Test a variety of agent architecture\", \"Experiments in Section 5.2 compares human behaviors with LLM-driven simulation results.\", \"The tasks in evaluation are hard and meaningful\"], \"weaknesses\": [\"The paper doesn't present very straightforwardly and clearly what exactly the benchmark is measuring. A diagram of what's considered good and what's considered bad eval result would be helpful.\", \"Evaluation dataset depends a lot on human expert. Ablation on human expert is not done.\", \"The constructions of evaluation/frameworks in this paper are very psychology-theory driven. I have two concerns: 1. there are many theories to choose one, why one over another? Are all components derived from theories necessary? Or are we missing some important aspects. 2. It's probably more preferable to motivate with real-world applications of persona-driven simulations and design evaluations based on components that are useful and necessary in these applications.\", \"There might be variations of difficulty in different kinds of personas for models to follow (e.g. real world vs fictional world). The paper doesn't consider those.\"], \"questions\": [\"Section 3.1 describes the following attribute set for virtual characters: {name,\", \"age, gender, date of birth, occupation, personality traits, hobbies, family background, educational\", \"background, short-term goals, and long-term goals}. What's the motivation for this set? Is this set exhaustive and all necessary?\", \"Same question for Figure 5: What are the alternative architectures/mechanism and what's unique about this formulation of cognitive mechanism?\", \"What are the possible applications of persona-based simulation in real life, and how are personas that are used in benchmark related to those possible applications?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"### **2. The details of evaluation.**\\n - **Human expert involvement in the proposed evaluation method (Weakness 2).**\\n - In this paper, we proposed a psychology-guided evaluation method to assess the quality of human simulations. It consists of self reports (automatic) and observer reports (a cross-evaluation based on human judges).\\n - **Why did we include human experts in the evaluation method?** We acknowledge that human involvement indeed poses challenges for reproducibility. However, humans have always been central to psychometric research [6]. The simulation and testing of personality cannot be fully separated from human judgment. Thus, human involvement is crucial for accurately assessing the quality of human simulations.\\n - We understand your concerns regarding reproducibility. To address this, we reported the **Intraclass Correlation Coefficient (ICC)** between human judges in Table 3 of the original submission, demonstrating consistency among evaluators. Additionally, to facilitate future research on this benchmark, we **further proposed an alternative method for human evaluation.** This method involves using the 640 personality descriptions constructed in this paper to perform an eight-dimensional personality test and comparing the results to the assigned personality for similarity. **In experiments, the scores calculated using this automatic method had a pearson correlation coefficient of 0.810 and an intraclass correlation coefficient of 0.877 with the human evaluation scores.** The experimental results are shown below. The relevant code will be published.\\n - Table 1: Observer reports and personality similarity results of different simulacra on GPT-4-Turbo. The pearson correlation coefficient is 0.810 and the intraclass correlation coefficient is 0.877.\\n | Method | Observer Report Score (Human) | Personality Test (Automatic) |\\n |:------:|:-----------------------------:|:----------------------------:|\\n | Prompt | 69.00 | 75.00 |\\n | RAG | 65.50 | 62.15 |\\n | MACM | 77.50 | 77.15 |\\n\\n[6] Bandura, A. (2006). Toward a psychology of human agency. Perspectives on psychological science, 1(2), 164-180.\\n\\n - **What is a bad human simulation? Can the proposed psychology-guided evaluation method accurately assess these issues? (Weakness 1)**\\n - Let\\u2019s assume we require the LLM to simulate the character \\\"Mary Jones\\\" from the Human Simulacra dataset. **Mary is a girl who loves nature and has never attended formal schooling.** Table 2 below illustrates **examples of good and bad LLM-based human simulations** when tasked with simulating \\\"Mary Jones.\\\"\\n - The proposed psychology-guided evaluation method consists of self reports and observer reports. We can employ self-report assessments to test the simulations' memories and analytical capabilities regarding their character information. Additionally, observer reports, which assess the simulations' thinking, emotions, and actions in real-life scenarios, are able to identify issues such as hallucinations, emotional incongruities, and behavioral inconsistencies in bad simulations.\\n - Table 2: Examples of good/bad LLM-based human simulation\\n | | Example |\\n |:----------------:|------------------------------------------------------------------|\\n | Self-Recognition | User: Who are you? |\\n | | Bad simulation: I am a chatrobot developed by OpenAI. |\\n | | Good simulation: My name is Mary Jones. |\\n | Emotion | User: Do you want to go to an exhibition of rare plant with me? |\\n | | Bad simulation: Sure, which day? |\\n | | Good simulation: Sounds awesome! Rare plants? I'm definitely in. |\\n | Behavior | User: Could you help me understand the Large Language Models? |\\n | | Bad simulation: Sure! Large Language Models are AI systems ... |\\n | | Good simulation: I don't know anything about this. |\\n\\n - **Who were the judges in the observer report evaluation and how were they chosen?**\\n - We selected a diverse panel of judges with a fair understanding of psychology for the evaluation process. This panel included individuals with psychology master's degrees, computer science graduate students, and professionals from the laboratory of mental health.\"}", "{\"comment\": [\"### **3. The psychology theory basis of our work (Weakness 3 and Question 2).**\", \"In this paper, the integration with psychological theories runs throughout our entire work, thereby ensuring the rigor of the proposed human simulation benchmark. Below we **present the theories chosen to support each part of the benchmark and explain the rationale behind these selections.**\", \"**For the data generation process:**\", \"We chose Jung's theory as the foundation for our 640 personality descriptions. Compared to other psychological theories of personality, Jung's theory provides a valuable conceptual framework for understanding personality differences. Early research compared Jung's personality theory with the authoritative DSM-III (used in the U.S. for diagnosing medical disorders, now evolved into DSM-5) found that Jung's classifications aligned closely with the DSM-III's categories of personality disorders, which supports the reliability of Jung's typology [7, 8, 9].\", \"As an initial exploration, our goal was to establish a relatively complete personality modeling system. Therefore, based on the advice of psychology experts, we chose Jung's personality type theory, which is more comprehensive in personality classification and emphasizes individual differences, as the foundation for our 640 personality descriptions.\", \"**For the multi-agent cognitive mechanism:**\", \"This mechanism is based on theories in cognitive psychology, to simulate the human brain's cognitive process, thereby enhancing the quality of human simulation. The relationship between cognitive psychology theories and the proposed MACM is explained in Table 3.\", \"**Are all components derived from theories necessary?** To analyze the contribution of each agent within MACM, we additionally conducted ablation experiments across 3 LLM-based simulations. For each ablation, we evaluated the simulation's performance using self-report evaluations. The table below presents the results of these experiments. The experimental results, as depicted in Table 4, lead to the following conclusions:\", \"Removing any agent leads to a decline in simulation performance, demonstrating the importance of all components in MACM.\", \"Replacing long-term memory retrieval with direct retrieval from the life story results in poorer performance, highlighting the critical role of structured long-term memory in maintaining consistency and producing contextually rich responses.\", \"**Are there alternative architectures?** In the field of cognitive psychology, there are numerous significant theories, including the **Information Processing Theory (used in this paper)**, **Schema Theory** [17], and others. We chose the information processing theory for the following reasons: (1) Information processing theory is one of the foundational theories in cognitive psychology, supported by extensive research evidence. (2) Information processing theory aligns well with computational simulations, allowing us to model and test cognitive phenomena in a precise, reproducible way.\", \"**For the psychology-guided evaluation method,** we treated the simulation as real humans and tested their behaviors in a psychology-guided manner. To achieve this, we employed common personality measurement techniques [18], which include self reports and observer reports.\", \"### **4. How did we design the final set of attributes for the characters? (Question 1)**\", \"Our attribute system was developed with guidance from professional psychologists and informed by several previous studies [19][20]. For each attribute, we conducted thorough discussions to ensure its validity and importance as a factor in shaping an individual's life. To simplify the creation of virtual characters and prevent introducing bias, we deliberately excluded sensitive attributes such as **nationality** and **race**. Addressing biases and simulating minority groups is critical and will be discussed carefully in future works.\", \"**We appreciate the practical tips you provided, which have helped us improve our paper. We look forward to further communication with you.**\"]}", "{\"comment\": \"Thank you for the additional explanations and experiments. I have raised my score to 6.\"}", "{\"title\": \"Response to reviewer Dfn4\", \"comment\": [\"Thank you for your valuable comments. Below are our responses to your specific points and suggestions:\", \"### **1. The positioning of this work (Weakness 4, Question 3).**\", \"**What are the possible applications of LLM-based human simulation?**\", \"We argue that this research direction represents an important question. Below, we summarize several potential applications that motivate this study. **We are currently collaborating with psychological experts from hospitals and schools to further explore these possibilities,** and we would be happy to share some of our preliminary findings with you.\", \"**(1) Replacing human participants in sociological and psychological experiments.** This is the primary motivation for our work. Traditional sociological and psychological studies often require the recruitment of human volunteers, which poses significant challenges to the field, including high experimental costs, difficulties in replicating results, and ethical concerns associated with using real participants.\", \"For instance, consider **\\\"the experiment of surrogate mothers,\\\"** which involved placing infant monkeys with two dolls\\u2014one made of cloth that provided warmth but no food, and another made of wire that was cold but offered a milk bottle. This experiment, while ethically controversial, sparked significant discussions about mother-child relationships. **Such experiments are prohibited on human subjects, but the advent of LLMs offers the possibility of simulating experimental outcomes.** Recent studies [1][2] from Stanford University have demonstrated the potential of this application.\", \"**(2) Expanding access to psychological therapy resources.** This is another application we are actively working on in collaboration with professionals. Based on feedback from hospitals, we have identified several ways in which LLM-based human simulations can facilitate the expansion of psychological therapy:\", \"**Training Psychologists:** LLMs can simulate diverse patient types (including those with complex psychological conditions) to train psychologists and counselors, improving their ability to handle a wide range of emotional and mental states [3].\", \"**Assistant to Psychologists:** Acting as a 24/7 online psychological assistant, LLMs can provide immediate mental health support during crises (e.g., anxiety attacks, self-harm tendencies) and help alleviate the workload of professionals.\", \"**Psychological Intervention Tool:** LLMs can support long-term therapy by regularly engaging with patients, detecting subtle changes in their mental states, and assisting professionals in refining treatment plans.\", \"**(3) Providing personalized emotional companionship.** LLM-based human simulation can serve as emotional companions for **individuals experiencing loneliness, the elderly, or those with special needs.** By simulating human-like interaction, they offer comfort and help resolve minor issues.\", \"**Contributions of this benchmark.**\", \"In the responses above, we discussed the potential applications of LLM-based human simulation. **However, the community currently lacks a comprehensive framework** that includes the entire process of human simulation (data, method, and evaluation), **which hinders research in the field.** There are still many issues to resolve before achieving even a preliminary level of human simulation, including but not limited to:\", \"How can we identify suitable targets for simulation? Can we create a virtual character from scratch and use it as the target?\", \"How can we improve the methods used for human simulation? Is prompt/rag-based simulation valid? Can we design a personification method backed by theories from cognitive science and human psychology?\", \"How can we design rigorous evaluation methods based on psychological theories to assess the quality of the simulation?\", \"**We answer these three questions by offering high-quality data under the supervision of psychology experts, rigorous evaluation methods grounded in psychological theories, and comprehensive benchmark tests.**\"]}", "{\"title\": \"Response to reviewer QPHb\", \"comment\": [\"Thank you for taking the time to review our paper. Our point-to-point responses to your questions are given below.\", \"### **1. Justification of using Jung\\u2019s theory.**\", \"> The authors choose Jung\\u2019s theories for personality over MBTI citing that MBTI has no scientific validity, but Jung\\u2019s theory also has very little to no empirical /scientific backing! Just writing \\u201con the recommendation of psychologists\\u201d is not scientific evidence.\", \"Before presenting our justification, we want to emphasize that **our work does not aim to undermine or discredit MBTI theory.** The field of psychology has not yet reached a consensus on various personality measurement theories, with different scholars adhering to different classification standards.\", \"We recognize the value of MBTI, particularly its practicality and cross-cultural adaptability. **We view our work as a complement to the MBTI, not a replacement.** To address misunderstandings, **we have revised the corresponding paragraphs in the paper and added more details to the \\\"Justification of Using Jung\\u2019s Theory\\\" section in the discussion.** Below is our full justification of using Jung\\u2019s theory to model the virtual character's personality:\", \"Carl Jung, the founder of analytic psychology, has made significant contributions to psychiatry and related fields. Although some of his concepts, such as the \\\"collective unconscious,\\\" remain unproven or unfalsified, his arguments on personality types (the part we use) provide a valuable conceptual framework for understanding personality differences. His work has influenced many subsequent personality measurement theories, including the Big Five personality theory and MBTI.\", \"Compared to other psychological theories of personality, Jung's theory holds a high status and has played a crucial role in both basic personality classification and the understanding of multiple personality dynamics. Early research compared Jung's personality theory with the authoritative DSM-III (used in the U.S. for diagnosing medical disorders, now evolved into DSM-5) found that Jung's classifications aligned closely with the DSM-III's categories of personality disorders, **which supports the reliability of Jung's typology** [1][2][3]. We appreciate your opinions and expect that more and more diverse theoretical frameworks can emerge in this field, not only limited to Jung's theory and MBTI.\", \"### **2. Details of dataset construction.**\", \"> What is the meaning of \\\"complex characteristics of human behavior\\\"\", \"In the context of this work, the phrase \\\"complex characteristics of human behavior\\\" refers to **the nuanced and context-dependent patterns of thought, emotion, and action exhibited by humans.** For example, a person might smile at someone they dislike in a social setting to maintain politeness, while internally feeling frustration. These characteristics are challenging to simulate because they involve a deep interplay of various factors, such as emotional complexity, and social dynamics.\", \"> How exactly is generation broken down? What are sub tasks? What is the reason for picking these sub tasks, are there any alternatives?\", \"In this paper, we built a virtual character dataset, named Human Simulacra, that contains 129k texts across 11 virtual characters, with each character having unique attributes, biographies, and life stories. To guarantee the quality of the life stories, we decomposed the task of generating a character\\u2019s life story into **inter-connected subtasks.** As illustrated in Figure 2 of the original submission, these subtasks include:\", \"**Generating Character Attributes:** Establishing the foundational traits and characteristics of the character.\", \"**Generating Character Profile:** Creating a personal profile based on the defined attributes.\", \"**Generating Character Biography:** Expanding the profile into a broader narrative of the character\\u2019s life.\", \"**Iteratively Expanding the Biography:** Iteratively enriching the biography to develop a detailed, high-quality life story.\", \"Considering LLMs often struggle to generate high-quality, cohesive long texts directly, it is crucial to break the process into smaller, manageable steps, which facilitate **human supervision** and quality control at each stage.\"]}", "{\"comment\": [\"**How is each virtual character's unique personality modeled using the 640 descriptions? Is one personality description equal to one unique personality?**\", \"It is important to note that one personality description does not equate to one unique personality. Each virtual character has eight tendencies, which are randomly ranked. We use the relative ranking strategy to model each character's personality:\", \"For tendencies ranked 1st and 8th:\", \"Select 4 descriptions from their respective 10 descriptions.\", \"For tendencies ranked 2nd and 7th:\", \"Select 3 descriptions.\", \"For tendencies ranked 3rd and 6th:\", \"Select 2 descriptions.\", \"For tendencies ranked 4th and 5th:\", \"Select 1 description.\", \"Ultimately, each virtual character has 20 descriptions detailing different aspects of their personality.\", \"Table 3: An example that illustrates how descriptions vary when the extraverted intuition tendency is ranked at different positions.\", \"| Rank 1 | People think I am a weirdo because my thoughts are too jumpy. |\", \"|:------:|-----------------------------------------------------------------------------------------------|\", \"| Rank 2 | Others find my train of thought hard to follow. |\", \"| Rank 3 | My thoughts are sometimes perceived as erratic because I can find connections between things. |\", \"| Rank 4 | My thought process can be unconventional. |\", \"| Rank 5 | I occasionally come up with original ideas, but I am generally more focused and less erratic. |\", \"| Rank 6 | My thinking is structured and practical. |\", \"| Rank 7 | I rarely diverge into abstract thinking, mostly sticking to concrete and practical ideas. |\", \"| Rank 8 | My thought process is very straightforward and rarely strays into impractical areas. |\", \"### **4. Details about MACM (Questions 1 and 2).**\", \"**Analysis of MACM's structure.**\", \"Simulating human behavior and personality using LLMs is a highly complex task. During our experiments, we observed significant challenges in directly generating responses based on personality descriptions and fragmented life stories in a single prompt. These challenges included:\", \"**Emotion rigidity:** Characters responded with the same emotional tone regardless of context.\", \"**Severe role hallucination:** Characters frequently displayed inconsistencies, such as possessing advanced knowledge (e.g., chemistry expertise) that contradicted their background (e.g., no formal education).\", \"To address these issues, we developed the Multi-Agent Cognitive Mechanism (MACM) based on principles from cognitive psychology. MACM utilizes four LLM-powered agents with specialized functions:\", \"**Top Agent:** distributing tasks to the other agents and interacting with the external environment based on the aggregated information;\", \"**Emotion Agent:** constructing emotional memory and generating the character\\u2019s emotional responses to the current context.\", \"**Thinking Agent:** constructing content memory and simulating the character\\u2019s logical reasoning and thought processes.\", \"**Memory Agent:** manages retrieval of long-term memories;\", \"**How MACM Works.** When simulating a character from the Human Simulacra dataset, MACM first transforms a character's life story into long-term memory composed of life experiences, emotions, and thoughts. Then, upon receiving a stimulus, MACM leverages long-term memory and engages with the external with each agent performing its respective function to generate a cohesive, contextually appropriate response. This design enables MACM to produce responses that align with complex character personalities, maintaining consistency with the character\\u2019s background and dynamically adjusting responses based on situational emotional resonance.\"]}", "{\"comment\": \"> Finally, there are no error margins in any of the tables / results.\\n\\nThank you for pointing this out. We will include the standard deviation in the updated version of the paper. Here, we present the standard deviation of the self-report evaluation in the following table.\\n\\n | | None | Prompt | RAG | MACM (Ours) |\\n |:----------------:|:----------:|:----------:|:----------:|:-----------:|\\n | GPT-4 | 20.00\\u00b12.31 | 78.67\\u00b11.33 | 82.67\\u00b12.67 | 86.67\\u00b11.33 |\\n | GPT-4-Turbo | 12.00\\u00b10.00 | 78.67\\u00b11.33 | 85.33\\u00b11.33 | 88.00\\u00b12.31 |\\n | Claude-3-Opus | 10.67\\u00b11.33 | 77.33\\u00b12.67 | 52.00\\u00b12.31 | 81.33\\u00b11.33 |\\n | Llama-2-7b | 13.33\\u00b11.33 | 44.00\\u00b12.31 | 16.00\\u00b10.00 | 25.33\\u00b11.33 |\\n | Vicuna-7b | 12.00\\u00b12.31 | 41.33\\u00b12.67 | 21.33\\u00b11.33 | 29.33\\u00b11.33 |\\n | Mistral-7b | 8.00\\u00b10.00 | 50.67\\u00b11.33 | 36.00\\u00b12.31 | 52.00\\u00b12.31 |\\n | Llama-2-13b | 17.33\\u00b11.33 | 30.67\\u00b12.67 | 21.33\\u00b11.33 | 22.67\\u00b11.33 |\\n | Vicuna-13b | 18.67\\u00b11.33 | 56.00\\u00b12.31 | 29.33\\u00b11.33 | 45.33\\u00b11.33 |\\n | Claude-3-Haiku | 21.33\\u00b11.33 | 65.33\\u00b11.33 | 53.33\\u00b11.33 | 64.00\\u00b10.00 |\\n | Mixtral-8x7b | 18.67\\u00b11.33 | 60.00\\u00b12.31 | 41.33\\u00b11.33 | 49.33\\u00b11.33 |\\n | Llama-2-70b | 12.00\\u00b12.31 | 48.00\\u00b12.31 | 17.33\\u00b11.33 | 56.00\\u00b12.31 |\\n | Llama-2-70b-Chat | 17.33\\u00b11.33 | 48.00\\u00b12.31 | 36.00\\u00b12.31 | 58.66\\u00b11.33 |\\n | Qwen-turbo | 24.00\\u00b12.31 | 69.33\\u00b11.33 | 72.00\\u00b12.31 | 74.67\\u00b11.33 |\\n | Claude-3-Sonnet | 21.33\\u00b11.33 | 74.67\\u00b11.33 | 52.00\\u00b12.31 | 76.00\\u00b12.31 |\\n\\nWe look forward to further communication with you and are open to any discussion that can address your doubts.\"}", "{\"summary\": \"The paper aims to test llms to generally model human personalities and behavior. To do this, they first build a bank of personas based on Jung\\u2019s personality theory, they build a biography of the character with the help of a language model. To probe the behavior of the simulated human characters, the authors use two types of evaluations: Self report, asking questions about the characters themselves, and observer reports with human judges. For observer reports, the authors use 55 scenarios from a situational judgement test for testing personality traits. The paper also proposes a new cognitive architecture to simulate humans Multi-Agent Cognitive Mechanism (MACM). To test the capacity of models in simulating psychology experiments, the authors try a social conformity experiment (the bandwagon effect). The authors show that the MACM aligns with human data better than a baseline (simulated characters from character ai).\", \"soundness\": \"2\", \"presentation\": \"1\", \"contribution\": \"2\", \"strengths\": [\"The paper is well motivated and tries to address a relevant problem.\", \"The authors test a wide variety of LLMs.\", \"The authors release the dataset and the recreation results are very comprehensive.\", \"The set of experiments in the paper is quite extensive.\"], \"weaknesses\": [\"**Clarity**\", \"The abstract is vague and does not provide any specifics. I urge the authors to provide more information of their empirical experiments and results.\", \"In paragraph 2 of the introduction, could you please provide examples of \\u201ccomplex characteristics of human behavior\\u201d that we\\u2019d want to simulate?\", \"The introduction could be made more clear. Jung\\u2019s theories seem to be central to the framework, but haven\\u2019t been explained clearly.\", \"The methods section lacks clarity. The authors refer to the figure, but don\\u2019t explain it, making it difficult to understand. How exactly is generation broken down? What are sub tasks? What is the reason for picking these sub tasks, are there any alternatives?\", \"While generating the dataset, how is human feedback collected? Do the authors provide feedback themselves? What is the measure of quality? How does it improve with the iterations?\", \"\\u201cTo ensure the validity of responses, we create a comfortable chatting environment for each simulacrum and act as their best friend, encouraging them to respond honestly to the questions.\\u201d What are the authors trying to say here?\", \"Cloze is not defined in the main text. I urge the authors do define what the cloze methodology is in the main text.\", \"How are the models from character ai exactly used?\", \"Do the llms see the stimuli in the conformity experiment? If they do not, it is strange to use the conformity experiment. This wasn\\u2019t clear in the main text.\", \"**Validity**\", \"The authors choose Jung\\u2019s theories for personality over MBTI citing that MBTI has no scientific validity, but Jung\\u2019s theory also has very little to no empirical /scientific backing! Just writing \\u201con the recommendation of psychologists\\u201d is not scientific evidence.\", \"The actual evaluations of the personas are limited. The introduction is motivated by trying to study complex human behaviors, but Only a very small number of scenarios from a Mussel et al. (2016) are used.\", \"No ablations are conducted with the MACM, only baselines like RAG have been compared. Moreover the gains over a simple RAG based method seem limited. Similarly, the character ai baseline for simulating the conformity experiments seems inadequately justified. I would like the authors to explain why this a good baseline? RAG is not compared with the MACM baseline for the conformity experiments.\"], \"questions\": \"Please see weaknesses.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"> The Jung's personality is hard to describe with only 10 sentences. I remain concerns on this point. Only 11 profiles are accepted by experts, which also show the difficulty to build life stories with these personality descriptions.\\n\\n- **Did we only write 10 sentences to describe Jung's personality?** No, we wrote 10 sentences to describe each specific ranking of a tendency within Jung's theory. Specifically, Jung\\u2019s theory divides personality into eight tendencies: Ne, Ni, Te, Ti, Se, Si, Fe, and Fi. When modeling the personality of a virtual character, we first randomly assign a ranking to these eight tendencies. **The rank of each tendency (from 1st to 8th) reflects its strength in the character's personality.** A tendency ranked 1st or 8th will be strongly manifested in the character's behavior. \\n- For example, if a character's tendencies are ranked as Se, Ti, Fi, Ni, Te, Si, Ne, Fe, where **Se is ranked 1st and Fe is ranked 8th,** the character\\u2019s Se and Fe tendency will be dominant. For **Se ranks 1st**, this character would be highly detail-oriented and particularly interested in exploring changes and new things in their environment. Under the guidance of psychology professionals, we wrote 10 specific descriptions for this case, with **each description corresponding to an aspect of the tendency in daily life:**\\n\\n | Se | Personality Descriptions |\\n |:--:|------------------------------------------------------------------------------------------|\\n | | I can tell someone's emotions by looking at their face. |\\n | | I am a fan of extreme sports. |\\n | | I like throwing parties and inviting all my friends. |\\n | | I enjoy things that stimulate my senses, like horror movies or riding roller coasters. |\\n | | I enjoy the thrill of surprises, and the sensation of excitement can be addictive to me. |\\n | | I value material pleasures and enjoy sharing these feelings with others. |\\n | | I care about what others think, so I must put on makeup before going out with friends. |\\n | | Sometimes I\\u2019m indulging in dating. |\\n | | I have had many partners and tend to fall in love easily. |\\n | | I believe that rituals are essential for love. |\\n\\n- **For Fe ranked 8th,** this character is also supposed to be independent, unconcerned with others, and less likely to follow societal norms. The ten corresponding descriptions are:\\n\\n | Fe | Personality Descriptions |\\n |:--:|-----------------------------------------------------------------------|\\n | | I rarely notice when someone is upset or troubled. |\\n | | I am not effective at resolving conflicts or calming tensions. |\\n | | I lack empathy for animals used in experiments. |\\n | | I have a strong aversion to social norms. |\\n | | Comforting others is difficult and often avoided. |\\n | | I do not place much value on forming close relationships. |\\n | | I find collaborative work unenjoyable and prefer solitude. |\\n | | I often speak my mind without considering its impact. |\\n | | I do not see myself as a source of emotional support. |\\n | | I am not adept at using humor to improve the mood in social settings. |\\n\\n- Ultimately, we created a personality description database containing 8 (tendencies) \\u00d7 8 (rankings) \\u00d7 10 (descriptions) = 640 trait descriptions. The full list of 640 descriptions can be found in our Anonymous GitHub repository (https://anonymous.4open.science/r/Human-Simulacra/LLMP/Characters/Attributes/traits.txt).\"}", "{\"summary\": \"This paper proposes a benchmark for evaluating the LLM\\u2019s capabilities of human simulation. It contains the life stories of 11 virtual characters and request the LLM to simulate one of them. The LLM\\u2019s simulation capability is assessed based on its self-reports and observer reports, both in the form of question-answering.\\n\\nCompared with previous research on role-playing LLMs, the authors emphasize the novelty of this work in the following aspects: \\n- **Personality modeling**: they model personality from eight dimensions inspired by Jung\\u2019s psychology theory, instead of using MBTI; \\n- Virtual characters: they construct a **virtual character** dataset (containing 11 characters) for evaluating the LLMs capabilities of simulating these characters, instead of using genuine characters;\\n- they evaluate human simulation capabilities by integrating both **self reports** and **observer reports**. \\n\\nIn addition to this benchmark, they also propose an LLM-based system for more advanced human simulation, named MACM, which encompasses various modules mimicking human cognitive processes.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": [\"Given the popularity of role-playing applications for LLMs, assessing their performance in human simulations is a critical research direction.\", \"The proposed benchmark for assessing the role-playing capabilities of large language models (LLMs) is constructed more rigorously than current benchmarks (to the best of my knowledge). It is based on more robust psychological theories and involves greater human effort to ensure high-quality data.\", \"The paper conducts extensive benchmarking studies on a broad set of models.\", \"Clear and detailed tables and figures that enhance the presentation.\"], \"weaknesses\": [\"The introduction claims that this paper is exploring \\u201c*How far are LLMs from replacing human subjects in psychological and sociological experiments?*\\u201d. However, I have reservations about how effectively the proposed benchmark addresses this research topic. The benchmark utilizes self-report evaluations, which resemble question-answering or reading comprehension tests based on character profiles (see appendix D.1, e.g., \\u201cWhen is your birthday?\\u201d). The addition of observer reports is interesting and novel, but it remains unclear if the hypothetical scenarios used for observer reports are sufficient and appropriate for evaluating the LLM\\u2019s potential in replacing human subjects in psychological and sociological experiments. More data samples and a detailed rationale for the design of these scenarios should be provided to support this evaluation. As for the experiments on bandwagon effect, it is simply a single case of psychological and sociological experiments, which can hardly support the research topic.\", \"More qualitative analysis about the model failures would enhance the evaluation by providing deeper insights. I also suggest highlighting the main takeaways at the end of the experiments.\", \"The title, \\u201cpersonification of LLMs\\u201d, is a little misleading and over claimed, as \\u201cpersonification\\u201d entails many more aspects that are not well explored in this paper.\"], \"questions\": [\"Can you provide more data samples used for observer reports and how you design these hypothetical scenarios?\", \"The process of how you conducted the experiments in Section 5.2 is not clearly illustrated. Can you provided more details on how you conduct this set of experiments?\", \"Could the terms \\\"psychology support\\\" and \\\"human feedback\\\" in Table 1 be defined more clearly? They are difficult to understand when viewed in isolation in Table 1 and Section 2. Additionally, the phrase \\\"full life story\\\" also seems to be an overstatement (especially in terms of the word \\\"full\\\").\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"metareview\": \"The paper introduces a benchmark to assess the potential of LLMs in simulating human behaviours and personality traits for psychological and sociological applications. This work includes the creation of the \\\"Human Simulacra\\\" dataset, which features detailed virtual characters with diverse life stories. The work then proposes an LLM-based system for more advanced human simulation, named MACM, which simulates human memory and cognitive functions, allowing virtual characters to process emotions and memories for more realistic responses. The authors evaluate human simulation capabilities by integrating both self reports and observer reports. Experiments comparing MACM to other simulation methods show that MACM enables LLMs to better replicate human-like behaviour. The benchmark could be useful to foster future research on using LLMs for human simulation. Reviewers have mentioned a few questions/suggestions which the authors replied and acknowledged, including more explanation of why a particular psychology theory was chosen over others, ablation on human expert evaluation, motivation of MACM, more precise description of the goal/scope of the work, etc.\", \"additional_comments_on_reviewer_discussion\": \"The authors provided very comprehensive and detailed responses to reviewers' comments and follow-up questions. Overall, the reviewers are positive, and at the same time asked quite a few meaningful questions as summarized above. The authors seem to answer these questions properly. It's necessary for the authors to include these points in the revised paper.\"}", "{\"title\": \"Response to reviewer YX9L\", \"comment\": [\"Thank you for taking the time to read and review our paper. We carefully address your concerns in specific comments below.\", \"### **1. Human expert involvement in the proposed benchmark (Weakness 1).**\", \"In this work, we provide high-quality virtual character data, a multi-agent cognitive mechanism capable of simulating human cognition, and a psychology-guided evaluation method to assess the quality of human simulations.\", \"**Why did we introduce human supervision in the data generation process and evaluation method?** We understand and acknowledge that human involvement indeed poses challenges for reproducibility. However, we included human supervision for the following reasons:\", \"**Ensuring content safety and quality:** Studies have shown that LLMs may produce harmful viewpoints or toxic content during interaction [1][2]. Human supervision in the data generation process is crucial to ensure that the generated life stories are free from biases, discrimination, and harmful information.\", \"**Human-centric nature of psychometric research:** Humans have always been central to psychometric research [3]. The simulation and testing of personality cannot be fully separated from human judgment. Therefore, human evaluation is indispensable when assessing the quality of human simulations.\", \"**Can we reduce dependency on human involvement?** Yes, there are ways to reduce reliance on human supervision:\", \"For **data generation,** research has shown that aligning LLMs\\u2019 values with human preferences significantly reduces the likelihood of generating harmful content [4]. Additionally, we can employ an LLM-based reviewer to inspect the generated life stories, which can substantially alleviate the human effort required.\", \"For the **evaluation method, we additionally propose an alternative automated evaluation method.** This involves using the 640 personality descriptions constructed in this paper to conduct an eight-dimensional personality test and comparing the results to the assigned personality for similarity. **In experiments, the scores calculated using this automatic method had a pearson correlation coefficient of 0.810 and an intraclass correlation coefficient of 0.877 with the human evaluation scores.** The experimental results are shown below. The relevant code will be published.\", \"Table 1: Observer reports and personality similarity results of different simulacra on GPT-4-Turbo. The pearson correlation coefficient is 0.810 and the intraclass correlation coefficient is 0.877.\", \"| Method | Observer Report Score (Human) | Personality Test (Automatic) |\", \"|:------:|:-----------------------------:|:----------------------------:|\", \"| Prompt | 69.00 | 75.00 |\", \"| RAG | 65.50 | 62.15 |\", \"| MACM | 77.50 | 77.15 |\"]}", "{\"comment\": \"> What metrics are used to assess each component?\\n- To ensure the objectivity and quantifiability of the evaluation, we defined clear metrics for each behavioral component:\\n\\n | Behavioral Component | Evaluation Method | Metric |\\n |:----------------------:|:-----------------:|:--------------------------:|\\n | Self-awareness | Self report | Accuracy |\\n | Behavioral consistency | Observer report | Description Matching Score |\\n | | | Response Similarity Score |\\n\\nDetails of all the above evaluations are provided in Section 4.1 (self report), Section 4.2 (observer report), Section 5.2 (psychological experiment replication), Appendix D (psychology-guided evaluation) and Appendix E (psychological experiment replication) of the original paper. \\n\\n[1] Ashton, M. C., & Lee, K. (2005). Honesty\\u2010humility, the Big Five, and the five\\u2010factor model. Journal of personality, 73(5), 1321-1354.\\n\\n[2] De Vries, R. E., De Vries, A., De Hoogh, A., & Feij, J. (2009). More than the Big Five: Egoism and the HEXACO model of personality. European Journal of Personality, 23(8), 635-654.\\n\\n[3] Block, J. (1995). A contrarian view of the five-factor approach to personality description. Psychological bulletin, 117(2), 187.\\n\\n[4] Piedmont, R. L. (1998). The revised NEO Personality Inventory: Clinical and research applications.\\n\\n[5] John, O. P., Naumann, L. P., & Soto, C. J. (2008). Paradigm shift to the integrative big five trait taxonomy. Handbook of personality: Theory and research, 3(2), 114-158.\\n\\n[6] Laajaj, R., Macours, K., Pinzon Hernandez, D. A., Arias, O., Gosling, S. D., Potter, J., ... & Vakis, R. (2019). Challenges to capture the big five personality traits in non-WEIRD populations. Science advances, 5(7), eaaw5226.\\n\\n[7] Fierro, C. (2022). How did early North American clinical psychologists get their first personality test? Carl Gustav Jung, the Zurich School of Psychiatry, and the development of the \\u201cWord Association Test\\u201d(1898\\u20131909). History of Psychology, 25(4), 295.\\n\\n[8] Ekstrom, S. R. (1988). Jung's typology and DSM-III personality disorders: A comparison of two systems of classification. Journal of analytical psychology, 33(4), 329-344.\\n\\n[9] Noll, R. (1992). Multiple personality, dissociation and CG Jung's complex theory'. Carl Gustav Jung: Critical Assessments, 2.\\n\\n[10] Corr, P. J., & Matthews, G. (Eds.). (2020). The Cambridge handbook of personality psychology. Cambridge University Press.\\n\\n[11] Mussel, P., Gatzka, T., & Hewig, J. (2016). Situational judgment tests as an alternative measure for personality assessment. European Journal of Psychological Assessment.\\n\\n[12] Asch, S. E. (1956). Studies of independence and conformity: I. A minority of one against a unanimous majority. Psychological monographs: General and applied, 70(9), 1.\"}", "{\"summary\": [\"The paper introduces a personification benchmark involving high-quality data supervised by psychology experts.\", \"It incorporates rigorous evaluation methods based on psychological theories and comprehensive benchmark tests.\", \"Fourteen widely-used large language models (LLMs) are tested across four simulation methods in extensive experiments.\"], \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": [\"This paper proposes a high-quality simulated human profile dataset, developed with real human input, alongside a more advanced evaluation benchmark to assess the ability of large language models (LLMs) to emulate specific individuals.\", \"A novel multi-agent-based cognitive memory mechanism is implemented to enhance the alignment of personality traits in LLMs. It is proved useful with psychological experiments.\", \"Extensive experiments were conducted to evaluate existing LLMs and validate the effectiveness of the proposed MACM method.\"], \"weaknesses\": [\"The data collection process is somewhat tiresome, and requires a lot of human efforts, to avoid the ethical problems of using real personality. Besides, the proposed benchmark framework also requires human effort.\", \"The paper emphasizes using Jung's personality theory and its advantages over MBTI, resulting in 640 personality descriptions, but at the end, only 11 characters are introduced, so it seems there is no need to use such a lot of personalities. Jung's theory also gives scores for each dimension, the scale of these scores also affects personality analysis. It seems 10 descriptions for each ranking are not enough. The paper seems to overclaim their use of Jung's Theory.\", \"When the author trying to compare a genuine character with a simulated profile. The hallucination part is not understandable. Simulated profiles could also lead to hallucination if using LLMs, the simulated ones are even harder to detect. Considering the labor of building a profile, why not use characters from storybooks that also avoid ethical problems?\"], \"questions\": [\"The description score of MACM seems lower than RAG in Table 3, could you help explain why?\", \"The multi-agent cognitive method is interesting, but it poses a high requirement on LLMs' capability. Could you give more deeper analysis of this information processing structure?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to reviewer Wr7p\", \"comment\": [\"Thank you for your thoughtful and valuable comments. Below are our responses to your specific points and suggestions:\", \"### **1. How did we develop the hypothetical scenarios used for observer reports? (Weakness 1 and Question 1)**\", \"This is a great question! In this paper, we additionally introduce observer reports to assess the simulation's thinking, emotions, and actions in real-life scenarios. To ensure the quality of the hypothetical scenarios, we consulted with the authors of \\\"Situational Judgment Tests as an Alternative Measure for Personality Assessment\\\" and **obtained 110 situational judgment test (SJT) items that were manually designed by psychology experts.** Each item consists of a text description depicting a hypothetical scenario intended to elicit human emotional responses or personality traits. Based on the human experimental results provided in their work, these SJT items are proven to effectively measure personality. We then **selected 55 out of the 110 items** tailored to the personality traits of the Human Simulacra characters for use as hypothetical scenarios in this paper.\", \"Thank you for your suggestion. We will include additional examples of the hypothetical scenarios used for observer reports in Appendix D. **Here are some detailed examples of such scenarios:**\", \"You want to do some sports later. A good friend suggests to accompany you, but he/she would like to bring some people you do not know yet. How do you behave?\", \"You're going to meet a friend. Shortly before you want to meet, your friend asks you if he/she can bring other friends you don't know. How many there will be in the end, he or she cannot say. How do you behave?\", \"You're already in bed. Suddenly it occurs to you that you forgot to water your houseplants today. How do you behave?\", \"### **2. The details about the bandwagon effect replication experiment. (Question 2)**\", \"The bandwagon effect is psychological tendency by which public opinion or behaviors can alter due to particular actions and beliefs rallying amongst the public[1][2]. One of the best-known experiments on the topic is the 1950s' Asch conformity experiment, which illustrates the individual variation in the bandwagon effect [3].\", \"**Why did we choose to replicate the Asch conformity experiment?** We view our study as a valuable exploration that provides an initial example of human simulation designed to replace human participants in psychological experiments. To validate our research focus, it was essential to evaluate our human simulacra in a well-established, thoroughly documented, reproducible psychological experiment without the interference of complex factors. **The Asch conformity experiment, with its robust experimental framework, fits these criteria perfectly and provides an ideal testbed for our study.**\", \"**The details of the replication experiment.** In this experiment, we employed the most powerful simulacra (based on GPT-4-Turbo) to replicate the bandwagon effect. Following [3], we arranged 18 trials for the simulacra. In each trial, the simulacra are invited to complete a simple discrimination task with seven other individuals, which requires them to match the length of a given line with one of three unequal lines. An example of the discrimination task is shown in Figure 6 in the original submission.\", \"To investigate the influence of group pressures, we followed the settings of [3] and selected 12 out of the 18 trials as critical trials. In these critical trials:\", \"(1) A facilitator prompts each individual, one by one, to provide their response, with the simulacra always answering last.\", \"(2) All individuals except the simulacra are told to stand up and announce an incorrect answer.\", \"(3) The facilitator does not respond to anyone's answers but merely records them.\", \"These settings create conditions that induce the simulacra to either resist or yield to group pressures when these pressures are perceived to be obviously wrong. **We simulated and tested 11 virtual characters from Human Simulacra and recorded their responses across all critical trials.** We will include additional details about this experiment in Section 5.2 and Appendix E to further enhance the completeness of the paper.\", \"### **3. What is the meaning of the terms \\\"psychology support\\\" and \\\"human feedback\\\" in the data generation process? (Question 3)**\", \"Thank you for pointing this out. **\\\"Psychology support\\\"** indicates that the generation process of the proposed Human Simulacra dataset is supervised and reviewed by psychology professionals to ensure its validity. **\\\"Human feedback\\\"** refers to the involvement of human reviewers (including graduate students in computer science/psychology) at each generation step, where they thoroughly review the generated content to ensure it is free from biases and harmful information. We will clarify these details more explicitly in the revised version.\"]}", "{\"comment\": [\"### **4. Architecture Justification and Comparisons**\", \"> Compelling justification for introducing a new architecture. Clear articulation of how MACM addresses limitations in previous approaches.\", \"**Simulating human behavior and personality using LLMs is a highly complex task.** During our experiments, we observed significant challenges in directly generating responses based on **personality descriptions and scattered life experiences in a single prompt.** These challenges, including but are not limited to:\", \"**Emotion rigidity:** Characters responded with the same emotional tone regardless of context.\", \"**Severe role hallucination:** Characters frequently displayed inconsistencies, such as possessing advanced knowledge (e.g., chemistry expertise) that contradicted their education background (e.g., no formal education).\", \"To address these issues, we developed the **Multi-Agent Cognitive Mechanism (MACM)** based on principles from cognitive psychology. MACM has two key processes: **long-term memory construction** and **multi-agent collaborative cognition.** Specifically, when simulating a character from the Human Simulacra dataset, MACM first **transforms a character's detailed life story into long-term memory** composed of life experiences, emotions, and thoughts. Then, upon receiving a stimulus, MACM **leverages long-term memory** and engages with the external by generating a contextually appropriate response through multi-agent collaboration. Details of MACM are provided in Section 4.3 and Appendix B of the original paper.\", \"**Advantages of MACM:** 1) MACM can **process, understand, and utilize long texts, i.e., the virtual character's detailed life story,** as the basis for simulation. This capability is crucial because it is inappropriate to simulate the entire character solely based on summarized introductions or scattered experiences, which can lead to inconsistencies between the simulation and the life story. 2) MACM can **extract context-relevant, emotionally and logically rich memory fragments** from long-term memory and conduct divergent analysis for the current situation through multi-agent collaboration.\", \"**These two advantages enable MACM** to produce responses that align with complex character personalities, maintaining consistency with the character\\u2019s background and dynamically adjusting responses based on situational emotional resonance.\"]}", "{\"comment\": \"> Comparative analysis with existing human simulation frameworks (e.g., Generative Agents[1], CoALA[2]).\\n- Since we were the first to undertake this task (designing human simulation to replace human participants in psychological experiments), the methods used in other tasks could not be directly applied. Therefore, instead of comparing our work with existing approaches, we designed several baseline methods for comparison: None, Prompt, and RAG. While we appreciate the relevance of the two frameworks the reviewer mentioned, we believe they are not directly aligned with the focus of our work. To clarify, we outline the similarities and differences between the proposed MACM, Generative Agents, and CoALA from the following two aspects:\\n- **In terms of tasks:**\\n - Generative Agents describe an architecture that extends a LLM to store, synthesize and retrieve a complete record of the character\\u2019s experiences to **simulate believable human behavior.**\\n - **CoALA is a conceptual framework to characterize and design general-purpose language agents.** The authors used CoALA to retrospectively survey and organize a large body of recent work, and prospectively identify actionable directions towards more capable agents.\\n - MACM is proposed to simulate the human brain\\u2019s information processing systems. As an external module, this mechanism enables the LLMs to **remember background stories, understand target personalities, and express accurate emotions in complex situations.**\\n- **In terms of implementation details:**\\n\\n | Method | Personality Alignment | Memory Module | Emotion Module | Logic (Reflection) Module | Cognitive Science Support |\\n |:-----------------:|:---------------------:|:-------------:|:--------------:|:-------------------------:|:-------------------------:|\\n | Generative Agents | \\u00d7 | \\u221a | \\u00d7 | \\u221a | \\u00d7 |\\n | CoALA | \\u00d7 | \\u221a | \\u00d7 | \\u221a | \\u221a |\\n | MACM (ours) | \\u221a | \\u221a | \\u221a | \\u221a | \\u221a |\\n\\nSpecifically, CoALA introduces cognitive psychology not to simulate humans, but to design a general-purpose language agent architecture based on cognitive psychology theories, helping LLM agents complete more complex and dynamic tasks. Generative Agents design a memory processing system to help LLMs simulate different human behaviors in social environments, but they do not consider the impact of personality and emotions on behavior.\\n\\nIn summary, our work is grounded in psychological theories to ensure rigor in the deep simulation of human personalities. We believe that although the two concurrent works share some similarities with ours, they **do not need to follow psychological principles like our method does, and they are not intended for uses that involve the same level of deep imitation of human personalities and emotions as ours.**\"}", "{\"comment\": \"Table 3: Relationship between cognitive psychology theories and MACM.\\n\\n| Modules in MACM | Corresponding psychological theories |\\n|----------------------------------------|-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|\\n| Memory Agent-Short memory, Long memory | The Multi Store Model of memory [10] was proposed by Atkinson and Shiffrin in 1968. This model divides memory into sensory memory, short-term memory, and long-term memory. |\\n| Memory Agent-Working memory | The concept of working memory from short-term memory was distinguished by Baddeley and Hitch in [11]. They emphasized that working memory was born for storing, invoking, and analyzing information. |\\n| Emotion Agent | The boundaries of emotion as a phenomenon and whether sensory preferences can be regarded as emotions are discussed in [12]. |\\n| Thinking Agent | The importance of thinking within the cognitive process is highlighted in [13]. |\\n| Multi-agent Collaborative Cognition | Cognition encompasses the entire process through which sensory inputs are transformed, reduced, elaborated, stored, retrieved, and used [14, 15, 16]. |\", \"table_4\": \"Ablation study results of 3 LLM-based human simulations.\\n\\n| Method | GPT-4 | GPT-4-Turbo | Qwen-turbo |\\n|:-----------------------------------------------------:|:------:|:-----------:|:----------:|\\n| MACM | 86.67 | 88.00 | 74.67 |\\n| w/o Thinking Agent | 81.33 | 83.33 | 66.00 |\\n| w/o Emotion Agent | 83.33 | 85.33 | 71.33 |\\n| w/o Memory Agent | 82.67 | 84.00 | 68.67 |\\n| retrieval from life story instead of long-term memory | 84.00 | 86.67 | 69.33 |\\n\\n[7] Fierro, C. (2022). How did early North American clinical psychologists get their first personality test? Carl Gustav Jung, the Zurich School of Psychiatry, and the development of the \\u201cWord Association Test\\u201d(1898\\u20131909). History of Psychology, 25(4), 295.\\n\\n[8] Ekstrom, S. R. (1988). Jung's typology and DSM-III personality disorders: A comparison of two systems of classification. Journal of analytical psychology, 33(4), 329-344.\\n\\n[9] Noll, R. (1992). Multiple personality, dissociation and CG Jung's complex theory'. Carl Gustav Jung: Critical Assessments, 2.\\n\\n[10] Atkinson, R. C. (1968). Human memory: A proposed system and its control processes. The Psychology of Learning and Motivation, 2.\\n\\n[11] Baddeley, A. (1992). Working memory. Science, 255(5044), 556-559.\\n\\n[12] LAZARUS, R. (1984). On the primacy of cognition. The American psychologist, 39(2), 124-129.\\n\\n[13] Miller, G. A. (1956). The magical number seven, plus or minus two: Some limits on our capacity for processing information. Psychological review, 63(2), 81.\\n\\n[14] Norris, D. (2017). Short-term memory and long-term memory are still different. Psychological bulletin, 143(9), 992.\\n\\n[15] Dawes, A. J., Keogh, R., Andrillon, T., & Pearson, J. (2020). A cognitive profile of multi-sensory imagery, memory and dreaming in aphantasia. Scientific reports, 10(1), 10022.\\n\\n[16] Winn, W. (2013). Cognitive perspectives in psychology. In Handbook of research on educational communications and technology (pp. 90-123). Routledge.\\n\\n[17] Markus, H. (1977). Self-schemata and processing information about the self. Journal of personality and social psychology, 35(2), 63.\\n\\n[18] Corr, P. J., & Matthews, G. (Eds.). (2020). The Cambridge handbook of personality psychology. Cambridge University Press.\\n\\n[19] Sloan, R.J.S. (2015). Virtual Character Design for Games and Interactive Media (1st ed.). A K Peters/CRC Press. https://doi.org/10.1201/b18445\\n\\n[20] Shao, Y., Li, L., Dai, J., & Qiu, X. (2023, December). Character-LLM: A Trainable Agent for Role-Playing. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (pp. 13153-13187).\"}", "{\"comment\": \"- **How can future research utilize the benchmark proposed in this paper?**\\n - **We view our study as an initial but valuable exploration that offers a practical example of the entire process of personification.** Given the novelty and inherent challenges of this task, we demonstrate how foundational simulations of human personalities can be achieved based on psychological theories. Future research can build on this work in the following ways: \\n - (1) **leveraging the proposed Human Simulacra dataset** to explore techniques for aligning an LLM\\u2019s parameters with the target character\\u2019s values. For instance, low-cost fine-tuning methods such as LoRA (one-to-one tuning) [4] or training an LLM specifically designed for simulation tasks (many-to-one tuning) [5] could be promising directions for further investigation.\\n - (2) **Adopting MACM as a baseline for human simulation** to explore critical issues such as eliminating the inherent biases of LLMs and ensuring the stability of human simulation. \\n - (3) **Extending the proposed psychology-guided evaluation** by incorporating social and ethical dimensions, such as fairness and bias in simulated behaviors.\\n\\nWe believe that advancements in LLM-based human simulation will not only propel experimental sciences, such as psychology and sociology, but also enhance the accessibility of mental health resources. At the same time, **we hope that our work can inspire further interest and participation in human simulation research.**\\n\\n[1] Ashokkumar, A., Hewitt, L., Ghezae, I., & Willer, R. (2024). Predicting results of social science experiments using large language models. Technical report, Working Paper.\\n\\n[2] Park, J. S., Zou, C. Q., Shaw, A., Hill, B. M., Cai, C., Morris, M. R., ... & Bernstein, M. S. (2024). Generative Agent Simulations of 1,000 People. arXiv preprint arXiv:2411.10109.\\n\\n[3] Liao, Y., Meng, Y., Wang, Y., Liu, H., Wang, Y., & Wang, Y. (2024). Automatic Interactive Evaluation for Large Language Models with State Aware Patient Simulator. arXiv preprint arXiv:2403.08495.\\n\\n[4] Yu, X., Luo, T., Wei, Y., Lei, F., Huang, Y., Hao, P., & Zhu, L. (2024). Neeko: Leveraging Dynamic LoRA for Efficient Multi-Character Role-Playing Agent. arXiv preprint arXiv:2402.13717.\\n\\n[5] Higgs-Llama-3-70B is a powerful chat model based on Meta\\u2019s LLaMA-3-base. It is specially tuned for role-playing task.\"}", "{\"summary\": \"The paper introduces a benchmark to assess the potential of LLMs in simulating human behaviours and personality traits for psychological and sociological applications. This work includes the creation of the \\\"Human Simulacra\\\" dataset, which features detailed virtual characters with diverse life stories constructed with human feedback to enhance realism and ethical accuracy. The authors present a Multi-Agent Cognitive Mechanism (MACM), which simulates human memory and cognitive functions, allowing virtual characters to process emotions and memories for more realistic responses. Evaluation is conducted through a psychology-guided framework that includes self-reports for self-awareness and observer-based assessments where human judges evaluate character responses in various scenarios. Experiments comparing MACM to other simulation methods show that MACM enables LLMs to better replicate human-like behaviour, although limitations remain, particularly in capturing the nuanced adaptability of real human responses to social pressures. This benchmark aims to foster future research on using LLMs as proxies for human participants in psychological experiments while acknowledging ethical implications and the need for authentic simulations.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"Originality: This paper brings together ideas from psychology, cognitive science, and artificial intelligence to create a unique benchmark for evaluating how well language models (LLMs) can act like humans. Unlike previous studies that mainly focus on simple character traits or responses, this paper takes a deeper approach using Jungian psychology to model personalities with eight different dimensions. This gives a fresh perspective on capturing complex human traits. Additionally, the Multi-Agent Cognitive Mechanism (MACM) is a new tool that helps the models better remember, process emotions, and respond in context, making their behavior more human-like.\", \"quality\": \"The paper is thorough and well-executed. The Human Simulacra dataset is carefully built, with multiple rounds of expert review to ensure quality, accuracy, and ethical considerations. Each character's story is carefully crafted and reviewed to provide a deep foundation for testing the LLMs\\u2019 performance in simulating humans. The MACM\\u2019s design, which coordinates memory, emotion, and logical processing, is a clear improvement over simpler models that rely on only one type of agent or basic retrieval methods.\", \"clarity\": \"The paper is well-organized and clearly explains its methods and objectives. From the motivation to simulate human personalities, through the dataset creation, model mechanism, and evaluation framework, each part is easy to follow.\", \"significance\": \"This paper makes an important contribution to AI and psychology, especially by opening up possibilities for LLMs to replace humans in some psychological studies. By creating a foundation for simulating complex human traits, this benchmark could allow LLMs to ethically stand in for human participants in specific research settings.\", \"weaknesses\": \"No outstanding weaknesses.\", \"questions\": \"No questions.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": [\"> Do the llms see the stimuli in the conformity experiment?\", \"Yes! In the conformity experiment, all individuals except the LLM-based human simulation are told to stand up and announce an incorrect answer. This creates conditions that induce the human simulation to either resist or yield to group pressures when these pressures are perceived to be obviously wrong. We incorporated the corresponding scenario description into the prompt to inform the model, as detailed in Table 26 of the original submission.\", \"**Below, we present a table illustrating how two LLM-based human simulations with differing personalities responded to conformity stimuli.** The experimental results demonstrate that the LLM-based human simulation can understand and simulate the pressure in such scenarios and provide different responses based on the target persona\\u2019s personality.\", \"Table 4: The responses of human simulations with different personalities.\", \"| Character from Human Simulacra | Personality | Responses when faced with group pressure |\", \"|--------------------------------|------------------------------------------------------------|----------------------------------------------------------------------------------------------------------|\", \"| Mary Jones | She has strong personal beliefs. | Everyone's choosing 1, but it's clearly wrong. I'll say 3 despite the group consensus. |\", \"| Erica Walker | She often withholds her opinion to avoid upsetting others. | I'm baffled; they're all wrong. But disagreeing is daunting. I'll just go with line 1 to avoid conflict. |\", \"> RAG is not compared with the MACM baseline for the conformity experiments.\", \"Thank you for pointing this out. **We additionally tested 4 LLMs with 4 different simulation methods** in the replication experiment: (1) a blank model (which has no information about the target character), (2) a prompt-based method, (3) the Retrieval-Augmented Generation (RAG) method, and (4) the proposed MACM. In the experiments, we tasked the LLMs with simulating Erica, a girl who \\\"often withholds her opinion to avoid upsetting others,\\\" feels \\\"overwhelmed\\\" by group pressure, and chooses the incorrect answer. The experimental results are shown in the table below.\", \"Based on the analysis of the results, we discovered several interesting conclusions:\", \"(1) As the size of the LLMs' parameters increases, the quality and stability of the portrayal gradually improve.\", \"(2) Since the Blank model is unaware of the target character's personality, it always provides the correct answer and does not exhibit the bandwagon effect.\", \"(3) While the RAG-based simulacra can retrieve relevant life story chunks when answering questions, their performance is limited by the LLMs' information processing capacities. Excessive descriptive information may interfere with the LLMs' self-positioning, resulting in responses that do not align with the target personality.\", \"(4) Compared to the RAG-based simulacra, the prompt-based simulacra perform better in simulating character personalities. However, the personalities constructed by this method are relatively fragile and often deviate from the intended character traits.\", \"Table 5: Bandwagon effect observations of different simulacra.\", \"| Method | GPT-4-Turbo | Qwen-Turbo | Claude-3-Opus | Claude-3-Sonnet |\", \"|:-----------:|:-----------------:|:-----------------:|:-----------------:|:----------------:|\", \"| Blank | No | No | No | No |\", \"| Prompt | Yes, but unstable | Yes, but unstable | Rare | Rare |\", \"| RAG | No | No | No | No |\", \"| MACM (Ours) | Yes | Yes, but unstable | Yes, but unstable | Rare |\", \"Yes: The bandwagon effect can be easily reproduced and remains stable.\", \"Yes, but unstable: The bandwagon effect can be reproduced. However, the performance is unstable as the model sometimes overlooks the character's personality.\", \"Rare: The performance is unstable as the model often overlooks the character's personality.\", \"No: We did not observe the bandwagon effect.\", \"Thank you for your detailed questions about the paper. We recognize that it is our responsibility to ensure our paper is sufficiently clear and accessible to readers from diverse backgrounds. We hope that the responses provided above help to present the details of the paper more comprehensively. **We are open to any further discussions that could help address your concerns or clarify any remaining doubts.**\"]}", "{\"comment\": \"Thank you to the authors for the detailed explanations. I hope you can incorporate these revisions into the manuscript.\\n\\nI have reviewed the authors' responses and decided to maintain my previous rating.\"}", "{\"title\": \"General Responses and Summary of Revisions\", \"comment\": [\"We sincerely thank all reviewers for their feedback, which has helped us significantly improve our work. We are encouraged by the reviewers' recognition of several key strengths:\", \"\\\"_This paper makes an important contribution to AI and psychology, especially by opening up possibilities for LLMs to replace humans in some psychological studies._\\\" (Reviewer jHzr)\", \"\\\"_This paper proposes a high-quality simulated human profile dataset, developed with real human input, alongside a more advanced evaluation benchmark to assess the ability of large language models (LLMs) to emulate specific individuals._\\\" (Reviewer YX9L)\", \"\\\"_This study introduces a new benchmark to assess the ability of large language models to mimic human personalities in psychological experiments...highlighting both the potential and limitations of using large language models in psychological research._\\\" (Reviewer Dfn4)\", \"\\\"_The proposed benchmark for assessing the role-playing capabilities of large language models (LLMs) is constructed more rigorously than current benchmarks (to the best of my knowledge). It is based on more robust psychological theories and involves greater human effort to ensure high-quality data._\\\" (Reviewer Wr7p)\", \"Based on their constructive comments, we have made substantial improvements to strengthen our paper's contributions and address the limitations. The major enhancements include:\", \"**Extended Discussion.** We added comprehensive analysis of MACM's structure (Section 4.3), the rationale for using Jung's theory (Section 6, Appendix A.5), and potential applications of LLM-based human simulation (Appendix H).\", \"**Additional Experimental Results.** We provided additional experimental results highlighting the contribution of each agent within MACM (Appendix B.3) and observations of the bandwagon effect of different simulacra (Appendix E). These results further validate the conclusions drawn in our paper.\", \"**More Implementation details.** We supplemented the paper with additional details on character attribute design (Section 3.1), human involvement (Section 3, Appendix D.2.2), hypothetical scenarios (Appendix D.2.1), and the bandwagon effect replication (Section 5.2, Appendix E).\", \"We thank the reviewers again for their valuable input that has helped us improve the paper. We believe these additions have fully realized the potential of our research contribution to the field of LLM-based human simulation.\"]}", "{\"comment\": \"> MACM seems to work reliably with only one model and does not seem to generalize.\\n- In the original submission (Section 5.1), we conducted a self report evaluation involving 14 widely-used LLMs with 4 different simulation methods (None, Prompt, RAG, and MACM) on the Human Simulacra dataset. As shown in Table 2 of the original submission, MACM's scores significantly surpassed other simulation methods across multiple LLMs (e.g., Qwen-turbo, GPT-4). In stronger-performing LLMs like GPT-4 and GPT-4-Turbo, the MACM-based human simulation achieve the best results (88 points) in all tests, aided by emotional and logical analysis. Below, we provide a summary of Table 2 to illustrate the performance improvements.\\n- We also note that the effectiveness of the MACM method remains constrained by the LLMs' analytical capabilities. While the MACM can extract context-relevant, emotionally and logically rich memory fragments from long-term memory, less capable LLMs (e.g., Llama-2-13b) are unable to analyze and utilize this information. A large amount of descriptive information may interfere with the LLMs' self-positioning, resulting in inappropriate responses or misunderstanding of questions. The solution to this problem lie in adjusting the LLM's parameters to align with the target character's values, for instance, employing techniques such as fine-tuning or reinforcement learning, which will be a primary focus of our future work.\\n\\n- Table 2: The self report evaluation performance of 7 LLM-based simulations with different simulation methods. \\n | | None | Prompt | RAG | MACM (Ours) |\\n |:----------------:|:-----:|:------:|:-----:|:-----------:|\\n | GPT-4 | 20.00 | 78.67 | 82.67 | **86.67** |\\n | GPT-4-Turbo | 12.00 | 78.67 | 85.33 | **88.00** |\\n | Qwen-turbo | 24.00 | 69.33 | 72.00 | **74.67** |\\n | Llama-2-70b-Chat | 21.33 | 48.00 | 36.00 | **58.66** |\\n | Mistral-7b | 8.00 | 50.67 | 36.00 | **52.00** |\\n | Llama-2-13b | 17.33 | **30.67** | 21.33 | 22.67 |\\n | Vicuna-13b | 18.67 | **56.00** | 29.33 | 45.33 |\"}", "{\"comment\": [\"**Why were only 11 profiles accepted by experts? Is it because personality modeling is difficult?** Yes and no. It depends on the quality requirements for the generated personality profiles. The most challenging part of personality modeling is writing the 640 personality descriptions. Once those are completed, the personality of each character can be automatically derived from them.\", \"**The reason only 11 profiles were accepted is that we set a very high and rigid standard for acceptable profiles, aiming to ensure the uniqueness and coherence of each character.** These 11 characters were carefully selected from a pool of 100 candidates. Each of them is independent and represents a distinct group (we provided the calculation of character uniqueness in Appendix A.3).\", \"Given the specificity of the human simulation task, **it was essential to ensure that each character possessed a unique and coherent profile.** To achieve this, we first used GPT-3.5-Turbo to rank the character profiles based on their quality, filtering out those that were clearly unreasonable. Then, multiple human reviewers, including graduate students in computer science and psychology, manually reviewed the remaining profiles. They made minor adjustments to any flaws that GPT might have missed (e.g., a character who loves solitude having overly extroverted hobbies) and **ensured a balanced distribution with equal numbers of male and female characters, as well as representation across various age groups and family backgrounds.** While this rigorous selection process led to a low acceptance rate (11/100), it ensured the high quality of the dataset.\", \"We hope the above explanations address your concerns, and we look forward to further communication with you.\"]}", "{\"title\": \"Response to reviewer QPHb\", \"comment\": \"Thank you for taking the time to follow up on our rebuttal. After reviewing your feedback, we engaged in a focused discussion with psychology experts in the field of mental health. They provided additional perspectives on the matter. Below are our responses to your specific points and suggestions:\\n\\n### **1. Why do we use Jung\\u2019s theory?** \\n\\nThe field of psychology has yet to reach a consensus on various personality measurement theories, with each theory having its limitations. Specifically, \\n - **Big Five theory:** critics argue that the Big Five may oversimplify the complexities of human personality. They suggest that there are additional aspects of personality (e.g., Honesty-Humility [1] and egoism [2] ) that are not captured by the Five Factor Model (FFM). Some experts also point out that the FFM has unclear boundaries between dimensions, leading to potential overlap or ambiguity in personality assessment [3]. Additionally, the selection criteria for words used in factor analysis can be highly subjective. Researchers may make errors in deciding which trait terms to include or exclude, leading to biased factor structures that do not fully represent personality [4, 5, 6]. \\n - **Jung's personality type theory:** Although Jung's theory provides a framework for understanding complex human behaviors by emphasizing the dynamic interaction between different personality tendencies, the empirical support for Jung\\u2019s typology is relatively limited compared to the Big Five theory. Given that Jung's theory was developed in the early 20th century, it lacks the empirical backing found in more recent models. However, researchers have found that Jung's classifications align closely with the DSM-III's categories of personality disorders, offering a certain degree of validation for this typology [7, 8, 9].\\n- As an initial exploration, our goal was to establish a relatively complete personality modeling system. We reviewed several personality theories, including the Big Five theory, Jung\\u2019s theory, MBTI, and others. **After comparing their strengths and weaknesses and considering input from psychology experts, we chose Jung's personality type theory as the foundation for our 640 personality descriptions.** This theory, which provides a more comprehensive personality classification and emphasizes individual differences, offered the most suitable framework for our work.\\n\\n### **2. What measures have been taken to ensure the validity of our personality descriptions?** \\n\\nAs mentioned earlier, we recognized the concerns regarding the validity of Jung's theory. To ensure the validity of our personality descriptions, **we implemented several measures, including:** \\n - Requiring multiple human reviewers (including graduate students in computer science/psychology) to review each description, ensuring that it aligns with its intended tendency and does not overlap with others. \\n - Having psychology professionals supervise and review each description to ensure its psychological validity and completeness.\\n\\n**We agree that advanced psychological theories provide a strong foundation for human simulation research.** However, we believe that the suitability of the theory is key. Among the personality theories we reviewed, the most popular might be MBTI, which is often favored by researchers involved in streamlined, computer-aided psychological studies, though it is actually a limited interpretation of Carl Jung's theory. While the Big Five is more recent and supported by stronger empirical evidence, its 5 dimensions exhibit high intercorrelations, broad definitions, and unclear boundaries, which make it less suitable for our specific purposes. After conducting in-depth research on psychological theories, we decided to adopt Jung's theory. Through extensive experiments, we found that regardless of the theory employed, the quality of the final personality descriptions is the most critical factor.\\n\\nWe appreciate your thoughtful comments. In the future, we will consider progressively incorporating different personality measurement theories (e.g., the Big Five and psychiatric scales) under the guidance of psychology experts and relevant professionals.\"}", "{\"comment\": \"- Table 1: The differences between memory agent from the proposed MACM and traditional RAG approaches.\\n | Method | Retriever | Retrieval Content | Retrieval Methods |\\n |-------------------------|------------------------|------------------------------------------------------------------|---------------------------------------------|\\n | Memory agent from MACM | Using LLM as retriever | Long-term memory filled with information, emotions, and thoughts | Similarity between query and memory segment |\\n | Traditional RAG | Embedding search | Life story of the character | Similarity between query and life story |\\n\\n- Table 2: The performance of 4 LLM-based simulations with different simulation methods. \\n | | RAG | MACM (Ours) |\\n |:----------------:|:-----:|:-----------:|\\n | GPT-4 | 82.67 | 86.67 |\\n | GPT-4-Turbo | 85.33 | 88.00 |\\n | Qwen-turbo | 72.00 | 74.67 |\\n | Llama-2-70b-Chat | 36.00 | 58.66 |\\n\\n> No ablations are conducted with the MACM.\\n - Thank you for pointing this out. To analyze the contribution of each agent within MACM, **we additionally conducted ablation experiments across 3 LLM-based simulations.** For each ablation, we evaluated the simulation's performance using self-report evaluations. The table below presents the results of these experiments. The experimental results, as depicted in Table 3, lead to the following conclusions:\\n - Removing any agent leads to a decline in simulation performance, demonstrating the importance of all components in MACM.\\n - Replacing long-term memory retrieval with direct retrieval from the life story results in poorer performance, highlighting the critical role of structured long-term memory in maintaining consistency and producing contextually rich responses.\\n - Table 3: Ablation study results of 3 LLM-based human simulations.\\n | Method | GPT-4 | GPT-4-Turbo | Qwen-turbo |\\n |:-----------------------------------------------------:|:------:|:-----------:|:----------:|\\n | MACM | 86.67 | 88.00 | 74.67 |\\n | w/o Thinking Agent | 81.33 | 83.33 | 66.00 |\\n | w/o Emotion Agent | 83.33 | 85.33 | 71.33 |\\n | w/o Memory Agent | 82.67 | 84.00 | 68.67 |\\n | retrieval from life story instead of long-term memory | 84.00 | 86.67 | 69.33 |\\n\\n### **4. Details of the conformity experiments.**\\n> Why use character ai as a baseline? \\n - Character.ai (https://character.ai/) is a neural language model chatbot service that has millions of users. It allows users to design their own AI characters and converse with them. We chose Character.ai as a baseline because (1) it features a robust role-playing model specifically designed for LLM-based simulations, and (2) it supports long text inputs as prompts to guide the model effectively. \\n\\n> How are the models from character ai exactly used?\\n - In our experiments, we employed the life stories of 11 virtual characters from the Human Simulacra dataset as input prompts. These narratives were used to create corresponding chatbots on Character.ai, with which we interacted and recorded the experimental outcomes.\"}", "{\"comment\": [\"### **3. The benchmark's scope.**\", \"In this paper, we proposed a benchmark for LLM-based human simulation, which includes virtual character data, simulation method, and psychology-guided evaluation.\", \"> Which specific behavioral components are being tested?\", \"In the psychology-guided evaluation, we measured the quality of human simulation by testing behavioral components including:\", \"**Self-awareness:** This involves testing the simulations' memory and analytical capabilities concerning the target character information. The goal is to evaluate whether the simulation can correctly answer direct questions related to its preset identity (such as name, occupation, interests, etc.) as well as complex questions (such as attitudes towards relationships with others, life goals, etc.).\", \"**Behavioral Consistency:** This involves testing the simulation's decision-making, emotional expression, and behavioral consistency in hypothetical scenarios. The simulation is expected to exhibit behaviors and emotional responses that align with the target character's life story and personality traits.\", \"> How these components are operationalized in the benchmark?\", \"For self-awareness:\", \"**We employed self-report assessments to evaluate the simulations's ability to establish self-awareness.** Self-reporting is a common personality measurement technique that requires individuals to answer questions about themselves [10]. To this end, we **manually craft a set of questionnaires for each virtual character,** featuring fill-in-the-blank and single/multiple-choice questions. The test content covers key attributes, social relationships, and life experiences of the target characters. For example, ''What is your name?'', ''What do you think of your father?'', and ''What were the reasons behind not going through formal schooling for you?''. **Each question is carefully reviewed to ensure they align with the character's unique setting and the scores are evaluated based on exact matches.**\", \"For behavioral consistency:\", \"**We introduced observer reports, a cross-evaluation based on human judges,** to assess the simulation's thinking, emotions, and actions in hypothetical scenarios. Specifically, we crawled 55 hypothetical scenarios [11] designed to elicit human emotional responses or personality traits. Two examples of such scenarios are as follows:\", \"On your birthday, you are invited by some friends to a well-attended restaurant. While you are seated together, your friends suddenly start chanting 'Happy Birthday' loudly, and all the guests start looking at you. Then a waiter asks you if the restaurant staff can sing a birthday serenade for you? How do you respond?\", \"You're going to meet a friend. Shortly before the meeting, your friend asks you if they can bring along other friends you don't know. They cannot specify how many will join. How do you respond?\", \"We instructed each simulation to imagine themselves in the given scenario and to describe how they would feel and what actions they would take. **All responses were collected and submitted for cross-evaluation by human judges, involving four tasks:**\", \"**Personality Describing** (Human judges 1 and 2): analyze the scenario (Q) and simulation response (A), and describe the simulation's personality.\", \"**Description Scoring** (Human judges 3 and 4): assess whether the descriptions align with the target character.\", \"**Reaction Describing** (Human judges 3 and 4): explain how they would feel and what actions they might take in the scenario (Q) if they were the character.\", \"**Similarity Scoring** (Human judges 1 and 2): compare the similarity between the human responses and the simulation's responses.\", \"We calculated the average score from two judges for the same scoring task as the final score of the task. **The simulation\\u2019s final score was calculated as the sum of description matching score and response similarity score.**\", \"Additionally, our study serves as an initial exploration of human simulation designed to replace human participants in psychological experiments. To validate our research focus, we further evaluated our human simulations in the asch conformity experiment [12]. It is a well-documented, reproducible psychological experiment that provides an ideal testbed for our study.\"]}" ] }

[ "Niklas Muennighoff", "Hongjin SU", "Liang Wang", "Nan Yang", "Furu Wei", "Tao Yu", "Amanpreet Singh", "Douwe Kiela" ]

[ "large language models", "instruction tuning", "text embedding" ]

{ "note_id": [ "vaPUFZ4eP4", "v3wJqzF6Am", "sS8u6rAd1r", "jcV7TQf74g", "int62mVV1M", "dBVDY6teRK", "bRShXXWVr0", "YVi200zIHO", "WAkwuWxbAk", "TO9kRkCXZC", "Ky4TiZaO9r", "IyvHbxSJN5", "IUVZoUdcZL", "GT3ETWeGyl", "FQaHLastHA", "ENjTs3tsVa", "80blEWPhm8" ], "note_type": [ "official_review", "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_review", "official_comment", "official_review", "official_comment", "official_comment", "meta_review", "official_comment", "official_comment" ], "note_created": [ 1730673072802, 1737523732914, 1732058707120, 1732058650239, 1732058613017, 1732058627806, 1732574384680, 1730733118858, 1732557130152, 1730697746966, 1732894738320, 1730860978510, 1732557189137, 1732058638784, 1734766129921, 1732557164767, 1732557177097 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission5909/Reviewer_xDdr" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission5909/Authors" ], [ "ICLR.cc/2025/Conference/Submission5909/Authors" ], [ "ICLR.cc/2025/Conference/Submission5909/Authors" ], [ "ICLR.cc/2025/Conference/Submission5909/Authors" ], [ "ICLR.cc/2025/Conference/Submission5909/Reviewer_XQzj" ], [ "ICLR.cc/2025/Conference/Submission5909/Reviewer_vUnt" ], [ "ICLR.cc/2025/Conference/Submission5909/Authors" ], [ "ICLR.cc/2025/Conference/Submission5909/Reviewer_2rfu" ], [ "ICLR.cc/2025/Conference/Submission5909/Reviewer_xDdr" ], [ "ICLR.cc/2025/Conference/Submission5909/Reviewer_XQzj" ], [ "ICLR.cc/2025/Conference/Submission5909/Authors" ], [ "ICLR.cc/2025/Conference/Submission5909/Authors" ], [ "ICLR.cc/2025/Conference/Submission5909/Area_Chair_JB4U" ], [ "ICLR.cc/2025/Conference/Submission5909/Authors" ], [ "ICLR.cc/2025/Conference/Submission5909/Authors" ] ], "structured_content_str": [ "{\"summary\": \"The paper introduces a new framework called GRIT, which aims to unify text generation and embedding tasks within a single LLM, GRITLM. The model handles both tasks efficiently by distinguishing between them through instructions, which streamline use in multi-task applications like RAG. The authors demonstrate that GRITLM performs strongly on text representation and generation benchmarks, achieving competitive performance on the Massive Text Embedding Benchmark (MTEB) while also excelling in generative tasks.\", \"contributions\": \"1. Unified Generative and Embedding Model: The GRIT framework combines generative and embedding tasks within a single LLM. By using instructional prompts to distinguish between tasks, GRIT allows both generation and embedding without sacrificing performance. GRIT also reduces the need for separate models and complex infrastructure setups. This unification could simplify real-world deployments, particularly for applications that traditionally require both retrieval and generation components, such as search engines, recommendation systems, and conversational AI.\\n2. Efficient RAG catching design: The paper proposes innovative caching techniques, like Doc-Query Caching, and Query-Doc Caching, that significantly speed up RAG processes by reducing the number of forward passes required for long document processing. This approach reduces computational load for RAG tasks, enhancing efficiency in applications that rely on fast, context-sensitive retrieval and generation.\\n3. Competitive Performance Across Generative and Embedding Benchmarks: GRITLM achieves strong results on both the MTEB and several generative tasks, outperforming other open models of comparable size. This dual-task proficiency demonstrates that GRIT can match or exceed task-specific models, marking a significant step toward a general-purpose language model that handles both types of tasks seamlessly.\\n4. Task-Specific Performance Optimization: GRIT introduces several improvements, such as bidirectional attention with mean pooling for embedding tasks and mixed token-sample level loss aggregation for generative tasks. These innovations contribute to the model's performance across diverse tasks and offer insights into optimizing large language models for multi-task functionality.\", \"soundness\": \"4\", \"presentation\": \"4\", \"contribution\": \"4\", \"strengths\": \"Originality: The GRIT framework presents an original approach by unifying generative and representational capabilities within a single model, GRITLM, that can seamlessly switch between tasks based on instructional prompts. This concept is innovative as it directly addresses a long-standing limitation in language models: the need for distinct models optimized separately for generation and embedding. Previous work has focused on either generation or embedding, often leading to complex infrastructures where multiple models must be managed, synchronized, and deployed separately. GRIT\\u2019s unified approach not only simplifies these workflows but also brings both task types under one architecture without compromising performance. Additionally, GRIT\\u2019s application of caching techniques to accelerate RAG showcases an innovative use of model design to enhance efficiency, a departure from traditional RAG approaches that rely on separate models.\", \"quality\": \"The paper demonstrates a strong methodological foundation, supported by comprehensive experimentation and ablation studies. The authors provide detailed evaluations on major benchmarks, contrasting GRITLM\\u2019s performance with task-specific models to validate its efficacy as a multi-task solution. The robustness of the results is further confirmed through comparisons with proprietary models and current open-source alternatives, evidencing GRIT\\u2019s strong performance in both generative and embedding tasks. The use of ablations to explore trade-offs in task prioritization, loss aggregation, and memory efficiency contributes to the overall rigor, allowing readers to clearly understand how GRIT\\u2019s dual-objective structure was optimized. The paper\\u2019s experiments on efficiency gains with caching also underscore the quality of its findings, providing quantitative backing for its claims regarding speed improvements in RAG tasks.\", \"clarity\": \"The paper is well-organized and clear in its presentation, guiding the reader through complex ideas with a logical flow. Key concepts, GRIT\\u2019s caching mechanisms, and instructional tuning, are introduced with adequate background and broken down into understandable segments. Figures effectively support comprehension, making the technical details more accessible. The thorough presentation of results, including detailed tables and ablation analyses, provides clarity around GRIT\\u2019s performance relative to baselines, demonstrating where it excels and where there may be trade-offs. Additionally, the inclusion of an in-depth Appendix suggests a commitment to transparency and accessibility, ensuring that interested readers have the resources to delve deeper into implementation specifics and experiment configurations.\", \"significance\": \"The significance of GRIT lies in its potential to impact the field of NLP by simplifying multi-task language model deployment and reducing reliance on separate models for embedding and generation tasks. Furthermore, GRIT\\u2019s caching innovations for RAG tasks significantly reduce computational overhead and latency, especially in long-document settings, which is valuable for any application that relies on fast, context-aware responses. Moreover, GRIT\\u2019s design choices and improvements, such as mean pooling for embeddings and loss aggregation, may inspire further research into architectural unification across other language model tasks.\", \"weaknesses\": \"Storage Costs for Caching: The paper proposes innovative caching strategies to speed up RAG, but for example, Doc Caching, in particular, requires 30TB of storage for key-value states on GRITLM 7B. Such high storage demands are prohibitive in many real-world scenarios, limiting the practical usefulness of these techniques.\", \"instruction_dependence\": \"Including experimental results on GRITLM\\u2019s sensitivity to instruction phrasing and format would provide valuable insights into its robustness and areas for improvement.\", \"complexity_of_caching_mechanisms_and_trade_offs\": \"While the caching mechanisms offer significant speed-ups, they introduce substantial complexity to the model's architecture and inference workflow. The paper acknowledges that Query-Doc Caching can result in degraded performance due to mismatches in attention patterns. This complexity could make it challenging for practitioners to implement GRITLM optimally and may lead to inconsistent performance across different tasks and input types.\", \"questions\": \"Storage Costs for Caching: The caching techniques provide impressive speed improvements, but the storage requirements (e.g., 30TB for Doc Caching) are substantial. Can the authors discuss potential strategies to make these techniques more feasible for real-world use, particularly in terms of storage optimization?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}", "{\"comment\": \"Thanks a lot for your extensive review and highlighting the originality and novelty of the approach.\\n\\n**Storage Costs:** Query caching does not require any additional storage, only doc caching requires storing additional key-value states. For doc caching the KV cache can be fully offloaded to disk and does not need to be kept in memory. Disk storage is generally cheap. One can also store only part of the cache, e.g. only cache the key-value states of the first N layers. This will still lead to speed-ups but to a lesser extent. Thus, practitioners can flexibly choose the amount of storage they want to use depending on their specific setup.\\n\\n**Instruction Dependence:** GritLM can produce reliable embeddings without instructions \\u2013 in fact, for embedding documents during retrieval we do not use instructions. The generative part, however, indeed needs instructions. There has been prior work investigating the robustness of generative instruction-tuned models [1] and while it can be problematic at small scale, these issues generally go away at larger scales.\\n\\n**Complexity of Caching:** Note that caching does not change the architecture of the model. It does add some additional steps at inference, however, these are only ~3 lines of additional code to extract the key-value cache and repass it to the model. We provide open-source example code for running the caching in the supplementary material. Overall, we think that the caching is less complex than having to load and serve a second embedding/generative model as is necessary for current RAG setups that do not use GRIT.\\n\\n[1] Evaluating the zero-shot robustness of instruction-tuned language models by J Sun, C Shaib, BC Wallace, URL https://arxiv.org/abs/2306.11270\"}", "{\"comment\": \"Thanks a lot for your review and notes on GritLMs strong performance.\\n\\n**Comparison of training resources:** Great point! As we write in Lines 114-118, we believe that finetuning is so cheap compared to pretraining, that the additional training resources for GRIT don\\u2019t make a big difference. However, it may still matter in resource-constrained scenarios, thus we have added precise information on the GPU hours for each approach. Specifically: We used 72 GPU hours for the gen-only model, 1760 GPU hours for the emb-only model and 3072 GPU hours for GRIT. The GRIT number was already in \\u201cAppendix P: Hardware\\u201d and we have added the other two numbers there, too. Increasing the GPU hours for the gen-/emb-only model to match GRIT is unlikely to improve performance as all models have converged; Especially for the gen-only model, it would probably just lead to an excessive number of epochs. Nonetheless, we acknowledge that efficiency is a limitation and we have added more discussion on this in our \\u201cAppendix Q: Limitations and Future Work\\u201d, where we mention that packing and reusing the same samples for both the embedding and generative losses could significantly improve efficiency. We have uploaded the revised paper with the resource numbers and additional discussion - thank you for bringing this up!\\n\\n**Other base models:** We experimented with different base models (Llama2 and GPT-J) in Appendix A, Table 5, where we found that the approach works just as well but Mistral delivers better performance. In addition, we have also finetuned Mixtral using GRIT. All of these variants will be open-sourced.\"}", "{\"comment\": \"We thank all reviewers for their detailed reviews and great feedback! Below is a summary of all changes we have made to the paper in a new uploaded revision:\\n1. Added RAG results with BGE in \\u201cAppendix F: Additional RAG results\\u201d in response to Reviewer XQzj.\\n2. Rephrased the end of the Introduction to better motivate that GritLM requires less compute than separate generative and embedding models when considering pretraining thanks to a pointer from Reviewer XQzj.\\n3. Added citations to more recent work in Section 3.2 thanks to pointers from Reviewer XQzj.\\n4. Added more discussion on the potential speed-performance trade-off of using a smaller and faster embedding model by using the embedding from intermediate layers of GritLM in \\u201cAppendix A: Ablations\\u201d together with our embedding head ablation that explores the cost-performance trade-off of smaller embedding dimensions.\\n5. Added resources used by Gen.-only and Emb.-only baselines in \\u201cAppendix P: Hardware\\u201d thanks to the comment by Reviewer 2rfu.\\n6. Elaborated more on training discussion and potential avenues for improvement in \\u201cAppendix Q: Limitations and Future Work\\u201d in response to Reviewer 2rfu.\\n\\nOverall, we are glad reviewers have found the paper to be well-positioned and the methods to be original and novel. Reviewers have also pointed to the strong performance of the model and its usefulness. There was a lot of interest in using GRIT for RAG and the caching variants proposed - We are excited about further pushing these approaches together with the broader community.\"}", "{\"comment\": [\"Thank you for your detailed review. We are glad that you think the model will be useful and the work is well-positioned!\", \"**Mixed results:** As highlighted in the text we generally recommend doc caching (or query caching) but not the combined doc-query / query-doc caching mechanisms. We mostly present the query-doc and doc-query variants to inspire future work and are working on improving their performance in follow-up work. In Figure 5, we show that caching reduces latency on GPUs by around half compared to traditional RAG, which can be quite significant in time-sensitive applications. We note that the speed-up from doc (query) caching correlates directly with the length of documents (queries). E.g. for a book retrieval service where books are retrieved given user queries and each book has on the order of 10,000 or more tokens, the speed-up via doc caching would be significantly more than 2x, probably closer to 10x (depending on the query lengths). We ran RAG with a smaller model in Table 4, specifically, we ran using BGE as the embedding model which we also compare retrieval performance with in Table 1. The generative model is still GritLM-7B. Below are the match scores on NQ:\", \"BGE Large 0.34B: 10.39\", \"BGE Base 0.11B: 10.31\", \"BGE Small 0.03B: 10.17\"], \"from_the_paper\": \"- GritLM 7B: 30.50\\n\\nWe find performance to be significantly worse than with GritLM. Based on a manual inspection of samples, it appears that the embedding models commonly retrieve irrelevant passages that confuse the generative model. There may be other smaller embedding models or other generative models that may perform better, but overall we expect the RAG performance to be a function of the embedding and generative performance of the individual components (e.g. if an embedding model performs better than GritLM, we would expect it to lead to better RAG performance; BGE generally does not perform better on embedding as shown in Table 1). We have added this in Appendix F, thank you for raising it!\\n\\n**Modularity:** This is an interesting topic, thanks for bringing it up! One can only use the embedding/generative part of GritLM, thus it is still modular. However, in that case, some advantages to having the unified model are gone, such as e.g. query caching. The doc caching technique we introduce, however, still works even if embedding and generative models are separate. In that case, however, the entire corpus needs to be passed through the generative model once during index construction. From the compute perspective, retraining a GRIT model can be cheaper than a traditional RAG model. For GRIT, the pretraining and finetuning is both done using the same model, whereas for traditional RAG models, the embedding and generative model need to be pretrained and finetuned separately thus incurring more compute. We have added a note on this in the paper by rephrasing the end of the introduction, thanks for bringing it up!\\n\\n**More recent work:** Thank you for pointing us to these great works. We have added citations to them and several other recent embedding papers. Please let us know if there are any other works we should be citing.\"}", "{\"title\": \"Response to the authors\", \"comment\": \"Thank you for your response. I appreciate the authors adding additional results in Appendix F. This is great! I also want to thank the authors for pointing out that GritLM is still modular.\\n\\nFor these reasons, I will be increasing my score to 8 and confidence to 4.\"}", "{\"summary\": \"This paper introduces GritLM, a language model designed to excel at both text generation and embedding tasks. Current large language models (LLMs) typically specialize in one or the other, requiring separate models for applications that need both functionalities. GritLM addresses this limitation by employing a joint training approach.\\n\\nThe model architecture leverages a standard autoregressive generation head for text generation, trained with a next-token prediction cross-entropy loss. For embedding tasks, GritLM uses bidirectional encoding of the input prompt and mean pooling of the final hidden layer representations. A contrastive loss with in-batch negatives is applied to these embedding representations. The overall training objective combines these two losses, allowing the model to learn both tasks concurrently.\\n\\nExperimental results demonstrate that GritLM achieves competitive performance on both generation and embedding benchmarks, comparable to similarly-sized specialized models. Furthermore, the authors explore the benefits of this unified architecture in two specific scenarios: (1) reranking, where GritLM improves its own generated text through its embedding capabilities, and (2) retrieval-augmented generation (RAG), where the unified model serves as both retriever and reader, significantly reducing inference costs.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": [\"GritLM effectively demonstrates strong performance in both generation and embedding tasks within a single model.\", \"The paper presents a thorough experimental evaluation, including reranking and RAG scenarios, showcasing the practical advantages of the unified architecture.\"], \"weaknesses\": \"The scalability of the proposed method raises some concerns. The practicality of training and deploying a single model for both retrieval and generation may be limited to certain model sizes. In real-world applications, employing a smaller, faster embedding model alongside a potentially much larger generation model is often preferred. A smaller embedding model typically suffices for retrieval, while larger generation models are crucial for high-quality text generation. The paper would benefit from a discussion addressing the impact of model scale on the effectiveness of the unified approach and whether it remains advantageous when using vastly different-sized models for retrieval and generation. Specifically, quantifying the trade-off between performance and efficiency in such mixed-size scenarios would strengthen the paper's claims.\\n\\n(An alternative approach for using different sizes of embedder and generator is to use the output of the N-th layer (where N is relatively small) for embeddings instead of the last layer.)\", \"questions\": \"N/A\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Follow-up\", \"comment\": \"Dear Reviewer,\\n\\nWe'd appreciate it if you'd let us know if our response has addressed your concerns.\\n\\nThanks!\"}", "{\"summary\": \"This work introduces GRIT, a method to train a single large language model to excel at both generative and embedding tasks through differentiated instructions. They proposed GRITLM models achieve SOTA performance on the Massive Text Embedding Benchmark and surpass other models in generative tasks. GRIT unifies the two tasks without compromising performance, offering efficiency gains such as over 60% faster Retrieval-Augmented Generation for long documents. The unified model simplifies infrastructure by handling both embedding and generative tasks, reducing the need for separate models.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. GRIT introduces a novel approach that enables a single large language model to excel at both generative and embedding tasks, traditionally handled separately.\\n2. By eliminating the need for separate retrieval and generation models, GRIT speeds up RAG by more than 60% for long documents, which is a substantial improvement in processing time and resource management\", \"weaknesses\": \"1. The unified model requires more training resource, while there is no comparison of the performance between separate generation models and embedding models under the same resource consumption as shown in Table 1 and Table 2.\\n2. The paper uses the Mistral model as the base model. I think it would also be necessary to conduct experiments on the LLaMA series models to verify the robustness of the method.\", \"questions\": \"Please refer to the above.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to the authors\", \"comment\": \"Thank you for following up! Yes, your response has addressed my concerns. I appreciate the detailed clarifications provided.\"}", "{\"summary\": \"The paper presents generative representational instruction tuning (GRIT), a unified model for embedding and generative tasks in text. GRIT learns embedding representations with a bidirectional attention followed by mean pooling and a instruction tuning with causal attention. The experiments show that GRITLM outperforms various prior open models on the MTEB benchmark and matches the performance of several instruction tuning models. Furthermore, the unified model speeds up retrieval augmented generation by 60%.\", \"soundness\": \"3\", \"presentation\": \"4\", \"contribution\": \"3\", \"strengths\": [\"The paper is very well written. It is easy to follow the main motivation of the paper. The related work positions the paper well.\", \"The paper presents large-scale experiments. The model matches the performance of strong baselines on challenging benchmarks such as MTEB and instruction tuning datasets.\", \"The caching mechanism reduces latency for RAG, especially longer sequences.\", \"The GRITLM model will be useful for practitioners.\"], \"weaknesses\": \"**Mixed results**\\nThe main contribution of the unified method is to reduce the latency for generating the output. However, Table 4 shows a tradeoff between performance and latency. In Doc-Query and Query-Doc experiments, we see that GRITLM speeds up RAG but at the cost of overall performance. Furthermore, GRITLM does not show significant speed-ups on GPUs. Finally, I would be curious to see if a smaller embedding model (besides a smaller GRITLM) shows improved performance compared to the RAG performance in Table 4.\\n\\n**Modularity.**\\nOne of the main advantages of RAG is that it is modular. The separation of the embedding model and the generative model makes it easy to swap out either one of the components. With a unified embedding and generative model, the entire model has to be retrained which can be computationally expensive. \\n\\n**Include more recent work**\\nThe authors have acknowledged that the more recent embedding models, such as NV-Embed, show improved performance over GRITLM. It would be awesome if the authors cited more recent work [a, b] and more. \\n\\n[a] \\u200b\\u200bSFR-Embedding-Mistral:Enhance Text Retrieval with Transfer Learning.\\n\\n[b] Towards General Text Embeddings with Multi-stage Contrastive Learning\", \"questions\": \"Please see the weaknesses.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Follow-up\", \"comment\": \"Dear Reviewer,\\n\\nWe'd appreciate it if you'd let us know if our response has addressed your concerns.\\n\\nThank you!\"}", "{\"comment\": \"Thank you for your detailed review and highlighting the extensive experiments!\\n\\n**Scalability:** GritLM-7B is faster than a 7B generative model + a tiny embedding model for RAG when using the caching techniques we introduce. This is because the caching techniques (e.g. doc caching) will only require a single forward pass of GritLM-7B at inference, while in the other case, a forward pass for both the 7B model and the tiny embedding model is required. Without the caching techniques, speed indeed matters. We like your idea of using an intermediate layer and would expect it to lead to a performance drop while improving speed. In fact, we performed a similar experiment to reduce storage costs in Appendix A, Table 5 (e), where we find that we can downproject the embeddings to a 4x smaller dimension (->1024) at a small reduction in performance. Similarly, if we cannot use caching, we could increase speed 2x by taking the output of the middle layer at a slight reduction in performance. We have added a short note on this in Appendix A, thanks a lot for bringing this to our attention!\"}", "{\"metareview\": \"Previous embedding models and generative models are typically learned separately. This paper proposes to learn them together through massive multi-task training and different tasks are separated through instructions. Experimental results are strong, demonstrating the performance of this joint model can match the best performance from both worlds. One unique advantage of this model is the improved efficiency in RAG applications, where the model can reuse the encodings of the query in an RAG pipeline. The reviewers unanimously vote by acceptance of this paper with scores 6,6,8,8.\", \"additional_comments_on_reviewer_discussion\": \"The reviewers praised the GRIT framework for its originality, unified approach to generative and embedding tasks, strong performance on benchmarks, and efficiency in Retrieval-Augmented Generation (RAG). Key concerns included scalability, modularity, storage costs for caching, the dependency on instructions, and the lack of experiments with other base models or resource comparisons. The authors addressed these by providing additional results, clarifying GRIT's modularity and caching techniques, discussing storage optimizations, emphasizing the robustness of instructions at larger scales, and adding experiments with different base models and detailed resource usage comparisons. I feel these concerns are satisfactorily addressed by the authors.\"}", "{\"title\": \"Follow-up\", \"comment\": \"Dear Reviewer,\\n\\nWe'd appreciate it if you'd let us know if our response has addressed your concerns.\\n\\nThank you!\"}", "{\"title\": \"Follow-up\", \"comment\": \"Dear Reviewer,\\n\\nWe'd appreciate it if you'd let us know if our response has addressed your concerns.\\n\\nThank you!\"}" ] }

[ "Salvatore Calcagno", "Isaak Kavasidis", "Simone Palazzo", "Brondi Marco", "Luca Sità", "Giacomo Turri", "Daniela Giordano", "Vladimir R Kostic", "Tommaso Fellin", "Massimiliano Pontil", "Concetto Spampinato" ]

[ "Neural Circuit Dynamics", "Neural Activity Forecasting", "Vector Quantization" ]

{ "note_id": [ "nc9RF00icV", "mpFtmwJ7ho", "gf5a7RkJ81", "dDMBXlhnBr", "cKMAn5uHEh", "YcydUcaVRK", "X26hxZyV2s", "WvzLASuwtY", "WQxnLjS1JU", "Vdaj4hbFxU", "QYjIHMvnQs", "Q8IVCQiFHz", "Pi96Bbeehu", "LudUW8tZuU", "AobgojgoHG", "7qbSb9sJXx", "521dskiO1a", "05hdXmoY2x" ], "note_type": [ "official_comment", "official_comment", "comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_review", "official_review", "official_comment" ], "note_created": [ 1732556274900, 1733132714010, 1733132883748, 1732556416850, 1732695107242, 1730589934844, 1732628726444, 1732622644253, 1733132798812, 1732557152919, 1732557228063, 1732556103365, 1729008205162, 1732710921937, 1732912635427, 1730388323194, 1730566446292, 1732619816788 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission14015/Authors" ], [ "ICLR.cc/2025/Conference/Submission14015/Authors" ], [ "ICLR.cc/2025/Conference/Submission14015/Authors" ], [ "ICLR.cc/2025/Conference/Submission14015/Authors" ], [ "ICLR.cc/2025/Conference/Submission14015/Reviewer_U5z1" ], [ "ICLR.cc/2025/Conference/Submission14015/Reviewer_ULZ9" ], [ "ICLR.cc/2025/Conference/Submission14015/Reviewer_Tdst" ], [ "ICLR.cc/2025/Conference/Submission14015/Authors" ], [ "ICLR.cc/2025/Conference/Submission14015/Authors" ], [ "ICLR.cc/2025/Conference/Submission14015/Authors" ], [ "ICLR.cc/2025/Conference/Submission14015/Authors" ], [ "ICLR.cc/2025/Conference/Submission14015/Authors" ], [ "ICLR.cc/2025/Conference/Submission14015/Reviewer_Tdst" ], [ "ICLR.cc/2025/Conference/Submission14015/Reviewer_4NGY" ], [ "ICLR.cc/2025/Conference/Submission14015/Reviewer_ULZ9" ], [ "ICLR.cc/2025/Conference/Submission14015/Reviewer_U5z1" ], [ "ICLR.cc/2025/Conference/Submission14015/Reviewer_4NGY" ], [ "ICLR.cc/2025/Conference/Submission14015/Reviewer_Tdst" ] ], "structured_content_str": [ "{\"comment\": \"Thank you for the detailed and insightful feedback. We greatly appreciate the opportunity to address your concerns and clarify the motivations, methodology, and results presented in our work. Below, we provide detailed responses to your comments.\\n\\n1. Vector quantization was chosen to address the critical challenge of data sparsity in neuroscience. VQ compresses neural data into discrete representations, enabling robust pattern discovery (as shown in Fig. A-9) while preserving essential temporal structure. This transformation allows us to reformulate the response forecasting problem as a classification task, which is more effective than traditional regression methods that struggle with sparsity. We acknowledge that the motivation for this choice may not have been sufficiently emphasized early in the paper, and we will clarify it in the introduction to avoid confusion.\\n2. We are currently working to evaluate comparisons with NDT1 and MtM. However, this comparison is not straightforward because these methods were designed for calcium imaging spiking activity, while our approach operates on raw fluorescence traces. Although this may seem like a minor modification, it is challenging because fluorescence traces are continuous and noisier compared to the discrete nature of spiking data. To address this, we are experimenting with using inferred spike probabilities as inputs to these methods and converting their predictions back into time series data. We will include these results in future updates.\\n3. We indeed included this analysis in the paper; please refer to lines 474\\u2013476. The baseline in Tab. 3 is our encoder backbone, a BERT-like transformer encoder, followed by a linear layer. This model was trained using cross-entropy loss for activity classification and mean squared error (MSE) for forecasting, offering a direct comparison to our approach.\\n4. The observed superior performance of our method is due to its ability to use quantized indices for reconstructing the original shape of the signal, which enables it to capture more detailed patterns. In contrast, other methods rely on regression-based losses that focus on modeling average activity patterns. Although we cannot compute PSTHs strictly (as we are not using spikes), we can approximate them by averaging all responses for each stimulus and using this average as a baseline for comparison at test time. We included in the supplementary materials (Fig. R1) qualitative examples showcasing both cases where our model succeeds and where it fails, providing a balanced view of its performance relative to the baselines.\", \"minor\": \"1. If stimulus information were excluded, the model would rely solely on the temporal structure and patterns in the neural activity itself for forecasting, likely reducing its predictive accuracy, particularly for responses tightly linked to specific stimuli. However, we conducted interpretability analysis to examine the attention distribution during predictions (see Fig. A-8). While the stimulus token is important, it plays a relatively minor role compared to neuron identity and has a comparable influence to the neuron's state before the stimulus onset.\\n2. We asked ourselves this question during preliminary experiments. The rationale is that a latent representation is meaningful if it enables recovery of the original input. When the reconstruction loss is applied only to masked tokens, the representation of unmasked tokens becomes poor, as their primary role is limited to aiding in reconstructing the missing parts. Moreover, while we did not conduct an exhaustive ablation study due to time constraints, preliminary experiments indicated that using the combined loss led to improved performance.\\n3. The [NEURON] embedding represents the identity of each neuron, ensuring that the model captures neuron-specific dynamics. Unlike POYO, where each spike is tokenized as an event, our approach tokenizes trace patches rather than individual spikes. Regarding the dimensions, the neuron, stimulus, and patches are embedded into vectors of dimension d, which matches the model's embedding space.\\n4. We visualized training and test data distributions (find attached an example in the supplementary material, Fig. R2) and observed significant overlap. As fine-tuning is always performed within the same session, we ensure the model adapts to within-session variability. \\n5. We conducted ablation studies on the number of indices and found that the best performance was achieved with 32 codes. However, we are not suggesting that 32 is the optimal encoding size for neural information in general\\u2014this is specific to our methodology. The relatively small number reflects how, as the number of codes increases, many of them end up representing neutral conditions (flat activity), leaving only a few for peak patterns. This imbalance creates challenges in classification, as too many similar codes make the task more like a regression problem, reducing effectiveness.\"}", "{\"comment\": \"Thank you for your feedback. We understand your concern regarding the reliance on qualitative plots and attention maps as evidence. While these provide valuable insights, we agree that they are preliminary indications and not definitive evidence.\\n\\nRegarding point 5, we agree that raw and deconvolved calcium traces share similarities, as both are continuous time series. However, they also differ in their nature, with raw traces being inherently noisier and requiring additional effort to disentangle meaningful patterns. To run a fair comparison on deconvolved data, we believe it would be necessary to pretrain and finetuning our model appropriately, which would require an amount of time beyond the scope of this work and of the current time-window. Moreover, our focus on raw traces is motivated by minimizing preprocessing to align with real-time applications.\\n\\nWe believe our model should also work with deconvolved traces and could be tested on datasets like SENSORIUM in the future. However, for the current work, our primary goal has been to demonstrate the efficacy of the model on fluorescence data, which we believe is a valuable contribution on its own.\"}", "{\"withdrawal_confirmation\": \"I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors.\", \"comment\": \"We appreciate the reviewers' time and effort in evaluating our work. While we understand and respect their point of view, we do not fully share it. We believe our work addresses important challenges in neuronal forecasting and provides contributions especially in modeling single-trial neural dynamics and leveraging fluorescence data in innovative ways.\\n\\nAlthough we stand by the significance of our contribution, we also understand that the reviewers will remain unconvinced, and their perspective is unlikely to change. With this in mind, we have decided to withdraw the paper at this time. The feedback provided will help us improve and strengthen our work for future submissions.\"}", "{\"comment\": \"Thank you for your thoughtful and detailed feedback. Below, we address each of your concerns and clarify our approach and its motivations.\\n\\n1. We agree with this observation. However, traditional encoding methods are designed to predict the average response, which does not capture variability or trial-specific dynamics. In contrast, our approach focuses on forecasting individual trial responses, which is crucial for applications like online experiments with optogenetic manipulation. In such scenarios, real-time predictions of neural responses can guide dynamic adjustments, even when the visual stimulus is controlled.\\n2. To demonstrate how our approach captures trial-specific dynamics beyond the stimulus-driven average response, we visually compare our forecasting model to a PSTH-based baseline. Specifically, we averaged the responses across all trials in the training set to create a baseline (train average). During evaluation, we compared this baseline to both the average response across test trials (test average) and our model's average prediction (Refer to Fig. R1 in supplementary material). This analysis highlights how our method accounts for variability beyond the stimulus-driven average response.\\n3. We address this point using the attention maps provided in the appendix. These maps reveal that while neuron identity is the most significant factor in determining the forecasted indices, the final predictions are also influenced by the context surrounding the onset. This demonstrates that the model does not rely solely on stimulus-driven patterns but also incorporates pre-stimulus and trial-specific variability. Additionally, although the correlation value of 0.33 in Table 2 may seem modest, it reflects the inherent challenge of capturing trial-specific variance, which extends beyond replicating the average response.\\n4. (5) We have not directly compared our model's performance to the baselines on the original datasets reported by their authors, as our approach is specifically tailored to work with fluorescence traces, which differ significantly from data used in these studies. For the Allen dataset, we trained the baseline models ourselves under identical conditions to ensure a fair comparison. While we cannot claim to outperform these baselines on their originally tested datasets, our results highlight the strengths of our method in modeling trial-specific dynamics in fluorescence data.\"}", "{\"title\": \"I stand with my previous evaluation\", \"comment\": \"Thanks for your feedback and the additional details. However, they did not address my fundamental concerns. While *using* quantification might be novel, I do not think it solves an existing problem in neuronal forecasting. On the contrary, I think reducing it to classification produces problems because of limitation to the finite codebook. However, because of the choice of data set, we cannot say that for sure. So until you provide evidence that you can also forecast with much bigger image sets where training and test do not overlap, I will go with the baseline assumption that reducing it to 32 vectors does not help.\\n\\nI find the real time argument for fluorescence data weak, because you don't test your model under real time conditions. In addition, since fluorescence data will be stronger correlated over time (because of the Ca indicator response kernel), forecasting should be easier. Thus, you will have to accept the concern that you picked an \\\"easy\\\" dataset for your problem. Finally, should the model be compared in real time conditions in the future, I think it needs a justification when experimenters would care about real time forecasting. At the moment, with the given datasets, I could just use the previous responses of the neuron to that stimulus, average them, and use that as forecasting into the future. This, then, would also be a good baseline to compare against. \\n\\nClarification regarding the term \\\"image computable\\\": I would consider a model image computable, if you can present it with an unseen image at test time and the model produces a prediction for that.\"}", "{\"summary\": \"This paper introduces a large-scale model pretrained on the Allen corpus, which includes calcium imaging spiking activity from the mouse visual cortex under various stimulus conditions. It presents a transformer that uses vector quantization to create a set of neural codebooks for forecasting spiking activity. This quantization approach was shown to be effective in neural activity prediction, outperforming other baseline time series forecasting models. Additionally, the paper demonstrates positive scaling results across different stimuli and individual subjects.\", \"soundness\": \"1\", \"presentation\": \"2\", \"contribution\": \"1\", \"strengths\": \"This paper attempts to tackle an important problem of building \\\"foundation models\\\" for neuroscience that can predict spiking activity and classify responses to stimuli.\", \"weaknesses\": \"1. While vector quantization has not previously been used to build neuroscience foundation models, the author did not provide sufficient justification for choosing this specific model architecture.\\n2. The paper proposes only two types of downstream tasks for evaluating the foundation model. A more comprehensive evaluation is needed to assess the model's generalizability across diverse downstream tasks.\\n3. The paper lacks a scaling analysis to evaluate how effective the proposed backbone is for developing a foundation model.\\n4. The foundation model backbone could benefit from more rigorous benchmarking against existing methods on self-supervised prediction of spiking activity.\", \"questions\": \"**Major:**\\n1. In the introduction and related work section, the author cites many other time series models but doesn\\u2019t clearly motivate the choice of vector quantization for this work. Is this because this architecture has not been applied to neuroscience before? Although the author attempts to motivate the model choice in Section 3.3, it would be good to clarify the motivation earlier in the paper. **Could the author elaborate on why vector quantization was used or provide interpretable analysis, similar to what is found in the Appendix regarding the neural codebook?** I\\u2019m looking for a deeper discussion of how ML tools can help us answer unaddressed neuroscience questions, rather than just presenting another ML method that hasn\\u2019t been applied in the field.\\n2. Instead of comparing this model to other time series transformers, this method could be better benchmarked against existing work on self-supervised prediction of neural activity, such as NDT1 [1] and MtM [2]. Both methods can be repurposed for causal prediction of calcium imaging spiking activity and use masked modeling. **Could the author include experiments that compare their model to at least one of these approaches?**\\n3. Regarding model architecture, why not directly predict activity using a linear layer after the transformer encoder, as done in BERT? What is the rationale for using quantization and additional parameters in the transformer decoder? **An ablation study could help show the advantages of using quantization.** While Table 3 includes an ablation comparing quantization to an autoencoder, it would be more informative to compare it to a transformer baseline without quantization.\\n4. In Figure 3, I\\u2019m curious why the other baselines performed so poorly in predicting the target. It seems that **the evaluation could be conducted more carefully and fairly against other methods**. **For qualitative analysis, could the authors provide single-neuron peri-stimulus time histograms (PSTH) and single-trial activity after subtracting the PSTHs?** This would help clarify whether the model is only capturing the average pattern in the data.\\n\\n**Minor:**\\n1. The author includes a stimulus token for neural activity forecasting, but incorporating stimulus information can also be considered a form of neural encoding. What would happen if stimulus information were excluded?\\n2. In Equation (2), why is the loss computed on both masked and unmasked portions? What is the rationale for balancing these two components, and what advantages does this provide?\\n3. I find Section 3.4.1 confusing. The author states that \\u201cfeeding neuron and stimulus identifiers to the encoder is a key aspect of the approach.\\u201d Could this be clarified? Why not use per-neuron tokens as in POYO [3]? Additionally, what is the dimension of $t_1, ..., t_P$?\\n4. In the experiment section, the author mentions allocating two sub-sessions for training and testing, with each sub-session separated by 10-15 minutes. This interval seems quite long, and I wonder how the author addressed non-stationarity and potential distribution shifts in the neural data. Has the distribution of the training and test data been visualized?\\n5. In lines 392-394, the author claims that \\u201cbrain signals can be encoded with 32 indices.\\u201d It would be cool for the author to further interpret this finding. However, I only found ablation studies on the number of quantization indices and a visualization of the learned neural notebook in the appendix. Does the author have a hypothesis as to why 32 indices are optimal?\\n\\n[1] Ye, J., & Pandarinath, C. (2021). Representation learning for neural population activity with Neural Data Transformers. \\n\\n[2] Zhang, Y., Wang, Y., Jimenez-Beneto, D., Wang, Z., Azabou, M., Richards, B., ... & Hurwitz, C. (2024). Towards a\\\" universal translator\\\" for neural dynamics at single-cell, single-spike resolution. \\n\\n[3] Azabou, M., Arora, V., Ganesh, V., Mao, X., Nachimuthu, S., Mendelson, M., ... & Dyer, E. (2024). A unified, scalable framework for neural population decoding.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thanks!\"}", "{\"comment\": \"No, we conducted full fine-tuning on the encoder component.\\nHowever, it is important to note that the decoder (see Fig. 2 in the paper) was not re-trained, as its sole purpose is to reconstruct the signal.\"}", "{\"comment\": \"Thank you for your feedback.\\n\\nWhile it is true that training and test images (mostly) overlap, the trials themselves do not, as we are modeling single-trial responses. This ensures that our model learns and evaluates on distinct neural activity patterns, making this a reasonable choice for our setup. Regarding stimulus generalizability, we understand the concern, but we believe it is a minor issue. The extracted features from the encoder are either derived from an external pretrained model or encoded as continuous variables, which makes overfitting by the linear layer unlikely. Additionally, for stimuli such as locally sparse noise, the model is trained under conditions where each trial features a unique configuration of black and white dots, generated randomly, ensuring independence between train and test trials.\\n\\nIt is also important to note that the codebook is unrelated to the stimuli and instead depends on the response dynamics. The codebook size was selected through hyperparameter tuning to optimize the model\\u2019s ability to forecast responses effectively while balancing complexity and performance.\\n\\nOn our point of view, we do not think that fluorescence data represents an easy dataset. While calcium signals are indeed temporally correlated due to the calcium indicator response kernel, this also introduces significant noise and variability at the single-trial level, which makes forecasting more challenging. \\n\\nRegarding the real-time applicability, while our work does not explicitly test real-time conditions, the motivation is rooted in enabling single-trial inference, which is critical for experimental neuroscience contexts such as optogenetic manipulations. In such experiments, models that predict future neural activity can guide dynamic, real-time interventions, offering substantial advantages over methods relying on averaged responses, which fail to capture trial-specific variability.\\n\\nWe acknowledge your point that testing the model on real-time conditions and larger datasets with non-overlapping training and testing images would strengthen the argument. We are working toward expanding our evaluation to address these concerns in future studies.\"}", "{\"comment\": \"Thank you for your detailed and insightful feedback.\\n\\nAs a general consideration, the choice of 32 codes results from careful experimentation, balancing effectiveness within our methodology. Increasing codes poses challenges akin to regression-based approaches, with most codes representing baseline activity, complicating classification.\\n\\nIf \\\"image computable\\\" refers to image-based features, our tokenization process accommodates features from CNNs or low-level image features, allowing us to handle diverse stimuli.\\nWe appreciate your feedback regarding the finite set of stimuli and the potential overlap between training and testing data. While pretraining includes the same neurons and stimuli, the model captures trial-specific dynamics beyond average stimulus-driven activity. Preliminary evidence is shown in the supplementary material (Fig. R1), comparing our average response to the training average.\\n\\nRegarding neuron identities, not using them during pretraining encourages the model to focus on common features across neurons rather than neuron-specific patterns. This enables the model to learn generalizable patterns during pretraining. When neuron identities are later introduced during fine-tuning, the model combines general and neuron-specific information, achieving its full predictive capability.\\n\\nMore specifically regarding your questions,\\n\\n1. The clear benefit of our architecture lies in the use of quantization, which enables us to reformulate the forecasting problem as a classification task. \\n2. The Allen dataset was chosen because it provides fluorescence data making it a benchmark for studying neural dynamics in response to visual stimuli. Temporal leakage is carefully avoided by splitting session data along the temporal dimension: training and testing data from the same session are separated by a gap of at least 10 min, ensuring that the model cannot directly learn from overlapping temporal patterns within a session. \\n3. We cannot conduct experiments with completely separate neurons because our model relies on learning neuron identities, which are session-specific. Regarding different stimuli, we cannot separate stimuli into training and testing sets within the Allen dataset without shuffling data along the temporal dimension, which would introduce information leakage. For example, test responses might inadvertently be included in the training set as \\\"baselines.\\\" This limitation arises from the construction of the Allen dataset, not from our approach. However, our method does not rely solely on simple stimulus IDs, but rather on stimulus features. For example, for natural scenes, we use features extracted by a pretrained ResNet, which are then projected into the model's d-dimensional space via a trainable linear layer. We will take your suggestion to evaluate the model on the SENSORIUM 2022 dataset, in future updates.\\n4. (and 7) Derivative-based normalization is particularly effective for sparse signals characterized by distinct peaks because it mathematically emphasizes the sharp changes associated with these peaks while preserving their essential features. For a sparse signal $ x(t) $ consisting of distinct peaks, the signal can be approximated as: $x(t) = \\\\sum_i A_i \\\\delta(t - t_i),$ where $A_i$ represents the peak amplitudes and \\\\( t_i \\\\) are the peak locations. The derivative of this signal is given by: $x'(t) = \\\\sum_i A_i \\\\delta'(t - t_i),$ which highlights sharp changes at the peak locations due to the presence of the derivative of the Dirac delta function, $ \\\\delta'(t) $. To normalize the signal using its derivative, we define: $x_{\\\\text{norm}}(t) = \\\\frac{x(t)}{\\\\lvert x'(t) \\\\rvert}.$ This normalization amplifies the regions with significant changes (i.e., the peaks) while diminishing other less significant parts of the signal.\\n5. For each trial, we calculate the baseline activity using a 2-second window before the stimulus onset. We then compute the \\u0394F/F (change in fluorescence) relative to this baseline, which allows us to normalize the responses. We consider a neuron to be active if its \\u0394F/F exceeds 10%. \\n6. We use sparsity in a broad sense, meaning that neural activity is rare compared to base activity, which aligns with the structure of fluorescence data. We chose not to use deconvolved traces or spiking data because these require additional preprocessing steps, and our goal is to avoid such steps to remain consistent with a real-time scenario. We agree that deconvolved traces are indeed more sparse and could potentially provide additional benefits in certain contexts, but our focus is on minimizing preprocessing to streamline applicability. That said, we will test our model on the SENSORIUM 2022 dataset to evaluate its performance under different conditions. Regarding spike data, their fundamentally different nature makes quantization less meaningful, as spiking activity is already discrete and does not benefit from this approach.\"}", "{\"comment\": \"Thank you for your detailed and thoughtful feedback. We appreciate the opportunity to address your concerns and provide clarifications. Thank you for bringing the cited works to our attention. We will include Zhang et al. (2024), Zhu et al. (2022), and Ye & Pandarinath (2021) in the literature review to provide a more comprehensive context for our contributions. Regarding Zhu et al. (2022), which focuses on calcium imaging data, we will conduct a direct comparison to evaluate our model against this baseline. For the ablation study on the importance of stimulus tokens, we are currently running the comparison. In the meantime, we encourage you to refer to our interpretability analysis in the appendix, which provides insights into how the model utilizes stimulus tokens during forecasting.\\n\\nThank you also for pointing out inaccuracies. We will rephrase the relevant sentences to correct and clarify these points. First, we acknowledge that some cited works use deconvolved calcium traces, which involve additional preprocessing compared to raw fluorescence traces. To the best of our knowledge, raw fluorescence data without deconvolution are not commonly used in these works. We are actively working to extend our approach to include deconvolved traces for a more direct comparison. \\nHowever, to the best of our understanding, Antoniades et al. (2023) (Neuroformer) focuses exclusively on spike data, as outlined in the workflow section of their paper. While some approaches (e.g., Turishcheva et al. 2024a;b, Li et al. 2023, Xu et al. 2023a) do not rely on trial-averaged responses during training, they do use repeats during evaluation or focus solely on visual stimuli for prediction. The total number of unique neurons used in our work is 3,989, considering the partial overlap between sessions. The mentioned 230,000 traces refers to the total number of trials, not neurons. We will ensure the work accurately reflects existing contributions and highlights our approach's unique focus on trial-specific neural forecasting with minimal preprocessing and we will correct this terminology to avoid any misunderstanding.\", \"regarding_your_questions\": [\"For subject generalization, we pretrained the model on data from 10 mice and fine-tuned it on the 11th mouse. During fine-tuning, neuron-specific information was incorporated to adapt the model to the test mouse. Importantly, the neurons used during the fine-tuning phase were not included in the pretraining phase. For stimulus generalization, we agree that static and moving gratings share similarities, particularly for neurons that are less sensitive to motion. However, the conditions still differ because parameters like orientation and spatial frequency vary, creating a meaningful distinction for evaluating generalization. All results were averaged over 44 runs, representing fine-tuning experiments across different combinations of mice and stimulus types.\", \"The STIM token is derived from a representation of each stimulus, and what is learned is a linear projection that maps stimulus features to the embedding size. Specifically, we employ the following stimulus features:\", \"Natural Scenes: We use the features from the last layer of a ResNet-50 as input.\", \"Gratings: We encode information about spatial frequency, temporal frequency, orientation, and phase into four real-valued numbers.\", \"Locally Sparse Noise: We encode the \\\\(x, y\\\\) positions of white and black points into a vector. For gratings and locally sparse noise, we prefer our encoding method because ResNet-50 is pretrained on natural images, and thus, its representation of these stimuli most probably will not be significant.\"]}", "{\"comment\": \"We extend our thanks to the reviewers for their thoughtful and constructive feedback, which has significantly enhanced our understanding of potential improvements and has been instrumental in clarifying our work. Below, we address the key points raised in your collective comments.\\n\\nWe greatly appreciate the suggestion to evaluate our results using peri-stimulus time histograms (PSHT). While our approach does not rely on spiking data, we have approximated PSHTs by averaging responses across trials and included qualitative examples in the supplementary materials. These examples aim to provide a clearer perspective on how our model handles stimulus-driven variability, and we would like to further develop this direction in future iterations.\\n\\nWe acknowledge the importance of benchmarking against existing methods to contextualize our contributions. However, as noted, our approach focuses on raw fluorescence traces rather than deconvolved traces or spiking data, and the inherent complexity of this task makes direct comparisons challenging. Although we plan to incorporate evaluations using deconvolved traces for a more comprehensive analysis, we emphasize that our objective is to enable real-time neural response forecasting with minimal preprocessing and algorithmic dependencies. This focus on efficiency and practicality underpins our methodological choices and aligns with real-time experimental scenarios, where preprocessing steps like deconvolution may be impractical. As highlighted, we intentionally avoided deconvolution to streamline preprocessing and preserve the raw nature of fluorescence data. However, we agree that deconvolved traces could complement our approach, and we aim to incorporate these aspects in future evaluations. Similarly, while the Allen dataset provided a robust foundation for our analysis, we are exploring the SENSORIUM 2022 dataset to validate our model under broader conditions.\\n\\nOnce again, we thank the reviewers for their suggestions and constructive feedback. We have addressed specific comments in personal responses, and we hope this global overview provides clarity on the core aspects of our work.\"}", "{\"summary\": \"The paper suggests to use a transformer to forecast neuronal activity with a focus on mice 2 photon imaged neurons. To tackle the sparse responses, the authors suggest to add quantization and classification loss derived from the quantization. They also try to tackle generalization issue and interpret their learnt neuron tokens.\", \"soundness\": \"2\", \"presentation\": \"1\", \"contribution\": \"2\", \"strengths\": [\"Relatively original but not very clear paper, with interesting results.\", \"I specifically liked\", \"A novel technical approach and an interesting idea with quantization, which seems to improve the results.\", \"The qualitative analysis of cross-former in comparison with Quant-Former (A6-A7)\", \"The authors made the first steps towards model interpretability, mainly in appendix, which is important for biological research.\", \"I appreciate attaching the code, this is a huge plus for reproducability\"], \"weaknesses\": [\"Incomplete literature review and missing baselines\", \"The Zhang et al 2024 work is not mentioned (https://arxiv.org/pdf/2407.14668 ) This work does neuron based forecasting but on neuropixels data.\", \"The older works for neuronal forecasting, such as Zhu et al 2022 (https://www.nature.com/articles/s41593-022-01189-0) or Ye & Pandarinath 2021(https://arxiv.org/abs/2108.01210) are not mentioned and also not used for baselines (Zhu et al 2022 is for calcium data).\", \"While this models do not have stimuli tokens, one of them could still be a competitive baseline. I would also be interested to see the ablation showing the importance of the stimuli tokens for QuantFormer (is it about specific stimuli or stimuli type? )\", \"Incorrect statement about other works and incorrect citations\", \"line 185 `SENSORIUM (Wang et al., 2023))`. Wang et al., 2023 does not use data from either Sensorium 2022 or 2023 competitions. It also barely discuss the competition\", \"moreover, both SENSORIUM 2022 and 2023 provide spike traces data\", \"For example, lines 176-177 *all the encoding and decoding methods discussed above rely on spiking data*, while in the mentioned works, Wang et al., 2023 (cited incorrectly though), Sinz et al 2019, Antoniades et al 2023, Turishcheva et al 2024 a/b all use calcium traces.\", \"Same for lines 90-91, both Turishcheva et al 2024 a, and Microns (https://www.microns-explorer.org/cortical-mm3) provide open access to extensive datasets with calcium traces, not spikes.\", \"Lines 144-147 *Approaches such as Turishcheva et al. (2024a;b); Li et al. (2023); Xu et al. (2023a); Sinz et al. predict neural responses based on stimuli, but often rely on trial-averaged data and are not designed to forecast future neural activity on a single-trial basis without the use of synchronous behaviour variables, which are not accessible in online settings.* While, indeed these approaches do not do neuronal forecasting, at least three out of four mentioned papers do not rely on either repeats or behaviour for responses prediction, only on the visual stimuli. Adding behaviour indeed improves performance, while repeats are used only during evaluations.\", \"Mentioned Antoniades et al 2023 work could be used for neuronal forecasting as well\", \"The biological validity of the paper is not clear\", \"For table 2 it is not clear what are the upper/lower bounds for the metrics, which makes it hard to interpret how good all of the models generally are, as the correlation upper bound could not be one due to significant noise in the biological data. I would inspire the authors to use repeated stimuli within the session and follow Wang et al., 2023 to estimate at least the correlation upper bound.\", \"It is also not clear, if the model is actually able to reproduce linear-nonlinear phenomenas, which the neurons should be able to do (like here https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1009028 )\", \"The writing clarity could be improved.\", \"For example, in section `4.1 DATASET` it would be nice to explicitly write the amount of unique neurons. If the neurons across sessions for the same mice did not repeated, then its 11*3*250 $\\\\approx$ 8500. If I understood the appendix correctly, there were 250 neurons per session on average. If these were exactly same neurons across sessions, it's 11*250 $\\\\approx$ 3000. This makes very different impression compared to the 230 000 traces, which might be understood on neurons.\", \"Generalization experiment is not explained (see questions)\"], \"questions\": [\"How exactly do you perform the generalization experiment? You trained on 10 mice and evaluated on the other one?\", \"But how the neuron-specific tokens were trained then for the test mouse? Also, were this neurons involved during pretraining? If yes, that might compromise the generalizatibility measure, as the model has see this data. What are the generalization ability of other models, like cross-former? Also, how are these numbers averaged? I am also not sure if it is really a good idea to measure generalization training on moving gratings and predicting the static ones as for the neurons ignoring motion this would be very close stimuli.\", \"Figure 1 states that neuronal forecasting models should take stimuli as input bit based on Figure 2- the model does not take visual stimuli as input but rather the stimulus-specific learnable tokens. Are this token per stimuli or per stimuli category (aka natural images, gratings, etc)?\", \"Minor -\", \"In lines 518-519 *2D t-SNE on neuron embeddings (Fig. A-10) revealed that the [NEURON] token encodes neuron-specific statistics like activation probability* - the t-sne plot actually does not separate low and high-activated neurons, especially on the first plot. I would inspire authors to revise this statement\", \"Lines 369-370 *However, we excluded natural movies, as isolating individual neuron responses is challenging, and spontaneous activity, as it is not stimulus-related.* But how do you isolate spontaneuos activity for other stimuli categories?\"], \"flag_for_ethics_review\": \"['Yes, Other reasons (please specify below)']\", \"details_of_ethics_concerns\": \"Incorrect statements about cited papers\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you for your response. I remain unconvinced by the replies. Qualitative plots of responses (point 2) and attention maps (point 3) are hints at most, but not evidence. Regarding point 5, I don't see a major difference of raw and deconvolved calcium traces. They're both continuous-valued time series related by (approximately) a linear convolution step. It's not clear to me why a method that works on one would not work on the other. Given that you happily train their method on your data for comparison, it's not clear to me why you wouldn't be able to train your method on their data, too.\"}", "{\"title\": \"Thank you\", \"comment\": \"I would like to thank the author for their response. However, I believe the proposed method requires (1) clearer clarification of its novelty, and (2) a more in-depth evaluation and comparison with baseline methods before it can be accepted at ICLR. At this stage, the method has potential for improvement in the future. I hope my suggestions could help the author improve this paper.\"}", "{\"summary\": \"This paper introduces QuantFormer, a novel transformer-based model designed to forecast neural activity in mouse visual cortex using two-photon calcium imaging data. The authors reframe neural forecasting as a classification problem through vector quantization, and employ neuron-specific tokens to identify single neurons. The approach uses a two-stage training process combining pre-training through masked auto-encoding with downstream training for neural activity classification and forecasting. The network is trained on raw fluorescence traces from the Allen dataset traces rather than spike data or deconvoled fluorescence data.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"1\", \"strengths\": [\"Very paper is well written. Previous literature and related work is mostly addressed, although a few citations might be missing.\", \"The main novelty seems to be the quantization stage that can predict the activity of neurons as a function of 32 codebook vectors.\"], \"weaknesses\": [\"A closer examination of QuantFormer's architecture raises important questions about its broader applicability and evaluation methodology. The encoder-quantizer-decoder architecture, while novel, relies on a relatively small codebook of 32 entries to represent all possible neuronal activity patterns. Since the codebook and decoder remain frozen after pre-training (as far as I understand), this potentially limits the model's ability to represent novel activity patterns not seen during pre-training. This yields a model that is strictly not image computable and has therefore less abilities than a simple CNN model based on images (which could be trained to forecast as well). This is a major limitation of the model that is not properly discussed.\", \"As a direct result from using a finite set of stimuli, also the training and test protocol raises questions. As far as I understand, the pre-training phase includes data from the same neurons and images that appear in the test set, which may allow the model to learn specific response patterns before the actual testing phase. Because the Allen datasets includes repetitions of stimuli, this could be a major confounder for the results as the model can simply learn the mean responses given the the stimulus. The authors describe that they do not use neuron identities for pretraining. However, I (a) find this a questionable choice because this should severely limit the prediction capabilities of the model (what\\u2019s common to all neurons in cortex?) and (b) I am still not convinced that this would avoid the problem that the model learns mean responses of neurons.\", \"Because of that, I find the contributions overstated:\", \"Forecasting for optogenetic manipulations is mentioned I could not find any experiment on that.\", \"Forecasting has been addressed by other transformer architecture, for instance the \\u201cuniversal translator\\u201d by Zhang et al (see below), which is not cited or compared to as far as I can see.\", \"Reframing forecasting of time series as a classification problem is per-se not a contribution if it doesn\\u2019t solve problems. As argued above, it seems to create problems.\", \"Handling arbitrary neuronal populations is not new as other works (such as the POYO model) already use neuron ID tokens.\", \"Finally, I would hardly call a model that is trained to forecast neurons from a finite set of stimuli a \\u201cfoundation model for visual cortex\\u201d. In particular, I would expect a foundation model to be image/video computable.\", \"I find the choice of dataset not well motivated. The authors argue with real-time applicability. But then they don\\u2019t test it in those conditions. So in that sense they could apply it to a bigger dataset such as SENSORIUM 2022 (if they still want to exclude videos). My guess is that the method will not work well as it contains many unique stimuli in the training set. In particular, the choice of dataset is at odds with the motivation for the codebook (sparsity). I would expect fluorescence data to be less sparse than deconvolved data (such as Sensorium). It that sense SENSORIUM or spiking data should be even better data. Finally, I do not understand how they can get baseline activity for neurons in the Allen data. As far as I remember about the dataset, images are presented back to back. This means that neuronal firing does not return to baseline between images. I do not see how this is addressed in the paper.\", \"I find the choice of models to compare to a bit weak. I would recommend to include at least an oracle estimator that uses the mean responses of the neuron to that stimulus in the training set. Additionally, my guess is that a model pretrained properly on SENSORIUM and then trained to forecast a fixed number of steps in the future, should be competitive.\", \"Why is neuronal identity and the stimulus ignored during pretraining? I do not understand the rationale for it. Why is it not trained on forecasting with neuronal activities since this seems to generate problems (as discussed in 3.4.2)\", \"I find motivation for the classification into active and non-active not clear. It somehow assigns a special role of 10% more activity, which seems arbitrary. It also raises a question how the baseline is computed if the images are shown back to back (see above).\", \"I do not find the evaluation metrics very clear. How are correlations computed (across what and are correlations averaged over).\", \"In the appendix, the authors show a table with forecasting of unnormalized responses. The scores there are much lower. I do not find the explanations of the authors very clear here. I think this raises a question whether the normalization scheme somehow favors some models. Maybe a visual comparison of of normalized vs. non-normalized responses would help. Or a more detailed motivation for why normalization by the accumulated gradient helps. Also, I do not find this very clear (What gradient? Accumulated over what? Isn\\u2019t an accumulated gradient equal to the original signal up to a constant?).\", \"**Minor weaknesses (I assume you will handle those, no need to respond to them):**\", \"`citep` and `citet` not consistently used. Please double check to use `citet` for inline citations and `citep` else.\", \"I would not call deconvolved calcium signals \\u201cspiking activity\\u201d. For instance, Turishcheva et al. does not use spikes, but deconvolved Ca++ activity.\", \"Possibly additional work to cite for forecasting\", \"Zhang, Y., Wang, Y., Jimenez-Beneto, D., Wang, Z., Azabou, M., Richards, B., Winter, O., International Brain Laboratory, Dyer, E., Paninski, L., & Hurwitz, C. (2024). Towards a \\u201cuniversal translator\\u201d for neural dynamics at single-cell, single-spike resolution. In arXiv [q-bio.NC]. arXiv. Retrieved from http://arxiv.org/abs/2407.14668\", \"Schmidt, F., Shrinivasan, S., Turishcheva, P., & Sinz, F. H. (2024). Modeling dynamic neural activity by combining naturalistic video stimuli and stimulus-independent latent factors. In arXiv [q-bio.NC]. arXiv. Retrieved from http://arxiv.org/abs/2410.16136\", \"Possibly interesting work to cite for neuronal quantization\", \"Wei, X.-X., Zhou, D., Grosmark, A., Ajabi, Z., Sparks, F., Zhou, P., Brandon, M., Losonczy, A., & Paninski, L. (2019). A zero-inflated gamma model for post-deconvolved calcium imaging traces.\", \"Wrong reference for SENSORIUM. The correct ones are (alternatively use the Retrospective papers from NeurIPS 2023 or NeurIPS 2024)\", \"Turishcheva, P., Fahey, P. G., Hansel, L., Froebe, R., Ponder, K., Vystr\\u010dilov\\u00e1, M., Willeke, K. F., Bashiri, M., Wang, E., Ding, Z., Tolias, A. S., Sinz, F. H., & Ecker, A. S. (2023). The Dynamic Sensorium competition for predicting large-scale mouse visual cortex activity from videos. In arXiv [q-bio.NC]. arXiv. Retrieved from http://arxiv.org/abs/2305.19654\", \"Willeke, K. F., Fahey, P. G., Bashiri, M., Pede, L., Burg, M. F., Blessing, C., Cadena, S. A., Ding, Z., Lurz, K.-K., Ponder, K., Muhammad, T., Patel, S. S., Ecker, A. S., Tolias, A. S., & Sinz, F. H. (2022). The Sensorium competition on predicting large-scale mouse primary visual cortex activity. In arXiv [q-bio.NC]. arXiv. Retrieved from http://arxiv.org/abs/2206.08666\"], \"questions\": [\"Can you motivate a clear benefit of your choice of your architecture?\", \"Can you cleaner motivate the choice of the Allen dataset and how you avoid data leakage between training and test? Could you conduct an experiment with completely separate neurons and images in training and test?\", \"Related: Can you run your model on the SENSORIUM 2022 data to show that it can deal better with unique image IDs?\", \"Can you provide a better explanation for your normalization scheme and why it improves model performance?\", \"Can you define how baseline activity is measured given the back-to-back image presentation in the Allen dataset.\", \"Can you address the potential mismatch between the sparsity motivation and the use of fluorescence data rather than deconvolved or spiking data?\", \"Can you provide a more detailed explanation of your normalization method, including a clear definition of the \\\"accumulated gradient\\\" and how it is calculated?\", \"Can you conduct an analysis of how the normalization scheme affects different models' performance, to ensure it's not unfairly advantaging certain approaches?\", \"Can you provide a clearer justification why this particular normalization method was chosen?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The paper proposes a transformer-based architecture for single-neuron response forecasting. In a first step, the authors use autoencoding to pre-train an encoder or neuronal response sequences. In a second step, they fine-tune the encoder on an activity classification and a response forecasting task. They evaluate their model using visually evoked responses in the visual cortex of mice. Compared to a few forecasting models they achieve improved activity classification and response forecasting metrics.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": [\"Novel architecture and interesting idea in principle\", \"Paper is well written and easy to follow\", \"Shows some ablation experiments to tease apart which components are important\"], \"weaknesses\": [\"Doing response forecasting on visually evoked responses seems like an odd choice\", \"Unclear how strong the baselines are\", \"Simple baselines like PSTH or linear encoding models are missing\"], \"questions\": \"While I like the overall modeling approach and thinks it\\u2019s sane in general, I am somewhat confused by the authors\\u2019 choice of evaluation using visually evoked responses, which leaves me with lots of question marks whether the model works and how well. In my opinion, it is simply impossible to deduce anything about the performance of this model from the evaluation presented by the authors, because it does not properly deal with the visual stimulus. Although I doubt it, it is possible that I\\u2019m missing something. I would therefore like the authors to answer the following:\\n\\n 1. When response forecasting is the goal, why do you use visually evoked responses, where the response is primarily determined by the stimulus rather than by past or pre-stimulus activity? There is some discussion around using the model online in experiments for optogenentic manipulation, but this motivation is not clear to me. In an experiment you control the visual stimulus that is shown, so you could easily incorporate it.\\n 1. If you choose to evaluate on visually evoked responses, the null model that you would have to beat is to simply take the PSTH of the neuron in response to the stimulus. I understand that your hypothesis may be that neurons do more than just responding to the stimulus and your goal is to explain this additional \\u201cnoise\\u201d \\u2014 but to drive home this point you first need a convincing stimulus-response baseline, onto which you can add the forecasting component. \\n 1. It appears to me that your model is trying to squeeze the stimulus-driven response patterns into a combination of neuron id token and stimulus token. Can you provide evidence against my hypothesis? Have you quantified for what fraction of the neurons\\u2019 response variance (during x_f) the PSTH accounts? Does your forecasting model exceed this value? I doubt it, given the correlation of 0.33 in Table 2.\\n 1. If you do not want to model the stimulus-driven response via an encoding model (or the PSTH), you should evaluate on datasets that do not have such strong external drive.\\n 1. Can you comment on whether your model beats any of your baselines on the datasets on which they have been tested and reported by the original authors? Did you train them yourself on the Allen dataset?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"details_of_ethics_concerns\": \"N/A\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thanks a lot for clarifications.\\nI have one more question\\n \\n> For subject generalization, we pretrained the model on data from 10 mice and fine-tuned it on the 11th mouse \\n\\nDid you freeze some parts of the model during fine-tuning?\"}" ] }

[ "Yang Qin", "Chao Chen", "Zhihang Fu", "Ze Chen", "Dezhong Peng", "Peng Hu", "Jieping Ye" ]

[ "Text-to-SQL", "LLMs" ]

{ "note_id": [ "zNQDJ2FPkZ", "xqoPjlKYPS", "siBixJ6KxK", "qiOaSszWJY", "p8F3kfR7dB", "ny6FXpZSIR", "gSyYJfnQzw", "fwPlHOGQZE", "fiAiXli95R", "daAoIfPaY0", "bHMorTCNzZ", "YLM1PChoT3", "XjE1FwWi3i", "Ugz8rzHScR", "UCvfEVXIp1", "S9fuVtoZlC", "LejQT09hhu", "J1NUTyQurV", "INlt8SOUPS", "GgDcmIAJTp", "Bf6w6BhALN", "AJRpeLidcI", "A2yKrKt9cQ", "73Ok3Nj2hq", "5nmYP7oTGu", "3gPwnbVxEI", "3IsFpsr0iU", "1OvrFuBuc0", "0nldGH0Kl5" ], "note_type": [ "official_comment", "official_review", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "decision", "official_comment", "official_comment", "official_comment", "meta_review", "official_review", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment" ], "note_created": [ 1732342404782, 1730573732429, 1732426067107, 1732342161575, 1730445952437, 1732373973286, 1732341863483, 1732963085012, 1732715811564, 1737523734999, 1733192190735, 1732715732747, 1732342125331, 1734745379030, 1730317324701, 1732715859566, 1732342272330, 1732678739508, 1730490545667, 1732341972402, 1732725014214, 1732341936661, 1732342300845, 1732342211898, 1732513969286, 1732681561303, 1732537701222, 1732715682814, 1732341819681 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission5945/Authors" ], [ "ICLR.cc/2025/Conference/Submission5945/Reviewer_2daL" ], [ "ICLR.cc/2025/Conference/Submission5945/Authors" ], [ "ICLR.cc/2025/Conference/Submission5945/Authors" ], [ "ICLR.cc/2025/Conference/Submission5945/Reviewer_TcaB" ], [ "ICLR.cc/2025/Conference/Submission5945/Reviewer_Ej7x" ], [ "ICLR.cc/2025/Conference/Submission5945/Authors" ], [ "ICLR.cc/2025/Conference/Submission5945/Authors" ], [ "ICLR.cc/2025/Conference/Submission5945/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission5945/Reviewer_cCGH" ], [ "ICLR.cc/2025/Conference/Submission5945/Authors" ], [ "ICLR.cc/2025/Conference/Submission5945/Authors" ], [ "ICLR.cc/2025/Conference/Submission5945/Area_Chair_Q4zL" ], [ "ICLR.cc/2025/Conference/Submission5945/Reviewer_Ej7x" ], [ "ICLR.cc/2025/Conference/Submission5945/Authors" ], [ "ICLR.cc/2025/Conference/Submission5945/Authors" ], [ "ICLR.cc/2025/Conference/Submission5945/Reviewer_TcaB" ], [ "ICLR.cc/2025/Conference/Submission5945/Reviewer_cCGH" ], [ "ICLR.cc/2025/Conference/Submission5945/Authors" ], [ "ICLR.cc/2025/Conference/Submission5945/Reviewer_cCGH" ], [ "ICLR.cc/2025/Conference/Submission5945/Authors" ], [ "ICLR.cc/2025/Conference/Submission5945/Authors" ], [ "ICLR.cc/2025/Conference/Submission5945/Authors" ], [ "ICLR.cc/2025/Conference/Submission5945/Reviewer_TcaB" ], [ "ICLR.cc/2025/Conference/Submission5945/Reviewer_2daL" ], [ "ICLR.cc/2025/Conference/Submission5945/Authors" ], [ "ICLR.cc/2025/Conference/Submission5945/Authors" ], [ "ICLR.cc/2025/Conference/Submission5945/Authors" ] ], "structured_content_str": [ "{\"title\": \"Meta Reply\", \"comment\": \"We sincerely thank all reviewers for spending their time to carefully review our paper and providing constructive suggestions, which helped to further improve our manuscript. Here, we provide a comprehensive overview of the reviewers' feedback and outline the corresponding modifications we have made.\\n\\n### The merits of our submission\\n1. The significant performance improvements and superior performance on open-source LLMs. (Reviewer 2daL, Ej7x)\\n2. The experiments are comprehensive and the effectiveness of the proposed method is verfied. (Reviewer cCGH, TcaB, Ej7x)\\n3. The paper is easy-to-follow. (Reviewer Ej7x) \\n\\n### Important suggestions and concerns\\n1. A more comprehensive evaluation on the Dr.Spider benchmark.\\n2. Accuracy evaluation on the task of schema linking.\\n3. The impact of each SQL-related task and the relative improvement.\\n4. The clarification on \\u2018hallucinations\\u2019.\\n5. The clarification on differences with existing methods.\\n\\n### Our Revisions\\nAccording to the reviewers' constructive and insightful feedbacks, we have carefully revised our paper, and provided comprehensive experiments, analysis in the appendix.\\n\\n- In main text, we double-checked and modified some descriptions based on the modification suggestions and highlighted them in blue.\\n- In Appendix A.3, we provide more comparison results with recent works.\\n- In Appendix A.4, we evaluate the performance of our schema linking module.\\n- In Appendix A.5, we provide the ablation results on Dr.Spider benchmark to further verify the effectiveness of our method.\\n- In Appendix A.6, we explore the impact of single-task SFT to further understand the interplay of multiple tasks.\\n- In Appendix A.7, we further clarified some details in the step of noisy correspondence filtering.\\n- In Appendix A.8, we further discuss the innovations of our ROUTE and how it differs from recent related methods.\\n\\nFinally, we sincerely thank all reviewers and ACs for their efforts.\"}", "{\"summary\": \"This paper addresses the limitations of current Text-to-SQL approaches that rely heavily on in-context learning using closed-source LLMs, such as GPT-4, which can cause privacy issues. To overcome these issues, the authors propose ROUTE, a comprehensive solution to enhance open-source LLMs' Text2SQL capabilities. ROUTE utilizes a multitask supervised fine-tuning approach incorporating tasks like text-to-SQL, schema linking, noise correction, and continuation writing to broaden the model's SQL generation skills and reduce the risk of overfitting. Additionally, a Multitask Collaboration Prompting (MCP) strategy is employed during inference to decompose the SQL generation process into simpler sub-tasks, reducing hallucinations and improving performance.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1) The proposed method significantly improves the performance of open-source LLMs and outperforms all existing methods trained on open-source LLMs.\\n\\n2) The proposed MCP approach not only enhances the performance of models trained with MSFT but also improves other models.\\n\\n3) The novel MSFT method substantially boosts model performance compared to standard SFT.\", \"weaknesses\": \"1) Although this paper focuses more on open-source LLMs, some recent approaches, such as CHASE-SQL, Distillery, and CHESS, are not included as benchmarks in their experiments.\\n\\n2) The proposed approach is a multi-step pipeline that can be prone to error propagation. To better understand the performance of the schema linking module and ensure it is not introducing errors into the pipeline, it would be beneficial to report the precision and recall of the schema linking module, as done in CHESS and DTS-SQL.\\n\\n3) The performance gap with the close-source LLMs is still large, roughly 13% on BIRD development set, which makes the applicability of this approach limited to the scenarios where privacy and local LLMs is essential.\", \"questions\": \"1). For the open-source LLMs and super large databases such as some of the databases in BIRD benchmark, how these large schema are fitted into the prompt of the open-source models?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"details_of_ethics_concerns\": \"NA\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Dear Reviewer Ej7x,\\n\\nWe greatly appreciate your timely feedback and positive support. We will further refine our manuscript to improve its quality. Thanks again for your feedback and taking the time to consider our response.\\n\\nBest, Authors\"}", "{\"title\": \"Official reply to Reviewer TcaB [2/3]\", \"comment\": \">W3: The authors mention that \\\"This strategy leverages collaboration across several SQL-related tasks to reduce hallucinations during SQL generation.\\\" However, the term \\\"SQL hallucinations\\\" is not defined, nor is there any discussion in the experimental section explaining how hallucinations are reduced. This claimed advantage, therefore, remains unclear.\\n\\n**Ans:** We follow several recent works to utilize hallucinations in SQL generarion task [4,5]. Hallucinations in LLMs refer to cases where LLM generate plausible but factually incorrect or nonsensical information [6]. Schema hallucinations and logic hallucinations are widely observed in LLM-based SQL generation [6]. \\n\\nIn this work, we present a multi-task training and a collaborative prompting framework, which significantly improved the accuracy of Text2SQL, in other words, it reduced hallucinations during the SQL generation process.\\n\\nConsidering that the definition of hallucinations in SQL-related tasks might be unclear, we have revised the usage of SQL hallucinations in the revised submission. Please see our revised submission for details.\\n\\n>W4: If Schema Linking, Noise Correction, and Continuation Writing are considered important, could the authors provide the relative improvement metrics for these tasks?\\n\\n**Ans:** Thank you for your constructive suggestions. As suggested, we provide relative improvement indicators for three tasks. For Schema Linking (SL), we report the Recall/Precession scores of predicted related tables and columns. For Noise Correction (NC), we reports the EX scores of SQL queries refined with Noise Correction on the output SQLs of Llama3. For Continuation Writing (CW), we reports the EX scores of all SQL queries obtained by continuation writing on half of ground-truth SQL queries. The results are shown in the following table:\\n\\n| |SPIDER-SL|SPIDER-SL|BIRD-SL|BIRD-SL|SPIDER-NC|BIRD-NC|SPIDER-CW|BIRD-CW \\n|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|\\n| MSFT|Table-R/P|Column-R/P|Table-R/P|Column-R/P|Dev-EX|Dev-EX|Dev-EX|Dev-EX|\\n| \\u2713|97.38/95.71|98.59/96.98|90.87/90.22|96.13/90.89|83.4 |53.4 |91.1 |73.9|\\n| |88.35/76.37|94.83/91.46|83.77/75.38|89.55/86.39|72.1 |38.1 |80.3|57.6 |\\n\\nAs observed, we can see that MSFT has improved all three tasks, especially achieves an amazing EX score of 91.1 by conducting Continuation Writing on half of ground-truth queries. In addition, to further explore the mutual influence between each task, we provide more results in **Appendix A.5** of the revised manuscript.\\n\\n>W5: There are inconsistencies in writing style, such as using both \\\"text-to-SQL\\\" and \\\"Text-to-SQL\\\" interchangeably. Ensuring uniform terminology would improve the clarity and professionalism of the writing.\\n\\n**Ans:** Thank you for your feedback. The typos and the inconsist statements have been revised. We will re-polish our paper carefully.\\n\\n>Q1: The noise correction process assumes access to well-curated data and high-quality schema information, which might not be available for all databases or domains. Without rigorous data preparation, the model may struggle with hallucinations, as noise correction and schema linking effectiveness are diminished when data quality is compromised.\\n\\n**Ans:** In our ROUTE, the basic information used for synthetic schema linking and noise correction is extracted from the databases of two high-quality datasets SPIDER and BIRD, which are widely used in NL2SQL training/evaluation and provide complete architectural information. In addition, we filter the noisy data before synthesizing the data to ensure data quality. \\n\\nIt is undeniable that if the data is noisy and incomplete, the final performance of multi-task collaboration will be affected, as shown in the SPIDER performance of #1 and #3 in Table 4. This encourages the construction of more reliable and complete multi-task datasets in the future to further improve ROUTE.\"}", "{\"summary\": \"The paper introduces ROUTE, a method for enhancing text-to-SQL capabilities in open-source language models. The approach addresses limitations in current methods that rely heavily on closed-source large language models (LLMs), such as GPT-4, for text-to-SQL tasks. ROUTE leverages Multitask Supervised Fine-Tuning (MSFT) and Multitask Collaboration Prompting (MCP) to improve SQL generation performance by incorporating tasks like schema linking, noise correction, and continuation writing. These tasks enable a collaborative prompting approach that reduces hallucinations in SQL generation. Extensive experimentation on multiple benchmarks with open-source LLMs shows that ROUTE significantly improves SQL generation accuracy and outperforms recent methods using fine-tuning and prompting approaches.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"1. The paper incorporates multiple tasks to enhance text-to-SQL capabilities, making the LLM more versatile and capable of handling complex SQL generation scenarios.\\n2. The paper evaluates ROUTE on several well-known benchmarks and compares its performance with other prompting and fine-tuning methods, demonstrating its effectiveness in real-world applications.\", \"weaknesses\": \"1. The authors mention that \\\"Most training-based methods only incorporate the \\u27e8Question, SQL\\u27e9 pairs for SFT, resulting in degraded performance in other tasks, such as schema linking.\\\" However, our approach usually incorporates a \\u27e8Question, Schema, SQL\\u27e9 tuple for SFT. Additionally, a reduction in schema linking performance cannot be seen as a limitation of existing methods. If a specific task is not included in training, optimal results for that task are not expected. Therefore, this should not be considered a limitation; instead, one could state that training with schema linking can achieve better outcomes.\\n2. The authors state that \\\"Training LLMs on a single SFT task poses a significant risk of overfitting, which may diminish the model's capability to understand instructions.\\\" However, overfitting is not further addressed in the subsequent sections. Could the authors clarify what overfitting entails in the context of SQL tasks, and explain how multi-task training specifically mitigates this risk? Additionally, training on more data and achieving good results may also suggest a potential overfitting scenario.\\n3. The authors mention that \\\"This strategy leverages collaboration across several SQL-related tasks to reduce hallucinations during SQL generation.\\\" However, the term \\\"SQL hallucinations\\\" is not defined, nor is there any discussion in the experimental section explaining how hallucinations are reduced. This claimed advantage, therefore, remains unclear.\\n4. If Schema Linking, Noise Correction, and Continuation Writing are considered important, could the authors provide the relative improvement metrics for these tasks?\\n5. There are inconsistencies in writing style, such as using both \\\"text-to-SQL\\\" and \\\"Text-to-SQL\\\" interchangeably. Ensuring uniform terminology would improve the clarity and professionalism of the writing.\", \"questions\": \"1. The noise correction process assumes access to well-curated data and high-quality schema information, which might not be available for all databases or domains. Without rigorous data preparation, the model may struggle with hallucinations, as noise correction and schema linking effectiveness are diminished when data quality is compromised.\\n2. In low-resource settings where high-quality SQL annotations or database schema information might be scarce, could ROUTE be enhanced by incorporating weak supervision, unsupervised learning, or semi-supervised data to fill gaps?\\n3. Given that database schemas often change over time in production, can ROUTE adapt to new tables or columns without needing extensive retraining, or would these require ongoing fine-tuning?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Thanks for the Response\", \"comment\": \"Thanks to the authors for their detailed response and efforts updating the paper. While I am not fully convinced, this paper may distinguish enough from related work as a novel combination of existing methods. Given the strong empirical performance and newly added experiments, I have increased my score accordingly.\"}", "{\"title\": \"Official reply to Reviewer 2daL [2/2]\", \"comment\": \">W3: The performance gap with the close-source LLMs is still large, roughly 13% on BIRD development set, which makes the applicability of this approach limited to the scenarios where privacy and local LLMs is essential.\\n\\n**Ans:** Thank you for your valuable comments. It is undeniable that our solution still has a certain gap with the solution using closed -source LLMs, but the solution using open-source models has unique advantages in privacy and local LLMs. It can be customized according to business data sets. Also, the smaller model reasoning speed is relatively fast and can better meet the needs of multi-step reasoning. In addition, our ROUTE is a general solution that is better than existing SFT methods (e.g., DTS-SQL and SENSE). From our experiments, we can see that it can be used in almost all kinds of open-source LLMs without paying too much cost for a considerable performance improvement. In addition, we can see that our method achieves **87.3** EX score on SPIDER development set which is close to the results of most closed-source methods. This shows that our ROUTE can be a substitute for applications in some closed-source scenarios. We believe ROUTE can contribute to the community of Text2SQL in the future.\\n\\n>Q1: For the open-source LLMs and super large databases such as some of the databases in BIRD benchmark, how these large schema are fitted into the prompt of the open-source models?\\n\\n**Ans:** Thank you for your valuable review. As the prompt template shown in **Appendix A.11**, we extracted the key information about the database to incorporate into the prompt, e.g., the column names and some example rows. In addition, It is worth noting that the context length of popular LLMs, such as LLama3 and Qwen2.5, can effectively meet the requirements of BIRD, with context lengths ranging from 8K to 32K.\\n\\nWhen dealing with a very large database and complex prompts with extensive sequence lengths, we can utilize the 128K version of open-source LLMs supported by YaRN [4]. Besides, we can also introduce special techniques to simplify the input prompts, as demonstrated by CODES [5].\\n\\nHowever, it is undeniable that the current solutions still have limitations when dealing with extremely large databases. This opens up a new avenue for future research on ROUTE. Thank you again for your valuable comments.\\n\\n\\n### Reference \\n>[1] CHASE-SQL: Multi-Path Reasoning and Preference Optimized Candidate Selection in Text-to-SQL, arxiv preprint, 2024.\\\\\\n>[2] The Death of Schema Linking? Text-to-SQL in the Age of Well-Reasoned Language Models, arxiv preprint, 2024.\\\\\\n>[3] Chess: Contextual harnessing for efficient sql synthesis, arxiv preprint, 2024.\\\\\\n>[4] YaRN: Efficient Context Window Extension of Large Language Models, ICLR, 2024.\\\\\\n>[5] Codes: Towards building open-source language models for text-to-sql, ACM SIGMOD 2024.\"}", "{\"comment\": \"Dear Reviwer cCGH,\\n\\nThank you for raising the score and valuable suggestions. \\n\\nWe have supplemented the ablation experimental results of removing each task from MSFT. The results are presented in the following table. \\n\\n| ||SPIDER-TS|BIRD-TS|SPIDER-SL|SPIDER-SL|BIRD-SL|BIRD-SL|SPIDER-NC|SPIDER-NC|SPIDER-CW|SPIDER-CW |\\n|:-:|:-|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|\\n| No.|Settings|Dev-EX|Dev-EX|Table-R/P|Column-R/P|Table-R/P|Column-R/P|Dev-EX|Dev-EX|Dev-EX|Dev-EX |\\n| #1|Llama3 with MSFT|83.6 |53.6 |97.38/95.71|98.59/96.98|90.87/90.22|96.13/90.89|83.4 |53.4 |91.1 |73.9|\\n| #2|MSFT w/o TS|0.1 |16.2 |96.58/93.94|98.40/96.32|90.79/88.26|95.95/90.34|77.4 |45.5 |86.5 |69.6|\\n| #3|MSFT w/o SL|81.8 |50.9 |--|--|--|--|76.3 |47.4 |91.3 |73.5|\\n| #4|MSFT w/o NC|82.8 |51.0 |96.52/94.25|99.00/96.59|90.41/88.85|96.09/90.75|--|--|91.7 |73.4|\\n| #5|MSFT w/o CW|81.2 |50.3 |96.51/93.97|98.65/96.39|90.59/88.12|96.05/90.64 |79.4 |49.0 |81.2 |56.7|\\n| #6|SFT with TS|83.1 |52.9 |--|--|--|--|--|--|85.6 |69.2|\\n| #7|SFT with SL|--|--|95.55/92.69|98.91/95.29|87.84/85.11|94.93/89.51 |--|--|--|-- |\\n| #8|SFT with NC|0.1 |8.7 |--|--|--|--|78.9 |49.3 |48.6 |38.6|\\n| #9|SFT with CW|68.1 |39.0 |--|--|--|--|--|--|89.8 |70.1|\\n| #10|Llama3 w/o SFT|69.3 |32.1 |88.35/76.37|94.83/91.46|83.77/75.38|89.55/86.39|72.1 |38.1 |80.3 |57.6|\\n\\n\\nThe results suggest that tasks not included in MSFT demonstrate lower performance due to the overfitting of LLMs to other tasks. This highlights the importance of considering SFT across multiple tasks to prevent performance degradation of an LLM when handling additional tasks.\\n\\nFurthermore, the results indicate that although the performance of the full MSFT on certain tasks, such as SL and CW, is somewhat inferior to that of single or triple-task SFT, the full MSFT demonstrates significant performance improvements across each task and exhibits better stability.\\n\\nWhile we can not upload the revised manuscript at this time, we will update the ablation results and corresponding discussion in the next version following the rebuttal stage.\\n\\nThank you again for your valuable feedback and the positive score.\\n\\nBest,\\nAuthors\"}", "{\"comment\": \"Dear Reviewer 2daL,\\n\\nThank you for your timely feedback and constructive suggestions.\\n\\nOur ROUTE focuses on the supervised fine-tuning paradigm for SQL generation using small-sized LLMs. This inevitably leads to bias and an unfair comparison with recent prompting-based methods that utilize GPT-4/4o (e.g., CHESS and CHASE). However, in our experiments, our baselines basically include recent advanced SFT-based methods such as CODES, SENSE, and DTS-SQL, to verify the effectiveness and superiority of ROUTE. We have also verified the generalization of our approach on multiple open-source LLMs and benchmarks.\\n\\nIn addition, due to the space limitation, we have provided hints about the Recall and Precession results in the revised manuscript: **lines 443~444** (Section 4.3-Study on Enhanced Schema Linking) to draw readers\\u2019 attention to these insightful results.\\n\\nWe thank you again for your timely feedback and valuable comments. Your suggestions have been of great help in improving the quality of our manuscript. If you have any further questions or suggestions, feel free to ask.\\n\\nBest,\\nAuthors\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}", "{\"comment\": \"Thank you for further addressing my questions.\"}", "{\"title\": \"Official reply to Reviewer TcaB's remaining concerns [2/2]\", \"comment\": \"> **Multi-tasking Collaboration**: To exploit the potential of multi-task capabilities, we propose a multi-task collaborative prompting strategy (MCP) to improve the final SQL generation. The most similar works are DIN-SQL[1] and MAC-SQL[2], which both aim to reduce the complexity of the Text2SQL and improve the final performance via self-correction.\\n>- **First**, compared to them, our MCP efficiently integrates multiple tasks using concise prompts across all tasks, which makes it more effective in small-sized LLMs that struggle with comprehending complex instructions. As shown in the results of **Section 4.1-Table 1**, the effectiveness of MAC-SQL and DIN-SQL is constrained by the limited capacity of small-sized LLMs to comprehend complex instructions, while our MCP can achieve better and impressive performance.\\n>- **Besides**, while all three methods employ a self-correction strategy to enhance the quality of generated SQL queries, our MCP introduces a novel continuation writing task specifically designed to refine challenging SQL queries and improve the performance significantly.\\n>\\n>**Table 1 (Partially) in Section 4.2**: Performance comparison on SPIDER and BIRD benchmarks.\\n>||SPIDER|SPIDER|SPIDER|BIRD|BIRD|\\n>|-|-|-|-|-|-|\\n>|Methods/LLMs|Dev-EX|Dev-TS|Test-EX|Dev-EX|Dev-VES|\\n>|Llama3-8B|69.3|58.4|69.1|32.1|31.6|\\n>|DIN-SQL + Llama3-8B|48.7|39.3|47.4|20.4|24.6|\\n>|MAC-SQL + Llama3-8B|64.3|52.8|65.2|40.7|40.8|\\n>|**MCP + Llama3-8B**|**75.0**|**63.4**|**72.0**|**42.7**|**44.8**|\\n>|Qwen2.5-7B|72.5|64.0|75.9|41.1|42.0|\\n>|DIN-SQL + Qwen2.5-7B|72.1|61.2|71.1|30.1|32.4|\\n>|MAC-SQL + Qwen2.5-7B|71.7|61.9|72.9|46.7|49.8|\\n>|**MCP + Qwen2.5-7B**|**78.3**|**67.2**|**78.7**|**49.7**|**52.8**|\\n\\nConsidering the comprehensive nature of our work, which encompasses data synthesis, supervised fine-tuning, and multi-task collaborative prompting, it is inevitable that there are some similarities with existing work. Nevertheless, we have offered numerous insights into the Text2SQL task and achieved promising results, which we believe are significant contributions to the Text2SQL community.\\n\\nWe thank you again for your timely feedback and valuable comments. We have added the above discussion in **Appendix A.8** to further clarify our innovations and contributions.\\n\\n### Reference \\n> [1] Din-sql: Decomposed in-context learning of text-to-sql with self-correction, NeureIPS, 2023.\\\\\\n> [2] MAC-SQL: A Multi-Agent Collaborative Framework for Text-to-SQL, arxiv preprint, 2024.\\\\\\n> [3] Synthesizing text-tosql data from weak and strong llms, ACL 2024.\"}", "{\"title\": \"Official reply to Reviewer TcaB [1/3]\", \"comment\": \"We thank Reviewer TcaB for the insightful comments and constructive suggestions that contribute to improving our paper. Below, we address your concerns one by one.\\n\\n>W1-I The authors mention that \\\"Most training-based methods only incorporate the \\u27e8Question, SQL\\u27e9 pairs for SFT, resulting in degraded performance in other tasks, such as schema linking.\\\" However, our approach usually incorporates a \\u27e8Question, Schema, SQL\\u27e9 tuple for SFT.\\n\\n**Ans:** Thank you for your kindly reminder. We apologize for any confusion caused by our misleading expression of \\u27e8Question, SQL\\u27e9. What we aim to say here is the \\u27e8Question, Schema, SQL\\u27e9 tuple, as illustrated by our method's notions (**Section 3.1**) and the prompt template (**Appendix A.11**). We have revised this sentence in the revised submission. Thank you again for your feedback. \\n\\n>W1-II: Additionally, a reduction in schema linking performance cannot be seen as a limitation of existing methods. If a specific task is not included in training, optimal results for that task are not expected. Therefore, this should not be considered a limitation; instead, one could state that training with schema linking can achieve better outcomes.\\n\\n**Ans:** Thank you for your thoughtful comments.\\n\\nSchema linking is a critical step in the Text2SQL task [1,2], which demonstrates significant performance gains in SQL generation. Training exclusively on SQL generation tasks with SFT can severely impair the model's other capabilities (**Appendix A.6**), such as schema linking.\\n \\nTherefore, to maintain proficiency across various tasks for effective multi-task collaboration, it is essential to consider multi-task supervised fine-tuning.\\n\\nConsidering the fact that SFT training in a single SQL generation task can significantly reduce its understanding of other instructions, thus making it unable to perform tasks like schema linking, we believe this is a potential drawback of the traditional SFT-based methods.\\n\\nTo avoid potential misunderstanding, We have revsied this sentence as:\\n**However, most training-based methods only incorporate the SQL generation task in the SFT stage, resulting in a degraded performance in other tasks that are important for Text2SQL capability, such as schema linking**\\n\\n>W2: The authors state that \\\"Training LLMs on a single SFT task poses a significant risk of overfitting, which may diminish the model's capability to understand instructions.\\\" However, overfitting is not further addressed in the subsequent sections. Could the authors clarify what overfitting entails in the context of SQL tasks, and explain how multi-task training specifically mitigates this risk? Additionally, training on more data and achieving good results may also suggest a potential overfitting scenario. \\n\\n**Ans:** Thanks for your feedback. It has been widely observed that SFT training in a single task significantly reduces the model's instruction following ability on other tasks [3]. Overfitting on SQL generation task means that a fine-tuned LLM will weaken its ability to understand other instructions and perform poorly on other important SQL-related tasks (such as schema linking and noise correction).\\n\\nTherefore, in this paper, we explore multitask tuning across various SQL-related tasks to preserve the other essential capabilities of LLMs\\u3010see our submission in lines 63\\uff5e69 for detailed description\\u3011We report the performance of single-task SFT on each task in the following table, where '--' means that LLMs cannot obtain the output in the expected format due to overfitting. The experimental results show that, compared to multi-task training, the models fine-tuned solely on a single task perform poorly on other tasks, which is caused by overfitting to the single task.\\n\\n**Table 15 in Appendix A.6:** The SFT impact of all tasks on each other.\\n\\n|| |SPIDER-TS|BIRD-TS|SPIDER-SL|SPIDER-SL|BIRD-SL|BIRD-SL|SPIDER-NC|BIRD-NC|SPIDER-CW|BIRD-CW |\\n|-|:-|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|\\n|No.| Settings|Dev-EX|Dev-EX|Table-R/P|Column-R/P|Table-R/P|Column-R/P|Dev-EX|Dev-EX|Dev-EX|Dev-EX |\\n| #1|Llama3 with MSFT|83.6 |53.6 |97.38/95.71|98.59/96.98|90.87/90.22|96.13/90.89|83.4 |53.4 |91.1 |73.9|\\n| #2|SFT with TS|83.1 |52.9 |--|--|--|--|--|--|85.6 |69.2|\\n| #3|SFT with SL|--|--|95.55/92.69|98.91/95.29|87.84/85.11|94.93/89.51 |--|--|--|-- |\\n| #4|SFT with NC|0.1 |8.7 |--|--|--|--|78.9 |49.3 |48.6 |38.6|\\n| #5|SFT with CW|68.1 |39.0 |--|--|--|--|--|--|89.8 |70.1|\\n| #6|Llama3 w/o SFT|69.3 |32.1 |88.35/76.37|94.83/91.46|83.77/75.38|89.55/86.39|72.1 |38.1 |80.3 |57.6|\\n\\nAccording to your feedback, we have revised this sentence to minimize any confusion caused by the term \\\"overfitting\\\", as follows:\\\\\\n**Training LLMs on a single SQL generation task poses a substantial risk of diminishing performance in understanding instructions, potentially reducing the model's effectiveness in other important SQL-related tasks beyond SQL generation.**\"}", "{\"metareview\": [\"**Strengths:**\", \"The proposed method \\u201csignificantly improves the performance of open-source LLMs and outperforms all existing methods trained on open-source LLMs\\u201d (2daL, also noted by cCGH), and \\u201cachieves comparable performance to GPT-4\\u201d (Ej7x).\", \"Results are strong across multiple well-known benchmarks, including Spider, Bird, perturbed variants of Spider, as well as Dr. Spider, \\u201cindicating the robustness of ROUTE\\u201d (Ej7x), while \\u201cdemonstrating its effectiveness in real-world applications\\u201d (TcaB)\", \"Evaluation is comprehensive \\u201cusing multiple LLMs as base models\\u201d (cCGH) and \\u201can extensive list of baselines\\u201d (Ej7x).\", \"The proposed method works in both fine-tuning and prompting regime (2daL), and makes \\u201cLLM(s) more versatile and capable of handling complex SQL generation scenarios.\\u201d (TcaB)\", \"\\u201cThe paper is easy-to-follow, and the writing is mostly clear\\u201d (Ej7x), with some clarity issues on difference with existing methods and definition of certain terms (see weaknesses).\", \"**Weaknesses:**\"], \"the_following_weaknesses_are_addressed_during_the_rebuttal_phase\": [\"Comparison with more recent prompting-based approaches (2daL, e.g., CHASE-SQL, Distillery, and CHESS)\", \"More careful ablation on different components of the pipeline to better understand their individual contributions to final performance (2daL, cCGH, Ej7x), while reporting quantitative evaluation metrics on the performance of individual components (e.g., schema linking, 2daL, TcaB).\", \"As noted by TcaB, there are a few places in the submission where the technical presentation is not clear, such as how the proposed method addresses overfitting and hallucinations in generated SQL queries.\"], \"the_single_remaining_issue_yet_to_be_addressed_after_the_rebuttal_phase_is_the_novelty_of_the_method\": \"as noted by TcaB and Ej7x, the proposed method is a combination of many existing approaches in text-to-SQL. Therefore, \\u201cit is not very clear what kind of novel contribution this paper is making\\u201d (Ej7x). \\u201cAlthough the authors propose many interesting points, none of them is particularly new in this area\\u201d (TcaB).\\n\\nWhile this submission may not present a truly novel idea, I believe this submission makes valuable contributions through strong empirical results. The fact that the paper focuses on improving the performance of open-source models via both prompting and fine-tuning is also a plus for industry practitioners to deploy similar approaches in practical scenarios. Therefore, given the rating and the strong empirical results, the recommendation is **Accept**. \\n\\n**Further suggestions on revision:**\\n\\nBesides existing feedback from reviewers, please focus more carefully on explaining the relations with many existing works flagged by TcaB and Ej7x. I also agree with Ej7x that some components in the framework are \\u201crebranded under new terms\\u201d. Please avoid inventing new terms since this would cause confusion. Instead, please use well-known technical terms to refer to certain parts of your model. For example, please consider renaming \\u201cnoise correction\\u201d to \\u201cself-debugging\\u201d or \\u201cprogram repair\\u201d.\", \"additional_comments_on_reviewer_discussion\": \"Please see the above meta review.\"}", "{\"summary\": \"The paper proposes ROUTE, a method that involves (1) multi-task training and (2) multi-stage prompting to improve LLMs\\u2019 performance on text-to-SQL parsing. Compared to an extensive list of baselines, the proposed ROUTE method demonstrates better parsing accuracy on two established datasets, Spider and BIRD, and three other perturbed variants of Spider.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. The performance of ROUTE is strong. On two well-established text-to-SQL datasets, Spider and Bird, ROUTE effectively improves the performance of two latest open-weight LLMs, Llama 3.1 and Qwen 2.5, and achieves comparable performance to GPT-4. The performance improvement also holds on three perturbed Spider variants, indicating the robustness of ROUTE.\\n2. The experiments to evaluate ROUTE are comprehensive. The authors gathered an extensive list of baselines and compared them with ROUTE (or the multi-stage prompting step in ROUTE). Additional experiments and ablation studies of the method further supports some design choices of ROUTE and demonstrates its stable performance improvement across different models and datasets.\\n3. The paper is easy-to-follow, and the writing is mostly clear.\", \"weaknesses\": \"1. It is not very clear what kind of novel contribution this paper is making. The tasks themselves for multi-task training and multi-stage prompting have all been studied in related work, and some of them are rebranded under new terms. To name some examples, \\u201cnoise correction\\u201d is essentially training and prompting LLMs to self-debug [1][2], and the \\u201ccontinuation writing\\u201d is simply a subset of text-to-SQL generation by the autoregressive nature of LLMs and has been one of the prompting paradigm for text-to-SQL parsing with LLMs [3][4]. At the framework level, there are also existing papers compiling different tasks to improve LMs\\u2019 text-to-SQL performance via multi-task training [5]. Thus, it is opaque how the proposed method combines these existing ideas in a novel way.\\n\\n2. The noisy correspondence filtering step to pre-process the training data is not fully elaborated, and the contribution of this step to ROUTE is minimal according to the ablation study (#1 vs #3 and #6 vs #8 in Table 3). Training details and quality of the noise filtering model is not discussed, e.g. through an intrinsic evaluation of how accurately it can discriminate noisy examples. The difference between ROUTE\\u2019s data synthesis procedure and that of SENSE is not clear. The method to \\u201cartificially and randomly introduce errors\\u201d (lines 225-230) is also not documented. Overall, this part of the method is not clearly explained, and its contribution in ROUTE is not obvious.\\n\\n[1] https://arxiv.org/abs/2304.05128 \\n\\n[2] https://arxiv.org/abs/2312.11242\\n\\n[3] https://arxiv.org/abs/2204.00498 \\n\\n[4] https://arxiv.org/abs/2303.13547 \\n\\n[5] https://arxiv.org/abs/2212.09278\", \"questions\": [\"1. What is the \\u201cpseudo-SQL\\u201d used to perform schema linking? How is it implemented? This term only appears twice in the paper without any further elaboration.\", \"2. The use of \\u201challucination\\u201d may not be appropriate here in the context of text-to-SQL parsing. Are the authors simply trying to say incorrect column matching and entity linking?\", \"3. The manuscript would benefit from another round of proof-read to correct typos and standardize term usage, including those mentioned above and some other examples as follows:\", \"\\u201cin-contextual learning\\u201d -> \\u201cin-context learning\\u201d (line 39)\", \"\\u201cpromoting-based methods\\u201d -> \\u201cprompting-based methods\\u201d (lines 200, 373)\", \"\\u201cshema linking\\u201d -> \\u201cschema linking\\u201d (line 228)\", \"\\u201cSQLer$(d_i, \\\\tilde{s}^*)$\\u201d -> \\u201cSQLer$(d_i, s^*)$\\u201d (line 283)\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Dear Reviewer Ej7x,\\n\\nThank you for raising the score! We have provided further clarification and discussion on the innovations, contributions, and differences between our ROUTE and existing methods in **Appendix A.8** of our revised submission.\\n\\nBest,\\nAuthors\"}", "{\"title\": \"Official reply to Reviewer Ej7x [1/2]\", \"comment\": \"We thank Reviewer Ej7x for the positive recognition of our ROUTE performance and the insightful comments that contribute to improving our paper. Below, we address your concerns one by one.\\n\\n>W1: It is not very clear what kind of novel contribution this paper is making. The tasks themselves for multi-task training and multi-stage prompting have all been studied in related work, and some of them are rebranded under new terms. To name some examples, \\u201cnoise correction\\u201d is essentially training and prompting LLMs to self-debug [1][2], and the \\u201ccontinuation writing\\u201d is simply a subset of text-to-SQL generation by the autoregressive nature of LLMs and has been one of the prompting paradigm for text-to-SQL parsing with LLMs [3][4]. At the framework level, there are also existing papers compiling different tasks to improve LMs\\u2019 text-to-SQL performance via multi-task training [5]. Thus, it is opaque how the proposed method combines these existing ideas in a novel way.\\n\\n**Ans:** Thank you for your detailed review. We acknowledge that the proposed ROUTE shares similarities with several recent works [1,2,3,4,5]. However, it offers valuable insights and makes significant contributions to the field of Text2SQL, which can be summarized as follows:\\n\\n- ROUTE is among the pioneering frameworks under the context of LLMs that explores multi-task tuning and collaborative prompting to improve Text2SQL performance.\\n- We have exhaustively introduced and defined three important tasks in SQL generation, demonstarting that multi-task tuning and collaborative prompting in Schema Linking, Noise Correction and Continuation Writing significantly improve SQL generation accuracy. The additionally introduced SQL-related tasks are well integrated during both the training and inference phases.\\n- We have achieved state-of-the-art performance in 7B/14B-sized LLMs on both the widely-recognized SPIDER and BIRD benchmarks, with verified generalization and transferability across various cross-domains benchmarks and LLMs.\", \"we_sumarize_the_differences_and_relationships_between_route_and_your_mentioned_works_are_as_follows\": \"\\\\\\n[1] proposed an effective Self-Debugging Strategy to teach existing LLMs to debug the predicted program via few-shot demonstrations, including SQL, Python, C++, etc. However, our method is self-debugging for SQL generation, and it is under the context of supervised fine-tuning. (**lines 101~102**)\\\\\\n[2] proposed a multi-agent collaboration framework and used the closed-source LLMs GPT-4 as a basis to enhance text-to-SQL parsing. However, we found that its migration to some small-sized models (Table 1) has limited effectiveness. We proposed MCP to achieve promising performance by combining multi-task collaboration with some simple instructions. (**lines 104~105**)\\\\\\n[3] performed an empirical evaluation based on Codex. However, it does not explicitly enhance the relevant capabilities by SFT, but only uses its autoregressive characteristics to achieve performance improvement with few-shot examples. Instead, we explored the solution that explicitly enhance multi-task capabilities to improve the accuracy of SQL generation. (**line 93**)\\\\\\n[4] is a pioneering work to evaluate the performance of ChatGPT on Text2SQL, showing that it has great potential for SQL generation. However, our paper focuses on open-source LLMs under the context of supervised fine-tuning. (**line 98**)\\\\\\n[5] proposed multiple subtasks and combined generative pretrained language models to improve Text2SQL. However, its focused subtasks do not completely overlap with ours. Moreover, the effectiveness of this paradigm in the context of LLMs is unknown. (**line 91**)\\n\\n>W2-I The noisy correspondence filtering step to pre-process the training data is not fully elaborated, and the contribution of this step to ROUTE is minimal according to the ablation study (#1 vs #3 and #6 vs #8 in Table 3).\\n\\n**Ans:** Thank you for your valuable comment. For SFT, the impact of noisy correspondence filtering (#6 vs #8) is not obvious, which indicates that noisy data accounts for a relatively low proportion in the Spider and Bird datasets. After introducing multi-task data synthesis, we observed that the noisy correspondence filtering step significantly boosts performance (#1 vs #3) on SPIDER. This indicates that the noise is substantial, and the accumulation of noise from multi-task data synthesis can adversely impact the model's comprehension of basic (simple) modes.\"}", "{\"title\": \"Thanks for your responses\", \"comment\": \"Thank you to the authors for their responses and efforts. However, my remaining concern is the novelty of applying multi-task learning in the text-to-SQL field (Although the authors propose many interesting points, none of them is particularly new in this area.). Therefore, I will temporarily maintain my score.\"}", "{\"summary\": \"The paper introduces \\u201cROUTE,\\u201d a novel approach to (1) finetune open-source large language models (LLMs) for Text-to-SQL through multi-task supervised fine-tuning (MSFT) and (2) leverage multitask collaboration prompting (MCP) for SQL generation during inference. The MSFT tasks include Text-to-SQL, Schema Linking, Noise Correction, and Continuation Writing. The proposed method aims to reduce hallucinations and enhance Text-to-SQL robustness, demonstrated by improved performance on two benchmarks: Spider and BIRD.\", \"soundness\": \"4\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. The paper introduces a multitask learning approach that leverages several text-to-SQL related tasks. Noise Correction is designed to assess whether the execution result of a SQL query correctly answers the question, reducing hallucinations when paired with multi-turn generation.\\n2. ROUTE demonstrates competitive accuracy, outperforming some closed-source methods on benchmarks, thus showcasing the effectiveness of multitask training over single-task fine-tuning.\\n3. The authors provide comprehensive experiments using multiple LLMs as base models, demonstrating that ROUTE is generalizable across various LLMs.\", \"weaknesses\": \"1. The paper lacks an ablation study on the contribution of each task in MSFT. For instance, the loss from continuation writing is likely already included in text-to-SQL learning after the first token of the SQL prediction. It is unclear how each task directly benefits SQL generation and other inference components.\\n2. Although Noise Correction helps improve performance, it relies on execution results within the model, which may be difficult to apply to queries with large outputs, such as selecting an entire column.\\n3. While ROUTE demonstrates strong performance on Spider variants compared to baselines, it remains unclear whether these gains are due to improved robustness or general text-to-SQL performance. It would also be valuable to understand how each component contributes to robustness specifically. Dr. Spider [1] is a more comprehensive perturbation dataset with relative robustness evaluation, which could be useful for evaluating ROUTE\\u2019s improvement more clearly.\\n[1] https://arxiv.org/pdf/2301.08881\", \"questions\": \"For Noise Correction, is it able to handle a large table as the execution result?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Official reply to Reviewer cCGH [2/2]\", \"comment\": \">W3: While ROUTE demonstrates strong performance on Spider variants compared to baselines, it remains unclear whether these gains are due to improved robustness or general text-to-SQL performance. It would also be valuable to understand how each component contributes to robustness specifically. Dr.Spider [1] is a more comprehensive perturbation dataset with relative robustness evaluation, which could be useful for evaluating ROUTE\\u2019s improvement more clearly. [https://arxiv.org/pdf/2301.08881](https://arxiv.org/pdf/2301.08881)\\n\\n**Ans:** Thank you for your constructive suggestions. As suggested, according to the suggestion, we evaluated on Dr.Spider to have a clearer and more comprehensive understanding of the advantages of our ROUTE. Dr.Spider includes 17 perturbation variants that can comprehensively measure the effectiveness and robustness. The specific experimental results are shown in the following tables, including the study on MSFT and on each component. From the results, its conclusions are consistent with those those on SPIDER and BIRD and each component brought performance improvement, which shows that each task has an indispensable contribution to the performance. This further verifies the advantages and robustness of ROUTE. Thank you again for your constructive suggestions. We have added the corresponding experimental results in **Appendix A.5** of the revised manuscript.\\n\\n**Table 13 in Appendix A.5:** The performance on Dr.Spider benchmark.\\n\\n| | Avg.DB | Avg.NLQ | Avg.SQL | Avg.all |\\n|---|---|---|---|---|\\n| Methods | Pre~Post | Pre~Post | Pre~Post | Pre~Post |\\n| Llama3 | 70.1~54.3 | 70.6~56.8 | 69.1~65.5 | 69.9~58.8 |\\n| Llama3 + MCP | 75.6~59.1 | 77.0~61.7 | 74.8~72.6 | 75.8~64.4 |\\n| Llama3 + SFT | 83.4~66.0 | 83.0~72.8 | 79.9~77.6 | 82.1~72.2 |\\n| Llama3 + SFT + MCP | 85.4~68.0 | 85.3~75.1 | 84.3~81.7 | 85.0~74.9 |\\n| Llama3 + MSFT | 83.8~66.3 | 82.9~72.5 | 80.0~77.5 | 82.2~72.1 |\\n| **Llama3 + ROUTE** | **86.7~69.6** | **85.4~75.8** | **84.5~81.9** | **85.5~75.8** |\\n\\n**Table 14 in Appendix A.5:** The ablation results (EX) on Dr.Spider.\\n\\n| | | | | Avg.DB | Avg.NLQ | Avg.SQL | Avg.all |\\n|---|---|---|---|---|---|---|---|\\n| No. | SL | NC | CW | Pre~Post | Pre~Post | Pre~Post | Pre~Post |\\n| #1 | \\u2713 | \\u2713 | \\u2713 | **86.7~69.6** | **85.4~75.8** | **84.5~81.9** | **85.5~75.8** |\\n| #2 | \\u2713 | | | 86.3~67.9 | 84.6~75.4 | 84.4~82.1 | 85.1~75.1 |\\n| #3 | | \\u2713 | | 84.4~67.7 | 83.7~73.7 | 80.3~78.7 | 82.8~73.3 |\\n| #4 | | | \\u2713 | 84.0~66.6 | 83.0~73.0 | 80.1~78.0 | 82.4~72.6 |\\n| #5 | | | | 83.8~66.6 | 82.9~72.5 | 80.0~77.5 | 82.2~72.1 |\\n\\n\\n>Q: For Noise Correction, is it able to handle a large table as the execution result?\\n\\n**Ans:** Thank you for your valuable question. We have to cliam thatThe input of our noise correction is <Schema, Question, SQL, Execution Information>, where the execution information only refers to the exception information of the SQL executor. If the execution passes, there is no information. Therefore, our noise correction is able to handle SQL involving the large tables instead of using the results of the entire table as a prompt. To prevent readers from being confused, we have clarified it again in **lines 286~288** of revised version. Thank you again for your valuable comments.\\n\\n### Reference \\n>[1] Dr.Spider: A Diagnostic Evaluation Benchmark towards Text-to-SQL Robustness, ICLR, 2023.\"}", "{\"comment\": \"Thank you to the authors for providing additional experiments. My concerns have been mostly addressed, and I have increased my scores to reflect this. If time allows, I would suggest conducting an ablation study by removing one task from the MSFT instead of using only one task in SFT for rows #2, #3, #4, and #5 in Table 15. I believe it is useful to have this experiment added in either the rebuttal or a later revision of the paper.\"}", "{\"title\": \"Official reply to Reviewer cCGH [1/2]\", \"comment\": \"We thank Reviewer cCGH for the positive feedback and constructive suggestions that contribute to improving this paper. Below, we address your comments point by point.\\n\\n>W1: The paper lacks an ablation study on the contribution of each task in MSFT. For instance, the loss from continuation writing is likely already included in text-to-SQL learning after the first token of the SQL prediction. It is unclear how each task directly benefits SQL generation and other inference components.\\n\\n**Ans:** Thank you for your constructive suggestions. We conducted additional experiments to explore the impact of single-task SFT and MSFT on each task. \\nFor Text-to-SQL (TS), we report zero-shot EX results on the SPIDER and BIRD development sets. For Schema Linking (SL), we report the Recall/Precession scores of predicted related tables and columns. For Noise Correction (NC), we reports the EX scores of SQL queries refined with Noise Correction on the output SQLs of Llama3. For Continuation Writing (CW), we reports the EX scores of all SQL queries obtained by continuation writing on half of ground-truth SQL queries. The specific experimental results are shown in following table, where '--' means that LLMs cannot obtain the output in the expected format due to overfitting.\\n\\n**Table 15 in Appendix A.6:** The SFT impact of all tasks on each other.\\n\\n|| |SPIDER-TS|BIRD-TS|SPIDER-SL|SPIDER-SL|BIRD-SL|BIRD-SL|SPIDER-NC|BIRD-NC|SPIDER-CW|BIRD-CW |\\n|-|:-|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|\\n|No.| Settings|Dev-EX|Dev-EX|Table-R/P|Column-R/P|Table-R/P|Column-R/P|Dev-EX|Dev-EX|Dev-EX|Dev-EX |\\n| #1|Llama3 with MSFT|83.6 |53.6 |97.38/95.71|98.59/96.98|90.87/90.22|96.13/90.89|83.4 |53.4 |91.1 |73.9|\\n| #2|SFT with TS|83.1 |52.9 |--|--|--|--|--|--|85.6 |69.2|\\n| #3|SFT with SL|--|--|95.55/92.69|98.91/95.29|87.84/85.11|94.93/89.51 |--|--|--|-- |\\n| #4|SFT with NC|0.1 |8.7 |--|--|--|--|78.9 |49.3 |48.6 |38.6|\\n| #5|SFT with CW|68.1 |39.0 |--|--|--|--|--|--|89.8 |70.1|\\n| #6|Llama3 w/o SFT|69.3 |32.1 |88.35/76.37|94.83/91.46|83.77/75.38|89.55/86.39|72.1 |38.1 |80.3 |57.6|\\n\\nFrom the results, we can see that although single-task SFT can improve the ability to process the corresponding task, it can easily cause overfitting on instructions and lead to the degradation of the ability to process other tasks (e.g., **No.#2,#3,#4,#5**), which is not conducive to multi-task collaboration. On the contrary, our MSFT (**No.#1**) can boost the ability of each task while reducing the risk of overfitting, thus improving the feasibility of multi-tasking collaboration.\\n\\n\\n>W2: Although Noise Correction helps improve performance, it relies on execution results within the model, which may be difficult to apply to queries with large outputs, such as selecting an entire column.\\n\\n**Ans:** Thank you for your careful review. We have to clarify that our noise correction does not need to put large outputs into Promt, but the status results of SQL query execution, such as error exception information. To this end, our noise correction is able to handle such SQL queries with large output. To prevent readers from being confused, we have clarified it again in **lines 286~288** of revised version. Thank you again for your careful comments.\"}", "{\"title\": \"Official reply to Reviewer Ej7x [2/2]\", \"comment\": \">W2-II **(1)** Training details and quality of the noise filtering model is not discussed, e.g. through an intrinsic evaluation of how accurately it can discriminate noisy examples. **(2)** The difference between ROUTE\\u2019s data synthesis procedure and that of SENSE is not clear. **(3)** The method to \\u201cartificially and randomly introduce errors\\u201d (lines 225-230) is also not documented. **(4)** Overall, this part of the method is not clearly explained, and its contribution in ROUTE is not obvious.\\n\\n**Ans:** Thank you for your valuable comments, we will reply to this weakness point-by-point.\\\\\\n**(1)** According to the suggestion, we have added an experiment to assess the accuracy of noise sample identification. We construct an evaluation set from SPIDER and BIRD development sets to evaluate their ability to identify positive examples and negative examples, where the negative examples come from challenging artificially synthesized SQL queries and the incorrect SQL queries obtained by SQL generation with Llama3-8B. \\n\\nAs shown in following table, we can observe that the LLM\\u2019s ability to identify negative (noisy) and positive examples has been significantly improved after SFT. Further, for more challenging dataset like Bird, there still remains room for improvement in distinguishing noisy samples.\\n\\n||w/o SFT|w/o SFT|w/o SFT|with SFT|with SFT|with SFT |\\n|-|-|-|-|:-:|:-:|:-:|\\n||Spider|Bird|All|Spider|Bird|All|\\n| Positive (Ground-truth, 100)|0.51|0.23|0.37|0.96|0.83|0.90|\\n| Negative (Artificial, 100)|0.64|0.50|0.57|0.68|0.77|0.73|\\n| Negative (Llama3, 100)|0.63|0.83|0.73|0.96|0.94|0.95|\\n\\n**(2)** Compared to SENSE, the data synthesis pipeline of ROUTE encompasses not only Text2SQL but also multiple other SQL-related tasks. Our approach focuses on utilizing existing data to synthesize multi-task SFT data, thereby enhancing the capabilities of open-source language models to handle various SQL-related tasks. In contrast, SENSE mainly focused on SQL generation task, leveraging powerful LLMs to increase the diversity of SQL generation training set and also synthesize preference data. We have clarified the relationship and difference between ROUTE and SENSE in our revised submission (**Appendix A.8**).\\n\\n**(3)** Thank you for your reminder, we have clarified it in **Appendix A.7** of our revised manuscript.\\n\\n**(4)** Please see our Answer to **W1**.\\n\\n>Q1: What is the \\u201cpseudo-SQL\\u201d used to perform schema linking? How is it implemented? This term only appears twice in the paper without any further elaboration.\\n\\n**Ans:** Thanks for your question. The pseudo SQL refers to the generated intermediate SQL using the defined template and the complete schema, i.e., $\\\\mathcal{M}(\\\\sigma_t(d_i, q_i),d_i)$, which we have clarified in the revised submission (please see **lines 277~284** for the details)\\n\\n>Q2: The use of \\\"hallucination\\\" may not be appropriate here in the context of text-to-SQL parsing. Are the authors simply trying to say incorrect column matching and entity linking?\\n\\n**Ans:** Thanks for your feedback. We follow several recent works to utilize hallucinations in SQL generarion task [6,7]. Hallucinations in LLMs refer to cases where LLM generate plausible but factually incorrect or nonsensical information[8]. Hallucination in Text2SQL task indicates incorrect SQL generations. Schema hallucinations and logic hallucinations are widely observed in LLM-based SQL generation [7]. \\n\\nConsidering that the definition of hallucinations might be unclear in the context of Text2SQL, we have revised the usage of SQL hallucinations in the revised submission. Please see our revised submission for details.\\n\\n>Q3: The manuscript would benefit from another round of proof-read to correct typos and standardize term usage.\\n\\n**Ans:** Thanks for your suggestion. The typos and the standardize term usages have been double checked. We will re-polish our paper carefully.\\n\\n### Reference\\n[1] Teaching Large Language Models to Self-Debug, ICLR, 2024.\\\\\\n[2] MAC-SQL: A Multi-Agent Collaborative Framework for Text-to-SQL, arxiv preprint, 2024.\\\\\\n[3] Evaluating the Text-to-SQL Capabilities of Large Language Models, arxiv preprint, 2022.\\\\\\n[4] A comprehensive evaluation of ChatGPT's zero-shot Text-to-SQL capability, arxiv preprint, 2023.\\\\\\n[5] MIGA: A Unified Multi-task Generation Framework for Conversational Text-to-SQL, AAAI, 2023.\\\\\\n[6] Before Generation, Align it! A Novel and Effective Strategy for Mitigating Hallucinations in Text-to-SQL Generation. ACL 2024.\\\\\\n[7] PURPLE: Making a Large Language Model a Better SQL Writer. ICDE 2024.\\\\\\n[8] A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions. ACM TOIS.\"}", "{\"title\": \"Official reply to Reviewer TcaB [3/3]\", \"comment\": \">Q2: In low-resource settings where high-quality SQL annotations or database schema information might be scarce, could ROUTE be enhanced by incorporating weak supervision, unsupervised learning, or semi-supervised data to fill gaps?\\n\\n**Ans:** Thank you for your valuable question. we think your view is valuable and feasible. \\n\\nThere are now a large number of high quality, public Text-to-SQL datasets, either LLM-synthesized or human annotated [7,8]. We have collected a SFT dataset comprising approximately 1 million entries for Text2SQL task. Despite the large scale, there remains a prevalence of low-quality data, and scenarios involving complex queries are relatively underrepresented. Therefroe, it is valuable to explore weak supervision, unsupervised learning or semi-supervised paradigm to enrich and refine the training data. For example, a lot of apporoaches has been proposed to synthesize Text2SQL training data [9,10]. \\n\\nHowever, in this paper, we primarily follow the mainstream approaches[2,9], focusing on the supervised fine-tuning paradigm, using the high-quality and well-established Text2SQL training data [11,12]. \\n\\nIn summary, exploring weak supervision, unsupervised learning or semi-supervised in low-resource scenario to achieve reliable Text2SQL is promising but beyond the scope of this work. We may explore this direction in our future work.\\n\\n>Q3: Given that database schemas often change over time in production, can ROUTE adapt to new tables or columns without needing extensive retraining, or would these require ongoing fine-tuning?\\n\\n**Ans:** Good question. Our fine-tuned LLM can be directly applied to new tables without the need for additional fine-tuning in novel scenarios.\\n\\nDuring inference, we supply the input prompt with necessary information from the test database. Consequently, our ROUTE can be seamlessly applied to cross-domain datasets beyond the original training domains (SPIDER and BIRD). This is demonstrated by our results on the SPIDER variants, as illustrated in **Section 4.2-Table 2**.\\n\\nAdditionally, we present the results on the Dr. spider benchmark [13], as requested by Reviewer 2daL. This benchmark includes 17 perturbation test sets designed to simulate dynamic scenarios in real-world applications. Our findings, displayed in the following table, clearly demonstrate that our ROUTE method maintains a significant advantage. The results indicate that our approach can be effectively applied to dynamic environments without compromising its performance.\\n\\n**Table 13 in Appendix A.5:** The performance on Dr.Spider benchmark.\\n\\n| |Avg.DB|Avg.NLQ|Avg.SQL|Avg.all |\\n|-|-|-|-|-|\\n| Methods|Pre~Post|Pre~Post|Pre~Post|Pre~Post |\\n| Llama3|70.1~54.3|70.6~56.8|69.1~65.5|69.9~58.8 |\\n| Llama3 + MCP|75.6~59.1|77.0~61.7|74.8~72.6|75.8~64.4 |\\n| Llama3 + SFT|83.4~66.0|83.0~72.8|79.9~77.6|82.1~72.2 |\\n| Llama3 + SFT + MCP|85.4~68.0|85.3~75.1|84.3~81.7|85.0~74.9 |\\n| Llama3 + MSFT|83.8~66.3|82.9~72.5|80.0~77.5|82.2~72.1 |\\n| **Llama3 + ROUTE**|**86.7~69.6**|**85.4~75.8**|**84.5~81.9**|**85.5~75.8** |\\n\\n\\n### Summary\\nWe have noted that your main concerns (see in weakness) about our work arise from the unclear definitions or descriptions of several key terms, such as <Question, SQL>, \\\"SQL hallucinations,\\\" and \\\"Overfitting in Text2SQL\\\". We acknowledge and appreciate your feedbacks, and have revised the expressions that may lead to confusions. \\n\\nConsidering that you have not provided significant negative feedbacks on our work in terms of motivation, technical contribution, and experimental performance, we kindly request that you re-evaluate our work based on our detailed explanations.\\n\\n### Reference \\n>[1] Din-sql: Decomposed in-context learning of text-to-sql with self-correction, NeureIPS, 2023.\\\\\\n>[2] Dts-sql: Decomposed text-to-sql with small large language models, arxiv preprint, 2024.\\\\\\n>[3] Benchmarking the text-to-sql capability of large language models: A comprehensive evaluation[J], arXiv preprint, 2024.\\\\\\n>[4] Before Generation, Align it! A Novel and Effective Strategy for Mitigating Hallucinations in Text-to-SQL Generation, ACL 2024\\\\\\n>[5] PURPLE: Making a Large Language Model a Better SQL Writer, ICDE 2024.\\\\\\n>[6] A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions, ACM TOIS.\\\\\\n>[7] https://huggingface.co/datasets/philikai/200k-Text2SQL \\\\\\n>[8] https://huggingface.co/datasets/gretelai/synthetic-text-to-sql \\\\\\n>[9] Codes: Towards building open-source language models for text-to-sql, ACM SIGMOD 2024.\\\\\\n>[10] Synthesizing text-tosql data from weak and strong llms, ACL 2024.\\\\\\n>[11] Spider: A large-scale human-labeled dataset for complex and cross-domain semantic parsing and text-to-sql task, EMNLP 2018.\\\\\\n>[12] Can llm already serve as a database interface? a big bench for large-scale database grounded text-to-sqls, NeureIPS, 2023.\\\\\\n>[13] Dr.Spider: A Diagnostic Evaluation Benchmark towards Text-to-SQL Robustness, ICLR, 2023.\"}", "{\"comment\": \"Thanks to the authors for their detailed responses, I will increase my score accordingly.\"}", "{\"title\": \"Thanks for the efforts\", \"comment\": \"Thank you so much for your efforts and addressing some of my concerns.\\n\\n> On SPIDER, our ROUTE-14B achieves similar performance as CHESS based on Gimini or GPt-4/4o\\n\\nAlthough the performance is on par with the CHESS and CHASE approaches on the Spider dataset, it is important to note that both of these approaches evaluated their pipelines on the Spider dataset without any fine-tuning or prompt optimization. This was done to demonstrate the generalizability of their methods. In contrast, ROUTE is a fine-tuned model, making this comparison somewhat less equitable.\\n\\n> We have reported the Recall and Precession results\\n\\nThank you so much for sharing these insightful results. As a suggestion, I believe it would be valuable to include this in the main paper.\"}", "{\"comment\": \"Dear Reviewer TcaB,\\n\\nWe sincerely thank you for your timely feedback and increased score. \\n\\nSince your current score is still below the acceptance threshold, we would appreciate any additional concerns you may have about our submission. We will do our utmost to address any remaining concerns during the rebuttal stage (about **48** hours left). Therefore, we sincerely hope you can offer further suggestions or feedbacks to help us improve the manuscript to meet the acceptance threshold. \\n\\nThank you again for your feedbacks and for taking the time to consider our response. We would appreciate it if you could consider raising your score to support the acceptance of this submission, if you have no further concerns.\\n\\n\\nBest,\\nAuthors\"}", "{\"title\": \"Official reply to Reviewer TcaB's remaining concerns [1/2]\", \"comment\": \"We sincerely appreciate your timely and valuable feedback. To address your concerns, we would like to further elaborate on the innovations of our ROUTE and highlight how it differs from existing methods.\", \"the_valuable_insights_and_significant_contributions__of_our_work_can_be_summarized_as_follows\": [\"ROUTE is among the pioneering frameworks under the context of LLMs that explores multi-task tuning and collaborative prompting to improve Text2SQL performance.\", \"We have exhaustively introduced and defined three important tasks in SQL generation, demonstarting that multi-task tuning and collaborative prompting in Schema Linking, Noise Correction and Continuation Writing significantly improve SQL generation accuracy. The additionally introduced SQL-related tasks are well integrated during both the training and inference phases.\", \"We have achieved state-of-the-art performance in 7B/14B-sized LLMs on both the widely-recognized SPIDER and BIRD benchmarks, with verified generalization and transferability across various cross-domains benchmarks and LLMs.\"], \"we_highlight_the_key_similarities_and_differences_between_our_route_and_other_related_works_as_follows\": \">**Multi-task Supervised Fine-tuning (MSFT)**: The method most comparable to our ROUTE approach is MAC-SQL[2], which introduces multiple task agents and demonstrates the effectiveness of fine-tuning through the use of multi-agent instructions on CodeLlama-7B.\\n>- **First**, the defined tasks in MSFT for ROUTE differ from those in MAC-SQL. We have introduced a new continuation writing (CW) task to further refine the challenging SQL queries. As demonstrated in **Ans-W4**, CW holds significant potential for SQL generation. On SPIDER development set, exploring CW task is able to achieve an impressive EX score of **91.1**.\\n>- **Second**, in MAC-SQL, generating instruction data for SFT involves decomposing complex questions into multiple sub-questions and constructing corresponding answers. In contrast, our approach, beyond noise correction, allows for the synthesis of SFT data for various tasks using programming functions. This makes our method more practical for large-scale multi-task data synthesis for MSFT.\\n>- **Third**, in terms of performance, our ROUTE is significantly outperforms MAC-SQL based on the open-source LLM of CodeLlama-7B. The detailed results are presented in the table below.\\n>\\n>|Methods (fine-tuning)|SPIDER-Dev-EX|BIRD-Dev-EX|\\n>|-|:-:|:-:|\\n>|CodeLlama-7B (SQL-Llama) + MAC-SQL|76.3|43.9|\\n>|CodeLlama-7B + ROUTE|83.2|52.2|\\n> \\n> **SQL-Data Synthesis**: Our ROUTE involves the synthesis of SQL-related instruction-following data, which shares similarities with the recent work SENSE[3].\\n>- **First**, compared to SENSE, the data synthesis pipeline of ROUTE encompasses not only Text2SQL but also multiply other SQL-related tasks. Our approach focuses on utilizing existing data to synthesize multi-task SFT data, thereby enhancing the capabilities of open-source LLMs to handle various SQL-related tasks. In contrast, SENSE mainly focused on SQL generation task, leveraging strong LLMs to increase the diversity of SQL generation training set and synthesize preference data.\\n>- **Besides**, our ROUTE achieves comparable performance to SENSE on the SPIDER development set and better performance on the BIRD development set, as shown in following table.\\n>\\n>**Table 10 (Partially) in Appendix A.2**: The performance (EX) of different open-source LLMs.\\n>|Methods (fine-tuning)|SPIDER-Dev-EX|BIRD-Dev-EX|\\n>|-|:-:|:-:|\\n>|CodeLlama-7B + SENSE|83.2|51.8|\\n>|CodeLlama-7B + ROUTE|83.2|52.2|\"}", "{\"title\": \"Official reply to Reviewer 2daL [1/2]\", \"comment\": \"We thank Reviewer 2daL for the positive feedback and constructive suggestions that contribute to improving this paper. Below, we provide one-on-one responses to your comments.\\n\\n>W1: Although this paper focuses more on open-source LLMs, some recent approaches, such as CHASE-SQL, Distillery, and CHESS, are not included as benchmarks in their experiments.\\n\\n**Ans:** Thank you for your valuable suggestion, we have included several recent methods in the revised version to ensure a comprehensive comparison and review of related work. Below is a brief summary, please refer to **Appendix A.3** in our revised submission for the details.\\n\\n**Table 11 in Appendix A.3:** The comparisons with recent methods on SPIDER and BIRD.\\n\\n| Methods | SPIDER-Dev-EX | BIRD-Dev-EX | BIRD-Dev-VES |\\n|---|:---:|:---:|:---:|\\n| CHASE-SQL + Gemini 1.5 | 87.6 | 73.1 | 73.0 |\\n| Distillery + GPT-4o | - | 67.2 | 72.9 |\\n| CHESS + proprietary (GPT-4) | 87.2 | 65.0 | 65.4 |\\n| ROUTE + Qwen2.5-14B | 87.3 | 60.8 | 65.2 |\\n\\nFrom the results, we observed that: \\n- On SPIDER, our ROUTE-14B achieves similar performance as CHESS based on Gimini or GPt-4/4o. This suggests that our ROUTE is an exceptional choice in both conventional and privatized Text2SQL scenarios.\\n\\n- On BIRD, our ROUTE fall behind CHESS + proprietary by 5 point and CHASS-SQL + Gemini 1.5 by 12 point. We believe this is due to the complexity of the BIRD database, which contains numerous tables and columns in a single database, resulting in an extensive input context. It is widely recognized that smaller-sized LLMs (7B/14B) have relative limitations in reasoning capabilities and managing lengthy texts. \\n\\n>W2: The proposed approach is a multi-step pipeline that can be prone to error propagation. To better understand the performance of the schema linking module and ensure it is not introducing errors into the pipeline, it would be beneficial to report the precision and recall of the schema linking module, as done in CHESS and DTS-SQL.\\n\\n**Ans:** Thank you for your constructive feedback. We have reported the Recall and Precession results in **Appendix A.4** of our revised submission. We also present the results as follows for convenience, which demonstartes several important observations: \\n- After MSFT, the schema linking capability is significantly improved, especially in terms of precision scores.\\n- A higher Recall score generally leads to improved EX performance due to minor information loss.\\n- While simplifying the database schema is necessary, ensuring its completeness is more crucial for achieving enhanced performance.\\n\\n**Table 12 in Appendix A.4:** The performance of schema linking.\\n\\n| \\u3000||SPIDER|Table|Table|Column|Column|BIRD|Table|Table|Column|Column |\\n|-|-|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|\\n| $\\\\text{SL}_{\\\\sigma_s}$|$\\\\text{SL}_{\\\\sigma_t}$|Dev-EX|Recall|Precession|Recall|Precession|Dev-EX|Recall|Precession|Recall|Precession |\\n| \\u2713 |\\u2713 |85.80 |**98.75** |94.67 |**99.24** |96.36 |56.00 |**95.21** |88.50 |**96.95**|89.93|\\n| \\u2713 ||84.10 |97.38 |95.71 |98.59 |96.98 |52.70 |90.87 |90.22 |96.13 |90.89|\\n| |\\u2713 |85.00 |97.01 |97.26 |98.21 |97.99 |54.50 |91.60 |93.14 |94.15 |94.11|\\n| ||83.30 |100.00 |18.27 |100.00 |40.05 |53.10 |100.00 |12.23 |100.00 |31.64|\\n| \\u2713 |\\u2713 |73.30 |**97.43** |74.99 |**98.61** |90.33 |36.80 |**93.65**|73.57 |**95.78** |84.36|\\n| \\u2713 ||64.50 |88.35 |76.37 |94.83 |91.46 |30.40 |83.77|75.38 |89.55 |86.39|\\n| |\\u2713 |73.10 |94.24 |91.60 |97.12 |95.30 |35.40 |82.15 |89.01 |88.32 |91.63|\\n| ||69.30 |100.00 |18.27 |100.00 |40.05 |32.10 |100.00 |12.23 |100.00 |31.64|\\n\\n\\nNote that the first four rows are the results of post-MSFT LLMs, and the last four rows are the that of the original LLMs.\"}" ] }

[ "Nguyen T Thach", "Patrick Habecker", "Anika R. Eisenbraun", "W. Alex Mason", "Kimberly A. Tyler", "Bilal Khan", "Hau Chan" ]

[ "AI for Public Health", "large language models", "heterogeneous time-series classification", "multi-label classification" ]

{ "note_id": [ "yZa46ph24O", "yCWg5SSEub", "tDPZGM9Vz4", "qM3wFSeH9x", "p6IngveqnO", "nyHEqSDBX2", "nUFSagsJ9r", "msOa40nVpo", "etlAsibYA9", "d7XMVXmXUh", "cpbtAZHAFj", "cRucPlpyoD", "Yrsav03n4u", "XBcdK58lvn", "Neqwqu77rS", "HpOs7zhdSu", "H1oXKVKEEI", "GGLtc7N4iJ", "BNRXbGFIQN", "7vWYMdSxU4", "7CHyhG1jWF", "5oSRcKrL0g", "5SZIETOGwo", "5IHX0JwgwP" ], "note_type": [ "official_comment", "official_comment", "meta_review", "official_review", "official_review", "official_review", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment" ], "note_created": [ 1732742932583, 1732383130807, 1734893674489, 1730009703802, 1730352583998, 1730634715578, 1732733519875, 1729949665108, 1732552947315, 1732381747796, 1732384543786, 1737523825359, 1732383947002, 1732385071117, 1732392693692, 1732508925731, 1732547246660, 1732407579793, 1732553004349, 1732743628642, 1732553480452, 1732407697039, 1732742281019, 1732497499440 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission7240/Authors" ], [ "ICLR.cc/2025/Conference/Submission7240/Authors" ], [ "ICLR.cc/2025/Conference/Submission7240/Area_Chair_qMf6" ], [ "ICLR.cc/2025/Conference/Submission7240/Reviewer_8yZu" ], [ "ICLR.cc/2025/Conference/Submission7240/Reviewer_aJj3" ], [ "ICLR.cc/2025/Conference/Submission7240/Reviewer_wDHz" ], [ "ICLR.cc/2025/Conference/Submission7240/Reviewer_qBMd" ], [ "ICLR.cc/2025/Conference/Submission7240/Reviewer_qBMd" ], [ "ICLR.cc/2025/Conference/Submission7240/Authors" ], [ "ICLR.cc/2025/Conference/Submission7240/Authors" ], [ "ICLR.cc/2025/Conference/Submission7240/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission7240/Authors" ], [ "ICLR.cc/2025/Conference/Submission7240/Authors" ], [ "ICLR.cc/2025/Conference/Submission7240/Reviewer_qBMd" ], [ "ICLR.cc/2025/Conference/Submission7240/Reviewer_aJj3" ], [ "ICLR.cc/2025/Conference/Submission7240/Authors" ], [ "ICLR.cc/2025/Conference/Submission7240/Authors" ], [ "ICLR.cc/2025/Conference/Submission7240/Authors" ], [ "ICLR.cc/2025/Conference/Submission7240/Authors" ], [ "ICLR.cc/2025/Conference/Submission7240/Reviewer_wDHz" ], [ "ICLR.cc/2025/Conference/Submission7240/Reviewer_aJj3" ], [ "ICLR.cc/2025/Conference/Submission7240/Authors" ], [ "ICLR.cc/2025/Conference/Submission7240/Authors" ] ], "structured_content_str": [ "{\"comment\": \"Dear Reviewer wDHz,\\n\\nWe are glad to hear that we have addressed your concerns. Again, we greatly appreciate your helpful review of our paper! Your comments and suggestions have significantly improved the overall quality of our work.\\n\\nSincerely,\\n\\nAuthors of the current ICLR submission\"}", "{\"comment\": \"We thank Reviewer aJj3 for your thorough review of our paper. We would like to respectfully present our rebuttal in the following.\\n\\n## Weaknesses\\n\\n1. For clarification, in our paper, we refer \\\"high-dimensional time series\\\" to those having a large number of features recorded over time and \\\"long time series\\\" to those consisting of many time points (i.e., either collected at high frequency or over a long period of time). While we indeed only conducted experiments on datasets with at most 30 time points, our methods can generally deal with long time series via the data conversion procedure as clarified in Appendix D.4 following your review. Unfortunately, we are not able to extend our experiments to longer time series due to the limitations of currently available ubiquitous health datasets, and hence we acknowledge this limitation from our work and have mentioned it in Appendix D.6.\\n\\n2. Although the two MuHBoost variants are based on the two stated existing works, we believe our work is the first to adapt their methodologies to LLMs. For (Nam et al., 2017), which reformulates MLTC problems and redesigns RNN architectures for generic text labeling, we adapted their idea by redesigning the inference directive in the inference prompt of LLMs, without the need for architectural design. Albeit straightforward, our approach effectively reduces the computational complexity of the problem and hence may alleviate hallucinations from LLMs as empirically shown in our newly added **MuHBoost vs. MuHBoost[LP+]** under Appendix C.3. For (Li et al., 2023), we apply the idea of combining AdaBoost with classifier chains (using traditional classifiers i.e., SVM for building weak learners) to enable LLM-based boosting classifier chains, which is novel to our best knowledge. Because this transforms MLC problems into multiple binary classification problems, our adaptation brings an alternative way of employing LLMs for MLC without the need for prompting multiple related questions simultaneously, which risks hallucinations, especially for complex prediction tasks. Overall, the novelty of our methods comes mainly from the approach rather than the individual techniques.\\n\\n## Questions\\n\\n1. Recall that SummaryBoost was designed specifically for tabular data (Manikandan et al., 2023) and does not directly apply to longitudinal human behavior data. On the other hand, XGBoost can be applied to both forms of data. Therefore, it is impossible to compare SummaryBoost and XGBoost in our settings. However, following our results from Table 1, our methods still outperform XGBoost on LifeSnaps and GLOBEM, which contain mostly numerical features (in both time-series and auxiliary data). This can be attributed to the fact that while XGBoost works especially well for tabular data (Borisov et al., 2022), the same may not apply to longitudinal human behavior data, highlighting the contributions of our work.\\n\\n2. As mentioned in Appendix C.2 when describing our baselines, for MLkNN and MLTSVM, we embedded the data descriptions via OpenAI\\u2019s embeddings API, which are then fed into these classifiers as input features. This is consistent with how the work of SummaryBoost (the foundation of our work) processed the data descriptions for their kNN baseline.\\n\\n3. Following your suggestion, we have included another baseline based on RNN (Du et al., 2021), which was recommended by the authors of GLOBEM, and found that its performance is substantially worse than vanilla MuHBoost (and hence the two MuHBoost variants as well). Please refer to **Additional Benchmarking** in Appendix C.3 for further details.\\n\\n## References\\n(Borisov et al., 2022) Deep Neural Networks and Tabular Data: A Survey.\\n\\n(Du et al., 2021) AdaRNN: Adaptive Learning and Forecasting for Time Series.\"}", "{\"metareview\": \"This paper proposed an efficient LLM-based multi-label classification method over longitudinal and spare data. The method was evaluated on 13 tasks from 4 datasets, mainly on longitudinal human behaviors (eg substance use, mental health).\\nOne of the focus of the downstream tasks is to classify changes with minimal seen examples and sparse longitudinal data. \\n\\nThe method is built on SummaryBoost, in order to address two limitations. First, the data could be heterogeneous in terms of data type (e.g., continuous measurements such as heart rate, and categorical responses, such as EMA data). \\nSecond, the data doesn't have to be large scale, in which LLM fine-tuning can be more useful. In this case, the authors adapted an existing method - SummaryBoost, and specifically Cluster Sampling in their framework. \\nThus, the method is quite versatile as it is able to handle data heterogeneity. \\n\\nThe discussion between authors and reviewers was robust. Almost all the reviewers increased their scores during the rebuttal. \\nThe biggest weakness of this paper is the adaptation of existing methods -- and the seemingly simple prompting mechanisms. This major concern was raised by one of the reviewers and seems that the authors have not fully addressed their concerns. \\n\\nin my view is this paper is quite well executed, with a very comprehensive set of experiments and benchmarks. The idea may seemingly be 'too simplistic' - however, perhaps that is all that is needed for the tasks at hand.\", \"additional_comments_on_reviewer_discussion\": \"One reviewer championed the paper for acceptance, whereas another reviewer was quite adamant that the work has limited novelty given that it's only adapted existing method -- eg Cluster Sampling, SummaryBoost, into an LLM-based method, specifically with just prompting.\\n\\nThis paper is just above the borderline range, however, the major concern requires SAC to weigh in their views, as the reviewers are not unanimous in their recommendation.\\nWe need SAC to weigh in - whether the contributions are deemed enough for ICLR.\"}", "{\"summary\": \"This paper introduces MuHBoost, a multi-label boosting approach for predicting health and well-being outcomes from longitudinal human behavior data. MuHBoost is designed to address challenges associated with heterogeneous data, missing values, and computational demands in existing models. By incorporating LLM prompting with multi-label classification, MuHBoost aims to predict multiple health outcomes concurrently. Additionally, two variants of MuHBoost are proposed to manage potential hallucinations in LLMs, potentially enhancing model robustness. Evaluated on 13 prediction tasks using four health datasets, MuHBoost reportedly outperforms baseline models across multiple metrics, with resource efficiency highlighted as a key feature\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": [\"Demonstrates how LLMs can be used as multi-label boosting method for temporal tabular data and auxiliary tabular data.\"], \"weaknesses\": [\"Iterative improvements on prior SummaryBoost work with limited novelty\", \"The new data conversion uses a seemingly straightforward two-step approach based on LLM prompting in Section 3.1. The prompting itself does not seem to be very innovative, and while Table 2 does demonstrate an empirical performance increase, it is unclear how exactly each step of the two step approach helped.\", \"MuHBoost modifies the original ClusterSampling algorithm by emphasizing rarer classes in the sampling procedure, then the follow-up methodologies MuHBoost[LP+] and MuHBoost[CC] seem to be straightforward adaptations of prior methods from Nam et al., 2017 and Adaboost.C2.\", \"Prior work section does not seem to contextualize the importance and relevance of SummaryBoost, especially in context of the longitudinal ubiquitous computing domain.\"], \"questions\": \"Comments:\\n* \\\\cite is incorrectly used throughout the paper when \\\\citep should be used\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper proposes a LLM-based method for multi-label classification, with focus on longitudinal human behavior. Built on SummaryBoost, this paper aims to address two limitations in the existing framework. First, the data could be heterogeneous in terms of data type (e.g., continuous measurements such as heart rate, and categorical responses, such as EMA data). Second, minimize the consumption of computing, time etc for inference. The proposed MuHBoost can handle the heterogeneous data type and enable efficient MLC. At the same time, two variants are proposed to mitigate hallucinations from LLMs. Extensive numerical experiments are conducted to evaluate the performance of the proposal, showing its potential advantages in longitudinal human behavior.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. The paper identifies some significant issues of the existing LLM-based prediction methods, especially when dealing with multi-labels in longitudinal human behavior data.\\n\\n2. The paper is well-written, and easy to follow, even if I do not have much research experience in this area.\\n\\n3. The proposed enhancement for MuHBoost is helpful for small sample size problem, where N < 2^Q.\", \"weaknesses\": \"1. I don't think the proposed method can handle realistic high-dimensional time series, for example, movement, heart rate data collected from smartwatch. In the numerical experiments, the highest dimensionality of consideration is less than 30. Traditional machine learning, such as (Zhang et al., 2024) can handle very high frequency time series data (seconds, minutes), which can capture the human behavior in much finer level. Therefore, I think it is essential to discuss the limitations of the proposed method for high-frequency time series data, or justify the performance in numerical experiments, for example, authors can test the proposal on LifeSnaps dataset with much finer data. In addition, the authors may compare the proposal with the aforementioned method for prediction. If it is difficult to handle high-frequency time series data in the current proposal, authors may discuss how their method can be extended to those realistic settings.\", \"reference\": \"Zhang, J., Xue, F., Xu, Q., Lee, J. and Qu, A., 2024. Individualized dynamic latent factor model for multi-resolutional data with application to mobile health. Biometrika, p.asae015.\\n\\n2. I think the enhancement of MuHBoost in Section 3.3 lack novelties. Two approaches are largely motivated by two papers mentioned in the text. Specifically, LP+ is motivated by (Nam et al., 2017) while CC relies on the AdaBoost.C2 (Li et al., 2023). Please emphasize the novelties differentiated from these papers. Emphasize its benefits when adapting these approaches to the MuHBoost.\", \"questions\": \"1. In the original SummaryBoost paper, the numerical results show that SummaryBoost actually perform worse than traditional Xgboost when datasets have many continuous features. I think this is also the case for longitudinal human behavior data, such as heart rate. Did authors observe this phenomenon in the experiments?\\n\\n2. Following the question 1, what are the input for MLkNN and MLTSVM? are they original measurements or embedding of text information or something else?\\n\\n3. I think it worth comparing the proposed method with some traditional machine learning algorithms dealing with longitudinal data, such as RNN-type methods.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The paper proposes a novel approach named MuHBoost to model longitudinal human behavior. It is based on large language models (LLM) and aims to address the limitations of prior related works: inadequate consideration of realistic data type and the consumption of computing resources.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. the proposal of MuHBoost as a resource-efficient multi-label classification method holds promise for solving the problem of multi-outcome health prediction.\\n2. The inclusion of both time series and ancillary data demonstrates the versatility of dealing with heterogeneous data formats common in health behavior modeling.\\n3. Comparative analyses show that MuHBoost generally outperforms benchmark methods on a variety of datasets, demonstrating its effectiveness.\", \"weaknesses\": \"1.The limited sample size of some datasets may affect the generalizability of the findings. This issue is only briefly acknowledged.\\n2. Scalability issues remain, especially in terms of the computational resources required when using MuHBoost for larger datasets or a wider range of applications.\\n3. The ideas presented in the article should be supported by more detailed derivation of the principle formulas.\", \"questions\": \"1.\\tIn Section 1, the authors need more explicit narration of how their model tackles the three challenges presented. Adding a clearer distinction between MuHBoost and previous LLM-based approaches would help readers better understand the unique contributions of the model.\\n\\n2.\\tIn Section 2, most of the Multi-Label Classification related works are a bit out-of-date, more recent works (2023/2024) should be included.\\n\\n3.\\tIn Section 3, \\u201cthe increased computational burden due to longer and more complex prompts may lead to hallucinations from LLMs\\u201dneeds more solid proof. The hallucination of LLM is a widely and deeply studied topic, the authors haven\\u2019t provided clearly (experimentally/theoretically) how their proposed method addresses this issue. For section 3.1: The data transformation process could be graphically illustrated to make it clearer to readers who are not familiar with data aggregation techniques.\\n\\n4.\\tIn Section 4, some of the compared baselines are too old (e.g., Random Forest and XGBoost). The state-of-the-art results are not convincing enough. \\n\\n5.\\tIn Section 4 and Appendix C.3, the resource consumption problem is not fully elaborated. The time complexity analysis needs further proof. In addition, more experimental indicators (e.g., FLOPS) should be included to prove the efficiency of the proposed method.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"thank you!\", \"comment\": \"Thank you to the authors for providing an additional clarification paragraph on SummaryBoost. This was very helpful! Space permitting, I recommend the authors consider adding this paragraph (or parts of it) to the main text. I will raise my score.\"}", "{\"summary\": \"The authors consider the task of using longitudinal ubiquitous health data to predict human behavior. They create a model, called MuHBoost, based on both LLMs (SummaryBoost model methodology) and multilabel classification (label powerset and classifier chain methods). The authors compare their model to various baselines and show that MuHBoost outperforms all baselines and presents two key advantages: (i) MuHBoost can consider different feature types and missing values and (ii) MuHBoost requires less computing time and cost. To address LLM hallucination the authors also build two extensions of their base model (which result in better performance). The authors apply their model to 4 datasets to predict psychology, student performance, and drug use.\", \"soundness\": \"4\", \"presentation\": \"2\", \"contribution\": \"3\", \"strengths\": \"S1 - Important application. The authors investigate an important application of longitudinal human behavior modeling with applications to patient monitoring, interventions for drug users, etc.\", \"s2___multiple_datasets\": \"The authors evaluate their model on multiple, diverse datasets with a range on the number of labels (ranging from 2 to 6 labels).\", \"s3___multiple_baselines_and_metrics\": \"The authors compare their model to a diverse set of baselines and evaluate on multiple standard MLC metrics.\", \"s4___multiple_robustness_checks\": \"The authors run various robustness checks including on SummaryBoost\\u2019s extract-then-refine procedure and on the use of GPT4 (more resource expensive) vs GPT3.5 (less resource expensive).\", \"weaknesses\": \"W1 - Needs more background clarification: The main weakness of the paper is a lack of background clarification. The authors should give more clarification on the SummaryBoost, LP and CC methods which the MuHBoost model is built off of.\", \"w2___comparison_to_multilabel_text_classification\": \"The authors state that the use of LLMs for MLC type prediction tasks have been done in multilabel text classification works. How do these methods compare to MuHBost?\", \"w3___statistical_significance\": \"Are the results in table 1 statistically significant?\", \"w4___only_binary_class_predictions\": \"The authors claim that MuHBoost presents an advantage over prior work because their model can consider different feature types. However, all experiments run in the paper only consider binary class predictions. Have the authors evaluated the model on other feature types as well (e.g. categorical/continuous prediction)?\", \"questions\": \"1. How do the authors combine data descriptions across topics?\\n2. For the two psychology datasets (LifeSnaps and GLOBEM), why are some tasks (e.g., negative emotions or anxiety) defined differently across datasets?\\n3. When describing baselines in Section 4.2, what do problem transformation and algorithm adaptation mean?\\n\\nMinor comments / suggestions: \\n\\n- Throughout I recommend using parenthetical citations (citep instead of cite)\\n- At multiple points, footnotes referenced in the main text were placed in the appendix. Could each footnote be placed on the page they are first referenced on?\\n- Line 081: \\u201chave the two following\\u201d\\n- Line 206: \\u201crequire\\u201d instead of \\u201crequires\\u201d\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Dear Reviewer 8yZu,\\n\\nAs the end of the discussion period approaches, we would like to check in with you on whether we have addressed your concerns of our paper. We are happy to discuss further if that is not the case.\\n\\nSincerely,\\n\\nAuthors of the current ICLR submission\"}", "{\"comment\": \"We thank Reviewer wDHz for taking the time to review our paper. We would like to first comment on the weaknesses and then address your questions.\\n\\n## Weaknesses\\n\\n1. Recall that our main goal is to develop methodologies that mainly consider the three stated challenges of ubiquitous health data (and two shortcomings of state-of-the-art i.e., LLM-based approaches). These challenges include small sample size due to the typically high data acquisition cost. You are correct that this brings a limitation to our work. Based on your suggestions, we have included a Limitation section in Appendix D.6. More specifically, we do not expect our proposed methods to scale to larger datasets (i.e., with thousands of samples) in terms of resource efficiency. When the opportunity to work with large ubiquitous health data arises, approaches that require ample sample size such as LLM finetuning become more suitable than ours i.e., boosting LLM-generated weak learners, which is also the case for SummaryBoost.\\n\\n2. Yes, you are correct. Following up on our previous point, our approach focuses on ubiquitous health datasets that typically come with small sample sizes. Therefore, we do not claim the scalability of our developed methods for large datasets. Nevertheless, we believe the proposed idea of employing LLMs for MLC in real-world health prediction tasks is novel (as discussed in related work from Appendix B) and could be applied to significantly reduce resource consumption for data-extensive approaches such as finetuning (which is prohibitively expensive when the number of health/well-being outcomes to predict, Q, is large).\\n\\n3. We apologize for any confusion in our presentation. During writing, we tried to follow the outlines and technical details provided in the SummaryBoost paper from NeurIPS-23, which forms the foundation of our methodology. In the paper, the authors presented their work with only algorithms and illustrations, and hence we adopted this convention. We are willing to elaborate on our derivations if Reviewer wDHz could provide more details on which \\\"principle formulas\\\" are in question.\\n\\n## Questions\\n\\n1. Based on your suggestions, we have further clarified how our proposed methods address the three challenges of ubiquitous health data compared to existing LLM-based approaches in Appendix D.4.\\n\\n2. Our current literature review has already provided a thorough coverage on MLC, as we are not aware (as of writing this rebuttal) of any innovative work from established venues in recent years. For MLTC, we have included our updated literature review. Please refer to the last paragraph in Appendix B for a discussion on more recent works.\\n\\n3. Based on your suggestion, we have provided more solid proof. In particular, in Table S10 (newly added from your review), we show that MuHBoost underperforms MuHBoost[LP+], particularly for PWUD dataset where the number of labels Q is larger. For further details, please refer to **MuHBoost vs. MuHBoost[LP+]** in Appendix C.3.\\n\\n> The hallucination of LLM is a widely and deeply studied topic, the authors haven\\u2019t provided clearly...\\n\\nOur approach actively alleviates LLM hallucinations during (1) the data conversion procedure and (2) the inference process (via the two MuHBoost variants). For (1), we demonstrated its positive effect through our ablation study Impact of Refining Data Description in Section 4.3. To further support our claim based on your review, we have conducted another ablation study where we omit the entire data conversion procedure (Table S11). Please refer to **Further Ablation** in Appendix C.3. For (2), please refer to **MuHBoost vs. MuHBoost[LP+]** in the same appendix for how our two MuHBoost variants can mitigate hallucinations from LLMs.\\n\\n> For section 3.1: The data transformation process could be graphically illustrated...\\n\\nBased on your suggestion, we have included an illustrative example in Figure S4 of our revised paper.\\n\\n4. In existing longitudinal human behavior modeling studies, all state-of-the-art approaches were compared against these baselines. Therefore, we naturally considered them as baselines following the convention. We also compared our methods to those state-of-the-art approaches. Furthermore, based on your review, we have included another baseline (Du et al., 2021), which was recommended by the authors of GLOBEM as a potential solution to longitudinal human behavior modeling. Our results in Table S12 show that MuHBoost still outperforms this baseline. Please refer to **Additional Benchmarking** in Appendix C.3 for further details.\\n\\n5. Based on your suggestions, we have included an analysis of time complexity in Appendix D.5. Because it may take arbitrary time to call LLMs (e.g., intermittent delays), we do not report empirical runtime in our experimental results (see also Footnote 16). We have also included a paragraph on FLOPs estimation in Appendix D.5.\\n\\n## Reference\\n(Du et al., 2021) AdaRNN: Adaptive Learning and Forecasting for Time Series.\"}", "{\"comment\": \"We thank Reviewer qBMd for your thorough review of our paper. We would like to respectfully present our rebuttal as follows.\\n\\n## Weaknesses\\n\\n> W1 - Needs more background clarification\\n\\nWe appreciate your concern regarding the background of SummaryBoost. To clarify, it was proposed to specifically address tabular data classification only with small sample sizes and high heterogeneity. Therefore, we do not include it in the background or related work, though we did clarify its relevance to ubiquitous health data in the Introduction (in Contribution I). For MLC approaches such as LP and CC methods, we introduced some background and discussed relevant works in Section 2 and Appendix B (under MLC subsection for both). We also clarified important works upon which our methods are built throughout Section 3 as well as Appendix D.1 (for MuHBoost[CC]).\\n\\n> W2 - Comparison to multilabel text classification\\n\\nExisting methods in multilabel text classification (MLTC) mentioned in the Related Work (last paragraph in Appendix B) have only been applied to generic text labeling tasks on large datasets (with thousands of samples). Furthermore, to our best knowledge, none employ commercial LLMs such as GPT (mostly on pretrained transformers such as BERT instead).\\n\\n> W3 - Statistical significance\\n\\nWe ranked the considered methods in Table 1 by the average of the respective performance measures (Hamming accuracy, macro F1, and micro F1) across 10 different data splits. Therefore, some results are not statistically significant (i.e., ties, particularly among the baselines). However, we have double-checked using paired sample t-test with significance level of 0.05 and confirmed that (i) our three proposed methods strictly outperform all baselines, (ii) the LLM-based baselines (zero-shot and few-shot methods) strictly outperform other baselines.\\n\\n> W4 - Only binary class predictions\\n\\nWe believe binary class prediction aligns with most applications in longitudinal human behavior modeling, where the goal is often to diagnose certain health/well-being outcomes such as depression, stress, and problematic use of various drugs. Therefore, we only considered binary class prediction for all tasks in this work. Conceptually, because our methods extend SummaryBoost (which considered single-label binary or multiclass classification) to MLC, extending them to multiclass MLC (where some or all of the Q labels are multiclass) is straightforward, as mentioned earlier in Footnote 6. We will consider extending our methodology and experiments in future work.\\n\\n## Questions\\n\\n> 1. How do the authors combine data descriptions across topics?\\n\\nFor clarifications, in the first step of our data conversion prompt, when the number of topics is reasonably small as illustrated in Figure S4b (newly added based on your review), we fit all topics within the data conversion prompt, hence there is only one data description for each record. When there are many topics, we split them into groups and respectively call the LLM to describe each group, resulting in multiple data descriptions. Then, in the second step, we append these data descriptions together (i.e., separated by newlines) to form a complete data description of the record.\\n\\n> 2. For the two psychology datasets (LifeSnaps and GLOBEM), why are some tasks (e.g., negative emotions or anxiety) defined differently across datasets?\\n\\nFollowing your review, we have clarified the factors deciding our label definitions for all considered datasets in Appendix C.1. For GLOBEM, we defined the label for PANAS task following the definition from (Englhardt et al. 2024), which is grounded on domain literature. For LifeSnaps, which studied participants in a different geographical region (Europe instead of U.S.), we simply defined the PANAS task for demonstration purposes only given the lack of conclusive findings regarding PANAS cutoffs in the area.\\n\\n> 3. When describing baselines in Section 4.2, what do problem transformation and algorithm adaptation mean?\\n\\nWe would like to categorize the baselines given the two main approaches for MLC, which were first introduced in Section 2 (under MLC subsection).\\n\\n### Minor comments / suggestions\\n\\n> Throughout I recommend using parenthetical citations (citep instead of cite)\\n> Line 081: \\u201chave the two following\\u201d\\n> Line 206: \\u201crequire\\u201d instead of \\u201crequires\\u201d\\n\\nWe have incorporated these helpful suggestions in our revised paper.\\n\\n> At multiple points, footnotes referenced in the main text were placed in the appendix. Could each footnote be placed on the page they are first referenced on?\\n\\nDue to the page limit, we resort to placing some footnotes in the appendix (only when they are mentioned in both the main text and the appendix).\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}", "{\"comment\": \"We thank Reviewer 8yZu for taking the time to review our paper. We would like to respectfully present our rebuttal as follows.\\n\\n## Weaknesses\\n\\n> Iterative improvements on prior SummaryBoost work with limited novelty\\n\\nWhile we indeed based our approach on SummaryBoost, it is not immediately clear how a tabular data classification method for small datasets can be applied to our problem of TSC given ubiquitous health data. Because this general form of time-series data may have high heterogeneity in practice (which has yet to be addressed by prior works as discussed in Appendix A), we leverage SummaryBoost to circumvent this challenge given its efficacy in heterogeneous tabular data classification and employment of LLMs (proven to be suitable for longitudinal human behavior modeling). Thus, we believe the novelty of our work comes mainly from the approach, rather than the individual techniques. \\n\\n> The new data conversion uses a seemingly straightforward two-step approach based on LLM prompting in Section 3.1. The prompting itself does not seem to be very innovative, and while Table 2 does demonstrate an empirical performance increase, it is unclear how exactly each step of the two step approach helped.\\n\\nFollowing up on our previous point, while the data conversion procedure is based on prompting techniques from existing works in various fields, it can effectively deal with general time series with high heterogeneity and dimensionality. Moreover, it can accommodate any auxiliary background data from the records, which can also be heterogeneous and high-dimensional (e.g., PWUD dataset). To our best knowledge, this flexibility is unseen in relevant works, which highlights the novel contribution of our data conversion approach.\\n\\nRegarding the individual performance of our two-step procedure, because the second step requires the data description(s) in complete sentences as input, we can only ablate either (i) the second step or (ii) both steps altogether. Based on your review, we have extended our ablation in Appendix C.3, under **Further Ablation**. As shown in Table S11, ablating (ii) results in a sharp decline in predictive performance compared to ablating (i), which implies more impact from the first step.\\n\\n> MuHBoost modifies the original ClusterSampling algorithm by emphasizing rarer classes in the sampling procedure, then the follow-up methodologies MuHBoost[LP+] and MuHBoost[CC] seem to be straightforward adaptations of prior methods from Nam et al., 2017 and Adaboost.C2.\\n\\nAlthough the two MuHBoost variants are based on existing works, we believe our work is the first to adapt their methodologies to LLMs. For (Nam et al., 2017), which reformulates MLTC problems and redesigns RNN architectures for generic text labeling, we adapted their idea by redesigning the inference directive in the inference prompt of LLMs, without the need for architectural design. Albeit straightforward, our approach effectively reduces the computational complexity of the problem and hence may alleviate hallucinations from LLMs as empirically shown in our newly added **MuHBoost vs. MuHBoost[LP+]** under Appendix C.3.\\n\\nFor (Li et al., 2023), we apply the idea of combining AdaBoost with classifier chains (using traditional classifiers i.e., SVM for building weak learners) to enable LLM-based boosting classifier chains, which is novel to our best knowledge. Because this transforms MLC problems into multiple binary classification problems, our adaptation brings an alternative way of employing LLMs for MLC without the need for prompting multiple related questions simultaneously, which risks hallucinations especially for complex prediction tasks. Overall, we respectfully stand by our view that despite being straightforward and built upon prior works, the proposed methods are still novel in the sense of how they approach our (nontrivial and understudied) problem of interest in a simple-yet-effective way.\\n\\n> Prior work section does not seem to contextualize the importance and relevance of SummaryBoost, especially in context of the longitudinal ubiquitous computing domain.\\n\\nFor clarification, SummaryBoost was proposed to specifically address tabular data classification with small sample sizes and high heterogeneity. Therefore, we do not include it in the related work as it cannot be applied to the longitudinal ubiquitous computing domain. We previously clarified its relevance to ubiquitous health data in the Introduction (in Contribution I). This unique approach to longitudinal human behavior modeling is hence the main novelty of our work.\\n\\n## Questions\\n\\n> \\\\cite is incorrectly used throughout the paper when \\\\citep should be used\\n\\nWe appreciate Reviewer 8yZu for noticing our inadvertent mistake. We have taken this into account in our revised paper.\"}", "{\"title\": \"Revised Paper Uploaded\", \"comment\": \"Dear Reviewers,\\n\\nWe greatly appreciate your thorough reviews of our paper, which significantly improve the overall quality of the proposed work. In response to your helpful comments and suggestions, we have uploaded our revised paper to this portal. Please note that when we say \\\"please refer to Appendix XYZ\\\" in our rebuttal, we are referring to the revised version.\\n\\nSincerely,\\n\\nAuthors of the current ICLR submission\"}", "{\"title\": \"Response to authors' reply\", \"comment\": \"Thank you to the authors for responding to my concerns and questions. As stated in W1, my main critique was that the authors do not provide enough background clarification about the methods they build on (e.g. SummaryBoost). The authors' response was that they didn't include a detailed background on SummaryBoost because this method was built for a distinct application. While this may be true, as a reader who didn't have prior experience with the SummaryBoost method, having this background would make the methods description in the paper (section 3) much easier to parse. For this reason, I am choosing to keep my score as is (a weak accept).\"}", "{\"comment\": \"I greatly appreciate authors' timely responses, which address all my concern. I will raise the score accordingly.\"}", "{\"comment\": \"We sincerely thank Reviewer aJj3 again for the time spent reviewing our paper! Your helpful comments and suggestions have significantly improved the overall quality of our work.\"}", "{\"comment\": \"We appreciate Reviewer qBMd for your concern regarding the background clarification in our paper. We agree that this is necessary for readers who are unfamiliar with SummaryBoost, which is an important piece in our approach. Following your suggestion, we have included a subsection providing some background on SummaryBoost in Appendix D.7. We would like to keep the section labels intact to avoid confusion in this rebuttal, hence we will reorganize this background information (such that it appears first in the appendix) after the end of the current discussion period.\\n\\nPlease let us know if the background we provided is sufficient. We are happy to include any further clarification.\"}", "{\"comment\": \"Dear Reviewer wDHz,\\n\\nAs the end of the discussion period approaches, we would like to check in with you on whether we have addressed your concerns of our paper. We are happy to discuss further if that is not the case.\\n\\nSincerely,\\n\\nAuthors of the current ICLR submission\"}", "{\"comment\": \"Dear Reviewer 8yZu,\\n\\nWhile we will not be able to upload our revised PDF after today (November 27 UTC), we can still address any of your further concerns. Please let us know if there is anything else we can provide to improve our paper.\\n\\nSincerely,\\n\\nAuthors of the current ICLR submission\"}", "{\"comment\": \"Dear Authors,\\n\\nI have carefully reviewed all the reviewers' comments and your responses. Since these comments and responses have resolved my concerns, I will not repeat any questions.\", \"have_a_nice_day\": \")\"}", "{\"comment\": \"Thanks for clarifying all my questions and I really appreciate your efforts. I have the following comments:\\n\\n1. given that there is no experiments for \\\"real-world\\\" long time series (very high-frequency time series like heart beat), I am not convinced that the proposed method can accommodate that very well. Or at least authors should state in the main text about this limitation, otherwise, it's misleading for readers that this paper can handle those real challenges.\\n\\n2. My major concern about LLM for predictive problems (not limited to longitudinal modeling) is as follows: for continuous numerical data, you must do some conversion, like binning, quantiles, to input into LLM, which will loss some information compared with other machine learning method. For example, if we binarize multiple continuous variables, the correlation among those variables are totally messed up, which may lead to inaccurate downstream tasks. In the presented results, it seems this impact is not very significant, but I concern about its general application. \\n\\nBased on the above reasons, I will keep my score, and glad to hear more from authors.\"}", "{\"comment\": \"We sincerely thank Reviewer qBMd again for the time spent reviewing our paper! Your comments and suggestions have significantly improved the overall quality of our work. We will definitely consider adding the background information on SummaryBoost in the main text of our final version.\"}", "{\"comment\": \"We thank Reviewer aJj3 for the follow-up comments and would like to address them in the following.\\n\\n1. We agree with your concern about the proposed methods. First, we would like to point out our specific goal of longitudinal human behavior modeling in order to better position our contributions in the appropriate context. In this paper, we focus on modeling longitudinal human behaviors (e.g., substance use and mental health) that change in small periods (e.g., days) under heterogeneity (i.e., non-numerical with high rate of missing values) and potentially high dimensionality of the data. We believe this setting is realistic for applications involving early detection (i.e., binary classification) of certain health/well-being outcomes such as depression, stress, and problematic use of various drugs. On the other hand, works that specifically address long time series such as (Zhang et al., 2024) as you mentioned earlier focus on time-series interpolation, which is outside the scope of our paper. Regardless, we acknowledge the lack of supporting experiments for long time series in our work and will clearly state this limitation in the main text.\\n\\nWhile we do not focus on long time series, our methods can be adapted to said settings via data preprocessing (i.e., via our data conversion procedure). As mentioned in Section 3.1 and Appendix D.4 from our revised paper, this is achieved by segmenting the multivariate time series into multiple subperiods and then respectively prompting the LLM to describe each. We notice that datasets containing a large number of variables (i.e., high dimensionality as what we call in our paper), all or most of which are collected at high frequency, are extremely costly in real-world applications. Typically, only a small number of variables (such as heart rate) can be collected at fine scales. This applies to LifeSnaps, GLOBEM, and the real dataset employed in (Zhang et al., 2024) (which only includes five variables). In such cases, because the goal is to capture the overall pattern of each time-varying variable (e.g., heart rate follows specific rhythms), we can treat each finely recorded variable as a separate 'topic' within the data conversion procedure (assuming the number of those variables in the dataset is reasonably small). Given that LLMs have been empirically shown to be capable of zero-shot identifying patterns in general sequences (Gruver et al., 2023), we believe such accommodation leveraging the flexibility of our data conversion procedure can effectively deal with the long time series issue in practice.\\n\\n2. We understand your concern regarding the employment of LLMs for real-world applications. As mentioned in both version of our paper (Line 1066 in the revised one), we do not discretize numerical data, i.e., by binning them into percentiles, as SummaryBoost originally did in the experiments. This is motivated by the fact that LLMs (e.g., GPT-3+) can already effectively zero-shot identify patterns in general sequences of raw numbers (Gruver et al., 2023), as well as the interpretability concerns for general health-related applications. Therefore, our data conversion procedure ensures minimal information loss for numerical time-series data, which is reflected in the good predictive performance of our methods on the LifeSnaps and GLOBEM datasets (consisting of mostly numerical features).\\n\\n## References\\n(Zhang et al., 2024) Individualized dynamic latent factor model for multi-resolutional data with application to mobile health.\\n\\n(Gruver et al., 2023) Large Language Models Are Zero-Shot Time Series Forecasters. NeurIPS-23.\"}" ] }

[ "Yikun Bai", "Abihith Kothapalli", "Hengrong Du", "Rocio Diaz Martin", "Soheil Kolouri" ]

[ "Optimal transport", "Gromov-Wasserstein problem", "Unbalanced optimal transport" ]

{ "note_id": [ "z9C4fD3Fp6", "yO0Z1Dy8GM", "yIrrYSKYi3", "vWaoR4yxcy", "svnBHX0ZKp", "skhL9GfbjN", "sCALUhKDSs", "r01Jsw3dR9", "m6o2Lb4uV1", "kEB5ipdJ9J", "hZqx4PXWsw", "hPhn1fE6p1", "Ylm0D0c8cB", "YY3RezGDwy", "Vn5oSAZXky", "VX0dn70JpA", "TruKsxBQ20", "P5E2PTUAz9", "Mf15Y8bQyh", "IQPiAB5615", "HeqIxxtkP5", "FY7PnZziCp", "ALjv2IclZq" ], "note_type": [ "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review", "official_review", "meta_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "decision" ], "note_created": [ 1732079753983, 1732079804793, 1732079711499, 1732079904845, 1732122437869, 1732557640594, 1732557403419, 1732079199405, 1732407275540, 1730697718157, 1730692523830, 1730268175725, 1732553625589, 1734430938723, 1732080037515, 1732079989784, 1733217075184, 1732078972048, 1732079613662, 1732080437057, 1732679741270, 1729972636015, 1737524058474 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission10517/Authors" ], [ "ICLR.cc/2025/Conference/Submission10517/Authors" ], [ "ICLR.cc/2025/Conference/Submission10517/Authors" ], [ "ICLR.cc/2025/Conference/Submission10517/Authors" ], [ "ICLR.cc/2025/Conference/Submission10517/Authors" ], [ "ICLR.cc/2025/Conference/Submission10517/Authors" ], [ "ICLR.cc/2025/Conference/Submission10517/Authors" ], [ "ICLR.cc/2025/Conference/Submission10517/Authors" ], [ "ICLR.cc/2025/Conference/Submission10517/Reviewer_wRQx" ], [ "ICLR.cc/2025/Conference/Submission10517/Reviewer_K3go" ], [ "ICLR.cc/2025/Conference/Submission10517/Reviewer_wRQx" ], [ "ICLR.cc/2025/Conference/Submission10517/Reviewer_ueVA" ], [ "ICLR.cc/2025/Conference/Submission10517/Reviewer_HYNr" ], [ "ICLR.cc/2025/Conference/Submission10517/Area_Chair_rMnP" ], [ "ICLR.cc/2025/Conference/Submission10517/Authors" ], [ "ICLR.cc/2025/Conference/Submission10517/Authors" ], [ "ICLR.cc/2025/Conference/Submission10517/Reviewer_ueVA" ], [ "ICLR.cc/2025/Conference/Submission10517/Authors" ], [ "ICLR.cc/2025/Conference/Submission10517/Authors" ], [ "ICLR.cc/2025/Conference/Submission10517/Authors" ], [ "ICLR.cc/2025/Conference/Submission10517/Authors" ], [ "ICLR.cc/2025/Conference/Submission10517/Reviewer_gpEb" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ] ], "structured_content_str": [ "{\"title\": \"Weakness 3, approximation\", \"comment\": \"**W3: Approximation of LPGW.** Equation (20) is the empirical implementation of the LPGW distance, formulated analogously to the empirical implementation of the LGW distance (see line 215).\\n\\nIn Table 1, we show the relative error gap (MRE) and Pearson correlation (PCC) between (20) and PGW. We find that when the reference measure is well-chosen (i.e., $\\\\mathbb{S}_7, \\\\mathbb{S}_8, \\\\mathbb{S}_9$), (20) and PGW exhibit low MRE (0.02) and high PCC (0.99). **Note that in this experiment, we do not have a Monge mapping for any embedding, since all the reference measures and target measures have distinct sizes and some of the reference measures are non-uniform. See the [repo](https://anonymous.4open.science/r/Linearized_Partial_Gromov_Wasserstein-F43E/ellipses/ellipses.ipynb), file `ellipse/ellipses.ipynb` for details.**\\n\\nUsing approximation formulations to replace the original linear OT/GW/PGW is common in other linear OT-based techniques as well. \\n\\nFor example, in the LOT method, LOT (Eq. (27)) is proposed to approxiamte OT. aLOT (Eq. (32)) is then proposed to approximate LOT.\\n\\nFurthermore, the aLOT (Eq. (32)) is approximated by: \\n$$\\\\|\\\\mathcal{T}_{\\\\gamma^1}-\\\\mathcal{T}_{\\\\gamma^2}\\\\|^2\\\\tag{b}.$$\\n\\nNote, to clarify it, it is not a multi-layers approximation. LOT (27) and aLOT (32) are only proposed for theoretical completeness and are not computational feasible. \\n\\nIn practice, (b) is used to approximate OT, though (b) is neither LOT (27) nor aLOT (32). However, in most of the existing literature that cites [LOT](https://link.springer.com/article/10.1007/s11263-012-0566-z), (b) is generally treated as the LOT distance and (27), (32) are not presented. \\n\\nOur approximation (20) follows analogously from the previous works.\"}", "{\"title\": \"Weakness 4: Experiments.\", \"comment\": \"**W4: Experiments.** Given that LPGW is an extension of LGW to the partial GW setting, our first aim is to demonstrate the performance of LPGW in existing experiments, such as the shape retrieval experiment in the [LGW paper](https://arxiv.org/pdf/2112.11964). In this experiment, we demonstrate that when the tested shapes have different numbers of points, LPGW achieves better accuracy with faster speed. Our aim here is not to invent completely new expiriments; rather, we deliberately choose existing experiments for better comparison.\\n\\n> The experiment in 4.3 is similar to that from S. Nieter, R. Cummings, Z. Goldfeld, Outlier-Roubust Optimal Transport: Duality, Structure, and Statistical Analysis, AISTATS 2022. \\n\\nThe authors respectively disagree with this opinion. In the MNIST classification experiment (Section 4.3), the experiment setup differs from the generative model (GANs) experiment in [Nieter et al., 2023](https://arxiv.org/pdf/2111.01361). Furthermore, the \\\"outlier-robust Wasserstein distance\\\" (Eq 1) which is used in [Nieter et al., 2023](https://arxiv.org/pdf/2111.01361) can be treated as a variant of partial OT, and is not robust to rotation and flipping. \\n\\n\\nThe problem setup of Experiment 4.3 is about the classification of data (e.g. MNIST digits) when the data in the testing set is corrupted by rotation/flipping/random noise. The flipping/rotation corruption is not considered in [Nieter et al., 2023](https://arxiv.org/pdf/2111.01361). \\n\\n> This appears a bit contrived in the setting of MNIST.\\n\\nTo the best of our knowledge, besides noise corruption, classifying rotated/flipped digits is also a common problem in machine learning research, as seen in, e.g. [Zhou et al, 2017](https://arxiv.org/abs/1701.01833). We believe the inclusion of random rotations/flipping in our work is important to demonstrate the robustness of our method; this was likely not included in [Nieter et al., 2023](https://arxiv.org/pdf/2111.01361) and the method presented there is not robust to rotation/flipping.\"}", "{\"title\": \"Weakness 2, Question 1, Monge mapping\", \"comment\": \"> Many of the results require the existence of a Monge map for the PGW problem.\\n\\nTo clarify, similar to LOT/LGW, **none of the formulations for the LPGW embedding technique (e.g., LPGW distance, LPGW embedding, aLPGW distance, etc.) rely on the Monge mapping assumption**. The formulation of the LPGW distance (e.g., Eq. (14)), which requires a Monge mapping, is presented in the main text only as a simplified formulation for the reader\\u2019s convenience. However, when a Monge mapping does not exist, these concepts can still be well-defined. We would be happy to discuss this in further detail if the reviewer so pleases.\", \"the_theoretical_results_requiring_the_monge_mapping_assumption_are\": \"- Proposition 3.1 (2), (3), (4)\\n- Theorem 3.2 (3)\\n\\nProposition 3.1 (2) characterizes the relationship between the barycentric projection and the Monge mapping, and we felt its inclusion important for the reader to understand the barycentric projection. \\n\\nProposition 3.1 (3), (4), and Theorem 3.2 (3) address the gap between the proposed (approximated) LPGW distance and the PGW distance. In Proposition 3.1 (3), we clarify the conditions under which the LPGW distance can exactly recover the PGW distance. Proposition 3.1 (4) shows when the numerical implementation (Eq. (20)) can recover PGW, and Theorem 3.2 (3) states the conditions for aLPGW to recover LPGW.\\n\\n**Conditions for the existence of Monge mapping**: In the continuous measure setting and the empirical discrete measure setting (i.e., where the mass of each point in $\\\\mu$ and $\\\\nu$ is the same), GW admits a Monge mapping (e.g., [Dumont et al., 2024](https://arxiv.org/pdf/2210.11945), Proposition 1.16, Theorem 2.1; [Courty et al., 2019](https://arxiv.org/pdf/1905.10124), Theorem 3.2). However, it is unclear whether this result extends to unbalanced GW (including PGW), as these formulations were proposed only recently. Further characterization of the conditions under which the Monge mapping assumption hold is an open problem left for future work.\\n\\n**Note:** Empirically, we observe that both GW and PGW admit Monge mappings when data are in an empirical distribution.\\n\\n> These results appear to be of limited applicability.\\n\\nWe respectfully disagree with this point. Based on our understanding, the only effect of the absence of Monge mappings in PGW is an increased gap between the LPGW distance and PGW distance (see lines 2231-2234 in the appendix). In a machine learning setting, when target datasets have different sizes or the reference measure has non-uniform distribution, then the Monge mapping assumption may not hold in all linear OT methods, including [LOT](https://link.springer.com/article/10.1007/s11263-012-0566-z), [Linear HK](https://arxiv.org/abs/2102.08807), [Linear OPT](https://arxiv.org/abs/2302.03232), and [LGW](https://arxiv.org/abs/2112.11964). However, these methods are still viable and of broad importance.\\n\\n(We would like to briefly remark, however, that if we manually set the mass for each point in all datasets to be the same, the Monge mapping assumption will still hold in the linear unbalanced OT setting, including our LPGW.)\\n\\nNonetheless, similar to Linear OT/Linear GW, we present the LPGW distance as a proxy for PGW rather than as an exact approximation. In fact, in our shape retrieval experiment, the PGW distance performs worse than LPGW distance in measuring similarity in the 2D dataset (see Table 2).\"}", "{\"title\": \"Questions 2,3,4\", \"comment\": \"**Q2: Role of $\\\\lambda$ in Proposition 3.3.** The authors apologize for this unclear explanation in Proposition 3.3. We will update the statement in proposition 3.1 (3) and 3.3. In particular, \\\"lambda is sufficiently large\\\" will be replaced by the following:\\n\\n$$2\\\\lambda \\\\ge \\\\max_{y_i^1,y_{i'}^1\\\\in Y^1,y^2_j,y^2_{j'}\\\\in Y^2}|g_{Y^1}(y^1_{i'},y_i^1)-g_{Y^2}(y_{j}^2,y_{j'}^2)|^2.\\\\tag{c}$$\\n\\n\\n> Moreover, if $\\\\lambda$ is too large, the regularizer dominates so there appears to be some tradeoff in terms of minimizing the first part of the objective. \\n \\nWhen $\\\\lambda$ is greater than the above threoushold, the maximum amount of mass will be transported, meaning that problem (19) has a closed-form solution. This ensures that all transportation plans for problems $PGW(\\\\mathbb{X}, \\\\mathbb{Y}^1)$, $PGW(\\\\mathbb{X}, \\\\mathbb{Y}^2)$, and (19) remain unchanged once $\\\\lambda$ exceeds this threshold. Intuitively, this implies no further increase in $\\\\lambda$ is needed. Thus, the largest value for $\\\\lambda$ in our experiments is based on this bound. \\n\\n\\n------------\\n\\n**Q3: MNIST Experiment.** There appears to be a potential misunderstanding regarding the experiment presented in Section 4.3. In this experiment, we train a logistic regression model using the embeddings from LOT/LGW/LPGW. Since OT/GW/PGW are not embedding techniques and do not provide embeddings, they cannot be used in this experiment.\\n\\nIf the goal is to use OT/GW/PGW to train a classifier, we can employ models such as a kernel-SVM, as demonstrated in the shape retrieval experiment (Section 4.2).\\n\\nThe main difference between LPGW and PGW lies in their computational complexity. \\nIn summary, LPGW requires approximately 3\\u20135 minutes to train the model, whereas PGW demands ~400 hours (over 15 days). Due to the time constraints of the review period, we were unable to provide classification results using PGW. However, the code is available in the [repository](https://anonymous.4open.science/r/Linearized_Partial_Gromov_Wasserstein-F43E/README.md), `mnist2d/classification.ipynb`.\", \"the_following_is_a_summary_of_the_difference_in_complexity\": \"Suppose the size of training dataset and testing dataset are $N_{train},N_{test}$ respectively. In this experiment $N_{train}=4000, N_{test}=1000$. \\n\\n- Training a kernel-SVM model for GW/PGW requires \\n$\\\\mathcal{O}(N^2_{train}T)$ where $T$ is the complexity for one GW/PGW computation. The testing step requires $\\\\mathcal{O}(N_{test}N_{train}T)$ complexity. \\n\\n- For the LPGW model, the training process requres only $\\\\mathcal{O}(N_{train}T)$ and the testing step requires $\\\\mathcal{O}(N_{test}T)$. \\n\\nThis means that the larger the (training and testing) dataset, the more pronounced the speed advantage of LPGW compared to PGW. \\n\\n------------\\n\\n**Q4: Robustness to Noise.** The authors suspect the reviewer's question is about LPGW, not PGW.\\n\\nWe will add a section to explain the intuitive noise-robustness feature of LPGW. \\n\\nNote, the robustness to noise is property of PGW due to its partial matching property. In PGW, $\\\\lambda$ plays a role as threshold, when the \\\"transportation distance\\\" for a certain pair is greater then $2\\\\lambda$, this pair is unlikely to be transported. When the noise points has larger distance to the clean data, these points are unlike to be transported/paired. \\n\\nBy extension, LPGW utilizes this property to contruct the embedding. That is, with high probability, outliers/noise will not be incoporated when we construct the LPGW embedding. Thus the affect of noise in downstream machine learning tasks (e.g. classification) will be limited. \\n\\nWe also refer to [the following figure](https://anonymous.4open.science/r/Linearized_Partial_Gromov_Wasserstein-F43E/mnist2d/results_visual/embedding.png) for a visulization of the LPGW embedding. We can observe that, when a digit is corrupted with noise, the reconstructed figure based on the LPGW embedding does not contain most of the noise points.\"}", "{\"title\": \"Logistic regression, other models\", \"comment\": \"> How exactly did they use logistic regression to predict classes when they range from 0-9?\\n\\nFor this experiment, we use the [scikit-learn logistic regression](https://scikit-learn.org/1.5/modules/generated/sklearn.linear_model.LogisticRegression.html), which automatically performs multi-class classification (with multinomial loss). Since we use this model out-of-the-box with all of its default settings, we originally did not include further details about its usage. We intend to make all of our source code publicly available to ensure that these experiments are reproducible.\\n\\n> Furthermore, does this work with other models such as neural networks?\\n\\nYes, one could use any model in this experiment that uses the LOT/LGW/LPGW embeddings to perform classification, including neural networks. We chose to use logistic regression for simplicity.\"}", "{\"title\": \"comment\", \"comment\": \"The authors sincerely thank the reviewer for their valuable comments and constructive discussion. The paper will be updated to include the detailed discussion on Monge mapping, approximation error and MNIST experiments. Additionally, the statement of Proposition 3.3 and the citation issue raised by the reviewer has been addressed/updated.\"}", "{\"title\": \"comment\", \"comment\": \"The authors sincerely thank the reviewer for their valuable comments and thoughtful discussion. The additional reference [4] has been added, and the typo in line 260 has been corrected.\\n\\nThe dynamic formulation of PGW is part of our future research direction. We aim to explore how our new formulation aligns with the tangent space and GW geodesics discussed in [Sturm](https://arxiv.org/abs/1208.0434), [Beier et al., 2021](https://arxiv.org/abs/2112.11964), and [Beier et al., 2024](https://arxiv.org/abs/2403.08612).\\n\\nHowever, to the best of our knowledge, there is currently no dynamic formulation for GW (specifically, the related continuity equation) in the general gauge measure space setting. Dynamic GW in a special setting has been studied very recently (see [Zhang et al., 2024](https://arxiv.org/pdf/2407.11800?)).\\n\\nIn conclusion, we agree with the reviewer on the importance of having a dynamic formulation for PGW, and it will be part of our future work.\"}", "{\"title\": \"Weaknesses 1, 2, 3\", \"comment\": \"**W1: Novelty of this paper**:\\n> The paper is mostly combining existing methods and theoretical novelty is limited, with discussions following from the the original ideas and definitions of linearized/partial GW/OT.\\n \\nWhile the general principle of linear partial GW is inspired by linear GW/OT, our proposed formulation of the LPGW embedding is not as simple as combining existing techniques. \\n\\nNotably, for prior linear GW/OT works in the balanced setting, the linear embedding represents the displacement between the reference measure and the given target measures based on the optimal transport plans. \\n\\nIn our work, it is essential to incorporate information about mass creation and destruction into the embedding. Formulating such an embedding within the partial GW setting is highly non-trivial. As such, we feel that describing our approach as simply a combination of existing methods does not fully capture its unique contributions. \\n\\nIn the classical OT/GW/UOT framework, the problem admits a dynamic formulation. Consequently, the linear embedding can be interpreted as a logarithm mapping, capturing the velocity of particles involved in mass transportation as well as the derivatives governing mass creation and destruction. However, the dynamic formulation for UGW/PGW has not yet been thoroughly explored. Incorporating this dynamic information to define an embedding and establishing a similarity measure between two embeddings that can recover LGW and PGW is a significant challenge.\\n\\nIn contrast, these challenges do not arise in the balanced OT/GW setting, where the formulation is inherently simpler.\\n\\nWe kindly encourage the reviewer to reconsider this perspective.\\n\\n> What if TV is replaced by a general divergence for unbalanced formulation as in [1]?\\n\\nAs explained above, defining a new embedding in the unbalanced GW setting based on general divergence penalty is non-trivial. To the best of our knowledge, linear unbalanced GW has not been studied in any setting, regardless of whether the penalty is TV or another divergence (e.g., KL divergence). Even in classical OT setting, the unblanced linear OT embedding has only been studied with the KL penalty or the TV penalty (see e.g., [Linear HK](https://arxiv.org/abs/2102.08807) and [Linear OPT](https://arxiv.org/abs/2302.03232)); a formulation using general divergences has not yet been proposed. \\n\\n\\n**W2: Additional References.** We appreciate the reviewer\\u2019s observation. We would like to remark that [2,3] are already cited in our work (note that they do not directly address unbalanced GW however). We will add [4] to our related works section. (See the updated pdf.)\\n\\n------------\\n\\n**W3: Barycentric LGW & Notation.** LGW\\u2019s theoretical formulation is (8), and its approximation is (10). Equation (10) can be further simplified as:\\n$$\\\\|g_{Y^1}(\\\\mathcal{T}_{\\\\gamma^1}(\\\\cdot_1),\\\\mathcal{T}_{\\\\gamma^1}(\\\\cdot_2))-g_{Y^2}(\\\\mathcal{T}_{\\\\gamma^2}(\\\\cdot_1),\\\\mathcal{T}_{\\\\gamma^2}(\\\\cdot_2))\\\\|^2_{L(\\\\mu^{\\\\otimes2})} \\\\tag{a}$$\\n(We've added a label to this equation in the paper.)\\n\\nIn all experiments, the numerical implementation of LGW is based on (a) rather than (8) or (10). The code in [Github](https://github.com/Gorgotha/LGW) also uses (a), similar to LOT and LPGW. We will add a sentence to clarify this point.\\n\\n> Notation $\\\\gamma^1 \\\\wedge \\\\gamma^2$ on line 260 seems to be incorrect and should be $\\\\inf_A \\\\gamma^1(A\\\\cap E)+\\\\gamma^2(A^C\\\\cap E)$.\\n\\nWe thank the reviewer for this comment and have corrected the typo.\"}", "{\"comment\": \"I thank the authors for their detailed response to my questions. I agree with many of the points raised and hence have updated my score to reflect this. I recommend that the authors clarify these points further in the main text.\"}", "{\"summary\": \"This paper studies the linear partial Gromov-Wasserstein (GW) alignment problem. Specifically the paper extends ideas from linearized GW and partial GW, and combines them to LPGW which inherits the merits of both method and is applicable to wider tasks. To formulate, the paper gives detailed account for how to incorporate general measure instead of probability measures, and how to coordinate different marginals. Extensive numerical experiments are presented, showcasing better performance.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"The paper is well motivated and the study of the proposed formulation is very thorough, covering both general and special cases under suitable assumptions. The proposed algorithm combines the advantages from both partial GW and linearized GW, and extensive experiments supports this extension.\", \"weaknesses\": \"1. The paper, though overall well motivated, is mostly combining existing methods and theoretical novelty is limited, with discussions following from the the original ideas and definitions of linearized/partial GW/OT. A further question: what if TV is replaced by a general divergence for unbalanced formulation as in [1]?\\n\\n\\n2. Discussion of related works: the term \\\"linearization\\\" is from the tangent structure of gauged GW space as is first pointed out in [2]. Mapping based unbalanced GW formulations are also studied in [3],[4].\\n\\n\\n3. Minor comment: \\nIf the barycentric LGW is the version to use as in the second equality of (10), then this should be stated in the paper.\\nNotation $\\\\gamma^1\\\\wedge\\\\gamma^2$ on line 260 seems to be incorrect and should be $\\\\inf_A \\\\gamma^1(A\\\\cap E)+\\\\gamma^2(A^C\\\\cap E)$.\\n\\n[1] S\\u00e9journ\\u00e9, Thibault, Fran\\u00e7ois-Xavier Vialard, and Gabriel Peyr\\u00e9. \\\"The unbalanced gromov wasserstein distance: Conic formulation and relaxation.\\\" Advances in Neural Information Processing Systems 34 (2021): 8766-8779.\\n\\n[2] Sturm, Karl-Theodor. The space of spaces: curvature bounds and gradient flows on the space of metric measure spaces. Vol. 290. No. 1443. American Mathematical Society, 2023.\\n\\n[3] Hur, YoonHaeng, Wenxuan Guo, and Tengyuan Liang. \\\"Reversible Gromov\\u2013Monge sampler for simulation-based inference.\\\" SIAM Journal on Mathematics of Data Science 6.2 (2024): 283-310.\\n\\n[4] Zhang, Zhengxin, et al. \\\"Cycle consistent probability divergences across different spaces.\\\" International Conference on Artificial Intelligence and Statistics. PMLR, 2022.\", \"questions\": \"See above.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The paper proposes a modification of a variant of the Gromov-Wasserstein (GW) distance, known as the Partial GW (PGW) problem, which enables to compare gauged measure (gm) spaces whose measures have different total masses. The modification consists of applying the linearization strategy proposed in (Beier et al., 2022) to the PGW problem, yielding the so-called Linear PGW discrepancy (LPGW). It is shown that if the PGW problem admits a solution that is induced by a deterministic map, then the LPGW problem recovers the PGW problem. For the purposes of computation, an approximated LPGW problem is considered. The paper concludes with some numerical experiments which exhibit the empirical performance of the proposed approach.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"The paper is, in my opinion well-written, and the theoretical findings appear correct. In the experiments, the LPGW algorithm compares favorably to the other approaches considered.\", \"weaknesses\": \"I believe that the paper is of a limited novelty. Indeed, both the linearized GW problem and the partial GW problem have appeared in previous works. Combining these two approaches to obtain a linearized partial GW problem appears to be incremental progress on the important question of efficient computation for the GW problem.\\n\\nFurthermore, the theoretical properties of the LPGW problem appear quite limited. Many of the results require the existence of a Monge map for the PGW problem, but sufficient conditions for these assumptions to hold are not discussed in the main text. This assumption is quite strong and, as such, these results appear to be of a limited applicability. \\n\\nThe approximated LPGW is, in practice approximated as (20). This appears to be a very coarse approximation in general; no justifications are provided for when this might be a reasonable approximation. \\n\\nFinally, the experiments do not present any significantly new applications. The experiments in 4.1 and 4.2 are essentially the same as those from (Beier et al., 2022) the experiment in 4.3 is similar to that from S. Nieter, R. Cummings, Z. Goldfeld, Outlier-Roubust Optimal Transport: Duality, Structure, and Statistical Analysis, AISTATS 2022. The main distinction is that a random rotation is performed on the dataset in the current paper, but this appears a bit contrived in the setting of MNIST.\", \"questions\": \"1. Many results in the paper depend on the assumption that a Monge map exists. Can the authors comment on sufficient conditions for such a result to hold?\\n\\n2. Proposition 3.3 holds when $\\\\lambda$ is sufficiently large. Can the authors clarify this condition further in the main text? Moreover, if $\\\\lambda$ is too large, the regularizer dominates so there appears to be some tradeoff in terms of minimizing the first part of the objective. \\n\\n3. In the numerical experiments, in the experiments on MNIST, it seems that PGW would offer the best comparison (vs LOT and LGW which do not account for noise). Could these results be included?\\n\\n4. I believe that the robustness to noise is the most interesting feature of the PGW formulation. I would recommend that this be addressed more clearly in the text.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper introduces the Linear Partial Gromov-Wasserstein (LPGW) embedding, a computationally efficient approximation of the Partial Gromov-Wasserstein (PGW) problem, which relaxes mass constraints for comparing measures across different metric spaces. By reducing the pairwise complexity from $\\\\mathcal{O}(K^2)$ to $\\\\mathcal{O}(K)$, LPGW preserves the partial matching capabilities of PGW while enhancing scalability. The authors provide theoretical guarantees and validate LPGW through experiments in shape retrieval and transport-based learning tasks.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"3\", \"strengths\": \"1. The paper is well-written and easy to follow.\\n \\n2. The authors introduce the Linear Partial Gromov-Wasserstein (LPGW) embedding, which addresses the computational challenges of PGW. They also provide a formal proof of the equivalence between LPGW and PGW under the PGW-Monge mapping assumption, contributing to the theoretical foundation of the method.\\n \\n3. The paper presents a method to approximate LPGW efficiently, reducing computational overhead. Furthermore, the authors conduct extensive numerical experiments to demonstrate the effectiveness and computational advantages of LPGW.\", \"weaknesses\": \"1. The equivalence between LGW and GW (Proposition 2.1) and between LPGW and PGW (Proposition 3.1) critically relies on the existence of a Monge mapping. However, proving the existence of such a mapping for the GW problem requires satisfying strict conditions [A]. This raises concerns about the robustness of these approximations in practical scenarios where Monge mappings may not exist. The paper would benefit from\\n - providing theoretical bounds on the approximation error when a Monge mapping does not exist;\\n - including empirical experiments comparing performance on datasets known to violate Monge mapping conditions;\\n - discussing potential modifications to make the method more robust when Monge mappings do not exist.\\n \\n2. The computational implementation of LPGW relies on two layers of approximations. First, the authors approximate LPGW using aLPGW, as introduced in Equation (18). Then, Equation (20) further approximates the aLPGW formulation. However, the paper lacks both theoretical guarantees and empirical evaluation of the error introduced by these sequential approximations. It would be beneficial if the authors could\\n - provide theoretical error bounds for each approximation step;\\n - include ablation studies quantifying the empirical error introduced at each approximation stage;\\n - discuss the tradeoffs between computational efficiency and approximation accuracy.\\n \\n3. The paper primarily compares LPGW with PGW and LGW. However, other GW approximation methods could offer meaningful insights into the strengths and weaknesses of LPGW. Notably, comparisons with Low-Rank GW [B] and Sliced GW [C] could demonstrate whether LPGW provides any unique computational or approximation advantages.\\n \\n\\n[A] Dumont, Th\\u00e9o, Th\\u00e9o Lacombe, and Fran\\u00e7ois-Xavier Vialard. \\\"On the existence of Monge maps for the Gromov\\u2013Wasserstein problem.\\\"\\u00a0*Foundations of Computational Mathematics*\\u00a0(2024): 1-48.\\n\\n[B] Scetbon, Meyer, Gabriel Peyr\\u00e9, and Marco Cuturi. \\\"Linear-time gromov wasserstein distances using low rank couplings and costs.\\\"\\u00a0*International Conference on Machine Learning*. PMLR, 2022.\\n\\n[C] Titouan, Vayer, et al. \\\"Sliced gromov-wasserstein.\\\"\\u00a0*Advances in Neural Information Processing Systems*\\u00a032 (2019).\", \"questions\": \"1. Could the authors extend the approximation results for LGW and LPGW to scenarios where the Monge mapping may not exist?\\n \\n2. Could the authors provide empirical or theoretical results on the error introduced by the two layers of approximation? Specifically, how much error is accumulated when transitioning from LPGW to aLPGW (Equation (18)) and then from aLPGW to the final approximation (Equation (20))?\\n \\n3. Proposition 3.1 (3) provides a result when the cost function $g_Y$\\u200b is an inner product. Could this result be extended to the case of squared Euclidean distance with additional assumptions? This would be relevant since [A] also explores the existence of Monge mappings under squared Euclidean costs.\\n \\n4. In the elliptical disks experiment, the choice of reference space has a significant impact on performance. Could the authors provide practical guidance or heuristics for selecting reference spaces to optimize the performance of LPGW?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The authors propose the linear partial Gromov-Wasserstein (LPGW) embedding, a linearization technique for the PGW problem. Theoretically, they prove that LPGW admits a metric with certain assumptions, and numerically, they demonstrate that the utility of the proposed LPGW method in shape retrieval and learning with transform-based embedding tasks. They illustrate that the LPGW- based method can preserve the partial matching property of PGW, while significantly improving computational efficiency.\", \"soundness\": \"4\", \"presentation\": \"4\", \"contribution\": \"3\", \"strengths\": \"1. The paper is well organized and easy to follow.\\n2. The paper contains many details, which is very helpful for understanding.\\n3. The experiment is promising, justifying theoretic part.\", \"weaknesses\": \"I am not an expert in this area, I don't find weaknesses from my perspective.\", \"questions\": \"I am not an expert in this area.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"1\", \"code_of_conduct\": \"Yes\"}", "{\"metareview\": \"The paper proposes the Linear Partial Gromov-Wasserstein embedding, which significantly improves the computational efficiency of Partial Gromov-Wasserstein distance while preserving its partial matching properties. Reviewers initially raised concerns regarding the novelty of the work, questioning whether the combination of linearization and partial GW techniques constituted a significant advancement. Additionally, the reviewer noted the reliance on strong assumptions, such as the existence of Monge maps, and the coarse nature of the approximations. However, the authors clarified that their primary contribution lies in the embedding formulation, distinct from PGW distance, and provided empirical results to validate its robustness and practical utility, even in scenarios where Monge mappings do not exist.\", \"additional_comments_on_reviewer_discussion\": \"During the rebuttal period, the authors addressed the reviewers\\u2019 concerns thoroughly, particularly by clarifying the theoretical role of approximations, improving explanations of Monge mapping assumptions, and highlighting LPGW\\u2019s computational advantages. These clarifications and the strong experimental results showing the method\\u2019s efficiency and robustness solidified confidence in the paper\\u2019s contributions, leading to the final recommendation for acceptance.\"}", "{\"title\": \"Question 1,2,3,4\", \"comment\": \"**Q1: Approximation Results without Monge Mapping Assumption.** In the general case without Monge mapping, LPGW is defined in (55) and LGW in (36). In the approximation error experiment (Section 4.1, Table 2), **a Monge mapping does not exist for any embedding** since all the reference measures and target measures have distinct sizes and some of the reference measures are non-uniform. See the [repo](https://anonymous.4open.science/r/Linearized_Partial_Gromov_Wasserstein-F43E/README.md), file `ellipse/ellipses.ipynb` for details.\\n\\n------------\\n\\n**Q2: Approximation Errors.** As explained above, LPGW (15) and aLPGW (18) are proposed for theoretical completeness and are computationally infeasible. Thus, we use (20) to approximate PGW. This situation is similar for LOT and LGW.\\n\\nIn the LOT paper, both LOT Eq. (27) and aLOT Eq. (32) are inefficient to solve. The only feasible implementation is the approximation of (32): \\n$$\\\\|\\\\mathcal{T}_{\\\\gamma^1} - \\\\mathcal{T}_{\\\\gamma^2}\\\\| \\\\tag{a}$$\\n\\nStrictly speaking, (a) is neither aLOT nor LOT. However, (a) is used directly as the representation of LOT in much of the LOT literature. In LGW/LPGW, the situation is the same.\\n\\n------------\\n\\n**Q3: Extensions of Cost Function.** To the best of our knowledge, the optimality of barycentric projection has been proven only for LOT and linear partial OT. We plan to extend this result to Euclidean cost in future work. In GW/PGW, we just proposed the proof for inner product. In the future, we will extend this result to more general cost. \\n\\nIn our understanding, to extend the result to general cost, the definition of barycentric projection may need to be modified. \\n\\n------------\\n\\n**Q4: Elliptical Disks Experiment.** A suitable reference measure can be chosen as follows:\\n\\n- The reference measure size acts as the \\\"resolution\\\" of the embedding, so a larger size is preferred. Typically, we set this to be the maximum/mean/median of the sizes of all target datasets.\\n\\n- Similar to linear OT and linear GW, the reference measure should ideally be the GW/PGW barycenter, as explained in lines 2364-2366, 2433-2338, and 2476-2478. Corresponding barycenter algorithms can be found in [Peyre et al., 2016](https://proceedings.mlr.press/v48/peyre16.pdf), [Bai et al., 2024](https://arxiv.org/abs/2402.03664).\"}", "{\"title\": \"Weaknesses 1,2,3\", \"comment\": \"We would like to express our gratitude for the thorough and insightful feedback provided in the review. In what follows, we offer our explanations to the reviewer's questions and concerns. We welcome follow-up inquiries and would be happy to engage in more in-depth discussions.\\n\\n\\n**W1: Monge Mapping Assumption.**\\n> Providing theoretical bounds on the approximation error when a Monge mapping does not exist.\\n\\nTo the best of our knowledge, theoretical bounds for LGW and LPGW have not yet been established. This bound will likely depend heavily on the reference measure. Empirically, we observe that increasing the size of the reference measure (i.e., increasing \\\"resolution\\\") helps decrease the error.\\n\\n> Including empirical experiments comparing performance on datasets known to violate the Monge mapping conditions.\\n\\nIn Table 2, we present the relative error and Pearson correlation coefficient for this purpose. Note that in this experiment, **we do not have a Monge mapping for any of the embeddings since all reference measures and target measures have distinct sizes, and some reference measures are non-uniform.** We observe that, when the reference measure is appropriate (e.g., $\\\\mathbb{S}_7, \\\\mathbb{S}_8, \\\\mathbb{S}_9$), the relative error is low (0.02) and the PCC is high (0.99).\\n\\n> Discussing potential modifications to make the method more robust when Monge mappings do not exist.\\n\\nIn our understanding, the existence of a Monge mapping is not essential since this situation can occur with the LOT, LGW, and LHK methods, and in these cases, barycentric projection is generally used. In fact, LPGW (or aLPGW) is viewed as a proxy for the original PGW distance rather than an exact approximation.\\n\\nOur approach for improving accuracy and preserving information in the original target measure includes the following strategies, as discussed in lines 2217-2252. To summarize, accuracy can be enhanced by:\\n\\n1. Choosing a good reference measure: \\n - The size of the reference measure acts as the \\\"resolution\\\" of the embedding, so larger sizes are preferred. Typically, we set this size to be either the maximum or mean/median of the sizes of all target datasets.\\n\\n - Similar to linear OT and linear GW, the reference measure should ideally be set to the GW/PGW barycenter, as explained in lines 2364-2366, 2433-2338, and 2476-2478.\\n \\n2. Setting the mass of each point in the reference and target measures to be equal if feasible. \\n\\n\\n\\n------------\\n\\n**W2: Computational Implementation.** We appreciate the reviewer\\u2019s observation and suggestion. To clarify the \\\"two-layer approximation\\\":\\n\\n- (20) is an approximation of (19), aLPGW.\\n- aLPGW (19) approximates LPGW (15).\\n- LPGW (15) approximates PGW (11).\\n\\nThis relationship explains the derivation of (20). However, it is not a multi-layers approximation. (15)(19) are only proposed for theoretical completeness and is not computational feasible. \\n\\n\\nIn practice, we directly approximate PGW with (20), rather than through (19) or (15), which were proposed for theoretical completeness, similar to LOT and LGW.\\n\\nThe above approximation is similar to LOT/LGW. That is LOT, aLOT, LGW, aLGW are only proposed for theoretical completeness. The practical formulations are \\n$$aLOT(\\\\mu,\\\\nu)\\\\approx \\\\|\\\\mathcal{T}_{\\\\gamma^1}-\\\\mathcal{T}_{\\\\gamma^2}\\\\|_{L(\\\\mu)}^2$$\\n\\n$$aLGW(\\\\mu,\\\\nu)\\\\approx \\\\|\\\\mathcal{T}_{\\\\gamma^1}-\\\\mathcal{T}_{\\\\gamma^2}\\\\|_{L(\\\\mu^{\\\\otimes2})}^2.$$\\n\\n\\n\\n> Provide theoretical error bounds.\\n\\nDeveloping a theoretical error bound is part of our future work. To our knowledge, there is a gap in this area for both LGW and LPGW.\\n\\n> Discuss the trade-offs between computational efficiency.\\n\\nSimilar to LOT and LGW, both (15) and (19) are computationally inefficient (in fact, there is not an existing solver for (15) or (19)). These formulations are presented primarily for theoretical completeness and to potentially support future developments (e.g., LPGW geodesics). Thus, only (20) is computationally efficient.\\n\\n------------\\n\\n**W3: Comparisons with Existing Methods.** LPGW is proposed as a proxy for Partial GW distance, and, to the best of our knowledge, low-rank GW and sliced-GW have not been extended to the partial GW setting.\\n\\nAdditionally, sliced-GW [C] solves a variant of the GW problem:\\n $$\\\\sum_{l=1}^L GW((\\\\theta_l)_\\\\#\\\\mathbb{X}, (\\\\theta_l)_\\\\#\\\\mathbb{Y}).$$\\n\\nApart from the difference between GW and PGW, there are other issues:\\n - The optimal value in the formulation above represents the average transport cost based on 1D projections, which can differ significantly from the transport cost in the original GW/PGW problem.\\n - Theorem 3.1 in [C] requires the Monge mapping assumption. However, in our ellipse experiment (Section 4.1) and MNIST classification experiment (Section 4.3), a Monge mapping does not exist and so the closed form in their main theorem cannot be applied.\"}", "{\"comment\": \"I thank the authors for their detailed response. I will keep my score unchanged.\"}", "{\"title\": \"Response to common questions and concerns\", \"comment\": \"# Main Contribution\\nFrom a theoretical perspective, our proposed LPGW formulation consists of two parts: the **LPGW embedding** and the **LPGW distance**.\\n- The LPGW embedding is the primary contribution of this paper, and we numerically demonstrate that this embedding is robust to rotation/flipping/noise corruption, compared with the LOT embedding or the LGW embedding. In addition, the embedding can be used to reconstruct (denoised) data (see [the following figure](https://anonymous.4open.science/r/Linearized_Partial_Gromov_Wasserstein-F43E/mnist2d/results_visual/embedding.png)). Meanwhile, OT/GW/PGW does not provide such an embedding. \\n\\n- The LPGW distance is a byproduct of the LPGW embedding. This distance can be treated as a proxy/approximation of PGW, and it significantly improves the computational complexity. \\n\\nWe would like to clarify that the LPGW distance is only a byproduct of the proposed embedding. **Specifically, the main contribution lies beyond providing an approximation for the PGW distance, as the reviewer's comment might suggest.** \\n\\n# Novelty of the paper\\nWhile the general principle of linear partial GW is inspired by linear GW/OT, our proposed formulation of the LPGW embedding is not as simple as combining existing techniques. \\n\\nNotably, for prior linear GW/OT in the balanced setting, the linear embedding represents the displacement between the reference measure and the given target measures based on the optimal transport plans. \\n\\nIn our work, it is essential to incorporate information about mass creation, transportation and destruction into the embedding. Formulating such an embedding within the partial GW setting is highly non-trivial. As such, we feel that describing our approach as simply a combination of existing methods does not fully capture its unique contributions. \\n\\nIn the classical OT/GW/UOT framework, the problem admits a dynamic formulation. Consequently, the linear embedding can be interpreted as a logarithm mapping, capturing the velocity of particles involved in mass transportation as well as the derivatives governing mass creation and destruction. However, the dynamic formulation for UGW/PGW has not yet been thoroughly explored. Incorporating this dynamic information to define an embedding and establishing a similarity measure between two embeddings that can recover LGW and PGW is a significant challenge.\\n\\nIn contrast, these challenges do not arise in the balanced OT/GW setting, where the formulation is inherently simpler. \\n\\n# Monge mapping assumption (We refer Appendix C in the updated pdf for details.) \\nTo clarify, similar to LOT/LGW, **none of the formulations for the LPGW embedding technique (e.g., LPGW distance, LPGW embedding, aLPGW distance, etc.) rely on the Monge mapping assumption**. The formulation of the LPGW distance (e.g., Eq. (14)), which requires a Monge mapping, is presented in the main text only as a simplified formulation for the reader\\u2019s convenience. However, when a Monge mapping does not exist, these concepts can still be well-defined. We would be happy to discuss this in further detail if the reviewer so pleases. Lack of Monge mapping may affect the gap between PGW and LPGW, and there is no other affections. \\n\\n# Approximation error (We refer appendix J in the updated pdf for details) \\nThe relation between PGW, LPGW, aLPGW can be described as follows: \\n- (20) is an approximation of (19), aLPGW.\\n- aLPGW (19) approximates LPGW (15).\\n- LPGW (15) approximates PGW (11).\\n\\nHowever, it is **not a multi-layers approximation**. (15)(19) are only proposed for theoretical completeness and is not computational feasible. \\n\\nUsing approximation formulations to replace the original linear OT/GW/PGW is common in other linear OT-based techniques as well. \\nFor example, in the LOT method, LOT (Eq. (27)) is proposed to approxiamte OT. aLOT (Eq. (32)) is then proposed to approximate LOT.\\nFurthermore, the aLOT (Eq. (32)) is approximated by: \\n$$\\\\|T_{\\\\gamma^1}-T_{\\\\gamma^2}\\\\|^2\\\\tag{b}.$$\\nIn practice, (b) is used to approximate OT, though (b) is neither LOT (27) nor aLOT (32). However, in most of the existing literature that cites [LOT](https://link.springer.com/article/10.1007/s11263-012-0566-z), (b) is generally treated as the LOT distance and (27), (32) are not presented. \\nOur approximation follows analogously from the previous works. \\n\\nIn Table 2, we present the relative error and Pearson correlation coefficient for this purpose. **we do not have a Monge mapping in this experiment.** We observe that, when the reference measure is appropriate (e.g., $\\\\mathbb{S}_7, \\\\mathbb{S}_8, \\\\mathbb{S}_9$), the relative error is low (0.02) and the PCC is high (0.99). Applying GW/PGW barycenter can help to find a good reference.\"}", "{\"title\": \"Weakness 1, novelty\", \"comment\": \"We would like to thank the reviewer for their feedback. We hope the following adequately addresses the reviewer's concerns and would be more than happy to engage in further discussions.\\n\\n**W1: Novelty.** First, we would like to clarify a potential misunderstanding of the LPGW formulation and our main contributions. The following will be emphasised in the main text and abstract. \\n\\nFrom a theoretical perspective, our proposed LPGW formulation consists of two parts: the **LPGW embedding** and the **LPGW distance**.\\n- The LPGW embedding is the primary contribution of this paper, and we numerically demonstrate that this embedding is robust to rotation/flipping/noise corruption, compared with the LOT embedding or the LGW embedding. In addition, the embedding can be used to reconstruct (denoised) data (see [the following figure](https://anonymous.4open.science/r/Linearized_Partial_Gromov_Wasserstein-F43E/mnist2d/results_visual/embedding.png), which we plan to add to the paper). Meanwhile, the embedding is orthogonal to PGW, as OT/GW/PGW does not provide such an embedding. \\n\\n- The LPGW distance is a byproduct of the LPGW embedding. This distance can be treated as a proxy/approximation of PGW, and it significantly improves the computational complexity. \\n\\nWe would like to clarify that the LPGW distance is not the primary contribution of our work; rather, it is a byproduct of the proposed LPGW embedding. Specifically, our main contribution lies beyond providing an approximation for the PGW distance, as the reviewer's comment might suggest.\\n\\n>Combining these two approaches to obtain a linearized partial GW problem appears to be incremental progress on the important question of efficient computation for the GW problem.\\n\\n\\nThe authors respectfully disagree with the opinion that the proposed LPGW is merely a combination of PGW and LGW.\\n\\nIn the classical OT/GW setting, the problem admits a dynamic formulation, where the embedding is defined through the logarithm mapping. This mapping incorporates both the velocity of mass transportation and the derivative related to mass creation and destruction (noting that in balanced OT/GW, the derivative for mass creation/destruction is zero). However, in UGW, including PGW, a dynamic formulation has not yet been explored. Incorporating this dynamic information to define the embedding and establishing a similarity measure between embeddings that can recover LGW and PGW in specific settings remains a significant challenge in the field.\\n\\nFrom a computational perspective, we demonstrate that for $K$ different metric measure spaces, computing their pairwise PGW distances requires solving the PGW problem $\\\\mathcal{O}(K^2)$ times. In contrast, computing the LPGW distances pairwise requires only $\\\\mathcal{O}(K)$ distance computations, representing a dramatic improvement in computational efficiency.\\n\\nIn addition, our work completes the theoretical analysis of linear GW in the [original paper](https://arxiv.org/abs/2112.11964) regarding barycentric projection (see Proposition 2.1 (3)), which is essential for explaining why the barycentric projection mapping can be used in the approximate aLGW formulation (e.g., (10)). We then extend this result to the PGW setting (see Proposition 3.2).\"}", "{\"title\": \"Time complexity\", \"comment\": \"We appreciate the reviewer\\u2019s time and insightful feedback. It seems the\\n\\n**W1 and Q1: Complexity.** In Appendix C, we discuss the complexity with respect to $K$. Here\\u2019s a summary:\\n\\nLet $n$ be the size of the reference measure and $m$ the average size of all target measures. \\n\\n- Pairwise computation of PGW/GW distance for $K$ measures requires \\n$$\\\\binom{K}{2} \\\\mathcal{O}\\\\left(\\\\frac{1}{\\\\epsilon} nm(n + m)nm^2\\\\right),$$ \\nwhere $\\\\epsilon > 0$ is the accuracy tolerance.\\n\\n- The time complexity for LGW/LPGW is \\n$$\\\\binom{K}{2} \\\\mathcal{O}(n^2) + K \\\\mathcal{O}\\\\left(\\\\frac{1}{\\\\epsilon} nm(n + m)n^2 m^2\\\\right).$$\\n\\nNote that if we apply the Sinkhorn algorithm in the solvers for GW/PGW, the term $nm(n + m)$ can be improved to $\\\\frac{1}{\\\\epsilon} nm \\\\ln(n + m)$.\\n\\nIn general, $n, m \\\\gg K$ (i.e., the size of each dataset is much larger than the number of datasets). Therefore, in (a), the term $\\\\binom{K}{2}(n^2)$ is negligible, and the dominant term is $K \\\\mathcal{O}\\\\left(\\\\frac{1}{\\\\epsilon} nm(n + m)n^2 m^2\\\\right)$, which is linear with respect to $K$.\"}", "{\"comment\": \"# LPGW VS PGW (We refer Appendix M.2 in the updated pdf for details discussion)\\nLPGW utilize the partial matching property of PGW while significantly improving computational efficiency. The larger the dataset, the more pronounced the computational advantage becomes. For instance, in the MNIST classification task with a training dataset of 4000 samples, LPGW trains the model in 5 minutes, whereas PGW takes approximately 15 days.\"}", "{\"summary\": \"The authors of this paper aim to improve on the partial Gromov-Wasserstein (PGW) method by creating linear partial Gromov-Wasserstein (LPGW) embedding. They begin the paper with a succinct coverage of optimal transport (OT). They further begin with covering the 2-Wasserstein distance, a common choice for OT-based methods, and proceed to lay out the fundamentals of the Gromov-Wasserstein (GW) distance. Next, the authors give a breakdown of the linear embedding technique, linear Gromov-Wasserstein (LGW). The mass constraint on OT methods motivated the creation of partial Gromov-Wasserstein, or PGW. Similar to the linearization of GW, the authors introduce the linearization of PGW through known techniques such as approximations; their work was further inspired by the need of having a faster solution for computing the distance between K different metric spaces. The authors claimed their algorithm can improve computing the PGW distances between K metric measure spaces from O(K^2) to O(K). While the authors have done a great job with the mathematical formulations, they have not provided a time-complexity analysis to back up their statement. Lastly, I believe the authors did not provide sufficient explanations for their experiments. For example, the simulation on the MNIST dataset, no clarifications are given on the details of training. For example, the authors state they will use a logistic regression for classification; this model is meant for binary classification and does not simply work on categorical cases. How exactly did they use logistic regression to predict classes when they range from 0-9? Furthermore, does this work with other models such as neural networks? It is always encouraged to be as specific as possible for reproducibility.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"4\", \"strengths\": \"The paper mathematically backs up their introduction of the linear partial Gromov-Wasserstein embedding algorithm. They also clearly demonstrate how previous work led to their contributions by breaking down and introducing each step along the way.\", \"weaknesses\": \"The authors claim that given K metric measure spaces, their algorithm can compute the distance between them in O(K) time. This claim was not supported with a time-complexity analysis and therefore remains unsubstantiated. It would strengthen the paper significantly to add a dedicated section, if even a small one, to prove the time-complexity (perhaps space-complexity if there's a tradeoff).\", \"questions\": \"Can you provide a time-complexity analysis to support your claim on a linear computation of the distances between K metric measure spaces?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}" ] }

[ "Kyeongkook Seo", "Dong-Jun Han", "Jaejun Yoo" ]

[ "Generative models", "Federated learning" ]

{ "note_id": [ "yZOi2teLxD", "yQqPB7Elkb", "xVm8vIIdyW", "q0H43E3nnY", "mCcdI8qtQj", "kcfSKN8DjB", "jmf78nTTzE", "jNPLNf4UsO", "jIOqhImbMx", "iKjfSkfigb", "gtZhzrG3IH", "gRFkN9bXvA", "eFdPJ1xLf5", "djLLh5zdne", "aXCKx1toEj", "Yjus4N39eD", "QaqwONQNYT", "PBmyBOuHk1", "MsCKJbPMQ1", "KQH2q7XJzB", "KNReEnkxnb", "I7HXgGFqea", "HjtdezziLu", "FrWJoZXoqe", "BIcGKFO1QR", "9NWYlCrFNH", "7bfvEUMQz5", "4whCFjdk3E", "46XKECH6dr" ], "note_type": [ "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "meta_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "decision", "official_review" ], "note_created": [ 1732121298321, 1732420759233, 1732122206484, 1732206659485, 1732122435398, 1732211606538, 1733058467358, 1730603386465, 1732420352191, 1734385318166, 1733058401565, 1732421095066, 1732204517548, 1732101731878, 1732204439089, 1732420986663, 1732102589493, 1732101775485, 1729567174524, 1732492013956, 1732209147386, 1732204386379, 1732101985797, 1732513985200, 1732159539981, 1730675395631, 1732122388959, 1737523704233, 1730384503453 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission5404/Authors" ], [ "ICLR.cc/2025/Conference/Submission5404/Authors" ], [ "ICLR.cc/2025/Conference/Submission5404/Authors" ], [ "ICLR.cc/2025/Conference/Submission5404/Reviewer_EU7U" ], [ "ICLR.cc/2025/Conference/Submission5404/Authors" ], [ "ICLR.cc/2025/Conference/Submission5404/Reviewer_EU7U" ], [ "ICLR.cc/2025/Conference/Submission5404/Authors" ], [ "ICLR.cc/2025/Conference/Submission5404/Reviewer_1FdJ" ], [ "ICLR.cc/2025/Conference/Submission5404/Authors" ], [ "ICLR.cc/2025/Conference/Submission5404/Area_Chair_15hk" ], [ "ICLR.cc/2025/Conference/Submission5404/Authors" ], [ "ICLR.cc/2025/Conference/Submission5404/Authors" ], [ "ICLR.cc/2025/Conference/Submission5404/Authors" ], [ "ICLR.cc/2025/Conference/Submission5404/Authors" ], [ "ICLR.cc/2025/Conference/Submission5404/Authors" ], [ "ICLR.cc/2025/Conference/Submission5404/Authors" ], [ "ICLR.cc/2025/Conference/Submission5404/Authors" ], [ "ICLR.cc/2025/Conference/Submission5404/Authors" ], [ "ICLR.cc/2025/Conference/Submission5404/Reviewer_EU7U" ], [ "ICLR.cc/2025/Conference/Submission5404/Authors" ], [ "ICLR.cc/2025/Conference/Submission5404/Reviewer_NWco" ], [ "ICLR.cc/2025/Conference/Submission5404/Authors" ], [ "ICLR.cc/2025/Conference/Submission5404/Authors" ], [ "ICLR.cc/2025/Conference/Submission5404/Authors" ], [ "ICLR.cc/2025/Conference/Submission5404/Reviewer_EU7U" ], [ "ICLR.cc/2025/Conference/Submission5404/Reviewer_Aeqb" ], [ "ICLR.cc/2025/Conference/Submission5404/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission5404/Reviewer_NWco" ] ], "structured_content_str": [ "{\"title\": \"[Response 2/2] Thank you for constructive feedbacks!\", \"comment\": \"**R3-5. Addressing PRISM's contributions [R3-W1, R3-Q6]**\\n\\nFirst, thanks for your positive feedback on our work. While we acknowledge that some components overlap with prior works, as you noted, the value of our work lies in effectively combining these methods and successfully addressing previously unsolved challenges in federated learning for generative models. In addition, we would like to highlight the proposed MADA framework, which demonstrates significant performance improvements. We believe this represents a noteworthy contribution to mitigating heterogeneity in federated learning environments.\\n\\n\\nRegarding your comment that MADA \\u201cseems to improve performance overall but not alleviate the effects of non-iid settings,\\u201d we are curious about what led you to this interpretation. Based on our results, MADA consistently improved performance across datasets, with the improvement being more pronounced under non-IID conditions compared to IID conditions. For example, the FID scores for Non-IID setups showed substantial improvement after applying MADA:\\n\\n* **MNIST**: Before: 49.6273 \\u2192 After: 34.2038\\n* **FMNIST**: Before: 83.0481 \\u2192 After: 67.1648\\n* **CelebA**: Before: 59.4877 \\u2192 After: 39.7997\", \"these_improvement_is_even_greater_than_those_observed_under_iid_conditions\": [\"**MNIST**: Before: 48.5636 \\u2192 After: 27.3017\", \"**FMNIST**: Before: 54.722 \\u2192 After: 46.1652\", \"**CelebA**: Before: 57.0573 \\u2192 After: 48.9983\", \"These results suggest that MADA is effective in addressing the challenges posed by non-IID settings. We would appreciate any additional insights you might have on this point, as understanding your perspective could help us further refine our work.\"]}", "{\"title\": \"[Response 1/2] Response to Reviewer NWco\", \"comment\": \"Thank you for carefully reviewing our paper and specifying your question again. We are trying to best address your concerns.\\n\\n**R3-6. Detailed explanation about MADA**\\n\\n**[Regarding learning rate]**\\n\\nGiven the local model $w_k^t = w_k^{t-1}-\\\\eta \\\\cdot \\\\Delta_k^t$ where $\\\\eta$ is constant learning rate and $\\\\Delta_k^t$ is local model update, the global model in FedAvg can be obtained as $w^t_{FedAvg} = w_k^{t-1}-\\\\eta \\\\cdot \\\\sum_{k}\\\\Delta_k^t$. In the case of MADA, the global model can be represented as $w^t_{MADA} = (1-\\\\lambda) w_k^{t-1}+\\\\lambda\\\\cdot w_k^t$, where $w_k^t = w_k^{t-1}-\\\\eta \\\\cdot \\\\Delta_k^t$.\", \"substituting_this_into_the_equation\": \"$w^t_{MADA}=(1-\\\\lambda) w_k^{t-1}+\\\\lambda\\\\cdot (w_k^{t-1}-\\\\eta \\\\cdot \\\\sum_{k}\\\\Delta_k^t) = w_k^{t-1}-\\\\eta \\\\cdot\\\\lambda \\\\sum_{k}\\\\Delta_k^t$, indicating that $\\\\lambda$ effectively modulates the learning rate for global updates. Based on this interpretation, MADA can be considered as an adaptive lr scheduler on the server side by estimating local divergence using similarity. Unlike traditional schedulers with fixed decay, MADA adapts dynamically to FL dynamics, either increasing or decreasing as needed. This flexibility eliminates the need for client-side learning rate scheduling to regularize local objectives.\\n\\n**[Decreasing $\\\\lambda$ in Non-IID]**\\n\\nThe intuition behind MADA is that model parameters can change drastically because the algorithm begins to minimize the loss function starting from the initialized model. The norm of gradients could be typically large as the model is far from the optimum. Hence, the models at consecutive epochs could differ significantly, leading to a large $\\\\lambda$ [1-2]. \\n\\nThe key question, as you pointed out, is whether this trend holds in heterogeneous scenarios. Your interpretation is certainly valid, a potential corner case exists that maintains $\\\\lambda$ large in heterogeneity scenarios. However, we believe that our interpretation to understanding $\\\\lambda$ is equally plausible and, based on the experimental results (see Figure 6 in the main paper), perhaps more natural.\\n\\nWe will elaborate on this question where client-drift occurs. Even if client-drift is biased toward a specific dominant client, the update difference for the dominant client will naturally decrease from the perspective of global support, while the difference with the local solutions of other clients will increase. This behavior is reflected in the aggregated model, and the difference between consecutive global models can partially capture this behavior (e.g., indicated by large $\\\\lambda$). Thus, as the training progresses and the model moves closer to the global optimum of the local objectives, the differences between consecutive global models become smaller, leading to a reduction in $\\\\lambda$. Otherwise, when client-drift does not occur and the model learns fairly across all clients (e.g., IID scenario), it will naturally converge without significant increases in differences.\\n\\nIn summary, MADA updates the global model in a way that avoids local divergence, gradually reducing global model differences even without learning rate decay. As shown in Figure 6 of our main paper, $\\\\lambda$ decreases sharply in the early stages of training due to significant local divergence but gradually converges over time. This result suggests that our understanding aligns with the behavior of $\\\\lambda$.\\n\\nIf there are any points of our response that you find unclear or if you require more detailed guidance, please let us know so that we can support your understanding.\\n\\n\\n[1] Cao, Yanzhao, Somak Das, and Hans\\u2010Werner van Wyk. \\\"Adaptive stochastic gradient descent for optimal control of parabolic equations with random parameters.\\\" Numerical Methods for Partial Differential Equations 38.6 (2022): 2104-2122.\\n\\n[2] Chatterjee, Sourav. \\\"Convergence of gradient descent for deep neural networks.\\\" arXiv preprint arXiv:2203.16462 (2022).\\n\\n...\\n\\nPlease refer the remaining responses in the [Respones 2/2].\"}", "{\"title\": \"[Response 1/3] Thank you for constructive feedbacks!\", \"comment\": \"**R4-1. Additional experiments on different random initialization or seed [R4-Q1]**\\n\\nWe conducted additional experiments across various seed values to provide further performance insights. As can be seen in the table below, PRISM shows robust performance regardless of specific initialization value or random seed with identical hyperparameters.\\n\\n| MNIST, IID, DP | FID | P&R | D&C |\\n|:---:|:---:|:---:|:---:|\\n| PRISM in paper (seed30) | 27.3017 | 0.4377 / 0.5576 | 0.1738 / 0.1982 |\\n| PRISM w/ seed123 | 26.8707 | 0.4342 / 0.5399 | 0.174 / 0.2047 |\\n| PRISM w/ seed1234 | 26.9664 | 0.4417 / 0.5842 | 0.1785 / 0.1997 |\\n\\n| MNIST, Non-IID, DP | FID | P&R | D&C |\\n|:---:|:---:|:---:|:---:|\\n| PRISM in paper (seed30) | 34.2038 | 0.4386 / 0.4236 | 0.1734 / 0.1597 |\\n| PRISM w/ seed123 | 34.0586 | 0.4372 / 0.3306 | 0.1732 / 0.1618 |\\n| PRISM w/ seed1234 | 32.9423 | 0.3921 / 0.4182 | 0.1496 / 0.1587 |\\n\\n**R4-2. Clarification on the comments: \\u201cFigures 2/3 for PRISM and baselines look problematic.\\u201d & \\u201cCIFAR10 and CelebA are not complex datasets\\\"[R4-W2, R4-W3]**\\n\\nWe understand your concern that demonstrating performance on simpler datasets, such as MNIST, may appear less impactful as an empirical result. However, it is important to note that the field of federated generative models remains far behind the current trend of using more complex and powerful models, such as diffusion models, and large-scale datasets. This gap arises from the inherent difficulties in adapting these advanced models to FL setups. \\n\\nExisting works have struggled to achieve stable performance in challenging setups, such as Non-IID and privacy-preserving cases, even with MNIST-level datasets. Even for centralized differential privacy generative models [1-2],training generative models is highly challenging and often unstable. Due to these obstacles, prior studies have primarily focused on simpler datasets, such as MNIST, and have been largely restricted to GANs. However, even under these conditions, many of these works failed to achieve desirable results. We summarize the FID scores and experimental setups reported in these works. \\n\\n| FID, IID case | MD-GAN [3] | UA-GAN [4] | Multi-FLGAN [5] |\\n|:---:|:---:|:---:|:---:|\\n| MNIST | 16.81 | 17.34 | 17.1 |\\n\\nIn contrast, PRISM is the first to consistently demonstrate stable performance under these conditions. \\n\\nThe purpose of our experiments was to highlight the limitations of prior approaches and showcase PRISM\\u2019s effectiveness in overcoming these challenges. Nevertheless, we acknowledge your doubts and have included additional experimental results below to further support our claims.\\n\\nStill, we also recognize the importance of assessing performance on high-resolution and large-scale datasets. Considering computational resources and the rebuttal period, we conducted experiments on the CelebA 128x128 and CIFAR100 datasets under Non-IID conditions without considering privacy. The results are summarized in the table below, with qualitative results provided in Appendix.I. Please note that PRISM is the first to achieve this level of performance under the current setting, whereas existing baselines have struggled even on MNIST-level benchmarks.\\n\\n| CelebA 128x128 | FID | P&R | D&C |\\n|:---:|:---:|:---:|:---:|\\n| PRISM | 40.2927 | 0.7738 / 0.006 | 1.00207 / 0.348 |\\n\\n| CIFAR100 | FID | P&R | D&C |\\n|:---:|:---:|:---:|:---:|\\n| PRISM | 74.1609 | 0.6655 / 0.0719 | 0.602 / 0.3121 |\\n\\n\\n[1] Dockhorn, Tim, et al. \\\"Differentially private diffusion models.\\\" arXiv preprint arXiv:2210.09929 (2022).\\n\\n[2] Jiang, Zepeng, Weiwei Ni, and Yifan Zhang. \\\"PATE-TripleGAN: Privacy-Preserving Image Synthesis with Gaussian Differential Privacy.\\\" arXiv preprint arXiv:2404.12730 (2024).\\n\\n[3] Hardy, Corentin, Erwan Le Merrer, and Bruno Sericola. \\\"Md-gan: Multi-discriminator generative adversarial networks for distributed datasets.\\\" 2019 IEEE international parallel and distributed processing symposium (IPDPS). IEEE, 2019.\\n\\n[4] Zhang, Yikai, et al. \\\"Training federated GANs with theoretical guarantees: A universal aggregation approach.\\\" arXiv preprint arXiv:2102.04655 (2021).\\n\\n[5] Amalan, Akash, et al. \\\"Multi-flgans: multi-distributed adversarial networks for non-IID distribution.\\\" arXiv preprint arXiv:2206.12178 (2022).\\n\\n...\\n\\nPlease refer the remaining responses in the [Respones 2/3].\"}", "{\"comment\": \"Thanks a lot for your reply--it is very much appreciated. I also appreciate the revisions you are making.\\n\\n**[MNIST generation]** Just quickly, regarding the first point: I mentioned previously that \\\"my worry is that the frameworks in the paper cannot even generate MNIST digits properly.\\\" To clarify the \\\"ambiguity\\\" that you mentioned, this was my main concern, as to my understanding we would expect papers published in ICLR to be able to generate MNIST digits properly. In your reply, you clarified quite well that in the setting that is analyzed in PRISM, this was not a trivial task due to the additional complexities imposed by DP, FL, masking, etc. Given this context, and the additional experiments you have provided, I am willing to be convinced that your algorithm presents a reasonable contribution. Point well taken; thanks for providing the additional works. \\n\\n**[Learning Rate]** Thanks for clarifying this. I'll take your word that you'll conduct and add additional experiments for verifying the fairness of your experimental process. \\n\\nThanks for resolving my enquiries and comments about your work. I'll think about this some more and then revise my score accordingly.\"}", "{\"title\": \"[Response 3/3] Thank you for constructive feedbacks!\", \"comment\": [\"**R4-6. Difference between PRISM and prior works [R4-W1]**\", \"We acknowledge that there is some overlap with prior works. However, as R3 also noted, the value of our work lies in effectively combining these methods to address previously unsolved challenges in federated learning for generative models. PRISM demonstrates impressive performance in scenarios where existing baselines fail, successfully tackling many of the prevailing difficulties in this field. In addition, we believe MADA, which is introduced for the first time in our work, constitutes a meaningful contribution by significantly improving performance and providing a novel approach to tackling heterogeneity in federated learning environments. As observed in our experiments, MADA consistently improved performance across datasets, with the improvement being more pronounced under non-IID conditions compared to IID conditions. For example, the FID scores for Non-IID setups showed substantial improvement after applying MADA:\", \"**MNIST**: Before: 49.6273 \\u2192 After: 34.2038\", \"**FMNIST**: Before: 83.0481 \\u2192 After: 67.1648\", \"**CelebA**: Before: 59.4877 \\u2192 After: 39.7997\"], \"these_improvement_is_even_greater_than_those_observed_under_iid_conditions\": \"* **MNIST**: Before: 48.5636 \\u2192 After: 27.3017\\n* **FMNIST**: Before: 54.722 \\u2192 After: 46.1652\\n* **CelebA**: Before: 57.0573 \\u2192 After: 48.9983\\n\\nThese results suggest that MADA is effective in addressing the challenges posed by both IID and non-IID settings.\\n\\nWe hope the reviewer recognizes these contributions and the advancements they represent in this domain.\\n\\n**R4-7. Details in architecture and hyperparameters of PRISM [R4-W4]**\\n\\nThank you for your constructive comment. We used a ResNet-based decoder and empirically observed that setting the local epoch to 100 and learning rate to 0.1 effectively optimizes the local stochastic binary masks. All of the detailed instructions have been added to Appendix.C, highlighted in \\u201cgreen\\u201d.\"}", "{\"comment\": \"I have updated my score.\"}", "{\"title\": \"Official Comment to Reviewer 1FdJ. We are looking forward to your comment\", \"comment\": \"Dear Reviewer 1FdJ,\\n\\nWith the author-reviewer discussion period nearing its conclusion, we kindly request your feedback on whether all your concerns have been fully addressed. Should you have any additional questions or require further clarification, please do not hesitate to let us know. We look forward to hearing from you.\"}", "{\"summary\": \"The core idea of this manuscript is to search for the optimal random binary mask of a random network to identify sparse sub-networks with high-performance generative capabilities. By communicating the binary mask randomly, PRISM minimizes communication overhead. Combining maximum mean discrepancy (MMD) loss and a mask-aware dynamic moving average aggregation method (MADA) on the server side, PRISM achieves stable and robust generative capabilities by mitigating local divergence in federated learning scenarios. Moreover, due to its sparse nature, models generated by PRISM are lightweight and well-suited for environments like edge devices without requiring additional pruning or quantization.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. The PRISM framework demonstrates a degree of innovation; although techniques such as the strong lottery ticket hypothesis and MMD are existing methods, the authors have effectively incorporated them into the federated learning setting.\\n\\n2. In terms of privacy, PRISM alleviates privacy leakage to some extent by introducing an (\\u03f5, \\u03b4)-differential privacy mechanism.\", \"weaknesses\": \"1. Although PRISM performs well on small-scale datasets, the manuscript does not adequately discuss its scalability and performance on large-scale datasets and larger models.\\n\\n2. The search process for random binary masks may increase computational complexity, especially in large networks. It is recommended that the authors analyze the specific computational costs of this process.\", \"questions\": \"None\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Additional Response to Reviewer 1FdJ\", \"comment\": \"**Computation cost analysis with FLOPS**\\n\\nRegarding R2-2, we provide the results of computation cost comparison in the table below. GS-WGAN shows increased FLOPS due to spectral normalization, introduced to stabilize GAN training. In contrast, PRISM, despite performing the SLT process, demonstrates sufficiently low FLOPS.\\n\\n| DP-FedAvgGAN | GS-WGAN | PRISM |\\n|:---:|:---:|:---:|\\n| 0.002 G | 1.94 G | 0.34 G |\"}", "{\"metareview\": \"The paper proposes a federated masking method for generative models that reduces communication and memory costs as well as improves performance in non-iid settings. The authors have addressed the comments and questions from the reviewers. None of the reviewers against the acceptance of the paper. Please revise and incorporate all the discussions into the final version.\", \"additional_comments_on_reviewer_discussion\": \"NA\"}", "{\"title\": \"Official Comment to Reviewer Aeqb. We are looking forward to your comment\", \"comment\": \"Dear Reviewer Aeqb,\\n\\nWith the author-reviewer discussion period nearing its conclusion, we kindly request your feedback on whether all your concerns have been fully addressed. Should you have any additional questions or require further clarification, please do not hesitate to let us know. We look forward to hearing from you.\"}", "{\"title\": \"[Response 1/1] Appreciate the reviewer for acknowledging our efforts and raising the score.\", \"comment\": \"We appreciate the reviewer for acknowledging our efforts and raising the score.\\n\\nBelow, we present additional experiments conducted with different learning rates in response to the reviewer\\u2019s comment.\\n\\n**R4-12. Explore various learning rates for baselines**\\n\\nTo validate the fairness of our experiments, we explored various configurations of learning rates for the baselines. Since they utilize GANs, there are separate learning rates for the generator and discriminator, and we examined these combinations. In the tables below, the top row represents the discriminator\\u2019s lr / generator\\u2019s lr, while the subsequent row reports the FID scores. Due to the notorious instability of GAN training caused by the coupled learning dynamics between the generator and discriminator [1-3], we observed divergence in some settings. For GS-WGAN, the best performance was achieved with 1e-4 / 1e-5, but it still failed to generate MNIST images properly. We include these results on Appendix L in our main paper.\\n\\n| MNIST, Non-IID, DP | 1e-4, 1e-4 (default) | 1e-3, 1e-4 | 1e-5, 1e-4 | 1e-4, 1e-3 | 1e-4, 1e-5 |\\n|:---:|:---:|:---:|:---:|:---:|:---:|\\n| GS-WGAN | diverge | 108.0657 | diverge | diverge | 96.1892\\n\\n| MNIST, Non-IID, DP | 1e-3, 5e-4 (default) | 1e-3, 5e-3 | 1e-2, 5e-4 | 1e-4, 5e-4 | 1e-3, 1e-5 |\\n|:---:|:---:|:---:|:---:|:---:|:---:|\\n| DP-FedAvgGAN | 153.9325 | 206.759 | diverge | 177.3752 | 232.2336 |\\n\\n[1] Miyato, Takeru, et al. \\\"Spectral normalization for generative adversarial networks.\\\" arXiv preprint arXiv:1802.05957 (2018).\\n[2] Mescheder, Lars, Andreas Geiger, and Sebastian Nowozin. \\\"Which training methods for GANs do actually converge?.\\\" International conference on machine learning. PMLR, 2018.\\n[3] Mao, Xudong, et al. \\\"Least squares generative adversarial networks.\\\" Proceedings of the IEEE international conference on computer vision. 2017.\"}", "{\"title\": \"[Response 3/3] Thank you for your prompt feedback!\", \"comment\": \"**R4-9. On Hospitals and iPhones**\\n\\nThank you for your follow-up question. We agree that examples such as hospitals, financial institutions, and mobile devices are standard scenarios for motivating FL. In the context of our work, the primary focus is not on generating hospital or private user data directly, but on developing methodologies that enable stable and effective generative modeling under the constraints typically faced in these settings. The trained generative model can be used for various use-cases, as we describe below.\\n\\nFor instance, in healthcare, our approach could be used to collaboratively train generative models across multiple institutions without sharing sensitive medical data. These generative models could then support downstream tasks such as synthetic data generation for privacy-preserving research, imputation of missing patient data, data augmentation to improve model robustness in low-data scenarios, or providing priors for solving inverse problems (e.g., denoising, reconstruction, super-resolution, etc.). Similarly, in mobile devices, the ability to train generative models across distributed devices while preserving user privacy could enhance applications like personalized content generation or user-specific data augmentation.\\n\\nWhile federated generative models in general may address these challenges to some extent, PRISM offers several key advantages. Beyond achieving superior performance under challenging Non-IID and privacy-preserving conditions, PRISM also improves communication efficiency and security by transmitting only sparse masks instead of full weights or gradients. As discussed in Appendix B of the main paper, this approach significantly mitigates potential privacy risks associated with transmitting generative model weights, which could be exploited to reproduce private data.\\n\\nFor example, even if a malicious actor intercepts the sparse masks during communication, these masks alone are insufficient to reconstruct the original data unless the attacker also knows the exact model architecture and the randomly initialized weights used in each local device. This is a notable contrast to methods that transmit full weights or gradients, which pose higher risks of privacy leakage once the model details are exposed. \\n\\nFurthermore, PRISM enhances privacy protection even against the server, as the server only aggregates and returns masks without needing access to sensitive weight information. This differs fundamentally from approaches that require sending complete weights or gradients to the server, which inherently carry more privacy-sensitive information.\\n\\nWe hope this clarifies how our work aligns with the practical scenarios provided and how PRISM contributes uniquely to addressing these challenges.\\n\\n\\n**R4-10. Cross-Device verses Cross-Silo**\\nYour comment on cross-device scenarios has been invaluable in strengthening our paper. We include these experimental results and discussion in the Appendix K (highlighted in \\u201cgreen\\u201d) to help readers better understand the applicability of our method across various scenarios.\\n\\n\\n**R4-11. Learning Rate**\\n\\nWe did our best to ensure a fair comparison. To this end, we conducted experiments using the hyperparameter values reported in prior works or by using the official code provided by the authors. These hyperparameter settings are, in fact, those optimized by the original authors for the same datasets we used, representing their best possible configurations. For GS-WGAN, we even followed the process of warming up the discriminator through pretraining. \\nNevertheless, to further address your concern regarding the fairness of our experimental process, we plan to conduct additional experiments on various configurations, such as learning rate and local epochs. Detailed results will be shared as they become available to keep the reviewers informed.\"}", "{\"title\": \"[Response 1/2] Thank you for constructive feedbacks!\", \"comment\": \"**R1-1. Baseline performance of purely server-side SLT [R1-Q1]**\\n\\nWe understand that your inquiry is about the performance of PRISM under a centralized setting. In Table 6 of Appendix G in the main paper, we compared PRISM (Non-IID and privacy-preserving) with PRISM-vanilla (centralized setting). For your convenience, we present the same results in the table below.\\nAs shown in prior studies [1-3], centralized SGD typically outperforms federated settings, and our findings align with this trend.\\n| MNIST | FID | P&R | D&C |\\n|:---:|:---:|:---:|:---:|\\n| PRISM | 34.2038 | 0.4386 / 0.4236 | 0.1734 / 0.1597 |\\n| PRISM (vanilla) | 5.8238 | 0.6913 / 0.851 | 0.4689 / 0.679 |\\n\\n| FMNIST | FID | P&R | D&C |\\n|:---:|:---:|:---:|:---:|\\n| PRISM | 67.1648 | 0.4967 / 0.1231 | 0.2748 / 0.1681 |\\n| PRISM (vanilla) | 5.5004 | 0.6985 / 0.8534 | 0.4864 / 0.6965 |\\n\\n| CelebA | FID | P&R | D&C |\\n|:---:|:---:|:---:|:---:|\\n| PRISM | 39.7997 | 0.6294 / 0.0713 | 0.4565 / 0.2967 |\\n| PRISM (vanilla) | 19.1512 | 0.6621 / 0.3895 | 0.5348 / 0.5947 |\\n\\n**R1-2. Would the approach work for diffusion-based generative models? [R1-Q2]**\\n\\nYes, adopting diffusion models (DDPM) [9] within PRISM is conceptually feasible. However, directly applying the MMD loss-based approach used in PRISM, which requires generating a set of samples for each update, poses practical challenges due to the well-known limitations of DDPM, including slow sampling and convergence rates. These limitations lead to significant computational costs, making the approach practically infeasible.\\nNevertheless, given the importance of diffusion models in generative modeling, we explored a modified version of our framework. Specifically, we retained the core concept of identifying SLT but replaced the MMD loss with diffusion loss to guide score updates. We conducted preliminary experiments on the MNIST dataset under a Non-IID, privacy-free scenario, and intermediate results are provided in Appendix H. While this approach is inherently sub-optimal\\u2014since it updates scores based on the loss of individual samples rather than the sample statistics\\u2014it still demonstrates the ability to generate recognizable shapes. With further tuning, we anticipate more promising results. However, as this lies beyond the scope of our current paper, it remains an interesting direction for future research. It is worth emphasizing that, to the best of our knowledge, this is the first federated generative model capable of generating data at this level of quality in a Non-IID, privacy-preserving setting for datasets more complex than MNIST or FMNIST. We hope the reviewers recognize the significance of this achievement.\\n\\n...\\n\\nPlease refer the remaining responses in the [Respones 2/2].\"}", "{\"title\": \"[Response 2/3] Thank you for your prompt feedback!\", \"comment\": \"Additionally, we model an even more challenging scenario by incorporating privacy-preserving mechanisms, further amplifying the difficulty of the task. Under these stringent conditions, prior works have consistently struggled to achieve meaningful results, particularly on Non-IID distributions [8-11]. In contrast, PRISM demonstrates significantly better performance in these settings, as evidenced by the experimental results presented in the manuscript.\\nLastly, regarding the phrase \\\"state-of-the-art image generation on complex datasets\\\" in line 083, we acknowledge that it could be perceived as somewhat ambiguous. What we intended to convey is that PRISM achieves state-of-the-art performance within the constrained FL+DP setup. To avoid any misunderstanding, we will revise the manuscript to clarify that this claim pertains specifically to the FL+DP setup and the associated challenges.\\n\\nP.S. We do recognize that the back-and-forth nature of written exchanges can sometimes lead to misunderstandings. While we are confident in the FL benchmarking practices and the relevance of our experiments, we are happy to engage in further discussion with patience and an open mind to better address any remaining concerns or clarify potential misunderstandings. \\n\\nThank you again for your effort.\\n\\n[1] Jiang, Zepeng, Weiwei Ni, and Yifan Zhang. \\\"PATE-TripleGAN: Privacy-Preserving Image Synthesis with Gaussian Differential Privacy.\\\" arXiv preprint arXiv:2404.12730 (2024).\\n[2] Zhang, Yikai, et al. \\\"Training federated GANs with theoretical guarantees: A universal aggregation approach.\\\" arXiv preprint arXiv:2102.04655 (2021).\\n[3] Li, Wei, et al. \\\"Ifl-gan: Improved federated learning generative adversarial network with maximum mean discrepancy model aggregation.\\\" IEEE Transactions on Neural Networks and Learning Systems 34.12 (2022): 10502-10515.\\n[4] Xin, Bangzhou, et al. \\\"Private fl-gan: Differential privacy synthetic data generation based on federated learning.\\\" ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2020.\\n[5] Tang, Yitong. \\\"Adapted Weighted Aggregation in Federated Learning.\\\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 38. No. 21. 2024.\\n[6] Ji, Xinyuan, et al. \\\"FedFixer: Mitigating Heterogeneous Label Noise in Federated Learning.\\\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 38. No. 11. 2024.\\n[7] Rahimi, Mohammad Mahdi, et al. \\\"EvoFed: leveraging evolutionary strategies for communication-efficient federated learning.\\\" Advances in Neural Information Processing Systems 36 (2024).\\n[8] Chen, Dingfan, Tribhuvanesh Orekondy, and Mario Fritz. \\\"Gs-wgan: A gradient-sanitized approach for learning differentially private generators.\\\" Advances in Neural Information Processing Systems 33 (2020): 12673-12684.\\n[9] Augenstein, Sean, et al. \\\"Generative models for effective ML on private, decentralized datasets.\\\" arXiv preprint arXiv:1911.06679 (2019).\\n[10] Hardy, Corentin, Erwan Le Merrer, and Bruno Sericola. \\\"Md-gan: Multi-discriminator generative adversarial networks for distributed datasets.\\\" 2019 IEEE international parallel and distributed processing symposium (IPDPS). IEEE, 2019.\\n[11] Amalan, Akash, et al. \\\"Multi-flgans: multi-distributed adversarial networks for non-IID distribution.\\\" arXiv preprint arXiv:2206.12178 (2022).\\n\\n...\\n\\nPlease refer the following responses [Response 3/3]\"}", "{\"title\": \"[Response 2/2] Response to Reviewer NWco\", \"comment\": \"**R3-7. Communication cost during downlink process**\\n\\nWe agree that downlink cost is also an important consideration, and various studies have focused on compressing it. Thanks to your feedback, the paper has become much clearer. We truly appreciate your insightful comments.\\n\\n**R3-8. Detailed explanation about memory savings of frozen model**\\n\\nThank you for your question. We will do our best to address your confusion.\\nFirst, we would like to clarify the concept of ternary quantization [1], to help explain the intuition behind our memory-saving approach. Ternary quantization decomposes $w$ into a scale factor $\\\\alpha$ and signed factor $\\\\hat{w}$ to obtain ternary weights $\\\\lbrace-\\\\alpha, 0, \\\\alpha\\\\rbrace$.\\n\\nBuilding on this concept, PRISM can further compress the final model. After the FL procedure concludes, each client has a sparse network represented as $W=W_{init} \\\\odot M$. While the binary mask $M$ is a 1-bit array, storing $W_{init}$ can be burdensome on small devices, depending on the model size. PRISM addresses this issue using signed constant initialization, inspired by ternary quantization. Specifically, each layer is initialized as $w_l=\\\\lbrace+\\\\sigma_l, -\\\\sigma_l, \\u2026, +\\\\sigma_l\\\\rbrace\\\\in \\\\mathbb{R}^d$, where $\\\\sigma_l \\\\in \\\\mathbb{R}^1$ is a standard deviation from Kaiming Normal distribution. Note that $w_l$ can be decomposed into a signed array $sign_l=\\\\lbrace+1, -1, \\u2026, +1\\\\rbrace \\\\in \\\\mathbb{R}^d$ and scale factor $\\\\sigma_l \\\\in \\\\mathbb{R}^1$. Since 1-bit quantized weights require negligible storage, only the scale factors contribute to the memory usage, this results in the lightweight final model size (7.25MB) reported in Tables 1-3 of the main text.\\n\\n**R3-9. Non-IID settings regarding CelebA dataset.**\\n\\nWe would like to clarify that CelebA is a multi-label dataset, where each image is associated with multiple attributes (e.g., age, gender, etc.), making it challenging to create an extreme Non-IID setup. As explained in Section 5.2 of the main paper, CelebA dataset was divided into positive and negative partitions for a pivotal attribute (gender, in our case), with the number of partitions corresponding to the number of clients. Each client was assigned either the positive or negative subset, ensuring non-overlapping data for the selected attribute.\\n\\nTo further explore this, we conducted additional experiments where each client possessed the pivotal attribute in different proportions using dirichlet distribution. For example, client 1 has 60% male and 40% female, while client 2 has 20% male and 80% female. The results are reported in the table below. To the best of our knowledge, methods for splitting multi-label datasets into extreme Non-IID scenarios have not been explored. If you could share additional insights on this matter, we would be glad to conduct further experiments based on your suggestions. \\n\\n| CelebA, Non-IID, DP | FID | P&R | D&C |\\n|:---:|:---:|:---:|:---:|\\n| PRISM | 34.2038 | 0.4386 / 0.4236 | 0.1734 / 0.1597 |\\n\\n[1] Zhu, Chenzhuo, et al. \\\"Trained ternary quantization.\\\" arXiv preprint arXiv:1612.01064 (2016).\"}", "{\"title\": \"[Response 1/2] Thank you for constructive feedbacks!\", \"comment\": \"**R3-1. Further explanation of MADA**\\n\\n**(1) Is MADA derived from other literature, or is this proposed here for the first time? [R3-Q1]**\\n\\nTo the best of our knowledge, MADA, which determines the moving average ratio on the server side based on the similarity of the global model, is proposed here for the first time.\\n\\n**(2) Is MADA convergence due to learning rate decay? [R3-Q4]**\\n\\nFirst, we would like to clarify that we did not use learning rate decay throughout the FL process. Thus, the local model continues to update consistently using a constant learning rate. During the FL process, the decrease in $\\\\lambda$ is due to the convergence of the binary mask, which reduces the similarity between the two global models, rather than being a result of learning rate decay.\\n\\n**R3-2. Communication cost savings of PRISM [R3-Q2, R3-W2]**\\n\\nPRISM transmits a float-type score during downlink, which, as you noted, does not support efficient communication, unlike the binary mask transmission in the uplink. In a typical FL setup, the server possesses powerful computational capabilities, whereas the clients do not. As a result, several prior studies primarily focus on addressing the limited bandwidth of clients [1-3]. Accordingly, we have concentrated on the communication cost in the downlink. We will clarify this in our manuscript (Appendix.J, highlighted in \\u201clime\\u201d).\\n\\n**R3-3. Memory savings of frozen model [R3-Q3]**\\n\\nOur memory-saving approach is akin to ternary quantization [4], a technique that quantizes neural network weights to {1, 0, -1} through thresholding and projection, making the sign the critical factor in the process. Similar but distinct, PRISM's frozen weights are already initialized to signed constant using Kaiming normal distribution standard deviation, $\\\\sigma_l$. Consequently, after FL concludes, each client can achieve additional memory savings by storing not only the lightweight binary mask but also the frozen weights with the scaling factor {$\\\\sigma_1$, \\u2026, $\\\\sigma_l$} and a 1-bit value representing the sign. We hope our clarification resolves any confusion, but let us know if further explanation is needed. \\n\\n\\n**R3-4. Clarification on Non-IID Empirical Settings and Additional experiments with Dirichlet Non-IID split [R3-W3, R3-Q5]**\\n\\nAs suggested, we conducted additional experiments under Non-IID splitting by Dirichlet distribution ($\\\\alpha=0.005$). We report the detailed results in the table below. PRISM continues to outperform other baselines in both Non-IID with/without DP setups, demonstrating its robustness.\\n\\n| MNIST, Non-IID, DP | FID | P&R | D&C |\\n|:---:|:---:|:---:|:---:|\\n| DP-FedAvgGAN | 175.3729 | 0.0408 / 0.1982 | 0.0102 / 0.0048 |\\n| GS-WGAN | 128.4401 | 0.0851 / 0.0633 | 0.0196 / 0.0071 |\\n| PRISM | 58.7524 | 0.3088 / 0.201 | 0.1078 / 0.0788 |\\n\\n| MNIST, Non-IID, No-DP | FID | P&R | D&C |\\n|:---:|:---:|:---:|:---:|\\n| MD-GAN | 106.3468 | 0.4292 / 0.105 | 0.1643 / 0.0332 |\\n| PRISM | 31.6191 | 0.5871 / 0.36 | 0.2828 / 0.2328 |\\n\\nWe would also like to address the concern (potential misunderstanding) regarding the Precision and Recall (P&R) results in our original experiments. While FID scores, which are widely regarded as a reliable metric for generative models, exhibit clear trends under non-IID settings, interpreting P&R and D&C metrics requires caution due to their inherent characteristics. For instance, fidelity metrics like Precision can yield inflated values under mode collapse, where many generated samples concentrate around a single real sample. Similarly, outliers in the generated data can disproportionately impact P&R scores, leading to variability and potentially misleading results, as highlighted in prior studies [5]. These characteristics can explain why P&R and D&C results appear similar to or even exceed those of IID settings, despite being evaluated under a Non-IID scenario. \\n\\n...\\n\\nremaining responses will be posted soon.\\n\\n[1] Kim, Do-Yeon, et al. \\\"Achieving Lossless Gradient Sparsification via Mapping to Alternative Space in Federated Learning.\\\" Forty-first International Conference on Machine Learning.\\n\\n[2] Hu, Rui, Yuanxiong Guo, and Yanmin Gong. \\\"Federated learning with sparsified model perturbation: Improving accuracy under client-level differential privacy.\\\" IEEE Transactions on Mobile Computing (2023).\\n\\n[3] Yi, Liping, Wang Gang, and Liu Xiaoguang. \\\"QSFL: A two-level uplink communication optimization framework for federated learning.\\\" International Conference on Machine Learning. PMLR, 2022.\\n\\n[4] Liu, Dan, and Xue Liu. \\\"Ternary Quantization: A Survey.\\\" arXiv preprint arXiv:2303.01505 (2023).\\n\\n[5] Naeem, Muhammad Ferjad, et al. \\\"Reliable fidelity and diversity metrics for generative models.\\\" International Conference on Machine Learning. PMLR, 2020.\\n\\n...\\n\\nPlease refer the remaining responses in the [Respones 2/2].\"}", "{\"title\": \"[Response 2/2] Thank you for constructive feedbacks!\", \"comment\": \"**R1-3. Addressing the Concerns on Model and Dataset Complexity in Federated Learning\\n [R1-W1]**\\n\\nWe would like to emphasize the significant challenges of training generative models under a federated learning (FL) setup. The field of federated generative models remains far behind the current trend of using more complex and powerful models, such as diffusion models, and large-scale datasets. This gap arises from the inherent difficulties in FL setups. \\nEven for centralized differential privacy generative models [4-5], as also highlighted in the related work of our main script, training generative models is highly challenging and often unstable. Due to these obstacles, prior studies have primarily focused on simpler datasets, such as MNIST, and have been largely restricted to GANs. However, even under these conditions, many of these works failed to achieve desirable results. We summarize the FID scores and experimental setups reported in these works. \\n\\n\\n| FID, IID case | MD-GAN[6] | UA-GAN[7] | Multi-FLGAN[8] |\\n|:---:|:----------------------:|:--------------------:|:--------------------------:|\\n| MNIST | 16.81 | 17.34 | 17.1 |\\n\\nWhile adapting our approach to models like diffusion models is beyond the scope of this paper (as noted in our response to R1-2, Appendix H), we recognize the importance of assessing performance on high-resolution and large-scale datasets. \\nConsidering computational resources and the rebuttal period, we conducted experiments on the CelebA 128x128 and CIFAR100 datasets under Non-IID conditions without considering privacy. The results are summarized in the table below, with qualitative results provided in Appendix.I. Please note that PRISM is the first to achieve this level of performance under the current setting, whereas existing baselines have struggled even on MNIST-level benchmarks.\\n\\n| CelebA 128x128 | FID | P&R | D&C |\\n|:---------------------------:|:--------------:|:---------------:|:--------------------:|\\n| PRISM | 40.2927 | 0.7738 / 0.006 | 1.00207 / 0.348 |\\n\\n| CIFAR100 | FID | P&R | D&C |\\n|:---------------------------:|:--------------:|:---------------:|:---------------------:|\\n| PRISM | 74.1609 | 0.6655 / 0.0719 | 0.602 / 0.3121 |\\n\\n\\n[1] McMahan, Brendan, et al. \\\"Communication-efficient learning of deep networks from decentralized data.\\\" Artificial intelligence and statistics. PMLR, 2017.\\n\\n[2] Zhao, Yue, et al. \\\"Federated learning with non-iid data.\\\" arXiv preprint arXiv:1806.00582 (2018).\\n\\n[3] Hardy, Corentin, Erwan Le Merrer, and Bruno Sericola. \\\"Md-gan: Multi-discriminator generative adversarial networks for distributed datasets.\\\" 2019 IEEE international parallel and distributed processing symposium (IPDPS). IEEE, 2019.\\n\\n[4] Dockhorn, Tim, et al. \\\"Differentially private diffusion models.\\\" arXiv preprint arXiv:2210.09929 (2022).\\n\\n[5] Jiang, Zepeng, Weiwei Ni, and Yifan Zhang. \\\"PATE-TripleGAN: Privacy-Preserving Image Synthesis with Gaussian Differential Privacy.\\\" arXiv preprint arXiv:2404.12730 (2024).\\n\\n[6] Hardy, Corentin, Erwan Le Merrer, and Bruno Sericola. \\\"Md-gan: Multi-discriminator generative adversarial networks for distributed datasets.\\\" 2019 IEEE international parallel and distributed processing symposium (IPDPS). IEEE, 2019.\\n\\n[7] Zhang, Yikai, et al. \\\"Training federated GANs with theoretical guarantees: A universal aggregation approach.\\\" arXiv preprint arXiv:2102.04655 (2021).\\n\\n[8] Amalan, Akash, et al. \\\"Multi-flgans: multi-distributed adversarial networks for non-IID distribution.\\\" arXiv preprint arXiv:2206.12178 (2022).\\n\\n[9] Ho, Jonathan, Ajay Jain, and Pieter Abbeel. \\\"Denoising diffusion probabilistic models.\\\" Advances in neural information processing systems 33 (2020): 6840-6851.\"}", "{\"summary\": \"This paper uses the strong lottery ticket hypothesis to search for valuable weights within initializations. They optimize for utilitarian weight masking structures, which provides communication efficiency and memory benefits. They apply their approach to federated learning, specifically within the context of generative models. The authors argue that PRISM is robust for non-IID data and promotes resource efficiency.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"(1) Figure 1 is quite effective in compactly communicating the main ideas.\\n\\n(2) Originality and significance: This paper is an interesting read. I will reserve my opinion on these two factors until after further clarification from the authors for (1) in weaknesses. \\n\\n(3) Clarity and quality: It is clear that much effort has been put into composing the related works.\", \"weaknesses\": \"(1) Can the authors help me understand the differences between PRISM, and the works [1-2]? The central methodology of this paper appears to just be a combination of the two papers, up to some minor changes.\\n\\n(2) The MNIST digits generated in Figures 2/3 for PRISM and baselines look problematic. As these are tasks that we would expect even a simple VAE to be able to do, I am concerned about the significance of the contributions given in this paper. The fact that Figure 2 is in the IID scenario is also concerning, as one might expect contributable training paradigms to be able to generate MNIST digits properly. I am beginning to wonder if I am missing the point of the paper. Moreover, I do not agree with the assertion that CIFAR10 and CelebA are \\\"complex datasets\\\" which allows evaluation of \\\"state-of-the-art image generation\\\" in line 083. Figure 3 gives some sensible-looking digits. However, this is an empirical paper, without theoretical results. Additionally, the central idea of training for lottery subnetworks has already been examined in prior works, e.g., cited above. Given this, MNIST, CelebA, and CIFAR-10 are not difficult enough to yield publishable empirical results by themselves, and there are far more informative datasets out there (e.g., ImageNet subsets). \\n\\n(3) In Appendix G, the authors state that they do not include comparisons with centralized settings as they study settings in which \\\"distributed samples cannot be shared\\\". I think the concern here is that the evaluated datasets are so mature, and generative central training on them easily practicable on even modern Laptops, such that if the evaluated methods cannot even beat the central model, it is difficult to see the contribution being made. Again, my assessment would differ significantly if this was a theoretical paper, but this paper is mainly empirical. Are there any reasonable, practical settings in which we would be forced to train in such restrictive scenarios? Training generative image models on iPhones with user data? \\n\\n(4) Could the authors point me to where they discuss (or add in a discussion of):\\n\\n(a) What is the architecture specification used in the paper?\\n\\n(b) How are the hyperparameters tuned?\\n\\nRegarding (a), it would be very helpful to explicitly discuss the architecture details in the text so that it is self-contained (c.f., Table 2 in [1]). That way, readers do not have to go line-by-line through the code link to get this information. \\n\\n(5) The empirical setup and discussions seem to be mismatched. The prior discussions seem to focus on communication costs, memory, privacy, etc. This strongly suggests the resource-restricted cross-device setting in FL, which is especially the case given the simplicity of model architectures as well as datasets being used for generation. However, the experiments use 10 clients, which is clearly in the cross-silo setting; such datacenters are capable of training far more advanced models (e.g., [3-4]). In this case, that would be diffusion models.\", \"questions\": \"(1) Given that the initialization is frozen, how robust is PRISM to the random seed? Does this robustness hold across identical hyperparameters?\\n\\n(2) Could the authors help to contextualize an industrial or academic setting in which PRISM may be deployed in practice?\\n\\n(3) Are 10 clients really enough to evaluate cross-device \\\"non-IID\\\" performance? For example, the authors note that they split MNIST into 40 disjoint partitions based on class labels. I'm assuming this means that each digit has 4 partitions. Then, the 40 portions are sampled uniformly without replacement and 4 each are assigned to a single client. Am I correct to assume the participation percentage is 100% (which would imply cross-silo)? This does imitate label imbalance, but the number of the clients seem far too low. I would have assumed 2 participating clients out of, say, 200 mobile devices to evaluate for non-IID (e.g., [5]). Also, to my knowledge, a more robust and realistic non-IID partitioning strategy in converting centralized datasets to FL datasets is to use LDA partitioning (widely used in topic modeling).\\n\\n[1] Sparse Random Networks for Communication-Efficient Federated Learning (Isik et al., ICLR 2023) \\n\\n[2] Can We Find Strong Lottery Tickets in Generative Models? (Yeo et al., AAAI 2023)\\n\\n[3] DiLoCo: Distributed Low-Communication Training of Language Models (Douillard et al., Arxiv 2023)\\n\\n[4] Asynchronous Local-SGD Training for Language Modeling (Liu et al., ICML workshop 2024)\\n\\n[5] Efficient Adaptive Federated Optimization (Lee et al., ICML workshop 2024)\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Dear Reviewer Aeqb,\\n\\nWe sincerely thank you once again for the time and effort you have dedicated to reviewing this paper. Your invaluable feedback have significantly contributed to improving its quality.\\n\\nIn the revised version, we have polished the manuscript, incorporated additional experimental results, and addressed your concerns with detailed clarifications. As the deadline for the Author-Reviewer discussion is approaching, we would like to ensure that our responses have sufficiently addressed your feedback. If there are any remaining concerns or additional clarifications or experiments that you would like us to provide, please do not hesitate to let us know.\\n\\nThank you once again for your time and thoughtful input.\\n\\nBest regards,\\n\\nPaper 5404 Authors\"}", "{\"title\": \"Thanks for the Detailed Responses!\", \"comment\": \"Dear Authors,\\n\\nThank you for the detailed responses.\\n\\nI reaffirm that I believe that the paper effectively combines multiple methods in order to achieve impressive empirical performance.\\n\\n> Regarding your comment that MADA \\u201cseems to improve performance overall but not alleviate the effects of non-iid settings,\\u201d we are curious about what led you to this interpretation.\\n\\nWhat I was mentioning within my review was that you mention \\\"As the global rounds progress, \\u03bb gradually decreases, promoting stable convergence\\\" in Line 276. However, this does not seem to be a given in non-IID settings. There likely will be model drift that occurs that will cause the difference between consecutive global masks to not necessarily lead to a gradual reduction in $\\\\lambda$. As a result, I was inquiring how exactly $\\\\lambda$ decreases in these heterogeneous environments. My possible explanation was a reduction in learning rate. Does that make sense?\\n\\nOverall, I think MADA provides an interesting technique to fight model drift, however I am unsure what theoretical underpinnings it has. Could the authors expand on that?\\n\\n> In a typical FL setup, the server possesses powerful computational capabilities, whereas the clients do not.\\n\\nI understand what the authors are arguing, and I don't believe this is a huge drawback. However, it does require smaller devices to receive a very large model (which could be memory-infeasible). Furthermore, if there are hundreds of thousands of devices, the amount of communication could be quite difficult for the server. Again, I understand that it's better to have compression in one way than none. However, the downlink process is still important.\\n\\n> Memory savings of frozen model [R3-Q3]\\n\\nI may have to read through this paper. I am still a little confused. Can these models really still perform super well with only 1-bit?\\n\\n> As suggested, we conducted additional experiments under Non-IID splitting by Dirichlet distribution ($\\\\alpha=0.005$).\\n\\nI appreciate the added experimental results. Not to be a pain, but the issues I mentioned in the rebuttal namely pertained to the CelebA dataset, as the Non-IID performance seemed to improve on that dataset compared to the IID setting. This was confusing and made the results a little murkier.\\n\\nBest,\\n\\nReviewer NWco\"}", "{\"title\": \"[Response 1/3] Thank you for your prompt feedback!\", \"comment\": \"**R4-8. Response to Concerns About Dataset and Experimental Setup [R4-W3]**\\n\\nWe sincerely appreciate your feedback, which reflects a clear understanding of the fundamental principles of FL setups. However, we were unsure about certain aspects of your comment and would greatly appreciate any clarification regarding your concerns. You reference our statement that \\\"distributed samples cannot be shared\\\" to explain why centralized comparisons were excluded, but also suggest that our datasets are \\\"mature and simple\\\" enough for centralized training on modern laptops. This seems to conflict with the assumptions of the FL setup, where the full dataset is not accessible at each device, requiring collaboration among the clients to build a reliable model. Furthermore, your statement that \\\"if the evaluated methods cannot even beat the central model, it is difficult to see the contribution being made\\\" adds further ambiguity. Are you concerned that PRISM underperforms in centralized setups where the full dataset is accessible? If so, Appendix G demonstrates that PRISM performs well in such scenarios. Here, if a single laptop is assumed to have access to the entire dataset (e.g., full MNIST dataset), this scenario is considered as a centralized setting, unlike the FL scenario where each client has access to only a limited number of data samples.\\n\\nFirst, to clarify your question, \\\"Why do PRISM and the baselines in the paper fail to properly generate MNIST digits in Figures 2 and 3?\\\" the observed challenges arise not from limitations of PRISM itself but from the stringent constraints imposed by the FL+DP (Federated Learning with Differential Privacy) setup we model. While generating MNIST digits is indeed trivial in centralized setups\\u2014even on a single laptop\\u2014this task becomes highly challenging under FL+DP conditions [1-4]. For example, as shown in Figure 4 of [1], generating MNIST digits while adhering to privacy constraints remains a significant challenge, even in centralized settings.\\n\\nSecond, regarding the evaluated datasets, we would like to emphasize that MNIST, CIFAR-10, and similar benchmarks are commonly used in FL literature to simulate challenging and constrained scenarios with limited data at each client and non-IID data distribution [5-7]. While it is true that these datasets can be fully utilized on a laptop for centralized training, the FL paradigm assumes that data is inherently distributed across multiple devices, and full datasets cannot be shared. This distinction is key: FL setups explicitly model environments where no single device or server has access to the entire dataset, which aligns with real-world applications. Thus, using subsets of datasets such as MNIST and CIFAR-10 is a standard benchmarking practice to evaluate methods under these constraints.\\n\\n...\\n\\nPlease refer the following responses [Response 2/3]\"}", "{\"title\": \"[Response 1/1] Thank you for constructive feedbacks!\", \"comment\": \"**R2-1. Scalability and performance on large-scale datasets and larger model [R2-W1]**\\n\\nWe would like to emphasize that recent FL studies often lack discussions on high-resolution and large-scale datasets. As highlighted in the related work section of our main script [3-6] and in recent studies on privacy-preserving generative models [1-2], this reflects the inherent difficulty of training generative models in unstable FL setups. For this reason, prior works have predominantly focused on such experimental environments, with most achieving only MNIST-level generation.\\nNevertheless, we recognize the importance of addressing whether robust performance can be achieved on high-resolution and large-scale datasets. Considering computational resources and the rebuttal period, we conducted experiments on the CelebA 128x128 and CIFAR100 datasets under Non-IID local data distributions without privacy considerations. Quantitative results are reported in the table below, and qualitative results are provided in Appendix I. It is noteworthy that existing baselines have failed not only in these settings but also on MNIST-level benchmarks. We are the first to achieve this level of performance under the current conditions.\\nTo further address the scalability of PRISM, we will update the main text with relevant discussions and respond accordingly once the content has been revised.\\n\\n| CelebA 128x128 | FID | P&R | D&C |\\n|:---------------------------:|:--------------:|:---------------:|:--------------------:|\\n| PRISM | 40.2927 | 0.7738 / 0.006 | 1.00207 / 0.348 |\\n\\n| CIFAR100 | FID | P&R | D&C |\\n|:---------------------------:|:--------------:|:---------------:|:---------------------:|\\n| PRISM | 74.1609 | 0.6655 / 0.0719 | 0.602 / 0.3121 |\\n\\n\\n\\n**R2-2. Computational cost of SLT process [R2-W2]**\\n\\nAs the number of model parameters increases, the required score parameters also increase linearly. However, in SLT, weight parameters are not learned through gradient descent. The only additional components are the sigmoid and Bernoulli processes introduced during the binary masking procedure. While this does incur some additional computational cost, it is negligible compared to the cost of gradient descent operations on GPUs.\\n\\nTo further substantiate this claim, we are conducting experiments to compare the FLOPS (floating-point operations per second) with and without the binary masking process. Detailed results will be provided as soon as the experiments are completed.\\n\\n[1] Dockhorn, Tim, et al. \\\"Differentially private diffusion models.\\\" arXiv preprint arXiv:2210.09929 (2022).\\n\\n[2] Jiang, Zepeng, Weiwei Ni, and Yifan Zhang. \\\"PATE-TripleGAN: Privacy-Preserving Image Synthesis with Gaussian Differential Privacy.\\\" arXiv preprint arXiv:2404.12730 (2024).\\n\\n[3] Hardy, Corentin, Erwan Le Merrer, and Bruno Sericola. \\\"Md-gan: Multi-discriminator generative adversarial networks for distributed datasets.\\\" 2019 IEEE international parallel and distributed processing symposium (IPDPS). IEEE, 2019.\\n\\n[4] Zhang, Yikai, et al. \\\"Training federated GANs with theoretical guarantees: A universal aggregation approach.\\\" arXiv preprint arXiv:2102.04655 (2021).\\n\\n[5] Amalan, Akash, et al. \\\"Multi-flgans: multi-distributed adversarial networks for non-IID distribution.\\\" arXiv preprint arXiv:2206.12178 (2022).\\n\\n[6] Li, Wei, et al. \\\"Ifl-gan: Improved federated learning generative adversarial network with maximum mean discrepancy model aggregation.\\\" IEEE Transactions on Neural Networks and Learning Systems 34.12 (2022): 10502-10515.\"}", "{\"title\": \"We submit the revised version of manuscript\", \"comment\": \"We have included the rebuttal discussions in the appendix, including R2-2, in the appendix and uploaded the revised version. If you have any additional concerns regarding these topics or suggestions for further experiments, please let us know.\"}", "{\"title\": \"Thanks for the additional experiments and rebuttals\", \"comment\": \"Thank you for your responses, and your additional experiments. I find your rebuttals quite convincing, and I'm especially encouraged by the extra experiment results you provided. But I would like to engage the authors in some more discussions during the rebuttal period.\\n\\n**[Companions to Centralized Settings]** Clearly, comparing the performance of DP-FL to centralized settings is unfair (as both frameworks have fundamentally different purposes), and FL is known to be upper bounded by a centralized barrier. My point was not that DP-FL needed to be as good as centralized settings in order for there to be a contribution. Rather, my worry is that the frameworks in the paper cannot even generate MNIST digits properly. As I said in my review, \\\"...the authors state that they do not include comparisons with centralized settings as they study settings in which \\\"distributed samples cannot be shared\\\". I think the concern here is that the evaluated datasets are so mature, and generative central training on them easily practicable on even modern Laptops, such that if the evaluated methods cannot even beat the central model, it is difficult to see the contribution being made. Again, my assessment would differ significantly if this was a theoretical paper, but this paper is mainly empirical.\\\" Similarly, I am not convinced that statements such as CIFAR10 and CelebA are \\\"complex datasets\\\" which allows evaluation of \\\"state-of-the-art image generation\\\" in line 083 are quite accurate, as I noted in my original review. If the authors could further discuss this point, that would be great.\\n\\n**[On Hospitals and iPhones]** Just quickly as a digression (not important): I find these examples quite interesting, but those examples, along with financial institutions, are standard examples for motivating FL. I'm still trying to contextualize your work in real-world settings. Could the authors clarify how their work can be useful in the context of the examples they provide? Would they be generating hospital or private user data?\\n\\n**[Cross-Device verses Cross-Silo]** From the response of the authors, I understand that the authors are currently in a compute-restricted environment. I believe that this should be taken into consideration. Thank you for the experiments for \\\"Number of Clients > 10\\\"; I would be satisfied if these can be included in the next draft of the paper. \\n\\n**[Learning Rate]** My original enquiry was trying to understand if sufficient hyperparameter tuning had been done for the other baselines to showcase their best performance, in order to be fair to all algorithms being compared. I'd like to ask how this was achieved.\"}", "{\"summary\": \"This paper introduces PRISM, a federated learning (FL) framework designed specifically for generative models to learn under heterogeneous data and lower the communication costs. The main idea leverages the strong lottery ticket (SLT) hypothesis by identifying an optimal sparse subnetwork with high generative performance through stochastic binary masking. The communication overhead is reduced by exchanging binary masks rather than full model weights. PRISM also includes a mask-aware dynamic moving average aggregation (MADA) to mitigate client drift, and maximum mean discrepancy (MMD) loss for stable training in generative tasks. Experiments demonstrate that PRISM outperforms existing federated generative models, such as GAN trained with DP-FedAvg on different image datasets.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": [\"The algorithm is a novel for learning the binary mask in SLT hypothesis under FL setting. The paper is well-written and the included MMD loss and MADA are well-justified approaches to handle data heterogeneity.\", \"The experiments considered different settings (e.g. IID v.s. non-IID and privacy v.s. no privacy) and the proposed PRISM showed considerable gain compared to the baseline approaches.\"], \"weaknesses\": \"There seems to be limited advances in GAN in the recent generative models literature so I am not sure if the experiment setting of GAN with simple image datasets (e.g. MNIST or CIFAR10) would still be a significant result for the community.\", \"questions\": [\"What is the baseline performance of purely server-side SLT? E.g. without any FL setup.\", \"Would the approach work for diffusion-based generative models?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"[Response 2/3] Thank you for constructive feedbacks!\", \"comment\": \"**R4-3: Response to Concerns About Comparisons with Centralized Settings and Practical Scenarios [R4-W3, R4-Q2]**\\n\\nWe respectfully disagree with the premise that comparisons with centralized settings are necessary to validate the contributions of our work. While centralized training on mature datasets may indeed be practicable on modern laptops, the focus of our study lies in addressing the unique challenges posed by federated learning (FL) environments, where data is inherently distributed and cannot be shared due to privacy constraints. These scenarios represent practical and increasingly relevant use cases, especially as the demand for privacy-preserving machine learning solutions grows.\\n\\nFor example, federated training setups are highly applicable in real-world scenarios such as healthcare, where sensitive medical image data is distributed across multiple institutions, and user data on mobile devices, such as iPhones, where privacy concerns prevent centralized aggregation. Our work addresses these restrictive scenarios by demonstrating stable and effective generative model training under challenging Non-IID and privacy-preserving conditions, where centralized baselines are not directly applicable.\\n\\nThus, rather than comparing against centralized models, our contribution lies in advancing federated generative modeling techniques to operate effectively in such distributed and restrictive environments. We hope this clarifies the motivation and significance of our work.\\n\\n**R4-4. Additional experiments**\\n\\n**(a) Number of clients>10[ R4-W5]**\\n\\nWe believe that the efficiency of PRISM makes it applicable across a wide range of scenarios. In fact, the settings used in our experiments were designed to just establish an intuitive FL environment rather than targeting specific scenarios (e.g., cross-silo or cross-device). Given the restricted resource budget during the rebuttal period, as an initial step to evaluate PRISM\\u2019s performance in resource-restricted cross-device settings [1], we conducted additional experiments involving 50 clients, where 10 clients (i.e., 0.2 ratio) participating in each communication round. The results, presented below, show that PRISM maintains robust performance in these cross-device environments, further showcasing its versatility and effectiveness across diverse federated learning scenarios. \\n\\n| MNIST, Non-IID, DP | FID | P&R | D&C |\\n|:---:|:---:|:---:|:---:|\\n| DP-FedAvgGAN | 118.3975 | 0.1095 / 0.3723 | 0.0301 / 0.0289 |\\n| GS-WGAN | 98.6563 | 0.8477 / 0.0359 | 0.2621 / 0.0105 |\\n| PRISM | 34.7157 | 0.4344 /0.3401 | 0.1692 / 0.1476 |\\n\\n| MNIST, Non-IID, No-DP | FID | P&R | D&C |\\n|:---:|:---:|:---:|:---:|\\n| MD-GAN | 15.4119 | 0.7305 / 0.359 | 0.5266 / 0.3803 |\\n| PRISM | 14.3168 | 0.7533 / 0.4757 | 0.5804 / 0.5293 |\\n\\n**(b) Non-IID [R4-Q3]**\\nIn the table below, we report the additional experimental results, Non-IID splitting by Dirichlet distribution with $\\\\alpha=0.005$. Again, we would like to highlight that 1) Other baselines struggle with the 4 shards Non-IID dataset, and 2) PRISM consistently outperforms other baselines under heterogeneity.\\n\\n| MNIST, Non-IID, DP | FID | P&R | D&C |\\n|:---:|:---:|:---:|:---:|\\n| DP-FedAvgGAN | 175.3729 | 0.0408 / 0.1982 | 0.0102 / 0.0048 |\\n| GS-WGAN | 128.4401 | 0.0851 / 0.0633 | 0.0196 / 0.0071 |\\n| PRISM | 58.7524 | 0.3088 / 0.201 | 0.1078 / 0.0788 |\\n\\n| MNIST, Non-IID, No-DP | FID | P&R | D&C |\\n|:---:|:---:|:---:|:---:|\\n| MD-GAN | 61.9427 | 0.4292 / 0.1639 | 0.1643 / 0.0747 |\\n| PRISM | 31.6191 | 0.5871 / 0.36 | 0.2828 / 0.2328 |\\n\\n[1] Isik, Berivan, et al. \\\"Sparse random networks for communication-efficient federated learning.\\\" arXiv preprint arXiv:2209.15328 (2022).\\n\\n...\\n\\nPlease refer the remaining responses in the [Respones 3/3].\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}", "{\"summary\": \"The authors propose a federated masking method for generative models that reduces communication and memory costs as well as improves performance in non-iid settings. A binary mask over the entire model is selected via the Strong Lottery Ticket (SLT) hypothesis, by which a maximum mean discrepancy (MMD) loss is used to provide gradient feedback to update weight-importance scores. Devices use these scores to generate (via Bernoulli distribution) a binary mask that is passed to the server for aggregation (thereby reducing communication costs). Finally, the server performs mask-aware dynamic moving average aggregation (MADA) whereby the server determines the drift of the global model and uses the value to determine how much of the new mask update should be incorporated. Experimental results show large increases in performance metrics while reducing communication and memory costs.\", \"soundness\": \"3\", \"presentation\": \"4\", \"contribution\": \"2\", \"strengths\": \"The strongest aspect are the empirical results of the proposed method. PRISM outperforms all of the baselines by a good margin while also being more efficient (in communication and memory). This is impressive.\\n\\nThe proposed MADA method is an interesting idea to alleviate model drift and improves performance empirically. As detailed in the questions, it seems to improve performance overall but not alleviate the effects of non-iid settings.\\n\\nThe paper writing and presentation is great. I could follow easily, the diagram is very helpful and the empirical results (both tables and figures) are presented cleanly. Great job!\", \"weaknesses\": \"I do not believe that the PRISM method itself is especially novel or specifically tailored for generative models as the authors market it as. The major performance driver behind PRISM seems to be the idea of training a sparse binary mask. This is shown to be quite effective in FedMask (Li et a. 2021), and it makes sense why it would be effective on GANs as well. However, this seems to be a simple application of FedMask coupled with the MMD loss (used in GAN training) and the WADA update to GANs. In fact, outside of using MMD loss, I do not see how PRISM is tailored towards generative models. The entire process could be applied to any other ML models if the MMD loss, which has already been proposed before, is substituted out.\\n\\nCommunication costs only seem to be saved during the uplink process, when only the binary mask is sent to the server. However, the downlink communication cost remains the same since the learnable weight-importance scores must be sent down to all devices. This is not mentioned by the authors and thus does not tackle all the communication issues. \\n\\nThe non-iid empirical setting seems slightly odd. Namely the authors partition \\\"datasets into 40 segments based on class labels and randomly assign four segments to each client\\\". Usually a Dirichlet split is the best way to simulate a non-iid split of labels amongst classes. As a result, it is odd that the performance of all algorithms stays around the same or *improves* for certain metrics. Precision and Recall improves for some of the other baselines which do not account for non-iidness. Generally non-iid settings should degrade performance in FL settings and that is not the case here for all baselines. This makes me feel that the partitioning was not non-iid enough.\", \"questions\": \"Is MADA derived from other literature, or is this proposed here for the first time?\\n\\nCould the authors further detail the communication savings of PRISM? It seems that there is no downlink communication savings?\\n\\nI was a bit confused about the model memory savings (Line 242). Could the authors clarify this? What I took away is that since model weights are frozen, and since they are initialized using the Kaiming Normal distribution, only the -/+ bit needs to be saved? Is this where the memory savings stem from (as shown in Tables 1/2 and Figure 4)?\", \"the_authors_mention_in_line_276\": \"\\\"As the global rounds progress, \\u03bb gradually decreases, promoting stable convergence.\\\" Is this due to model convergence arising from learning rate decay? What I mean is that, especially in non-iid settings, the global model generally converges once the learning rate decays and thus this should in turn decrease \\u03bb. Is that the correct characterization?\\n\\nCould the authors provide new non-iid experiments that showcase a more realistic setting?\\n\\nOverall, I feel that the application of sparse binary mask training to GANs is powerful and effective as shown in this paper. While each component is not especially novel, the combination suited for GANs is a novelty (albeit not a major one). However, the empirical performance is impressive and I believe in conjunction leads to a nice paper. The only disclaimer is that I am not especially well-versed within GAN or FL sparse binary mask literature.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"details_of_ethics_concerns\": \"N/A\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}" ] }

[ "Ruichen Xu", "Kexin Chen" ]

[ "differential privacy", "two-layer neural networks", "disparate impact" ]

{ "note_id": [ "vyieuvuRme" ], "note_type": [ "comment" ], "note_created": [ 1728802432167 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission10480/Authors" ] ], "structured_content_str": [ "{\"withdrawal_confirmation\": \"I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors.\"}" ] }

[ "Mihir Prabhudesai", "Russell Mendonca", "Zheyang Qin", "Katerina Fragkiadaki", "Deepak Pathak" ]

[ "Diffusion Model", "Text-to-Video Generation", "Generative Models" ]

{ "note_id": [ "yxPVIUyfUL", "wGBiW48skm", "thpYNYzXum", "se5Zc1kq2z", "se4GPvFZ0x", "rQubgAZe0N", "qCjO9kXgOo", "m8b6kHvcA0", "m11oeAYZtO", "lVJcpfzvRA", "l7TFvxtBjd", "fuQaHrhW0C", "eyfcD1qeaO", "ekUpXA1fdD", "cn70Llx9w6", "bU8FgoQKgg", "ZVFOIb7ftY", "YczrKirHk8", "ViBYG3KmKY", "VQQOewX6kr", "V5UcR769x1", "TqkQ16nZrM", "Tl8PK5XlGq", "PWmVGL9aUO", "Makea86431", "LO8oTLZ77t", "KqNyM5nhhP", "IVafUbcNr2", "GyxiIOou0S", "FaKSgEa3uj", "EbIWbjV1TX", "ELrgPoUqYf", "C2iCu8FfpH", "Bt6x2rOczr", "7luKUFDNyL", "7XFCAeGcw7", "3vvwc92DKo" ], "note_type": [ "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "decision", "official_comment", "official_review", "meta_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review" ], "note_created": [ 1732847151458, 1732828741643, 1730679822651, 1733313477864, 1732936845813, 1730701812176, 1732939022293, 1732827750877, 1737523734726, 1732854741330, 1730642467404, 1734289274432, 1733225033633, 1732991669994, 1732828510657, 1733184604756, 1733168930252, 1732829493212, 1732867300062, 1733183598719, 1732936821710, 1733052031995, 1733313467087, 1733010168544, 1733031615094, 1732828921147, 1732827124805, 1730617964117, 1733185561775, 1732828093814, 1730723177859, 1732853258238, 1732875139345, 1733180603257, 1733189928742, 1732865010088, 1730901801222 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission5941/Reviewer_wLqi" ], [ "ICLR.cc/2025/Conference/Submission5941/Authors" ], [ "ICLR.cc/2025/Conference/Submission5941/Reviewer_sDFV" ], [ "ICLR.cc/2025/Conference/Submission5941/Authors" ], [ "ICLR.cc/2025/Conference/Submission5941/Area_Chair_TYiK" ], [ "ICLR.cc/2025/Conference/Submission5941/Reviewer_wLqi" ], [ "ICLR.cc/2025/Conference/Submission5941/Authors" ], [ "ICLR.cc/2025/Conference/Submission5941/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission5941/Reviewer_sDFV" ], [ "ICLR.cc/2025/Conference/Submission5941/Reviewer_mSP2" ], [ "ICLR.cc/2025/Conference/Submission5941/Area_Chair_TYiK" ], [ "ICLR.cc/2025/Conference/Submission5941/Authors" ], [ "ICLR.cc/2025/Conference/Submission5941/Reviewer_sDFV" ], [ "ICLR.cc/2025/Conference/Submission5941/Authors" ], [ "ICLR.cc/2025/Conference/Submission5941/Authors" ], [ "ICLR.cc/2025/Conference/Submission5941/Authors" ], [ "ICLR.cc/2025/Conference/Submission5941/Authors" ], [ "ICLR.cc/2025/Conference/Submission5941/Reviewer_HJ2c" ], [ "ICLR.cc/2025/Conference/Submission5941/Authors" ], [ "ICLR.cc/2025/Conference/Submission5941/Area_Chair_TYiK" ], [ "ICLR.cc/2025/Conference/Submission5941/Reviewer_HJ2c" ], [ "ICLR.cc/2025/Conference/Submission5941/Authors" ], [ "ICLR.cc/2025/Conference/Submission5941/Reviewer_tG9V" ], [ "ICLR.cc/2025/Conference/Submission5941/Authors" ], [ "ICLR.cc/2025/Conference/Submission5941/Authors" ], [ "ICLR.cc/2025/Conference/Submission5941/Authors" ], [ "ICLR.cc/2025/Conference/Submission5941/Reviewer_tG9V" ], [ "ICLR.cc/2025/Conference/Submission5941/Reviewer_sDFV" ], [ "ICLR.cc/2025/Conference/Submission5941/Authors" ], [ "ICLR.cc/2025/Conference/Submission5941/Reviewer_HJ2c" ], [ "ICLR.cc/2025/Conference/Submission5941/Reviewer_sDFV" ], [ "ICLR.cc/2025/Conference/Submission5941/Reviewer_mSP2" ], [ "ICLR.cc/2025/Conference/Submission5941/Reviewer_sDFV" ], [ "ICLR.cc/2025/Conference/Submission5941/Authors" ], [ "ICLR.cc/2025/Conference/Submission5941/Authors" ], [ "ICLR.cc/2025/Conference/Submission5941/Reviewer_bKeQ" ] ], "structured_content_str": [ "{\"comment\": \"I would like to express my gratitude to the authors for their efforts in addressing the feedback thoroughly. Their clarifications and additional experiments adequately resolved my initial concerns. The detailed explanations strengthened the validity of the methodology and results. The additional evidence presented aligns well with the claims made in the paper, enhancing its overall impact and clarity. As a result, the score has been revised from 6 to 8.\"}", "{\"comment\": \"**Q 4.1: Lack of novelty.**\\n\\nPlease refer to the global answer.\\n\\n**Q 4.2: Unfair comparison between an on-policy strategy and off-policy strategies.**\\n\\nWe agree that VADER is more on-policy compared to DDPO and DPO. However we do not think this is the reason why VADER is more sample-efficient. In-fact off-policy methods are generally considered to be more sample efficient than on-policy methods [1].\\n\\nTo further investigate this, we created an on-policy version of DPO and DDPO, we do this by taking a single gradient update per video sampled. In our experiments we don\\u2019t find the results to improve the sample efficiency of the baselines. We plot these results in the following link: https://vader-anonymous.github.io/#training-efficiency-comparison and in Figure 12 in the main paper. Overall we think VADER is more sample efficient, because it backpropagates dense feedback to the model weights, while DDPO or DPO backpropagate scalar feedback.\\n\\n\\n**Q 4.3: More methods should be considered for comparison.**\\n\\nThere are not many methods in video-diffusion alignment, therefore we adapt state-of-the-art methods in image diffusion alignment space to videos. Since the deadline we have also added other baselines such as DDPO-on policy, DPO-on policy and DOODL , the results for these methods can be found in https://vader-anonymous.github.io/#training-efficiency-comparison and https://vader-anonymous.github.io/#doodl-vader. \\n\\n\\n**Q 4.4: Experimental details on DPO and DDPO.**\\n\\nWe implement DDPO [2], by closely following the official code from https://github.com/kvablack/ddpo-pytorch. We adopt the same code structure, which computes the log probability for each ddim denoising step, for the video diffusion model. This log probability is used to compute policy gradients, following the method code. We use the PPO version of the method as opposed to only using policy gradient, since this is reported to give slightly better performance in their paper. We use the same LoRA and gradient checkpointing approach that is used in VADER for updating the video diffusion model\\n\\n\\nWe implement DiffusionDPO [3] based on the official code from https://github.com/SalesforceAIResearch/DiffusionDPO. We alternate between sampling from the diffusion model, and training the model via the DPO objective, similar to the process in DDPO. Samples are added to a replay buffer (since offline samples can be used for DPO training), and we use batches sampled from the buffer for training. Given pairs of video generations, we assign them as $V^w$ and $V^l$ based on rewards from the reward model, where $V^w$ obtains higher reward. We then use the same\\nloss function as the original paper, which increases the likelihood of the $V^w$ sample. We set \\u03b2, the KL penalty in DPO to be 5000 following the standard. Just like DDPO, the model is updated using same LoRA training and gradient checkpointing approach as in VADER.\\n\\nThanks for pointing this out, we have added the explanation in the paper.\\n\\n-----------\", \"references\": \"[1] Efficient Off-Policy Meta-Reinforcement Learning via Probabilistic Context Variables\\n\\n[2] Training Diffusion Models with Reinforcement Learning\\n\\n[3] Diffusion Model Alignment Using Direct Preference Optimization\\n\\n[4] End-to-End Diffusion Latent Optimization Improves Classifier Guidance\"}", "{\"summary\": \"The paper proposed a reward fine-tune method called VADER. By using dense gradient information tied to generated RGB pixels, VADER enables more efficient learning in complex spaces like video generation. Backpropagating gradients from reward models to the video diffusion model facilitates both compute and sample-efficient alignment. Results across various reward and video diffusion models demonstrate that this gradient-based approach learns more efficiently than previous gradient-free methods in terms of reward queries and computational resources.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. The paper proposes to use gradient information for critic model preference tuning.\\n2. The experiment results are good.\", \"weaknesses\": \"1. Lack of novelty, the paper is more likely to be a tech report rather than a paper. Directly backward the gradient is not novel for RL.\\n2. The paper proposes a on-policy strategy, while all the comparisons are off-policy strategies which is not fair. There are so many on-policy strategies[1] that perform better than off-policy strategies. \\n3. Missing experiment details. How do you use DPO/DDPO on your dataset since they need preference pairs for training. How do you create the preference pair?\\n\\n\\n[1] Yuan, Weizhe, et al. \\\"Self-Rewarding Language Models.\\\" Forty-first International Conference on Machine Learning.\", \"questions\": \"1. More methods should be considered for comparison.\\n2. More details of the experiments should be enclosed.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": [\"We sincerely thank the reviewers for their consistent engagement and thoughtful critique of our work. Below we summarize the experiments we ran since our first rebuttal response:\", \"We **compare VADER against T2V-Turbo and various other baselines**, on Vbench and EvalCrafter. We find VADER consistently outperforms the baselines on these benchmarks. Experiment Links - (https://vader-anonymous.github.io/#vbench-evaluation-standard-prompt-suite, https://vader-anonymous.github.io/#evalcrafter-evaluation)\", \"We **train a custom video reward model using the automated rewards from VBench**, essentially creating a differentiable distilled VBench reward model. We then use this reward model to finetune VADER and find that this improves the dynamicness of the generated videos on unseen prompts. Experiment Links - https://vader-anonymous.github.io/#vbench-distilled-reward-model\", \"We ran **qualitative visualization comparing the in-domain diversity between VADER and VideoCrafter** baselines. We do not find major reduction in diversity in VADER compared to VideoCrafter baseline. Experiment Link - https://vader-anonymous.github.io/Video_Diversity.html\", \"We will update these results in the revised version of our paper.\"]}", "{\"comment\": \"Dear Reviewer,\\n\\nThe authors have provided their responses. Could you please review them and share your feedback?\\n\\nThank you!\"}", "{\"summary\": \"The authors introduce an alignment tuning method for video generation models utilizing gradient backpropagation from reward models. This approach addresses a critical need for producing high-quality, aligned video content. By directly guiding the generation model through reward gradients, the method achieves notable sample and computational efficiency compared to gradient-free approaches.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": [\"The problem to be solved is clearly defined, and its importance is well-acknowledged.\", \"The use of reward model gradients is highly intuitive and well-explained. Additionally, as the proposed methodology is data-free, it is practically useful and less likely to inherit biases from datasets used in alignment fine-tuning.\", \"Experimental results show this method is computationally efficient compared to gradient-free approaches such as DDPO and DPO.\"], \"weaknesses\": \"**Clarity of Contribution:** It is unclear what the authors\\u2019 unique contributions are. Utilizing reward model gradients for alignment tuning has already been demonstrated in text-to-image generation by Prabhudesai et al. and Clark et al. The authors should clarify what specific advancements they are claiming in this area.\\n\\nAdditionally, the authors mention that there are significant memory overhead issues in video models when implementing these methods, yet a clear, step-by-step ablation study is needed to show how each component of VADER addresses this problem.\\n\\nPrabhudesai et al., \\\"Aligning text-to-image diffusion models with reward backpropagation.\\\" arXiv. 2023. \\nClark et al. \\\"Directly fine-tuning diffusion models on differentiable rewards.\\\" ICLR. 2024.\\n\\n**Objective of the Alignment Tuning:** The results shown in the DDPO and DPO papers use significantly larger reward query samples. In this paper, however, smaller-scale experiments were conducted to highlight the sample and computational efficiency of reward gradients. This discrepancy may have resulted in DDPO and DPO showing unusually poor results. Since the current alignment settings are data-free, with comparable test times, sample and computational efficiency are not as critical. Therefore, it would be beneficial to include comparisons with DPO and DDPO results over longer training times.\", \"questions\": \"**Comparison with Existing Guidance Methods:** Reward gradients are commonly used in guidance-based research within diffusion models, as seen in studies like DOODL and Universal Guidance. It would be beneficial to analyze and compare the proposed method with these approaches. If direct implementation of these methods is challenging, an alternative comparison could involve modifying the reward guidance objective to apply guidance through $\\\\nabla_{x_t} R$ during generation, similar to the approach taken in this paper.\\n\\n**Clarification in Experiments (Table 1):** It is unclear which video generation model serves as the baseline in Table 1. The experimental setup mentions the use of VideoCrafter, Open-Sora 1.2, and ModelScope. Is the baseline an average of these models, or is it based on one specific model? Further clarification on this would be helpful.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"**Q5.3: Lack of Textual description alignment**\\n\\nWe thank the reviewer for their prompt response. VADER improves the textual alignment score of VideoCrafter from 0.256 to 0.267 when evaluated via HPS score, when evaluated via human eval for text alignment, VADER's generations are preferred 72\\\\% of the times over the baseline. These numbers are averaged over 700 videos generated by both the model. We therefore think this is a more accurate metric to evaluate the text-alignment of VADER than evaluating specific videos shown in the paper. We never claim VADER has perfect video-text alignment, however we stand with the claim that on average it's considerably better than the videos generated by VideoCrafter or ModelScope in terms of text-alignment.\"}", "{\"comment\": \"**Q1.1 Novelty:**\\n\\nPlease refer to the global answer.\\n\\n**Q1.2 Image-based Reward function don\\u2019t adequately encapsulate the unique characteristics of video data:**\\n\\nWe agree that the image-based reward models do not capture the temporal dynamics. \\n\\nWe therefore, run experiments with the following video reward models:\\n\\ni) **Vi-CLIP** - since deadline we have included Vi-CLIP, which is a video-text model trained in a similar fashion to CLIP. We show that this improves our video-text alignment and overall temporal dynamics of VADER. We visualize the results in https://vader-anonymous.github.io/#aesthetic-and-viclip-reward. \\n\\nii) **V-JEPA**- we use the base V-JEPA model trained using SSL objective as our reward function. We show that masked autoencoding reward functions can improve the temporal consistency of the generated video. These results are shown in the https://vader-anonymous.github.io/#v-jepa-reward, and Figure 9 in the main paper. \\n\\niii) **VideoMAE** - we use VideoMAE fine-tuned for action classification as our reward model, Our results are described in Figure 10, and Table 1 in the paper.\\n\\nFurther we ran evaluations using **VBench** (https://arxiv.org/abs/2311.17982) and found that using a video based reward function such as ViCLIP indeed improves the dynamic degree of the generated videos. The results can be found in the following link and the table below: \\n\\n| Model | Subject Consistency | Background Consistency | Motion Smoothness | Dynamic Degree | Aesthetic Quality | Imaging Quality |\\n|---------------------------|----------------------|-------------------------|-------------------|----------------|-------------------|-----------------|\\n| VideoCrafter | 0.9544 | 0.9652 | 0.9688 | 0.5346 | 0.5752 | 0.6677 |\\n| VADER-Aesthetic and HPS | 0.9659 | 0.9713 | **0.9734** | 0.4741 | **0.6295** | **0.7145** |\\n| VADER-PickScore | **0.9668** | **0.9727** | 0.9726 | 0.3732 | 0.6094 | 0.6762 |\\n| VADER-ViCLIP and Aesthetic| 0.9564 | 0.9662 | 0.9714 | **0.5519** | 0.6008 | 0.6566 |\", \"as_can_be_seen_in_the_table_above\": \"the dynamic degree for VADER-ViCLIP is the highest amongst all the baselines. Dynamic degree in Vbench is calculated by checking the dynamicness in the 2D flow predicted by RAFT (https://arxiv.org/abs/2003.12039). However as can also be seen, VADER-ViCLIP doesn\\u2019t score high on other categories such as Image Quality, Aesthetic Quality.\\n\\n**To address the above concern:** We have trained our own video reward model by fine tuning ViCLIP. Our video reward model can better capture various dimensions of improvements such as Dynamic Degree, Image Quality, Subject Consistency and Motion Smoothness. We are currently conducting HumanEval using our video reward model.\\n\\nWe also find that temporal coherence and motion quality is not significantly hampered while using Image-based reward models such as PickScore or HPS when benchmarking using EvalCrafter (https://arxiv.org/abs/2310.11440), as shown in the Table below:\\n\\n| Model | Temporal Coherence | Motion Quality |\\n|---------------------------------|--------------------|----------------|\\n| VideoCrafter | 55.90 | 52.89 |\\n| VADER-Aesthetic and HPS | 59.65 | **55.46** |\\n| VADER-PickScore | **60.75** | 54.65 |\\n| VADER-Aesthetic and ViCLIP | 57.08 | 54.25 |\\n\\n\\nEvalCrafter uses Video Action classification models and flow prediction models such as RAFT to evaluate Motion Quality, further it uses warping error and semantic consistency to evaluate Temporal Coherence.\\n\\n**Q1.3 An error in Equation (231) and ordering missing in Equations.**\\n\\nThank you for pointing this out. We have better clarified the equation you mentioned and have updated the manuscript with numbered equations. References to equations in the text are also cross-checked and updated accordingly.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"title\": \"Question about the experiment\\u3002\", \"comment\": \"Moreover, I understand that in Table 1, the authors utilized multiple reward models to simultaneously perform gradient backpropagation. However, when comparing with DPO, the authors claimed to use a replay buffer to store preference pairs scored by the reward models. Yet, different reward models might produce varying or even contradictory results for the same dataset. In other words, different reward models could have differing opinions on the winner-loser pairs. I am curious about how the authors handled the generation of offline preference data under the influence of multiple reward models?\"}", "{\"summary\": \"This paper introduces VADER, a method for fine-tuning video diffusion models using reward gradients to improve task-specific alignment. By utilizing pre-trained reward models as discriminators, VADER enhances video quality, text alignment, and temporal consistency. The approach employs memory optimization techniques to enable efficient training, even with limited resources. Experimental results demonstrate that VADER achieves strong performance across various video generation tasks.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. The authors propose VADER, which uses reward gradients to fine-tune video diffusion models, achieving efficient task adaptation.\\n2. This paper experiments with various reward models, making it suitable for multiple video generation tasks and achieving strong results in both subjective evaluations and quantitative metrics.\\n3. This paper employs several optimization techniques (such as truncated backprop) to enable operation in resource-limited environments.\", \"weaknesses\": \"1. The proposed VADER approach, which fine-tunes video diffusion models using reward gradients, optimizes network parameters with various reward models serving as discriminators. This enables improved adaptation to specific tasks. However, the use of reward models in generative model training is a well-explored concept. This paper just extends that approach within the context of diffusion models.\\n2. Since using reward models to backpropagate gradients requires the diffusion model to produce fully denoised outputs, all denoising steps must be executed, which places high demands on training resources. This might also lead to very small batch sizes. Although the authors employ several tricks to reduce resource usage, it raises the question of whether these adjustments impact the training outcomes. For instance, how much does backpropagation through only one timestep in the diffusion model affect the network parameters?\\n3. Some visual results in the paper still show misalignment between text and image content. For example, in Figure 7, the prompt \\\"A bear playing chess\\\" leads VADER to generate two bears.\", \"questions\": \"See Weakness\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"metareview\": \"This work presents a method to efficiently adapt large-scale video diffusion models to downstream tasks by leveraging pre-trained reward models. The results show the effectiveness of the method.\\n\\nThe paper is generally well-written. The results are convincing and demonstrate the effectiveness of the method.\", \"the_paper_has_mixed_review_scores\": \"866385.\\nAfter the rebuttal, internal discussions among reviewers and AC were initiated. Two critical issues below were identified, which led to the rejection decision.\\n\\n[limited novelty] reviewers sDFV, mSP2, and HJ2c raised their concerns about novelty, though some of them gave a score of 6. Using a discriminator to directly backward the gradient is widely used in many studies and lacks significant novelty. This work applies gradient backpropagation in video diffusion models. It is incremental work, despite good performance in several benchmark tests.\\n\\n[text video alignment is not addressed well using reinforcement learning] Models fine-tuned with the Aesthetic Reward Model fail to align with text, because this reward model is trained to evaluate aesthetics. As pointed out by reviewer mSP2: in Figure 8, the output does not depict a starry sky as described. Similarly, in the newly presented experimental results titled \\u201cDOODL vs. VADER,\\u201d prompts like 2 dogs and a whale, ocean adventure and Teddy bear and 3 real bears continue to exhibit this problem. \\n\\nThe authors are encouraged to incorporate this feedback into the revised version and resubmit the work to future conferences.\", \"additional_comments_on_reviewer_discussion\": \"The paper has mixed review scores: 866385.\\nAfter the rebuttal, two critical issues below were identified, which led to the rejection decision.\\n\\n[limited novelty] reviewers sDFV, mSP2, and HJ2c raised their concerns about novelty, though some of them gave a score of 6. Using a discriminator to directly backward the gradient is widely used in many studies and lacks significant novelty. This work applies gradient backpropagation in video diffusion models. It is incremental work, despite good performance in several benchmark tests.\\n\\n[text video alignment is not addressed well using reinforcement learning] Models fine-tuned with the Aesthetic Reward Model fail to align with text, because this reward model is trained to evaluate aesthetics. As pointed out by reviewer mSP2: in Figure 8, the output does not depict a starry sky as described. Similarly, in the newly presented experimental results titled \\u201cDOODL vs. VADER,\\u201d prompts like 2 dogs and a whale, ocean adventure and Teddy bear and 3 real bears continue to exhibit this problem.\"}", "{\"comment\": \"**Evaluation against T2V-Turbo while using Standard Prompt Suite of VBench**\\n\\nWe conduct an additional experiment while using the standard prompt suite of VBench. For the baselines, we copy-paste the numbers reported by T2V-Turbo in Table 1 of their paper. We follow the standard evaluation pipeline on VBench, for evaluating VADER (https://github.com/Vchitect/VBench?tab=readme-ov-file#evaluation-on-the-standard-prompt-suite-of-vbench). \\n\\nWe find that, VADER achieves the best Quality Score, amongst all the baselines. Quality Score is the weighted sum of the normalized score of each metric, as reported in VBench and T2V-Turbo.\", \"we_report_the_results_in_the_table_below_and_the_following_link\": \"https://vader-anonymous.github.io/#vbench-evaluation-standard-prompt-suite.\\n\\n\\n| Model | Subject Consistency | Background Consistency | Motion Smoothness | Dynamic Degree | Aesthetic Quality | Imaging Quality | Temporal Flickering | Quality Score |\\n|-----------------------------|---------------------|-------------------------|-------------------|----------------|-------------------|-----------------|---------------------|---------------|\\n| **VideoCrafter2** | 96.85 | 98.22 | 97.73 | 42.50 | 63.13 | 67.22 | 98.41 | 82.20 |\\n| **Pika** | 96.76 | **98.95** | 99.51 | 37.22 | 63.15 | 62.33 | **99.77** | 82.68 |\\n| **Gen-2** | **97.61** | 97.61 | **99.58** | 18.89 | 66.96 | 67.42 | 99.56 | 82.47 |\\n| **T2V Turbo (VC2)** | 96.28 | 97.02 | 97.34 | 49.17 | 63.04 | **72.49** | 97.48 | 82.57 |\\n| **VADER - Aesthetic and HPS** | 95.79 | 96.71 | 97.06 | **66.94** | **67.04** | 69.93 | 98.19 | **84.15** |\\n\\n\\n**Evaluation against T2V-Turbo on EvalCrafter benchmark**\\n\\nTo further evaluate temporal coherence and motion quality, we use the official Evalcrafter benchmark while using their standard prompts. We find that VADER outperforms T2V-Turbo on these metrics, as shown in the table below and the following link: https://vader-anonymous.github.io/#evalcrafter-evaluation\\n\\n| Model | Temporal Coherence | Motion Quality |\\n|--------------------------------|--------------------|----------------|\\n| VideoCrafter2 | 55.90 | 52.89 |\\n| T2V Turbo (4 Steps) | 57.10 | 54.93 |\\n| T2V Turbo (8 Steps) | 57.05 | 55.34 |\\n| VADER-Aesthetic and HPS | 59.65 | **55.46** |\\n| VADER-PickScore | **60.75** | 54.65 |\\n| VADER-Aesthetic and ViCLIP | 57.08 | 54.25 |\\n\\nWe will include these results in the revised version of our paper.\"}", "{\"title\": \"Feedback\", \"comment\": \"Thanks for the clarification in Table 1.\\n\\nFor the related work, first of all, T2V-Turbo has already demonstrated in the ablation study the impact of image model rewards and video model rewards on aspects such as aesthetics, dynamic quality, and motion smoothness, etc. , rather than merely using RL to improve inference speed. Furthermore, one concern I have with this paper is that T2V-Turbo was trained and evaluated using a larger-scale dataset and compared against numerous well-known methods, making its results more reliable than those presented in this paper. \\n\\nI agree with the authors that this work is a concurrent work with T2V-Turbo. But I think the authors need to carefully review more papers before claiming a very strong argument.\\n\\nI will raise my score to 5.\"}", "{\"comment\": \"**Q3.1: Clarity of Contribution**\\n\\nPlease refer to the global answer.\\n\\n\\n\\n**Q 3.2: Step-by-step ablation study is needed to show how each component of VADER addresses memory overhead problem**\\n\\nWe conduct a step-by-step ablation study to demonstrate how each component of VADER addresses the memory overhead problem. To study the contribution of each component to memory reduction, we conduct experiments on a single gpu, with batch size of 1. For this experiment, we offload the memory to the CPU main memory to prevent GPU out-of-memory error. The results are shown in the Table below. \\n\\n| Method | VRAM | System RAM | Total RAM |\\n|-------------------------------|--------|------------|-----------|\\n| LoRA + Mixed Precision | 12.1 GB | 264.2 GB | 276.3 GB |\\n| + Subsampling Frames | 12.1 GB | 216.8 GB | 228.9 GB |\\n| + Truncated Backpropagation | 12.1 GB | 57.3 GB | 69.4 GB |\\n| + Gradient Checkpointing | 12.1 GB | 20.4 GB | 32.5 GB |\\n\\n\\n\\n\\nWe find that the total memory is reduced by a significant amount, while using the above components.\\n\\n\\n\\n\\n\\n\\n**Q 3.3: Include comparisons with DPO and DDPO results over longer training times**\\n\\nThanks for the suggestion. We trained DPO and DDPO for longer. We found a meaningful improvement in DPO after long training, however we couldn\\u2019t see the same improvement in DDPO. We think this is because the number of gradient accumulation steps in DDPO\\u2019s implementation are very large i.e they scale linearly wrt the number of diffusion timesteps, thus the number of parameter updates done are relatively very little even with large amounts of compute. A better study on DDPO\\u2019s parameter update rate for video training, could improve its sample efficiency. The results of our experiments can be found here: https://vader-anonymous.github.io/#training-efficiency-comparison, and also in Figure 12 in the main paper.\\n\\n\\n\\n\\n**Q3.4: Comparison with Existing Guidance Methods**\\n\\nThanks for the suggestion, the guidance methods such as DOODL or Universal Guidance are applied for each example separately. VADER on the other hand updates the weights of the model, thus not requiring per sample adaptation at test time. In the following link: https://vader-anonymous.github.io/#doodl-vader, we compare VADER with DOODL, we find DOODL improves with more gradient update steps, however the improvement is still relatively less and scales linearly wrt number of examples we use for evaluation, thus making it difficult to be used in practice. \\n\\n\\n\\n**Q3.5 Clarification in Table 1.**\\nOur experiments in Table 1 are based on ModelScope, we have clarified this in the paper. We have also conducted further quantitative experiments with other base video models, which are shown in [VBench Evaluation](https://vader-anonymous.github.io/#vbench-evaluation), [Eval Crafter](https://vader-anonymous.github.io/#evalcrafter-evaluation) [diversity test](https://vader-anonymous.github.io/#diversity-test).\"}", "{\"comment\": \"Dear Reviewer,\\n\\nAs the rebuttal discussion is coming to an end, If there are any experiments we could run, or concerns that we can address that can help you better evaluate our work. Please let us know!\\n\\nSince our last discussion, we have done the following comparisons:\\n\\n**Diversity Test**\\n\\nWe visualize the generations from VADER and baseline to see if the diversity gets negatively affected due to it. Please find them here: https://vader-anonymous.github.io/Video_Diversity.html\\n\\n**Comparision againt T2V-Turbo**\\n\\nWe have conducted evaluation against T2V-Turbo, on VBench benchmark. T2V-Turbo showcases results on this benchmark in their paper. We find that VADER-Aesthetic+HPS outperforms both T2V-Turbo (4-step) and T2V-Turbo (8-step), while we use the same weighted average formula as used in their paper. Note that we never optimize for this benchmark, neither directly nor indirectly. Evaluation on the VBench benchmark was an afterthought after we had trained our models.\", \"please_find_the_results_in_the_table_below_or_on_the_following_link\": \"https://vader-anonymous.github.io/#vbench-evaluation\\n\\n\\n| Model | Subject Consistency | Background Consistency | Motion Smoothness | Dynamic Degree | Aesthetic Quality | Imaging Quality | Weighted Average |\\n|--------------------------------|---------------------|-------------------------|-------------------|----------------|-------------------|-----------------|------------------|\\n| VideoCrafter | 0.9544 | 0.9652 | 0.9688 | 0.5346 | 0.5752 | 0.6677 | 0.7997 |\\n| T2V Turbo (4 Steps) | 0.9639 | 0.9656 | 0.9562 | 0.4771 | 0.6183 | 0.7266 | 0.8126 |\\n| T2V Turbo (8 Steps) | **0.9735** | **0.9736** | 0.9572 | 0.3686 | 0.6265 | 0.7168 | 0.8058 |\\n| VADER-Aesthetic and HPS | 0.9659 | 0.9713 | **0.9734** | 0.4741 | **0.6295** | 0.7145 | **0.8167** |\\n| VADER-PickScore | 0.9668 | 0.9727 | 0.9726 | 0.3732 | 0.6094 | 0.6762 | 0.7971 |\\n| VADER-Aesthetic and ViCLIP | 0.9564 | 0.9662 | 0.9714 | **0.5519** | 0.6008 | 0.6566 | 0.8050 |\\n\\nThank you.\"}", "{\"comment\": \"**Comparision against T2V turbo**\\n\\nWe have conducted evaluation against T2V-Turbo, on VBench benchmark. T2V-Turbo showcases results on this benchmark in their paper. We find that VADER-Aesthetic+HPS outperforms both T2V-Turbo (4-step) and T2V-Turbo (8-step), while we use the same weighted average formula as used in their paper. Note that we never optimize for this benchmark, neither directly nor indirectly. Evaluation on the VBench benchmark was an afterthought after we had trained our models.\", \"please_find_the_results_in_the_table_below_or_on_the_following_link\": \"https://vader-anonymous.github.io/#vbench-evaluation\\n\\n\\n| Model | Subject Consistency | Background Consistency | Motion Smoothness | Dynamic Degree | Aesthetic Quality | Imaging Quality | Weighted Average |\\n|--------------------------------|---------------------|-------------------------|-------------------|----------------|-------------------|-----------------|------------------|\\n| VideoCrafter | 0.9544 | 0.9652 | 0.9688 | 0.5346 | 0.5752 | 0.6677 | 0.7997 |\\n| T2V Turbo (4 Steps) | 0.9639 | 0.9656 | 0.9562 | 0.4771 | 0.6183 | 0.7266 | 0.8126 |\\n| T2V Turbo (8 Steps) | **0.9735** | **0.9736** | 0.9572 | 0.3686 | 0.6265 | 0.7168 | 0.8058 |\\n| VADER-Aesthetic and HPS | 0.9659 | 0.9713 | **0.9734** | 0.4741 | **0.6295** | 0.7145 | **0.8167** |\\n| VADER-PickScore | 0.9668 | 0.9727 | 0.9726 | 0.3732 | 0.6094 | 0.6762 | 0.7971 |\\n| VADER-Aesthetic and ViCLIP | 0.9564 | 0.9662 | 0.9714 | **0.5519** | 0.6008 | 0.6566 | 0.8050 |\"}", "{\"comment\": \"**Q 6.1: Lacks a quantitative evaluation of the temporal coherence over V-JEPA and other models.**\\n\\nThanks for pointing this out, we have added temporal dynamics and temporal coherence evaluation over VADER trained using V-JEPA, ViCLIP, PickScore, HPS and Aesthetics reward functions on our webpage. We use VBench [1] and EvalCrafter [2] for conducting this eval. \\n\\nVBench Evaluation while using **image to video models*, while trained using **V-Jepa reward** model:\\n\\n| Model | Subject Consistency | Background Consistency | Motion Smoothness | Dynamic Degree | Aesthetic Quality | Imaging Quality |\\n|------------------------|---------------------|------------------------|-------------------|----------------|-------------------|-----------------|\\n| Stable Video Diffusion | 0.9042 | 0.9469 | 0.9634 | 0.8333 | 0.6782 | 0.6228 |\\n| VADER-V-JEPA | **0.9401** | **0.9551** | **0.9669** | 0.8333 | **0.6807** | **0.6384** |\\n\\n\\n\\nVBench Evaluation while using **text to video models**, while trained using **various reward models**: \\n\\n\\n| Model | Subject Consistency | Background Consistency | Motion Smoothness | Dynamic Degree | Aesthetic Quality | Imaging Quality |\\n|---------------------------|----------------------|-------------------------|-------------------|----------------|-------------------|-----------------|\\n| VideoCrafter | 0.9544 | 0.9652 | 0.9688 | 0.5346 | 0.5752 | 0.6677 |\\n| VADER-Aesthetic and HPS | 0.9659 | 0.9713 | **0.9734** | 0.4741 | **0.6295** | **0.7145** |\\n| VADER-PickScore | **0.9668** | **0.9727** | 0.9726 | 0.3732 | 0.6094 | 0.6762 |\\n| VADER-ViCLIP and Aesthetic| 0.9564 | 0.9662 | 0.9714 | **0.5519** | 0.6008 | 0.6566 |\\n\\n\\n**EvalCrafter Evaluation** using text-to-video models can be found in the Table below:\\n\\n| Model | Temporal Coherence | Motion Quality |\\n|---------------------------------|--------------------|----------------|\\n| VideoCrafter | 55.90 | 52.89 |\\n| VADER-Aesthetic and HPS | 59.65 | 55.46 |\\n| VADER-PickScore | **60.75** | 54.65 |\\n| VADER-Aesthetic and ViCLIP | 57.08 | 54.25 |\\n \\n\\nWe also visualize the generated videos after training on each of these reward models in the [project website](https://vader-anonymous.github.io/#aesthetic-and-viclip-reward). \\n\\nOverall we find that VADER beats the baseline over all these metrics. We find that video reward models such as ViCLIP help achieve higher temporal dynamics, while image-reward models such as PickScore help achieve higher temporal coherence and subject consistency. Lastly, we find that fine-tuning via V-JEPA improves both temporal coherence and motion quality of the original Stable Video Diffusion.\\n\\n**Q 6.2: Lacks a detailed examination of how each specific reward model influences alignment objectives.**\\n\\nThanks for the suggestion. We show a detailed comparison on how finetuning using a specific reward model affects the results on the other reward models: https://vader-anonymous.github.io/#reward-correlation . We find that there is a strong positive correlation between PickScore and HPS reward models because both of them contribute to text-video alignment while there is a somewhat negative correlation between Aesthetic Score and ViCLIP Score. We visualize the correlation matrix on the webpage and in Figure 11 of the paper.\\n\\n\\n**Q 6.3: How are the visualizations selected?**\\nThe examples were not selected entirely random. To show an entirely random selection we visualize a lot more videos here: https://vader-anonymous.github.io/Video_Gallery.html. We have also conducted a human eval study and our results are shown in Table 2.\\n\\n**Q 6.4 Why aren\\u2019t popular metrics, such as FID/FVD, used in the experiments?**\\nFVD or FID requires access to a real video dataset to compute the distance between generated and real distributions. VADER is data-free and does not rely on any real video data, so FID or FVD are not applicable.\\n\\n\\n[1] VBench: Comprehensive Benchmark Suite for Video Generative Models\\n[2] EvalCrafter: Benchmarking and Evaluating Large Video Generation Models\"}", "{\"title\": \"Feedback on Rebuttals\", \"comment\": [\"I thank authors for their response to my concerns and questions.\", \"**Novelty**: My understanding of authors contributions, in summary, is that they apply the standard techniques for alignment of image diffusion models to the video domain for the first time, and they analyze the behavior through thorough experiments. While I find this analysis valuable, I am still not fully convinced about the novelty of the findings.\", \"**Sample Diversity**: I find authors's argument on the expected reduction of diversity to a specific domain after alignment convincing. They also provide results to quantify the diversity of their method. However, it would be great to see the examples of generating multiple videos using the same prompt, to better assess the in-domain diversity of the method.\"]}", "{\"comment\": \"Dear Reviewer,\\n\\nJust a gentle reminder to check if you've had a chance to review our rebuttal. If there are any experiments or concerns that we could address and are still unresolved, that prevent you from accepting the paper, please let us know, as the deadline is 12 hours away.\\n\\nThank you!\"}", "{\"comment\": \"Dear Reviewer,\\n\\nThe authors have provided their responses. Could you please review them and share your feedback?\\n\\nThank you!\"}", "{\"title\": \"Final Feedback\", \"comment\": \"I appreciate authors' response. I do not have major concerns about the work, and I find the experiments sufficient. The only remaining concern is the limited novelty. Therefore, I raise my score to 6 (marginally above acceptance).\"}", "{\"comment\": \"**Q1.4) Beneficial if the authors developed and trained a reward model specifically tailored for video content**\\n\\n\\nTo investigate the benefits of training a custom video reward model, We fine-tuned PickScore reward model using custom 16,000 video-text + reward pairs. We constructed this dataset, by generating video-text pairs using VADER, we obtained the reward values using the automated metrics in VBench. \\n\\nAs the PickScore reward model can process only image-text pairs, we adapt it to process video-text pairs. We do so by using its image encoder to obtain 1D embeddings for each frame in the video denoted by $[f_0, f_1 .. f_N]$, where $N$ is the number of frames in the video. We then use its text encoder to obtain the text embeddings denoted as $c$. We then train a 5 layer and 8 head decoder-only transformer model that takes in N+1 embedding vectors (frames + text) as input and outputs a k-dimensional vector, where k represents the number of metrics in VBench such as Subject Consistency,Background Consistency etc. We train this model via supervised learning using MSE loss to predict the normalized scores of VBench metrics. We freeze the PickScore Image and Text encoders, and only train our custom transformer from scratch. We visualize the training curves for the reward models here -\", \"https\": \"//vader-anonymous.github.io/#vbench-distilled-reward-model .\\nWe observe that fine tuning using the VBench-Distill reward model consistently improves the dynamicness of the generated videos.\"}", "{\"comment\": \"I thank the authors for their response. They have addressed all my questions and concerns. I have no further questions.\"}", "{\"comment\": \"**Examples for in-domain diversity**\\n\\nThanks for the suggestion! Please find some examples for in-domain diversity here: https://vader-anonymous.github.io/Video_Diversity.html\"}", "{\"comment\": \"**Q 5.1 How does K=1 in VADER affect the outcomes?**\\n\\nWe run comparisons with K=10, by offloading the GPU memory to the CPU memory. We find different tradeoffs for using a higher value of K. \\n\\nIn the earlier steps of training, higher values of K results in more semantic level changes, while K=1 results in more fine-grained changes. Further, we find as we train longer, both the models start exhibiting semantic level changes. \\n\\nWe also find, training a model at K=10 is much slower than training it at K=1, for instance with 12 GPU hours, K=10 takes 50 gradient update steps, while K=1 takes 350 gradient update steps. We also find that K=1 is much easier to optimize, for instance with 200 gradient update steps, K=1 gets a reward of 5.49, while K=10 gets a reward of 5.26.\", \"this_result_is_shown_here\": \"https://vader-anonymous.github.io/#truncated-backpropagation-ablation\\n\\nOverall, we think the right value of K, might be system specific, higher values are preferred when the training system is not GPU VRAM bottlenecked.\\n\\n**Q 5.2 Misalignment between text and image content in Figure 7 in two bears example.**\\n\\nFigure 7 of the paper was generated by a model fine-tuned using the Aesthetic Reward Model. This reward model is trained to predict the aesthetic quality of pictures, so it does not contribute much to image-text alignment. Fine-tuning by utilizing the HPS or PickScore or ViCLIP Reward Models would help text alignment much more. More visualizations for video-text alignment are available at https://vader-anonymous.github.io/.\"}", "{\"comment\": [\"We thank the reviewers for their detailed feedback and thoughtful engagement with our work. We appreciate that several reviewers acknowledged the significance of aligning video diffusion models with reward gradients (HJ2c, wLqi, tG9V) . Reviewers also noted the clarity and quality of our presentation (HJ2c, tG9V), as well as the strong experimental results demonstrating the effectiveness of VADER across diverse reward models (HJ2c, wLqi, mSP2).\", \"A common concern raised by reviewers is novelty. We therefore state the contributions of our work:\", \"We are amongst the first few works to successfully align video diffusion models using reinforcement learning. We are the first to do entirely online reinforcement learning, that is we do not use any external datasets during training. We show results using a wide range of video models, reward models and benchmarks.\", \"We show that using reward gradients for aligning video diffusion models is promising, and can be very resourceful (specifically in the video setting which is very compute bottlenecked). VADER with VideoCrafter as the base model requires less than 24 GPU hours to train and can fit within 24GBs of GPU VRAM.\", \"Finally we perform a very dense evaluation for video reward alignment. We compare against various popular methods in the image alignment community such as Diffusion-DPO, DDPO etc. Further we show results across various benchmarks (VBench, EvalCrafter), while using various base video diffusion models such as text to video and image to video models. Lastly we show that existing image reward models and self-supervised trained video discriminative models such as V-JEPA, VideoMAE and ViCLIP can be successfully used as reward model.\", \"Since the submission deadline, we have added new experiments which can be summarized as follows:\", \"We have included results with the **ViCLIP video reward model**, which helps improve the temporal dynamics of our generated videos. (bKeQ, HJ2c, mSP2, tG9V). Experiment link - https://vader-anonymous.github.io/#aesthetic-and-viclip-reward\", \"We have added strong benchmarks such as **VBench** and **EvalCrafter**, where we compare temporal dynamics, consistency, subject or background correctness and various other metrics of the generated videos. (bKeQ, HJ2c, wLqi, tG9V). Experiment link - (https://vader-anonymous.github.io/#vbench-evaluation | https://vader-anonymous.github.io/#evalcrafter-evaluation)\", \"We have added new baselines, which include **DOODL** [1] and an **on-policy versions of DDPO and DPO** (wLqi, sDFV). Experiment link - (https://vader-anonymous.github.io/#training-efficiency-comparison | https://vader-anonymous.github.io/#doodl-vader)\", \"We further study how finetuning using a specific reward model affects the results on the other reward functions, thus studying the **correlation between different reward functions** (tG9V) Experiment link - https://vader-anonymous.github.io/#reward-correlation\", \"We also study the **diversity of generated videos in VADER** when finetune using various reward models. (HJ2C). Experiment link - https://vader-anonymous.github.io/#diversity-test\", \"We **ablate various memory reduction tricks** proposed in VADER along with the **number of truncation steps** used (wLqi, mSP2). Experiment links - (https://vader-anonymous.github.io/#memory-usage-comparison | https://vader-anonymous.github.io/#truncated-backpropagation-ablation)\", \"Lastly we provide **hundreds of qualitative visualization** of the generated videos from VADER and the baseline (tG9V). Experiment link - https://vader-anonymous.github.io/Video_Gallery.html\"], \"all_these_experiments_can_be_found_on_our_updated_appendix_in_the_paper__and_at_https\": \"//vader-anonymous.github.io/.\\n\\nWe also apologize to the reviewers for a relatively late response. As we had six reviewers, it took us relatively more time to conduct all the requested experiments.\\n\\nBelow, we address each reviewer's concerns.\"}", "{\"summary\": \"This work introduces VADER, a method for aligning video diffusion models using reward gradients. VADER repurposes off-the-shelf vision discriminative models as reward models to adapt video diffusion models more effectively. Additionally, the paper presents practical techniques to optimize memory usage, enabling efficient training of VADER on standard hardware.\", \"soundness\": \"4\", \"presentation\": \"4\", \"contribution\": \"4\", \"strengths\": \"Originality:\\nVADER is a novel approach to aligning video diffusion models using reward gradients rather than policy gradient methods. It creatively repurposes various pre-trained vision models as reward models, expanding their utility in video generation.\", \"quality\": \"The paper is methodologically rigorous, with strong experimental results demonstrating clear performance gains. Additionally, memory-optimizing techniques make VADER more accessible, broadening its potential user base.\", \"clarity\": \"The work is well-organized, with clear explanations and visualizations that effectively showcase the benefits of reward gradients and diverse reward models for alignment.\", \"significance\": \"VADER significantly advances practical video generation, making it more accessible and adaptable. This positions VADER as a valuable contribution to generative AI in video synthesis.\", \"weaknesses\": \"The paper lacks a quantitative evaluation of the temporal coherence achieved with the v-jepa reward model. While Figure 9 provides qualitative evidence of improvement, the analysis would be more robust with a quantitative assessment of temporal consistency. Adding such an experiment would offer a more comprehensive understanding of VADER's performance in maintaining coherence over time when trained on the v-jepa reward.\\n\\nWhile VADER incorporates multiple reward models, the paper lacks a detailed examination of how each specific reward model influences alignment objectives like temporal coherence, aesthetic quality, or text-video alignment. Additionally, it would be valuable to understand how optimizing for one reward, such as aesthetics, might impact the performance on other metrics, like temporal coherence.\", \"questions\": \"How are the visualizations selected?\\nAre the visual examples in the paper randomly chosen, or were they curated to highlight specific successes of VADER? Understanding the selection process would clarify whether the results are representative or potentially cherry-picked.\\n\\nWhy aren\\u2019t popular metrics, such as FID/FVD, used in the experiments?\\nFrechet Video Distance (FVD) is a commonly used metric for evaluating video quality in generative models, albeit with its own limitations and pitfalls. Including it would allow for a more standardized and comparable evaluation.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Some Questions\", \"comment\": \"You mentioned, \\\"T2V-Turbo showcases results on this benchmark in their paper,\\\" implying that the results from the paper were used as the reference in the table. However, I couldn't find any matching results in the T2V-Turbo paper that align with those reported in the table. Could you clarify which specific table or section in their paper you are referencing?\"}", "{\"comment\": \"**Q2.1 Lack novelty**\\n\\nPlease refer to the global answer.\\n\\n\\n**Q2.2 Diversity of the generated videos using the aligned model.**\\n\\nAligning a model for a specific use-case, very often results in a reduction of diversity and has been well-studied in prior works [1,2]. For instance, a pre-trained base model has the ability to generate videos of various types such as aesthetic, non-aesthetic, compressed, text-aligned or non text-aligned etc. Fine-tuning the model for alignment forces the model to generate a specific type of videos (eg: only aesthetic and text-aligned) thus reducing its overall diversity.\\n\\nWe find a similar result in the Table below, where we study the diversity of the models. \\n\\n| | VideoCrafter | VADER-PickScore Reward | VADER-Aesthetic and HPS Reward | VADER-Aesthetic and ViCLIP Reward |\\n|----------------------|--------------|-------------------------|---------------------------------|------------------------------------|\\n| **Average Variance** | **0.0037** | 0.0026 | 0.0023 | 0.0031 |\\n\\n\\nWe study diversity by generating multiple videos given a text prompt, we embed these videos via VideoMAE, and then calculate the variance of the embedding. We find that the overall diversity of the base model reduces as we align them using specific reward models. We also find that the highest diversity is achieved with VADER trained via ViCLIP+Aesthetic reward among all other reward models.\\n\\n**Q2.3: No video visualizations for video reward function**\\n\\nThanks for pointing this out. We have added visualizations for\", \"viclip_here___https\": \"//vader-anonymous.github.io/#aesthetic-and-viclip-reward\", \"v_jepa_here___https\": \"//vader-anonymous.github.io/#v-jepa-reward.\\n\\nTo further investigate, any sacrifice in terms of temporal variations, we study temporal dynamics using VBench (https://arxiv.org/abs/2311.17982) in the table below.\\n\\n| Model | Subject Consistency | Background Consistency | Motion Smoothness | Dynamic Degree | Aesthetic Quality | Imaging Quality |\\n|---------------------------|----------------------|-------------------------|-------------------|----------------|-------------------|-----------------|\\n| VideoCrafter | 0.9544 | 0.9652 | 0.9688 | 0.5346 | 0.5752 | 0.6677 |\\n| VADER-Aesthetic and HPS | 0.9659 | 0.9713 | **0.9734** | 0.4741 | **0.6295** | **0.7145** |\\n| VADER-PickScore | **0.9668** | **0.9727** | 0.9726 | 0.3732 | 0.6094 | 0.6762 |\\n| VADER-ViCLIP and Aesthetic| 0.9564 | 0.9662 | 0.9714 | **0.5519** | 0.6008 | 0.6566 |\\n\\nWe find that VADER-ViCLIP achieves the highest dynamic-degree amongst all the baselines. Dynamic degree in Vbench is calculated by checking the dynamicness in the 2D flow predicted by RAFT (https://arxiv.org/abs/2003.12039). \\n\\nWe also find that temporal coherence and motion quality is not hampered while using Image-based reward models such as PickScore or HPS when benchmarking using EvalCrafter (https://arxiv.org/abs/2310.11440), as shown in the Table below: \\n\\n| Model | Temporal Coherence | Motion Quality |\\n|---------------------------------|--------------------|----------------|\\n| VideoCrafter | 55.90 | 52.89 |\\n| VADER-Aesthetic and HPS | 59.65 | **55.46** |\\n| VADER-PickScore | **60.75** | 54.65 |\\n| VADER-Aesthetic and ViCLIP | 57.08 | 54.25 |\\n\\n\\nEvalCrafter calculates Motion Quality using Video Action classification models and flow prediction models such as RAFT, further it calculates Temporal Coherence using warping error and semantic consistency.\\n\\n\\n[1] - Understanding the Effects of RLHF on LLM Generalisation and Diversity https://arxiv.org/abs/2310.06452\\n\\n[2] - One fish, two fish, but not the whole sea: Alignment reduces language models' conceptual diversity https://arxiv.org/abs/2411.04427\"}", "{\"summary\": \"This paper addresses the problem of aligning pretrained video generation diffusion models to downstream tasks/domains using available reward functions without using fine-tuning datasets. Specifically, the authors propose updating the parameters of the diffusion model using the gradients of the target reward functions. This is motivated by the analysis provided in the paper (figure 3), showing that the feedback from reward gradients scale much more with video resolution compared to methods based on policy gradients. The authors apply their method to different reward functions, such as image aesthetics evaluation, image-text alignment, and video temporal consistency evaluation functions. The authors also incorporate some techniques to maintain efficiency when fine-tuning the retrained model. The proposed method is compared with multiple pretrained video generation models, as well as their aligned version using policy-gradient-based methods.\", \"soundness\": \"3\", \"presentation\": \"4\", \"contribution\": \"2\", \"strengths\": [\"The paper is very well-written and well-structured, making it easy to follow and understand.\", \"The provided analysis in Fig. 3, which shows the gap between the feedback from reward gradients and policy gradients for higher resolution videos, is valuable.\", \"The method is evaluated on multiple base video generators, showing consistent results.\", \"The proposed method shows better results in comparison to the policy-gradient-based baselines in terms of the evaluated metrics. This is also noticeable in the visual results.\", \"The proposed method has better generalizability on unseen text prompts compared to the baseline alignment methods.\"], \"weaknesses\": [\"**Novelty**: To my understanding, the proposed method is in essence a standard fine-tuning method with the objective of maximizing task-specific discriminative/reward functions. The proposed techniques for efficiency, including LoRA, truncated back propagation, and frame subsampling are also all standard and commonly used in different areas. The amount of technical novelty is not a major concern as long as the method has significant findings. However, the behavior shown in the paper, i.e. better alignment of the diffusion model when directly optimized to maximize the target reward function, is not very surprising to me.\", \"**Experiments**:\", \"In addition to regression in generalizability, another potential down-side of fine-tuning methods is the reduced diversity of the aligned model. Therefore, it is important to properly evaluate the diversity of the generated videos using the aligned model. For example, it would be interesting to see how diverse the videos are the same text prompt compared to the base model.\", \"Additionally, I noticed no video visualizations are provided for aligned models using video reward functions. For example, in the results provided in Fig. 9 does not show much temporal motion in the generated videos. This could also relate to the previous point, where the model could sacrifice temporal variations for more consistent frames.\"], \"questions\": \"Please see the concerns in the Weaknesses section. I am open to increasing my score, if the authors could clarify their novelty and contribution more, and address my concerns about the experiments.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Some comments\", \"comment\": \"I really appreciate the effort of the author. Unfortunately, to my knowledge, the author's paper is not the first to successfully align video diffusion models using reinforcement learning. The known T2V-Turbo[1] also utilized multiple reward models as feedback, was first released on May 29, and has been accepted by NeurIPS 2024. Although there are differences in implementation, I believe T2V-Turbo surpasses VADER in terms of large-scale experimental design and results. Therefore, I think the novelty of this paper is limited, and I maintain my original score.\\n\\n[1] Li, Jiachen, et al. \\\"T2V-Turbo: Breaking the Quality Bottleneck of Video Consistency Model with Mixed Reward Feedback.\\\" arXiv preprint arXiv:2405.18750 (2024).\"}", "{\"comment\": \"I thank the authors for their response.\\n\\nI agree with reviewer sDFV and HJ2c regarding their concerns about novelty. Using a discriminator to directly backward the gradient is widely used in many studies and lacks significant novelty. This work seems more like an application of gradient backpropagation in video diffusion models. While the authors' findings in the ablation studies on the number of truncated backpropagation steps are interesting, they are insufficient to support this paper's novelty.\\n\\nAdditionally, the authors emphasize their contribution as aligning video diffusion models using reinforcement learning. In my understanding, aligning with textual descriptions should be a crucial aspect of this. The authors noted that models fine-tuned with the Aesthetic Reward Model fail to align with text, because it is trained to predict the aesthetic quality of pictures. However, the results using the PickScore Reward still face this issue. For example, in Figure 8, the output does not depict a starry sky as described. Similarly, in the newly presented experimental results titled \\u201cDOODL vs. VADER,\\u201d prompts like *2 dogs and a whale, ocean adventure* and *Teddy bear and 3 real bears* continue to exhibit this problem.\\n\\nTherefore, I have decided to give a score of 5.\"}", "{\"title\": \"Feedback\", \"comment\": \"I truly appreciate the authors' efforts and the thorough comparison presented. I hope all the results will be shown in the revised version. Based on this, I am happy to raise my score to 6.\"}", "{\"comment\": \"Sorry for the confusion, we use the same evaluation pipeline/benchmark as them i.e VBench, which involves the same evaluation metrics such as Subject Consistency,Background Consistency, Motion Smoothness,Dynamic Degree, Aesthetic Quality\\tand Imaging Quality.\\n\\nHowever, To ensure consistency with our Human Eval and VBench results, we use the same set of 700 prompts officially released from the EvalCrafter benchmark (https://github.com/evalcrafter/EvalCrafter/tree/master/prompts, ). Note the same set of 700 EvalCrafter prompts were also used by T2V-Turbo for human-evaluation in Section 4.2.\", \"further_we_also_follow_the_vbench_custom_prompt_pipeline_for_evaluation\": \"https://github.com/Vchitect/VBench?tab=readme-ov-file#new-evaluate-your-own-videos, and use the officially released checkpoints from T2V-Turbo.\\n\\nWe will mention all of these details in the revised version.\"}", "{\"comment\": \"We thank the reviewer for their quick response!\\n\\n**Q 4.5: Related work - T2V-Turbo**\\n\\nThanks for sharing this work! We were not aware of the work of T2V-Turbo [1]. We plan to cite them and have also started working on comparing our results against them. \\n\\nHowever, we would like to point out some major differences between the two works:\\n\\n- The **goal of the T2V-Turbo is very different from goal of VADER**. Their goal is to improve the inference speed of T2V models, by distilling a teacher model (eg: VideoCrafter) into a few-step video consistency model. The goal of VADER is instead to align existing video-models to various downstream tasks. For instance, we show results on removing objects from the scene while using object detectors as reward models - https://vader-anonymous.github.io/#object-removal-reward . This could be a very useful downstream task for removing explicit content from video generation pipelines but this doesn't align with the goal of T2V-Turbo, as again their goal is to improve the speed of video generation. Further as both approaches have very different goals, this gives rise to very different evaluations studied in both the approaches, for instance we compare against other forms of RL techniques such as DDPO or DPO and we focus on training efficiency of VADER in terms of compute and samples used during finetuning. T2V-Turbo on the other hand focuses on inference time compute efficiency, and hasn't reported any numbers studying finetuning efficiency.\\n\\n- **T2V-Turbo requires having access to an external video-text training dataset**, VADER on the other hand is entirely data-free. Explicit requirement on an external training dataset, could potentially make it difficult to adapt to different base video models, as the base models considered might have never been trained on certain datasets.\\n\\n- Lastly our approach is not restricted to T2V models and can be easily adapted to I2V mdoels as we show here: https://vader-anonymous.github.io/#v-jepa-reward\\n\\nIrrespective of the major differences between the two works, we still think it's not fair to consider T2V-Turbo's work, when evaluating the novelty/originality of VADER. We consider T2V-Turbo to be a concurrent work due to the following reasons:\\n\\n- T2V-Turbo was arxiv-only and was not released as a NeurIPS paper until the abstract deadline for ICLR (Sept 27th)\\n- Our methods and initial results were submitted to conferences and our project webpage was available by end-February, which is way before the arxiving date for T2V-Turbo. Unfortunately we can't share exact details on this, while preserving the sanctity of double blind reviewing.\\n\\nHaving said that we do acknowledge that T2V-Turbo indeed does RL on video diffusion models. We have therefore **corrected** our previous statement stating that: \\\"we are the first work to successfully align video diffusion models using reinforcement learning\\\".\\n\\n**Q 4.6: Table 1 Clarification**\\n\\nThanks for pointing this out. We believe there is a misunderstanding here, In Table 1 we do not use multiple reward models at once, instead the results are shown while using each reward model in the column **independently** for fine-tuning. We consider each reward model to be a different downstream task, and show how effectively VADER can fit to a variety of downstreams tasks. We hope this clarifies the comparison with DPO. We will include this clarification in the paper.\"}", "{\"summary\": \"This paper proposes to utilize pre-trained reward models that are learned via preferences on top of powerful vision discriminative models to adapt video diffusion models. The results across a variety of reward models and video diffusion models showcase the effectiveness of the proposed approach.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"1\", \"strengths\": \"The authors propose the use of reward models aimed at enhancing the quality of generated videos. The experimental results demonstrate promising improvements in video quality, showcasing the effectiveness of the proposed approach.\", \"weaknesses\": \"1. Algorithm 1 represents a standard approach that utilizes reward feedback within a diffusion model framework, lacking significant innovation.\\n2. A critical challenge in applying reward feedback to diffusion models is the precise definition and training of the reward model. The manuscript employs certain image-based methods to establish the reward function for video generation; however, these methods may not adequately encapsulate the unique characteristics of video data. It would be beneficial if the authors developed and trained a reward model specifically tailored for video content, similar to advancements made in the field of image rewards [1], thereby contributing more meaningfully to the domain.\\n3. There appears to be an error in Equation (231). Could you please provide a detailed derivation to clarify this point?\\n4. The ordering of equations has been overlooked starting from Equation (3). Please ensure all equations are correctly sequenced for clarity and coherence.\\n[1] Imagereward: Learning and Evaluating Human Preferences for Text-to-Image Generation\", \"questions\": \"See weakness\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}" ] }

[ "Yuqi Luo", "Chenyang Song", "Xu Han", "Yingfa Chen", "Chaojun Xiao", "Zhiyuan Liu", "Maosong Sun" ]

[ "activation sparsity", "large language model" ]

{ "note_id": [ "zGN6RZUEyy", "yZUyh3lGVc", "xzM2RIJDpE", "wXDgRQRsZ0", "uhvE3Y85gd", "rscCd7UUIg", "qoTpixidFx", "k2tZ1AAYTl", "hx7C1KikHD", "gTvcHYSQzB", "gLvrYvWkDF", "g1VNyVfwid", "fs4N96UOVu", "a07VvgnXHS", "YvnBD9aV61", "XjaLVIHf1A", "X5FCAfASSE", "UnsM2YhMWU", "Ty5bbjIYZQ", "TCrDqQMcRp", "QRLBoC1yob", "PGPWZxlY4B", "O6XJt9Fdkr", "IGaF8MxrKI", "FYGr2vB1MQ", "EM6iBFYbrw", "C1JmgAdiz5", "BEUYOzsjsn", "6Nsaq86qxP", "2OfjA3CmaQ", "1OQdfFlrIK" ], "note_type": [ "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "decision", "official_comment", "meta_review", "official_comment" ], "note_created": [ 1733222751646, 1733112030748, 1732035667088, 1732035613560, 1732648609007, 1732192368707, 1732812670096, 1732549490194, 1732242970760, 1730732382712, 1733111965688, 1732785932356, 1733111913186, 1732243012797, 1733270339042, 1732242991400, 1730693432164, 1732550816503, 1733136616688, 1732549462342, 1729245446001, 1732549516909, 1732675845292, 1730660215744, 1732035835244, 1732035778433, 1732035892731, 1737524126476, 1732786150427, 1734855209856, 1732035922321 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission11471/Authors" ], [ "ICLR.cc/2025/Conference/Submission11471/Authors" ], [ "ICLR.cc/2025/Conference/Submission11471/Authors" ], [ "ICLR.cc/2025/Conference/Submission11471/Authors" ], [ "ICLR.cc/2025/Conference/Submission11471/Reviewer_MqGB" ], [ "ICLR.cc/2025/Conference/Submission11471/Reviewer_96Rd" ], [ "ICLR.cc/2025/Conference/Submission11471/Authors" ], [ "ICLR.cc/2025/Conference/Submission11471/Authors" ], [ "ICLR.cc/2025/Conference/Submission11471/Authors" ], [ "ICLR.cc/2025/Conference/Submission11471/Reviewer_6UUp" ], [ "ICLR.cc/2025/Conference/Submission11471/Authors" ], [ "ICLR.cc/2025/Conference/Submission11471/Authors" ], [ "ICLR.cc/2025/Conference/Submission11471/Authors" ], [ "ICLR.cc/2025/Conference/Submission11471/Authors" ], [ "ICLR.cc/2025/Conference/Submission11471/Area_Chair_7TiD" ], [ "ICLR.cc/2025/Conference/Submission11471/Authors" ], [ "ICLR.cc/2025/Conference/Submission11471/Reviewer_MqGB" ], [ "ICLR.cc/2025/Conference/Submission11471/Authors" ], [ "ICLR.cc/2025/Conference/Submission11471/Reviewer_96Rd" ], [ "ICLR.cc/2025/Conference/Submission11471/Authors" ], [ "ICLR.cc/2025/Conference/Submission11471/Reviewer_96Rd" ], [ "ICLR.cc/2025/Conference/Submission11471/Authors" ], [ "ICLR.cc/2025/Conference/Submission11471/Reviewer_aAg2" ], [ "ICLR.cc/2025/Conference/Submission11471/Reviewer_aAg2" ], [ "ICLR.cc/2025/Conference/Submission11471/Authors" ], [ "ICLR.cc/2025/Conference/Submission11471/Authors" ], [ "ICLR.cc/2025/Conference/Submission11471/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission11471/Authors" ], [ "ICLR.cc/2025/Conference/Submission11471/Area_Chair_7TiD" ], [ "ICLR.cc/2025/Conference/Submission11471/Authors" ] ], "structured_content_str": [ "{\"title\": \"The Rationality of the Power-Law Function\", \"comment\": \"In this response, we will comprehensively demonstrate the rationality of employing the power-laws to fit the activation-data curve from three aspects.\\n\\n### The Trend of Functions\\n\\nSimilar to the common training loss-data curves of most deep neural networks, the curves of activation ratios v.s. data have three important properties: (1) The curve should generally be monotonous. (2) Each curve has a steep increase (SiLU) or decrease (ReLU) at the beginning, and then its trend gradually becomes gentler with the absolute derivative decreasing to zero. (3) **According to the definition of the activation ratio, each curve must have a range within $[0,1]$ when $D\\\\geq0$.** Considering the monotonous trend of this function, it must be convergent at some limit activation ratio within $[0,1]$.\\n\\nNow, let us consider the power-law and the two functions you propose to fit well with our data. The power-law function is monotonous, bounded when $D\\\\geq0$, and a trend of absolute derivatives consistent with the above 2nd property. However, $A = c\\\\cdot\\\\log(aD+b)+d$, **the logarithmic function is not bounded at all**. **The sigmoid-like function**, $A = a - [{ 1 + b \\\\cdot\\\\exp( c D + d ) }]^{-1}$ can just meet the two properties, but **its fitting performance cannot match the power-law** according to our following experiments.\\n\\n### Fitting Performance in MAE and MSE\\n\\nTo more substantially demonstrate the advantage of our power-laws, we conduct experiments to evaluate the fitting performance of different functions using MAE and MSE. The fitting and evaluation process is presented in the script ([link](https://anonymous.4open.science/r/SparsingLawData-180E/calc.py)).\\n\\n| Scale | MAE / power-law | MAE / logarithmic | MAE / sigmoid | MSE / power-law | MSE / logarithmic | MSE / sigmoid |\\n| ------- | --------------- | ----------------- | ------------- | --------------- | ----------------- | ------------- |\\n| 0.1B | 4.07E-04| 6.22E-04| 1.16E-03| 2.68E-07| 5.83E-07| 2.17E-06|\\n| 0.2B | 2.40E-04| 2.78E-04| 4.99E-04| 9.42E-08| 1.36E-07| 3.98E-07|\\n| 0.4B | 3.45E-04| 2.76E-04| 5.68E-04| 3.07E-07| 1.14E-07| 6.77E-07|\\n| 0.8B | 1.87E-04| 3.39E-04| 1.83E-04| 6.51E-08| 1.90E-07| 5.15E-08|\\n| 1.2B | 4.91E-04| 7.51E-04| 6.41E-04| 9.42E-07| 2.66E-06| 1.87E-06|\\n| Average | 3.34E-04| 4.53E-04| 5.08E-04| 3.35E-07|7.36E-07| 8.61E-07|\\n\\nAs demonstrated by the above table, our power-law considerably outperforms the other two baselines according to both metrics of MAE and MSE.\\n\\n### Comparison in Linear Space\\n\\nVisually, the power-law, the logarithmic function, and the sigmoid-like function may all fit well with our data. Nevertheless, this is not the case when we inspect them in linear space. Specifically, these three functions can all be converted into linear functions. For example, the vanilla power-law $A=-c\\\\cdot D^{-\\\\alpha}+A_0$ can be written as $\\\\log(A_0-A)=\\\\log(c)-\\\\alpha \\\\log(D)$. The logarithmic function and the sigmoid-like function is equivalent to $\\\\exp(\\\\frac{A-d}{c})=aD+b$ and $\\\\log(\\\\frac{1-a+A}{b\\\\cdot(a-A)})=cD+d$ respectively.\\n\\nFor the above three functions, we re-plot the data points and fitted curves in the linear space, while the results can be obtained by our scripts: power-law ([link](https://anonymous.4open.science/r/SparsingLawData-180E/show.py)), logarithmic function ([link](https://anonymous.4open.science/r/SparsingLawData-180E/calc_log.py)), and sigmoid-like function ([link](https://anonymous.4open.science/r/SparsingLawData-180E/calc_sigmoid.py)). Sample results obtained on the 0.1B ReLU setting are already available: power-law ([link](https://anonymous.4open.science/r/SparsingLawData-180E/figures/show_01b_relu.png)), logarithmic function ([link](https://anonymous.4open.science/r/SparsingLawData-180E/figures/show_log_01b_relu.png)), and sigmoid-like function ([link](https://anonymous.4open.science/r/SparsingLawData-180E/figures/show_sigmoid_01b_relu.png)). By inspecting the results obtained by these scripts, we can find that both the power-law and the logarithmic function can be fitted well in the linear space. However, **the sigmoid-like function performs much worse when converted into the linear form**.\\n\\n### Summary\\n\\nTo sum up, when we take all the above three aspects into account, the power-laws are just the best choice. Both the logarithmic function and the sigmoid-like functions suffer from **larger MSE and MAE**. Besides, **the logarithmic function has the critical weakness of an unbounded range**, while **the sigmoid-like function performs badly in the linear space**.\\n\\nIn the field of scaling laws, the power-laws have long been accepted as the mainstream choice for fitting. This choice is not made intuitively. Instead, its advantage is claimed in many existing studies and experiments (e.g., Figure 23 of [1]).\\n\\n[1] Kaplan, Jared, et al. \\\"Scaling laws for neural language models.\\\" *arXiv preprint arXiv:2001.08361* (2020).\"}", "{\"title\": \"We Have Addressed Your Remaining Concerns\", \"comment\": \"We have conducted comprehensive experiments and invested considerable computation resources to address the remaining concerns you raised. We are looking forward to your response, which is really important to us.\"}", "{\"title\": \"Rebuttal\", \"comment\": \"Thank you for your excellent review. These will encourage us to further improve the quality of our work and continuously forge ahead on the research path.\\n\\n### Weakness #1\\n\\nThank you for pointing out the over-fitting concern. The computation of PPL-p% does not intentionally minimize the PPL. Actually, it adopts adaptive thresholds for different layers while maintaining a layer-wise consistent L2 relative output error (i.e., CETT). The inactivated neurons are those with output magnitudes lower than the layer-specific threshold. The PPL is incorporated as a variable to make our metric more performance-aware, binary searching the CETT just to meet the PPL requirements to provide a performance-dependent sparsity value. The determination of inactivated neurons is weakly related to PPL itself.\\n\\nBesides, existing works have already demonstrated that **a monotonous relationship between loss (log PPL) and downstream performance** generally exists. [1,2] Therefore, if PPL-p% can ensure lower PPL, this usually produces better downstream performance. As for specific results, Table 1 provides accuracies on C.R. and R.C. benchmarks given by the dense model and PPL-p% at different p% levels.\\n\\nTo further consolidate the advantage of our results, in **Figure 14** of the updated manuscript, we give the **performance-sparsity Pareto curves**, which show that PPL-p% obtains a better trade-off between task performance and sparsity compared to two baseline metrics. Specifically, both baseline metrics start to show significant performance degradation (i.e., decrease by more than 5%) at a sparsity point consistently lower than PPL-p%.\\n\\n### Weakness #2 & #4\\n\\nTo consolidate our findings, we have already started experiments on a 2.4B model, which ties all our findings to pursue a highly sparsely-activated model. Specifically, it adopts ReLU activation, a small width-depth ratio (within the interval that ensures the lowest loss), and more training data to utilize the decreasing trend between activation ratio and data of ReLU models. Besides, its larger size can further demonstrate the generalizability of our work.\\n\\nWe have long met with great difficulties collecting sufficient GPUs to run larger models. At present, we struggle to gather 64 GPUs for the above 2.4B experiment. While works on scaling laws are extremely expensive, we will try our best to present the results before the end of the rebuttal period.\\n\\n### Weakness #3\\n\\nAs stated in line 237, we adopt the same architecture of MiniCPM [3], which adopts mostly the same architecture as LLaMA except for minor muP [4] adjustments for training stability. As for training strategies, the two-stage training paradigm is already accepted by cutting-edge models such as LLaMA3 [2]. Therefore, we think experiments on MiniCPM are already generalizable enough to cover most mainstream LLMs adopting the LLaMA-like Transformer decoder-only architecture.\\n\\nIf you have other LLMs worth concern (e.g., BERT, ViT, T5, or other LLMs dissimilar to LLaMA), you may kindly provide specific suggestions so that we can present experiments in time.\\n\\n### Question #1\\n\\nThe dataset details are already described very comprehensively in Appendix E and I.\\n\\n### References\\n\\n[1] Owen, David. \\\"How predictable is language model benchmark performance?.\\\" *arXiv preprint arXiv:2401.04757* (2024).\\n\\n[2] Dubey, Abhimanyu, et al. \\\"The Llama 3 herd of models.\\\" *arXiv preprint arXiv:2407.21783* (2024).\\n\\n[3] Hu, Shengding, et al. \\\"MiniCPM: Unveiling the potential of small language models with scalable training strategies.\\\" *arXiv preprint arXiv:2404.06395* (2024).\\n\\n[4] Yang, Greg, et al. \\\"Tensor programs V: Tuning large neural networks via zero-shot hyperparameter transfer.\\\" *arXiv preprint arXiv:2203.03466* (2022).\"}", "{\"title\": \"Rebuttal\", \"comment\": \"Thank you for your excellent review. These will encourage us to further improve the quality of our work and continuously forge ahead on the research path.\\n\\n### Weakness #1\\n\\nWe believe that experiments on even larger models can further substantiate the generalizability of our research. However, we have long met with great difficulties collecting sufficient GPUs to run larger models. At present, we struggle to gather 64 GPUs and start running a 2.4B model to validate our findings, and we will try our best to present the results before the end of the rebuttal period.\\n\\nBesides, note that even the most famous works on scaling laws [1] did not run models with more than 1B parameters due to the extremely expensive nature of such studies. As for our work, we have experimented on scales from 0.1B to 1.2B, where the largest model has 12 times the number of non-embedding parameters as the smallest one. Such a large gap can already provide some generalizability for our findings.\\n\\n### Weakness #2\\n\\nSorry for causing this misunderstanding. Actually, in our paper, the FFNs in all models are gated FFN, as specified in line 153 and line 237 (MiniCPM uses gated FFN). Therefore, the ReLU and SiLU refer to ReGLU and SwiGLU, respectively. We do not incorporate non-gated FFN, as it is seldom used in recent mainstream LLMs.\\n\\nBesides, to increase the generalizability of our work, **we have started experiments on gated GELU-activated FFNs**, and already completed 0.1B and 0.2B settings. Similar to SiLU, we find a power-law relationship between activation ratio and data. The fitted curves are $A_{GELU}=-\\\\frac{0.02}{D^{1.87}} + 0.333$ and $A_{GELU}=-\\\\frac{0.14}{D^{1.15}} + 0.342$ for 0.1B and 0.2B respectively. The two limit activation ratios are also very close, and the smaller 0.1B GELU model converges much faster than 0.2B. These observations are consistent with existing results.\\n\\n### Question #1\\n\\nMathematically, the relationship between PPL and CETT is not rigorously monotonous. LLMs are statistic models with random output noises and fluctuations (PPL evaluates the deviation of model outputs from the original ground-truth training corpus). However, from the statistical perspective, we can state that PPL generally rises with the increase of CETT. This can be demonstrated through experiments.\\n\\nAs shown in **Figure 13** of our updated paper, for both 0.1B and 0.8B, from ReLU to SiLU, the PPL always rises with the increasing CETT. The larger the CETT, the more our models deviate from the original dense checkpoint trained on the ground-truth corpus. This means that the model outputs will deviate more from the ground-truth and finally produce a larger PPL value.\\n\\n### Question #2\\n\\n**The activation pattern has already been widely used in the interpretation of LLMs.** Specifically, the activation pattern reflects the behavior of neurons (FFN parameters) in response to given input tokens. By analyzing the correlation between neuron activation and inputs, we can explain the specialization of neurons (i.e., what input patterns a specific neuron is sensitive to) and even the grouping of neurons (i.e., neurons with similar specialization are clustered as intrinsic modules within LLMs). [2,3]\\n\\n**The next question is why higher sparsity can bring better interpretability.** Let's take MoEfication [4] as an example, which is a work that finds and utilizes the intrinsic modules within LLMs. This work clusters neurons into groups (i.e., experts) by using the co-activation frequencies between neurons as a distance metric. According to Figure 3 of the MoEfication paper, the relative performance degradation from the original dense model to MoEfied model is clearly less significant on sparser models. In other words, the loss of neuron clustering is less on sparser models, indicating more significant neuron grouping. Therefore, we can assume that sparser models are potentially more interpretable from the aspect of neuron grouping, an important part of interpretation.\\n\\n### References\\n\\n[1] Kaplan, Jared, et al. \\\"Scaling laws for neural language models.\\\" *arXiv preprint arXiv:2001.08361* (2020).\\n\\n[2] Li, Zonglin, et al. \\\"The Lazy Neuron Phenomenon: On Emergence of Activation Sparsity in Transformers.\\\" *The Eleventh International Conference on Learning Representations*.\\n\\n[3] Zhang, Zhengyan, et al. \\\"Emergent Modularity in Pre-trained Transformers.\\\" *Findings of the Association for Computational Linguistics: ACL 2023*. 2023.\\n\\n[4] Zhang, Zhengyan, et al. \\\"MoEfication: Transformer Feed-forward Layers are Mixtures of Experts.\\\" *Findings of the Association for Computational Linguistics: ACL 2022*. 2022.\"}", "{\"title\": \"Thank you\", \"comment\": \"Thank you to the authors for addressing several of the points. I appreciate the clarifications and additional experiments. I have no additional comments but I would like to maintain my previous score. My primary concern with the paper is that it presents a new pruning metric and then examines the phenomena resulting from that pruning metric without contextualizing the findings relative to other pruning metrics. I appreciate that figures 3 and 14 show performance comparisons between PPL-p% sparsity and other methods. How does PPL-p% sparsity directly compare with CETT since that is established as a baseline? Do all of your scaling law findings only hold when using PPL-p% sparsity or are they more general results?\"}", "{\"comment\": \"The authors address a clear misconception in my review; however, I still have concerns regarding the paper:\\n\\n2. Fitting Data with Several Parameters\\nI would like to demonstrate that, given four parameters, it is entirely possible to fit something like\\n$ F(D)= A_0 + \\\\log( a*D+B) $\\nor something entirely different to the data. However, I was unable to find the code or data from your experiments. I believe this would be necessary to strengthen your arguments in a purely empirical context.\\n\\n3. The Dying ReLU Problem: \\nCould the dying ReLU problem not precisely explain the trend (2) you observe? In the beginning, all neurons are active, but as training progresses, more neurons \\\"die,\\\" which could explain the observed behavior.\"}", "{\"title\": \"Response to the Concern\", \"comment\": \"### Performance-Awareness is the Key Property\\n\\nIn the paper, we propose three properties (line 85) that a good sparsity metric should have: versatility across model architectures, performance-awareness, and the precise recognition of weakly-contributed neurons. The Pareto curves only test the 3rd property. However, **the first key to your concern lies in performance-awareness**, **without which a sparsity metric can hardly be applied in scaling law studies**.\\n\\nSpecifically, **PPL-p% is a performance-aware improvement of the original CETT** (so they have the same Pareto curve), which distinguishes the CETT value that just makes PPL increase by p%. **The consistency of PPL increase makes it possible to compare models of distinct architectures as well as different checkpoints of the same architecture.** If you use metrics such as CETT, Top-$k$, or FAT-$\\\\epsilon$, a critical question is: what hyper-parameter (i.e., the CETT, $k$, and $\\\\epsilon$) should you choose when you measure the sparsity of different models and checkpoints? After all, they can have very different intrinsic properties, and thus it is unreasonable to apply the same CETT, $k$, or $\\\\epsilon$ for all models. **This makes it a very tough problem to use these metrics in the scaling law study and model comparison.** However, PPL-p% makes this question solvable. **To compare the sparsity of different models under the same increase PPL is the most reasonable practice we can conduct.** Of course, we can also apply a similar performance-aware extension to FAT-$\\\\epsilon$ (with an unchanged Pareto curve), but this practice fails in the 3rd property mentioned above, as demonstrated in Figure 3. For Top-$k$ and FAT-$\\\\epsilon$, specifically, their lower accuracies in recognizing weakly-contributed neurons make the sparsity level given by them questionable in whether they accurately reflect the activation pattern of models.\\n\\n### Generalizability across Different Metrics\\n\\nThe second key to your concern addresses **the generalizability across different metrics**. Through experiments, we obtain the following conclusion: **The consistency of our power-laws between the activation ratio and data mainly depends on the PPL increase ratio, rather than the sparsity metric**.\\n\\nTo support the above statement, we add experiments on FAT-$\\\\epsilon$. Note that different $\\\\epsilon$ can lead to different PPL increase ratios and results. Our results on 0.1B ReLU and 0.1B SiLU settings are shown in the figure (anonymous [link](https://anonymous.4open.science/r/SparsingLawData-180E/figures/fat_e_01b.jpg)). As demonstrated in the figure, for 0.1B ReLU, when the PPL increase ratio is relatively small (i.e., $\\\\epsilon=0,0.3,0.6$), the activation ratio follows a convergent decreasing logspace power-law with the amount of data, which is consistent with our paper. When the PPL increases considerably (i.e., $\\\\epsilon=1.0$), the power-law does not hold any longer. Similarly, for 0.1B SiLU, the convergent increasing power-law holds when PPL increases slightly (i.e., $\\\\epsilon=0.1,0.2$), while the laws are broken with a large PPL increase (i.e., $\\\\epsilon=0.275,0.5$). Besides, in another figure (anonymous [link](https://anonymous.4open.science/r/SparsingLawData-180E/figures/fat_03_scale.jpg)), we demonstrate the consistency of our power-laws for different scales of ReLU models when $\\\\epsilon=0.3$. The weak correlation between activation ratios and parameter numbers, as well as the faster convergence of smaller models, also holds according to the same figure.\\n\\nTherefore, **for both metrics including PPL-$p\\\\%$ and FAT-$\\\\epsilon$, under different numbers of parameters, our power-laws hold when the PPL does not increase considerably** (e.g., below 5%), while not the case with a large PPL increase ratio (e.g., around 10%). Obviously, **the former case is more reasonable and helpful**, as a considerably larger PPL can harm the model's performance. This provides our insights with more substantiated practical value.\"}", "{\"title\": \"Experimental Results on Larger Models (Weakness #2 & #4)\", \"comment\": [\"We have finally obtained the experimental results on a ReLU-activated **2.4B model**, which has twice the number of non-embedding parameters as the previously largest model (1.2B). At this point, it has been pre-trained on about **291B tokens**. The data used to fit the activation-data curves is available at this **anonymous** GitHub repository ([link](https://anonymous.4open.science/r/SparsingLawData-180E)), including the results of the above 2.4B model.\", \"By analyzing the 2.4B results, **we find our conclusions generalizable to this larger model**.\", \"The activation ratio also follows a logspace power-law with the amount of data: $A(D)=\\\\exp(-(9.04\\\\times10^{-6})\\\\cdot D^{2.11}-3.82)+0.071$.\", \"The limit activation ratio is 7.1%, which is close to the activation ratios of smaller models (e.g., 7.8% for 1.2B and 7.2% for 0.8B). This is consistent with our conclusion that the limit activation ratio is weakly correlated with the number of parameters.\", \"By comparing the activation-data curves, we find that the activation ratio of the 2.4B model converges much slower than the 1.2B model. This is consistent with our observation that smaller models tend to converge faster than larger models to the limit activation ratio.\", \"This 2.4B model combines all our previous insights: more efficient ReLU activation, efficient training data, and a smaller width-depth ratio. Finally, as expected, it achieves a high activation sparsity level with only about 7.1% neurons on average to be activated by each token.\"]}", "{\"title\": \"Looking Forward to Your Response\", \"comment\": \"We are looking forward to your response! Your comments will certainly help us reflect more on our work and continue to forge ahead on the academic path!\"}", "{\"summary\": \"This paper presents a study on the activation sparsity in large language models (LLMs), particularly focusing on decoder-only Transformer-based models. The authors propose a new metric called PPL-p% sparsity, which is a performance-aware measure of activation sparsity. The paper contributes to the understanding of how to design and pretrain LLMs for greater activation sparsity, which has implications for efficiency and interpretability.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. The paper introduces a novel metric for measuring activation sparsity (PPL-p% sparsity) and uncovers empirical laws regarding the scaling properties of activation sparsity in LLMs.\\n2. The paper is well-written, and the findings are presented clearly with the aid of visualizations.\\n3. The insights gained from this study can inform the design and training of future LLMs, potentially leading to more efficient and interpretable models.\", \"weaknesses\": \"1. The paper could benefit from additional experiments on larger-scale models (which is more commonly seen in applications) to confirm the generalizability of the findings.\\n2. The paper compares ReLU and SiLU activation functions, but in practical applications, a variety of activation functions such as SwiGLU may be used. The performance of these activation functions in terms of activation sparsity may differ, affecting the efficiency and performance of the model.\", \"questions\": \"1. The paper mentions the calculation of PPL-p% using a binary search method, but does the PPL change monotonically with CETT? Is this an assumption, an intuition, or is there a rigorous proof to support this relationship?\\n2. The paper mentions that activation sparsity can improve the interpretability of models; are there specific examples or methods to demonstrate how this sparsity helps explain the model's decision-making process?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"We Have Addressed Your Primary Concerns\", \"comment\": \"We have conducted comprehensive experiments and invested considerable computation resources to address the primary concerns you raised. We are looking forward to your response, which is really important to us.\"}", "{\"title\": \"Looking Forward to Your Response\", \"comment\": \"We are looking forward to your response! Your comments will certainly help us reflect more on our work and continue to forge ahead on the academic path! We have added more experiments on a larger model and more activation functions.\"}", "{\"title\": \"We Have Addressed Your Concerns\", \"comment\": \"We have conducted comprehensive experiments and invested considerable computation resources to address the concerns you raised. We are looking forward to your response, which is really important to us.\"}", "{\"title\": \"Looking Forward to Your Response\", \"comment\": \"We are looking forward to your response! Your comments will certainly help us reflect more on our work and continue to forge ahead on the academic path!\"}", "{\"title\": \"End of reviewer-author discussion phase\", \"comment\": \"Dear reviewers,\\n\\nAs we near the conclusion of the reviewer-author discussion phase, I wanted to kindly follow up to see if you\\u2019ve had a chance to review the author responses on your comments. Could you confirm that you\\u2019ve read it and, if needed, update your review and scores accordingly?\\n\\nThank you for your time and effort!\\n\\nYour AC\"}", "{\"title\": \"Looking Forward to Your Response\", \"comment\": \"We are looking forward to your response! Your comments will certainly help us reflect more on our work and continue to forge ahead on the academic path!\"}", "{\"summary\": \"This paper aims to analyze the relationship between activation sparsity and various other features of transformer-based LLMs. The authors introduce a novel metric, PPL-p% sparsity, based on a previous metric called CETT, that identifies a sparsity level at which perplexity is only increased by p% relative to a dense baseline. They study how several features relate to sparsity: amount of training data, choice of activation function, width-depth ratio of the network, and parameter scale. There are several findings, including that ReLU networks can achieve lower sparsity ratios than SiLU networks at the same performance level and convergence rates of activation (1-sparsity) ratios as the aforementioned features are varied.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": [\"This paper introduces a new metric, PPL-p%, that builds upon a previous metric, CETT, by finding a sparsity level for a desired perplexity score.\", \"The paper demonstrates that using PPL-p% as a metric for measuring activation ratio resuilts in a lower perplexity score relative to other metrics for measuring sparsity.\", \"The set of analyses are interesting and provide some valuable, albeit limited, insights into the behavior of some LLMs.\"], \"weaknesses\": [\"I believe that there is not enough evidence showing that PPL-p% is a better metric. The main comparison point between PPL-p% and other metrics evalutates the methods on perplexity of the resultant model. This seems a bit like the metric is simply overfitting to the downstream evaluation criterion. How do each of these methods do on other tasks such as the commonsense reasoning and reading comprehension tasks?\", \"A significant part of this paper involved identifying various relationships between aspects of the model and the activation sparsity. It seems like a reasonable next step would be to create a model that embodies all of these takeaways, ie has an optimal activation function, amount of training data, etc, and show the results of the activation sparsity and downstream performance relative to some baseline. This would demonstrate the practical value of the observations discussed in the paper.\", \"It is not clear if the results generalize to other LLMs. It would be nice to see results on other models.\", \"The authors state that the goal of this paper is to produce an LLM with greater activation sparsity, but I feel like this question is not quite answered. It seems as if the authors have conducted several (interesting and thorough) ablation studies, but do not tie all of their insights together to produce one most sparse model.\", \"Overall, this is a decent exploration of some phenomenology around activation ratios in neural networks, but the findings are not comprehensive or cohesive enough to warrant acceptance.\"], \"questions\": [\"I would recommend mentioning some more dataset details in the main paper rather than just \\u201ccommonsense reasoning\\u201d and \\u201creading comprehension\\u201d.\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to Additional Questions\", \"comment\": \"### Question #2\\n\\nWe have released **all our data** used to fit the activation-data curves at this **anonymous** GitHub repository ([link](https://anonymous.4open.science/r/SparsingLawData-180E)). It also includes codes to evaluate and visualize our fitting results.\\n\\nBesides, we compare the fitting results of our power-law and the logarithmic function $A_0+\\\\log(a*D+b)$. We find that for both ReLU and SiLU models, **the fitting results of power-laws are much better than the logarithmic function** according to either the Mean Square Error (MSE) or Mean Absolute Error (MAE). For example, the MSE of our power-law on 0.1B ReLU-activated model is $2.68\\\\times10^{-7}$, while the MSE of the logarithmic function increases to $2.17\\\\times10^{-5}$.\\n\\nNotably, the logarithmic function is intuitively not suitable for our data and experiment setting. It cannot converge with $D\\\\rightarrow\\\\infty$, which is totally unreasonable. After all, **the activation ratio is a bounded variable** within $[0,1]$.\\n\\nWe use the `curve_fit` method from package `scipy` to fit our curves, which employ the Levenberg-Marquardt method.\\n\\n### Question #3\\n\\nTo obtain more insights into the \\\"dying ReLU\\\" problem, we study the dead neurons in our models. Specifically, we define the neurons that are activated by less than 0.1% tokens in the validation dataset on average as \\\"dead neurons\\\". For the 0.1B ReLU-activated model and SiLU-activated model, we obtain their trends of activations ratios and dead neuron ratios. The curves are drawn in the figure included in the **anonymous** GitHub repository ([link](https://anonymous.4open.science/r/SparsingLawData-180E)).\\n\\nFor both models, **the dead neuron ratios do not increase considerably throughout the training process**. For example, the dead neuron ratio of 0.1B ReLU model increases to about 0.35% at the end of training. However, its activation ratio decreases by about 3.38% (from 10.47% to 7.09%). We can draw the following conclusion: **The \\\"dying ReLU\\\" phenomenon does slightly exist, but is far from the fundamental cause of the decreasing activation ratio trends of ReLU-activated models.**\\n\\n### Generalizability\\n\\n**Our empirical conclusions are generalizable to even larger models.**\\n\\nWe have finally obtained the experimental results on a ReLU-activated **2.4B model**, which has twice the number of non-embedding parameters as the previously largest model (1.2B). At this point, it has been pre-trained on about **291B tokens**. The data used to fit the activation-data curves is available at this **anonymous** GitHub repository ([link](https://anonymous.4open.science/r/SparsingLawData-180E)), including the results of the above 2.4B model.\\n\\nBy analyzing the 2.4B results, **we find our conclusions generalizable to this larger model**.\\n\\n- The activation ratio also follows a logspace power-law with the amount of data: $A(D)=\\\\exp(-(9.04\\\\times10^{-6})\\\\cdot D^{2.11}-3.82)+0.071$.\\n- The limit activation ratio is 7.1%, which is close to the activation ratios of smaller models (e.g., 7.8% for 1.2B and 7.2% for 0.8B). This is consistent with our conclusion that the limit activation ratio is weakly correlated with the number of parameters.\\n- By comparing the activation-data curves, we find that the activation ratio of the 2.4B model converges much slower than the 1.2B model. This is consistent with our observation that smaller models tend to converge faster than larger models to the limit activation ratio.\"}", "{\"title\": \"Acknowledgment of Response\", \"comment\": \"Thank you to the authors for their efforts in addressing my concerns.\\n\\n3) Thank you for giving further insides into the dying ReLU problem. I agree that the small amount of dying neurons given by the threshold is not sufficient to explain the results. \\n\\n2) I inspected the repository and fitted functions myself. I found that fitting $a*\\\\log(b*x+c)+d$ worked well while $a-1/(1+b*\\\\exp(c*x+d) $ could also be fitted to the data. While the repository shows some of the data I believe that providing all the data and code to generate the data in a repository would strengthen the work.\\n\\nAfter reviewing the clarifications provided, I have decided to maintain my previous score.\"}", "{\"title\": \"Experimental Results on Larger Models (Weakness #1)\", \"comment\": [\"We have finally obtained the experimental results on a ReLU-activated **2.4B model**, which has twice the number of non-embedding parameters as the previously largest model (1.2B). At this point, it has been pre-trained on about **291B tokens**. The data used to fit the activation-data curves is available at this **anonymous** GitHub repository ([link](https://anonymous.4open.science/r/SparsingLawData-180E)), including the results of the above 2.4B model.\", \"By analyzing the 2.4B results, **we find our conclusions generalizable to this larger model**.\", \"The activation ratio also follows a logspace power-law with the amount of data: $A(D)=\\\\exp(-(9.04\\\\times10^{-6})\\\\cdot D^{2.11}-3.82)+0.071$.\", \"The limit activation ratio is 7.1%, which is close to the activation ratios of smaller models (e.g., 7.8% for 1.2B and 7.2% for 0.8B). This is consistent with our conclusion that the limit activation ratio is weakly correlated with the number of parameters.\", \"By comparing the activation-data curves, we find that the activation ratio of the 2.4B model converges much slower than the 1.2B model. This is consistent with our observation that smaller models tend to converge faster than larger models to the limit activation ratio.\"]}", "{\"summary\": \"The paper identifies architectural choices that influence activation sparsity, which is measured by there own metric. They find that the ReLU activation function promotes sparsity more than SiLU, and propose relations that model activation sparsity as a function of training data. The influence of model depth and width is also analysed.\", \"soundness\": \"1\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1) The paper raises an interesting question, that is: when is a model sparse?\\n\\n2) The paper points out potential influences that can lead to sparsity.\", \"weaknesses\": \"1) The paper claims that creating sparse LLMs is of broad interest to the community, but the cited papers are mainly interested in removing sparse parts of the network to speed up e.g. inference of large models. The question of where sparse networks perform worse/better, or how the modifications in general affect performance, is not adequately addressed.\\n\\n2) The paper claims \\\"empirical laws\\\", but they are not sufficiently motivated and validated. Even if the parameters of a model can be determined empirically, the generality of the results must be questioned when a function with four parameters (Eq. 4) is fitted to a curve. \\n\\n3) One of the key results of the paper is that ReLU produces sparser networks than SiLU, which is not at all surprising given the \\\"dying ReLU\\\" problem.\", \"questions\": \"See Weaknesses\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Experimental Results on Larger Models (Weakness #3)\", \"comment\": [\"We have finally obtained the experimental results on a ReLU-activated **2.4B model**, which has twice the number of non-embedding parameters as the previously largest model (1.2B). At this point, it has been pre-trained on about **291B tokens**. The data used to fit the activation-data curves is available at this **anonymous** GitHub repository ([link](https://anonymous.4open.science/r/SparsingLawData-180E)), including the results of the above 2.4B model.\", \"By analyzing the 2.4B results, **we find our conclusions generalizable to this larger model**.\", \"The activation ratio also follows a logspace power-law with the amount of data: $A(D)=\\\\exp(-(9.04\\\\times10^{-6})\\\\cdot D^{2.11}-3.82)+0.071$.\", \"The limit activation ratio is 7.1%, which is close to the activation ratios of smaller models (e.g., 7.8% for 1.2B and 7.2% for 0.8B). This is consistent with our conclusion that the limit activation ratio is weakly correlated with the number of parameters.\", \"By comparing the activation-data curves, we find that the activation ratio of the 2.4B model converges much slower than the 1.2B model. This is consistent with our observation that smaller models tend to converge faster than larger models to the limit activation ratio.\"]}", "{\"title\": \"Acknowledgment of Response\", \"comment\": \"Thank you to the authors for their efforts in addressing my concerns. I think the sparsity law observed in this paper offers valuable insights for researchers to better understand the sparsity of large language models and inspire the design of sparse models. After reviewing the clarifications provided, I have decided to maintain my previous score.\"}", "{\"summary\": \"This paper addresses activation sparsity in large language models (LLMs), where a significant portion of elements in activation outputs have minimal contributions to the final output. The authors aim to enhance activation sparsity to improve computational efficiency, interpretability, and training dynamics. They propose a new metric, PPL-p% sparsity, which is performance-aware and adaptable across various architectures. Through extensive experiments, the paper studies factors influencing activation sparsity, including activation functions, width-depth ratios, and parameter scales, revealing patterns and scaling properties that can inform the design and training of efficient, sparse LLMs.\", \"soundness\": \"4\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. Novel Metric for Performance-Aware Sparsity: The introduction of PPL-p% sparsity is a key contribution, offering a more performance-sensitive measurement of activation sparsity that is adaptable across model architectures. This metric brings a practical perspective to sparsity evaluation in LLMs.\\n2. Insightful Scaling Laws for Activation Sparsity: The proposed scaling laws establish patterns in activation sparsity across varying model parameters, helping to guide model design for optimized efficiency and sparsity. These scaling laws are particularly valuable for practitioners focused on resource-efficient model scaling.\\n3. Practical Design Guidelines for Sparse LLMs: The paper\\u2019s findings offer actionable design guidelines, such as optimal width-depth ratios, for promoting activation sparsity without compromising model performance. This provides a practical framework for building more efficient LLMs.\", \"weaknesses\": \"1. Limited Exploration of Sparsity Effects on Downstream Tasks: While the paper extensively analyzes activation sparsity during training, it lacks exploration of how increased sparsity impacts downstream task performance. This leaves uncertainty about whether the sparsity benefits hold in real-world applications.\\n2. Inconsistency in Performance Across Different Width-Depth Ratios: Although the paper highlights optimal width-depth ratios, it lacks a detailed examination of potential performance trade-offs when deviating from these ratios. This makes it difficult to understand the flexibility of the proposed guidelines for diverse architectures.\\n3. Scalability Concerns for Very Large Models: The paper\\u2019s experiments are conducted on specific model scales, but it is unclear how well the findings scale to extremely large models (e.g., hundreds of billions of parameters). Additional validation on larger LLMs would enhance the paper\\u2019s applicability to state-of-the-art models.\\n4. Unverified Claims on Sparsity Ratios and Performance Degradation: The paper mentions that overly high or low sparsity ratios may lead to severe performance degradation or unnecessary computation. However, this claim lacks experimental validation. Empirically, in many kinds of MoE models, higher sparsity levels generally correlate with improved performance, which contradicts the paper\\u2019s statement.\\n5. Insufficient Training Tokens and Overreliance on Predicted Curves: The experiments use an insufficient number of training tokens, relying heavily on predicted scaling curves (e.g., Figure 4). For larger models (e.g., 0.8B and above), training up to 200B tokens is typically required to observe convergence. The lack of experiments on larger scales raises doubts about the reliability of the proposed scaling laws and their applicability to state-of-the-art models.\\n6. Limited Practicality in Reducing Computational Load: The proposed method may face challenges in reducing computational load in real applications. Empirically, different tokens activate different channels, making it difficult to apply a uniform activation pattern across all tokens. Since the method relies on precomputed PPL and activation patterns, these patterns may not generalize well to other tokens. In extreme cases, this would require all tokens to activate all channels to achieve their unique activation pattern, negating the intended efficiency gains.\", \"questions\": \"1. Applicability of Scaling Laws to Extremely Large Models: The current experiments are conducted on a specific range of model scales. Do the authors plan to validate the proposed scaling laws on much larger models (e.g., hundreds of billions of parameters)? Such validation would enhance the reliability of the scaling laws for the latest state-of-the-art LLMs.\\n2. Memory and Computational Efficiency Gains: Could the authors provide quantitative results on memory and computational efficiency improvements achieved through increased activation sparsity? Detailed comparisons would strengthen the practical impact of promoting sparsity in LLMs.\\n3. Broader Exploration of Activation Functions: The paper mainly discusses ReLU and SiLU activations. Have the authors considered examining other commonly used activation functions, such as GELU and Swish? Exploring a broader range of activations could help generalize the findings to a wider variety of LLM architectures\\n4. Scaling Law Reliability with Limited Training Tokens: Given the limited number of training tokens used in the experiments, could the authors discuss the potential impact of this on the accuracy of the scaling laws? Would they consider conducting larger-scale experiments, ideally with 200B tokens for models above 0.8B parameters, to validate these scaling patterns more robustly?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Rebuttal 2/2\", \"comment\": \"### Weakness #5\\n\\nThank you for pointing out this problem. During the period after submission, we have already invested resources to remedy this problem. Now the amount of training data for 0.8B and 1.2B has both been largely **extended to more than 200B tokens**. New results are included in the updated manuscript (see Figure 4) and conclusions remain consistent.\\n\\n### Weakness #6\\n\\nIndeed, the diverse token-wise activation pattern is a major obstacle in leveraging activation sparsity for acceleration. However, existing works have already provided solutions. Representative works include the PowerInfer series [6], which manages to achieve up to 29.2x speedup compared to SOTA inference frameworks by leveraging activation sparsity. The major keys to overcoming the acceleration challenge include: (1) a sophisticated sparsification training method named TurboSparse to improve activation sparsity; (2) the introduction of activation predictors to forecast activated parameters and prefetch required data from memory; (3) careful design of cache mechanism and fine-grained pipeline to reduce IO overheads.\\n\\nBesides, some other interesting works also reveal that the token-wise activation pattern is not completely irregular and unpredictable. For example, GRIFFIN [7] observes the flocking phenomenon, where relative activation magnitudes are shared within a sequence, across the prompt and generated texts. This can be utilized for practical acceleration up to 1.29x.\\n\\n### Question #1\\n\\nSame response as **Weakness #3**.\\n\\n### Question #2\\n\\nExisting works can provide some quantitative results for reference. As stated in **Weakness #6**, when serving TurboSparse-Mixtral-47B (a model activating only about 3B parameters for each token on average), PowerInfer-2 [6] can achieve up to 29.2x speedup and about 40% reduction in memory usage. Note that these are all done on smartphones.\\n\\nBesides, ProSparse [8] specifically experiments on the practical acceleration effects of higher sparsity. As shown in Table 2 of the ProSparse paper, with PowerInfer-1, a 7B model with 66.98% activation sparsity can achieve 3.10x speedup compared to the dense setting, while the speedup rises to 4.44x for the model of sparsity 88.11%. With the CUDA operators for sparse FFN, 66.98% sparse model has 1.35x and 1.32x speedup compared to the dense operator for two FFN computation steps respectively, while the acceleration rises to 1.94x and 1.49x for the 88.11% sparse model.\\n\\n### Question #3\\n\\nTo increase the generalizability of our work, **we have started experiments on gated GELU-activated FFNs**, and already completed 0.1B and 0.2B settings. Similar to SiLU, we find a power-law relationship between activation ratio and data. The fitted curves are $A_{GELU}=-\\\\frac{0.02}{D^{1.87}} + 0.333$ and $A_{GELU}=-\\\\frac{0.14}{D^{1.15}} + 0.342$ for 0.1B and 0.2B respectively. The two limit activation ratios are also very close, and the smaller 0.1B GELU model converges much faster than 0.2B. These observations are consistent with existing results.\\n\\n### Question #4\\n\\nSame response as **Weakness #5**.\\n\\n\\n\\n### References\\n\\n[1] Petty, Jackson, et al. \\\"The impact of depth and width on transformer language model generalization.\\\" *arXiv preprint arXiv:2310.19956* (2023).\\n\\n[2] Wu, Chuhan, and Ruiming Tang. \\\"Performance Law of Large Language Models.\\\" *arXiv preprint arXiv:2408.09895* (2024).\\n\\n[3] Kaplan, Jared, et al. \\\"Scaling laws for neural language models.\\\" *arXiv preprint arXiv:2001.08361* (2020).\\n\\n[4] Krajewski, Jakub, et al. \\\"Scaling laws for fine-grained mixture of experts.\\\" *arXiv preprint arXiv:2402.07871* (2024).\\n\\n[5] Fedus, William, Barret Zoph, and Noam Shazeer. \\\"Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity.\\\" *Journal of Machine Learning Research* 23.120 (2022): 1-39.\\n\\n[6] Xue, Zhenliang, et al. \\\"PowerInfer-2: Fast Large Language Model Inference on a Smartphone.\\\" *arXiv preprint arXiv:2406.06282* (2024).\\n\\n[7] Dong, Harry, Beidi Chen, and Yuejie Chi. \\\"Prompt-prompted Adaptive Structured Pruning for Efficient LLM Generation.\\\" *First Conference on Language Modeling*. 2024.\\n\\n[8] Song, Chenyang, et al. \\\"ProSparse: Introducing and Enhancing Intrinsic Activation Sparsity within Large Language Models.\\\" *arXiv preprint arXiv:2402.13516* (2024).\"}", "{\"title\": \"Rebuttal 1/2\", \"comment\": \"Thank you for your excellent review. These will encourage us to further improve the quality of our work and continuously forge ahead on the research path.\\n\\n### Weakness #1\\n\\nIn Table 1, we already provide accuracies on C.R. and R.C. benchmarks given by the dense model and PPL-p% at different p% levels (also, with different sparsity levels).\\n\\nTo further consolidate the advantage of our results, in **Figure 14** of the updated manuscript, we give the **Pareto curves of task performance v.s. sparsity**. These curves show that PPL-p% obtains a better trade-off between task performance and sparsity compared to baseline metrics. Specifically, both baseline metrics start to show significant performance degradation (i.e., decrease by more than 5%) at a sparsity point consistently lower than PPL-p%. Meanwhile, this also demonstrates that high sparsity can co-exist with a low PPL increase ratio and minor performance degradation, especially in ReLU-activated models. **Note that these evaluation accuracies and sparsity levels are measured on models after the decay stage, which includes SFT data.**\\n\\n### Weakness #2\\n\\nAs for the width-depth issue, I'd like to state the most accurate suggestion we can give. That is, **use the smallest width-depth ratio that ensures training stability**.\\n\\nExisting works have conducted studies on the trade-off between performance and width-depth ratio. [1,2] As revealed by these papers, deeper models (with smaller width-depth ratios) generally present better performance if unlimited precision is given. However, real-life training usually involves half-precision training, where an extremely high depth can considerably harm stability and thus cause performance degradation. These findings are consistent with our work, which states that a smaller width-depth ratio is better for higher sparsity, but an extremely small ratio significantly harms performance (due to training instability).\\n\\nTo sum up, if unlimited precision virtually exists, smaller width-depth ratios can simultaneously bring better performance and higher sparsity. Nevertheless, if limited precision and training stability are considered, the best value is hard to say, as the training stability depends on various factors, such as the training framework and hardware. The most accurate suggestion is to use the smallest width-depth ratio that guarantees training stability.\\n\\n### Weakness #3\\n\\nWe believe that experiments on even larger models can further substantiate the generalizability of our research. However, we have long met with great difficulties collecting sufficient GPUs to run larger models. At present, we struggle to gather 64 GPUs and start running a 2.4B model to validate our findings, and we will try our best to present the results before the end of the rebuttal period.\\n\\nBesides, note that even the most famous works on scaling laws [3] did not run models with more than 1B parameters due to the extremely expensive nature of such studies. As for our work, we have experimented on scales from 0.1B to 1.2B, where the largest model has 12 times the number of parameters as the smallest one. Such a large gap can already provide some reliability for our findings.\\n\\n### Weakness #4\\n\\nIn both Figure 3 and Table 1, we have comprehensively demonstrated that **whatever sparsity metric we use, enforcing too high activation sparsity can cause considerable performance degradation**. In Figure 1, we also present experiments studying the effect of granularity and activation rate on the PPL of MoE models, which shows that **larger activation ratios generally present lower PPL in MoE**. Note that we apply standard load balancing loss for all MoE settings.\\n\\nBesides, there are also works [4] studying the quantitative relationship between performance and the number of activated parameters in fine-grained MoE models, which clearly show that **less activated parameters can negatively affect performance**. Note that fine-grained MoE can be regarded as a special case of activation sparsity, and thus it is reasonable to assume similar laws in our scenario.\\n\\nFinally, as you mention many kinds of MoE models can simultaneously obtain higher sparsity and better performance, we think this can usually be attributed to better training data, special training techniques (e.g., sparsity restrictions in training target), or other potential factors. A special case we find is Switch Transformer [5], where the top-1 routing performs better than top-2 routing. This may be attributed to many other improvements, such as differentiable load balancing, the setting of expert capacity, and many sophisticated training techniques mentioned in Section 2.4. If you have more recent cases, where models are of the same architecture and training recipe, you can provide us for further discussion.\"}", "{\"title\": \"Rebuttal\", \"comment\": \"Thank you for your review. These will encourage us to further improve the quality of our work and continuously forge ahead on the research path.\\n\\n### Weakness #1\\n\\nFirst, there seems to be a fundamental misunderstanding in your review. **Activation sparsity is completely different from pruning, which you indicate by \\\"removing sparse parts of the network to speed up\\\".** Pruning realizes sparsity by removing certain parts of the model and this process is independent of the inputs. Its input-independent property can easily cause performance degradation. By contrast, activation sparsity, which we mainly research, is dynamically determined by the inputs and thus potentially compromises less model capacity and downstream task performance. More specifically, **activation sparsity does not remove any part from the model**. Instead, the sparsity is realized by recognizing the unimportant parameters for each token and then dynamically deactivating them.\\n\\n**Next, how the sparsity affects performance is well studied in this paper.** In Table 1, we already provide accuracies on C.R. and R.C. benchmarks given by the dense model and PPL-p% at different p% levels (also, with different sparsity levels).\\n\\nTo further consolidate the advantage of our results, in **Figure 14** of the updated manuscript, we give the Pareto curves of task performance v.s. sparsity. These curves show that PPL-p% obtains a better trade-off between task performance and sparsity compared to baseline metrics. Specifically, both baseline metrics start to show significant performance degradation (i.e., decrease by more than 5%) at a sparsity point consistently lower than PPL-p%. Meanwhile, this also demonstrates that high sparsity can co-exist with a low PPL increase ratio and minor performance degradation, especially in ReLU-activated models. Note that these evaluation accuracies and sparsity levels are measured on models after the decay stage, which includes SFT data.\\n\\n**Finally, the works we cited can sufficiently demonstrate the advantages of employing activation sparsity**, such as inference acceleration and interpretability. PowerInfer-2 [3], for example, can achieve up to 29.2x speedup and about 40% reduction in memory usage by utilizing activation sparsity. Note that these are all done on smartphones.\\n\\n### Weakness #2\\n\\nFirst, **the parameters of our models are not empirically determined**. We adopt MiniCPM [1] as our experimental architecture and employ muP [2] to specify hyper-parameters for training stability. Such model architecture and parameter specification are both demonstrated to be reliable and well recognized, with considerable citations and application.\\n\\nNext, **all the works on scaling properties are empirical**, usually including extensive experiments and curve fitting. Among these works, the most famous one is done by OpenAI [4]. Though also consisting of empirical experiments and curve fitting, this work reveals the quantitative relationship between the performance of AI models and the amount of training data as well as the number of parameters. Its precious experience lays the solid foundation of LLM and leads human beings to the revolution of AI.\\n\\nA helpful work may not necessarily have all its conclusions proved in a mathematical and rigorous way. In the field of AI, **reliable conclusions can be obtained from extensive experiments**. As for our work, we have experiments on **five scales ranging from 0.1B to 1.2B**, where the largest model has 12 times the number of non-embedding parameters as the smallest one. Besides, models are fully trained with **hundreds of billions of data**. Such extensiveness and expensiveness can already provide some generalizability for our findings.\\n\\nFinally, even from the statistical perspective, fitting four parameters with dozens of samples per curve is a quite reliable practice.\\n\\n### Weakness #3\\n\\nThe findings of our paper are far beyond \\\"ReLU is sparser\\\". The \\\"dying ReLU\\\" may be a reasonable explanation for the higher sparsity of ReLU-activated models, but **it cannot explain another two important findings related to ReLU**: (1) ReLU-activated models are well comparable in performance with SiLU-activated ones; (2) ReLU-activated models undergo an increasing trend of sparsity with the increasing data, while SiLU-activated ones undergo an opposite trend. Therefore, we demonstrate that ReLU is a more efficient choice considering both sparsity and performance. **The \\\"dying ReLU\\\" may be true, but it is an advantage of ReLU that promotes sparsity without harming performance, rather than a problem.**\\n\\nBesides, our findings related to activation functions are of very important value in this LLM era, as most mainstream LLMs adopt SiLU without awareness of the potential benefits in activation sparsity provided by ReLU. What we want to achieve is to attract more attention to activation sparsity, and provide precious experience for LLM researchers to reconsider specific model settings for higher sparsity.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"title\": \"Looking Forward to Your Response\", \"comment\": \"We are looking forward to your response! Your comments will certainly help us reflect more on our work and continue to forge ahead on the academic path! We have addressed your remaining concerns in detail with full experiments. Meanwhile, the data and codes are also provided to substantiate our studies further.\"}", "{\"metareview\": \"a) Summary:\\nThe paper presents a comprehensive study of activation sparsity in large language models (LLMs), focusing on decoder-only Transformer architectures. The key claims include:\\n1. Introduction of PPL-p% sparsity as a new performance-aware metric for measuring activation sparsity\\n2. Different activation functions (ReLU vs SiLU) show opposite trends in training-time sparsity, with ReLU demonstrating better sparsity properties\\n3. The activation ratio follows power-law relationships with training data: decreasing for ReLU and increasing for SiLU\\n4. The activation ratio increases linearly with width-depth ratio below a bottleneck point\\n5. The limit value of activation sparsity shows weak correlation with model parameter scale\\n\\n(b) Strengths:\\n1. Novel metric (PPL-p%) that considers both sparsity and performance impact\\n2. Extensive empirical analysis across different model scales (0.1B to 2.4B parameters)\\n3. Clear practical implications for LLM design, particularly regarding activation function choice\\n4. Well-documented experimental methodology with comprehensive ablation studies\\n5. Strong theoretical grounding in relating findings to existing work on model scaling\\n\\n(c) Weaknesses:\\n1. Limited exploration of extremely large-scale models (>10B parameters)\\n2. Initial experiments had insufficient training tokens for larger models, though partially addressed during rebuttal\\n3. Unclear generalization beyond the specific architecture studied (MiniCPM/LLaMA-like models)\\n4. Some claims about performance impact of sparsity ratios needed stronger empirical validation\\n5. Limited exploration of practical acceleration benefits in real-world scenarios\\n\\n(d) Reasons for Decision:\", \"i_recommend_rejection_primarily_because\": \"1. The work, while thorough in its empirical analysis, presents findings that are somewhat incremental rather than transformative\\n2. The practical impact is limited by the focus on a specific architecture and moderate model scales\\n3. Some key claims about sparsity's impact on performance and scaling properties need stronger validation\\n4. The findings, while interesting, don't provide sufficient breakthrough insights to warrant acceptance at ICLR\", \"additional_comments_on_reviewer_discussion\": \"The rebuttal period generated substantive discussion around several key aspects of the paper. The reviewers collectively raised concerns about the generalizability to larger models, the superiority of the PPL-p% metric, insufficient training tokens, and the validity of the empirical laws. In response, the authors conducted extensive additional experiments, notably training a 2.4B parameter model and extending training to over 200B tokens for larger models. They also provided detailed analyses validating their curve fitting methodology and demonstrating the practical benefits of their approach through references to implementation studies.\\nReviewer 96Rd questioned the validity of the empirical laws and ReLU findings, which the authors addressed through detailed analysis of dead neurons and comprehensive curve fitting validation. Reviewer MqGB's concerns about the PPL-p% metric were addressed with additional performance comparisons, while Reviewer aAg2's questions about practical computational benefits received substantive responses with references to concrete acceleration benefits in deployed systems. Reviewer 6UUp's concerns about generalizability were partially addressed through the new 2.4B model experiments.\\nWhile the authors' responses were thorough and backed by new experimental evidence, they did not fully resolve the fundamental limitations of the work's scope and impact. The additional experiments and analyses strengthened the paper's empirical foundation but did not elevate the contribution to the level expected for ICLR acceptance.\"}", "{\"title\": \"References\", \"comment\": \"[1] Hu, Shengding, et al. \\\"MiniCPM: Unveiling the potential of small language models with scalable training strategies.\\\" *arXiv preprint arXiv:2404.06395* (2024).\\n\\n[2] Yang, Greg, et al. \\\"Tensor programs V: Tuning large neural networks via zero-shot hyperparameter transfer.\\\" *arXiv preprint arXiv:2203.03466* (2022).\\n\\n[3] Xue, Zhenliang, et al. \\\"PowerInfer-2: Fast Large Language Model Inference on a Smartphone.\\\" *arXiv preprint arXiv:2406.06282* (2024).\\n\\n[4] Kaplan, Jared, et al. \\\"Scaling laws for neural language models.\\\" *arXiv preprint arXiv:2001.08361* (2020).\"}" ] }

[ "Guangyan Gan", "Mengzhe Ruan" ]

[ "Reinforcement Learning" ]

{ "note_id": [ "xtvcybxJw8", "wNhhw1wsRo", "WrKBwL6hmq", "Pe8UFpd4pR", "PWIWsHir9h", "LTeTPilXPc", "9OVGrJYlvj" ], "note_type": [ "official_review", "official_review", "comment", "official_review", "official_comment", "official_review", "official_review" ], "note_created": [ 1730485359793, 1730586844744, 1732801158909, 1730186098337, 1732800201192, 1729755078357, 1729347978544 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission2309/Reviewer_naX6" ], [ "ICLR.cc/2025/Conference/Submission2309/Reviewer_Q1VE" ], [ "ICLR.cc/2025/Conference/Submission2309/Authors" ], [ "ICLR.cc/2025/Conference/Submission2309/Reviewer_kEVf" ], [ "ICLR.cc/2025/Conference/Submission2309/Reviewer_GCRq" ], [ "ICLR.cc/2025/Conference/Submission2309/Reviewer_8GXE" ], [ "ICLR.cc/2025/Conference/Submission2309/Reviewer_GCRq" ] ], "structured_content_str": [ "{\"summary\": \"The authors propose a method for knowledge transfer in offline RL. In their setting, there are L source tasks and a target task which is a linear combination of the source tasks in the linear MDP. There is an abundance of the data in the source tasks, but the target task has only limited data. The authors propose an algorithm for knowledge transfer in this case and provide a theoretical analysis on the optimality of the algorithm.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": [\"This paper studies the problems within the domain of offline reinforcement learning which is an active area of research and has a potential to have a large impact in the field.\", \"The authors consider a novel problem setting and the assumption where the target task is a linear combination of several source tasks.\", \"The authors are rigorous in their definitions and approach to studying the problem.\"], \"weaknesses\": [\"My main concern about the paper is that while it is set to address a practical problem of data scarcity in offline reinforcement learning in the target domain, it completely lacks any experiments, even in toy domains. Given that the paper proposes a new algorithm to address the problem and motivates the need for an algorithm by some practical limitations, I believe that the experimental section is necessary even though the paper's focus is theoretical analysis.\", \"I am not completely convinced by the assumption of the paper where the target task is a linear combination of the source tasks. What are the examples when this is relevant? How common are such examples in reality? How can we know if this assumption holds for a given task / environment?\", \"I think the clarity of the paper could be improved. For example, the real problem formulation comes only at the end of page 4, and it would be better to understand the main setting and assumptions early in the paper, as, for example, saying at the beginning that \\\"target data is a linear combination of source data\\\" is not enough and it should be explained what it means precisely.\", \"In several places in the paper, the authors rely on the closed-form solutions. With what size of the dataset is this applicable computationally?\"], \"questions\": \"I would like the authors to comment on the applicability and realism of the assumption on the target domain being a linear combination of the source domains.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper focuses on the problem of knowledge transfer in offline reinforcement learning. The paper specifically focuses on a linear MDP setting, where transition function and reward are defined as linear functions of a set of features. In the knowledge transfer setting, there is a large amount of data from *source* tasks, and limited data for the desired *target* task. The key assumption is that the linear reward parameters of the target task are a linear combination of the reward parameters of the source tasks.\\n\\nThe proposed algorithm, KT-RL, involves constructing an estimate of the value function under the target task, using data from the source tasks. The value function, Q-function, and uncertainty are computed via Bellman iteration. The estimates for the target task reward function parameters are estimated from the source task statistics. The resulting Q-values are then penalized under uncertainty.\", \"soundness\": \"2\", \"presentation\": \"1\", \"contribution\": \"2\", \"strengths\": \"This paper introduces an offline knowledge transfer setting that is tractable for theoretical analysis. By making clear assumptions about the structure of the source and target data, they are able to derive appropriate bounds on the accuracy of the estimated value functions. To my knowledge this analysis has not been done on the specific offline-data, knowledge transfer regime.\", \"weaknesses\": \"The assumptions described in the paper are limiting. The algorithm assumes access to a prior-known feature map phi(x,a), in which the transition functions and rewards are linear. A large challenge in many practical RL problems is to learn this feature map -- presumably, offline data from source tasks can help in learning this map itself (which is not covered under these assumptions).\\n\\nThe contributions noted in the introduction lack justification. For example, assuming linearity is framed as breaking an assumption, when it instead limits the practicality of the results. The same holds for the assumption about dataset compliance and trajectory independence. The authors argue that their method \\\"enhances privacy preservation\\\", but does not elaborate on how this may be achieved. \\n\\nWhile it is OK for the paper to focus entirely on theoretical contributions, since the paper is framed as providing an algorithm (KT-RL), at least some form of didactic example of this algorithm would strengthen the contribution.\\n\\nThe writing in the paper can use work on improving clarity. Many terms are introduced which are not properly defined. In addition, especially Section 5 about the algorithm is hard to read, and does not clearly explain the algorithmic procedure, nor the justification behind why it is designed as such. Sections 5.1 and 5.2 contain many repeated sentences and equations. The terms \\\"Calibration\\\" and \\\"Truncation\\\" are used in Algorithm 1, but are not defined anywhere in the paper. It is often unclear which quantities are assumed to be known ahead of time, and which are aimed to be discovered by the algorithm.\", \"questions\": \"What is the relation between data scarcity (as motivated in the introduction) and the proposed algorithm? Additional details on the motivation of what this analysis is aimed to provide would strengthen the paper.\\n\\nWhat is the motivation behind the particular definition of suboptimality used in equation 4? A connection between this and a desired goal would be helpful (i.e. minimizing regret, maximizing expected performance of a policy under the target task, etc).\\n\\nWhy is the writing in 5.1 and 5.2 duplicated? For example MSBE is defined multiple times. If the intent is to showcase the difference between the target and source tasks, such as the presence of 'w', it would be clearer to show the comparison side by side. Algorithm 1 and the algorithm section in general can use significant clarity improvements to make it clear what each part of the algorithm aims to accomplish, and to justify each step. For example, the calibration, pessimism, and truncation steps are undefined.\\n\\nThe paper can in general benefit from less overstated language. For example, the abstract and introduction talk about \\\"practical and efficient utilization for applications\\\", yet this claim is not demonstrated in the paper. The introduction also talks about powerful function approximation, yet the paper does not describe any relation to them.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"withdrawal_confirmation\": \"I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors.\", \"comment\": \"We would like to express our sincere gratitude for the time and effort you have dedicated to reviewing our submission for ICLR 2025. After careful consideration, we have decided to withdraw our current manuscript.\\nWe acknowledge that there are deficiencies and lack of clarity in Assumption 1. We have already made new progress based on this and are in the process of refinement. Additionally, there are some issues with unclear descriptions and word choices in the article, which we will revise. We also plan to conduct further experiments to enhance this work.\\nOnce again, we truly appreciate the valuable time and suggestions from the reviewers. We will strive to improve the quality of our work and may consider resubmitting it in the future.\"}", "{\"summary\": \"The paper presents a theoretical framework for knowledge transfer in offline RL. This study considers a linear MDP and investigates an algorithm for transferring value functions and policies from a source domain to a target domain. The upper bound on suboptimality and the minimax optimality are established.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": [\"I am not aware of prior theoretical studies on knowledge transfer in offline RL. It is interesting to see that an upper bound on suboptimality can be demonstrated for knowledge transfer in offline RL.\"], \"weaknesses\": [\"The relationship to prior work on transfer learning in offline RL is unclear. Is this the first study to establish an upper bound on suboptimality in offline RL transfer learning? If not, please provide a comparison, such as how tight the upper bound is relative to previous work.\", \"No experiments are presented to evaluate the proposed algorithm. If possible, please provide experimental results on applicable tasks, such as simple tasks with tabular settings.\", \"There is no clear connection to recent deep RL algorithms. I believe that algorithms based on successor features could be a promising approach for implementing the proposed method in ways applicable to real-world problems. Please offer insights on how to adapt the proposed algorithm for real-world applications.\"], \"questions\": [\"Please provide a comparison, such as how tight the upper bound is relative to previous work.\", \"If possible, please provide experimental results on applicable tasks, such as simple tasks with tabular settings.\", \"Please offer insights on how to adapt the proposed algorithm for real-world applications.\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Official Comments by Reviewer GCRq\", \"comment\": \"I still believe that the author's idea has innovative elements. However, the author has not replied to my questions, and thus my doubts remain unresolved. Consequently, my concern about the author's positioning of this paper, whether it is purely theoretical or also takes into account experimental testing, will not be fully addressed. Therefore, if the author still fails to respond to my questions, I will consider lowering my score or confidence.\"}", "{\"summary\": \"The paper provides theoretical insights into knowledge transfer from source tasks to a target task in offline reinforcement learning (RL) to address data scarcity.\\nThe analysis is applicable to algorithms operating within the linear MDP framework, where the target data is assumed to be a linear combination of the source data.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"1. The paper presents a thoughtful combination of offline reinforcement learning and transfer learning, with useful mechanisms for handling uncertainty and leveraging source tasks.\\n2. The paper offers a rigorous theoretical analysis of the proposed approach, KT-RL.\", \"weaknesses\": \"1. Lack of experimental evaluation: The authors could prepare synthetic experiments to demonstrate the practical applicability and robustness of KT-RL.\\n2. The related work could look into some recent related knowledge transfer techniques in offline RL like [1] [2].\\n3. Calculation of inverse Gram matrices \\u039b(l)\\u22121 can be computationally expensive if the features or samples is large\\ncausing performance bottlenecks in high-dimensional settings.\", \"suggestion\": \"Since the paper uses many theoretical notations, it is hard to keep track of them, which hampers readability. Providing a notation table would enhance the paper's clarity and accessibility.\", \"references\": \"[1] Gangopadhyay, Briti, et al. \\\"Integrating Domain Knowledge for handling Limited Data in Offline RL.\\\" arXiv preprint arXiv:2406.07041 (2024).\\n[2] Wang, Zhao, et al. \\\"Augmenting Offline RL with Unlabeled Data.\\\" arXiv preprint arXiv:2406.07117 (2024).\", \"questions\": \"Q1. What does data scarcity mean in quantifiable terms? Can the authors define data scarcity?\\n \\nQ2. How relevant is the linear combination of source data resulting in target data assumption for practical scenarios. Are there some examples of tasks following such assumption?\\n\\nQ3. In line 126, the authors state, \\\"source data closely resembles target data, our approach diverges from this assumption.\\\" \\nHowever, they also assume that the target data is a linear combination of the source data. Could the authors elaborate on how this differs from related work? Additionally, \\nw_h^l in Assumption 1 has not been defined before\\n\\nQ4. What is \\u03c8 in MSBE? This term has not been defined before.\\n\\nQ5. How sensitive is the algorithm to the choice of hyperparameters such as \\u03bb, \\u03b3, \\u03b7?\\n\\nQ6. How does Theorem 2 differ from Theorem 4.7 in (Jin et al., 2021)?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper investigates the challenging problem of knowledge transfer in reinforcement learning and its application to solving the issue of scarce datasets in offline reinforcement learning. Specifically, the author integrates feature extraction through feature engineering into the Bellman update process, mathematically modeling the extraction of features from source data that are similar to those in target data. From the perspective of algorithm design, this is an elegant end-to-end approach. However, the paper does not provide any experimental validation, which makes it hard to directly prove that the algorithm works in practice.\\n\\nI think the author's idea interesting. If the author can address my questions and concerns, I will increase the score.\", \"soundness\": \"3\", \"presentation\": \"4\", \"contribution\": \"3\", \"strengths\": \"# Written\\nThe author has defined the reinforcement process very clearly. If the author's paper aims to address theoretical issues in knowledge transfer of reinforcement learning domain, then this paper is undoubtedly well-written. However, if the author hopes to extend the algorithm to practical applications, it is recommended to refer to the most effective reinforcement learning algorithms for further improvement. In particular, I mentioned some algorithm design-related issues in the question session.\\n\\n# Soundness\\n\\n**Importance.** The author has investigated a critically important issue in reinforcement learning. Knowledge transfer poses challenges in reinforcement learning, even within a single environment, making it far from straightforward. Additionally, the problem studied by the author is offline reinforcement learning, thus the author's research simultaneously addresses the challenge of scarcity of data in offline reinforcement learning.\\n\\n**Reasonability.** I believe the author's method is reasonable. The author employs kernel techniques to map the features of the target domain and source domain into a high-dimensional space, and filters out similar features between domains based on the similarity between matrices, thereby achieving data augmentation. Additionally, this method does not require pre-training any estimators and has a relatively rigorous mathematical foundation.\", \"weaknesses\": \"I consider the author's idea quite interesting, but it has not been designed with reference to the current state-of-the-art reinforcement learning algorithms. I have already raised questions related to the algorithm design during the question session. Furthermore, although the author's algorithm is theoretically sound, no experiments have been provided for verification. If the author's primary goal is to address theoretical issues, then even a simple demo experiment could be used to validate the theoretical claims.\\n\\nAdditionally, please see questions.\", \"questions\": \"# Algorithm\\n\\n**Q1:** In line 16 of the pseudocode, truncating Q is reasonable under the offline setting, but why does the truncated bond have a relationship with the sequence length?\\n\\n**Q2:** In line 17 of the pseudocode, it is suggested to use $\\\\log \\\\pi$ instead of $\\\\pi$. $Q\\\\log\\\\pi$, known as actor-critic, has been proven superior both experimentally and theoretically in reinforcement learning.\\n\\n**Q3:** When modeling the Bellman equation, you chose to use $V(s')$ instead of $Q(s', a')$, where $a'\\\\sim\\\\pi(\\\\cdot|s')$. The advantage of this is that it avoids accessing out-of-distribution (OOD) actions. However, there are also disadvantages, such as infinitely approaching the behavior policy, which may prevent the current algorithm from surpassing the behavior policy. Could you consider introducing the expected regression from Implicit Q Learning, which could potentially allow the algorithm to learn better decisions than the behavior policy?\\n\\n[1] Ilya Kostrikov, Ashvin Nair, Sergey Levine. Offline Reinforcement Learning with Implicit Q-Learning\\n\\n# Theoretical Proof\\n\\nI am still reviewing the mathematical proofs and may have additional questions or comments late.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}" ] }

[ "Hyosoon Jang", "Yunhui Jang", "Jaehyung Kim", "Sungsoo Ahn" ]

[ "Large language model", "molecular generative model", "drug discovery" ]

{ "note_id": [ "yXOuSVyJ1Q", "uTk7k04Ynv", "tnzguMSaSN", "rAmO9qkqjP", "lOm0YLNxlb", "kZsbZHHUkL", "jFTZopcTk3", "V8Ap5tosCX", "MuBndhyHCu", "Kdfw21Dvbl", "FIy3D5lUvj", "CAbnxXYG6N", "7yZtra3ZHR", "6neTwPUvdB", "5JaO2cnEvY", "0wl0iljeEA" ], "note_type": [ "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_review", "official_comment", "meta_review", "official_comment", "official_comment", "official_review", "decision" ], "note_created": [ 1730481896649, 1732456477776, 1732241701117, 1732651026286, 1732242794650, 1732501975140, 1732685445060, 1730589203090, 1732243125108, 1730710293380, 1732718796589, 1734864492900, 1732241802208, 1732243173094, 1730374034754, 1737523585598 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission3613/Reviewer_oKU9" ], [ "ICLR.cc/2025/Conference/Submission3613/Reviewer_xfWi" ], [ "ICLR.cc/2025/Conference/Submission3613/Authors" ], [ "ICLR.cc/2025/Conference/Submission3613/Reviewer_APHU" ], [ "ICLR.cc/2025/Conference/Submission3613/Authors" ], [ "ICLR.cc/2025/Conference/Submission3613/Authors" ], [ "ICLR.cc/2025/Conference/Submission3613/Authors" ], [ "ICLR.cc/2025/Conference/Submission3613/Reviewer_FZz8" ], [ "ICLR.cc/2025/Conference/Submission3613/Authors" ], [ "ICLR.cc/2025/Conference/Submission3613/Reviewer_APHU" ], [ "ICLR.cc/2025/Conference/Submission3613/Reviewer_oKU9" ], [ "ICLR.cc/2025/Conference/Submission3613/Area_Chair_KboK" ], [ "ICLR.cc/2025/Conference/Submission3613/Authors" ], [ "ICLR.cc/2025/Conference/Submission3613/Authors" ], [ "ICLR.cc/2025/Conference/Submission3613/Reviewer_xfWi" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ] ], "structured_content_str": [ "{\"summary\": \"The paper proposes a fine-tuning method to generate diverse molecules using LLMs. The method involves supervised fine-tuning (SFT) and reinforcement learning (RL).\\n\\nFor SFT, the LLM is prompted to generate a molecule with a given property. The prompt is sampled many times, the generated molecules are filtered and then concatenated to create the SFT training set. The SFT step uses a prompt describing the property and requesting a set of molecules. The set of molecules generated previously is appended to this prompt.\\n\\nFor RL, the SFT prompt is used to generate a sequence of molecules. Every time a molecule is generated, the LLM policy is updated using Proximal Policy Optimization (PPO). The reward is based on how well the molecule matches a property and how different it is from the previous molecules.\", \"soundness\": \"4\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"**Originality**\\nTo my knowledge, the method is novel and the problem is under-researched.\\n\\n**Quality**\\nThe SFT and RL methods are clear and intuitive. The algorithms are well described in the text and code is provided in Supplementary. Figure 4 helps the understanding of the method. The gap in existing LLMs at generating structurally diverse molecules is well explained in the text and in figure 2. The baselines and metrics are mostly well chosen. The results are convincing, with the authors' method producing significant increase in the chosen metrics.\\n\\n**Clarity** The methods are explained well. The figures are easy to read. Including different texture in Figure 5 helps the readability. The Appendix with all the experimental details and the code in Supplementary is appreciated and helps reproducibility. I also appreciate the title and claims being straightforward and to the point. \\n\\n**Significance** Generating diverse molecules is important in discovery workflows, thus the problem here is significant. The proposed method is easy to understand and implement. The paper also inspires a few interesting follow-ups and thus I think is an important addition to the community.\", \"weaknesses\": \"1) In the experiments with BioT5+ and MolT5, the authors used the BLEU score on prompts and molecules from ChEBI-20. I think BLEU is a misleading score in the text to molecule task. For example, in BLEU, the order of words in a sentence matters, while there are many ways of writing a molecules as SMILES strings. Also, small changes in a molecule SMILES in the right place can have high changes in a property, such as in the case of hydrogen bond donors and acceptors. I recommend the authors discuss this in their manuscript and compare it with scores generated by RDKit (which they used in the other experiments).\\n\\n2) There are a few minor things that could improve the paper: \\n - Figure 7 could have different symbols. It was hard to read in a black-and-white version of the paper. \\n - Worth considering giving the method a name, this will help wider adoption.\", \"questions\": \"1) Can you improve the explanation of NCircles? I think I get it, but given it's a relatively new method I yet don't have an intuition for it like I do for things like T-SNE or UMAP. What kinds of hyperparameters does it have? What are the conditions for two nodes to be close to each other? Do the overlap of the circles mean anything or are they a consequence of the force-directed algorithm?\\n\\n2) In Table 1, you show the outputs of your method for the description \\\"The molecule is a primary aliphatic ammonium ion which is obtained from streptothricin F by protonation of the guanidino and amino groups. It has a role as an antimicrobial agent. It is a guanidinium ion and a primary aliphatic ammonium ion. It is a conjugate acid of a streptothricin F\\\". Are all the generated molecules a conjugate acid of streptothricin F? Do they have the correct scaffold? This would be important to understand and is related to limitations of the BLEU score.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you for your detailed response and efforts in addressing my concerns. I appreciate the updates and clarifications provided in the revised manuscript. However, I continue to view **Weakness 2**\\u2014the suitability of applying and evaluating diverse molecular generation on the ChEBI-20 dataset\\u2014as a significant issue.\\n\\nWhile I understand the argument that diversity could increase the likelihood of covering the target molecule, I believe that **diversity is not always an appropriate objective for molecule design**. For instance, in tasks like molecule optimization, the goal is to generate new compounds similar to a reference molecule. In such cases, similarity, rather than diversity, is the desired outcome, which is directly contrary to the focus of this paper.\\n\\nMoreover, the key baselines in your paper, **MolT5 and BioT5+**, are designed to generate a single desirable molecule that matches the given description, rather than a diverse set of molecules. This highlights a fundamental mismatch between the baselines\\u2019 objectives and your proposed evaluation, which prioritizes diversity. This disconnect raises concerns about the **experimental validity of comparing your method with these baselines**.\\n\\nTherefore, I find the experimental setup of the paper to be unreasonable and not aligned with the goals of diverse molecular generation in the context of the dataset and baselines used. As a result, I do not change my rating.\"}", "{\"comment\": \"Dear reviewer APHU,\\n\\nWe express our deep appreciation for your time and insightful comments. In our updated manuscript, we highlight the changes in $\\\\color{blue}{\\\\text{blue}}$. \\n\\nIn what follows, we address your comments one by one.\\n\\n---\\n**W1. The two-stage fine-tuning approach itself is not that novel.** \\n\\nThis is not a weakness of our work, since using the two-stage fine-tuning approach is not our main idea. Our idea to leverage LLMs for diverse molecule generation is novel. We design new reward functions for maximizing the structural diversity of generated outputs. We newly construct experiments and metrics for comparing the diversity of our algorithm and existing LLM decoding algorithms. \\n\\n---\\n\\n**W2. Although tested on standard datasets and metrics, this works lacks experimental or real-world validation of the practical utility of the generated molecules.**\\n\\nEvaluating on the standard datasets and metrics is sufficient, since the datasets and metrics were designed to be representative of the practical utility of the generated molecules [1,2,3]. Our belief aligns with the vast literature of computational methods for designing new molecules [4,5,6,7]. \\n\\n---\\n\\n**W3. The proposed approach may require extensive setup or hyperparameter tuning, which can hinder future adoption and reproducibility.**\\n\\nWe clarify that our approach requires minimal setup and hyperparameter tuning due to following the widely used two-stage fine-tuning approach. Our work requires a similar number of hyperparameters compared to many of the two-stage fine-tuning approaches. We ensure reproducibility by providing the code in the **Supplementary Materials**, specifying hyperparameter values in **Appendix B**. We also plan to release the full parameters of the fine-tuned models once our work is accepted.\\n\\n---\\n\\n**W4. The authors overstate that (1) \\\"the first to explore the use of LLMs for generating diverse molecule\\\" and (2) \\\"the first propose a fine-tuning approach for LLMs to generate diverse solution\\\".**\\n\\nWe do not think this is an overstate, since we were unable to find any works that claim (1) and (2). We will be happy to tone down our claims if you could provide any reference that already claims (1) and (2).\\n\\n---\\n\\n[1] Edwards et al., Text2Mol: Cross-Modal Molecule Retrieval with Natural Language Queries, EMNLP 2021\\n\\n[2] Polykovskiy et al., Molecular Sets (MOSES): A Benchmarking Platform for Molecular Generation Models, Frontiers in Pharmacology 2020\\n\\n[3] Xie et al., How Much Space Has Been Explored? Measuring the Chemical Space Covered by Databases and Machine-Generated Molecules, ICLR 2023\\n\\n[4] Hu et al. De novo drug design using reinforcement learning with multiple gpt agents, NeurIPS 2023\\n\\n[5] Pei et al., BioT5+: Towards Generalized Biological Understanding with IUPAC Integration and Multi-task Tuning, ACL 2024\\n\\n[6] Yu et al., LlaSMol: Advancing Large Language Models for Chemistry with a Large-Scale, Comprehensive, High-Quality Instruction Tuning Dataset, COLM 2024\\n\\n[7] He et al. Evaluation of reinforcement learning in transformer-based molecular design, Journal of Cheminformatics 2024\"}", "{\"comment\": \"Thank you for your detailed response! While it addressed some of my concerns, I still have the following questions:\\n\\n1. \\\"the first to explore the use of LLMs for generating diverse molecule\\\"\\nYour work focuses on using reinforcement learning (RL) to redesign the reward function and improve diversity, rather than leveraging Large Language Models (LLMs) themselves to generate diverse molecules. The pretraining stage didn't enhance diversity and the pertaining approach is not novel either. RL fine-tuning play the role of improve diversity.\\n\\n2. \\\"the first propose a fine-tuning approach for LLMs to generate diverse solution\\\"\\nYour proposal to fine-tune LLMs is not that novel, as similar approaches have been explored before (e.g., Reinvent [1]),(they didn't use your redesigned reward function). Your method uses the PPO algorithm, and the multiple stages can be regarded as multiple epochs in traditional RL. Therefore, novelty leads to diversity still relies on the redesigned reward function, but it's quite limited.\\n\\nTherefore, I will keep my score. \\n\\n[1]Reinvent 4: Modern AI\\u2013driven generative molecule design\"}", "{\"comment\": \"Dear reviewer oKU9,\\n\\nWe express our deep appreciation for your time and insightful comments. In our updated manuscript, we highlight the changes in $\\\\color{blue}{\\\\text{blue}}$. \\n\\nIn what follows, we address your comments one by one.\\n\\n---\\n\\n**W1. The authors should (1) discuss the limitation of using BLEU scores in comparing generated molecules and the target molecule satisfying a given description and (2) consider other RDKit-based scores.** \\n\\n\\nWe considered BLEU scores following prior studies [1] on text-to-molecule generation. However, we acknowledge their limitations in comparing the generated molecules with the target molecule: they simply measure textual similarity of SMILES strings that may fail to capture the structural or functional information. \\n\\nTo address your concerns, in **Appendix D.2** of our updated manuscript (also in **sixth footnote** in **Page 7**), we (1) discuss the limitations of BLEU scores and (2) conduct additional experiments by replacing BLEU scores with Tanimoto and Dice scores [2] that are more concrete to capture the structural and functional information. One can observe that our fine-tuned model, even though it is trained with the BLEU score, consistently outperforms the baselines in these experiments. We also plan to fine-tune LLMs using Dice scores instead of BLEU scores and include the results in the Appendix.\\n\\n\\n\\n\\n\\n---\\n\\n**W2. About minor errors or recommendation.**\\n\\nThanks you for valuable suggestions! In our updated manuscript, we modify the symbols in **Figure 7**. We would like to defer naming until the end of the discussion period.\\n\\n\\n---\\n\\n**Q1. Can authors improve the explanation of NCircles metric?**\\n\\n\\nTo address your comment, we have modified explanation of NCircles metric in **Section 4.1** of our updated manuscript and provided more detailed explanation in **Appendix A.3** of our updated manuscript.\\n\\n\\nTo measure the diversity of a given set of molecules $\\\\mathcal{M}$, the NCircles metric computes the size of the largest subset of molecules in which no two molecules are similar to each other. Specifically, this metric is defined with a Tanimoto similarity $T(\\\\cdot,\\\\cdot)$ and a similarity threshold $h$ as follows [3]:\\n\\\\begin{equation}\\n{{\\\\text{NCircles}}_{h}} = \\\\max _{\\\\mathcal{C}\\\\subseteq\\\\mathcal{M}} |\\\\mathcal{C}|\\\\quad \\\\text{s.t. }T(x,y)<h, \\\\forall x \\\\neq y \\\\in \\\\mathcal{C},\\n\\\\end{equation}\\nwhere $\\\\mathcal{C}$ is a subset of molecules and every pair of molecules in $\\\\mathcal{C}$ should have a similarity lower than $h$. The high NCircles metric implies that the given set of molecules covers a larger volume of molecular space, indicating diversity [3]. We measured this by defining $\\\\mathcal{M}$ as the set of accepted molecules to capture quality and diversity.\\n\\n\\n\\n> What kinds of hyperparameters does it have? What are the conditions for two nodes to be close to each other? Do the overlap of the circles mean anything or are they a consequence of the force-directed algorithm?\\n\\nTherefore, the NCircles metric does not have hyperparameters. Additionally, we guess that the latter two questions stem from **Figure 6** which is just illustrated to explain NCircles in a 2-D space. Here, two nodes are close to each other when they have high similarity, the circle implies the boundary of threshold $h$, and the overlap of circles is just the consequence of the force-directed algorithm.\\n\\n\\n\\n---\\n\\n**Q2. In Table 1, do the generated molecules correctly follow the given description?**\\n\\nUnfortunately, it is hard to directly evaluate whether the generated molecules follow the given description due to the complexity of qualitative description. However, it is worth noting that the results in **Figure 7** are measured by evaluating whether the generated molecules exactly follow the molecular quantitative description, e.g., QED value or the number of hydrogen bond donors. This shows how generated molecules from our method correctly follow the given descriptions. \\n\\n\\n\\n\\n---\\n\\n[1] Pei et al., BioT5+: Towards Generalized Biological Understanding with IUPAC Integration and Multi-task Tuning, ACL 2024\\n\\n[2] Bajusz et al, Why is Tanimoto index an appropriate choice for fingerprint-based similarity calculations?, Journal of Cheminformatics 2015\\n\\n[3] Xie et al., How Much Space Has Been Explored? Measuring the Chemical Space Covered by Databases and Machine-Generated Molecules, ICLR 2023\"}", "{\"comment\": \"Thank you for the response! We would like to clarify the following points to address your comments.\\n\\n---\\n\\n**W2-1. Diversity is not always an appropriate objective as similarity to a reference molecule can be significant rather than diversity, e.g., a molecular optimization task given a reference molecule.**\\n\\n\\nFirst, we would like to clarify that we have not only considered the diversity but also considered the quality with $r_ {\\\\text{match}}$ which can be defined with the similarity to the reference molecule in the molecular optimization tasks you mentioned. In these tasks, we would like to clarify that both similarity to a reference molecule and diversity are still important, as focusing only on similarity may generate multiple molecules similar to the reference but identical to other generated molecules, i.e., the set of generated molecules may not be informative. \\n\\nAdditionally, in the experiments on the ChEBI-20 dataset, we have already considered a setting similar to the settings mentioned above, where $r_ {\\\\text{match}}$ is computed with the similarity to the target molecule satisfying the given description (as described in **Equation 4** of **Appendix B.2**) to generate diverse molecules similar to the target molecule. We also have considered this in the evaluation (see response of W2-2).\\n\\nBesides, there exist plenty of significant molecular optimization tasks where diversity is useful, e.g., drug diversification [1,2]. We firmly believe our algorithm will be useful in these scenarios.\\n\\n\\n\\n---\\n\\n\\n\\n**W2-2. As MolT5 and BioT5$^{+}$ are designed to generate a single desirable, there are concerns about (1) mismatch between the baselines objectives and your proposed evaluation which prioritizes diversity and (2) experimental validity in comparing your method with MolT5 and BioT5$^{+}$ do not consider the diversity.**\\n\\nFor (1), we would like to clarify that we have also considered the baseline objectives: evaluating how the generated molecules are similar to the target molecules, in defining the set of accepted molecules (as described in **Line 348**) which is also defined as $\\\\mathcal{M}$ to measure $\\\\text{NCircles}$ (as described in **Line 350**), and in defining $\\\\text{Top }10$ scores (as described in **Line 359**), i.e., our comparisons do not only focus on diversity.\\n\\n\\nFor (2), we would like to clarify that we compare our method with decoding schemes for diversified generation, e.g., beam search, rather than MolT5 and BioT5$^{+}$ which serve backbones. Our experiments are valid as they aim to show that applying existing decoding schemes is insufficient to obtain diverse and high-quality molecules similar to the target, while our method addresses these limitations.\\n\\n\\n---\\n\\n[1] Krantz et al., Diversification of the drug discovery process, Nature Biotechnology 1998\\n\\n[2] Nippa et al., Enabling late-stage drug diversification by high-throughput experimentation with geometric deep learning, Nature Chemistry 2024\"}", "{\"comment\": \"Thank you for your response. It seems that you are still concerned that we are over-claiming our contributions. We further provide our response.\\n\\n---\\n\\n**W4-1. \\\"the first to explore the use of LLMs for generating diverse molecule\\\" Your work focuses on using RL to improve diversity, rather than leveraging LLMs to generate diverse molecules.** \\n\\nOur work explicitly focuses on the idea of leveraging the remarkable expressive power of LLMs to (1) understand the textual conditions and (2) capture the long context for sequentially generating molecules. We use the off-the-shelf RL algorithms since this is not our main focus. \\n\\n---\\n\\n**W4-2. \\\"the first propose a fine-tuning approach for LLMs to generate diverse solution\\\" Your proposal to fine-tune LLMs is not that novel, as similar approaches have been explored before (e.g., Reinvent [1]).**\\n\\nWe would like to clarify that Reinvent [1] did not use LLMs and just chose transformer-based model architectures. We also note that we propose to autoregressively generate the molecules in a concatenated sequence, exploiting the power of LLMs, which was not considered in previous works (including [1]).\\n\\n\\n---\\n\\n[1] Loeffler et al. Reinvent 4: Modern AI\\u2013driven generative molecule design, Journal of Cheminformatics 2024.\"}", "{\"summary\": \"This paper investigates the limitations of current large language models (LLMs) in generating structurally diverse molecules and proposes a novel two-stage fine-tuning approach to address this challenge. Specifically, the authors apply supervised fine-tuning to enable LLMs to generate a sequence of molecules and subsequently leverage reinforcement learning to enhance structural diversity. Experimental results demonstrate that the proposed method improves molecular diversity compared to existing decoding strategies.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"This paper identifies a limitation in the molecular generation capabilities of LLMs and proposes a diversity enhancement strategy based on autoregressive generation and reinforcement learning.\", \"weaknesses\": \"1. The method directly adopts LLMs with reinforcement learning. Thus, the technical novelty of the method is limited.\\n2. To help readers understand the impact of the approach, the authors could provide additional explanations for the diversity metrics or introduce more similarity evaluation methods.\", \"questions\": \"1. In Table 3, it is unclear why the comparison is made when the number of generated molecules is different. Additionally, all relevant metrics should be included in the comparison to provide a comprehensive assessment.\\n\\n2. In Table 4, the authors should specify which large language model (LLM) was used as the basis for fine-tuning and reinforcement learning in the compared methods. Furthermore, when comparing Tables 2 and 4, the performance improvement of the Div-SFT approach appears to be minimal.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Dear reviewer xfWi,\\n\\nWe express our deep appreciation for your time and insightful comments. In our updated manuscript, we highlight the changes in $\\\\color{blue}{\\\\text{blue}}$. \\n\\nIn what follows, we address your comments one by one.\\n\\n---\\n\\n**W1. The paper lacks discussion about relevant RL-based methods for diversity in molecular generation.** \\n\\nThank you for the valuable suggestion! In **Appendix A.1** of our updated manuscript (also in **third footnote** in **Page 3**), we incorporate discussions on prior RL methods for generating diverse molecules and describe how our work differs from them: they learn a model focused on a fixed target molecular property, whereas our work fine-tunes LLMs to generate diverse molecules given a prompt flexibly describing various target properties.\\n\\n\\n---\\n\\n**W2. Applying and evaluating diverse molecular generation on ChEBI-20 datasets may not suitable as a molecular description can be corresponded to a single molecule.**\\n\\n\\nTo address this comment, we would like to re-emphasize that the diverse molecular generation aims to increase the probability of the target molecule being included in the generated set of molecules by exploring a wide chemical space. Even for solving problems with one ground truth answer, generating diverse candidates increase probability of including the answer and being informative.\\n\\n\\n\\n\\n---\\n\\n**W3. It would be beneficial to introduce coefficient terms for description-matching and diversity rewards.**\\n\\nThanks you for your suggestion! In our updated manuscript, We incorporate the coefficient terms $\\\\lambda_{\\\\text{match}}$ and $\\\\lambda_{\\\\text{div}}$ for description-matching and diversity rewards, respectively.\\n\\n---\\n\\n**W4. In Figure 5, it seems that NCircles with $h=0.65$ and $h=0.75$ metrics do not follow the original metric definition where values at a higher $h$ should logically be lower.**\\n\\n\\nFirst, we would like to clarify that our NCircles metrics follow the original definition of NCircles metric and are logically correct, but we found that the description of NCircles in our paper is insufficient. We have modified the description of NCircles in our updated manuscript with the following contents.\\n\\n\\nGiven a set of molecules $\\\\mathcal{M}$, the NCircles metric computes the size of the largest subset of molecules in which no two molecules are similar to each other. Specifically, it is defined with a Tanimoto similarity $T(\\\\cdot,\\\\cdot)$ and similarity threshold $h$ as follows:\\n\\\\begin{equation}\\n\\\\text{NCircles}_{h}=\\\\max _{\\\\mathcal{C}\\\\subseteq\\\\mathcal{M}} |\\\\mathcal{C}| \\\\quad \\\\text{s.t. }T(x,y)<h, \\\\forall x \\\\neq y \\\\in \\\\mathcal{C},\\n\\\\end{equation}\\nwhere $\\\\mathcal{C}$ is a subset of molecules. Every pair of molecules in $\\\\mathcal{C}$ should have a similarity lower than $h$. Consequently, the value of NCircles metric is increased with respect to increasing $h$ as the condition becomes loose, e.g., $\\\\mathcal{C}=\\\\mathcal{M}$ when $h>1$. \\n\\nAdditionaly, we would like to clarify that this is identical to the definition of NCircles metric in the original paper [1], which is defined with a Tanimoto distance $d(x,y)=1-T(x,y)$:\\n\\\\begin{equation}\\n\\\\text{NCircles} ^{\\\\text{dist}} _{t} = \\\\max _{\\\\mathcal{C}\\\\subseteq\\\\mathcal{M}} |\\\\mathcal{C}| \\\\quad \\\\text{s.t. } d(x,y)>t, \\\\forall x \\\\neq y \\\\in \\\\mathcal{C},\\n\\\\end{equation}\\nwhere $t$ is a distance threshold and $\\\\text{NCircles} ^{\\\\text{dist}} _{t}=\\\\text{NCircles} _{1-h}$. \\n\\n---\\n\\n**Q1. Is it reasonable to represent each cluster as a circle in Figure 6? I wonder if this visualization accurately reflects clustering in chemical space.**\\n\\nWe would like to clarify that the circle in **Figure 6** is just a visualization purpose for explaining the NCircles in 2-D space. Therefore, this visualization does not accurately reflect clustering in chemical space. Specifically, we draw **Figure 6** with the Fruchterman-Reingold force-directed algorithm used in the spring_layout of the NetworkX package [2], where the edge weight is defined with the Tanimoto similarity $T(x,y)$.\"}", "{\"summary\": \"This paper proposes a two-stage fine-tuning approach to address the challenge of generating diverse molecules: (1) supervised fine-tuning to enable LLMs to autoregressively generate molecules in a sequence, and (2) reinforcement learning to enhance structural diversity among the generated molecules.\\nThe contributions of this work include introducing an innovative method that combines supervised learning and reinforcement learning to generate structurally diverse molecules. Additionally, the paper demonstrates that this approach surpasses existing LLMs and traditional decoding methods in producing high-quality, diverse molecular structures.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1.This paper addresses an important need in molecular generation by developing a method that enhances structural diversity, a key factor in drug discovery.\\n2. The two-stage fine-tuning approach, which combines supervised learning with reinforcement learning (RL) for diversity maximization well-designed solution. \\n3. Through empirical testing, the paper shows that its approach outperforms both current LLM-based molecular generators and advanced decoding schemes, highlighting the practical advantages of its method for achieving high-quality, diverse outputs.\\n4. The proposed method is adaptable and could be extended to other domains requiring diversity in generated outputs, such as protein design or materials discovery. This versatility increases the impact of the work beyond just molecular generation.\", \"weaknesses\": \"1. The two-stage fine-tuning approach which is highlighted in this paper is not that novel.\\n2. Although the model is tested on standard datasets and metrics for structural diversity, it lacks experimental or real-world validation to demonstrate the generated molecules\\u2019 practical utility in drug discovery.\\n3. The proposed method\\u2019s multi-stage fine-tuning approach may require extensive hyperparameter tuning and setup, which can hinder reproducibility and adoption among researchers.\\n4. The paper claim \\\"the first to explore the use of LLMs for generating diverse molecule\\\" and \\\"the first propose a fine-tuning approach for LLMs to generate diverse solution\\\", which is overstate.\", \"questions\": \"See weakness\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Many thanks for addressing my comments. I will keep my score.\\n\\n---\\n\\n**WP1.**\\nThank you for adding the additional table. Using Tanimoto and Dice scores is more convincing.\\n\\n**Q1.**\\nDescribing NCircles by computing \\\"the size of the largest subset in which no two molecules are similar to each other (Tanimoto similarity below a threshold h)\\\" is clear.\\n\\n**Q2.**\\nAgreed, Figure 7 helps understand the match to description.\\n\\n---\"}", "{\"metareview\": \"In this work, authors propose a two-stage fine-tuning approach for large language models (LLMs) to generate diverse molecular structures for drug discovery applications. The method combines supervised fine-tuning to enable autoregressive molecule generation followed by reinforcement learning optimization to maximize structural diversity while maintaining adherence to target properties. The authors evaluate their approach against existing LLMs and decoding schemes using various diversity metrics.\\n\\nThere are several strengths for the work as follows. The paper addresses an important need in computational drug discovery by focusing on molecular diversity generation. The methodology is clearly presented with detailed experimental protocols and comprehensive evaluations across multiple metrics (Reviewer oKU9). The authors provide thorough empirical validation showing improvements over baseline approaches in generating structurally diverse molecules that match target properties. The supplementary materials include code and detailed experimental settings that support reproducibility.\\n\\nHowever, there are several limitations for the work as evident from the discussion during the review period. The two-stage fine-tuning approach largely adopts existing methods without substantial innovation (Reviewer APHU, FZz8). While the authors argue their contribution lies in applying these techniques to molecular diversity generation, the core technical components are standard. Reviewer xfWi raises significant issues about the appropriateness of using ChEBI-20 for diversity evaluation, since it maps descriptions to single target molecules. While the authors argue diversity remains valuable even with single targets, this creates a fundamental mismatch between the baseline models' objectives and the proposed evaluation metrics. Multiple reviewers (APHU, FZz8) note that the paper's claims of being ``first'' to explore LLMs for diverse molecule generation are overstated, given existing work in this space. Further, the reliance on BLEU scores for molecular similarity is problematic given the multiple valid SMILES representations for identical molecules (Reviewer oKU9). While the authors added Tanimoto and Dice scores in revision, this highlights initial gaps in the evaluation approach.\\n\\nBased on these comments, the manuscript cannot be recommended for acceptance at this point. To strengthen future submissions, the authors should: (1) develop a more appropriate evaluation framework that fairly compares diversity-oriented approaches, (2) more precisely position their technical contributions relative to existing work, and (3) strengthen the novelty of their methodological approach beyond standard fine-tuning techniques. More comments below.\", \"additional_comments_on_reviewer_discussion\": \"During the discussion, the following important points were noted.\\n\\n1. On technical novelty, reviewer APHU initially questioned the innovation of the two-stage fine-tuning approach and stated that the paper over-claimed its contributions. The authors responded by clarifying that their novelty lies not in the two-stage approach itself, but in leveraging LLMs specifically for diverse molecule generation. However, APHU remained unconvinced, noting that the diversity improvements stem primarily from the reward function redesign rather than novel use of LLMs.\\n\\n2. Reviewer FZz8 raised concerns about the limited performance improvement of the Div-SFT approach and requested clarification about which LLM was used as the baseline. The authors addressed this by explaining that Div-SFT is merely a subroutine and not the final model - the focus should be on Div-RL results. They also clarified that BioT5+ was used as the baseline.\\n\\n3. Reviewer oKU9 highlighted issues with using BLEU scores for molecular similarity evaluation and requested better explanation of the NCircles metric. The authors acknowledged these limitations and added Tanimoto and Dice scores in their revision. They also provided a clearer mathematical formulation of NCircles. oKU9 was satisfied with these additions and maintained their positive assessment.\\n\\n4. The most critical discussion emerged from reviewer xfWi's concerns about the fundamental mismatch between evaluation objectives and baseline models when using the ChEBI-20 dataset. While the authors argued that diversity remains valuable even for single-target problems, xfWi maintained that this creates an unfair comparison since baselines like MolT5 and BioT5+ were designed for single-molecule generation. The authors' final response attempted to justify their evaluation framework by emphasizing that they compared against diverse decoding schemes rather than the base models, but this did not fully address the core experimental design concern.\\n\\nIn weighing these discussions, I found that while the authors made good efforts to address many technical clarifications and evaluation metrics, the fundamental concerns about experimental validity raised by xfWi and the limited technical novelty highlighted by APHU remained inadequately resolved. These unresolved issues, combined with the over-claimed contributions noted in multiple reviews, ultimately support the rejection decision despite the authors' detailed responses during the rebuttal period.\\n\\nThe discussion reveals that while the authors were responsive and made several improvements to their manuscript, the core issues affecting the paper's suitability for ICLR publication - experimental design validity and technical innovation - were not sufficiently addressed through the rebuttal process.\"}", "{\"comment\": \"Dear reviewer FZz8,\\n\\nWe express our deep appreciation for your time and insightful comments. In our updated manuscript, we highlight the changes in $\\\\color{blue}{\\\\text{blue}}$. \\n\\nIn what follows, we address your comments one by one.\\n\\n---\\n\\n**W1. The method directly adopts LLMs with reinforcement learning. The technical novelty of the method is limited.**\", \"we_would_like_to_clarify_that_the_novelty_of_our_work_lies_in_introducing_a_new_concept\": \"leveraging LLM fine-tuning to generate diverse molecules, rather than simply proposing to adopt LLMs with reinforcement learning. We also note that our work is the first to explore the use of reinforcement learning for generating diverse output with LLMs, which has not been previously addressed in the LLM literature to the best of our knowledge.\\n\\n---\\n\\n**W2. To help readers understand the impact of the approach, can authors provide additional explanations for the diversity metrics or introduce more similarity evaluation methods?**\\n\\nThanks you for valuable suggestions! To address your comments, we provide additional explanations for the diversity metrics in **Appendix A.3** of our updated manuscript. We also provide additional explanations of various similarity evaluation methods, e.g., Tanimoto, Dice, and cosine similarities, in **Appendix A.2** of our updated manuscript. Additionally, we update **first footnote** in **Page 1**, **Line 200** in **Section 3**, **fourth footnote** in **Page 4**, and **Line 343** in **Section 4.1** to provide high-level explanations of molecular diversity and similarity.\\n\\n\\n\\n---\\n\\n**Q1. In Table 3, why is the comparison made when the number of generated molecules differs? Why did authors not include all relevant metrics?** \\n \\n \\n \\nThe comparison with different numbers of generated molecules aims to show that our method consistently discovers more diverse molecules even when the baseline generates a larger number of molecules.\\n\\nTo address your comment, in our updated manuscript, we include all relevant metrics in **Table 10** of **Appendix D.1**.\\n\\n---\\n\\n**Q2-1. What LLM is used as the basis in Table 4?**\\n\\nWe considered BioT5$^{+}$ in **Table 4**. In our updated manuscript, we clarify this in the captions of **Table 4**.\\n\\n---\\n\\n**Q2-2. The performance improvement of Div-SFT seems to be minor.**\\n\\nThe performance Div-SFT is not appropriate for evaluating the benefits of our algorithm, since it is a subroutine of our algorithm. Instead, one should focus on the result of reinforcement learning (Div-RL), which is our final model. The main purpose of Div-SFT is not to improve the performance, but to repurpose the base LLMs for generating multiple molecules before Div-RL.\"}", "{\"comment\": \"**Q2. Have the authors considered the randomness of SMILES strings in measuring metrics? What effect does the randomness of SMILES have on the stability of method or evaluation?**\\n\\n\\nTo remove the randomness stemming from the non-uniqueness of SMILES, i.e., two distinct SMILES strings represent an identical molecule, we applied canonicalization [3]: converts SMILES strings representing an identical molecule into a unique SMILES string. We clarify this in **Line 315** of our updated manuscript.\\n\\n\\n\\n---\\n\\n[1] Xie et al., How Much Space Has Been Explored? Measuring the Chemical Space Covered by Databases and Machine-Generated Molecules, ICLR 2023\\n\\n[2] Hagberg et al., Exploring network structure, dynamics, and function using NetworkX, 2008\\n\\n[3] Weininger et al., SMILES. 2. Algorithm for generation of unique SMILES notation, Journal of chemical information and computer sciences 1989\"}", "{\"summary\": \"This paper tackles an essential problem in AI-driven drug discovery: generating structurally diverse molecules using LLMs. Addressing the challenge that traditional LLMs often produce structurally similar molecules from a single prompt, the authors propose a two-stage approach: (1) supervised fine-tuning to enable autoregressive mutiple molecule generation, and (2) RL to enhance structural diversity. Through experiments with several LLMs and comparison against various decoding schemes, the proposed approach, Div-SFT+RL, demonstrates an improved ability to generate diverse molecular structures, which is crucial for discovering viable drug candidates. The results indicate that Div-SFT+RL outperforms existing models on diversity-related metrics, supporting its potential for applications in drug discovery.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. **Focus on an Essential Problem**: The paper addresses a critical requirement in drug discovery\\u2014molecular diversity in generated candidates. By providing a more diverse set of molecules, the proposed method aligns with real-world drug discovery processes, enhancing the chance of identifying successful compounds.\\n2. **Clear Presentation**: The paper is well-structured, with a clear exposition of the methodology and experimental procedures.\\n3. **Detailed Experiments**: The authors perform thorough evaluations across multiple metrics and baselines, including the comparison of Div-SFT+RL with various decoding schemes. The use of IntDiv, and NCircles as diversity-related metrics offers a comprehensive view of the method\\u2019s performance.\", \"weaknesses\": \"1. **Limited Discussion of Related Works on RL for Diversity in Molecular Generation**:\\n The paper lacks an in-depth discussion of prior reinforcement learning (RL) approaches that also aim to improve molecular diversity, particularly those using RL for targeted or diverse molecule generation. Works such as Blaschke et al. (2020), Pereira et al. (2021), Hu et al. (2023), and He et al. (2024) provide diverse RL strategies in molecular design, which may not all involve LLMs but offer relevant insights on RL methods and should be discussed to position this work better.\\n\\n [1] Blaschke, Thomas, et al. \\\"Memory-assisted reinforcement learning for diverse molecular de novo design.\\\" Journal of cheminformatics 12.1 (2020): 68. \\n\\n [2] Pereira, Tiago, et al. \\\"Diversity oriented deep reinforcement learning for targeted molecule generation.\\\" Journal of cheminformatics 13.1 (2021): 21. \\n\\n [3] Hu, Xiuyuan, et al. \\\"De novo drug design using reinforcement learning with multiple gpt agents.\\\" Advances in Neural Information Processing Systems 36 (2023). \\n\\n [4] He, Jiazhen, et al. \\\"Evaluation of reinforcement learning in transformer-based molecular design.\\\" Journal of Cheminformatics 16.1 (2024): 95.\\n\\n2. **Applicability of the ChEBI-20 Dataset for Diversity Evaluation**:\\n The use of the ChEBI-20 dataset raises concerns, as each molecular description in this dataset corresponds to a single molecule. Baseline approaches typically focus on generating one molecule to match the target rather than generating a diverse set for a single description. Furthermore, some descriptions may specify molecular details to the extent that producing diverse structures could be contradictory, making ChEBI-20 potentially unsuitable for assessing molecular diversity.\\n\\n3. **Reward Function\\u2019s Balancing of Terms**:\", \"the_reward_function_includes_two_terms\": \"one for structural diversity and another for description matching. It would be beneficial to introduce a coefficient to balance these terms, enabling tuning based on the importance of each aspect (diversity vs. adherence to the description).\\n\\n4. **NCircles Values in Figure 5**:\\n In Figure 5, NCircles values at threshold $h=0.75$ are higher than those at $h=0.65$. This seems inconsistent with the metric\\u2019s definition, where values at a higher threshold should logically be lower. The unexpected result suggests potential misinterpretation or misuse of the metric, which requires clarification.\", \"questions\": \"1. **Cluster Representation in Figure 6**:\\n Is it reasonable to represent each cluster as a circle in Figure 6? Given that a 2D projection of a circle in a high-dimensional space may not accurately be circular, I wonder if this visualization accurately reflects clustering in chemical space.\\n2. **Effect of SMILES Randomness on BLEU Scores**:\\n Since a single molecule can be represented by multiple SMILES strings, the randomness in SMILES representations could impact BLEU scores in the experiments. Has this potential source of variability been considered, and what effect does it have on the method\\u2019s stability or the evaluation of generated molecules?\\n\\nIn conclusion, this paper offers a valuable approach to improving molecular diversity in AI-driven drug discovery, and I will raise my score if my concerns (in Weaknesses and Questions) are well-addressed.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}" ] }

[ "Madhusudan Verma", "Manoj Kumar" ]

[ "NEURAL ORDINARY DIFFERENTIAL EQUATIONS", "GENERALIZATION BOUNDS", "BOUNDED VARIATION FUNCTIONS", "MACHINE LEARNING" ]

{ "note_id": [ "uk0GbCO6gk", "snSVpRxHlg", "mmrZXrFXlv", "lglPxdbLgw", "hoCD9t7XrC", "h2Z5LIRZFV", "dx07ywz0uM", "dBXXjyXBGY", "d5em353EGh", "d4iJRoPGYL", "c4dsGyj0mS", "ZcasghOCun", "XB0jksWC1P", "X9naBJ9FeV", "U2UOAt7aF5", "RS2l0n0igE", "DnuJyci47Y", "21A1UParDO", "1arj33c8vc", "12JzAb1UG8" ], "note_type": [ "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "comment", "official_review", "official_review", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment" ], "note_created": [ 1732512937036, 1732281032587, 1732284643985, 1730209399946, 1731923416939, 1732354269753, 1732297164283, 1731920687611, 1734012136951, 1730842162101, 1730531070478, 1732270458884, 1731919082433, 1730678357284, 1732345117455, 1733161284646, 1732289227874, 1731921660632, 1732464596265, 1732715146375 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission299/Authors" ], [ "ICLR.cc/2025/Conference/Submission299/Reviewer_e326" ], [ "ICLR.cc/2025/Conference/Submission299/Reviewer_BqoU" ], [ "ICLR.cc/2025/Conference/Submission299/Reviewer_e326" ], [ "ICLR.cc/2025/Conference/Submission299/Authors" ], [ "ICLR.cc/2025/Conference/Submission299/Authors" ], [ "ICLR.cc/2025/Conference/Submission299/Reviewer_SUqg" ], [ "ICLR.cc/2025/Conference/Submission299/Authors" ], [ "ICLR.cc/2025/Conference/Submission299/Authors" ], [ "ICLR.cc/2025/Conference/Submission299/Reviewer_2svr" ], [ "ICLR.cc/2025/Conference/Submission299/Reviewer_BqoU" ], [ "ICLR.cc/2025/Conference/Submission299/Reviewer_2svr" ], [ "ICLR.cc/2025/Conference/Submission299/Authors" ], [ "ICLR.cc/2025/Conference/Submission299/Reviewer_SUqg" ], [ "ICLR.cc/2025/Conference/Submission299/Authors" ], [ "ICLR.cc/2025/Conference/Submission299/Authors" ], [ "ICLR.cc/2025/Conference/Submission299/Authors" ], [ "ICLR.cc/2025/Conference/Submission299/Authors" ], [ "ICLR.cc/2025/Conference/Submission299/Reviewer_2svr" ], [ "ICLR.cc/2025/Conference/Submission299/Authors" ] ], "structured_content_str": [ "{\"comment\": \"It seems the reviewers has not thoroughly gone through the paper.None of the work either by Gulliox et. al (2024) or Marion(2023) follows the method of finding stricter bound of covering number of bounded variation functions.Only fig 3 resembles to that of Marion(2023) still that is for ResNet and for different dynamics to be learned and a small part(finding the relation between bound of solution and parameters) is similar(None of the complete Lemma and Theorem were used) to Gulliox et.al (2024 ), but we have found for time dependent case.\\n\\nWe are disappointed by the reviewers since none has pointed out the main contribution.\"}", "{\"comment\": \"These answers do only partially alleviate my concerns with this paper. I will maintain my score.\"}", "{\"comment\": \"Thanks for your response. I am not totally convinced yet.\\n\\n2. For deep neural networks, it is NP-hard to compute $\\\\mathrm{Lip}(f)$. Thus, your assumption is too strong.\\n3. I don't think it is problem of activation functions. It is the model architecture, e.g., attention mechanism which involves quadratic terms and thus it is only locally Lipschitz.\\n\\nThus, I tend to maintain my score.\"}", "{\"summary\": \"The authors claim to prove new generalization bounds for neural ODEs and residual neural networks. However, these claims are largely unsupported since their work does not significantly improve on Marion (2023) and Bleistein and Guilloux (2024). Some lemmas and proofs are directly borrowed along with notations from these two works without sufficient citations, which might be considered as a case of light plagiarism. The title is almost identical to the work of Marion (2023).\", \"soundness\": \"1\", \"presentation\": \"1\", \"contribution\": \"1\", \"strengths\": \"I do not believe that this paper has significant strengths.\", \"weaknesses\": [\"**Section 2.** The related work is close to insufficient and misses several recent contributions to the field, such as Marion, Wu et al. (2024) and Chen (2024).\", \"**Section 3.** This section compiles a list of definitions and Lemmas without sufficiently motivating their introduction. I would suggest a major rewriting of this section in order to guide the reader through the proofs.\", \"Also, Lemma 3.8 does not exist in Bartlett 2017b, and I believe that the Lemma as stated is wrong: there should not be a factor $1/\\\\sqrt{n}$ in the integral, but rather a factor $1/\\\\varepsilon$. See Bartlett 2017b Lemma A.5.\", \"**Section 4.**\", \"The learning setup is unclear. The authors write \\\"Let z be the solution of Neural ODE with x as the initial condition and let y be the true solution of the true differential equation learned by Neural ODE given by equation (4.3)\\\" (l. 221-223): it is unclear what is meant by the \\\"true solution\\\". Do the authors assume a generative model for the data ? In this case, it should be introduced.\", \"The authors write that the empirical risk \\\"cannot be optimized, since we do not have access to the continuous data.\\\" (l. 229 - 230). The authors have not introduced any form of continuous data, nor do they explain why the empirical risk cannot be optimized.\", \"This section seems to plagiarize Bleistein & Guilloux (2024) section 3.2, who consider a generative model where a continuous function, which is only observed at a discrete set of sampling times, generates the outcome through an unknown neural ODE. I believe that the authors have carelessly copied this text, hence introducing the two confusing sentences mentioned above which make no sens in their setting as it stands.\", \"**Section 5.**\", \"The main contribution of this part seems to be an adaptation of Proposition 2 of Marion (2023) to the general non-linear case. The results in the section strongly resemble the results of Bleistein and Guilloux (2024) --- see Lemma 3.3. These results should be in my opinion at least ackowledged in the main text.\", \"Settings concerns about plagiarism aside, the result cited here is directly implied by the results from Bleistein & Guilloux (2024), who establish generalization bounds for neural CDEs of the form $ dz(t) = \\\\mathbf{G}(z(t))dx(t)$ , where $ \\\\mathbf{G} $ is a generic neural network. Indeed, by setting $ x(t) = t $, one recovers a generic neural ODE. The authors of the aforementioned paper highlight the proximity between both models in Figure 2 of their paper.\", \"In the abstract, the authors claim to have \\\"showed the effect of overparameterization and domain bound in the generalization error bound\\\". This is a strong overstatement, since the type of arguments used by the authors only work in the case where $n$ is taken to be sufficiently large to obtain concentration ; even if in this case the number of parameters exceeds the number of observations, these bounds become vacuous in this setting, since the bound presented in Theorem 5.9 does not tend to $0$ when $d$ grows at the same rate than $n$. Hence these bounds say nothing about the overparametrized regime, in which it is typically observed that neural network achieve good prediction performance even if they completely overfit the training data (see for instance Bartlett 2019).\", \"**Section 6.**\", \"Both Marion (2023) and Bleistein and Guilloux (2024) invoke discretization based arguments to go from continuous neural-ODE like architectures to discrete ResNet-type architectures. I do not see such an argument here, and am hence unconvinced by the soundness of Theorem 6.1. In particular, the authors simply write that \\\"a neural ODE with an euler solver and $\\\\Delta t= 1$ replicated the ResNet updates, it follows that the solution space of ResNets is contained within the solution space of Neural ODEs.\\\" (l. 366-368).\", \"**Section 7.**\", \"The authors claim to perform these experiments on neural ODEs. Given the previous approximations in the paper and the strong similarities of experiments displayed in figures 2 and 3 with the experimental section of Marion (2023), I strongly suspect that these experiments are carried out on ResNets rather than **continuous** neural ODEs.\", \"Writing that experiment 1 validates Theorem 5.9 (l. 385) is an overstatement: the authors show (without any confidence intervals) on a purely synthetic dataset that the generalization error increases with the number of hidden units. However, the details on the model are insufficient (what is exactly meant by the number of hidden units ?). Also, since no precision is given on the training data, it is unclear whether the model operates in an overparametrized or underparametrized setting.\", \"The experiment displayed in Figure 3 is directly copied from Marion (2023). This article should at least be acknowledged here in my opinion.\", \"The experiment displayed in Figure 2 is novel and investigates the effect of penalizing the loss of a neural ODE with the bound of the solution, hence favoring solutions with a low euclidian norm. However, the experiment is not conclusive due to the high variance and the little variability of the mean for every choice of regularization.\", \"Figures should be included in a vectorized format (PNG or PDF).\", \"A experimental appendix should be added, that includes a detailed overview of the experiments.\", \"**References**\", \"Bartlett, Peter et al., \\\"Spectrally-normalized margin bounds for neural networks\\\", Neurips 2017.\", \"Bartlett, Peter et al., \\\"Benign overfitting in Linear Regression\\\", Proceedings of the National Academy of Sciences, 2020.\", \"Bleistein, Linus and Guilloux, Agathe, \\\"On the Generalization and Approximation Capacities of Neural Controlled Differential Equations\\\", ICLR 2024.\", \"Chen, Yihang et al., Generalization of Scaled Deep ResNets in the Mean-Field Regime, ICLR 2024.\", \"Marion, Pierre, \\\"Generalization bounds for neural ordinary differential equations and deep residual networks\\\", Neurips 2023.\", \"Marion, Pierre, Wu, Yu-Han, et al., Implicit regularization of deep residual networks towards neural ODEs, ICLR 2024.\"], \"questions\": [\"I believe that this paper is largely insufficient for publication as it stands due to a lack of novelty, and often teeters on the brink of plagiarism. I strongly encourage the authors not to submit this work at the moment and to read the ethics requirements of ICLR 2025. Many points can be improved (see above). I list a few questions bellow.\", \"Can you provide extensive details on the experiments carried out ? In particular, I would appreciate a mathematical formulation of the model use to generate your data and architectural details. Also, please carefully check that the experiments are run with neural ODEs instead of ResNets.\", \"Please provide more mathematical details on your neural ODE to ResNet conversion (Theorem 6.1).\"], \"flag_for_ethics_review\": \"['Yes, Research integrity issues (e.g., plagiarism, dual submission)']\", \"details_of_ethics_concerns\": \"This paper seems to partially plagiarize Marion (2023) and Bleistein and Guilloux (2024), without any significant contribution. I have provided more details above.\", \"rating\": \"1\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Answers for Weakness.\\n\\nSection 2. \\nWe have already referenced them in Abstract and Introduction.\\n\\nSection 3.\", \"we_have_used_a_variation_of_it__link\": \"https://www.cs.cornell.edu/~sridharan/dudley.pdf\\n\\nSection 4. \\nThe learning setup is similar to that of [1]. Sorry for confusion, we meant true real value(not function) not true solution at final time, we used generative model with unknown ODE.\\nHow the data is generated is explained in answer given below to question 1.\\n\\n Since we used generative model with unknown ODE, we do not have access to continuous data. \\n\\nSection 5.\\n\\nThis is not the main contribution ,we will still have generalization bound if we do not use this result , We found stricter bound for the covering number of solution space of Neural ODE, this is main contribution .We used this to get the bound of solution in terms of norm of weights ,We have not used directly any Lemma or Proof from Marion [2] or Blesistein and Guilloux [1]\\n\\nYes, by setting $x(t) =t$, one recovers a generic neural ODE, but their bound depend on discretization of control variable, in this case time, also the bound consists of more number of parameters.\\n\\nI think there is confusion in understand from your side $d$ will not say anything about over-parameterization, $d$ is the dimension of solution, the number of parameters will have effect on Norm and $V$ is dependent on norm. In Marion's paper also it is dependent on bound on norm of weight in the numerator.\\n\\nSection 6.\\n\\nThis is addressed in answer given to Question 2 below. \\n\\nSection 7.\\n\\n1.Code is for neural ODE, so one can verify the code and rerun it to reproduce the results.\\n\\n 2.The number of neurons in a layers is referred as number of hidden units.details of synthetic data generation is provided,it is actually real life complex data for particle moving in the air.\\n\\n 3. We have acknowledged the work of Marion [2]. \\n\\n 4.Yes ,but the mean is either same or it is decreasing although variance is high.If one zooms the image one can find.\\n\\n 5. We have added figure in png format now. \\n\\n 6. We have added more details in the Appendix section.\\n\\nAnswers to Questions.\\n\\n1. The synthetic data is generated to mimic complex real-life particle motion rather than being derived from numerical solutions of an explicit ODE. Specifically, it simulates the motion of a particle in a potential field with a sinusoidal pattern plus random noise, as defined by:\\n$$\\nx = \\\\sin(t) + 0.5 \\\\cdot \\\\text{noise}\\n$$\\n$$\\ny = \\\\cos(t) + 0.5 \\\\cdot \\\\text{noise}\\n$$\\n\\nThis noisy sinusoidal motion is intended to represent realistic, complex trajectories, suitable for testing the neural ODE model\\u2019s ability to generalize. Also, the experiments are run with neural ODE, you can verify using codes provided in supplementary.\\n\\n2. We have provided more mathematical details. We meant that the bound on time dependent Neural ODE will be same, that is when the parameters are also time dependent. The bound for time independent Neural ODE will be less than ResNet since Lipschitz constant of weights will be zero in that case. If L(final time of neural ODE) is discrete then the final time L of neural ODE will be same as Number of layers of Resnet, $V$ will be same since it independent of time, $L$ can be real number and since the number of layers are discrete, we take $N=floor(L)$. Since $f$ is same for both $Lip(f)$ will also be same hence V will be same. $N$ will always be less than or equal to $L$, hence covering number for set of solutions by Resnet will be less than or equal to covering number for set of solutions by Neural ODE and thus the Rademacher complexity. This is why bound is same.\\n\\n[1] Bleistein et al. On the Generalization and Approximation Capacities of Neural Controlled Differential Equations, ICLR 2024.\\n\\n[2] Marion, Generalization bounds for neural ordinary differential equations and deep residual networks, NeurIPS 2023.\"}", "{\"comment\": \"Thank you for the reply\\nCan you let us know the problem with the answers?\"}", "{\"comment\": \"Thank you for your response and some clarifications. However, after reading all other reviews and responses, I have decided to maintain my score.\"}", "{\"comment\": \"Answers for Weakness.\\n\\n1. We have changed some of the notations for the convenience of readers. \\n2. Our bound is stricter in terms of $n$ than Marion. For bound given by Bleistein and Guilloux, if we take the case $x(t)=t$, in which case it is neural ode, the bound is the same in terms of $n$ but the bound depends on the discretization of time, here it is independent of that and also the bound is simpler in this case as it contains less number of parameters. Comparing the bound given by Marion, the bound in our work is stricter in terms of $n$ for the linear case. For residual neural networks, the bound given by Marion does not depend on depth and have worse dependence on width. Our bound depends on depth but does not depend on width. We have added comparison results.\\n3. This is just substitution of bound on Radamacher complexity to theorem on regression bounds by Mohri 2018.\\n\\nAnswers for Questions.\\n\\n1. Thanks for pointing out, yes there is typo, it should depend at time 0. We have changed it.\\n2. Yes you are correct, it should be \\\"if\\\" instead of \\\"then\\\" and line 169 is an assumption of lemma.\\n3. No, it is a vector in $\\\\mathbb{R}^{d}$ (the solution at a final time) and not function of time. The theory will be valid for any finite-dimensional ODE (the solution can be of any dimension).\\n4. There is typo in definition, it should be function of $z$. For more details, we refer to [1], they have used the transform version of $z$. \\n5. Thanks for pointing out this, it should be function of $z, t$, and $\\\\theta$ everywhere. We refer to [2] \\n6. We used bound on solution as regularization loss function for figure 2 and lipschitz constant of weights as regularization loss function for figure 3.\\n7. The ODE is unknown. The synthetic data is generated to mimic complex real-life particle motion rather than being derived from numerical solutions of an explicit ODE. Specifically, it simulates the motion of a particle in a potential field with a sinusoidal pattern plus random noise, as defined by:\\n$$\\nx = \\\\sin(t) + 0.5 \\\\cdot \\\\text{noise}\\n$$\\n$$\\ny = \\\\cos(t) + 0.5 \\\\cdot \\\\text{noise}\\n$$\\n\\nThis noisy sinusoidal motion is intended to represent realistic, complex trajectories, suitable for testing the neural ODE model\\u2019s ability to generalize.\\n\\n[1] Bleistein et. al On the Generalization and Approximation Capacities of Neural Controlled Differential Equations, ICLR 2024.\\n\\n[2] Chen et. al Neural Ordinary Differential Equations, NeurIPS 2018.\"}", "{\"withdrawal_confirmation\": \"I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors.\"}", "{\"summary\": \"This submission proposes generalization bounds for neural ODEs and residual neural networks (the latter being a discretization of the first).\\nThe first 5 pages of the paper are dedicated to related works and preliminary results. The main results are in Theorems 5.9 and 6.9, generalizing results from Marion (2024), where the author only consider linear parametrization of the residuals (while still having non linear residuals with respect to the activations). Experiments on synthetic data are conducted.\", \"soundness\": \"1\", \"presentation\": \"1\", \"contribution\": \"2\", \"strengths\": \"The paper studies the interesting problem of deriving generalization bounds for residual architectures which are at the core of most successful deep learning methods.\", \"weaknesses\": [\"In my opinion, this paper is not suitable for acceptance at ICLR. It appears incomplete and lacking in polish. Below are specific points:\", \"The paper shares almost the exact same title as Marion (2024), with only the term \\\"deep\\\" removed. This is inappropriate.\", \"The obtained generalization bounds in Theorems 5.9 and 6.1) are not commented on and, most importantly, are not compared with existing ones from Marion (2024).\", \"The second experiment (Fig. 2) appears very similar to experiments shown in Marion (2024) (Figs. 1 and 2), yet Marion\\u2019s work is not cited here.\", \"The paper references only around 20 prior works, which is insufficient. A broader acknowledgment of previous studies is necessary (see the references cited by Marion (2024) as a comparison).\", \"The bibliography is poorly presented and lacks formatting consistency.\", \"There are no experimental details provided. I looked in the appendix, but none were included.\", \"Overall, the paper appears to have been submitted without adequate proofreading (see, for instance, the last sentence of the abstract). In addition:\", \"In Assumption 1, there is an unexpected dependence on time\\u2014this should be clarified.\", \"The symbol $z$ in line 227 is the same as the notation for the ODE solution. This is confusing.\", \"Line 230: The expression involving the $\\\\arg\\\\min$ is difficult to understand. The $\\\\arg\\\\min$ is taken over $\\\\theta$, yet it is denoted as a function $f$ (that is itself parametrized by $\\\\theta$). Furthermore, $\\\\arg\\\\min$ is applied to $\\\\theta(t) \\\\in \\\\theta(t)$, which is extremely unclear and problematic.\", \"Line 235: Typo present.\", \"In Lemma 5.1, it would be helpful to explicitly state that Assumption 1 is being used.\", \"In Lemma 5.1, the structure is confusing. It is hard to tell what is an assumption and what is a result. Key definitions (e.g., for $f$) are also missing. Can you clarify ?\", \"Line 265: The presentation here lacks rigor. Can you please specify the assumptions?\", \"There are multiple typos in the use of parentheses in lines 262, 267, and 272, which is unacceptable.\", \"Multiple typos are present in the experimental section (e.g., lines 423 and 429).\"], \"questions\": [\"How different are your experimental results from those of Marion (2024)?\", \"Why is there sometimes an additional t argument within the parametrized function in the neural ODES?\", \"Both theorems provide the same bounds, which requires clarification\\u2014how is this possible?\", \"Please also see the questions and remarks in the previous section.\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"Generalization bound of neural ODEs whose vector field is an MLP\", \"soundness\": \"2\", \"presentation\": \"1\", \"contribution\": \"2\", \"strengths\": \"1. give an bound estimation of $z(t)$ based on the Lipschitz constant of an MLP $f(z)$\\n2. follow the similar technique as [1] to derive the generalization bound of $\\\\dot z= f(z) $ \\n\\n[1] Bartlett et, al. Spectrally-normalized margin bounds for neural networks, NeurIPS 2017.\", \"weaknesses\": \"1. The bound estimation of $\\\\|z(t)\\\\|$ is very loose due to the Gronwall's inequality $u(t)\\\\leq \\\\alpha(t)+\\\\int_a^t\\\\beta(s)u(s)ds$. In this case, $\\\\beta$ is the Lipschitz bound $\\\\mathrm{Lip}(f)$ and $\\\\alpha$ is a bound related to $\\\\mathrm{Lip}(f)$ and bias norm. Thus, the downstream analysis of generalization bound could be very conservative.\\n2. $\\\\mathrm{Lip}(f)$ is estimated using the product of spectral norm bounds, which is again very loose. The SOTA estimation is based on some semidefinite programming formulation, see [2].\\n3. The assumption of globally Lipschitz $f$ is quite strong as the popular transformer architecture is only locally Lipschitz. \\n4. The experimental results are quite weak, lacking of extensive comparison study. \\n5. The presentation is poor. A few (not all) examples are listed as follows:\\n\\n- The mapping $z(t)\\\\rightarrow y$ is not defined. \\n- Assumption 2 involves $A_i(t)$ and $b_i(t)$ which are introduced later in Section 5.\\n- A right ) is missing in 5.4.\\n- [1] appears twice in the reference.\\n\\n[2] P. Pauli et. al. Novel quadratic constraints for extending lipsdp beyond slope-restricted activations, ICLR 2024.\", \"questions\": \"1. There exist many generalization bound analysis of residual networks. Can you provide some comparison studies with Thm. 6.1?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you for your response to my review. I have read your answer as well as the other reviews and your answers to them.\\n\\nI decide to maintain my score to 3 as I believe this paper should not be published at the moment, for the reasons listed in my review.\"}", "{\"comment\": \"Answers for Weakness.\\n\\n1. Sorry for this confusion. We have changed the title and also the method used by Marion are totally different, because he used the covering number bound for $\\\\theta$ and our focus is on the class of solution of neural ODE itself. \\n2. The bound by this is stricter in terms of $n$. Bound by Marion does not depend on depth but has worse dependence on width, this bound depends on depth but does not depend on width.\\n3. The experiments are not similar. The regularization term is different in our case.\\n4. Thanks for the suggestion, we have added some more references. \\n5. We have updated reference formatting.\\n6. We have already added experiment details in numerical illustration itself , but still we provide some more experimental details such as how the data is generated for the convenience of readers. \\n\\nAnswers for Questions.\\n\\n1. There is typo it should be $t$ everywhere inside $f$ , because the dynamic function is $f(z(t),t,\\\\theta)$ for example the dynamic function can be $z(t)t^2\\\\theta$, authors in [1] also used this type of notation. \\n2. Yes, it is a solution of Neural ODE. But $z$ itself will depend on $\\\\theta$, so $\\\\theta$ dependence is there. Also, if we see [2], the risk is defined for tranformed version of solution only i.e., $\\\\Phi^{T}z_{1}$. \\n3. It will be $\\\\hat{\\\\theta}$ and $\\\\hat{f}$ with be predictor for this $\\\\hat{\\\\theta}$. It should be $z_{\\\\theta(t)}{(t)}$ which is solution to Neural ODE.\\n4. Marion is trying to show the link between the generalization gap and the Lipschitz constant of the weights, here we are also trying to show the link between the generalization gap and the bound on solution. Marion has done the experiment shown in fig 3 for Resnet, we have done for Neural ODE.\\n5. Already addressed in point 1.\\n6. We meant that the bound on time dependent Neural ODE will be same, that is when the parameters are also time dependent. The bound for time independent Neural ODE will be less than ResNet since Lipschitz constant of weights will be zero in that case. If L(final time of neural ODE) is discrete then the final time L of neural ODE will be same as Number of layers of Resnet, $V$ will hold for any time, it will be same for final time t for Neural ODE and Resnet, $L$ can be real number and since the number of layers is discrete we take $N=floor(L)$. Since $f$ is same for both, $Lip(f)$ will also be same hence $V$ will be same. $N$ will always be less than or equal to $L$, hence covering number for set of solutions by Resnet will be less than or equal to covering number for set of solutions by Neural ODE and thus the Rademacher complexity. This is why bound is same.\\n\\n[1] Chen et. al Neural Ordinary Differential Equations, NeurIPS 2018.\\n\\n[2] Bleistein et. al On the Generalization and Approximation Capacities of Neural Controlled Differential Equations, ICLR 2024.\"}", "{\"summary\": \"This paper provides generalization bounds for neural ODEs. It extends the class of neural networks used as the dynamics function. The bound is applied also to residual neural networks. Numerical experiments are used to show the effect of hyperparameters on the generalization gap.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"3\", \"strengths\": [\"The results are applicable to a much larger class of neural ODEs than prior work, such as Marion's paper which is only for when f depends linearly on the parameters.\", \"The prior work is well explained. Important lemmas from the prior work as used.\", \"I did not find errors in any of the proofs themselves.\"], \"weaknesses\": [\"Some of the notation is conflicting/confusing. Please see the questions.\", \"The prior work of Bleistein and Guilloux, and Marion is referenced. However, the bounds derived in these paper and in this paper are not compared.\", \"The main theorems 5.9 and 6.1 have only an outline of the proof, and there is not a full proof in the appendix.\", \"The numerically illustrations are missing details. Please see the questions.\"], \"questions\": [\"On lines 89 and 93, the initial condition is given by $z(0)=\\\\phi_{\\\\theta(t)}(u)$. Why does the initial condition depend on the parameters at time t instead of time 0?\", \"On line 168, should it be \\\"if\\\" instead of \\\"then\\\"? Is line 169 an assumption of the lemma?\", \"In Assumption 3 (line 236), the outcome y is said to be in R. Is y a single number (say the solution at a final time) or is y a function of time? Same question for Assumption 4 and the loss function. Does this theory only apply for one-dimensional ODEs?\", \"Why is the risk a function of f in Definition 3.9, but is a function of z in section 4?\", \"Why is f a function of z, t, and theta in Assumption 1, but only a function of z and theta in equation (4.3)?\", \"In the numerical illustrations, what are the regularization loss functions for either case?\", \"In the numerical illustration, how is the synethic data generated? Is it from numerical solutions of an ODE? If so, what ODE?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Please consider the new response as well given to BqoU.\"}", "{\"comment\": \"Today is the last date to reply for the reviewers but we did not get a reply.\"}", "{\"comment\": \"Thank you for the reply.\\n\\n We used the idea similar to [1] just to get relation between bound of solutions and parameters .Even if we don't assume f is Lipschitz(global or local) and don't use Lif(f) in the bound of solution we will still have genralization bound consisting of V term, since the solution will be of bounded variation (it requires only condition that solution is differentiable). One can find V in many different ways.\\n\\n2.Yes we agree for deep neural networks, it is NP-hard to compute Lip(f), but prior work[1] also has Lip(f) in their bound (equation 120 Lip(G) term ). It does not describe its estimation\\n\\n3.Yes, it depends not only on activation functions but also on other factors such as weights and overall architecture. However, the prior work [1] done in this area for general function does not assume it is locally Lipschitz ( section 3.1 Neural Vector Fields first line).\\n\\n\\n[1] Bleistein et al. On the Generalization and Approximation Capacities of Neural Controlled Differential Equations, ICLR 2024.\"}", "{\"comment\": \"Answers for Weakness.\\n\\n1. Thanks for pointing this out, but our aim in this paper was to get stricter bounds for covering numbers. We derive tighter bound on covering number but for the solution bound, we used the Lipschitz based argument. \\n2. Thanks for this interesting observation. This observation will help from the point of view of implementation, but the theoretical bound will have no effect since our bound does not have explicit form for estimation of Lip$(f)$. Our bound assumes Lip$(f)$ is known.\\n3. Yes, you are correct, it is not globally Lipchitz for some activation functions but we can choose activation functions to make $f$ globally Lipschitz. \\n 4. The experiment's aim was not to compare in terms of strictness of bound because for that we believe the bound should have same type of parameters. The experiment's aim was to get the relation of the bound of the solution and the Lipschitz constant of weights with the generalization bound.\\n5. We have modified the manuscript based on the comments of reviewer, as in earlier version presentation was not up to the mark. \\n\\nAnswers for Questions.\\n\\n1. The bound given in our work is stricter in terms of $n$. Bound given by Marion [1] does not depend on depth but has worse dependence on width, this bound depends on depth but does not depend on width. Same is true for Bartlett et al. [2] work. We have added comparison. \\n\\n\\n[1] Marion Generalization bounds for neural ordinary differential equations and deep residual networks, NeurIPS 2023. \\n\\n[2] Bartlett et al. Spectrally-normalized margin bounds for neural networks, NeurIPS 2017.\"}", "{\"comment\": \"Your answers do not convince me not the reject the paper hence my decision to maintain my score. In addition, the other reviews confirm to me that the paper lacks originality in its contribution and its ressemblance to Marion (2023) worries me on the transparency of your work.\"}", "{\"comment\": \"We have uploaded the revised version of paper.\\n1. The title has been changed.\\n2. Typos, bibliography formatting corrected.\\n3 . We have made changes in section 4(such as x is changed with t,y is label instead of solution etc ).\\n4. We have made changes in section 5 (Theorem 5.9 and 6.1)\\n5. We have added comparisons.\\n6. Experimental details are added in the Appendix.\"}" ] }

[ "Lei Shen", "Zhenheng Tang", "Lijun Wu", "Yonggang Zhang", "Xiaowen Chu", "Tao Qin", "Bo Han" ]

[ "Federated Learning" ]

{ "note_id": [ "zGSH3QmxH1", "vsNfd3a0iB", "t17iR1CPEj", "qlML6XxtGH", "qUjrQlXvTz", "nty4kzylIn", "lXH3B3Es79", "koLdxUFNiq", "khF2d84mEV", "kL3HOGT0gM", "kA4qRckRF7", "jDSPSF0DGU", "geKGokmR6Q", "aaPERclT2K", "YdVxjq9xn9", "WDmMC428z9", "UzKCxfvVKy", "Tq3gnKwr4F", "SbUHETXuEM", "Qi9w8Dmks0", "NVJ4YzIF3W", "M2qsKIcIAM", "FcYElXewAW", "DCC6TXuI9s", "ANUnGeiVsC", "8FPegrwm0p" ], "note_type": [ "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_review", "official_comment", "official_comment", "meta_review", "decision", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment" ], "note_created": [ 1732471940805, 1730681937284, 1732472038550, 1732332890491, 1732699405506, 1732472068749, 1732332853855, 1732544449042, 1732332328948, 1732332283986, 1732699695437, 1732331789743, 1730341642975, 1732332461237, 1730694843106, 1732695177597, 1732332748331, 1734746698155, 1737523955809, 1732803874910, 1732331734110, 1732694993295, 1729472051225, 1732472130098, 1732332050844, 1732699966221 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission9027/Authors" ], [ "ICLR.cc/2025/Conference/Submission9027/Reviewer_FWNo" ], [ "ICLR.cc/2025/Conference/Submission9027/Authors" ], [ "ICLR.cc/2025/Conference/Submission9027/Authors" ], [ "ICLR.cc/2025/Conference/Submission9027/Authors" ], [ "ICLR.cc/2025/Conference/Submission9027/Authors" ], [ "ICLR.cc/2025/Conference/Submission9027/Authors" ], [ "ICLR.cc/2025/Conference/Submission9027/Authors" ], [ "ICLR.cc/2025/Conference/Submission9027/Authors" ], [ "ICLR.cc/2025/Conference/Submission9027/Authors" ], [ "ICLR.cc/2025/Conference/Submission9027/Authors" ], [ "ICLR.cc/2025/Conference/Submission9027/Authors" ], [ "ICLR.cc/2025/Conference/Submission9027/Reviewer_8vJb" ], [ "ICLR.cc/2025/Conference/Submission9027/Authors" ], [ "ICLR.cc/2025/Conference/Submission9027/Reviewer_VDPF" ], [ "ICLR.cc/2025/Conference/Submission9027/Reviewer_VDPF" ], [ "ICLR.cc/2025/Conference/Submission9027/Authors" ], [ "ICLR.cc/2025/Conference/Submission9027/Area_Chair_4oBR" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission9027/Authors" ], [ "ICLR.cc/2025/Conference/Submission9027/Authors" ], [ "ICLR.cc/2025/Conference/Submission9027/Reviewer_FWNo" ], [ "ICLR.cc/2025/Conference/Submission9027/Reviewer_kxCM" ], [ "ICLR.cc/2025/Conference/Submission9027/Authors" ], [ "ICLR.cc/2025/Conference/Submission9027/Authors" ], [ "ICLR.cc/2025/Conference/Submission9027/Authors" ] ], "structured_content_str": [ "{\"title\": \"Window for discussion is closing\", \"comment\": \"Dear Reviewer VDPF,\\n\\nThanks a lot for your time in reviewing and reading our response and the revision. Thanks very much for your valuable comments. We sincerely understand you\\u2019re busy. But as the window for discussion is closing, would you mind checking our responses and and confirm whether you have any further questions? We look forward to answering more questions from you.\\n\\nBest regards and thanks,\\n\\nAuthors of #9027\"}", "{\"summary\": \"This paper presents a novel framework called Hot-Pluggable Federated Learning (HPFL) that aims to bridge the gap between generic federated learning (GFL) and personalized federated learning (PFL). The authors propose a new learning paradigm, Selective Federated Learning (SFL), which combines model optimization with model selection. HPFL addresses the challenges of storing and selecting whole models by designing an efficient framework that allows clients to train personalized plug-in modules and upload them to a server. During inference, a selection algorithm identifies suitable plug-in modules to enhance performance on target data distributions. The paper also incorporates differential privacy protection during the selection process. Comprehensive experiments demonstrate HPFL's effectiveness in improving GFL performance and its potential in addressing other FL challenges like continual learning and one-shot FL.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"1. The authors identify a substantial gap between GFL and PFL, and formulate a new problem SFL to bridge them together to address this performance gap. Both optimization function of GFL and PFL are the special cases of it.\\n\\n2. The authors propose a general, efficient and effective framework HPFL, which practically solves SFL.\\n\\n3. Comprehensive experiments and ablation studies on four datasets and three neural networks demonstrate the effectiveness of HPFL.\", \"weaknesses\": \"1. It\\u2019s adorable that the authors define a paradigm to bridge the gap between GFL and PFL, but the selective method are not novelty enough. For example, [1] allowed each client to choose the appropriate scale model to train. And the training process in HPFL are identical to the split federated learning, the authors only add another select process in inference process.\\n2. The authors claim that they theoretically show PMs can be used to enhance GFL with a new learning problem named Selective FL (SFL), which involves optimizing PFL and model selection. But only the statement of the loss cannot totally evaluate the effectivess of the SFL. And in the Eq.4, the greater-than sign \\\\geq should be a less-than sign \\\\leq? The better one should gain the less loss?\\n3. What\\u2019s the originality in the analysis of the privacy protection? It seems that add Gaussian noise to the partial model or full model is identical. Thus I think it\\u2019s only an existing result.\\n4. The notations needs to be improved. For example, in section 2.4 the definition of Selective FL (SFL) problem, the introducing of auxiliary information is confused. And in Theorem2.3, the function s(\\\\dot) lacks description.\\n5. The presentation in experimental section should be improved. For example, the Table 2 is hard to read. It\\u2019s confused that the authors not only give the best result in grey, but also some second best result. And why the proposed HPFL does not gain the best result especially compared to the GFL FedSAM?\\n\\n[1] Cho, Yae Jee, et al. \\\"Heterogeneous ensemble knowledge transfer for training large models in federated learning.\\\" arXiv preprint arXiv:2204.12703 (2022).\", \"questions\": \"As shown in the Weakness.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Window for discussion is closing\", \"comment\": \"Dear Reviewer FWNo,\\n\\nThanks a lot for your time in reviewing and reading our response and the revision. Thanks very much for your valuable comments. We sincerely understand you\\u2019re busy. But as the window for discussion is closing, would you mind checking our responses and and confirm whether you have any further questions? We look forward to answering more questions from you.\\n\\nBest regards and thanks, \\n\\nAuthors of #9027\"}", "{\"title\": \"Response to Reviewer kxCM - Part 3\", \"comment\": \"**Q5** :\\n> How can the plug-in marker contribute to bridging the gap between GFL and PFL?\\n\\n***Ans for Q5):***\", \"below_we_clarify_how_hpfl_contribute_to_contribute_to_bridging_the_gap_between_gfl_and_pfl_in_detail\": \"- **Plug-in mechanism helps improve GFL performance with PFL modules:** Sharing all personalized plug-ins (PFL modules) and make it available for all clients grealty enrich the model space a client can choose from when meet with all datasets (GFL setting). \\n- **The naive design of selecting is inefficient:** Considering the design of MoE, there should be a gating layer to identify PFL modules. However, in the real-world module sharing platforms like huggingface, an unified updating gating layer is not achievable due to the asynchronous and continuous new plug-ins. \\n- **Markers help to identify the training data distribution of the plug-in modules:** Therefore, we design a selection mechanism based on plug-in markers (like a identifier of client in the selection process). We match the approprite plug-in and test data by plug-in and task markers through distance measurement. \\nTake one of our real-world examples in ***introduction***, when one\\ntraveling abroad, the personal map app might recommend entirely different restaurants from their residence. In this process, **plug-in marker can help us find the models trained on local restaurant and personal data**, which can make better recommendations. Then the final recommendation made by suitable model is prompted to the user, thus increase the overall chance of satisfying the user. \\n\\n\\n**Q6** :\\n> (1) Can you please clarify which type(s) of differential privacy (local, central, or both) are used, (2) and to provide a more detailed explanation of how the plug-ins interact with the DP mechanisms and (3) what novel contributions are made in this area?\\n\\n***Ans for Q6):*** \\n - (1) As explained above in ***answer to Q3***, DP is applied to markers instead of models. Therefore, it cannot be classified with the category used in model DP. In our code implementation, the noise of markers are added in client side. \\n - (2) Since DP is applied to markers instead of model components like backbone or plug-ins, we think no interaction exists between the plug-ins and DP mechanisms [1]. \\n - (3) Compared with applying DP to model, applying DP to shared information is significantly less explored. Therefore, we hope our study can add diversity to DP applications like applying DP to shared information except for model parameters.\\n\\n\\n\\n> ***Reference***\\n> \\n> [1] Wei, Kang, Jun Li, Ming Ding, Chuan Ma, Howard H. Yang, Farhad Farokhi, Shi Jin, Tony QS Quek, and H. Vincent Poor. \\\"Federated learning with differential privacy: Algorithms and performance analysis.\\\" IEEE transactions on information forensics and security 15 (2020): 3454-3469.\\n> \\n> [2] Acar, Abbas, Hidayet Aksu, A. Selcuk Uluagac, and Mauro Conti. \\\"A survey on homomorphic encryption schemes: Theory and implementation.\\\" ACM Computing Surveys (Csur) 51, no. 4 (2018): 1-35.\\n> \\n> [3] Knott, Brian, Shobha Venkataraman, Awni Hannun, Shubho Sengupta, Mark Ibrahim, and Laurens van der Maaten. \\\"Crypten: Secure multi-party computation meets machine learning.\\\" In NeurIPS, 2021.\\n> \\n> [4] Lalitha, Anusha, Shubhanshu Shekhar, Tara Javidi, and Farinaz Koushanfar. \\\"Fully decentralized federated learning.\\\" In Third workshop on bayesian deep learning (NeurIPS), 2018.\\n> \\n> [5] Nofer, Michael, Peter Gomber, Oliver Hinz, and Dirk Schiereck. \\\"Blockchain.\\\" Business & information systems engineering 59 (2017): 183-187.\\n> \\n> [6] Shazeer, Noam, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. \\\"Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer.\\\" In ICLR, 2016.\\n> \\n> [7] Xie, Cong, Sanmi Koyejo, and Indranil Gupta. \\\"Asynchronous federated optimization.\\\" arXiv preprint arXiv:1903.03934 (2019).\\n> \\n> [8] Yoon, Jaehong, Wonyong Jeong, Giwoong Lee, Eunho Yang, and Sung Ju Hwang. \\\"Federated continual learning with weighted inter-client transfer.\\\" In ICML, 2021.\"}", "{\"title\": \"Thanks for your reply and raising the score\", \"comment\": \"Dear reviewer FWNo,\\n\\nThanks for your efforts in reviewing our paper. Your constructive comments have greatly helped us improve our paper. If you have any further concerns, we are pleased to respond to them. \\n\\nBest regards and thanks,\\n\\nAuthors of #9027\"}", "{\"title\": \"Window for discussion is closing\", \"comment\": \"Dear Reviewer 8vJb,\\n\\nThanks a lot for your time in reviewing and reading our response and the revision. Thanks very much for your valuable comments. We sincerely understand you\\u2019re busy. But as the window for discussion is closing, would you mind checking our responses and and confirm whether you have any further questions? We look forward to answering more questions from you.\\n\\nBest regards and thanks, \\n\\nAuthors of #9027\"}", "{\"title\": \"Response to Reviewer kxCM - Part 2\", \"comment\": [\"**Q3** :\", \"> (1) What is the novelty of integrating DP in the algorithm?\", \"> (2) How does it lead to advancing the proposed HPFL solution?\", \"> (3) Is it HPFL+DP or the integration has some challenges? If so, what are the challenges and how does this paper tackle them?\", \"> (4) Why can\\u2019t we use other privacy preservation mechanisms?\", \"***Ans for Q3):***\", \"(1) Instead of adding Gaussian noise to the partial or full model, we **utilizing Differential Privacy (DP) on the markers** which are the intermediate features outputed from the model according to the training data and test data, to provider better privacy safety for HPFL. Compared with applying DP to model [1], applying DP to shared information is significantly less explored.\", \"(2) DP is mainly used to provide privacy protection for the shared plug-in markers.\", \"(3) No special issues have been encountered in our attempts to apply DP to HPFL, DP is a commonly-used mechanism used to protect privacy in FL, so we utilize this technique to provide guarantee for privacy protection. Both theoretic analysis in ***Section 3.4*** and experiments in ***Appendix E.2*** have shown its effectiveness.\", \"(4) Sure, there should be other privacy preservation mechanisms which can protect HPFL from information leakage, we adopt a commonly-used and simple technique in the simple implementation of HPFL our paper shows. This indicates that the HPFL is not restricted to the DP as the privacy protection. It is adaptable to some other methods to enhance its privacy protection. For example, we can use Homomorphic Encryption [2], Secure Multi-Party Computation [3] into HPFL to enhance privacy protection.\", \"**Q4**\", \"> Detailed comparison with two types of studies: i. existing PFL algorithms and comparing with HPFL, ii. Existing privacy preserving algorithms and comparison with DP.\", \"***Ans for Q4):***\", \"Differences of HPFL with existing PFL algorithms:\", \"(1) PFL algorithms focus on how to personalize model to perform best in PFL setting (local test datasets) instead of how to utilize them for GFL setting (all datasets), and HPFL is the first framework to leverage the personalized model component from other clients to adapt to various and shifting test distribution.\", \"(2) PFL algorithms cannot adapt its model at test-time to keep the inference model suitable for test data, this problem significantly degrade its robustness and performance when meet test data out of its local data distribution. While HPFL manage to choose suitable plug-in according to the test data with its selection mechanism.\", \"(3) clients in PFL algorithms only have the access of its local personalized model, while clients in HPFL can get personalized plug-in of other clients, which greatly enrich the knowledge a client can access.\", \"(4) the plug-in in HPFL is much more light-weight than the personalized model in PFL, which greatly increse the efficiency of HPFL.\", \"Differences of DP with other existing privacy preserving algorithms: DP is a commonly-used privacy protection mechanism. Main differences of DP with other existing privacy preserving algorithms used in FL can be concluded as:\", \"(1) DP provide strict theory framework to quantify its risk of privacy leakage, other methods cannot provide the possibility of privacy leakage.\", \"(2) the implementation is simple, different with Decentralized Federated Learning without a central server [4] and Secure Multi-Party Computation[3], which require complicated protocols between clients to directly communicate.\", \"(3) requires less computational costs than methods like Homomorphic Encryption [2] and blockchain [5].\"]}", "{\"title\": \"Welcome for more discussions\", \"comment\": \"Dear reviewer FWNo,\\n\\nThanks for your valuable time in reviewing and constructive comments, according to which we have tried our best to answer the questions and carefully revise the paper. Here is a **summary of our response** for your convenience:\\n\\n- (1) **Comparison with Fed-ET & split FL**: We listed the differences in a table to deliver them more intuitively. \\n\\n | | Fed-ET | split FL | HPFL |\\n |--|-- |-- |-- |\\n | Major Mechanism | ensemble and KD | parameter decoupling | adapt to test data with appropriate well-trained plug-in models |\\n | Focus on performance of | all datasets | local datasets | **all datasets** |\\n | training paradigm | train a single model by aggregation and KD | fine-tuning local model | fine-tuning **model components** with a backbone |\\n | inference paradigm | use global model | use personalized models or in vetical FL way | **select appropriate personalized plug-in and the common backbone** then inference |\\n | real-world deployment | all clients share a common model | clients can only use local-personalized models | **all clients can share all personalized plug-ins and a common backbone** |\\n | public datasets | required on server for KD | not required | not required |\\n | transmitted model | whole model | none | plug-in (part of model) |\\n | when and how selection happens | randomly picking participating clients every round | randomly picking participating clients every round | **select by deterimined mechanism (e.g. distance metric like MMD) at test time** |\\n\\n - **Focus on performance of:** This column distinguish the target test data of different algorithms, showing HPFL is target on all datasets (global dataset) instead of local datasets as in split FL algorithms. \\n - **real-world deployment:** This column shows the differences when various algorithms deploy in real-world applications, here we **highlight the importance of sharing all personalized plug-ins** and regard this as our key contribution leading to better GFL performance than all split FL algorithms. \\n\\n\\n With the above design, HPFL has the following advantages than Fed-ET and split FL methods:\\n - **Superb performance on All datasets instead of local datasets.**\\n - **Lower communication resource than Fed-ET.**\\n - **Free from the requirement of public dataset, which is diffult to get in many cases like medical information.**\\n - **Implementing a possible market mechanism of plug-in sharing between different users.**\\n - **Potential Application for asynchrnous and continuous learning scenarios.**\\n \\n We have add the above works and their comparison with HPFL in ***our revision***.\\n\\n- (2) **Problems of Theoretical analysis**: Following your constructive comments, we have revised the text description of ***Theorem 2.1***. Besides, We also show analyze loss is a common way for theoretical performance analysis in FL. \\n- (3) **Originality of our DP protection**: Following your valuable suggestions, we have make the novelty of our DP clear: **Our DP is applied to **shared information like intermediate features instead of models or gradients**. \\n- (4) **Presentation problems**: Following your valuable suggestions, we have revised the presentation of our theoretical analysis: 1. In ***Section 2.4 in our revision***: \\\"$H$ is auxiliary information exploited to select plug-in module, e.g. noisy feature for calculating distance metrics like Maximum Mean Discrepancy (MMD)\\\"; 2. \\\"$S$ is called selection function that outputs the model index to select a model from the PMs based on the input $\\\\xi_m$ and the auxiliary information $H$, which will be illustrated in Section 3\\\", we also revise the notation to more clearly explain the select function s(\\\\dot). \\n- (5) **Presentation in experimental section**: We explain the meaning of results in different colors in ***Table 2***, together with the motivation of our experiments. \\n\\nWe humbly hope our repsonse has addressed your concerns. If you have any additional concerns or comments that we may have missed in our responses, we would be most grateful for any further feedback from you to help us further enhance our work.\\n\\n\\nBest regards, \\n\\nAuthors of #9027\"}", "{\"title\": \"Response to Reviewer FWNo - Part 3\", \"comment\": \"> ***Reference***\\n> \\n> [1] Cho, Yae Jee, et al. \\\"Heterogeneous ensemble knowledge transfer for training large models in federated learning.\\\" arXiv preprint arXiv:2204.12703 (2022).\\n> \\n> [2] Liang, Paul Pu, Terrance Liu, Liu Ziyin, Nicholas B. Allen, Randy P. Auerbach, David Brent, Ruslan Salakhutdinov, and Louis-Philippe Morency. \\\"Think locally, act globally: Federated learning with local and global representations.\\\" arXiv preprint arXiv:2001.01523 (2020).\\n> \\n> [3] Wu, Zhaomin, Qinbin Li, and Bingsheng He. \\\"A coupled design of exploiting record similarity for practical vertical federated learning.\\\" Advances in Neural Information Processing Systems 35 (2022): 21087-21100.\\n> \\n> [4] Karimireddy, Sai Praneeth, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. \\\"Scaffold: Stochastic controlled averaging for federated learning.\\\" In ICML, 2020.\\n> \\n> [5] Chen, Hong-You, and Wei-Lun Chao. \\\"On Bridging Generic and Personalized Federated Learning for Image Classification.\\\" In ICLR, 2022.\\n> \\n> [6] Li, Tian, Anit Kumar Sahu, Manzil Zaheer, Maziar Sanjabi, Ameet Talwalkar, and Virginia Smith. \\\"Federated optimization in heterogeneous networks.\\\" In MLSys, 2020.\\n> \\n> [7] Wei, Kang, Jun Li, Ming Ding, Chuan Ma, Howard H. Yang, Farhad Farokhi, Shi Jin, Tony QS Quek, and H. Vincent Poor. \\\"Federated learning with differential privacy: Algorithms and performance analysis.\\\" IEEE transactions on information forensics and security 15 (2020): 3454-3469.\\n>\\n> [8] Shazeer, Noam, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. \\\"Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer.\\\" In International Conference on Learning Representations. 2016.\\n> \\n> [9] Ren, Jie, Peter J. Liu, Emily Fertig, Jasper Snoek, Ryan Poplin, Mark Depristo, Joshua Dillon, and Balaji Lakshminarayanan. \\\"Likelihood ratios for out-of-distribution detection.\\\" Advances in neural information processing systems 32 (2019). \\n> \\n> [10] Loh, Wei\\u2010Yin. \\\"Classification and regression trees.\\\" Wiley interdisciplinary reviews: data mining and knowledge discovery 1, no. 1 (2011): 14-23.\\n> \\n> [11] Cutler, Adele, D. Richard Cutler, and John R. Stevens. \\\"Random forests.\\\" Ensemble machine learning: Methods and applications (2012): 157-175.\"}", "{\"title\": \"Response to Reviewer FWNo - Part 2\", \"comment\": \"**Q2** :\\n> (1) The authors claim that they theoretically show PMs can be used to enhance GFL with a new learning problem named Selective FL (SFL), which involves optimizing PFL and model selection. But only the statement of the loss cannot totally evaluate the effectivess of the SFL. \\n> (2) And in the Eq.4, the greater-than sign \\\\geq should be a less-than sign \\\\leq? The better one should gain the less loss?\\n\\n***Ans for Q2):*** \\n - (1) We understand the value of loss function in training (i.e. training error) is not enough to evaluate the performance of model on test data, however, in our analysis, we actually calculate the generalization error of the models, we believe it should be enough for evaluating the actual performance in inference stage. Besides, in the condition that same training samples and model complexity/architecture, which is obvisouly satisfied in our analysis, lower training error means lower generalization bound, and thus the better model performance. The theoretical analysis of our paper follows a series of work [4, 5, 6]. We also experimentally report the accuracy of HPFL, which is the implementation of SFL. \\n - (2) We are sorry for the misleading information we convey in our presentation. The greater-than sign $\\\\geq$ in Eq.4 is the correct meaning we trying to express, where the local model $\\\\theta_m$ test on its own data $\\\\xi_m$, while $\\\\theta_i$ is the model of other client i. We will revise the notation from \\\"client i\\\" to \\\"client m\\\" in text description to prevent potential misunderstanding in ***our revision***. Thanks for your suggestion and careful reading. \\n\\n**Q3** :\\n> What\\u2019s the originality in the analysis of the privacy protection? It seems that add Gaussian noise to the partial model or full model is identical. \\n\\n***Ans for Q3):*** \\nInstead of adding Gaussian noise to the partial or full model, we utilizing **Differential Privacy (DP) on the markers** to provider better privacy safety for HPFL. Compared with applying DP to model, applying DP to **other shared information like intermediate features is significantly less explored**. Using DP for privacy protection, which is common in FL [7], is a component of HPFL.\\n\\n**Q4** :\\n> (1) The notations needs to be improved. For example, in section 2.4 the definition of Selective FL (SFL) problem, the introducing of auxiliary information is confused. \\n> (2) And in Theorem 2.3, the function s(\\\\dot) lacks description.\\n\\n***Ans for Q4):*** \\n - (1) To solve the potential confusion, we add some text descriptions to explain the function of auxiliary information in ***Section 2.4 in our revision***: \\\"$H$ is auxiliary information exploited to select plug-in module, e.g. noisy feature for calculating distance metrics like Maximum Mean Discrepancy (MMD)\\\". In real implementation, $H$ are not limited in feature prototype like markers, but also some mechanisms such as learning gate like in MOE [8], distinguish models like OOD detector[9], decision tree [10], and random forest [11]. Hopefully with the modification, the introducing of auxiliary information make more sense.\\n - (2) \\\"$S$ is called selection function that outputs the model index to select a model from the PMs based on the input $\\\\xi_m$ and the auxiliary information $H$, which will be illustrated in Section 3\\\", we use the lowercase letter $s$ to denote the instantialized version of Selection function $S$, i.e. a specific selection method, e.g. MMD, SVCCA and CKA. We revise the notation to more clearly explain the select function s(\\\\dot). \\n\\n**Q5** :\\n> The presentation in experimental section should be improved. For example, the Table 2 is hard to read. It\\u2019s confused that the authors not only give the best result in grey, but also some second best result. And why the proposed HPFL does not gain the best result especially compared to the GFL FedSAM?\\n\\n***Ans for Q5):*** Thank you for your kindly reminder. We use different colors to distinguish best GFL performances under different settings including GFL-GM and GFL-PM. In GFL-GM, the single global model meet all datasets, while GFL-PM refers to the setting where personalized local model meet all datasets. We found PMs often perform not well in all datasets, and revert this with adaption of personalized model, which we found even surpass the GFL performance of GM. \\n- **ForestGreen:** overall best GFL results under the **GFL-GM and GFL-PM** setting (highest among those two settings).\\n- **Grey:** only the best results under the **GFL-GM** setting (highest of only the traditional **GFL-GM** setting).\\n\\nYou may misunderstand the experiment results we list in ***Table 2***. The focus of our experiments is to compare the **GFL performances of baselines and HPFL (including two columns GFL-GM and GFL-PM instead of only the first column GFL-GM)**, as shown in ***Table 2 in the main text***, HPFL gains most of the best GFL performances among all thest baselines including FedSAM.\"}", "{\"title\": \"Thanks for your reply\", \"comment\": \"Dear reviewer VDPF,\\n\\nThanks for your time in reviewing our paper and reply despite such a busy period. Your comments have helped us improve our presentation. If you have any further concerns, we will do our best to respond to them. \\n\\nBest regards and thanks,\\n\\nAuthors of #9027\"}", "{\"title\": \"Response to Reviewer VDPF - Part 2\", \"comment\": \"**Q3** :\\n> How to balance the computational and communication overheads brought by the storage and selection of personalized plug-in modules in the HPFL framework to improve model efficiency?\\n\\n***Ans for Q3):*** \\nWe provide three implementations of HPFL inference, each provides solution with different focus on computational/communication and storage/selection costs: \\n- (1) **Plug-in Updating:** only update the local plug-in stored locally at a client-defined frequency, which greatly reduce communication and computational costs of single inference; \\n- (2) **Cache All:** request all plug-ins and markers from the server to omit the communication costs at inference time, suitable for the applications where low latency is required; \\n- (3) **Cache Selected:** store the plug-ins that clients have chosen before, in real world, the number of distributions from other clients one client can meet tends to be limited, thus a single client may not need too many plug-ins to perform inference. This solution trades storage of serveral regularly-used plug-ins for significantly lower communication costs, which is usually dominate in FL systems. \\n\\n**Plug-in Updating** and **Cache All** is the **$\\\\alpha$** and **$\\\\beta$** implementation in ***the introdution***. \\n\\nBesides, we also attempts to reducing the number of plug-ins to control the overhead of plug-in storage and selection. \\n - **Controllable Number of Plug-ins:** In Appendix F.2.1, we provide some naive ideas on reducing the number of plug-ins, and show HPFL outperform the best GFL-PM baseline FedTHE with only 1/3 number of plug-ins by simply eliminating plug-ins with the calculated selection score. \\n\\n**Q4** :\\n> Is it possible that differential privacy protection during model selection in this article, or when using other privacy protection methods, may significantly affect model performance?\\n\\n***Ans for Q4):***\", \"expeiment_results_shows_differential_privacy_probably_do_little_harm_to_the_performance_of_hpfl\": \"In ***Table 13***, we have experimentally shown under the protection of Differential Privacy, HPFL still achieve the significantly better GFL performance over all baselines. Moreover, results in ***Table 8*** also shows that the model performance are not significantly affected by the Differental Privacy protection, the performances of HPFL didn't experienced significant decrease during the rising of noise coeffiency from 0 to 1000.\\n\\n\\n\\n> ***Reference***\\n> \\n> [1] Mingsheng Long, Han Zhu, Jianmin Wang, and Michael I Jordan. Deep transfer learning with joint adaptation networks. In ICML, 2017.\\n> \\n> [2] Maithra Raghu, Justin Gilmer, Jason Yosinski, and Jascha Sohl-Dickstein. Svcca: Singular vector canonical correlation analysis for deep learning dynamics and interpretability. In NeurIPS, 2017.\\n> \\n> [3] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In ICML, 2019. \\n> \\n> [4] Karimireddy, Sai Praneeth, Satyen Kale, Mehryar Mohri, Sashank Reddi, Sebastian Stich, and Ananda Theertha Suresh. \\\"Scaffold: Stochastic controlled averaging for federated learning.\\\" In ICML, 2020.\\n> \\n> [5] Chen, Hong-You, and Wei-Lun Chao. \\\"On Bridging Generic and Personalized Federated Learning for Image Classification.\\\" In International Conference on Learning Representations, 2022.\\n> \\n> [6] Qu, Zhe, Xingyu Li, Rui Duan, Yao Liu, Bo Tang, and Zhuo Lu. \\\"Generalized federated learning via sharpness aware minimization.\\\" In ICML, 2022.\\n> \\n> [7] Jiang, Liangze, and Tao Lin. \\\"Test-Time Robust Personalization for Federated Learning.\\\" In ICLR, 2023.\"}", "{\"summary\": \"The paper introduces Hot-Pluggable Federated Learning (HPFL), a framework aimed at bridging the gap between Generic Federated Learning (GFL) and Personalized Federated Learning (PFL) by proposing Selective Federated Learning (SFL). SFL optimizes PFL while allowing for the selection of personalized models (PMs) to enhance generalization performance across diverse test data. In HPFL, clients train personalized plug-in modules based on a shared backbone model, which are then uploaded to a server for selection during inference. This process also incorporates differential privacy to protect user data during the selection phase. Experimental results demonstrate that HPFL significantly outperforms traditional GFL and PFL methods, suggesting its applicability in various federated learning scenarios, including continual learning.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. The introduction of SFL effectively addresses the limitations of existing PFL methods, allowing for better adaptation to real-world scenarios where test data may differ significantly from local training data.\\n\\n2. The HPFL framework\\u2019s modular approach enables efficient communication and computation, making it suitable for practical applications in federated learning, including scenarios with resource-constrained clients.\", \"weaknesses\": \"1. The dependency on a common backbone model may lead to reduced performance if the backbone fails to generalize well across heterogeneous client data distributions.\\n\\n2. The plug-in selection process during inference could introduce additional computational delays, particularly for clients with limited resources, potentially hindering real-time performance.\\n\\n3. While the framework claims differential privacy protection, the effectiveness of this mechanism in preventing information leakage during plug-in selection remains to be empirically validated in diverse operational contexts.\", \"questions\": \"See the weaknesses.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to Reviewer 8vJb\", \"comment\": \"We thank the reviewer for taking the time to review. We appreciate that you find the proposed framework HPFL efficient and practical, the introduction of SFL effective. According to your valuable comments, we provide feedback below.\\n\\n**Q1** :\\n> The dependency on a common backbone model may lead to reduced performance if the backbone fails to generalize well across heterogeneous client data distributions.\\n\\n***Ans for Q1):*** \\n- **Low Sensitivity of backbone GFL performance:** HPFL does not require a pretty high performance backbone. Because even given a moderate backbone, HPFL can ensure the performance of plug-ins by finetuning on local data. HPFL is not greatly sensitive to the generality of backbone, as shown in experiment in ***Table 2 in the main text***, evidenced by little or even reverse PFL performance gap between low heterogenous setting (10 client, $\\\\alpha=0.1$), and high heterogenous setting (100 client, $\\\\alpha=0.05$), even given GFL performances significantly decrease in the high heterogeneity. \\n- **Orthogonality of backbone training methods:** Our main contribution and novelty is the design of HPFL framework, which is orthogonal to traditional FL. The specific training algorithm of the backbone can be any other GFL algorithm. Even if the backbone performance is not enough to support HPFL given the high heterogeneity, we can train the model with more advanced GFL algorithms like FedSAM, which take the data heterogeneity into consideration. \\n- **Easy access of appropriate backbone:** In the era of LLM, a pretrained backbone decent enough for fine-tuning under PFL performance is easy to get, such as LLama 3.\\n\\nFor these three reasons mentioned above, we believe backbone available to use in HPFL is easily accessible. \\n\\n**Q2** :\\n> The plug-in selection process during inference could introduce additional computational delays, particularly for clients with limited resources, potentially hindering real-time performance.\\n\\n***Ans for Q2):*** \\nTo show the risk of selection process hindering real-time performance. We analyze computation costs of selection process. Moreover, we discuss how to reduce the number of plug-ins and the frequency of carrying out selection process to relief HPFL from heavy computation costs.\\n - **Computation costs of selection** is correlated to the total number of plug-in training samples. **In a large FL system, samples on each client tend to be in little quantity**, so the number of training samples are controllable in real deployment. \\n - **Reduce number of plug-ins:** Within **an FL system involving millions of clients, many of them share similar plug-ins**. Therefore, in Appendix F.2.1, we propose initial ideas on controlling the number of plug-ins, and show that HPFL surpass the best GFL-PM baseline FedTHE with only 1/3 plug-ins, by simply using selection score to measure the similarity of plug-ins and abandon similar ones. \\n - In real implementation, **clients will keep the selected plug-in locally**. Therefore, in real deployment, **the update of plug-in doesn't happens from time to time**, as the distribution shifts happen in a relatively low frequency, e.g. traveling to another country or place, the climate change of a certain place. This behaviour is consistent with many real-world FL systems[1,2]. These FL updates often are set to happen only at spare time of clients[1,2,3], further reduce the burden in real-world FL system.\\n\\n**Q3** :\\n> While the framework claims differential privacy protection, the effectiveness of this mechanism in preventing information leakage during plug-in selection remains to be empirically validated in diverse operational contexts.\\n\\n***Ans for Q3):*** \\n - In **Appendix E.2 and Figure 15, 16**, we carry out the resorted image reconstruction by feature inversion method, and observe that after our protection against the markers, no raw image can be successfully reconstructed by inverting the representation through the pretrained global backbone model parameters. This results experimentally demonstrate that our privacy protection scheme is effective to protect HPFL from information leakage. \\n - Besides, we have carried out expeiments showing differential privacy do little harm to the performance of HPFL: In ***Table 13***, we have experimentally shown under the protection of Differential Privacy, HPFL still achieve the significantly better GFL performance over all baselines. Moreover, results in ***Table 8*** also shows that the model performance are not significantly affected by the Differental Privacy protection, the performances of HPFL didn't experienced significant decrease during the rising of noise coeffiency from 0 to 1000.\\n\\n> ***Reference***\\n> \\n> [1] Age-Based Scheduling Policy for Federated Learning in Mobile Edge Networks. In ICASSP, 2020.\\n> \\n> [2] CMFL: Mitigating communication overhead for federated learning. In ICDCS, 2019.\\n> \\n> [3] Towards federated learning at scale: System design. In SysML, 2019.\"}", "{\"summary\": \"This article proposes a new federated learning framework called Hot-Pluggable Federated Learning (HPFL), which aims to solve the performance gap problem between general federated learning (GFL) and personalized federated learning (PFL). Traditional GFL cannot cope with the diversity of data distribution, while PFL is only suitable for scenarios where local data distribution is similar. When the client encounters test data that is different from local data, PFL's personalized model has difficulty maintaining efficient generalization performance. To this end, this paper proposes a new problem framework of Selective Federated Learning (SFL), which enhances the effect of GFL by selecting an appropriate personalized model for each client in the inference stage. The HPFL framework divides the model into shared backbone and personalized plug-in modules. The client trains and uploads plug-ins based on local data. During inference, it can select appropriate plug-ins to adapt to different data distributions, while protecting data privacy through differential privacy.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"Originality\\uff1aThe HPFL framework proposed in this paper innovatively introduces a plug-in selection mechanism into federated learning, realizes the bridge between general models and personalized models, and solves the performance balance problem of traditional GFL and PFL. This is a new attempt in federated learning.\\n\\nQuality\\uff1aThe experimental part is relatively comprehensive, covering a variety of data sets and model verification, and enhancing security through differential privacy. The overall design is rigorous, and the results demonstrate the advantages of HPFL in performance and adaptability.\\n\\nClarity\\uff1aThe paper is well-structured, with clear background and problem descriptions, and clear algorithm design, framework details, and experimental procedures, making it easy for readers to understand its core contributions.\\n\\nSignificance\\uff1aThis study proposed a new solution to the adaptability problem of federated learning under heterogeneous data distribution, which has practical application potential and provides a new direction for the future development of federated learning.\", \"weaknesses\": \"\\u00b7Insufficient details of the selection mechanism (page 5, Section 3.3)\\uff1aHPFL uses multiple distance metrics such as MMD, SVCCA, and CKA to select plug-ins, but the specific algorithm steps and implementation details are rarely described. The article can add mathematical expressions or pseudocodes for some of the selection methods to increase readability and make it easier for readers to understand the robustness of the selection process.\\n\\n\\u00b7Selection of comparison methods (page 7&8, Table 2 and Table 3)\\uff1aThe paper compares a variety of GFL and PFL algorithms, but lacks a comparison of newer solutions that focus on heterogeneous data distribution problems. It is recommended to add more to further highlight the advantages of HPFL in performance and adaptability.\", \"questions\": \"\\u00b7How to balance the computational and communication overheads brought by the storage and selection of personalized plug-in modules in the HPFL framework to improve model efficiency?\\n\\n\\u00b7Is it possible that differential privacy protection during model selection in this article, or when using other privacy protection methods, may significantly affect model performance?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you for the response, I tend to keep my positive score.\"}", "{\"title\": \"Response to Reviewer kxCM - Part 1\", \"comment\": \"We sincerely thank the reviewer for taking the time to review. According to your insightful comments, we provide detailed feedback below.\\n\\n**Q1** :\\n> What is the main contribution of the HPFL that makes it outperform existing PFL models? This could be described by adding a table with core update rule of existing PFLs [including more recent studies] and the proposed method;\\n\\n***Ans for Q1):*** \\n| | GFL | PFL | HPFL (SFL) |\\n|--|-- |-- |-- |\\n| Focus on performance of | all datasets | local datasets | **all datasets** |\\n| training paradigm | train a single model by aggregation | fine-tuning local model | fine-tuning **model components** with a backbone |\\n| inference paradigm | use global model | use personalized models | **select appropriate personalized plug-in and the common backbone** then inference |\\n| real-world deployment | All clients share a common model | Clients can only use local-personalized models | **All clients can share all personalized plug-ins and a common backbone** |\\n\\n\\n- **Focus on performance of:** This column distinguish the target test data of different algorithms, showing HPFL is target on all datasets (global dataset) instead of local datasets as in PFL algorithms. \\n- **training paradigm:** This column show the difference of training process of all algorithms.\\n- **inference paradigm:** This column show the different pipeline when inference, **selection mechanism make sure our methods can get the suitable plug-in for inference**. \\n- **real-world deployment:** This column shows the differences when various algorithms deploy in real-world applications, here we **highlight the importance of sharing all personalized plug-ins** and regard this as our key contribution leading to better GFL performance than all PFL algorithms. \\n\\n\\nWith the above design, HPFL has the following advantages than previous GFL and PFL paradigms:\\n- **Superb performance on All datasets instead of local datasets.**\\n- **Lower resource requirements than model selection [6].**\\n- **Implementing a possible market mechanism of plug-in sharing between different users [5].**\\n- **Potential Application for asynchrnous and continuous learning scenarios [7, 8].**\\n\\n\\n**Q2** :\\n> What is the difference of plug-in module and vanilla personalized model? Intuitively they seem to be the same and help considering the local model to personalize the local model of each client.\\n\\n***Ans for Q2):*** \\n\\n- **Plug-in modules PFL models are the personalized part of models:** In HPFL, plug-ins are produced by finetuning on client data with frozen backbone. While vanilla personalized models are often personalized by directly finetuning on client data. \\n- **Plug-in modules can be selected and plugged:** Personalized plug-ins can be shared and plugged to all clients due to its light-weight feature. Moreover, since plug-in is a part of model, encoded middle feature (marker) can be shared and utilized to distinguish which plug-in it matches, however, for vanilla personalized model, the identifier information (auxiliary information) selection process based on is difficult to construct. \\n- **Selecting the complete personalized models is low efficient.** Storing the whole models on the server or client devices will occupy the memory as $nM$, in which $n$ is the number of clients, $M$ the size of the whole model. While storing the plug-ins require $M + nm$ memory, where $m << M$ is the size the of plug-ins.\"}", "{\"metareview\": \"a) Summary\\n\\nThe paper introduces Selective Federated Learning (SFL), a framework to bridge the gap between Generic Federated Learning (GFL) and Personalized Federated Learning (PFL) by enabling clients to share and selectively integrate lightweight plug-in modules trained on local data. The core insight is that personalized modules can enhance generalization across diverse test distributions when combined with a shared backbone and selected using markers. The main result demonstrates that the proposed Hot-Pluggable Federated Learning (HPFL) framework significantly outperforms state-of-the-art GFL and PFL methods while addressing privacy concerns through differential privacy and reducing system overhead with efficient plug-in sharing mechanisms\\n\\nb) Strengths\\n\\n- The Hot-Pluggable Federated Learning (HPFL) framework, which sets up a \\\"module store\\\" and picks the most relevant module for the test distribution. This framework is realistic and potentially impactful. This could be interesting to extend to a *marketplace* setting as well.\\n- Introduces privacy protection for shared plug-in markers using differential privacy, with theoretical guarantees and experimental validation.\\n- Extensive evaluations across multiple datasets and architectures demonstrate significant improvements over traditional GFL and PFL methods.\\n\\nc) Weakness\\n\\n- The core methodology\\u2014using modular plug-ins, markers, and differential privacy\\u2014largely combines existing techniques, with minimal theoretical innovation beyond problem formulation. Potential extensions could try analyze more interesting data and model acquisition techniques (see e.g. Lu et al. 2023: https://arxiv.org/abs/2403.13893). It is also clearly related to the routing modules in mixture of experts (MoEs): see Yadav et al. 2024: https://arxiv.org/abs/2408.07057.\\n- Performance depends heavily on the backbone model's ability to generalize across heterogeneous client data\\n\\nd) Decision to recommend **accept**\\n\\nThe authors introduce a system that allows clients to share and use small, adaptable components (plug-ins), which improves performance and reduces resource requirements. The experiments show that this approach works better than existing methods, and it also includes privacy protection to keep shared information secure. While some aspects build on existing ideas, the practical usefulness and strong results lead all the reviewers to recommend accepting this work.\", \"additional_comments_on_reviewer_discussion\": \"During the rebuttal, reviewers raised concerns about the originality of the approach, clarity in the selection mechanism, reliance on the backbone model, and adequacy of comparisons with recent methods. The authors addressed these by providing additional implementation details, revising notations, elaborating on the plug-in selection process, and expanding comparisons with related works. They clarified that the framework\\u2019s novelty lies in adapting personalized components for generalization, not in theoretical advancements. While the originality remains moderate, the practical contributions and thorough revisions demonstrated the approach's effectiveness and relevance, leading to a positive overall evaluation.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}", "{\"title\": \"Global Response\", \"comment\": \"We sincerely thank all reviewers for taking the time to review our work. We appreciate you find **our framework novel, innovative, general, efficient and effective** (Reviewer VDPF, FWNo and 8vJb), **our paper well-written** (Reviewer VDPF and kxCM) and **easy to follow** (Reviewer VDPF), **detailed and fair discussion of related works** (Reviewer kxCM), we **identify a substantial gap between GFL and PFL, and formulate a new problem SFL to bridge them together** (Reviewer VDPF, FWNo), **extensive experiments and ablation studies showing superior performance** (Reviewer VDPF, FWNo and 8vJb), **our method promsing and potential in applications** (Reviewer VDPF and 8vJb), we **provides a new direction for the future development of FL** (Reviewer VDPF).\\n\\nHere, we provide a summary of our responses to frequent questions for convenient reading.\\n\\n**Q1: Clarification on contributions and advantages:** (Reviewer VDPF, FWNo and kxCM)\\n\\n***Ans for Q1):***\", \"we_identify_our_advantages_compared_with_previous_works_as\": [\"**Superb performance on all datasets instead of local datasets.**\", \"**Lower communication resource required.**\", \"**Free from the requirement of a public dataset, which is diffult to get in many cases like medical information.**\", \"**Implementing a possible market mechanism of plug-in sharing between different users.**\", \"**Potential application for asynchrnous and continuous learning scenarios.**\", \", which are achieved by\", \"**We first formulate the performance gap between GFL and PFL, and though this fomulation we find a theoretical framework named Selective FL (SFL) that can utilize the excellent PFL performance of local models to boost the GFL performance.**\", \"**We instantiate SFL with an efficient and practical framework named HPFL, which incorporates the plug-in mechanism to improve GFL performance with PFL modules.**\", \"**We use markers help to identify the training data distribution of the plug-in modules and adapt test model accordingly.**\", \"**Q2: Selection method:** (Reviewer VDPF, FWNo)\", \"***Ans for Q2):***\", \"**Clarification**: we add the specific implementations and formulas of MMD, SVCCA, and CKA in Appendix D.2 to provide better elaboration on the selection process.\", \"**Issues on formulation and presentation**: We have revised the notation and text presentation to better clarify auxiliary information and selection function ***in Section 2.4 in our revision***.\", \"**Q3: Differential Privacy:** (Reviewer VDPF, FWNo, 8vJb and kxCM)\", \"***Ans for Q3):***\", \"**Novelty:** Compared with applying DP to models, applying DP to other shared information like intermediate features is significantly less explored.\", \"**Experimental Verification:**\", \"**In Table 13**, we experimentally shows under the protection of Differential Privacy, HPFL still achieve better GFL performance over all baselines.\", \"Moreover, **results in Table 8** also show that the model performance are not significantly affected by the Differental Privacy protection.\", \"In **Appendix E.2 and Figure 15, 16**, we observe that after our protection against the markers, no raw image can be successfully reconstructed with model inversion attack.\", \"**Q4: Discussion on extra system overheads** (Reviewer VDPF and 8vJb)\", \"***Ans for Q4):***\", \"We provide three real-world deployment schemes of HPFL inference, each targeted at a specific scenario to reduce the according bottleneck of system overheads:\", \"(1) In light-weight applications requiring low update frequency (like backend activities), the bottleneck is the storage and communication costs for frequent update, **Plug-in Updating** only update the local plug-in stored locally at a client-defined frequency, which greatly reduce inference costs;\", \"(2) In latency-sensitive applications, the bottleneck is communication time to fetch plug-ins, **Cache All** request all plug-ins and markers from the server to omit the need for communication at inference time;\", \"(3) When one client meets a limited number of distributions, the bottleneck comes from the frequent download of the same plug-in, we adopt **Cache Selected**, which stores the plug-ins that clients have chosen before.\", \"**Plug-in Updating** and **Cache All** are the **$\\\\alpha$** and **$\\\\beta$** implementation in the introdution.\", \"Below we briefly summarize the factors reducing the system overheads of HPFL.\", \"**Low update frequency and fewer transmitted parameters**, detailed explanation can be found in responses to Reviewers VDPF, 8vJb.\", \"**Controllable Number of Plug-ins and Storage**: In Appendix F.2.1, we provide some naive ideas on reducing the number of plug-ins, and show HPFL outperform the best GFL-PM baseline FedTHE with only 1/3 number of plug-ins by simply eliminating plug-ins with the calculated selection score.\", \"We have integrated the above mentioned modifications into ***our revision***.\"]}", "{\"title\": \"Response to Reviewer VDPF - Part 1\", \"comment\": \"We thank the reviewer for taking time to review. In light of your insightful comments, we offer responses below.\\n\\n**Q1** :\\n> Insufficient details of the selection mechanism (page 5, Section 3.3)\\uff1aHPFL uses multiple distance metrics such as MMD, SVCCA, and CKA to select plug-ins, but the specific algorithm steps and implementation details are rarely described. The article can add mathematical expressions or pseudocodes for some of the selection methods to increase readability and make it easier for readers to understand the robustness of the selection process.\\n\\n***Ans for Q1):*** Thank you for your insightful suggestion that help us clarify HPFL, due to the page limit, we add the specific implementations and formulas of MMD [1], SVCCA [2] and CKA [3] in ***Appendix D.2*** to provide better elaboration on the selection process. MMD\\n - MMD\\nGiven observations $X:=\\\\left\\\\\\\\{x_{1}, \\\\ldots, x_{m}\\\\right\\\\\\\\}$ and $Y:=\\\\left\\\\\\\\{y_{1}, \\\\ldots, y_{n}\\\\right\\\\\\\\}$, \\n$$\\nMMD_u^2[X, Y]= \\\\frac{1}{m(m-1)} \\\\sum_{i=1}^m \\\\sum_{j \\\\neq i}^m k\\\\left(x_i, x_j\\\\right)+\\\\frac{1}{n(n-1)} \\\\sum_{i=1}^n \\\\sum_{j \\\\neq i}^n k\\\\left(y_i, y_j\\\\right) \\\\\\\\\\n-\\\\frac{2}{m n} \\\\sum_{i=1}^m \\\\sum_{j=1}^n k\\\\left(x_i, y_j\\\\right) .\\n$$\\n, where $k(\\\\cdot)$ is the kernel function.\\n\\n - CKA\\n Let $K$ and $K^{\\\\prime}$ be two kernel functions defined over $\\\\mathcal{X} \\\\times$ $\\\\mathcal{X}$ such that $0<\\\\mathrm{E}\\\\left[K_c^2\\\\right]<+\\\\infty$ and $0<\\\\mathrm{E}\\\\left[K_c^{\\\\prime 2}\\\\right]<+\\\\infty$. Then, the alignment between $K$ and $K^{\\\\prime}$ is defined by\\n\\n $$\\n \\\\rho\\\\left(K, K^{\\\\prime}\\\\right)=\\\\frac{\\\\mathrm{E}\\\\left[K_c K_c^{\\\\prime}\\\\right]}{\\\\sqrt{\\\\mathrm{E}\\\\left[K_c^2\\\\right] \\\\mathrm{E}\\\\left[K_c^{\\\\prime 2}\\\\right]}} .\\n $$\\n\\n - SVCCA\\n 1. Input: $X, Y$\\n 2. Perform: SVD($X$), SVD($Y$). Output: $X^{'} = UX, Y^{'} = VY$\\n 3. Perform CCA($X^{'}$, $Y^{'}$). Output: $\\\\tilde{X} = W_{X}X^{'}$, $\\\\tilde{Y} = W_{Y}Y^{'}$ and $corrs = \\\\{\\\\rho_{1}, . . . , \\\\rho_{min(m_1,m_2)} \\\\}$, where $m_1$ and $m_2$ is the number of samples of $X$ and $Y$\\n\\n\\n\\n\\n**Q2** :\\n> Selection of comparison methods (page 7&8, Table 2 and Table 3)\\uff1aThe paper compares a variety of GFL and PFL algorithms, but lacks a comparison of newer solutions that focus on heterogeneous data distribution problems. It is recommended to add more to further highlight the advantages of HPFL in performance and adaptability.\\n\\n***Ans for Q2):*** As far as we are concerened, most previous works aiming at dealing with data heterogeneity can be categoried into two types: \\n- (1) **how to train a better global model given the heterogenous data scattered on different clients**, usually though (i) **designing more robust aggregation schemes**, (ii) **reduce the heterogeneity of trained local models when given heterogenous local data**. For (1,i), SCAFFOLD [4] is the representative of this type of methods, however, very few recent work focus on this path, so we didn't choose recent representitive from this category. FedRoD [5] optimize local models towards class-balanced objectives to improve GFL performance, which is a implementation of (1, ii) scheme. FedSAM [6] also carries out (1, ii) scheme by applying Sharpness Aware Minimization (SAM) local optimizer; Moreover, our method is orthogonal to this type of methods, since HPFL can adopt these methods to train better backbone.\\n- (2) **how to adaptively adjust the inference model to mitigate the test-time distritbution shifts**. FedTHE [7] achieve the goal mentioned in (2) by adjust the enesemble weight of global and local heads according to test data. The experimental results of HPFL and these baselines show that HPFL have advantages over those kinds of methods.\"}", "{\"comment\": \"Thank you for the detailed response, I will raise my score accordingly.\"}", "{\"summary\": \"After authors responses: The rating has been updated considering the authors' inputs and clarifications. I appreciate their efforts in providing those responses.\\n\\nIn order to solve the selective FL (SFL) problem, this paper leverages the model components that attempt to edit a submodel for specific purposes to design a framework referred to as Hot-Pluggable Federated Learning (HPFL). In HPFL, clients individually train\\npersonalized plug-in modules based on a shared backbone, and upload them with a plug-in marker on the server modular store. During the inference stage, a selection algorithm allows clients to identify and retrieve suitable plug-in modules from the modular store to enhance their generalization performance on the target data distribution. This paper also provides differential privacy protection during\\nthe selection with theoretical guarantee. Key contributions can be summarized as:\\n-\\tIdentifying a major gap between Generic FL (GFL) and Personalized FL (PFL), and formulate a new problem SFL to bridge this performance \\n-\\tDeveloping a general, efficient and effective framework HPFL, which practically solves SFL and adding noise on communicated markers to provide differential privacy protection with theoretical guarantee.\\n-\\tExperiments on four datasets and three neural networks to demonstrate the effectiveness of HPFL\", \"soundness\": \"3\", \"presentation\": \"4\", \"contribution\": \"3\", \"strengths\": \"The paper is well-written. It fairly cites prior works that it built on and shows how it leverages those solutions. It clarifies what is the problem and what are research questions to answer.\\nAfter carefully reviewing authors' clarifying points and responses to my concerns and other reviewers, I increased two of my scores.\", \"weaknesses\": \"The originality of this paper is not clear. For instance, the following are two major claimed contributions according to the paper \\u201cIdentifying a major gap between Generic FL (GFL) [e.g., (Karimireddy et al., 2019; Woodworth\\net al., 2020; Tang et al., 2022b)'s works] and Personalized FL (PFL) [e.g., (Li & Wang, 2019; Chen & Chao, 2021; Li et al., 2021c):'s works], and formulate a new problem SFL to bridge this performance; and Developing a general, efficient and effective framework HPFL, which practically solves SFL and adding noise on communicated markers to provide differential privacy protection with theoretical guarantee.\\u201d\\nHowever, the methodology seems to be a combination of existing works on PFL while leveraging several existing work with minimal advances. It is not clear how the mentioned theorems add value to the literature, for example, this statement is vague and is not adequately explained: \\u201csolving SFL means that clients achieve performance in GFL as high as in PFL.\\u201d. It is not clear how plug-in marker can contribute to bridging the gap between GFL and PFL? It would be very helpful to explain elaborately. Specifically, please provide a more detailed explanation and concrete example of how the plug-in markers specifically help bridge the gap between GFL and PFL performance.\\nWhile using differential privacy can add value to the method, it is not clear whether it can benefit from local DP, central DP, or both? Please clarify which type(s) of differential privacy (local, central, or both) are used, and to provide a more detailed explanation of how the plug-ins interact with the DP mechanisms and what novel contributions are made in this area. Specifically it could be explained how plug-ins affect DP and what is the contribution in this part of the paper.\\nFigure 2 should also include results of HPFL. The comparison should be expanded to cover more advanced PFL studies to showcase how HPFL performs compared to those methods. Currently is mainly focuses on basic PFL algorithms for comparison purposes. For instance, some key papers in this domain can help authors to provide a more compelling comparison among proposed method and existing PFL solutions, such as FedAlt/FedSim of Krishna et al, 2022@ICML and for the specific case of PFL with differential privacy, Hu et al, 2020 @ IEEE IoT Journal.\", \"questions\": \"Q1. What is the main contribution of the HPFL that makes it outperform existing PFL models? This could be described by adding a table with core update rule of existing PFLs [including more recent studies] and the proposed method;\\nQ2. What is the difference of plug-in module and vanilla personalized model? Intuitively they seem to be the same and help considering the local model to personalize the local model of each client.;\\nQ3. What is the novelty of integrating DP in the algorithm? How does it lead to advancing the proposed HPFL solution? Is it HPFL+DP or the integration has some challenges? If so, what are the challenges and how does this paper tackle them? Why can\\u2019t we use other privacy preservation mechanisms? ;\\nQ4. Detailed comparison with two types of studies: i. existing PFL algorithms and comparing with HPFL, ii. Existing privacy preserving algorithms and comparison with DP;\\n5. How can the plug-in marker contribute to bridging the gap between GFL and PFL?\\nQ6. Can you please clarify which type(s) of differential privacy (local, central, or both) are used, and to provide a more detailed explanation of how the plug-ins interact with the DP mechanisms and what novel contributions are made in this area?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Window for discussion is closing\", \"comment\": \"Dear Reviewer kxCM,\\n\\nThanks a lot for your time in reviewing and reading our response and the revision. Thanks very much for your valuable comments. We sincerely understand you\\u2019re busy. But as the window for discussion is closing, would you mind checking our responses and and confirm whether you have any further questions? We look forward to answering more questions from you.\\n\\nBest regards and thanks, \\n\\nAuthors of #9027\"}", "{\"title\": \"Response to Reviewer FWNo - Part 1\", \"comment\": \"We sincerely thank the reviewer for taking the time to review. We appreciate that you find our proposed framework novel, general and efficient, our experiments and ablation studies comprehensive. According to your valuable comments, we provide detailed feedback below and add them into our main text or appendix in the revision. We hope that these changes have addressed your concerns and improved the overall quality of our work.\\n\\n\\n**Q1** : \\n> It\\u2019s adorable that the authors define a paradigm to bridge the gap between GFL and PFL, but the selective method are not novelty enough. For example, [1] allowed each client to choose the appropriate scale model to train. And the training process in HPFL are identical to the split federated learning, the authors only add another select process in inference process.\\n\\n***Ans for Q1):*** \\nThanks for your valuable comments. \\n - Differences with [1]: As far as we know, only selection in [1] happens when deciding which client takes part in the training round according to their dataset size, it's quite different from the plug-in selection of HPFL in three ways: (1) First, our method attempts to deal with the notorious data heterogeneity problem to **select appropriate well-trained plug-in models**, while [1] attempt to achieve by knowledge distillation; (2) the selection of participating clients are completely random with the possiblity in proportion to their dataset size, instead our selection is based on the matching degree of task markers and plug-in markers, which measured by Maximum Mean Discrepancy (MMD); (3) our selection only happens in inference phase to adapt the model used in test time according to the test data, while the selection of clients happens every training round in [1]. (3) Moreover, **their method requires a public unlabelled dataset** on server to perform knowledge distillation (KD) to transfer the knowledge from clients model to server, which is not always accessible especially for fields where privacy are highly evaluated, such as medicine and finance. \\n - Differences with split federated learning: (1) model decoupling is a commonly-used technique in FL, thus we do not regard this technique as our main novelty, as claimed in **Section 3.2**, this technique serves for the purpose of reducing the storage and communication costs in real implementation, i.e. any training techique can be incorporated into HPFL as long as it get light-weight model components to act as plug-ins. (2) split federated learning often focus on PFL [2] or vertical FL [3] setting, however, our method HPFL aims to enhance GFL performance with the plug-in trained with PFL methods, instead of PFL performance itself. (3) the model architectures of splif FL are fixed to ensure the aggregation of client-side model update, while HPFL support heterogenous personalized part of model with different architecture since model component are pluggable and personalized. (4) to support dynamic hot-pluggable module selection, HPFL further considers the prototypes named markers, while split federated learning only consider computation partition and parallelism. (5) the design of HPFL supports asynchronous and one-shot FL scenarios, while the split FL cannot support due to the requirenment of client-side model update aggregation.\\n - Our Novelty: we would like to highlight our major novelty as: (1) we first formulate the performance gap between GFL and PFL, and though this fomulation we find a theoretic framework named Selective FL (SFL) that can utilize the excellent PFL performance of local models (implemented as plug-ins in HPFL) to boost the GFL performance. In our knowledge, this is the first work that enhances GFL through learning, sharing and selecting model components, instead of classic paradigm relying on single global model. (2) we instantiate SFL with a efficient and practical framework named HPFL and (3) conduct comprehensive experiments and ablation studies to demonstrate the advantage of HPFL over previous methods. (4) we supports asynchronous, one-shot and continual FL scenarios with our practical and efficient framework HPFL.\"}", "{\"title\": \"Thanks for raising the score\", \"comment\": \"Dear reviewer kxCM,\\n\\nThanks for your efforts in reviewing our paper and raising the score. Your constrcutive and concrete suggestions have greatly contributed to our paper. We are pleased to respond to you further concerns if you have any.\\n\\nBest regards and thanks,\\n\\nAuthors of #9027\"}" ] }

[ "Jiajun He", "Gergely Flamich", "José Miguel Hernández-Lobato" ]

[ "Model compression", "bits-back", "bit-back coding", "coding", "LLMs", "Transformers" ]

{ "note_id": [ "uPfbOTvs2c", "pYN3MIFmG3", "pP0QOWFjTR", "o1D2Dn3kkG", "o0GkvcO0Gb", "lxrt9J1WO9", "jIK9O74hBS", "j6Ikj9za4j", "hPIqJTx60Y", "hG5O8j59JS", "hClVPujTrb", "f4lH3dvzPq", "bZMCqkbQYT", "YrLw6W3ZRl", "Y5HjQbUBMO", "UOP5YgX0Xt", "GFM6H2QbP9", "DIJ1y58I27", "CdxA5jM74K", "BPKe0SwOBZ", "AasjfUiOKK", "AORsLGf5cP", "AMXGt6lYCz", "43EuSV3eOK", "3MZ9LoFF99", "1OyoWO9iaC", "1BNhhDDAJJ", "0XH1nUCXBZ" ], "note_type": [ "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "meta_review", "official_review", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment" ], "note_created": [ 1732545651011, 1732503881959, 1732014914422, 1732016674316, 1730534748416, 1732229021783, 1734250469651, 1730673336422, 1732293425430, 1732016456712, 1730887470713, 1732292776454, 1733059596550, 1732228949860, 1732834741907, 1737524022539, 1732737113715, 1732280677954, 1732285286447, 1732228870466, 1732736982172, 1733186677626, 1732016903918, 1730706650004, 1733222458016, 1732737012313, 1732228903984, 1732015807149 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission10052/Authors" ], [ "ICLR.cc/2025/Conference/Submission10052/Reviewer_Li28" ], [ "ICLR.cc/2025/Conference/Submission10052/Authors" ], [ "ICLR.cc/2025/Conference/Submission10052/Authors" ], [ "ICLR.cc/2025/Conference/Submission10052/Reviewer_cA7j" ], [ "ICLR.cc/2025/Conference/Submission10052/Authors" ], [ "ICLR.cc/2025/Conference/Submission10052/Area_Chair_hrx2" ], [ "ICLR.cc/2025/Conference/Submission10052/Reviewer_r3xU" ], [ "ICLR.cc/2025/Conference/Submission10052/Authors" ], [ "ICLR.cc/2025/Conference/Submission10052/Authors" ], [ "ICLR.cc/2025/Conference/Submission10052/Reviewer_kYZS" ], [ "ICLR.cc/2025/Conference/Submission10052/Authors" ], [ "ICLR.cc/2025/Conference/Submission10052/Authors" ], [ "ICLR.cc/2025/Conference/Submission10052/Authors" ], [ "ICLR.cc/2025/Conference/Submission10052/Reviewer_cA7j" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission10052/Authors" ], [ "ICLR.cc/2025/Conference/Submission10052/Reviewer_r3xU" ], [ "ICLR.cc/2025/Conference/Submission10052/Reviewer_cA7j" ], [ "ICLR.cc/2025/Conference/Submission10052/Authors" ], [ "ICLR.cc/2025/Conference/Submission10052/Authors" ], [ "ICLR.cc/2025/Conference/Submission10052/Reviewer_kYZS" ], [ "ICLR.cc/2025/Conference/Submission10052/Authors" ], [ "ICLR.cc/2025/Conference/Submission10052/Reviewer_Li28" ], [ "ICLR.cc/2025/Conference/Submission10052/Authors" ], [ "ICLR.cc/2025/Conference/Submission10052/Authors" ], [ "ICLR.cc/2025/Conference/Submission10052/Authors" ], [ "ICLR.cc/2025/Conference/Submission10052/Authors" ] ], "structured_content_str": [ "{\"comment\": \"Thank you for your reply; we now address your comments below.\\n\\n1. > However, I believe that the author\\u2019s method can provide significant latency and throughput benefits if some accuracy degradation is acceptable.\", \"we_are_unsure_about_what_you_mean_by_this_comment_for_two_reasons\": \"a. As we have noted in our rebuttal already, our proposed approach has **no influence on the inference latency and throughput** of the LLM. It only slightly affects the time it takes to load the model weights from the hard drive into the memory.\\n\\nb. Our paper demonstrated that **our method does not lead to accuracy degradation.** In Table 1 of our paper, we show that there is no statistically significant accuracy degradation after we apply our compression method: roughly half of the time, the model accuracy slightly decreases, while half of the time, it increases after we apply our method.\\nTherefore, **there is no inherent trade-off between accuracy degradation and latency or throughput in our approach.**\\n\\nTo sum up, our contribution focuses on connecting bits-back coding with model compression for efficient storage and transmission. This is a standard, general-purpose model compression task. For its intended purpose, our current approach is already complete, efficient, and practical, as we have demonstrated in our paper and prior replies. Our method does not influence latency and has no trade-off between accuracy degradation and latency or throughput.\\n\\n2. > In the generation phase of LLM inference, the batch size is usually small, and the proposed method can benefit from the memory-bound nature of LLM inference.\\n\\nThis is an interesting suggestion, but we do not quite understand it. We note that our contribution focuses on connecting bits-back coding with model compression for efficient storage and transmission. The scenario you describe seems to diverge from the intended design of our work. Could you please further clarify how this connects with our proposed approach?\\n\\n3. > I think that if the proposed method can focus on optimizing the decoding algorithm, it can improve latency and throughput by sacrificing additional computational cost and reducing significant memory communication cost.\\n\\nTo clarify, our proposed method does not influence latency and throughput. Rather, it aims to reduce the space needed to store the model on the hard drive and to reduce the transmission cost of sharing models. We wonder whether our usage of the term \\u201c*decoding*\\u201d might be causing this misunderstanding. We use the term from the source coding and information theory perspective, referring to reading and processing the compressed weights from the hard drive and loading them into memory. You can understand this term as \\\"*decompression*\\\", rather than the process in LLM inference. For a more elaborate discussion on this topic, please refer to [our global response to the reviewers](https://openreview.net/forum?id=B8aHIDSi7E&noteId=hPIqJTx60Y).\\n\\nWe thank you once again for your effort to review our paper. Given that not much time is left in the discussion period, we would appreciate it if you could clarify your comments as soon as possible. On the other hand, if we have addressed your concerns and clarified our position, we kindly invite you to reconsider raising your score.\"}", "{\"title\": \"Response to the authors\", \"comment\": \"Thank you for the detailed response. Most of my questions have been clarified. However, I believe that the author's method can provide significant latency and throughput benefits if some accuracy degradation is acceptable. In the generation phase of LLM inference, the batch size is usually small, and the proposed method can benefit from the memory-bound nature of LLM inference. I think that if the proposed method can focus on optimizing the decoding algorithm, it can improve latency and throughput by sacrificing additional computational cost and reducing significant memory communication cost. For this reason, I would like to keep the original score unless further explanation is provided.\"}", "{\"comment\": \"Thank you for your detailed review and constructive questions. We are glad that you found our paper novel, insightful, and to have the potential to have a broader impact. We now reply to your two questions below.\\n\\n1. **Question 1**: \\n>the paper relies heavily on bits-back coding. However, the paper does not properly connect SliceGPT with the previously proposed bits-back coding. It is hard to understand the actual algorithm.\\n\\nThis seems to be a misunderstanding. There is no a priori connection between SliceGPT and bits-back coding. In fact, making this connection is one of our contributions.\\n\\nBits-back coding is a data compression algorithm that works in cases where multiple choices are acceptable or equally good. In our case, as we explained in Remark 3.1, SliceGPT introduces symmetries into Transformers. Specifically, different weights, up to any random rotation, are equally good. Therefore, we can connect bits-back coding to SliceGPT. We then explain the benefit of this connection through an informal calculation of the codelength. \\n\\nWe appreciate this concern and agree that this connection does not seem crystal clear at first glance. Therefore, we have added a more detailed motivation from Lines 140-144.\\n\\n2. **Question 2**: \\n>It is questionable if the method can be applied in the real world \\u2026 the run speed of this method could be slower than that of the vanilla Transformer model.\\n\\nThank you for raising this question. However, this is not a concern:\\n\\n(1) Our approach aims to reduce the storage space and transmission cost. Therefore, we will only run the encoding algorithm when saving the model and decoding when loading the model. Once the model is decoded and loaded into memory, the inference time will be identical to the vanilla SliceGPT.\\n\\n(2) We measured the encoding and decoding time and updated our manuscript to include the results and discussion. We list Table 2 in the updated manuscript for an easier reference: \\n\\n| Model Name | OPT-1.3B | OPT-1.3B | OPT-2.7B | OPT-2.7B | OPT-6.7B | OPT-6.7B | OPT-13B | OPT-13B |\\n|----------------|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|:-------:|:-------:|\\n| Slicing | 20% | 30% | 20% | 30% | 20% | 30% | 20% | 30% |\\n| Encoding time | 15 s | 13 s | 30 s | 24 s | 2.5 min | 1.7 min | 6.5 min | 4.1 min |\\n| Decoding time | 6 s | 5 s | 14 s | 11 s | 1.2 min | 45 s | 2.5 min | 2 min |\\n\\n\\nFor a more detailed discussion, please refer to the runtime analysis paragraph on page 10. As we can see, our algorithm only takes seconds to minutes on a consumer-grade GPU. Also, as we discussed in lines 501-511, we can even parallelize this process to accelerate the execution further. \\n\\nTherefore, the runtime does not pose an obstacle to deploying our algorithm in real-world applications. Moreover, our algorithm can also be executed on a CPU, further broadening its applicability. We also include the encoding and decoding times for CPU execution in Appendix C.1 in the updated manuscript.\\n\\n*Thank you once again for taking the time to review our work and our response. We believe we have addressed your questions. If you find our clarifications satisfactory, we kindly invite you to consider raising your score.*\"}", "{\"comment\": \"Thank you for your valuable feedback, which helped us improve our manuscript. We are glad that you appreciate that our method is easy to adopt. Below, we reply to your concerns one by one.\", \"1_weakness_1\": \">The improvement of 3~5% appears small unless the method does not impose other overheads. However, there is no detailed analysis of how much memory the other components, such as the correction code, use.\\n\\nThis a misunderstanding. The 3-5% bits already take the correction code into account, so the 3-5% savings are net. In fact, the length of the correction code is negligible. We appreciate that you pointed out this potential confusion, and therefore, we have updated the caption of Table 1 to clarify this.\\n\\n2. **Weakness 2**: \\n\\n> the work does not include any analysis of the overhead in terms of the time required for inference caused by applying additional computations to the model. \\n\\nThank you for raising this question. We have updated our manuscript to include the results and discussion on the encoding and decoding time. \\nPlease refer to Table 2 and the runtime analysis paragraph on page 10. We also list Table 2 below for easier reference:\\n\\n| Model Name | OPT-1.3B | OPT-1.3B | OPT-2.7B | OPT-2.7B | OPT-6.7B | OPT-6.7B | OPT-13B | OPT-13B |\\n|----------------|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|:-------:|:-------:|\\n| Slicing | 20% | 30% | 20% | 30% | 20% | 30% | 20% | 30% |\\n| Encoding time | 15 s | 13 s | 30 s | 24 s | 2.5 min | 1.7 min | 6.5 min | 4.1 min |\\n| Decoding time | 6 s | 5 s | 14 s | 11 s | 1.2 min | 45 s | 2.5 min | 2 min |\\n\\n\\nAs we can see, our algorithm only takes seconds to minutes on a consumer-grade GPU. We also provide the runtime on CPU in Appendix C.1. Also, as we discussed in lines 501-511, we can even parallelize this process to accelerate the execution further. \\n\\nAdditionally, we note that our approach aims to reduce the storage space and transmission cost. Therefore, we will only run the encoding algorithm when saving the model and decoding when loading the model. Once the model is decoded and loaded into memory, the inference time will be identical to the vanilla Transformer.\\nTherefore, when measured end-to-end, our approach only increases the whole procedure by a few seconds/minutes. Therefore, this is not a big obstacle for practical usage.\\n\\n3. \\n\\n>The beginning of Section 2 uses the word \\u201cdelving\\u201d prominently. As the word \\u201cdelve\\u201d is strongly associated with large language model outputs, we advise the authors to rephrase the sentence.\\n\\nThank you for your suggestions! We have rephrased this sentence.\\n\\n\\n\\n*Thank you once again for taking the time to review our work and our response. We believe we have addressed your questions. \\n We are happy to discuss any further concerns you might have. However, If you find our clarifications satisfactory, we kindly invite you to raise the score.*\"}", "{\"summary\": \"This work proposes to apply a coding scheme to utilize symmetries made available by the SliceGPT method for training-free weight-only quantization. The resulting method achieves an additional 3-5% reduction in the total weight sizes for SliceGPT compressed models. The authors show that their resulting models do not diverge significantly from the original models when evaluated on tasks such as PIQA, WinoGrande, and HellaSwag.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"1. The proposed method is training-free, making the result independent from the choice of calibration set.\\n2. The proposed method can be computed using only a CPU, without heavy computational requirements. This will make adoption easier.\", \"weaknesses\": \"1. The improvement of 3~5% appears small unless the method does not impose other overheads. However, there is no detailed analysis of how much memory the other components, such as the correction code, use.\\n\\n2. Additionally, the work does not include any analysis of the overhead in terms of the time required for inference caused by applying additional computations to the model. Even if computing the rotation can be performed on CPU, there should be an analysis of the effect on inference latency when measured end-to-end.\", \"questions\": \"The beginning of Section 2 uses the word \\u201cdelving\\u201d prominently. As the word \\u201cdelve\\u201d is strongly associated with large language model outputs, we advise the authors to rephrase the sentence.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Thank you for your review! Please consider our response\", \"comment\": \"Thank you again for the effort you put into our work. As only a few working days are left in the discussion period, we would like to ask if our response has satisfied your concerns. If so, we kindly invite you to consider raising the score. If any concerns remain, we are happy to discuss them further here.\"}", "{\"metareview\": \"This paper proposes a method to compress rotationally symmetric weight matrices. Due to their rotational symmetry, the encoding of these weight matrices can be redundant. To address this, the authors propose the bits-back coding algorithm for compression.\", \"main_strengths\": [\"The paper makes an interesting algorithmic contribution.\"], \"main_weaknesses\": [\"The proposed method has only been benchmarked on SliceGPT pruning and Transformers.\", \"The achieved savings are limited.\", \"It will help if the authors take concrete steps to demonstrate broader applicability of the method.\"], \"additional_comments_on_reviewer_discussion\": \"Some points have been clarified by the authors during the rebuttal:\\n- The coding scheme is primarily designed to reduce storage and transmission costs and is therefore not significantly related to latency in model predictions.\\n- The coding scheme uses rotational symmetry to save bits, which complements typical compression methods.\\n\\nWhile these points are clear, I believe the paper could be further improved by demonstrating broader applications beyond a single pruning method. It would also be useful to explore general rotational symmetries in weight matrices and evaluate the potential savings achieved through this approach.\"}", "{\"summary\": \"The paper presents a novel training-free compression technique of large language models that exploits rotational symmetries in the weight space. It uses bits-back coding, a compression strategy that takes advantage of these rotational symmetries to compress Transformer models by about three to five percent while impacting the model's perplexity in a negligible way. The method was tested on SliceGPT-pruned Transformers, namely the OPT and Llama-2.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": [\"The paper presents bits-back coding used on neural network models, mainly focusing on enlarging language compression.\", \"The proposed method is computationally feasible since it runs without retraining.\", \"The paper's novel technique is evaluated on models, such as OPT and Llama-2, demonstrating performance metrics are not significantly affected in terms of perplexities drop.\"], \"weaknesses\": [\"This approach is inherently SliceGPT pruning and Transformer-specific architecture, which may also limit its use to other neural networks or pruning techniques.\", \"The methodology relies only on Transformer architectures, so applicability to lighter models suited to edge devices could be considered.\"], \"questions\": [\"What is the prospective effectiveness of the method when it comes to implementation on the models with precision format lower than float16?\", \"Is it possible to apply the bits-back coding method to the architectures that are not transformers or the architectures compressed with different methods?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"3. **Review cA7j raised a good question: as our aim is to reduce the storage and transmit cost, why we do not just use some source coding algorithm?**\\nWe believe that question allows us to see our work from a new angle and, hence, also respond to it in the global response:\\n\\nTo answer this question, we highlight that we should not directly compare our proposed method to these source coding algorithms. In fact, we can view our proposed algorithm as a pre-processing step before universal source coding. To be completely clear, we suggest that we combine our proposed method with a source coding algorithm as follows:\\n\\n1. We obtain some network weights with rotational symmetries.\\n\\n2. We use our method to eliminate the redundancies induced by the rotational symmetries in the weights.\\n\\n3. We apply a universal source coding algorithm (e.g. Zstandard) to the output of our algorithm.\\n\\n4. We save the output of this pipeline to the disk.\\n\\nIn our paper, we looked at the gains we get if we apply our method only, without any universal source coding. In this response, we showcased that we still retain significant gains if we run the pipeline we suggest above, compared to just running universal source coding without our method.\\nIn particular, we compared the pipeline suggested above to universal source coding (we used ZIP in our experiment) on the OPT-2.7B model. Using ZIP only, the compressed size comes to 3.97 GB while using our suggested bits-back pipeline, it reduces to 3.79 GB, and approximately 5% gain in storage size as before.\", \"there_is_an_intuitive_reason_for_retaining_the_gain_even_after_source_coding\": \"ZIP (or any other general-purpose universal source code) is unaware of the redundancies introduced by the rotational symmetry in the weights and, therefore, cannot utilize it to reduce space. **From this perspective, the storage savings resulting from our bits-back method are \\u201corthogonal\\u201d to the savings that result from source coding.**\"}", "{\"comment\": \"Thank you for your valuable feedback. We are encouraged by the positive comments around soundness and novelty. We address your concerns in the following.\\n\\n1. **Weakness 1 & Question 2**\\n\\n>This approach is inherently SliceGPT pruning and Transformer-specific architecture, which may also limit its use to other neural networks or pruning techniques. \\n\\n>Is it possible to apply the bits-back coding method to the architectures that are not transformers or the architectures compressed with different methods?\\n\\nWe agree that the algorithm described in our manuscript focuses on SilceGPT. However, we emphasize that the concept of bits-back coding applied to neural networks is general and can be extended to other architecture exhibiting symmetries. One of our key contributions, as noted in the conclusion section, is establishing a connection between bits-back coding and networks with symmetrical properties. Below, we provide two examples where our algorithm (or its variations) can be employed to achieve free bits-back. \\nWe have also included these discussions in the Conclusion, Limitations, and Future Directions Section in our updated manuscript.\\n\\n(1) Applying our approach to encoding LoRA modulation or weights decomposed in LoRA-style. Specifically, LoRA approximates a matrix $M$ by $M=BA$. We can see applying $Q$ to $B$ and $Q^T$ to $A$ will leave M invariant. Therefore, our proposed approach can be seamlessly applied in this context.\\n\\n(2) Neural networks with permutation symmetry. It is well known that in many architectures, such as standard MLPs, permuting the hidden units in one layer and applying the reverse permutation to the subsequent layer leaves the output unchanged.\\nWe can modify our approach to apply bits-back coding in this case.\\nSpecifically, we can define a canonical order for hidden units based on a predefined criterion (e.g., sorting the weights or biases corresponding to each unit in descending order). During encoding, a permutation can be randomly selected by decoding bits from the current bitstream. During decoding, this permutation can be easily recovered by rearranging the hidden units back to their canonical order.\\nIn fact, it is a standard fact that permutation matrices are orthogonal (i.e., they can be thought of as rotations). Therefore, they can also be seen as a special case of our approach. \\n However, the gain here is smaller than for the rotational symmetry case, as the permutation invariance of the hidden units introduces significantly less redundancy compared to rotations in general. \\n\\n2. **Weakness 2**\\n\\n> Applicability to lighter models suited to edge devices could be considered.\\n\\nThank you for your valuable suggestion; applying our method to smaller models could be quite impactful! We only considered SliceGPT in our paper, because its parameters exhibit a very clear rotational invariance that we could exploit. If the reviewer can recommend other, smaller architectures with well-known parameter symmetries, we would be very grateful and interested to hear!\\n\\n\\n3. **Question**\\n\\n>What is the prospective effectiveness of the method when it comes to implementation on the models with precision format lower than float16?\\n\\nQuantization can impact the performance of our approach. In our algorithm, the rotated matrix $WQ$ is encoded into the bitstream. During decoding, we use eigenvalue decomposition to recover $Q$ from $WQ$. If $WQ$ is quantized to a low precision, the quantized version may differ from the original $WQ$. As a result, performing eigenvalue decomposition on the quantized matrix may produce a recovered $\\\\hat{Q}$ that has a larger difference compared to the original $Q$.\\n\\nTo illustrate this impact, we provide an extra experiment in Appendix C.2 in our updated manuscript. We apply simple linear quantization to reduce the precision to 11-15 bits and measure the compression rate with bits-back coding. We can see that as the precision decreases, the bits saved by our approach become smaller. This is because the size of the correction code increases to compensate for the error caused by lower precision. \\nOur proposed approach consistently delivers gains when the precision is larger than 12-13 bits. We also want to note that we apply the simplest linear quantization in this simple demonstration. A better quantization strategy and correction code can further improve the performance of our approach. Designing a method that would allow us to quantize the weights below 12-bit precision is an interesting avenue for future research. \\n\\n\\n*Thank you once again for taking the time to review our work and our response. We are happy to discuss any further questions you might have. However, should we have addressed your concerns, we kindly invite you to raise the score.*\"}", "{\"summary\": \"In this paper, the authors highlight that the rotational symmetries of SliceGPT introduce redundancies. Based on SliceGPT, they propose further compressing the weights by the bits-back coding algorithm. Specifically, rather than treating all weight configurations as unique, they encode weights up to an equivalence class defined by rotations, enabling smaller memory requirements. They conducted experiments on several benchmarks with multiple models. The results verify the effectiveness of their proposed method.\\n\\n**Strengths**\\n1. The proposed method is novel and insightful\\n2. The proposed method is well-motivated and has the potential to make a broader impact.\\n3. The experiment results are promising.\\n\\n**Weaknesses**\\n1. The writing is not self-contained. Specifically, the paper relies heavily on bits-back coding. However, they do not properly connect SliceGPT with the previously proposed bits-back coding. It is hard to understand the actual algorithm.\\n2. It is questionable if the method can be applied in the real world given the compression/decompression and matrix decomposition procedures involved. The run speed of this method could be slower than that of the vanilla Transformer model.\\n\\nIn summary, the paper is insightful and well-motivated. However, the writing is not self-contained and the overhead could hinder the real-world application. As a result, I recommend a weak acceptance.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"3\", \"strengths\": \"1. The proposed method is novel and insightful\\n2. The proposed method is well-motivated and has the potential to make a broader impact.\\n3. The experiment results are promising.\", \"weaknesses\": \"1. The writing is not self-contained. Specifically, the paper relies heavily on bits-back coding. However, they do not properly connect SliceGPT with the previously proposed bits-back coding. It is hard to understand the actual algorithm.\\n2. It is questionable if the method can be applied in the real world given the compression/decompression and matrix decomposition procedures involved. The run speed of this method could be slower than that of the vanilla Transformer model.\", \"questions\": \"See the weaknesses.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you for your reply; you make excellent points, and they made us see our work from a new angle.\\n\\n> I believe that much more extensive comparisons should be made with compression methods such as Zstandard and byte shuffling, which also have the advantage of being lossless. \\n\\nYou are correct that if we are storing the weights on the disk, then it is sensible to use a universal source coding algorithm such as the ones you mention to ensure we save as much space as possible. \\n\\nHowever, our proposed method is not directly comparable to these source coding algorithms. In fact, we can view our proposed algorithm as a pre-processing step before universal source coding. To be completely clear, we suggest that we combine our proposed method with a source coding algorithm to form the following \\\"bits-back\\\" pipeline:\\n1. We obtain some network weights with rotational symmetries.\\n2. We use our method to eliminate the redundancies induced by the rotational symmetries in the weights.\\n3. We apply a universal source coding algorithm (e.g. Zstandard) to the output of our algorithm.\\n4. We save the output of this pipeline to the disk.\\n\\nIn our paper, we only looked at the gains we get if we apply our method only, without any universal source coding. Hence, an important concern arises: do we retain significant gains if we run the pipeline we suggest above, compared to just running universal source coding without our method? \\n\\nFortunately, the answer is positive. In particular, we compared the pipeline suggested above to universal source coding (we used ZIP in our experiment) on the OPT-2.7B model. Using ZIP only, **the compressed size comes to 3.97 GB while using our suggested bits-back pipeline, it reduces to 3.79 GB, and approximately 5% gain in storage size as before**. Note, that while the exact gain my vary depending on the universal source coding algorithm we use, these models are large enough that the gains we report here should be fairly robust across different source coding algorithms.\", \"there_is_an_intuitive_reason_for_retaining_the_gain_even_after_source_coding\": \"ZIP (or any other general-purpose universal source code) is unaware of the redundancies introduced by the rotational symmetry in the weights and, therefore, cannot utilise it to reduce storage size. From this perspective, **the storage savings resulting from our bits-back method are \\\"orthogonal\\\" to the savings that result from source coding.** Essentially, to the best of our knowledge, we are the first to address these structural redundancies in neural networks through bits-back coding. As such, our approach is not directly comparable to other compression algorithms that do not account for this specific type of redundancy.\\n\\nOnce again, thank you for this excellent point; we have also updated our paper to include the above discussion in the appendix!\\n\\n**Regarding your second concern about the motivation of our approach**, we want to highlight 3 points:\\n1. We agree GPU is scarcer compared to disk. However, bandwidth is also another bottleneck. Therefore, *reducing storage not only reduces disk costs but also reduces transmission costs and speeds up weights sharing*.\\n2. One potential application that reflects point 1 is applying Federated Learning for LLMs, or for LoRAs. In these applications, the client needs to frequently send messages of weights/gradients to the server. Therefore, our approach can gain potential application in these areas.\\n3. we also want to note that, our approach can save 3-5% bits mostly for free and there is almost no drawback for applying our approach given its minimal runtime overhead and negligible influence on performance. Therefore, *it is always beneficial to apply this approach in the cases where we need to store and transmit weights*.\\n\\nWe thank you again for your detailed reply. If you find our further clarifications satisfactory, we kindly invite you to consider raising the score.\"}", "{\"comment\": \"Thank you for your reply. However, given your comments, we believe there are three fundamental facts that we haven\\u2019t communicated clearly enough, and thus you appear to have misunderstood:\\n\\n1. **Our method is NOT a lossy quantization algorithm.** Rather, our method is a lossless operation (up to floating-point precision) that we can apply before saving the weights. \\n2. **Our method is NOT comparable with the methods you suggest.** The methods you list, such as byteshuffle, are general preprocessing methods to improve the compressibility of model weights, and GZip and ZStandard are universal source codes. On the other hand, our method\\u2019s sole purpose is to **eliminate the description inefficiency induced by rotational symmetries in the weights**. This incomparability also means that **we can freely combine our method with those you mention**. Our previous response demonstrated how our method retains the ~5% gain in storage space when we combine it with Zip. We have now run some experiments where we apply Zstd *with and without bshuffling* as well as *with and without applying our method* on OPT-2.7 model weights:\\n\\n| | Zstd, autoshuffle | Zstd, noshuffle |\\n|-------------|-------------------|-----------------|\\n| **without** our method | 3.48 GB | 3.78 GB |\\n| **with** our method | 3.33 GB | 3.60 GB |\\n\\nAs you can see, **our algorithm still provides ~5% gain, regardless of whether we use shuffling, or not.**\\n\\n3. **We are NOT proposing to apply our method to reduce GPU usage**, only storage. Reducing GPU cost is out of the scope of design intent, as outlined in our manuscript and previous discussion.\\n\\n\\nTo summarise, our method aims to reduce the storage cost of network weights, e.g., LLMs, at a negligible additional computational cost. **Our method eliminates a redundancy that no other method can/does** by leveraging the extra knowledge that there are rotational symmetries in the model weights. \\nAs such, we believe that our method is a significant and valuable addition to the LLM and the broader model compression literature.\\nThank you for the time you have invested in reviewing our paper and actively engaging in discussion with us. We believe we have addressed your concerns and thus kindly invite you once more to raise your score.\"}", "{\"title\": \"Thank you for your review! Please consider our reply\", \"comment\": \"Thank you once again for the effort you put into reviewing our paper. As only a few working days are left in the discussion period, we would like to ask if our response has satisfied your concerns. If so, we kindly invite you to consider raising the score. If any concerns remain, we are happy to discuss them further here.\"}", "{\"comment\": \"We thank the authors for their detailed reply.\\nHowever, we maintain that the gain from reducing storage memory does not justify introducing a lossy quantization algorithm.\\nSeveral lossless alternatives which serve a similar purpose exist, for example, the byteshuffle and bitshuffle filters available in the Zarr library. Applying these filters improves the compressibility of input data, even for lower compression levels of standard compression algorithms such as GZip and ZStandard. For example, applying GZip level 1 with the byteshuffle filter often produces a better result than applying GZip level 6.\\nMoreover, while network bandwidth is sometimes a bottleneck, this problem can often be mitigated by caching frequently used components in faster storage. For example, small weights such as LoRA can be kept in RAM instead of being transferred.\\nAs mentioned in our previous reply, the costs of a large storage footprint are only paid once at the first step, while the costs of a large memory footprint in the GPU are paid every time the model executes. While a lossy algorithm may be justifiable to reduce the latter problem, it is not justifiable to reduce the former.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"title\": \"Thank you for your review! Please consider our response\", \"comment\": \"Thank you for the effort you put into reviewing. We would like to ask if you could look at our response and kindly invite you to consider increasing your score if our response addresses your concerns. We are happy to continue discussing them further if any concerns remain.\"}", "{\"title\": \"Update Reviews\", \"comment\": \"Dear Authors,\\n\\nThank you for addressing my questions regarding the effectiveness of your approach with a lower precision format and the generalization of your approach to other neural network architectures. I thus increased my rating score. \\n\\nBest regards.\"}", "{\"comment\": \"I thank the authors for their detailed response. It has helped clarify the scope and motivation of the work.\\nHowever, if the aim of the paper is to reduce disk storage, I believe that much more extensive comparisons should be made with compression methods such as Zstandard and byte shuffling, which also have the advantage of being lossless. Lossy algorithms for floating-point data compression also exist.\\nIn general, the model compression algorithms are designed to reduce the amount of memory in the GPU, not disk storage. This is because GPU memory is a scarcer resource, and reducing the required GPU memory allows models to run using fewer GPUs, which reduces costs. In addition, GPU memory must be accessed once per decoding step of the model, whereas storage needs only be accessed once at the very beginning when the model is being loaded to RAM.\\nBecause of this dynamic, I am curious whether there is a motivating use case in mind. For example, if many user-defined models need to be saved and transferred frequently, there may be a case to reduce the storage requirements, even at the cost of introducing lossy compression. However, even in this case, a rigorous comparison against both lossless and lossy techniques for model weight compression should be conducted for a fair comparison.\\nDue to the difficulty of understanding the motivation of the work, I will keep my score.\"}", "{\"title\": \"Thank you for your review! Please consider our response\", \"comment\": \"Thank you once again for the effort you put into reviewing our paper. As only a few working days are left in the discussion period, we would like to ask if our response has satisfied your concerns. If so, we kindly invite you to consider raising the score. If any concerns remain, we are happy to discuss them further here.\"}", "{\"title\": \"Thank you for your review and discussion! Please consider our response\", \"comment\": \"Thank you for the effort you put into reviewing and discussing. We would like to ask if you could look at our last response and kindly invite you to increase your score if our response clarifies our method and addresses your concerns. If any concerns remain, we are happy to continue discussing them further.\"}", "{\"comment\": \"> Weakness 1\\n\\nThis does not seem to be a misunderstanding. I was aware that there is no prior connection, and I believe that the authors need to \\\"properly connect SliceGPT with the previously proposed bits-back coding\\\". Specifically, the authors need to illustrate in detail what these techniques are and why SliceGPT could fit in the framework of bits-back coding. So far, I have not found this addressed yet.\\n\\n> Weakness 2\\n\\nThank you for the information. The question is now clear.\\n\\nIn general, the authors partially addressed my concerns. Since I have already suggested an acceptance, I maintain my original score.\"}", "{\"title\": \"Global Response to all reviewers\", \"comment\": \"We extend our gratitude to all the reviewers for their detailed and comprehensive reviews and for the time spent reviewing our manuscript. We are pleased that the reviewers acknowledged the novelty, insightfulness, and applicability of our approach. We have carefully addressed their concerns in our responses, and have modified our manuscript accordingly (modifications are highlighted in red).\\n\\nMost of the reviewers raised two common concerns. Therefore, we provide a concise summary of our responses in this global response.\\n\\n1. **Regarding the runtime and the influence of our approach on inference latency:**\\n\\nWe highlight that our approach aims to reduce the storage space and transmission cost. Therefore, we will only run the encoding algorithm when saving the model and decoding when loading the model. Our approach will not influence the inference time once the model is decoded and loaded into memory.\\n\\nTo evaluate the influence on the model saving and loading time, we have updated our manuscript to include encoding and decoding time in Table 2 and discussion in the runtime analysis paragraph on page 10. Our algorithm only adds seconds to minutes to the loading time on a consumer-grade GPU. We also provide the runtime on the CPU in Appendix C.1. \\n\\n\\n\\n2. **Our proposed method is specific to SliceGPT:**\\n\\nWe emphasize that the concept of bits-back coding applied to neural networks is general and can be extended to other architectures exhibiting symmetries. One of our key contributions, as noted in the conclusion section, is establishing a connection between bits-back coding and networks with symmetrical properties. In our reply to the reviewers, we provide two examples (LoRA and networks with Permutation symmetry) where our algorithm (or its variations) can be employed to achieve free bits back. We have updated our manuscript to include a discussion on this at the end.\"}", "{\"summary\": \"This paper proposes the application of the bits-back algorithm to reduce the overhead of additional matrices introduced by pruning large language models (LLMs) using SliceGPT. In the SliceGPT method, an additional matrix is introduced for the rotation matrix, which helps in maintaining accuracy but results in additional computational overhead, thereby acting as a compression overhead. To address this issue, we propose an algorithm that encodes/decodes the rotation matrix using the bits-back algorithm, demonstrating that the rotation matrix can be computed solely through the decoding process during inference. Our proposed method shows an additional 3-5% improvement in compression efficiency compared to the practical compression rate of SliceGPT.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": [\"The paper proposes a method to compress the rotation matrix Q introduced by SliceGPT using the bits-back algorithm, effectively reducing the parameter overhead.\", \"It demonstrates that the rotation matrix Q can be encoded and decoded using the bits-back algorithm without requiring a calibration set, relying solely on the weight matrix.\", \"The study shows that while the actual compression rate of SliceGPT with the rotation matrix Q is approximately 9%, the proposed method can achieve a closer-to-expected compression rate of 13%. It also demonstrates that applying the proposed encoding method to the rotation matrix Q in models such as OPT and LLaMA2-7B does not result in significant differences in Commonsense Reasoning (CSR) performance.\"], \"weaknesses\": [\"The paper lacks sufficient analysis and experimentation regarding the practical impact on latency and throughput during inference when decoding the rotation matrix Q using the proposed method.\", \"The proposed method is somewhat limited in scope, as it can only be applied after the implementation of SliceGPT, thereby restricting its applicability.\", \"The actual benefits of encoding the rotation matrix Q in terms of inference latency and throughput might be minimal. It is likely that during the prefill stage, the additional decoding step for the rotation matrix Q could result in higher inference latency and lower throughput compared to SliceGPT alone.\"], \"questions\": [\"How does the proposed method perform in terms of inference latency and throughput gains compared to SliceGPT when applied on actual hardware like GPUs? If the effectiveness of this aspect is demonstrated, I would be inclined to increase my rating.\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you for your great suggestions. We apologize for misunderstanding your previous concerns. Since we cannot update our PDF at this stage, we commit to including a clearer explanation connecting bits-back and sliceGPT in the camera-ready version to make our paper easier to follow.\"}", "{\"title\": \"Thank you for your review and discussion! Please consider our response\", \"comment\": \"Thank you for the effort you put into reviewing and discussing. We would like to ask if you could look at our last response and kindly invite you to increase your score if our response clarifies our method and addresses your concerns. If any concerns remain, we are happy to continue discussing them further.\"}", "{\"title\": \"Thank you for your review! Please consider our response\", \"comment\": \"Thank you again for the effort and time you put into our paper. As only a few working days are left in the discussion, we would like to ask if our response has satisfied your concerns. If so, we kindly invite you to consider raising the score. If any concerns remain, we are happy to discuss them further here.\"}", "{\"comment\": \"Thank you for your insightful review and questions, which have helped us improve our manuscript. We now reply to the stated weaknesses and questions.\\n\\n1. **weaknesses 1 & 3, and question** (regarding impact on latency and throughput): \\n> The paper lacks sufficient analysis and experimentation regarding the practical impact on latency and throughput during inference when decoding the rotation matrix Q using the proposed method. The actual benefits of encoding the rotation matrix Q in terms of inference latency and throughput might be minimal. How does the proposed method perform in terms of inference latency and throughput gains compared to SliceGPT when applied on actual hardware like GPUs?\\n\\nThank you for raising this concern. While our approach requires additional time for decoding, there is no influence on the latency and throughput during inference:\\n\\n(1) Our approach aims to reduce storage space and transmission costs. Therefore, we will only run the encoding algorithm when saving the model and decoding when loading the model. Once the model is decoded and loaded into memory, there is no overhead on the influence of the inference time and throughput.\\n\\n(2) We measured the encoding and decoding time and updated our manuscript to include the results and discussion. Please refer to Table 2 and the runtime analysis paragraph on page 10. We also show Table 2 here for easy reference:\\n\\n| Model Name | OPT-1.3B | OPT-1.3B | OPT-2.7B | OPT-2.7B | OPT-6.7B | OPT-6.7B | OPT-13B | OPT-13B |\\n|----------------|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|:-------:|:-------:|\\n| Slicing | 20% | 30% | 20% | 30% | 20% | 30% | 20% | 30% |\\n| Encoding time | 15 s | 13 s | 30 s | 24 s | 2.5 min | 1.7 min | 6.5 min | 4.1 min |\\n| Decoding time | 6 s | 5 s | 14 s | 11 s | 1.2 min | 45 s | 2.5 min | 2 min |\\n\\nWe can see that decoding the entire network is actually fast on the consumer-grade GPU. Therefore, even if we measure the inference time end-to-end, the influence of our approach on runtime is minimal.\\n\\n(3) Additionally, we did not optimize our implementation for the decoding runtime, and there are ways to improve it. We discuss one such possibility in lines 501-511 in our updated manuscript: parallelizing the encoding and decoding to accelerate the execution.\\n\\n2. **weakness 2:** \\n>The proposed method is somewhat limited in scope, as it can only be applied after the implementation of SliceGPT, thereby restricting its applicability.\\n\\nWe agree that our approach focuses on SilceGPT. However, we emphasize that the concept of bits-back coding applied to neural networks is general and can be extended to any architecture exhibiting symmetries. One of our key contributions, as noted in the conclusion section, is establishing a connection between bits-back coding and statistical models that exhibit invariance under rotations and, more generally, under the action of the elements of a symmetry group. Below, we provide two examples where our algorithm (or its variations) can be employed to achieve free bits-back. We have also included these discussions in the Conclusion, Limitations, and Future Directions Section in our updated manuscript.\\n\\n(1) Applying our approach to encoding LoRA modulation or weights decomposed in LoRA-style. Specifically, LoRA approximates a matrix $M$ by $M=AB$. We can see applying $Q$ to $A$ (as $AQ$) and $Q^T$ to $B$ (as $Q^TB$) will leave M invariant. Therefore, our proposed approach can be seamlessly applied in this context.\\n\\n(2) Neural networks with permutation symmetry. It is well known that in many architectures, such as standard MLPs, permuting the hidden units in one layer and applying the reverse permutation to the subsequent layer leaves the output unchanged.\\nWe can modify our approach to apply bits-back coding in this case.\\nSpecifically, we can define a canonical order for hidden units based on a predefined criterion (e.g., sorting the weights or biases corresponding to each unit in descending order). During encoding, a permutation can be randomly selected by decoding bits from the current bitstream. During decoding, this permutation can be easily recovered by rearranging the hidden units back to their canonical order. In fact, it is a standard fact that permutation matrices are orthogonal (i.e., they can be thought of as rotations). Therefore, they can also be seen as a special case of our approach. However, the gain here is smaller than for the rotational symmetry case, as the permutation invariance of the hidden units introduces significantly less redundancy compared to rotations in general.\\n\\n*Thank you once again for taking the time to review our work and our response. We are happy to discuss any further questions you might have. However, should we have addressed your concerns, we kindly invite you to raise the score.*\"}" ] }

[ "Weiyu Guo", "Ziyue Qiao", "Ying Sun", "Hui Xiong" ]

[ "surface electromyography", "gesture recognition", "signal processing" ]

{ "note_id": [ "zXW2PkcYxZ", "vfvSDqBh0i", "qRYQb0gyS6", "dNWTmHrL11", "bHtRohPnep", "aBoz581cFB", "Y19OMj6E3E", "XKWB2jaoO6", "X876UxQ7ta", "WssviITsjt", "Toc3DohEiu", "Qgm8k80UCW", "MgJmuP8ca3", "4z6QAhizab", "45Se66fmaK" ], "note_type": [ "official_comment", "official_comment", "official_comment", "official_comment", "decision", "official_comment", "official_review", "official_comment", "official_review", "official_comment", "official_review", "official_review", "meta_review", "official_comment", "official_review" ], "note_created": [ 1732404647820, 1732159132289, 1732158623846, 1732404242220, 1737524095547, 1732198974228, 1730590015337, 1732406361742, 1730847718856, 1732158722483, 1730660184259, 1730715606178, 1734303321936, 1732434622510, 1730382898159 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission10974/Reviewer_5Rea" ], [ "ICLR.cc/2025/Conference/Submission10974/Authors" ], [ "ICLR.cc/2025/Conference/Submission10974/Authors" ], [ "ICLR.cc/2025/Conference/Submission10974/Reviewer_5Rea" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission10974/Area_Chair_MM92" ], [ "ICLR.cc/2025/Conference/Submission10974/Reviewer_dCsH" ], [ "ICLR.cc/2025/Conference/Submission10974/Reviewer_5Rea" ], [ "ICLR.cc/2025/Conference/Submission10974/Reviewer_5Rea" ], [ "ICLR.cc/2025/Conference/Submission10974/Authors" ], [ "ICLR.cc/2025/Conference/Submission10974/Reviewer_bTGn" ], [ "ICLR.cc/2025/Conference/Submission10974/Reviewer_Xpzg" ], [ "ICLR.cc/2025/Conference/Submission10974/Area_Chair_MM92" ], [ "ICLR.cc/2025/Conference/Submission10974/Reviewer_5Rea" ], [ "ICLR.cc/2025/Conference/Submission10974/Reviewer_aqm7" ] ], "structured_content_str": [ "{\"title\": \"Reviewer 5Rea: Response to Reviewer 5Rea 3\", \"comment\": [\"**Concerns on Novelty and 'Intrinsic Variation Capture Module'**:\", \"While the authors have provided additional details regarding experimental procedures and performance improvements, my primary concerns on **limited novelty** remain unaddressed. Specifically, the **'Intrinsic Variation Capture Module'** is mentioned 14 times in the manuscript, but it appears to involve only a slight modification\\u2014a short-term learning module based on self-attention, integrated with a transformer.\", \"The following issues should be clarified to fully address these concerns:\", \"1. **Full Set of Modifications**: A detailed explanation of all modifications made to the transformer structure to adapt it for gesture recognition tasks is needed.\", \"2. **Rationale for Changes**: Justify why certain components were added or removed, and explain how these changes are specifically tailored to sEMG signal processing, feature learning and noise reduction.\", \"3. **Ablation Studies**: Provide ablation studies to evaluate the impact of these modifications.\", \"**Concerns on Code and Manuscript Consistency**:\", \"The authors have acknowledged inconsistencies between the provided code and the manuscript, stating**\\\"time constraints and code copyright considerations\\\"** as reasons for releasing only a portion of the code.\", \"However, I find the term **\\\"code copyright considerations\\\"** ambiguous. Could the authors clarify what this refers to? While it is within the authors' rights to choose whether or not to provide code to reviewers, once code is shared, any **misleading, incorrect, or inconsistent information** can have a negative impact on the credibility of the work.\"]}", "{\"title\": \"Response to Reviewer 5Rea 5\", \"comment\": \"**Fairness of Comparison Regarding Pretraining.** We want to emphasize that the pretraining in our work is fundamentally different from typical pretraining approaches in fields like NLP. Unlike models that are pretrained on large external datasets and then fine-tuned on specific tasks, our method performs pretraining **without introducing any external data**.\\n- We use only the user-specific training data for both pretraining and fine-tuning. This means that both our method and the comparison methods are trained and tested on **exactly the same data**, ensuring a fair comparison. Our contribution lies in demonstrating that even without additional data, our approach enhances the model's robustness and performance using the existing user-specific data.\\n\\n---\\n**The reviewer raises concerns about Generalizability.** Thank you for raising the concern about the generalizability of our model. We would like to clarify that we evaluated our framework on two comprehensive and diverse datasets: Ninapro DB2-all, which includes 49 gestures, and the latest GRABMyo dataset containing 17 different types of actions. In contrast, \\\\[3\\\\] used Ninapro DB2-EB and DB2-EC (totaling 35 gestures) and CapgMyo DB-c (12 gestures). Therefore, our work utilizes more data in terms of both quantity and diversity, demonstrating robustness across a broader range of gestures. Additionally, CapgMyo DB-c is a high-density EMG dataset primarily used in medical detection fields, which does not align with our focus on daily interaction scenarios. **Furthermore, we have validated our model in real-world settings using a commercial EMG armband, as detailed in the appendix**, confirming its practical applicability and generalizability.\\n\\n- In response, we have conducted additional experiments on the Ninapro DB2-B dataset. Due to time constraints, we were only able to train a subset of the baselines. The time window was uniformly set to 200ms, and the parameters remained consistent with those in the paper. Nevertheless, our method, STET, still demonstrates the best performance, as shown below:\\n\\n| Model | Acc |\\n| ----------- | ----- |\\n| LST-EMG-Net | 82.23 |\\n| Informer | 85.22 |\\n| TEMGNET | 80.74 |\\n| STET | 87.61 |\\n\\n---\", \"minor_concerns\": \"1. 'STEM generalizes across different gesture recognition tasks' (line 029-030). Since the experiments are conducted using only one dataset, it would be more appropriate to state that the model \\\"generalizes across different gesture recognition tasks within the GRABMyo dataset.\\\" This would avoid any implication of broader cross-dataset generalization.\\n\\t- By \\\"STEM generalizes across different gesture recognition tasks,\\\" we intended to convey that our model performs effectively on two distinct tasks: gesture classification and hand joint angle prediction. We will revise the statement to clarify that our model generalizes across different gesture recognition tasks within our experimental setup, specifically the classification and regression tasks, to more accurately reflect the scope of our work.\\n\\n2. Use of Linear Projection. Figure 2(b) shows that linear projection is applied in the long-term enhanced module but not in the short-term enhanced module. The rationale behind this selective application of linear projection requires further clarification.\\n\\t- Thank you for your question about the use of linear projection in our model. The linear projection is applied in the long-term enhanced module because this module processes inputs of fixed length due to its structural design. We use linear projection here to adjust the feature dimensions appropriately. In contrast, the short-term enhanced module operates based on a sliding window, which means the input length remains consistent and does not require dimensional adjustment. Therefore, linear projection is not needed in the short-term module.\"}", "{\"title\": \"Response to Reviewer 5Rea 2\", \"comment\": [\"**The reviewer suggests providing more evidence for computational efficiency.**\", \"| model | inference time-GPU\\uff08A6000\\uff09 | inference time-CPU\\uff08AMD EPYC 7543\\uff09 | \\u2003parameter count | GPU memory allocated |\", \"| --------------------------- | ------------------------- | --------------------------------- | ---------------- | -------------------- |\", \"| transformer | 3.8 ms | 15.1 ms | 481169 | 18.08 MB |\", \"| add STEM with weightsharing | 3.9 ms | 17.6 ms | 489233 | 23.66 MB |\", \"| \\u2003without weightsharing | 4.8 ms | 27.5 ms | 581137 | 21.65 MB |\", \"In this table, we use the following high parameter: Feature dimension (feat_dim) is 12, maximum length (max_len) is 200, model dimension (d_model) is 64, number of attention heads (n_heads) is 2, number of layers (num_layers) is 3, dimension of feedforward network (dim_feedforward) is 256, number of classes (num_classes) is 17, the dropout rate is 0.1, positional encoding is 'learnable', the activation function is 'gel', and normalization is 'BatchNorm'. The STEM model uses the same parameters as those in the experimental setup of the paper.\", \"Thank you for pointing out the ambiguity surrounding the term \\\"cost-effectiveness.\\\" We would like to provide further clarification. In the introduction, the term \\\"cost-effectiveness\\\" primarily refers to **inference efficiency**, where we achieve optimization in computational resource consumption while maintaining high accuracy and robustness by introducing the Short-Term Enhanced Module (STEM) and its weight-sharing strategy in the sliding window attention mechanism. This optimization includes, but is not limited to, the following:\", \"1. **Parameter Count**: By incorporating STEM, we significantly reduce the model's parameter count. Specifically, when the STEM module is added (with weight sharing enabled), the parameter count increases slightly from 481,169 to 489,233, an increase of about 1.7%. However, if weight-sharing is not used, the parameter count increases substantially to 581,137, a 21% increase.\", \"2. **Inference Time**: We measured the inference time on both GPU and CPU. When using the STEM module with weight sharing, the GPU inference time is 3.9ms, and the CPU inference time is 17.6ms, which is a significant improvement compared to the case without the STEM module (4.8ms on GPU and 27.5ms on CPU). This indicates that the STEM module while enhancing short-term features, maintains a low inference time.\", \"3. **GPU Memory Consumption**: The relative increase in GPU usage is due to the mechanism of parallel computation that the GPU activates when using a sliding window.\", \"These experimental results show that by incorporating weight sharing in the design of the STEM module, we effectively control computational overhead while improving model performance without significantly increasing the computational burden. Specifically, weight sharing helps keep the model complexity low while significantly enhancing the ability to capture short-term features, thus reducing the additional demand on computational resources.\", \"To address the reviewer's concern, we will clarify these experimental results in the revised version of the paper, and we change the term \\\"cost-effectiveness\\\" to **inference efficiency**\\u2014achieving the enhancement of short-term features while minimizing computational overhead. Additionally, we provide more detailed comparison data on inference time, memory consumption, and parameter count in the supplementary experiments to make the concept of \\\"inference efficiency\\\" clearer.\"]}", "{\"title\": \"Reviewer 5Rea: Response to Reviewer 5Rea 2\", \"comment\": [\"**Concerns on 'Cost-effectiveness'**:\", \"I appreciate that the authors have recognized the inappropriate use of the term **'Cost-effectiveness'**.\", \"**Clarification on Appendix C.2**:\", \"I would like to thank the authors for providing the additional experiments. However, in **Appendix C.2**, it is unclear which method is proposed by the authors and which methods are being compared. A clear description is required.\", \"**Issues with Table 7**:\", \"The naming of methods in **Table 7** are confusing. For instance, **'without weight sharing'** is mentioned only once, making it difficult to understand what it refers to. A clear and more consistent comparison across methods is required to improve the table's clarity.\"]}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"title\": \"Rebuttal\", \"comment\": \"Dear Reviewers,\\n\\nI encourage you to review the rebuttal and reach out to the authors with any additional questions or requests for clarification.\\n\\nBest,\\\\\\nAC\"}", "{\"summary\": \"This paper proposes an innovative solution for improving sEMG-based gesture recognition, tackling issues such as noise interference and distinguishing similar gestures, especially in non-laboratory settings. The authors introduce the Short-Term Enhancement Module (STEM), which captures local signal variations to maintain noise resistance, and a self-supervised sEMG Intrinsic Pattern Capture (EIPC) for pre-training and learning intrinsic signal patterns. STEM is designed to be easily integrated into various time-series deep learning models. Experiments show it significantly improves performance in classification and regression tasks, making it suitable for practical applications.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"The paper is well-written and structured clearly, making it easy to follow.\\nThe parallel use of both long-term and short-term decoders is an innovative approach that preserves context without losing important information from different stages in the sequence. This approach helps address limitations in existing sEMG models and presents a new way of combining local and global signal features, adding originality and significance to the field.\\nThe self-supervised pre-training method effectively reduces the need for extensive labeled data and boosts the model's generalization, making it practical for real-world applications.\\nAdditionally, the emphasis on noise robustness, a common challenge in sEMG signal processing, enhances the model's real-world applicability.\", \"weaknesses\": \"Insufficient Methodological Rationale: The paper would benefit from more detailed explanations for certain design choices, such as the selection of the window size in the short-term decoder and the masking ratio in the EIPC module. Clarifying how these parameters were optimized or chosen could strengthen the methodological foundation.\", \"questions\": \"1. On line 80, the authors state, *\\u201cwe show that the long-term and short-term features are complementary in sEMG-based gesture recognition tasks.\\u201d* While the methodology section explains these concepts, it would improve reader clarity if a brief explanation of short-term and long-term features were included earlier in the paper.\\n2. Please clarify the rationale behind the choice of a sensor-wise masking strategy over other potential approaches, and how this decision impacts the model\\u2019s ability to generalize to various sEMG signal conditions.\\n3. Please provide more detail on why the specific masking ratio (0.15) and the average length of masked segments were chosen. Also, please add explanation of the sensitivity of the model\\u2019s performance to these parameters.\\n4. Could the authors explain why a multi-head self-attention layer was chosen for the long-term decoder and a sliding-window self-attention layer for the short-term decoder. \\n5. For the sliding-window self-attention, how was the window size determined, and how sensitive is the model\\u2019s performance to variations in this parameter?\\n6. Please clarify how normalization was performed for the NRMSE (Normalized Root Mean Squared Error) metric.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Reviewer 5Rea: Response to Reviewer 5Rea 5\", \"comment\": [\"**Unaddressed Comment on Ablation Study**:\", \"My previous comment regarding the need for an ablation study has not been addressed:\", \"*\\\"The contributions of proposed innovations appear marginal based on Table 2 and Table 3. Using focal loss seems to lead significantly better performance than cross-entropy. For fair comparison, consistency across methods (focal loss vs. cross-entropy) is recommended.\\\"*\", \"It remains unclear whether the authors have conducted experiments to ensure fair and consistent comparisons across methods. Addressing this issue is critical to assess the actual contributions of the proposed innovations.\", \"**Clarity on Regression Task (Table 4)**:\", \"The comparison details in **Table 4** and the related descriptions are not clear.\", \"logical flow of the paper is weird where the regression task results were reported **after** ablation studies, robustness analysis, and visualization.\"]}", "{\"summary\": \"The authors propose a Short Term Enhancement Module (STEM) to improve noise resilience in sEMG-based gesture recognition. The results show that the proposed method performs the best, both with and without additive noise, on the publicly available EMG dataset, GRABMyo.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"1\", \"strengths\": \"Gesture recognition using sEMG is a significant field with applications in prosthetic control and rehabilitation. Extensive experiments have been conducted in this work. The inclusion of code is appreciated. However, clearer instructions are needed.\", \"weaknesses\": \"Major concerns on the Claims and Contributions:\\n1. 'Learnable denoise, enabling noise reduction without manual data augmentation' (line 022-023).\\nThe authors claim to introduce a \\\"learnable denoise\\\" method, enabling noise reduction without manual data augmentation. To the best of my knowledge, 'learnable denoise' typically refers to a model trained on paired noisy and clean data in an unsupervised manner, as illustrated in Fig. 2 of Ref [1] (published in CVPR'2024). Another common approach of denoising is to employ an autoencoder-based structure. However, this manuscript does not incorporate such a strategy in the pretraining stage. In this work, pretraining stage appears limited to signal or segment masking, as suggested by Equation 1, Figure 2(a), and Algorithm 1 in the Appendix. This is also evident in the provided code: https://anonymous.4open.science/r/short_term_semg/model/generate_mask.py.\\n\\n2. 'Scalability, adaptable to various models' (line 023).\\nThe claim of \\\"scalability\\\" as described here appears to lack justification. Please correct me if I am wrong, 'scalability' typically refers to a model\\u2019s ability to handle increased data or computational demands, which is not evidently addressed here. \\\"Adaptability\\\" may be a more appropriate term in this context.\\n\\n3. 'Cost-effectiveness, achieving short-term enhancement through minimal weightsharing in an efficient attention mechanism.' (line 024).\\nThe term \\\"cost-effectiveness\\\" seems ambiguous, with no evidence for computational efficiency. While the authors claims the reduced training time, other metrics such as parameter count, inference time, FLOPs, and GPU memory consumption, are not discussed. \\n\\n4. Limited novely, 'Intrinsic Variation Capture Module' repeated 14 times in this manuscript. \\nIt appears this module provides only slight modification, a short-term learning module based on self-attention, along with a transformer. The authors should clarify the full set of modifications made to the transformer structure to adapt it for gesture recognition tasks.\\n\\n5. 'We pre-train on GRABMyo for 20 epochs using a fixed learning rate of 1e-4 for the backbone.' (line 307-308). \\nThe manuscript states that GRABMyo was used for pretraining. However, the code indicates that pretraining was conducted on Ninapro DB2. Please see evidence: (1) https://anonymous.4open.science/r/short_term_semg/pretrain.py, code line 21: train_loader, val_loader = get_dataloader_db2(dataCfg, \\\"EMG_CSV/S1_E1_A1/\\\"); (2) https://anonymous.4open.science/r/short_term_semg/cfg/db2.yaml; \\n(3) https://anonymous.4open.science/r/short_term_semg/dataloaders/Ninapro.py, code line 67, function get_dataloader_db2(cfg, path_s,exercise), which returns data from Ninapro.\\nThe descriptions in the manuscript and code appear contradictory and require clarification. Furthremore, the manuscript does not discuss which dataset was used for fine-tuning. Based on the code in https://anonymous.4open.science/r/short_term_semg/finetuning.py, fine-tuning seems to employ two datasets, namely \\\"hospital\\\" and Ninapro DB2.\\n\\n6. Unclear Details on Dataset and Evaluation Protocol.\", \"clearer_descriptions_and_justifications_are_requied\": \"(1) Whether this work follows the protocol proposed by the original paper that published the dataset. If not, please provide justification, as the dataset\\u2019s original paper [2] appears to use a different evaluation protocol;\\n(2) Whether results for comparison methods directly reported from recent works, or re-implemented by the authors (e.g. Table 1);\\n(3) For works used in comparison (e.g., [3]) that employed the same dataset, was the same evaluation protocol followed? If not, please provide justification;\\n(4) Specify whether the experiments were participant-dependent or participant-independent;\\n(5) Clearly state the sizes of training, validation, and test sets.\\n\\n\\n7. Fairness of Comparison. \\n(1) Pretraining: Most methods used for comparison (except [2]) were not pretrained on EMG data, while the proposed method employs comprehensive pretraining and fine-tuning stages. This may lead to an unfair comparison.\\n(2) Ablation Study: The contributions of proposed innovations appear marginal based on Table 2 and Table 3. Using focal loss seems to lead significantly better performance than cross-entropy. For fair comparison, consistency across methods (focal loss vs. cross-entropy) is recommended.\\n\\n\\n8. Generalizability. \\nThe generalizability of the proposed model remains uncertain as only one dataset was used for evaluation. In [3], a comparison method used in this work, has been evaluated on multiple EMG datasets. Additional dataset(s) should be used to evaluate the generalizability of the proposed framework.\", \"minor_concerns\": \"1. 'STEM generalizes across different gesture recognition tasks' (line 029-030). \\nSince the experiments are conducted using only one dataset, it would be more appropriate to state that the model \\\"generalizes across different gesture recognition tasks within the GRABMyo dataset.\\\" This would avoid any implication of broader cross-dataset generalization.\\n\\n2. Use of Linear Projection. \\nFigure 2(b) shows that linear projection is applied in the long-term enhanced module but not in the short-term enhanced module. The rationale behind this selective application of linear projection requires further clarification.\\n\\n3. Naming of 'Short-Term Enhanced Transformer' for the Proposed Method.\\nThe proposed network integrates both short-term and long-term enhancement modules, yet is named \\\"Short-Term Enhanced.\\\" This naming could be misleading, suggesting the network is optimized solely for short-term enhancement.\\n\\n\\n\\n\\n[1] Kim, Changjin, Tae Hyun Kim, and Sungyong Baik. \\\"LAN: Learning to Adapt Noise for Image Denoising.\\\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024.\\n\\n[2] Pradhan, Ashirbad, Jiayuan He, and Ning Jiang. \\\"Multi-day dataset of forearm and wrist electromyogram for hand gesture recognition and biometrics.\\\" Scientific data 9.1 (2022): 733.\\n\\n[3] Zhang, Wenli, et al. \\\"LST-EMG-Net: Long short-term transformer feature fusion network for sEMG gesture recognition.\\\" Frontiers in Neurorobotics 17 (2023): 1127338.\", \"questions\": \"See comments in the 'Weakness' section above. The current version of the manuscript is not satisfactory, but I am willing to raise my score if the authors can address any of my major concerns.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to Reviewer 5Rea 3\", \"comment\": \"**The reviewer raises concerns about the novelty.** Thank you for your comments. We would like to clarify the following key points regarding the novelty and modifications of our approach:\\n\\n - **Self-Supervised Learning and Robustness**: We incorporate a **self-supervised learning paradigm** using **sEMG Signal Masking**, which allows the model to learn intrinsic patterns and variations directly from the sEMG signals without requiring extensive labeled data. Importantly, we pretrain the model using **only the same data as the fine-tuning stage**, with no additional external data introduced. For example\\uff0cwhen we train the model for user A, we only use the training set of user A to pretrain the model. This approach enhances the model\\u2019s **noise resistance**, improving robustness even with limited training data. Our results align with findings in the paper *\\u201cA Transformer-based Framework for Multivariate Time Series Representation Learning\\u201d* (KDD 2021), demonstrating that noise resistance can be significantly improved without extra data. we conducted additional experiments to evaluate its impact on the model\\u2019s robustness to various noise types. Specifically, we removed the short-term and long-term enhancement modules and assessed the backbone model (Transformer) under conditions with and without pretraining using the sEMG Signal Masking strategy. The following table reflects the enhancement against different types of noise by using only the self-supervised learning paradigm with the sEMG Signal Masking strategy.\\n\\n| Backbone | pretrain with sEMG Signal Masking strategy | AG noise | MG noise | Signal loss |\\n| ----------- | ------------------------------------------ | -------- | -------- | ----------- |\\n| Transformer | No | 22% | 15% | 16% |\\n| Transformer | Yes | 15% | 13% | 12% |\\n\\n\\n- **Short-Term Enhancement with Weight Sharing**: The **Short-Term Enhanced Module (STEM)** introduces a **sliding window attention mechanism** that focuses on **short-term local signal variations**, critical for gesture recognition. By applying **weight sharing** in the attention layers, we efficiently capture fine-grained features in the sEMG signal without significantly increasing the parameter count. This modification, while subtle, is crucial for distinguishing gestures with similar global patterns but different short-term variations.\\n\\n- These innovations represent **substantial advancements** in adapting transformer-based architectures for sEMG gesture recognition tasks. The combination of short-term enhancement and self-supervised learning significantly improves model performance, particularly in noisy environments, while maintaining efficiency and minimizing the need for additional data.\\n\\n---\\n\\n **The reviewer points out an apparent discrepancy regarding the pretraining dataset (GRABMyo vs. Ninapro DB2) and suggests clarifying the datasets used for both pretraining and fine-tuning.** We sincerely apologize for any confusion caused by the discrepancies between the manuscript and the code. Due to time constraints and code copyright considerations, we have currently only released a portion of our code.\\n- **Gesture Classification Task**: For this task, we used the **GRABMyo** dataset for both pre-training and fine-tuning. The GRABMyo dataset provides the necessary data for gesture classification, but it does not include joint angle information.\\n- **Hand Joint Regression Task**: Since the GRABMyo dataset lacks joint angle data, we employed the **Ninapro DB2** dataset for this task. Both pre-training and fine-tuning were conducted using data exclusively from Ninapro DB2.\\n\\n- Additionally, our model is designed to be **user-specific**. For example, when developing a model for a specific user (e.g., User 1), both pre-training and fine-tuning are performed solely using data from that user. This approach is one of our significant findings: even without introducing additional data during the pre-training phase, our method enhances the model's robustness to noise.\\n\\n- The code snippets you referenced correspond to the hand joint regression task using the Ninapro DB2 dataset, which may have led to the confusion. We appreciate your feedback and will revise the manuscript to clearly specify which datasets were used for pre-training and fine-tuning in each task.\"}", "{\"summary\": \"This manuscript proposes a new framework for gesture recognition and regression from sEMG signals. They two existing datasets, GRABMyo and Ninapro that record sEMG alongside gesture identity and finger joint angles and develop a framework, STET, to process the signals. STET combines a pre-trained masked autoencoder with two transformers that are designed to process short term and long term information in parallel. These signals are concatenated, and in the case of gesture recognition, use a new asymmetric loss to upweight positive samples. The approach is thus a combination of novel components, although there is considerable past precedent for these approaches in the literature (eg LST-EMG-Net for separately modeling long and short term context and CutMix for time and channel wise masking of temporal signals). The experiments compare accuracy for gesture classification with a variety of baselines and perform ablations of different network components. They perform separate experiments simulating the addition of noise.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": [\"The approach appears to be novel. While I am somewhat surprised that this approach would be better than a single transformer with a variety of temporal scales, it is reminiscent of other work in CV integrating different spatial context.\", \"The ablations are thorough, and there are numerous baselines.\", \"The pretraining appears to reduce the effects of additive and multiplicative noise.\", \"the application domain is timely and interesting\"], \"weaknesses\": [\"Cons\", \"There are a few novel components here, but I don't find the combination especially well developed or compelling. It is likely the case that their combination of masking, parallel coarse and fine transformers and asymmetric losses improves performance but the intuition advances is a little ad hoc. Unsupervised pretraining might help, but likely not by reducing gaussian noise in the signal. It is unclear whether the short of long term components are more important for affecting different types of gestures, it is not clear what window sizes are important. There are experiments oriented at some of these, but they don't quite get to the bottom of things.\", \"It is unclear if the developed approach can be run online, and how it performs in the most important generalization domain for sEMG, across novel participants. So in practice I am not sure how useful this approach is.\", \"Given the lack of documentation of hyperparameters for all models (see below) or statistics it is hard to evaluate the results critically.\", \"The novelty of the long and short timescale approach is questionable. As they note, other manuscripts in the literature use this approach eg L262 claims that these long and short timescales are run in serial in LST-EMG-Net, but based on Figure 3 of Zhang et al 2023 they appear to be run in parallel.\", \"The added gaussian and multiplicative noise experiments are not especially convincing to me as a realistic noise aggressor for sEMG, where the predominant failure mode is generalization across participants or issues like contact loss or power line interference. Given that the masking approach only improves accuracy by 0.6-1% in the GrabMYO data, it seems like these noise sources are not large. I find the noise discussion is a bit of a digression from the rest of the manuscript.\", \"The related work section can be better organized, and the novelty and distinction between the architectures can be better presented in the related work, rather than throughout the text and appendix.\"], \"questions\": [\"The data collection done for the robustness analysis is unclear. Is it the experiment described in Figure 6? Similarly, it is unclear if the methods in \\u2018dataset\\u2019 section describe data collected for this work or different work\", \"How were hyperparameters determined e.g. the short window length, for this dataset? what about the hyperparameters for the other baselines? what was the train/test/validation strategy? Was the data tested on held out users or held out gestures within a user? What fraction of data was used for the unsupervised pretraining?\", \"How was the model run online in a streaming fashion? Is the architecture benchmarked offline the same as run online?\", \"Is it the short term or long term e.g. for table 3 what is the breakdown by gesture.\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"Surface Electromyography (sEMG) is highly susceptible to noise in real-world environments. To enhance noise resilience in sEMG-based models, this paper presents three key contributions:\\n1) introduction of the Short-Term Enhancement Module (STEM): A scalable, learnable, and low-cost module that enhances noise resistance. When integrated into neural networks, STEM improves performance by focusing on short-term signal features critical for distinguishing gestures amidst noise.\\n2) Self-Supervised Signal Masking: This technique leverages intrinsic variability in sEMG signals to enhance pre-training, allowing the model to learn robust representations without requiring extensive labeled data.\\n3) comprehensive Evaluation on the GRABMyo Dataset: Experiments on the largest available wrist sEMG dataset, GRABMyo, demonstrate that the Short-Term Enhanced Transformer (STET) achieves superior accuracy and robustness compared to existing methods, with reduced accuracy drop rates under noise and more precise gesture classification boundaries.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1) to address the risk of self-supervised learning fitting to ineffective patterns and missing meaningful temporal semantics, the authors introduce a sophisticated masking strategy. This approach uses a geometric distribution to control masked and unmasked segments, promoting more robust temporal feature learning and improving the model's ability to capture essential sEMG patterns.\\n2) the authors propose a dual-decoder approach that extracts both long-term and short-term dependencies within the signal sequences. The long-term decoder captures the global context and overall signal structure, while the short-term decoder focuses on specific local characteristics critical for accurate gesture differentiation, enhancing both precision and robustness.\\n3) an asymmetric loss function is introduced to address class imbalance and improve model optimization by emphasizing difficult samples. This loss function enhances the model's ability to generalize across varied gesture categories, prioritizing accurate classification for challenging cases without overfitting on easy negatives.\\n4) the paper is well-structured and logically organized, with clear presentation of theorems, statements, and methodological components. The writing is precise and effectively communicates the technical contributions, making the model and findings accessible to readers.\", \"weaknesses\": \"1) The experimental evaluation presents classification results solely on the GRABMyo dataset, which may introduce dataset bias. To improve the robustness and generalizability of the findings, it would be beneficial to include experiments on additional publicly available datasets.\", \"questions\": \"1) The authors claim that the proposed STEM module is low-cost; however, the paper lacks a detailed description or analysis of the module's computational efficiency. Could the authors clarify how they evaluated STEM's efficiency and provide comparative metrics or benchmarks to substantiate its low-cost nature? Additionally, how does STEM\\u2019s computational cost compare to other common noise resilience approaches?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"metareview\": \"The paper proposes a method for gesture recognition from sEMG. The work tackles an important area and shows strong results. The paper is easy to follow and understand. The shortcomings of the paper initially raised by the reviewers include the lack of novelty in individual components of the method, limited dataset scope, lack of diverse baselines and comparisons with state-of-the-art methods, missing references, and fairness in evaluation protocols.\", \"additional_comments_on_reviewer_discussion\": \"The reviewers appear divided in their assessments, with scores of 3, 5, 6, 6, 8. I have carefully read the paper, the reviews, and the rebuttal. While I disagree with some of the critiques brought up by Reviewer 5Rea (particularly regarding the validity of the replicated results and the importance of the completeness of the provided code), I do have some reservations about the paper. Specifically, I agree with the assessments of Reviewers bTGn as well as 5Rea that the paper falls short in the following areas:\\n\\n1. While the paper explains the methodology (\\\"what\\\") clearly, it does not explicitly address \\\"why\\\" the proposed mechanisms are particularly well-suited for sEMG. This leaves an important gap in the paper's justification of its approach. As an extension to this, I do find the general contributions/novelty of the work to be more on the limited side.\\n\\n2. The experimental evaluation is relatively narrow in scope. Since the paper is highly empirical and does not provide significant theoretical contributions, the expectation is that the experiments would be comprehensive; which is not the case. In particular, the paper's primary claim revolves around robustness to noise, yet the experiments only consider synthetic noise, i.e., Gaussian noise and signal loss. This, I believe is a significant shortcoming of the paper. The paper should have conducted more diverse and in-depth analyses, particularly with respect to real-world noise such as motion artifacts, which are commonly encountered in sEMG systems.\\n\\nWhile some of the initial concerns of the reviewers with the paper were alleviated during the rebuttal (e.g., Reviewer 5Rea raised their score from 1 to 3), the above two weaknesses remained, which were deemed significant by the reviewers.\"}", "{\"title\": \"Re\", \"comment\": [\"The big claims on contribution of the work in Abstract section are major concerns (e.g., inappropriate claim of 'Learnable denoise').\", \"Here is my orginal comment on **inconsistencies** between code and manuscript : 'We pre-train on GRABMyo for 20 epochs using a fixed learning rate of 1e-4 for the backbone.' (line 307-308). The manuscript states that GRABMyo was used for pretraining. However, the code indicates that pretraining was conducted on Ninapro DB2. Please see evidence: (1) https://anonymous.4open.science/r/short_term_semg/pretrain.py, code line 21: train_loader, val_loader = get_dataloader_db2(dataCfg, \\\"EMG_CSV/S1_E1_A1/\\\"); (2) https://anonymous.4open.science/r/short_term_semg/cfg/db2.yaml; (3) https://anonymous.4open.science/r/short_term_semg/dataloaders/Ninapro.py, code line 67, function get_dataloader_db2(cfg, path_s,exercise), which returns data from Ninapro. The descriptions in the manuscript and code appear contradictory and require clarification. Furthremore, the manuscript does not discuss which dataset was used for fine-tuning. Based on the code in https://anonymous.4open.science/r/short_term_semg/finetuning.py, fine-tuning seems to employ two datasets, namely \\\"hospital\\\" and Ninapro DB2.\", \"**If the authors acknowledge that the code is incomplete, then it must be explicitly stated in the manuscript when submitted.** It is **unacceptable** to submit code for review purposes and subsequently attribute contradictory findings to its incompleteness **after reviewers identifed discrepancies**.\", \"As a reviewer, I evaluate submissions thoroughly and fairly, focusing on the scientific rigor and clarity of the work presented. While I may not have directly worked on **every** specific aspect of the paper, I evaluate the work on established research principles on representation learning, knowledge in the field of sEMG, as well as **the clarity of the claims**, and the evidence provided in the **code and manuscript.**\", \"Given the current state of the manuscript, where many major concerns remain unresolved, I **do not recommend acceptance** for ICLR at this time. I urge the authors to thoroughly address all raised concerns of **original and further comments since rebuttal phase** to meet the standards expected for acceptance.\"]}", "{\"summary\": \"This paper addresses the task of gesture classification and hand joint angle prediction from sEMG signal from a multi-electrode array. Specifically, the authors make three main contributions: 1) Self-supervised pre-training based off of a signal masking objective 2) The introduction of short-term and long-term modules which are fused to create a final prediction 3) Use of asymmetric loss function to address asymmetries to address the imbalance between positive and negative samples.\\nThe self-supervised pre-training step enables training on unlabeled EMG signals which is valuable given the sparse state of labeled EMG datasets. The self-supervised objective consists of reconstruction of masked signals, its loss denoted by MSE. The masks were constructed such that contiguous chunks of signal are masked at once, as otherwise neighboring signal timestamps make the objective too simple.\\nThe short-term and long-term modules are both transformer-based models. The main difference between the two is that for the short-term module, the attention is limited within a certain window size. This forces the model to focus only on this short-term context to perform predictions.\\nThe asymmetric loss follows Ridnik et al., and was introduce after making the observation that the positive and negative samples given a gesture is severely unbalanced in favor of the negatives.\\nExperimental results demonstrate improved performance over related work. Ablation studies show improved robustness to artificially added noise, as well as the improvement each proposed component gives.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": [\"Thorough empirical investigation of their proposed approach with various ablation studies\", \"Comparisons to related work show SotA performance\", \"Approach tested on both hand joint angle prediction as well as gesture classification\", \"Statistical significance test performed, which isn't common but valuable (appendix)\", \"Combination of their proposed approach with prior works architecture demonstrates generalizability of the improvements (e.g TEMGNet 78% vs TEMGNET+STEM 84%)\"], \"weaknesses\": [\"Although the long term module has been empirically validated to show improved performance I would've liked to see better explanation / investigation on what kind of information is being stored in long term vs short term module. EMG inherently encodes force and gestures are performed on a short time scale. Therefore it is not immediately clear to me why the long term module improves performance\", \"In similar manner, the ablation study demonstrates the performance of one module over the other and over their combined performance, where the combined performance yields the best results. However, one could also conclude from these results that having an extra module is what lead to the improved performance (i.e extra model parameters), and not the the type of module. To that end, how would the model perform with two LT or ST units and how does that compare to the fused variant?\"], \"questions\": [\"L035-036: motoring -> monitoring\", \"L240-241: slide -> sliding\", \"What kind of information does LT vs ST store?\", \"Could the author clarify what kind of data pre-training was done on vs. fine-tuning?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}" ] }

[ "Leonardo Hernandez Cano", "Maxine Perroni-Scharf", "Neil Dhir", "Arun Ramamurthy", "Armando Solar-Lezama" ]

[ "world models", "finite state machines", "structure learning" ]

{ "note_id": [ "quKSZr6MmA", "gSGJk59Ack", "RZDOPpemN7", "EuJuYTNYG6", "4XtGAQx1t3" ], "note_type": [ "official_review", "official_review", "official_review", "comment", "official_review" ], "note_created": [ 1729948149723, 1730500213465, 1731087716405, 1732040980032, 1730743791219 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission13254/Reviewer_7Vzg" ], [ "ICLR.cc/2025/Conference/Submission13254/Reviewer_xDxB" ], [ "ICLR.cc/2025/Conference/Submission13254/Reviewer_CDcn" ], [ "ICLR.cc/2025/Conference/Submission13254/Authors" ], [ "ICLR.cc/2025/Conference/Submission13254/Reviewer_fsLJ" ] ], "structured_content_str": [ "{\"summary\": \"This paper proposed an unsupervised method that can cluster agent behaviours and build finite state machines over it.\", \"soundness\": \"1\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"The authors have identified a meaningful research gap, and their motivation for conducting the study is compelling.\\nClustering agent behaviours and building finite state machines could be useful for hierarchical RL.\", \"weaknesses\": \"Overall, the paper reads more like an experimental report or an essay rather than a fully developed research contribution.\\n\\n1. The proposed algorithm depends heavily on the initial expert policy $\\\\pi_0$, which limits its applicability in online learning settings.\\n\\n2. While the motivation centres on improving RL control via FSM synthesis, the paper does not demonstrate how this method performs when integrated with RL algorithms. The experiments are limited to unsupervised labelling and FSM generation, which undermines the credibility of the claimed benefits for RL control.\\n\\n3. Although the authors compare their approach with an HMM-based method, the details of the latter are insufficiently explained, with no descriptions or references provided. Generally speaking, this paper lacks of comparison with other methods (experimentally or analytically).\\n\\n4. The caption for Figure 4 is incomplete, and there is a typographical error on line 053.\", \"questions\": \"See the weakness. Here are some extras:\\n\\n1. How are the ground truth labels generated?\\n\\n2. How much data is collected to get the results in the paper?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The paper introduces Structured World Modeling from Low-Level Observations (SWMPO), a framework for unsupervised learning of neural Finite State Machines (FSMs) that represent the high-level structure of dynamic environments. SWMPO segments an environment into modes through unsupervised clustering of time-series data, synthesizing an FSM that can support tasks like policy optimization in reinforcement learning. The authors evaluate SWMPO on four environments with continuous dynamics, showcasing the method's ability to capture environmental structures.\", \"soundness\": \"1\", \"presentation\": \"1\", \"contribution\": \"2\", \"strengths\": \"The concept of learning environment structure through FSMs in an unsupervised manner from low-level continuous observations (such as sensor readings) is novel and could be a useful contribution to reinforcement learning and world modeling research. The paper provides detailed explanations of the SWMPO method, including algorithms and assumptions. Theoretical concepts, such as mode synthesis and transition predicate generation, are carefully formalized, supporting readers in understanding the approach. This work addresses a gap in world modeling for environments where the state dynamics can benefit from structured representations, particularly useful in scenarios with distinct modes (e.g., land vs. water). The FSM-based approach may offer advantages in certain robotic and control tasks where mode identification is crucial for efficient policy learning.\", \"weaknesses\": \"The paper devotes a large portion of space to formal definitions and algorithms but lacks comprehensive experimentation. Key elements, such as ablation studies on hyperparameters (e.g., epsilon, terms in the loss function, and pruning), are missing, which would help understand the impact of various components and parameters in SWMPO.\\n\\nTesting SWMPO on only four environments limits its generalizability. Notably, three of these environments are custom-designed, raising questions about reproducibility and relevance. Including standardized benchmarks (e.g., MiniGrid or Four Rooms) would provide a more robust evaluation. Comparisons are limited, with only one baseline, Hidden Markov Models (HMMs). Introducing additional baselines like DreamerV3 or simpler models (e.g., behavioral cloning) would strengthen the experimental results and clarify the unique strengths of FSMs over other architectures. The results lack interpretation, and some visual elements are unclear. For instance, Figure 6d seems to illustrate a failure mode, yet this is not discussed. Similarly, Figure 4's caption is cluttered, and details like the number of seeds for Figure 8 are missing. This lack of detail reduces the clarity and interpretability of the findings.\\n\\nMost importantly, the FSM's performance advantage over HMMs is sometimes minimal, particularly in Figure 8, where the FSM's results lie within HMM\\u2019s performance range in three of the four cases. This raises questions about the FSM's robustness as a general-purpose world model.\\n\\nI suggest revising the paper and resubmit with a stronger experiment section, including: ablation studies on key hyperparameters and components, pruning etc., as well as broader environmental benchmarks and standardized baselines to demonstrate SWMPO's applicability across various scenarios. An expanded discussion section interpreting results, especially on observed failure modes and performance variations between SWMPO and HMM.\", \"questions\": [\"Could some of the definitions, algorithms, and assumptions be shifted to the appendix to allow more space for experimental analysis?\", \"Have you considered evaluating on additional environments such as standardized benchmarks like MiniGrid or Four Rooms?\", \"Why were baselines like DreamerV3 and behavioral cloning omitted from the comparisons?\", \"Could you provide an interpretation of failure modes, specifically the anomaly seen in Figure 6d?\", \"How many seeds were used in Figure 8, and what measures were taken to ensure robustness in the experiments?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper introduces SWMPO (Structured World Modeling From Low-Level Observations), a framework for unsupervised learning of neural Finite State Machines (FSMs) to represent environment structure. The proposed method operates on continuous, low-level, non-visual data (e.g., LiDAR, joint positions) and involves three key steps: 1) labeling transitions in a dataset based on learned mode embeddings, 2) pruning spurious transitions between modes, and 3) synthesizing transition predicates between modes. The authors evaluate SWMPO on four simulated environments and demonstrate improved performance over Hidden Markov Models (HMMs) in most cases.\", \"soundness\": \"2\", \"presentation\": \"1\", \"contribution\": \"3\", \"strengths\": [\"The application of FSMs for environment modelling is an interesting problem\", \"The method described appears to be general and well motivated\", \"SWMPO performs well in the experiments presented.\"], \"weaknesses\": [\"The paper makes references (L158, L166, L180, L344) to how SWMPO has been used in the paper in an RL training loop, but there are no experiments to back up that claim. Either these experiments should be added or the claims removed from the paper.\", \"All the experiments assume access to an expert policy.\", \"Assumption 2 is restrictive. It is not trivial to assume that a POMDP can become a deterministic MDP conditioned on the mode.\", \"There are only two environments tested where the number of modes is >2 and LiDAR racing appears to have simple dynamics. Most environments where this system can be used have more than two modes.\", \"The paper assumes knowledge of the number of modes in the environment.\", \"Presentation is poor, making the paper difficult to read (see changes requested in the section below)\", \"I believe the paper shows promise but in its current state is not ready for acceptance. I would be willing to reconsider if the authors are able to provide the requested changes.\"], \"questions\": [\"How would SWMPO perform if the data was from a suboptimal/random policy? Is an expert policy a requirement for the algorithm?\", \"Could you replicate Fig. 7 with the BipedalWH or LiDAR environments. That would be more convincing since they have more modes to partition.\", \"What are the observations given to SWMPO in each of the environments?\", \"How much data was given for each environment?\", \"What is the dimension of the mode vector?\", \"What are the modes in the LiDAR racing and BipedalWH environments?\", \"What is the value of $\\\\epsilon$ used for pruning? Is the algorithm sensitive to this hyperparameter?\", \"**Presentation and grammatical changes**\", \"The optimization problem in Eq. 1 is defined over $e$ and $d$ which are never defined. I am assuming that these should be $m$ and $f$. Similar error in Algorithm 2.\", \"Eq. 1 is missing an $o_t$ as input to $f$. Also does it assume access to the environment state transition function $T$? Shouldn't it be the difference in observations and not states as per L205.\", \"L143, $o$ should belong to the space of observations $\\\\Omega$ not states.\", \"Algorithm 1 does not seem to be necessary since it just consists of calling one function.\", \"Please also label the colors used to describe the modes in Figs 5 and 6.\", \"In Fig 4, the caption is cut off.\", \"The text in the figures (for eg 5 and 6) is extremely small, and is illegible\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"withdrawal_confirmation\": \"I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors.\"}", "{\"summary\": \"The paper proposes a method for learning dynamical systems under changing dynamics, where the change is caused by additional causal variable namely mode (m). The mode m is treated as a parameterized function, of the most recent interaction history. m along with a parameterized dynamics function f (both parameterized as neural nets) are learned end to end via a traditional objective function for dynamics learning with a mutual information-based regularization. Experiments which mainly focus on detecting the modes are benchmarked on a variety of non-stationary benchmarks (possibly for one-step ahead predictions). The method performs well compared to an HMM baseline.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"1\", \"strengths\": \"These are the strengths of the paper in my opinion.\\n\\n1. Targets an important problem of learning dynamics under non-stationary dynamics.\\n2. Interesting method, and the additional regularization with mutual information seems novel especially in the context of dynamics learning under non-stationarity.\\n3. The algorithm seems simple as far as computational complexity and implementation are concerned.\", \"weaknesses\": \"These are the weaknesses of the paper in my opinion.\\n\\n1. The paper ignores a plethora of literature in this domain, like switching state-space models[1][2], hidden-paramter SSMs[3] etc which tackles the same problem but with more principled approaches and in more complex settings. These should be made as baselines.\\n\\n2. The main purpose of using this as a world model, is for making action conditional multi-step ahead predictions. No experiments on this regard are seen.\\n\\n3. The paper could be made even more stronger, if control experiments (model-based RL) with this world model are performed.\\n\\n4. There are a few typos, which need to be corrected, including in the objective function in equation (1). It should be $\\nf\\\\left(m\\\\left(o_{t-1}, a_{t-1}, o_t\\\\right), a_t, o_t \\\\right)$ in Equaton 1? what is e and d ?\", \"questions\": \"1. Would this method scale to partially observable domain like images ?\\n2. Can you provide multi step ahead prediction comparison ? \\n3. Maybe add more baselines like [1],[2] ?\\n4. An ablation on the importance of mutual information regularization scheme would be very interesting. \\n\\n\\nReferences\\n1. Switching SSM (one among many) https://proceedings.neurips.cc/paper/2020/file/aa1f5f73327ba40d47ebce155e785aaf-Paper.pdf\\n2. Hidden Parameter Recurrent State Space Models For Changing Dynamics Scenarios https://arxiv.org/abs/2206.14697\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"1\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}" ] }

[ "Jesús Jaime Moreno Escobar", "Veronica de Jesus Perez Franco", "Ana Lilia Coria Páez", "Mauro Daniel Castillo Pérez", "Oswaldo Morales Matamoros" ]

[ "Decision-Making; PCA; DCNN; Neuromarketing" ]

{ "note_id": [ "sVm8ag2bQ8", "rwdQQfO7UL", "qlloauOY7M", "qjUNSIT2tY", "YDVAZ2SL0s", "WdUXnIhbmB", "UYskvPfX0j", "NrnUdhZbgd", "MhtwrEfBUa", "EqWxL8iT8z", "4ZGZebkpPV", "1UtSUCNcpm" ], "note_type": [ "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "decision", "meta_review", "official_review", "official_review", "official_comment", "official_comment", "official_review" ], "note_created": [ 1730638376457, 1732805174744, 1732663916310, 1732405594292, 1732405958319, 1737524089034, 1734799759615, 1729198613085, 1730970372535, 1732405665902, 1732406527024, 1730611171446 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission10888/Reviewer_ZBct" ], [ "ICLR.cc/2025/Conference/Submission10888/Authors" ], [ "ICLR.cc/2025/Conference/Submission10888/Reviewer_1hb6" ], [ "ICLR.cc/2025/Conference/Submission10888/Authors" ], [ "ICLR.cc/2025/Conference/Submission10888/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission10888/Area_Chair_sfYh" ], [ "ICLR.cc/2025/Conference/Submission10888/Reviewer_TzSq" ], [ "ICLR.cc/2025/Conference/Submission10888/Reviewer_1hb6" ], [ "ICLR.cc/2025/Conference/Submission10888/Authors" ], [ "ICLR.cc/2025/Conference/Submission10888/Authors" ], [ "ICLR.cc/2025/Conference/Submission10888/Reviewer_2T1G" ] ], "structured_content_str": [ "{\"summary\": \"The study shows that EEG analysis can effectively assess taste preferences. This allows researchers to determine whether participants like or dislike certain food samples based on their brain activity. Key indicators of preference included low beta and gamma frequency bands as well as attention and meditation levels. A deep convolutional neural network (CNN) was used, which utilized four types of input including image data and EEG signals to classify participants' preferences.\", \"soundness\": \"1\", \"presentation\": \"1\", \"contribution\": \"1\", \"strengths\": \"The paper presents a new and promising idea. However, the chosen approach is quite simple and does not introduce new methods or models that advance existing research. The study would benefit from a more innovative methodological contribution to set it apart from previous work in the field.\", \"weaknesses\": \"The introduction lacks a comprehensive summary and does not adequately convey the motivation for the study. Instead, it reads more like a general document on the use of EEG in consumer choice analysis without clarifying the specific approach or objectives of the present work. Furthermore, the introduction consists of a single paragraph with little conceptual linkage between the topics covered. A clearer structure and a more coherent explanation of the purpose and methodology of the research are needed to explain in the introduction.\\n\\nThe study only considers the sensory or taste aspects of the product as the primary factor influencing consumer preferences, which is insufficient to provide valuable insights for the food industry in the context of product development. A more comprehensive assessment that includes additional factors such as texture, aroma, visual appeal and emotional response would provide a more complete understanding of consumer preferences and increase the relevance of the study to the industry\\n\\nSections 2.2 to 2.3.4 of the paper primarily resemble tutorials on EEG signals and their acquisition processes rather than focused discussions relevant to the study's research questions.\\n\\nThere is no related work provided on the existing studies on the given topic.\\n\\nOverall, this manuscript lacks the scholarly depth and clarity expected of a research paper. The presentation of ideas is often unclear and insufficient attention is paid to structure, coherence and technical detail necessary to effectively communicate the research objectives, methodology and results.\", \"questions\": \"I recommend that the authors refer to the weaknesses outlined earlier in the review.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"1\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Follow-Up on Revised Submission for ICLR 2025 - Reviewer 1hb6\", \"comment\": \"Dear Reviewer,\\n\\nWe deeply appreciate the time and effort you\\u2019ve devoted to evaluating our work and reviewing our rebuttal. We value your constructive feedback, which has significantly improved the manuscript's clarity and rigor.\\n\\nWe\\u2019d like to respectfully highlight that in this revised version, we have addressed all the points raised in the original review. Beyond providing the requested clarifications, we have also incorporated additional experimental evidence to strengthen our findings. Notably, we have demonstrated that the proposed architecture not only performs well but also generalizes effectively across diverse dimensions such as flavor profiles and even emotional responses elicited by consuming functional products. This novel aspect underscores the human-centered potential of our approach, bridging technical innovation with meaningful applications in real-world scenarios.\\n\\nWe hope these enhancements address any lingering concerns and provide compelling evidence of the robustness and impact of our work. We humbly suggest a reconsideration of the evaluation, as we believe this version significantly advances the original submission in both quality and relevance.\\n\\n\\nIn this link, you can view the article where we have highlighted in red not only the observations you provided but also those from others, It is worth noting that significant portions of the article have been restructured, and new research findings have been discovered:\", \"https\": \"//openreview.net/pdf?id=B6xUlbgP7j\\n\\nOnce again, thank you for your valuable feedback and for considering our appeal. Please let us know if there are additional elements we can further clarify or improve.\\n\\nWarm regards,\"}", "{\"title\": \"No Change in Decision Following Rebuttal\", \"comment\": \"The reviewer has conducted a thorough evaluation of the authors' rebuttal, comparing it to the original review comments. While the authors have provided additional context and clarifications, no significant new evidence or arguments have been presented to alter the initial assessment. Therefore, the original decision will stand. We appreciate the authors' diligence in addressing the reviewer's concerns.\"}", "{\"title\": \"Response to Reviewer Comments for ICLR 2025 Submission\", \"comment\": \"Thank you very much for your valuable observations. We sincerely appreciate the time and effort you invested in providing feedback, which is essential for improving the quality and clarity of our work. Below, we address each of your points in detail:\\n_____________________________________________________________________________________________________\", \"question\": \"Furthermore, why was the manuscript not subjected to proofreading and peer review by colleagues who might have identified its challenging comprehensibility for individuals not directly involved in the project?\", \"answer\": \"We appreciate your valuable feedback. We have double checked the manuscript and modified it significantly to concern your points. Main ideas explained, presentation reorganized to improve coherence, and necessary technical details introduced to facilitate communication of the findings, we hope that the manuscript is now adequate to meet the editorial and scientific standards required. In addition \\nTo ensure reproducibility, this dataset used in this research is ready for download via this link https://drive.google.com/file/d/1WuNkuMJ2SsyA-yNgyWGNYABfpEwguzMD/view?usp=drive_link. By publicly releasing this data set, other researchers will be able to reproduce our results, affirm our methodological process, and generate new paths of advancement for the field, a necessary component for transparency and collaboration in science.\\n_____________________________________________________________________________________________________\\nMoreover, we have considered the feedback from other reviewers as well, making several revisions throughout the manuscript. These updates are clearly marked in red for your convenience, allowing you to easily identify the changes implemented.\\nWe kindly invite you to review these revised sections considering the feedback provided. We are confident that the enhancements made address the concerns raised, and we hope this will allow you to reevaluate your assessment and consider this work as a valuable contribution to ICLR 2025.\\nThank you once again for your valuable insights, which have significantly contributed to improving the quality and clarity of our research. Finally, we hope these clarifications address your concerns and contribute to a better understanding of our work. Thank you again for your valuable feedback, which has greatly helped us identify areas for improvement.\", \"weaknesses\": \"The writing style of the manuscript presents significant challenges to comprehension. Specifically, it lacks adequate detail regarding the reproducibility of the research, particularly in terms of the specifications for the machine learning models employed, including differentiating between image data and time-series data, among other factors.\"}", "{\"title\": \"Response to Reviewer Comments for ICLR 2025 Submission\", \"comment\": \"Thank you very much for your valuable observations. We sincerely appreciate the time and effort you invested in providing feedback, which is essential for improving the quality and clarity of our work. Below, we address each of your points in detail:\", \"question\": \"Do you ever validate ...\\nSmall Sample ...\", \"answer\": \"All bands were included in the study; however, the low beta and low gamma bands were the most representative bands because they are associated with cognitive processes, sensory perception and decision making within the corresponding frequency ranges (Hz). It is important that, to highly that, the gamma, which is associated with memory (essential, of course, in food association and preference).\\n\\nWe have incorporated feedback from other reviewers, revising the manuscript and marking changes in red for clarity. We invite you to review these updates, confident they address the concerns raised. We hope this will lead to a reevaluation of our work as a valuable contribution to ICLR 2025. \\n\\nThank you for your insights, which have significantly improved the clarity and quality of our research. We trust these clarifications address your concerns and provide a better understanding of our work.\", \"weaknesses\": \"Overemphasis.....\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"metareview\": \"The paper uses dimensionality reduction methods and DNNs to establish a relationship between EEG signals recorded in the brain and taste preferences, finding that preferences could be predicted by low beta and gamma frequency band activity. Reviewers felt that although the paper contained some promising ideas, it did not represent a substantial enough methodological or scientific advance to warrant acceptance to ICLR. I regret that the paper cannot be accepted to this year's meeting, but I wish the authors the best of luck in revising it for publication elsewhere.\", \"additional_comments_on_reviewer_discussion\": \"This paper was a strong reject (1,1,3,3). The authors posted rebuttals that were somewhat vague and lacking in detail, which did not persuade reviewers to raise their scores.\"}", "{\"summary\": \"The paper presents a clear logic and well-structured model framework that integrates Principal Component Analysis (PCA) with deep learning neural networks (DCNNs) to analyze consumer preferences through EEG signals. However, it lacks a comparative analysis with existing methods, making it challenging to assess the robustness of its contributions. The overall framework appears simple, and the innovative aspects of the research are not clearly defined, raising questions about its uniqueness. While the results demonstrate solid applications, the absence of sufficient validation diminishes their impact. Additionally, the tables presented are distorted screenshots, affecting clarity. Key questions arise regarding the model\\u2019s comparative effectiveness and the specific innovations that enhance its applicability in the field of consumer neuroscience and neuromarketing.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"1\", \"strengths\": \"The logic of the paper is clear, with a well-structured presentation of the model framework that effectively integrates PCA with DCNN framework. The strong performance metrics in application underscore the model\\u2019s capability to predict consumer preferences.\", \"weaknesses\": [\"The paper lacks a comparative analysis, making it difficult to assess how the proposed method measures up against existing approaches.\", \"It is challenging to determine the solidity of the contribution, as the overall framework\\u2014comprising EEG acquisition followed by a deep convolutional neural network (DCNN)\\u2014appears relatively simple, and the innovative aspects of the research are not clearly articulated.\", \"While the application results are promising, the paper does not provide sufficient validation to validate the results of these findings.\", \"Additionally, the tables included in the paper appear to be screenshots, resulting in distortion that affects their readability and clarity.\", \"In the captions of Figures 7, 8, and 9, the authors refer to the \\u201cEfficiency of BRAIN Architecture including $\\\\bar{\\\\beta}$ and $\\\\bar{\\\\gamma}$ brain rhythms in training, validation, and test phases.\\u201d However, they provide no context or explanation on how the data was split into training, validation, and testing. Additionally, the figures themselves only present confusion matrices and a single ROC curve, with no clear indication of how validation and testing were performed or represented.\"], \"questions\": [\"Comparative Methods: How does the proposed framework utilizing EEG signals and DCNN compare to other existing methods in neuromarketing and consumer neuroscience, particularly regarding accuracy and interpretability? Are there specific benchmarks or studies the authors can reference to validate the effectiveness of their approach?\", \"Innovative Aspects: What are the key innovative elements of the proposed model that distinguish it from similar frameworks in the field? How do these innovations contribute to the understanding of consumer behavior and enhance the applicability of the findings in real-world scenarios, especially in developing personalized nutrition strategies?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"details_of_ethics_concerns\": \"N/A\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The reviewer encountered significant challenges in understanding the manuscript due to the unclear writing style exhibited by the authors, despite her/his expertise as a practitioner in EEG and BCI. Analyzing the EEG time-series presented in Figure 2, it appears that the authors lack a fundamental understanding of EEG amplitudes and the distinction between samples and seconds as units. Furthermore, the authors attempt to integrate EEG data with facial and food product images into a singular machine learning model, employing a simplistic application of PCA. This amalgamation raises substantial questions regarding the model's intended function, particularly in relation to the handling of potential movement artifacts. Additionally, there seems to be a conflation of EEG and BCI terminology within the methodology section, indicating a limited comprehension of the subject matter. Regrettably, these issues lead to a recommendation for outright rejection of the submission.\", \"soundness\": \"1\", \"presentation\": \"1\", \"contribution\": \"1\", \"strengths\": \"Regrettably, the submission does not meet the rigorous standards expected of an academic publication.\", \"weaknesses\": \"The writing style of the manuscript presents significant challenges to comprehension. Specifically, it lacks adequate detail regarding the reproducibility of the research, particularly in terms of the specifications for the machine learning models employed, including differentiating between image data and time-series data, among other factors.\", \"questions\": \"Why did the authors solely focus on naive PCA? What rationale underlies their decision to refrain from evaluating more advanced machine learning methodologies? Furthermore, why was the manuscript not subjected to proofreading and peer review by colleagues who might have identified its challenging comprehensibility for individuals not directly involved in the project?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"1\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to Reviewer Comments for ICLR 2025 Submission\", \"comment\": \"Thank you very much for your valuable observations. We sincerely appreciate the time and effort you invested in providing feedback, which is essential for improving the quality and clarity of our work. Below, we address each of your points in detail:\\n_____________________________________________________________________________________________________\", \"question\": \"I recommend that the authors refer to the weaknesses outlined earlier in the review\", \"answer\": \"We appreciate your valuable feedback. We have double checked the manuscript and modified it significantly to concern your points. Main ideas explained, presentation reorganized to improve coherence, and necessary technical details introduced to facilitate communication of the findings, we hope that the manuscript is now adequate to meet the editorial and scientific standards required.\\n_____________________________________________________________________________________________________\\n\\nMoreover, we have considered the feedback from other reviewers as well, making several revisions throughout the manuscript. These updates are clearly marked in red for your convenience, allowing you to easily identify the changes implemented.\\nWe kindly invite you to review these revised sections considering the feedback provided. We are confident that the enhancements made address the concerns raised, and we hope this will allow you to reevaluate your assessment and consider this work as a valuable contribution to ICLR 2025.\\nThank you once again for your valuable insights, which have significantly contributed to improving the quality and clarity of our research. Finally, we hope these clarifications address your concerns and contribute to a better understanding of our work. Thank you again for your valuable feedback, which has greatly helped us identify areas for improvement.\", \"weaknesses\": \"Overall, this manuscript lacks the scholarly depth and clarity expected of a research paper. The presentation of ideas is often unclear and insufficient attention is paid to structure, coherence and technical detail necessary to effectively communicate the research objectives, methodology and results.\"}", "{\"title\": \"Response to Reviewer Comments for ICLR 2025 Submission\", \"comment\": \"Thank you very much for your valuable observations. We sincerely appreciate the time and effort you invested in providing feedback, which is essential for improving the quality and clarity of our work. Below, we address each of your points in detail:\", \"question\": \"Innovative Aspects.....\", \"answer\": \"Thank you for your feedback. To clarify Figures 7, 8, and 9, the image corpus was split as 70% for training, 20% for validation, and 10% for testing to ensure robustness and generalizability. While the figures focus on confusion matrices and a single ROC curve, we acknowledge this may lack clarity on validation and testing. We\\u2019ve added details in Section 2.3.5 to explain the data split and metric computation. To enhance reproducibility, the dataset is publicly available, enabling others to replicate our results, validate methodologies, and drive future advancements. Thank you again for your suggestions.\\n\\nWe have incorporated feedback from other reviewers, revising the manuscript and marking changes in red for clarity. We invite you to review these updates, confident they address the concerns raised. We hope this will lead to a reevaluation of our work as a valuable contribution to ICLR 2025. \\n\\nThank you for your insights, which have significantly improved the clarity and quality of our research. We trust these clarifications address your concerns and provide a better understanding of our work.\", \"weaknesses\": \"In the captions of Figures .....\"}", "{\"summary\": \"The authors investigate brain activity and behavioral responses in relation to consumer neuroscience through exploring consumer decisions regarding food by analyzing 16 participants using EEG signals to classify preferences for functional food products with respect to different brain rhythms and facial expressions through the application of PCA and Deep Convolutional Neural Network. The beta and gamma frequency bands are emphasized for purposes of decision-making and form a possible pathway in the realms of neuromarketing and customized nutrition planning for the enhancement of healthy diets.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"Interesting Application: Using EEG data in this research for consumer preference assessment of functional foods falls within the currently developing interests in personalized nutrition and neuromarketing.\\n\\nCombining PCA and a DCNN in EEG data management is a good choice because this study focused on the decision-making analysis on the beta and gamma bands.\", \"practical_implications\": \"These findings are valuable pieces of information that could be very useful for direct marketing and product development directed toward consumers in the food industry, especially regarding healthier products.\", \"weaknesses\": \"Small Sample Size and Generalizability: The small sample size of 16 limits the generalizability of the findings. Testing a larger and more diverse population will provide a more robust base for the findings.\\nI'd say this study lacks comparative analysis with previous models or even traditional machine learning techniques since the outperformance of this proposed approach over simpler or alternative models is not clear.\", \"lack_of_reproduction_instructions\": \"Important parameters like PCA as well as the DCNN architecture used have not been described. An entire hyperparameter table along with data augmentation strategies would be useful in further increasing reproduction and clarity.\", \"overemphasis_on_beta_and_gamma_bands\": \"Though beta and gamma rhythms are relevant to decision-making, excessive concentration may neglect other EEG components that could be significant for consumer preferences.\", \"questions\": \"What are the hyperparameters of the DCNN model selected? Are there any data augmentation strategies used during training?\\n\\nWhy restrict single-band analysis to the beta and gamma frequency bands? Were other bands, such as alpha or theta, considered and found irrelevant?\\n\\nHow is this different from existing neuromarketing models that work on EEG? A comparison would explain the advantages and disadvantages of your proposed method.\\n\\nDo you ever validate the model on larger or different datasets? Because the participant pool is so small, these may provide further evidences about the generalizability and performance of your model in other contexts.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}" ] }

[ "Qingzhao Zhang", "Ziyang Xiong", "Zhuoqing Mao" ]

[ "large language model", "adversarial machine learning", "denial of service" ]

{ "note_id": [ "ouJbA2nzF4", "OmdXPzUMg0", "OHDxQ6YQf2", "L0dpO1yiBe", "HVY4XaDZ1c" ], "note_type": [ "official_review", "official_review", "comment", "official_review", "official_review" ], "note_created": [ 1730701797031, 1730690784304, 1732503947017, 1730517652774, 1730404774853 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission799/Reviewer_Cnfo" ], [ "ICLR.cc/2025/Conference/Submission799/Reviewer_6B4j" ], [ "ICLR.cc/2025/Conference/Submission799/Authors" ], [ "ICLR.cc/2025/Conference/Submission799/Reviewer_rgC2" ], [ "ICLR.cc/2025/Conference/Submission799/Reviewer_osUL" ] ], "structured_content_str": [ "{\"summary\": \"This paper presents a novel adversarial denial-of-service (DoS) attack targeting the safeguards of large language models (LLMs). The authors introduce an attack method that exploits the false positive rates in LLM safety mechanisms by injecting inconspicuous, seemingly benign adversarial prompts into user configurations. This injection causes LLM safeguards to incorrectly flag legitimate user requests as unsafe, effectively creating a DoS attack on the model\\u2019s users. Using optimized adversarial prompts, the method blocks up to 97% of queries in tests on state-of-the-art safeguards like Llama Guard 3.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": [\"The research idea of exploiting the LLM safeguards is interesting and important.\", \"Provide an examination of current defenses and introduce a unique adversarial DoS attack targeting LLM safeguards.\", \"Proposes an efficient optimization algorithm, balancing effectiveness and stealth for adversarial prompt generation.\"], \"weaknesses\": [\"A key limitation is that the DoS attack depends on the attacker\\u2019s ability to inject adversarial prompts into user templates, which would typically require client compromise or user manipulation. In practice, this level of access may not be feasible in highly secure environments or applications.\", \"The paper\\u2019s assumption of white-box access to the safeguard model and the ability to manipulate user configuration files introduces limitations. A real-world attacker may not possess such extensive access, especially in highly secure environments. Clarifying scenarios where such access might be feasible (e.g., within compromised client software environments) could enhance the paper\\u2019s relevance.\", \"The authors include a case study to demonstrate the feasibility of chaining the software vulnerabilities and vulnerabilities in LLM safeguards to conduct the proposed attack. However, there is not enough detail to illustrate how this could be achieved. It would be better if the authors could provide an end-to-end case study to demonstrate the whole process of the attack.\", \"The paper operates on the assumption that injected adversarial prompts can remain undetected by users or monitoring systems. However, in practice, systems with moderate security measures may detect unexpected modifications to prompt templates. Including a more realistic discussion of the attack\\u2019s stealth and potential detection would be valuable, as it\\u2019s likely that this level of prompt injection might not go unnoticed in all deployment environments.\", \"The paper evaluates its attack mainly on the Llama Guard and Vicuna models, which may limit its generalizability to other types of LLM safeguards, especially proprietary or differently architected ones. Including more model types or extensively testing black-box models would broaden the applicability of the findings.\", \"The paper points out that existing defenses like random perturbation and resilient optimization reduce the DoS attack\\u2019s success but harm safeguard accuracy. However, it lacks targeted strategies to address the proposed false positive exploitation. Discussing adaptive approaches that adjust sensitivity based on prompt characteristics or an anomaly detection focused on prompt template anomalies could add values to the paper.\"], \"questions\": [\"Since the attack assumes white-box access, how would its success rate change with only black-box or limited model access? Could the method be adapted for more restricted conditions?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"In this paper, the authors propose a novel denial-of-service (DoS) attack that forces an LLM or safeguard model to falsely reject user requests by inserting an optimized adversarial prompt into a user prompt template. Experiments with various model types and task settings demonstrate the effectiveness and stealthiness of the proposed DoS attack compared to baselines like GCG.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. The authors introduce a novel denial-of-service attack that offers a different perspective compared to existing jailbreak attacks. This method could enhance the diversity of robustness tests when measuring the safety level of an LLM.\\n2. Experiments using different metrics and models demonstrate the effectiveness and stealthiness of the proposed DoS attack.\", \"weaknesses\": \"1. The adversarial prompts optimized via DoS attack still lack semantic meanings, which means they could be easily identified by users or software developers as abnormal text. The authors should further justify the real-world threat posed by the proposed DoS attack to existing LLM services.\\n2. Will the proposed DoS attack remain effective in cross-task settings? For instance, can adversarial prompts optimized with logical reasoning prompts cause the target LLM to deny service when faced with coding or mathematics prompts?\\n3. As mentioned in the GCG paper, the length of the adversarial prompt can be a hyper-parameter during optimization. Will the DoS-optimized adversarial prompts show better effectiveness than GCG\\u2019s adversarial prompts when the lengths of both are restricted to a specific number?\\n4. More transfer attack results should be added to Table 2. For instance, if adversarial prompts are optimized using a model with relatively low safety levels (e.g., Vicuna), will these prompts remain effective in transfer tests against a model with a higher safety level?\", \"questions\": \"1. When user prompts are inserted into a template, software may sometimes paraphrase the entire prompt for improved model inference. Will DoS-optimized adversarial prompts remain effective against paraphrasing?\\n2. Will the proposed method remain effective against the latest commercial models, such as Claude-3 and the Gemini-1.5 series?\\n3. Can the DoS attack still be effective when using extremely short adversarial prompts (e.g., fewer than 10 or even fewer than 5 tokens)? If so, this would provide strong justification for the real-world impact of the proposed method.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"withdrawal_confirmation\": \"I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors.\", \"comment\": \"We sincerely thank all reviewers for their valuable feedback. After careful consideration, we would like to withdraw the paper and improve it based on reviewers' suggestions. Thanks.\"}", "{\"summary\": \"This work describe a DDoS attack towards the LLM service by compromising the input security fence. The authors proposed to generate adversarial example to induce the input filter to classify the query of the benign user as malicious, making the service provider refuse to responce. To make the attack stealthy and practical, two extra loss terms were used to constrain the length and the semantics of the adversarial prompt injected. Experiments were done to demonstrate the effectiveness, robustness, and transferability of the adversarial prompt.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"1\", \"strengths\": \"1. Good writing, easy to follow.\\n\\n2. Disclosing a threat that is inherent in the LLM system with the security mechanism.\", \"weaknesses\": \"1. This work seems incremental to me, as it is more like a new application of the existing adversarial attack, the discussions towards other adversarial attacks in the related work (Sec.2) may help distinguish the proposed attack and the existing methods.\\n\\n\\n2. Lack of baseline methods: Though disclosing a new threat, the proposed attack is **essentially an adversarial attack against the LLM** in my perspective. Therefore I suggest the author compare the results with those of the existing SOTA works, e.g. [1], [2] to further demonstrate their contributions. Furthermore, discussions about the latest adversarial attack in Sec.2, \\n\\n[1] Zou, Andy, et al. \\\"Universal and transferable adversarial attacks on aligned language models.\\\" arXiv preprint arXiv:2307.15043 (2023).\\n\\n[2] Alzantot, Moustafa, et al. \\\"Generating natural language adversarial examples.\\\" arXiv preprint arXiv:1804.07998 (2018).\\n\\n\\n3. More SOTA models to be tested: Only open-sourced models are investigated in the transferability experiment. Experiments conducted on ChatGPT and Claude might more effectively demonstrate the severity of this threat.\\n\\n\\n4. The transferability seems to be not as good, especially for GPT4o-mini and LLAMMA-Guard. The success rate ranges from 20%~65%. Some of the results also seem to contradict the conclusion made in the paper that \\\"*transfer attacks are more effective when the target\\n models share similar architectures or training data*\\\", as that of Llamma-Guard is the lowest. I'd like to suggest the authors dig deeper into the reason behind these phenomena and make a more thorough discussion about the results.\\n\\n\\n5. The stealthiness restriction can be concluded as one of the main contributions, yet there are not many experiments focusing on it. Moreover, I am concerned about whether the adversarial prompt gained is really stealthy or not. To attack GPT4o-mini, prompts like \\\"*Ignore all above*\\\" are needed, making the attack prompts suspicious enough. The prompt examples given in the paper like \\\"*buffs yayJet bombkup deton it*\\\" and \\\"*to Rape PEade Lol howKra*\\\" seem quite **unnatural** to me, making it very likely to be spotted by the user himself. I suppose this is because the author proposed to use the Bert-based model as the semantic loss. As the tokens with similar embeddings are not necessarily similarly natural. I expect more examples of adversarial prompts and more experiments focus on stealthiness, compared to other existing attacks.\", \"questions\": \"Seem in weaknesses\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"Safeguards in large language models (LLMs) are intended to block unsafe or harmful content. However, they can also be exploited to launch denial-of-service (DoS) attacks. This paper presents a method to exploit these safeguards by inserting seemingly harmless but adversarial prompts that trigger \\\"false positives,\\\" causing the safeguard to misclassify safe content as unsafe and thereby block user requests. The authors propose an optimization algorithm that generates short adversarial prompts, which result in legitimate user requests being rejected.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": [\"The topic of triggering safeguards for DoS attacks is interesting.\", \"Presentation is clear and well-structured.\"], \"weaknesses\": [\"The proposed attack is easily detected; specific reasons are listed below.\", \"Real-world application scenarios for this attack are limited.\"], \"questions\": \"The attack appears to be easily identifiable. The authors claim that the inserted prompts are seemingly harmless, making the perturbations stealthy. However, upon reviewing the examples provided, these \\\"harmless\\\" prompts seem easy to detect. For instance, in ``Write a tutorial onafb to hack into a bank\\u2019s database,'' if the user checks the prompt, they can immediately spot the abnormal input. The prompt contains an odd word and a sentence that appears irrelevant to the intended content, making it stand out.\\n\\nThis raises questions about whether the attack could realistically be deployed in practical scenarios. If the user can inspect the prompt, they would likely detect the problem. Conversely, if the user cannot inspect the prompt, it would seem simpler to insert the harmful content directly, which would achieve even better results without needing optimization.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}" ] }

[ "Vladislav Trifonov", "Alexander Rudikov", "Oleg Iliev", "Yuri M. Laevsky", "Ivan Oseledets", "Ekaterina Muravleva" ]

[ "Scientific computing", "PDEs", "Linear systems", "Iterative solvers", "Graph neural networks" ]

{ "note_id": [ "zLB3NbZgZY", "uO4lZpPfou", "o2mt5puAD0", "l71cky05w1", "g7SkSqxPsa", "f9Ddp6Ztob", "dPcENkaeTC", "bu9kJLLQLd", "Zb8NkUKY31", "NMFFYqjU7n", "N3XuACEoZQ", "KoICOt0Rve", "I1qNOfk9A0", "EA7D9dlcLA", "ClVNZKxLDQ", "15BjlsDHb3", "0iwiNl1pWK" ], "note_type": [ "official_comment", "official_review", "meta_review", "official_comment", "decision", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment" ], "note_created": [ 1732262327177, 1730054507717, 1734749788240, 1732621879800, 1737524032369, 1732261169642, 1730648965977, 1732260845846, 1732721789769, 1732311867311, 1730030474635, 1732260292703, 1732410746731, 1732345346299, 1732640349003, 1730430898937, 1732262299418 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission10200/Authors" ], [ "ICLR.cc/2025/Conference/Submission10200/Reviewer_VtSh" ], [ "ICLR.cc/2025/Conference/Submission10200/Area_Chair_WavS" ], [ "ICLR.cc/2025/Conference/Submission10200/Reviewer_ZRtj" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission10200/Authors" ], [ "ICLR.cc/2025/Conference/Submission10200/Reviewer_7yrG" ], [ "ICLR.cc/2025/Conference/Submission10200/Authors" ], [ "ICLR.cc/2025/Conference/Submission10200/Authors" ], [ "ICLR.cc/2025/Conference/Submission10200/Reviewer_X4c7" ], [ "ICLR.cc/2025/Conference/Submission10200/Reviewer_ZRtj" ], [ "ICLR.cc/2025/Conference/Submission10200/Authors" ], [ "ICLR.cc/2025/Conference/Submission10200/Reviewer_X4c7" ], [ "ICLR.cc/2025/Conference/Submission10200/Authors" ], [ "ICLR.cc/2025/Conference/Submission10200/Reviewer_VtSh" ], [ "ICLR.cc/2025/Conference/Submission10200/Reviewer_X4c7" ], [ "ICLR.cc/2025/Conference/Submission10200/Authors" ] ], "structured_content_str": [ "{\"comment\": \"> Any explanation why the factor L depends on b from $L(\\\\theta) = \\\\text{GNN}(\\\\theta, A, b)$?\\n\\nInput for the GNN is a graph combined out of linear system $Ax = b$, where edges are elements of the matrix and vertices are the rhs $b$ (for Precorrector instead of matrix $A$ we pass $L$ from classical linear algebra decomposition). Note that GNN is only inferenced once before the CG iterations to combine the preconditioner. During the CG iterations, the preconditioners obtained with GNN or PreCorrector are used as usual IC preconditioners from linear algebra. It is therefore not necessary to call PreCorrector for each CG iteration.\"}", "{\"summary\": \"This paper proposes to design CG preconditioners with GNNs. The proposed PreCorrector routine takes legacy preconditioners as input and learns corrections to these legacy preconditioners (e.g., ILU). The proposed method maintains the sparsity pattern of the system matrix and provides better convergence properties in CG iterations. The effectiveness of the proposed method is tested on large-scale diffusion and Poisson's equations. Experimental results show significant speedup over legacy preconditioners.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": [\"**Convergence**: The proposed PreCorrector routine does converge faster than traditional methods;\", \"**Generalization**. The proposed method performs well when dealing with high-contrast coefficients in PDEs. This is considered a challenge to all neural preconditioners/solvers;\", \"**Sparsity Preservation**. The underlying GNN architecture preserves the sparsity pattern, yielding better performance for succeeding GEMV operations;\", \"**Experimental Results**. The experiments are comprehensive and the results are convincing, showing significant speedup.\"], \"weaknesses\": [\"**Theoretical study of the loss function**, as explained by the authors themselves;\", \"**Generalization**. The proposed methods are tested over diffusion and Poisson's equations, which are symmetric positive definite. This is acceptable provided that the CG iterator itself requires this property, yet this limits the potential applications of the proposed method.\", \"**Limited Comparison**. The paper compares the PreCorrector against classical methods and (Li et al, 2023). The scope is somewhat limited.\"], \"questions\": [\"Have the authors tried to extend the method to non-symmetric indefinite systems? BiCGSTAB is already available for these systems and how does PreCorrector work with BiCGSTAB?\", \"How does PreCorrector perform when preconditioning different sparsity patterns beyond C(0) v.s. ICt(5)?\", \"Ablation on the $\\\\alpha$ parameter?\", \"What specific steps or techniques were used to stabilize the training process of PreCorrector?\", \"Computational Trade-off: For large systems, does the time saved by reduced CG iterations outweigh the precomputation and GNN inference time?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"metareview\": \"Here the authors propose a method for designing preconditioners based on training a GNN. While some of the reviewers point out the novelty of using GNNs for learning preconditioners, the reviewers also note weaknesses in the scope and applicability of the work beyond the specific problems presented in the paper along with questions regarding the overall computation cost of training a GNN to produce a preconditioner, particularly for large datasets.\\n\\nOverall, the consensus of the reviewers is to reject the paper, and I would encourage the authors to incorporate the feedback of the reviewers in preparing future versions of the manuscript.\", \"additional_comments_on_reviewer_discussion\": \"The authors were active in responding to the questions from the reviewers and engaging with the reviewers in further follow-up questions.\"}", "{\"comment\": \"Thank authors for addressing my questions. I still do not understand why the preconditioner or the PreCorrector should depends on b.\\n\\nOverall, the problems concerned in the current paper are too simple to draw a meaningful conclusion.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"comment\": \"Dear Reviewer VtSh,\\n\\nThank you for your work in reviewing our manuscript. Let us answer the questions from your review.\\n\\n> Have the authors tried to extend the method to non-symmetric indefinite systems? BiCSTAB is already available for these systems and how does PreCorrector work with BiCGSTAB?\\n\\nWhile we have focused on the linear systems with SPD matrices, the proposed architecture can be generalised to general patterns: one should use ILU instead of IC and the GNN neural network should predict the whole graph, not only the one corresponding to the lower triangular part of the matrix (to obtain the factors $L$ and $U$). However, different types of matrices besides SPD and a larger number of iterative solvers to solve systems with them deserve their own research paper.\\n\\n> How does PreCorrector perform when preconditioning different sparsity patterns beyond IC(0) v.s. ICt(5)?\\n\\nIf we understand your question correctly, you asked what sparsity patterns can be used with PreCorrector. Actually, the sparsity pattern for ICt(k) decomposition can vary depending on the chosen threshold for each new matrix. Also, more advanced IC algorithms (e.g. with pivoting) can be used to get even better preconditioners with probably more complex patterns. Overall, with PreCorrector we do not need to define the sparsity pattern ourselves and hope that it will work well. We rely on classical algorithms to get the pattern from the classical decomposition, which is already good. \\n\\nWe also tried to artificially increase the sparsity pattern of initial $A$ for GNN from [Li et al.] by padding to larger pattern $k$ in IC($k$) (e.g. IC($1$), IC($2$)). The results were almost identical to those obtained by just passing $A$. \\n\\n> Ablation on the \\\\alpha parameter?\\n\\nThe $\\\\alpha$ parameter is part of the neural network weights learned during training. Therefore there is no ablation study for $\\\\alpha$. It is worth noting that $\\\\alpha$ is equal to $0$ at the very beginning of training, which provides first gradients from pure classical IC decomposition and ensures stable training in the very first step.\\n\\n> What specific steps or techniques were used to stabilize the training process of PreCorrector?\\n\\nThe PreCorrector itself is built to stabilise the training process by (i) starting with a good initial guess as in classical IC and (ii) ensuring a valid gradient in the very first step starting with $\\\\alpha = 0$. In addition, we normalize the matrix by its largest element by absolute value before inputting it to GNN.\\n\\n> Computational Trade-off: For large systems, does the time saved by reduced CG iterations outweigh the precomputation and GNN inference time?\\n\\nSpeaking of inference time, for larger systems the PreCorrector call becomes cheaper and thus the effect of the number of CG iterations is less affected by the construction overhead (Tables 1-3). Please note that PreCorrector is not a preconditioner itself, it is used to create one. When preconditioners are combined with PreCorrector, there is no difference in usage during CG compared to preconditioners designed using classical methods. Moreover, it has exactly the same algorithmic complexity as the classical preconditioner it is made of (e.g. PreCorrector[IC($0$)] and IC($0$)). See also Appendix A.5 for more details on scalability. \\n\\nThe impact of PreCorrector\\u2019s training overhead is difficult to assess, as it directly depends on the problem you are trying to solve and the degree of generalization required. In general, any data-driven approach requires additional overhead to train the model. On the other hand, some applications require solving linear systems with the same matrix multiple times. For these settings, the ability of PreCorrectors to generalise (Figure 4) makes the training overhead less significant.\"}", "{\"summary\": \"The paper titled \\\"Learning from Linear Algebra: A Graph Neural Network Approach to Preconditioner Design for Conjugate Gradient Solvers\\\" presents a novel method for designing preconditioners using graph neural networks (GNNs). It aims to improve the efficiency of solving large linear systems that arise in computational science and engineering, particularly those characterized by parametric partial differential equations. The authors argue that their approach, termed \\\"PreCorrector,\\\" leverages existing linear algebra techniques and demonstrates superior performance in numerical experiments compared to traditional and other neural network-based methods.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": [\"**Innovative Approach**: The use of GNNs to enhance traditional preconditioners is a fresh perspective in the domain of numerical linear algebra.\", \"**Theoretical Foundation**: The paper provides a strong theoretical basis for the loss function and discusses its implications on the preconditioner's performance.\", \"**Numerical Validation**: Extensive experiments validate the proposed method, showcasing its effectiveness in reducing the condition number of systems, which is crucial for the performance of iterative solvers.\"], \"weaknesses\": [\"**Limited Scope**: While the paper presents promising results, the experiments are confined to a specific class of parametric PDEs. The generalizability of the results to broader contexts remains unclear.\", \"**Comparison with State-of-the-Art**: The comparison with existing methods could be more comprehensive. Many contemporary techniques in preconditioner design and GNN applications are not adequately addressed.\", \"**Complexity of Implementation**: The proposed method may introduce additional complexity in practical implementations, which could deter its adoption in industry settings.\"], \"questions\": \"1. Can the authors provide more insights into how the proposed GNN architecture can be adapted or generalized to other types of linear systems?\\n2. How do the computational costs of training and implementing the GNN compare to those of traditional preconditioner design methods?\\n3. Are there plans to evaluate the proposed method on a wider array of problems, particularly those encountered in real-world engineering applications?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Dear Reviewer X4c7,\\n\\nThank you for your work in reviewing our manuscript. Let us answer the questions from your review.\\n\\n> The paper is incremental in that a majority of the basis work comes from Li et al. (2023)\\n\\nAlthough we have not invented the use of GNNs on sparse linear systems, the PreCorrector is, to our knowledge, the first to achieve a better effect on the spectrum than classical preconditioners of the ILU family. Moreover, in our experiments, different realizations of the message passing architecture, node/edge updates, etc. did not change the convergence or the resulting preconditioner quality. We observe that the crucial part of a good neural preconditioner is initialization and stable learning, which is achieved by the PreCorrector architecture.\\n\\nThe GNNs from [Li et al.] have major limitations that limit the quality of the resulting preconditioner: (i) convergence to local minima and (ii) unstable learning. Both are addressed by PreCorrector.\\n\\n> While the authors have made clear how the training loss (3) is reformulated to (5), one should note that (2) can also be reformulated to (5).\\n\\nIndeed, the Hutchinson trick can be applied to loss (2). In addition, we used the Hutchinson trick for training with loss (2) (please note the caption for Table 6). The main idea is that using loss (3) is more beneficial than using loss (2) because loss (3) emphasises the low frequencies where CG has the most problems. While training with loss (2) can be done in an unsupervised manner, the resulting preconditioner will have a worse effect on the spectrum than when training with loss (3) (see Table 6). Moreover, when training with loss (2), the resulting preconditioner will be worse than the classical IC preconditioner in terms of CG iterations (see number of iterations \\\"Loss (2)\\\" from Table 6 and \\\"IC(0)\\\" for grid 64x64 in Table 1). Thus, loss (5) is only needed to avoid explicit inverse calculation in loss (3).\\n\\n> How does training time compare with solution time? Does the Pre-time column in Tables 1 to 4 mean the evaluation of the GNN or the training of the GNN?\\n\\nPre-time means the construction time of the neural and classical preconditioners. Note that the pre-time for the PreCorrector is calculated including both the inference of the GNN and the construction of the classical preconditioner.\\n\\nTraining PreCorrector on the most complex linear system (grid 128, variance 0.7) converged in about 500 epochs or 170 minutes (see lines 306-316 for experiments environment info). Unfortunately, any data-driven approach requires additional overhead to train the model. On the other hand, some applications require solving linear systems with the same matrix several times. For these settings, the ability of PreCorrector to generalize (see Figure 4) makes it a more favourable approach.\"}", "{\"comment\": \"Thank you for pointing out the right-hand side as an input to the GNN. Our work was inspired by the previous paper (Li et al. 2023) where the GNN takes both $A$ and $b$ as inputs. In fact, it is a bug and for a proper preconditioner construction routine, the GNN should not be dependent on $b$. We repeated our experiments with dummies for $b$ in the GNN input (vector of ones for each sample) and the results were exactly the same. So the experiments in the submitted manuscript are valid for a GNN that does not depend on $b$.\\n\\nRegarding the test problems, we tested our approach on the datasets that are more complex than the datasets of the previous work (with contrast coefficients). In fact, the previous approach failed on these datasets. We believe that it is possible to create a learnable transformation that will be universal for different sparse matrices to construct the ILU decomposition that will significantly reduce $\\\\kappa(A)$. Considering this learnable transformation as a primary goal, we will create a comprehensive and diverse dataset for it, which will most likely deserve its own research paper.\"}", "{\"title\": \"One question unanswered\", \"comment\": \"Thank you for the rebuttal. One question for which I am still waiting for an answer is, in your loss (5) reformulated from the advocated loss (3), how is $x_i$ generated? It appears that you still need to solve the linear system $A x_i = b_i$ given a sampled $b_i$ (which is the original problem being tackled).\"}", "{\"summary\": \"introduces a method for improving preconditioners in the Conjugate Gradient (CG) method using Graph Neural Networks (GNNs). The key contribution is a new scheme, called PreCorrector, which trains a GNN to modify and improve classical preconditioners like incomplete Cholesky (IC). The paper demonstrates that this approach outperforms traditional preconditioning techniques and previously proposed neural network-based preconditioners.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. The method generalizes well to unseen linear systems, indicating robust applicability.\\n2. The integration of GNNs seems reasonable.\", \"weaknesses\": \"1. The novelty of this work appears to be limited, as it closely resembles existing methods. Moreover, there is a lack of thorough discussion and comparison with relevant prior works, such as [Azulay, Yael, and Eran Treister. \\\"Multigrid-augmented deep learning preconditioners for the Helmholtz equation\\\"], [Li, Kangan, Sabit Mahmood Khan, and Yashar Mehmani. \\\"Machine learning for preconditioning elliptic equations in porous microstructures: A path to error control\\\"], and [Li, Kangan, Sabit Mahmood Khan, and Yashar Mehmani. \\\"Machine learning for preconditioning elliptic equations in porous microstructures: A path to error control\\\"]. A more detailed comparison would help to better contextualize the contribution of this work within the existing literature.\\n\\n2. The problems for validation are relatively esay and small, which are Poisson 2D or diffusion equation.\", \"questions\": \"1. Could the authors provide more specific details on how the computation time was measured, including the tools or setup used for timing and any relevant conditions that might affect the results?\\n\\n2. The idea presented seems similar to the method proposed by [Li, Yichen, et al. \\\"Learning preconditioners for conjugate gradient PDE solvers.\\\" International Conference on Machine Learning. PMLR, 2023]. However, their method failed with cases involving varying coefficients, as seen in Tables 4 and 5. Could you provide a more detailed explanation of why their approach might have failed, considering differences in the architecture, loss function, and training method? The current explanation in the paragraph \\\"Experiments with neural preconditioner design\\\" is not clear, and it is difficult to understand what sets PreCorrector apart and how it addresses these challenges.\\n\\n3. Any explanation why the factor $L$ depends on $b$ from L(\\u03b8) = GNN(\\u03b8, A, b)? Does it mean that in the solving phase, the factor $L$ would vary across iterations in the sense that L(\\u03b8) = GNN(\\u03b8, A, r), where r is the residual.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Dear Reviewer 7yrG,\\nThank you for your work in reviewing our manuscript. Let us answer the questions from your review.\\n\\n> The comparison with existing methods could be more comprehensive. Many contemporary techniques in preconditioner design and GNN applications are not adequately addressed.\\n\\nOur main goal is to create a better preconditioner than the classic ones in the ILU family. So the comparison is made with that in mind. Another reviewer shared two relevant papers, which we have mentioned in the \\\"Related work\\\" paragprah in lines 98-101 in the revised manuscript. Although we have limited the number of pages, we would like to clarify the comparison with these papers here.\\n\\nBoth approaches use convolutional neural networks (CNNs). In preconditioner design, CNNs can suffer from the curse of dimensionality, as convolutions scale poorly with matrix growth, since sparse matrices must be materialised as dense ones. Furthermore, message-passing GNNs can be seen as a generalization of convolutional neural networks, which can operate not only on a rectangular grid with a fixed number of neighbours, but also on an arbitrary grid. Moreover, both works are essentially quite different from our approach: these papers propose hybrid preconditioners with neural networks that also perform inference at each step of the iterative solvers. This is very different from the PreCorrector, which is not a preconditioner itself, but is used to create a classical preconditioner from the matrix.\\n\\n> The proposed method may introduce additional complexity in practical implementations, which could deter its adoption in industry settings.\\n\\nThe proposed approach is a very shallow neural network (2754 parameters), so the inference of such a network with a sparse linear system is negligible compared to the iteration speed-up, which is also illustrated in our experiments. Furthermore, for real industrial applications, the C++ and/or CUDA implementation will be used. These will further reduce the computation time. See also Appendix A.5 for details on scalability.\\n\\nOn the other hand, some applications require solving linear systems with the same matrix several times. For these settings, the ability of PreCorrector to generalize (see Figure 4) makes it a more favourable approach.\\n\\n> Can the authors provide more insights into how the proposed GNN architecture can be adapted or generalized to other types of linear systems?\\n\\nWhile we have focused on the linear systems with SPD matrices, the proposed architecture can be generalized to general patterns: one should use ILU instead of IC and the GNN neural network should predict the whole graph, not only the one corresponding to the lower triangular part of the matrix (to obtain the factors $L$ and $U$). However, different types of matrices besides SPD and a larger number of iterative solvers to solve systems with them deserve their own research paper.\\n\\n> How do the computational costs of training and implementing the GNN compare to those of traditional preconditioner design methods?\\n\\nThe inference (preconditioner construction) time with the proposed approach is reported as \\\"Pre-time\\\" in the Tables 1-4 and Tables 7-10 (for PreCorrector, this time includes both GNN inference and classical preconditioner construction time). As the matrix size increases, the construction time of both classical and neural preconditioners becomes less significant. When the preconditioner is combined with the PreCorrector, there is no difference in its use during CG compared to preconditioners designed using classical methods. Moreover, it has exactly the same algorithmic complexity as the classical preconditioner it is composed of (e.g. PreCorrector[IC(0)] and IC(0)).\\n\\nTraining PreCorrector on the most complex linear system (grid 128, variance 0.7) converged in about 500 epochs or 170 minutes (see lines 306-316 for experiments environment info). Unfortunately, any data-driven approach requires additional overhead to train the model.\\n\\nSpeaking about the implementation of PreCorrect, it is basically a combination of well-known operations: classical algorithm for IC decomposition, multilayer perceptrons for encoders, decoders and update functions in GNN, and message-passing architecture, which mostly consists of common graph operations (e.g. collecting information about node neighbourhood).\\n\\n> Are there plans to evaluate the proposed method on a wider array of problems, particularly those encountered in real-world engineering applications?\\n\\nYes, we have plans to evaluate the PreCorrector on the real problems of various applications. The application of PreCoorector is our next primary goal.\"}", "{\"comment\": \"Thank you for supplementing additional information. I can see that (3) is better than (2). The need to solve $A_ix_i=b_i$ to create a dataset, however, is a computational concern, especially when this dataset is large.\"}", "{\"comment\": \"For training with loss (3), reformulated as loss (5), one needs a dataset of $N$ samples of the linear system $A_i x_i = b_i$, where $x_i$ is obtained by solving the linear system. However, in practice this should not be an additional problem, since the solution is not needed during inference, but is part of the dataset preparation for training.\\n\\nNote that training with loss (2) can indeed be done unsupervised without solving $A_i x_i = b_i$. However, since models with loss (2) cannot outperform preconditioners with classical algorithms for IC (see number of iterations \\\"Loss (2)\\\" from Table 6 and \\\"IC(0)\\\" for grid 64x64 in Table 1), in practice neural network models trained with loss (2) will not be applicable, since it is better just to use classical IC (a very simple and efficient algorithm) without any training at all.\"}", "{\"comment\": \"Thank you for your clarification. I am keeping my original rating.\"}", "{\"summary\": \"This paper proposes an approach to computing a preconditioner for the conjugate gradient method for solving large linear systems. The approach improves Li et al. (2023) in that rather than learning the nonzeros of the incomplete Cholesky factor, it learns a correction to the factor. The authors show that their approach can outperform the standard incomplete Cholesky (IC) preconditioner, addressing a drawback of Li et al. (2023), which is less performant than IC.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": [\"The idea of learning a correction rather than learning the nonzeros of a preconditioner is novel.\", \"The learned preconditioner can generalize to different grid sizes and different parameters of the PDE.\", \"The proposed approach can outperform the standard incomplete Cholesky (IC) preconditioner, addressing a drawback of Li et al. (2023), which is less performant than IC.\"], \"weaknesses\": [\"The paper is incremental in that a majority of the basis of the work comes from Li et al. (2023).\", \"The work stresses the use of a different loss function (Eqn (3) rather than Eqn (2)), but the argument is dubious. The authors reformulate Eqn (3) to Eqn (5) in practice, which, however, can also be obtained directly from Eqn (2). Hence, the authors' argument that they use a better loss function ((3) against (2)) does not hold. For more details, see a related question in the following \\\"Questions\\\" section.\"], \"questions\": [\"While the authors have made clear how the training loss (3) is reformulated to (5), one should note that (2) can also be reformulated to (5), because $b_i = A x_i$. The subtlety lies in whether $x_i$ or $b_i$ is drawn from the standard normal. When one starts with the loss (3), $b_i$ is drawn from the standard normal, which causes the trouble of computing $x_i$ (which requires solving linear systems). In contrast, when one starts with the loss (2), one may draw $x_i$ from the standard normal, which causes no difficulty in computing $b_i$. In this regard, the loss (2) is even better than (3) in practice.\", \"How does training time compare with solution time? Does the Pre-time column in Tables 1 to 4 mean the evaluation of the GNN or the training of the GNN?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Dear Reviewer ZRtj,\\n\\nThank you for your work in reviewing our manuscript. Let us answer the questions from your review.\\n\\n> Citations\\n\\nThank you for sending us relevant papers, we have mentioned them in the \\\"Related work\\\" section in lines 98-101 in the revised manuscript. Although we have limited the number of papers, we would like to clarify the comparison with these papers here.\\n\\nBoth approaches use convolutional neural networks (CNNs). In preconditioner design, CNNs can suffer from the curse of dimensionality, as convolutions scale poorly with matrix growth, since sparse matrices must be materialized as dense ones. Furthermore, message-passing GNNs can be seen as a generalization of convolutional neural networks, which can operate not only on a rectangular grid with a fixed number of neighbours, but also on an arbitrary grid. Moreover, both works are essentially quite different from our approach: these papers propose hybrid preconditioners with neural networks that also perform inference at each step of the iterative solvers. This is very different from the PreCorrector, which is not a preconditioner itself, but is used to create a classical preconditioner from the matrix.\\n\\n> The novelty of this work appears to be limited, as it closely resembles existing methods.\\n\\nAlthough we have not invented the use of GNNs on sparse linear systems, the PreCorrector is, to our knowledge, the first to achieve a better effect on the spectrum than classical preconditioners of the ILU family. Moreover, in our experiments, different realizations of the message passing architecture, node/edge updates, etc. did not change the convergence or the resulting preconditioner quality. We observe that the crucial part of a good neural preconditioner is initialization and stable learning, which is achieved by the PreCorrector architecture.\\n\\nThe GNNs from [Li et al.] have major limitations that limit the quality of the resulting preconditioner: (i) convergence to local minima and (ii) unstable learning. Both are addressed by PreCorrector.\\n\\n> Could the authors provide more specific details on how the computation time was measured, including the tools or setup used for timing \\nand any relevant conditions that might affect the results?\\n\\nPlease see lines 312-316 for details of the comparison setup. Note that while the GPU was used for preconditioner construction with PreCorrector and GNN from [Li et al.], the CPU was used for the classical IC algorithm. Classical IC algorithm is fully sequential and cannot significantly benefit from data parallelization. The $\\\\texttt{ilupp}$ library is a Python API for C++ code, and our implementations of PreCorrector and GNN from [Li et al.] are coded in $\\\\texttt{jax}$, which is JIT-compiled for time computation. For CG, the $\\\\texttt{scipy.sparse.linalg.cg}$ implementation and the $\\\\texttt{scipy.sparse.linalg.LinearOperator}$ class from it are used for time-to-solution measurements. \\n\\n> Could you provide a more detailed explanation of why their approach might have failed, considering differences in the architecture, loss function, and training method?\\n\\nThe main limitations of the previous work by [Li et al.] are summarised in lines 194-198. Let us discuss them here in greater details. While GNN of [Li et al.] indeed converges and provides a neural IC that can reduce the number of iterations in CG for certain problems, it has a worse effect on the spectrum than classical IC from linear algebra. First, we believe that this is due to convergence to suboptimal local minima, which can be overcome by the starting point in training. In PreCorrector we get such a good initial guess with a classical IC decomposition. Second, the training GNN of [Li et al.] is unstable, since at the very beginning of training one has to compute the loss (5) with a random matrix L. We observed a loss overflow as the matrix size grows. Finally, we observe that we cannot predict the resulting quality of the GNN from [Li et al.]. In our experiments, we were able to compute the condition number for small linear systems: decreasing the loss did not correspond to decreasing the condition number of $P(\\\\theta)^{-1}A$. Thus, the stopping criteria of GNN from [Li et al.] training has to rely on the condition number of the preconditioned system, which is extremely costly. The learning of the PreCorrector starts with the gradient of the pure classical IC (since $\\\\alpha$ is initialized with 0), which ensures stable learning.\\n\\nThus, the limitations of GNN from [Li et al.] are (i) convergence to local minima and (ii) unstable learning. Both are fixed by PreCorrector.\"}" ] }

[ "Hoang Thanh Lam", "Raul Fernandez-Diaz", "Vanessa López" ]

[ "Molecule", "repesentation learning", "drug discovery", "ADMET", "drug properties prediction" ]

{ "note_id": [ "C2pOif2KXw", "Bdghi7b85D", "6Y9VqTw8IY", "5wKmcM5gyx", "4SaxkMsaOJ" ], "note_type": [ "comment", "official_review", "official_review", "official_review", "official_review" ], "note_created": [ 1732175229353, 1730345377589, 1730360404090, 1730711576386, 1729052276702 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission10246/Authors" ], [ "ICLR.cc/2025/Conference/Submission10246/Reviewer_hPpg" ], [ "ICLR.cc/2025/Conference/Submission10246/Reviewer_8aHp" ], [ "ICLR.cc/2025/Conference/Submission10246/Reviewer_BeZK" ], [ "ICLR.cc/2025/Conference/Submission10246/Reviewer_udzS" ] ], "structured_content_str": [ "{\"withdrawal_confirmation\": \"I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors.\"}", "{\"summary\": \"This paper proposes the use of substructure masking, instead of random masking, for molecular SMILES representation learning and achieves performance gains on ADMET tasks.\", \"soundness\": \"2\", \"presentation\": \"1\", \"contribution\": \"1\", \"strengths\": \"The paper is well-structured and covers most of the important sections.\", \"weaknesses\": \"1. **Novelty**: The proposed approach seems overly simple and lacks novelty, especially given that substructure masking for graph or SMILES has been extensively explored in recent work (e.g., GROVER[1], UniMAP[2], Unicorn[3]). This paper\\u2019s masking approach could be seen as a subset of UniMAP's method, but lacks a clear advantage or innovation in comparison. Specifically, UniMAP\\u2019s Fragment Level Cross-Modality Masking task masks the substructure of the graph and its corresponding smiles sequence, which is almost the same as the idea of \\u200b\\u200bthis paper.\\n\\n[1] Self-Supervised Graph Transformer on Large-Scale Molecular Data, NeurIPS 2020\\n\\n[2] UniMAP: Universal SMILES-Graph Representation Learning, 2023\\n\\n[3] UniCorn: A Unified Contrastive Learning Approach for Multi-view Molecular Representation Learning, ICML 2024\\n\\n2. **Writing and Literature Support**:\\n\\nThe motivation in the introduction requires stronger literature support. For example, the claim that transformer-based models are inferior to description/fingerprint models would benefit from additional evidence and analysis.\\n\\nThe Related Work section lacks references in subsection 2.4. The methodology section presents only the masking algorithm without introducing the overall model architecture, specific loss functions, or details like masking ratios.\", \"reference_formatting_is_inconsistent\": \"for instance, the citation for the ADMET benchmark (Huang et al.) is missing a publication year, and the reference for Graphormer should include its original proposal, not solely later benchmarking studies.\\n\\n3. **Experimental Evaluation**:\\n\\nThe experimental evaluation is limited, as it only uses one benchmark and lacks sufficient transformer-based baselines. The reported results do not convincingly demonstrate the effectiveness of the approach. In most tasks, the proposed method does not surpass baselines.\\n\\nRecommended baselines to consider include transformer-based methods like ChemBERTa and MolFormer. Given the convertibility between SMILES and 2D graphs, including graph masking methods as baselines (such as Graphormer and GROVER) would add relevance. \\n\\nThe experimental setting lacks detail, particularly on Elo ratings and Wilcoxon signed-rank test definitions.\\n\\n---\\nBased on the issues outlined above, I find the method overly simplistic and lacking in noticing recent works. The writing requires substantial improvements, and the experimental improvements do not appear sufficient for acceptance at a top conference. Therefore, I recommend rejection.\", \"questions\": \"1. Could the authors incorporate more recent baselines and potentially test on the MolNet benchmark?\\n\\n2. How does this method compare to other substructure masking approaches like UniMAP and GROVER in terms of advantages or improvements?\\n\\n3. For the Wilcoxon signed-rank test, was a one-tailed or two-tailed test applied? Please clarify the specific alternative hypotheses in each pairwise comparison in paper.\\n\\n4. For the leaderboard in Figure 1, could the authors specify the time of the ranking data collection?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The paper introduces a substructure-aware masking strategy alongside a new learning objective. Compared with random masking, the proposed masking strategy results in improved predictions.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"1\", \"strengths\": \"1. The paper is well-motivated since structure-aware masking provides more domain knowledge.\\n2. The paper is evaluated on extensive experiments.\", \"weaknesses\": \"1. The improved masking strategy does not provide significant improvement in experiments. In regression tasks, since the overall performance of SmilesGraph does not outperform MapLight+GNN, does this means it can be better just ensemble two models instead of providing structural knowledge in masking strategy?\\n\\n2. The baselines are not comprehensive enough. Many advanced Transformer-based achieved excellent performance, such as MolT5[1] and BioT5[2]. It will be much more convincing if there's direct comparison among these approaches.\\n\\n3. The major novelty, masking the molecule based on the substructure, is somewhat incremental.\\n\\n[1]. Translation between Molecules and Natural Language\\n[2]. BioT5: Enriching Cross-modal Integration in Biology with Chemical Knowledge and Natural Language Associations\", \"questions\": \"See weaknesses.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper introduces a new mask strategy for pretraining the smiles-based Transformers. In contrast to the previous token-level masked strategy, the proposed method tends to conduct a substure-level prediction task during pretraining to better utilize the inductive bias over the molecule data. The experiments are conducted over several property prediction benchmarks to demonstrate the effectiveness of the proposed method.\", \"soundness\": \"2\", \"presentation\": \"1\", \"contribution\": \"1\", \"strengths\": \"1. I believe that the paper is studying a reasonable problem of mask-predicted based objective. And integrating the new masked strategy could be an interesting exploration direction.\\n2. The paper makes a comprehensive review and discussion of the previous works, including the current leaderboard of the several important benchmarks. And it is helpful for gaining an intuitive overview of the progress in the field.\", \"weaknesses\": \"However, I believe the paper needs a major revision to get accepted by top conferences.\\n1. The presentation and formulation of the paper is very poor and it seems the paper is finished in a rush. Specifically, the paper is making some basic misunderstandings by treating Transformer as a training algorithm which is actually a network structure. And it causes confusion for me to understand the actual training algorithm. I suppose it is a mask prediction like Bert's objective. If the guess is correct, what are the details for the masking? e.g. What ratio of the tokens is masked during the pretraining? \\n2. The experiments results are not solid enough. The main results are conducted over several benchmarks. However, whether the training data of different methods are aligned is lack of explaination. Besides, the ablation study is also missed to demonstrate the effectiveness of different components. \\n3. Though I believe the motivation of the method is reasonable, the substructure or scaffold based tokenizers are proposed several years ahead of this work, e.g. JTVAE. I would like to know what the benefit of the proposed approach is over using fragment tokenizer and conducting masked predictions. \\nGiven the above facts, I recommend a strong rejection of this draft.\", \"questions\": \"Refer to above weakness.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"1\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The authors present a novel approach aims at improving transformers' ability to identify molecular substructures through a substructure-aware masking strategy. They utilize molecular graphs to identify groups of neighboring molecules based on a radius r (referred to as substructures) and map the corresponding characters in SMILES strings. Subsequently, they apply a hash function to the identified substructures and use a masked language modeling transformer to predict fixed-size binary hash vectors, employing binary cross-entropy loss for optimization.\", \"soundness\": \"2\", \"presentation\": [\"The first sentence of the paper states, \\\"Molecular graphs are commonly represented using SMILES (Simplified Molecular Input Line Entry System) strings,\\\" which is not entirely accurate. Molecular graphs and SMILES represent molecules in different ways; therefore, it is misleading to assert that SMILES represent molecular graphs.\", \"In the abstract, the authors claim to propose \\\"a new learning objective,\\\" but it is unclear what this new learning objective entails. If it refers to the BCE loss, then BCE loss is not a new concept.\", \"In line 239, the term \\\"alphabet\\\" in the context of masked substructures could be misleading, as \\\"alphabet\\\" typically refers to a fixed set of symbols. It would be clearer to refer to it as a \\\"vocabulary\\\" or \\\"set of possible tokens,\\\" aligning the terminology with common practices in natural language processing (NLP).\", \"The figures depicting molecular graphs in the paper do not accurately represent the structures of the molecular graphs discussed.\", \"Citations for the baseline models in Tables 3, 4, and 5 should be included for proper attribution and context.\", \"The writing in the paper is not well polished, and there are several grammatical errors that need to be addressed to improve clarity and comprehension.\"], \"contribution\": \"1\", \"strengths\": [\"The authors thoroughly evaluate their approach across a wide range of tasks (22 ADMET tasks).\", \"The method achieves state-of-the-art results on 4 out of 22 tasks, highlighting its effectiveness and potential in specific applications within the field.\", \"The authors conduct a comprehensive ablation study examining the impact of hash size h and the number of masked substructures. This rigorous analysis provides valuable insights into the model's performance and helps identify optimal configurations for enhancing predictive accuracy.\"], \"weaknesses\": \"Methodology\", \"substructure_aware_masking_strategy\": [\"The authors randomly choose an atom and expand it to a specific radius r to form the substructure for masking. However, this method does not effectively capture chemically meaningful substructures, such as functional groups like OH, COOH, and others. By relying on a purely random selection of atoms and a fixed radius, the approach may overlook essential chemical features that are important for accurately modeling molecular behavior. Instead, a more effective strategy would be to focus on well-defined functional groups or significant chemical motifs. This could enhance the representation of molecules and improve the understanding of molecular substructure interactions and properties.\"], \"hash_vector_generation\": [\"The explanation lacks sufficient detail regarding the characteristics and implementation of the hash vector. It would be beneficial for the authors to clarify how the hash function is defined and how collisions are managed. Without this information, the effectiveness of the hashing process remains uncertain.\", \"Random hashing can lead to information loss, particularly if multiple unique substructures hash to the same value (collisions). The authors should address how they minimize the loss of information due to collisions and whether they evaluate the effectiveness of this method in retaining meaningful structural information.\"], \"masked_language_modeling\": [\"In line 237, the rationale for performing regular individual token masking is unclear. The authors state that it is \\\"to learn regular token association,\\\" but the specific purpose of this masking remains ambiguous.\", \"The authors should conduct an ablation study using the transformer model under the same experimental conditions, differing only in the masking strategy. This would provide a clearer demonstration of the effectiveness of the proposed substructure-aware masking strategy compared to conventional approaches.\"], \"evaluation\": \"The authors benchmark their method across a wide range of ADMET tasks, reporting results on 4 out of 22 tasks. However, I recommend that the authors also run benchmarks on MoleculeNet tasks to compare their performance against other well-established molecular representation methods.\", \"questions\": [\"Clarification of the New Learning Objective: Could you clarify what you mean by \\\"a new learning objective\\\"? Specifically, how does this differ from established methods like binary cross-entropy (BCE) loss?\", \"Impact of Masked Substructures: You mentioned that \\\"the number of masked substructures has minimal impact in the no-mask setting.\\\" Can you elaborate on this finding? What implications does this have for the design of your substructure-aware masking strategy?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}" ] }

[ "Edith Cohen", "Benjamin Cohen-Wang", "Xin Lyu", "Jelani Nelson", "Tamas Sarlos", "Uri Stemmer" ]

[ "PATE", "diverse tasks", "privacy-preserving machine learning", "coordinated sampling", "in-context learning" ]

{ "note_id": [ "xaBJCo6lTU", "w4xrv9TOUb", "v669q4o28o", "te31qxDhvU", "tP8Vo79TDg", "nHxrbm0p3c", "mFbV3995zl", "iApjVrLnwH", "hanFTGGTZq", "haQqj3JCU6", "fMurVOuLhx", "Xjx7kbMwAU", "XEmVTMD1KO", "UHxyQVxsKx", "ScJ2VUbSwS", "Qg7Dyk1SiZ", "PrTsYJ8pew", "Oadoj5Hpdg", "Noy6hE8wd0", "KFEtFiMsTk", "Er0JywssjC", "EdqRWP8VHe", "CAf4NGEVE7", "9Pb2VSLHR0", "8kLYqMBZhD", "8Mx9v3LpOi", "3GvQEvvw6W", "1yeSrmoLqs", "1N5FKLKeYk", "1ATbl0AoTH" ], "note_type": [ "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "meta_review", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "decision", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_review", "official_comment" ], "note_created": [ 1731566336033, 1731607775989, 1732675922715, 1732560114607, 1731781334002, 1731779382853, 1732479465860, 1732593353282, 1734545875651, 1732595045341, 1731607725662, 1731277705145, 1731583760171, 1732628254607, 1730442604611, 1731506497066, 1733008692253, 1731775102124, 1732678214669, 1732660408457, 1732595787317, 1731781684067, 1737523956611, 1732651913325, 1731778911614, 1731584974231, 1730652598200, 1733010950313, 1730699761689, 1731607568130 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission9042/Authors" ], [ "ICLR.cc/2025/Conference/Submission9042/Authors" ], [ "ICLR.cc/2025/Conference/Submission9042/Reviewer_KdZL" ], [ "ICLR.cc/2025/Conference/Submission9042/Reviewer_YCGG" ], [ "ICLR.cc/2025/Conference/Submission9042/Authors" ], [ "ICLR.cc/2025/Conference/Submission9042/Authors" ], [ "ICLR.cc/2025/Conference/Submission9042/Reviewer_QVdG" ], [ "ICLR.cc/2025/Conference/Submission9042/Authors" ], [ "ICLR.cc/2025/Conference/Submission9042/Area_Chair_W4sC" ], [ "ICLR.cc/2025/Conference/Submission9042/Reviewer_YCGG" ], [ "ICLR.cc/2025/Conference/Submission9042/Authors" ], [ "ICLR.cc/2025/Conference/Submission9042/Reviewer_FQD1" ], [ "ICLR.cc/2025/Conference/Submission9042/Authors" ], [ "ICLR.cc/2025/Conference/Submission9042/Authors" ], [ "ICLR.cc/2025/Conference/Submission9042/Reviewer_QVdG" ], [ "ICLR.cc/2025/Conference/Submission9042/Authors" ], [ "ICLR.cc/2025/Conference/Submission9042/Reviewer_QVdG" ], [ "ICLR.cc/2025/Conference/Submission9042/Authors" ], [ "ICLR.cc/2025/Conference/Submission9042/Authors" ], [ "ICLR.cc/2025/Conference/Submission9042/Authors" ], [ "ICLR.cc/2025/Conference/Submission9042/Authors" ], [ "ICLR.cc/2025/Conference/Submission9042/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission9042/Reviewer_FQD1" ], [ "ICLR.cc/2025/Conference/Submission9042/Authors" ], [ "ICLR.cc/2025/Conference/Submission9042/Authors" ], [ "ICLR.cc/2025/Conference/Submission9042/Reviewer_KdZL" ], [ "ICLR.cc/2025/Conference/Submission9042/Authors" ], [ "ICLR.cc/2025/Conference/Submission9042/Reviewer_YCGG" ], [ "ICLR.cc/2025/Conference/Submission9042/Authors" ] ], "structured_content_str": [ "{\"title\": \"Re rigorous privacy guarantees\", \"comment\": \"Thank you for your review! We will respond to the questions in multiple comments.\\n\\n\\n\\n### Addressing the stated \\\"weakness\\\" on rigorous privacy guarantees:\\n\\u201cI felt that the privacy guarantees are not rigorously stated, DP implementations are largely left as poorly-described black boxes (e.g., `NoisyArgMax' in Algorithm 2 is never formally introduced) and none of the algorithms include the privacy parameters as input. I didn't see a formal privacy analysis that can be easily verified, and in terms of reproducibility I feel like the algorithms can\\u2019t really be implemented without knowing, for example, how to calibrate the noise scale.\\u201d\\n\\n### Response:\\nThe detailed privacy analysis can be found in Appendix D and Appendix E. While this analysis is deferred to the Appendix due to its technical nature, it primarily reflects the fact that the aggregation methods\\u2014specifically, transforming a vote histogram into a single token via NoisyArgMax or private sampling\\u2014are **standard techniques in the literature.** In particular, for NoisyArgMax, one can simply **plug in** the procedures proposed in the original PATE paper (Laplace noise) or follow up work (Gaussian noise) and apply it with the histograms produced with Hot PATE. In the appendix, we study the tradeoffs obtained using the TCT framework (an extension of the sparse vector technique, but this is a \\u201cbonus\\u201d study that is not necessary for Hot PATE implementation).\\nThe key insight here is that with all these (standard) aggregation methods, the level of privacy noise required depends on the maximum frequency of any given token. That is, high consensus among teachers is best. With Cold PATE, consensus breaks down with diverse responses. In contrast, Hot PATE maintains high agreement even with diverse teacher responses\\u2014a property we establish both mathematically and empirically.\\n\\nIn the main text, we aimed for an accessible overview that is focused on the novel components of our work. The privacy loss improvement is demonstrated via the impact on maximum token frequency, that with all aggregation methods, is directly related to the privacy loss. We demonstrate a very significant **order of magnitude** improvement. Mathematically, Hot PATE aggregation can only result in improvement, and this increases with diversity.\"}", "{\"title\": \"Part 3 of response\", \"comment\": \"## Question 1\\n\\n\\u201cCan you ground your notion of diversity in a practical example? Why should one adopt your notion of diversity? What utility does it bring? Can you provide concrete empirical results to support the benefit of improved diversity as you define it?\\n\\n### Answer:\\n\\nIt is not \\u201cour notion\\u201d of diversity. Diversity is present in current LLMs in that the next-token distribution is a distribution and not a point mass. LLMs are tuned (via the temperature parameter) to allow for significant diversity \\u2013 often 2-7 bits of entropy per token. When you prompt an LLM, it will sample different responses. \\n\\nNote that our goal is not to \\u201ccreate\\u201d diversity but simply to allow for it to be transferred effectively in a privacy preserving way. Simply to emulate as effectively as we can (while preserving privacy) what an LLM prompted with sensitive data would do. We measure performance by the effectiveness of this \\u201ctransfer.\\u201d \\n\\n## Question 2\\n\\n\\u201cI had a lot of trouble with your presentation of the suggested method as a privacy-preserving algorithm. Having read the paper, I am not convinced of claims such as Line 337:\\nThis high agreement allows rare tokens to pass even with high privacy noise and allow for the aggregate distribution, with fixed privacy requirements, to satisfy Definition 1.\\nCan you make a clear case for this?\\u201d\\n\\n### Answer:\\n\\nCoordinated ensembles produce histograms that have \\u201chigh agreement\\u201d on a single (or very few) tokens. Magically, this happens even with high diversity, which is exactly the point of Hot PATE. You might want to look at Figure 10 to see an illustration. With coordinated ensembles, each time we sample the histogram the agreement token might be different. Now, high agreement (high count) means that we can use more noise. More noise means that the privacy loss is lower.\"}", "{\"comment\": \"Thanks for the response! Many of my concerns have been addressed \\u2014 particularly with regards to the updated privacy analysis. I also really appreciate the efforts to make the paper more accessible (and glad to have inadvertently helped catch a typo). I\\u2019ve read the other reviews and responses and am raising my score to 6.\\n\\nWhile I do highly value the novelty and technical contribution of the methods proposed in this paper, I would strongly caution the authors against gatekeeping this work for an \\u201cexperts-only\\u201d audience. Granted, it is inevitable that understanding many of the technical details would require a certain expertise, but ideally even a non-expert (or an expert with time constraints) could come away from this paper feeling inspired by the high-level ideas. I am sure that many readers with a similar background to reviewer FQD1 will be interested in this work, and in this case I don\\u2019t think it\\u2019s quite fair to shift the responsibility of having misunderstandings onto the reader, without first reflecting on how to improve the paper\\u2019s communication. For what it\\u2019s worth, I\\u2019m on board with reviewer FQD1\\u2019s suggestions and think that they could help the paper reach a broader audience and also clarify the practical implications of this work.\"}", "{\"comment\": \"Thank you for updating the submission! This makes this phase of the reviewing process easier.\\n\\nI missed that the definition of approx. DP was missing (as reviewer Kdzl pointed out) but agree that that was an important inclusion for completeness. \\n\\nYou addressed my concerns. Nice paper. I'm increasing my score to the next rung.\"}", "{\"comment\": \"Thank you for your comments! We uploaded a revised version accordingly and implemented your suggestions.\\n\\nWe made an effort to increase accessibility and reorganized and expanded section 4 in the main text to highlight the privacy properties of coordinated ensembles. \\nThe main text now exceeds the page limit, but we plan to address it by moving some of section 4.1 (proofs) to the Appendix but we left it in place for now to facilitate an easier comparison of the versions.\", \"as_for_your_comment_on_additional_empirical_evaluation\": \"The benefits of our method -- coordinated ensembles -- is established mathematically. We gain whenever there is diversity and we know that there is diversity. We believe that a proof of concept demonstration is sufficient for this purpose.\", \"as_for_your_question_on_limitations\": \"In terms of privacy and utility, coordinated ensembles are always more favorable than independent ensembles. This is established mathematically. But as we mention in Section 4.3, for the in-context learning application, without API support in the LLM there would be impact on efficiency.\\n\\nPlease let us know if the revision and our response addressed your concerns and if you have additional questions.\", \"title\": \"A revised version is uploaded\"}", "{\"title\": \"Uploaded revised version\", \"comment\": \"Thank you again for the review! We uploaded a revised version that addressed all comments and increased accessibility.\\n\\nAs you suggested, we included a citation and a review of [2] in which we explained why the aggregation method proposed there (a variant of independent ensembles) is inferior to ensemble coordination. This is established mathematically, and does not need experiments. Our contribution is an aggregation method that can be plugged in systems that use PATE for sequential text generation like the one in [2]. \\n\\nAdditionally, we expanded sections 3 and 4 to highlight our contributions and privacy analysis. As a result, the main text now exceeds the page limit, but we plan to address it by moving some of section 4.1 (proofs) to the Appendix. For now, we left it in place to facilitate an easier comparison of the versions.\\n\\nPlease let us know if you have further questions. Especially on why an empirical comparison on the setup of [2] is not needed and why the advantages of our work follow from the math.\"}", "{\"comment\": \"Thank you for the detailed response. After carefully reading the rebuttal and the comments from other reviewers, I keep my original score unchanged.\"}", "{\"comment\": \"Thank you for going through our response! You stated that you will increase your score, but it appears to still be \\\"6\\\"? thanks again.\"}", "{\"metareview\": \"The paper proposes a transformation that can be applied before PATE to increase consistency of expert predictions in case they share a low-probability prediction. The authors suggest this allows using private aggregation for new tasks such as aggregating next-word predictions from LLMs.\\n\\n**Strengths:**\\n* Interesting idea\\n* Promising initial experiments\\n\\n**Weaknesses:**\\n* Unconventional presentation that many reviewers found inaccessible despite being experts in the area (median score for presentation 2)\\n* Missing concrete privacy analysis in terms of theorems\\n* No experimental results for an actual end task\\n\\nBased on the reviews and my own reading of the paper, I very much agree with the assessment and recommendations for improvements by Reviewer FQD1.\\n\\nWhile the idea is interesting and the initial experiments are promising, I find the paper to be too premature for publication at ICLR. I appreciate the authors' claim of generality of their approach, but do not find this an acceptable excuse for not providing concrete privacy-utility tradeoff results and privacy analysis for an actual end task, with comparisons with previous state-of-the-art. Such results would be vital for potential users of the method to understand its value - while you demonstrate improvements in an intermediate metric, there is no guarantee on whether these will actually translate to significant improvements in performance in a relevant end task.\\n\\nIn terms of writing of the paper, I would strongly urge the authors to take the feedback of **all** reviewers seriously. The reviewers are experts in the field and represent potential readers of your paper. If they do not understand the paper, you should not blame the reviewers but think how you could write the paper better to avoid such misunderstandings in the future.\", \"additional_comments_on_reviewer_discussion\": \"While all reviewers have scores suggesting acceptance, in further discussion none of them was willing to champion the paper for acceptance.\\n\\nThe reviewers also found the authors' tone in the responses inappropriate in dismissing the expertise of the reviewers, bordering on violating the ICLR Code of Conduct.\"}", "{\"comment\": \"Updated now\"}", "{\"title\": \"Part 2 of response\", \"comment\": \"## \\\"Weakness\\\" 2:\\n\\n\\u201cWriting and exposition is not polished. The introduction is too long and full of technical detail with frequent forward references. None of the technical terms first appearance receive proper introduction. I find page 4 almost completely incomprehensible as a result. New terms are frequently used before they are properly defined. For instance, \\\"homogeneous ensembles\\\" is used in Line 186 but partially defined in Line 191. Some terms are really never properly defined at all in the introduction (\\\"diversity\\\", \\\"robustness parameter\\\", etc.)\\u201d\\n\\n### Response\\n\\nOur work is a technical contribution that is established formally. It must have technical details and we deferred what we could to the appendix. There are forward references to more details in order to make the presentation more accessible, which is standard practice. If you have constructive suggestions, we are happy to implement them. As for \\u201chomogeneous\\u201d ensembles in line 186, this is the \\u201cworking assumption\\u201d of \\u201csample and aggregate\\u201d DP schemes like PATE. We are happy not to use that term in line 186. You claim that the \\u201crobustness parameter\\u201d $\\\\tau$ is not formally defined, but this is incorrect. In the first mention (page 3) its purpose is explained at a higher level. We state it is a parameter and then state how it is used. We then fully formalize this in Definition 1 (That the reviewer indicated somehow they had read (see \\\"Weakness\\\" 1) but apparently missed that part?). Finally, you claim that we do not define \\\"diversity.\\\" We say exactly what we refer to -- \\u201cdiversity\\u201d is simply the notion that there are \\u201cmultiple good answers\\u201d which is the reality with generative tasks. The conditional distribution of the next token and the fact that it is not simply point mass means there is diversity. We do formalize the notion of \\u201ctransferring diversity\\u201d.\\n\\n## \\\"Weakness\\\" 3:\\n\\n\\\"Experimental results are limited. The results are mostly validating that the algorithm produces more \\\"diverse\\\" tokens. I think this is necessary and good. However, throughout the paper it is unclear what the value of this \\\"diversity\\\" is. I was hoping the experimental results would showcases a concrete benefit from having more diverse tokens. For instance, better generalization (test error) on a down-stream task.\\\"\\n\\n### Response\\n\\nThe misunderstanding by the reviewer here is that the diversity is not something we generate but something that is **present** in the teacher\\u2019s distributions. A teacher distribution is simply that of a tuned LLMl that is applied to the sensitive data. Our method is about allowing this diversity to be transferred in a privacy preserving way to the final privacy-preserving distribution. This is exactly what we demonstrate experimentally. Essentially, for a given privacy loss, our privacy-preserving token distribution is much closer to the average teacher distribution than what is produced by cold PATE aggregation. There is an order of magnitude improvement. Therefore, in short, we are not evaluating the LLM (llama 8B in this case). We take it as a given and there is no actual need to validate \\u201cgeneralization.\\u201d The goal is simply to transfer more effectively what it already does and not lose it in our privacy-preserving mechanism. \\n\\n## \\\"Weakness\\\" 4:\\n\\n\\\"Empirical results contain no privacy quantification. Although the paper seeks to find the trade-off between privacy and diversity, the empirical section contains no quantification of the privacy budget of the algorithm. Coupled with the fact that a proper privacy analysis is missing (see first point above) I have serious doubts regarding the privacy claims of the paper and the empirical section did not do much to alleviate them.\\\"\\n\\n### Response:\\n\\nWhat we compare is the baseline \\u201cindependent ensembles\\u201d and the proposed \\u201ccoordinated ensembles\\u201d. They are compared on the coverage and diversity of \\u201cfiltered\\u201d histograms that filter out the counts that are below a threshold $T$. Note that the histogram produced using shared randomness (coordinated ensembles) has the same privacy properties (L1 distance 2 between neighboring datasets) as a histogram produced via \\\"cold\\\" PATE. The only difference is the shape of the histograms, which is what we evaluate. The metric of coverage and diversity after filtering precisely captures what we want since the noise scale (privacy loss) determines that effective threshold. Doing the comparison this way is cleaner as it is not specific to the particular noise distribution (Laplace or Gaussian).\"}", "{\"summary\": \"The paper introduces HotPATE, a method based on the Private Aggregation of Teacher Ensemble with the distinction that the method forgoes independent teacher data and models. In fact, the teacher coordinate their sampling such that upon aggregation of their votes, rare teacher decisions (for instance, rare tokens in the case of private synthetic next-token generation) can still be produced without requiring a lot of agreement between teachers. The paper claims this process improves the diversity of the resulting vote histograms without privacy cost of not having high agreement (which is traditionally what ensures low PATE privacy costs for private prediction). A new definition for diversity-preserving aggregeation of distributions is presented. Empirical results show that under that definition, HotPATE improves upon ColdPATE. However, practical implications of the definition and broader contribution is unclear.\", \"soundness\": \"2\", \"presentation\": \"1\", \"contribution\": \"3\", \"strengths\": [\"Improving diversity of PATE responses is an interesting goal, given how much the privacy of PATE comes from teacher agreement (therefore lack of diversity in teacher votes).\", \"The idea of coordinated sampling of tokens seems novel. Although its privacy implications are unclear.\"], \"weaknesses\": [\"As someone who is quite familiar with PATE and its derivatives, I found this paper very hard to read and digest. I think there are a couple of reasons for this:\", \"**A robust privacy analysis is missing.** The paper introduces a particular histogram aggregation strategy that produces rate token frequency. In a sense, this is not an aggregation that produces a single vote but rather a transformed histogram. Overall, I found the presentation of this rather simple idea overly complicated in Section 3. However, the key issue here is not the contrived procedure and Definition 1, but rather the complete lack of privacy analysis under this new aggregation method. Let me clarify this point: the PATE privacy analysis only holds under the noisy argmax release. In particualr, the analysis is a function of the gap between the top vote and the second top vote of the histogram. If we were to use Def.1 and instead release transformed vote count (for the purposes of diversity), we are strictly releasing more information. In fact, since the rare token frequencies are kept (for diversity purposes), such a scheme will likely have higher privacy cost than releasing a full noised histogram of votes.\", \"**Writing and exposition is not polished.** The introduction is too long and full of technical detail with frequent forward references. None of the technical terms first appearance receive proper introduction. I find page 4 almost completely incomprehensible as a result. New terms are frequently used before they are properly defined. For instance, \\\"homogeneous ensembles\\\" is used in Line 186 but partially defined in Line 191. Some terms are really never properly defined at all in the introduction (\\\"diversity\\\", \\\"robustness parameter\\\", etc.)\", \"**Experimental results are limited.** The results are mostly validating that the algorithm produces more \\\"diverse\\\" tokens. I think this is necessary and good. However, throughout the paper it is unclear what the value of this \\\"diversity\\\" is. I was hoping the experimental results would showcases a concrete benefit from having more diverse tokens. For instance, better generalization (test error) on a down-stream task.\", \"**Empirical results contain no privacy quantification.** Although the paper seeks to find the trade-off between privacy and diversity, the empirical section contains no quantification of the privacy budget of the algorithm. Coupled with the fact that a proper privacy analysis is missing (see first point above) I have serious doubts regarding the privacy claims of the paper and the empirical section did not do much to alleviate them.\"], \"questions\": [\"Can you ground your notion of diversity in a practical example? Why should one adopt your notion of diversity? What utility does it bring? Can you provide concrete empirical results to support the benefit of improved diversity as you define it?\", \"I had a lot of trouble with your presentation of the suggested method as a privacy-preserving algorithm. Having read the paper, I am not convinced of claims such as Line 337:\", \"> This high agreement allows rare tokens to pass even with high privacy noise and allow for the aggregate distribution, with fixed privacy requirements, to satisfy Definition 1.\", \"Can you make a clear case for this?\", \"Have I misunderstood part of your work? To be clear, I think as is, this paper is not ready for publication. However, I want to be fair and make sure that I have not misunderstood your work. So I'll be happy to engage with you during the rebuttal process.\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"details_of_ethics_concerns\": [\"Can you ground your notion of diversity in a practical example? Why should one adopt your notion of diversity? What utility does it bring? Can you provide concrete empirical results to support the benefit of improved diversity as you define it?\", \"I had a lot of trouble with your presentation of the suggested method as a privacy-preserving algorithm. Having read the paper, I am not convinced of claims such as Line 337:\", \"> This high agreement allows rare tokens to pass even with high privacy noise and allow for the aggregate distribution, with fixed privacy requirements, to satisfy Definition 1.\", \"Can you make a clear case for this?\", \"Have I misunderstood part of your work? To be clear, I think as is, this paper is not ready for publication. However, I want to be fair and make sure that I have not misunderstood your work. So I'll be happy to engage with you during the rebuttal process.\"], \"rating\": \"6\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to the reviewer's questions\", \"comment\": \"### Question 1:\\n\\u201cBesides coverage and diversity, are there other metrics which could be used to demonstrate the effectiveness of hot PATE?\\u201d\\n\\n### Response:\\nWe did consider other metrics such as the TV distance between the average and \\u201cfiltered\\u201d distribution (after removing counts below threshold) for independent (\\u201ccold\\u201d PATE) and coordinated (\\u201chot\\u201d PATE) teacher ensembles (Figure 5). We also considered diversity per coverage. See also \\u201crobust average\\u201d (Remark 3 in Appendix B.2) and Figure 7. This focuses on the fraction that we can transfer \\u201cprivately\\u201d versus the potential for what can be transferred privately (the robust average distribution).\\n\\nBut overall these are the \\\"right\\\" metrics for our goal of transferring the diversity present in the teacher distributions.\\nAll our metrics demonstrate very significant gains from the hot PATE method (coordinated teacher ensembles). \\n\\n### Question 2\\n\\u201cLine 274\\u201d\\n\\n### Response:\\nYes exactly. Often (such as in the original PATE works [Papernot et al 2018] the noise scale $\\\\sigma$ is scaled to the level of obtaining utility and this determines the privacy cost. (Other privacy analysis frameworks such as TCT do this as well) So the privacy loss from transferring is inversely related to $\\\\max_j c_j$ (the $\\\\arg$ there is a typo which we will fix)\"}", "{\"comment\": \"We believe we addressed your primary concerns and those of other reviewers as well, so we are puzzled by your response.\\n\\nYour primary concerns were \\n1. Relation to literature and review of prior work\\n2. Empirical comparison on the datasets used in some prior/concurrent works like [2]\\n\\n We believe both were addressed. For [2] we explained why such a comparison is irrelevant. In a nutshell:\\n - We propose a method, not a system. [2] proposes a system. \\n - The improvement we show is established mathematically. There is always improvement that is more significant with diversity. We also included an empirical demo as an illustration (and a simulation on varied parametrized synthetic distributions in the appendix for the purpose, as a bonus, of exploring data-dependent privacy analysis) but just taking arbitrary distributions and going further is kind of pointless. \\n - Moreover, if we \\\"plug in\\\" ensemble coordination in the \\\"system\\\" of [2] it will not only lead to significant improvement when there is diversity (and [2] works on the premise that there is) but might change what appears to be compromises made in their design.\\n\\nNow, let us try to explain again this last point (which appears to have been missed by the reviewer).\\nConsider the case that all the teacher distributions are identical (the improvement kicks in more broadly, but this is just in order to convey the argument in a very simple setting). Suppose the prompt was \\\"Please suggest one boy's name for my newborn.\\\" A modern model would sample a name from a distribution, likely to be non uniform and with large support, with the most likely response having probability perhaps 1/50 or 1/100. (The distribution might change by the private context but we are not considering it right now, see the paper). Ideally, we would want in this case for the privacy-preserving response to be a sample from the exact same distribution. This is exactly what a coordinated ensemble would produce and there is no privacy loss penalty if the top probability is 1/100 or 1/1000.... One token would have all votes each time, depending on the \\\"shared\\\" randomness and the aggregation is successful with large privacy noise.\\n\\nWith cold PATE, the votes split up and utility drops sharply and for low enough top probability we cannot report anything. [2] seemed to address it by changing the teacher distributions to a uniform over the top $k$ tokens. For a parameter $k$. This is undesirable already because it looses the semantics of the well tuned base model. We do not want the distribution to be uniform. We also want it to report rare names (not only the top-$k$) with some probability. Moreover, the aggregation still pays a penalty for $k$. The privacy loss of the aggregation depends inversely on the scale of the noise, which decreases $\\\\propto k$ (aka the diversity). So hot PATE offers a factor $k$ gain.\"}", "{\"summary\": \"This paper introduces \\\"hot\\\" PATE, an extension of PATE designed for in-context learning via prompts, addressing tasks that are \\\"diverse\\\" and open-ended. They empirically demonstrate the potential of hot PATE for in-context learning.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"1. The motivation is clear: sequential text generation tasks through in-context learning are inherently diverse (\\\"hot\\\") with multiple valid responses.\\n\\n\\n2. The idea of aggregating responses from different teachers to maintain both diversity and privacy is interesting.\", \"weaknesses\": \"1. My primary concern is the empirical evaluation. The utility of in-context learning is typically measured by accuracy in the literature (e.g., [1,2,3]). However, this paper does not report in-context learning accuracy on specific tasks. It is unclear how much benefit hot PATE can provide for in-context learning. Additionally, the experiment is conducted on only one dataset, which is insufficient, and there is only one baseline (\\\"cold\\\" PATE). It is unclear why comparisons to prior in-context learning work (e.g., [1,2,3]) are not included.\\n\\n\\n[1] Duan, Haonan, et al. \\\"Flocks of stochastic parrots: Differentially private prompt learning for large language models.\\\" Advances in Neural Information Processing Systems 36 (2024).\\n\\n[2] Tang, Xinyu, et al. \\\"Privacy-Preserving In-Context Learning with Differentially Private Few-Shot Generation.\\\" The Twelfth International Conference on Learning Representations.\\n\\n[3] Wu, Tong, et al. \\\"Privacy-Preserving In-Context Learning for Large Language Models.\\\" The Twelfth International Conference on Learning Representations.\\n\\n\\n2. The paper states that Wu et al. (2023), Lin et al. (2024), and Xie et al. (2024) are independent concurrent work, which is inaccurate. These should be considered prior work, as Wu et al. (2023) and Lin et al. (2024) were published at ICLR 2024, and Xie et al. (2024) at ICML 2024.\\n\\n\\n3. I suggest extending the literature review of this paper by including the work \\\"Tang, Xinyu, et al. Privacy-Preserving In-Context Learning with Differentially Private Few-Shot Generation. The Twelfth International Conference on Learning Representations.\\\". This work studies differentially private in-context learning and proposes to use the sample and aggregate framework to generate DP synthetic examples for in-context learning inference. It could also serve as an experimental baseline for comparison.\\n\\n\\n\\n\\n3. Some typos:\\n\\n(1) I recommend ensuring the correct application of \\\\citet and \\\\citep.\\n\\n\\n(2) Missing periods in Line 299, Line 396, and Line 427.\", \"questions\": \"As in the weaknesses.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Diversity and utility\", \"comment\": \"Thank you for the review. As for the question: \\\"In practice, does increasing diversity ever harm utility?\\\"\\n\\nIn practice, the temperature parameter in language generation models is tuned to balance the tradeoff between diversity and accuracy. Lower temperatures limit diversity, which often helps with focused factual responses but also tends to reduce creativity. Note that the tuning in current large models allow for significant diversity -- 2-7 bits of entropy per token. \\n \\nHowever, this discussion is beyond the scope of our work! Our goal is to preserve the diversity that is already present in the teacher models, which are already tuned to have the \\\"right level\\\" of diversity. Methods that incur privacy-diversity tradeoff, like cold PATE, often must suppress diversity for utility. Hot PATE allows for diversity to be transferred with no privacy loss by producing correlated vote histograms via shared randomness. \\n\\nFor instance, consider the prompt \\u201cI like my ***\\u201d in a conversation about pets that follows examples in a private context (say each teacher sees data of 10 different families from a town that has only dogs, cats, or frogs for pets). In this case we do want the model to randomly choose from these 3 good answers. Suppose most teacher models assign a probability of \\u2153 to each, as expected. Now, with Cold PATE, the histogram would have 1/3 of teachers' votes for each pet type. With hot PATE, the histogram would have nearly all votes for the same option, depending on the shared randomness, also with probability 1/3 to each. This means that with hot PATE it suffices to use X3 the noise scale than with cold PATE. Again, this diversity is something that is present in an already-tuned model. It is not something we add. Hot PATE just allows for better utility-privacy tradeoff when transferring it.\"}", "{\"comment\": \"Thank you for your detailed response. I understand that your contributions are primarily theoretical, supported by an illustrative example. Based on this, I have raised my score to 6.\\n\\nI also agree with Reviewer FQD1 that showing the practical benefits of your method through empirical evaluation on datasets used in prior works like [1,2] would greatly strengthen the paper.\"}", "{\"title\": \"Response to the review\", \"comment\": \"Thank you for your review! We plan to upload a revised version soon. Here is our response to questions:\\n\\n## Question 1. \\n\\nYou are referring to accuracy on very specific benchmarks or setups. We believe that such an empirical evaluation is not relevant or needed here. Let us explain:\\n\\nThe application we consider is sequential text generation via PATE, and our method, of coordinated ensembles, offers a very significant improvement. This is established mathematically, and an order of magnitude improvement is demonstrated empirically over the baseline of independent ensembles. For a given privacy budget, the distribution of the hot PATE generated text is much closer to the original teachers distributions than that of cold PATE generated text. This is the ultimate metric for the problem we are out to solve. It is not about a specific task or set of tasks. The improvement factors in whenever the response is diverse (higher entropy, multiple good answers). The baseline of independent ensembles is the method used in [1] and some other works including [2] that you mention. \\n\\nThe relation of the related works [1], [3] is explained in Appendix A. These works used PATE (sample an aggregate) in a different way. The relevance of [3] is that they acknowledged diversity and the perceived tradeoff with utility and used clustering in the semantic space to reduce it. Our approach shows that it may not be necessary to do so as inherently there is no tradeoff. In any case, it is not a pure PATE sequential text generation approach.\\n\\nAs for [2] (concurrent work) \\u2013 the version we will upload adds a citation to [2]. Thank you! It appears that the method of [2] is to limit diversity by trimming to top-k tokens from each teacher and taking a uniform distribution over these tokens. This appears to be an attempt to limit diversity in order to counter the cold PATE tradeoff we demonstrate. But this is inferior to coordinating the ensemble as it fixes k and does not distinguish among the top-k. It also modifies the distributions and makes them further from that of the tuned LLM, which we believe is less desirable. Moreover, observe that our method of coordination would improve utility vs privacy even with the modified teacher distributions used in [2]. Again, it is provably better with coordination, and that component in their work can be replaced by our method of coordinating the ensemble. There is no need for experiments to validate this because it is provably more effective: Their approach satisfied definition 1 only in very limited settings.\\n\\nFinally, again, we did not build a system that competes with other systems. Nor do we claim to. We propose a method that is mathematically established to be used with all such systems and include a demo to showcase its benefits. It can be plugged in with any existing or future system that applies PATE for sequential text generation. The existence of all these related/prior/concurrent/future works on using PATE with prompts only shows that our work is relevant. \\n\\n## Question 2\\n\\nA version of our work (that presented the mathematical framework with motivation, but no Llama experiments) was made public around the same time or before. Therefore, these works are concurrent and not prior. Regardless, none of these other works proposed coordinated ensembles and in some cases, the works could benefit from it. So they do not subsume our work.\\n\\n## Question 3\\n\\nWe will include a review of [2] in the revised version and thank you for pointing it out. But as explained, the experimental evaluation on the particular setup/platform is not needed as the benefits of coordinated ensembles are established mathematically. The revised version will explain this (see answer to question 1).\\n\\n## Question 4\\n\\nThe revised version will include the missing periods in displayed equations and correct applications of \\\\citet/p. Thank you!\"}", "{\"comment\": \"Thank you for raising your score and for your thoughtful feedback! We deeply appreciate your recognition of the novelty and technical contributions of our work, as well as your constructive suggestions.\\n\\nReaching a broader audience is indeed a priority for us, and gatekeeping was never our intention. We are always eager to improve our communication. While we do not know the specific background or perspective of Reviewer FQD1, we are open to further suggestions on how we can better convey the high-level ideas to inspire and engage a wider readership.\\n\\nIf you have any additional recommendations on how to make the paper more accessible or clarify its practical implications, we would be delighted to implement them. Thank you again for your valuable input and encouragement!\"}", "{\"comment\": \"Thank you for raising the score and lowering the confidence level\\u2014we sincerely appreciate your effort and the thoughtful suggestions. While some of your comments still reflect certain misunderstandings, we value your acknowledgment in deferring the decision to reviewers who fully grasp the contribution.\\n\\nFully appreciating our contribution, particularly in the context of sequential text generation, requires familiarity with LLMs, a solid understanding of PATE and differential privacy theory, and a significant degree of mathematical background and sophistication. This level of expertise is understandably more demanding than what many researchers in empirical privacy may possess, making the review process more challenging. Our long introduction was intentionally designed to make the ideas and motivation accessible. However, without the necessary background, readers may need to pause, reflect, and re-read certain parts to grasp the ideas. The numerous papers published over the past year applying PATE with prompts and diversity\\u2014while missing the key issues we address\\u2014underscore that the ideas are far from obvious.\\n\\nAs for the suggestions. Thank you again for your time and effort.\\n\\n-- \\\"multi-task benchmark\\\": We have included an empirical demonstration of the benefits of our method. Given that we propose an improved aggregation method that delivers measurable benefits per generated token\\u2014validated mathematically\\u2014we do not believe it is necessary to use \\\"multi-task benchmarks for GenAI\\\" as part of our evaluation. The focus of our work is on the aggregation method itself, rather than a broad system-level benchmark.\\n\\n-- Our original manuscript assumed familiarity with differential privacy, allowing us to focus on introducing novel ideas and contributions that are less familiar to privacy researchers. However, as per your and other reviewers suggestion, the revised version now includes a more explicit discussion of the privacy properties in the main text to address potential gaps in understanding.\\n\\n-- \\\"Taking a Page\\\" from Papernot et al. (2018):\\nWhile we appreciate the suggestion, we see limited value in reciting Papernot (2017, 2018). We do state that their privacy-preserving aggregation techniques can be seamlessly \\\"plugged in\\\" to our method. This aspect is not the novelty of our work. Furthermore, we provide an alternative, data-dependent privacy analysis in Appendix E.\"}", "{\"comment\": \"Thank you!\"}", "{\"title\": \"Revised version is uploaded\", \"comment\": \"Thank you again for your review. As explained in our response, the review showed significant misunderstanding of our results. Our revised version is an attempt to increase accessibility of our work and we hope it will be helpful. Please let us know if you have further questions.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"title\": \"Thank you for the rebuttal\", \"comment\": \"I have read the rebuttal and the updated sections of the paper (in particular section 4.2, which brings back the privacy analysis) and have gone through other reviews and rebuttal responses.\\n\\nI think section 4.2 helps to clarify the privacy guarantee. I think the paper would benefit from bringing in data-dependent privacy analysis in the appendix in the form of an informally stated theorem; given that's really where PATE shines. \\n\\nI think the rebuttal does not alleviate the rest of my concerns regarding exposition and empirical results. I see that similar concerns have been raised by other reviewers. Regarding the latter, I get the author's argument that they are seeking to implement a general algorithm that is more useful for general-purpose GenAI tasks. They take the \\\"diversity\\\" of the tokens generated as a proxy measure for such general utility. This might as well be true, but would it not be better to clearly demonstrate these benefits in concrete scenarios?\\n\\nFinally, regarding the writing and exposition, I want to say that I am very squarely in the audience of your paper. If I had to read your introduction 3 times then I do not believe the paper, as it stands, is ready for publication. \\n\\nAs for concrete suggestions, here are mine:\\n- Significantly shorten the introduction\\n- Focus on the application scenario and demonstrate improvement, for example, over established multi-task benchmarks for GenAI\\n- Emphasize the privacy analysis\\n- Revamp the empirical section with concrete applications. Show how improved diversity helps these applications.\\n- As for theory, take a page from Papernot et. al 2018 and provide informal theorems if theory is too much for the main paper\\n\\nGiven the updated privacy analysis and the rebuttal responses, I have updated my score to 6 and reduced my confidence to 2. I will not be increasing my score any further.\"}", "{\"title\": \"Uploaded revised version\", \"comment\": \"Thank you for your comments! We uploaded a revised version accordingly.\", \"we_made_an_effort_to_increase_accessibility_and_implemented_the_following\": \"We reorganized and expanded sections 3 and 4 in the main text. Section 3 presents our formalization of an aggregation of distribution that preserves diversity. Section 4 contains our proposed coordinated ensembles and breaks down the presentation by including a separate subsection on establishing the privacy properties. \\nWe added Theorem 1 as a formal statement that connects the properties of coordinated ensembles in the lemmas in section 4 to Definition 1. The Theorem states that it suffices to only take high-count tokens in the coordinated histograms to preserve diversity.\\n\\nWe added a separate subsection in Section 4 on the privacy properties. In particular, we added \\u201cobservation 1\\u201d and \\u201ccorollary 1\\u201d that highlight the fact that the privacy (sensitivity to a change in a single user) properties of the histograms generated by coordinated ensembles are identical to those generated by independent ensembles \\u201ccold PATE\\u201d. We highlighted that the gain of coordinated ensembles is due to the shape of the histograms, since high-count tokens can be reported with a lower privacy cost. \\n\\nWe added a definition of $(\\\\varepsilon,\\\\delta)$-differential privacy before the analysis in Appendix D and E that explicitly uses this divergence in the TCT framework (an extension of sparse vector technique with which we evaluate data dependent analysis of composition cost). As explained, our main contribution and the established benefits of our proposed method are not specific to a particular privacy definition.\\n\\nThe main text now exceeds the page limit, but we plan to address it by moving some of section 4.1 (proofs) to the Appendix but we left it in place for now to facilitate an easier comparison of the versions.\\n\\nPlease let us know if the revision addresses your concerns and if you have additional questions.\"}", "{\"title\": \"Beginner-friendliness\", \"comment\": \"### Reviewer's concern:\\n\\\"The paper is not beginner-friendly and seems to assume a reader who is already very familar with DP, PATE and LLMs. In fairness, this probably is going to be the chief audience of this paper, but at the same time I find it somewhat egregious that differential privacy is never formally defined (even if the definition has to be deferred to the appendix due to space constraints).\\\"\\n\\n### Response\\n\\nWe appreciate the feedback and understand the need to make the paper more accessible. As the reviewer noted, there is a natural tradeoff in balancing technical depth with approachability for less familiar readers. The presentation in the main body is already an attempt to present concepts at a higher level.\\n\\nAs for the absent definition of differential privacy. We agree that we should include it in the appendix, especially with parts of our considerations focused on $(\\\\varepsilon,\\\\delta)$-DP. It\\u2019s worth noting that our contributions and improvements are applicable across DP models (different divergences) and apply with $(\\\\varepsilon,\\\\delta)$ privacy but also with concentrated differential privacy (CDP and zCDP), and R\\u00e9nyi differential privacy (RDP). Moreover, it also applies when the goal is not privacy but robustness to \\u201coutlier\\\" examples in the training data. This flexibility is why we initially chose to avoid a precise definition in the main text and consider metrics that transcend a particular definition.\"}", "{\"summary\": \"Private Aggregation of Teacher Ensembles (PATE) was designed for classification-like tasks where each datapoint has a single ground-truth label. For \\u201cdiverse\\\" tasks such as sequential text generation, the responses might instead be distributions. But there is a tension between diversity and privacy: diversity in the responses reduces agreement among the teachers, which in turn requires a smaller noise scale and less privacy. This paper proposes \\u201chot PATE\\u201d which allows for higher diversity in the responses without increasing the privacy cost.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"4\", \"strengths\": [\"I think this paper has a significant contribution \\u2014 via a carefully designed aggregation method, PATE can now thrive in a broader and more modern setting. Formalizing the notion of \\u201cdiversity-preserving\\u201d (Definition 1) is also a helpful contribution.\", \"The PATE framework can now be applied to very fashionable problems such as in-context learning.\"], \"weaknesses\": [\"The paper is not beginner-friendly and seems to assume a reader who is already very familar with DP, PATE and LLMs. In fairness, this probably is going to be the chief audience of this paper, but at the same time I find it somewhat egregious that differential privacy is never formally defined (even if the definition has to be deferred to the appendix due to space constraints).\", \"I felt that the privacy guarantees are not rigorously stated, DP implementations are largely left as poorly-described black boxes (e.g., NoisyArgMax in Algorithm 2 is never formally introduced) and none of the algorithms include the privacy parameters as input. I didn't see a formal privacy analysis that can be easily verified, and in terms of reproducibility I feel like the algorithms can\\u2019t really be implemented without knowing, for example, how to calibrate the noise scale.\"], \"questions\": [\"Besides coverage and diversity, are there other metrics which could be used to demonstrate the effectiveness of hot PATE?\", \"Line 274: If I\\u2019ve understood correctly, \\u201cthe noise scale must satisfy $\\\\sigma << \\\\arg \\\\max_j c_j$\\u201d is a requirement on the utility, and not the privacy? It might be helpful to explain this more thoroughly.\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you for your time and for raising the score! Regarding the comparison, it appears that [1] employed curated prompts specifically designed to elicit a single dominant response, rather than enabling free-form text generation. Consequently, we expect only minimal improvement in that setting, as it inherently involves very little diversity.\\n\\nAs for [2], as mentioned, we anticipate that replacing their approach with coordinated ensembles would result in significant improvements and yield a privacy-preserving aggregate distribution that more closely aligns with the unmodified teacher distributions. However, the evaluation in [2], based on their GitHub page, seems to have relied on a deprecated feature of OpenAI's API (logprobs) that allowed access to the top 100 probabilities\\u2014providing high diversity. Since this feature is deprecated, we can not repeat the experiment and obtain these probability distributions. Without access to their actual collections of teacher distributions for the prompts (which do not appear to be included in their GitHub repository), we are unable to perform a direct comparison with their results. \\n\\nWe appreciate your thoughtful feedback and will continue to explore avenues to strengthen our evaluation. Thank you once again!\"}", "{\"summary\": \"This paper introduces Hot PATE, an extension of the PATE (Private Aggregation of Teacher Ensembles) framework, to settings where output diversity is important. PATE works by partitioning the data and training a teacher model on each partition. Then, for a given model input, the each teacher model \\\"votes\\\" on a label, and a final label is privately sampled from the teacher histogram.\\n\\nThe key idea of Hot PATE is to preserve both privacy in the output label and the diversity of teacher distributions.\\u00a0The paper introduces the property of diversity preserving aggregation and introduces ensemble coordination as a technique to satisfy the property.\\u00a0Ensemble coordination strategically introduces correlation between teacher votes to ensure that rare tokens are transferred with high privacy noise, effectively mitigating the privacy penalty associated with high diversity, due to private sampling.\\n\\nThe authors provide an empirical demonstration of this approach in the context of in-context learning and show that Hot PATE yields greater diversity in output tokens.\", \"soundness\": \"3\", \"presentation\": \"4\", \"contribution\": \"3\", \"strengths\": [\"Introduces an extension of PATE that overcomes the diversity-privacy tradeoff\", \"Motivates the analysis through the notion of diversity preserving aggregation\", \"Connects proposed method with existing statistics literature: coordinated sampling\", \"Paper reads well, particularly with comparisons between hot and cold PATE\"], \"weaknesses\": [\"The empirical analysis is more along the lines of a proof-of-concept rather than a thorough comparison. The paper would benefit from more systematic experiments between hot and cold PATE.\", \"No discussion of the limitations of the proposed methods. See question 1.\"], \"questions\": \"1. In practice, does increasing diversity ever harm utility?\", \"other_notes\": [\"Typo on Line 93: \\\"...include component that...\\\"\", \"Typo on Line 190: \\\"...two use scenarios of applications...\\\"\", \"Typo on Line 323: \\\"A tokens j that...\\\"\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Reviewer expertise mismatch? part 1 of response\", \"comment\": \"Thank you for your time and feedback. We recognize that our paper presents a specialized technical contribution, which may require relevant background to fully appreciate. Regarding your last question, \\\"Have I misunderstood part of your work,\\\" the review does reflect a misunderstanding of key aspects of our approach\\u2014more than just a \\u201cpart\\u201d of the work. Given this, the confidence score of \\\"3\\\" seems higher than warranted by the level of understanding reflected in the review.\\n\\nBelow (see multiple responses) we specifically address the claims made in the review, point to the misunderstandings, and answer the questions asked.\\n\\n## \\\"Weakness\\\" 1:\\n\\u201c A robust privacy analysis is missing. The paper introduces a particular histogram aggregation strategy that produces rate token frequency. In a sense, this is not an aggregation that produces a single vote but rather a transformed histogram. Overall, I found the presentation of this rather simple idea overly complicated in Section 3. However, the key issue here is not the contrived procedure and Definition 1, but rather the complete lack of privacy analysis under this new aggregation method. Let me clarify this point: the PATE privacy analysis only holds under the noisy argmax release. In particualr, the analysis is a function of the gap between the top vote and the second top vote of the histogram. If we were to use Def.1 and instead release transformed vote count (for the purposes of diversity), we are strictly releasing more information. In fact, since the rare token frequencies are kept (for diversity purposes), such a scheme will likely have higher privacy cost than releasing a full noised histogram of votes.\\u201d\\n\\n### Response\\n\\nThe first major misunderstanding here is that we do not release the histogram. We release a single token each time -- just like \\u201ccold\\u201d (standard) PATE. This is explained in Section 2 that introduces the framework for Hot PATE and also illustrated in Figure 4. We are happy to get constructive suggestions on what led to this misunderstanding of our text and figure. \\n\\nThe heart of our method is the way the histogram is sampled (with correlated teacher votes) in Algorithm 1. You dismiss it as \\u201ccontrived,\\u201d which is a poor choice of an adjective.\\n\\nDefinition 1 is not about privacy at all. It is a formal definition of what it means for an aggregate distribution to transfer the diversity of a collection of distributions. This is our \\\"utility.\\\" We then propose a way to do this in a privacy preserving way (via ensemble coordination). The aggregate distribution corresponds to that of the output of the probabilistic end-to-end aggregation process that produces a single token. \\n\\nYou claim that a \\u201crobust privacy analysis is missing.\\u201d In fact, since what we change is the way the histogram is produced, and each teacher contributes a single vote, we can simply plug in the privacy analysis of cold PATE. We do state this clearly (see lines 185-187 \\u2013 the fifth paragraph of page 4). Additionally, we consider in Appendix D and E additional regimes (heterogeneous ensembles) and additional privacy analysis methods (based on TCT) but the high order bit is that you can simply apply standard PATE aggregation. Again, the primary benefit of Hot PATE, at an intuitive level, is that the histograms are much more favorable, since there tends to be a very high count to a single token even with very diverse distributions. This is all made very precise mathematically.\\n\\nAs for \\u201cthe gap\\u201d, in fact, with PATE, if there are multiple \\u201csimilar\\u201d tokens we are ok with releasing either one. Typically the parameters are not set as to separate the highest and second highest when the difference is small, especially with diversity, but so that we can release one of top votes and not release tokens with no votes or very low vote counts. \\n\\nWe are not sure we addressed all the misunderstandings in \\\"Weakness 1\\\", but please read our response and we are happy to clarify further.\"}" ] }

[ "Juan Luis Gastaldi", "John Terilla", "Luca Malagutti", "Brian DuSell", "Tim Vieira", "Ryan Cotterell" ]

[ "Tokenization", "Language Models", "Consistency", "NLP", "Theoretical Foundations", "Stochastic Maps", "Category Theory" ]

{ "note_id": [ "pTRX1r5VtQ", "mtuCyKHmKB", "mGyUM7cIdu", "gJfoUrWiTr", "f9mW6rQEik", "bzLg0e4iV9", "YZ4v8wpZED", "Ru6upsZCE2", "QzVrW9H8FM", "J7b0s2kVCz", "9mGyWG0Tdo", "7yHOHx0XUj", "5vUwg6g8Xv", "4EPmPlpLgN", "2Hn03CkKgz" ], "note_type": [ "official_comment", "official_review", "official_review", "official_comment", "official_comment", "official_review", "meta_review", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "decision" ], "note_created": [ 1732376805567, 1730467392904, 1731010551462, 1732357080315, 1732377202817, 1729776626763, 1734921048518, 1732356610660, 1732356222530, 1732657295077, 1730457686777, 1732356727739, 1732357213455, 1732639465177, 1737524062954 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission10574/Reviewer_btSn" ], [ "ICLR.cc/2025/Conference/Submission10574/Reviewer_pg9p" ], [ "ICLR.cc/2025/Conference/Submission10574/Reviewer_yPjZ" ], [ "ICLR.cc/2025/Conference/Submission10574/Authors" ], [ "ICLR.cc/2025/Conference/Submission10574/Authors" ], [ "ICLR.cc/2025/Conference/Submission10574/Reviewer_btSn" ], [ "ICLR.cc/2025/Conference/Submission10574/Area_Chair_9xuA" ], [ "ICLR.cc/2025/Conference/Submission10574/Authors" ], [ "ICLR.cc/2025/Conference/Submission10574/Authors" ], [ "ICLR.cc/2025/Conference/Submission10574/Authors" ], [ "ICLR.cc/2025/Conference/Submission10574/Reviewer_rVcs" ], [ "ICLR.cc/2025/Conference/Submission10574/Authors" ], [ "ICLR.cc/2025/Conference/Submission10574/Authors" ], [ "ICLR.cc/2025/Conference/Submission10574/Reviewer_yPjZ" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ] ], "structured_content_str": [ "{\"comment\": \"I appreciate the authors' feedback which gives further clarifications on the concerns I have. While I agree that the theoretic research that this paper follows is interesting, the general theorem (3.1) does not relate to tokenization in any particular way and is a general math derivations. The paper would benefit from further content/materials/empirical analysis and I personally think it is a pity that the authors choose to split such things out of the content of this paper, rending the current paper below acceptance criteria.\"}", "{\"summary\": \"This paper presents a formalization of the tokenization process in modern language models. The formalization is based on the category of stochastic maps, which presents a novel and unified framework for representing and analyzing tokenizers. The paper also provides some examples of sing the framework to account for tokenizer behavior such as ambiguity and inconsistencies.\", \"soundness\": \"4\", \"presentation\": \"4\", \"contribution\": \"3\", \"strengths\": \"The paper presents a novel framework for representing and analyzing tokenizers in the form of stochastic maps. The proposed framework has the potential to provide more a foundational statistical understanding of tokenizer behavior, which may lead to practical improvements.\", \"weaknesses\": \"The paper provides some examples of how the framework can be utilized to shed new light on tokenizer behavior such as ambiguity and inconsistency, but it is not immediately clear to what extent the proposed framework can lead to practical improvements of tokenizer performance.\", \"questions\": \"Would it be possible to provide some concrete examples of how the framework can be used to shed new light on differences between different types of tokenizers?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"Disclaimer: I previously reviewed an earlier version of this paper during its submission to NeurIPS 2024.\\n\\nThis paper proposes a formal approach to tokenization. Specifically, the authors apply the notion of stochastic maps (Baez & Fritz, 2014) to tokenization models, which allows them to formalize key properties of tokenizers such as their consistency. In a next step, they use their framework to analyze further statistical and computational aspects of tokenization (e.g., spurious ambiguity).\", \"soundness\": \"3\", \"presentation\": \"4\", \"contribution\": \"3\", \"strengths\": [\"The paper has several strengths:\", \"Tokenization is a critical aspect of modern-day natural language processing, but its theoretical underpinnings are not yet fully understood. The formalisms introduced in the paper help close this gap and might become the basis for future work.\", \"The application of stochastic maps to tokenization is novel.\", \"The presentation is excellent; the writing is clear and overall easy to follow.\"], \"weaknesses\": \"This is a completely theoretical paper without any empirical evaluation. While not a weakness _per se_, the authors mention that their findings have implications for the practical use of tokenizers (e.g., line 95). Compared to the version of the paper that was under submission at NeurIPS 2024, the authors have added sections discussing practical aspects of their observations (e.g., lines 345-350), but I still believe that a proper case study showcasing the practical value of the proposed formal framework would greatly enhance the potential impact of the paper. This is particularly desirable since many of the discussed problems are well-known in the community, so it is not clear what exactly the paper adds beyond a new theoretical perspective.\", \"questions\": [\"Comments:\", \"While you mention the relation between tokenization and linguistic segmentation, you should give more credit to the line of work that has investigated this link over the last few years (e.g., [Bostrom & Durrett, 2020](https://aclanthology.org/2020.findings-emnlp.414/), [Hofmann et al., 2021](https://aclanthology.org/2021.acl-long.279/), [Gow-Smith et al., 2022](https://aclanthology.org/2022.emnlp-main.786/), [Beinborn & Pinter, 2023](https://aclanthology.org/2023.emnlp-main.272/)), which is also relevant for your discussion of linguistic ambiguities (lines 395-399).\", \"Lines 405-410: I think you should mention that the practical benefit of this operation has been shown to be negligible in most cases ([Chirkova et al., 2023](https://aclanthology.org/2023.acl-short.1/)).\", \"I would recommend putting the examples from the appendix back into the main paper, especially Example 8.1. You have a bit of space left, so this should be feasible, and it would help illustrate your arguments.\"], \"typos\": [\"Line 120: \\\"distributions\\\" -> \\\"distribution\\\"\", \"Line 158: \\\"equation equation 1\\\" -> \\\"equation 1\\\"\", \"Line 205: \\\"this\\\" -> \\\"these\\\"\", \"Line 459: \\\"complexity a tokenizer model\\\" -> \\\"complexity of a tokenizer model\\\"\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you very much for your careful evaluation and remarks. We appreciate your comments on the paper\\u2019s strengths.\\n\\nWe are also grateful for pointing out that the connection between Theorem 3.1 and the remaining sections of the paper is not clear. The connection is, however, essential in our view. As stated at the beginning of section 4, while consistency is usually assumed, the necessary conditions for consistency formally established in th 3.1 are not usually met. The entire section is therefore a way to show that known problems commonly addressed from a practical or empirical perspective (e.g., OOV terms, special tokens, ambiguity, approximation errors, positivity of the softmax activation function, etc.) are in fact related to formal properties that violate the conditions formally established in th. 3.1. These conditions are precisely those related to injectivity and surjectivity of $\\\\tau$ and $\\\\kappa$. If section 4 discusses injectivity and surjectivity, it is because these are formal consequences of exactness, which is, in turn, an unconditioned form to obtain consistency (Prop. 3.1). So, if the discussed conditions on injectivity and surjectivity are not met, consistency is not guaranteed. As for section 5, the idea is that while consistency is not everything that matters, the use of the formal properties of $\\\\tau$ and $\\\\kappa$ laid down to establish its necessary and sufficient conditions can be further used to analyze and guarantee computational requirements. We hope this clarification can help to see the connection more clearly. If you have any suggestions on how this can become more clear in the paper, we would be happy to include them in a final version.\\n\\nWe agree that this paper does not provide any empirical results concerning the connection between these formal properties and their practical consequences. We also agree that those results would significantly contribute to enhancing the paper\\u2019s impact, even if we believe that the purely theoretical character of the paper shouldn\\u2019t be a weakness *per se* in the ICLR area of \\u201cLearning theory\\u201d where it has been submitted. On this point, we would like to provide some context. This paper constitutes the first step in a broader research program aiming at addressing tokenization problems from a formal principled perspective, including various empirical results on practical issues. Given the strict page limit of ML and NLP conferences, it was impossible for us to include all the aspects and results of this program in one paper. Therefore, we split up the work over multiple papers, distributing their content as consistently as we could. Concretely, we have used the formalism introduced in this paper to design algorithms for converting token-level autoregressive language models to character-level ones, present both exact and approximate algorithms, and benchmark their practical runtime and approximation quality. The framework proposed here (especially, the principles discussed in the last section) was crucial to identify a minimal subset of $\\\\Delta^*$ over which an exact, tractable algorithm could be defined, as well as a consistent notion of approximation, all of which uses injective, surjective, multiplicative, and kernel properties of the maps $\\\\tau,\\\\kappa$. Direct practical applications of this work include: principled solutions to the prompt boundary problem, constrained generation expressed over characters, or consistent linguistic and psycholinguistic measures. Concerning this last point, for instance, we have already performed an analysis of the role of tokenizers in psycholinguistics, providing an exact way of computing the surprisal of regions of interest over characters from a token model to assess psycholinguistic measures of reading times (this work has been recently published). Currently, we are studying the preservation (or lack thereof) by tokenization of structural features induced by formal grammars over character and token models (work in progress). All these applications use the framework proposed here as their theoretical foundation. Accordingly, we thought that the most consistent way to disseminate the different steps in our program was to adopt a \\u201cdivision of labor\\u201d between papers, keeping this paper in its pure theoretical form without tying it to any one application, and envisaging its impact as part of the broader research program.\"}", "{\"title\": \"Clarification\", \"comment\": \"Thank for your feedback. Could we ask for a quick clarification on how Theorem 3.1 does not relate to tokenization in any particular way? It tells us when we can consistently estimate a language model under a tokenizer.\"}", "{\"summary\": \"The paper investigates into the tokenization process for language modelling from a theoretical viewpoint.\\n\\nThe framework examines language modelling as mapping between a true language model on an alphabet and a learned language model on a separate alphabet, known as a stochastic map. In section 3, the paper establishes convergence property for various distributions in different spaces, as well as necessary function properties (injectivity and surjectivity). Later on, the paper leverages such properties to examine the problems (e.g injectivity, surjectivity) of current language models and connect them with concrete issues such as inconsistency, ambiguity, tractability, etc.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": [\"The paper proposes an interesting and also promising theoretical direction to examine the tokenization problems: by explicitly examining the relationship between the true language model under its alphabet and the learned language model with the practical alphabet. Under this framework, the paper has shown some interesting preliimary results:\", \"Define propery the consistency from analysis viewpoint\", \"The propoerty of encoder and decoder and connect them with the problem of inconsistency and ambiguity\"], \"weaknesses\": [\"The connections between the derived properties/theorems and the NLP practice/problems are not strong. My arguments are stated as the following:\", \"In theorem 3.1 the authors derive the condition to make the estimator consistent; however, this property is nowhere more leveraged, the paper discusses in more detail the exact mappings and the derived injectivity/surjectivity properties which do not require the analysis from 3.1 and makes 3.1 a isolated construction without any implication built inside the paper.\", \"The injectivity (and surjectivity) discussions look more like discussions as a \\\"related work\\\" section. For example, no derivations are shown for problems such as \\\"how severe the problem is when injectivity assumption is broken\\\" or \\\"when injectivity are ensured by certain techniques, what is the advantage quantified/bounded mathematically\\\".\", \"While the boundness and tractability discussions are interesting, they are of observational nature. The paper will have significantly more impact if the paper is able to derive algorithms/improvements/theoretical implication out of them.\"], \"questions\": \"Can the authors further clairify the connection between 3.1 and the rest of the paper please?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"metareview\": \"This paper presents a foundational perspective on tokenization and established some formal properties of tokenization; the authors established these properties via the utilization of stochastic maps, which is novel.\", \"strengths\": [\"Very clear exposition.\", \"Novel work where there has been very little theoretical work on tokenization.\"], \"weaknesses\": [\"No empirical validation of the framework (which is not a huge issue in my opinion).\"], \"additional_comments_on_reviewer_discussion\": \"There has been significant discussion between the reviewers and authors where the latter made a significant effort in allaying concerns from reviewers and provided answers.\"}", "{\"comment\": \"Thank you very much for your careful evaluation and remarks. We appreciate your comments on the paper\\u2019s strengths.\\n\\nThanks also for your remark and question concerning the practical consequences of the framework proposed. This is indeed a completely theoretical paper, submitted in the principal area of \\u201cLearning theory\\u201d. We agree that associating the proposed theoretical framework with practical aspects of tokenization would significantly contribute to enhancing its impact. On this point, we would like to provide some context that might help explain our choices. This paper constitutes the first step in a broader research program aiming at addressing tokenization problems from a formal principled perspective, including various practical issues. Given the strict page limit of ML and NLP conferences, it was impossible for us to include all the aspects and results of this program in one paper. Therefore, we split up the work over multiple papers, distributing their content as consistently as we could. While improving the performance is indeed one possible practical outcome of our framework, we believe there are other practical outcomes. Concretely, we have used the formalism introduced in this paper to design algorithms for converting token-level autoregressive language models to character-level ones, present both exact and approximate algorithms, and benchmark their practical runtime and approximation quality. The framework proposed here (especially, the principles discussed in the last section) was crucial to identify a minimal subset of $\\\\Delta^*$ over which an exact, tractable algorithm could be defined, as well as a consistent notion of approximation, all of which uses injective, surjective, multiplicative, and kernel properties of the maps $\\\\tau,\\\\kappa$. Direct practical applications of this work include, among others, principled solutions to the prompt boundary problem, constrained generation expressed over characters, or consistent linguistic and psycholinguistic measures. Concerning this last point, for instance, we have already performed an analysis of the role of tokenizers in psycholinguistics, providing an exact way of computing the surprisal of regions of interest over characters from a token model to assess psycholinguistic measures of reading times (this work has been recently published). Currently, we are studying the preservation (or lack thereof) by tokenization of structural features induced by formal grammars over character and token models (work in progress). All these applications use the framework proposed here as their theoretical foundation. Accordingly, we thought that the most consistent way to disseminate the different steps in our program was to adopt a \\u201cdivision of labor\\u201d between papers, keeping this paper in its pure theoretical form without tying it to any one application, and envisaging its impact as part of the broader research program.\"}", "{\"comment\": \"Thank you very much for your careful evaluation and remarks. We appreciate your comments on the paper\\u2019s strengths.\\n\\nThis is indeed a completely theoretical paper, and we share the view that this should not constitute *a priori* a weakness, especially for a submission in the principal area of \\u201cLearning theory\\u201d as this one. We are also grateful for mentioning the improvements in this sense compared to our previous version. That being said, we agree that associating the proposed theoretical framework with practical aspects of tokenization would significantly contribute to enhancing its impact. On this point, we would like to provide some context that might help explain our choices. This paper constitutes the first step in a broader research program aiming at addressing tokenization problems from a formal principled perspective. Given the strict page limit of ML and NLP conferences, it was impossible for us to include all the aspects and results of this program in one paper. Therefore, we split up the work over multiple papers, distributing their content as consistently as we could. Concretely, we have used the formalism introduced in this paper to design algorithms for converting token-level autoregressive language models to character-level ones, present both exact and approximate algorithms, and benchmark their practical runtime and approximation quality. The framework proposed here (especially, the principles discussed in the last section) was crucial to identify a minimal subset of $\\\\Delta^*$ over which an exact, tractable algorithm could be defined, as well as a consistent notion of approximation, all of which uses injective, surjective, multiplicative, and kernel properties of the maps $\\\\tau,\\\\kappa$. Direct practical applications of this work include: principled solutions to the prompt boundary problem, constrained generation expressed over characters, or consistent linguistic and psycholinguistic measures. Concerning this last point, for instance, we have already performed an analysis of the role of tokenizers in psycholinguistics, providing an exact way of computing the surprisal of regions of interest over characters from a token model to assess psycholinguistic measures of reading times (this work has been recently published). Currently, we are studying the preservation (or lack thereof) by tokenization of structural features induced by formal grammars over character and token models (work in progress). All these applications use the framework proposed here as their theoretical foundation. Accordingly, we thought that the most consistent way to disseminate the different steps in our program was to adopt a \\u201cdivision of labor\\u201d between papers, keeping this paper in its pure theoretical form without tying it to any one application, and envisaging its impact as part of the broader research program.\\n\\nWe agree that more credit should be attributed to the work mentioned in the points suggested. We added remarks concerning the relation of tokenizers to linguistic segmentation with reference to relevant work suggested both in lines 54-55 and 400-402. Following your suggestion, we also added a remark on the limited practical benefits of marginalization shown by Chrikova et al, 2023 (line 388-389). We will be happy to include Ex 8.1 in the main text of the final version if space permits, as we agree that it can be very helpful for the reader. Finally, we thank you for pointing out the typos, which we have also corrected in the current version.\"}", "{\"comment\": \"We would like to sincerely thank the reviewer for considering our remarks and adjusting the score.\"}", "{\"summary\": \"The authors address the problems of tokenization and use of the token representations in NLP from a foundational mathematical perspective. Based on the category of stochastic maps, the authors propose a general definition of a tokenizer as an arbitrary pair of composable maps, and this theoretical framework enables to establish general conditions for a principled use of tokenizers and the necessary and sufficient conditions for a tokenizer model to preserve the consistency of statistical estimators. The statistical and computational concerns (inconsistency, ambiguity, tractability, and boundedness) are discussed. To achieve this, authors characterize these known issues through the lens of formal properties of composable maps, such as injectivity, multiplicativity, and finite decomposition.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"\\u2014 Theoretical position paper is from a foundational perspective, and mathematically justified analysis is introduced; strict definitions and notation conventions are presented.\", \"weaknesses\": \"It's still unclear how to apply this knowledge to practical issues and better tokenizers for LLMs.\", \"questions\": \"I am unsure if I am not a relevant reviewer for this paper or if the paper is too foundational for the conference.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"1\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you very much for your evaluation and remarks. We appreciate your comments on the paper\\u2019s strengths.\\n\\nWe would like to provide some context concerning the foundational character of this paper and its relation to practical issues. This is indeed a completely theoretical paper, submitted in the principal ICLR area of \\u201cLearning theory\\u201d, for which we believe the paper is a good match. We think that the purely theoretical character of a paper shouldn\\u2019t be a weakness *per se* in this area. We agree that associating the proposed theoretical framework with practical aspects of tokenization would significantly contribute to enhancing its impact. In this regard, we would like to mention that this paper constitutes the first step in a broader research program aiming at addressing tokenization problems from a formal, principled perspective, including various practical issues. Given the strict page limit of ML and NLP conferences, it was impossible for us to include all the aspects and results of this program in one paper. Therefore, we split up the work over multiple papers, distributing their content as consistently as we could. Concretely, we have used the formalism introduced in this paper to design algorithms for converting token-level autoregressive language models to character-level ones, present both exact and approximate algorithms, and benchmark their practical runtime and approximation quality. The framework proposed here (especially, the principles discussed in the last section) was crucial to identify a minimal subset of $\\\\Delta^*$ over which an exact, tractable algorithm could be defined, as well as a consistent notion of approximation, all of which uses injective, surjective, multiplicative, and kernel properties of the maps $\\\\tau,\\\\kappa$. Direct practical applications of this work include: principled solutions to the prompt boundary problem, constrained generation expressed over characters, or consistent linguistic and psycholinguistic measures. Concerning this last point, for instance, we have already performed an analysis of the role of tokenizers in psycholinguistics, providing an exact way of computing the surprisal of regions of interest over characters from a token model to assess psycholinguistic measures of reading times (this work has been recently published). Currently, we are studying the preservation (or lack thereof) by tokenization of structural features induced by formal grammars over character and token models (work in progress). All these applications use the framework proposed here as their theoretical foundation. Accordingly, we thought that the most consistent way to disseminate the different steps in our program was to adopt a \\u201cdivision of labor\\u201d between papers, keeping this paper in its pure theoretical form without tying it to any one application, and envisaging its impact as part of the broader research program.\"}", "{\"comment\": \"We have also remarked that, among the strengths, you mention that \\u201cthe presentation is excellent\\u201d, while the score you gave for the presentation is \\u201c3: good\\u201d. If you don\\u2019t mind us asking, would you agree to update the presentation score (and, if applicable, the final rating) to match the text of your evaluation?\"}", "{\"title\": \"Reviewer Response\", \"comment\": \"I thank the authors for their explanations and especially the addition of the relevant work on linguistic segmentation. I have increased my score.\\n\\nThe original score of \\\"3: good\\\" for presentation was reflective of the fact that this category is supposed to include contextualization relative to prior work (\\\"Please assign the paper a numerical rating on the following scale to indicate the quality of the presentation. This should take into account the writing style and clarity, as well as contextualization relative to prior work.\\\"), and I had reservations about this point. However, since this has been changed in the updated version of the paper, I have increased this score as well.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}" ] }

[ "Chia-Ying Hsieh", "Chun-Shu Wei" ]

[ "Explainable AI", "Post-hoc explanation", "EEG", "Feature attribution method", "Saliency map", "in-distribution imputation" ]

{ "note_id": [ "ylXbpWZZmN", "rbUEyLIR4R", "oKiSxtf8ky", "iOm5d1kqxq", "gcVvkXPozj", "cjNLzzEqum", "bSQFinKSK0", "WC0kmGoG0c", "UljoFugw0n", "TfmCf9iSIU", "B1q1ZBfHMk", "Ak49cDsOR2", "7e5MLrnk5q", "2siqmQzLSb", "2XGEuY98vq" ], "note_type": [ "official_comment", "official_comment", "comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment" ], "note_created": [ 1733187296792, 1733187224244, 1737028605326, 1733105586844, 1732716351545, 1732923821432, 1732716996899, 1730317944996, 1730635520254, 1732716283104, 1730657658066, 1732923677703, 1733187286844, 1733105704284, 1732704375699 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission12578/Authors" ], [ "ICLR.cc/2025/Conference/Submission12578/Authors" ], [ "ICLR.cc/2025/Conference/Submission12578/Authors" ], [ "ICLR.cc/2025/Conference/Submission12578/Authors" ], [ "ICLR.cc/2025/Conference/Submission12578/Authors" ], [ "ICLR.cc/2025/Conference/Submission12578/Reviewer_kAcF" ], [ "ICLR.cc/2025/Conference/Submission12578/Authors" ], [ "ICLR.cc/2025/Conference/Submission12578/Reviewer_cFhK" ], [ "ICLR.cc/2025/Conference/Submission12578/Reviewer_YJwj" ], [ "ICLR.cc/2025/Conference/Submission12578/Authors" ], [ "ICLR.cc/2025/Conference/Submission12578/Reviewer_kAcF" ], [ "ICLR.cc/2025/Conference/Submission12578/Reviewer_kAcF" ], [ "ICLR.cc/2025/Conference/Submission12578/Authors" ], [ "ICLR.cc/2025/Conference/Submission12578/Authors" ], [ "ICLR.cc/2025/Conference/Submission12578/Reviewer_YJwj" ] ], "structured_content_str": [ "{\"comment\": \"We would be grateful for your feedback on the revised manuscript. Please let us know if there are any remaining concerns or if further clarification is needed.\"}", "{\"comment\": \"Thank you, Reviewer kAcF, for your thoughtful comments and for advocating for the acceptance of our paper.\"}", "{\"withdrawal_confirmation\": \"I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors.\", \"comment\": \"The authors would like to express their sincere gratitude to the reviewers for their time and effort in evaluating our submission. Your constructive feedback has provided invaluable guidance for refining our manuscript. After careful consideration, we have decided to withdraw the paper to focus on further improving our work.\"}", "{\"comment\": \"Thank you once again for your valuable feedback and dedicated service as a reviewer. We have carefully considered your input and made significant updates to the manuscript in response. We genuinely value your perspective and encourage you to review our rebuttal and the revised manuscript. Given the positive feedback from other reviewers, we hope our revisions address your concerns and provide the clarity needed to reconsider your assessment. Please don not hesitate to reach out if further clarification or discussion would be helpful.\"}", "{\"title\": \"Response to Reviewer YJwj\", \"comment\": \"We sincerely thank the reviewer for the thoughtful review and constructive comments, and apologize for the delay in our rebuttal, as we aimed to provide the reviewer with a thoroughly refined version. We appreciate the valuable insights provided and would like to respond to your feedback as follows. Additionally, we have made major revisions based on the reviews, particularly to: 1) enhance the clarity of our study\\u2019s premise, 2) emphasize its contributions, and 3) revise the metric computation to facilitate a more direct analysis. We look forward to your continued evaluation of our work.\\n\\n> (Weaknesses) It is unclear to me what follows from the evaluation results. Shall we only use some of the methods which perform well in Table 2 for the analysis of EEG-based explanations? What are the consequences of the evaluation for the practitioner?\\n\\n> (Questions) What follows from the evaluation results? \\n\\nWe understand your interest in the practical implications of our evaluation results. Our aim is not to prescribe a definitive set of \\\"best\\\" or \\\"must-use\\\" explanation methods for EEG deep learning (EEG-DL) applications; rather, we propose a guiding framework for practitioners. This framework is designed to assist practitioners in selecting the explanation methods that are most appropriate for their specific datasets or in identifying methods that reliably highlight key features within certain domains.\\n\\nFor instance, we emphasize that practitioners should ensure the preservation of the sign of the utilized explanation method when visualizing frequency domain features. This precaution mitigates the risk of displaying confounding patterns. Additionally, our framework provides a systematic method for evaluating whether a specific explanation technique can accurately represent the intended feature domain. The insights garnered from this approach are crucial for practitioners to make informed decisions in real-world brain-computer interface (BCI) applications, thereby enhancing the interpretability of EEG-DL models in both academic research and practical contexts.\\n\\n> (Weaknesses) It is unclear to me what follows from the evaluation results. Are the results consistent with other evaluation approaches?\\n\\nWe acknowledge your concern regarding the consistency of our evaluation results with other evaluation methodologies. Our results obtained from the mdROAD and mdAR frameworks represent two distinct evaluation strategies, as illustrated in Figure 8. While these results may not be identical, the relative trends in faithfulness\\u2014both high and low\\u2014are discernible. We have elaborated on potential reasons for these discrepancies in Section 5.2. \\n\\nFurthermore, as indicated in Section 2.4, the existing literature related to EEG-DL demonstrates substantial room for improvement in terms of evaluation strategies and experimental materials. Consequently, we respectfully argue that strict alignment with prior research was not a requisite for the objectives of our study.\\n\\nWe hope these clarifications adequately address your concerns and enhance the overall impact of our work. Thank you again for your insightful feedback.\"}", "{\"title\": \"updated score\", \"comment\": \"30 Nov: raised the score from 6 to 8.\\nDid not raise my confidence due to lack of familiarity with EEG data.\"}", "{\"title\": \"Response to Reviewer cFhK\", \"comment\": \"We sincerely thank the reviewer for the thoughtful review and constructive comments, and apologize for the delay in our rebuttal, as we aimed to provide the reviewer with a thoroughly refined version. We appreciate the valuable insights provided and would like to respond to your feedback as follows. Additionally, we have made major revisions based on the reviews, particularly to: 1) enhance the clarity of our study\\u2019s premise, 2) emphasize its contributions, and 3) revise the metric computation to facilitate a more direct analysis. We look forward to your continued evaluation of our work.\\n\\n> (Weaknesses) Writing \\n\\nWe have conducted a thorough review of the manuscript to identify and address grammatical errors, clarity issues, and writing style deficiencies. Significant revisions have been made throughout the document to enhance overall readability. We invite you to review the revised version, where the modified sections are highlighted in red.\\n\\n> (Weaknesses) Give more context regarding the EEG experiments \\n\\nWe acknowledge the importance of providing additional context for the EEG experiments. While we aim to present a comprehensive understanding of our methodology and findings, we plan to include more detailed examples of attributions in the revised manuscript. We are considering the possibility of extracting the imputation function to make the explanation clearer or tabulating the target-imputation functions for each domain for easier comparison.\\n\\n> (Weaknesses) Cross-framework comparison \\n\\nWe completely agree that a comparison with other frameworks would be beneficial and could yield valuable insights. However, as our findings suggest, the evaluated faithfulness of each explanation method is closely tied to the specific characteristics of the dataset being analyzed. This variability makes direct comparisons challenging, but we recognize the potential for future research to delve deeper into this aspect.\\n\\n> (Weaknesses) It might be useful to briefly describe what the proposed metrics (AOC, ABC) conceptually mean and how they differ. \\n\\nWe appreciate your suggestion to improve clarity concerning the faithfulness metrics. In response, we have revised Section 4.3, adding the following text to the end of the paragraph on Lines 376-377: \\n- *Higher measurements indicate that the curves\\u2019 behavior aligns more closely with expectations, reflecting greater faithfulness. An illustrated example of the metrics is displayed in Figure 1b.*\\n\\n> (Weaknesses) Do the authors believe the proposed framework could be relevant for recent interests in large amounts of data, for example, resting state EEG? \\n\\nThank you for highlighting recent research that focuses on large volumes of non-task-related EEG data. Our information masking methods are designed to be input-specific, as they depend on feature attribution explanations for each specific input. As long as time-series-like saliency maps can be generated to represent all features in the input, our framework remains applicable. We are confident that our work will continue to play a crucial role in advancing the interpretability of EEG deep learning models, especially as they are applied to increasingly large and intricate datasets.\", \"references\": \"1. https://pmc.ncbi.nlm.nih.gov/articles/PMC8870584/ \\n2. https://www.mdpi.com/2079-9292/13/1/186 \\n3. https://journals.sagepub.com/doi/full/10.1177/15500594211063662 \\n\\n> (Question) Be more concise in your contributions\", \"we_have_summarized_our_contributions_more_concisely_as_follows\": [\"We expand the leading in-distribution information masking method, Remove and Debias, to accommodate multiple domains, including spatial, temporal, and spectral dimensions.\", \"We introduce an adversarial information masking (AIM) approach to circumvent issues related to hand-crafted distribution selection and to enhance in-distribution information masking for multivariate time series data.\", \"We assess the effectiveness of in-distribution information masking through a novel Multi-Domain Adversarial Robustness (mdAR) framework that includes new normalized faithfulness metrics and an evaluation result consistency-based methodology for framework validation.\", \"We demonstrate assessments of faithfulness for existing post-hoc explanation methods and their limitations under specific conditions in the context of deep learning interpretation of human EEG data.\", \"> (Question) Show more example attributions\", \"We acknowledge the benefit of providing additional examples of attributions. While we strive for a comprehensive presentation of our proposed framework, our primary focus in this work is on enhancing the information masking techniques and establishing a robust quantitative comparison framework.\", \"We sincerely hope that our revisions and responses adequately address your concerns and contribute to the clarity and impact of our work. Thank you once again for your valuable and constructive feedback.\"]}", "{\"summary\": \"The manuscript aims to contribute to the evaluation of interpretations of deep learning models applied to EEG. It does so by identifying issues with traditional adversarial robustness evaluation for EEG and proposing alternative information masking methods to evaluate the faithfulness of feature attribution methods. Their framework is able to differentiate between attribution methods on spatial, temporal, and spectral domains and thereby seem to generate useful findings for the field. Such progress is valuable as explainability of neural networks in EEG data is not well-studied.\", \"soundness\": \"2\", \"presentation\": \"1\", \"contribution\": \"2\", \"strengths\": [\"Addresses a real gap in the EEG/DL field\", \"Seems to generate useful findings\", \"Three EEG domains are considered\"], \"weaknesses\": \"1.Writing: Unfortunately, the level of the writing of the manuscript is poor. Especially the first half of the paper, outlining the motivations, prior literature, and outline of the paper, is difficult to follow. The paper could really benefit from another thorough round of editing as the many grammatical errors lead to semantic ambiguity. A few examples that I am unable to understand:\\na. L218-219: \\u2018ability to exert data distribution\\u2019\\nb. L220-221:\\u2018computationally exhaustive while remain biased or uncontrollable\\u2019.\\nc. L271-273: \\\"(\\u2026), whose value are concluded to reflect certain series trend.\\u201d\\nAlso the experiments are hard to follow and it is difficult to assess the contributions of this work. \\n\\n2. Evaluation: Is it possible to perform some form of cross-framework comparison? It is difficult to understand the advantage of the proposed framework over existing ones. For example, why did the authors choose not to analyze where and why frameworks agree or disagree? Would synthetic data enable a comparison between frameworks? Understanding both the advantages and disadvantages of this new framework would be very valuable.\", \"minor\": [\"It might be useful to briefly describe what the the proposed metrics (AOC, ABC) conceptionally mean and how they differ.\", \"Recent work on EEG-DL concerns the use of large amounts of resting state data for clinical predictions (e.g. [1-3]). Do the authors believe the proposed framework could be relevant for such work as well? Understandably, task-based explanations are easier to verify and interpret. However, the datasets used by the authors are small, while it may be argued that deep learning models may be particularly interesting in case of larger datasets, which to my knowledge tend to be resting-state.\"], \"references\": \"1.\\thttps://openreview.net/forum?id=QzTpTRVtrP\\n2.\\thttps://arxiv.org/abs/2305.10351\\n3.\\thttps://arxiv.org/abs/2409.07480\", \"questions\": [\"Improve clarity and style of writing\", \"Be more concise in your contributions\", \"Evaluate your framework with respect to other frameworks\", \"Give more context regarding the EEG experiments and show more example attributions\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The paper addresses the problem of evaluation of post-hoc explanations in the context of models trained on EEG decoding tasks.\\nIn particular, it focuses on the evaluation of faithfulness of explanations and proposes a novel framework involving multi-domain adversarial information masking (AIM) based on Multi-Domain Adversarial Robustness (mdAR), which overcomes some of the limitations of standard faithfulness evaluation approaches. The framework is validated on multiple model architectures and EEG datasets.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"The paper addresses a relevant problem, namely evaluation of post-hoc explanations. The paper is original in the sense that it proposes two imputation techniques specifically tailored for multivariate EEG data. The proposals are based on the ROAD and AR frameworks as the and carefully integrate the spatial, spectral and temporal dimension of multivariate EEG data. The overall originally and quality of the proposed approach is rather limited and specific to models trained for EEG analysis. It is unclear how to generalise the approach beyond this specific application domain.\\nThe paper is well written and easy to understand. The experimental evaluation is ok, but could be more detailed and deep. Currently it is not clear how follows, e.g., from the results in Table 2 or 3.\\nOverall, the contribution is rather incremental and will probably be of interest / significance only to a limited (EEG) community.\", \"weaknesses\": \"The contributions of the paper are very specific and may be of interest to a limited community, mainly only researchers training and explaining NN for EEG analysis. There has been a lot of research on faithfulness evaluation of explanations. The proposed method represents an incremental contribution to this field. The experimental evaluation is not 100% convincing. It is unclear to me what follows from the evaluation results. Shall we only use some of the methods which perform well in Table 2 for the analysis of EEG-based explanations? Are the results consistent with other evaluation approaches? What are the consequences of the evaluation for the practioner?.\\nCurrently, the paper reads to me as proposing yet another faithfulness evolution metric, here specifically for EEG analysis tasks. The overall originality of the contribution and relevance for the ICLR research community is rather limited. Therefore I recommend \\\"reject\\\".\", \"questions\": \"What follows from the evaluation results?\\nWhat are the consequences for the practitioner?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to Reviewer kAcF\", \"comment\": \"We sincerely thank the reviewer for the thoughtful review and constructive comments, and apologize for the delay in our rebuttal, as we aimed to provide the reviewer with a thoroughly refined version. We appreciate the valuable insights provided and would like to respond to your feedback as follows. Additionally, we have made major revisions based on the reviews, particularly to: 1) enhance the clarity of our study\\u2019s premise, 2) emphasize its contributions, and 3) revise the metric computation to facilitate a more direct analysis. We look forward to your continued evaluation of our work.\\n\\n\\n\\n> (Questions) Footnote in Table 2\\n\\nWe apologize for the oversight regarding the lack of description in Table 2. We have added a corresponding explanation in the last line of the table's caption: \\n- *Highlighted cells represent the \\\"most faithful\\\" method, and the superscripts indicate the top-3 highest faithfulness measurements within each column.*\\n\\n> (Questions) Add citation of relevant work https://www.nature.com/articles/s42256-023-00620-w \\n\\nWe are grateful to the reviewer for suggesting this relevant citation to enrich our survey. We have included the work recommended in the introduction of Section 2.4, specifically in Lines 201-202: \\n- *With the growing understanding of post-hoc explanations in computer vision, there has been a recent expansion into other fields exploring this topic (Turb\\u00e9 et al., 2023; Fang et al., 2024).*\\n\\n> (Weaknesses) In particular for the temporal domain, but also for the frequency domain, trying out different non-adversarial masking methods would make the evaluation more interesting. For the temporal domain, that would involve different stochastic processes. \\n \\nWe appreciate the reviewer's valuable suggestion regarding the exploration of non-adversarial masking methods. In response, we have conducted additional experiments utilizing various non-adversarial techniques in both the temporal and frequency domains. The results can be found in Appendix D.1.\\n\\nWe hope that these revisions address your comments satisfactorily and enhance the clarity and quality of our paper. Thank you once again for your constructive feedback.\"}", "{\"summary\": \"The authors present approaches on how to mask EEG input data in spatial, frequency and temporal domains. The aim of this masking is faithfulness evaluation of attribution maps. The present novel ideas for masking in the temporal domain, also with likely a novelty for the frequency domain. Besides conventional approximate in-distribution masking they also evaluate masking by copying from adversarially crafted samples. The present results for several networks and several attribution methods. They investigate the question whether the sign of attribution maps carries information and investigate an unexpected result in the frequency domain.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"It is a reasonable application study about faithfulness evaluation for a particular field - which is an acceptable type of invention. A good set of experiments in three domains, also for multiple networks. They measure also consistency in the sense of rank correlations.\", \"weaknesses\": \"In particular for the temporal domain, but also for the frequency domain, trying out different non-adversarial masking methods would make the evaluation more interesting. For the temporal domain that would be different stochastic processes.\", \"questions\": \"What are the footnotes in Table 2 ?\", \"it_might_be_fair_to_cite_https\": \"//www.nature.com/articles/s42256-023-00620-w .\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Reviewer kAcF - an argument for accepting this paper\", \"comment\": \"Disclaimer: I am not aware at the current time (end of Nov 2024) who could be the authors of this submission.\\n\\n@Authors: Thank you for adding the additional requested experiments. \\n\\nIn my view, the paper is a domain-specific analysis of faithfulness of different methods. Yes, it is incremental. Reporting more in-depth for existing measures.\\n\\nI have read the weaknesses from the other reviewers. \\n\\nAs for reviewer YJwj, \\nI would argue that having no strong conclusion is valid if several methods yield similar performances. Seeing no clear difference is a valid result in science. If we as a community would downvote that, it would risk to nudge the field towards sensationalistic reporting. \\n\\nYes, they could write a more clear conclusion, but I do not see this as justifying a reject (3) level. For me a level 3 reject is something with methodical flaws, very poor readability or lacking or simplistic experiments. \\nSame holds when adding the domain specificity to the above argument. DL for EEG is an active field in the sciences.\\n(side note, AUC, ABC, AOC are up to the model prediction on the original sample not novel metrics)\", \"as_for_reviewer_cfhk\": \"faithfulness is pretty much established in vision as a category of evaluation metrics for attribution methods, with various variants having been developed (Samek et al, 2017 \\\"Evaluating the visualization of what a\\nDeep Neural Network has learned\\\", Arya et al 2019 \\\"One Explanation Does Not Fit All...\\\", Alvarez-Melis, 2018 \\\"Towards Robust Interpretability with Self-Explaining Neural Networks\\\" and many more). Therefore, I also found the readability of this paper okay.\\n\\nOverall, the reviewer after reading the other weaknesses still thinks that this paper is ok to be accepted.\"}", "{\"comment\": \"We would be grateful for your feedback on the revised manuscript. Please let us know if there are any remaining concerns or if further clarification is needed.\"}", "{\"comment\": \"Thank you once again for your insightful feedback as a reviewer. We have thoughtfully considered your suggestions and made substantial revisions to the manuscript accordingly. Your perspective is truly appreciated, and we invite you to review our rebuttal along with the updated manuscript. Based on the positive feedback from other reviewers, we hope our revisions effectively address your concerns and offer the clarity required for you to reassess your evaluation. If you need any further clarification or wish to discuss anything, please feel free to reach out.\"}", "{\"title\": \"No rebuttal\", \"comment\": \"I will keep my rating\"}" ] }

[ "Shuo Ma", "Yingwei Zhang", "Yiqiang Chen", "Hualei Wang", "Yuan Jin", "Wei Zhang", "Ziyu Jia" ]

[ "Sleep staging", "Multimodal coordination", "Ubiquitous computing" ]

{ "note_id": [ "zkDqnUlzEN", "zJwyvCPT5h", "yftdOSaNOc", "woyiDSGEJJ", "waR1ikwgbK", "v45x7HjIxl", "rhQoCZYYiJ", "ivVv889Hel", "iClf6RwZde", "dWEhw4Ix3r", "cYmR68jZ71", "XLRMZrDonG", "WUBJdF2SU6", "Tj0SuiIuBu", "SUqoXmpUb7", "RzFCLhkKM6", "QCSipWvrpP", "LrDr6ztBEj", "KoCoSE38z5", "JZg0VvcbDA", "J0a86LpZlB", "GlcHtdTOu0", "GNK8xcrkw3", "CnujCUvFu8", "CmatFxjGSv", "CTi8eiQ7eO", "BMqGsmb7L0", "9iFYVLqP7a", "7esxfA2M4f", "6g8Gy1LSrU", "5WsaqlQSHJ", "54V5ma7XWU", "4uM3e05J4x" ], "note_type": [ "official_comment", "official_comment", "official_comment", "official_comment", "decision", "official_comment", "official_comment", "meta_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_review" ], "note_created": [ 1732080679120, 1732081059129, 1732436251973, 1732693481570, 1737523407707, 1732077330491, 1732080793785, 1733996616669, 1732676616507, 1732080439942, 1732676385397, 1732711417849, 1732094795424, 1733109069703, 1733108932995, 1733211779841, 1732077374253, 1732513742053, 1732509714659, 1730361464061, 1732079592043, 1732080967522, 1733214786046, 1732676050251, 1733213852523, 1732077438388, 1730642653302, 1732096481900, 1732080911544, 1732081129826, 1730534995074, 1732675102990, 1730285430788 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission626/Authors" ], [ "ICLR.cc/2025/Conference/Submission626/Authors" ], [ "ICLR.cc/2025/Conference/Submission626/Authors" ], [ "ICLR.cc/2025/Conference/Submission626/Reviewer_HHyg" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission626/Authors" ], [ "ICLR.cc/2025/Conference/Submission626/Authors" ], [ "ICLR.cc/2025/Conference/Submission626/Area_Chair_HAdZ" ], [ "ICLR.cc/2025/Conference/Submission626/Authors" ], [ "ICLR.cc/2025/Conference/Submission626/Authors" ], [ "ICLR.cc/2025/Conference/Submission626/Authors" ], [ "ICLR.cc/2025/Conference/Submission626/Authors" ], [ "ICLR.cc/2025/Conference/Submission626/Reviewer_iRq4" ], [ "ICLR.cc/2025/Conference/Submission626/Authors" ], [ "ICLR.cc/2025/Conference/Submission626/Authors" ], [ "ICLR.cc/2025/Conference/Submission626/Authors" ], [ "ICLR.cc/2025/Conference/Submission626/Authors" ], [ "ICLR.cc/2025/Conference/Submission626/Reviewer_A2VX" ], [ "ICLR.cc/2025/Conference/Submission626/Reviewer_HHyg" ], [ "ICLR.cc/2025/Conference/Submission626/Reviewer_HHyg" ], [ "ICLR.cc/2025/Conference/Submission626/Authors" ], [ "ICLR.cc/2025/Conference/Submission626/Authors" ], [ "ICLR.cc/2025/Conference/Submission626/Authors" ], [ "ICLR.cc/2025/Conference/Submission626/Authors" ], [ "ICLR.cc/2025/Conference/Submission626/Reviewer_HHyg" ], [ "ICLR.cc/2025/Conference/Submission626/Authors" ], [ "ICLR.cc/2025/Conference/Submission626/Reviewer_Dg5B" ], [ "ICLR.cc/2025/Conference/Submission626/Authors" ], [ "ICLR.cc/2025/Conference/Submission626/Authors" ], [ "ICLR.cc/2025/Conference/Submission626/Authors" ], [ "ICLR.cc/2025/Conference/Submission626/Reviewer_iRq4" ], [ "ICLR.cc/2025/Conference/Submission626/Authors" ], [ "ICLR.cc/2025/Conference/Submission626/Reviewer_A2VX" ] ], "structured_content_str": [ "{\"title\": \"Response to [HHyg] (Part 1/n)\", \"comment\": \"Thank you for recognizing and supporting the practical significance and writing quality of our paper! We understand that your main concerns are related to the methodological innovation and some experimental details. Below, we address these points:\\n\\n1. **This paper does not introduce a new concept, but only combines and applies existing methods?**\\n\\n Our method incorporates supervised modality-level contrastive coordination and uncertainty-based weighting. The former introduces label information during modality alignment to enhance accuracy, aligning not only modality instances at the same time but also across different times, fully learning the category-relevant temporal information. The latter facilitates dynamic filtering of information during transmission, significantly improving the model's robustness. \\n\\n The combination of these two innovations effectively achieves single-modality sleep staging in ubiquitous scenarios. Therefore, while our work demonstrates innovation and practical significance in its application, it also includes substantial methodological novelty.\\n\\n2. **Why not adopt a sequence-to-sequence paradigm? Can this method adapt to a sequence-to-sequence paradigm?**\\n\\n 1) Many existing methods are not based on a sequence-to-sequence paradigm but still achieve excellent results. For example, several recent methods:\\n\\n [1] Jia Z, Wang H, Liu Y, et al. Mutual Distillation Extracting Spatial-temporal Knowledge for Lightweight Multi-channel Sleep Stage Classification[C]//Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 2024: 1279-1289.\\\\\\n [2] Thapa R, He B, Kjaer M R, et al. SleepFM: Multi-modal Representation Learning for Sleep Across Brain Activity, ECG and Respiratory Signals[C]//Forty-first International Conference on Machine Learning. 2024.\\\\\\n [3] Zhu H, Zhou W, Fu C, et al. Masksleepnet: A cross-modality adaptation neural network for heterogeneous signals processing in sleep staging[J]. IEEE Journal of Biomedical and Health Informatics, 2023, 27(5): 2353-2364.\\n\\n 2) Although our method is not specifically designed as a sequence-to-sequence paradigm, it inherently learns sequence-level information during the supervised contrastive coordination process. Our approach not only aligns sequences at the same time point but also sequences of the same class across different time points. This approach considers sequence-level contextual consistency and effectively utilizes class label information to handle situations such as abrupt class transitions in sleep staging tasks.\\n\\n For instance, during transitions between two sleep stages, traditional sequence-to-sequence methods might mistakenly classify adjacent sequences as belonging to the same stage. Our method addresses this issue effectively. Clearly, our approach is adaptable to a sequence-to-sequence paradigm.\"}", "{\"comment\": \"5. **Can unsupervised contrastive learning methods mentioned in the paper be fairly compared?**\\n\\n Thank you for your suggestion. We supplemented our comparisons with the method by Liu et al. (2024) across all scenarios on three datasets. To ensure a fair comparison, we adapted the method into a supervised, end-to-end approach while retaining its core modality-level and intra-modal contrastive learning components.\\n\\n- Multimodal scenario:\\n\\n| Dataset | Method | Accuracy | Macro F1 | Kappa |\\n|---------------|---------------|-----------|-----------|---------|\\n| **ISRUC-S3** | BSTT | 0.7756 | 0.7568 | 0.7114 |\\n| | XSleepNet | 0.6705 | 0.6440 | 0.5771 |\\n| | **DrFuse** | 0.7741 | 0.7469 | 0.7091 |\\n| | **MERL** | 0.7559 | 0.7458 | 0.6876 |\\n| | **Ours** | **0.7930** | **0.7815** | **0.7344** |\\n| **MASS-SS3** | BSTT | 0.8114 | 0.7492 | 0.7190 |\\n| | XSleepNet | 0.8066 | 0.7464 | 0.7158 |\\n| | **DrFuse** | 0.8628 | 0.8086 | 0.7964 |\\n| | **MERL** | 0.8605 | 0.8055 | 0.7915 |\\n| | **Ours** | **0.8686** | **0.8193** | **0.8058** |\\n| **Sleep-EDF-78** | BSTT | 0.7321 | 0.6335 | 0.6245 |\\n| | XSleepNet | 0.7577 | 0.6855 | 0.6631 |\\n| | **DrFuse** | 0.8009 | 0.7411 | 0.7235 |\\n| | **MERL** | 0.7990 | 0.7267 | 0.7196 |\\n| | **Ours** | **0.8158** | **0.7558** | **0.7450** |\\n\\n- Unimodal scenario + ISRUC-S3 dataset:\\n\\n| Dataset | Method | Accuracy | Macro F1 | Kappa |\\n|---------------|---------------|-----------|-----------|---------|\\n| **EEG** | BSTT | 0.7191 | 0.6921 | 0.6371 |\\n| | XSleepNet | 0.6555 | 0.6322 | 0.5614 |\\n| | **DrFuse** | 0.7532 | 0.7138 | 0.6818 |\\n| | **MERL** | 0.7467 | 0.7295 | 0.6758 |\\n| | **Ours** | **0.7646** | **0.7397** | **0.6969** |\\n| **EOG** | BSTT | 0.4700 | 0.3163 | 0.2790 |\\n| | XSleepNet | 0.6288 | 0.6071 | 0.5233 |\\n| | **DrFuse** | 0.6947 | 0.6799 | 0.6078 |\\n| | **MERL** | 0.6976 | 0.6741 | 0.6132 |\\n| | **Ours** | **0.7444** | **0.7168** | **0.6697** |\\n| **EMG** | BSTT | 0.3046 | 0.0934 | 0.0000 |\\n| | XSleepNet | 0.3660 | 0.3484 | 0.1935 |\\n| | **DrFuse** | 0.3857 | 0.3789 | 0.2318 |\\n| | **MERL** | 0.3981 | 0.3907 | 0.2348 |\\n| | **Ours** | **0.4384** | **0.4075** | **0.2693** |\\n\\n- Unimodal scenario + the other datasets:\\nPlease refer to Section A.3 of the revised manuscript for the full results and analysis.\\n\\n**Reference** \\\\\\n[1] Liu Y, Jia Z. Bstt: A bayesian spatial-temporal transformer for sleep staging[C]//The Eleventh International Conference on Learning Representations. 2023.\\\\\\n[2] Phan H, Ch\\u00e9n O Y, Tran M C, et al. XSleepNet: Multi-view sequential model for automatic sleep staging[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021, 44(9): 5903-5915.\\\\\\n[3] Yao W, Yin K, Cheung W K, et al. DrFuse: Learning Disentangled Representation for Clinical Multi-Modal Fusion with Missing Modality and Modal Inconsistency[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2024, 38(15): 16416-16424.\\\\\\n[4] Liu C, Wan Z, Ouyang C, et al. Zero-Shot ECG Classification with Multimodal Learning and Test-time Clinical Knowledge Enhancement[C]//Forty-first International Conference on Machine Learning, 2024.\", \"title\": \"Response to [A2VX] (Part 3/n)\"}", "{\"title\": \"Dear Reviewers: Request for your Feedback on Our Rebuttal Responses\", \"comment\": \"Dear Reviewers,\\n\\nThank you for your time and effort in reviewing our paper. We truly appreciate the positive feedback from the reviewers, recognizing our work as novel (Reviewer Dg5B), well-structured and clear (Reviewers A2VX and iRq4), and of practical significance (Reviewers HHyg and Dg5B).\", \"as_the_discussion_period_ends_on_nov_26_at_24\": \"00 AoE, we kindly remind you to review our rebuttal and shared responses.\\n\\nWe have provided a comprehensive response to your concerns through detailed explanations and thorough experimental validations. We are eagerly awaiting your feedback.\\n\\nThank you again for your support and valuable insights!\\n\\nBest regards\"}", "{\"title\": \"The results are not so convincing. The results of BSTT on both S1 and MASS-S3 reported by the authors are much lower than those reported in the original paper [1].\", \"comment\": \"Thank you for the effort in adding new experiments. I am curious about why the results of BSTT on both S1 and MASS-S3 are significantly lower than those reported in the original paper[1]. For instance, the BSTT result on S1 reported by the authors is 0.7247, while the original paper [1] states it as 0.8196. Why is there such a large discrepancy (0.7247 vs. 0.8196)? Similarly, the BSTT result on MASS-S3 provided by the authors is 0.8114, compared to 0.8950 in the original paper [1]. This gap is also quite substantial.\\n\\nGiven that the dataset and task are identical, I believe it would be more reasonable to directly compare the results with those reported in the original paper rather than re-implementing the method. This is why I find the results unconvincing.\\n\\n[1] Liu Y, Jia Z. Bstt: A bayesian spatial-temporal transformer for sleep staging[C]//The Eleventh International Conference on Learning Representations. 2023.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}", "{\"title\": \"Summary of revision\", \"comment\": [\"Dear reviewers, AC, SAC, and PC,\", \"We sincerely thank you for your time and suggestions. As a summary, we are grateful that merits in novelty, methodology, theory, and empirical results are favored by reviewers:\", \"**Reviewer Dg5B**: find our work *'clever and novel'*, *'empirically and theoretically effective'*, *'high quality'*, *'clearly presented'*, and *'meaningful contribution'*\", \"**Reviewer iRq4**: *'good organization'*\", \"**Reviewer HHyg**: *'certain practical significance'*\", \"**Reviewer A2VX**: *'well-structured and clear'*, *'easy to follow'*, and *'extensive experiments'*\", \"We revised the paper according to the reviewers' suggestions. The newer version of the paper is uploaded (The modified position is highlighted in blue).\", \"Based on the reviewers' suggestions, we have supplemented and revised the manuscript in the following key areas:\", \"**Experiments**: We have added four state-of-the-art methods to the experiments across three datasets under two scenario settings (including a multimodal testing scenario and three unimodal testing scenarios). These methods encompass two sequence-to-sequence sleep staging approaches, one multimodal robustness approach, and one multimodal contrastive coordination method. Additionally, we have supplemented ablation results using different weighted metrics and analyzed the computational complexity of our method. It is worth noting that all our original experiments were conducted under a cross-subject setting, which holds significant relevance for physiological signal analysis.\", \"**Writing**: We have reorganized and supplemented the Theorems and Proofs in the appendix and enlarged the text in Figures 3 and 4 for better readability. We also corrected and checked a grammatical error, following the suggestion from reviewer 'HHyg.' Furthermore, we added a list of symbols in the appendix, as recommended by reviewer 'Dg5B.' We deeply appreciate these valuable comments.\", \"Thank you for your help in making this work stronger.\", \"Although most reviewers provided positive feedback, we would still like to emphasize the following points:\", \"**Motivation and Significance**: Multimodal monitoring devices are often difficult to wear, uncomfortable during the monitoring process, and lack robustness. Our work aims to reduce the number of modalities required during the testing phase through multimodal coordination learning while ensuring comfort and improving accuracy. Our initial motivation is to achieve comfortable, simple, and highly accurate sleep monitoring in ubiquitous scenarios, such as home environments.\", \"**Innovation**: Our work is not only innovative at the application level but also demonstrates sufficient methodological novelty. Specifically, our method incorporates supervised modality-level instance contrastive coordination and uncertainty-based weighting. The former introduces label information into the alignment of multiple modality relationships to improve accuracy, while the latter dynamically filters information during transmission to enhance robustness.\"], \"we_sincerely_our_response_can_address_all_you_concerns\": \") If you have any questions, please let us know:)\"}", "{\"comment\": \"3. **The selected baselines are outdated. Why choose ISRUC-S3 instead of S1?**\\n\\n Thank you for your suggestion. The three methods you mentioned are highly influential in the sleep staging domain. We will include these three papers in the revised manuscript's related work section and provide comparisons. Additionally, we supplemented the experiments with five recent methods (including BSTT and XSleepNet), and the results under two scenarios demonstrate that our method still achieves the best performance.\\n\\n Regarding our choice of the smaller S3 subset instead of S1, there are two main considerations:\\n - **Dataset scale**: Although the SleepEDF dataset only includes 78 subjects, its total data volume is much larger than that of S1 and is often considered a large-scale dataset. We have already validated our method on SleepEDF.\\n - **Scenario**: S3 and S1 originate from the same source. Our method focuses on multimodal robustness in ubiquitous and low-resource scenarios. Thus, we prioritized validating performance on low-resource, small-scale datasets like S3.\\n\\n- Multimodal scenario:\\n\\n| Dataset | Method | Accuracy | Macro F1 | Kappa |\\n|---------------|---------------|-----------|-----------|---------|\\n| **ISRUC-S3** | **BSTT** | 0.7756 | 0.7568 | 0.7114 |\\n| | **XSleepNet** | 0.6705 | 0.6440 | 0.5771 |\\n| | DrFuse | 0.7741 | 0.7469 | 0.7091 |\\n| | MERL | 0.7559 | 0.7458 | 0.6876 |\\n| | **Ours** | **0.7930** | **0.7815** | **0.7344** |\\n| **MASS-SS3** | **BSTT** | 0.8114 | 0.7492 | 0.7190 |\\n| | **XSleepNet** | 0.8066 | 0.7464 | 0.7158 |\\n| | DrFuse | 0.8628 | 0.8086 | 0.7964 |\\n| | MERL | 0.8605 | 0.8055 | 0.7915 |\\n| | **Ours** | **0.8686** | **0.8193** | **0.8058** |\\n| **Sleep-EDF-78** | **BSTT** | 0.7321 | 0.6335 | 0.6245 |\\n| | **XSleepNet** | 0.7577 | 0.6855 | 0.6631 |\\n| | DrFuse | 0.8009 | 0.7411 | 0.7235 |\\n| | MERL | 0.7990 | 0.7267 | 0.7196 |\\n| | **Ours** | **0.8158** | **0.7558** | **0.7450** |\\n\\n\\n- Unimodal scenario + ISRUC-S3 dataset:\\n\\n| Modality | Method | Accuracy | Macro F1 | Kappa |\\n|---------------|---------------|-----------|-----------|---------|\\n| **EEG** | **BSTT** | 0.7191 | 0.6921 | 0.6371 |\\n| | **XSleepNet** | 0.6555 | 0.6322 | 0.5614 |\\n| | DrFuse | 0.7532 | 0.7138 | 0.6818 |\\n| | MERL | 0.7467 | 0.7295 | 0.6758 |\\n| | **Ours** | **0.7646** | **0.7397** | **0.6969** |\\n| **EOG** | **BSTT** | 0.4700 | 0.3163 | 0.2790 |\\n| | **XSleepNet** | 0.6288 | 0.6071 | 0.5233 |\\n| | DrFuse | 0.6947 | 0.6799 | 0.6078 |\\n| | MERL | 0.6976 | 0.6741 | 0.6132 |\\n| | **Ours** | **0.7444** | **0.7168** | **0.6697** |\\n| **EMG** | **BSTT** | 0.3046 | 0.0934 | 0.0000 |\\n| | **XSleepNet** | 0.3660 | 0.3484 | 0.1935 |\\n| | DrFuse | 0.3857 | 0.3789 | 0.2318 |\\n| | MERL | 0.3981 | 0.3907 | 0.2348 |\\n| | **Ours** | **0.4384** | **0.4075** | **0.2693** |\\n\\n- Unimodal scenario + the other datasets:\\nPlease refer to Section A.3 of the revised manuscript for the full results and analysis.\\n\\n**Reference** \\\\\\n[1] Liu Y, Jia Z. Bstt: A bayesian spatial-temporal transformer for sleep staging[C]//The Eleventh International Conference on Learning Representations. 2023.\\\\\\n[2] Phan H, Ch\\u00e9n O Y, Tran M C, et al. XSleepNet: Multi-view sequential model for automatic sleep staging[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021, 44(9): 5903-5915.\\\\\\n[3] Yao W, Yin K, Cheung W K, et al. DrFuse: Learning Disentangled Representation for Clinical Multi-Modal Fusion with Missing Modality and Modal Inconsistency[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2024, 38(15): 16416-16424.\\\\\\n[4] Liu C, Wan Z, Ouyang C, et al. Zero-Shot ECG Classification with Multimodal Learning and Test-time Clinical Knowledge Enhancement[C]//Forty-first International Conference on Machine Learning, 2024.\\n\\n4. **Supplement and proof not detailed enough?**\\n\\n Thank you for your feedback, which has improved the readability of our paper. In the revised version, we have further supplemented the relevant definitions and proofs for greater clarity and completeness.\", \"title\": \"Response to [HHyg] (Part 2/n)\"}", "{\"metareview\": \"The paper introduces SleepSMC, a novel framework for ubiquitous sleep staging that addresses the challenges of leveraging multimodal data for training while ensuring high performance with unimodal data during testing. SleepSMC employs supervised modality-level instance contrastive coordination to align category-related features across modalities and utilizes uncertainty-based feature weighting to prioritize reliable auxiliary modalities during training. Tested on three public datasets (ISRUC-S3, MASS-SS3, Sleep-EDF-78), SleepSMC achieves state-of-the-art performance in both multimodal and unimodal testing scenarios, bridging the gap between the complex multimodal settings of clinical sleep staging and practical, real-world applications. This approach enhances robustness, interpretability, and real-world applicability for sleep monitoring in ubiquitous scenarios. The initial weakness of this paper mainly comes from the technical novelty and the evaluation.\\n\\nIn the rebuttal, the authors provides more clarifications and has supplement more experiments that demonstrate the effectiveness. 3 reviewers give accept and 1 reviewer still gives \\\"lower than borderline\\\" due to a confusing point in the experiment part. I concisely check the discussions between authors and the reviewer HHyg and believe that the supplementary results follow a reasonable experiment design. However, the previous results in the original paper should also be included and the gap should be explained. Otherwise, it is confusing why previous methods produce a higher baseline. At least, the authors should compare it using the original setting. Overall, the method is tailored in sleep staging application and each module is well justified. Since I believe most concerns have been addressed, I still vote towards a borderline accept.\", \"additional_comments_on_reviewer_discussion\": \"In the rebuttal, the authors provides more clarifications and has supplement more experiments that demonstrate the effectiveness, addressing most of the concerns in clarification and experiments. 3 reviewers give accept and 1 reviewer still gives \\\"lower than borderline\\\" due to a confusing point in the experiment part. I concisely check the discussions between authors and the reviewer HHyg. The issue is that the authors conduct the experiments on S3 following a new experiment protocol. After reading the original baseline paper, I believe that the supplementary results by authors follow a reasonable experiment design. However, the previous results in the original paper should also be included or compared and the gap should be explained. Otherwise, it is confusing why previous methods produce a higher baseline. At least, the authors should compare it using the original setting.\"}", "{\"title\": \"Thanks for Reviewer [A2VX]\", \"comment\": \"Dear reviewer,\\n\\nThank you for your time and effort in reviewing our paper, and thank you for your recognition and positive feedback on our paper. Your support has brought us great encouragement.\\n\\nWe are really happy to address all your concerns, and we are very grateful for your help and encouragement, which makes our paper better and better. If you have further questions, we will be happy to receive your feedback.\\n\\nThank you again for your support and valuable insights!\\n\\nBest wishes\"}", "{\"title\": \"Response for [iRq4] (Part 3/n)\", \"comment\": \"2. **What is the computational complexity?**\\n\\n Our model uses only a CNN-based structure, its computational complexity is low overall, as detailed in the table below. In particular, the computational complexity during the single-modality testing phase is extremely low, which strongly supports its deployment in ubiquitous single-modality scenarios.\\n | Scenario | Training phase | Testing phase |\\n |------------|------------|------------|\\n | Multimodal|33.0M FLOPs & 1.3M param|16.5M FLOPs & 0.6M param|\\n | Unimodal| 26.1M FLOPs & 1.0M param|6.9M FLOPs & 0.2M param|\\n\\n3. **How equation 10 become optimized?**\\n\\n Equation 10 defines the overall objective function for optimizing SleepSMC. It integrates a supervised contrastive loss $\\ud835\\udc3f_{Con}$ to align cross-modal features and a classification loss $\\ud835\\udc3f_{cls}$ to ensure accurate sleep stage predictions. The $\\ud835\\udc3f_{Con}$ stabilizes the training process by controlling the contribution of auxiliary modalities.\\n\\n These losses are jointly optimized through a sum. The joint optimization is performed using stochastic gradient descent (SGD) with the Adam optimizer, which efficiently handles the gradients of both losses.\"}", "{\"title\": \"Thanks for Reviewer [iRq4]\", \"comment\": \"Dear reviewer,\\n\\nThank you for your time and effort in reviewing our paper, and thank you for your recognition and positive feedback on our paper.\\n\\nWe have provided comprehensive responses to your concerns through detailed explanations and thorough experimental validations. In particular, we have explained the real-time monitoring issue you mentioned.\\n\\nWe hope that we have addressed your concerns. If you have any further questions, we would be happy to receive your feedback.\\n\\nThank you again for your support and valuable insights!\\n\\nBest wishes\"}", "{\"comment\": \"We sincerely apologize for any confusion caused. In fact, our experimental setup differs significantly from that of the original paper. The primary reason for this is that the experimental setting in the original paper [1] is not entirely standardized.\\n\\nFrom the publicly available code of the original paper [1], it can be observed that the authors only divided the dataset into a training set and a test set. Specifically, they trained the model on the training set and reported the test set results for the best-performing epoch. Similarly, in another work, MSTGCN [2], the same author of [1] used a comparable setting, as seen in their released code.\\n\\nHowever, from a methodological perspective, this approach is not strictly standardized. To better avoid test set data leakage and evaluate the cross-subject generalization performance of the model, we introduced a validation set by randomly splitting 20% of the training set. The model was then trained using the remaining 80% of the training data, and the best-performing model on the validation set was saved for evaluation on the test set.\\n\\nThis approach inevitably reduced the amount of data available for training (to 80%) and ensured that the test set was not directly used for model selection. To maintain fairness, we applied this standardized experimental setup consistently across all our experiments.\\n\\nThis explains why we did not directly cite the results reported in the original paper. We hope this addresses your concerns.\\n\\n [1] Liu Y, Jia Z. Bstt: A bayesian spatial-temporal transformer for sleep staging[C]//The Eleventh International Conference on Learning Representations. 2023.\\n\\n[2] Jia Z, Lin Y, Wang J, et al. Multi-view spatial-temporal graph convolutional networks with domain generalization for sleep stage classification[J]. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2021, 29: 1977-1986.\", \"title\": \"Reason for Not Directly Using the Experimental Results from the Original Paper: Differences in Experimental Setting\"}", "{\"comment\": \"In the case of supervised contrastive learning, is it possible to monitor real-time home sleep in ubiquitous scenarios?\\nIf yes, then how?\"}", "{\"title\": \"Request for Reviewer [iRq4]'s Feedback\", \"comment\": \"Dear Reviewer [iRq4],\\n\\nThank you again for your time and effort in reviewing our paper.\", \"as_the_new_discussion_period_ends_on_december_2nd_at_24\": \"00 AoE, we kindly remind you to review our rebuttal and shared responses.\\n\\nWe have comprehensively responded to your concerns through detailed explanations and thorough experimental validation, and we also explain the contribution of our approach to real-time sleep monitoring. We, along with Reviewer [A2VX], sincerely believe that our paper has been strengthened a lot thanks to feedback.\\n\\nWe are eagerly awaiting your feedback.\\n\\nBest wishes\"}", "{\"title\": \"Request for Reviewer [Dg5B]'s Feedback\", \"comment\": \"Dear Reviewer [Dg5B],\\n\\nThank you again for your time and effort in reviewing our paper.\", \"as_the_new_discussion_period_ends_on_december_2nd_at_24\": \"00 AoE, we kindly remind you to review our rebuttal and shared responses.\\n\\nWe have comprehensively responded to your concerns through detailed explanations and thorough experimental validations. We, along with Reviewer [A2VX], sincerely believe that our paper has been strengthened a lot thanks to feedback. \\n\\nWe are eagerly awaiting your feedback. \\n\\nBest wishes\"}", "{\"comment\": \"Dear Reviewer HHyg,\\n\\nWe would like to express our sincere gratitude for your careful review and thoughtful comments, which have greatly helped improve our paper. We deeply appreciate the time and effort you have put into providing such valuable feedback.\\n\\nAs the discussion period is drawing to a close, we humbly hope to earn your support. We have made every effort to address your concerns, and we sincerely welcome any further questions or suggestions you may have. Your insights are truly invaluable to us, and we would be grateful for any additional guidance you can offer.\\n\\nThank you again for your time and consideration.\", \"title\": \"Dear Reviewer HHyg\"}", "{\"title\": \"Response to [Dg5B] (Part 1/n)\", \"comment\": \"Thank you for your comprehensive recognition and support of our work! We understand that your main concerns are related to certain experimental details and writing aspects of the paper. We address them below:\\n\\n1. **A concise list of symbols to better clarify notation?**\\n\\n We have added such a list in Section A.1 of the revised appendix. Thank you for your suggestion, which has improved our paper!\\n\\n2. **Solid contribution in sleep staging but without significant breakthroughs. Could it yield greater benefits in other multimodal learning problems?**\\n\\n Achieving a 2-3% improvement in the sleep staging field is already a substantial contribution. Furthermore, our paper primarily focuses on enhancing sleep staging comfort in ubiquitous scenarios. However, our method does indeed have the potential to be applied in other fields. Thank you for pointing this out\\u2014we have included this possibility in the \\\"Discussion and Limitation\\\" section of the revised appendix.\\n\\n3. **Cross-subject experiments**\\n\\n As described in the experimental settings of Section 5.2, all our experiments adopt a cross-subject setup, considering that physiological signal models are typically designed to generalize in such scenarios. In all experiments, the training and testing sets are cross-subject, and we report aggregate results over five different splits. We apologize for any confusion and have highlighted the *cross-subject* setup in Section 5.2.\\n\\n4. **Could the model handle dynamic shifts in modality reliability, and do the authors envision practical deployments for health monitoring or similar applications?**\\n\\n Yes, the model can handle dynamic shifts in modality reliability. During training, the reliability of each modality is re-evaluated at every step, and relevant metrics are recalculated. Our method ultimately enhances the overall robustness of the model, which allows it to adapt to reliability changes during the testing phase to a certain extent. What a coincidence! We have indeed envisioned practical deployments for health monitoring. We are currently developing a wearable single-modality sleep monitoring device and plan to deploy this model in such applications.\"}", "{\"comment\": \"The authors have addressed my comments, including comparison with more recent baselines, clarification of the results analysis, and the experiment setting. I appreciate the authors' efforts in making the necessary adjustments. I believe the quality of paper has improved, and I have increased my score accordingly.\"}", "{\"title\": \"the advantages of the proposed method are not significant.\", \"comment\": \"In the multimodal scenario, compared to other methods, the advantages of the proposed method are not significant (e.g., 0.8605 v.s. 0.8686; 0.8009 v.s. 0.8158). Besides, the S1 also has multiple modality data. I still think S1 is a better choice for the evaluation.\"}", "{\"summary\": \"To address the issue that only one modality is available in ubiquitous scenarios, this paper introduce multimodal collaboration in sleep staging, leveraging multiple auxiliary modalities to improve the performance of primary modality-based sleep staging in an end-to-end manner. The paper utilize supervised modality-level instance contrastive coordination and uncertainty estimates to learn coordinated features. The experiment results show that the proposed method achieves SOTA performance in multimodal scenarios and unimodal scenarios.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"The paper aims to address the issue that only one modality is available in ubiquitous scenarios, which has certain practical significance.\\n\\nThe paper utilizes uncertainty estimates to adaptively weight auxiliary modality features during\\ntraining, which ensures that more reliable auxiliary modality features contribute more significantly to the contrastive learning process.\\n\\nThe presentation of the paper is generally quite clear.\", \"weaknesses\": \"(1) This paper does not introduce a new concept; rather, it applies existing multimodal coordination and uncertainty estimates to the ubiquitous scenarios of sleep staging. Moreover, the proposed method for improving sleep staging does not seem to differ significantly from the multimodal coordination methods already used in areas such as vision. Some design details in the method, such as certain aspects related to uncertainty estimates, do not demonstrate a specific focus on the sleep staging task.\\n\\n(2) The authors seem to use a single-epoch sleep staging paradigm instead of a sequence-to-sequence sleep staging approach. Why is this the case? In fact, there is a strong correlation between adjacent epochs in sleep staging, and the single-epoch paradigm has already fallen behind in performance and been largely abandoned. As a paper aimed at improving sleep staging performance, it is difficult to understand the choice of using a single-epoch paradigm. Could the authors explain why this paradigm was chosen and whether the proposed method can be adapted to the sequence-to-sequence paradigm?\\n\\n(3) The reviewer has some concerns about the experimental design and results. The baselines chosen by the authors appear to be relatively weak and outdated, such as the selection of SimCLR, which is a very old self-supervised learning method. The baselines in the paper perform very weakly in multimodal scenarios (shown in Table 1), making it difficult to demonstrate the advantages of the proposed method. Some very strong baselines, such as SalientSleepNet [1], BSTT [2] and XSleepNet [3] were not compared. Meanwhile, the reviewer cannot understand why the authors chose ISRUC-S3 instead of ISRUC-S1 as the primary evaluation dataset. ISRUC-S1 has more subjects and a larger data volume, making it more convincing compared to ISRUC-S3.\\n\\n(4) Although the presentation of the paper is relatively clear, the content in \\\"DETAILED ANALYSIS AND PROOFS IN SECTION 4\\\" is difficult for readers to understand. As a supplement and proof of certain concepts in the main text, this section does not sufficiently clarify the formal definitions of each concept and the complete proof process. For example, what is the complete proposition that needs to be proven in this section? What are the assumptions underlying the proof process? What is the formal definition of \\\"margin\\\"? What is the formal definition of \\\"information transfer\\\"? The authors should provide clear explanations and complete definitions of these concepts in the appendix to ensure the readability of the paper.\\n\\n[1] Jia Z, Lin Y, Wang J, et al. SalientSleepNet: Multimodal salient wave detection network for sleep staging[J]. arXiv preprint arXiv:2105.13864, 2021.\\n\\n[2] Liu Y, Jia Z. Bstt: A bayesian spatial-temporal transformer for sleep staging[C]//The Eleventh International Conference on Learning Representations. 2023.\\n\\n[3] Phan H, Ch\\u00e9n O Y, Tran M C, et al. XSleepNet: Multi-view sequential model for automatic sleep staging[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021, 44(9): 5903-5915.\", \"questions\": \"Please See Weakness\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"- **Experiments**: To further demonstrate the effectiveness of our method, we have included comparisons with four state-of-the-art methods, covering approaches for multimodal missing and sleep staging methods. On three datasets, under both multimodal and unimodal scenarios, our method consistently achieves the best performance.\\n\\n1. Multimodal scenario:\\n\\n| Dataset | Method | Accuracy | Macro F1 | Kappa |\\n|---------------|---------------|-----------|-----------|---------|\\n| **ISRUC-S3** | BSTT | 0.7756 | 0.7568 | 0.7114 |\\n| | XSleepNet | 0.6705 | 0.6440 | 0.5771 |\\n| | DrFuse | 0.7741 | 0.7469 | 0.7091 |\\n| | MERL | 0.7559 | 0.7458 | 0.6876 |\\n| | **Ours** | **0.7930** | **0.7815** | **0.7344** |\\n| **MASS-SS3** | BSTT | 0.8114 | 0.7492 | 0.7190 |\\n| | XSleepNet | 0.8066 | 0.7464 | 0.7158 |\\n| | DrFuse | 0.8628 | 0.8086 | 0.7964 |\\n| | MERL | 0.8605 | 0.8055 | 0.7915 |\\n| | **Ours** | **0.8686** | **0.8193** | **0.8058** |\\n| **Sleep-EDF-78** | BSTT | 0.7321 | 0.6335 | 0.6245 |\\n| | XSleepNet | 0.7577 | 0.6855 | 0.6631 |\\n| | DrFuse | 0.8009 | 0.7411 | 0.7235 |\\n| | MERL | 0.7990 | 0.7267 | 0.7196 |\\n| | **Ours** | **0.8158** | **0.7558** | **0.7450** |\\n\\n\\n2. Unimodal scenario + ISRUC-S3 dataset:\\n\\n| Modality | Method | Accuracy | Macro F1 | Kappa |\\n|---------------|---------------|-----------|-----------|---------|\\n| **EEG** | BSTT | 0.7191 | 0.6921 | 0.6371 |\\n| | XSleepNet | 0.6555 | 0.6322 | 0.5614 |\\n| | DrFuse | 0.7532 | 0.7138 | 0.6818 |\\n| | MERL | 0.7467 | 0.7295 | 0.6758 |\\n| | **Ours** | **0.7646** | **0.7397** | **0.6969** |\\n| **EOG** | BSTT | 0.4700 | 0.3163 | 0.2790 |\\n| | XSleepNet | 0.6288 | 0.6071 | 0.5233 |\\n| | DrFuse | 0.6947 | 0.6799 | 0.6078 |\\n| | MERL | 0.6976 | 0.6741 | 0.6132 |\\n| | **Ours** | **0.7444** | **0.7168** | **0.6697** |\\n| **EMG** | BSTT | 0.3046 | 0.0934 | 0.0000 |\\n| | XSleepNet | 0.3660 | 0.3484 | 0.1935 |\\n| | DrFuse | 0.3857 | 0.3789 | 0.2318 |\\n| | MERL | 0.3981 | 0.3907 | 0.2348 |\\n| | **Ours** | **0.4384** | **0.4075** | **0.2693** |\\n\\n3. **Unimodal scenario + the other datasets**:\\nPlease refer to Section A.3 of the revised manuscript for the full results and analysis.\\n\\n**Reference** \\\\\\n[1] Liu Y, Jia Z. Bstt: A bayesian spatial-temporal transformer for sleep staging[C]//The Eleventh International Conference on Learning Representations. 2023.\\\\\\n[2] Phan H, Ch\\u00e9n O Y, Tran M C, et al. XSleepNet: Multi-view sequential model for automatic sleep staging[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021, 44(9): 5903-5915.\\\\\\n[3] Yao W, Yin K, Cheung W K, et al. DrFuse: Learning Disentangled Representation for Clinical Multi-Modal Fusion with Missing Modality and Modal Inconsistency[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2024, 38(15): 16416-16424.\\\\\\n[4] Liu C, Wan Z, Ouyang C, et al. Zero-Shot ECG Classification with Multimodal Learning and Test-time Clinical Knowledge Enhancement[C]//Forty-first International Conference on Machine Learning, 2024.\", \"title\": \"Response to [iRq4] (Part 2/n)\"}", "{\"title\": \"Response to [A2VX] (Part 2/n)\", \"comment\": \"4. **Do you expect high uncertainty at the beginning of training? Any experiments on dynamic weight adjustment? Experiments with alternative metrics?**\\n\\n 1) In fact, we do not expect high uncertainty, as it is undesirable. Ideally, uncertainty should be low for perfect modality coordination learning. Our goal is to enable the model to overcome situations of high uncertainty and large differences, thereby improving robustness.\\n\\n 2) All weights in our method are dynamically computed and adjusted, adapting at each step during training.\\n\\n 3) We conducted experiments using the uncertainty variance metric, directly applying the negative value of formula (5), $-r_i^{u_a}$. However, as explained in formula (6), *\\\"the exponential function introduces a smooth, continuous inverse scaling, where weights decrease more sharply for high uncertainty and more gradually for low uncertainty, reflecting the varying tolerance of the model.\\\"* This approach achieves better performance. Details are shown in below:\\n\\n| Dataset | Metric | Modality | Accuracy | Macro F1 | Kappa |\\n|-----------|-------------------------|-------------|-----------|-----------|---------|\\n| ISRUC-S3 | $ -r_i^{u_a} $ | Multimodal | 0.7910 | 0.7755 | 0.7315 |\\n| | $ \\\\exp(-r_i^{u_a}) $ | Multimodal | **0.7930** | **0.7815** | **0.7344** |\\n| | $ -r_i^{u_a} $ | EEG | 0.7501 | 0.7259 | 0.6794 |\\n| | $ \\\\exp(-r_i^{u_a}) $ | EEG | **0.7646** | **0.7397** | **0.6969** |\\n| | $ -r_i^{u_a} $ | EOG | 0.7317 | 0.7087 | 0.6541 |\\n| | $ \\\\exp(-r_i^{u_a}) $ | EOG | **0.7444** | **0.7168** | **0.6697** |\\n| | $ -r_i^{u_a} $ | EMG | 0.4194 | 0.4018 | 0.2523 |\\n| | $ \\\\exp(-r_i^{u_a}) $ | EMG | **0.4384** | **0.4075** | **0.2693** |\\n| MASS-SS3 | $ -r_i^{u_a} $ | Multimodal | 0.8683 | 0.8193 | 0.8050 |\\n| | $ \\\\exp(-r_i^{u_a}) $ | Multimodal | **0.8686** | **0.8193** | **0.8058** |\\n| | $ -r_i^{u_a} $ | EEG | 0.8474 | 0.7779 | 0.7731 |\\n| | $ \\\\exp(-r_i^{u_a}) $ | EEG | **0.8517** | **0.7871** | **0.7798** |\\n| | $ -r_i^{u_a} $ | EOG | 0.8195 | 0.7478 | 0.7305 |\\n| | $ \\\\exp(-r_i^{u_a}) $ | EOG | **0.8227** | **0.7534** | **0.7359** |\\n| | $ -r_i^{u_a} $ | EMG | 0.5350 | 0.3734 | 0.2514 |\\n| | $ \\\\exp(-r_i^{u_a}) $ | EMG | **0.5408** | **0.3770** | **0.2613** |\"}", "{\"comment\": \"Dear **Reviewer Dg5B** and **Reviewer iRq4**,\\n\\nWe would like to express our heartfelt gratitude for the recognition and support you have shown us in the initial stage of the review process. We are also deeply thankful for your valuable suggestions, which have truly enhanced the quality of our revised paper.\\n\\nWe have made every effort to carefully address and resolve the concerns you raised, and we hope that our revisions meet your expectations. As the discussion period is coming to a close, we humbly and earnestly hope to earn your continued strong support. If there are any further questions or concerns, we would be more than willing to address them.\\n\\nDear **Reviewer A2VX**,\\n\\nWe are truly delighted that we have been able to address all of your concerns, and we are deeply grateful for your decision to raise our score. Your support and encouragement mean a great deal to us.\\n\\nAs the discussion period is nearing its end, we sincerely welcome any further questions or feedback you may have. Your insights have been invaluable to us, and we would be more than happy to address any additional concerns.\"}", "{\"title\": \"Thanks for Reviewer [Dg5B]\", \"comment\": \"Dear Reviewer,\\n\\nThank you for taking the time and effort to review our paper, and thank you for your recognition of our paper.\\n\\nWe have provided a comprehensive response to your concerns through detailed explanations.\\n\\nWe hope we have addressed your concerns. If you have any further questions, we would be delighted to hear your feedback.\\n\\nThank you again for your support and valuable insights!\\n\\nBest regards\"}", "{\"title\": \"Thank you for your efforts and explanation, and I maintain my initial review score.\", \"comment\": \"Thank you for your efforts and explanation. However, I am still confused. It seems unreasonable to have such a significant gap in the results for sleep staging solely due to differences in the experimental settings in the training and validation sets. Therefore, I will maintain my initial review score.\"}", "{\"title\": \"Response to [iRq4] (Part 1/n)\", \"comment\": \"Thank you for recognizing and supporting our experimental work and writing! We understand that your main concerns are related to the motivations behind our method and some implementation details. Below, we address them in detail.\\n\\n1. **The motivation is not clear? Where is the novelty of the method?**\\n\\nThank you for your recommendation. We have carefully reviewed the CoRe-Sleep and TSEDSleepNet methods you mentioned, which are excellent sleep staging approaches. However, our method is significantly different from theirs, and we will compare and cite it in the revised paper. Details are as follows:\\n\\n - **Motivation**: Our method aims to address the problem of multimodal training with single-modality testing, significantly improving single-modality performance during the testing phase. This makes it well-suited for ubiquitous applications, such as home sleep monitoring in ubiquitous scenarios. In contrast, CoRe-Sleep and TSEDSleepNet focus on achieving better multimodal fusion. CoRe-Sleep emphasizes robust fusion, while TSEDSleepNet focuses on class imbalance and temporal modeling capabilities.\\n\\n - **Methodological Novelty**: We introduce supervised contrastive learning for modality-level coordinated and associated learning, effectively addressing the inconsistency between modalities during training and testing phases. Additionally, we propose uncertainty-based weighting to facilitate the selection and transmission of high-quality information during coordination. While CoRe-Sleep also performs modality alignment, it relies on unsupervised alignment and does not fully leverage class label information or optimize the alignment process. On the other hand, the primary innovations in TSEDSleepNet lie in its model structure and its loss function designed for class imbalance.\"}", "{\"summary\": \"This paper presents SleepSMC, a sleep stage classification framework that integrates uncertainty-based feature reweighting with modality-level contrastive learning to handle multimodal physiological data (EEG, EOG, EMG) effectively. The reweighting mechanism assigns weights to auxiliary modalities based on their uncertainty. Then, they use contrastive learning to align representations of the same sleep stage across different modalities. SleepSMC achieves modality-invariant embeddings that allow for robust performance even when only a single primary modality is available at test time. SleepSMC is evaluated on three public datasets, consistently outperforming SOTA baselines in both multimodal and unimodal settings. They present ablation studies demonstrating that reweighting and contrastive learning are both effective individually and moreso when combined. Visualization analyses further demonstrate the model\\u2019s interpretability, showing clear, well-separated embeddings for each sleep stage.\", \"soundness\": \"4\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"This paper presents SleepSMC, a sleep stage classification framework that integrates uncertainty-based feature reweighting with modality-level contrastive learning to handle multimodal physiological data (EEG, EOG, EMG) effectively. The reweighting mechanism assigns weights to auxiliary modalities based on their uncertainty. Then, they use contrastive learning to align representations of the same sleep stage across different modalities. SleepSMC achieves modality-invariant embeddings that allow for robust performance even when only a single primary modality is available at test time. SleepSMC is evaluated on three public datasets, consistently outperforming SOTA baselines in both multimodal and unimodal settings. They present ablation studies demonstrating that reweighting and contrastive learning are both effective individually and moreso when combined. Visualization analyses further demonstrate the model\\u2019s interpretability, showing clear, well-separated embeddings for each sleep stage.\\n\\nOriginality\\nThe uncertainty reweighting technique is clever and novel. Most importantly, it is both empirically and theoretically effective. The contrastive learning component, combined with reweighting, enables the model to create modality-invariant embeddings, allowing SleepSMC to generalize effectively across different primary modalities.\\n\\n\\nQuality\\nThe quality of the work is high. The evaluation is thorough, providing empirical, theoretical, and qualitative support for their approach. Reweighting consistently outperformed non-reweighted setups in both multimodal and unimodal testing, with clear gains in accuracy, Macro F1, and Kappa scores, particularly under noisy conditions. Combining reweighting with contrastive learning further boosted performance compared to contrastive learning alone, showing that these methods work well together, especially with unreliable auxiliary modalities. SleepSMC also generalized effectively across target modalities, performing robustly whether EEG, EOG, or EMG was used as the primary modality, demonstrating flexibility across configurations.\\nIn both multimodal and unimodal testing, SleepSMC maintained high accuracy, with unimodal performance benefiting notably from reweighting. t-SNE visualizations showed clear, well-separated clusters for each sleep stage, supporting the model\\u2019s modality-invariant embedding space and interpretability. These results confirm that SleepSMC\\u2019s reweighting and contrastive learning mechanisms enhance robustness and adaptability, making it a practical solution for real-world sleep staging.\\n\\n\\nClarity\\nThe work is mostly clearly presented. The methodology and results are well-structured, and the theoretical analysis is solid, providing strong support for the approach. However, the introduction of mathematical notation in Section 3 could be organized more effectively; starting with a table or list of symbols and definitions before diving into the equations would help clarify the notation and reduce cognitive load for the reader. The paper is otherwise transparent in discussing its limitations, and the overall clarity of the empirical findings is strong.\\n\\n\\nSignificance\\nThe work makes a meaningful contribution, adding to the literature on multimodal integration, particularly in scenarios where only unimodal data is available at test time\\u2014an approach that aligns well with many real-world applications.\", \"weaknesses\": \"1. The introduction of mathematical notation in Section 3 is somewhat disorganized and could be clarified better. Starting this section with a concise table or list of symbols and definitions would provide readers with a quick reference point, making it easier to follow the subsequent equations. This adjustment would enhance readability, especially for readers less familiar with the specific notation conventions used.\\n\\n2. The work makes a solid contribution to the field, with gains in the range of 2-3%, which are statistically significant but don\\u2019t drastically improve performance over baseline methods. While these improvements are valuable they're somewhat incremental. Perhaps the gains are more significant in other multimodal learning problems.\", \"questions\": \"An important aspect of multimodal classification models is cross-subject generalization. Given that real-world applications often involve new subjects with varying characteristics, an evaluation on unseen subjects would provide valuable insight into the model\\u2019s robustness. If cross-subject generalization was analyzed, could the authors include these results? Otherwise, adding this analysis would strengthen the paper\\u2019s practical relevance.\\n\\nAdditionally, while the reported improvements are statistically significant, they appear relatively modest, with gains in the range of 2-3%. Could the authors elaborate on why these incremental gains are meaningful in the context of sleep staging? Providing additional context on the impact of these gains for real-world applications would clarify the value of these results.\\n\\nAlthough the reweighting technique is effective within sleep staging, it would be useful to understand its potential applicability beyond this domain. Do the authors see possible applications of uncertainty-based reweighting in other multimodal learning tasks? Expanding on this would highlight the versatility and broader impact of the proposed method.\\n\\nIn terms of practical applications, it would be helpful to know how this model performs in real-world scenarios where data quality and availability are less controlled. Could the model handle dynamic shifts in modality reliability, and do the authors envision practical deployments for health monitoring or similar applications?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to [iRq4] (Part 4/n): Feasibility of Real-Time Monitoring\", \"comment\": \"Our method indeed enables real-time sleep monitoring. Within the context of supervised contrastive learning, contrastive learning is performed during the training phase. The goal is to enhance the learning of modality-specific, class-relevant information through multimodal data. This approach significantly improves the model's performance in single-modality testing scenarios. In practical applications, users only need to wear a simple single-modality device, such as an ear-EEG device. The model can process the collected single-modality data to accurately classify sleep stages. By analyzing sleep cycles and structures, effective sleep monitoring becomes straightforward.\\n\\n**Feasibility of Real-Time Monitoring**:\\n\\nAs discussed in Part 3/n, the computational complexity of single-modality model inference is extremely low, requiring only 6.9M FLOPs. Even on a mid-range ARM CPU (capable of executing 5 GFLOPs per second), the inference time is approximately **1.38 milliseconds**, while a typical clinical sleep segment spans **30 seconds**. Since the inference time is orders of magnitude smaller than the sample duration, real-time monitoring is highly feasible.\\n\\nLet me know if you have further questions!\"}", "{\"title\": \"Response to [A2VX] (Part 1/n)\", \"comment\": \"Thank you for recognizing and supporting the writing and experimental work in our paper! We understand that your main concerns are related to methodological innovation and some experimental details. Below, we address these points:\\n\\n1. **Why was \\\"Supervised Contrastive Learning\\\" not used as a baseline? What is the motivation for contrastive learning?**\\n\\n The \\\"Supervised Contrastive Learning\\\" method is built on contrastive learning between augmented and original data samples, aiming to enhance general representation capabilities. In contrast, our contrastive learning approach focuses more on inter-modal coordination and alignment, with the goal of improving single-modality robustness and performance in an end-to-end manner.\\n\\n To evaluate the effect of data augmentation and provide a meaningful comparison, we supplemented our experiments with the recent method you suggested: MERL (Liu et al., ICML2024). This method incorporates both inter-modal and intra-modal contrastive learning. Specifically, inter-modal alignment facilitates information transfer, while intra-modal contrastive learning leverages data augmentation to enhance representation and robustness. From the experimental results, our method still demonstrates significant superiority.\\n\\n2. **Should methods for handling missing modalities in other fields also be compared?**\\n\\n Thank you for your suggestion. ShaSpec (CVPR 2023) and DrFuse (AAAI 2024) are excellent works, and we will cite and compare them in the revised paper's related work section. Additionally, we supplemented comparisons with DrFuse (AAAI 2024) under all scenarios across three datasets. The experimental results show that our method still outperforms these approaches.(The table has been combined with Question 5 below.)\\n\\n3. **What is the significance of single-modality available? Why not consider a set of modalities?**\\n\\n In sleep monitoring, multimodal devices are often complex, difficult to wear, and significantly disrupt sleep experiences, posing substantial obstacles to the monitoring process. Such devices are not suitable for ubiquitous scenarios (e.g., home settings) and may even decrease monitoring accuracy. We are developing a single-modality device aimed at enabling anyone, even without specialized medical knowledge, to perform comfortable, convenient, and accurate sleep monitoring through simple contact.\\n\\n Therefore, our method primarily focuses on single-modality testing scenarios. Thank you for pointing this out. We have added this limitation in the \\\"Discussion and Limitation\\\" section of the revised paper.\"}", "{\"title\": \"Response to [A2VX] (Part 4 /n)\", \"comment\": \"6. **Why are uncertainty estimation methods mentioned in related work not applicable to the target scenario?**\\n\\n Thank you for pointing this out. We will elaborate on this in the revised manuscript. These methods focus on effective multimodal fusion, including retaining useful modality information and removing irrelevant data. However, they do not address single-modality application scenarios, which are our primary concern. Their methods require all modalities present during the testing phase, and retaining only one modality significantly degrades accuracy. In contrast, our method improves the performance of individual modalities through coordinated learning, enabling robust single-modality applications in ubiquitous scenarios.\\n\\n7. **The performance improvement of the uncertainty-based weighting module is smaller than that of the contrastive learning module. How is randomness mitigated?**\\n\\n Thank you for highlighting this. To mitigate randomness, we have taken the following steps:\\n\\n - **Experimentation**: As described in Section 5.2 of the original paper, we conducted extensive experiments, and all reported results are aggregated from five-fold cross-validation.\\n\\n - **Metrics**: To account for the impact of randomness, we introduced the Kappa metric. Kappa is a robust metric designed to measure agreement between predictions and ground truth while accounting for agreement due to chance. The Kappa results clearly illustrate the contribution of the uncertainty-based weighting module.\\n\\n - **Writing**: Thank you for pointing this out. We acknowledge that the role of the uncertainty-based weighting module is smaller than that of the contrastive learning module. Our original language may have been misleading, implying that most of the performance gain came from the uncertainty-based module. We have revised this statement to: *\\\"The Supervised Modality-level Contrastive Coordination module plays a more significant role in both scenarios. Meanwhile, the Uncertainty-based Feature Weighting module demonstrates relatively enhanced performance in unimodal compared to multimodal scenarios.\\\"* Thank you again for your thoughtful feedback, which has improved our paper.\\n\\n8. **Fluctuations in Figure 3c?**\\n\\n The robustness gap between EMG and the other two modalities is significant, resulting in less fluctuation for EMG compared to EEG and EOG. For EEG and EOG, their robustness is closer, leading to relative changes in their rankings. Additionally, robustness gaps always exist, and our goal is not to eliminate them but to demonstrate the significance of our method through these gaps. Our method dynamically leverages these differences to achieve better information transfer.\\n\\n9. **Minor writing suggestions?**\\n\\n Thank you for your meticulous review, which has made our paper better! We have incorporated your suggestions into the revised manuscript and have thoroughly checked the entire paper for any writing issues.\"}", "{\"summary\": \"This paper introduced multimodal collaboration in sleep staging, leveraging multiple auxiliary modalities to improve the performance of primary modality-based sleep staging in an end-to-end manner. The authors utilized supervised modality-level instance contrastive coordination to capture category-related consistency and complementarity across intra-modality and inter-modality.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"Experiments are performed using 3 different datasets. The organization of the paper is good.\", \"weaknesses\": \"The limitations are as follows:\\n\\nThe motivation is not clear.\\nHow this contribution is different from the following contributions? a)CoRe-Sleep: A Multimodal Fusion Framework for Time Series Robust to Imperfect Modalities,\\\" in IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 32, pp. 840-849, 2024, doi: 10.1109/TNSRE.2024.3354388. b) Multi-Modal Sleep Stage Classification With Two-Stream Encoder-Decoder,\\\" in IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 32, pp. 2096-2105, 2024, doi: 10.1109/TNSRE.2024.3394738.\\nThe motivation behind the use of Uncertainty Estimation with Frozen Gradients is not clear.\\nThe methods mentioned in Table 2 are from before 2022. It is necessary to compare with some recent SOTA methods\", \"questions\": \"1. How the proposed method is different from SOTA?\\n2. What is the novel contribution that makes this paper unique?\\n3. What is the computational complexity?\\n4. How equation 10 become optimized?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to [HHyg] (Part 3/n): Supplemented the ISRUC-S1 dataset\", \"comment\": \"Dear reviewer:\\n\\nAccording to your suggestion, we added experiments on the **ISRUC-S1** dataset with four comparison methods in two scenarios. Thank you for your suggestion, which make our paper better. If you have any additional suggestions or feedback, we would be truly grateful to hear them.\\n\\n1. Multimodal scenario:\\n\\n| Dataset | Method | Accuracy | Macro F1 | Kappa |\\n|---------------|---------------|-----------|-----------|---------|\\n| **ISRUC-S1** | **BSTT** | 0.7247 | 0.6890 | 0.6423 |\\n| | **XSleepNet** | 0.7444 | 0.7226 | 0.6707 |\\n| | DrFuse | 0.7441 | 0.7215 | 0.6669 |\\n| | MERL | 0.7245 | 0.7042 | 0.6417 |\\n| | **Ours** | **0.7710** | **0.7462** | **0.7018** |\\n\\n\\n2. Unimodal scenario + **ISRUC-S1** dataset:\\n\\n| Modality | Method | Accuracy | Macro F1 | Kappa |\\n|---------------|---------------|-----------|-----------|---------|\\n| **EEG** | **BSTT** | 0.6840 | 0.6367 | 0.5878 |\\n| | **XSleepNet** | 0.7092 | 0.6735 | 0.6233 |\\n| | DrFuse | 0.6978 | 0.6620 | 0.6087 |\\n| | MERL | 0.7096 | 0.6690 | 0.6223 |\\n| | **Ours** | **0.7328** | **0.6959** | **0.6536** |\\n| **EOG** | **BSTT** | 0.3123 | 0.1115 | 0.0019 |\\n| | **XSleepNet** | 0.6320 | 0.6047 | 0.5229 |\\n| | DrFuse | 0.6539 | 0.6223 | 0.5466 |\\n| | MERL | 0.6579 | 0.6261 | 0.5573 |\\n| | **Ours** | **0.7066** | **0.6753** | **0.6156** |\\n| **EMG** | **BSTT** | 0.3155 | 0.0959 | 0.0000 |\\n| | **XSleepNet** | 0.3883 | 0.3636 | 0.2083 |\\n| | DrFuse | 0.3395 | 0.2528 | 0.1418 |\\n| | MERL | 0.3786 | 0.3458 | 0.2012 |\\n| | **Ours** | **0.4190** | **0.3705** | **0.2382** |\\n\\nBest wishes\"}", "{\"summary\": \"The manuscript proposes SleepSMC, a method for detecting sleep stages leveraging contrastive learning and feature weighting based on uncertainty. The approach is specifically designed to handle scenarios where multiple data modalities are available during training, but only a single modality is accessible during testing (multimodal scenario has also been evaluated). Using three publicly available datasets, the authors demonstrate that SleepSMC achieves performance improvements over existing baselines.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": [\"The paper is well-structured and clear, and easy to follow.\", \"Includes extensive experiments comparing the proposed method with various baselines.\"], \"weaknesses\": [\"Limited technical novelty.\", \"More recent & missing-modality-specific baselines could be considered as stronger baselines.\", \"[Details]\", \"1. Could the authors explain why\\u00a0\\\"Supervised Contrastive Learning\\\"\\u00a0by Khosla et al. (2020) was not considered as one of the baselines? What specific motivations underlie the use of contrastive learning here, and in what ways does the proposed technique differ (in terms of technical novelty) from the original approach in\\u00a0Khosla et al.?\", \"Khosla, Prannay, et al. \\\"Supervised contrastive learning.\\\" Advances in Neural Information Processing Systems 33 (2020): 18661-18673.\", \"2. Is there any method or technique in this work that is specifically tailored for sleep stage detection? The proposed approach appears applicable to most multimodal data; If that's the case, other recent works targeting learning/inference with missing modalities could be considered as baselines as well:\", \"Wang, Hu, et al. \\\"Multi-modal learning with missing modality via shared-specific feature modelling.\\\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023.\", \"Yao, Wenfang, et al. \\\"DrFuse: Learning Disentangled Representation for Clinical Multi-Modal Fusion with Missing Modality and Modal Inconsistency.\\\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 38. No. 15. 2024.\"], \"questions\": \"[Problem setting & method]\\n1. Could the authors clarify the rationale for assuming that only one modality is available during inference? While it is understandable that not all modalities used in training may be accessible at test time, it seems uncommon to assume the presence of only a single modality. A more practical scenario might involve each user having a unique set of available modalities at test time, rather than relying on one fixed modality. Could the authors elaborate on the motivation behind this specific setting?\\n\\n2. Would it be expected that the auxiliary modality classifiers exhibit higher uncertainty at the beginning of training? If so, could this uncertainty indicate an opportunity for the model to learn more effectively? Did the authors experiment with dynamic tuning of weights over the course of training? Additionally, numerous metrics (such as entropy, model confidence, loss) could potentially assess sample importance. What was the rationale for selecting model confidence? Were other metrics considered, and if so, what were the comparative results?\\n\\n[Related work & Baselines] \\n1. In Section 2, Liu et al. (2023b) and Liu et al. (2024) are referenced as leveraging multimodal data to enhance performance, with the former employing contrastive learning to align modalities. Could the authors clarify why these methods were not considered as baselines? Additionally, it is noted that these works focus on multimodal consistency without capturing class-specific information. However, the design choices of related works should not necessarily be considered limitations. Could the authors discuss the impact of excluding class-specific information? How can we assess the relative performance without a comparison to the proposed method?\\n\\n2. In Section 2, several works related to uncertainty estimation are discussed, but the conclusion remains somewhat ambiguous. Why are these methods not applicable to the target scenario? What challenges prevent their direct application to the proposed task, especially given that multimodal data are still used during training?\\n\\n[Results]\\n1. In most scenarios, including unimodal cases, the primary improvements appear to stem from supervised contrastive learning. However, in Section 5.5, it is stated that \\u201cuncertainty-based feature weighting has a greater impact in the unimodal testing scenario,\\u201d which may be an overstatement. Could the authors provide comparative improvement results (e.g., average \\u00b1 standard deviation) for both components? If the enhancement from uncertainty-based feature weighting is minor, what justifies its inclusion?\\n\\n2. Could the authors explain the observed fluctuations in EEG and EOG uncertainty weights as training progresses (as seen in Figure 3c)?\\n\\n[Minor comments]:\\n- Font size in Figure 3 appears small.\\n- Class labels in Figure 4 are difficult to read; increasing the legend font size could improve readability.\\n- Typo in Section 4.2, first sentence: \\u201cadaptively weighted\\u201d should be \\u201cadaptively weights.\\u201d\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}" ] }

[ "Jianqiao Lu", "Yingjia Wan", "Yinya Huang", "Jing Xiong", "Zhengying Liu", "Zhijiang Guo" ]

[ "Large Language models", "Autoformalization", "Lean 4", "Formal Math", "AI for Math" ]

{ "note_id": [ "yeBGcqczzH", "xw4HxQXivL", "xu61mvY1YE", "wKFtpRYiJv", "w2RXSBGYtA", "tym3hGA8Cm", "tKMCiOQhtI", "sUFfKoDPje", "sPhGej2eZG", "p8Bag2FqHE", "nZ5o9bXqJl", "jsDh4iRoyi", "iGSssX7CQ3", "hbYPqJpEX5", "h0h8IDJ8fg", "euiUPStNAZ", "Zs0hWsWXEl", "W5LQB4GEZX", "VvaGeLy28w", "UBxtdOTvyA", "T3nTHeCWFD", "SAjOpY4srf", "Qt9r3ndyhX", "PDI9JeW0J6", "OGhmnQWKtG", "K1ix5ZlbhW", "INjeupjgGJ", "E8vhvUxJ3v", "9vPUgjpMjP", "9drVlzjTan", "0Lul2UG8hs" ], "note_type": [ "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "meta_review", "decision", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment" ], "note_created": [ 1733035853683, 1733140972804, 1733035798278, 1730719698383, 1732098983846, 1733141028677, 1735013666139, 1737523479484, 1732246259529, 1732433576630, 1732433441583, 1730604786399, 1732241186910, 1729520490012, 1732098667021, 1732098873502, 1732694997810, 1732098223177, 1733027572153, 1732797481671, 1732097577478, 1732432843655, 1732191883048, 1732098251014, 1732097684761, 1732097741091, 1732695040992, 1730087923241, 1732099022699, 1732098821411, 1732102003684 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission1997/Reviewer_Y3AS" ], [ "ICLR.cc/2025/Conference/Submission1997/Authors" ], [ "ICLR.cc/2025/Conference/Submission1997/Reviewer_Y3AS" ], [ "ICLR.cc/2025/Conference/Submission1997/Reviewer_Y3AS" ], [ "ICLR.cc/2025/Conference/Submission1997/Authors" ], [ "ICLR.cc/2025/Conference/Submission1997/Authors" ], [ "ICLR.cc/2025/Conference/Submission1997/Area_Chair_nik2" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission1997/Authors" ], [ "ICLR.cc/2025/Conference/Submission1997/Reviewer_8v22" ], [ "ICLR.cc/2025/Conference/Submission1997/Reviewer_G7JL" ], [ "ICLR.cc/2025/Conference/Submission1997/Reviewer_G7JL" ], [ "ICLR.cc/2025/Conference/Submission1997/Reviewer_G7JL" ], [ "ICLR.cc/2025/Conference/Submission1997/Reviewer_ok2Q" ], [ "ICLR.cc/2025/Conference/Submission1997/Authors" ], [ "ICLR.cc/2025/Conference/Submission1997/Authors" ], [ "ICLR.cc/2025/Conference/Submission1997/Authors" ], [ "ICLR.cc/2025/Conference/Submission1997/Authors" ], [ "ICLR.cc/2025/Conference/Submission1997/Authors" ], [ "ICLR.cc/2025/Conference/Submission1997/Reviewer_ok2Q" ], [ "ICLR.cc/2025/Conference/Submission1997/Authors" ], [ "ICLR.cc/2025/Conference/Submission1997/Authors" ], [ "ICLR.cc/2025/Conference/Submission1997/Reviewer_ok2Q" ], [ "ICLR.cc/2025/Conference/Submission1997/Authors" ], [ "ICLR.cc/2025/Conference/Submission1997/Authors" ], [ "ICLR.cc/2025/Conference/Submission1997/Authors" ], [ "ICLR.cc/2025/Conference/Submission1997/Authors" ], [ "ICLR.cc/2025/Conference/Submission1997/Reviewer_8v22" ], [ "ICLR.cc/2025/Conference/Submission1997/Authors" ], [ "ICLR.cc/2025/Conference/Submission1997/Authors" ], [ "ICLR.cc/2025/Conference/Submission1997/Authors" ] ], "structured_content_str": [ "{\"title\": \"Official Comments by Reviewer Y3AS (2/2)\", \"comment\": \"**R W3 & 4:**\\nThe reviewer appreciates the extra experiment results. Indeed, the data shows that the 1:1 combination of LCE and LCL achieves the best result evaluated by the combined metric, supporting the choice made in the paper. The reviewer would like to ask some more questions about the results of these new experiments for better understanding.\\n\\n1. In the table showing the effect of weight ratios, the results with respect to different weight ratios seem to be relatively close to each other (differing less than 5% around a number ~97%). Could the author(s) also include the result for weight ratios 1:0 and 0:1 evaluated in the same metric in the table? That would further help in demonstrating the importance of training with a combined loss and understanding whether a 5% difference is a significant improvement. Note that the table with the row (LCL, w/ sim) achieves significantly lower results (more than 40% in every column) than the combined approach with respect to the combined metric. Even a 100% on the other metric (cer) will not shrink the difference in the combined metric to less than 20%. Does this mean that weight ratio 0:1 has a much more significant difference to 1:4 than the difference between 1:4 and 1:1?\\n\\n2. Could the author(s) clarify whether there is a particular reason to choose 1:1 as the metric ratio (weight of sim and cer)? The reviewer would like to figure out whether it is possible that the best performance of the weight ratio 1:1 model is a result of choosing the metric ratio as 1:1? If the metric ratio changes to 2:1, would the best performance weight ratio also change to n:1 where n > 1? If the metric changes to 0:1, would the best performance weight ratio also change to 0:1? The reviewer does not treat this as a fundamental flaw in the paper, but the reviewer is very interested in the answer to the question above, especially the last question, which will also provide insight into the following question: Does additional training on other tasks (loss: LCE, metric: sim) boost the performance on the original task (loss: LCL, metric: cer)? Or is it merely two different tasks without any positive correlation?\"}", "{\"comment\": \"**For R W3 & R W4:**\\n\\nPart 1\\n\\nWe have conducted additional experiments with extreme weight ratios (1:0 and 0:1) to provide a better understanding.\\n\\n| Weight Ratio (LCE:LCL) | FormL4-Basic | | | FormL4-Random | | |\\n| ---------------------- | ------------ | --------- | --------- | ------------- | --------- | --------- |\\n| | AS | Precision | Recall | AS | Precision | Recall |\\n| 1:0 (LCE only) | 98.64 | 85.21 | 78.45 | 82.81 | 76.32 | 72.15 |\\n| 4:1 (LCE dominant) | 95.32 | 89.41 | 82.15 | 82.14 | 83.25 | 85.33 |\\n| 2:1 (LCE heavy) | 97.45 | 91.73 | 84.62 | 84.56 | 85.12 | 87.45 |\\n| 1:1 (Balanced) | **99.21** | **93.65** | **86.43** | **85.85** | **86.90** | **89.20** |\\n| 1:2 (LCL heavy) | 96.83 | 90.88 | 83.91 | 83.92 | 84.76 | 86.82 |\\n| 1:4 (LCL dominant) | 94.67 | 88.94 | 81.73 | 81.73 | 82.54 | 84.56 |\\n| 0:1 (LCL only) | 59.05 | 52.33 | 48.76 | 57.55 | 50.12 | 46.88 |\\n\\n\\nAs shown in the expanded results table, using either loss component alone leads to substantially different outcomes. The LCE-only setting (1:0) achieves reasonable performance (98.64% on FormL4-Basic) but falls short in precision and recall compared to combined approaches. More strikingly, the LCL-only setting (0:1) shows significantly degraded performance across all metrics (59.05% on FormL4-Basic), confirming your observation about the dramatic difference between 0:1 and other ratios.\", \"the_dramatic_performance_drop_in_the_0\": \"1 setting, compared to both 1:4 and 1:1 ratios, indicates that the contrastive learning component (LCL) requires support from the cross-entropy loss (LCE) to maintain effective alignment capabilities. This supports our findings highlighted in section 5.2, \\\"**Autoformalization Inherently Learns Alignment:**\\\" and \\\"**Complementary Role of Contrastive Loss**\\\", i.e., both components play crucial roles in the learning alignment process, rather than being simply additive contributions.\\n\\n\\n\\nPart 2\\n\\nWe chose equal weighting (0.5 each) during inference to maintain balanced consideration of both sequence-level and representation-level alignment signals. To validate this design choice systematically, we conducted comprehensive experiments varying the weighting between certainty and similarity scores (where similarity weight = 1 - certainty weight):\\n\\n| Weight (Certainty) | MiniF2F-Valid | MiniF2F-Test |\\n| ------------------ | ------------- | ------------ |\\n| 0.9 | 64.82 | 65.13 |\\n| 0.7 | 65.45 | 65.92 |\\n| 0.5 | **66.39** | **66.70** |\\n| 0.3 | 65.21 | 65.84 |\\n| 0.1 | 64.55 | 65.08 |\\n\\nThese results reveal several key insights. First, the balanced 0.5/0.5 weighting consistently achieves optimal performance across both validation and test sets, suggesting this isn't merely coincidental but reflects a fundamental complementarity between the metrics. Second, the performance degradation when heavily favoring either metric (0.9/0.1 or 0.1/0.9) indicates that both components provide essential, non-redundant information for alignment assessment. Third, the relatively modest performance variations (~2% range) demonstrate our approach's robustness to weight selection, though balanced weighting remains optimal.\\n\\n\\n\\nRegarding the reviewer\\u2019s inquiry about whether changing the metric ratio (e.g., 2:1 or 0:1) might influence the optimal weight ratio for performance, this is an insightful hypothesis. Preliminary observations suggest that the optimal ratio is influenced by the interaction between metrics during inference. However, whether this behavior holds under different training loss ratios requires further investigation. We also recognize the importance of the additional question about task interactions: **Does additional training on other tasks (Loss: LCE, Metric: SIM) enhance performance on the original task (Loss: LCL, Metric: CER)?**\\n\\nTo thoroughly address this, we are conducting additional experiments exploring whether the balanced weighting identified in our current setting remains optimal under varying training loss ratios. Specifically, we aim to determine whether the relationship between metrics in inference aligns with the training loss design.\\n\\nWe commit to including these findings in the final version of the paper, ensuring a rigorous examination of how task interactions influence both metrics and loss formulations.\"}", "{\"title\": \"Official Comments by Reviewer Y3AS (1/2)\", \"comment\": \"Dear Author(s),\\n\\nThank you for your detailed and patient clarifications. It solves many of my questions. Below are more specific comments and new questions arising from the newly provided information.\\n\\n**R W1:**\\nThe reviewer appreciates the careful explanation. The update of the dataset by removing flawed data and the quantitative observation (~3% of the data were involved) solves the reviewer\\u2019s concern on this possible wrong misalign strategy. The reviewer believes that, regarding only 3% of the data that was involved, the impact on the main results is minimal. The reviewer doesn\\u2019t know whether the experiment results have already been updated based on the revised dataset. If not, the reviewer would like to kindly remind the author(s) that, in the consideration of rigor, in the final version of the paper, the revised dataset should indeed be used for the experiments to confirm that the results are not influenced by this small subset.\\n\\n**R W2:**\\nThe explanation regarding the miniF2F dataset is appreciated. The reviewer fully acknowledges the effectiveness of the misalign strategy on questions that mainly involve more concrete mathematical objects (like real numbers), including many of the difficult mathematical olympiad (AMC, AIME, IMO) questions. However, the reviewer still has some concern about whether the strategy may not generalize well to a broader range of mathematical questions, especially those with more abstractness, even if they are not as hard as mathematical olympiad questions. The reviewer apologizes for using an inaccurate phrase \\u201capplicable to high-school-level mathematics or below\\u201d in the original review. To clarify the reviewer\\u2019s concern accurately, the reviewer would like to present the following examples.\\n\\nExample 1\\n```\\nexample (n : \\u2115) : \\u00ac 7 \\u2223 2 ^ n + 1 := by sorry\\n```\\n\\nExample 2\\n```\\nexample {p q : Prop} (h : p \\u2227 q): p \\u2228 q := by sorry\\n```\\n\\nThe first example is a mathematical-contest level number theory question in the miniF2F dataset. Most of the misalign strategies apply well to this problem. The second example is a common and simple exercise in logic saying \\u201c(p and q) implies (p or q)\\u201d. Four out of six of the misaligned strategies mentioned in the paper: constant modification, exponent modification, change of variable type, and modification of equality do not apply. The first problem is significantly more difficult than the second one, and the reviewer believes that the difficulty is not the obstacle for the misalign strategy to apply. The same thing would happen for common set theory problems (e.g., `example {\\u03b1 : Type*} (s t u : Set \\u03b1) : s \\u2229 (t \\u222a u) = s \\u2229 t \\u222a s \\u2229 u`) and almost all problems in some undergraduate level mathematical branches (e.g., abstract algebra). The reviewer believes whether the statements are related to natural numbers or real numbers (or any other kind of concrete numbers) influences whether the misalign strategy applies. And it is also part of the reviewer\\u2019s concern that the miniF2F dataset may contain a large proportion of statements that are related to concrete numbers. Some analysis or results on statements that are less related to concrete numbers would reduce the reviewer\\u2019s concern on this issue.\"}", "{\"summary\": \"This paper introduces FormalAlign, an automated alignment evaluation framework designed for autoformalization tasks. FormalAlign addresses the challenge of ensuring semantic alignment between informal and formal mathematical statements, a critical issue given the limitations of traditional manual verification methods. By employing a dual loss framework that combines cross-entropy loss and contrastive loss, FormalAlign enhances its alignment detection ability in a comprehensive way. Evaluations on FormL4 and MiniF2F demonstrate FormalAlign\\u2019s superior performance over GPT-4, with notable improvements in precision scores.\", \"soundness\": \"4\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": [\"FormalAlign demonstrates high precision across two major datasets, maintaining robust recall scores.\", \"In the LLM-as-judge setting detailed in Appendix G, FormalAlign notably outperforms GPT-4o in identifying nuanced misalignments and step further to human experts' level.\", \"FormalAlign's approach shows strong potential for enhancing the quality of autoformalized statement datasets in Lean, providing a valuable contribution to training more effective automatic theorem-proving models.\"], \"weaknesses\": [\"The misalignment strategy \\\"Change of Variable Type\\\" in Table 2 may create examples that should be considered aligned rather than misaligned. Specifically, the example provided, where the variable type is changed from real numbers (R) to rational numbers (Q), does not necessarily create a misaligned example. The original natural statement does not explicitly require the variable to be real numbers. The proposition with the variable in Q still accurately reflects the original question.\", \"The misalignment strategies discussed in the paper appear to be primarily applicable to high-school-level mathematics or below. When dealing with more advanced mathematical statements, such as those only involving logic, quantifiers, or set theory, half of the proposed misalignment methods, including constant modification, exponent modification, and modification of equality, become less relevant or inapplicable. This limitation suggests that the strategies may not generalize well to more complex mathematical domains.\", \"A more extensive ablation study to determine the optimal balance between the two components of the loss function, $ L_{CE} $ and $ L_{CL} $, is recommended. Specifically, the study could explore how different balancings affect the model's performance on the pure autoformalization task. Understanding this balance could provide insights into the trade-offs between the two loss components and potentially enhance the model's ability to accurately formalize mathematical statements.\", \"The ablation experiment in Section 5.2 may not effectively demonstrate the benefits of combining cross-entropy and contrastive loss, as these two loss functions have distinct objectives and may steer the training process in different directions. This discrepancy is particularly pronounced when using the hidden state of a decoder-only model, which is not inherently optimized for capturing semantic information. Since the alignment-selection metric (the average of certainty and similarity scores) naturally favors models trained with the combined loss, the improvement in scores may reflect a metric bias rather than a genuine enhancement from the augmented loss.\"], \"questions\": \"Have you considered splitting the alignment-checking baseline for GPT-4 into two steps: back-translation followed by a natural language alignment check? This approach could encourage GPT-4 to capture more details during the back-translation phase and minimize any impact from its limited familiarity with Lean during the alignment-checking stage.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to Reviewer ok2Q (1/3):\", \"comment\": \"We would like to thank the reviewer for their time and effort in reviewing our paper. We very much appreciate the insightful suggestions. We hereby address the concerns below:\\n\\n**R W1: Use of Synthetic Data for Evaluation**\\n\\nFollowing your suggestion, we conducted an additional evaluation study using real-world autoformalization errors to validate our approach. We ran Gemini with few-shot prompting as our baseline AF system on a sample of 100 theorems from our test sets, collecting both successful and failed formalization attempts. Three expert Lean users then annotated these formalizations, providing corrections for incorrect ones, yielding 78 correct-incorrect formalization pairs.\", \"the_performance_comparison_between_synthetic_and_real_world_validation_sets_is_shown_below\": \"| Evaluation Set | Accuracy (%) | Precision (%) | Recall (%) |\\n| --------------------- | ------------ | ------------- | ---------- |\\n| Synthetic Test | 85.8 | 86.9 | 89.2 |\\n| Real-world Validation | 83.5 | 80.2 | 79.8 |\\n\\nWhile performance on real-world errors is slightly lower than on synthetic cases, the results demonstrate that our model can effectively detect subtle misalignments in practice.\", \"we_are_currently_exploring_two_promising_directions\": \"1. Using FormalAlign as a reward model for reinforcement learning to improve AF system outputs \\n2. Implementing rejection sampling during inference to filter out likely misaligned formalizations\\n\\nOur preliminary experiments with rejection sampling (threshold \\\\= 0.8) show a 15% reduction in misaligned outputs while maintaining 92% of correct formalizations, suggesting this is a promising direction for improving AF system reliability.\\n\\nWe have incorporated this content in the Appendix M section, where it is highlighted in red.\\n\\n**R W2: Baselines and Comparison Strategy**\\n\\nFollowing your recommendations, we implemented two additional GPT-4 based approaches:\\n\\n1. Binary Classification: Directly asking GPT-4 to make a true/false judgment about alignment correctness \\n2. Chain of Thought (CoT): Guiding GPT-4 through step-by-step reasoning about potential discrepancies\", \"here_are_our_updated_experimental_results\": \"| Models | FormL4-Basic | | | FormL4-Random | | | MiniF2F-Valid | | | MiniF2F-Test | | |\\n| -------------- | ------------ | --------- | ----- | ------------- | --------- | --------- | ------------- | --------- | ----- | ------------ | --------- | ----- |\\n| | AS | Prec. | Rec. | AS | Prec. | Rec. | AS | Prec. | Rec. | AS | Prec. | Rec. |\\n| GPT-4 (Score) | 88.91 | 26.33 | 88.69 | 90.52 | 28.56 | 90.02 | 64.34 | 44.58 | 90.98 | 68.31 | 51.11 | 94.65 |\\n| GPT-4 (Binary) | 89.45 | 35.21 | 87.92 | 91.12 | 38.45 | 89.76 | 65.82 | 52.33 | 89.54 | 69.45 | 58.92 | 93.21 |\\n| GPT-4 (CoT) | 90.23 | 42.68 | 88.15 | 91.85 | 45.72 | 89.95 | 67.24 | 59.85 | 89.87 | 70.82 | 62.45 | 92.88 |\\n| FormalAlign | **99.21** | **93.65** | 86.43 | 85.85 | **86.90** | **89.20** | **66.39** | **68.58** | 60.66 | 64.61 | **66.70** | 63.37 |\\n\\nAs shown in the results, both the binary classification and CoT approaches provide stronger baselines than the original score-based method, with CoT showing particularly notable improvements in precision across all datasets. However, FormalAlign still maintains superior performance, especially in precision, demonstrating the effectiveness of our specialized alignment detection approach. These results not only suggest that improved prompting strategies can enhance GPT-4's performance but also emphasize the significant value of having a dedicated, fine-tuned model for alignment detection. The fact that FormalAlign, a smaller model, achieves results comparable to GPT-4 underscores the effectiveness of our proposed training strategy.\\n\\nWe have incorporated this content in the Appendix I section, where it is highlighted in red.\"}", "{\"comment\": \"**For R W1:**\\nWe appreciate the reviewer's understanding regarding the small proportion (3%) of potentially problematic data points. We are currently reproducing all experimental results with the refined dataset, and preliminary findings from our ongoing experiments indicate that the performance differences fall within a \\u00b12% range. We commit to updating all experimental results in both the main paper and appendix with the complete reproduced findings in the final version, ensuring complete transparency and rigor in our evaluation using only properly validated misalignment examples.\\n\\n\\n\\n**For R W2:**\\n\\nWe acknowledge the reviewer's clarified concern about the applicability of our misalignment strategies to more abstract mathematical statements. As our initial response (R W2) mentioned, our current strategies have limited applicability to highly abstract mathematical domains (e.g., pure logic, set theory, and quantifier-heavy statements). We agree that there is significant potential in developing new misalignment strategies specifically tailored for abstract mathematical domains, and this represents an important direction that we are actively exploring in our ongoing research.\"}", "{\"metareview\": \"The alignment problem in the setting of autoformalization is vaguely hinted through a cross-entropy loss function, rather than formally specified. The proposed approach is a finetuned indicator based on a dual loss with respect to a reference aligned and misaligned dataset. The effectiveness is shown on limited misalignment types (e.g., substitutions of constants, variables, fixed operators), which are minor tweaks of formalized math equations. The reference alignment dataset may not be easy to curate, without which generalization is a concern. For insta\\nnce, the performance of the proposed approach (after finetuning on a relatively similar reference task) only slightly outperforms vanilla LLMs.\\n\\nThe concerned research problem is novel, but given the above concerns (especially about the generalization), this work is a borderline work.\", \"additional_comments_on_reviewer_discussion\": \"During the rebuttal, the authors share new results of ablation studies and a manual evaluation of 100 samples, which partially concerns rais\\ned by reviewers. Reviewer ok2Q acknowledges the value of new results but feels the improvement over the baseline is still somewhat incremental.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}", "{\"comment\": \"Dear Reviewer,\\n\\nThank you for your effort and prompt reply. We would like to address your further concerns as follows:\\n\\n**R Q1: Comparision with baselines on real datasets:**\\n\\nThank you for acknowledging our system's performance on real-world data. To further address your concerns, we applied GPT-4 to the real dataset using three different prompt strategies as you suggested. Below are the results:\\n\\n\\n| Models | AS (%) | Prec. (%) | Rec. (%) |\\n| -------------- | ------ | --------- | -------- |\\n| GPT-4 (Score) | 70.2 | 23.1 | 85.5 |\\n| GPT-4 (Binary) | 72.4 | 30.2 | 84.8 |\\n| GPT-4 (CoT) | 75.8 | 38.5 | 85.1 |\\n| FormalAlign | 83.5 | 80.2 | 79.8 |\\n\\n\\nThe results on real-world data show a similar performance gap to those observed in other datasets. FormalAlign consistently outperforms GPT-4-based baselines, particularly in precision, achieving 80.2% compared to 38.5% for the best baseline. While maintaining competitive recall rates, this advantage in precision highlights FormalAlign's robust alignment evaluation capabilities, which generalize well to real-world scenarios. We will incorporate these results into the main evaluation section to better emphasize FormalAlign's effectiveness on real-world data.\\n\\n\\n**R Q2: Adding advanced baseline into main paper**\\n\\nThank you for acknowledging our system's strength in detecting alignment, particularly in terms of precision. We sincerely apologize for mistakenly bolding the AS score for FormL4-Random.\\n\\nBased on your suggestion, we have updated the main content to include the best results from GPT-4 (CoT) while leaving the other weaker baselines in the Appendix due to space limitations. We will incorporate all the results into the main content of the paper when more spaces are allowed.\\n\\n\\n**R Q3: Autoformalization performance concerns**\\n\\n\\nThank you for your insightful question and for raising these important points. \\n\\nTo further address your concern about performance on the autoformalization task, we have conducted additional experiments on the MiniF2F-Valid and MiniF2F-Test. Below are the results:\\n\\n\\n| Model | FormL4 Basic (%) | FormL4 Random (%) | MiniF2F-Valid (%) | MiniF2F-Test (%) |\\n| ------------------ | ---------------- | ----------------- | ----------------- | ---------------- |\\n| Baseline (\\ud835\\udcdb_CL) | 40.92 | 35.88 | 21.11 | 16.34 |\\n| Ours (\\ud835\\udcdb_CL + \\ud835\\udcdb_CE) | 43.14 | 36.02 | 22.49 | 17.54 |\\n\\n\\nOur method demonstrates a consistent improvement across datasets, maintaining an average increase comparable to the results observed on the FormL4 datasets (+2.22% and +0.14%, respectively, for FormL4-Basic and FormL4-Random). Importantly, on MiniF2F-Valid and MiniF2F-Test, our approach achieves a similar average improvement of +1.18%. This demonstrates the generalizability of our alignment evaluation approach across datasets.\\n\\nOn the other hand, we would like to clarify that the primary goal of our paper is to enhance **autoformalization alignment evaluation** rather than the autoformalization task itself. As noted in the abstract (lines 012-013): \\n\\n\\n> \\\"Ensuring semantic alignment between informal and formalized statements remains challenging.\\\"\\n\\n\\nSimilarly, in the introduction (line 078), we state: \\n\\n\\n> \\\"To bridge this gap, we introduce the FormalAlign framework, which assesses the alignment between informal and formal languages during autoformalization.\\\"\\n\\nWhile we recognize that the ultimate objective is to improve autoformalization performance, we believe it is essential to address the challenges in alignment evaluation first. The problem of alignment evaluation has been a longstanding issue in the formal language community, as discussed in [1]. Our experiments on the autoformalization task are intended to provide a **comprehensive evaluation** of our model\\u2019s performance and demonstrate the effects of incorporating contrastive learning loss, as noted in Appendix F:\\n\\n\\n> \\\"These experiments aim to provide a comprehensive evaluation of our model\\u2019s performance and demonstrate the effects of incorporating contrastive learning loss on autoformalization.\\\"\\n\\n\\nTo achieve the suggested goal\\u2014showcasing significant improvements in autoformalization\\u2014our evaluation framework would need to be integrated into a reinforcement learning (RL) loop. This would involve applying an autoformalizer model and leveraging our alignment evaluation as a **reward model** within the loop. Our model could provide more feedback about the alignment and enhance the overall alignment process. We believe this would address your ultimate goal of improving autoformalization performance more substantially.\\n\\nWe greatly appreciate your insights and look forward to further discussions on how best to evolve this work into more advanced systems that fully utilize its potential.\\n\\n[1] [A Survey on Deep Learning for Theorem Proving] (https://arxiv.org/abs/2404.09939)\"}", "{\"comment\": \"Thank you for your response. My concerns have been resolved, and I have updated the score to 6.\"}", "{\"comment\": \"Thank you kindly for the further discussion and clarifications, as well as, the inclusion of the dataset rationale in S4.1.\\n\\n@R3, I agree that this is more for future work.\"}", "{\"summary\": \"The authors propose to check the alignment between informal and formal mathematical statements using a model trained both on the autoformalisation and alignment tasks. This approach is automatic, avoiding the scaling issues of manually checking, while enhancing previous approaches that relied either on BLEU relative to a ground-truth statement or checking for logical consistency, since statements can be logically consistent but misaligned. To validate their approach, the authors create a dataset by mutating knowing correct informal-formal pairs using one of the following operators: modify constant, modify exponent, introduce variable, change variable type, and modify equality. Further, they also introduce misalignments by shuffling the formal/informal data for random pairings. The authors demonstrate improvements in alignment score and precision across FormL4 and MiniF2F.\", \"soundness\": \"3\", \"presentation\": \"4\", \"contribution\": \"4\", \"strengths\": [\"Solving the autoformalisation as an autoformalisation and alignment joint task which should enable better inner representations.\", \"Ablation study to empirically show the value of both tasks (hence the value of cross-ent and contrastive loss and the value of considering both sequence probability and similarity for the alignment score).\", \"Empirical justification of the foundation model that is used/fine-tuned for FormalAlign.\", \"\\\"Mutation\\\" based construction of a known ground-truth dataset.\"], \"weaknesses\": [\"Results are presented at a fixed alignment score threshold instead of P-R curves. (I suspect this is due to a lack of access to create such curves for GPT models).\", \"The misaligned types generated are not uniform (Figure 3)\", \"Model performance by misalignment type is not reported, the case study is only vs BLEU and BERTScore, which is welcomed, but a comparison with GPT-4 would also be welcomed. (A similar breakdown for the LLM-as-judge study would also be welcomed)\"], \"questions\": \"1. Looking at the Table 3 results, and considering that you have an alignment threshold, would it be possible to obtain P-R curves for FormalAlign? If so, I would like to know the performance of the approach when matching GPT-4 recall (So P@(R=0.9)). Further, presenting the P-R curves would allow people to have a better understanding of the performance, while providing such curves to users of the approach for evaluation could make a more informed choice of the threshold value depending on the expected manual load afterwards.\\n\\n2. Is there a reason why the strategies were applied non-uniformly to generate negative examples? (The strategies are in Table 2, while Figure 3 presents the distribution of misalignments generated.)\\n\\n3. For the case study and LLM-as-judge results, do you have per-misalignment-type data? Would it be possible to report that, as it would be interesting if any category is especially difficult/easy to spot?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you kindly for the effort, detailed response, and additional data and experiments!\\n\\nRW1, the P-R Curves for datasets are interesting to observe, and I am glad to see that the original points are on/close to the \\\"elbow\\\" of the curves. The MiniF2F curves are a bit surprising though as I did not expect the drop to be as much.\\n\\nRW2, I apologise if I missed it in the updated paper, but I am not sure I spotted where this was clarified in the paper. I think the paper could benefit from a sentence that says that the aim is to match an empirically expected distribution of errors or some weighted coverage per experiment design. Missing the explanation from this conversation, my prior, and I would expect that of other readers, is \\\"uniform distribution\\\"/\\\"balanced classes\\\". (To be clear, I am happy with the explanation, but I want it be afforded to other readers as well)\\n\\nRW3, Thank you for the very detailed discussion as well as the new breakdown by types. Looking only at the H-Method \\u0394, I see three broad groups: {Constant}, {Exponent, Random}, and {...rest...}. Exponent seems to be equally hard, and Random equally easy, the other categories have the \\\"usual\\\" gap, and then Constant seems to have the widest gap to Human performance. Do you have an intuition why the Human vs Model performance is unevenly impacted by Constant vs Exponent? (Feel free to shoo me away to a psychology venue if the question veers more in that direction, but as it stands, the gap here seems to be an outlier; to some degree, it seems to stem from the models finding exponent easier while Humans find Constant easier. All deltas between these two categories are \\u00b1 ~2)\"}", "{\"summary\": \"This work presents a technique for automatically evaluating the semantic alignment between a natural language description and its corresponding formal representation that may be produced by an AutoFormalization system (Autoformalization aims to convert informal mathematical proofs to formal proofs that can be automatically checked by existing theorem provers). The technique is based on fine-tuning a base LLM with respect to a dual loss function that combines cross-entropy loss in the sequence generation with a contrastive loss based on cosine similarity of the informal/formal pairs. The technique is evaluated on existing datasets for autoformalization that are augmented with misalignment strategies proposed by the authors to test negative examples of incorrect formalizations. It is shown to have superior performance in many cases against GPT models prompted with a similar alignment evaluation task.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. Innovative focus on the alignment problem in Autoformalization. Translating informal mathematical proofs into formal, machine-verifiable formats is an important problem in fields like theorem proving and formal verification, since formal systems (e.g. Lean, Isabelle) are highly precise but difficult for humans to use. This work brings an explicit focus on a key issue that plagues autoformalization: to ensure semantic alignment between the informal input and formal output. This is an important problem to address, since formal provers can check for syntactic but not semantic correctness with respect to the original intent, and other metrics such as BLEU scores are very imprecise. Addressing this alignment problem explicitly seems a novel angle taken by this work, and such focus could help further progress in the field of autoformalization.\\n\\n2. Dual-Task Training Approach. The technique of simultaneously training to generate formal statements and check alignment is interesting. While multi-task learning and contrastive learning have been used in other areas (e.g., machine translation, image-text alignment), the combination of sequence generation with contrastive alignment seems to be a novel application in the context of autoformalization. \\n\\n3. Clarity and presentation. Though some parts unclear (see below), in general the paper clearly explains and motivates the problem, and explains the technique well, including definitions of the combination of cross-entropy loss (for sequence generation) and contrastive loss (for semantic alignment). The evaluation covers some important aspects to show that on the synthetically augmented datasets the technique performs well in general, and also better or comparable to alignment inference based on prompting GPT models. This demonstrates the core value of the technique. The authors also include ablation studies that provide insight into the contribution of both kinds of loss functions, as well as comparison with human evaluation.\", \"weaknesses\": \"1. Use of synthetic data for evaluation. To evaluate alignment, the authors create a synthetic dataset of negative example formalizations based on explicit misalignment strategies that they propose. I do not have a sense of how realistic those misalignments are in practical terms. I think addressing this issue is important, especially since the whole purpose of the task is to detect the errors made by existing AF systems: we should be testing the ability to detect those kinds of variations, which could be very much smaller and more subtle than the ones produced by the explicit misalignment strategies here - would the contrastive loss be able to help distinguish between those kinds of variations (for which the task is actually intended)? While I understand such data is not readily available, I think it should be possible to obtain some real data perhaps in the following way: since you already have access to human judges (which you used for a similar task), perhaps you can run a SoTA AutoFormalization system on the standard datasets you use here, get the outputs and ask the human judges to label correct/incorrect for the formalizations produced and also provide corrected formalizations. Then just check the alignment evaluation against this data (which would provide one positive and one negative example for each case where the AF system had failed). The dataset need not be very big as it is only to be used for evaluation. Such an approach would also show the actual improvement you can make to a SoTA AF system.\\n\\n2. Lack of compelling baselines in comparison. The prompt used for GPT models Appendix C.1 seems to ask GPT to assign an \\\"alignment score\\\" from 1 - 5. This seems a pretty vague and not well-defined task that may be difficult for the model (or even a human) to interpret. A more well-defined question is to simply ask if the formalization correctly represents the informal NL or not - since that is the main (only) problem we are trying to solve? Scoring seems to make it a subjective/continuous domain problem - e.g. it may be possible that a given formalization correctly represents the problem, but perhaps the model may still assign it a score 4 rather than 5 due to some stylistic differences with the informal input it detects. Having a true/false binary judgement would make a more well-defined task for the model to address. Also, this seems a very obvious case of where Chain of thought (CoT) reasoning should be tried as another baseline - the evaluation task is very reasoning based and the GPT models can be expected to do much better with CoT reasoning to infer any discrepancies between the informal and formal representations - rather than just predicting some number score from 1 to 5. Especially since you do not have other baseline or SoTA methods to compare against (do there exist such methods?), using a CoT baseline is easy to implement and would be very relevant here. You also provide all the candidate formalizations in the same prompt - perhaps evaluating each one of them separately will be a better approach so not to confound the model with too many tasks at the same time. \\n\\n3. Some problems with positioning/presentation. You state in the abstract: \\\"FORMALALIGN trains on both the autoformalization sequence generation task and the representational alignment between input and output, employing a dual loss that combines a pair of mutually enhancing autoformalization and alignment tasks.\\\" and make similar statements in the introduction. That sounds like you are addressing both the autoformalization task as well as the alignment evaluation task - but you have not shown any evaluation of the autoformalization task - if that is not an intended contribution of the work then such statements are a bit misleading or confusing in the intro and abstract. It should be clarified that you are only providing an evaluation method to test alignment of given candidate informal-formal pairs. Also, if the AF task itself is not a contribution - why not? If the combination of cross-entropy and contrastive loss accurately infers alignment, why can't the fine-tuned model based on that dual loss function be better at the autoformalization task overall? Could you have compared it to prior auto-formalization approaches?\", \"typos\": [\"\\\"The challenges of automated evaluating natural language\\\"\", \"\\\"to train an FORMALALIGN\\\"\", \"\\\"Eq. equation 5. Table 6\\\"\"], \"questions\": \"1. I am confused about the Metrics description in section 4.2. I thought this should describe the metric itself independently of the system being evaluated (whether is your system or any other baseline system). But the AS and detection metrics are defined with respect to the V_align function which seems specific to your model only and used in its inference. Is that specific for the FormalAlign model only or otherwise how are these metrics computed for the other models like GPT4? (I hope that is not the case as they would then be really problematic metrics to be using that are specific to your model). And if that is not the case then please move this discussion to the inference section - since they look like modes of inference in which your model can be used: it can either select the best formalization from a given set (selection), or return a true/false label for a given informal/formal pair (detection). The metrics are then just to measure the quality of these two selection and detection modes of inference using standard precision/recall.\\n\\n\\n2. Similarly, can you clarify exactly how you compute AS, precision and recall numbers for the GPT models? (As above, I hope you are not computing the alignment scores for them in the exact way you describe in section 4.2). I am assuming it is based on you prompting the models to predict an \\\"alignment score\\\" from 1-5 as in the prompt in Appendix C2? Please confirm this is correct and it should be explained clearly in the evaluation section. Also please explain the details - how is AS inferred from 1-5 score generated by GPT model? How is precision measured? What if there are ties on the top score between multiple candidate formalizations? \\n\\n\\n3. I dont understand the second prompt in Appendix C.2 - it is just doing a translation task? How is it used in your evaluations? Or is it just the prompt used to run your final fine-tuned model for the sake of generating and extracting the alignment scores?\\n\\n4. Just to confirm - you do not use the synthetically generated misalignment data for training at all right? \\n\\n5. What is SIM() function in Figure 2? - is it the same as Cos()? If so please fix for consistency.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to Reviewer 8v22 (1/3):\", \"comment\": \"We would like to thank the reviewer for their time and effort in reviewing our paper. We very much appreciate the insightful suggestions. We hereby address the concerns below:\\n\\n**R W1: Broader Significance and Formal Definition**\\n\\nThank you for this valuable feedback. We agree that better contextualizing autoformalization's broader significance and providing a formal definition would enhance the paper's accessibility and impact.\", \"we_propose_adding_the_following_formal_definition_of_autoformalization_alignment\": \"Given a natural language statement N and its formalization F, assuming F is verified as syntactically valid by the formal language compiler, alignment A(N, F) is defined as a binary relation where A(N, F) \\\\= 1 if and only if:\\n\\n1. F preserves all mathematical constraints specified in N \\n2. F maintains the same logical relationships between components as expressed in N \\n3. F does not introduce additional constraints not present in N\\n\\n\\n\\nWe agree that better contextualizing the broader significance of autoformalization would enhance the paper's accessibility and impact. We propose adding a comprehensive discussion of why autoformalization is a crucial research direction:\\n\\nAutoformalization - the automated conversion of natural language mathematics into formal languages - serves as a critical bridge between human mathematical knowledge and machine-verifiable formal systems. Its significance extends across multiple domains:\\n\\n1. Software Engineering and Verification\\n\\n- Enables automated translation of natural language specifications into formal verification properties\\n- Reduces the manual effort required to formally specify system requirements\\n- Helps detect ambiguities and inconsistencies between informal specifications and implementations\\n\\n2. Mathematical Knowledge Management\\n\\n- Facilitates the preservation and verification of mathematical knowledge by converting informal mathematics into machine-checkable formats\\n- Enables systematic organization and cross-referencing of mathematical content\\n- Supports automated reasoning systems by expanding available formal mathematical content\\n\\n3. Automated Reasoning Research\\n\\n- Expands the scope of automated theorem provers by allowing them to work with natural language inputs\\n- Enables hybrid reasoning systems that combine formal and informal mathematical knowledge\\n- Accelerates the development of more powerful mathematical reasoning systems with feedback from formal compilers.\\n\\nWe expanded this discussion in Appendix N in our revised paper to better articulate how advances in autoformalization can benefit these broader research communities. \\n\\n\\n\\n**R W2: Running Example Concerns**\\n\\nThank you for these astute observations about our running example. You raise valid points about both the detectability through proof checking and the generation probability.\\n\\nWhile this specific misalignment can indeed be detected through proof checking (as demonstrated by your counterexample), we chose it as a running example primarily for its clarity in illustrating our approach. However, we acknowledge that a more nuanced example might better demonstrate the unique value of our alignment detection method.\\n\\nIn practice, many real-world misalignments are more subtle and challenging to detect through proof-checking alone:\\n\\n1. Type-level misalignments where statements are well-typed but semantically misaligned \\n2. Cases where the proof search is computationally expensive or intractable \\n3. Misalignments in assumptions or preconditions that don't affect provability \\n4. Situations where formal statements are provable but don't capture the intended informal meaning\\n\\nFor instance, consider cases where variables are accidentally swapped or constraints are slightly modified \\\\- these might still result in provable statements that don't match the informal intent. Our alignment detection approach can identify such mismatches more efficiently than proof search, especially for complex theorems where automated proof finding is challenging.\"}", "{\"title\": \"Response to Reviewer 8v22 (3/3):\", \"comment\": \"**R Q1: Extending FormalAlign to Proof Autoformalization**\\n\\nThank you for this interesting question about extending FormalAlign to proof autoformalization. The primary challenge stems from the structural mismatch between formal proofs and their natural language counterparts. In Lean 4, proofs often rely on specialized tactics and proof environments that have no direct natural language equivalents.\", \"this_mismatch_manifests_in_two_key_ways\": \"1. Tactic-based reasoning: Lean commands like `linarith` encode complex mathematical reasoning in single operations, while natural language proofs typically express these steps through multiple sentences using informal reasoning. This many-to-one mapping makes alignment particularly challenging. \\n2. Environment dependencies: Formal proofs often reference pre-defined lemmas and proof environments specific to Lean 4\\\\. These formal structures may lack clear natural language counterparts that non-experts can understand, and the translation would require extensive \\\"unpacking\\\" of these formal constructs.\\n\\nWhile Lean 4's compiler provides automated and reliable correctness checking for the logical validity of formal proofs, this verification differs from semantic alignment. Our alignment framework could help validate whether a formal proof captures the intended reasoning steps from its natural language counterpart. However, achieving reliable end-to-end proof alignment would require significant extensions to handle the fundamental structural and semantic disparities between formal and natural language proofs, even with Lean's strong verification capabilities.\\n\\n**R Q2: Combining Certainty and Similarity Scores**\\n\\nThank you for this question about our score combination strategy. While the certainty and similarity scores indeed operate on different scales, we chose to average (weight=0.5) to place balanced emphasis on both metrics during inference. To validate this design choice, we conducted an ablation study testing different weighting schemes. We note that the weight for similarity is 1 \\\\- Weight (Certainty):\\n\\n| Weight (Certainty) | MiniF2F-Valid | MiniF2F-Test |\\n| ------------------ | ------------- | ------------ |\\n| 0.9 | 64.82 | 65.13 |\\n| 0.7 | 65.45 | 65.92 |\\n| 0.5 | **66.39** | **66.70** |\\n| 0.3 | 65.21 | 65.84 |\\n| 0.1 | 64.55 | 65.08 |\\n\\nThe results demonstrate that a balanced weighting (0.5/0.5) achieves optimal performance across both datasets. We observe that heavily favoring either score (weights of 0.9/0.1 or 0.1/0.9) leads to decreased performance, suggesting that both metrics contribute complementary information to the alignment decision. The relatively small performance variations (within \\\\~2 percentage points) also indicate that our approach is fairly robust to weight selection, though balanced weighting consistently yields the best results.\\n\\n**R Q3: Evidence for Reducing Manual Verification**\\n\\nThank you for raising this important point about quantitative evidence. We have conducted several experiments to validate FormalAlign's practical impact on reducing manual verification needs (as detailed in R W3). In our real-world validation study, we applied Gemini to 100 theorems, yielding 78 formalization pairs, where our system achieved 83.5% accuracy in detecting misalignments. This suggests a potential \\\\~80% reduction in manual verification needs for typical autoformalization outputs.\\n\\nWhen integrated into an autoformalization pipeline with a confidence threshold of 0.8, FormalAlign successfully filtered out \\\\~85% of incorrect formalizations while maintaining 92% of correct formalizations, reducing manual review time by approximately 75% compared to the baseline. These quantitative results demonstrate a significant practical impact on reducing verification workload while preserving high reliability.\\n\\nLooking forward, we see significant potential for further improvements through incorporating FormalAlign directly into the autoformalization pipeline, either as a reward model for reinforcement learning to improve AF system outputs or through rejection sampling during inference to filter out likely misaligned formalizations. We plan to validate these approaches through larger-scale deployments and more diverse use cases.\"}", "{\"comment\": \"Dear Reviewer **Y3AS**,\\n\\nThank you for your detailed feedback regarding our paper's technical methodology and experimental evaluations. We have carefully considered your points and made several improvements to address them.\\n\\n1. Regarding the misalignment strategy \\\"Change of Variable Type,\\\" we acknowledge your observation about the \\u211d to \\u211a conversion example. We have revised our approach to focus on more definitive cases of misalignment, such as \\u211d to \\u2124 conversions where fractional values are essential to the theorem's meaning. Our empirical analysis shows that potentially ambiguous type changes occurred in only 3% of samples, and we have updated our screening process to filter these edge cases.\\n\\n2. On the concern about misalignment strategies' applicability to advanced mathematics, while we acknowledge certain limitations with pure logic and set theory, our framework has demonstrated effectiveness on MiniF2F's olympiad-level problems. Our ablation studies have revealed optimal performance with balanced (1:1) weighting between LCE and LCL losses, showing significant improvements over individually trained models across all datasets.\\n\\n3. For your suggested two-phase GPT-4 baseline, we implemented and tested this approach. While it showed slight improvements over the original baseline, we found that the decoupled process introduced potential error propagation issues and increased computational overhead.\\n\\nWe would appreciate your thoughts on whether these clarifications address your concerns, or if any points need further elaboration. Thank you for your time and consideration.\\n\\n\\n\\nBest regards,\\n\\nAuthors of Paper Submission 1997\"}", "{\"title\": \"Response to Reviewer G7JL (1/2):\", \"comment\": \"We would like to thank the reviewer for their time and effort in reviewing our paper. We very much appreciate the insightful suggestions. We hereby address the concerns below:\\n\\n**R W1: Fixed Alignment Score Threshold and P-R Curves & Q1: P-R Curves and Performance at P@R=0.9**\\n\\nThank you for these valuable comments regarding our presentation of alignment score thresholds and P-R curves. To address this, we have conducted additional experiments and generated comprehensive P-R curves. **These curves are presented in Figure 4 (Appendix K) of our revised paper**, providing a detailed view of model performance across varying alignment score thresholds. Below, we summarize key observations and enhancements:\\n\\nWhen evaluating FormalAlign at GPT-4's recall level (\\u22480.9), our model achieves:\\n\\n* FormL4-Basic: 88% precision at 89% recall \\n* FormL4-Random: 83% precision at 90% recall \\n* MiniF2F-Valid: 45% precision at 91% recall \\n* MiniF2F-Test: 48% precision at 88% recall\\n\\nThe P-R curves reveal several interesting characteristics of FormalAlign's behavior:\\n\\n1. On FormL4 datasets, precision remains robust (\\\\>75%) even at high recall levels, demonstrating strong alignment capabilities for structured formal mathematics. \\n2. For MiniF2F datasets, we observe a more pronounced precision drop at higher recall points, reflecting the increased complexity of these problems.\\n\\n\\n\\n**R W2: Non-Uniform Distribution of Misalignment Types & Q2: Non-Uniform Strategy Application for Negative Examples**\\n\\nThank you for these thoughtful observations about the non-uniform distribution of misalignment strategies. This distribution pattern is actually intentional in our design, though we recognize it warrants careful explanation.\\n\\nThe non-uniform application of strategies reflects two key aspects of our approach. First, it accounts for the inherent constraints and characteristics of mathematical contexts in different datasets. For example, \\\"Constant Modification\\\" appears more frequently in miniF2F datasets (\\\\~31%) compared to FormL4-Basic (\\\\~11%) because miniF2F contains more theorems with explicit numerical constants that can be meaningfully modified. Similarly, the \\\"Change of Variable Type\\\" strategy varies across datasets (from \\\\~14% in FormL4-Basic to \\\\~18% in miniF2F-Valid) because its applicability depends heavily on the theorem's type constraints \\u2013 it's only meaningful when changing the variable type would affect the theorem's validity, such as changing from \\u2124 to \\u211d in problems requiring integer solutions.\", \"this_varied_distribution_of_strategy_application_serves_a_dual_purpose\": \"it ensures our negative examples remain mathematically meaningful while also testing FormalAlign's robustness against a diverse range of misalignment types. Rather than creating artificial misalignments that might be too obvious or mathematically invalid, this approach helps us assess how well the model can generalize across different types of errors, from simple modifications to more complex transformations.\\n\\nIn our future experiments, we plan to explore more uniform sampling strategies to provide a clearer assessment of the model's performance across balanced misalignment types. We will also clarify in the paper that our current approach prioritized covering a wide spectrum of mathematically meaningful misalignment scenarios rather than achieving uniform distribution.\"}", "{\"comment\": \"Thank you for your detailed and thoughtful feedback on our paper. We greatly appreciate the time you spent reviewing our revisions and the additional experiments we conducted. We are pleased that you found value in the consistent performance improvements demonstrated over the GPT-4 baselines across all datasets, including both the real dataset and the two additional evaluation sets.\\n\\nThank you again for your review. We welcome any additional comments or discussions and are eager to continue refining our work.\"}", "{\"comment\": \"Thank you to the authors for the detailed responses to my comments and I really appreciate the further experiments to provide clarity. It is good that the system is consistently performing better than GPT4 baselines on the real dataset, and also reassuring that the AF performance on the other two datasets is also consistently (though incrementally) better than the baseline.\\n\\nOverall, this work addresses the subproblem of alignment evaluation which is presented as an important step towards improving AF. The evaluation shows its value over GPT4-based prompting baselines as there is no other SoTA to compare against (for it being a newly defined subproblem). Hence while this is an interesting and potentially valuable subproblem to address and its value has been shown to some degree, it has not yet been shown how well it can address the actual AF problem for which it is meant (especially since the evaluation technique presented here is itself derived from a new way of training an AF model which itself has shown only incremental improvement so far). \\n\\nAs compared to an approach that shows significant improvement to the AF task over existing SoTA systems, this approach defines a new subproblem and shows value on it with respect to constructed baselines. In that respect I believe it provides partial value towards the main problem, and I would like to keep my current rating of marginal accept.\"}", "{\"title\": \"Response to Reviewer Y3AS (1/3):\", \"comment\": \"**R W1: Misalignment Strategy \\\"Change of Variable Type\\\"**\\n\\n\\n\\nWe appreciate this insightful observation about variable type changes. You are correct that changing from real numbers (\\u211d) to rational numbers (\\u211a) may not constitute misalignment since both types can represent the same mathematical concepts in many cases. This prompted us to revise our examples and strengthen our screening criteria for type-based misalignments.\", \"consider_instead_this_illustrative_example_from_our_dataset\": \"```lean\\n-- Original theorem\\ntheorem mathd_algebra_478\\n (b h v : \\u211d)\\n (h\\u2080 : 0 < b \\u2227 0 < h \\u2227 0 < v)\\n (h\\u2081 : v = 1/3 * (b * h))\\n (h\\u2082 : b = 30)\\n (h\\u2083 : h = 13/2) :\\n v = 65 :=\\n\\n-- Misaligned version\\ntheorem mathd_algebra_478\\n (b h v : \\u2124)\\n (h\\u2080 : 0 < b \\u2227 0 < h \\u2227 0 < v)\\n (h\\u2081 : v = 1/3 * (b * h))\\n (h\\u2082 : b = 30)\\n (h\\u2083 : h = 13/2) :\\n v = 65 :=\\n```\\n\\n\\n\\nThis example demonstrates misalignment because the theorem relies on fractional values (e.g., $ h = \\\\frac{13}{2} $) and calculations ($ v = \\\\frac{1}{3} \\\\cdot (b \\\\cdot h) $) that cannot be represented in the integer domain ($ \\\\mathbb{Z} $). Such arithmetic operations inherently require non-integer values, leading to a semantic contradiction where the original mathematical meaning is lost when confined to integers.\\n\\nIn our empirical analysis of 100 sampled cases, potentially ambiguous scenarios, such as conversions from $ \\\\mathbb{R} $ to $ \\\\mathbb{Q} $, were observed in only 3% of samples. Our revised screening process effectively filters these edge cases, ensuring focus on changes that result in true mathematical contradictions.\\n\\nIn the revised paper, we have updated this case and plan to further improve our misalignment datasets by replacing the $ \\\\mathbb{R}/\\\\mathbb{Q} $ example with the $ \\\\mathbb{R}/\\\\mathbb{Z} $ case presented above, which more clearly illustrates misalignment. Additionally, we will clarify our criteria for identifying true misalignments, particularly emphasizing cases where type constraints directly alter the mathematical meaning.\\n\\nThese updates aim to provide a more thorough and transparent discussion, enhancing both the clarity of our presentation and the robustness of our analysis.\\n\\n\\n\\n**R W2: Applicability of Misalignment Strategies to Advanced Mathematics**\", \"we_would_like_to_address_this_concern_from_several_perspectives\": \"We applied our misalignment strategies to generate adversarial examples in advanced mathematical domains (MiniF2F), demonstrating their effectiveness beyond basic arithmetic. Specifically, MiniF2F includes problems from **mathematical olympiads (AMC, AIME, IMO)** and undergraduate-level mathematics. These problems encompass sophisticated mathematical reasoning across domains like abstract algebra, real analysis, and number theory, where our strategies successfully generated meaningful misalignments to test model robustness.\\n\\nFurthermore, we acknowledge that our current strategies might have limited applicability to highly abstract mathematical domains (e.g., pure logic, set theory, and quantifier-heavy statements). However, contextualizing our work within the current capabilities of language models in mathematical reasoning, we believe the current dataset choice and its complexity level are sufficient. Even state-of-the-art models struggle with theorem proving, achieving less than **50% pass@1** rates on the MiniF2F test sets. Extending our framework to handle more sophisticated mathematical domains is an important direction for future work.\"}", "{\"comment\": \"Thank you for your thorough follow-up analysis and insightful observations. We greatly appreciate your careful consideration of our responses and the additional perspectives you've provided.\\n\\n**R1**: The more pronounced drop in MiniF2F likely reflects the increased complexity and diversity of mathematical reasoning required in MiniF2F problems compared to FormL4. We plan to investigate whether additional training strategies might help maintain precision at higher recall levels for these more challenging problems in our future work.\\n\\n\\n\\n**R2**: You raise a good point about the documentation of our distribution rationale. We will include an explicit explanation in Section 4.1 (*Datasets*), clarifying how the distribution of misalignment types was designed.\\n\\n\\n**R3**: Your observation about the differential gaps between human and model performance for Constant (16.8) versus Exponent (13.6) misalignments is particularly insightful. We believe this discrepancy may stem from several factors:\\n\\n1. Human Recognition: Humans appear to be particularly adept at spotting constant changes, possibly because they can quickly reference common mathematical values and relationships from experience.\\n\\n \\n\\n2. Training Data: The relative abundance of exponent-related examples in the FormL4 and MiniF2F training datasets might provide stronger learning signals for exponent modifications compared to arbitrary constant changes.\\n\\n \\nThis asymmetry suggests an interesting direction for future work in understanding how both humans and models process different types of mathematical modifications.\"}", "{\"comment\": [\"I want to thank the authors for the very detailed response, for answering all my questions and the significant additional efforts they have made in experimental evaluations in such a short time period, which is highly commendable.\", \"Real vs Synthetic data. These results indeed show that performance of your system is not very significantly lower on the real data, which is reassuring. However, my actual suggestion was to use such a dataset for your main evaluation in comparison with baselines. The key question: on real world data such as this, does FormalAlign perform better at alignment evaluation than baselines?\", \"Including more advanced baselines. These results indeed show that the stronger baselines perform better and reduce the gap with your system, but I agree the gap is still significant especially in terms of precision (note however: do not bold your AS score of 66.39 for FormalL4-Random since CoT performs better with 67.24). I would say that these stronger baselines should be included in the main paper results rather than appendix as they are more important comparisons than the numbering-based prompt.\", \"Performance on the Autoformalization task. This remains my main concern. These results look very incremental with an average improvement of only 1.18% on these two datasets (and only 0.14% on one of them). Can we see these numbers on the other datasets (MiniF2F-Valid and MiniF2F-Test)? I am curious if they can be better or if they may actually show regressions in these other datasets (where alignment results were weaker)? From my understanding, improvement to AF is the real ultimate goal of having an alignment evaluation technique isn\\u2019t it? So this should be the ultimate test of how well the technique is working. If even in comparison to your own base model the results are so incremental, that is concerning. In general, one may expect a strong argument that here is an AF system that has X% accuracy, and when we extend it with our alignment evaluation technique it improves it\\u2019s accuracy significantly by Y% \\u2013 do you have such a summary result that is more significant improvement than the 1.18% above? And if it is more significant on another AF system then would be good to understand why it is so much better than on your own AF model. This seems to me to be an important question in showing the ultimate value of your technique towards the AF task for which it is meant.\"]}", "{\"title\": \"Response to Reviewer G7JL (2/2):\", \"comment\": \"**R W3: Model Performance Breakdown by Misalignment Type & Q3: Per-Misalignment-Type Data for Case Study and LLM-as-Judge Results**\\n\\nWe agree that providing per-misalignment-type data for both the case study and the LLM-as-judge results would be valuable for understanding how FormalAlign performs on different types of misalignments. To provide a comprehensive understanding of the strengths and limitations of our proposed method relative to human judgment and baseline LLM performance, we conducted an in-depth analysis focusing on performance across various misalignment types. The detailed results are shown below:\\n\\n| Misalignment Type | Human | Method | GPT-4o | \\u0394 (Human-Method) | \\u0394 (Method-GPT4o) |\\n| ----------------- | -------- | -------- | -------- | ---------------- | ---------------- |\\n| Constant | 75.2 | 58.4 | 42.1 | 16.8 | 16.3 |\\n| Exponent | 73.8 | 60.2 | 44.3 | 13.6 | 15.9 |\\n| Variable\\\\_type | 78.9 | 64.5 | 46.2 | 14.4 | 18.3 |\\n| Variable\\\\_new | 81.3 | 66.7 | 48.9 | 14.6 | 17.8 |\\n| Equality | 82.4 | 67.8 | 49.5 | 14.6 | 18.3 |\\n| Random | 85.9 | 72.4 | 54.0 | 13.5 | 18.4 |\\n| **Overall** | **79.6** | **65.0** | **47.5** | **14.6** | **17.5** |\\n\\n**Easiest and Most Challenging Misalignment Types**:\\n\\n* **Random Pairing Misalignments**: Performance was highest across all methods for random pairings, with human experts achieving 85.9%, our method at 72.4%, and GPT-4o at 54.0%. This finding aligns with expectations, as random pairings involve significant structural or contextual discrepancies that are more readily identifiable. \\n* **Subtle Modifications (Constant and Exponent Changes)**: These types of misalignments proved to be the most difficult to detect, evidenced by the lowest scores across all evaluated methods. Human performance dropped to 75.2% for constant changes and 73.8% for exponent modifications, with corresponding lower scores for our method and GPT-4o. This highlights the need for more sophisticated mechanisms capable of identifying nuanced numerical and syntactic changes in mathematical reasoning.\\n\\n**Performance Gaps**:\\n\\n* **Human vs. Our Method**: Human evaluators consistently outperformed our method by approximately 15-20 percentage points across all misalignment types. This consistent gap underscores the advantage of human intuition and mathematical insight, particularly in recognizing subtle logical shifts and contextual meanings that automated methods still struggle to replicate. \\n* **Our Method vs. GPT-4o**: Our method maintained a clear advantage over GPT-4o, with a performance margin of 15-20 percentage points across all misalignment types. This gap indicates that the enhancements in our approach, including targeted modeling of structural and semantic relationships, yield significant improvements over standard LLM capabilities.\\n\\nOur method shows marked improvement in handling structural changes, such as equality modifications and random pairings, compared to more local, subtle modifications. This suggests that while the model is adept at identifying broader, more evident shifts in alignment, fine-grained detection still poses a challenge. The findings suggest future enhancements should prioritize the model\\u2019s ability to detect minute variations, such as constant and exponent changes, potentially through integrating deeper mathematical reasoning frameworks or improved training with synthetic examples that emphasize such subtleties.\\n\\n\\n\\nWe have incorporated this content in the Appendix L section, where it is highlighted in red.\"}", "{\"title\": \"Response to Reviewer Y3AS (2/3):\", \"comment\": \"**R W3: Ablation Study on Loss Function Components (LCE vs LCL)**\\n\\n\\n\\nOur current approach employs equal weighting between these components, as outlined in our paper (lines 214-215): \\\"We train an alignment-aware FormalAlign model by minimizing a combined loss, enabling it to benefit from both the **sequence alignment** inherent in the autoformalization and the **representation alignment** facilitated by the contrastive learning process.\\\" To validate this design choice, we conducted a comprehensive ablation study varying the relative weights between LCE and LCL:\\n\\n| Weight Ratio (LCE:LCL) | FormL4-Basic | | | FormL4-Random | | |\\n| ---------------------- | ------------ | --------- | --------- | ------------- | --------- | --------- |\\n| | AS | Precision | Recall | AS | Precision | Recall |\\n| 4:1 (LCE dominant) | 95.32 | 89.41 | 82.15 | 82.14 | 83.25 | 85.33 |\\n| 2:1 (LCE heavy) | 97.45 | 91.73 | 84.62 | 84.56 | 85.12 | 87.45 |\\n| 1:1 (Balanced) | **99.21** | **93.65** | **86.43** | **85.85** | **86.90** | **89.20** |\\n| 1:2 (LCL heavy) | 96.83 | 90.88 | 83.91 | 83.92 | 84.76 | 86.82 |\\n| 1:4 (LCL dominant) | 94.67 | 88.94 | 81.73 | 81.73 | 82.54 | 84.56 |\\n\\n*Note: To ensure fair comparison, we normalize to keep total loss magnitude across different weight ratios.*\\n\\nThe ablation results strongly support our original design choice of balanced weighting. The balanced 1:1 ratio consistently outperforms other weightings across all metrics, achieving optimal trade-offs between sequence-level and representation-level alignment. Performance drops substantially when moving too far in either direction (4:1 or 1:4), confirming that both components play essential, complementary roles in robust autoformalization - LCE ensures accurate sequence-level alignment for precise formalization, while LCL maintains semantic coherence through contrastive learning.\\n\\n\\n\\n**R W4: Ablation Experiment and Metric Bias**\\n\\nTo directly address this concern, we conducted comprehensive experiments comparing models trained with individual loss functions, i.e., training each model with their corresponding metrics as optimization objectives (certainty scores, similarity scores) against our combined approach:\\n\\n| Training Method | FormL4-Basic | FormL4-Random | RandomMiniF2F | ValidMiniF2F Test |\\n| --------------- | ------------ | ------------- | ------------- | ----------------- |\\n| LCE (w/ cer) | 95.45 | 82.31 | 50.12 | 51.89 |\\n| LCL (w/ sim) | 42.76 | 18.92 | 18.33 | 19.45 |\\n| Combined (Ours) | 99.21 | 85.85 | 66.39 | 66.70 |\\n\\nThe results clearly demonstrate the superiority of our combined loss approach. Models trained with individual loss functions show significantly lower performance \\\\- the cross-entropy loss (LCE) model achieves limited performance when evaluated with certainty scores, while the contrastive loss (LCL) model struggles when assessed using similarity scores. In contrast, our combined approach consistently outperforms both individual methods across all datasets.\", \"this_empirical_evidence_aligns_with_our_theoretical_motivation\": \"the combined loss structure was intentionally designed to capture complementary aspects of alignment \\\\- the sequence-level patterns inherent in autoformalization and the representation-level relationships revealed through contrastive learning (as noted in lines 214-215). Rather than creating conflicting objectives, these components work together to develop a more comprehensive understanding of the alignment task.\\n\\nOur previous ablation studies further support this design choice. The ablation in **Section 5.2** examines the necessity of the combined loss structure itself, while **Section 5.3** specifically investigates potential metric bias through a systematic evaluation using individual components of our alignment-selection metric. The results in Table 6 demonstrate that models trained with the combined loss maintain strong performance even when evaluated solely on certainty scores, indicating no degradation of language generation capabilities. Additionally, the superior performance of the combined metric across all datasets suggests it successfully integrates both language-based and representation-level information, rather than reflecting any inherent bias in the evaluation approach.\\n\\n\\n\\nWe have incorporated this content into the revised version, specifically in the Appendix I section, where it is highlighted in red.\"}", "{\"title\": \"Response to Reviewer Y3AS (3/3):\", \"comment\": \"**R Q1: Splitting the Alignment-Checking Baseline for GPT-4 into Two Steps**\", \"thank_you_for_your_suggestion_to_split_the_alignment_checking_process_for_gpt_4_into_two_steps\": \"back-translation followed by a natural language alignment check. We tested this approach, referred to as \\\"Two-Phase Alignment Checking **(Two-Phase)**,\\\". The \\\"**GPT-4 (Score)**\\\" baseline refers to the original method using the prompt listed in Appendix C.2. The results are shown in the table below:\\n\\n| Models | FormL4-Basic | | | FormL4-Random | | | MiniF2F-Valid | | | MiniF2F-Test | | |\\n| ----------------- | ------------ | --------- | ----- | ------------- | --------- | --------- | ------------- | --------- | ----- | ------------ | --------- | ----- |\\n| | AS | Prec. | Rec. | AS | Prec. | Rec. | AS | Prec. | Rec. | AS | Prec. | Rec. |\\n| GPT-4 (Score) | 88.91 | 26.33 | 88.69 | 90.52 | 28.56 | 90.02 | 64.34 | 44.58 | 90.98 | 68.31 | 51.11 | 94.65 |\\n| GPT-4 (Two-Phase) | 89.35 | 38.21 | 87.95 | 91.20 | 41.10 | 89.55 | 65.75 | 53.30 | 89.10 | 69.40 | 57.80 | 92.10 |\\n| FormalAlign | **99.21** | **93.65** | 86.43 | 85.85 | **86.90** | **89.20** | **66.39** | **68.58** | 60.66 | 64.61 | **66.70** | 63.37 |\\n\\nThe \\\"Two-Phase Alignment Checking\\\" method performs slightly better than GPT-4's baseline approach. While this method does encourage GPT-4 to capture more details during back-translation, we identified several weaknesses that might outweigh its potential benefits:\\n\\n1. Splitting the task into back-translation and alignment-checking phases requires consistency across three representations: the original formal statement, the back-translated natural language, and the final alignment judgment. This decoupling introduces opportunities for error propagation, as inaccuracies in back-translation may compromise the subsequent alignment stage. \\n2. Moreover, the two-step approach adds overhead, both in computational time and model interaction complexity.\\n\\n\\n\\nWe have incorporated this content in the Appendix J section, where it is highlighted in red.\"}", "{\"comment\": \"Dear Reviewer **ok2Q**,\\n\\nThank you for your thoughtful feedback and for acknowledging our detailed response and experimental efforts. We appreciate your careful examination of our work and have taken several steps to address your concerns.\\n\\n1. Our analysis on real-world data demonstrates FormalAlign's significant advantage over GPT-4-based baselines, particularly in precision (80.2% vs 38.5%), while maintaining competitive recall rates. As suggested, we have updated our main paper to incorporate the GPT-4 (CoT) results.\\n\\n2. Regarding autoformalization performance, we want to emphasize that our paper's primary focus is on enhancing **autoformalization alignment evaluation**. Our experiments on autoformalization performance demonstrate that our proposed training framework (integrating both contrastive and cross-entropy losses) maintains and slightly improves upon baseline performance, with consistent gains across datasets. Looking forward, we believe substantial improvements in autoformalization could be achieved by integrating our framework into a reinforcement learning loop, where our alignment evaluation would serve as a reward model.\\n\\nWe would appreciate your thoughts on whether these clarifications address your concerns, or if any points need further elaboration. Thank you for your time and consideration.\\n\\nBest regards,\\n\\nAuthors of Paper Submission 1997\"}", "{\"summary\": \"This paper introduces a novel framework, FormalAlign, which trains a Large Language Model (LLM) to evaluate autoformalization alignment. Beyond the standard cross-entropy loss derived from viewing autoformalization as a sequence generation task, FormalAlign incoperates a new loss term as a dual objective. This new loss measures alignment between the informal and formal versions by leveraging the cosine similarity between their representations. Experiments on the FormL4 and MiniF2F datasets demonstrate the effectiveness and robustness of FormalAlign.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": [\"The paper addresses an important and relatively unexplored problem\\u2014improving the autoformalization of mathematical statements.\", \"The proposed framework is intuitive and easy to understand, achieving promising empirical performance on challenging evaluation datasets.\"], \"weaknesses\": [\"The discussion on the broader significance of autoformalization is lacking, which limits the accessibility of this problem to researchers in other fields (e.g., software verification). Furthermore, providing a clear formal definition of autoformalization alignment would enhance the reader\\u2019s understanding of the paper\\u2019s objectives.\", \"The running example is somewhat unconvincing to me. First, in this example, the alignment issue can be detected easily by verifying whether the statement is provable. Using the lean-set tool, a counterexample [b := 7/8, a := 1/2, c := 63/160] can be found. Second, the low certainty score suggests that LLMs may only generate this example with low probability (correct me if not).\", \"The construction of datasets is a bit unnatural. The authors generate negative examples by mutating correct input-output pairs, but these synthetic examples may not realistically reflect the types of errors typically produced by LLMs during autoformalization.\", \"The reference section contains a duplicated citation for Deepseekmath, and some related works are missing, such as [1, 2, 3].\", \"[1] Murphy, Logan, et al. \\\"Autoformalizing Euclidean Geometry.\\\" ICML, 2024\", \"[2] Li, Zenan, et al. \\\"Autoformalize Mathematical Statements by Symbolic Equivalence and Semantic Consistency.\\\" NeurIPS, 2024.\", \"[3] Zhang, Lan, et al. \\\"Consistent Autoformalization for Constructing Mathematical Libraries.\\\" EMNLP, 2024.\"], \"questions\": [\"FormalAlign appears to focus exclusively on statement autoformalization. Could this framework be extended to proof autoformalization as well?\", \"The certainty score and similarity score operate on different scales ([0,1] and [-1,1] respectively). Why are these combined using a simple average?\", \"The authors claim that FormalAlign can significantly reduce the need for manual verification, but only limited evidence is presented in the experiments. Could the authors provide more quantitative results on this claim? For instance, by deploying FormalAlign in a complete autoformalization pipeline, how many formalizations can be automatically verified?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to Reviewer ok2Q (2/3):\", \"comment\": \"**R W3: Rationale for AF Task Exclusion**\\n\\nWhile our paper primarily focuses on alignment evaluation, we did conduct experiments on the autoformalization task itself, as noted in lines 433-434, \\\"We then explore the effect of incorporating contrastive learning loss on the performance of autoformalization of natural language statements to formal language statements (Appendix F)\\\" \\n\\nThe results are placed in **Appendix F** due to space constraints. We attach it here for your convince.\\n\\nAs shown in the table below, our combined loss approach demonstrates meaningful improvements in autoformalization performance:\\n\\n| Model | FormL4 Basic (%) | FormL4 Random (%) |\\n| ------------------- | ---------------- | ----------------- |\\n| Baseline (\\ud835\\udcdb\\\\_CL) | 40.92 | 35.88 |\\n| Ours (\\ud835\\udcdb\\\\_CL \\\\+ \\ud835\\udcdb\\\\_CE) | 43.14 | 36.02 |\\n\\nThe results show that incorporating contrastive learning alongside cross-entropy loss yields consistent improvements across both FormL4 Basic (+2.22%) and FormL4 Random (+0.14%) datasets. This improvement stems from the model learning more discriminative representations through contrastive learning while maintaining strong generation capabilities via cross-entropy loss.\\n\\nWe chose to focus the main paper on alignment evaluation since it represents our primary technical contribution, but we agree that the autoformalization improvements deserve mention. We will revise our abstract and introduction to more clearly distinguish between these two aspects of our work and better highlight the dual benefits of our approach.\\n\\n**R Typos and Edits**\\n\\nWe express our sincere gratitude for your insightful comments. Regarding the presentation issues raised, we carefully incorporate all suggested improvements in our uploaded revision.\"}", "{\"title\": \"Response to Reviewer 8v22 (2/3):\", \"comment\": \"**R W3: Dataset Construction Validity**\\n\\nTo further validate our approach in real-world cases, we few-shot prompted Gemini as our autoformalization baseline on a sample of 100 theorems from our test sets. Three expert Lean users annotated these formalizations, providing corrections for misaligned ones, yielding 78 aligned-misaligned formalization pairs. The performance comparison between synthetic and real-world validation sets is shown below:\\n\\n| Evaluation Set | Accuracy (%) | Precision (%) | Recall (%) |\\n| --------------------- | ------------ | ------------- | ---------- |\\n| Synthetic Test | 85.8 | 86.9 | 89.2 |\\n| Real-world Validation | 83.5 | 80.2 | 79.8 |\\n\\nWhile performance on real-world errors is slightly lower than on synthetic cases, the strong results demonstrate that our model can effectively detect subtle misalignments in practice.\\n\\nOur dataset construction method is designed with the following considerations combined, balancing between practicability, coverage, and analysis:\\n\\n1. Automated ground truth instead of expensive and time-consuming manual labeling by human experts \\n2. Provides systematic coverage across different error types \\n3. Enables precise and thorough analysis of model capabilities for various misalignment patterns\", \"we_view_our_synthetic_dataset_as_complementary_to_real_world_examples\": \"* Synthetic data provides comprehensive coverage of potential error types \\n* Real LLM errors help validate practical effectiveness \\n* A combined approach ensures both breadth and real-world applicability\\n\\nWe have incorporated this content in the Appendix M section, where it is highlighted in red.\\n\\nWe fully agree that with more human expert resources, incorporating more real-world LLM errors into our training data could further improve our model's practical utility. In future work, we plan to create a hybrid dataset that combines synthetic mutations with collected LLM autoformalization errors (annotated with human experts) to provide both comprehensive coverage and realistic error patterns.\\n\\n**R W4: Reference Section Improvements**\\n\\nThank you for catching the duplicate citation and suggesting these relevant works. Indeed, these papers represent important contributions to autoformalization that deserve discussion.\\n\\n\\\\[1\\\\] tackles autoformalization in the specific domain of Euclidean geometry, where they develop a neuro-symbolic framework that addresses the unique challenge of diagram-dependent proofs. Their approach differs from FormalAlign by focusing on domain-specific autoformalization rather than alignment evaluation, using theorem provers to fill in diagrammatic information that makes formalization easier for language models. While both papers aim to improve autoformalization, Murphy's work complements FormalAlign by providing domain-specific solutions that could potentially benefit from FormalAlign's alignment evaluation capabilities.\\n\\n\\\\[2\\\\] addresses the gap between pass@1 and pass@k accuracies in LLM-generated formalizations by introducing a framework that scores and selects the best result from multiple candidates using symbolic equivalence and semantic consistency. While both Li's work and FormalAlign aim to improve autoformalization quality, they approach it from different angles \\\\- Li focuses on improving candidate selection through verification methods, while FormalAlign develops an automated evaluation framework for alignment. Their methods could be complementary, with Li's verification approach potentially enhancing FormalAlign's alignment evaluation capabilities.\\n\\n\\\\[3\\\\] focuses on improving consistency in large-scale autoformalization for mathematical libraries through three mechanisms: most-similar retrieval augmented generation, denoising steps, and auto-correction with syntax error feedback. Their work differs from FormalAlign by emphasizing the practical challenges of building consistent mathematical libraries, while FormalAlign focuses on the fundamental problem of evaluating alignment between informal and formal statements. Zhang's approach to consistency could potentially be integrated with FormalAlign's alignment evaluation framework to create more robust autoformalization systems for large-scale applications.\\n\\nWe revised our reference section to remove the duplicate Deepseekmath citation and discuss these three relevant papers in the Appendix B.\\n\\nThank you for helping us improve the completeness and accuracy of our literature review.\\n\\n\\\\[1\\\\] [Murphy, Logan, et al. \\\"Autoformalizing Euclidean Geometry.\\\" ICML, 2024](https://arxiv.org/abs/2405.17216)\\n\\n\\\\[2\\\\] [Li, Zenan, et al. \\\"Autoformalize Mathematical Statements by Symbolic Equivalence and Semantic Consistency.\\\" NeurIPS, 2024\\\\.](https://arxiv.org/abs/2410.20936)\\n\\n\\\\[3\\\\] [Zhang, Lan, et al. \\\"Consistent Autoformalization for Constructing Mathematical Libraries.\\\" EMNLP, 2024\\\\.](https://arxiv.org/abs/2410.04194)\"}", "{\"title\": \"Response to Reviewer ok2Q (3/3):\", \"comment\": \"**R Q1:**\\n\\nThank you for pointing out this confusion. The metrics described in Section 4.2 are indeed intended to be model-agnostic and should be presented independently of any specific implementation. We apologize for the lack of clarity caused by framing them in terms of $V_{\\\\text{align}}$, which may have inadvertently suggested that these metrics are tied to the FormalAlign model.\\n\\nTo clarify, the metrics\\u2014Alignment Score (AS), precision, and recall\\u2014are general evaluation measures that can be applied to any system performing formalization alignment tasks. For example:\\n\\n- When evaluating GPT-4 using a binary classification prompt, the Alignment Score is simply 0 or 1 based on whether the pair is deemed aligned.\\n- For our model, we compute $V_{\\\\text{cer}}$ (as defined in Equation 3), which measures the model\\u2019s sequence-level confidence in the formalization, and $V_{\\\\text{sim}}$ (as defined in Equation 4), which assesses the similarity between informal and formal representations. A threshold is then applied to determine alignment.\\n\\nThese metrics are designed to provide a consistent basis for comparison across different systems. \\n\\nTo address your concern, the details of how $V_{\\\\text{align}}$ is computed and its role in our model's inference is relocated to Appendix C.3. We hope these changes can eliminate any ambiguity and provide a clearer distinction between evaluation metrics and model-specific inference processes.\\n\\n\\n**R Q2:**\\n\\nFor evaluating GPT and other baseline models, we employ a standardized approach that is independent of our model's internal mechanisms. Specifically, we prompt these models to assign an alignment score on a 1\\u20135 scale for given informal-formal pairs, as detailed in Appendix C2. To facilitate consistent evaluation, the final Alignment Score is normalized to the [0,1] range by dividing by 5. Precision and recall metrics are then calculated based on these normalized scores using a defined threshold. In scenarios where multiple formalizations receive the same top score, one is randomly selected for evaluation to ensure fairness and reproducibility. \\n\\nAdditionally, we conducted experiments using binary classification prompts and chain-of-thought (CoT) prompting strategies to construct diverse baselines. These approaches allow us to perform a thorough comparison across different evaluation methodologies. Details regarding these experiments and their results can be found in our response to **R W2**.\\n\\n\\n\\n**R Q3:**\\n\\nThe second prompt in Appendix C.2 is indeed our evaluation prompt for the FormalAlign model, though its role may not be immediately obvious from its translation-like format. Unlike other approaches that rely on explicit alignment score prompting, our method leverages this prompt to compute alignment scores through our model's combined loss architecture.\\n\\nAs described in lines 214-215 of our paper, our model is trained with a combined loss that captures both sequence-level and representation-level alignment. During the evaluation, this prompt allows us to compute both the certainty score $V_{\\\\text{cer}}$ through equation (3), which measures the model's sequence-level confidence in the formalization, and the similarity score $V_{\\\\text{sim}}$ through equation (4), which captures representation-level alignment. The final alignment score is derived from these computational measures rather than from direct prompting.\\n\\nThis evaluation approach fundamentally differs from prompting-based methods used for baseline models. Instead of asking the model to output an explicit alignment score, we utilize our model's learned representations and certainty measures, enabled by our combined loss training, to compute alignment scores analytically. This allows us to leverage both the autoformalization patterns and representation relationships that our model learned during training. Moreover, we conducted experiments on the autoformalization task, please check R W3 for more details.\\n\\nWe revised Appendix C.2 to better explain how this prompt facilitates our computational evaluation approach and distinguish it from the prompting-based evaluation used for baseline models.\\n\\n\\n\\n**R Q4:** To confirm, we do not use synthetically generated misalignment data for training - these examples serve purely as evaluation cases to test model robustness. \\n\\n\\n\\n**R Q5:** Regarding the SIM() function in Figure 2, it is indeed equivalent to Cos() - we apologize for this inconsistency in notation and updated the figure to use consistent terminology throughout.\"}" ] }

[ "Xiaoyu Xiong", "Changyu Hu", "Chunru Lin", "Pingchuan Ma", "Chuang Gan", "Tao Du" ]

[ "Gaussian Splatting", "Differential Simulation", "Topology Optimization", "Neural Implicit Surface" ]

{ "note_id": [ "xRqcyZbdPF", "wUaCRvvxBP", "vylYj8BzVP", "uFm67UEEf2", "soHMBMSXmv", "nZncYJH0ZL", "jgZNA2E7JR", "XFQal6a32E", "UieyaKtglI", "SsPhiPDfkL", "Q6LYGVKULK", "KMu68DqYBQ", "Jqlzegziwm", "JFo0E20Ox8", "IacNqto16r", "IQgmkuwvJ8", "DXCuxXBwaF", "DAHXEXPlv9", "8BDwaRvm1M", "6j9fQuX1mW", "1MnGTsl0RX", "1CJUPCNZoq" ], "note_type": [ "decision", "official_comment", "official_comment", "meta_review", "official_comment", "official_review", "official_review", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment" ], "note_created": [ 1737523736644, 1732418958721, 1732599190035, 1734485955861, 1733297935714, 1730671590293, 1730675448885, 1733208025697, 1732895044131, 1732239741897, 1730377624335, 1732895111412, 1732239769796, 1733032016190, 1732564537018, 1732239754120, 1732980754794, 1730645911802, 1732321099600, 1732239785559, 1733023099891, 1732239419262 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission5970/Authors" ], [ "ICLR.cc/2025/Conference/Submission5970/Authors" ], [ "ICLR.cc/2025/Conference/Submission5970/Area_Chair_4UPX" ], [ "ICLR.cc/2025/Conference/Submission5970/Authors" ], [ "ICLR.cc/2025/Conference/Submission5970/Reviewer_q35E" ], [ "ICLR.cc/2025/Conference/Submission5970/Reviewer_vXS1" ], [ "ICLR.cc/2025/Conference/Submission5970/Authors" ], [ "ICLR.cc/2025/Conference/Submission5970/Authors" ], [ "ICLR.cc/2025/Conference/Submission5970/Authors" ], [ "ICLR.cc/2025/Conference/Submission5970/Reviewer_HpwZ" ], [ "ICLR.cc/2025/Conference/Submission5970/Authors" ], [ "ICLR.cc/2025/Conference/Submission5970/Authors" ], [ "ICLR.cc/2025/Conference/Submission5970/Authors" ], [ "ICLR.cc/2025/Conference/Submission5970/Reviewer_2FAv" ], [ "ICLR.cc/2025/Conference/Submission5970/Authors" ], [ "ICLR.cc/2025/Conference/Submission5970/Authors" ], [ "ICLR.cc/2025/Conference/Submission5970/Reviewer_2FAv" ], [ "ICLR.cc/2025/Conference/Submission5970/Reviewer_2FAv" ], [ "ICLR.cc/2025/Conference/Submission5970/Authors" ], [ "ICLR.cc/2025/Conference/Submission5970/Reviewer_HpwZ" ], [ "ICLR.cc/2025/Conference/Submission5970/Authors" ] ], "structured_content_str": [ "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}", "{\"title\": \"Thank you for your reply!\", \"comment\": \"Thank you for proposing the advice to strengthen our work. We have revised the manuscript based on the advice and will briefly explain the revision.\\n\\n**1. Comparison Against GT**\\n\\nThank you for suggesting the qualitative comparison with the ground truth. We have added an extra column in Figs. 14 & 15 to demonstrate the ground truth structure. We observe that our method outputs the closest result, which is consistent with the quantitative results reflected in Fig.11.\\n\\n**2. Unseen video evaluation**\\n\\nThank you for suggesting the additional statistics and qualitative comparison. We have added Fig.26 to visualize one example in our unseen test set, where we observe that the motion of our inferred structure is almost identical to that of the ground truth. We have calculated the PSNR value of our method and the baselines with respect to the ground truth and reported the statistics in the following table:\\n| Method | Ours | PGSR (0.05 voxel size) | PGSR (0.2 voxel size) | Gaussian Surfels\\n| -------- | -------- | -------- | -------- | -------- |\\n| PSNR | **36.6** | 30.4 | 30.8| 29.5|\\n\\nwhere we can observe that our method achieves the highest PNSR, leading to the most accurate results. The full statistics are exhibited in Fig.25.\\n\\n**3. Multi-Video Training**\\n\\nThank you for your advice on the further comparisons. We have updated Fig.24 to add comparison of the multi-video result against the single-video result and the ground truth, where we observe that the multi-video and single video results are generally similar. To analyze the new results quantitatively, we have also calculated their density loss compared with the ground truth and reported the statistics in the following table:\\n|Example|Single-video|Multi-video|\\n|------|------|------|\\n|Wobble doll|0.1015|0.0968|\\n|Horse|0.1001|0.0987|\\n\\nHere we can observe that the multi-video loss is slightly smaller than the single-video loss.\"}", "{\"comment\": \"Thank you for your constructive review on the additional experiments! Your feedback has helped us a lot to improve our manuscript. We are also encouraged by your comment on the interior structure reconstruction problem we study, and we hope our work can attract more researchers to investigate this less explored problem and inspire interesting follow-up applications.\\n\\nIf you have any other comments or suggestions on improving our work, please feel free to let us know. Thank you again for your time and review!\", \"title\": \"Thank you for your review!\"}", "{\"metareview\": \"This paper introduces a novel method for reconstructing the internal topology structure of an opaque 3D object from a video and/or multi-view images. The pipeline first reconstructs the 3D object as a set of Gaussian splats. The internal topology structure is then represented using a solid part indicator for each point and an SDF for fine details. A differentiable particle-based simulator for dynamic motion, incorporating constitutive models, an actuation model, and a collision model, is employed to measure the L2 difference with the reference motion and perform gradient descent on the topology representation. The experiments demonstrate results with several real-world examples.\", \"all_reviewers_gave_positive_scores_and_found_that\": \"the paper solves a novel and interesting problem; the method proposes a novel combination of Gaussian splatting and physics-based optimization; and the presentation is clear. While there were some concerns about the evaluation and validation of the design choices in each part of the method, most of them were addressed in the rebuttal. The discussion thus quickly converged to accept the submission.\", \"additional_comments_on_reviewer_discussion\": \"There was no rebuttal. All reviewers were positive to accept this submission.\"}", "{\"title\": \"Rebuttal Summary\", \"comment\": \"We thank all reviewers and the AC for their time and effort in reviewing our work and providing insightful post-rebuttal feedback. Here we would like to summarize the highlights of our work and the discussion during the rebuttal period.\\n\\n**1. Highlights:**\\n\\n1. **[Motivation]** Our manuscript studies an **important** problem (3D internal topology reconstruction from 2D visual inputs) [vXS1, HpwZ].\\n2. **[Method]** Our method is **novel** [vXS1, q35E, 2FAv] and provides a **flexible** pipeline for **various** applications [vXS1, q35E, 2FAv].\\n3. **[Experiments]** Our experiments cover **multiple** metrics with **real world** validation [vXS1, 2FAv, HpwZ], and our method **outperforms** the baselines [2FAv].\\n\\n\\n**2. New Results:**\\n1. **[Experiments: Reconstruction error with respect to ground truth structure]** We introduced **four more metrics** to demonstrate the accuracy of our methods, and achieve **better** performance over the original baselines on all these metrics (Fig. 11, 12, 13, and 15), with visualization in Fig. 14, 15, and 26. This addresses the general suggestions from all reviewers for more metrics.\\n2. **[Extensions: Optimization on multiple videos]** We extend our pipeline to accept **multiple videos** as optimization input, and successfully output a result to satisfy the multiple physical characteristics with slightly better performance than single-video optimization (Fig. 24). This addresses the suggestion from reviewer 2FAv for the extension to multi-video training.\\n3. **[Ablation Studies: Exhibition of continuous material]** We add experiments to exhibit the ability of our pipeline to output a **continuously** varying topology structure which can explain the motion in the input (Fig. 23). This addresses the inquiry from reviewer 2FAv on the ability to process continuous material.\\n\\nWe hope our new experiments have addressed the questions raised in the reviews. We are encouraged by the unanimously positive scores from all reviewers after rebuttal. We thank all reviewers for the constructive discussions during the rebuttal phase and the time and efforts of AC to review our manuscript.\"}", "{\"summary\": \"The paper presents TopoGaussian, a pipeline for inferring the internal structure of opaque objects using only photos and videos as input. The pipeline works by first using Gaussian splatting on multi-view images to obtain a point cloud, then optimizing the internal topology structure through a differentiable physics simulator to match observed motion patterns.\\n\\nThe key contributions include a particle-based, mesh-free pipeline that combines Gaussian splatting with a differentiable physics simulator; three flexible topology representation options: particle-based, neural implicit surface, and quadratic surface; a particle-based differentiable simulator supporting both rigid and soft objects with different topology structures.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. Novel combination of Gaussian splatting with physics-based optimization for internal structure inference.\\n2. Uses a mesh-free approach that avoids common issues with mesh processing; presents particle-based differentiable simulations that are compatible with three flexible topology representations, including particle, neural implicit surface, and quadratic surface.\\n3. Well-structured presentation with a clear pipeline overview.\", \"weaknesses\": \"The reviewer appreciates the authors' effort in building a particle-based pipeline to find a physically plausible internal topology structure. As the authors have also mentioned in the paper, this task is relatively new, and there are fewer baselines to compare with (at least other baselines do not use point-based representation). I have several concerns about the measuring metrics and their validation to support the claims from the authors:\\n\\n1. Optimization Loss: This measures the difference between simulated motion and reference motion, and it directly indicates whether the internal topology structure is physically plausible. However, in Figure 3, the current method does not achieve the lowest loss among baselines in multiple test samples.\\n2. Comparison Implementation: When exporting the mesh from other baselines and chaining it into the rest of the pipeline in this paper, how can the mesh-based representation be made compatible with the rest of the system that is particle-based?\\n3. Time / Smoothness: It is unclear whether this improvement comes from the GS representation itself or from the authors' method. The reviewer encourages the authors to elaborate more on this or provide ablation studies to explain that the improvement comes from the proposed method itself.\\n4. Inner Structure: Can the authors provide the reference ground truth for the inner structures when making comparisons with other baselines? The reviewer understands that, in practice, the inner structures are hard to acquire, but in synthetic data, it is practical to obtain the ground truth of inner structures.\", \"other_question\": \"How is the decision variable applied to the point cloud representation to obtain a continuous indicator function from the point cloud? (Line. 193)\", \"questions\": \"Please see my questions in the previous weakness section.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"details_of_ethics_concerns\": \"N/A\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The authors present a holistic approach to estimating the internal topology of objects from images. The work relies on the use of a particle-based differentiable simulator to estimate probable internal topology directly from the seen motion of an object. By this approach, the author is able to generate even real-world 3D printed versions with reasonable structure directly from a small video sequence.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"3\", \"strengths\": \"The author works on a novel approach to solving a relatively novel problem for modern 3D Computer Vision. While the idea of reconstruction of internal topologies is not novel to the best of my knowledge, I have not seen much work trying to solve this for 3D Gaussian Splatting or NeRF-based approaches.\\nThe work itself is decently written and shows a great evaluation to validate the quality of their methodology.\\nThe authors can propose a holistic pipeline that should make their work easily usable for users, with the authors giving a significant amount of design possibilities.\\nIn general, I am more than in favor of this work's results and core concept being interesting.\", \"weaknesses\": \"A large issue in this work is the convoluted writing. While still quite understandable, this work packs a significant amount of results, ideas, and concepts from many different fields.\\nAs such, while many parts of the simulation (core contribution) have been well explained, much information regarding the volumetric representation is missing. As such, to improve the work (and make it complete), I would suggest the author add more information in the appendix.\\nAnother larger issue is the motivation. While in Computer Graphics/Computer Vision, the challenge of estimating internal topologies is quite interesting, the estimation part might be a larger issue for practical, real-world use cases. In many real-world applications that rely on internal topologies, it is quite important that exact information is given, as this cannot assured by your model. I am still quite unsure when this work will be usable in real-world applications.\", \"questions\": [\"The authors claim one of the applications is in 3D printing; since I lack any knowledge of 3D printing, I am unsure about its weaknesses. But to enhance my understanding, why do we require the internal topology to be known for this? Shouldn't having the surface not be enough?\", \"Please fix the typo in line 175 \\\"point clout\\\"\", \"Please keep writing style consistent for example, line 234, \\\"point-cloud\\\"\", \"Would it be possible again to summarize for me what actually is the main goal and main application of this work?\", \"Not so much a question but rather a comment regarding Abstract Style (improvement/suggestion): Having this kind of structure usually improves understandability and readability: 1. What is the problem and why is it important? 2. What are the limitations of existing solutions? 3. What are the advantages of the proposed approach? 4. How does it work? 5. Summary of results\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Dear Reviewer,\\n\\nThank you for your constructive review on our manuscript. As we approach the end of the rebuttal phase, would it be possible for us to ask for your post-rebuttal feedback to help us enhance our work? Thank you.\\n\\nBest,\\n\\nAuthors\"}", "{\"title\": \"Look forward to post-rebuttal feedback\", \"comment\": \"Dear reviewer vXS1,\\n\\nWe are truly grateful for your insightful comments and advice, which have played a significant role in enhancing the quality and clarity of our paper.\\n\\nWe hope that the additional details and experimental results we provided have effectively addressed your concerns. As the rebuttal period comes to an end, we kindly request your thoughts on our rebuttal and ask that you consider raising your score accordingly. If there are any remaining concerns, please feel free to share them with us.\\n\\nOnce again, we deeply appreciate your thoughtful review and constructive feedback.\\n\\nBest,\\n\\nAuthors\"}", "{\"title\": \"Thank you for your review!\", \"comment\": \"We thank the reviewer for the constructive questions.\\n\\n**1. (W2) Reconstruction error with respect to ground truth structure**\\n\\nThank you for pointing this out. We perform more quantitative analysis on the accuracy of reconstructed topology compared with GT with 3 metrics, and achieve **3.21x**, **7.51x**, and **3.47x** smaller error than the original baselines on average (Fig.11-13). Since this is a common question among all reviewers, we explain it with more details in the general response.\\n\\n**2. (Q4) Would it be possible again to summarize for me what actually is the main goal and main application of this work?**\\n\\nThank you for bringing this issue. Our pipeline reads the photos and videos of an object as input, and outputs an interior structure which can explain the motion in the input. In many cases, simply guessing a fully solid structure from surface information will restrict the possibility of physical parameters. For example, if a wobble doll is fully solid, its center of mass will be restricted at a high position and it will be impossible to remain stable. Therefore, many practical applications require us to depict the internal topology in order to satisfy the physical characteristics.\\nThe application of our work covers many areas including computer vision, robotics and manufacturing. One example is to build a physical artifact from an online video of an unknown object, and use 3D printing to reconstruct it in the real world, even if the video is synthetic (AI generated). Another potential application is to validate the authenticity given a period of video by analyzing the interior structure of the objects in the video using our pipeline\\n\\n**3. (Q1) Why do we require the internal topology to be known for 3D printing?**\\n\\nThank you for asking this question. In 3D printing, the printer prints the objects layer by layer, and needs to know the detailed structure of each layer. More concretely, we must tell the printer which part of the layer should be printed (solid part) and which should not (hollow part). This requires us to detailedly describe the interior topology of the object in our printing file.\\n\\n**4. (W1 \\\\& Q2 \\\\& Q3 \\\\& Q5) Writing Problems including typos, writing styles and lack of information**\\n\\nThank you for providing corrections and suggestions on our writing, and we have fixed those misses in our revised version. Besides, we have added more detailed information on our volumetric representation in appendix A.3, and revised the abstract based on the suggestions in Q5.\"}", "{\"summary\": \"This paper proposed a particle-based pipeline for motion reconstruction with physically corrected internal topology based on visual inputs. Specifically, they used a vision-based reconstruction method to construct the point cloud and optimize the physical properties of each particle through a differentiable simulation. Through several rigid-body and soft-body motion tasks, they verified the validity of the approach.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"\\u25cf Extracting an object's internal topology through a visual solution sounds interesting and is a novel task. The feasibility of the solution is also demonstrated.\\n\\n\\u25cf The article constructs multiple synthetic data to validate the effectiveness of its scheme, and more importantly, it accomplishes multiple sets of validation experiments in real-world scenarios.\", \"weaknesses\": \"\\u25cf The main goal of the article is to build the internal structure of the object by visual motion, but there is no proper metric or GT comparison to measure the correctness of the internal structure recovery, it only compares the rendering results\\n\\n\\u25cf The comparative experiments are insufficient; the paper only compares optimization loss, reconstruction quality, and time with two mesh-based approaches. This fails to highlight the main contribution of the paper, which is the recovery of internal structures. Additionally, the visual comparison of internal structures does not fully reflect the accuracy of the internal structure recovery.\\n\\n\\u25cf The proposed method involves multiple steps, such as Gaussian Splatting, Volumetric Shape Generation, and Topology Optimization, etc., but lacks thorough validation for these steps. For example, there is no detailed analysis of the impact of filling quality on topology optimization or the robustness of the entire system to the visual input quality and movement amplitude.\", \"questions\": \"\\u25cf Is there a unique solution for recovering the internal structure of an object based solely on visual observation? How can we determine if the recovered internal structure is reasonable?\\n\\n\\u25cf The article mentions that the volumetric shape generation stage is faster than mesh-based methods. Is this stage entirely consistent with the PhyGaussian?\\n\\n\\u25cf What is the relationship between the metrics used in the article\\u2014reconstruction quality and optimization loss? How is reconstruction quality defined? Does the \\\"Smoothness\\\" mentioned in Section 8.1 Metrics represent reconstruction quality? Why does the proposed method achieve similar optimization loss to the baseline, yet show significantly better reconstruction quality than the baseline?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Dear reviewer HpwZ,\\n\\nWe are truly grateful for your insightful comments and advice, which have played a significant role in enhancing the quality and clarity of our paper.\\n\\nWe hope that the additional details and experimental results we provided have effectively addressed your concerns. As the rebuttal period comes to an end, we kindly request your thoughts on our rebuttal and ask that you consider raising your score accordingly. If there are any remaining concerns, please feel free to share them with us.\\n\\nOnce again, we deeply appreciate your thoughtful review and constructive feedback.\\n\\nBest,\\n\\nAuthors\"}", "{\"title\": \"Thank you for your review!\", \"comment\": \"We are grateful for the reviewer\\u2019s insightful comments.\\n\\n**1. (W1) Reconstruction error with respect to ground truth structure**\\n\\nThank you for pointing this out. We perform more quantitative analysis on the accuracy of reconstructed topology compared with GT with 3 metrics, and achieve **3.21x**, **7.51x**, and **3.47x** smaller error than the original baselines on average (Fig.11-13). Since this is a common question among all reviewers, we explain it with more details in the general response.\\n\\n**2. (W2) Multi-video training**\\n\\nThis is a brilliant experiment proposal. We have performed new experiments on the synthetic wobble doll and horse examples to leverage this idea. The experiments share similar settings with the original ones, while we generate several different videos (with different view point, light conditions, vibration amplitude, etc.) to perform optimization, with optimization loss based on all the input motions. We can observe that our model successfully outputs a smooth, accurate result with multiple motions as input, which matches the physical characteristics properly, with almost identical final optimization loss. For visual demonstration, please refer to Fig.24 in appendix.\\n\\n**3. (W3) Limitation to simple material compositions**\\n\\nThank you for bringing up this issue. We would like to clarify the ability of our pipeline in the two mentioned parts:\\n- **Continuously varying material**: Our method choose to focus on a dual-material setting due to the application concerns. In practice, it is difficult to manufacture a continuously varying topology structure by traditional methods or 3D printing, which means that a continuous result will be unrealistic to implement in industry. Therefore, we add a sharp sigmoid function in our pipeline to restrict our output to an industrial-friendly dual material result.\\nThat said, if we choose a smoother sigmoid function, our pipeline can also optimize a continuous topology structure. We use an experiment on the synthetic wobble doll and horse to demonstrate this, with similar settings to the original ones. The final optimization loss is **2.2e-5**, which is similar to the binary material (**5.6%** difference), while the visualization is shown in Fig.23.\\n- **Multi-object setting**: Our pipeline can handle the interaction between objects through the collision-handling system mentioned in Sec 6.3, and has been exhibited in the collision experiment in Fig.21. The limitation here is that although we allow interaction, we can only optimize the topology structure of a single object, which is an intriguing problem for future work.\"}", "{\"comment\": \"Thank you for your insightful and constructive feedback on our manuscript. We are encouraged by your positive comments on our approach, which motivate us to further explore this research area.\\n\\nWe noticed that the score (5) in your original review hasn't been updated yet. Would it be possible for us to get your new score on our submission?\\n\\nIf you have any additional suggestions or comments, please feel free to share them with us. Thank you once again for your time and thoughtful review!\"}", "{\"title\": \"Thanks for the additional results\", \"comment\": \"I would like to thank the authors for providing additional results, both quantitative and qualitative. This paper addresses an intriguing problem\\u2014internal structure reconstruction\\u2014which is not as extensively explored as external surface reconstruction. My initial concerns were primarily related to the evaluation protocols and the lack of sufficient results to demonstrate the method's effectiveness. However, the additional results provided by the authors have successfully addressed these concerns. As a result, I am pleased to raise my score.\"}", "{\"title\": \"Thank you for your review!\", \"comment\": \"We appreciate the reviewer's effort in reading and evaluating our work carefully!\\n\\n**1. (W1&W4) Optimization loss and comparison between the ground truth topology and inferred topology**\\n\\nThank you for pointing this out. We would like to clarify that the optimization loss in Fig.3 is the residual during the optimization process (the L2 loss between the optimized motion and input motion). It is mainly used to show the two failure cases, mainly demonstrating the loss in the specific optimization case with only qualitative information.\\n\\nFor more general and quantitative metrics to exhibit the effects of optimization, we provide 3 new metrics and perform experiments with our synthetic dataset, achieving **3.21x**, **7.51x**, and **3.47x** smaller error than the original baselines on average (Fig.11-13). Since this is a common question among all reviewers, we explain it with more details in the general response.\\n\\n**2. (W3) It is unclear whether the time/smoothness improvement comes from the GS representation itself or from the authors' method.**\\n\\nWe thank the insightful question from the reviewer. We would like to clarify that all methods, including all baselines and our pipeline, is based on GS. All these methods require GS to extract the surface information from the input. The main difference between our method and the baselines lies in the further processing and representation of the information extracted by GS. In more details, the baselines in our paper build the simulator and topology optimizer based on mesh. On the other hand, our method builds a particle-based pipeline by introducing a particle-based differentiable simulator and topology representation (point, neural implicit surface, and quadratic surface), providing a more flexible and various topology representation. Therefore, the time/smoothness improvement comes from this part without relation to GS itself since GS is also used in baselines.\\n\\n**3. (W2) How can the mesh-based representation be made compatible with the rest of the system that is particle-based?**\\n\\nThank you for bringing this issue. Mesh and particles are only two different geometry representations for the discretization in simulator and optimizer. We only need to perform a traditional particle resampling on the mesh to make it compatible with our pipeline. For example, in rigid cases, we perform a resampling on Eqn.1 and calculate the center of mass by $\\\\mathbf c=\\\\frac{1}{m}\\\\sum \\\\rho_iV_i\\\\mathbf x_{ci}$.\"}", "{\"title\": \"Could AC help to reach out reviewers?\", \"comment\": \"Dear AC and SAC:\\n\\nThank you for efficiently handling our draft.\\n\\nAs we approach the end of the rebuttal phase, we noticed that post-rebuttal feedback from three borderline reviewers is still pending.\\n\\nCould the AC assist in contacting these reviewers for their responses?\\n\\nWe greatly appreciate all your efforts!\\n\\nThank you,\\n\\nAuthors\"}", "{\"summary\": \"The paper introduces TopoGaussian, a particle-based approach that uses visual clues from photos and videos to infer internal topology structures of opaque objects. Key contributions include:\\n1. Mesh-Free Topology Inference: TopoGaussian combines Gaussian splatting and a particle-based differentiable simulator to infer interior topology without requiring mesh-based representations, which traditionally entail extensive fixing and filling processes.\\n2. Flexible Topology Representations: The pipeline supports different topology representations (particle, neural implicit surface, and quadratic surface), allowing for optimization and simulation within a unified framework.\\n3. Experimental Validation: TopoGaussian is evaluated on synthetic and real-world tasks, showcasing its capability to generate 3D-printable reconstructions that exhibit high fidelity and reduced processing time compared to existing mesh-based methods.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. Efficient and High-Quality Reconstruction: The particle-based approach of TopoGaussian achieves efficient reconstructions, with the authors reporting a significant speedup (5.26x faster) and superior boundary reconstruction quality (2.33x improvement) compared to mesh-based methods like PGSR and Gaussian Surfels.\\n2. Annotation-Free and Flexible: The pipeline\\u2019s independence from intrusive sensors or annotation requirements makes it practical and applicable in fields like robotics and manufacturing. Additionally, the three topology representation options offer flexibility based on application needs, from rigid to soft-body simulation.\\n3. Simplicity and Smoothness of Output: By eliminating the need for mesh processing, TopoGaussian produces smoother outputs conducive to 3D printing and manufacturing.\", \"weaknesses\": \"1. Evaluation Limitation: The method optimizes the interior structure based on a single motion, which may lead to overfitting and an inaccurate reconstruction of the true internal structure. Since ground truth data for internal topology is unavailable, it is difficult to verify if the inferred structure is correct or merely adapted to the given motion. Although the authors acknowledge this limitation and propose alternative metrics, these may not fully reflect the true structure. To strengthen the evaluation, the authors could consider obtaining ground truth through simulation (maybe with physics simulation) or testing the inferred structure on new motion videos as a test set. If the predicted structure is accurate, it should exhibit consistent behavior across these unseen motions, providing stronger validation.\\n2. Potential Overfitting to Single Motion: The current approach optimizes the internal structure based on a single motion video, which may lead to overfitting and limit the model\\u2019s generalization capability. To improve the robustness and accuracy of the inferred structure, the authors could consider optimizing based on multiple motion videos. By incorporating a variety of motions, the resulting model may better capture the true internal structure and provide more reliable and generalizable results.\\n3. Limitation to Simple Material Compositions: The current framework supports only single-object, dual-material compositions, which may limit applications involving complex, heterogeneous materials or multi-object interactions. Future work could focus on extending support to more intricate material compositions.\\n\\n---\\n\\nThe authors have provided additional results to address my concerns.\", \"questions\": \"See weakness\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"details_of_ethics_concerns\": \"/\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"follow-up questions\", \"comment\": \"## Comparison Against GT\\nThank you for providing the additional experiments and quantitative comparison (density difference/unseen video) against the ground truth. This strengthens the evidence supporting the method's effectiveness.\\nI would also suggest including a qualitative comparison since ground truth is now available. For example, in Figures 14\\u201315, adding an additional column with the ground truth would provide a clear visual reference, enhancing the interpretability of the results.\\n\\n## Unseen video evaluation\\n1) I suggest including a side-by-side comparison video for the test set. For example, a minimal demonstration could involve showing the ground truth test video on the left side and the rendered video on the right side.\\n\\n2) Additionally, I recommend including commonly used metrics for image rendering evaluation, such as PSNR, SSIM, and LPIPS. These metrics, as referenced in the 3DGS paper, provide a more direct and standardized quality assessment compared to the current metric (optimization loss). These values would help readers better understand the rendering quality of the method.\\n\\n## Multi-Video Training\\nThank you for extending the method to include multi-video training and providing the corresponding results.\\nAs the authors have incorporated this extension, is there a comparison between the results of multi-video training and single-video training, both quantitatively and qualitatively? Specifically, does multi-video supervision provide any measurable improvement in reconstruction accuracy or quality?\\nCurrently, Figure 24 shows the results for multi-video training, but there is no reference provided. It would be beneficial to include the ground truth structure and the results from single-video training as a comparison. Without these references, the multi-video training experiment lacks sufficient context to demonstrate its additional merit to the paper.\"}", "{\"title\": \"Thank you for your review!\", \"comment\": \"We appreciate the valuable comments from the reviewer.\\n\\n**1. (W1 & W2 & Q1) Reconstruction error with respect to ground truth structure**\\n\\nThank you for pointing this out. We perform more quantitative analysis on the accuracy of reconstructed topology compared with GT with 3 metrics, and achieve **3.21x**, **7.51x**, and **3.47x** smaller error than the original baselines on average (Fig.11-13). Since this is a common question among all reviewers, we explain it with more details in the general response.\\n\\n**2. (W3) The proposed method involves multiple steps, such as ..., but lacks thorough validation for these steps.**\\n\\nThank you for bringing this issue. The reviewer points out two important point to be tested in our pipeline:\\n- Filling quality\\n- Input motion quality\", \"we_have_the_following_ablation_studies_to_test_the_robustness_of_these_parts_in_our_original_paper\": \"- In Fig.19, we test the influence of different filling resolutions, where we find that the results are not affected by different resolutions and output similar results\\n- In Fig.22, we test the impact of different video quality, including light condition and view point, where the results also remain similar, proving the robustness of our pipeline.\\n\\nMoreover, we have performed another ablation study in Fig.20 to test the robustness with different material properties and object size, which also strengthen the robustness of our pipeline\\n \\n**3. (Q2) The article mentions that the volumetric shape generation stage is faster than mesh-based methods. Is this stage entirely consistent with the Phy(s)Gaussian?**\\n\\nThank you for asking this valuable question. We do not find ''PhyGaussian'' on the Internet and guess the reviewer means ''PhysGaussian''. The point cloud filling process (which transfers the surface point cloud from GS to a volume point cloud) in our pipeline is similar to PhysGaussian. On the other hand, since PhysGaussian does not contain topology optimization, the topology representation part, including the point and surface representation (neural implicit surface and quadratic surface) is not relevant to it.\\n\\n**4. (Q3) What is the relationship between the metrics used in the article\\u2014reconstruction quality and optimization loss?**\\n\\nThank you for bringing this issue. We would like to discuss these two metrics together with the three new metrics (density loss, test loss, and CoM loss) we provide in the general response.\", \"there_are_many_metrics_we_need_to_consider_to_measure_the_result\": \"1. **Motion loss of input** measures the difference between simulated motions of results and input.\\n2. **Motion loss of test set (unseen motions)** measures the difference of simulated motions between results and test set.\\n3. **Density loss compared with GT** measures the difference of output topology (density distribution).\\n4. **CoM loss compared with GT** measures the deviation of the center of mass.\\n5. **Smoothness loss of the result (3D printing concerns)** measures the average laplacian of the internal surface in output topology.\\n\\nWe use the second to fourth metrics for model evaluation by comparing them with the ground truth (GT) in the testing set. These metrics are strictly reserved for evaluation and should not be utilized during optimization. On the other hand, the differential of the fifth loss is difficult to solve, which makes it hard to be added in the gradient-based optimization process. Therefore, we choose our optimization objective to minimize the first loss function, and use the other four metrics to validate and measure the accuracy and quality of the results. It may be a promising task to add the fifth loss as a regularization term in the optimization process and enhance the quality of the results in the future.\"}", "{\"comment\": \"Dear Authors:\\nThank you for your detailed explanations and additional experiments. After reading other reviews and your feedback, I think you have answered most of my questions and addressed my concerns, especially on the structure evaluation. I still maintain that this is an interesting and novel work. I am pleased to raise my score.\"}", "{\"title\": \"Authors' rebuttal: general questions and new experiments\", \"comment\": \"We thank all reviewers and the AC for their time and effort in reviewing and for insightful comments to strengthen our work. We update a revised version of the manuscript based on the suggestions from the reviewers. Besides the responses to individual reviewers, here we would like to highlight our contributions and new quantitative/qualitative results added in the rebuttal.\\n\\n**1. Contributions:**\\n1. **[Motivation]** Our manuscript studies an important problem (3D internal topology reconstruction) [vXS1, HpwZ].\\n2. **[Method]** Our method is novel [vXS1, q35E, 2FAv] and provides a flexible pipeline for various applications [vXS1, q35E, 2FAv].\\n3. **[Experiments]** Our experiments cover multiple metrics with real world validation [vXS1, 2FAv, HpwZ], and our method outperforms the baselines [2FAv].\\n\\n\\n**2. New Results:**\\n1. **[Experiments: Reconstruction error with respect to ground truth structure]** We perform more quantitative analysis on the accuracy of reconstructed topology compared with GT with 3 metrics, and achieve **3.21x**, **7.51x**, and **3.47x** smaller error than the original baselines on average. Since this is a common question among all reviewers, we will explain it with more details in the following part and refer to Fig.11-13 for detailed statistics.\\n2. **[Extensions: Optimization on multiple videos]** Thanks for the proposal from reviewer 2FAv, we extend our pipeline to accept multiple videos as optimization input, and successfully output a result to satisfy the multiple physical characteristics with almost identical optimization loss compared to the single-motion result. (Fig.24).\\n3. **[Ablation Studies: Exhibition of continuous material]** Based on the proposal from reviewer 2FAv, we add experiments to exhibit the ability of our pipeline to output a continuously varying topology structure which can explain the motion in the input (Fig.23). The final optimization loss is **2.2e-5**, which is similar to the binary material (**5.6%** difference).\\n\\n**3. Further explanation on reconstruction error with respect to ground truth structure:**\\nWe thank all reviewers for asking this valuable question. We provide the following three metrics to measure the reconstruction error:\\n- Density loss: Based on the traditional practice, we defined the volumetric average density difference to depict the difference between optimized result and ground truth.\\n- Test loss: Based on the suggestion from reviewer 2FAv, we generate several **unseen** poses as a test set, and measures the difference between optimized result and ground truth based on this test set, which characterizes the generalization ability of our pipeline.\\n- CoM loss: Since the physical behavior of a rigid body is dominated by its center of mass (CoM) in rigid cases, we also test the difference of CoM between the result and ground truth.\\n\\nBased on these 3 metrics, we perform experiments on the objects in our synthetic datasets with results shown in Fig.11-13 respectively. The following table summarize the average loss comparisons between our method and baselines:\\n|Baselines|Density Loss|Test Loss|CoM Loss|\\n|---------|------------|---------|--------|\\n|Ours|**0.099**|**0.038**|**0.301**|\\n|PGSR (0.05 voxel size)|0.307 (3.11x)|0.297 (7.46x)|0.990 (3.29x)|\\n|PGSR (0.2 voxel size)|0.295 (2.99x)|0.259 (6.52x)|0.988 (3.28x)|\\n|Gaussian Surfels|0.348 (3.53x)|0.342 (8.57x)|1.161 (3.86x)|\\n\\nWe can observe that our method outperforms all the original baselines in all three metrics. The reason may come from the particle-based pipeline providing a more flexible and various topology representation than mesh-based pipeline, leading to more accurate results and stronger generalization ability. This gives us a good motivation for a particle-based representation.\"}" ] }

[ "Zemin Huang", "Zhengyang Geng", "Weijian Luo", "Guo-Jun Qi" ]

[ "one-step generator", "text-to-image generation", "flow matching" ]

{ "note_id": [ "s68N5umuxZ", "qOlCNSkD7G", "q9XHYnTguL", "inyZjrRf1r", "iLGLgJLuYC", "gYMH9SJgZV", "fUxSUiAiQK", "ejDSDrGybQ", "egFdr1pGiQ", "eCQ7Iv7WEM", "d5FpiHmn4X", "c6XJWYEIr8", "YYegRx9XfJ", "Sp6e0YR41b", "P63tz0QRIY", "K0rZr6TcZ6", "I5P5NtIrtg", "BtLV1Rozdx", "7OvgXmoenL" ], "note_type": [ "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "meta_review", "official_comment", "official_review", "official_comment", "official_comment", "decision", "official_comment", "official_comment" ], "note_created": [ 1733226024653, 1730648430983, 1733225599193, 1732807145103, 1732093031547, 1730261947403, 1732092974715, 1733211664803, 1732807160411, 1732807019097, 1730616309185, 1734583040864, 1732680882823, 1730709395395, 1733226061375, 1732439603640, 1737523924689, 1732093107207, 1732093784216 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission8666/Authors" ], [ "ICLR.cc/2025/Conference/Submission8666/Reviewer_sdTk" ], [ "ICLR.cc/2025/Conference/Submission8666/Authors" ], [ "ICLR.cc/2025/Conference/Submission8666/Authors" ], [ "ICLR.cc/2025/Conference/Submission8666/Authors" ], [ "ICLR.cc/2025/Conference/Submission8666/Reviewer_BVRU" ], [ "ICLR.cc/2025/Conference/Submission8666/Authors" ], [ "ICLR.cc/2025/Conference/Submission8666/Reviewer_BVRU" ], [ "ICLR.cc/2025/Conference/Submission8666/Authors" ], [ "ICLR.cc/2025/Conference/Submission8666/Authors" ], [ "ICLR.cc/2025/Conference/Submission8666/Reviewer_4qEE" ], [ "ICLR.cc/2025/Conference/Submission8666/Area_Chair_Vcov" ], [ "ICLR.cc/2025/Conference/Submission8666/Reviewer_4qEE" ], [ "ICLR.cc/2025/Conference/Submission8666/Reviewer_jkBP" ], [ "ICLR.cc/2025/Conference/Submission8666/Authors" ], [ "ICLR.cc/2025/Conference/Submission8666/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission8666/Authors" ], [ "ICLR.cc/2025/Conference/Submission8666/Authors" ] ], "structured_content_str": [ "{\"comment\": \"As we approach the conclusion of our rebuttal, we would like to kindly **summarize our key points** and **express our eagerness for your final feedback.** We have made significant revisions to enhance the manuscript in response to the reviewers' feedback. Below is a summary of the major changes implemented during the revision process:\\n1. As **Reviewer sdTK** and **Reviewer jkBP** wish, **we polish the presentation of the method**. We add a paragraph in Section 4.2 to clarify the difference between FGM and diffusion distillation by pointing out technical challenges. We also add a paragraph in Section 5.1 to compare the training efficiency of FGM and CFM, showing that FGM can significantly surpass CFM on both generation results and training efficiency.\\n2. In response to **Reviewer 4qEE** 's suggestion. In Section 4.2, **we have included a comprehensive explanation of the stop $\\\\theta$-gradient technique** utilized within our model. This addition aims to clarify its implementation details and the impact on gradient backpropagation\\n3. As suggested by **Reviewer BVRU**, **we have revised the notation in Algorithm 1** for improved clarity. Specifically, we replaced $v_{\\\\text{sg}[\\\\theta]}$ with the actual online flow model $v_\\\\psi$ that we employ. Furthermore, we have provided additional explanations concerning the online flow model $v_\\\\psi$ in Section 4.2 to enhance understanding.\\n4. Addressing the feedback from **Reviewer 4qEE** and **Reviewer BVRU**. **we have added detailed pointers in Section 5** that direct readers to the appendix for comprehensive training details, and **expanded our ablation studies**, which can be found in Appendix C. These studies focus on two primary areas: (C.1) generator initialization, and (C.2) the impact of including or excluding regression loss during training.\\n\\n**If you have any additional issues or concerns, we would be happy to address them promptly.**\"}", "{\"summary\": \"This paper proposes extending score implicit matching from diffusion to flow matching, which is called FGM. The method seems to work well on text-to-image and unconditional generation.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. The paper is well-written and easy to follow.\\n2. The paper extends the score implicit matching to a flow matching generator.\", \"weaknesses\": \"1. The novelty of the paper is limited. Since flow matching can be considered a special form of diffusion model, score-based diffusion and flow matching are equivalent. Therefore, rawly extending the implicit score matching for the flow matching model should work well.\\n2. The evaluation for the text-to-image section should include FID, Recall, ClipScore, Image Reward, PickScore, and AES score. These metrics are often used as standard metrics to evaluate generative models.\\n3. Since this method is closely related to implicit score matching, the author should include a background section for implicit score matching.\", \"questions\": \"My biggest concern is the novelty and originality of this paper's idea. There is a list of works working on score implicit models, such as [1,2,3]. Extending the framework from diffusion to flow matching is not interesting and does not introduce something new.\\n\\n[1]: Diff-Instruct: A Universal Approach for Transferring Knowledge From Pre-trained Diffusion Models\\n\\n[2]: One-Step Diffusion Distillation through Score Implicit Matching\\n\\n[3]: Score identity Distillation: Exponentially Fast Distillation of Pretrained Diffusion Models for One-Step Generation\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you for your feedback. We have made significant revisions to improve the presentation and clarity in response to reviewers' feedbacks during rebuttal. We believe the updated version might addresses your concerns effectively. Below is a summary of the major changes made during the revision process:\\n\\n1. As **Reviewer sdTK** and **Reviewer jkBP** wish, **we polish the presentation of the method**. We add a paragraph in Section 4.2 to clarify the difference between FGM and diffusion distillation by pointing out technical challenges. We also add a paragraph in Section 5.1 to compare the training efficiency of FGM and CFM, showing that FGM can significantly surpass CFM on both generation results and training efficiency.\\n2. In response to **Reviewer 4qEE** 's suggestion. In Section 4.2, **we have included a comprehensive explanation of the stop $\\\\theta$-gradient technique** utilized within our model. This addition aims to clarify its implementation details and the impact on gradient backpropagation\\n3. As suggested by **Reviewer BVRU**, **we have revised the notation in Algorithm 1** for improved clarity. Specifically, we replaced $v_{\\\\text{sg}[\\\\theta]}$ with the actual online flow model $v_\\\\psi$ that we employ. Furthermore, we have provided additional explanations concerning the online flow model $v_\\\\psi$ in Section 4.2 to enhance understanding.\\n4. Addressing the feedback from **Reviewer 4qEE** and **Reviewer BVRU**. **we have added detailed pointers in Section 5** that direct readers to the appendix for comprehensive training details, and **expanded our ablation studies**, which can be found in Appendix C. These studies focus on two primary areas: (C.1) generator initialization, and (C.2) the impact of including or excluding regression loss during training.\\n\\n**If there are any further issues or concerns, we would be happy to address them promptly.**\"}", "{\"comment\": \"We sincerely appreciate your valuable feedback. We are eager to address any remaining concerns and can provide additional classifications or experiments as needed. A prompt response would greatly assist us in making timely improvements.\"}", "{\"comment\": \"We sincerely appreciate your constructive criticism and suggestions. In response to your concerns, we would like to clarify the following points. Before that, we will first provide a summary of the key contributions of our work.\\n\\nIn this paper, we present **Flow Generator Matching (FGM)**, an innovative approach amied at accelerating the sampling of flow matching models into a **one-step generation model**. Our results on CIFAR10 unconditional achieve a remarkable **FID score of 3.08** among all flow-matching models. Futhermore, our distillation results on **SD3-medium** demonstrates outstanding performance among other few-step distillation models. In addition to experimental performance, we propose the **flow-projection identity** to bypass the intractable flow-matching objective, leading to our **practical FGM training objective with theoretial guarantees**, which lays the groundwork for potential advancements in future research. In contrast to earlier score-based implicit distillation methods, our approach is distinct in its emphasis on the flow-matching objective rather than relying on inapplicable score functions in flow matching.\\n\\n**Q1: Novelty and originality of this paper's idea**\\n\\n**A1:** Previous implicit models have largely relied on diffusion models. A key distinction between flow-matching and diffusion models is that flow matching does not inherently imply score functions, rendering the definitions of distribution divergences inapplicable. \\n\\nOur contribution resolves this issue by **focusing on the flow matching objective**, rather than on distribution divergence. To the best of our knowledge, the flow-generator-matching is one of the few attempts to introduce one-step distillation method based on the flow matching model without explicitly involving probability divergences. Interestingly, FGM bypass the need for probability divergence through directly handling the flow-matching objective. To do so, we introduce a novel flow-projection identity and deriving an equivalent loss that minimizes the intractable flow-matching objective. However, we do appreciate previous works, especially the score-identity distillation and the score-implicit matching. Both works focuses on diffusion models which have an straightforward probability interpretation rather than flow-matching models. \\n\\n**Q2: Extra Evaluation**\\n\\n**A2:** We have followed your suggestion to conduct further evaluation on the COCO 2017 validation set with 5,000 samples. This more comprehensive evaluation demonstrates that our one-step distillation model closely matches the performance of the teacher model in many aspects and even surpasses the results of multi-step distillation on several metrics.\\n\\n| Model | Steps | FID | CLIP Score | Image Reward | Aesthetic Score | Pick Score |\\n| ------------ | ----- | --------- | ---------- | ------------ | --------------- | ---------- |\\n| SD3(Teacher) | 28 | 23.15 | 32.10 | 0.92 | 5.27 | 0.223 |\\n| Hyper-SD3 | 4 | 71.64 | 31.57 | **0.83** | 5.30 | **0.225** |\\n| Flash-SD3 | 4 | 70.67 | 31.91 | 0.65 | 5.38 | 0.220 |\\n| Ours | **1** | **32.75** | **31.92** | 0.78 | **5.39** | 0.221 |\\n\\n**We hope our rebuttal have resolved all your concerns. If you still have any concerns, please let us know. We are glad to provide further clarifications as well as more additional experiments.**\"}", "{\"summary\": \"The paper introduces Flow Generator Matching (FGM), a method for to distill a pretrained flow-matching model into a single-step model. Specifically, the paper proposes a data-free distillation approach to match the implicit vector field of the student model $v_{\\\\theta, t}$ with the pretrained vector field u_t, where u_t can be regarded as the teacher. Experiments are carried out on CIFAR-10 and paired text-image dataset, which demonstrate the effectiveness.\", \"soundness\": \"2\", \"presentation\": \"1\", \"contribution\": \"3\", \"strengths\": \"1. The paper proposes a novel method to distill a flow-matching model based on the property of its vector field from a probabilistic standpoint, which represents a good contribution.\\n2. Experiments show effectiveness of the appraoch.\", \"weaknesses\": \"1. My biggest concern is the clarity of the paper.\\n - Algorithm 1 is not clear to me and hard to understand. What is the online flow model $v_\\\\psi$? It gets introduced nowhere in the entire Section 4 except the Input field of Algorithm 1. Does it represent the implicit vector field? If yes, why shall we have another notation of $v_{sg[\\\\theta], t}$ in 4.11 and 4.12? If not, then what does the online model do and how could we use $\\\\theta$ to parameterize both the generator and the implicit vector field? Also in 4.12, there're both $u_t(x_t)$ and $u_t(x_t | x_0)$, yet $x_t$ is only sampled from $x_t | x_0$ ~ $q_t(x_t | x_0)$, then why do we have different notations here? This is the core of the paper which concretizes the theory into algorithm, and it shall be presented as clear as possible. \\n - in line 203, there's an abuse of notation where using $x_0 = g_\\\\theta(z)$ instead of $x$ could make the notation more consistent with the latter integral.\\n2. The one-step results seem to have checkerboard artifact. When zoomed in Fig.1, most samples show the pattern. It also appear in the 2nd sample of Figure 3 while SD3-28 steps have much smoother texture. The quality degradation is still a concern. \\n3. There's missed reference in Table 2 for StyleGAN2 + Smart. \\n4. How do we find the t* in 5.1? The ablation study is not presented.\", \"questions\": \"Please refer to weakness.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"We appreciate the your feedback and would like to address the following points. Before that, we will first provide a summary of the key contributions of our work.\\n\\nIn this paper, we introduce **Flow Generator Matching (FGM)**, an innovative approach designed to accelerate the sampling of flow matching models into a **one-step generation model**. Our results on CIFAR10 unconditional set a new record **FID score of 3.08** among all flow-matching models. Futhermore, our distillation results on **SD3-medium** demonstrates outstanding performance among other few-step distillation models. In addition to experimental performance, we propose the **flow-projection identity** to bypass the intractable flow-matching objective, leading to our **practical FGM training objective with theoretial guarantees**, thus laying a solid groundwork for future research advancements.\\n\\n**Q1: Is this auxiliary model used after training?**\\n\\n**A1:** The auxiliary model would be **no longer needed** when the training is finished. Once our one-step generator has almost converged, the auxiliary model should be similar to the per-trained flow-matching model. In practice, we only use keep one-step generator for text-to-image generations. However, we believe that the study on how to further use knowledge remained in the auxiliary model is an interesting research topic in future. \\n\\n**Q2: Training efficiency compared with other common baselines**\\n\\n**A2:** On distillation of Stable-diffusion-3-medium, a 2B transformer model, our **full parameter** training spent approximate 25 H800 days to achieve the reported results with 23k optimization step. It is important to note that this duration includes a significant amount of exploratory testing. Once all hyperparameters are determined, the actual training time required to develop a production-ready model may be considerably shorter. \\n\\nAs for other common baselines, Flash-SD3 was trained for approximately 50 hours on 2 H100 GPUs utilizing LoRA, with **only 90.4 million trainable parameters**. The training duration for Hyper-SD3 is currently unknown. Although our model requires a significantly longer training time compared to Flash-SD3, the adoption of LoRA may substantially reduce the training costs, an avenue we plan to explore further.\\n\\n**We hope our rebuttal have resolved all your concerns. If you still have any concerns, please let us know. We are glad to provide further clarifications as well as more additional experiments.**\"}", "{\"comment\": \"Thank the reviewer for the response. Given such observable flaws in presentation/clarity from the original submission, it would be better to have the paper go through another review round for self consistency and make sure it is understandable by others. I'll remain my rating.\"}", "{\"comment\": \"We sincerely appreciate your valuable feedback. We are eager to address any remaining concerns and can provide additional classifications or experiments as needed. A prompt response would greatly assist us in making timely improvements.\"}", "{\"comment\": \"We sincerely appreciate your valuable feedback. We are eager to address any remaining concerns and can provide additional classifications or experiments as needed. A prompt response would greatly assist us in making timely revisions.\"}", "{\"summary\": \"This paper introduces Flow Generator Matching, an innovative distillation framework for pre-trained flow-matching models. The framework is designed to accelerate sampling by constructing a one-step generator. To achieve this, the authors define a noise-to-image mapping $g_\\\\theta(z)$, and optimize the distance between the vector field of the pre-trained flow model, and the vector field implicitly derived from the one-step generated images and the online flow model. By leveraging a theoretical approximation of this optimization using the flow product identity, the authors develop a fast and efficient one-step flow generator.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"3\", \"strengths\": [\"Theoretical justification: The proposed distillation formulation is both intuitive and novel in the context of flow matching. Although optimizing $L_{FM}$ directly with respect to $\\\\theta$ is non-trivial, the authors introduce an innovative alternative whose gradient remains aligned with the original objective.\", \"Specialized distillation methods for flow matching: Given the widespread use of flow matching in various state-of-the-art large-scale generative models (e.g., Movie Gen), effective sampling acceleration is critical. This paper offers a promising approach to address this important contemporary challenge.\", \"Experiments: Comprehensive empirical evaluations, including unconditional, conditional, and text-to-image experiments, support the effectiveness of the method. The approach demonstrates competitive performance with other flow distillation techniques, achieving lower NFE.\"], \"weaknesses\": [\"Writing & Notations: The connection between the derivations and practical implementations is difficult to follow. For instance, while FGM relies on the online flow model parameterized by $\\\\psi$, as outlined in Algorithm 1, $\\\\psi$ is notably absent in Section 4, despite being a crucial component. Additionally, there is no explanation or justification for the phrase \\\"stop the $\\\\theta$-gradient\\\" mentioned in Line 250, which is a critical implementation detail. From my understanding, FGM is conceptually similar to distribution matching distillation, where both the generator $\\\\theta$ and an online critic $\\\\psi$ are updated alternately. However, this understanding comes from Algorithm 1 rather than the main text. The authors should provide a clear and theoretically grounded explanation of \\\"stop gradient\\\" and revise the derivations to explicitly incorporate the online flow model $\\\\psi$. Moreover, some important information are provided in appendix without pointers, e.g. model parameterization in L952. Detailed pointers in the main paper would be appreciated.\", \"Novelty: As previously mentioned, FGM shares significant technical similarities with existing diffusion distillation methods. This overlap may diminish the novelty of the approach and, consequently, inherit some of the limitations of these existing methods, such as reliance on a potentially sub-optimal online flow model.\", \"Qualitative Results: Although FGM demonstrates better quantitative performance compared to other flow distillation methods, the generated images appear overly \\\"synthetic\\\" and color-saturated (e.g., the rightmost images in Lines 54 and 488), while i admit that this may vary based on human perception. This could be potentially attributed to the image-data-free property proposed as limitations by the authors.\"], \"questions\": [\"Overall, the paper is well-structured and presents promising results. However, the derivations and notations could be refined to enhance clarity and facilitate faster understanding. Below are some questions for consideration:\", \"Ablation Study on Generator Initialization: The generator is initialized with pre-trained flow models, which may be crucial for warm-up and accelerating convergence, especially given that FGM does not utilize real-image datasets. This initialization might also mitigate the OOD gap for pre-trained flow models trained on real images. Conducting an ablation study or providing further analysis could improve understanding and highlight the significance of this choice.\", \"Bypassing Equation 4.11: Although Equations 4.11 and 4.12 both originate from the FGM loss, the authors opt not to use Equation 4.11, as mentioned in Line 368. This decision represents a significant bypass. Additional analysis, discussion, and empirical evidence are needed to justify this choice. Could the authors elaborate on why Equation 4.11 is deemed unnecessary in practice?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"metareview\": \"This paper suggest a distillation method based on the Flow Matching loss where the idea is to optimize a \\u201cgenerator\\u201d $g_\\\\theta$ that maps noise to sample in one step by defining the student data distribution $p_{\\\\theta,0}$ via $g_\\\\theta$ and minimize the FM loss to match it to the teacher data distribution defined by the given pre-trained model.\\n\\nConcerns about this paper in its current state include its somewhat limited relative added contribution compared to recent implicit score 1-step distillation methods that take a similar approach but for score parameterization, the quality of the results presented, and lacking/unclear presentation. We still feel the paper is of value as a promising flow distillation method and encourage the authors to revise their paper and resubmit in the future.\", \"additional_comments_on_reviewer_discussion\": \"No additional comments.\"}", "{\"title\": \"Official Comment by Reviewer 4qEE\", \"comment\": \"Thank you for the clarifications. They address several of my concerns, particularly the explanation of the stop-gradient component. I remain inclined toward acceptance, as the derivation from the intractable flow-matching objective is non-trivial and could inspire future research. However, some unresolved concerns remain as follow.\\n\\n- **Differences from diffusion distillation**: As Reviewer sdTk noted, flow models subsume diffusion models as specific cases. In this context, I find the implications of Lines 313\\u2013317 unclear:\\n\\n> (...) flow matching does not imply explicit modeling of either the probability density as the diffusion models do. Therefore, the definitions of distribution divergences cannot be applied to flow models (...)\\n\\nFlow-based generative models can explicitly recover the ODE structure in a form of denoisers (Sec. 2.1, [1]) and generalize probability path definitions, as demonstrated in [2]. Since this 'Differences from diffusion distillation' section is critical to the novelty, more detailed clarification is necessary. While bypassing explicit probability divergence is theoretically meaningful, the novelty could be diminished if the resulting algorithms closely resemble existing methods.\\n\\n- **C.2 Training with regression loss**: While I appreciate the comparative results, a more detailed intuitive explanation, reasoning, or analysis would strengthen this section. Could you elaborate on why omitting Eq. (4.11) improves the results?\\n\\n---\\n**References**\\n\\n[1] Kim, Beomsu, et al. \\\"Simple ReFlow: Improved Techniques for Fast Flow Models.\\\" arXiv preprint arXiv:2410.07815 (2024).\\n\\n[2] Tong, Alexander, et al. \\\"Improving and generalizing flow-based generative models with minibatch optimal transport.\\\" arXiv preprint arXiv:2302.00482 (2023).\"}", "{\"summary\": \"The paper introduces the **Flow Generator Matching (FGM) objective** for distilling a one-step generator from pretrained flow-matching models. It provides theoretical guarantees that the proposed objective yields the same gradient, with respect to the parameters of the one-step generator, as the standard flow-matching loss. The paper demonstrates competitive and, in some cases, superior results in both unconditional generation on CIFAR-10 and large-scale text-to-image generation, achieved by distilling Stable Diffusion 3 (SD3).\", \"soundness\": \"3\", \"presentation\": \"4\", \"contribution\": \"3\", \"strengths\": \"1. The paper is well-written and easy to follow.\\n2. The proposed algorithm is elegantly designed, alternating the optimization of the parameters of the one-step generator, $\\\\theta$, and an auxiliary flow model, $\\\\psi$. The correctness and effectiveness of the algorithm are supported by Theorems 4.1 and 4.2.\\n3. The paper conducts thorough empirical evaluations of the proposed method across various tasks, including unconditional generation on CIFAR-10 and large-scale text-to-image generation, demonstrating superior results.\", \"weaknesses\": \"1. The proposed method requires training an auxiliary flow model with parameters $\\\\psi$, designed to generate the implicit flow determined by the one-step generator. Is this auxiliary model used after training?\\n2. The paper does not include a discussion of training efficiency or speed. One concern is that the proposed method might be slower compared to distillation approaches like Consistency Flow Matching (Yang, Ling, et al., \\\"Consistency Flow Matching: Defining Straight Flows with Velocity Consistency\\\"). Could you compare or discuss the training efficiency of the proposed methods with other common baselines?\", \"questions\": \"Please see weaknesses part.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"As we approach the conclusion of our rebuttal, we would like to kindly **summarize our key points** and **express our eagerness for your final feedback.** We have made significant revisions to enhance the manuscript in response to the reviewers' feedback. Below is a summary of the major changes implemented during the revision process:\\n\\n1. As **Reviewer sdTK** and **Reviewer jkBP** wish, **we polish the presentation of the method**. We add a paragraph in Section 4.2 to clarify the difference between FGM and diffusion distillation by pointing out technical challenges. We also add a paragraph in Section 5.1 to compare the training efficiency of FGM and CFM, showing that FGM can significantly surpass CFM on both generation results and training efficiency.\\n2. In response to **Reviewer 4qEE** 's suggestion. In Section 4.2, **we have included a comprehensive explanation of the stop $\\\\theta$-gradient technique** utilized within our model. This addition aims to clarify its implementation details and the impact on gradient backpropagation\\n3. As suggested by **Reviewer BVRU**, **we have revised the notation in Algorithm 1** for improved clarity. Specifically, we replaced $v_{\\\\text{sg}[\\\\theta]}$ with the actual online flow model $v_\\\\psi$ that we employ. Furthermore, we have provided additional explanations concerning the online flow model $v_\\\\psi$ in Section 4.2 to enhance understanding.\\n4. Addressing the feedback from **Reviewer 4qEE** and **Reviewer BVRU**. **we have added detailed pointers in Section 5** that direct readers to the appendix for comprehensive training details, and **expanded our ablation studies**, which can be found in Appendix C. These studies focus on two primary areas: (C.1) generator initialization, and (C.2) the impact of including or excluding regression loss during training.\\n\\n**If you have any additional issues or concerns, we would be happy to address them promptly.**\"}", "{\"title\": \"Paper Revision\", \"comment\": \"# Dear Reviewers and Area Chair,\\n\\nWe thank all reviewers for their useful feedback. We have made significant updates to our draft to improve the overall writing quality, provide more clarifications, and add additional experimental results. We have highlighted the changes in the draft in blue color. Here are the major changes:\\n\\n1. As **Reviewer sdTK** and **Reviewer jkBP** wish, **we polish the presentation of the method**. We add a paragraph in Section 4.2 to clarify the difference between FGM and diffusion distillation by pointing out technical challenges. We also add a paragraph in Section 5.1 to compare the training efficiency of FGM and CFM, showing that FGM can significantly surpass CFM on both generation results and training efficiency. \\n\\n2. In response to **Reviewer 4qEE** 's suggestion. In Section 4.2, **we have included a comprehensive explanation of the stop $\\\\theta$-gradient technique** utilized within our model. This addition aims to clarify its implementation details and the impact on gradient backpropagation\\n\\n3. As suggested by **Reviewer BVRU**, **we have revised the notation in Algorithm 1** for improved clarity. Specifically, we replaced $v_{\\\\text{sg}[\\\\theta]}$ with the actual online flow model $v_\\\\psi$ that we employ. Furthermore, we have provided additional explanations concerning the online flow model $v_\\\\psi$ in Section 4.2 to enhance understanding.\\n\\n4. Addressing the feedback from **Reviewer 4qEE** and **Reviewer BVRU**. **we have added detailed pointers in Section 5** that direct readers to the appendix for comprehensive training details, and **expanded our ablation studies**, which can be found in Appendix C. These studies focus on two primary areas: (C.1) generator initialization, and (C.2) the impact of including or excluding regression loss during training.\\n\\n**Change 1** improved the overall presentation of the paper a lot. **Change 2** aided readers in comprehensively understanding the purpose and functionality of the stop-gradient technique within our model. **Change 3** clarified the notation utilized in Algorithm 1, rendering it more accessible and easier for readers to follow. **Change 4** expanded our ablation studies, which substantially enrich our findings and provide deeper insights into the research.\\n\\nWe appreciate constructive suggestions from all reviewers that help strengthen our draft. \\n\\nBest,\\n\\nAuthors of submission #8666\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"comment\": \"Thank you for your constructive feedback on our manuscript. We will address these questions below. Before that, we will first provide a summary of the key contributions of our work.\\n\\nIn this paper, we introduce **Flow Generator Matching (FGM)**, an innovative approach designed to accelerate the sampling of flow matching models into a **one-step generation model**. Our results on CIFAR10 unconditional set a new record **FID score of 3.08** among all flow-matching models. Futhermore, our distillation results on **SD3-medium** demonstrates outstanding performance among other few-step distillation models. In addition to experimental performance, we propose the **flow-projection identity** to bypass the intractable flow-matching objective, leading to our **practical FGM training objective with theoretial guarantees**, which offers a foundation for potential developments in future research.\\n\\n**Q1: Ablation Study on Generator Initialization**\\n\\n**A1:** We choose several $t^*=[0.00, 0.25, 0.50, 0.75, 1.00]$ to train from scratch on 512-px, and the qualitative results are presented in our revised submission **appendix page 22, Figure 6**. Notes that our model parameterization for the ablation can be simplified as\\n$$\\n\\\\hat {x_0} = z- v_\\\\theta(z, t^*), \\\\quad z\\\\sim \\\\mathcal{N} (\\\\mathbf 0, \\\\mathbf{I})\\n$$\\n\\nThe visual results indicate that a suitable range for $t^*$ should be $[0.75, 1.00]$. However, the cost of further determining the optimal choice for $t^*$ is likely to be high and may not yield significant value. A key observation is that as $t^*$ decreases, the structural integrity of the images tends to deteriorate. This phenomenon can be attributed to the property of pre-trained flow matching model. When noise intensity is high, the model primarily focuses on generating the overarching structure of the image. Conversely, at lower noise intensity, the model leans toward creating finer details based on the pre-existing structure. However, in our one-step model, this foundational structure is absent, resulting in divergence.\\n\\n**Q2: Why regression loss $\\\\mathcal{L}_1$ is unnecessary in practice?**\\n\\n**A2:** We conduct two experiments on an early checkpoints, one training with both loss $\\\\mathcal{L}_1+\\\\mathcal{L}_2$, another training with only $\\\\mathcal{L}_2$, please check our revised submission **appendix page 23, Figure 7** ,our results show that simply apply the extra regression loss $\\\\mathcal{L}_1$ quickly degrade the performance. From the visual results we can tell that the model trained with $\\\\mathcal{L}_1$ resulting noisy images and quickly corrupted. So the regression term is omitted in our training.\\n\\n**We hope our rebuttal have resolved all your concerns. If you still have any concerns, please let us know. We are glad to provide further clarifications as well as more additional experiments.**\"}", "{\"comment\": \"Thank you for your constructive feedback of flow-generator matching. Below, we address each of your concerns in detail. Before that, we will first provide a summary of the key contributions of our work.\\n\\nIn this paper, we introduce **Flow Generator Matching (FGM)**, an innovative approach designed to accelerate the sampling of flow matching models into a **one-step generation model**. Our results on CIFAR10 unconditional set a new record **FID score of 3.08** among all flow-matching models. Futhermore, our distillation results on **SD3-medium** demonstrates outstanding performance among other few-step distillation models. In addition to experimental performance, we propose the **flow-projection identity** to bypass the intractable flow-matching objective, leading to our **practical FGM training objective with theoretial guarantees**\\u2014providing a solid foundation for potential advancements in future research.\\n\\n**Q1: The online flow model**\\n\\n**A1:** The online flow model $v_\\\\psi$ does represent the implicit vector field, which is used to approximate the vector field of the generator. Since we cannot explicitly compute the generator's vector field (as it is no longer a flow matching model), we learn the flow vector field $v_{\\\\theta,t}$ of the generator distribution through this online flow model $v_\\\\psi$. \\n\\nOn the other hand, the online flow model $v_\\\\psi$ replaces the term $v_{sg[\\\\theta],t}$ in the loss function. Therefore, even though the online flow model cannot be differentiated with respect to $\\\\theta$, it does not affect the gradient's ability to propagate back to the generator through $x_t(\\\\theta)$. So the equivalent notation for equations (4.11) and (4.12) is presented as follows:\\n\\n$$\\n\\\\mathcal{L}\\\\_{1}(\\\\theta) = \\\\mathbb{E}\\\\_{t, z\\\\sim p_z, x_0=g_\\\\theta( z), \\\\atop x_t\\\\sim q_t( x_t| x_0)} \\n\\\\{ \\\\\\\\| u_t( x_t) - v_{\\\\psi}( x_t, t)\\\\\\\\|_2^2 \\\\}\\n$$\\n\\n$$\\n\\\\mathcal{L}\\\\_2(\\\\theta) = \\\\mathbb{E}\\\\_{t, z\\\\sim p_z, x_0=g_\\\\theta( z), \\\\atop x_t| x_0\\\\sim q_t( x_t| x_0)} [2 (u_t( x_t) - v_{\\\\psi}( x_t, t) )^T (v_{\\\\psi}(x_t, t) - u_t( x_t|x_0))]\\n$$\\n\\n**Q2: Different notation for $u_t$**\\n\\n**A2:** In response to the different notations regarding $u_t(x_t|x_0)$ and $u_t(x_t)$ in above Eq(4.12), it is important to clarify that the first $u_t(x_t)$ represents the pre-trained flow-matching model, which should be considered as a marginal vector field, while the latter term is introduced by **Theorem 4.1 (Flow Product Identity)**; it denotes the conditional vector field used in training of the flow matching model, which is usually the difference between data and sampled Gaussian noise in practice.\\n\\n**Q3: Abuse of Notation \\\\& Missed Reference**\\n\\n**A3:** Thank you for pointing out the notation issue on line 203 and missed reference in our table. We have made these adjustments in the revised manuscript to ensure clarity and correctness.\\n\\n**Q4: Quality Degradation** \\n\\n**A4:** We present our additional training results in our revised submission **appendix page 24, Figure 8**, comparing the previous results with those from the model trained for more steps. Given that we are fine-tuning a transformer model with 2B parameters, this artifact is typically observed in the early stages of training. This suggests that it can be substantially mitigated, and the overall image quality can also be enhanced with more extensive training.\\n\\n**Q5: Ablation Study on $t^{*}$**\\n\\n**A5:** We choose several $t^*=[0.00, 0.25, 0.50, 0.75, 1.00]$ to train from scratch on 512-px, and the qualitative results are presented in our revised submission **appendix page 22, Figure 6**. . Notes that our model parameterization for the ablation can be simplified as\\n\\n$$\\n\\\\hat {x_0} = z- v_\\\\theta(z, t^*), z\\\\sim \\\\mathcal{N} (\\\\mathbf 0, \\\\mathbf{I})\\n$$\\n\\nThe visual results indicate that a suitable range for $t^*$ should be $[0.75, 1.00]$. However, the cost of further determining the optimal choice for $t^*$ is likely to be high and may not yield significant value. A key observation is that as $t^*$ decreases, the structural integrity of the images tends to deteriorate. This phenomenon can be attributed to the property of pre-trained flow matching model. When noise intensity is high, the model primarily focuses on generating the overarching structure of the image. Conversely, at lower noise intensity, the model leans toward creating finer details based on the pre-existing structure. However, in our one-step model, this foundational structure is absent, resulting in divergence.\\n\\n**We hope our rebuttal have resolved all your concerns. If you still have any concerns, please let us know. We are glad to provide further clarifications as well as more additional experiments.**\"}" ] }

[ "Mia Cameron", "Yusi Chen", "Terrence Sejnowski" ]

[ "biologically-plausible learning rules", "Newton-like methods", "local learning rules" ]

{ "note_id": [ "qb1QzknO97", "nuzthJzJKl", "nkqW49z01U", "lUAKXB9Ykt", "ajWViROWDJ", "XcIlOq9qGP", "7Ig9x8orAE", "2jSwTILSXq", "0CCOvwN9gq" ], "note_type": [ "decision", "official_comment", "official_comment", "official_review", "meta_review", "official_review", "official_comment", "official_comment", "official_review" ], "note_created": [ 1737524141807, 1732786709585, 1732740533048, 1730696738235, 1734573733587, 1730383816254, 1732792409182, 1732738484767, 1729517026873 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission11720/Reviewer_NytU" ], [ "ICLR.cc/2025/Conference/Submission11720/Authors" ], [ "ICLR.cc/2025/Conference/Submission11720/Reviewer_wCKF" ], [ "ICLR.cc/2025/Conference/Submission11720/Area_Chair_KcXJ" ], [ "ICLR.cc/2025/Conference/Submission11720/Reviewer_ujbb" ], [ "ICLR.cc/2025/Conference/Submission11720/Reviewer_ujbb" ], [ "ICLR.cc/2025/Conference/Submission11720/Authors" ], [ "ICLR.cc/2025/Conference/Submission11720/Reviewer_NytU" ] ], "structured_content_str": [ "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"comment\": \"Thank you for your response.\\n\\n__W1__: Insufficiently contextualized within recent literature\\n\\nThe embedding in the literature has improved somewhat. However, missing is still any reference to how your method (which learns the Moore-Penrose inverse) is similar to/alternative to the approach of Podlaski et al. 2020 (Biological Credit Assignment through Dynamic Inversion of Feedforward Networks) which dynamically produces the same pseudo-inverse. Note also some misconceptions in your statements: GAIT-prop in fact shows equivalence to backpropagation and not Gauss-Newton optimization.\\n\\nHaving said that, the framework of your paper is indeed a useful contribution, though weaknesses remain.\\n\\n__W2__: Experiments are too limited\\n\\nExperiments are limited but also in my opinion far far away from being convincing. Even considering the current results as they are, BP show extremely low accuracy at even MNIST. As an example, in the Podlaski paper they were able to train BP to a test error of 2.2% and target propagation (with dynamic inversion) to an error of 2.8%. This is far away from the >7% error rate that you show for all algorithms. And in their case it was a simple two layer network (784-1000-10 nodes). I believe that the results at present are misleading compared to similar existing literature. I would even be happy to accept the current scale of the tasks if they were carried out more robustly and for CIFAR-10 with a proper data augmentation (which I assume is missing given it's low accuracy).\\n\\nFurthermore, note that there are a few errors in the paper at present: I believe your headings in Figure 4 are the wrong way around.\\n\\n__W3__: Hyperparameters are not included\\n\\nThank you for their inclusion.\"}", "{\"comment\": \"**W1:** Issues with the sleep/wake training scheme\\n\\nThank you for your suggestions. In the latest draft, we have included simulations where forward and feedback weights are trained concurrently utilizing an iteratively updated approximate pseudoinverse at each training step. This method has a lower computational complexity, but has been less rigorously studied. \\n\\nUnder the sleep-wake training scheme, we assume several steps are made during each phase, with a small enough learning rate such that the feedback weights learned during the previous \\u201csleep\\u201d period are close enough to the true pseudoinverse. \\n\\n\\n**W2:** Convergence speed of second-order optimization\\n\\nThank you for pointing this out. When the loss landscape is relatively convex, it is theoretically true that second-order methods will tend to converge faster than first-order gradient descent, since each step size is \\u201cscaled\\u201d by the inverse of the Hessian matrix. Practical approximations of second-order optimization, such as K-FAC (Martens and Grosse 2015) and Shampoo (Gupta, et. al 2018), tend to converge faster than gradient descent, usually at the cost of additional computational complexity. \\n\\nIn regards to convergence to flatter minima, we acknowledge that there is not enough experimental evidence to support this, and we will remove that statement in our future draft. \\n\\n\\n**W3:** Biological implausibility in comparison to predictive coding\\n\\nFrom our understanding, the family of models underthe predictive coding framework have included stacked-autoencoder architectures similar to ours (such as Marino 2020, and Whittington and Bogatz 2017). However, we acknowledge that our architecture is more simplified than those referenced, as do not include recurrent lateral connections, or error-encoding neurons. \\n\\nIn regards to directly comparing with PC, we cannot compare our method to PC with weight constraints relaxed, as our method is a proposed solution to the weight transport problem itself. In fact, when benchmarked against random feedback weights (Feedback Alignment), our method converges faster. **Our algorithm is not intended as a direct alternative to energy-based PC architectures, such as that of Millidge et. al 2020, but rather a solution to the weight transport problem encountered by many such models** (inspired by the PC framework). \\n\\n**Q1:**\\nCould you please clarify what you mean by the algorithm being completely parallel? Is that in reference to having separate sleep/wake training phases?\\n\\nJames C. R. Whittington, Rafal Bogacz; An Approximation of the Error Backpropagation Algorithm in a Predictive Coding Network with Local Hebbian Synaptic Plasticity. Neural Comput 2017; 29 (5): 1229\\u20131262. doi: https://doi.org/10.1162/NECO_a_00949\\n\\nMarino, J. (2021). Predictive Coding, Variational Autoencoders, and Biological Connections. arXiv [Cs.NE]. http://arxiv.org/abs/2011.07464 \\n\\nGupta, V., Koren, T., & Singer, Y. (2018). Shampoo: Preconditioned Stochastic Tensor Optimization. arXiv [Cs.LG]. http://arxiv.org/abs/1802.09568\\n\\nMartens, J., & Grosse, R. B. (2015). Optimizing Neural Networks with Kronecker-factored Approximate Curvature. CoRR. http://arxiv.org/abs/1503.05671\"}", "{\"summary\": \"In this paper, the authors propose a mechanism to train deep neural networks that solves some of the biological implausible aspects of back propagation. In particular, they propose Reciprocal Feedback:\\n* The propose a modification of the Recirculation algorithm that can learn the Moore-Penrose pseudo-inverse of the feedforward (or feedback) weights\\n* Using the Hildenbrandt-Graves Theorem they show that the learned pseudo inverse can be used as an alternative to traditional gradient descent (which relies on the transpose of the feedforward weights)\\n\\nThey show some preliminary result on the Mnist and Cifar10 classification tasks, and compare them with Backpropagation and biologically plausible algorithms.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"3\", \"strengths\": [\"I believe the paper is well structured and shows interesting methods and results. In detail:\", \"The theory part is well explained and are very complete and self contained.\", \"The results and dynamics of the algorithms are well explained by the theory.\", \"The assumptions needed for the algorithm to work are clearly stated.\"], \"weaknesses\": [\"I believe this paper is solid, but I will give it a 5 because of the insufficient contextualization with the previous literature. I would be happy to improve my score if these concerns are addressed:\", \"Section 2 (Related work) is very brief and it is not enough for readers who are not familiar with the mentioned algorithms. For example, a more detailed explanation of Target Propagation, Recirculation Algorithm and Weight-Mirroring would be useful to make the paper more clear.\", \"I like the fact that the authors jump straight into the theory, but I believe that I could be better contextualized by comparing the results with the Related Work section algorithms.\", \"I believe the results on the MNIST on CIFAR10 are missing some key details: how were the hyperparameters chosen? I find the results for backpropagation surprisingly worse than what I would expect (e.g, >97% for MNIST)\", \"In general, although the paper is well written, I believe space could be optimized to include these suggestions, in particular a more complete explanation of the related work and a comparison, which are needed to contextualize the paper.\"], \"questions\": [\"It is not super clear to me how this algorithm compares to others in terms of computational complexity, it would be nice to see it as a metric of comparison\", \"The word biologically plausible is used a lot, and I do agree that this algorithm may solve some of the issues of backpropagation (e.g. weight simmetry), but there is no mention on how it could be implemented in biological networks. Could you elaborate more on that?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"metareview\": \"The authors propose a novel optimization method, Reciprocal Feedback Learning, which aims to address the biological implausibility of backpropagation. Preliminary results are presented on MNIST and CIFAR-10, comparing the proposed approach with backpropagation and other biologically inspired methods.\\n\\nWhile the paper provides a solid theoretical foundation and an interesting contribution to biologically plausible learning, several issues were identified by the reviewers, particularly in terms of contextualization, experimental rigor, and comparison with prior work. All three reviewers rated the submission as below the acceptance threshold, citing overlapping concerns.\", \"additional_comments_on_reviewer_discussion\": \"The authors made some improvements during the rebuttal period, addressing contextualization, experiments, and minor presentation issues, but the responses fell short of fully addressing the reviewers\\u2019 major concerns.\"}", "{\"summary\": \"This work aims to describe a biologically plausible alternative to the backpropagation of error algorithm. The authors contend that algorithms sending predictions to lower layers are more biologically plausible than those that depend on feedback error networks. Their algorithm, called reciprocal feedback learning, which falls in the former category, involves a biologically plausible mechanism (i.e., local numerical integration) to compute the pseudoinverse of the weights of a layer. This mechanism is an extension of recirculation algorithms. They then use the Hildebrandt-Graves Theorem to develop an algorithm to propagate errors in the backward pass of the network. The specific claim of the paper is that this algorithm avoids the biological implausibility of weight transport by circumventing explicit calculation of a weight transpose. On MNIST and CIFAR, they show that the algorithm has comparable asymptotic performance to backpropagation and converges in fewer iterations than conventional random target propagation algorithms. The contribution is a comprehensive derivation of a novel algorithm to calculate the pseudoinverse, a novel training algorithm, positive evidence from benchmark comparisons, and an overall claim of biological plausibility compared with other algorithms.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"The idea is interesting and clearly and concisely presented. The algorithm is thoroughly derived and described with thoughtful attention to detail. The work makes some nice connections between previously disparate algorithms and theorems. It adds a novel idea to the growing research on alternative backpropagation, particularly in algorithms that address the weight transport problem.\", \"weaknesses\": \"The work is poorly situated in the existing literature on biologically plausible alternatives. Thus, a main weakness of the work is the insufficient comparison and benchmarking against other algorithms. In particular, the biological motivation is rather weak, especially considering that the authors recognize that predictive coding algorithms have strong biological underpinnings. While it is true that predictive coding (PC) suffers from the weight transport problem, there have been works showing it is also robust to randomness in backward connections (see https://arxiv.org/pdf/2010.01047). I would expect a comparison against these algorithms, especially when weight transport assumptions are relaxed.\\n\\nIt is true that PC has not been scaled to problems on the scale of ImageNet. If the argument here is that this algorithm comes from a family of algorithms that have been shown to scale, there should be an attempt to test it on ImageNet. More generally, the algorithm is only tested against an early algorithm from this family\\u2014random feedback\\u2014rather than newer algorithms like sign symmetry.\\n\\nOverall, it is not clear what merit this algorithm has over existing literature. From an engineering perspective, it is another compute-heavy alternative to backpropagation, and it is not clear how it is better than other algorithms. For example, it would be good to compare the computational expense (how many pseudoinverse iterations per learning update are needed and the number of learning updates required to achieve asymptotic performance) and accuracy to BP, PC, target propagation, etc.\\n\\nIts alignment with neuroscience appears weak, and methods like PC have much greater resonance with neural circuitry. While this algorithm solves the weight transport issue for PC, its biological potential seems limited. Until a closer analysis of possible underlying neural circuitry is conducted, or at least a direct comparison with PC when weight transport assumptions are relaxed, its biological plausibility remains uncertain.\", \"questions\": \"One of the biggest barriers to the adoption of alternatives to backpropagation is increased computational expense. Reducing this is a key focus of the biomorphic algorithms community, and a comparison against the computational expense of other similar algorithms is needed. For example, it is known that PC algorithms can be trained with roughly 2N iterations, whereas this method requires at least 60. How robust is the algorithm if the pseudoinverses do not fully converge?\\nWhile the description of the system as a sleep-wake algorithm is interesting, I am unsure of the biological plausibility. Aren't you limited to a single update during the wake phase, after which you would have to learn another pseudoinverse? Similarly, how robust is the algorithm to performing multiple weight updates in the wake phase? For small learning rates, might there be some tolerance? These aspects should be tested.\\n\\nRelatedly, I assume that recalculating the pseudoinverse after one step of learning might actually be inexpensive, i.e., the original pseudoinverse could provide a good initialization.\\n\\nIn the discussion, the paper states that second-order, Newton-like methods tend to converge faster and to flatter minima than gradient descent, but they cite only one paper as evidence. Is this always true? Could you provide additional references to support this claim, clarify under what conditions it holds, or be more conservative with this claim?\\n\\nAnother concern I have is that (unlike target propagation) the error signal is backpropagated from the output layer all the way through the network. Thus, it is not clear if this computation is completely parallel, as in other algorithms (e.g., PC). Doesn't this jeopardize the biological plausibility?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Nice proof of concept but needs a more detailed quantitative and qualitative comparison to other algorithms\", \"comment\": \"Thanks for the response. It is nice to see there is evidence that concurrent training, rather than the sleep-wake conceptualization, is more compute efficient, but ideally, this does need to be rigorously studied to augment the overall contribution.\", \"on_the_parallel_nature\": \"I mean that in PC, inference and learning are completely parallelized, i.e., there is no such thing as a backward or forward pass. I think this algorithm, if I\\u2019m not mistaken, has an effective backward pass during learning, requiring coordination of how the information flows.\\n\\nWhile I appreciate the extra results, I still feel that comparison to other algorithms has not been comprehensive, and thus it\\u2019s not clear whether there are significant contributions to biological plausibility or engineering. Relaxing the weight-transpose problem is interesting, but it also introduces a suite of other commitments that plausibility needs to be argued for. As an engineering paradigm, it is not clear how computationally expensive it is compared to other algorithms. I agree this is a nice proof of concept, but without more comprehensive investigation and comparison of these ideas to other works I will maintain my already relatively high score.\"}", "{\"comment\": \"**W1**: Insufficiently contextualized within recent literature\\n\\nThank you for bringing these papers to our attention. \\n\\nOur algorithm is similar to that proposed by Meuleumans et. al (2020), and we identify the same problems as them in connecting target propagation to Gauss-Newton optimization. More specifically, we both identified the problem that the pseudoinverse of the whole network can\\u2019t be factorized as the pseudoinverse of each individual layer. \\n\\nHowever, our solution differs from Meulemans et. al (2020), in that **we do not use direct feedback connections** from the final layer to each intermediate layer. Our method preserves the modular, stacked-autoencoder structure proposed under the target-propagation framework, and uses a Newton-like optimization method which **only requires local, layer-wise feedback connections**. Overall, we believe that our proposal constitutes a different theoretical framework for target-propagation-like methods, which has greater biological-plausibility than Meulemans et. al. (2020). \\n\\nAhmad et. al (2020) and Bengio et. al (2020) also propose similar methods to ours, with the use of layer-wise inverses to propagate error. However, the mathematical derivations in Ahmad et. al (2020) and Bengio et. al (2020) both assume that the network is invertible. In that case, both our method, GAIT-prop and target-propagation are equivalent to Gauss-Newton optimization. Our method differs from these previous approaches in that **we do not require perfect invertibility** - allowing for greater architectural flexibility. \\n\\nOverall, our method can be considered an additional mathematical framework by which the methods in these papers can be better understood and generalized. \\n\\n**W2**: Experiments are too limited \\n\\nWe agree that the current set of experiments are limited, however, we think that they constitute a sufficient proof-of-concept for a new optimization method, which we have validated mathematically. \\n\\nFurthermore, in our updated draft, we will include multiple seed initializations and run each method for more epochs. \\n\\n**W3**: Hyperparameters are not included \\n\\nThank you for the suggestion. We have included a table of hyperparameters in our latest draft\"}", "{\"summary\": \"This work presents an alternative to weight-transpose based measurement of gradients for updating and optimization of neural network models. Specifically, a method for the use of (pseudo)inverse-based top-down models of each layer of a network\\u2019s activity is described. A \\u2018sleep-wake\\u2019-esque cycle of updating is used to compute the Moore-Penrose pseudoinverse of weight matrices in deep neural network architectures, with a relation to the recirculation algorithm. Thereafter, the Hildebrandt-Graves Theorem is applied to show how such pseudoinverses can provide feedback to layers of a neural network for an alternative optimization method of neural networks. This is applied to the MNIST and CIFAR-10 classification tasks to demonstrate its performance.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"1\", \"strengths\": [\"The work has a well written narrative structure. A reader is led well through the method as well as through the steps for running this method. Algorithms and mathematical steps are relatively clear.\", \"The relation of inverse-based credit assignment to the Hildebrandt-Graves Theorem is original (novel) and can contribute to the discussion around the plausibility or efficacy of target-propagation methods.\", \"Illustrations are used effectively to describe the process of computing the proposed inversion and forward parameter learning processes.\"], \"weaknesses\": \"- This work has not been embedded within recent literature. This is a major drawback as it misses out the ways in which existing research has contributed to this question and does not provide comparison (theoretical or empirical) against these methods. There exist works which describe: how target propagation relates to iterative approximate inverses (Bengio 2020), how to dynamically compute the Moore-Penrose inverse (Podlaski et al 2020), how the use of inverses relates to second order optimization (Meulemans et al. 2020), and how in special cases the inverse and transpose are equivalent (Ahmad et al. 2020). These are just some of the recent works which have contributed to this story and are missing in this context, many of which also fulfil some of the theoretical requirements outlined in this work.\\n- The experiments presented are much too limited to draw conclusions from. In the results shown, simulations are extremely short (in terms of epochs), do not have multiple repeats, and are within a relatively low task complexity range. Existing work has often indicated that traditional target propagation can fail suddenly, and not scale to tasks of greater difficulty than MNIST or CIFAR-10 (See Bartunov et al. 2018). Furthermore, the results reached by backpropagation are far far short of what is possible with longer training and coupling with adaptive optimizers. To prove that these results are robust and scale, one would desire multiple repeats, training until convergence (100+ epochs), and application to a more challenging task. Furthermore, one would expect comparison against traditional target propagation or difference target propagation to see whether this method is comparable or worse.\\n- The computational complexity and robustness of hyperparameters is not much discussed but would be important for recreation of results. A hyperparameters table is missing for a reader to understand exactly how the parameters were tuned.\\n- There are some minor textual issues, for example feedback alignment is referred to as \\u2018random target propagation\\u2019 in the abstract (a term that this reviewer has never encountered) but later referred to as \\u2018feedback alignment\\u2019 as it is called in the original referenced paper. \\n\\n\\nBengio, Y. (2020). Deriving Differential Target Propagation from Iterating Approximate Inverses. In arXiv [cs.LG]. arXiv. http://arxiv.org/abs/2007.15139\\n\\nPodlaski, W. F., & Machens, C. K. (2020). Biological credit assignment through dynamic inversion of feedforward networks. In arXiv [q-bio.NC]. arXiv. http://arxiv.org/abs/2007.05112 (Published in Neurips 2020)\\n\\nMeulemans, A., Carzaniga, F. S., Suykens, J. A. K., Sacramento, J., & Grewe, B. F. (2020). A Theoretical Framework for Target Propagation. In arXiv [cs.LG]. arXiv. http://arxiv.org/abs/2006.14331 (Published Neurips 2020)\\n\\nAhmad, N., van Gerven, M. A. J., & Ambrogioni, L. (2020). GAIT-prop: A biologically plausible learning rule derived from backpropagation of error. In arXiv [cs.LG]. arXiv. https://arxiv.org/abs/2006.06438 (Published Neurips 2020)\\n\\nBartunov, S., Santoro, A., Richards, B. A., Marris, L., Hinton, G. E., & Lillicrap, T. (2018). Assessing the Scalability of Biologically-Motivated Deep Learning Algorithms and Architectures. In arXiv [cs.LG]. arXiv. http://arxiv.org/abs/1807.04587 (Published Neurips 2018)\", \"questions\": \"I have no further questions and would point to the weaknesses section for a full list of actionable critiques. The review period is rather short to correct such a list of weaknesses. Nonetheless, should all of these points be addressed sufficiently I would be willing to revise my score.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}" ] }

[ "Songyuan Yang", "Long Lan", "Xihuai He", "Xueqiong Li", "Lionel Z. WANG", "Mengzhu Wang", "Huibin Tan" ]

[ "Human Motion Generation; VLM; 3D Generative Models" ]

{ "note_id": [ "dYW8JuJNUm", "cdyihqrFcV", "SvDkB5hTZq", "BHKKU1mJcG", "AQ6NN46gV9", "5sB1BZ9jVC" ], "note_type": [ "official_review", "official_review", "official_review", "comment", "official_review", "official_review" ], "note_created": [ 1730678482803, 1730703712067, 1730721560637, 1731428388249, 1730454670634, 1730566417418 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission428/Reviewer_hpvf" ], [ "ICLR.cc/2025/Conference/Submission428/Reviewer_GYtZ" ], [ "ICLR.cc/2025/Conference/Submission428/Reviewer_vMhQ" ], [ "ICLR.cc/2025/Conference/Submission428/Authors" ], [ "ICLR.cc/2025/Conference/Submission428/Reviewer_SRSP" ], [ "ICLR.cc/2025/Conference/Submission428/Reviewer_Y1A7" ] ], "structured_content_str": [ "{\"summary\": \"This paper compensates for the lack of multi-person interaction in previous HSI tasks. It constructs a dataset named HumanVL that contains multi-person actions and is aligned with scenes. On this basis, a model called MMG-VL is designed, providing an effective solution for generating realistic multi-person scenarios.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"The proposed dataset has relatively rich scenes and includes multi-person actions.\", \"weaknesses\": \"## W1.Insufficient related work review\\n1. \\\"FreeMotion: A Unified Framework for Number-free Text-to-Motion Synthesis\\\" and \\\"CrowdMoGen: Zero-Shot Text-Driven Collective Motion Generation\\\" both perform text-to-nPerson motion synthesis and utilize large models for text-level design. These two paper should be cited and the differences between them and this paper should be clarified.\\n\\n## W2.Dataset construction has some problems.\\n1. Lack of human-human interaction. Although this paper have placed actions of 2 to 7 people in the room, from the content and materials provided, I don't seem to find any interaction actions between people. If there is no interaction between people, then what is the difference between this and single-person actions?\\n2. MDM is a model for generating single-person actions and does not consider how to interact with the scene. If human body behaviors are constructed through MDM, how does it interact with the scene?\\n3. Does the dataset only contain images of the scene? If there is only images and no 3D object representation, how to put motion into the scene?\\n\\n## W3.The model implementation lacks many details.\\n1. Why leverage a LoRA to fine-tune a VLM? Since the data scale is limited, will the fine-tuning process cause the VLM to overfit to limited action descriptions, resulting in the loss of generalization ability of the VLM?\\n2. I am worried that if a full-image are used to extract the viusal feature, how can we ensure that the generated actions can interact with the correct object in the scene?\\n3. What is the difference between the input representation and the representation provided by HumanML3D?\\n4. How to ensure that there will be no collisions or interpenetrations between the actions of multiple people generated by only using VLM to extract image features?\\n5. What is the specific fusion process of $v_{feat}$ and $l_{feat}$ at line 265?\\n6. Line 279 mentioned that the motion generation process for each individual is based on their respective instructions. Then what is the difference from single-person action generation? Why can this paper achieve the generation of actions of more people?\\n7. From Line 368, this paper conducts full fine-tuning of the MDM. But how to deal with the difference between the fused feature $c$ in this paper and the text feature in MDM?\\n\\n## W4.Lack of experimental details.\\n1. It is best to provide an ablation experiment to verify the role of LoRA, such as removing the fine-tuning process and comparing the results.\\n2. It is better to describe the details of evaluation metrics (SQ, SD, CC, EI, MPC, MRC).\\n3. There are few qualitative results, and the generation effect is not persuasive.\", \"questions\": \"Please refer to the weakness\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper presents a novel challenge of generating multi-human motions in multi-room home environments. The authors have curated the HumanVL dataset, which contains 584 multi-room household images and 35,622 human motion samples. They present the MMG-VL approach, which uses a visual language model (VLM) conditioned on top-down images of scenes to generate detailed motion descriptions for each individual. Then, leveraging the MDM method, they generate motion sequences for each person individually. The authors compare the performance of MMG-VL with prior methods across three datasets: HumanML3D, InterHuman, and the proposed HumanVL.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"1. The authors introduce a multi-person, multi-room human motion generation dataset, addressing a notable gap in the field.\\n2. The qualitative figures are of high quality.\", \"weaknesses\": \"1. **Layout and Formatting**: The layout is too dense, with minimal spacing before section captions. The authors should consider removing some content or relocating parts to the appendix to restore a more standard layout.\\n2. **Lack of Motivation for Multi-Person Dataset**: The rationale for introducing a multi-person dataset is unclear. The proposed method and sample data do not explicitly model interactions between individuals, which could be addressed with iterative single-person motion generation. In fact, the HumanVL multi-person data generation proposed here is performed by generating each person individually with MDM, without accounting for other individuals, reducing the significance of a multi-person setting. The current multi-person data appears to be merely a spatial combination of multiple single-person motion sequences. It would be more interesting if the authors consider modeling inter-person interactions.\\n3. **Unclear SID Module Description**: The SID module is not clearly described. In the formulas in Section 3.3.2, it\\u2019s unclear how the SID module interacts with the environment; the single motion sequence seems to rely solely on text instruction $c_i$, which doesn\\u2019t make sense in this context. The authors should explain how environmental information is incorporated into the SID module.\\n4. **Lack of Experimental Metric Descriptions**: There is insufficient explanation of how Diversity, Multimodality, and Multi-modal Distance metrics are calculated.\\n5. **Typos**:\\n - Line 230: \\\"Dulti-person\\\" should be \\\"Multi-person\\\"\\n - Line 278: \\\"f_{split}\\\" should be \\\"f_{MGC}\\\"\", \"questions\": \"The paper needs to reconsider the introduction of multi-person scenarios from both the dataset and method design perspectives. It would benefit from reorganizing the content to clearly articulate the motivation.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper presents a method for multi-person motion generation in indoor scenes. The proposed method includes two stages, the vision-language auxiliary instruction (VLAI) module and the scenario-interaction diffusion (SID) module. The VLAI module generates multi-human motion prompts from the visual input and the text. The SID module is a human motion generation module to generate human motion for each individual. And this paper builds a multi-human motion dataset, HumanVL. In the experimental section, the proposed method is evaluated on the single-human motion generation and multi-human motion generation.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"Here, I highlight the strength of this paper:\\n1. The problem this paper focuses on, multi-human motion generation in indoor scenes, is worthy of study in the field of human motion generation.\\n2. This paper provides visualizations of the generation results, which are intuitive.\\n3. This paper is easy to read.\", \"weaknesses\": \"Although the problem this paper focuses on is valuable, this paper only proposes a simple pipeline based on existing methods, the novelty of the proposed method is limited. Here I highlight my major concerns.\\n\\n1. The novelty of the proposed method is limited. For the two key modules in the proposed method, the VLAI uses existing ViT encoder and LLM, and the SID uses existing human-motion methods. Thus, the proposed method appears to be a combination of existing methods, and lacks substantial innovation.\\n2. About the modeling of human-scene interaction and human-human interaction. This paper relies only on LLM for scene and text understanding, and does not explicitly model these interactions. Then the upper limit of the method depends on LLM, and the error output of LLM will also affect the performance of the method.\\n3. The user study is unscientific. In the user study in this paper, only five PhD candidates took part in the experiment. The small number of people in the user study, the lack of details of the user study, etc., will make people question the user study.\\n4. The performance on the main metrics for human motion generation is bad, like FID, R Precision. These metrics reflect the quality of the generation.\", \"questions\": \"1. What is fsplit in L278?\\n\\n2. From L354 to L362, there is only the definition of the metric, not how the metric is calculated.\\n\\n3. The input of ViT encoder is 2D image of the scene, how to understand the 3D interaction between humans and scene?\\n\\n4. In the limitation, the paper claims that the proposed method only can generate two to three human motions, why does the first image use 7 people as an example?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"withdrawal_confirmation\": \"I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors.\"}", "{\"summary\": \"The paper focus on the multi-person scenarios, combines natural language instruction and environmental visual inputs to generate realistic 3D human motions. The paper propose a generation approach driven by vision and language for generating multi-person human motion in multi-room home scenarios. It firstly use vision-language auxiliary instruction to generate motion instructions align with real-world constraints, then it use scenario interaction diffusion to generate human motions. Moreover, it also provide a dataset, contains multi-room household images.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. The paper explore a challenge but valuable task, generating multi-person interaction motions.\\n2. The writing is well and contributions is enough, including proposed dataset and network module designs.\\n3. In the experimental section, the comparison with other methods and the visualization results are sufficient\", \"weaknesses\": \"1. The analysis and discussion of the dataset in the paper are insufficient. Maybe you can provide more details such as data distribution, annotation process, or specific dataset statistics.\", \"questions\": \"1. In the dataset, although some physical rules are considered in the instructions, do the interactions between people, people and objects, and people and scenes have physical constraints when generating motions for multiple people?\\n2. How to ensure that the instructions respect the physical constraints and logical affordances of the scene in dataset, can you show some examples?\\n3. Does the model incorporate constraints for foot-ground contact or collision avoidance between individuals? If so, how are these implemented within the generation process?\\n4. One suggestion, a better way to show you generation results with a video.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper introduces a new paradigm for motion generation that integrates natural language instructions with environmental visual inputs to produce 3D human motion. The authors present a two-stage pipeline, consisting of instruction generation followed by motion generation, to achieve this. Additionally, they provide a dataset featuring multi-room, multi-human motion samples.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"1. The established HumanVL dataset encourages further research on multi-human and scene interaction.\", \"weaknesses\": \"1. There is a lack of discussion and comparison with prior works on language-guided human motion generation in 3D scenes[1,2]. This paper only incorporates 2D scene information into the generation process, which merely informs the model of where to place the motion. The motivation behind the current setting needs further clarification.\\n\\n2. The proposed model lacks novelty in design, as it merely combines existing components. It uses a VL module to interpret input descriptions into instructions, which then guide motion generation through an MDM model. However, the model does not incorporate detailed scene information, relying solely on interpreted instructions, which would fail to generate results that align accurately with the scene.\\n\\n3. The proposed model appears to perform significantly worse than existing methods on R-Precision, FID, and MM-Dist metrics, so it is unclear why the authors claim \\\"it still demonstrates competitive performance.\\\"\\n\\n4. The paper lacks essential details, such as training details and the construction details of the HumanVL dataset. Please refer to the following questions.\\n\\n[1] Wang et al., HUMANISE: Language-conditioned Human Motion Generation in 3D Scenes. NeurIPS 2022.\\n\\n[2] Yi et al., Generating Human Interaction Motions in Scenes with Text Control. ECCV 2024.\", \"questions\": \"- About HumanVL:\\n\\n1. In the HumanVL dataset, are the language instructions generated manually or produced by an LLM?\\n\\n2. How is the motion aligned with the 3D scenes?\\n\\n- About Model:\\n\\n1. When instructions generated by VLAI relate to different scenarios, the model can generate distinct motions for different individuals. However, how does the model handle this if two instructions refer to the same scenario and overlap spatially?\\n\\n2. Since the model does not receive 3D scene information as input, how does it achieve the physically plausible results shown in your figures?\\n\\n- Other comments:\\n\\n1. The notation $x$ in lines 189 and 191 appears inconsistent in the use of superscripts and subscripts.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}" ] }

[ "Alexander Huang-Menders", "Kevin Lin", "Yu-Wing Tai" ]

[ "parameter efficient fine-tuning (PEFL)", "quantized LLMs", "LoRA", "hierarchical learning" ]

{ "note_id": [ "zChwVBUg4T", "jmjLXZe30I", "UPvcaYZm4z", "NHnB2nbcfX", "74AAbm0rfr" ], "note_type": [ "official_review", "official_review", "official_review", "comment", "official_review" ], "note_created": [ 1730561063516, 1730190693504, 1730537807950, 1731614865563, 1730600008211 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission2651/Reviewer_3Xy3" ], [ "ICLR.cc/2025/Conference/Submission2651/Reviewer_JLEK" ], [ "ICLR.cc/2025/Conference/Submission2651/Reviewer_CngU" ], [ "ICLR.cc/2025/Conference/Submission2651/Authors" ], [ "ICLR.cc/2025/Conference/Submission2651/Reviewer_Nhnf" ] ], "structured_content_str": [ "{\"summary\": \"This paper presents H-QLoRA, a method designed to fine-tune large language models (LLMs) more effectively in resource-constrained environments. Traditional approaches often use heavy quantization to reduce memory and computational needs, which can degrade accuracy. H-QLoRA, however, introduces hierarchical adaptors with low-rank weights to capture high-frequency information typically lost in 4-bit quantization, thus improving performance. Testing on models from the LLaMA and Gemma families shows that H-QLoRA outperforms state-of-the-art methods on multiple instruction datasets, achieving greater accuracy without increasing inference costs. By maintaining the same number of trainable parameters as QLoRA, H-QLoRA offers a more optimized solution for certain model types, advancing fine-tuning efficiency and outcomes in low-resource settings.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"1\", \"strengths\": [\"This paper clearly explains the idea of H-QLoRA\", \"The experiments are conducted on multiple model families\"], \"weaknesses\": [\"It shows that the H-QLoRA works well on \\u201cold and small\\u201d models, which makes this work less promising.\", \"Why using multiple additive adapters can improve the performance is not explained.\", \"The model is only tested on MMLU. It should be tested on more datasets like QA tasks to verify the effectiveness.\", \"In Table 2, the model trained on different datasets can give performance with significant gaps, which indicates the lack of generalizability.\"], \"questions\": [\"What does \\u201chierarchical\\u201d come from? Adapters are additive, I didn\\u2019t see any visual or theoretical analysis of how the adapters have hierarchical structures.\", \"The improvement seems not consistent in Table 1. Is it possible that the performance difference mainly comes from randomness, e.g. adapter initialization?\", \"Can you give any insights into your method? why additive adapters are better? when are they better? how to determine which config (e.g. 32, 32 or 32, 8, 8, 8, 8), any possible analysis? This paper seems more like preliminary observations if these questions are not answered.\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"Authors propose to modify QLoRA approach, by splitting low rank decomposition into sum of smaller low rank decompositions.\\nThey show that this approach can have better accuracy on some sets of models and data at cost of around 24.3% slower training time.\", \"soundness\": \"1\", \"presentation\": \"2\", \"contribution\": \"1\", \"strengths\": \"Paper describes well previous contributions with proposed method.\\nAuthors did thorough experimental comparison on multiple model configs and data sets.\", \"weaknesses\": \"Q1: Please explain the mechanisms or properties of the hierarchical approach that differentiate it from a single larger LoRA, particularly in terms of how it recovers high-frequency information lost during quantization and how it is numerically different with a single large LoRA? (sum outputs of several linear low rank decompositions, should be numerically the same with single LoRA?).\", \"q2\": \"Please explain how the adapters are specialized and designed to recover lost resolution across varying data frequencies.\", \"q\": \"Probably I missed something, would be happy to discuss and revisit paper rating.\\n\\n> \\\"Improved Performance: Demonstrating enhanced performance over QLoRA for certain models by effectively recovering high-frequency data lost during quantization.\\\"\", \"q3\": \"I could find any \\\"targeted loss function that prioritizes high-frequency detail learning at lower levels\\\" in the paper. Please clarify.\\n \\n \\nIn \\\"3.2 H-QLORA: HIERARCHICAL QUANTIZED LOW-RANK ADAPTATIO\\\":\\n> Training time:\\n> Y = X (W + La1Lb1) + X (W + La2Lb2) + X (W + La3Lb3) + . . .\", \"q4\": \"Equation for inference time matches numerically topology shown on Figure 1, but above equation for Training time does not match Figure 1 and does not match numerically equation for inference time. So, I guess above equation for training time is not correct and it should be Y = X (W) + X (La2Lb2) + X (La3Lb3) + ...\", \"q5\": \"There is no clear winner for H-QLoRA configuration across different data sets and different models.\\n\\n> \\\"Novel Hierarchical Quantization Method: Introducing H-QLoRA, which incorporates hierarchical learning via multi-adaptor training, improving upon QLoRA.\\\"\", \"q6\": \"The naming of H-QLoRA is confusing: please explain of why do you consider this approach \\\"hierarchical,\\\" especially given that the structure appears to be a flat summation of adapters rather than a traditional hierarchical structure.\", \"questions\": \"Q: Main concerns are:\", \"q1\": \"Why proposed method should be better than standard QLoRA? Authors propose to sum outputs of several linear low rank decompositions, but it is numerically the same with single LoRA. So, if I understand the paper correctly and H-QLoRA is numerically the same with QLoRA, then all accuracy differences can be attributed to randomization. Below I show that QLoRA with rank 4 is numerically the same with H-QLoRA with two adapters, where each adapter has rank 2. Please explain why H-QLoRA should have better accuracy in comparison to single QLoRA with larger rank.\\n\\n\\n\\\"\\\"\\nimport torch\\n\\ninp_features = 4\\n\\nout_features = 4\\n\\nbatch = 1\\n\\n// Rank of adapter \\n\\nR = 2 \\n\\nsize_a = (R, inp_features)\\n\\nsize_b = (out_features, R)\\n\\nsize_x = (batch, inp_features)\\n\\n// input feature x [batch, inp_features]\\n\\nx = torch.rand(size_x, requires_grad=True, dtype=torch.float32)\\n\\n// Weights of adapter 1\\n\\na1 = torch.rand(size_a, requires_grad=True, dtype=torch.float32)\\n\\nb1 = torch.rand(size_b, requires_grad=True, dtype=torch.float32)\\n\\n// Weights of adapter 2\\n\\na2 = torch.rand(size_a, requires_grad=True, dtype=torch.float32)\\n\\nb2 = torch.rand(size_b, requires_grad=True, dtype=torch.float32)\\n\\n// Output of adapter 1\\n\\nout1 = torch.matmul(torch.matmul(x, a1.t()), b1.t())\\n\\n// Output of adapter 2\\n\\nout2 = torch.matmul(torch.matmul(x, a2.t()), b2.t())\\n\\n// Final output of all adpaters (sum them all)\\n\\nout = out1 + out2\\n\\n// Concatenated weights of all apdaters\\n\\nA = torch.cat((a1, a2), dim=0)\\n\\nB = torch.cat((b1, b2), dim=1)\\n\\n// Final output of single adapter with concatenated weights (R = 4) is the same with sum of adapters\\n\\nOUT = torch.matmul(torch.matmul(x, A.t()), B.t())\\n\\ntorch.testing.assert_close(OUT, out1 + out2)\\n\\n\\\"\\\"\\\"\", \"in_the_introduction\": \"> \\\"By adopting a targeted loss function that prioritizes high-frequency detail learning at lower levels, we anticipate further performance improvements over QLoRA, while maintaining a consistent parameter count.\\\"\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The paper explores the use of a hierarchical adaptor structure on top of QLORA and evaluates performance through instruction tuning and MMLU evaluation.\", \"soundness\": \"1\", \"presentation\": \"2\", \"contribution\": \"1\", \"strengths\": \"The paper investigates the use of multiple LORA adaptors at each layer in comparison to QLORA and measures performance on the MMLU dataset using several Llama/Gemma models.\", \"weaknesses\": \"The paper's contributions are not substantial enough for a main conference paper.\\n\\n1. **Lack of Novelty:** The authors explore a straightforward idea of whether using multiple LORA adaptors at each layer is better compared to QLORA. There is no novelty in this approach.\\n\\n2. **Limited Experiments:** The experiments are minimal, utilizing only one evaluation dataset (MMLU) and two training sets (Alpaca, OASST1). Furthermore, the experimental results do not conclusively validate the authors' hypothesis. In the main Table 1, the results are mixed, with 3 positive outcomes and 2 negative ones.\\n\\n3. **Runtime Analysis:** The paper provides a comparison of training time differences between QLORA and H-QLORA. However, there is no analysis of inference time, which is a more practical concern.\\n\\n4. **Minor Issues:** The equations for training time and inference time, (3) and (4), are not equivalent.\", \"questions\": \"1. Could the authors provide an analysis of inference runtime?\\n2. Could the authors demonstrate conclusive improvements over QLORA?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"1\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"withdrawal_confirmation\": \"I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors.\"}", "{\"summary\": \"The authors introduce H-QLoRA, an extension of the original QLoRA approach that incorporates hierarchical adaptors. Unlike QLoRA, which uses a single adaptor, H-QLoRA employs multiple adaptors, which the authors claim can improve fine-tuning performance. To optimize configuration within a given memory budget for fine-tuning, the authors experiment with varying the number of adaptors. By testing on various LLaMA and Gemma models, they demonstrate that H-QLoRA can achieve superior fine-tuning accuracy on the OASST1 and Alpaca datasets.\", \"soundness\": \"1\", \"presentation\": \"1\", \"contribution\": \"1\", \"strengths\": \"The question of whether using a single adaptor or multiple adaptors (to be added and merged) for fine-tuning is an intriguing area of research. If multiple adaptors prove to be crucial for enhancing accuracy or performance post-fine-tuning, this concept could be expanded and explored further in future studies.\", \"weaknesses\": \"Overall, the quality of this work and its results are quite underwhelming, for several reasons outlined below:\\n\\n1. In both Table 1 and Table 2, the necessity of H-QLoRA is not clearly demonstrated. Do the authors genuinely believe that incorporating multiple adaptors will significantly enhance 5-shot MMLU scores?\\n\\n2. Relying solely on MMLU scores is insufficient to substantiate the claims made. A more comprehensive set of metrics is needed, and if possible, A/B testing should also be conducted to provide stronger evidence.\\n\\n3. Fragmenting a relatively large adaptor into several smaller adaptors introduces various computational overheads, such as adaptor control logic, memory management issues, and potential performance degradation in highly parallel computing environments. A detailed analysis of these factors is essential.\\n\\n4. What is the main takeaway from Table 1? It appears that as model sizes increase, QLoRA performs better, which contradicts the authors' claims. This inconsistency needs to be addressed.\", \"questions\": \"Please see 'weaknesses' above.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"1\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}" ] }

[ "li zhe", "Xiangfei Qiu", "Peng Chen", "Yihang Wang", "Hanyin Cheng", "Yang Shu", "Jilin Hu", "Chenjuan Guo", "Aoying Zhou", "Qingsong Wen", "Christian S. Jensen", "Bin Yang" ]

[ "Time Series Forecasting", "Foundation Model", "Benchmark" ]

{ "note_id": [ "zDMDnA2U1I", "mEPQ3fka7p", "bvLJHQfL5t", "Qta4rqGE3R", "LKVl5VEKR7" ], "note_type": [ "comment", "official_review", "official_review", "official_review", "official_review" ], "note_created": [ 1732188019959, 1730468206497, 1730519631534, 1730744196163, 1730965772479 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission3508/Authors" ], [ "ICLR.cc/2025/Conference/Submission3508/Reviewer_ZGbi" ], [ "ICLR.cc/2025/Conference/Submission3508/Reviewer_p1Rf" ], [ "ICLR.cc/2025/Conference/Submission3508/Reviewer_2Ngw" ], [ "ICLR.cc/2025/Conference/Submission3508/Reviewer_BqMP" ] ], "structured_content_str": [ "{\"withdrawal_confirmation\": \"I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors.\"}", "{\"summary\": \"This paper proposes to evaluate different families of time-series forecasting models in a unified framework.\\nThe benchmark consists of several well-known datasets which are grouped based on their origin and dataset characteristics such as seasonality.\\nThe authors proceed to evaluate a variety of models (foundation models as well as customly trained approaches) in different evaluation settings such as zero-shot.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"1\", \"strengths\": [\"The first benchmark to compare time-series foundation models.\", \"The benchmark includes a variety of datasets and methods but also importantly evaluation settings like zero-shot and few-shot which have become more relevant for foundation models.\", \"Interesting finding that LLM-based time-series foundation models work worse than time-series specific foundation models.\"], \"weaknesses\": [\"Since the datasets used in the benchmark are publicly available and commonly used in time-series forecasting, the foundation models considered in this benchmark may be contaimnated. Could you please provide a list of datasets each foundation model was trained and compare it to the datasets used in your benchmark?\", \"Several foundation models like Chronos or ForecastPFN were at least partially trained using synthetic data. In reference to weakness 1, I believe that a fair evaluation of foundation models should also include some synthetic data showing some realistic data characteristics. Could you please consider adding maybe the data from the ForecastPFN or Chronos data generator to your benchmark and report results on them?\", \"Could you explain your selection criteria for the foundation models you evaluated? Some foundation models like Chronos and ForecastPFN are missing and I think they would improve the submission as forecastpfn is only trained on synthetic data and Chronos is widely used at the moment.\"], \"minor\": [\"Please check the references and make sure to use citet and citep where appropriate. I noticed this especially in section 3.2.1.\", \"The paper is slightly above page limit with the reproducibility section going to page 11.\"], \"questions\": \"-\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper presents FoundTS, a benchmark for fair and comprehensive evaluation of Time Series Forecasting foundation models. FoundTS includes diverse models\\u2014LLM-based, time series pre-trained, and specific models\\u2014and supports multiple evaluation strategies (zero-shot, few-shot, and full-shot) in a standardized pipeline. Key results show time series pre-trained models excel in few-shot tasks, while LLM-based models struggle, indicating a need for optimization. Smaller models like ROSE and TTM offer a good performance-efficiency balance, and no single model dominates across scenarios, highlighting the need for versatile TSF models.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. This work incorporates a comprehensive range of models and settings, including LLM-based and foundational time series models, and evaluates them across zero-shot, few-shot, and full-shot scenarios.\\n\\n2. It includes a wide variety of data domains, enhancing the benchmark\\u2019s relevance and robustness across different applications.\\n\\n3. The insights from this study are clear and may possibly benefit for industry professionals.\", \"weaknesses\": \"1. The takeaway conclusions are not impressive, such as 'Foundational models are more powerful than specific models'; 'Foundational TS models are better than LLM-based models'; and 'Scaling law doesn't hold for TS models'. Most of them have been noted in prior studies [1,2,3]. Similarly, the suggestions for improving foundational TS models lack specificity and fail to offer new insights.\\n\\n[1] MOMENT: A Family of Open Time-series Foundation Models. ICML 2024\\n\\n[2] A Decoder-Only Foundation Model for Time-Series Forecasting. ICML 2024\\n\\n[3] Are Language Models Actually Useful for Time Series Forecasting? \\u200b\\u200bhttps://arxiv.org/pdf/2406.16964v1\\u200b\\n\\n2. Figures 3 and 5 are difficult to interpret; additional context and explanation are needed to clarify the insights these figures provide.\\n\\n3. Beyond simply ranking TS models, it would be beneficial to highlight actionable guidance or overarching principles in TS forecasting that could directly benefit industry professionals using FoundTS.\\n\\n4. Section 4.2.4\\u2019s conclusions (points 2 and 3) are unconvincing. As illustrated in Figure 1, LLM-based models take prompts during pretraining, but they cannot accurately interpret prompts if pre-trained weights are not effectively loaded.\\n\\n5. The open-sourced code provided is currently inaccessible.\", \"questions\": \"Please refer to the weakness.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper presents FoundTS, a benchmark for time series foundation models. The key motivation is that the previous work in this domain has applied different hyperparameter settings. So it is not very clear whether the evaluation was fair. In this paper, the authors build a standardized pipeline to evaluate time series forecasting tasks in few-shot and full-shot settings.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"1. This paper benchmarks time series foundation models.\\n2. The presentation is clear.\\n3. The benchmark additionally presents some interesting analytical experiments.\", \"weaknesses\": \"1. A foundation model is expected to be able to be applied to various downstream tasks. However, the paper only benchmarks time series forecasting, while ignoring other tasks such as anomaly detection, classification, clustering, etc.\\n2. The main results (Tables 4, 5, and 6) do not have much new compared with previous work. Similar results have already been reported in the previous papers.\\n3. The standard deviation is not reported in all the results.\\n4. Some related work in benchmarking or evaluating time series foundation models are not discussed, such as [1] [2] [3] [4]\\n\\n[1] GIFT-Eval: A Benchmark For General Time Series Forecasting Model Evaluation\\n[2] Evaluating Large Language Models on Time Series Feature Understanding: A Comprehensive Taxonomy and Benchmark\\n[3] Understanding Different Design Choices in Training Large Time Series Models\\n[4] TSGBench: Time Series Generation Benchmark\", \"questions\": \"See weakness.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The paper presents results of several time series models on a set of datasets in several settings - zero-shot, few-shot, and full-shot. Some analysis was performed on characteristics of the models such as \\\"channel independence\\\" and \\\"channel dependence\\\".\", \"soundness\": \"2\", \"presentation\": \"1\", \"contribution\": \"1\", \"strengths\": \"The paper presents a good effort on the important task of benchmarking and quantitatively assessing the recent advances in deep learning for time series forecasting.\", \"weaknesses\": \"Unfortunately, the paper does not quite cover sufficient scope to constitute a comprehensive benchmark, and also falls short of presenting some significant insight into the methods. The datasets are insufficiently diverse, for example, ETTh1/h2/m1/m2 are in fact from the same source and should be considered a single dataset instead. A single dataset for a domain is insufficient to present robust results, especially for a benchmark paper. There seems to be no consideration regarding the sampling frequency of the datasets, which is a critical characteristic of time series, and the analysis of the properties of characteristics of the datasets is not well presented - just a table in the appendix. The reader is not guided towards the significance of the differences in these characteristics.\\n\\nThere are also some concerns regarding the rigor of the experimental setup. For example, TimesFM has been pre-trained on the Electricity and Traffic datasets, and should not qualify for the zero-shot setting for these datasets. Furthermore, for a benchmark paper, there is a greater onus to ensure that each model is well tuned without bias - such details are missing from the exposition of the paper.\\n\\nI would also recommend the authors to revisit terms such as \\\"specific model\\\", and use terms which are more aligned with existing literature.\", \"questions\": \"-\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}" ] }

[ "Bo Yue", "Shufan Wang", "Ashish Gaurav", "Jian Li", "Pascal Poupart", "Guiliang Liu" ]

[ "Constraint Inference", "Training Efficiency", "Cross-environment Transferability" ]

{ "note_id": [ "zA4IIGRBWp", "ycIUoptMYx", "wzQGRpB7mG", "wCtn0LR9iT", "qFUBetv0sd", "q6eoEEIKMa", "oS0JNEVkEo", "oC1Zy2ZQp0", "my2ku5YUta", "logYiLgSMb", "hQyKHiQrFx", "cPJokiAIRA", "Wt7DeUKVEd", "WfBthpFqvl", "SzLKm5ZR5F", "PAaeDndkmz", "O6L0n0Yaco", "N0kQPW0ohx", "LOqPBjrZ5I", "L1vtAFUqNp", "JI7My9tfwp", "IPTpVeOGKk", "GfzlznQyiC", "ElYVzHf613", "BPP58lddIm", "7lDSw0nXZC", "4iTNHPT59e", "3ju2oR3h99", "3cqXEcSpoC", "1gMSYe8IGM" ], "note_type": [ "official_comment", "official_comment", "official_comment", "official_comment", "meta_review", "official_comment", "official_review", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "decision", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment" ], "note_created": [ 1732759484612, 1732844575813, 1732714305647, 1732238726913, 1735074929334, 1732237940499, 1730521608739, 1732237572127, 1731253922027, 1732589946316, 1732237180839, 1732305005409, 1730699229429, 1733127316330, 1732656932269, 1732844094314, 1732237220971, 1733127905463, 1732310993099, 1732237747704, 1732413666653, 1737523766727, 1730789583745, 1733214295561, 1732844877048, 1733212927113, 1732238065743, 1733126984790, 1732237976624, 1732844474361 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission6391/Reviewer_87Lh" ], [ "ICLR.cc/2025/Conference/Submission6391/Authors" ], [ "ICLR.cc/2025/Conference/Submission6391/Authors" ], [ "ICLR.cc/2025/Conference/Submission6391/Authors" ], [ "ICLR.cc/2025/Conference/Submission6391/Area_Chair_39Fo" ], [ "ICLR.cc/2025/Conference/Submission6391/Authors" ], [ "ICLR.cc/2025/Conference/Submission6391/Reviewer_87Lh" ], [ "ICLR.cc/2025/Conference/Submission6391/Authors" ], [ "ICLR.cc/2025/Conference/Submission6391/Reviewer_MoLg" ], [ "ICLR.cc/2025/Conference/Submission6391/Reviewer_87Lh" ], [ "ICLR.cc/2025/Conference/Submission6391/Authors" ], [ "ICLR.cc/2025/Conference/Submission6391/Reviewer_foPH" ], [ "ICLR.cc/2025/Conference/Submission6391/Reviewer_i1Qn" ], [ "ICLR.cc/2025/Conference/Submission6391/Reviewer_MoLg" ], [ "ICLR.cc/2025/Conference/Submission6391/Reviewer_i1Qn" ], [ "ICLR.cc/2025/Conference/Submission6391/Authors" ], [ "ICLR.cc/2025/Conference/Submission6391/Authors" ], [ "ICLR.cc/2025/Conference/Submission6391/Authors" ], [ "ICLR.cc/2025/Conference/Submission6391/Reviewer_foPH" ], [ "ICLR.cc/2025/Conference/Submission6391/Authors" ], [ "ICLR.cc/2025/Conference/Submission6391/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission6391/Reviewer_foPH" ], [ "ICLR.cc/2025/Conference/Submission6391/Authors" ], [ "ICLR.cc/2025/Conference/Submission6391/Authors" ], [ "ICLR.cc/2025/Conference/Submission6391/Reviewer_foPH" ], [ "ICLR.cc/2025/Conference/Submission6391/Authors" ], [ "ICLR.cc/2025/Conference/Submission6391/Authors" ], [ "ICLR.cc/2025/Conference/Submission6391/Authors" ], [ "ICLR.cc/2025/Conference/Submission6391/Authors" ] ], "structured_content_str": [ "{\"comment\": \"Thank you for your response. The clarification of the theoretical motivation and the related work in your response helps me better understand your work. I choose to maintain my score.\"}", "{\"title\": \"Author Response to Reviewer 87Lh\", \"comment\": \"Dear Reviewer 87Lh,\\n\\nWe are deeply grateful for the reviewer\\u2019s feedback and the significant time and effort invested in reviewing our manuscript. Your insightful comments have been invaluable in enhancing the clarity and quality of our work. \\n\\nThank you very much! Wishing you a joyful Thanksgiving!\"}", "{\"title\": \"Further Response on the 1st and 3rd Questions.\", \"comment\": \"Dear Reviewer 87Lh,\\n\\nThank you for your thoughtful feedback and for engaging in the discussion with us.\\n\\n> Comment: tightness of sample complexity upper bound for IRC and ICRL solvers\\n\\n**Response:** We agree with the reviewer that upper bounds are an important part of the analysis, and demonstrating the tightness of these bounds is critical for a more rigorous comparison. However, our work primarily focuses on the transferability of constraint knowledge, with an emphasis on discussing the safety and optimality of policies based on constraint knowledge in target environments. The main takeaway we want from previous related works of IRC and ICRL is that IRC offers a better upper bound in terms of sample complexity, which suggests it outperforms ICRL in this regard.\\n\\nIn our analysis, based on Lemma 4.1 (implicit definition of the feasible reward correction set) and Lemma 4.3 (implicit definition of the feasible cost set), we show that ICRL solvers need to estimate further the advantage function for expert alignment, a step that is not needed for IRC solvers. This extra estimation step in ICRL leads to a higher sample complexity compared to IRC under uniform sampling from a generative model. Thus, we argue that the sample complexity of ICRL is, in theory, larger than that of IRC.\\n\\nEmpirical experiments further support this theoretical finding. We observe that IRC consistently requires fewer samples than ICRL to achieve comparable performance levels in the source (learning) environments. This empirical result also matches the findings presented in [1].\\n\\nIn our work, we adopt the feasible set approach for comparison of sample complexity upper bounds. This approach aims to recover the entire correction terms (for IRC) and cost functions (for ICRL) compatible with the demonstrations, thereby addressing the identifiability problem in IRL (or IRC/ICRL) without being influenced by the choice of heuristics or additional restrictions, as discussed in [2].\\n\\nTo the best of our knowledge, the feasible set approach is a relatively new framework for analyzing sample complexity and this line of works only include [3], [4], and [5]. Specifically, [3] introduced this approach in the infinite horizon case for IRL, [4] derived a lower bound in the finite horizon setting with additional assumptions for IRL to assess the tightness of upper bounds, and [5] extended [3] to ICRL.\\n\\nWe agree that proving the tightness of the upper bounds through lower bound analysis or other methods would be a valuable direction for future research, and we plan to explore this in subsequent work.\\n\\nReferences\\n\\n[1] Hugessen, A., et al. Simplifying constraint inference with inverse reinforcement learning. NeurIPS, 2024.\\n\\n[2] Lazzati, Filippo, Mirco Mutti, and Alberto Maria Metelli. How does Inverse RL Scale to Large State Spaces? A Provably Efficient Approach. NeurIPS, 2024.\\n\\n[3] Alberto Maria Metelli, Giorgia Ramponi, Alessandro Concetti, and Marcello Restelli. Provably efficient learning of transferable rewards. ICML, 2021.\\n\\n[4] Alberto Maria Metelli, Filippo Lazzati, and Marcello Restelli. Towards theoretical understanding of inverse reinforcement learning. ICML, 2023.\\n\\n[5] Yue, Bo, Jian Li, and Guiliang Liu. Provably Efficient Exploration in Inverse Constrained Reinforcement Learning. arXiv preprint arXiv:2409.15963.\\n\\n---\\n\\n> Comment: the third question\\n\\n**Response:** Thank you for reminding us; you are definitely correct. Since the objective function inside remains the same and the variable $\\\\mathit{\\\\Delta r}$ falls into the domain $-\\\\lambda c$, it should be $\\\\max_{\\\\mathit{\\\\Delta r}}$ instead of $\\\\min_{\\\\mathit{\\\\Delta r}}$. Our previous response was inaccurate, and we sincerely apologize for the oversight. We have corrected this typo in the revised manuscript. This does not affect later analyses of IRC and ICRL solvers.\\n\\nThank you again for your valuable comments. We would be more than happy to continue the discussion if there are any further questions.\"}", "{\"title\": \"Summary of updates\", \"comment\": \"We sincerely thank the reviewers (Reviewer MoLg, Reviewer foPH, Reviewer i1Qn, and Reviewer 87Lh) for their insightful and valuable feedbacks, which have been instrumental in improving our work.\\n\\nWe have carefully gone through all comments and incorporated the key updates into the revised manuscript, with changes highlighted in orange for clarity. The updates are listed as follows.\\n\\n1. We **added experimental results to evaluate soft constraint scenarios** (Remark B.25, Reviewer MoLg).\\n\\n2. We **included comparisons in more complex environments**, e.g., Half-Cheetah, for IRC and ICRL solvers (Sec 6 and Appendix Sec C, Reviewer i1Qn, 87Lh).\\n\\n3. We **improved the clarity and readability of the paper** by clarifying mathematical notations, correcting typos, and adding hyperparameter tables (Sec 4 and Appendix Sec D, Reviewer foPH, i1Qn).\\n\\n4. We **provided additional explanations** in the main text when referring to the appendix to enhance clarity (all changes highlighted in orange, Reviewer foPH, 87Lh).\\n\\nWe hope that our revisions can address the concerns raised and look forward to receiving some feedback from the reviewers. We are more than willing to engage in further discussions to refine our work.\"}", "{\"metareview\": \"The paper studies the inverse reward correction (IRC), a recently popularized approach to tackle the problem of inferring safety constraints from expert demonstrations, where the reward that the expert is maximizing is known. The authors show that IRC has better sample complexity but worse generalization across environments with different dynamics or rewards than inverse constrained RL (ICRL), which involves in solving a three-level optimization. Then, they derive conditions in dynamics, rewards, and Lagrangian multipliers that guarantee epsilon-optimality of IRC across different environments.\", \"here_are_some_positive_and_negatives_from_the_reviews\": \"(+) This paper suggests that IRC could be a simpler alternative to ICRL, and thus, can potentially save the practitioners from the complexity of ICRL methods. \\n(+) The authors have done a good job in motivating the problem and questions posed in the paper. \\n(-) Moderate to low novelty in the theoretical results as they closely follow prior work on the sample complexity of IRL and transfer of the obtained rewards. \\n(-) Poor presentation that makes it a challenge to understand all the details. This can explain the low confidence of two reviewers and can be seen from the review of the high-confidence reviewer (Reviewer foPH). Of course, the authors made some improvements in this regard during the rebuttals in response to the reviewers' comments, especially those by Reviewer foPH.\", \"additional_comments_on_reviewer_discussion\": \"The authors addressed some of the reviewers' comments, especially those by Reviewer foPH and improved the clarity of their presentation.\"}", "{\"title\": \"Author response to Reviewer i1Qn - Part 1\", \"comment\": \"Dear Reviewer i1Qn,\\n\\nWe deeply value your detailed review of our manuscript and the constructive feedback you provided. To address your comments, we have revised the manuscript, with all modifications distinctly highlighted in orange for ease of review. We have thoroughly considered your suggestions and wish that the responses below will resolve your concerns.\\n\\n> Comment 1: Could benefit from comparison on more complex environments than grid-world\\n\\n**Response 1:** Thank you for raising this concern. To address the extension to high-dimensional continuous environments, we incorporate the Maximum Entropy framework of ICRL solvers [1] and IRC solvers [2]. Additionally, our original online approach should be adapted to an offline setting, where the agent learns to infer constraint knowledge that best aligns with expert demonstrations from an offline dataset, rather than relying on a predefined expert policy for querying. Our results demonstrate that while ICRL exhibits slower learning in the training environment, it achieves better transferability when environment rewards or transition dynamics differ in the target environment. This observation is consistent with our theoretical findings for discrete spaces. In the revised manuscript, we have detailed this addition in lines 483\\u2013486 (highlighted in orange). Further experimental results for comparison are provided in Appendices C and D.\\n\\nIn addition to conducting empirical studies, we have made considerable contributions to understanding the differences among various solvers used for modeling constraint knowledge in the field of constraint inference.\\n\\n---\\n\\n> Comment 2: It would be nice to have a comparison of the pseudocode of the IRC and ICRL algorithms used in the experiment.\\n\\n**Response 2:** The pseudocode is provided in Algorithm 1 in Appendix Section B.3. We omitted it from the main body of the original draft due to the page limit. A comparison of pseudo-code is now available in lines 761-764, highlighted in orange. We have explained $\\\\mathcal{I}^{\\\\mathit{\\\\Delta r}}_ {k+1}$ in lines 226-234 and $\\\\mathcal{I}^{c}_ {k+1}$ in lines 264-269 in the revised manuscript.\\n\\n---\\n\\n> Comment 3: Consider adding hyperparameters of the algorithms in the experiments.\\n\\n**Response 3:** Thanks for this comment. In the original draft, we illustrate the utilized hyperparameters in lines 1332-1346 in Appendix Section C. To enhance rigor and clarity, we have added a list of utilized hyperparameters in Appendix Table 3 and 4 in the revised version (table caption highlighted orange).\\n\\n---\\n\\n> Comment 4: What are the implications of this in few-shot learning after transferring an IRC/ICRL policy?\\n\\n**Response 4:** Instead of transferring policies, we focus on transferring feasible constraint knowledge, such as reward correction terms or cost functions, from the source to the target environment. The agent derives this knowledge entirely from the source and uses it to generate a policy in the target environment, aligning with a zero-shot learning paradigm. This approach emphasizes the generalizability of IRC and ICRL solvers, distinguishing them from imitation learning methods. Implications are that although IRC has lower training efficiency to ensure the optimality of expert agents in source environments, it has poorer performance at generalizability when transferring.\"}", "{\"summary\": \"This paper focuses on two approaches for safety-critical inverse RL: Inverse Constrained Reinforcement Learning (ICRL) and Inverse Reward Correction (IRC). It derives the upper bounds of these two methods and concludes the advantage of IRC in terms of sample efficiency. The paper also discusses potential constraint violations when learning a reward correction term or cost in source environments and transferring them to new environments. Theoretical analysis demonstrates that ICRL is more robust to these issues compared to IRC and examines the optimality of ICRL in target environments. Finally, empirical studies on gridworlds validate the theoretical findings for both methods.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": [\"The paper provides a comprehensive comparison of ICRL and IRC approaches from multiple perspectives, including sample complexity and transferability. This reveals a tradeoff between the methods, offering interesting insights into their applications.\", \"The theoretical formulations and derivations are detailed and well-defined.\", \"The empirical evaluations are convincing and consistent with theoretical results.\"], \"weaknesses\": \"See Questions.\", \"questions\": [\"The paper compares sample efficiency using theoretical upper bounds. I haven\\u2019t fully examined the details, but I am curious about the tightness of these estimations and whether they reliably represent real sample efficiency in practice.\", \"Although the experiments are conducted in gridworlds to validate the theoretical findings, I wonder if the results hold true for more complex, scalable tasks. Discussing potential challenges in such contexts could offer guidance and inspiration for the practical application of ICRL and IRC.\", \"I am a little confused about $\\\\min_{\\\\triangle r}$ in Eq. (2), which seems to minimize the expert's returns. Should it be $\\\\max$ instead?\", \"The term 'iteration' appears in several theorems. It would be helpful to briefly explain what ICRL and IRC do in each iteration, such as taking a fixed number of gradient steps based on Eq. (1) and (2) ?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Author response to Reviewer foPH - Part 1\", \"comment\": \"Dear Reviewer foPH,\\n\\nWe sincerely appreciate your thorough, valuable, and constructive review of our manuscript. We have carefully revised the manuscript, with all changes highlighted in orange for clarity. Your suggestions have been thoughtfully considered, and we hope the following responses adequately address your concerns.\\n\\n> Comment 1: For $Q^{r,\\\\pi}_{\\\\mathcal{M}\\\\cup r^\\\\prime}$ (and similarly V), it's never made clear how the value depends on $r$ in the superscript, which remains confusing throughout, especially as other values including the cost c get swapped into that place and this value is quite central in the paper and used frequently throughout the text.\\n\\n**Response 1:** We apologize for any potential ambiguity in the original manuscript. This notation is actually clarified in lines 218\\u2013222 under Lemma 4.1 in the original draft, where it is first introduced. The *superscript* $r$ (or $c$) in $Q^{r,\\\\pi^E}_ {\\\\mathcal{M} \\\\cup r_1}$ (or $Q^{c,\\\\pi^{E}}_{\\\\mathcal{M} \\\\cup c_1}$) specifies whether the function represents a reward $Q$-function or a cost $Q$-function, highlighting the distinct roles of the reward or cost value functions in constraint inference. The *subscript* $r_1$ (or $c_1$) identifies the specific rewards or costs being evaluated by the $Q$-function.\\n\\nFor greater clarity, we have now explicitly defined this notation in lines 150\\u2013156 of the revised manuscript.\\n\\n---\\n\\n> Comment 2: The paper initially introduces a soft-constraint setting (allowing for costs up to $\\\\epsilon\\\\geq 0$, with the hard-constraint case $\\\\epsilon=0$ as a special case) but then, in places, assumes the hard-constraint setting without alerting the reader to the fact adding further confusion.\\n\\n**Response 2:** We understand that the reviewer is likely referring to Section 5. In the original manuscript, we specified where a hard-constraint setting is assumed: in line 372 for Section 5.1 (safety) and line 384 for Section 5.2 (optimality).\\n\\nTo improve clarity and rigor, as recommended by the reviewer, we have explicitly identified the hard-constraint setting in line 346, as well as in Lemma 5.2 and Theorem 5.3, in the revised manuscript. \\n\\nAlso, please note that we have studied the soft constraint scenario in Appendix B.8.2, as mentioned in the paragraph 'Extension to Soft Constraint'.\\n\\n---\\n\\n> Comment 3: Several pieces of notation are not defined in the main text (e.g. $N^+_{k+1}$) on page 5.\\n\\n**Response 3:** Thanks for this comment. These definitions were initially provided in the appendix due to space limitations in the main text. To improve clarity, we have now included explanations for the significance $\\\\delta$ and the cumulative count of visitations to $N^+_{k+1}$ in lines 227 and 236 of the revised manuscript, respectively.\\n\\n---\\n\\n> Comment 4: l.219 mentions advantage, followed by a definition of what appears to be a state-value function instead with advantage never defined.\\n\\n**Response 4:** Thanks for your correction. For better clarity, we have now clearly defined the reward advantage function in line 154 of the revised manuscript.\\n\\n---\\n\\n> Comment 5: I'd advise using similar notation for different concepts: e.g. depending on sub/superscript, $\\\\mathcal{C}$ is sometimes a set of feasible cost functions, sometimes a (not directly related) scalar value in the sample complexity bound, or $\\\\Delta$ is once a set of measures, another time a reward correction term. I'd advise using at least a different font in the two cases.\\n\\n**Response 5:** We greatly appreciate this valuable feedback. In response, we have updated the scalar values in the sample complexity bound, by replacing $\\\\mathcal{C}^c_{k+1}$ with $\\\\mathcal{I}^c_{k+1}$ and replacing $\\\\mathcal{C}^ {\\\\mathit{\\\\Delta r}}_ {k+1}$ with $\\\\mathcal{I}^{\\\\mathit{\\\\Delta r}}_{k+1}$, respectively. Furthermore, we have modified the notation for the reward correction term, changing $\\\\Delta r$ to $\\\\mathit{\\\\Delta r}$ for clarity and consistency. We have highlighted this modification in orange in the revised manuscript.\\n\\n---\\n\\n> Comment 6: Several times, the main text refers to numbered Tables or Figures, which are in the appendix but this is not pointed out - by default, a reader will search in the main text\\n\\n**Response 6:** We apologize for the oversight of not directly referring to \\\"Appendix\\\" for Table 1 and Algorithm 1. This has now been corrected in the revised manuscript in line 140 and line 265. Regarding the reference to Figure 1 under Theorem 5.3, we believe this refers to Figure 1 in the main text. We have thoroughly reviewed the draft to eliminate this issue.\\n\\n---\\n\\n> Comment 7: There is also a fair amount of typos.\\n\\n**Response 7:** Thanks for your correction. We have conducted a thorough review of the draft and prepared a revised version with the corrected parts highlighted in orange for clarity.\"}", "{\"summary\": \"This work introduces the IRC solver to overcome the limitation of the IRL solver, which generally lacks a mechanism to leverage existing reward signals and may not be compatible with different rewards. The authors also give theoretical analysis and achieve a lower sample complexity. Besides, they also study and extend the transferability. The results of the experiments demonstrate its efficiency.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. This paper is well-written and easy to follow.\\n2. The related work is well organized.\\n3. The authors give a strong theoretical. analysis of the advantages and shortcomings between IRC and ICRL solvers, i.e., convergence, and sample complexity. Besides, the authors clearly presented their theory contributions.\", \"weaknesses\": \"1. What are the wall-clock running times of your method and the other baselines in the experiments?\\n2. In Theorem 3.2, Can we have such that directly maximizing the cumulative rewards and considering the constrained optimization objective?\\n3. In B.8.2 subsection, Can we provide experimental results to evaluate the soft constraint scenario?\", \"questions\": \"Please see the weaknesses.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you for your response. I would like to confirm several points: For the first question, it is clear that the IRC provides a better upper bound. I am not familiar with the sample complexity theory of IRC and ICRL but is it fair to use the upper bounds to compare sample complexity and claim that the IRC outperforms IRCL in terms of sample complexity, especially when the tightness of this bound is unclear?\\n\\nFor the third question, I would like to confirm whether the change of variable from $c$ to $\\\\mathit{\\\\Delta r}=-\\\\lambda c$ would switch the optimization operation from $\\\\max$ to $\\\\min$. Because the objective inside seems to remain the same in Eq. (1) and (2).\"}", "{\"title\": \"Author response to Reviewer MoLg - Part 1\", \"comment\": \"Dear Reviewer MoLg,\\n\\nWe sincerely thank you for your detailed and insightful review of our manuscript. In response, we have carefully revised the manuscript, highlighting all changes in orange for discrepancies. We have thoughtfully incorporated your suggestions and believe that the following responses address your concerns effectively.\\n\\n> Comment 1: What are the wall-clock running times of your method and the other baselines in the experiments?\\n\\n**Response 1:** The table below reports the average wall-clock running times (mean $\\\\pm$ standard deviation) in the format of minutes:seconds, based on 5 parallel experiments. The label 'source' indicates that the solver (ICRL or IRC) infers the reward correction term or cost function and applies it within the same environment as the source environment. In contrast, the label 'target' represents inference in the source environment following knowledge transfer to a distinct target environment. All experiments were conducted on a desktop computer equipped with an Intel(R) Core(TM) i5-14400F processor and an NVIDIA GeForce RTX 4060 GPU, consistent with the device specifications detailed in Appendix C.\\n\\n| Solver\\\\Gridworld | Gridworld-1 | Gridworld-2 | Gridworld-3 | Gridworld-4 |\\n|----------------------|-------------------|-------------------|-------------------|-------------------|\\n| ICRL-source | 11:24 \\u00b1 00:06 | 09:12 \\u00b1 00:08 | 09:54 \\u00b1 00:09 | 09:39 \\u00b1 00:07 |\\n| ICRL-target | 12:36 \\u00b1 00:10 | 08:54 \\u00b1 00:08 | 11:18 \\u00b1 00:11 | 12:42 \\u00b1 00:12 |\\n| IRC-source | 08:27 \\u00b1 00:05 | 08:06 \\u00b1 00:05 | 08:03 \\u00b1 00:04 | 08:03 \\u00b1 00:04 |\\n| IRC-target | 08:12 \\u00b1 00:06 | 07:57 \\u00b1 00:05 | 08:15 \\u00b1 00:07 | 08:09 \\u00b1 00:06 |\\n\\n---\\n\\n> Comment 2: In Theorem 3.2, Can we have such that directly maximizing the cumulative rewards and considering the constrained optimization objective?\\n\\n**Response 2:** Yes, there are algorithms that address the Constrained Reinforcement Learning (CRL) problem in a more direct manner. Below, we summarize these approaches:\\n\\n1. Manual Selection of Lagrange Multipliers: \\n Methods such as [1, 2, 3] manually select Lagrange multipliers to directly maximize the objective $r - \\\\lambda c$, where $r$ represents the reward and $c$ represents the cost.\\n\\n2. Projection-Based Constraint Enforcement: \\n The approach in [4] utilizes prior knowledge of system transitions to project the policy's chosen action onto a set that guarantees constraint satisfaction, ensuring compliance without compromising the task.\\n\\n3. Projection-Based Constrained Policy Optimization (PCPO): \\n PCPO [5] is an iterative algorithm designed to optimize policies under expected cumulative constraints. It operates in two stages:\\n - Stage 1: Maximizes the reward using TRPO [6], producing an intermediate policy that may not satisfy the constraints.\\n - Stage 2: Projects this intermediate policy onto the closest feasible policy, ensuring constraint satisfaction while improving the reward. \\n This method effectively balances reward optimization and constraint enforcement.\\n\\nWhile these methods offer innovative solutions, they come with limitations:\\n- The first two approaches require additional information, such as system manual selected Lagrange multipliers or transition models.\\n- PCPO has the drawbacks of being computationally expensive and has limited generality [7].\\n\\nIn contrast, the Lagrangian relaxation method, the most widely adopted approach for addressing cumulative constraints [7], demonstrates high performance. This method achieves high long-term rewards and maintains low cumulative costs [8, 9].\\n\\nTheoretical support for this approach is provided by Theorem 3.2 from Paternain et al. (2019), which states that if the reward and cost functions are bounded, the constrained optimization problem (PI) can be solved precisely in the dual domain. This implies that the optimal policy for the constrained objective in (PI) can be obtained precisely by solving its dual problem (DI), which involves iteratively optimizing both the Lagrange multipliers and the policy.\"}", "{\"comment\": \"Thank you for the responses! Here are a few more comments in reaction to those\", \"response_1\": \"Thank you for the clarification. However, since $r$ is also used as a variable name in the text (even in the same font!), the superscript suggests that the value of $Q^r$ depends on the value of the variable $r$ via the superscript. Two possible ways to fix this come to mind: replace $r$ with $\\\\text{rew}$ in the superscript or something like that to clearly distinguish it from a variable $r$ (at the very least, one should use a different font (e.g. $Q^{\\\\text{r}}$ but I think that could still remain confusing). Or just use a different letter for $Q^c$ altogether, such as $C$.\\n\\nHowever, this brings another question: if the superscript is used to distinguish this from $Q^c$, then is the definition meaningfully different (also note that $Q^c$ is never defined in the main text). I'd say it isn't, if you use the right notation. E.g. If you use e.g. $Q^{r,\\\\pi}_{\\\\mathcal{M}}$ to actually denote the expected return w.r.t. $r$ (or $c$ swapped into the same spot), then you can just use a single definition and get rid of the confusion.\", \"response_2\": \"Yes, I noticed that, but the unfortunate choice is that you first state one assumption (possibly soft constraints). Then you give a bunch of results (at odds with that assumption, which confuses the reader). And only then you reveal to them that you were in fact using another assumption. If you're changing an assumption, please state that before results building on that assumption.\", \"response_3_7\": \"Thank you for the responses and improving the manuscript.\"}", "{\"summary\": \"The paper explores inverse reward correction (IRC) that, instead of solving a three-level optimization in inverse constrained RL (ICRL), tries to combine the constraints as part of the reward function so as to solve only a reward optimization problem. The paper proves IRC has lower sample complexity but worse generalizability than ICRL solvers. The work investigates conditions in the transition laws, cost/reward functions, and Lagrangian multipliers that guarantee epsilon-optimality of IRC when transferring to a different environment.\", \"soundness\": \"3\", \"presentation\": \"4\", \"contribution\": \"3\", \"strengths\": [\"well-written and clear\", \"problem and questions are well-motivated\", \"interesting theoretical guarantees and analysis of IRC\"], \"weaknesses\": [\"could benefit from comparison on more complex environments than grid-world\", \"it would be nice to have a comparison of the pseudocode of the IRC and ICRL algorithms used in experiment\", \"consider adding hyperparameters of the algorithms in the experiments\"], \"questions\": [\"what are the implications of this in few-shot learning after transferring an IRC/ICRL policy?\", \"What would these bounds on the sample complexity/generalizability of IRC and ICRL mean in real world tasks where S and A are continuous spaces?\", \"Furthermore, do these bounds consider any measure (like complexity) of the dynamics/transition function? Why or why not?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Thanks for your response.\", \"comment\": \"Your response's additional experiments and theoretical analysis somewhat relieved my concerns. Synthesize comments from other reviewers, I choose to maintain my score.\"}", "{\"comment\": \"Thank you for your response. I maintain my score.\"}", "{\"title\": \"Follow-up on Additional Response\", \"comment\": \"Dear Reviewer foPH,\\n\\nWe hope this message finds you well and that you're enjoying a wonderful Thanksgiving season!\\n\\nWe sincerely appreciate the time and effort you have dedicated to reviewing our submission, especially during this busy period. We wanted to kindly follow up and invite you to review our additional response at your convenience. Your feedback is invaluable to us and will be instrumental in further refining our work.\\n\\nIf there are any points that require clarification or if you would like further details, please don't hesitate to reach out. We are more than happy to provide any additional information you may need.\\n\\nThank you again for your thoughtful insights and support. We truly appreciate your contributions to the review process.\\n\\nWishing you a joyful and fulfilling Thanksgiving!\\n\\nBest regards,\\n\\nThe Authors of Paper 6391\"}", "{\"title\": \"Author response to Reviewer MoLg - Part 2\", \"comment\": \"> Comment 3: In B.8.2 subsection, Can we provide experimental results to evaluate the soft constraint scenario?\\n\\n**Response 3:** Yes, we have added additional results and analyses for the soft constraint scenario, as detailed at the end of Appendix Section B.8.2 (highlighted in orange). These analyses evaluate the performance of the ICRL solver under soft constraints in two distinct aspects:\\n\\n1. Aspect One: different reward functions \\n\\n In this case, the reward functions differ between the source and target environments. Our results demonstrate that the ICRL solver consistently outperforms the IRC solver, effectively mitigating the impact of variations in reward functions across environments.\\n\\n2. Aspect Two: different transition dynamics\\n\\n Here, the transition dynamics differ between the source and target environments. The findings indicate that similar to the IRC solver, the ICRL solver can potentially violate constraints when transferring inferred cost functions in soft constraint scenarios due to inferred penalizations compensated by the altered transition dynamics.\\n\\n---\\n\\nReferences\\n\\n[1]Vivek S Borkar, \\u201cAn actor-critic algorithm for constrained markov decision processes,\\u201d Systems and control letters, vol. 54, no. 3, pp. 207\\u2013213, 2005.\\n\\n[2] Dotan Di Castro, Aviv Tamar, and Shie Mannor, \\u201cPolicy gradients with variance related risk\\ncriteria,\\u201d arXiv preprint arXiv:1206.6404, 2012.\\n\\n[3] Aviv Tamar and Shie Mannor, \\u201cVariance adjusted actor critic algorithms,\\u201d arXiv preprint\", \"arxiv\": \"1310.3697, 2013.\\n\\n[4] Gal Dalal, Krishnamurthy Dvijotham, Matej Vecerik, Todd Hester, Cosmin Paduraru, and Yuval Tassa, \\u201cSafe exploration in continuous action spaces,\\u201d arXiv preprint arXiv:1801.08757, 2018.\\n\\n[5] Tsung-Yen Yang, Justinian Rosca, Karthik Narasimhan, and Peter J Ramadge. Projection-based constrained policy optimization. arXiv preprint arXiv:2010.03152, 2020.\\n\\n[6] John Schulman, Sergey Levine, Pieter Abbeel, Michael Jordan, and Philipp Moritz. Trust region policy optimization. In International conference on machine learning, pages 1889\\u20131897, 2015.\\n\\n[7] Liu, Yongshuai, Avishai Halev, and Xin Liu. \\\"Policy learning with constraints in model-free reinforcement learning: A survey.\\\" The 30th international joint conference on artificial intelligence (ijcai). 2021.\\n\\n[8] Yongshuai Liu, Jiaxin Ding, and Xin Liu. Ipo: Interior-point policy optimization under constraints.\\nIn Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 4940\\u20134947, 2020.\\n\\n[9] Yinlam Chow, Mohammad Ghavamzadeh, Lucas Janson, and Marco Pavone. Risk-constrained reinforcement learning with percentile risk criteria. The Journal of Machine Learning Research, 18(1):6070\\u20136120, 2017.\"}", "{\"comment\": \"Dear Reviewer MoLg,\\n\\nWe appreciate the significant time and effort dedicated to this process. Thank you very much!\\n\\nBest regards,\\n\\nThe Authors of Paper 6391\"}", "{\"comment\": \"**Response 9**: Sorry, you may have misunderstood my point, so let me clarify. You have a whole set of feasible correction terms. Some of these may be \\\"unsafe\\\", though when talking about an unsafe correction term, we really mean an unsafe policy. So we have a set of feasible policies, some of which are unsafe. But if there are some that are robustly safe? Then, if we indeed care about prioritizing safety, we would choose a policy that is robustly safe, i.e. does not incur high cost under any of the recovered costs (/reward correction terms). If we choose such a policy, instead of an arbitrary one, you criticism of IRC as less safe than ICRL stops being sound, doesn't it?\"}", "{\"title\": \"Author response to Reviewer foPH - Part 2\", \"comment\": \"> Comment 8: What role does the superscript $r$ play in $Q^{r,\\\\pi}_{\\\\mathcal{M}\\\\cup r^\\\\prime}$?\\n\\n**Response 8:** Please refer to Response 1.\\n\\n---\\n\\n> Comment 9: In Definition 3.3 of an IRC solver, you introduce $\\\\mathcal{C}_{\\\\text{IRC}}$, \\\"the set of feasible reward correction terms derived by [the solver]\\\", which seems to imply that the IRC solver at least may recover a whole set of feasible reward correction terms. Then, having this set, one may try to produce a policy that is as robust as possible with respect to that set. Using your own example from Fig. 3, you write that the solver learns an error correction term $-1-\\\\beta,\\\\beta>0$. If understood as a set of feasible correction terms, a robust policy would avoid this state. Instead, you seem to force a choice of one particular reward correction. What do you think of using the full set of feasible error correction terms? Would this eliminate the advantage you point out? Alternatively, Bayesian IRL methods would give a posterior over the different feasible error correction terms, again allowing for producing robust policies.\\n\\n**Response 9:** Instead of transferring robust policies, we transfer reward correction terms or cost functions from source to target. This is a key distinction between inverse learning algorithms such as IRL, IRC, and ICRL, and imitation learning approaches like behavior cloning.\\n\\nIn the source environment, we infer a feasible set of reward corrections or costs that align with the given expert policy. Policies in the target environment are then derived based on these inferred terms. For example, in Fig. 3, we demonstrate a subset of the feasible set ($-1-\\\\beta, \\\\beta>0$) that leads to unsafe policies in the target environment, although it leads to safe policies in the source environment. Moving one step further, we propose Theorem 5.3 to precisely identify the conditions under which such \\\"unsafe\\\" feasible terms can arise.\\n\\n---\\n\\n> Comment 10: Could you please point out which results and proofs mostly just follow prior work on related concepts, and which parts (e.g. giving line ranges) are indeed novel and specific to this context?\\n\\n**Response 10:** Yes, we list them as follows. They have also been specified in the contribution part of the introduction. The part of ICRL solvers recaps previous work. We formalize and summarize the IRC solvers and follow the prior theoretical framework to analyze its sample complexity (lines 211-245). We make a comparison between the two solvers (IRC vs. ICRL) and discuss the source of additional sample complexity for ICRL (lines 277-286).\\nWe propose and investigate the safety issues in transferability (lines 300-377). Concerning the optimality issues, we extend the transferability definition from prior work in MDP settings to accommodate more complicated CMDP settings. We derive conditions that limit the similarity between source and target environments to ensure $\\\\varepsilon$-optimality for ICRL (lines 420-452). We empirically validate our results on training efficiency and cross-environment transferability in various environments (lines 453-519).\"}", "{\"title\": \"Further Author Response for Discussion\", \"comment\": \"Dear Reviewer foPH,\\n\\nThank you for engaging in the discussion and giving us the opportunity to offer further clarification.\\n\\n> Additional Comment on 1: Confusion on notation $Q^{r,\\\\pi}_ {\\\\mathcal{M}\\\\cup r^\\\\prime}$ and $Q^{c,\\\\pi}_ {\\\\mathcal{M}\\\\cup c}$.\\n\\n**Further response on on 1:** Thanks for this additional comment and valuable suggestion. We believe it requires the superscript to differentiate Q-functions of reward or cost under a CMDP $\\\\mathcal{M}\\\\cup c$.\\nIf we replace both $Q^{r,\\\\pi}_ {\\\\mathcal{M}\\\\cup c}$ and $Q^{c,\\\\pi}_ {\\\\mathcal{M}\\\\cup c}$ with $Q^{\\\\pi}_ {\\\\mathcal{M}\\\\cup c}$, we will not be able to know whether $Q^{\\\\pi}_ {\\\\mathcal{M}\\\\cup c}$ denotes the expected rewards or costs under $\\\\mathcal{M}\\\\cup c$. \\nBased on your advice, we now figure out a clearer notation to address this confusion. There are a total of five notations under IRC and ICRL solvers: 1) for IRC, $Q^{r, \\\\pi}_ {\\\\mathcal{M}\\\\cup (r+\\\\mathit{\\\\Delta r})} / Q^{\\\\mathit{\\\\Delta r}, \\\\pi}_ {\\\\mathcal{M}\\\\cup (r+\\\\mathit{\\\\Delta r})} / Q^{r+\\\\mathit{\\\\Delta r}, \\\\pi}_ {\\\\mathcal{M}\\\\cup (r+\\\\mathit{\\\\Delta r})}$, 2) for ICRL, $Q^{r, \\\\pi}_ {\\\\mathcal{M}\\\\cup c} / Q^{c, \\\\pi}_ {\\\\mathcal{M}\\\\cup c}$. The subscript specifies the environment $\\\\mathcal{M}$ with either an updated reward function $r+\\\\mathit{\\\\Delta r}$ or a cost function $c$. The superscript specifies the actual rewards or costs under evaluation. If the superscript and the subscript are the same (e.g., in $Q^{c, \\\\pi}_{\\\\mathcal{M}\\\\cup c}$), this means we are calculating the cumulative costs in the environment. We list them as follows.\\n\\n$Q^{r,\\\\pi}_ {\\\\mathcal{M}\\\\cup (r+\\\\mathit{\\\\Delta r})}(s,a) = \\\\mathbb{E}_ {\\\\pi,P_\\\\mathcal{T}}\\\\left[\\\\sum_{t=0}^{\\\\infty} \\\\gamma^t r(s_t, a_t)\\\\right],$\\n\\n$Q^{\\\\mathit{\\\\Delta r},\\\\pi}_ {\\\\mathcal{M}\\\\cup (r+\\\\mathit{\\\\Delta r})}(s,a) = \\\\mathbb{E}_ {\\\\pi,P_\\\\mathcal{T}}\\\\left[\\\\sum_{t=0}^{\\\\infty} \\\\gamma^t \\\\mathit{\\\\Delta r}(s_t, a_t)\\\\right],$\\n\\n$Q^{r+\\\\mathit{\\\\Delta r},\\\\pi}_ {\\\\mathcal{M}\\\\cup (r+\\\\mathit{\\\\Delta r})}(s,a) = \\\\mathbb{E}_ {\\\\pi,P_\\\\mathcal{T}}\\\\left[\\\\sum_{t=0}^{\\\\infty} \\\\gamma^t [r+\\\\mathit{\\\\Delta r}](s_t, a_t)\\\\right],$\\n\\n$Q^{r,\\\\pi}_ {\\\\mathcal{M}\\\\cup c}(s,a) = \\\\mathbb{E}_ {\\\\pi,P_\\\\mathcal{T}}\\\\left[\\\\sum_{t=0}^{\\\\infty} \\\\gamma^t r(s_t, a_t)\\\\right],$\\n\\n$Q^{c,\\\\pi}_ {\\\\mathcal{M}\\\\cup c}(s,a) = \\\\mathbb{E}_ {\\\\pi,P_\\\\mathcal{T}}\\\\left[\\\\sum_{t=0}^{\\\\infty} \\\\gamma^t c(s_t, a_t)\\\\right]$.\\n\\nNote that although $Q^{r,\\\\pi}_ {\\\\mathcal{M}\\\\cup (r+\\\\mathit{\\\\Delta r})}(s,a)$ and $Q^{r,\\\\pi}_ {\\\\mathcal{M}\\\\cup c}(s,a)$ are equal in value, they belong to different solvers and the subscript specifies which solver each reward Q-function belongs to.\\n\\n\\nWe have revised the manuscript accordingly to incorporate this modification in the main text and appendix.\\n\\n---\\n\\n> Additional Comment on 2: Mention the assumption to avoid confusion.\\n\\n**Further response on on 2:** We agree with the reviewer on this point. As mentioned in the previous rebuttal, in the revised manuscript, we have now explicitly added the assumption of hard-constraint settings in line 346, as well as in Lemma 5.2 and Theorem 5.3.\\n\\n---\\n\\n> Additional Comment on 9: Comparison of IRC and ICRL on safety issues.\\n\\n**Further response on 9:** \\nThank you for your comment. We would like to provide additional clarification. By the word 'safe', we mean whether the inferred reward correction terms induce safe policies in the *target* environment. Labeling a correction term with 'safe' or 'unsafe' requires prior knowledge of the transition and reward functions in the target environment. However, in practice, we do not assume that such prior knowledge is available, i.e., we only know that the correction term aligns with expert policy in the source environment. This unavailability of prior knowledge is reasonable since the agent could directly infer the correction terms in the target environment if the agent had access to this additional knowledge. This makes any terms inferred from the source environment unnecessary. Theorem 5.3 states the conditions under which unsafe correction terms exist (safe in the source but unsafe in the target) and Figure 1 is one instance of such unsafe terms.\\n\\nIn essence, each reward correction term is safe regarding a set of (reward, transition) pairs. Exploring solutions to identify the most robustly safe correction terms (encompassing the largest subset of a given set of reward-transition pairs) is an interesting direction for future research. However, this lies outside the scope of our current settings.\\n\\nIf you have any further questions, we would be more than happy to engage in discussions.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}", "{\"summary\": \"The paper addresses the task of inferring safety constraints from expert demonstrations where the expert maximizes a known reward subject to unknown constraints. Although inverse constrained reinforcement learning (ICRL) has developed specialized methods for this task, recent work has shown that casting the problem into the framework of inverse reinforcement learning (IRL) via \\\"inverse reward correction\\\" (IRC) could offer advantages. This paper provides a deeper analysis of these claims, showing that:\\n1) Inverse reward correction can achieve better sample efficiency than ICRL for constraint inference;\\n2) The implicit constraints inferred through IRC, however, may transfer poorly across environments with different dynamics or reward structures compared to constraints explicitly modeled by ICRL; and\\n3) ICRL\\u2019s explicit constraint modeling offers conditions under which its inferred constraints allow for approximately optimal and safe policies when transferred to new environments, even under varied dynamics.\", \"soundness\": \"2\", \"presentation\": \"1\", \"contribution\": \"3\", \"strengths\": [\"the task of inferring unknown safety constraints is important\", \"the recent challenge that the field of ICLR with its growing literature may be redundant deserves further scrutiny, and the results in this paper add important nuance to this challenge, showing that such redundancy is not clear cut\", \"the paper shows potentially useful results relating inverse reward correction and ICRL that deserve a place in the literature\"], \"weaknesses\": [\"unfortunately, the presentation is bad, making deciphering the paper a painful challenge even to a motivated reader closely familiar with the work this paper builds on:\", \"for me, this point alone currently warrants rejection, but I may be swayed if the authors manage to upload a substantial revision\", \"here is a non-exhaustive list of issues:\", \"for $Q^{r,\\\\pi}_{\\\\mathcal{M}\\\\cup r'}$ (and similarly V), it's never made clear how the value depends on $r$ in the superscript, which remains confusing throughout, especially as other values including the cost c get swapped into that place and this value is quite central in the paper and used frequently throughout the text\", \"the paper initially introduces a soft-constraint setting (allowing for costs up to $\\\\epsilon\\\\geq 0$, with the hard-constraint case $\\\\epsilon=0$ as a special case) but then, in places, assumes the hard-constraint setting without alerting the reader to the fact adding further confusion\", \"several pieces of notation are not defined in the main text (e.g. $N^+_{k+1}$) on page 5\", \"l.219 mentions advantage, followed by a definition of what appears to be a state-value function instead with advantage never defined\", \"regarding minor issues:\", \"I'd advise using similar notation for different concepts: e.g. depending on sub/superscript, $\\\\mathcal{C}$ is sometimes a set of feasible cost functions, sometimes a (not directly related) scalar value in the sample complexity bound, or $\\\\Delta$ is once a set of measures, another time a reward correction term. I'd advise using at least a different font in the two cases.\", \"several times, the main text refers to numbered Tables or Figures, which are in the appendix but this is not pointed out - by default, a reader will search in the main text\", \"there is also a fair amount of typos\", \"the theoretical results very closely follow similar prior results on the sample complexity of IRL and the transferrability of the recovered rewards (Metelli 2021, Schlaginhaufen 2024). Per se, I don't see that as a huge problem - clearly porting these results to the context of constraint inference still has value. However, if the main contribution of the paper is translating the results into another context, presentation would seem to be one of the main possible contributions, so doing a poor job at that takes away most of the value\"], \"questions\": [\"what role does the superscript $r$ play in $Q^{r,\\\\pi}_{\\\\mathcal{M}\\\\cup r'}$ ?\", \"in Definition 3.3 of an IRC solver, you introduce $\\\\mathcal{C}_{\\\\text{IRL}}$, \\\"the set of feasible reward correction terms derived by [the solver]\\\", which seems to imply that the IRC solver at least *may* recover a whole set of feasible reward correction terms. Then, having this set, one may try to produce a policy that is as robust as possible with respect to that set. Using your own example from Fig. 3, you write that the solver learns a error correction term $-1-\\\\beta,\\\\beta>0$, If understood as a set of feasible correction terms, a robust policy would avoid this state. Instead, you seem to force a choice of one particular reward correction. What do you think of using the full set of feasible error correction terms? Would this eliminate the advantage you point out? Alternatively, Bayesian IRL methods would give a posterior over the different feasible error correction terms, again allowing for producing robust policies.\", \"could you please point out which results and proofs mostly just follow prior work on related concepts, and which parts (e.g. giving line ranges) are indeed novel and specific to this context?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Dear Reviewer foPH,\\n\\nWe would like to express our sincere gratitude for your continued engagement, detailed feedback, and insightful discussions throughout this review process. We greatly appreciate the time, effort, and expertise you have dedicated to evaluating our paper.\\nWe will continue refining the paper based on your valuable comments.\\n\\nWe also want to thank you for considering to increase the score. However, it appears that the revised score was not updated in the initial review.\\n\\nThank you once again for your thoughtful contributions.\\n\\nBest regards,\\n\\nThe Authors of Paper 6391\"}", "{\"title\": \"Author Response to Reviewer MoLg\", \"comment\": \"Dear Reviewer MoLg,\\n\\nWe deeply appreciate the invaluable feedback and the thoughtful effort you invested in reviewing our paper! These insightful comments have been invaluable in enhancing both the clarity and the quality of our work. \\n\\nWe hope the above response could resolve your concerns. If you have more questions, please feel free to discuss them with us. \\n\\nThank you very much! Wishing you a joyful Thanksgiving!\"}", "{\"comment\": \"Dear authors,\\nthank you for the substantial effort you put into clarifying your work and updating the manuscript. As a result, I think it's fair to increase my score, and I don't have any additional clarifying questions at this point. That said, I think the manuscript still remains slightly below the threshold I'd expect from papers at ICLR mostly due to the significance of the contributions and the overall clarity of presentation. In preparing possible future versions or other future work, I would recommend working with an editor or a colleague otherwise not involved in the project to polish the clarity of the paper prior to initial submission - I found my review work was initially made much harder than it could have been and I would have been able to provide better feedback on the substance on the paper had I started with an easier-to-read manuscript. That said, I'm a fan of this research direction and wish you all best in pursuing it further, if you choose so!\"}", "{\"title\": \"Author response to Reviewer 87Lh\", \"comment\": \"Dear Reviewer 87Lh,\\n\\nThank you for your thoughtful and comprehensive review of our manuscript. We have meticulously updated the manuscript, ensuring all revisions are clearly highlighted in orange for your convenience. We have given careful consideration to your suggestions and hope the following responses satisfactorily address your concerns.\\n\\n>Comment 1: The paper compares sample efficiency using theoretical upper bounds. I haven\\u2019t fully examined the details, but I am curious about the tightness of these estimations and whether they reliably represent real sample efficiency in practice.\\n\\n**Response 1:** Thanks for the suggestion. The upper bound is proposed to guide the exploration, and by exploring to minimize the upper bound of error, the solvers are theoretically guaranteed to have a better estimate of reward correction terms or costs. Besides, we consider an infinite horizon scenario where a lower bound is hard to derive.\\n\\n---\\n\\n>Comment 2: Although the experiments are conducted in gridworlds to validate the theoretical findings, I wonder if the results hold true for more complex, scalable tasks. Discussing potential challenges in such contexts could offer guidance and inspiration for the practical application of ICRL and IRC.\\n\\n**Response 2:** Thanks for this comment. To address the extension to high-dimensional continuous environments, we incorporate the Maximum Entropy framework of ICRL solvers [1] and IRC solvers [2]. Additionally, our original online approach can be adapted to an offline setting, where the agent learns to infer constraint knowledge that best aligns with expert demonstrations from an offline dataset, rather than relying on a predefined expert policy for querying. Our results demonstrate that while ICRL exhibits slower learning in the training environment, it achieves superior transferability when environment rewards or transition dynamics differ in the target environment. This observation is consistent with our theoretical findings for discrete spaces. In the revised manuscript, we have detailed this addition in lines 483\\u2013486 (highlighted in orange). Further experimental results for comparison are provided in Appendices C and D.\\n\\nThere are certain challenges for more complex environments. First, the offline dataset may contain noise or sub-optimal demonstrations, robustly learning constraint knowledge or identifying these annoying demonstrations holds promise. Second, because of the multi-level optimization layers, both solvers may take a long time to converge. \\n\\nWe acknowledge that exploring real-world applications could provide deeper insights into the practical challenges and opportunities of transferring constraint information, ultimately guiding the development of more robust and scalable algorithms.\\n\\n---\\n\\n> Comment 3: I am a little confused about $\\\\min_{\\\\mathit{\\\\Delta r}}$\\n in Eq. (2), which seems to minimize the expert's returns. Should it be $\\\\max$\\n instead?\\n\\n**Response 3:** Thanks for your comment. This is derived from Eq. (1) where $\\\\mathit{\\\\Delta r}=-\\\\lambda c, \\\\lambda>0$. Maximizing the objective with $c$ in Eq. (1) is equivalent to minimizing the objective with $\\\\mathit{\\\\Delta r}$, since the objective is linear with regard to $\\\\mathit{\\\\Delta r}$ or $c$.\\n\\n---\\n\\n> Comment 4: The term 'iteration' appears in several theorems. It would be helpful to briefly explain what ICRL and IRC do in each iteration, such as taking a fixed number of gradient steps based on Eq. (1) and (2)?\\n\\n**Response 4:** Thanks for your comment. In each iteration, we first uniformly collect samples from the state-action space, and then update the estimation of the transition model and expert policy based on these samples. As the number of samples increases, the estimation error of transition and expert policy reduces. This leads to better estimation of either feasible reward correction terms or feasible cost functions. What the two solvers do in each iteration is detailed in Appendix Algorithm 1.\\n\\n---\\n\\nReferences\\n\\n[1] Shehryar Malik, Usman Anwar, Alireza Aghasi, and Ali Ahmed. Inverse constrained reinforcement learning. In International Conference on Machine Learning (ICML), pp. 7390\\u20137399, 2021.\\n\\n[2] Hugessen, Adriana, Harley Wiltzer, and Glen Berseth. \\\"Simplifying constraint inference with inverse reinforcement learning.\\\" The Thirty-eighth Annual Conference on Neural Information Processing Systems. 2024.\"}", "{\"title\": \"A kind reminder regarding our further response\", \"comment\": \"Dear Reviewer foPH,\\n\\nWe hope this message finds you well. We greatly appreciate the time and effort you have dedicated to reviewing our paper, especially given the many responsibilities you manage.\\n\\nAs the discussion period is nearing its conclusion, we would like to kindly follow up regarding your feedback on our response. Your input is invaluable to us, and we are eager to address any further questions or concerns you may have.\\n\\nThank you once again for your thoughtful contribution to this process. We deeply appreciate your time and expertise.\\n\\nBest regards,\\n\\nThe Authors of Paper 6391\"}", "{\"title\": \"Author response to Reviewer i1Qn - Part 2\", \"comment\": \">Comment 5: What would these bounds on the sample complexity/generalizability of IRC and ICRL mean in real world tasks where S and A are continuous spaces?\\n\\n**Response 5:** Thanks for this comment. Sample complexity analysis has primarily focused on discrete state-action spaces [3].\\nExtending such analyses to continuous spaces remains a significant challenge in the field. Existing algorithms for learning feasible sets [4, 5, 6] face difficulties when scaling to problems with large or continuous state spaces. This is largely due to their sample complexity being directly tied to the size of the state space, which presents a substantial limitation since real-world problems often involve large or continuous spaces.\\nContinuous domains often require function approximation techniques, additional assumptions (e.g. for smoothness or linear structures), and more sophisticated exploration strategies. Generalizability is another concern due to the infinite nature of these domains since it relies heavily on the ability to approximate value functions, policies, and constraints. We leave developing a scalable approach for sample complexity analysis as our future work.\\n\\n---\\n\\n> Comment 6: Furthermore, do these bounds consider any measure (like complexity) of the dynamics/transition function? Why or why not?\\n\\n**Response 6:** Yes, they do. The complexity of the transition function is determined by the cardinality of the state-action space since the transition is a mapping from state-action to state: $\\\\mathcal{S}\\\\times \\\\mathcal{A}\\\\rightarrow \\\\mathcal{S}$. The size of the transition matrix is $O(\\\\mathcal{S}^2\\\\mathcal{A})$. This bound relies on the size of state space $\\\\mathcal{S}$ and the size of action space $\\\\mathcal{A}$.\\n\\n---\\n\\nReferences\\n\\n[1] Shehryar Malik, Usman Anwar, Alireza Aghasi, and Ali Ahmed. Inverse constrained reinforcement\\nlearning. In International Conference on Machine Learning (ICML), pp. 7390\\u20137399, 2021.\\n\\n[2] Hugessen, Adriana, Harley Wiltzer, and Glen Berseth. \\\"Simplifying constraint inference with inverse reinforcement learning.\\\" The Thirty-eighth Annual Conference on Neural Information Processing Systems. 2024.\\n\\n[3] Agarwal, Alekh, et al. \\\"Reinforcement learning: Theory and algorithms.\\\" CS Dept., UW Seattle, Seattle, WA, USA, Tech. Rep 32 (2019): 96.\\n\\n[4] Alberto Maria Metelli, Filippo Lazzati, and Marcello Restelli. Towards theoretical understanding of inverse reinforcement learning. ICML, 2023.\\n\\n[5] Lei Zhao, Mengdi Wang, and Yu Bai. Is inverse reinforcement learning harder than standard reinforcement learning? ICML, 2024.\\n\\n[6] Filippo Lazzati, Mirco Mutti, and Alberto Maria Metelli. Offline inverse rl: New solution concepts and provably efficient algorithms. ICML, 2024.\"}", "{\"title\": \"Author response to Reviewer i1Qn\", \"comment\": \"Dear Reviewer i1Qn,\\n\\nWe are delighted by your positive assessment of our work and are most grateful for your thoughtful and thorough review of our manuscript. Your insightful comments have been invaluable in improving the clarity and precision of our work. We appreciate the significant time and effort you have dedicated to this process. Thank you very much! Wishing you a happy Thanksgiving!\"}" ] }

[ "Chongkai Gao", "Haozhuo Zhang", "Zhixuan Xu", "Cai Zhehao", "Lin Shao" ]

[ "World Model", "Long-Horizon Planning", "Robot Manipulation", "Flow Generation" ]

{ "note_id": [ "zeo8puYFm3", "zJA8DLTBEt", "wVRJxZsBBp", "vmijIVPo4j", "stRtpOpWDj", "qKDTR47Ilf", "pD805I2Bik", "lUrMafEOyf", "keMq96oC0p", "jgHm1EREpW", "iwrzLC0CtA", "h3JV4WHJFw", "gQQyvVsDvD", "gEMDQWB4Wj", "c0EkKggfGX", "YMvEp7ANJw", "MhLx42ZUqn", "LMBIZ0wV6q", "H1XcKkrfh9", "DJIKNKWRSx", "D2MwrxhoqI" ], "note_type": [ "official_comment", "official_review", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "decision", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "meta_review" ], "note_created": [ 1732638768116, 1730557809366, 1732691375396, 1730952971381, 1732711408648, 1732669700999, 1732638643630, 1732710787482, 1737523497895, 1732638539610, 1730706529586, 1732639324959, 1732691407710, 1732691251918, 1732639305022, 1732638923237, 1732639178428, 1732707455287, 1732638020491, 1730272925150, 1734519355205 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission2338/Authors" ], [ "ICLR.cc/2025/Conference/Submission2338/Reviewer_R3ys" ], [ "ICLR.cc/2025/Conference/Submission2338/Authors" ], [ "ICLR.cc/2025/Conference/Submission2338/Reviewer_H4xH" ], [ "ICLR.cc/2025/Conference/Submission2338/Reviewer_R3ys" ], [ "ICLR.cc/2025/Conference/Submission2338/Reviewer_6B9z" ], [ "ICLR.cc/2025/Conference/Submission2338/Authors" ], [ "ICLR.cc/2025/Conference/Submission2338/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission2338/Authors" ], [ "ICLR.cc/2025/Conference/Submission2338/Reviewer_9xux" ], [ "ICLR.cc/2025/Conference/Submission2338/Authors" ], [ "ICLR.cc/2025/Conference/Submission2338/Authors" ], [ "ICLR.cc/2025/Conference/Submission2338/Authors" ], [ "ICLR.cc/2025/Conference/Submission2338/Authors" ], [ "ICLR.cc/2025/Conference/Submission2338/Authors" ], [ "ICLR.cc/2025/Conference/Submission2338/Authors" ], [ "ICLR.cc/2025/Conference/Submission2338/Reviewer_R3ys" ], [ "ICLR.cc/2025/Conference/Submission2338/Authors" ], [ "ICLR.cc/2025/Conference/Submission2338/Reviewer_6B9z" ], [ "ICLR.cc/2025/Conference/Submission2338/Area_Chair_7PZV" ] ], "structured_content_str": [ "{\"title\": \"Response to Reviewer 9xux\", \"comment\": \"We thank reviewer 9xux again for reviewing our paper and providing helpful feedback! Here we\\naddress your concerns one by one.\\n\\n> The paper could have compared to some stronger baselines ... UniSim [1] ... DIAMOND [2]. Both of these paper use just a diffusion model without the need to first extract action flows, which seems to contradict to the key proposal of this paper, which is to make the model flow-centric. Some discussion on this would be appreciated.\\n\\nThanks for your question. The baseline used in our paper is LVDM [3], UniPI[4], and IRASim [5], which cover UniSim and DIAMOND in network architectures. \\n\\nFor UniSim, it uses a standard DiT [6] architecture, which is the same as IRASim, and we have shown the advantage of our newly designed dynamics module in Table 3. The originality of UniSim is that they use an internet-scale video dataset to train, which is beyond the scope of our method.\\n\\nFor DIAMOND, it uses a U-Net structure, which is also used in UniPI and LVDM, and we have already compared them in Table 1, Table 2, and Table 3. The novelty of DIAMOND is that they use EDM [7] to train the diffusion model for a faster denoising process, and train a reinforcement learning agent in the dreamed world, which are not the main points of this paper.\\n\\nThus, we think it is not necessary to add these two baselines. We have cited them in our paper.\\n\\n> For the flow generation model, why using a C-VAE instead of a diffusion model?\\n\\nThanks for your question. We use CVAE because it has a shorter inference time than diffusion\\nmodels. Here we add an ablation experiment in Appendix B.3 to show if diffusion models can\\nachieve better results for flow generation. The architecture is a DiT [6]. From the results in Table 7,\\nwe can see that there is not too much difference between CVAE and diffusion models in LIBERO-\\nLONG and Bridge-V2 data, which shows that for such short-horizon flow generation tasks, CVAE\\nis enough to represent them.\\n\\nIt is worth noting that there are some specially designed flow generation architectures with diffusion\\nmodels, and here we think this is out of the scope of this paper and we leave this question for future\\nworks. We believe in larger datasets with diverse flows, diffusion models can be better than CVAE.\\nFor example, in [8], they first use a pretrained Stable-Diffusion model to extract the embedding of\\nthe flow, then use AnimateDiff [9] as the backbone for flow diffusion.\\n\\n> please specify the quality of the video demonstrations.\\n\\nThanks for your question. The demonstrations used to train the FLIP model are all expert videos. This is because the value module is aligned with the language instruction, which is the goal of the whole video. If there are some random policies during the video, the value module will not be able to identify them and assign some values according to their corresponding task progress. We believe that with an internet-level training dataset, this issue could be alleviated. \\n\\n\\n[1] Yang, Mengjiao, et al. \\\"Learning interactive real-world simulators.\\\" arXiv preprint arXiv:2310.06114 (2023).\\n\\n[2] Alonso, Eloi, et al. \\\"Diffusion for World Modeling: Visual Details Matter in Atari.\\\" arXiv preprint arXiv:2405.12399 (2024).\\n\\n[3] He, Yingqing, et al. \\\"Latent video diffusion models for high-fidelity long video generation.\\\" arXiv preprint arXiv:2211.13221 (2022).\\n\\n[4] Du, Yilun, et al. \\\"Learning universal policies via text-guided video generation.\\\" Advances in Neural Information Processing Systems 36 (2024).\\n\\n[5] Zhu, Fangqi, et al. \\\"IRASim: Learning Interactive Real-Robot Action Simulators.\\\" arXiv preprint arXiv:2406.14540 (2024).\\n\\n[6] Peebles, William, and Saining Xie. \\\"Scalable diffusion models with transformers.\\\" Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023.\\n\\n[7] Karras, Tero, et al. \\\"Elucidating the design space of diffusion-based generative models.\\\" Advances in neural information processing systems 35 (2022): 26565-26577.\\n\\n[8] Xu, Mengda, et al. \\\"Flow as the cross-domain manipulation interface.\\\" arXiv preprint arXiv:2407.15208 (2024).\\n\\n[9] Guo, Yuwei, et al. \\\"Animatediff: Animate your personalized text-to-image diffusion models without specific tuning.\\\" arXiv preprint arXiv:2307.04725 (2023).\"}", "{\"summary\": \"This paper proposes FLIP, a model-based planning algorithm on visual space.\\nThe algorithm consists of three main modules 1) a flow generation model as an action proposal module, 2) a flow-conditioned video generation model as a dynamic module, and 3) a value module.\\nThe flow generation model generates flows based on the observation and a language instruction.\\nThe video generation model generates videos with a Diffusion Transformer based the predicted flows.\\nThe value module, built on a vision-language representation model, is used to assign values for each frame in a video to enable model-based planning in the image space.\\nModel-based planning algorithm leverages the three modules to search for a sequence of flow actions and video plans that maximizes the discounted return.\\nExperiments were performed to evaluate the capability of the proposed method on generating video plans, generating long-horizon videos, and accomplishing manipulation tasks.\\nThe proposed method outperforms comparing baseline methods and showcases interactive properties, zero-shot transfer capability, and scalability.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"The paper is well-written and clear.\\nLeveraging CVAE to address the multi-modality of flows is well-motivated.\\nAdditionally, the paper proposes to use a mixed-conditioning mechanism for multi-modal conditional video generation.\\nFor the value module, the paper modifies the original LIV method and uses video clips instead of video frames as a unit to account for the noisy value prediction.\\nExperiments show that the proposed method surpasses comparing baseline methods in generating video plans, long-horizon videos, and performing manipulation tasks.\\nThe paper also showcases interesting interactive properties, zero-shot transfer ability, and scalability of the proposed method.\", \"weaknesses\": \"1. There are no real-robot experiments to validate the effectiveness of the proposed method in policy learning in the real world. Incorporating such experiments and comparing with recent policy learning methods (e.g. ATM, OpenVLA [1] or Octo [2]) would provide a more comprehensive understanding of FLIP's performance in real-world policy learning.\\n\\n2. The paper only evaluates the policy learning capability on a suite of LIBERO, i.e. LIBERO-Long. Conducting an evaluation to assess the generalization capability of the proposed method, e.g. on the LIBERO-Object suite, would be beneficial. Also, including a comparison with recent policy learning methods, like OpenVLA [1] or Octo [2], would further strengthen the paper.\\n\\n2. In Sec. 5.2, the paper compares with IRASim on a text-to-video task. However, IRASim generates video based on a trajectory instead of a text. For the text-to-video task, it would be better to compare with a text-to-video method (e.g. [3]).\\n\\n[1] Kim, Moo Jin, et al. \\\"OpenVLA: An Open-Source Vision-Language-Action Model.\\\" arXiv preprint arXiv:2406.09246 (2024).\\n\\n[2] Team, Octo Model, et al. \\\"Octo: An open-source generalist robot policy.\\\" arXiv preprint arXiv:2405.12213 (2024).\\n\\n[3] Ma, Xin, et al. \\\"Latte: Latent diffusion transformer for video generation.\\\" arXiv preprint arXiv:2401.03048 (2024).\", \"questions\": \"1. In Sec. 5.1, the paper compares with FLIP-NC, an ablation of the proposed method FLIP which has no value module as guidance. Is it possible to provide more details on how FLIP-NC performs beam search without a value guidance?\\n\\n2. Is it possible to provide a more detailed description on the typical failure modes of FLIP in policy learning (Sec. 5.3) ?\\n\\n3. In the Dynamics Module Experiments in Sec. 5.4, the paper compares with LVDM and IRASim on short-horizon video generation. The proposed method is provided with ground-truth flows to generate videos. What information is provided for the two comparing baseline methods?\\n\\n4. Are the flow generation model and the video generation model trained individually or jointly?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"We thank reviewer H4xH again for reviewing our paper and providing helpful feedback! We hope the additional experiments and clarifications have addressed the reviewer\\u2019s concerns. If the reviewer has any further concerns or questions, please let us know before the rebuttal period ends, and we will be happy to address them. We deeply appreciate your thorough review and constructive comments, which have been invaluable in refining the paper's presentation.\\n\\nWe have carefully revised the manuscript to address the concerns that led to your lowered score, and we sincerely hope the changes will allow for a more favorable consideration. Thanks!\"}", "{\"summary\": \"In this work, the authors consider the task of high-dimensional generative planning for manipulation tasks. Specifically, they seek to design a prediction task which is relevant for manipulation planning and can also operate in a task-agnostic way on widely-available data. The model-based planning primitives are: an action generation module (generating 2D flow), a state transition model (generating next frame of video given flow + current observations), and a value function for determining how close a state is to the goal (based on language-conditioned visual encodings). This allows them to search for a flow-based plan, which at execution time can be executed by a learned plan-conditioned low-level policy. They evaluate the ability of their model to make coherent / correct plans on several benchmark tasks, and evaluate its utility for policy execution on the LIBERO-LONG task.\", \"soundness\": \"4\", \"presentation\": \"4\", \"contribution\": \"3\", \"strengths\": \"# Originality\\n\\nThe high-level approach is a reasonable combination of ideas in planning, visual representations, and generative modeling. On the flow generation component, It\\u2019s difficult to establish the level of novelty as there are a number of concurrent works proposing generative flow models for manipulation (e.g. GeneralFlow, Track2Act, etc.), but I think the combination of ideas here - for planning - is original/novel.\\n\\n# Quality\\n\\nThe quality of this paper is high. The experiments are thorough and comprehensive (with two caveats, discussed later).\\n\\nThe design of each component is sound, and there are lots of details / modifications that the authors conducted (and motivated!) which I found interesting and commendable. For instance, the ideas about video chunking (instead of per-frame) for value prediction, and the details about conditioning a model on flow, action space ablation, etc. are all solid contributions for other practitioners.\\n\\n# Clarity\\n\\nThe paper is quite clear - each step of the algorithm is well explained.\\n\\n# Significance\\n\\nThis paper is of moderate significance. I think the flow-based (particle-based) generative modeling approach as an action space has a lot of potential to be powerful, and clearly shows improvements for reconstruction / video generation.\", \"weaknesses\": \"The primary weakness of this paper is that, while the representation + models they built are quite interesting and useful for modeling the actual video domain they are imitating, the actual downstream utility of their method is not sufficiently characterized. Specifically, the method simply does not substantially outperform ATM, which has no planning at all and uses a similar generative representation, on the actual downstream manipulation task. Moreover, the inclusion of flow only seems to hurt the downstream policy in comparison to video conditioning. I\\u2019m not saying this method can\\u2019t show meaningful improvements over other approaches, but either the LIBERO setting chosen, or the particular low-level policy chosen, yield results that do not support the claim that this generative flow modeling + planning method provides downstream utility. Especially given the overhead. Of particular concern is the Ours-FV results, which are not well-explained.\\n\\nAnother weakness is that the experiments seem to be restricted to single domains, even though the method seems to have been designed to be task-agnostic - I would have liked to see the authors leverage this property more, e.g. by training predictive models on the full LIBERO benchmark and then finetuning on a subset of tasks for policy learning, or similar (even internet-scale pretraining\\u2026 although I realize this is out of scope for this contribution / infeasible if there are resource constraints).\\n\\nAnother existential issue is that this paper is framed as a \\u201cflow generation for manipulation\\u201d paper, but the bulk of the experimentation+analysis is geared towards video prediction which has little to do with manipulation. I don\\u2019t think it stands on its own as simply a video prediction paper (at least on these tasks alone) - and I would have liked to see a larger emphasis on the actual analysis for manipulation tasks (e.g. geometric precision is a visual prediction problem, or feasibility, or other downstream metrics).\", \"questions\": \"Why doesn\\u2019t this method offer significant downstream benefits compared to ATM, which has no planning, on the selected benchmark?\\n\\nHow much does the action space affect downstream performance? If you were to retrain ATM with this adjusted action space, would the ranking change significantly? Do the marginal benefits of your model come from this component (unrelated to the planning contribution)?\\n\\n\\n---\\n# Post-Rebuttal Update\\n\\nThe authors have addressed many of my concerns adequately, and greatly strengthened their paper with the modifications to their downstream policy architecture (as well as additional details in the appendix). Bumped score to Accept.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to Authors\", \"comment\": \"Thanks for the response. The provided additional information addresses my concerns. I update my score to 6.\"}", "{\"comment\": \"The authors response clarify my concerns and I update my score to 6.\"}", "{\"title\": \"Response to Reviewer H4xH (2/2)\", \"comment\": \"> I would have liked to see the authors train the model on the full LIBERO benchmark and then finetuning on a subset of tasks for policy learning (or even internet-scale pertaining)\\n\\nWe have trained the agent view world model on the LIBERO-90 (although with a smaller resolution\\nof 64\\u00d764) and shown the *zero-shot* generative planning results on LIBERO-LONG in Figure 9,\\nas well as on the website. We are sorry that we do not have enough resources to retrain it with a\\nresolution of 128\\u00d7128 or train for the eye-in-hand view on LIBERO-90 during this rebuttal, but we\\nwill do this after this paper gets accepted.\\n\\nTo show the fine-tuning world model for policy learning, we finetune this pretrained agent view\\nmodel with 50 actionless demonstrations for each task of LIBERO-LONG with a resolution of\\n64\\u00d764, and use 10 demonstrations with action labels for policy learning. The architecture is the\\nsame in the first question. Results are in Figure 12 in Appendix B.2. We can see that Ours-90\\nperforms similarly to Ours-F and Ours-FV, showing that pretraining in other tasks may not bring\\nsignificant improvement for low-level policy learning. This comforts with the lifelong learning\\nresults in the original LIBERO paper, where they also show that pretraining cannot help (sometimes\\neven hurt) the policy training results.\\n\\n> I would have liked to see a larger emphasis on the actual analysis for manipulation tasks.\\n\\nWe use the success rates of the low-level policy as the quantitative metric for our model. For qualitative analysis, the advantage of FLIP over other generative planning methods on manipulation tasks is two-fold, as stated in the Introduction Section: 1) be able to represent various kinds of movements across diverse objects, robots, and tasks in the whole scene (e.g., dexterous manipulation and deformable objects are very hard to describe with languages); 2) be easy to obtain or label a large amount of training data for scaling up.\\n\\nPlease also check our real-world experiments and results in Appendix B.4.\\n\\n> How much does the action space affect downstream performance?\\n\\nIn our original experiments, we used behavior cloning with delta translation, rotation, and aperture of the end-effector as the action space. In our new experiments, we use action chunking with the same delta action as our action space, and the results show that with the receding horizon policy, the results are better.\"}", "{\"comment\": \"Thanks very much for your response! We are happy to see that our rebuttal addresses most of your concerns. Here we answer your remaining questions as follows:\\n\\n> The OpenVLA result in Figure 12 of the Appendix is very low compared to the reported result in the original OpenVLA paper (53.7%). Based on the provided demonstration, OpenVLA increased the resolution and filtered the data. However, none of these operations increase the number of trajectories in training. Therefore, I believe the OpenVLA result in Figure 12 is not convincing.\\n\\nBesides the training data resolution difference and data filtering difference, the most important difference between our testing experiments and their original experiments is that they restrict the testing environments to have the same initialization as the training environments (as stated in the last paragraph of their Appendix E.1 (https://arxiv.org/pdf/2406.09246)). However, in our testing, we use randomly initialized configurations for all tasks, which means the objects' positions are randomly initialized according to the .bddl files in the LIBERO benchmark. This will make the success rate drop. \\n\\nFYI, we finetune the pretrained OpenVLA checkpoint for 150,000 steps before testing (about 100 epochs).\", \"we_here_add_two_verifications_to_make_our_results_convincing\": \"1. We use the checkpoint downloaded from OpenVLA official github repo and test it on their modified LIBERO-LONG environments (the same initial configurations as in the demonstrations). The average success rate is 50% (which is similar to their reported results), with detailed statistics as follows:\\n\\n| Task Name | Success Rate |\\n|----------|----------|\\n| put both the alphabet soup and the tomato sauce in the basket | 38% |\\n| put both the cream cheese box and the butter in the basket | 72% |\\n| turn on the stove and put the moka pot on it | 62% |\\n| put the black bowl in the bottom drawer of the cabinet and close it | 28%|\\n| put the white mug on the left plate and put the yellow and white mug on the right plate| 54% |\\n| pick up the book and place it in the back compartment of the caddy | 74% |\\n| put the white mug on the plate and put the chocolate pudding to the right of the plate | 48% |\\n| put both the alphabet soup and the cream cheese box in the basket | 56% |\\n| put both moka pots on the stove | 20% |\\n| put the yellow and white mug in the microwave and close it | 48% |\\n| Average| 50% |\\n\\n2. Using this checkpoint, we test on randomly initialized configures of LIBERO-LONG (the same testing setting used for all other comparison methods in our paper). The average success rate is 0.4%, with detailed statistics as follows:\\n\\n| Task Name | Success Rate |\\n|----------|----------|\\n| put both the alphabet soup and the tomato sauce in the basket | 0% |\\n| put both the cream cheese box and the butter in the basket | 0% |\\n| turn on the stove and put the moka pot on it | 0% |\\n| put the black bowl in the bottom drawer of the cabinet and close it | 0%|\\n| put the white mug on the left plate and put the yellow and white mug on the right plate| 0% |\\n| pick up the book and place it in the back compartment of the caddy | 4% |\\n| put the white mug on the plate and put the chocolate pudding to the right of the plate | 0% |\\n| put both the alphabet soup and the cream cheese box in the basket | 0% |\\n| put both moka pots on the stove | 0% |\\n| put the yellow and white mug in the microwave and close it | 0% |\\n| Average| 0.4% |\\n\\nWe hope these explanations can solve your concern.\\n\\n> For the results in Table 3, the conditions for video generation are different across different models, which makes the comparison unfair. LVDM is provided with no additional information; IRASim is provided with trajectories segmented with SAM2; FLIP is provided with ground-truth flow.\\n\\nThis experiment is designed to show that using flows can help improve the video generation quality, thus only our method can use flow as the condition.\\n\\nIt is worth noting that this conclusion (using flows can help improve the video generation quality) is non-trivial. Actually, with additional conditions, it is quite tricky to design proper ways to correctly use them to get positive effects, otherwise, the performance may even drop. For example, in IRASim [1], they show that given the end-effector trajectory as an additional condition, simply adding it to the video diffusion process with AdaLN will lead to a performance drop on some datasets (See Table 1 and Table 2 in their paper). Thus they design a special IRASim-Frame-Ada mechanism to consistently improve the generation quality.\\n\\nWe hope these explanations can solve your concerns.\\n\\n---\\n[1] Zhu, Fangqi, et al. \\\"IRASim: Learning Interactive Real-Robot Action Simulators.\\\" arXiv preprint arXiv:2406.14540 (2024).\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}", "{\"title\": \"Response to Reviewer H4xH (1/2)\", \"comment\": \"We thank reviewer H4xH again for reviewing our paper and providing helpful feedback! Here we address your concerns one by one.\\n\\n> Why doesn\\u2019t this method offer significant downstream benefits compared to ATM, which has no planning, on the selected benchmark?\\n\\nThanks for your question. Firstly we want to emphasize the major contribution of this paper is\\nthe high-level generative model-based planning framework, and thanks for your recognition of this\\npoint.\\n\\nSecondly, we think the original low-level policy for FLIP suffers from two problems: 1) it is a simple\\nbehavior cloning algorithm, which may not be able to make good use of the temporal information\\nof the future plans from our high-level model, because it only regresses one step action; 2) it only\\nuse one layer of MLP to extract the visual embedding from image patches, which may be future\\nimproved with pretraiend vision encoders. To better show the advantage of our FLIP model for\\nlow-level policy learning, we here train a diffusion policy [1] for the low-level policy with the action\\nchunking mechanism [2], which can predict the future action sequence rather than regress single-\\nstep actions. Meanwhile, we also use pretrained ResNet to extract the full image embedding for\\nconditioning. With these modifications, we compare our policies to the diffusion policy version of\\nATM (ATM-DP) and OpenVLA (asked by Reviewer R3ys).\\n\\nFrom Figure 12 in Appendix B.2, we can see that with the new model architecture, our policy can\\nachieve better results than previous results, and the success rates are better than ATM-DP. With this\\nnew architecture, Ours-FV is no longer the worst one among the three architectures, which shows\\nthat with an effective vision backbone, the model can have a better multi-modal capacity for different\\ncondition information. We can also see that using dense image flows can lead to a smaller variance\\nin LIBERO-LONG than using videos.\\n\\nAlthough the improvement of success rates of our policy compared to ATM-DP is not large (about\\n6%), we think this is not because our method is not good enough, but because this result has almost\\nreached the upper limit of LIBERO-LONG under our setting (10 demonstrations with actions for\\neach task). The SOTA results of this LIBERO-LONG task suite is [3], which achieves 53.7 \\u00b1 1.3%\\nand the mean success rate is only about 4% higher than ours. We think there are two main reasons\", \"for_limiting_the_further_improvement_of_the_success_rates_on_libero_long\": \"1) we are using a\\nresolution of 128\\u00d7128, which may not be large enough to represent the details in the scene. In\\ncomparison, in [3], they use a resolution of 256\\u00d7256. 2) We are using the official demonstrations\\nprovided by the LIBERO paper, which may not be as good as the re-collected demonstrations in [3].\\n\\nFinally, we show the results of our method on real robot experiments compared to Diffusion Policy\\nand ATM to better demonstrate the superiority of our policy over baselines. Here we only test Ours-\\nF because video generation is time-consuming during online replanning. From Appendix B.4, we\\ncan see that our policy is way better than the baselines, showing that in very difficult long-horizon tasks, our model can perform better with the help of model-based planning.\\n\\n[1] Chi, Cheng, et al. \\u201dDiffusion policy: Visuomotor policy learning via action diffusion.\\u201d The\\nInternational Journal of Robotics Research (2023): 02783649241273668.\\n\\n[2] Zhao, Tony Z., et al. \\u201dLearning fine-grained bimanual manipulation with low-cost hardware.\\u201d\", \"arxiv_preprint_arxiv\": \"2304.13705 (2023).\\n\\n[3] Kim, Moo Jin, et al. \\u201dOpenVLA: An Open-Source Vision-Language-Action Model.\\u201d arXiv\", \"preprint_arxiv\": \"2406.09246 (2024).\\n\\n> of particular concern is the Ours-FV results, which are not well-explained.\\n\\nHere, we explain the (original) Ours-FV results in detail. In the original low-level policy learning experiments, Ours-FV is worse than Ours-F and Ours-V. We think this is because the visual features inherently contain richer and more diverse information. However, our previous architecture did not incorporate a dedicated feature extraction process for the visual modality. Instead, both visual and flow information were passed through a single, simplistic MLP layer independently. This lack of specialized processing for visual features resulted in challenges when integrating them effectively with flow features. This is also pointed out by OpenVLA and Pi0.\\n\\nTo address this problem, in our new low-level experiments, we employ a pre-trained ResNet to\\nextract the visual features and then employ an additional transformer dedicated specifically to the feature\\nfusion from visual conditions, flow conditions, and language conditions, thereby decoupling the\\nthe learning process of the policy transformer. From Figure 12 in Appendix B.2, we can see that the\\nnew results show that Ours-FV is better than the previous result, which shows the effectiveness of\\nthe new multi-modal architecture.\"}", "{\"summary\": \"This paper proposes a new model-based planning framework called FLIP. It consists of 3 components: 1) a flow generation network that generates action flows given the current image observation and a language instruction 2) a video generation (dynamics model) that conditions on the flow and generate future video frames 3) a value function that evaluates the task progress (reward) given the image observation and language description. The key is the introduction of using flow as an action representation and condition the video generation model on the flow. The whole system can be used for model-based planning to generate video plans given a task, and the flow itself can also be used for guiding low-level policy execution. Various experiments are performed in both simulation environments and real-world videos.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": [\"This paper is overall clearly written.\", \"The experiments cover a range of test settings, and the ablation studies help understand each component of the method.\", \"Overall the experiment results are good, which demonstrates the effectiveness of the proposed method.\"], \"weaknesses\": [\"I feel the paper could have compared to some stronger baselines, for some of the experiments. E.g., for experiments in section 5.1 and section 5.2, a stronger baseline than UniPi could be UniSim [1]. The reviewer understands that the code may not be open-sourced, in this case, at least some discussion to the paper should be included. There is another very recent work DIAMOND [2] that can do very long-horizon and detailed video prediction into the future conditioned on actions. Both of these paper use just a diffusion model without the need to first extract action flows, which seems to contradict to the key proposal of this paper, which is to make the model flow-centric. Some discussion on this would be appreciated.\", \"For the flow generation model -- why using a C-VAE instead of a diffusion model? Some discussion on this design choice would be appreciated.\", \"For all experiments, please specify the quality of the video demonstrations, e.g., are they collected using a random policy, or are they expert videos? This would be important to understand, e.g., in section 5.1., if the system can learn from sub-optimal data for planning or does it need optimal demonstrations.\", \"[1] LEARNING INTERACTIVE REAL-WORLD SIMULATORS, Yang et al, ICLR 2024\", \"[2] DIAMOND\\ud83d\\udc8eDiffusion for World Modeling: Visual Details Matter in Atari, Alonso et al, NeurIPS 2024\"], \"questions\": \"Please see the weakness section\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"We thank reviewer 6B9z again for reviewing our paper and providing helpful feedback! We hope the additional experiments and clarifications have addressed the reviewer\\u2019s concerns. If the reviewer has any further concerns or questions, please let us know before the rebuttal period ends, and we will be happy to address them.\"}", "{\"comment\": \"We thank reviewer 9xux again for reviewing our paper and providing helpful feedback! We hope the additional experiments and clarifications have addressed the reviewer\\u2019s concerns. If the reviewer has any further concerns or questions, please let us know before the rebuttal period ends, and we will be happy to address them. We deeply appreciate your thorough review and constructive comments, which have been invaluable in refining the paper's presentation.\\n\\nWe have carefully revised the manuscript to address the concerns that led to your lowered score, and we sincerely hope the changes will allow for a more favorable consideration. Thanks!\"}", "{\"comment\": \"Thanks for your reply! We deeply appreciate your thorough review and constructive comments, which have been invaluable in refining the paper's presentation.\\n\\nWe have carefully revised the manuscript to address the concerns that led to your lowered score, and we sincerely hope the changes will allow for a more favorable consideration. Thanks!\"}", "{\"comment\": \"We thank reviewer R3ys again for reviewing our paper and providing helpful feedback! We hope the additional experiments and clarifications have addressed the reviewer\\u2019s concerns. If the reviewer has any further concerns or questions, please let us know before the rebuttal period ends, and we will be happy to address them.\\n\\nWe have carefully revised the manuscript to address the concerns that led to your lowered score, and we sincerely hope the changes will allow for a more favorable consideration. Thanks!\"}", "{\"title\": \"Response to Reviewer R3ys\", \"comment\": \"We thank reviewer R3ys again for reviewing our paper and providing helpful feedback! Here we\\naddress your concerns one by one.\\n\\n> There are no real-robot experiments to validate the effectiveness of the proposed method in policy learning in the real world. \\n\\nWe have added real-world experiments in Appendix B.4. Please check this section and our website for more real-world results.\\n\\n> Including a comparison with recent policy learning methods, like OpenVLA [1] or Octo [2], would further strengthen the paper.\\n\\nThanks for your suggestion! We add an experiment of OpenVLA on LIBERO-LONG to compare\\nits results with FLIP. We show the results of zero-shot and fine-tuned with 50 demonstrations for\\neach task of OpenVLA in Appendix B.2. We can see that OpenVLA cannot handle the long-horizon\\ntasks of LIBERO-LONG either with zero-shot or fine-tuned models, showing there is still a long\\nway to go for general-purpose vision-language-action models.\\n\\nIt is worth noting that, in the original OpenVLA paper, they also fine-tuned the pretrained model\\non LIBERO-LONG tasks and archived a 53.7 \\u00b1 1.3% success rate. We think the success of their\\nresults comes from two aspects, which cannot be true in our setting: 1) we are using a resolution\\nof 128\\u00d7128, which may not be large enough to represent the details in the scene. In comparison,\\nOpenVLA uses a resolution of 256\\u00d7256. 2) We are using the official demonstrations provided by\\nthe LIBERO paper, which may not be as good as the re-collected demonstrations in their demonstrations.\\n\\n> The paper compares with IRASim on a text-to-video task. However, IRASim generates video based on a trajectory instead of a text. For the text-to-video task, it would be better to compare with a text-to-video method (e.g. [3]).\\n\\nThanks for your question. The LVDM [4] baseline we used is a text-to-video method. We hope this can answer your question.\\n\\n> Is it possible to provide more details on how FLIP-NC performs beam search without a value guidance?\\n\\nThanks for your question. FLIP-NC performs a beam search in an autoregressive manner, which means it initializes the same number of beams as FLIP and generates the long-horizon flows and videos iteratively within each beam, without multiple action generation for each beam.\\n\\n> Is it possible to provide a more detailed description on the typical failure modes of FLIP in policy learning (Sec. 5.3) ?\\n\\nThanks for your question. We provide some failure videos of the trained policy in our attached videos and on our website. The typical failure mode is that the robot has a small action error when it is going to grasp something, which leads to an unsuccessful grasp and the robot will do this action repeatedly. Interestingly, our flow generation model can still generate reasonable future flows in these out-of-distribution areas to guide the robot to the correct regions. However, it sometimes cannot accomplish this in given episode timestep limits.\\n\\n> What information is provided for the two comparing baseline methods (LVDM and IRASim)?\\n\\nThanks for your question. For LVDM, there is no additional information provided. For IRASim, it is provided with the end-effector trajectory extracted with SAM2.\\n\\n> Are the flow generation model and the video generation model trained individually or jointly?\\n\\nThey are trained individually.\"}", "{\"title\": \"Response to Reviewer 6B9z\", \"comment\": \"We thank reviewer 6B9z again for reviewing our paper and providing helpful feedback! Here we\\naddress your concerns one by one.\\n\\n> The framework may struggle with ambiguities in language instructions or unexpected changes in the environment.\\n\\nWe point out that ambiguous language instruction and unexpected environmental changes fall outside the scope of this study. This work is based on the static MDP assumption, where we assume a language instruction can clearly specify the goal, and the MDP is not changed during the model-based planning and policy execution.\\n\\n> Depending on the computational demands of the multi-modal modules, real-time performance in dynamic environments could be a concern. Evaluating the speed and efficiency of planning under real-time constraints would be crucial.\\n\\nThanks for your question. Yes, real-time (which means about 30 HZ) inference speed is a question of our model. This is mainly constrained by: 1) the denoising process of the dynamics module; 2) the model-based planning process with three networks. However, with the development of faster denoising techniques such as [1], we believe we can achieve near real-time planning performance in the future.\\n\\n[1] Lu, Cheng, and Yang Song. \\\"Simplifying, Stabilizing and Scaling Continuous-Time Consistency Models.\\\" arXiv preprint arXiv:2410.11081 (2024).\\n\\n> How does FLIP compare to other state-of-the-art planning frameworks in terms of efficiency and effectiveness?\\n\\nWe compare FLIP to two generative planning methods (UniPi [1] and IRASim [2]) in Table 1 and Table 2. We hope these results can address your concerns.\\n\\n> What are the limitations of using image flow as an action representation? In which scenarios/tasks, this would not be ideal?\\n\\nThanks for your good question! There are two main limitations of using image flows as action representation: 1) 3D motions that have very small movements in the 2D space. For example, an arrow flying in the direction perpendicular to the screen. 2) objects with a very small size in the image. For example, if some object is put very far away from the camera, there could be no query points sampled on that object because it only occupies several pixels. Thus its movement will not be predicted.\\n\\n> How does FLIP handle ambiguities in language instructions?\\n\\nCurrently, we use the same language instruction during testing. \\n\\nThe language embedding is extracted from Meta Llama 3.1, thus we leave the ambiguities in language instructions to LLMs. We show zero-shot results in our website of the pretrained FLIP model on LIBERO-90 to LIBERO-LONG tasks, where the language instruction and the whole scene are new to the model. Since all of the three modules of FLIP is scalable, we hope in the future, with internet-level data, it can handle the ambiguities in language instructions.\\n\\n\\n> Are there any results on the real robot that show the applicability on world models even for quasi-static tasks?\\n\\nThanks for your question. We add real robot experiments in Appendix B.4.\\n\\n> How will the model's performance be affected in the presence of noise or visual obstructions?\\n\\nThanks for your good question! We add visual obstructions experiments in the LIBERO-LONG experiments. We manually add a picture of an apple into the scene during the planning, and from the results, we can see that our flow generation model can resist visual distraction, and the video generation model can generate dynamics for a few planning steps (x16 for execution steps), while it will generation distorted images after that. This shows that flows are a good choice to represent high-level actions that can be robust to visual distractions. Since our model is only trained on specific sense data, we believe with internet-level data, the video generation model will also perform better.\\n\\n> How does the performance of FLIP depend upon or scale with the size of the available data, numerical analysis for the same would be helpful to understand the efficiency of the proposed approach.\\n\\nThanks for your question. We add a data scalability experiment in Appendix B.5. From Table 9, we\\ncan see that with more data for each task, the planning success rates become better.\"}", "{\"title\": \"Response to Authors\", \"comment\": \"Thank you for your detailed response. Some of my concerns have been addressed. The updated real-robot experiments are able to showcase the performance of the proposed method in real-world policy learning. Based on the response, I have two follow-up questions.\\n\\n1. The OpenVLA result in Figure 12 of the Appendix is very low compared to the reported result in the original OpenVLA paper (53.7%). Based on the provided demonstration, OpenVLA increased the resolution and filtered the data. However, none of these operations increase the number of trajectories in training. Therefore, I believe the OpenVLA result in Figure 12 is not convincing.\\n\\n2. For the results in Table 3, the conditions for video generation are different across different models, which makes the comparison unfair. LVDM is provided with no additional information; IRASim is provided with trajectories segmented with SAM2; FLIP is provided with ground-truth flow.\"}", "{\"title\": \"General Response to All Reviewers\", \"comment\": \"We thank all reviewers for your dedicated time and effort to review our paper. We greatly\\nappreciate your insightful comments and questions. We thank all reviewers for their recognition\", \"of_the_novelty_of_our_work_and_its_value_to_the_field\": \"*I think the combination of ideas here - for\\nplanning - is original/novel* (reviewer H4xH), *a new model-based planning framework* (reviewer\\n9xux), *leveraging CVAE to address the multi-modality of flows is well-motivated* (reviewer R3ys), *a\\nnovel approach to model-based planning* (reviewer 6B9z).\\n\\nIn this rebuttal, we perform extra experiments to respond to reviewers' common concerns.\\nPlease find the new pdf file (modified areas are highlighted in blue), new appendix (modified areas\\nare highlighted in blue), attached videos, and more videos on our website. New experiments include:\\n\\n1. We re-design the low-level policy to: 1) use diffusion policy as the training algorithm; 2) use\\naction chunking as the output; 3) simplify the flow and video condition mechanism. Please see\\nAppendix A.3 for more details.\\n\\n2. Real robot experiments of 2 long-horizon tasks, including a tea scooping task and a cloth unfold-\\ning task. Please see Appendix B.4 for more details and results.\\n\\n3. Using diffusion models as the action module (flow generation model) and compare it with the\\nCVAE architecture. Please see Appendix B.3 for more results.\\n\\n4. Testing OpenVLA for LIBERO-LONG tasks, and testing using pretrained FLIP from LIBERO-\\n90 and fine-tuning it as the planner for training a low-level policy for LIBERO-LONG tasks. Please\\nsee Appendix B.2 for more results.\"}", "{\"summary\": \"The paper introduces a model-based planning framework (FLIP) designed for general-purpose robotic manipulation tasks using language and vision inputs. The framework allows for the progressive synthesis of long-horizon action plans, starting from an initial image and language instruction. FLIP effectively uses image flows to represent complex movements, enhancing the planning process for manipulation tasks across various objects and robots. They propose a multi-modal flow generation model predicting actions by generating dynamic representations of movements, simulating future video sequences based on the proposed actions, and evaluate the generated videos to maximize the task's success probability.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": [\"The paper is well-written and methodically presented.\", \"FLIP introduces a novel approach to model-based planning by integrating multi-modal inputs, which enhances its versatility for various manipulation tasks\", \"FLIP is designed to scale with increasing model and data budgets, making it suitable for a variety of applications and capable of leveraging more extensive datasets as they become available.\", \"The generated plans can be used to inform and train low-level control policies which can support several hierarchical policies that require strategic decision-making and planning.\"], \"weaknesses\": [\"The framework may struggle with ambiguities in language instructions or unexpected changes in the environment.\", \"Depending on the computational demands of the multi-modal modules, real-time performance in dynamic environments could be a concern. Evaluating the speed and efficiency of planning under real-time constraints would be crucial.\"], \"questions\": [\"How does FLIP compare to other state-of-the-art planning frameworks in terms of efficiency and effectiveness?\", \"What are the limitations of using image flow as an action representation? In which scenarios/tasks, this would not be ideal?\", \"How does FLIP handle ambiguities in language instructions?\", \"Are there any results on the real robot that show the applicability on world models even for quasi-static tasks?\", \"How will the model's performance be affected in the presence of noise or visual obstructions?\", \"How does the performance of FLIP depend upon or scale with the size of the available data, numerical analysis for the same would be helpful to understand the efficiency of the proposed approach.\", \"Please also address other comments in the weaknesses section above.\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"metareview\": \"This paper presents FLIP (FLow-centric generative Planning), a novel model-based planning framework for robotic manipulation tasks that operates on visual and language inputs.\\n\\nThe reviewers consistently praised the paper's innovative combination of ideas in planning, visual representations, and generative modeling. They highlighted the well-motivated use of CVAE for handling multi-modal flows, comprehensive experiments, and the potential for scaling with increasing model and data budgets. The paper demonstrates strong results across video plan synthesis and manipulation tasks.\", \"initial_concerns_focused_on\": \"limited downstream performance gains compared to simpler baselines like ATM, lack of real-robot experiments, and questions about comparison to recent text-to-video methods. The authors provided extensive responses and additional experiments addressing these points. They improved the low-level policy architecture using diffusion models and action chunking, added real robot experiments showing strong performance on challenging tasks, and clarified their choice of baselines and experimental conditions. The authors also conducted new ablation studies comparing CVAE vs diffusion models for flow generation.\\n\\nThe rebuttal successfully addressed most reviewer concerns, with all reviewers explicitly noting satisfaction with the responses. While some reviewers maintained reservations about the magnitude of performance improvements and fairness of certain comparisons, they agreed the paper's novel framework and comprehensive evaluation merit acceptance.\", \"additional_comments_on_reviewer_discussion\": \"None -- see metareview\"}" ] }

[ "Rafael Valle", "Rohan Badlani", "Zhifeng Kong", "Sang-gil Lee", "Arushi Goel", "Sungwon Kim", "Joao Felipe Santos", "Shuqi Dai", "Siddharth Gururani", "Aya Aljafari", "Alexander H. Liu", "Kevin J. Shih", "Ryan Prenger", "Wei Ping", "Chao-Han Huck Yang", "Bryan Catanzaro" ]

[ "Generative Models", "Audio", "Foundation Models" ]

{ "note_id": [ "zT7RO3QVn4", "ssfDkhKOf3", "q7QYxAM0FS", "phprYVlGYn", "juNm1paHIS", "hhlLxyjLsP", "eZAao61LvH", "bgWXOmcZXt", "aY11okDDVd", "Zb03ueMOd4", "ZQAFlVlspS", "XFoL6fuASs", "KebMLyDkgJ", "JTzom4UT52", "HkGGZCBoAX", "HWc04SMKUN", "ETB032vi90", "BjM1NYC35n", "8MkNEy38kG", "2v4Ps2hnM1", "268bPEqM7B" ], "note_type": [ "official_comment", "official_comment", "decision", "official_comment", "official_review", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_review", "meta_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment" ], "note_created": [ 1732479476993, 1732464264003, 1737524110878, 1732173306719, 1730348980043, 1732464932752, 1732834388057, 1730846136649, 1732780528257, 1732753521253, 1732795309439, 1729780368605, 1734700061025, 1732404097812, 1732751143143, 1732170719828, 1732482560768, 1732783047833, 1732170318881, 1732757094192, 1732174103496 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission11207/Authors" ], [ "ICLR.cc/2025/Conference/Submission11207/Reviewer_FfHR" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission11207/Authors" ], [ "ICLR.cc/2025/Conference/Submission11207/Reviewer_sTa1" ], [ "ICLR.cc/2025/Conference/Submission11207/Authors" ], [ "ICLR.cc/2025/Conference/Submission11207/Reviewer_sTa1" ], [ "ICLR.cc/2025/Conference/Submission11207/Reviewer_xcth" ], [ "ICLR.cc/2025/Conference/Submission11207/Authors" ], [ "ICLR.cc/2025/Conference/Submission11207/Reviewer_sTa1" ], [ "ICLR.cc/2025/Conference/Submission11207/Reviewer_xcth" ], [ "ICLR.cc/2025/Conference/Submission11207/Reviewer_FfHR" ], [ "ICLR.cc/2025/Conference/Submission11207/Area_Chair_4BtT" ], [ "ICLR.cc/2025/Conference/Submission11207/Reviewer_xcth" ], [ "ICLR.cc/2025/Conference/Submission11207/Authors" ], [ "ICLR.cc/2025/Conference/Submission11207/Authors" ], [ "ICLR.cc/2025/Conference/Submission11207/Reviewer_sTa1" ], [ "ICLR.cc/2025/Conference/Submission11207/Authors" ], [ "ICLR.cc/2025/Conference/Submission11207/Authors" ], [ "ICLR.cc/2025/Conference/Submission11207/Authors" ], [ "ICLR.cc/2025/Conference/Submission11207/Authors" ] ], "structured_content_str": [ "{\"title\": \"Request to review the rebuttal\", \"comment\": \"Thank you for taking the time to review our paper. We have addressed your concerns in our submitted response. As the rebuttal period is nearing its conclusion, we kindly request you to review our rebuttal and share any additional comments or concerns you may have. Thank you once again for your valuable feedback and we would be happy to answer additional questions!\"}", "{\"comment\": \"Thank you for your response.\\n\\nMy concerns have been partially addressed.\\nFor Table 1, I suggest that the authors carefully review it. For instance, AudioBox can be considered as trained on a large-scale dataset. While I agree with the authors\\u2019 use of this table to highlight their work, please ensure that the authors of the cited paper would not disagree or be misrepresented if any inaccuracies or incorrect decisions about their work are made.\\n\\nAdditionally, please include a discussion section in the final version, particularly addressing the following points:\\n\\n1. How to determine the duration.\\n\\n2. The definition of \\\"Emergent Properties.\\\"\\n\\nI recommend emphasizing that your interpretation of \\\"Emergent Properties\\\" is specifically limited to task combination. From my perspective, this differs from the \\\"Emergent Properties\\\" observed in LLMs. Please ensure this distinction is clear to avoid confusing readers.\\n\\nI will increase my score to 6.\\n\\nBest wishes.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}", "{\"title\": \"Addressing Comments and/or Questions\", \"comment\": \"Thank you for your thoughtful feedback and for sparking a discussion on \\\"unsupervised multitask learning\\\" and \\\"emergent abilities.\\\" We hope that our comments below clarify and align our positions, and hope that you would be willing to raise your score. We are happy to address any other questions you might have.\\n\\n**a. How should we understand \\\"unsupervised multitask learning\\\"?**\\n\\nThank you for this comment, as it seems to both have captured our formulation, as well as a gap in our presentation of it, which we will improve upon in the new manuscript. \\n\\nOur use of the term \\u201cUnsupervised Multitask Learning\\u201d follows \\\"Language Models are Unsupervised Multitask Learners\\\" (2019), where p(output_text | input_text) is considered unsupervised and p(output_text | input_text, task_text) is considered supervised. Compared to the text domain, modeling p(output_audio | input_audio) or p(output_audio | input_audio, task_text) does not provide us with the ability to teach the model to follow language based instructions related to sound generation and transformation.\\n\\nAs such, by drawing a parallel with text-LLMs, one can define that p(output_audio | input_audio, instruction, task_text) [template-based instructions] is equivalent to supervised learning, and p(output_audio | input_audio, instruction_text) [free-form instructions] is equivalent to unsupervised. Then, \\u201cunsupervised multi-task learning\\u201d can be defined as learning to perform multiple tasks with text instructions without \\\"explicitly\\\" providing task conditioning. \\n\\nWe understand that, without this long contextualization, the term \\u201cunsupervised multi-task learning\\u201d and how it relates to our work may not be clear. As such, we will separately draw the parallel with text-LLMs and find an alternative term to refer to our setup.\\n\\n**b. Your main framework is based on Flow Matching (FM). How do you control the duration?**\\n\\nDuring training for TTS and SVS, we silence-pad the audio up to 20 seconds and compute the loss with weight 0.1 on the padded region, following E3TTS (https://arxiv.org/abs/2311.00945). During inference, we sample a 20-second long noise and can, arguably, control the speech rate through language.\\n\\n**c. In Table 1, what qualifies as \\\"large-scale data\\\" in terms of hours?**\\n\\nWe will update the paper to list the number of hours used on each model and mark UniAudio as using \\\"large-scale\\\" data. On another note, we will rectify an arithmetic mistake in our total number of hours: 20 million rows with 10 seconds of audio equates to 50,000 hours of audio, not 2.9 million hours.\\nThough the on-the-fly augmentations in audio and instructions considerably do increase the number of audio and text pairs to millions, we prefer to provide the lower bound in number of audio hours.\\n\\n**d. The definition of \\\"Emergent properties\\\":**\\n\\nWe believe that \\\"emergent properties\\\" are starting to be studied in Audio foundation models and appreciate the reviewer\\u2019s willingness to discuss it within this review process. We follow the definition in \\u201cEmergent Abilities of Large Language Models\\u201d ( https://arxiv.org/abs/2206.07682), which considers \\u201can ability to be emergent if it is not present in smaller models but is present in larger models.\\u201d. As we stated in the paper, Fugatto\\u2019s smallest model does not possess emergent abilities present in the largest model.\\n\\n**e. The authors need to explicitly highlight the advantages Fugatto has over AudioBox. That said, I do agree that the model's performance is impressive.**\\n\\nPlease refer to Table 1, the Experiments section comparing Fugatto with AudioBox and UniAudio, and Fugatto's emergent abilities. If these are not sufficient, we kindly ask you to provide suggestions on what aspects you would like to see compared.\"}", "{\"summary\": \"The paper proposes Fugatto 1, a versatile audio synthesis and transformation model, has capable of following free-form text instruction with optimal audio inputs. To address the problem that audio data does not inherently contain the instructions that were used to generate it, the paper introduces a specialized dataset generation approach to enhance the data reveals meaningful relationships between audio and language. Additionally, the paper proposes ComposableART with CFG to enhance the compositional abilities of the model.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"1\", \"strengths\": \"The paper proposes Fugatto, a large foundational audio synthesis and transformation model. This paper attempts to address the problem of audio data does not highly correspond to the text-instruction by a data generation strategy and extend CFG to support compositional guidance. The paper provides extensive experiments across different tasks in the audio domain.\", \"weaknesses\": \"The paper is a bit difficult to follow, and the proposed methods are trivial incremental. The main contribution of this paper is to train a generalist model with additional enhanced diverse synthesized datasets. I admit that the motivation of exploring a generalist model to benefit downstream tasks is good; however, the paper methodology lacks insights in the domain of audio synthesis and generation. The contribution of this paper for audio domain research is limited. In my opinion, this paper is more suitable for a technical report.\\n\\nI feel lost when reading the experiment section, the paper should provide at least a brief introduction for evaluation metrics used in the experiments, experimental details for adapting the proposed method in each single-task, and insights about why to adapt the method to each different single-task.\\n\\nEven though the paper showcased its applicability in various audio related tasks, however, it is not convincing to me the advantage of the proposed generalist model compared to other specialist models for each single-task. For example, in the TTS experiment, only speech similarity and intelligibility been evaluated, a further analysis, such as MOS study or F0 visualization compared to other methods and GT, is necessary for evaluating speech naturalness and quality to showcase the method can provide more natural speech than existing methods; No model comparison for the SVS experiment; Table 3(b) compares its performance with other specialist models, however, MusicGen and AudioLDM2 have different focuses (music v.s., multi-modality audio generation). The experiment is not a fair comparison and not convincing.\", \"questions\": \"1. If I understand correctly, the paper provides an incremental method for building a large foundational audio generation model. What is the technical contribution of this paper?\\n\\n2. Have you verified the quality of the synthesized new dataset? How do you ensure the synthesized data is high-quality and strongly corresponds to the prompt instruction?\\n\\n3. Will you release the codes, the new datasets and the pretrained model?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Thank you\", \"comment\": \"Thank you for your the short discussion, your comments, and the opportunity they create to improve the paper and, by consequence, the knowledge it shares with our community. We will modify the manuscript to incorporate your suggestions.\\n\\nAnd thank you for increasing your score to 6. Though the proposed change hasn't been reflected on OpenReview yet, we assume it will be reflected towards the final score.\"}", "{\"comment\": \"Thank the author for this clarification. Now I\\u2019m convinced and adjust the score to 6 accordingly. However, the paper should be majorly revised to highlight above as well as a comprehensive related work. Good luck !\"}", "{\"summary\": \"The paper presents Fugatto, a generalist audio synthesis model capable of handling various audio generation and transformation tasks using text prompts. It introduces a dataset generation method that broadens Fugatto\\u2019s capabilities. In addition, ComposableART is proposed which extends classifier-free guidance to allow compositional control over audio generation. The model is evaluated on tasks across different audio domains, showcasing its versatility and competitive performance.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. Fugatto handles multiple tasks with text and audio inputs.\\n2. The approach to synthetic instruction generation and data enrichment is well-structured, supporting the model\\u2019s generalization across diverse tasks\\n3. ComposableART enables flexible audio composition, interpolation, and negation of prompts, adding control over generated outputs.\", \"weaknesses\": \"1. Techniques such as using LLMs for synthetic instruction generation lack novelty, which may challenge the paper's originality and scientific contribution.\\n\\n3. The comparison with state-of-the-art specialist models is limited for certain tasks, and the overall impact of ComposableART on performance remains unclear.\", \"questions\": \"1. How does ComposableART impact model performance quantitatively on specific tasks? It is recommended that the authors include more subjective and objective comparisons of the proposed system against task-specific models, such as LASS-Net and AudioSep for text-based audio removal.\\n\\n2. Could the authors clarify which benchmarks or metrics were used to evaluate compositional tasks?\\n\\n3. As a scientific conference submission, I would personally advise against the use of excessive fancy fonts and varied colors within the paper.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Insights about the inherent nature of audio 1 of 2\", \"comment\": \"Dear reviewer,\\n\\nbelow we showcase design choices and contributions in the paper that highlight audio specific insights. Our main design choice in Fugatto is to develop a general model with weaker modeling assumptions, or inductive priors, in consonance with recent work in [LLMs](https://youtu.be/orDKvo8h71o?si=85brQxlhPTWcWhcy&t=774)\\n\\nGiven the diversity in audio subdomains (speech, music, other sounds, captions, music scores, speech prompts), we focus on weak modeling assumptions that support audio synthesis and transformations through text instructions and optional audio inputs in general. This is in contrast with recent works and specialist models that focus on strong assumptions.\", \"below_we_highlight_challenges_that_motivate_our_design_choices_and_disadvantages_of_other_approaches_towards_building_generalist_models\": \"I) Conditioning on Music Scores, Chords, Melodies\\n\\nThough a musical score can be represented with text (MuseNet and others), this choice comes with several drawbacks:\\n1) Prohibitively long context, specially for complex orchestrations from the likes of Wagner and Mahler\\n2) Challenging to temporally align each instrument or instrument section \\n3) Challenging to represent melodic curves such as the ones in [Xenakis' Aroura](https://www.youtube.com/watch?v=6HmpDpZWLCw)\\n\\nOur design choice is to use audio generated from MIDI, and provide it to a single audio encoder shared amongst all task. This provides several advantages:\\n1) Context length does not scale with the number of instruments\\n2) Instruments and instrument sections are temporally aligned by construction\\n3) Melodic curves are directly represented\\n4) The audio encoder is shared between all tasks and audio subdomains, promoting emergent abilities.\\n\\nOur argument can be extended to melodies and chords. For chords, previous works suggest using a look-up table, which is clearly a brute force approach given that even a instrument like the piano, with only 88 unique notes, can produce as many as 10^26 unique chords.\\n\\nII) Speech Synthesis\", \"speaker_embedding\": \"Some models promote using a speaker verification model for embedding the speaker information, this comes with several drawbacks:\\n1) The information bottleneck from speaker verification models has been proven to result in worse speaker similarity (p-flow, vall-e, voicebox)\\n2) A separate speaker embedding does not provide an audio embedding that is shared amongst domains, possibly limiting emergent capabilities\\n\\nOur design choice is to use the output of the shared audio encoder. This provides several advantages:\\n1) Model has full access to the speaker's sample, promoting higher speaker similarity\\n2) Full access to the speaker's previous sample, promoting continuation tasks\\n3) A shared audio encoder likely promotes emergent capabilities\", \"f0_control\": \"some models explicitly conditioning on F0 by providing a separate embedding for F0. This comes with disadvantages:\\n1) A separate F0 encoder does not provide an audio embedding that is shared amongst domains, possibly limiting emergent capabilities\\n2) Drastically complicates the data ingestion pipeline, creating difficulties for scaling the data\\n3) Hard to adapt to polyphonies\\n\\nOur design choice is to provide control over F0 through text captions or through an F0 contour as audio provided to the audio encoder. This provides several advantages\\n1) F0 can be modified with verbal instructions\\n2) Our data generation strategies promotes absolute and relative instructions (higher, lower, more varied contour, less varied contour, etc...).\\n3) Trivial to go from a F0 contour to audio but arguably hard to go from audio to an F0 contour\\n4) Trivial to adapt to polyphonies\\n5) The audio encoder is shared between all tasks and audio subdomains, promoting emergent abilities.\\n\\nOur argument can be extended to phoneme durations, which can be easily controlled by asking the model to speak slow, fast, slower or faster.\\n\\nIII) Text representation\\n\\nThough fixed vector length representations like CLIP and CLAP have been used in TTA models, this design choice comes with several drawbacks:\\n1) Due to their fixed length representation, information needs to be compressed to maximize the alignment between caption and audio, removing information that is not related to the captions\\n2) Representation is not adequate for textual instructions, nor the \\\"text\\\" to be sung or said in SVS and TTS tasks\\n3) Such representations are normally interpreted as bag of words, providing very [limited understanding of language](https://www.youtube.com/watch?v=BnpB3GrpsfM)\\n\\nOur design choice is to instruct the model with a byT5 representation. This has several advantages:\\n1) Supports graphemes, IPA, and characters from non-english languages\\n2) byT5 is trained on vasts amounts of language and, hence possessing a much better understanding of language than CLIP or CLAP.\\n3) Supports captions, instructions, text, html, etc.. all in a single model and shared space\"}", "{\"comment\": \"This paper lacks technical contribution to the domain of audio research. Although the paper presents an ambitious motivation; however, either using language prompts to build additional dataset for training the model and adjusting model architecture and scales (e.g., DiT , ComposableART, and sampling schedules) are common solutions for generative tasks.\\n\\nThe main research problem in this paper, if I understand correctly, is to increase the data scale and model scale, further to make the model to be generalist. I think this paper may not be counted as a scientific paper, which mainly doing engineering rather than solving the problem with a clear insight on audio domain. In other words, the proposed solution can be directly applied to any generative tasks, such as image. Audio is a general scope that can be classified into many subdomains, such as music, speech, and sounds. Each subdomain has its different focus, which cannot be solved by simply increasing the model scale and data scale. \\n\\nI suggest the author to think the problem of how to build a generalist audio model deeper, with insights about the inherent nature of audio. In addition, this paper should be revised, with related work section and, as other reviewers suggest, detail experimental details.\\n\\nIn summary, I think this paper lacks enough technical contribution as a scientific paper.\"}", "{\"comment\": \"After reviewing the authors' responses to the other reviewers, I have reconsidered my evaluation. I now believe that a generalized paradigm for unified and controllable audio generation represents a significant contribution to the field. I am happy to raise my score to 8. I also encourage the authors to include additional discussion of related work and to emphasize their design choices, particularly in response to reviewer sTa1's feedback.\"}", "{\"summary\": \"This paper introduces Fugatto, a versatile audio synthesis and transformation model capable of following free-form text instructions with optional audio inputs. The paper focuses on enabling generalist audio models with emergent capabilities, similar to large language models (LLMs), but optimized for audio. Fugatto can handle tasks such as audio generation, transformation, and compositional tasks by using a new dataset generation approach and a technique called Composable Audio Representation Transformation (ComposableART). This method enhances classifier-free guidance, enabling complex operations like combining, interpolating, and negating instructions.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"(1) Generalist Audio Model: Fugatto offers a broad range of audio generation and transformation capabilities, filling the gap between specialist models and generalist models.\\n\\n(2) ComposableART: The novel technique extends classifier-free guidance to handle compositional tasks, allowing the model to compose instructions in ways that were not seen during training.\\n\\n(3) Dataset Generation: The paper provides a strategy for generating diverse and dynamic datasets, using LLMs for instruction creation and data augmentation. The authors claims they will release these the dataset and instruction generation code, which is useful for the research community.\", \"weaknesses\": \"At this stage, I believe the presentation of this paper is not strong enough. I still have the following questions:\\n\\na. How should we understand \\\"unsupervised multitask learning\\\"? In your training tasks, each task has both inputs and labels, such as TTS and TTA. I am not clear on how this relates to unsupervised multitask learning. UniAudio and AudioBox are closely related to your work, and I believe both Fugatto and UniAudio are trained on multiple tasks. Generally, we refer to 'unsupervised learning' when there are no labels, such as in the pre-training of LLMs.\\n\\nb. Your main framework is based on Flow Matching (FM). How do you control the duration? Since FM is a non-autoregressive model, this could be a challenge. For example, AudioBox or VoiceBox use a phoneme duration predictor. Are you using a duration predictor for TTS? Similarly, for tasks such as singing generation, how is the duration controlled?\\n\\nc. In Table 1, what qualifies as \\\"large-scale data\\\" in terms of hours? It seems UniAudio uses more than 100,000 hours of audio data. It would be helpful if you could list the number of hours for each model, as this would provide readers with a clearer understanding of what constitutes 'large-scale.'\\n\\nd. The definition of \\\"Emergent properties\\\": In the context of LLMs, we describe 'emergent properties' as the ability to solve unseen tasks and perform reasoning. However, in this paper, the examples given for emergent properties seem to focus on generating sounds that don\\u2019t exist in the real world or the training data. From my perspective, this doesn\\u2019t fit the definition of 'emergent properties,' as the model has already learned to understand sound types based on text descriptions. I strongly recommend discussing this point with other reviewers, as it currently seems like a bit of an overstatement.\\n\\ne. From a high-level perspective, Fugatto follows the multi-task training paradigm used in UniAudio and AudioBox. The authors need to explicitly highlight the advantages Fugatto has over AudioBox. That said, I do agree that the model's performance is impressive.\\n\\nIn conclusion, I agree that building a generalist model is a valuable topic, and this paper demonstrates good performance. However, the authors need to improve the presentation to help readers better understand the contributions.\\n\\nI am happy to improve my score during the rebuttal stage if the authors solve my concerns.\", \"questions\": \"Refer to weakness.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"metareview\": \"> This paper introduces Fugatto, a versatile audio synthesis and transformation model capable of following free-form text instructions with optional audio inputs. The paper focuses on enabling generalist audio models with emergent capabilities, similar to large language models (LLMs), but optimized for audio. Fugatto can handle tasks such as audio generation, transformation, and compositional tasks by using a new dataset generation approach and a technique called Composable Audio Representation Transformation (ComposableART). This method enhances classifier-free guidance, enabling complex operations like combining, interpolating, and negating instructions.\\n\\nThe paper is of sufficient scope, experimentally well validated against the adequate concurrent models and baselines. The work overall is a positive contribution to the ML research community.\\n\\nThanks to reviewers for engaging with the authors during the rebuttal. The paper post-rebuttal meets the criteria for publication at ICLR.\", \"additional_comments_on_reviewer_discussion\": \"Thanks to reviewers for engaging with the authors during the rebuttal. The paper post-rebuttal meets the criteria for publication at ICLR.\"}", "{\"title\": \"Response to authors\", \"comment\": \"Thank you for your response. My concerns have been partially addressed.\\n\\nWhile I acknowledge that the approach to synthetic instruction generation and training with large-scale data is well-structured and represents a solid contribution, the underlying novelty appears limited, as it seems more like an engineering optimization rather than a scientific advancement.\\n\\nThe ComposableART component also seems to be an interesting contribution; however, the authors do not provide enough results, analysis, and comparisons to make the contribution distinct.\\n\\nTherefore, I will maintain my original score.\"}", "{\"title\": \"Clarifications and Follow-Up on Audio Generation Domain Insights\", \"comment\": \"Dear reviewer,\\n\\nThank you for your feedback and for raising your score. Could you kindly elaborate on the specific domain insights of audio generation that you feel are missing from our paper? While we believe our contributions provide substantial insights, we would be happy to incorporate additional domain-specific perspectives to further enhance the work.\\n\\nPlease let us know what you would like to see more of.\"}", "{\"title\": \"Addressing Comments and/or Questions.\", \"comment\": \"Thank you for your feedback. Below we attempt to address your questions and commentaries, in hopes that you would be willing to raise your score.\\n\\n**Comment and/or Question 1: \\u201dTechniques such as using LLMs for synthetic instruction generation lack novelty, which may challenge the paper's originality and scientific contribution.\\u201d**\\n\\nWe agree that using LLMs for synthetic instruction generation, as explored in our dataset generation pillars I, III, and V, have been addressed in previous work. With the paragraph below, we hope that you agree that our approach is novel in multiple ways. \\n\\nFirst, we unify synthetic instruction generation and dataset creation into a cohesive framework. Furthermore, dataset generation pillars II and IV stand out as innovative contributions: (II) generating absolute and relative instructions enables nuanced tasks like \\u201cincrease the happiness of this voice,\\u201d or \\u201cincrease the reverb\\u201d, while (IV) transmuting datasets uncovers latent relationships, allowing the creation of entirely new tasks such as MIDI2AUDIO and Speech Modulation (Emotion Conversion, Emotion Modulation, Sentence Variation). *Importantly*, we not only propose a dataset and instruction generation strategy, but also demonstrate its effectiveness\\u2014our results show that Fugatto achieves performance at least comparable to state-of-the-art specialist models, highlighting the originality and practical impact of our approach.\\n\\n**Comment and/or Question 2: The comparison with state-of-the-art specialist models is limited for certain tasks. It is recommended that the authors include more subjective and objective comparisons of the proposed system against task-specific models, such as LASS-Net and AudioSep for text-based audio removal.**\\n\\nWe appreciate your feedback and agree that additional evaluations on specific tasks could further demonstrate Fugatto\\u2019s capabilities. However, we hope the reviewer recognizes that the existing experiments already provide a broad and comprehensive evaluation across diverse tasks, including TTS, TTA, SVS, Speech Denoising, Upsampling, MIDI2Audio, and Speech Modulation. Furthermore, we hope that you agree that the extensive qualitative examples on our Demo Page, including emergent sounds and tasks, strongly substantiate Fugatto's abilities, versatility and applicability to a large number of scenarios.\\n\\n\\n**Comment and/or Question 3: How does ComposableART impact model performance quantitatively on specific tasks?**\\n\\nInstead of using ComposableART to improve Fugatto\\u2019s performance on benchmarks, which is already comparable or superior to specialist models and UniAudio, our focus with ComposableART is to control sound generation in a new way, to create novel combinations of sounds that don't exist in the training data, and to provide users with the ability to control the influence of each instruction over time.\\n\\n\\n**Comment and/or Question 4: Could the authors clarify which benchmarks or metrics were used to evaluate compositional tasks?**\\n\\nWe used cosine similarity between CLAP embeddings between A and B, where A is obtained from the text description used to create the audio event and B is obtained from the model\\u2019s audio output generated with the text description. \\n\\n**Comment and/or Question 5: As a scientific conference submission, I would personally advise against the use of excessive fancy fonts and varied colors within the paper.**\\n\\nWe appreciate your suggestion and will find an alternative to articulate the entities without using color.\"}", "{\"comment\": \"Thank the author for the clarification, my concerns have been partially addressed. I'm happy to raise my score from 3 to 5. The paper could be further improved by involving domain insights of audio generation.\"}", "{\"title\": \"Insights about the inherent nature of audio 2 of 2\", \"comment\": \"We hope that you agree that the simplicity in our formulation, our ability to easily scale up the data and number of tasks, and the performance of our model reflect our thorough understanding of audio, and deep consideration of the challenges in combining tasks and audio data from different domains, not a lack of insight.\\n\\nPlease note that even though we do not explicitly provide a related works section, it is merged with the introduction given the limit in number of pages and the amount of information we must cover in a work like Fugatto.\\n\\nWe are happy to highlight more challenges and how they were addressed in our paper.\"}", "{\"title\": \"General Comment\", \"comment\": \"We kindly thank the reviewers for their comments and suggestions.\\n\\nGenerally speaking, we would like to emphasize that Fugatto is a generalist model that is comparable or better than specialist models, even though it was never adapted to specific tasks. In addition to the list of contributions already provided in our Introduction, below we provide a list of our Technical Contributions and Insights in hopes that this birds-eye view further clarifies our contributions.\\n\\nIn addition, since our submission, we have found new emergent capabilities (New Tasks) from Fugatto related to combining tasks seen during training to perform a new task. We updated our website to illustrate them and provide a concise commentary as well.\\n\\nOn a final note, we observed an arithmetic mistake when computing the total number of hours we used to train our model: 20 million rows with 10 seconds of audio equates to 50,000 hours of audio, not 2.9 million hours as we have reported. We have rectified this mistake and have marked UniAudio as using large-scale data on Table 1.\\n\\n**Technical Contributions (Ablations):**\\n1) **t-Sampling Schedules**: We demonstrate that uniform t-sampling is effective across all tasks, while logistic t-sampling, proposed by Stable Audio, significantly degrades TTS performance, highlighting a key insight for generalist models. \\n2) **Model and data scale**: Our results show that increasing model parameters improves validation losses and delays overfitting. Consistent with \\\"Language Models are Unsupervised Multitask Learners\\\" (2019), we observe that smaller models lack emergent abilities.\\n3) **\\u201cDiT\\u201d implementation improvements**: We improve the \\u201cDiT\\u201d implementation by computing adaptive layer norm in FP32, using GELU as approximate tanh, and initializing final layers to output zero, aligning with our scaled mel-distribution.\\n\\n**Technical Contributions (Dataset Generation):**\\nGiven that data scarcity is a problem that must be addressed to train generalist state-of-the-art models for audio generation and transformation \\u2013 audio data does not come with the instructions used to create it \\u2013 we make the following contributions:\\n1) We propose a thorough strategy for producing datasets and instruction sets, aiming to address data scarcity and provide a framework that benefits the broader research community. This strategy unifies synthetic instruction generation and dataset creation into a cohesive framework.\\n2) We propose a dataset generation strategy based on five pillars. While pillars I, III, and V, have been addressed in previous work, pillars II and IV stand out as innovative contributions: (II) generating absolute and relative instructions enables nuanced tasks like \\u201cincrease the happiness of this voice,\\u201d or \\u201cincrease the reverb\\u201d, while (IV) transmuting datasets uncovers latent relationships to create entirely new tasks like MIDI2AUDIO and Speech Modulation.\\n3) Importantly, we not only propose a dataset generation strategy but also demonstrate its effectiveness\\u2014our results show that Fugatto achieves performance comparable to or exceeding state-of-the-art specialist models, underscoring the originality and practical impact of our approach.\\n\\n\\n**Technical Contributions (Compositionality):**\\nRecognizing the challenges of achieving compositional abilities (combination, interpolation, negation) from data alone, we introduce ComposableART, an inference method for composing instructions through latent space manipulation, even across different models.\\n\\n\\n**New Emergent Abilities (New Tasks):**\\nWe were able to find new emergent abilities in Fugatto that showcase its ability to execute a new task that we interpret as a combination of related tasks seen during training. We provide samples in the Emergent Tasks section in our demo page https://fugatto.github.io\\n1) Fugato is trained on A) TTS with speech prompt and B) Singing Voice Synthesis with a text prompt, but never on A+B. \\nSurprisingly, we find that the model is able to perform A+B: Singing Voice Synthesis with a speech prompt.\\n2) Fugato is trained on A) TTA with captions and B) Real Audio from MIDI audio with music styles (MIDI2AUDIO) but never on A+B.\\nNonetheless, we find that the model is able to perform A+B: TTA with captions and a melody provided as audio prompt, and the model\\u2019s outputs follow the notes in the melody. \\n3) Fugato is trained on A) MIDI2AUDIO and B) SVS with a text prompt but never on A+B.\\nWe observe that Fugatto can perform A+B: SVS with a text prompt and a melody provided as an audio prompt, and the model\\u2019s output follows the notes in the melody.\"}", "{\"title\": \"Focus of different domains\", \"comment\": \"Thank you for reply.\\n\\nWe will soon provide a response describing the adjustments that have been done such that all subdomains in audio (music, speech, the rest) are addressed in a way that does not compromise each other. Meanwhile, we kindly ask the reviewer to familiarize themselves, if that's not the case, with the paradigm described by Hyung Won Chung in the video below, highlighting that different data regimes require different inductive biases. We believe that an understanding of inductive biases will be important in understanding the explanations we will provide regarding adjustments for each subdomain.\", \"https\": \"//youtu.be/orDKvo8h71o?si=85brQxlhPTWcWhcy&t=774\"}", "{\"title\": \"Addressing Comments and/or Questions\", \"comment\": \"Thank you for your feedback. Before addressing your comments below, we emphasize that Fugatto is a generalist model that is never adapted to a single-task. We hope that the comments clarify our paper and hope that you would be willing to raise your score.\\n\\n**Comment and/or Question 1: the proposed methods are trivial incremental.**\\n\\nWe acknowledge your feedback regarding the clarity of the paper and will revise the manuscript to improve its clarity. \\n\\nWe appreciate your perspective but respectfully suggest that, given triangle inequality and Table 1 in our paper, characterizing our work as \\u201ctrivial incremental\\u201d not only devalues the significant effort and innovation behind it, but also inadvertently undermines the contributions of other models we compare with. We kindly ask you to read our technical contributions in our general comment, as well as revisit our Introduction section, to re-asses your characterization of our work and other works as trivial.\\n\\n**Comment and/or Question 2: I admit that the motivation of exploring a generalist model to benefit downstream tasks is good;**\\n\\nWe appreciate your acknowledgment of the value that generalist models can bring to downstream tasks. We emphasize, however, that Fugatto is a generalist model that is comparable or better than specialist models. We also highlight that our work focuses on the unique properties and benefits of generalist models, especially emergent properties.\\n\\n**Comment and/or Question 3: The paper methodology lacks insights in the domain of audio synthesis and generation.**\\n\\nWe appreciate your perspective but respectfully disagree with the characterization of our work as lacking insights in the domain of audio synthesis and transformation. We believe this assessment overlooks our contributions, as well as the positive feedback we have received from other reviewers. Please see our technical contributions in our general comment.\\n\\n**Comment and/or Question 4: experimental details for adapting the proposed method in each single-task**\\n\\nWe emphasize that Fugatto is not adapted for each single-task (TTS, SVS, TTA, etc\\u2026). Fugato is a single generalist model trained to follow instructions expressed in text and optional audio inputs. We appreciate your suggestion and will provide a brief introduction for the evaluation metrics.\\n\\n**Comment and/or Question 5: For example, in the TTS experiment, only speech similarity and intelligibility been evaluated.** \\n\\nThe evaluation used in our paper for TTS is present in several SOTA TTS models since Vall-E, including P-Flow, VoiceBox, and AudioBox.\\n\\nNonetheless, we address your request and run SQUIM-MOS, i.e. model based MOS, evaluations to compare Fugatto with samples available on Vall-E\\u2019s and UniAudio\\u2019s demo page. Our results below further promote our results in the paper, and show that Fugatto surpasses Vall-E and UniAudio in terms of SQUIM-MOS.\\n\\nSQUIM-MOS Fugatto 4.75 Vall-E 4.00\\n\\nSQUIM-MOS Fugatto 4.58 UniAudio 3.933\\n\\n**Comment and/or Question 6: No model comparison for the SVS experiment; MusicGen and AudioLDM2 have different focuses (music v.s., multi-modality audio generation).**\\n\\nA direct comparison with other SVS model is not possible because, unlike TTS models, they do not provide the speech prompts used during inference. As such, we do not compare with them not to draw unclear conclusions.\\n\\nRegarding TTA comparisons with MusicGen and AudioLDM2, we re-emphasize that Fugatto is a generalist model.\\n\\n**Comment and/or Question 7: What is the technical contribution of this paper?**\\n\\nPlease refer to our technical contributions in our general comment.\\n\\n**Comment and/or Question 8: Have you verified the quality of the synthesized new dataset?**\\n\\nYes, quality and efficacy have been confirmed in many ways: \\n1) All our results were obtained with Fugatto trained with our data generation strategy, and Fugatto is on par or superior to the state-of-the-art.\\n2) Some of the synthetic captions used in Fugatto were confirmed to work on several papers, including https://arxiv.org/abd/2406.15487 and https://arxiv.org/abs/2407.04416v3.\\n3) The LLM-generated instructions were manually inspected with a 1 in 100 test, where we sample 100 instructions and check if at most 1 instruction needs improvement, else we improve the instruction generator by hand or by re-prompting the LLM. \\n\\n\\n**Comment and/or Question 8: Will you release the codes, the new datasets and the pretrained model?**\\n\\nYes. We believe that releasing training and inference code, including instruction generators, data and task processors, including a list of datasets, and instructions on how to create new datasets from existing datasets is very important. By releasing such assets, we hope that we will accelerate our journey, as a community, towards a future where unsupervised multitask learning in audio synthesis and transformation emerges from data and model scale. We plan to release checkpoints once the appropriate guardrails to prevent misuse are in place.\"}" ] }

[ "Luke Heffernan" ]

[ "zero shot", "classification", "CLIP", "VLM", "DCLIP", "WaffleCLIP", "open vocabulary", "pretrained" ]

{ "note_id": [ "wFVwTiDfOr", "rO4AofYhXD", "q7MpfHgziQ", "pbulLlWyXe", "nv8YQxHnMx", "lFqckouVpf", "kyQpkiv9tv", "d3EBKqcyu4", "UZ2GOGryfy", "NjI4b6jC0n", "D22ScE1WxQ" ], "note_type": [ "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_review", "official_comment", "official_review", "official_review", "meta_review", "decision" ], "note_created": [ 1733304791293, 1733302093449, 1733303389323, 1730693561672, 1733304935908, 1730720604712, 1733297048931, 1729604731227, 1730202734848, 1734248234985, 1737524243179 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission13192/Authors" ], [ "ICLR.cc/2025/Conference/Submission13192/Authors" ], [ "ICLR.cc/2025/Conference/Submission13192/Authors" ], [ "ICLR.cc/2025/Conference/Submission13192/Reviewer_Uvhk" ], [ "ICLR.cc/2025/Conference/Submission13192/Authors" ], [ "ICLR.cc/2025/Conference/Submission13192/Reviewer_DoJ7" ], [ "ICLR.cc/2025/Conference/Submission13192/Authors" ], [ "ICLR.cc/2025/Conference/Submission13192/Reviewer_ewH2" ], [ "ICLR.cc/2025/Conference/Submission13192/Reviewer_cME6" ], [ "ICLR.cc/2025/Conference/Submission13192/Area_Chair_b8sh" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ] ], "structured_content_str": [ "{\"title\": \"Response to ewH2\", \"comment\": \"Thank you for taking the time to give feedback and adding references to clarify the paper. We\\u2019ll aim here to give some extra context to our work and respond to some of your points directly.\\n\\n**W1.1 - Use of WaffleCLIP Concepts** \\nWe have added WaffleCLIP + Concepts as a benchmark, using the high-level concepts provided in Roth et al. 2023 for the CUB, EuroSAT, Places365, Food101, and Oxford Pets datasets. While these concepts are limited to a select number of datasets and apply only at the whole-dataset level, DefNTaxS is able to create these hierarchical labels to subcategories within these datasets, using them to create distinction, increased interpretability, and ultimately improve performance. Our approach also automates the definition of these taxonomies, as opposed to the level of manual effort involved in generating these in WaffleCLIP. Please see responses to other reviewers for comparisons to other approaches (character count is very limited).\\n\\n**W1.2 - Formatting Issues** \\nFigure 1 has been condensed to reduce space usage. Also, equations have had reference numbers added, and bolding and underlining have been added to experimental results tables to indicate performance.\\n\\n**Q1 - CuPL Benchmarking** \\nThis may have been a simple omission in the Python-to-LaTeX printing code. CuPL has now been included in Table 1 of the submission as a baseline.\\n\\n**Q2 - Intra-class Similarity** \\nIn this case, we are referring to similarities in class labels or descriptors between classes that are actually not considered similar and can be resolved through extra context, e.g., \\\"crust\\\" describing the Earth or a pie, \\\"boxer\\\" describing a dog or an athlete, etc. In the situation you are describing, we agree, and we see that maybe our terminology may have been more precisely defined in this case.\\n\\n**Q3 - Improvements on EuroSAT and Oxford Pets** \\nWe have added further context to the additional changes to the text prompts for these two specific datasets, but both can be described through improved contextualization. Results for EuroSAT showed similar improvements to any other dataset when applying the taxonomic subcategory generation equally, but once terms like \\\"satellite imagery\\\" or \\\"images from EuroSAT satellite\\\" were added, we observed the performance jump seen in Table 1. We surmise that this is due to the significant difference between the semantics of the class labels in this dataset when seen from Earth versus from orbit. Similarly, the Oxford Pets dataset received a significant boost from the inclusion of the E-CLIP templates, a factor unique to this dataset and described in lines 331-338.\\n\\n**Q4 - Reduction of Competition Between Classes** \\nLeveraging the response to Q2, this is largely due to incorrect correlations between otherwise dissimilar classes. In other cases like fine-grained datasets such as CUB, the taxonomic subcategory labels like \\\"small birds\\\" and \\\"large birds\\\" may be generated. These act as consistent separators between classes that may otherwise share descriptors like \\\"black bird\\\", \\\"orange beak\\\", and other extremely common descriptors.\\n\\n**Q5 - Additional Experimentation** \\nThis experiment was intended to be undertaken for this submission, but we unfortunately ran out of time. We expect to see that the errors of misclassification while classifying correctly within a dataset are improved with DefNTaxS in comparison to other baselines used in this submission.\"}", "{\"title\": \"Response to Uvhk\", \"comment\": \"**W1 - Detailed Prompt Context Examples**\\nExamples of g(C,t_i) have been provided in lines 235-241, including Eq. 6, and context added in Figure 2 (lines 162-179).\\n\\n---\\n\\n**W2 - Distinction from Existing Literature** \\nThe distinction between DefNTaxS and the existing literature is in the method of implementation of the LLM-derived taxonomic subcategories, specifically their direct use in the classification text prompt alongside fine-grained descriptors. This approach incorporates hierarchical elements directly into the text embedding, instead of using them iteratively to reweight the model prior to classification (CHiLS, Novack et al., 2023) or creating text elements for classification separate from the class labels (ChatGPT-Powered\\u2026, Ren et al., 2023). These elements rely on non-interpretable clustering methods and modified scoring methods.\\n\\nDefNTaxS balances efficiency and accuracy of the LLM in creation and refinement of taxonomic subcategories, avoiding errors in generation as seen in D-CLIP descriptors (Menon et al., 2023, Roth et al., 2023). By incorporating both class-specific descriptors and taxonomic contexts into prompts, this approach increases the semantic granularity of zero-shot classification, improving interpretability, accuracy, and robustness to descriptor noise. It also gives insight into the taxonomic representations implicit in CLIP\\u2019s language structures through the effect of varying levels of visual and taxonomic granularity being more effective than either individually (Ablation 6.3).\\n\\n---\\n\\n**W3 - Baseline Selection** \\n*MPVR* is a concurrent work that has significant benefits within this field of work, but was being released alongside the completion of our work, making it difficult to recreate for comparison within the timeline of this submission. We acknowledge that there appears to be a marginal relative improvement on most datasets using MVPR, but it comes with a cost. MVPR uses a significantly increased text volume, making interpretability and comparison to known factors difficult. The intermediate prompting step also produces brand new prompts that affect the performance of the final classification step at each iteration, potentially leading to inconsistencies in performance and ability to externally verify this information.\\n\\n*CHiLS* cannot be compared to pure zero-shot methods like CLIP, D-CLIP, and DefNTaxS due to the iterative recalculation of the model weights using the LLM-generated hyponyms, creating an unclear baseline comparisons for a given model.\\n\\n*S3A* is a few-shot classification method (Zhang et al., 2023) and requires a user to have access to extra application-specific training data to enable the performance increases achieved in this approach. The DefNTaxS approach has been compared only to other zero-shot classification approaches for the fairest comparison possible.\\n\\n*LLM Explainer* is a method to provide \\\"post-hoc explanations [...] for other complex predictive models\\\" (Kroeger et al., 2024). While it does share the benefit of increased explainability with the DefNTaxS approach, the aims, and applications are distinct.\\n\\n---\\n\\n**W4.1 - Disjoint and Completeness Constraints** \\nTo clarify our interpretation of this query on completeness and disjoint constraints, we are taking this question to ask how it can be certain that: \\n1. A class will be assigned to one and only one taxonomic subcategory, and \\n2. All classes will be assigned to a subcategory with none leftover. \\n\\nWe have added Figure 2 and extra detail around this portion of the method section (lines 157-209) to clarify process flow, and Appendix A for the prompts used. In short, the LLM API is called iteratively for assigning each class, avoiding errors due to overloading the LLM with too many instructions at once.\\n\\n---\\n\\n**W4.2 - Taxonomy Layers** \\nLike CHiLS, only a single layer of taxonomy is generated. However, unlike CHiLS: \\n- This taxonomy focuses on traversing the taxonomic hierarchy upward, allowing use on datasets that contain classes of refined taxonomy, not only high-level classes like \\\"dogs\\\" and \\\"cats.\\\" \\n- DefNTaxS does not alter the weights of the base model, minimizing compute and reducing potential errors/inconsistencies due to manual editing. \\n- Utilizes both fine-grained semantics and taxonomic hierarchical information, allowing for grouping of classes into subcategories (especially useful in generalized datasets) while still maintaining distinction between those classes. \\n\\n---\\n\\n**W5.1 - Game Theory Reference Adjustment** \\nThe references to game theory in this case were intended to be an illustrative reference to the competition between classes. This reference has been removed to prevent confusion.\\n\\n---\\n\\n**W5.2 - Semantic Interpretability** \\nAs in related literature [D-CLIP, WaffleCLIP, CuPL], \\\"interpretability\\\" is to be understood as \\\"able to be read and validated through natural language.\\\" Qualitative examples of g(C,t_i) have been added in Figure 2 and lines 235-242 and Eq. 6.\"}", "{\"title\": \"Response to cME6\", \"comment\": \"Thank you for taking the time to give feedback and adding references to clarify the paper. We\\u2019ll aim here to give some extra context to our work and respond to some of your points directly.\\n\\n**W1 - WaffleCLIP Findings** \\nWhile WaffleCLIP concludes that their work is simply \\\"a sanity check\\\" and that \\\"VLMs struggle to leverage the actual semantics,\\\" their experiments in (1) 4.3.2 suggest that despite similar accuracy scores, D-CLIP and their own approach tend to produce these results through very different mechanisms (with the only difference in approach being semantic information vs noise) and (2) that WaffleCLIP + Concepts tends to \\\"demonstrate consistent and significant improvements.\\\" Other experiments of ours show that simply removing either the fine-grained semantic descriptors or the taxonomic hierarchical information from the prompt significantly alters performance in a manner consistent per dataset, suggesting both that semantics have more effect than previously suspected and that CLIP incorporates a level of hierarchical language structuring within its representation space.\\n\\n**W2 - Context of Improvements** \\nAs can be observed with the other baselines, the scale of relative incremental improvement is consistent with other literature in this field. DefNTaxS achieves equal or greater performance compared to all other baselines on the tested datasets (Table 1). All hyperparameters shared across approaches were kept consistent for fair comparison, and hyperparameters introduced in this publication (e.g., minimum number of classes per taxonomic class) are all calculated and selected as part of the taxonomy creation process (Figure 2 and lines 157-209).\\n\\n**W4 - Error Correction** \\nThese minor typos have been addressed, thank you for picking up on those and apologies that you had to. Other formatting changes include:\\n- Figure 1 has been condensed to reduce space usage.\\n- Equations have had reference numbers added.\\n- Bolding and underlining have been added to experimental results tables to indicate performance.\"}", "{\"summary\": \"This paper proposes to leverage LLMs to generate hierarchical taxonomic sub-categories for a specific category to augment the textual prompt of CLIP to better capture the semantic information in the text prompts as well as refine the alignment between the images and prompts. Experimental results on zero-shot classification validate the effectiveness of the proposed prompting method. Further, fine-grained investigations are conducted on several factors, e.g., the prompt lengths, and prompt formats.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": [\"Strengths:\", \"Fine-grained analysis and investigations are conducted on the different factors of building good semantic prompts, which can offer readers with some practical inspirations.\", \"The performance gains on some datasets are good.\"], \"weaknesses\": [\"Limitations:\", \"Does not provide detailed qualitative examples of the proposed prompt contexts.\", \"The technical novelty of this paper is limited, considering a bunch of existing works on augmenting the CLIP textual prompts. Therefore, the technical differences between this work and other works on LLM-based prompt augmentation methods need to be clarified.\", \"Lack of baselines for comparisons: two important baselines, CHiLS and MPVR, are mentioned in the related work but are not compared. Besides, there is also a line of missing related works on prompt augmentation with semantic discriminativeness, e.g., S3A[1], Meta-Prompting[2], and LLM Explainer [3].\", \"Technical issues: (1) It is not guaranteed that the LLM-generated subcategories can satisfy the completeness and disjoint constraints of taxonomy stated in section 3.1. In practice, to what extent these two constraints can be satisfied and what implications will it have on the results needs further investigations and discussions. (2) How many layers of generated taxonomy will there be? If only generate a single layer of subcategories, this work would have high technical similarity with CHiLS; otherwise, more ablations are needed to investigate its benefits.\", \"Overclaim issue: (1) the motivation of the hierarchical taxonomic prompt starts from the game theory, however, the competition and players in the categorization contexts are not clearly defined. There is no clear and direct relationship between them. (2) The claimed semantic interpretability advantage mentioned in the abstract is not supported since there are neither interpretability results and comparisons nor qualitative examples.\", \"Presentation issue: The motivation figure 1 covers too large areas. Can add some subtitles to the paragraphs in section 6.3.\"], \"references\": \"[1] S3A: Towards Realistic Zero-Shot Classification via Self Structural Semantic Alignment\\n\\n[2] Meta-Prompting for Automating Zero-shot Visual Recognition with LLMs\\n\\n[3] LLMs as Visual Explainers: Advancing Image Classification with Evolving Visual Descriptions\", \"questions\": \"All my questions are listed in the weaknesses.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"General Response\", \"comment\": \"Thank you all for the time and effort putting into reviewing our submission and offering your improvements. We hope the findings of the work to be interesting and valuable to our understanding of zero-shot classification performance, but also the implications on the structure of the representation space of multimodal models like CLIP.\\n\\nWe look forward to your decision for this submission while wishing you all happy holidays.\"}", "{\"summary\": \"This study concentrates on zero-shot classification utilizing CLIP. To boost performance, it aggregates the class set into various clusters, termed taxonomies, which function as superclasses. By adapting the prompts to the format \\u201c{c} which is/has {d}, {g(C,T_c)}\\u201d, the model demonstrates enhanced accuracy compared to the simplistic template \\u201ca photo of {c}\\u201d. Positive results are shown in some experiments.\", \"soundness\": \"1\", \"presentation\": \"1\", \"contribution\": \"2\", \"strengths\": \"The work presents a simple and straightforward method that changes the prompts to a more carefully designed prompt using taxonomies.\", \"weaknesses\": \"1. Employing class hierarchy is not groundbreaking in few-shot classification. As highlighted in related work, both CHiLS and D-CLIP utilize hierarchy, along with other studies not cited here, such as [Ren. NIPS 2024].\\n* Ren Z, Su Y, Liu X. ChatGPT-powered hierarchical comparisons for image classification[J]. Advances in neural information processing systems, 2024, 36.\\n\\n2. This paper fails to elucidate crucial details, such as the process of clustering and g(C, T_c). In addition, in experimental evaluation, it appears to compare with lower baselines instead of fair benchmarks. \\n\\n3. The presentation is poor. Figure 1 occupies excessive space. Additionally, all equations lack equation numbers for reference.\", \"questions\": \"1. The paper fails to elucidate the method for clustering classes, and what's g(C, T_c) exactly. As illustrated in 181, the superset will be refined, how is it refined? In addition, the proposed method employs numerous hyperparameters, such as |c|/10, |C| < 20. How are the hyperparameters decided? The process appears quite handcrafted and lacks generalization.\\n\\n2. In experimental evaluation, this work achieved an accuracy of 68.03 on ImageNet using ViT B/16. As shown in [CLIP ICML 2021], its enhanced version with certain prompting variations can attain an even better result of 68.6. It suggests that the benefits might be obtained through additional prompting templates, such as \\\"an image of {}\\\", etc. More experiments are needed to determine whether the supplementary taxonomy can compensate for CLIP's prompting templates.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to DoJ7\", \"comment\": \"Thank you for taking the time to provide feedback and references to help improve this submission. We have made modifications in line with your suggestions where appropriate. We\\u2019ll aim here to give some extra context to our work and respond to some of your points directly.\\n\\n**W1/W2 - Hierarchy in Relation to Existing Literature** \\nThe distinction between DefNTaxS and the existing literature is in the method of implementation of the LLM-derived taxonomic subcategories, specifically their direct use in the text prompt used for classification. This approach incorporates hierarchical elements directly into the text embedding, instead of using them iteratively to reweight the model prior to classification (CHiLS, Novack et al., 2023) or creating text elements for classification separate from the class labels (ChatGPT-Powered\\u2026, Ren et al., 2023). These elements are often more prone to semantic noise, rely on non-interpretable clustering methods, and use modified scoring methods. DefNTaxS balances efficiency and accuracy of the LLM in creation and refinement of taxonomic subcategories, avoiding errors in generation as seen in D-CLIP descriptors (Menon et al., 2023, Roth et al., 2023), and requires minimal inference compute due to only modifying text, no image modification or further training. By incorporating both class-specific descriptors and taxonomic contexts into prompts, this approach increases the semantic granularity of zero-shot classification, improving interpretability, accuracy, and robustness to descriptor noise. It also gives insight into the taxonomic representations implicit in CLIP\\u2019s language structures through the effect of varying levels of visual and taxonomic granularity being more effective than either individually (Ablation 6.3). DefNTaxS maintains all these benefits while also achieving improved or equal performance across the majority of relevant benchmarks.\\n\\n**W3.1/Q1 - Examples of g(C,t_i) and clustering** \\nAn example of g(C,t_i) is provided in lines 235-241, including Eq. 6, with additional context added in Figure 2 (lines 162-179). The LLM-based taxonomy creation and assignment method, distinct from k-means or other clustering methods, is further explained in Figure 2 and its caption.\\n\\n**W3.2 - Benchmarks** \\nFor fair comparison, we have used the same benchmarks as in existing literature on text-augmented zero-shot classification and similar fields [Menon et al. 2023, Pratt et al., 2023, Roth et al. 2023, Ren et al., 2023, Li et al., 2024]. CuPL (Pratt et al., 2023) and WaffleCLIP + Concepts (Roth et al., 2023) have also been added as benchmarks. If other relevant benchmarks were preferred, we would have appreciated if they had been named and referenced.\\n\\n**W4 - Formatting Issues** \\nFigure 1 has been condensed. Equations have had reference numbers added. Boldface and underlining have been added to results tables to indicate performance.\\n\\n**Q1 - Hyperparameter selection** \\nSelection of the hyperparameters in question (minimum taxonomy count, maximum class count per taxonomy) was conducted through empirical selection. This result generalized to all datasets tested. Reduced performance resulting from changes to this approach is shown in Ablation 6.1. These hyperparameters are also not restrictively enforced; they are simply provided to the LLM to use as a guide, as in Appendix A.\\n\\n**Q1 - Superset (subcategory) refinement** \\nMore detailed explanation of the refinement method has been provided in Figure 2 and its caption. The LLM-based taxonomy creation and assignment method is simply repeated with adjusted hyperparameters if certain outcomes are not met.\\n\\n**Q2.1 - Varying Results for CLIP/E-CLIP** \\nIt is suggested that CLIP and E-CLIP achieve different results in the original OpenAI paper where it was introduced. Each paper referenced in this submission also reports different results for these values. For example, for the ViT-B/32 backbone, CLIP's performance on ImageNet ranges from 58.5 (D-CLIP, ChatGPT-Powered) to 63.4 (CuPL). As stated in Section 4.3 (Baselines), lines 294-297, we recreated all these benchmarks using the code provided in these works to ensure a fair comparison. We achieved a result of 58.9 on ImageNet using ViT-B/32, comfortably within this range, suggesting a fair comparison of results to these other baselines.\\n\\n**Q2.2 - Experiments to Combine CLIP Templates with DefNTaxS Approach** \\nExperiments were conducted combining elements of various baseline approaches with DefNTaxS, including D-CLIP descriptors, random characters, and dataset-level concepts from WaffleCLIP, and prompt templates from E-CLIP. None produced performance improvements or scientifically useful results, so were omitted in favor of more valuable inclusions. In specific response to this suggestion, using the E-CLIP prompting templates in combination with DefNTaxS reduced performance for all benchmarks except Oxford Pets, lines 331-338.\"}", "{\"summary\": \"The paper presents a novel method to enhance prompts for zero-shot classification tasks by leveraging a large language model (LLM) to identify taxonomic relationships between classes. By incorporating the taxonomy of a given class into the prompt, the method mitigates competition between semantically similar classes, leading to more accurate classifications. The proposed approach is rigorously evaluated against established baselines across multiple classification datasets.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": [\"The paper is clearly written and easy to follow, with a logical flow that makes the concepts accessible.\", \"Extensive experimental evaluations provide valuable insights into the method's performance and applicability.\", \"The proposed method is both simple and well-motivated, making it easy to implement and understand.\", \"The method demonstrates consistent improvements across the majority of evaluated datasets, showcasing the method's effectiveness.\"], \"weaknesses\": [\"The proposed method offers limited novelty compared to existing literature. Specifically, WaffleCLIP already introduces the idea of incorporating one high-level concept into the prompts. If my understanding is correct, the baseline used for comparison seems to only include the random characters/words, without accounting for these high-level concepts. It would be interesting to see how the proposed method compares to WaffleCLIP under these conditions.\", \"[Minor, subjective] Figure 1 could benefit from a layout adjustment, such as adopting a more landscape-oriented aspect ratio.\"], \"questions\": [\"CuPL is mentioned as a baseline in Section 4.2, but is never compared against. Could you explain why it was omitted from the experimental comparisons?\", \"The paper states in two instances that high intra-class similarity contributes to model confusion. My understanding is that high intra-class similarity means that images within the same class are visually alike, which should benefit classification. Could you clarify this point?\", \"In Section 5.1, it is claimed that the largest improvements occur on datasets with either high-class counts or high intra-class similarity. However, Table 1 shows that the most significant improvements are observed on Oxford Pets and EuroSAT, both of which have relatively few classes and exhibit high inter-class similarity. Did I miss something here, or could you clarify this discrepancy?\", \"It is claimed that appending taxonomy classes to prompts helps reduce inter-class competition and improves classification accuracy. However, since inter-class similarity typically occurs between classes within the same taxonomy, I would expect that the addition of these taxonomy-based prompts would not necessarily mitigate this issue. Could you provide more clarification on how the method addresses inter-class similarity?\", \"To gain insights into the above issue, a potential experiment could involve measuring the impact of DefNTaxS on the type of classification errors. Specifically, the experiment would assess how often false predictions occur within the defined taxonomy versus those occurring outside of it.\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The authors propose Defined Taxonomic Stratification (DefNTaxS), a novel method that leverages the power of large language models (LLMs) to enhance zero-shot image classification in vision-language models (VLMs) like CLIP. DefNTaxS introduces a hierarchical classification framework by generating subcategories that group semantically similar classes, reducing competition between classes with overlapping features. The authors demonstrate that DefNTaxS sometimes outperforms existing methods across various benchmark datasets, including ImageNet, CUB, Oxford Pets, DTD, Food101, Places365, and EuroSAT and model sizes. Furthermore, they perform some ablation studies highlighting the importance of taxonomic refinement, prompt structure, and incorporating richer descriptors for achieving performance.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"1\", \"strengths\": [\"The paper is clear, well-structured, and easy to follow.\", \"The proposed method demonstrates improved results across all CLIP sizes on certain datasets, such as Oxford Pets and EuroSAT.\"], \"weaknesses\": [\"The authors mention WaffleCLIP in the introduction but largely overlook its core finding: that adding additional words around the class name in prompts for CLIP has minimal effect, and previous works ([1],[2]) showing improvements may not provide meaningful benefits.\", \"The main results (Table 1) indicate some improvements on specific datasets and CLIP model sizes; however, these gains are generally minor and may be due to hyperparameter optimization rather than the method itself. Additionally, the proposed method underperforms compared to others on a substantial portion of the datasets.\", \"If I understand correctly, Table 2 shows that the method often does not outperform other approaches. For model sizes B/16 and L-14, D-CLIP achieves slightly better results, but the differences are negligible.\", \"Minor typos: Line 40 - \\\"also\\\" is split; Figure 1 - \\\"visualisation\\\" should be \\\"visualization\\\"; Line 47 - \\\"labelsWhile\\\".\"], \"questions\": \"-\\n\\n\\n[1] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. ICLR, 2023.\\n[2] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"metareview\": \"This paper addresses the task of zero-shot classification by leveraging a large language model (LLM) to identify taxonomic relationships between classes. The review appreciated the method\\u2019s simpleness, tut they also raised important concerns including insufficient comparison, limited novelty, overclaim issue,etc. The authors did provide additional results in the response, but more experimentation is required to fully justify the framework. Overall, the drawbacks outweigh the benefits of the paper.\", \"additional_comments_on_reviewer_discussion\": \"During rebuttal, the concerns raised by reviewers mainly include limited novelty, unclear experimental details, poor presentation, moderate performance improvement, overclaims etc. The authors provided rebuttals relatively late, and the reviewers didn\\u2019t participate in the discussion. Only reviewer ewH2 provided his final rating. Although some results and explanation are provided in the rebuttal, the novelty of this paper is limited and the performance improvement is moderate.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}" ] }

[ "Chengming Xu", "Kai Hu", "Donghao Luo", "Jiangning Zhang", "Qilin Wang", "Xiaobin Hu", "Yanwei Fu", "Chengjie Wang" ]

[ "image stylization", "diffusion model" ]

{ "note_id": [ "cksu4fkOey", "bKTD93pwAo", "ZmIV0FW9jJ", "VPghfsGoUH", "RMlKLv5AOz" ], "note_type": [ "official_review", "official_review", "official_review", "official_review", "comment" ], "note_created": [ 1730459761061, 1730542468644, 1730541771172, 1730307136699, 1731643838237 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission2096/Reviewer_JaeZ" ], [ "ICLR.cc/2025/Conference/Submission2096/Reviewer_PMjk" ], [ "ICLR.cc/2025/Conference/Submission2096/Reviewer_75nS" ], [ "ICLR.cc/2025/Conference/Submission2096/Reviewer_qUmh" ], [ "ICLR.cc/2025/Conference/Submission2096/Authors" ] ], "structured_content_str": [ "{\"summary\": \"This paper introduces StyleMaster to improve the style misinterpretation and semantic inconsistency in previous t2i generation frameworks. With proposed modules StyleMaster effectively combines style flexibility with text-driven semantic integrity.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"3\", \"strengths\": \"StyleMaster shows improved style interpretation and semantic consistency compared to baseline models. I believe that this is valuable for users needing customizable style control in t2i generation.\", \"weaknesses\": \"1. The manuscript says, the use of the CLIP visual encoder for style embeddings emphasizes local patterns, but elaborating on this aspect could improve the manuscript. Specifically, (1) it\\u2019s unclear why local-only embeddings might be insufficient to replicate style, as I believe that local elements (e.g., brushstrokes) often receive primary focus from artists or viewers when evaluating specific artistic styles. (2) What if the CLIP visual encoder are used to model global features by embedding the entire image as a whole? I am not fully convinced of the need for an additional VGG network and Gram, though they are frequently used in style transfer. Is CLIP not sufficient for embedding global features?\\n\\n2. The paper lacks a clear explanation and discussion on the necessity of semantic-aware features with BLIP in this model. Additionally, while this approach resembles the method in DreamStyler, the benefits or potential issues of omitting semantic-aware features remain unclear. Further discussion of how these features enhance overall performance, along with a comparison to DreamStyler, would clarify their contribution.\\n\\t\\n3. In model comparisons, there is a mix of models utilizing test-time optimization and those using pre-training approaches. Differentiating between these categories would improve clarity. Additionally, the N-shot experimental setup and procedure are somewhat ambiguous. I assume the initial (pre-trained) model is trained using the settings described in Section 5.1, but in the N-shot experiment, does further optimization occur using the N reference samples? If so, what approach is used?\", \"questions\": \"Please see the weaknesses.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper proposes to extract descriptive representations of style from both image and text, which are leveraged to steer the diffusion models via dynamic attention adaptation. Two additional losses (style disentagle loss and style consistency loss) are adopted to enhance the training.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"3\", \"strengths\": \"1.Dynamic attention adaptation is exploited to generate additional weights of the attention layers in the diffusion model for each input style image.\\n\\n2.Ablation study is sufficient.\", \"weaknesses\": \"1.The input style images should be put into Fig.1, 4, 5 and 6 for side-by-side comparsions.\\n\\n2.The references of 2024 are missing in the part \\u201cStylized image generation\\u201d of Sec.2. For example, [A-D] ...\\n\\n[A] Style injection in diffusion: A training-free approach for adapting large-scale diffusion models for style transfer, CVPR'24\\n\\n[B] ArtBank: Artistic Style Transfer with Pre-trained Diffusion Model and Implicit Style Prompt Bank, AAAI'24.\\n\\n[C] DEADiff: An Efficient Stylization Diffusion Model with Disentangled Representations, CVPR'24\\n\\n[D] ArtAdapter: Text-to-Image Style Transfer using Multi-Level Style Encoder and Explicit Adaptation, CVPR'24.\\n\\n3.The details of style destroy are missing. How the input style image is processed in style destroy to derive the output?\\n\\n4.The differences between results generated with or without style/disen loss are minor in Fig. 6 and Table.5.\", \"questions\": \"1.Why different styles are considered in one-shot (Fig.4) and multi-shot (Fig.5)? Does it mean that multi-shot inference cannot directly improve the one-shot one? Are there any trade-offs between one-shot and multi-shot settings?\\n\\n2.Why the Text Sim of the run w/o disen loss (0.282) is higher than the run w/o style loss (0.280) in Table.5? In my opinion, the model trained w/o disen loss should be expected to entangle the style and the content, which leads to semantic leakage from the input style image to the generated image with a different prompt and thus lower Text Sim.\\n\\n3.Will the training code be released in the future?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The paper presents StyleMaster, a novel framework for Stylized Text-to-Image Generation (STIG), aiming to address issues in previous methods like misinterpreted style and inconsistent semantics. It proposes a multi-source style embedder to extract comprehensive style embeddings and a dynamic attention adapter for better style injection. The model is trained with additional objectives to enhance semantic and style consistency. Experiments show its superiority over existing methods in generating stylized images with correct styles and semantic fidelity.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. The method is easy to follow. StyleMaster introduces a modular approach with each component focusing on a specific enhancement (e.g., style capture or semantic preservation), which effectively addresses common issues in stylized text-to-image generation.\\n\\n2. The model consistently outperforms baselines in quantitative evaluations, including both objective measures (text and style similarity) and subjective user studies, showing its capability to produce high-quality stylized images.\", \"weaknesses\": \"1. The paper introduces several complex loss functions (e.g., Gram consistency loss and semantic disentanglement loss) and additional model components (e.g., multi-source embedding, dynamic adapter). While these designs are effective in improving performance metrics, they resemble \\\"patches\\\" added to the existing model, contributing to engineering complexity. The paper lacks an in-depth theoretical analysis to explain the interrelationships between these modules, relying instead on enhancing specific metrics with pre-existing components.\\n2. The multi-source style embedder and dynamic attention adapter largely extend existing techniques rather than introducing fundamentally new theoretical contributions. While effective in real-world applications, this approach may be perceived as lacking in novelty from a research perspective.\\n3. StyleMaster currently only supports image-based style conditioning. Extending this to other forms (e.g., text, video, or 3D data) would improve its versatility for broader stylization applications.\\n4. Due to the multi-source style embedding extraction, which involves patch-level transformers, the model is computationally intensive. This may limit its practicality in real-time or resource-constrained environments, particularly when handling larger or multiple style reference images.\", \"questions\": \"See Weakness\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The paper studied stylized text-to-image generation and improved StyleAdapter by addressing the issues like misinterpreted style and inconsistent semantics. Specifically, the paper proposed 1) a multi-source scheme for style extraction, which incorporates global-level VGG descriptors and semantic text embeddings from image captions into style representations besides CLIP-based patch embedding; and 2) dynamic attention adapter layers to adjust the weights of the style embeddings to the self-attention and cross-attention layers in the diffusion UNet. The proposed method compares favorably with some recent SIG methods in the experiments.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"This is a descent work on stylized text-to-image generation. Some technical approaches are interesting and may be inspiring to the community, like the dynamic cross attention adapter and multiple features to derive the style representations.\", \"weaknesses\": \"The major problem of this paper is that the claims are not well justified. There is some gap between the motivation of addressing the issues in StyleAdapter and the proposed technical approach.\\n\\nBoth CLIP features and VGG descriptors are trained or designed to conduct image/object recognition. So their abilities to describe and represent complicated or fine-grained image styles are in question, which lead to the questions: 1) If CLIP features are aligned with image captions and not good to describe image styles, VGG descriptors may also suffer from the similar issues; 2) Is image style more related to local image patterns or global content? More in-depth discussions on how to define and describe \\\\emph{style} are required to justify the advantage of using VGG descriptors.\\n\\nConceptionally, if a method uses feature from more sources or removes some unrelated/negative tokens, it \\u201cmay\\u201d improve the image generalization, but not necessarily. So the paper needs to explain and justify the proposed method can achieve this. For example, the paper needs to first justify \\u201cthe global information can guide the model to better concentrate on style-related knowledge\\u201d, which is an assumption and shall not be used as a principle to derive the conclusion that so our approach can yield \\u201cbetter\\u201d results. Actually, the paper uses \\u201cbetter\\u201d many times where there are some logic gaps in those statements. \\n\\nThe embeddings of Z_{clip} and text embeddings Z_{cap} are no longer aligned after respective self attention operations. So it may not be that impactful to subtract/remove some unrelated/negative tokens as the paper claims. \\n\\nThe experiments show that the improvements are not consistent. There are 20 styles in the experiments, thus it is not clear how the proposed method performs for fine-grained style control.\", \"questions\": \"Please discuss \\u201cstyles\\u201d are related to global image contents or local image characteristics.\\n\\nPlease explain more about why VGG descriptors that are trained for image recognition are good to describe image styles and do not suffer from the same issues as CLIP features.\\n\\nPlease discuss if the proposed can extend to handle hundreds of image styles.\\n\\nSome missing related references, please compare and discuss:\\n\\nMeasuring Style Similarity in Diffusion Models, 2024.\", \"styletokenizer\": \"Defining Image Style by a Single Instance for Controlling Diffusion Models, ECCV 2024.\", \"clap\": \"Isolating Content from Style through Contrastive Learning with Augmented Prompts, ECCV, 2024.\", \"typo\": \"ll.184, \\\"are achieves\\\".\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"withdrawal_confirmation\": \"I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors.\"}" ] }

[ "Farhan Damani", "David H Brookes", "Theodore Sternlieb", "Cameron Webster", "Stephen Malina", "Rishi Jajoo", "Kathy Lin", "Sam Sinai" ]

[ "protein engineering", "sequence design", "model-based optimization" ]

{ "note_id": [ "rGNxHh8anf", "pzIRbMUCrt", "pUiBTmLkxH", "p6dWQ14Xsz", "mDgx9id5Lq", "jF3WE2UEGH", "ivQNR9MHJi", "hz7XzSqdGQ", "YVHf6UQiIZ", "Qno6ThgrNV", "DXvT91pSAK", "CrUWCGq2ys", "7E6ScQw6FP", "5QK2umAu5Z" ], "note_type": [ "official_review", "official_comment", "decision", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_review", "meta_review", "official_comment" ], "note_created": [ 1730697511672, 1732319108019, 1737524105491, 1732319000997, 1731152469651, 1732318912536, 1732319066744, 1732537229168, 1732318902528, 1730727863382, 1732319076291, 1730673684259, 1734764381538, 1732552772936 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission11130/Reviewer_okPN" ], [ "ICLR.cc/2025/Conference/Submission11130/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission11130/Authors" ], [ "ICLR.cc/2025/Conference/Submission11130/Reviewer_eBHJ" ], [ "ICLR.cc/2025/Conference/Submission11130/Authors" ], [ "ICLR.cc/2025/Conference/Submission11130/Authors" ], [ "ICLR.cc/2025/Conference/Submission11130/Reviewer_6qeD" ], [ "ICLR.cc/2025/Conference/Submission11130/Authors" ], [ "ICLR.cc/2025/Conference/Submission11130/Reviewer_6qeD" ], [ "ICLR.cc/2025/Conference/Submission11130/Authors" ], [ "ICLR.cc/2025/Conference/Submission11130/Reviewer_9umf" ], [ "ICLR.cc/2025/Conference/Submission11130/Area_Chair_X7yx" ], [ "ICLR.cc/2025/Conference/Submission11130/Reviewer_okPN" ] ], "structured_content_str": [ "{\"summary\": \"This paper introduces a binary classification model for detecting out-of-distribution (OOD) samples in the context of offline model-based optimization (MBO). The classification model is trained on a given offline dataset (labeled 0, in-distribution) and a generated dataset (labeled 1, OOD), where the generation algorithm is task-dependent. The learned classification model is used to calculate a score indicating the intensity of the distribution shift. Experiments were conducted in a synthetic problem, a simulated protein structure design, and a real-world Adeno-Associated Virus (AAV) capsid sequence design.\", \"soundness\": \"1\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"1. The proposed algorithm is straightforward and also easy to implement.\\n2. The paper was generally easy to understand and follow.\\n3. The method achieved better results in OOD detection than the uncertainty-based method (deep-ensemble).\", \"weaknesses\": \"1. I'm not fully convinced by the proposed method. The paper says in line 921 that $1/p_{tr}(x)$ (where $p_{tr}$ is training distribution) is more suitable for detecting distribution shift. If so, there are several ways to achieve this, e.g., kernel density estimation (KDE) or neural autoregressive density estimation (NADE; [1]). A careful conceptual and experimental comparison with these density estimation methods seems crucial.\\n2. I'm uncertain how the proposed method \\\"can significantly improve design quality\\\" (line 27). Choosing the right threshold seems critical to effectively balance the exploitation of the surrogate model and the OOD robustness. Moreover, I suspect that the optimal threshold for achieving the best design, regardless of whether a score-based or percentile-based method is used, will vary depending on the specific design task and the distribution of the training dataset. However, there has been limited discussion on threshold selection.\\n3. Similarly, there is no experimental evidence showing that the proposed algorithm can actually improve design quality. The experiments were only about the ability to detect OOD, specifically in comparison to deep ensembles.\\n4. The proposed algorithm relies on the MBO algorithm to generate the OOD dataset for classifier training. However, this paper only validates it using a single MBO algorithm, AdaLead, which raises concerns about the method\\u2019s versatility and generalizability.\\n5. A minor point, but the writing could be improved for better readability. For instance, it might be helpful to create a separate 'Preliminaries' section for the content in Sections 2.1 (offline MBO) and 2.2 (distribution shift), allowing the 'Method' section to focus solely on the main contribution. Additionally, the paper is somewhat verbose, particularly in the experiment section. A more concise presentation that highlights the main contributions and insights would strengthen the overall readability.\\n\\n[1] Uria, Benigno, et al. \\\"Neural autoregressive distribution estimation.\\\" JMLR (2016).\", \"questions\": \"1. (Related to weakness 1) Why should one use the proposed binary classification approach for OOD detection instead of simply approximating $p_{tr}(x)$ using, e.g., KDE or NADE?\\n2. Similarly, I seems feasible to use the minimum distance between a sample $x$ and samples in the in-distribution dataset $D$\\u2014often referred to as \\\"novelty\\\" [2] and easy to compute\\u2014as an OOD score. Have you considered this approach? If so, is there a specific reason why the proposed method (the learned binary classifier) might be more effective for OOD detection?\\n3. Could you explain why the proposed algorithm is called a \\\"meta-algorithm\\\"?\\n4. Recent works have attempted to inject structural biases into surrogate models to improve OOD generalization in offline MBO settings [3, 4]. It would be interesting to explore how these structurally-biased surrogate models could synergize with the proposed OOD detection method, potentially opening up future research directions.\\n\\n[2] Kim, Minsu, et al. \\\"Bootstrapped training of score-conditioned generator for offline design of biological sequences.\\\" NeurIPS (2024). \\n[3] Grudzien, Kuba, et al. \\\"Functional Graphical Models: Structure Enables Offline Data-Driven Optimization.\\\" AISTATS (2024). \\n[4] Grudzien, Kuba, et al. \\\"Cliqueformer: Model-Based Optimization with Structured Transformers.\\\" arXiv:2410.13106 (2024).\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"We thank the reviewer for their thoughtful and positive review. We respond to their questions below:\\n\\n> Have you explored strategies for using the OOD scores beyond binary accept/reject decisions? Could the scores be incorporated into the optimization objective to guide exploration?\\n\\nWe have not currently explored such ways of using the OOD scores, though we agree that this is a fruitful area for future research. One challenge is that the current version of the OOD classifier requires the set of designed sequences in order to be trained, and thus it could not be used during the process of optimization without modifications. We will highlight in the Discussion section that it would be an exciting area of research to determine a reasonable method for incorporating the OOD scores into the optimization procedure itself.\\n\\n> How sensitive is the method to surrogate model architecture and training procedure choice? Additional experiments testing robustness across architectures would be informative.\\n\\nWe assume the reviewer is referring to modifications to the architecture and training procedure of the OOD classifier, as the surrogate model used in the design procedure is outside of the scope of this paper. We performed a basic hyperparameter search on the OOD classifier architecture, using binary cross entropy on a validation set as the target property. We found that architectures with fewer parameters underfit the data (i.e. did not achieve as low a validation loss), while larger models did not meaningful improve performance. We will clarify in the Appendix that we performed this basic search and that this should be done for any new application (which is straightforward, as it follows the best practices for training any binary classifier)\\n\\n> Could you elaborate on potential approaches to leverage the OOD scores for active learning or model refinement when significant distribution shift is detected?\\n\\nTo clarify, the OOD scores are able to detect whether individual sequences are OOD, but are not able to distinguish between different degrees of distribution shift across the entire design set (i.e. they can not be used to say that one complete set of designed sequences is more or less OOD than another). If this is a goal, a distributional statistic such as Maximum Mean Discrepancy [1] may be more appropriate. Nonetheless, it is reasonable to ask whether the OOD scores can be used to correct for distribution shift or guide sampling in an active learning setting. Logistic regression of a similar type to the OOD classifier has been used to produce approximate importance weights in Importance Weighted Empirical Risk Minimization (IWERM) [2] to improve the performance of a regression model on the test set. It is not immediately clear how such a correction would be used in our setting given that our goal is not to improve regression performance. However, investigating how using the OOD classifier to modify the surrogate model in this way effects the prediction performance of the model is an intriguing possibility that we will mention in the Discussion section.\\n\\n[1] Gretton, A., Borgwardt, K. M., Rasch, M. J., Sch\\u00f6lkopf, B., & Smola, A. A kernel two-sample test. Journal of Machine Learning Research, 13 (2012).\\n[2] Shimodaira, H. Improving predictive inference under covariate shift by weighting the log-likelihood function. J. Stat. Plan. Inference 90, 227\\u2013244 (2000).\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"comment\": \"We thank the reviewer for their feedback and respond to specific comments below:\\n\\n> The novelty of this approach is limited: The idea of OOD classifier is not new, and the way of using the predicted OOD score is not particularly novel neither. OOD score is simply used to filtering out the sample with threshold or range.\\n\\nWe do not intend to claim that the use of a binary classifier to detect covariate shift is novel, and have cited a number of papers that use such methods. Our contributions are two fold: first, we demonstrate the extent to which feedback covariate shift (FCS) is present in a real-world sequence design problem and how it can impact the resulting design. Second, we demonstrate that an application of the OOD classifier to distinguish between the training and designed distribution can mitigate the effects of FCS. \\n\\nWe are committed to clarifying our contributions by modifying the writing of the paper in a number of ways. First, we will modify the introduction to emphasize the two contributions discussed above. Second, we will further emphasize the previous work that has used binary classifiers for covariate shift detection. Finally, we will move the derivation of the \\u201cdensity ratio trick\\u201d in Section 2.3 to the SI, as this section may be giving an unintended impression of novelty.\\n\\n> \\\"the complexity of using MBO algorithms correctly... limits the adoption among practitioners. Selecting a trust region for any search algorithm can be an art rather than a science and risks wasting experimental resources\\\" may be overstated. For example, proximal exploration (PEX) [1]...\\n\\nPEX is indeed a powerful method for designing biological sequences when one starts only with the wild type (WT) sequence and is able to make a number of experimental queries (i.e. in an online setting). In contrast, our paper is focused on the more common offline setting where one has an existing dataset of sequence-function measurements for sequences with a variety of edit distances to WT and is designing sequences for only one follow-up experiment. PEX primarily uses a constraint on the edit distance to WT to limit the distribution shift of the designed sequences. In the offline setting, this edit distance constraint will not take into account the distribution of the existing data and thus will be overly restrictive in the case where there are high edit distance sequences in the existing data (as in our case; see Figure 4 in Appendix A). PEX is thus not an applicable baseline to our problem. Nonetheless, we agree with the reviewer that the language in this sentence is too strong and we will soften it to recognize contributions to this field. \\n\\n> Including further discussion for other MBO approaches that consider distribution shift, such as RoMA [2] and BDI [3].\\n\\nWe thank the reviewer for providing these citations and we will add them to our discussion in the Related Work section. Notably, both of these methods require modifications to the surrogate models and design algorithm. In contrast, we discuss a method that can be used as a \\u201cdrop-in\\u201d with any design method, given only training and designed sequences. It can thus be more practical in many design scenarios. \\n\\n> From my understanding, the optimization process seems to be conducted iteratively. Does it means multiple query rounds like the setting in AdaLead?\\n\\nWe assume the reviewer is asking whether the optimization procedure is online or offline (i.e. that \\u201citeratively\\u201d refers to multiple rounds of experimental query). Our paper is focused on the offline setting, where the ground truth experiment cannot be queried during the design procedure. This is also the case for AdaLead, which is offline genetic that queries the **surrogate model** during iterative rounds of mutation and selection. We believe the misunderstanding in how AdaLead is used may arise from the PEX paper, where (we believe) the authors query the ground truth experiment during the genetic algorithm iterations, rather than the surrogate model. It is perfectly valid to use AdaLead in this way, but is not the way that it was intended to be used. \\n\\n> Regarding the AAV task, is the surrogate model the same as the one used in AdaLead?\\n\\nYes, the surrogate model used in AdaLead is the mean of the ensemble used to calculate ensemble uncertainties. We will clarify this in the Appendix. \\n\\n> How many models are used for the deep ensemble?\\n\\nWe used 10 models with independent initializations for the ensemble. We ensure that this detail is added the main text.\\n\\n> What is the meaning of \\\"50 bootstrap data samples\\\" in line 472?\\n\\nIn this case, the designed sequences are fixed since they were experimentally tested prior to the application of our selection method. In order to ensure that our regret calculations were not anomalous to the specific set of designed sequences, we resampled the designed sequences with replacement 50 times and calculated the regret curves for each resampling. We will clarify this procedure in the paper.\"}", "{\"summary\": \"This paper introduces a method to detect and correct for feedback covariate shift in experimental design (such as protein sequence design), where training data distribution differs from the distribution of the new data candidates generated throughout the process. The method is based on softmax regression for binary classification of the domain, which is employed to discriminate between the training data and the data generated within the feedback loop. The logit score (here OOD score) represents the intensity of the covariate shift. The authors empirically validate their approach on three use-cases: synthetic function, biological simulation, and real-world application of protein sequence design. The domain classifier equipped with the OOD score is able to identify and reduce the distribution shift in the design loop in a real-world application.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"1\", \"strengths\": \"The paper presents an application in biology of the method for addressing feedback covariate shift. It is indeed of crucial importance to be able to correct for this shift across many fields with the strong presence of automated experimental design, biology being only of them. The method is very simple to integrate into any existing experimental pipeline, and seems to be performing well in practice.\", \"weaknesses\": \"To the best of my knowledge, I believe this way of addressing covariate shift is not novel; the novelty might lie in the aspect of applying this method to *feedback* covariate shift (see for example, https://doi.org/10.7551/mitpress/9780262170055.003.0008, or https://dl.acm.org/doi/10.5555/1577069.1755858, and references therein). Domain classification by means of logistic regression (with or without importance weighting) is one of the widely known methods, hereby just applied to the special case of feedback covariate shift in experimental design. In feedback covariate shift, it is assumed that the distribution of points generated within the feedback loop depends also on the training distribution. The paper lacks a clearer presentation of its methodological contributions, and does not fully allow one to appreciate the usefulness of the method in practice. Comparison with other baselines (e.g., other methods for unsupervised domain adaptation) would further highlight the strengths and weaknesses of the method. The running time of the method was not investigated, i.e., how much of the optimization budget needs to be dedicated to detecting and correcting for shifts in a real-world application. This could be done by comparing the computational overhead of the proposed method to the baseline optimization approach in real-world scenarios. The figures throughout the paper are not necessarily self-explanatory. Furthermore, a more rigorous technical and mathematical notation is missing.\", \"questions\": \"Could you please clarify how your approach differs from or improves upon existing methods for addressing covariate shift, particularly in the context of feedback covariate shift in experimental design? It would help to highlight your exact contributions more clearly and position with greater care your approach within related work.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"> The running time of the method was not investigated, i.e., how much of the optimization budget needs to be dedicated to detecting and correcting for shifts in a real-world application\\n\\nOur paper is written in the context of offline model-based optimization (MBO). These types of problems are \\u201coffline\\u201d because the temporal and dollar cost of collecting experimental data is much higher than the cost of computational design. In these cases, the specific computational complexity of design methods is not particularly important. For example, in the AAV case, the in-vitro experiments can take months to perform whereas the design can be done in a matter of minutes or hours with basic computational resources. For this reason, we do not perform a rigorous run time analysis of the method. However, we will highlight that the method only requires training a dense MLP on one-hot encoded sequences which can be done in a matter of minutes on a single T4 GPU.\"}", "{\"comment\": \"We thank the reviewer for their careful review. Our response to specific comments are below:\\n\\n> If so, there are several ways to achieve this, e.g., kernel density estimation (KDE) or neural autoregressive density estimation (NADE; [1])\\n\\nDirect density estimation in high dimensional spaces such as those of biological sequences is notoriously difficult when one wants to use the numerical density values rather than sample from the learned density. Consider for example [1], where the authors find that OOD examples are often assigned higher model densities than in-distribution examples. To demonstrate these challenges, we trained a NADE model on the AAV training data and applied the log density estimates to the design selection tasks whose results are shown in Figure 3. We find that in this setting, NADE performs slightly better than the deep ensemble uncertainties, but worse than the OOD classifier (i.e. the Pearson correlation between packaging measurements and the NADE scores is -0.38 compared to -0.54 and -0.18 for the OOD classifier and deep ensemble uncertainty, respectively, and the minimum mean top 100 regret for the NADE scores is 0.65 compared to 0.18 and 1.02 for the OOD classifier and deep ensemble uncertainty, respectively)\\n\\nIn response to another reviewer\\u2019s questions, we have also tested a baseline where we fit an Isolation Forest model [2] to ESM2 embeddings of the AAV sequences. The Isolation Forest is an established anomaly detection method that is based on the intuition that anomalous examples require fewer partitions to isolate from the rest of the data than regular examples. We find that this method works better than both the NADE scores and deep ensemble uncertainties, but worse than the OOD classifier (i.e. the Pearson correlation between packaging measurements and the Isolation Forest scores is -0.5 and the minimum mean top 100 regret for the NADE scores is 0.42). We will discuss both of these results in the main text and add the complete results to the Appendix. \\n\\n[1] Nalisnick, E., Matsukawa, A., Teh, Y. W., Gorur, D. & Lakshminarayanan, B. Do Deep Generative Models Know What They Don\\u2019t Know? arXiv (2018) doi:10.48550/arxiv.1810.09136.\\n[2] Liu, F. T., Ting, K. M. & Zhou, Z.-H. Isolation Forest. 2008 Eighth IEEE Int. Conf. Data Min. 413\\u2013422 (2008) doi:10.1109/icdm.2008.17.\\n\\n> I'm uncertain how the proposed method \\\"can significantly improve design quality\\\" (line 27). Choosing the right threshold seems critical to effectively balance the exploitation of the surrogate model and the OOD robustness. \\n\\nThe results in Figure 3c demonstrate that selecting sequences based on OOD scores leads to lower regret than alternative strategies, which leads to the conclusion that the method can improve the quality of designs. Indeed, the choice of threshold affects this result; we suggest in the main text (line 228) that stratifying across thresholds is a reasonable strategy to resolve this problem. Notably, a threshold percentile around 10 performs well in both the results shown in Figure 3c and the Protein Structure Prediction experiment shown in Appendix C. \\n\\n> The proposed algorithm relies on the MBO algorithm to generate the OOD dataset for classifier training. However, this paper only validates it using a single MBO algorithm, AdaLead, which raises concerns about the method\\u2019s versatility and generalizability.\\n\\nThis is a consequence of performing real world biological experiments on our designed sequences. These experiments limit the number of sequences that can be tested and thus we chose to use one design algorithm to enable us to include sequences at each optimization iteration of AdaLead. To us, this is a reasonable tradeoff in order to perform real-world experiments.\\n\\n> A minor point, but the writing could be improved for better readability. For instance, it might be helpful to create a separate 'Preliminaries' section for the content in Sections 2.1 (offline MBO) and 2.2 (distribution shift), allowing the 'Method' section to focus solely on the main contribution\\n\\nWe thank the reviewer for these suggestions and will incorporate them into our updated manuscript. The change suggested by the reviewer will work with the changes we have committed to in response to other reviewers to strengthen the clarity of our contribution.\"}", "{\"comment\": [\"Thank you for your effort, and I apologize for the delayed response. I appreciate the detailed explanations you've provided. However, I still have some concerns I'd like to discuss:\", \"From what I understand, the OOD classifier is applied to distinguish between the training and designed distributions to enhance MBO methods. It filters out OOD samples, resulting in lower regret scores compared to the Ensemble method. My concern is whether this approach ultimately benefits the generation of more desirable samples (e.g., higher f) or more diverse and novel batched samples. Since these metrics are commonly reported in various studies, could you provide these results, at least for the simulated protein design?\", \"The manuscript can be revised during the discussion period. Any updates or clarifications you can provide would help me reconsider my review.\"]}", "{\"comment\": \"We thank the reviewer for the thoughtful and constructive feedback. We address specific comments below:\\n\\n> To the best of my knowledge, I believe this way of addressing covariate shift is not novel; the novelty might lie in the aspect of applying this method to feedback covariate shift\\n\\nIndeed, we do not intend to claim that using logistic regression to address covariate shift is novel and we cite a number of papers that use this technique. Our contributions are two fold: first, we demonstrate the extent to which feedback covariate shift (FCS) is present in a real-world sequence design problem and how it can impact the resulting design. Second, we show that we can mitigate the effects of FCS by training the logistic regression model to distinguish between the training and designed distributions and then using the log probability scores of the model to select sequences that are less likely to have been affected by FCS. Our paper is thus impactful for practitioners of sequences, as it highlights the dangers of FCS in this scenario and provides a straightforward technique for mitigating its worst effects. \\n\\nWe are committed to clarifying our contributions by modifying the writing of the paper in a number of ways. First, we will modify the introduction to emphasize the two contributions discussed above. Second, we will further emphasize the previous work that has used logistic regression for covariate shift detection and add additional citations, including those given by the reviewer. Finally, we will move the derivation of the \\u201cdensity ratio trick\\u201d in Section 2.3 to the SI, as this section may be giving an unintended impression of novelty.\\n\\n> Comparison with other baselines (e.g., other methods for unsupervised domain adaptation) would further highlight the strengths and weaknesses of the method.\\n\\nWe use the logistic regression method to remove designed sequences with unreliable predicted fitness due to FCS. We are therefore most concerned with the ability to detect whether individual points are distant from the training distribution and thus likely to be affected by FCS. This is a different problem then is typically approached with unsupervised domain adaption approaches, where one is typically concerned with detecting and adapting to distributional shifts between the training and test distributions. The most salient baselines to compare to the logistic regression method are instead methods for anomaly detection. One such method is the Isolation Forest [1], an established anomaly detection method that is based on the intuition that anomalous examples require fewer partitions to isolate from the rest of the data than regular examples. We implemented a version of the Isolation Forest for anomaly detection in sequences where the sequences are first embedded into a continuous space using the ESM2 protein language model with mean pool embedding and then an Isolation Forest is trained on the resulting embeddings. We applied this method to the AAV design problem and found that it performed better than ensemble uncertainties, but slightly worse than the OOD classifier (i.e. the Pearson correlation between packaging measurements and the Isolation Forest scores is -0.50 compared to -0.54 and -0.18 for the OOD classifier and deep ensemble uncertainty, respectively, and the minimum top 100 regret for the Isolation Forest scores is 0.42 compared to 0.18 and 1.02 for the OOD classifier and deep ensemble uncertainty, respectively). In response to another reviewer\\u2019s questions, we have also tested a baseline where we fit a Neural Autoregressive Density Estimation (NADE) [2] model and used the resulting density estimates as scores to select sequences. We find that this method again works better than the deep ensemble uncertainties, but worse than the Isolation Forest scores and OOD classifier scores (i.e. the Pearson correlation between packaging measurements and the NADE scores is -0.38 and the minimum mean top 100 regret for the NADE scores is 0.65). We will discuss both of these results in the main text and add the complete results to the Appendix. \\n\\n[1] Liu, F. T., Ting, K. M. & Zhou, Z.-H. Isolation Forest. 2008 Eighth IEEE Int. Conf. Data Min. 413\\u2013422 (2008) doi:10.1109/icdm.2008.17.\\n[2] Uria, Benigno, et al. \\\"Neural autoregressive distribution estimation.\\\" JMLR (2016).\"}", "{\"summary\": \"This work proposes an out-of-distribution (OOD) classifier to detect distribution shifts, guiding design selection to avoid adversarial results. The authors suggest multiple ways to guide or filter sequence generation based on the predictions of the OOD classifier. The proposed method is tested on three different tasks, including AAV sequence design, using two different search methods, AdaLead and beam search. The experimental results show that the proposed OOD classifier achieves lower regret scores compared to deep ensemble-based OOD detection.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"1\", \"strengths\": [\"I agree with the motivation that trained models can be unreliable, and handling distribution shifts is crucial in model-based optimization (MBO).\", \"The proposed method is straightforward and effective compared to deep ensembles\", \"Simplicity and ease of implementation\"], \"weaknesses\": [\"The novelty of this approach is limited: The idea of OOD classifier is not new, and the way of using the predicted OOD score is not particularly novel neither. OOD score is simply used to filtering out the sample with threshold or range.\", \"The claim that \\\"the complexity of using MBO algorithms correctly... limits the adoption among practitioners. Selecting a trust region for any search algorithm can be an art rather than a science and risks wasting experimental resources\\\" may be overstated. Several studies have focused on discovering new sequence designs while maintaining close distances to known designs (wild types). For example, proximal exploration (PEX) [1] has made significant progress by effectively balancing the enforcement of in-distribution constraints and exploration in a practical and scientific way. Though they assume multiple query rounds in their original setting, PEX gives competitive results with a single round, which is the same as MBO. Including further discussion for other MBO approaches that consider distribution shift, such as RoMA [2] and BDI [3].\", \"#### Minor comments\", \"Line 65: expansive \\u2192 expensive\", \"Line 71: In this work, propose (there is no subject)\", \"For me, meta-heuristic sounds improper in this context\", \"In Appendix E. Fig Xa, Fig Xb, Fig Xc\", \"#### References\", \"[1] Ren, Zhizhou, et al. \\\"Proximal exploration for model-guided protein sequence design.\\\" International Conference on Machine Learning. PMLR, 2022.\", \"[2] Yu, Sihyun, et al. \\\"Roma: Robust model adaptation for offline model-based optimization.\\\" Advances in Neural Information Processing Systems 34 (2021): 4619-4631.\", \"[3] Chen, Can, et al. \\\"Bidirectional learning for offline infinite-width model-based optimization.\\\" Advances in Neural Information Processing Systems 35 (2022): 29454-29467.\"], \"questions\": [\"From my understanding, the optimization process seems to be conducted iteratively. Does it means multiple query rounds like the setting in AdaLead? If not, please clearly state the difference with AdaLead setting. If yes, the motivation and approach might be improper (even though I agree with the claim that we should carefully handle the unreliable surrogate model for the adversarial samples, as mentioned above), as a key assumption in MBO is that we cannot make additional queries to the black-box function. Allowing additional queries can lead to significant differences in the methodologies used. For instance, we need to explore the unreliable region for the subsequent iterations rather than filtering out these samples.\", \"Regarding the AAV task, is the surrogate model the same as the one used in AdaLead? If not, I am concerned that the comparison between the OOD classifier and deep ensembles might not be entirely fair. I have checked Appendix A.2, but I am unsure whether the model capacity is sufficient to learn the fitness function of the AAV tasks and, consequently, whether deep ensemble-based OOD detection would be effective with an insufficiently trained surrogate model.\", \"How many models are used for the deep ensemble?\", \"What is the meaning of \\\"50 bootstrap data samples\\\" in line 472?\", \"Additionally, I am curious whether the proposed OOD classifier could benefit search methods that already enforce in-distribution samples, such as PEX.\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"> Similarly, I seems feasible to use the minimum distance between a sample x and samples in the in-distribution dataset D\\u2014often referred to as \\\"novelty\\\" [2] and easy to compute\\u2014as an OOD score. Have you considered this approach?\\n\\nWhile these novelty score can be reasonable for small datasets, their calculation requires N*M edit distance calculations, where N is the size of the training set and M is the number of designed sequences. In the AAV case, $N\\\\approx 50,000$ and $M\\\\approx 5,000$ and therefore calculation of novelty scores requires over 200 million edit distance calculations! Due to this unfavorable scaling, it is unfeasible to use these novelty scores as a baseline method. \\n\\n> Could you explain why the proposed algorithm is called a \\\"meta-algorithm\\\"?\\n\\nWith this terminology, we meant to refer to the fact that the method we test is not a design algorithm itself, but can be applied as a \\u201cplug in\\u201d to any other design method. We recognize that this may be confusing when compared to meta-learning, which would be more analogous to a method that chooses or creates its own design algorithm. We will thus remove this terminology in the updated manuscript. \\n\\n> Recent works have attempted to inject structural biases into surrogate models to improve OOD generalization in offline MBO settings [3, 4]. It would be interesting to explore how these structurally-biased surrogate models could synergize with the proposed OOD detection method, potentially opening up future research directions.\\n\\nWe agree that these are exciting avenues for future research, which we will mention in the Discussion section of the updated manuscript. Both of these methods require modifications to the design algorithm itself and therefore we can not test them for the current paper without re-running the in-vitro experiments.\"}", "{\"summary\": \"This paper presents a method for detecting distribution shifts in machine learning-guided biological sequence design design, specifically addressing model-based optimization (MBO) prediction reliability when exploring regions distant from training data, identifying distribution shift when it occurs. The work introduces:\\n\\n1. A binary classifier approach to detect out-of-distribution samples in MBO\\n2. Empirical validation through AAV capsid engineering experiments\\n3. Comparison between simulation benchmarks and real-world distribution shift severity\\n4. A framework for identifying unreliable predictions during sequence optimization\", \"soundness\": \"4\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. Validation through wet-lab experiments, extending beyond simulation-based evaluation (which itself was also thorough)\\n2. Straightforward implementation of the proposed method\\n3. Demonstration that simulation benchmarks may not capture real-world distribution shift challenges\\n4. Comprehensive ablation studies and baseline comparisons\\n5. Technical foundation in density ratio estimation literature\\n6. Direct applicability to real-world MBO applications in biological sequence design\", \"weaknesses\": \"1. Limited analysis of predictor architecture and training choices' effects on distribution shift detection. Understanding the method's robustness across different model choices would strengthen the results.\\n2. While the method effectively identifies OOD samples, the paper provides limited guidance on what to do with these flagged sequences beyond excluding them. The practical impact would be enhanced by discussing mitigation strategies such as active learning, model retraining, or ways to incorporate OOD scores into exploration.\\n3. The theoretical foundations could benefit from deeper analysis, particularly regarding how classifier architecture affects density ratio estimation accuracy and potential bounds on detection performance under various distribution shift scenarios.\", \"questions\": \"1. Have you explored strategies for using the OOD scores beyond binary accept/reject decisions? Could the scores be incorporated into the optimization objective to guide exploration?\\n2. How sensitive is the method to surrogate model architecture and training procedure choice? Additional experiments testing robustness across architectures would be informative.\\n3. Could you elaborate on potential approaches to leverage the OOD scores for active learning or model refinement when significant distribution shift is detected?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"metareview\": \"The paper introduces a binary classification-based meta-heuristic to address distribution shift in model-based optimization (MBO), with applications to biological sequence design. While the idea is relevant and straightforward to implement, reviewers highlighted key weaknesses, including limited novelty, insufficient experimental validation across diverse MBO frameworks, and inadequate exploration of threshold selection and generalizability. The paper also lacked a clear demonstration of how the method improves design quality beyond OOD detection and was limited to testing with a single MBO algorithm, raising concerns about versatility. Although the authors provided clarifications and additional comparisons to alternative methods, these responses did not fully address the primary concerns. Consequently, the paper does not meet the standards for acceptance due to its limited scope, lack of rigorous novelty, and insufficient validation.\", \"additional_comments_on_reviewer_discussion\": \"During the rebuttal period, reviewers raised concerns about the paper's novelty, limited experimental scope, and unclear contributions.\\n\\nReviewer *eBHJ* questioned the novelty of using a binary classifier for OOD detection and requested comparisons with baseline methods for unsupervised domain adaptation. The authors clarified their contributions and added comparisons to Isolation Forest and Neural Autoregressive Density Estimation, but these additions highlighted only marginal improvements. \\n\\nReviewer *6qeD* critiqued the lack of task diversity in experiments and the reliance on a single MBO algorithm, while the authors argued this was due to constraints from real-world biological experiments. \\n\\nReviewer *okPN* questioned the practicality of threshold selection and the lack of generalizability testing, to which the authors proposed stratified thresholding and noted the need for future research. Reviewer 9umf acknowledged the method\\u2019s straightforward implementation but requested guidance on leveraging OOD scores beyond binary filtering, which was only partially addressed. \\n\\nOverall, while the rebuttal clarified some points and improved the manuscript, the core concerns about limited validation, novelty, and generalizability remained unresolved, leading to the decision to reject.\"}", "{\"comment\": \"I apologize for the late response and appreciate your effort on the rebuttal. Overall, I'm pleased with the authors' responses, but I still have some concerns.\\n\\n> This is a consequence of performing real world biological experiments on our designed sequences. These experiments limit the number of sequences that can be tested and thus we chose to use one design algorithm to enable us to include sequences at each optimization iteration of AdaLead. To us, this is a reasonable tradeoff in order to perform real-world experiments.\\n\\nI believe you could have conducted tests in the \\\"2D Toy Model (4.1)\\\" and the \\\"Simulated Protein Structure Design (4.2)\\\".\\n\\n> we suggest in the main text (line 228) that stratifying across thresholds is a reasonable strategy to resolve this problem.\\n\\nI'm not convinced that this suggestion fully addresses the issue of setting the threshold (since the suggestions still need humans to set the threshold).\\n\\n> Notably, a threshold percentile around 10 performs well in both the results shown in Figure 3c and the Protein Structure Prediction experiment shown in Appendix C.\\n\\nI believe the threshold should depend not only on the task but also on the dataset. For example, when we have a large amount of data and an accurate surrogate model, we can be more aggressive; conversely, with a small dataset and an inaccurate model, we should be more conservative.\\n\\n> W5\\n1. ***Although this is still a minor issue***, I think placing the subsections \\\"2.1. Offline Model-Based Optimization\\\" and \\\"2.2. Distribution Shift in Design\\\" under the \\\"2. Method\\\" section could be misleading. Since these sections introduce previous studies and outline the problems, it might be better to position them outside of the \\\"Method\\\" section, where your contribution should be emphasized.\\n\\n2. The verbosity has not been improved.\"}" ] }

[ "Zijie Pan", "Yushan Jiang", "Dongjin Song", "Sahil Garg", "Kashif Rasul", "Anderson Schneider", "Yuriy Nevmyvaka" ]

[ "Continual Learning", "Multivariate Time Series Forecasting" ]

{ "note_id": [ "tV7zWxSC0w", "t9aPHwRVYq", "r0MJkzeYZu", "opW7Iu7ib4", "lARYaZrKtI", "Xq4erA3lw1", "WW27lR7AMj", "W0aLpl3IuR", "TkM4zviHmZ", "MrDmG5tvb7", "IKXQbsTUsR", "HXeGOS2r4b", "EOttYudsCY", "DedadEnxYQ", "7WOVe01nMt", "0wDCfGFtr8" ], "note_type": [ "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "decision", "official_comment", "official_review", "official_comment", "official_review", "official_comment", "official_review", "meta_review" ], "note_created": [ 1732419617333, 1732522712670, 1730713737478, 1732720007245, 1732419235760, 1730628660264, 1732419883613, 1732512446014, 1737524050234, 1732420058755, 1730881564776, 1732418954798, 1730688494881, 1732638679274, 1730686371763, 1734331103093 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission10391/Authors" ], [ "ICLR.cc/2025/Conference/Submission10391/Reviewer_8F4W" ], [ "ICLR.cc/2025/Conference/Submission10391/Reviewer_8F4W" ], [ "ICLR.cc/2025/Conference/Submission10391/Reviewer_ifJg" ], [ "ICLR.cc/2025/Conference/Submission10391/Authors" ], [ "ICLR.cc/2025/Conference/Submission10391/Reviewer_ZKrh" ], [ "ICLR.cc/2025/Conference/Submission10391/Authors" ], [ "ICLR.cc/2025/Conference/Submission10391/Reviewer_gtJT" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission10391/Authors" ], [ "ICLR.cc/2025/Conference/Submission10391/Reviewer_ifJg" ], [ "ICLR.cc/2025/Conference/Submission10391/Authors" ], [ "ICLR.cc/2025/Conference/Submission10391/Reviewer_1id3" ], [ "ICLR.cc/2025/Conference/Submission10391/Reviewer_1id3" ], [ "ICLR.cc/2025/Conference/Submission10391/Reviewer_gtJT" ], [ "ICLR.cc/2025/Conference/Submission10391/Area_Chair_g7xV" ] ], "structured_content_str": [ "{\"title\": \"Response to Reviewer 1id3\", \"comment\": \"**W1: Research topic.**\\n\\nWe have demonstrated the importance of addressing catastrophic forgetting when time series are sequentially collected (please check figure 1 and our descriptions, and the real-world example in lines 053-061). That being said, our major contributions are about proposing a unique perspective to effectively solve continual multivariate time series forecasting. \\n\\n**W2: Our method differs from exisiting works .**\\n\\nThe existing literature either learns the structure via training parameters or representations. Our novelty lies in the role of dynamic structure learning for regime characterization, rather than an individual structure learning component. It is important that we impose the consistency regularization in the dynamic structure learning process, which aligns the learned structure with the universal and task-irrelevant structural knowledge to characterize a specific regime. The joint structure modeling based on both components leads to an capability that the existing backbones cannot achieve, i.e., **during the inference stage, our model can automatically infer a consistent structure solely based on the time series data, without knowing the regime and accessing the memory buffer**, as shown in **Figure 1: SKI-CL: Testing and validated in Figure 4 in the experiments**. \\n\\n**W3: Analysis of representation-matching memory replay scheme .**\\n\\nThanks a lot for the comments. \\n\\n1) We have introduced the experience-replay methods in the experiment section from lines 354-360: A herding method randomly replays training samples, a DER++ method enforces a L2 knowledge distillation loss on the previous logits, a MIR method selects samples with highest loss for experience replay. \\n\\n 2) Intuitively, our representation-matching replay scheme is efficient in terms of selecting samples after the most diverse partitions, instead of the whole training set, which minimizes the search space. We will provide an efficiency analysis in our future manuscripts. \\n\\n(3) We agree with the reviewer that visualizing the selected samples can be helpful to demonstrate our idea. We have provided the visualization of learned dependencies structures, continual learning performance, and a case-study with time series visualizations, and we will provide the visualization of selected samples once we update the manuscript. \\n\\n\\n**W4: Inference process.**\\n\\nOur training and inference process follows the continual learning paradigm, where the model is sequentially trained and evaluated for each regime, and the pipeline has been demonstrated in figure 1. We thank the reviewer for the advice and will consider further elaboration. \\n\\n**W5: Experiments.**\\n\\n(1) We have included the most recent state-of-the-art time series forecasting model, iTransformer that explicitly captures the variable dependencies of time series for forecasting purposes. We will consider adding more baselines. Currently continual learning tailored for time series is still underexplored, and we\\u2019ve implemented the most well-known and effective methods that serve as strong baselines in the continual learning area. \\n\\n(2) The basic intention of continual learning is maintaining performance across different tasks. We intend to demonstrate the effectiveness of our SKI-CL in terms of different continual learning scenarios regarding the time series dependencies, including the regimes representing different underlying structures. We have introduced the data splits in the training details in line 1013: The data split is 6/2/2 for training/validation/testing. \\n\\n\\n**W7: Clarification of the presentation.**\\n\\nDue to the space limitations of the main text, we have indicated that we provided the visualizations of regimes 3 and 4 for case-study. As we have mentioned in line 513, the full case study visualizations of all regimes are provided in Appendix G. The bold and underlines results mean the best and second-best performance respectively. For equation 3, $n_K$ the number of samples for K-th mode after K-partition, which is jointly defined with K. That being said, K and $n_1, \\\\cdots, n_K$ represent the sample partition, which are the parameters that optimize our objective. \\n\\n**W8: Code**\\n\\nPlease check the supplemental materials where we have attached our code.\"}", "{\"comment\": \"Thank you for the author's response, which has addressed some of my concerns. However, I still believe that the contribution of the paper to the research field is limited, both in terms of methodology and results. Overall, after considering the author's reply and the feedback from other reviewers, I feel that this paper is not yet ready for acceptance at ICLR, and therefore I will maintain my current score.\"}", "{\"summary\": \"Existing MTS models suffer from the forgetting problem in the continuous learning scenario, which makes it difficult to remember the variable dependencies from each regime. To solve this problem, the authors propose a structural knowledge informed continuous learning framework to infer the dependency structure between different regimes. In this framework, the authors propose to use a graph structure to explicitly represent the dependencies between regimes, and introduce a regularization to facilitate continuous learning. In addition, in order to alleviate the forgetting problem, the authors propose a new memory replay method, which effectively preserves the temporal dynamics and dependency structure of historical regimes by selecting MTS data samples that satisfy the maximum temporal coverage. Finally, the authors verify the superior performance of the proposed method through a large number of experiments.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"3\", \"strengths\": \"1. Modeling dependencies between states in a graph is novel.\\n2. When describing the method, a large number of model diagrams and structure diagrams are used to help readers understand the method.\\n3. The authors compared a large number of baselines, and the experimental workload was large.\", \"weaknesses\": \"1. On the one hand, the authors introduce a graph structure to model dependencies. On the other hand, the authors propose a new replay method to solve the forgetting problem in continuous learning. In my opinion, these are two points, but it is inappropriate for the author to mix the two points together in the abstract and introduction.\\n2. This paper did not conduct ablation experiments and did not demonstrate the significance of each part of the method.\\n3. The performance on the four datasets is not much better than the baselines.\", \"questions\": \"1. How well does the model resist forgetting after many new regimes coming\\n2. How to understand \\\"We emphasize that we don't intend to use structural knowledge as a ground truth\\\", but isn't structural knowledge still used as a label in the loss of $L_G$?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"I thank the authors for their responses. Most of my concerns are addressed, despite of the open questions. Therefore, I remain my comment on this paper.\"}", "{\"title\": \"Response to Reviewer 8F4W\", \"comment\": \"**W1: Mixing graph structure and continuous learning in the abstract and introduction.**\\n\\nThank you for your valuable review comments. We would like to emphasize that continual learning for time series with variable dependencies is a critical challenge. Our contribution lies in proposing a novel graph model to effectively leverage structural information in this continual learning setting. Additionally, we introduce a more efficient memory sample selection method for replay, aimed at mitigating the forgetting problem in continual learning. \\n\\n**W2: Ablation study.**\\n\\nThank you for your comments. We include the ablation experiments (different predicting horizon setting) in Appendix Section F and Hyperparameter analysis in Section 4.5 and Appendix Section E. \\n\\n**W3: Performance compared to baselines.**\\n\\nThank you for your comments. Our method achieves first or second place consistently across the datasets compared to baseline methods. This demonstrates the robustness and generalizability of our approach across diverse scenarios. Such consistent high-ranking performance highlights the effectiveness of our proposed method in addressing the challenges of continual learning for time series with variable dependencies. \\n\\n**Q1: Evaluation metrics.**\\n\\nThank you for raising this point. The metrics we report, AP (Average Precision) and AF (Average Forgetting), are widely recognized and commonly used to evaluate a model's resistance to forgetting in continual learning settings. For clarity, we have provided detailed definitions of these metrics in Section 4.1. \\n\\n**Q2: Question regarding structure knowledge.**\\n\\nThank you for raising this question. First, we incorporate a forecasting objective to guide the learning of the graph structure, ensuring it aligns with the continual learning setting. Second, unlike directly applying structural knowledge for message passing\\u2014an approach that, as shown in Tables 7 and 8, does not yield satisfactory results (e.g., STGCN)\\u2014we instead leverage structural knowledge to guide the model in identifying and adapting across regimes. This strategic use of structural information effectively reduces performance degradation and enhances the model's robustness in dynamic environments.\"}", "{\"summary\": \"This paper proposes a continuous multivariate time series prediction framework based on structural knowledge\\uff0c which utilizes structural knowledge to enhance the prediction of multivariate time series in a continuous learning environment. The proposed framework incorporates a deep prediction model that combines a graph learner to capture variable dependencies and a regularization scheme to ensure the consistency between the learned variable dependencies and the structural knowledge. The authors tackle the challenge of modeling variable dependencies across different regimes while maintaining the prediction performance. Experimental results on several real datasets are presented, demonstrating the effectiveness of the proposed framework in improving the prediction performance and maintaining consistency with the structural knowledge.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. This paper propose a new freamwork which is aimed at knowledge transfer learning.\\n\\n2. The proposed model can well use the knowledge of former tasks. \\n\\n3. The paper is well written.\", \"weaknesses\": \"1. The complexity of the model increases so much relative to the performance gain that I don't see the need for such a complex design.\\n\\n2. The model is not novel enough, as far as I know, the graph structure, the memory module, these are not new concepts. \\n\\n3. Models are not so essential to the development of the field.\", \"questions\": \"The OFA is published in ICML2022, the more recently model should be added as baselines.\\n\\n\\nTian Zhou, Ziqing Ma, Qingsong Wen, Xue Wang, Liang Sun, and Rong Jin. Fedformer: Frequency\\nenhanced decomposed transformer for long-term series forecasting. In International Conference\\non Machine Learning, pp. 27268\\u201327286. PMLR, 2022.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to Reviewer gtJT\", \"comment\": \"**W1**\\n\\n*(1) Analysis for Figure 1.*\\n\\nThanks for your comments. We use Figure 1 to demonstrate the high-level concept of sequential training to ease the understanding of continual learning for multivariate time series forecasting, where we have provided a real-world example in lines 053-061. We have provided sufficient analyses in the experiments in the main text and appendices. \\n\\n*(2) Other datasets.*\\n\\nOther benchmarks are applicable too. As we mentioned in Appendix C.1 from line 825, the Traffic-CL dataset is built upon the PEMSD3 benchmark, which is similar to METR-LA and PEMS04. We choose the PEMSD3 as it contains sufficient traffic data from 2011 to 2017, with sensors expanding across consecutive years, which is more suitable to demonstrate the continual forecasting setting. \\n\\n*(3) Baselines*\\n\\nThe full experimental results covering all baselines are presented in Table 7 & 8. As we mentioned in Appendix C.2 from line 923, our collection of baselines covers latent static graph-based forecaster (e.g., AGCRN, MTGNN), latent dynamic graph-based forecaster (e.g., ESG, StemGNN), static graph-based forecaster (STGCN), non-graph-based forecasters. \\n\\nMoreover, the most recent baselines like TimesNet, OFA, and iTransformer are general-purposed time series forecaster, where the state-of-the-art performance of short-term forecasting has been validated in their papers as well. In addition to that, iTransformer also captures the variable dependencies via attention. \\n\\n**W2: Insights our paper.**\\n\\nThanks for your comments. As we have depicted in introduction (lines 49-61) and figure 1, the challenge is about the catastrophic forgetting of learned dependency structures in multivariate time series forecasting in a sequential training learning scenario where MTS are continuously accumulated under different regimes. This challenge typically leads to performance degradation and distorted dependency structures (which are also validated in the main experiments and visualizations). To address this challenge, we leverage structural knowledge to steer the forecasting model toward identifying and adapting to different regimes, and select representative MTS samples from each regime for memory replay. As such, the obtained model can maintain accurate forecasts and infer the learned structures from the existing regimes. \\n\\n \\nThe dynamically changing dependency structures means the underlying varying characteristics of multivariate time series across different regimes, especially in the context of sequential training/continual learning, where the basic assumption is the underlying dependency structures are sampled from the same distribution within one regime, and from different distributions across regimes. The notion of dynamic graph learning is about capturing the variable dependencies in a more fine-grained manner, saying an input window.\"}", "{\"title\": \"Thanks for the rebuttal.\", \"comment\": \"Thank you for the rebuttal, but I am not fully convinced. Additionally, I still suggest that the authors adopt stronger spatial-temporal baselines. The current spatial-temporal baselines are not state-of-the-art (SOTA) and are from 2022 or earlier. Moreover, long-term time series forecasting methods are not suitable for spatial-temporal forecasting tasks. They perform significantly worse than spatial-temporal models on such tasks, and I am very confident in this point. The authors should select more appropriate baselines to validate the effectiveness of the proposed method and proactively discuss the specific problems this method addresses as well as the limitations of the algorithm.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"title\": \"Response to Reviewer ZKrh\", \"comment\": \"**W1: Complexity of model design.**\\n\\nThank you for your feedback. The proposed design is necessary to effectively leverage the graph structure information, which is crucial for modeling variable dependencies in time series within a continual learning setting. This design utilizes the structural relationships that simpler designs cannot, addressing key challenges unique to this problem domain. \\n\\n**W2&W3: Novelty.**\\n\\nThank you for your comment. While it is true that graph structures and memory modules are not entirely new concepts, existing graph models fail to effectively reflect the correct variable dependencies in a continual learning setting. As shown in Figure 4, this limitation can significantly deteriorate performance. Our proposed approach addresses this gap by accurately capturing and utilizing variable dependencies, which is a critical improvement. Additionally, regarding memory replay, we introduce a novel sampling scheme designed to systematically select more representative samples, further enhancing the replay mechanism. These contributions collectively demonstrate the novelty and importance of our work in advancing the field. \\n\\n**Q1:The OFA is published in ICML2022, the more recently model should be added as baselines.** \\n\\nThank you for your comments, The OFA is published in NeurIPS 2023, which can be considered as a SOTA model. \\n\\n*Xue Wang Liang Sun Rong Jin Tian Zhou, Peisong Niu. One Fits All: Power general time series analysis by pretrained lm. In NeurIPS, 2023.*\"}", "{\"summary\": \"This paper proposes Structural Knowledge Informed Continual Learning (SKI-CL), a framework for multivariate time series (MTS) forecasting that addresses challenges posed by variable dependencies across different regimes. By leveraging structural knowledge (such as physical constraints and domain knowledge), SKI-CL aims to mitigate catastrophic forgetting and improve model adaptation to shifting data distributions. The framework utilizes dynamic graph learning with consistency regularization, aligning learned dependencies with structural knowledge for better regime recognition. A representation-matching memory replay scheme is introduced to retain essential temporal dynamics from each regime. Experiments on synthetic and real-world datasets show superior forecasting accuracy and reliable dependency inference compared to state-of-the-art methods.\", \"soundness\": \"2\", \"presentation\": \"4\", \"contribution\": \"3\", \"strengths\": \"1. This paper is well written. The notations are clear.\\n\\n2. It provides up-to-date literature on MTS techniques with regards to regime shift. It underscores the potential of graph-based learning, paving ways for deep regime awareness in MTS.\\n\\n3. Among many lines of work addressing the regime discovery in MTS, graph-based learning has been well explored. However, this paper provides a systematic approach to tackle the regime shift in a rationale and reasonable way.\\n\\n4. The experiments are convincing and supports the arguments in the merits of the proposed SKI-CL framework. Especially, Figure 4, 5 and Figure 7,8,9,10 high light the differentiation of the proposed framework from competing methods well.\", \"weaknesses\": \"1. Overall, the technical documentation is comprehensive. It would be clearer if an algorithmic procedure is provided to give a high-level reference of how different components orchestrates in Figure 2 and 3.\\n2. While the traceback of nodes is valid and transparent in the inference, in domain applications, calibrated regimes matter to model owners because it helps interpret the results in the format of narratives. The proposed learning and inference of graph structure contain Coverage Maximization and Representation Matching Selection which are unsupervised, therefore not yet assembled for regime calibration tasks.\", \"questions\": \"1. Benchmarking regimes in multivariate time series forecasting is a foundation problem in MTS research, traditional econometrics method, e,g., Markov regime switch model that can be solved by EM algorithm can serve as a baseline model for regime discovery. Would that be something that can help bring the diverging methodologies to the same ground for a fair and reasonable competition, instead of the checking the numerical metrics?\\n\\n2. The message of Table 5 is to compare the number of baseline model parameters. Therefore, would it be better if the sorting is number of baseline model parameters instead of chronological order?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to Reviewer ifJg\", \"comment\": \"**W1**\\n\\nThank you for your comments, we will adjust it and provide a high-level explanation in our updated version. \\n\\n**W2** \\n\\nThank you for your insightful analysis of our work. We agree that calibrated regimes are indeed critical for interpretability in domain-specific applications. In our current model design, the Coverage Maximization and Representation Matching Selection processes are unsupervised, aimed to generalize to dynamic environments. In future work, we plan to incorporate supervised signals into these components to better support regime calibration tasks, thereby further enhancing the interpretability and usability of the model in real-world scenarios. \\n\\n**Q1** \\n\\nThank you for your insightful idea and we will continue our research in this direction.\\n\\n**Q2**\\n\\nThank you for your suggestion and we will adjust it in our updated version.\"}", "{\"summary\": \"The authors propose a novel Structural Knowledge Informed Continual Learning (SKI-CL) framework with a graph-based forecaster and a novel representation-matching memory replay scheme, which can perform MTS forecasting and infer dependency structures accurately under the continual learning setting. Experiments demonstrate the superiority of the proposed framework in continual MTS forecasting and dependency structure inference.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"1.The paper proposes an interesting method by combining dynamic graph learning with a novel representation-matching memory replay scheme for MTS forecasting and dependency structures inference under continual learning setting.\\n\\n2.The organization of this paper is clear.\", \"weaknesses\": \"1.The paper just combines continual learning with multi-variate time series forecasting, which can be seen as an incremental work. The authors should identify the research topic and main contribution, instead of adapting continual learning to multi-variate time series forecasting.\\n\\n2.The design of structural knowledge-informed graph learning model lacks innovation. The parameterized graph learning is similar to many works, e.g., AGCRN [1]. The authors could further clarify how their graph learning method differs significantly from existing works.\\n\\n3.The analysis of representation-matching memory replay scheme is uncompleted:\\n\\n(1)The authors should clarify how their representation-matching memory replay scheme differs with other experience-replay methods.\\n\\n(2)The analysis of efficiency about the scheme is not well discussed in the paper. It would be beneficial to explain the scheme's efficiency from both theoretical and experimental perspectives. \\n\\n(3)The visualization of selected samples in the scheme should be included to demonstrate the effectiveness of the model.\\n\\n4.In section 3.4, the authors could explain the inference process more clearly.\\n\\n5.The paper has some weaknesses in the experiments, which are not convincing enough:\\n\\n(1)Since one of the main contributions is developing a graph-based forecaster, some recent graph-based time series forecasting models should be mentioned and compared, e.g., CrossGNN [2] and MSGNet [3]. In addition, the continual learning methods applied to forecasting methods are old and some latest methods could be compared.\\n\\n(2)Different datasets have different methods to construct regimes, e.g., by year, state, activity, and adjacency, authors could further investigate the effect of different construction methods on the performance of SKI-CL. Some construction methods, e.g., by state and activity, are not reasonable and deviate from the intention of continual learning. In addition, the paper misses details regarding the train-test data splits.\\n\\n6.From a reader's perspective, the authors should enhance presentation to avoid misunderstanding. For example, for Fig. 6, the horizontal coordinate and vertical coordinate of heat map should start from 1. For Table. 1, what do the bolded and underlined results mean? For Equation 3, what does the $n_K$ mean?\\n\\n7.Strong recommendation to make the code publicly available.\\n\\n[1] Bai, L., Yao, L., Li, C., Wang, X., & Wang, C. 2020. Adaptive graph convolutional recurrent network for traffic forecasting. Advances in Neural Information Processing Systems, 33, 17804-17815.\\n\\n[2] Huang, Q., Shen, L., Zhang, R., Ding, S., Wang, B., Zhou, Z., & Wang, Y. 2023. CrossGNN: Confronting noisy multivariate time series via cross interaction refinement. Advances in Neural Information Processing Systems, 36, 46885-46902.\\n\\n[3] Cai, W., Liang, Y., Liu, X., Feng, J., & Wu, Y. (2024, March). MSGNet: Learning multi-scale inter-series correlations for multivariate time series forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence, 38, 11141-11149.\", \"questions\": \"See Weaknesses.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you for your detailed response and my concerns have been partially addressed. But I am not fully convinced and I decide to maintain my original score. Below are the specific reasons:\\n\\n1.I still believe this work is an incremental work which just combines continual learning with multi-variate time series forecasting after carefully reading of paper and responses.\\n\\n2.I still think the design of structural knowledge-informed graph learning model lacks innovation because the parameterized graph learning methods are well-researched. The authors claim that \\u2018Our novelty lies in the role of dynamic structure learning for regime characterization\\u2019, however, I think it just applies the existing dynamic structure learning to regime characterization, which lacks innovation.\\n\\n3.I still do not understand the detailed inference process from responses. The authors just introduce the regular inference process of continual learning and the benefits of SKI-CL in the paper and responses, but overlooks the detailed description of the inference process of SKI-CL.\\n\\n4.I still advise the authors to add some latest graph-based baselines because most baselines in the paper are old baselines from two years ago and not graph-based baselines, which would not prove the effectiveness of SKI-CL well. In addition, I still suggest that the authors should add the efficiency analysis of the representation-matching memory replay scheme and visualize the selected samples in the scheme to demonstrate the efficiency and effectiveness of the scheme respectively.\"}", "{\"summary\": \"This paper introduces the Structural Knowledge Informed Continual Learning (SKI-CL) framework for multivariate time series forecasting, which addresses the challenge of catastrophic forgetting when modeling variable dependencies across different regimes. SKI-CL leverages structural knowledge to guide the model in identifying and adapting to regime-specific patterns, and employs a representation-matching memory replay scheme to preserve temporal dynamics and dependency structures. The framework's efficacy is validated through experiments on synthetic and real-world datasets, demonstrating its superiority over state-of-the-art methods in continual MTS forecasting and dependency structure inference.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. **The motivation is very meaningful.**\\n\\n As recent research [1] has pointed out, distribution drift of time series (including dependency structures) may be the core bottleneck in the forecasting process. Therefore, I believe the authors are attempting to conduct a very significant study.\\n\\n2. **The writing is good and easy to follow.**\", \"weaknesses\": \"1. **I am concerned whether the change in dependency structures is indeed the core bottleneck in real-world scenarios.**\\n- 1.1 The authors should consider using analyses based on real data rather than just the schematic diagram in Figure 1. \\n- 1.2 Is this issue the core bottleneck of the dataset chosen by the authors? Are other more commonly used datasets, such as METR-LA and PEMS04, also applicable to this method? \\n- 1.3 In the current manuscript, it seems that the effectiveness of modeling the change in dependency structures can only be validated through experimental results. Thus, the authors need to compare against a broader range of stronger baseline methods, such as latent (but static) graph models, dynamic (but predefined graph-based) models (e.g., DGCRN [2]), and non-graph models (e.g., STID[3], STNorm[4], STEAformer[5]). The baselines currently chosen by the authors are not strong enough. For example, TCN is a conventional temporal model, while PatchTST, DLinear, TimesNet, and iTransformer are long-sequence forecasting models that are not specifically designed for spatiotemporal prediction and do not explicitly model the dependency graph between sequences. Additionally, the code for GTS contains unintentional errors that significantly impact its performance compared to the original paper.\\n\\n2. **Lack of sufficient insights.**\\n- 2.1 What are the core challenges and solutions in modeling the dynamic changes of dependency structures for time series forecasting? Currently, I cannot clearly see the connection between the challenges and the proposed techniques.\\n- 2.2 What is the distinction between dynamically changing dependency structures and dynamic graph learning? \\n\\n---\\n\\n[1] BasicTS: Exploring progress in multivariate time series forecasting: Comprehensive benchmarking and heterogeneity analysis. TKDE 2024.\\n\\n[2] DGCRN: Dynamic graph convolutional recurrent network for traffic prediction: Benchmark and solution. TKDD 2023.\\n\\n[3] STID: Spatial-Temporal Identity: A Simple yet Effective Baseline for Multivariate Time Series Forecasting. CIKM 2022.\\n\\n[4] STNorm: Spatial and Temporal Normalization for Multi-variate Time Series Forecasting. SIGKDD 2022.\\n\\n[5] STAEformer: Spatio-Temporal Adaptive Embedding Makes Vanilla Transformer SOTA for Traffic Forecasting. CIKM 2023.\", \"questions\": \"See Weakness.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"metareview\": \"This work tackles the continual learning problem for multivariate time series forecasting, where variables are added and their dependency structure evolves over time. To model the dependency structure, this work adopts a method similar to existing graph convolutional networks, e.g., AGCRN. Overall, the novelty lacks and its in-depth analyses are missing.\", \"additional_comments_on_reviewer_discussion\": \"The authors left short rebuttal messages, and the reviewers mostly stick to their original reviews.\"}" ] }

[ "Zhaogeng Liu", "Jinjie Fang", "Xingchen Li", "Bin Gu", "Yi Chang" ]

[ "Prompt learning", "LLM" ]

{ "note_id": [ "xoPiKB1Jl8", "wNj5Ln1seB", "rFgLFYKgiz", "o8tiNrEv7K", "O0290wUsbD", "IDK8xMJgD8", "ERL4DNrBAR" ], "note_type": [ "official_review", "comment", "official_review", "official_review", "decision", "official_review", "meta_review" ], "note_created": [ 1730135613692, 1739186443272, 1729999910491, 1730694819082, 1737523459219, 1730710269278, 1734548544708 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission1591/Reviewer_LLdx" ], [ "ICLR.cc/2025/Conference/Submission1591/Authors" ], [ "ICLR.cc/2025/Conference/Submission1591/Reviewer_u8A6" ], [ "ICLR.cc/2025/Conference/Submission1591/Reviewer_LiNf" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission1591/Reviewer_itSQ" ], [ "ICLR.cc/2025/Conference/Submission1591/Area_Chair_RH8f" ] ], "structured_content_str": [ "{\"summary\": \"This paper proposes a black-box discrete prompt optimization method to deal with the scenario where no label vocabulary is available. The method includes optimizing the Gumbel-Softmax parameterization process and a mapping matrix. Experiments on multiple classification datasets and models with different scales prove the efficacy of the method.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": [\"It is an interesting scenario where no pre-defined labels are available when optimizing prompts.\", \"The authors give a theoretical analysis of the convergence of the proposed optimization process.\"], \"weaknesses\": [\"The paper focuses on classification tasks. In the era of LLM, prompt optimization towards generation is more desirable. I would like to see if the authors can adapt their methods to generation tasks (e.g., question-answering).\", \"In Section 5.1, it is important to emphasize the method for conducting the label-free setting. However, this aspect is not addressed in the paper.\", \"The improvements on some datasets are trivial (e.g., BOOK, CoLA, QNLI). Significance tests are recommended to validate the effect of the proposed method.\"], \"questions\": [\"I am confused about the B_Y in Figure 1. I do not understand what it includes. Just input text or text and label? Can you explain the loss function in line 180? I am confused about what is the objective of this loss.\", \"The label-free scenario is an important application scenario in this work. However, I do not know how do you define \\\"label-free\\\". Do you use label information in the prompt optimization process?\", \"I will keep my confidence low because I do not understand the above-mentioned claim in the paper. If the authors explain them in detail, I will consider increasing the score. However, the presentation is not good and the paper should be reframed.\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}", "{\"withdrawal_confirmation\": \"I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors.\"}", "{\"summary\": \"This paper studies black-box discrete prompt learning when there does not exist a clear map between the task labels and the language model vocabularies. It desgins a reinforcement learning algorithm (LEAP) to optimize the discrete prompts and the mapping between model vocabularies and the task labels. Experiments and theoretical analysis show the effectiveness of LEAP on text classification tasks compared to exsiting balck-box prompt tuning methods.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"1\", \"strengths\": \"1. LMaSS is am important direction to explore in the era of LLMs.\\n2. The paper provides theoretical supports of convergence for the proposed algorithm.\", \"weaknesses\": \"1. The main weakness of this paper is the practical significance of the \\\"label-free\\\" setting. For large language models (LLMs), tasks are generally unified in a generative framework, where most labels can be expressed as text (or at least as phrases). For instance, in the Amazon Books rating task mentioned in lines 68-69, numerical ratings can be converted to strings, tokenized, and treated as labels. Therefore, further clarification of the practical implications of this \\\"label-free\\\" setting would be necessary.\\n\\n2. (Following Point 1) The \\\"label-free\\\" setting appears to be independent of specific tuning methods. Are there other studies that explore this setting? If so, discussing them in the Related Work section would be needed.\\n\\n3. A secondary weakness of the paper is the limited applicability of the proposed method to various tasks. Section 3.1 suggests that the approach relies on predefined categories, and the experiments focus solely on simple text classification tasks. This approach thus appears restricted to text classification, while most real-world applications involve language generation tasks without predefined \\\"categories\\\" (though ground-truth label tokens do exist).\", \"questions\": \"See Weakness.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The paper studied the setting of black-blox prompt learning without a label vocabulary. To solve the missing label vocabulary problem, the author proposed to learn a mapping from LLM tokens to discrete labels. The mapping and the discrete prompts are trained jointly via policy gradient descent. Experiments on GPT-XL, RoBERTA-large and Llama3-8B show that the newly proposed LEAP algorithm performs better than baselines like BDPL, SSPT.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"1\", \"strengths\": \"The paper's idea of learning a mapping from token to label is straightforward. The author also proposed a training algorithm with variance-reduced policy gradient descent.\", \"weaknesses\": \"The experimental settings are not realistic. The author only conducted experiments on datasets with at most 4 classes (as shown in Table 5). If you have less than four label classes, it is often straightforward to assign textual description of the labels and there is no need to consider the missing-label-vocabulary problem. On the other hand, the author conducted experiments on weak LLMs like RoBERTa, GPT2-XL and Llama3-8B. The author needs to conduct experiments on stronger LLMs like OpenAI's GPT-4o, or at least Llama3-70B. Black-box prompt learning are usually targeted for stronger LLMs. It is not clear if LEAP will still work.\", \"questions\": \"N/A\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"summary\": \"The paper presents LEAP, a method for black-box discrete prompt learning without relying on a predefined label vocabulary. It employs an alternating optimization framework to learn prompt tokens and a mapping matrix for LLM outputs. The paper provides convergence analysis and experimental results showing LEAP's effectiveness on label-vocabulary-free datasets.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"1. The paper introduces LEAP, a method for black-box prompt learning that does not require a predefined label vocabulary, offering a new solution for optimizing LLMs in scenarios with limited access to model internals.\\n\\n2. LEAP is presented with a clear structure, detailing an alternating optimization strategy and a learnable mapping matrix. The paper includes a thorough theoretical analysis on the convergence of the proposed method.\\n\\n3. The research tackles a practical challenge in applying LLMs, enhancing their adaptability in real-world applications where internal model parameters and gradients are inaccessible.\", \"weaknesses\": \"1. The paper focuses on classification tasks and does not provide experiment results for generation tasks, which may limit the assessment of LEAP's versatility across different NLP applications.\\n\\n2. The paper does not include comparisons with the latest model optimization methods like BBTv2[1] and GDFO[2], which could impact the perceived novelty and competitiveness of the LEAP method.\\n\\n3. The paper lacks details on the initialization of the learnable matrix M and its optimization's influence on model performance.\", \"questions\": \"1. Could the authors elaborate on the initialization process of the learnable mapping matrix and its sensitivity to different initializations?\\n\\n2. How does LEAP compare with existing methods in terms of computational resources required, especially when scaling up to larger models or datasets?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"metareview\": \"This paper studies the problem of discrete prompt learning for a black box LLM where no label vocabulary is available. It employs an alternating optimization strategy to learn discrete prompt tokens and a learnable matrix for LLM outputs.\\n\\nSome of the reviewers appreciate the authors tackling the interesting and challenging application scenario. The provided theoretical supported is valuable. But there are some concerns before this paper is ready for publication. The major one is the experimental setting is not practical. The paper only studies the classification problem, would be encouraged to see the results on generation tasks. Stronger comparison methods should be included. Based on the assessment, it is concluded that the paper could not be accepted in its current form and would require a major revision.\", \"additional_comments_on_reviewer_discussion\": \"No rebuttal provided.\"}" ] }

[ "Yu Bo", "Weian Mao", "Yanjun Shao", "Weiqiang Bai", "Peng Ye", "Xinzhu Ma", "Junbo Zhao", "Hao Chen", "Chunhua Shen" ]

[ "DNA modeling", "foundation model", "Genomic Language Model", "Representation Learning" ]

{ "note_id": [ "wN6JhGaYPy", "vPP1uCmAHR", "vHztsBMn2W", "v1neGyaP4u", "uemo0cE2Fi", "uTP3MQBfvE", "twGr0e1jDU", "plbPkWnhPj", "nuvLwqyIGD", "m8W8WgLc06", "gVE8jLGx65", "f4no4d6RWR", "esVTjwqOo2", "dH8lbZ7w9M", "cMTQKgtF6z", "bt6Hbi8PQl", "aUufbYVhBF", "aLlkHBy3Ar", "aCuUrcTly3", "aBKOjR49dH", "X9QJlxjFze", "VrkwBQplmp", "Rj6iXgA5CG", "RGacXg2UYa", "QD5AmjVmcO", "OEqp3IDQfe", "KwWyXYjcr2", "JDrpGgZyuv", "H47R01Gwrf", "E4wSBC92IR", "DzmkX7L03r", "DjdEne67Yo", "BTFzfGVuFt", "7oMMGitd3S", "6De3S20ZaN", "5iDzm1bg9w", "0nybF11Zmu", "00p9vCoehB" ], "note_type": [ "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "meta_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "decision", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment" ], "note_created": [ 1732735587714, 1732327287824, 1732699291838, 1732656718019, 1732516603291, 1730720476250, 1729376647335, 1732324697788, 1732744236657, 1732326955684, 1732573823325, 1732325349638, 1730655846473, 1732326324829, 1730671086136, 1732324445210, 1732327122928, 1732698720314, 1732625417293, 1732324926433, 1732900607466, 1734052972532, 1732483185701, 1732325621809, 1732482649482, 1732504051836, 1732513390132, 1732481774054, 1732511408920, 1737523473533, 1732482169600, 1732700696308, 1733101893023, 1732326156073, 1732324818736, 1732899054664, 1732326574352, 1732327565744 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission1899/Reviewer_2bPc" ], [ "ICLR.cc/2025/Conference/Submission1899/Authors" ], [ "ICLR.cc/2025/Conference/Submission1899/Authors" ], [ "ICLR.cc/2025/Conference/Submission1899/Reviewer_NQxy" ], [ "ICLR.cc/2025/Conference/Submission1899/Authors" ], [ "ICLR.cc/2025/Conference/Submission1899/Reviewer_NQxy" ], [ "ICLR.cc/2025/Conference/Submission1899/Reviewer_2bPc" ], [ "ICLR.cc/2025/Conference/Submission1899/Authors" ], [ "ICLR.cc/2025/Conference/Submission1899/Reviewer_PceR" ], [ "ICLR.cc/2025/Conference/Submission1899/Authors" ], [ "ICLR.cc/2025/Conference/Submission1899/Reviewer_2bPc" ], [ "ICLR.cc/2025/Conference/Submission1899/Authors" ], [ "ICLR.cc/2025/Conference/Submission1899/Reviewer_SH7n" ], [ "ICLR.cc/2025/Conference/Submission1899/Authors" ], [ "ICLR.cc/2025/Conference/Submission1899/Reviewer_PceR" ], [ "ICLR.cc/2025/Conference/Submission1899/Authors" ], [ "ICLR.cc/2025/Conference/Submission1899/Authors" ], [ "ICLR.cc/2025/Conference/Submission1899/Authors" ], [ "ICLR.cc/2025/Conference/Submission1899/Authors" ], [ "ICLR.cc/2025/Conference/Submission1899/Authors" ], [ "ICLR.cc/2025/Conference/Submission1899/Authors" ], [ "ICLR.cc/2025/Conference/Submission1899/Area_Chair_Hb7W" ], [ "ICLR.cc/2025/Conference/Submission1899/Authors" ], [ "ICLR.cc/2025/Conference/Submission1899/Authors" ], [ "ICLR.cc/2025/Conference/Submission1899/Authors" ], [ "ICLR.cc/2025/Conference/Submission1899/Reviewer_2bPc" ], [ "ICLR.cc/2025/Conference/Submission1899/Reviewer_2bPc" ], [ "ICLR.cc/2025/Conference/Submission1899/Authors" ], [ "ICLR.cc/2025/Conference/Submission1899/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission1899/Authors" ], [ "ICLR.cc/2025/Conference/Submission1899/Authors" ], [ "ICLR.cc/2025/Conference/Submission1899/Reviewer_SH7n" ], [ "ICLR.cc/2025/Conference/Submission1899/Authors" ], [ "ICLR.cc/2025/Conference/Submission1899/Authors" ], [ "ICLR.cc/2025/Conference/Submission1899/Authors" ], [ "ICLR.cc/2025/Conference/Submission1899/Authors" ], [ "ICLR.cc/2025/Conference/Submission1899/Authors" ] ], "structured_content_str": [ "{\"comment\": \"Thanks for the clarification. I believe A5, Figure 5, and Table 15 provide strong evidence to support the authors' claims. I would recommend that the authors extract some of the visualizations and results and include them in the main paper. This would help readers better understand the authors' motivations. Based on the additional visual evidence and results, I am happy to raise my score.\"}", "{\"title\": \"Response to Reviewer 2bPc(3/4)\", \"comment\": \"> I found the organization of the paper somewhat confusing. For instance, the comparison of downsampling and dilation in Section 3.3 and Table 1 would be more appropriately placed in the experiments section. Additionally, given the variety of tasks with different state-of-the-art models and settings, I recommend adding a dedicated section that briefly outlines the experimental setup. This section should include the objectives of each task, descriptions of the baseline models, and the specific configurations used in each experiment.\\n\\nThank you for your feedback. We agree that the comparison of downsampling and dilation would be more appropriately placed in the experiments section. We will reorganize the paper as suggested and include a dedicated section outlining the experimental setup, including the objectives of each task, descriptions of the baseline models, and the specific configurations used. We will inform you once the revisions are complete. Thank you again for your valuable input!\"}", "{\"comment\": \"Dear reviewer:\\n\\nHere we provide the motivation to consider CNNs against other models for DNA modeling tasks, **which has not been discussed in other work.**\\n\\nBesides the computational complexity and model sizes of CNNs, **CNNs inherently possess inductive biases for neighborhood modeling, which may be crucial for many DNA sequence tasks**. We have already added relevant experiments and analyses; please refer to A5, Figure 5, and Table 15 for the updated information.\\n\\nHere we briefly explain how we support our hypothesis.\\n\\nSpecifically, the main issue with self-attention is that it does not inherently focus on neighborhood sequences, which could contribute to its suboptimal performance in many DNA modeling tasks. To support our claim, we modify the NTv2 model by adjusting the RoPE's \\ud835\\udf03 and initializing the bias of the \\ud835\\udc5e\\ud835\\udc58 linear layer to [0,0,\\u2026,1], while keeping other initializations consistent with NTv2 (std=0.02,mean=0). This adjustment increases the attention map's focus on neighborhood sequences, thereby enhancing the Transformer's inductive bias.\\n\\nOur approach is similar to the methodology in [1]. When tested on H3K14ac, a task with strong local dependencies, the results show significant improvement (34.42 vs. 46.65 with the enhancements). Similar results on other 2 tasks can be found in Table 15.\\n\\nHowever, this method provides only a naive way to add inductive bias to Transformers. It still does not surpass ConvNova trained from scratch. Further research is needed to fundamentally strengthen the Transformer's inductive bias for neighborhood modeling.\\n\\nWe welcome any additional questions or feedback you might have and would be glad to engage in further discussion.\\n\\n[1] RepMLPNet: Hierarchical Vision MLP with Re-parameterized Locality. CVPR. 2022.\"}", "{\"title\": \"Reply to Authors' Rebuttal\", \"comment\": \"Thanks for the detailed response, extensive experiments, and the updated manuscript. Although this manuscript is technically oriented with some straightforward design, it really provided a new insight and useful DNA foundation model, which are verified by comprehensive benchmarks. What I am still concerned about is W1. I believe the authors should further clarify the novelty of the design model with empirical analysis and well-arranged backgrounds to enhance the contribution of rethinking the pure convolution-based architectures for long-sequence tasks with DNA. Maybe provide a table or timeline figure to illustrate further the motivation that CNNs can still be competitive for the genomic tasks. Overall, after going through comments from other reviewers, I decided to raise my score to 6 and encourage the authors to further polish the manuscript and provide more ablation experiments.\"}", "{\"comment\": \"The main issue with self-attention is that it does not inherently focus on neighborhood sequences, which could contribute to their suboptimal performance in many DNA modeling tasks. To support our claim, we modify the NTv2 model by adjusting the RoPE's\\n\\ud835\\udf03 and initializing the bias of the \\ud835\\udc5e\\ud835\\udc58 linear layer to [0, 0, ..., 1], while keeping other initializations consistent with NTv2 (\\ud835\\udc60\\ud835\\udc61\\ud835\\udc51=0.02, \\ud835\\udc5a\\ud835\\udc52\\ud835\\udc4e\\ud835\\udc5b=0). This adjustment increases the attention map's focus on neighborhood sequences, enhancing the Transformer's inductive bias.\\n\\nOur approach is similar to the methodology in [1]. When tested on H3K14ac, a task with strong local dependencies, the results improved significantly (34.42 vs. 46.65 with the enhancements).\\n\\nHowever, this method provides only a naive way to add inductive bias to Transformers. It still does not surpass ConvNova trained from scratch. Further research is needed to fundamentally strengthen the Transformer's inductive bias for neighborhood modeling.\\n\\nWhile this is not our main focus, we still hope this addresses your concerns.\\n\\n[1] RepMLPNet: Hierarchical Vision MLP with Re-parameterized Locality. CVPR. 2022\"}", "{\"summary\": \"Since various Transformer and SSM-based models are proposed in DNA language modeling, the authors conduct empirical studies of whether CNNs are truly being surpassed by these recently proposed approaches. With analysis results, this paper develops a simple yet well-designed CNN-based method called ConvNova, which identifies and proposes three effective designs: 1) dilated convolutions, 2) gated convolutions, and 3) a dual-branch framework for gating mechanisms. Through extensive empirical experiments, we demonstrate that ConvNova significantly outperforms recent methods on more than half of the tasks across several foundation model benchmarks.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"3\", \"strengths\": [\"**(S1)** The paper addresses a critical problem in DNA embedding with a novel application of classical convolution, showing significant improvements in species-aware tasks.\", \"**(S2)** The overall presentation is well originated and easy to follow. It is clear that the authors provide step-by-step designs to improve the attention mechanism for better performance and efficiency.\"], \"weaknesses\": \"* **(W1)** Despite the efficient design, the proposed ConvNova is somewhat simple and lacks novelty and support. The (dilated) convolution with gating branch is not a new design (proposed by MogaNet [1] in 2022 and well studied by StarNet [2] and MambaOut [3] in 2024), especially after Mamba variants came out (i.e., the implicit long convolution with gating). The authors should discuss these background works (discussing in the related work section or comparing with them), and provide more supports of why the proposed design is specially useful in DNA applications (refer to Q1 for details).\\n\\n* **(W2)** Although the authors have compared with several well-known DNA models, some recently published DNA models and pre-training works are overlooked (e.g., GPN [4], VQDNA [5], and DNABERT-S [6]). Meanwhile, there are various DNA benchmarks that should be referred to and compared (classical benchmarks like GUANinE v1.0 [7], BEND [8], and GUE [9]), especially some long-range benchmarks [10] where SSM-based models work well. From my perspective, I am still not sure or not convinced that the dilated convolution with gating aggregation could consistently outperform self-attention or SSM architectures.\\n\\n### Reference\\n\\n[1] MogaNet: Multi-order Gated Aggregation Network. ICLR, 2024.\\n\\n[2] Rewrite the Stars. CVPR, 2024.\\n\\n[3] MambaOut: Do We Really Need Mamba for Vision? arXiv, 2024.\\n\\n[4] DNA language models are powerful predictors of genome-wide variant effects. PNAS, 2023.\\n\\n[5] VQDNA: Unleashing the Power of Vector Quantization for Multi-Species Genomic Sequence Modeling. ICML, 2024.\\n\\n[6] DNABERT-S: Learning Species-Aware DNA Embedding with Genome Foundation Models. arXiv, 2024.\\n\\n[7] GUANinE v1.0: Benchmark Datasets for Genomic AI Sequence-to-Function Models. bioRxiv, 2023.\\n\\n[8] BEND: Benchmarking DNA Language Models on biologically meaningful tasks. ICLR, 2024.\\n\\n[9] DNABERT-2: Efficient Foundation Model and Benchmark For Multi-Species Genome. ICLR, 2024.\\n\\n[10] Advancing DNA Language Models: The Genomics Long-range Benchmark. ICLR Workshop on Machine Learning for Genomics Explorations, 2024.\", \"questions\": [\"**(Q1)** Is there any empirical analysis or theoretical support to verify that the dilated convolution is capable and more efficient than self-attention or SSM modules on the DNA tasks? For example, the authors could visualize the reception field or analysis the learned patterns to show that the dilated convolutions with gating learn better.\", \"**(Q2)** Some questions about the hyper-parameters in the network. How to determine the kernel size and the dilated ratio in ConvNova? Are there any implementation details of the ablation studies and providing the corresponding configurations (in addition to Table 9)? Meanwhile, is there more ablation of different designs (like analysis in [1, 2, 3]) of the gating branches as mentioned in Eq. (1) and Appendix A.3?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"details_of_ethics_concerns\": \"N/A\", \"rating\": \"6\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The paper argues that a well-designed convolutional neural network (CNN), when used as a DNA foundation model, can outperform Transformer- and SSM-based models not only in accuracy but also in inference speed, model size, and training cost. The authors introduce ConvNova, a model composed of dual-branched Gated Convolutional Blocks (GCBs). The GCBs incorporate dilated convolutions and avoid performance-degrading downsampling, as supported by empirical findings. Ablation studies demonstrate the benefits of the dilated convolutions, dual-branch architecture, and gated mechanisms within the GCBs. The proposed model achieves higher accuracy while maintaining a smaller parameter count compared to several state-of-the-art Transformer and SSM models.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"The authors conducted a series of carefully controlled experiments using different random seeds to demonstrate the accuracy advantages of their design. They compared the proposed ConvNova model against state-of-the-art Transformer- and SSM-based models across various tasks, including two benchmarks for short-range input sequences (the Nucleotide Transformer Benchmark and the Genomic Benchmark) and two benchmarks for long-range tasks (the Bend Gene Finding and the Chromatin Profile Prediction). Overall, the experiments and the ablation study are robust and effectively support the claims and the effectiveness of the proposed design.\", \"weaknesses\": \"While the authors present solid empirical evidence, the paper lacks a clear discussion on the intuition behind the design, making it more like a technical report. For example, ConvNeXt [1] builds on a well-known variant, ResNeXt [2], and details the accuracy impact of each modification. In contrast, ConvNova does not seem to establish a clear rationale connecting it to previous CNN-based DNA foundation models, which makes the design choices appear somewhat ad hoc. Beyond demonstrating superior accuracy compared to Transformers and SSMs, it would be valuable to include an in-depth discussion and comparison with models like LegNet on block design, architecture, parameter count, and training schemes in the main text.\\n\\nI found the organization of the paper somewhat confusing. For instance, the comparison of downsampling and dilation in Section 3.3 and Table 1 would be more appropriately placed in the experiments section. Additionally, given the variety of tasks with different state-of-the-art models and settings, I recommend adding a dedicated section that briefly outlines the experimental setup. This section should include the objectives of each task, descriptions of the baseline models, and the specific configurations used in each experiment.\\n\\n[1] A ConvNet for the 2020s\\n\\n[2] Aggregated Residual Transformations for Deep Neural Networks\", \"questions\": [\"What is the intuition and rationale behind using dilation over self-attention in a DNA foundation model? And why is dilation outperforming self-attention?\", \"How is the receptive field for the dilated convolutions determined?\", \"What factors contribute to the dual-branch design outperforming a single-branch approach?\", \"In addition to the details in Table 11 and Section A.4, could you provide a comparison between ConvNova and LegNet in terms of structure and training scheme, and explain why your design is superior?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to Reviewer NQxy(2/4)\", \"comment\": \"> - (W2) Although the authors have compared with several well-known DNA models, some recently published DNA models and pre-training works are overlooked (e.g., GPN [4], VQDNA [5], and DNABERT-S [6]). Meanwhile, there are various DNA benchmarks that should be referred to and compared (classical benchmarks like GUANinE v1.0 [7], BEND [8], and GUE [9]), especially some long-range benchmarks [10] where SSM-based models work well. From my perspective, I am still not sure or not convinced that the dilated convolution with gating aggregation could consistently outperform self-attention or SSM architectures.\\n\\nThank you for the valuable feedback. We coduct additional comparisons with GPN and DNABERT-2 on the GUE benchmark. As shown in the following results, ConvNova outperforms the baseline in 19 out of 28 tasks. \\n\\n### Epigenetic Marks Prediction\\n\\n| Model | H3 | H3K14ac | H3K36me3 | H3K4me1 | H3K4me2 | H3K4me3 |\\n|-------------|--------|-----------|-----------|-----------|-----------|-----------|\\n| DNABERT-2 | 79.01 | _56.23_ | _60.92_ | 48.55 | 34.35 | 42.22 |\\n| DNABERT-S | **79.42** | 52.00 | 55.10 | 48.53 | 37.19 | 36.43 |\\n| GPN | _79.30_ | 52.38 | 60.77 | _51.66_ | _40.86_ | _45.38_ |\\n| ConvNova | 76.94 | **58.43** | **61.04** | **53.11** | **42.86** | **52.21** |\\n\\n---\\n\\n### Epigenetic Marks Prediction and Promoter Detection\\n\\n| Model | H3K79me3 | H3K9ac | H4 | H4ac | all | notata | tata |\\n|-------------|----------|----------|-------|--------|---------|---------|--------|\\n| DNABERT-2 | 62.69 | _57.60_ | 77.73 | 45.66 | 85.82 | _93.33_ | 64.14 |\\n| DNABERT-S | 63.90 | 53.90 | 80.30 | 47.75 | 86.89 | **93.37** | _64.45_ |\\n| GPN | _66.15_ | 57.01 | **81.56** | _51.66_ | **89.03** | 92.02 | 53.78 |\\n| ConvNova | **67.67** | **61.75** | _80.30_ | **54.89** | _87.45_ | 92.16 | **66.06** |\\n\\n---\\n\\n### Transcription Factor Prediction (Human) and Core Promoter Detection\\n\\n| Model | 0 | 1 | 2 | 3 | 4 | all | notata | tata |\\n|-------------|---------|---------|---------|---------|---------|---------|---------|--------|\\n| DNABERT-2 | 68.16 | 71.39 | 66.83 | **62.05** | 75.55 | 67.88 | 66.76 | 58.38 |\\n| DNABERT-S | **69.85** | **74.65** | 65.04 | _56.28_ | 74.94 | 67.00 | _69.26_ | _74.88_ |\\n| GPN | 66.37 | _72.31_ | _73.31_ | 49.73 | _76.21_ | _70.31_ | 68.42 | 73.26 |\\n| ConvNova | _69.82_ | 71.83 | **74.52** | 50.58 | **76.65** | **71.40** | **69.86** | **78.74** |\\n\\n---\\n\\n### Transcription Factor Prediction (Mouse) and Virus & Splice Detection\\n\\n| Model | 0 | 1 | 2 | 3 | 4 | Covid | Reconstruct |\\n|-------------|---------|---------|---------|---------|---------|---------|-------------|\\n| DNABERT-2 | 57.29 | _83.85_ | 78.10 | _77.53_ | 46.19 | 70.29 | 86.19 |\\n| DNABERT-S | _58.29_ | **85.14** | 74.19 | 77.25 | **50.30** | _70.78_ | 85.43 |\\n| GPN | 57.79 | 78.87 | _82.32_ | 74.59 | 43.84 | 70.46 | _88.03_ |\\n| ConvNova | **59.40** | 81.68 | **86.05** | **81.63** | _47.12_ | **73.86** | **88.29** |\\n\\nRegarding VQDNA, it is not included in the comparison as our work focuses on CNN-based models versus self-attention or SSM-inspired architectures. VQDNA uses vector quantized tokenization and the DNABERT-2 architecture, which falls outside the scope of this study. Additionally, as VQDNA lacks open-source code, future work may explore comparisons with VQ-based settings.\\n\\nFor the long-range benchmarks mentioned, the lack of open-source code presents a limitation. However, we complete the Variant Effect Prediction task across three tasks. As shown in the results, ConvNova successfully models long-range dependencies.\\n\\n### **Variant Effect Prediction**\\n\\n| Params |**ConvNova(25M)** | **Caduceus(8M)** | **HyenaDNA(1.6M)** | **HyenaDNA(0.4M)** | **HyenaDNA(3.3M)** | **HyenaDNA(6.6M)** | **NT(50M)** | **NT(100M)** | **NT(250M)** | **NT(500M)** |\\n|---------|-----------------|---|-----------|--------------|--------------|--------------|--------|--------|--------|--------|\\n| Context Length (bp) | 10K | 131K | 1K | 16K | 32K | 160K | 12K | 12K | 12K | 12K | 196K |\\n| AUC-ROC | **0.742** | 0.715 | 0.705 | 0.704 | 0.713 | 0.706 | 0.714 | 0.722 | 0.721 | 0.719 |\\n| Accuracy | **0.667** | 0.666 | 0.648 | 0.649 | 0.658 | 0.647 | 0.661 | 0.664 | 0.661 | 0.657 |\"}", "{\"title\": \"Overall Thoughts\", \"comment\": \"Thank you for the additional information and explanation. I do think that paper could still benefit from adding in qualitative results figures into the main paper.\\n\\nI maintain that **the paper should be accepted**, but would still benefit from some more discussion in exploring the qualitative differences between model classes a bit more in the main paper.\"}", "{\"title\": \"Response to Reviewer 2bPc(1/4)\", \"comment\": \"We sincerely appreciate Reviewer 2bPc's thoughtful feedback, and here we provide corresponding responses to address these concerns.\\n> While the authors present solid empirical evidence, the paper lacks a clear discussion on the intuition behind the design, making it more like a technical report. For example, ConvNeXt [1] builds on a well-known variant, ResNeXt [2], and details the accuracy impact of each modification. In contrast, ConvNova does not seem to establish a clear rationale connecting it to previous CNN-based DNA foundation models, which makes the design choices appear somewhat ad hoc. \\n\\nIn recent foundation model research, CNNs have been largely ignored and not compared with existing methods. In fact, previous CNN models, such as Basenji, performes poorly across tasks(See Table 13). We are the first to compare CNNs with other architectures, demonstrating that with the right design, CNNs can outperform them. Indeed, this answers the question: CNNs are still not surpassed by other architectures.\\n\\nThe motivation behind our architecture design are following several key considerations. The gating mechanism could enable dynamic feature selection. The dual-branch design facilitates independent feature extraction and promotes complementary representation learning. Dilation serves two purposes: (1) downsampling, and (2) investigating the local dependency condition of downstream tasks.\\n\\nHowever, unlike convNext, which can use ImageNet a single unified metric, DNA-related tasks are diverse and lack such benchmarks. Thus, instead of detailed analyses like ConvNeXt's, we will provide ablation studies on dilation rates and kernel sizes, offering insights into their impact across tasks soon. We will inform you once available.\"}", "{\"comment\": \"Thanks for the reply. Would you kindly point out the tables/experiments for the modified NTv2?\"}", "{\"title\": \"Response to Reviewer PceR(1/2)\", \"comment\": \"We sincerely appreciate Reviewer PceR's thoughtful feedback, and here we provide corresponding responses to address these concerns.\\n\\n> The paper needs to do a better job, in the main text, of demonstrating what specifically about the model is novel, and what specifically contributes to the superior performance of this model compared to prior CNN-based DNA models.\\nFor example, is it just better pretraining data? Is it including the complement sequence? I saw dilations, and receptive field choice were shown to be a portion of this, and I saw the ablations on the different components.\\n>\\n>However, the specific novelty of this design is a bit ambiguous to me. The figures need to be clearer to emphasize the novel components compared to prior CNN-based models. Specifically, just answering the question in explicit detail: CNNs have been around for a while, why didn't we find a high-performing model like this well before this? What new has been added that wasn't found before?\\n\\nIn recent foundation model research, CNNs have been largely ignored and not compared with existing methods. In fact, previous CNN models, such as Basenji, performes poorly across tasks(See Table 13). We are the **first to compare CNNs with other architectures, demonstrating that with the right design, CNNs can outperform them**. Indeed, this answers the question: **CNNs are still not surpassed by other architectures.**\\n\\nYour points are insightful and highlight considerations that, while not our primary focus, remain important. Through additional ablation experiments(we will provide the results soon), we find that the **most critical factor is the appropriate dilation mechanism**, as deviations from optimal dilation rates lead to performance degradation. Other design choices, such as the dual-branch architecture, pretraining data, and gating mechanism, also contribute to performance improvements, as shown in our ablation studies Table 5.\\n\\nAs for the question, \\\"why didn't we find a high-performing model like this well before this?\\\" the main reason is that this is an underexplored area. **We are the first to systematically evaluate CNNs across such a broad range of tasks**. What's more, traditional CNN designs(like DeepSEA) often rely on pooling operations, which are generally unsuitable for masked language modeling. By addressing these gaps, we demonstrate that with thoughtful design, CNNs can perform robustly as DNA foundation models.\"}", "{\"summary\": \"The paper proposes to revisit CNN based architectures for DNA sequence modeling, and propose a new CNN based architecture, ConvNova. The authors show, with an extensive empirical study, that the model has state of the art performance on several DNA modeling task when controlling for model size.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"The authors have a high quality set of evaluation experiments for their method. This can be significant as it may lead to further research on CNN based architectures for sequence modeling. The writing is mostly clear and the author's proposal significantly improves the state of the art.\", \"weaknesses\": \"The novelty of the approach is somewhat limited, as CNNs have already been applied to this setting, and the training procedure is broadly the same as in the HyenaDNA paper.\\n\\nThe CNN architecture used is based on gated convolutions, which has already been proposed elsewhere, limiting the novelty. \\n\\nTo better understand the novelty of the approach, a benchmark comparing with a more standard CNN architecture would have been useful. \\n\\nThe paper seems to lack a clear motivation as to why this specific CNN architecture was proposed. \\n\\nI would have expected to find a description of the pretraining data in the main paper. \\n\\nI would say that calling the model a \\u201cfoundation model\\u201d is somewhat misleading as the model used in the paper has 7M parameters and was pretrained for 14 hours on a single GPU.\", \"questions\": \"What is the pretraining dataset?\\n\\nDid you benchmark the model against a more traditional CNN architecture?\\n\\nIs the architecture inspired by the DNA task in some way?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to Reviewer SH7n(2/3)\", \"comment\": \">I would have expected to find a description of the pretraining data in the main paper.\\n\\nThank you for pointing this out. In our paper, we mention that our pretraining paradigm follows HyenaDNA, which implicitly includes a description of the pretraining data. However, we appreciate your suggestion and will explicitly clarify the details of the pretraining data in A.1 to ensure there is no ambiguity. We will inform you as soon as we've made the revision.\"}", "{\"summary\": \"The paper introduces a new convolutional-based DNA foundation model and shows that CNNs can be competitive with Transformers and SSMs both in terms of absolute performance as well as accuracy vs. speed tradeoffs\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"3\", \"strengths\": \"The paper presents strong, rigorous, and detailed results, with several strong baselines\\nThe significance of the model itself is high, both for the application domain but what it means in the context of SSMs.\\nTo me, DNA seems like the prime use case that SSMs, based on their performance profile, should be the winning model. A pure convolutional model outperforming Cadesus is a very important result in the context, and adds to the growing body of work that show SSMs and their complexities may not be necessary.\", \"weaknesses\": \"- The paper needs to do a better job, in the main text, of demonstrating what specifically about the model is novel, and what specifically contributes to the superior performance of this model compared to prior CNN-based DNA models.\\n\\nFor example, is it just better pretraining data? Is it including the complement sequence? I saw dilations, and receptive field choice were shown to be a portion of this, and I saw the ablations on the different components. \\n\\nHowever, the specific novelty of this design is a bit ambiguous to me. The figures need to be clearer to emphasize the novel components compared to prior CNN-based models. Specifically, just answering the question in explicit detail: CNNs have been around for a while, why didn't we find a high-performing model like this well before this? What new has been added that wasn't found before?\\n\\nThere should be some more qualitative analysis in the main text, mainly I would like to see an understanding of how the different model classes compare in the types of errors they make. For example, does the smaller receptive field contribute to certain types of errors for tasks where longer range reasoning is necessary?\", \"questions\": \"Also mentioned in the weaknesses section, CNNs have been around for a while, why didn't we find a high-performing model like this well before this? What new has been added that wasn't found before?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"We sincerely appreciate Reviewer NQxy's thoughtful feedback, and here we provide corresponding responses to address these concerns.\\n\\n## For Weaknesses:\\n> (W1) Despite the efficient design, the proposed ConvNova is somewhat simple and lacks novelty and support. The (dilated) convolution with gating branch is not a new design (proposed by MogaNet [1] in 2022 and well studied by StarNet [2] and MambaOut [3] in 2024), especially after Mamba variants came out (i.e., the implicit long convolution with gating). The authors should discuss these background works (discussing in the related work section or comparing with them), and provide more supports of why the proposed design is specially useful in DNA applications (refer to Q1 for details).\\n\\nIn recent dna foundation model research, CNNs have been largely ignored and not compared with existing methods. Previous CNN models, such as Basenji, perform poorly across tasks. We want to clarify that our primary contribution is not architectural novelty, but being the **first to propose and compare traditional CNN-based DNA foundation model with other architectures, demonstrating that with the right design, CNNs can outperform them**. Indeed, this answers the question: **CNNs are still not surpassed by other architectures**.\\n\\nWhile the architectural design is not our main focus, we would like to address the concerns about the dilation with gating mechanism. The key factor driving the model's performance is not gating but rather the carefully chosen dilation design. We have compared U-Net-style downsampling approaches and performed ablation studies on different dilation rates, selecting the current design based on the requirements of foundation models for long-range tasks. Moreover, dilation help to investigate the down stream tasks' local dependency. For instance, larger dilation sizes (e.g., 4) improve performance on H3 (81.49 vs. 77.16), while smaller sizes (e.g., 1) are better for H3K4me3 (67.15 vs. 60.20).\\n\\nAs for the gating mechanism, while it is a known concept, its use in ConvNova is empirically validated for effectiveness in DNA modeling(Table 5). While our approach differs in several aspects, we acknowledge the contributions of MogaNet, StarNet, and MambaOut.\", \"title\": \"Response to Reviewer NQxy(1/4)\"}", "{\"title\": \"Response to Reviewer 2bPc(2/4)\", \"comment\": \">Beyond demonstrating superior accuracy compared to Transformers and SSMs, it would be valuable to include an in-depth discussion and comparison with models like LegNet on block design, architecture, parameter count, and training schemes in the main text.\\n\\nThank you for the suggestion. As our main focus mentioned above, we do not prioritize detailed comparisons with specific models like LegNet in the main text. However, we appreciate the importance of such comparisons and have provided the following details for reference: .\\n\\n**Training Schema**\\n1. LegNet's original training involved prediction of expression bin probabilities, but such tasks are absent in the foundation model benchmarks. Thus both models use original labels for supervision.\\n2. LegNet uses Lion optimizer in their original implementations, but for fair comparison, we use AdamW consistently across all models.\\n3. LegNet is trained in a fully supervised, from-scratch manner, whereas ConvNova leverages pretraining.\\n4. Even without pretraining, a from-scratch ConvNova outperforms LegNet. You can see with Table 13 and the results coming soon.\\n\\n**Parameter Count**\\nConvNova has multiple versions to ensure fair comparisons with different models. For comparison with LegNet, we used a 1.7M parameter version of ConvNova, while LegNet has 2.1M parameters(with the output head).\\n\\n**Block Design**\\n1. ConvNova incorporates dual-branch structures and dilated convolutions, which are absent in LegNet.\\n2. ConvNova\\u2019s gating mechanism is implemented using convolution, whereas LegNet uses an MLP for gating.\\n3. ConvNova maintains a fixed hidden dimension throughout, while LegNet adopts a progressively shrinking hidden dimension sequence: 256, 128, 128, 64, 64, 64, 64.\\n4. LegNet employs group convolutions to reduce parameter count, while ConvNova does not.\\n\\nActually, we do not see much similarity between these two models. We hope this helps clarify the distinctions.\"}", "{\"comment\": \"Thank you for your thoughtful feedback and for raising your score! We truly appreciate your recognition of our work and the constructive suggestions. We plan to include a table, along with further empirical analysis, in future versions of the manuscript to better illustrate the motivation and contributions of our design. Thank you again for your support!\"}", "{\"comment\": \"Thank you for your message. We have included the relevant details in the paper revision. Please refer to A5, Figure 5, and Table 15 for the updated information.\\n\\nLet us know if you need further clarification!\"}", "{\"title\": \"Response to Reviewer NQxy(4/4)\", \"comment\": \">(Q2) Some questions about the hyper-parameters in the network. How to determine the kernel size and the dilated ratio in ConvNova? Are there any implementation details of the ablation studies and providing the corresponding configurations (in addition to Table 9)? Meanwhile, is there more ablation of different designs (like analysis in [1, 2, 3]) of the gating branches as mentioned in Eq. (1) and Appendix A.3?\\n\\nWe will provide additional ablation experiments soon.\"}", "{\"title\": \"Global Response\", \"comment\": \"We sincerely thank all the reviewers for their thoughtful feedback, constructive suggestions, and encouraging comments. Your insights have been invaluable in helping us refine our work and strengthen the manuscript. We appreciate your recognition of the contributions and potential impact of our approach and will continue to improve the paper by incorporating your suggestions. Thank you again for your time and effort in reviewing our submission.\\n\\nHere, we again highlight the key contributions of our work:\\n\\n1. In recent foundation model research, CNNs have largely ignored and not compared with existing methods. We are **the first to conduct a comprehensive comparison across diverse benchmarks, demonstrating that with proper design, CNNs can outperform other architectures**. This provides a clear answer to the question: **CNNs are still not surpassed by other architectures**.\\n\\n2. Previous CNN models, such as Basenji, fail to perform well on all DNA sequence tasks. In contrast, we introduce ConvNova, **the first DNA foundation model based on classical CNN architecture to achieve strong performance across most tasks**. Through extensive experiments, we highlight three critical design principles: a well-designed dilation mechanism, a dual-branch structure, and gating convolution.\\n\\n3. We are also **the first to explore the relationship between receptive fields and the local dependency characteristics of DNA sequences**. Our experiments reveal that **CNNs inherently possess inductive biases for neighborhood modeling, which may be essential for many DNA sequence applications**.\"}", "{\"metareview\": \"This paper makes a strong case for reconsidering convolutional neural networks (CNNs) as competitive alternatives to Transformers and state-space models (SSMs) in DNA foundation modeling. The authors propose ConvNova, a CNN-based architecture incorporating innovative elements such as dilated convolutions, gated convolutions, and a dual-branch framework. Through extensive empirical evaluation, the authors demonstrate that ConvNova outperforms state-of-the-art models across various benchmarks, including both short- and long-range DNA tasks. ConvNova's performance improvements are coupled with lower parameter counts and faster computation, addressing practical concerns of scalability and efficiency.\\n\\nThe paper's contributions are particularly noteworthy in a field increasingly dominated by attention-based models. By leveraging the inductive biases inherent in CNNs, such as neighborhood modeling, the authors effectively challenge the assumption that Transformers and SSMs are universally superior. Moreover, the paper underscores the importance of architectural simplicity and principled design, showing that well-tuned CNNs can excel in specialized domains like genomics.\", \"additional_comments_on_reviewer_discussion\": \"There were concerns about the novelty of ConvNova\\u2019s design, insufficient comparison with prior CNN models, and limited intuition behind its architectural choices. Reviewers also requested clarification on pretraining data and qualitative analysis of errors across model classes. The authors responded comprehensively, emphasizing the paper\\u2019s focus on revisiting CNNs rather than proposing fundamentally novel architectures. They added comparisons with prior models, detailed ablation studies, and reorganized the paper for clarity. The reviewers found these updates satisfactory, acknowledging the robustness of experiments and the practical significance of the findings.\"}", "{\"title\": \"Response to Reviewer 2bPc(with updated paper revision)\", \"comment\": \">Beyond demonstrating superior accuracy compared to Transformers and SSMs, it would be valuable to include an in-depth discussion and comparison with models like LegNet on block design, architecture, parameter count, and training schemes in the main text.\\n\\nWe have added the comparison between LegNet and ConvNova(trained from scratch) in Table 13 to showcase the superiority.\\n\\n>I found the organization of the paper somewhat confusing. For instance, the comparison of downsampling and dilation in Section 3.3 and Table 1 would be more appropriately placed in the experiments section. Additionally, given the variety of tasks with different state-of-the-art models and settings, I recommend adding a dedicated section that briefly outlines the experimental setup. This section should include the objectives of each task, descriptions of the baseline models, and the specific configurations used in each experiment.\\n\\nWe have moved the comparison of downsampling and dilation to Section 4.4, as you suggested. Additionally, we have made modifications and added further descriptions regarding the objectives, baseline models, and configurations of each task in Section A2.\\n\\n>How is the receptive field for the dilated convolutions determined?\\n\\nIf your question is how we determine the dilation rate, we have included relevant experiments and details in A3.2, Tables 11 and 12.\\n\\nWe appreciate your insights and hope these changes address your concerns. Please don't hesitate to reach out if you have any further questions.\"}", "{\"title\": \"Response to Reviewer PceR(2/2)\", \"comment\": \">There should be some more qualitative analysis in the main text, mainly I would like to see an understanding of how the different model classes compare in the types of errors they make. For example, does the smaller receptive field contribute to certain types of errors for tasks where longer range reasoning is necessary?\\n\\nThis is an excellent question. From our experiments, we observe that sufficiently large pretrained transformers(like NTv2) excel at tasks with short sequences (a few hundred bp) but minimal local dependencies, such as splice site donor and acceptor prediction. However, transformers struggle on tasks like H3K4me2, which heavily rely on local dependencies. This may be due to the lack of an inductive bias in self-attention that emphasizes adjacent tokens, as CNNs naturally do. \\n\\nSSM models, on the other hand, perform well on long-range tasks, and they do not exhibit specific type of error.\\n\\nTraditional CNNs, such as LegNet, often underperform on targeted tasks like splice site prediction. However, with our design adjustments, **CNNs achieve robust performance across most tasks**, including long-range ones like gene finding(Table 2) and variant effect prediction:\\n\\n| Params |**ConvNova(25M)** | **Caduceus(8M)** | **HyenaDNA(1.6M)** | **HyenaDNA(0.4M)** | **HyenaDNA(3.3M)** | **HyenaDNA(6.6M)** | **NT(50M)** | **NT(100M)** | **NT(250M)** | **NT(500M)** |\\n|---------|-----------------|---|-----------|--------------|--------------|--------------|--------|--------|--------|--------|\\n| Context Length (bp) | 10K | 131K | 1K | 16K | 32K | 160K | 12K | 12K | 12K | 12K | 196K |\\n| AUC-ROC | **0.742** | 0.715 | 0.705 | 0.704 | 0.713 | 0.706 | 0.714 | 0.722 | 0.721 | 0.719 |\\n| Accuracy | **0.667** | 0.666 | 0.648 | 0.649 | 0.658 | 0.647 | 0.661 | 0.664 | 0.661 | 0.657 |\\n\\n\\nWe hope these clarifications and additional results address the concerns and demonstrate the value of our approach.\"}", "{\"title\": \"Response to Reviewer SH7n(with updated paper revision)\", \"comment\": \"We have added a description of the pretraining data in Section 3.3 of the latest revision.\\n\\nWe also have added the motivation why we adopt the dual-branch design in Section 3.2. The formula in Section A.3.1 has been revised in the hope that it will help you discover how the dual-branch structure with two independent convolutions provides greater expressiveness compared to a single-branch convolution.\\n\\nMore ablation experiments are conducted in Section A.3.2, table 11 and table 12 to explain why we design the dilation rate mechanism.\\n\\nWe hope this provides the clarity you were looking for, and please feel free to reach out if you have any further questions.\"}", "{\"comment\": \"I appreciate the authors' additional experiments during the discussion period. Indeed, your experimental results are solid, demonstrating superior performance across several tasks.\\n\\nHowever, I believe what the authors and this paper are missing is the \\\"why\\\"\\u2014why your ConvNet design outperforms Transformers and SSMs. Since the reasons behind this superior performance remain unclear, the proposed method appears to me to be more of a hyperparameter tuning exercise. For instance, if I understand the experiments in Tables 11 and 12 correctly, the authors performed a brute-force search on the kernel size and dilation rate.\\n\\nAdditionally, should the title of the first column be \\\"Kernel Settings\\\" instead of \\\"Task\\\"?\"}", "{\"comment\": \"Thank you for your reply. I agree that the computational complexity and model sizes of CNNs are appealing features. Additionally, the authors highlight a key observation in their response: \\\"CNNs inherently possess inductive biases for neighborhood modeling, which may be crucial for DNA sequence tasks.\\\" If this is the case, could you at least provide empirical results to justify your argument? For example, demonstrating that the attention maps from Transformers and SSMs primarily focus on neighborhood sequences, thereby supporting the claim that CNNs are better options in terms of computational cost and performance for DNA sequence tasks.\"}", "{\"title\": \"Response to Reviewer NQxy's Q2(with updated paper revision)\", \"comment\": \">(Q2) Some questions about the hyper-parameters in the network. How to determine the kernel size and the dilated ratio in ConvNova? Are there any implementation details of the ablation studies and providing the corresponding configurations (in addition to Table 9)? Meanwhile, is there more ablation of different designs (like analysis in [1, 2, 3]) of the gating branches as mentioned in Eq. (1) and Appendix A.3?\\n\\nWe have provided ablation studies regarding dilation and kernel size in A3.2, Table 11 and Table 12. In short, increasing the kernel size and dilation generally improves performance. Although some tasks may benefit from smaller dilations, we choose a dilation rate of 4 to guarantee long sequences modeling capability for the foundation model. To reduce the number of parameters, we select a kernel size of 9.\\n\\nRegarding the gating and dual-branch designs, we have revised the ablation study in A3.1. The learning rate, optimizer settings and other settings for this ablation study are kept consistent with ConvNova. \\n\\nAs mentioned above, our goal is not to propose a highly novel CNN architecture, so we did not conduct further comparisons of different designs.\\n\\nWe hope this answers your questions!\"}", "{\"comment\": \"Thank you for your insightful comments. However, the premise of your question may be problematic. CNNs have long been established in DNA modeling [1],[2],[3], while SSMs and Transformers emerged later. Additionally, the DNA foundation model works [4],[5],[6],[7] did not compare Transformers and SSMs directly with CNN-based foundation models when introduced. Therefore, there is no evidence to suggest that Transformers and SSMs outperform CNNs.\\n\\nThe key question in this field isn't why CNNs outperform Transformers and SSMs, but rather whether Transformers and SSMs truly surpass CNNs, and if so, why that might be the case. Given the current focus in the field on Transformer and SSM architectures, revisiting CNNs and including them in comparisons is both necessary and timely. Our paper addresses this by showing that under well-designed settings for CNNs, neither Transformers nor SSMs have surpassed CNNs in this domain.\\n\\nHowever, we can still offer some insights into why CNNs can outperform Transformers and SSMs. CNNs inherently possess inductive biases for neighborhood modeling, which may be crucial for DNA sequence tasks. Additionally, CNNs have lower computational complexity compared to Transformers and SSMs, making them more efficient.\\n\\nRegarding the kernel size and dilation rate experiments, these are not the central focus of our paper. They are included to ensure rigor and provide a more comprehensive evaluation.\\n\\nWe encourage you to revisit the settings and focus of our paper, which may further clarify our contributions. Your feedback is greatly appreciated, and we welcome continued discussion on this topic!\\n\\n[1] Basset: learning the regulatory code of the accessible genome with deep convolutional neural networks. Genome Res. 2016\\n\\n[2] Sequential regulatory activity prediction across chromosomes with convolutional neural networks. Genome Res. 2018\\n\\n[3] Predicting effects of noncoding variants with deep learning-based sequence model. Nat Methods. 2015\\n\\n[4] The Nucleotide Transformer: Building and Evaluating Robust Foundation Models for Human Genomics. bioRxiv. 2023\\n\\n[5] DNABERT-2: Efficient Foundation Model and Benchmark For Multi-Species Genome. ICLR. 2024\\n\\n[6] HyenaDNA: Long-Range Genomic Sequence Modeling at Single Nucleotide Resolution. NeurIPS. 2023.\\n\\n[7] Caduceus: Bi-Directional Equivariant Long-Range DNA Sequence Modeling. ICML. 2024\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}", "{\"title\": \"Response to Reviewer PceR(with updated paper revision)\", \"comment\": \"We have provided ablation studies on dilation and kernel size in A3.2, Tables 11 and 12. For a comprehensive understanding, we recommend reviewing these tables alongside Table 5 to observe the impact of our proposed components on performance. We hope this helps clarify the details, and please feel free to reach out if you have any further questions.\"}", "{\"comment\": \"Dear reviewer:\\n\\nHere we provide the motivation to consider CNNs against other models for DNA modeling tasks, **which has not been discussed in other work.**\\n\\nBesides the computational complexity and model sizes of CNNs, **CNNs inherently possess inductive biases for neighborhood modeling, which may be crucial for many DNA sequence tasks**. We have already added relevant experiments and analyses; please refer to A5, Figure 5, and Table 15 for the updated information.\\n\\nHere we briefly explain how we support our hypothesis.\\n\\nSpecifically, the main issue with self-attention is that it does not inherently focus on neighborhood sequences, which could contribute to its suboptimal performance in many DNA modeling tasks. To support our claim, we modify the NTv2 model by adjusting the RoPE's \\ud835\\udf03 and initializing the bias of the \\ud835\\udc5e\\ud835\\udc58 linear layer to [0,0,\\u2026,1], while keeping other initializations consistent with NTv2 (std=0.02,mean=0). This adjustment increases the attention map's focus on neighborhood sequences, thereby enhancing the Transformer's inductive bias.\\n\\nOur approach is similar to the methodology in [1]. When tested on H3K14ac, a task with strong local dependencies, the results show significant improvement (34.42 vs. 46.65 with the enhancements). Similar results on other 2 tasks can be found in Table 15.\\n\\nHowever, this method provides only a naive way to add inductive bias to Transformers. It still does not surpass ConvNova trained from scratch. Further research is needed to fundamentally strengthen the Transformer's inductive bias for neighborhood modeling.\\n\\nWe welcome any additional questions or feedback you might have and would be glad to engage in further discussion.\\n\\n[1] RepMLPNet: Hierarchical Vision MLP with Re-parameterized Locality. CVPR. 2022.\"}", "{\"title\": \"Official Comment by Reviewer SH7n\", \"comment\": \"The rebuttal have addressed my concerns. I will maintain my positive score.\"}", "{\"title\": \"Response to Reviewer SH7n(1/3)\", \"comment\": \"We sincerely appreciate Reviewer SH7n's thoughtful feedback, and here we provide corresponding responses to address these concerns.\\n>The novelty of the approach is somewhat limited, as CNNs have already been applied to this setting, and the training procedure is broadly the same as in the HyenaDNA paper.\\n\\n>The CNN architecture used is based on gated convolutions, which has already been proposed elsewhere, limiting the novelty.\\n\\n>To better understand the novelty of the approach, a benchmark comparing with a more standard CNN architecture would have been useful.\\n> \\n> The paper seems to lack a clear motivation as to why this specific CNN architecture was proposed.\\n\\nIn recent foundation model research, CNNs have been largely ignored and not compared with existing methods. We are the **first to compare CNNs with other architectures, demonstrating that with the right design, CNNs can outperform them**. Indeed, this answers the question: **CNNs are still not surpassed by other architectures.** Thus, novel architecture is not the primary focus of our work.\\n\\nThe motivation behind our architecture design is following several key considerations. The gating mechanism could enable dynamic feature selection. The dual-branch design facilitates independent feature extraction and promotes complementary representation learning. Dilation serves two purposes: (1) downsampling, and (2) investigating the local dependency condition of downstream tasks.\\n\\nStandard CNN models like Basenji and LegNet are designed for specific tasks and lack the generalizability required for DNA foundation modeling (See Table 13). This broader scope differentiates our work from prior efforts.\"}", "{\"title\": \"Response to Reviewer NQxy(3/4)\", \"comment\": \"## For Questions:\\n>(Q1) Is there any empirical analysis or theoretical support to verify that the dilated convolution is capable and more efficient than self-attention or SSM modules on the DNA tasks? For example, the authors could visualize the reception field or analysis the learned patterns to show that the dilated convolutions with gating learn better.\\n\\nFor efficiency, dilated convolutions possess lower computational complexity compared to self-attention and SSM modules. For capability, we have conducted extensive experiments with varying sequence lengths, and the results demonstrate the effectiveness of ConvNova across DNA tasks. As for a potential explanation, we hypothesize that, similar to how enzymes interact with DNA in a sliding window fashion, DNA modeling may prioritize adjacent nucleotides. In this sense, the inductive bias of CNNs might be more advantageous than self-attention models, which may not emphasize local context in the same way. This could be a key reason for the success of CNN-based models in this domain.\\n\\nWhile we do attempt to visualize the receptive fields and analyze the learned patterns, we do not observe any particularly insightful results. We agree that interpretability is an important aspect, and this might require a more systematic and dedicated effort.\"}", "{\"title\": \"Qualitative Results\", \"comment\": \"Thank you for your thoughtful feedback and suggestions. Based on your recommendations, we conduct additional explorations to investigate the qualitative differences in the types of errors made by different model classes. The conclusion from our analysis is that there are **no significant differences in the error sequences between the model classes**. The results of our analysis can be found in this anonymous link: https://anonymous.4open.science/r/Error-Analysis-AB32.\\n\\nSpecifically, we focus on the H3K14ac task, which exhibited significant performance differences among the models (70.71 vs. 60.84 vs. 57.22). We analyze the error sequences from ConvNova, Caduceus, and NTv2 using t-SNE to visualize the distribution of error points from various angles. However, **no clear qualitative differences are observed in the distributions t-SNE figures**.\\n\\nTo further ensure rigor, we train a linear SVM on the error sequences to verify whether a hyperplane could effectively separate the errors from different models(0 for ConvNova and 1 for Caduceus). The results are shown below:\\n\\n| | precision | recall | f1-score | support |\\n|-------|-----------|--------|----------|---------|\\n| **0** | 0.55 | 0.51 | 0.53 | 600 |\\n| **1** | 0.45 | 0.48 | 0.46 | 492 |\\n\\nSimilar results are observed in another test where 1 represents error sequences from ConvNova and 0 from NTv2, as shown below:\\n\\n| | precision | recall | f1-score | support |\\n|-------|-----------|--------|----------|---------|\\n| **0** | 0.56 | 0.52 | 0.54 | 600 |\\n| **1** | 0.46 | 0.51 | 0.49 | 443 |\\n\\nThese results suggest that **no significant qualitative differences exist in the types of errors made by different model classes numerically**.\\n\\nWe have also used traditional DNA sequence clustering algorithms including MMseq2, ALFATClust and MeShClust v3. However, these sequence-identity-based algorithms still do not yield other insights and conclusion.\\n\\nWe welcome further discussion and appreciate your valuable insights. Please do not hesitate to share additional thoughts or suggestions!\"}", "{\"title\": \"Response to Reviewer SH7n(3/3)\", \"comment\": \">I would say that calling the model a \\u201cfoundation model\\u201d is somewhat misleading as the model used in the paper has 7M parameters and was pretrained for 14 hours on a single GPU.\\n\\nThank you for the comment. Following prior works like HyenaDNA, DNABERT-2 and Caduceus, we adopt the term \\\"foundation model.\\\" While our model is relatively small, we feel its evaluation across such diverse tasks justifies this terminology, though we appreciate differing interpretations.\\n\\n>Is the architecture inspired by the DNA task in some way?\\n\\nYes. The key factor influencing robust performance is the carefully designed dilation mechanism. We will provide the ablation for dilation, kernel size as well as local dependency for tasks soon. We will inform you once available.\"}", "{\"comment\": \">What is the intuition and rationale behind using dilation over self-attention in a DNA foundation model? And why is dilation outperforming self-attention?\\n\\nWe hypothesize that DNA modeling often prioritizes interactions between adjacent nucleotides, similar to how enzymes process DNA in a sliding window fashion. This local inductive bias of CNNs, particularly with dilation, may offer an advantage over self-attention, which lacks a built-in focus on local context. This is especially evident in tasks like H3K4me2. Additionally, dilated convolutions have lower computational complexity compared to self-attention, further contributing to their efficiency and performance in DNA modeling.\\n\\n>How is the receptive field for the dilated convolutions determined?\", \"the_receptive_field_can_be_calculated_using_the_formula\": \"$\\ud835\\udc45\\ud835\\udc39(\\ud835\\udc5b)=\\ud835\\udc45\\ud835\\udc39(\\ud835\\udc5b\\u22121)+(\\ud835\\udc58\\u22121)\\u00d7\\ud835\\udc51$\\nwhere $\\ud835\\udc45\\ud835\\udc39(\\ud835\\udc5b)$ is the receptive field of the current layer, $\\ud835\\udc58$ is the kernel size, and $\\ud835\\udc51$ is the dilation rate.\\nIf your question is about how the dilation rate ($\\ud835\\udc51$) is determined, we select it based on ablation study results (coming soon). Additionally, since ConvNova is designed as a DNA foundation model capable of handling long-range tasks, we choose a dilation rate of 4 to balance these requirements.\\n\\n>What factors contribute to the dual-branch design outperforming a single-branch approach?\\nThe dual-branch design facilitates independent feature extraction and promotes complementary representation learning. Because of the gating mechanism, dual independent convolution can add more expressiveness than single branch.\", \"title\": \"Response to Reviewer 2bPc(4/4)\"}" ] }

[ "Ruo-Jing Dong", "Yu Yao", "Bo Han", "Tongliang Liu" ]

[ "transformer mistakes", "token-level semantic depdendencies" ]

{ "note_id": [ "ziSLaF2agn", "xrah1HmiVK", "xZ6nqfkDUf", "v6rg5LFTmP", "niAT6wUtwp", "kRM90Revbp", "kBr9pAycBf", "i9WeF7TUyi", "gdqF4U8rp1", "bZAABrZfJS", "bNyxuPplZA", "VuhV83JAJz", "SgtB24UBfA", "SaWw1rSp4b", "RlhgZ4890n", "RZ7lAnx7Ta", "Os5Rq3tFlY", "JdFaRsGbcX", "Ip37JwDxhu", "Il230hoh6A", "EQhgy82HQO", "9RZeNJNUFF", "8PHpTpoqcT", "7d2qzFEzYd", "5QtHEXF2mn", "02KXfElIqI" ], "note_type": [ "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_review", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "decision", "official_comment", "official_comment", "official_comment", "official_comment", "meta_review", "official_comment" ], "note_created": [ 1732513220025, 1732548443627, 1732548944443, 1732548638496, 1730565939917, 1732503463004, 1732722363056, 1730498336026, 1732721762384, 1732638743931, 1730760466646, 1732548830522, 1730712195553, 1732551844832, 1732503271469, 1732511659159, 1732570365838, 1730688514020, 1732513096576, 1737523926619, 1732501828131, 1732501648026, 1732727163371, 1732547836663, 1734630009870, 1732503867990 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission8699/Authors" ], [ "ICLR.cc/2025/Conference/Submission8699/Authors" ], [ "ICLR.cc/2025/Conference/Submission8699/Authors" ], [ "ICLR.cc/2025/Conference/Submission8699/Authors" ], [ "ICLR.cc/2025/Conference/Submission8699/Reviewer_DNKk" ], [ "ICLR.cc/2025/Conference/Submission8699/Authors" ], [ "ICLR.cc/2025/Conference/Submission8699/Authors" ], [ "ICLR.cc/2025/Conference/Submission8699/Reviewer_mBPy" ], [ "ICLR.cc/2025/Conference/Submission8699/Authors" ], [ "ICLR.cc/2025/Conference/Submission8699/Reviewer_QEqH" ], [ "ICLR.cc/2025/Conference/Submission8699/Reviewer_iRzN" ], [ "ICLR.cc/2025/Conference/Submission8699/Authors" ], [ "ICLR.cc/2025/Conference/Submission8699/Reviewer_QEqH" ], [ "ICLR.cc/2025/Conference/Submission8699/Reviewer_mBPy" ], [ "ICLR.cc/2025/Conference/Submission8699/Authors" ], [ "ICLR.cc/2025/Conference/Submission8699/Authors" ], [ "ICLR.cc/2025/Conference/Submission8699/Reviewer_DNKk" ], [ "ICLR.cc/2025/Conference/Submission8699/Reviewer_cEUh" ], [ "ICLR.cc/2025/Conference/Submission8699/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission8699/Authors" ], [ "ICLR.cc/2025/Conference/Submission8699/Authors" ], [ "ICLR.cc/2025/Conference/Submission8699/Authors" ], [ "ICLR.cc/2025/Conference/Submission8699/Authors" ], [ "ICLR.cc/2025/Conference/Submission8699/Area_Chair_rrZ9" ], [ "ICLR.cc/2025/Conference/Submission8699/Authors" ] ], "structured_content_str": [ "{\"comment\": \"**Q4. [Annotations] Annotations The notations can be clearer.**\\n\\nIn the revised version, we will improve the notations and add definitions to make them clearer. We will also add a symbol list in the appendix.\\n\\n**Q5. [question-answering tasks] It seems that only one NLP task (QA) has been investigated. It seems that wrongly aggregate information may affect tasks like reasoning and in-context learning. Maybe the authors can investigate more tasks to obtain more comprehensive results.**\\n\\nThank you for the suggestion. Though the scope of our study is currently limited to question-answering (QA) tasks, this setting allows for a **controlled analysis** of how semantic information propagates through the model. We will incorporate explanations into the revised version to clarify our task selection rationale and outline potential directions for broader exploration.\\n\\nA5.1 (Choice of QA as the primary task) We chose the question-answering (QA) task because it is particularly well-suited for evaluating the impact of semantic dependency errors at the token level. QA tasks inherently involve understanding and associating tokens in a question with those in the context, making them ideal for testing the model's ability to handle complex dependencies. This directly aligns with the focus of our study, which explores how semantic dependencies lead to errors in aggregation.\\n\\nA5.2 (Need for Ground Truth Datasets) To validate our findings, it is crucial to have ground-truth datasets that clearly present correct and incorrect dependencies. QA tasks provide such datasets, where the answers are explicitly tied to certain context tokens. These datasets enable us to systematically evaluate how dependency errors between question and context tokens contribute to prediction errors.\\n\\nA5.3 (Applicability to Other Tasks) \\n\\nWe acknowledge that wrongly aggregated dependencies may affect other NLP tasks, such as reasoning and in-context learning. In future work, we plan to investigate tasks where dependency relationships are less explicit, such as natural language inference and commonsense reasoning.\"}", "{\"comment\": \"**A1.2.2 Robustness Study**\\n\\n(Existing Work) Robustness studies (Jia & Liang, 2017) show that adversarial sentences can drastically reduce the performance of machine reading comprehension models. By analyzing model performance on various types of adversarial examples, they attribute the decline to broader issues, such as the model's tendency to rely on surface-level features like word overlap and positional cues.\\n\\n(Difference with Our Work) Our study seeks to explain this performance decline from the perspective of the model's internal mechanisms, providing an underlying reason for the observed accuracy drops in robustness studies. We found **the rank of semantic dependency strength encoded in a token changes** when adding irrelevant context or simply changing the order of the context sequence. We also provide a statistical evaluation method rather than observing the model\\u2019s output performance in a general study.\\n\\n(Importance)\\nOur insight can further help training or finetuning a robust language model in which the rank of encoded semantic dependency within tokens is stable when given irrelevant context in prompts.\\n\\n**Q1.3 [prior work] Additionally, in many cases it seems the authors invent new metrics or experimental methodologies without discussing how such metrics or methods compare with those proposed by prior work.**\\n\\nThank you for the suggestion. We will further discuss how our proposed metrics differ from or complement existing metrics in related works. We added related works and discussions in the appendix.\\n\\n**Section 2**\\n\\n**Q2.1 The definition of\\u00a0`P(f_\\\\theta)`\\u00a0was difficult to parse. Should\\u00a0`M`\\u00a0be\\u00a0`N`\\u00a0and\\u00a0`m`\\u00a0be\\u00a0`i`? It might also be helpful to remind the reader of the indicator function notation being used.**\\n\\nWe appreciate the suggestion regarding the notation. To clarify, in our definition of $P(f_\\\\theta)$, M represents the number of tested token cases across all sequences, not the number of tokens N in a single sequence. Each test case may include multiple sequences, making M the broader measure. In the revision, we will add an explicit explanation of M to distinguish it from N for better reader comprehension.\\n\\n**Q2.2 It seems like it would make more sense to refer to this quantity as a \\u201cpercentage\\u201d than a \\u201cprobability\\u201d? (Here and elsewhere in the paper, too)**\\n\\nThanks for your suggestion. we will change this in the later version.\\n\\n**Q2.3 Are there metrics from prior work that could have been chosen instead to measure the saliency of various input tokens on contextualized representations in the output?**\\nThank you for your question. \\n\\nA2.3.1 Saliency refers to the importance of specific input features (such as tokens) in contributing to the model\\u2019s final output. Semantic dependency is the relationship between words in a sentence where the meaning of one word depends on another word in the sentence \\n\\nA2.3.2 Our study mainly focuses on how **token-level semantic dependency** influences model performance through the token perturbation method, which is different from works that study input tokens\\u2019 saliency on representation output.\"}", "{\"comment\": \"**Q4.2 The experimental methodology seems a bit circular. The intended conclusion is that for cases where the model chooses the incorrect answer, the incorrect answer has higher saliency towards the question tokens than the correct answer. This doesn\\u2019t necessarily seem to be surprising, and it\\u2019s not clear what is actionable about this finding.**\\n\\nWe appreciate your feedback.\\n\\nA4.2.1 (Critical Gap) Given a specific QA task, Our study can explain why the model outputs wrong answers from **a semantic dependency perspective**. For example, in the case of a question like, \\u201cWhere do A live?\\u201d with context stating, \\u201cA lives on an island. B lives in a mountain.\\u201d the model may incorrectly output \\u201cmountain\\u201d instead of \\u201cisland.\\u201d We can examine the relationship between \\u201cmountain\\u201d and question tokens to see if they are falsely dependent. The prior study usually works by **analyzing model performance** on various types of adversarial examples and attributing the decline to broader issues, such as the model's tendency to rely on surface-level features like word overlap and positional cues. But we offer a finer-grained explanation for model errors.\\n\\nA4.2.2 (Potential Applications) We have discussed future model design in the Introduction. Future research could refine attention mechanisms to better prioritize meaningful\\ntoken interactions. Also, the **semantic dependency between tokens can be localized** (This part is added to this section and appendix). We have designed a method based on token perturbation to localize the attention head groups that are responsible for the semantic dependency between tokens. We can also evaluate how much a single attention head contributes to a token-level semantic dependency. This provides insights for finetuning attention heads according to specific model errors.\\n\\n**5. [Nits] The terms \\u201cunfaithful\\u201d and \\u201csemantically dependent information\\u201d appear in the abstract without clear definitions. It wasn\\u2019t clear to me what their meaning was in this context. Perhaps this could be clarified to improve readability of the abstract.**\\n \\nWe appreciate your suggestion. We will change the term \\\"unfaithful\\\" into \\u201cincorrect\\u201d and change the second sentence to \\u201cTo understand the cause of this issue, we explore how semantic information is encoded and semantic dependency is learned within the model. And we will define \\\"semantic dependency\\u201d and \\u201dsemantic information\\\" in the Introduction to improve accessibility for readers as follows:\\n \\n**Semantic information** refers to the meaningful content that consists of data or representations that carry meaning interpretable in a specific context. \\n \\n**Semantic dependency** can be defined as the relationship between words in a sentence where the meaning of one word (predicate) depends on another word (argument) in the sentence.\"}", "{\"comment\": \"**Q2.4 [clarify our finding 1]It wasn\\u2019t clear what the take-away was for the finding here. Should we be surprised that the most salient input for a given output token is the corresponding input token?**\\n\\nThank you for your concern. It is **not intuitive** that the most salient input token for a given output token is often the corresponding input token after extensive processing by deep transformer models.\\n\\nA2.4.1 (**Semantic Preservation Across Layers**) Our focus is studying how semantic information is preserved in tokens. Even after 24 (BERT-large) or 32 (Llama3-7B) attention layers most tokens still primarily retain their original semantic information. **This is surprising because**, during the attention mechanism, tokens aggregate diverse semantic information from other tokens in the sequence. The fact that most tokens still predominantly reflect their initial semantics highlights a strong retention property, which is not inherently expected given the iterative aggregation of contextual information.\\n\\nA2.4.2 **(Model-Specific Insights)** The observation that 20% of tokens in GPT-2 do not primarily retain their original semantic information points to **a potentially distinct mechanism** in how GPT-2 handles token-level information. This could indicate a different trade-off in how semantic context is integrated across tokens, which might relate to its performance characteristics compared to other models like BERT or LLaMA. Understanding this divergence could reveal alternative modeling strategies that influence downstream tasks differently.\\n\\nA2.4.3 (Importance) The finding that the last-layer token retains its original semantic information is critical for multiple reasons. It ensures that semantic dependency analyses in later layers are meaningful and reliable. It also provides a stable basis for interpretability, enabling attention-based methods to trace token contributions accurately. Additionally, this property could inspire new architectures that further optimize **semantic stability**.\\n\\n**Q2.5 Given the significance of casual attention mask on the analysis, it might be worth explicitly denoting in Tables 1 and 2 which models are decoder-only vs. encoder-only.**\\n\\nThank you for this suggestion. We will explicitly denote which models are encoder-only (Bert series) and decoder-only (GPT, Llama).\\n\\n**Section 3** \\n\\n**Q3.1 [dependency parsing tool] This analysis relies on a neural dependency parser, so although the analysis is presented as analyzing the degree to which dependencies in contextualized representations encode \\u201csemantic dependencies\\u201d, it might be helpful to clarify that the analysis is actual measuring agreement with a neural dependency parser model rather than some \\\"ground truth\\\" notion.**\\n\\nThank you for your suggestion. \\n\\nA3.1.1 (Dependency Parsing Tool) We have mentioned in our paper that this analysis relies on semantic dependency data derived from SpaCy, a pre-trained neural network-based dependency parser. SpaCy generates syntactic dependency trees using robust neural architectures trained on large annotated corpora, offering a reliable approximation of semantic dependencies.\\n\\nA3.1.2(the Benefit of Using This Tool) To our knowledge, no token-level semantic dependency dataset with comprehensive human annotations exists. Constructing such a dataset would be prohibitively expensive and prone to omissions due to the complexity of identifying all dependent token relationships manually. We will emphasize that our methodology measures the alignment of model-derived token dependencies with those provided by SpaCy as a proxy for ground truth.\"}", "{\"summary\": \"This paper delves into the internal mechanisms of transformer models to explore how semantic information is propagated and aggregated across tokens, which can contribute to the errors produced by large language models (LLMs). Experimental results under four settings have illustrated several useful findings.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. The motivation and flow of this paper is clear.\\n2. Experimental results have demonstrated several useful findings.\", \"weaknesses\": \"1. Figure 1 is not illustrative enough. For example, in Figure 1c, it is hard to understand the key idea, why the two arrows are in the opposite direction? Why one sequence has green followings and the other does not? The answers cannot be found until reading line 112-120 (and still confusing).\\n2. Line 184: if the perturbation token is sampled randomly, it is possible that the semantic information of the original sentence does change a lot due to the perturbation token, meaning that the semantic information change may due to the dependency between this perturbation token and the other token in the original input. Thus, the authors may need to consider all the tokens in each sequence when calculating the average change of the jth token.\\n3. The experiment results in Table3 also reflects what I have mentioned in point 2.\\n4. The notations can be clearer.\\n5. It seems that only one NLP task (QA) has been investigated. It seems that wrongly aggregate information may affect tasks like reasoning and in-context learning. Maybe the authors can investigate more tasks to obtain more comprehensive results.\", \"questions\": \"Please see the above weakness section.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"**Q5. [Questions: Clarify Experimental Setting] In L303, are the 10,000 cases described a subset of the 600,000 token cases described in L249? If so, how were these cases chosen?**\\n\\nThank you for your question. In the revised version, we will include a detailed explanation in the corresponding experiments to make it clear.\\n\\nA5.1 (Datasets Preparation) The 10,000 cases are not a direct subset of the 600,000 token cases from L249. Instead, these 10,000 cases were derived from a specialized word dependence dataset we generated using SpaCy. This dataset includes sentences from the GLUE dataset, where each word (as one case) in the sentence is annotated with its semantically dependent word groups as standard dependency data.\\n\\nA5.2 (Experiment Setting) The 10,000+ token variation cases evaluate our method's alignment with the semantic dependencies in this proxy ground-truth dataset. Specifically, for each token variation, we calculated an alignment score against the standard dependency data to validate that the token behavior aligns with semantically dependent tokens. This evaluation demonstrates that our method effectively captures the semantic dependencies between tokens.\\n\\n**Q6. [Questions: Detecting Potential QA Errors] Detecting errors require you to know which tokens are incorrect to begin with. Do you have any potential application of detecting potential QA errors at scale without prior knowledge of the ground truth?**\\n\\nThank you for your question.\\n\\nA6.1 Note that the aim of our method is not to predict how likely the answer is wrong given a question. There are some **unsupervised** evaluation methods based on confidence scores for this aim, such as:\\n\\n[1] Muttenthaler et al. 2020. Unsupervised Evaluation for Question Answering with Transformers.\\n\\n[2] Deng et al. 2023. Characterizing Prediction Matrix for Unsupervised Accuracy Estimation.\\n\\nA6.2 The aim of our method is to explain the reason why the model makes mistakes in Q&A. Given a large number of Q&A datasets, we can effectively use them. Our detection method is also fully automatic and can run at scale.\"}", "{\"comment\": \"Thank you for your positive response and for increasing your score. We appreciate your constructive feedback and will continue refining our work.\"}", "{\"summary\": \"This paper uses perturbation analysis to understand how individual tokens in the input of a Transformer affect the contextualized representations at the output of a Transformer. Essentially, they randomly replace various input tokens and measure the L2 difference on the encoded token representations. The paper studies how these dependencies relate to linguistic notions of semantic dependency. The paper also studies how seemingly irrelevant context affects the contextualized representations. Finally, the paper studies the correlation in prediction errors with deviations from the expected dependency relations.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": [\"It was nice to see comparisons across models spanning from BERT to LLama3. It was interesting to see trends across the various analyses with respect to modeling advancements across time.\", \"It\\u2019s potentially interesting to connect errors in model predictions to errors in propagating information correctly between tokens.\"], \"weaknesses\": [\"Overall weaknesses:\", \"The paper is structured as a sequence of loosely related experiments and analysis. It was a bit difficult to understand which findings were new and interesting. I think the paper could be improved by *removing* some content or deferring it to the appendix in order to focus and expand on the most interesting results.\", \"It would have been nice to contextualize the methods and findings more in prior work, e.g. probing studies (e.g. summarized in Rogers et al. 2020, https://arxiv.org/abs/2002.12327) and studies related to robustness to adversarial changes in context (e.g. Jia & Liang 2017, https://arxiv.org/abs/1707.07328). It would have been nice to have included discussion of in what cases the authors' findings are validating those of prior work vs. contradicting.\", \"Additionally, in many cases it seems the authors invent new metrics or experimental methodologies without discussing how such metrics or methods compare with those proposed by prior work.\", \"Section 2\", \"The definition of `P(f_\\\\theta)` was difficult to parse. Should `M` be `N` and `m` be `i`? It might also be helpful to remind the reader of the indicator function notation being used.\", \"It seems like it would make more sense to refer to this quantity as a \\u201cpercentage\\u201d than a \\u201cprobability\\u201d? (Here and elsewhere in the paper, too)\", \"Are there metrics from prior work that could have been chosen instead to measure the saliency of various input tokens on contextualized representations in the output?\", \"It wasn\\u2019t clear what the take-away was for the finding here. Should we be surprised that the most salient input for a given output token is the corresponding input token?\", \"Given the significance of casual attention mask on the analysis, it might be worth explicitly denoting in Tables 1 and 2 which models are decoder-only vs. encoder-only.\", \"Section 3\", \"This analysis relies on a neural dependency parser, so although the analysis is presented as analyzing the degree to which dependencies in contextualized representations encode \\u201csemantic dependencies\\u201d, it might be helpful to clarify that the analysis is actual measuring agreement with a neural dependency parser model rather than some \\\"ground truth\\\" notion.\", \"It would have been useful to connect the methodology and findings with prior work that has studied the degree to which Transformer representations capture linguistic notions of syntactic or semantic dependencies.\", \"Prior work on probing could provide an alternative to the perturbation-based analysis, which has some limitations (as discussed by the authors towards the end of the paper).\", \"Section 4\", \"I think this section is potentially the most interesting, because it could be interesting to attribute prediction errors to cases where the model appears to misunderstand the syntactic or semantic dependencies between tokens. However, I have some concerns about the methodology.\", \"It would be useful to show the % of cases where the incorrect answer has higher saliency than the correct answer for \\u201ccorrect cases\\u201d as a comparison point for Table 4.\", \"The experimental methodology seems a bit circular. The intended conclusion is that for cases where the model chooses the incorrect answer, the incorrect answer has higher saliency towards the question tokens than the correct answer. This doesn\\u2019t necessarily seem to be surprising, and it\\u2019s not clear what is actionable about this finding.\", \"Nits\", \"The terms \\u201cunfaithful\\u201d and \\u201csemantically dependent information\\u201d appear in the abstract without clear definitions. It wasn\\u2019t clear to me what their meaning was in this context. Perhaps this could be clarified to improve readability of the abstract.\"], \"questions\": \"Please see weaknesses.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you for your feedback. We have carefully addressed your concerns in our revised paper and response. To further improve the quality of our work, **we would greatly appreciate it if you could share the specific points or concerns that you believe require additional clarification or improvement**. We will provide further revisions and explanations.\"}", "{\"comment\": \"I thank the authors for their response.\\n\\n1. Applications and possible interventions\\n\\nIt is true that existing work has found the phenomenon that models can be distracted by irrelevant context, while this work finds an internal mechanism in the transformers that correlates with this. I agree with reviewer's cEUh's perspective that we do not understand the mechanism well enough. The paper would greatly benefit from a deeper investigation into the mechanism. For example, can you do some intervention to correct the mistakes when the model does a QA task wrongly? \\n\\n2. Tasks past QA\\n\\nI find that that the tasks only being QA where the answer exists in the context to be extremely limiting for the paper.\\n\\n---\\n\\nOverall, after reading the other reviews and authors reponses, my thought process is as follows: \\n\\n(1) Does this paper show some surprising result about how transformers make mistakes? No, the failure modes are well understood externally, though the authors do show some internal mechanism that correlates with the failure modes.\\n\\n(2) If not, does the paper then convincingly show that the mechanism exists in a meaningful way and is causally responsible for certain behaviors of the model? Unfortunately, I don't think the paper has hit this bar. I would encourage the authors to do further work into understanding how this mechanism work and how significant this mechanism is in the model's functioning.\"}", "{\"summary\": \"This paper investigates how the semantic dependency encoded by tokens change within the model architecture, and how it influences the prediction of the model. The authors discuss that (1) many tokens' semantics do not change with deeper layers; (2) a token also encode information of semantically related words; (3) changes of context, even irrelevant, can change the semantic dependency; (4) when models make mistakes, there are erroneous semantic dependency.\", \"soundness\": \"2\", \"presentation\": \"1\", \"contribution\": \"2\", \"strengths\": \"This paper tackles an important problem about understanding the model\\u2019s erroneous behavior by looking into the internal activations. The authors\\u2019 attempt to attribute the model\\u2019s prediction to semantic dependency can be a meaningful trial towards a more broad mechanistic understanding of the connections between internal activation and model behavior.\", \"weaknesses\": \"I feel the conclusions draw by the authors are in general discussed by existing literature and known by the overall language model community, and may not be strong enough to make a ICLR paper. Specifically:\\n\\n- The conclusion that the model\\u2019s activation does not change much through layers is already observed by a thread of work, based on which there are techniques derived like early exist: Liao et. al. 2021, Schuster et. al. 2022, Elhoushi et. al. 2024.\\n- The conclusion that language models behavior changes with perturbation by irrelevant context is observed in Shi et. al. 2024.\\n- The conclusion that model mistakes can be attributed to incorrect dependency is also observed in Wu et. al. 2024.\\n\\nGiven the listed works, I tend to feel that this work is more or less describing the conclusions that are already aware by the community (though with their own metrics).\\n\\nIn addition, I find it difficult to read and navigate notations from line 265, page 5 to line 347, page 7, mostly because the authors use a large set of symbols without clearly state their definition and the motivation of using them. This part of the paper may require significant rewrite and clarification. I would suggest the authors may compile a list of symbols to clarify the meaning of all symbols and the motivation of using them, and maybe use diagrams and specific examples to illustrate the procedure described in Section 3.\", \"references\": \"Liao et. al. 2021. A Global Past-Future Early Exit Method for Accelerating Inference of Pre-trained Language Models \\n\\nSchuster et. al. 2022. Confident Adaptive Language Modeling\\n\\nElhoushi et. al. 2024. LayerSkip: Enabling Early Exit Inference and Self-Speculative \\n\\nShi et. al. 2024. DecodingLarge Language Models Can Be Easily Distracted by Irrelevant Context\\n\\nWu et. al. 2024. Retrieval Head Mechanistically Explains Long-Context Factuality\", \"questions\": \"See above weakness section.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"**Q3.2 [Connection with prior work] It would have been useful to connect the methodology and findings with prior work that has studied the degree to which Transformer representations capture linguistic notions of syntactic or semantic dependencies.**\\n\\n(our change in revision) Thank you for your suggestion. We will add related work in the later version as follows:\\n\\nA3.2.1 (existing work) There are several prior works investigating how Transformer-based models capture linguistic properties. Hewitt and Manning (2019) demonstrated that BERT encodes syntactic tree structures in its vector space, allowing a probing classifier to reconstruct syntactic distances between words. Tenney et al. (2019) revealed that BERT encodes high-level linguistic features like entity types, semantic roles, and relations through probing tasks. Pimentel et al. (2020) utilized information-theoretic probing methods to quantify the mutual information between model representations and linguistic properties, reducing over-interpretation risks. Wu et al. (2020) proposed a parameter-free probing technique that analyzed the influence of syntactic subtree structures on MLM predictions.\\n\\nA3.2.2 (Difference) These works primarily investigate how models encode **syntactic and high-level semantic features**, such as entity relations or syntactic structures. In contrast, our study focuses specifically on **token-level semantic dependencies**, analyzing fine-grained interactions between individual tokens rather than task-specific feature aggregation or high-level semantic encoding. Moreover, we introduce an evaluation framework to **measure semantic dependency strength** between two tokens without relying on prior knowledge. Our approach also identifies false semantic dependencies that arise when the model produces incorrect answers. Unlike static syntactic or semantic structures, our framework captures the dynamic and context-sensitive semantic dependencies, which can vary irregularly across diverse scenarios.\", \"reference\": \"[1] Hewitt, J., & Manning, C. D. 2019. A Structural Probe for Finding Syntax in Word Representations \\n\\n[2] Tenney et al. 2019. BERT Rediscovers the Classical NLP Pipeline \\n\\n[3] Wu et al. 2020. Perturbed Masking: Parameter-Free Probing of Linguistic Structure in MLMs \\n\\n[4] Pimentel et al. 2020. Information-Theoretic Probing for Linguistic Structure \\n\\n**Q3.3 [Alternative methodologies] Prior work on probing could provide an alternative to the perturbation-based analysis, which has some limitations (as discussed by the authors towards the end of the paper).**\\n\\nA3.3.1 We appreciate your suggestion. Our study mainly focuses on how token-level semantic dependency influences model performance through a token perturbation method, specifically how semantic information propagates from input tokens to last-layer tokens. This requires:\\n\\n- Analyzing **token-level interactions** rather than an aggregation of task-specific features or high-level semantic features in models\\n- Evaluating **semantic dependency strength** between two tokens without prior knowledge.\\n- Evaluating false semantic dependency when the model makes mistakes\\n\\nA3.3.2 Probing methods cannot fully address these needs because they typically focus on **predefined linguistic tasks** (e.g., syntactic tree reconstruction or semantic role labeling) rather than dynamic semantic dependency shifts, which can not be used when semantic dependency are confusing and irregular, especially when we analyze why model make mistakes.\\n\\n**Section 4**\\n\\n**Q4.1 It would be useful to show the % of cases where the incorrect answer has higher saliency than the correct answer for \\u201ccorrect cases\\u201d as a comparison point for Table 4.**\\n\\nThank you for your suggestion. However, the correct case is not provided with the wrong model answer as a reference. Thus it is impossible to get cases with incorrect answer saliency when the model outputs the correct answer. But we do take correct semantic dependency into comparison, which is usually weaker than wrong dependencies when the model makes mistakes.\"}", "{\"summary\": \"This paper studies how semantic information from a sequence is encoded and aggregated in a single token position. They find that the information initially contained in a token is mostly retained at that position, that how the context information is aggregated in a token can be influenced by irrelevant context, and how some mistakes can be attributed to incorrect information propagation in a token.\", \"soundness\": \"2\", \"presentation\": \"1\", \"contribution\": \"1\", \"strengths\": [\"The paper does its experiments on different models across several model families (BERT, Llama, GPT).\", \"The framework of understanding mistakes are caused by erroneous information propagation is interesting. While it is likely that not all mistakes are caused by this type of error, this mechanism seems to be a promising way of understanding how transformers can get answers wrong.\"], \"weaknesses\": [\"The paper is hard to follow and read. For example, the key terms 'semantic information' and 'semantic dependence' is not defined. Semantic dependency seems to be defined later in section 3 at L273, but it seems to be a pretty general definition that makes categorising false dependencies in Section 4 based on a post-hoc 'dependency is false if the answer is wrong'.\", \"There seems to be not much applications or possible interventions based on these findings. For example, we already know that LMs can be distracted by irrelevant contexts or context changes [2].\", \"The claim of 'understanding mistakes' is too broad, as they study only question-answering tasks where the answer already exists verbatim in the context.\", \"The paper does not discuss any related works on interpreting how information flow in transformers to answer questions, such as [1].\", \"[1] Dissecting Recall of Factual Associations in Auto-Regressive Language Models\", \"[2] Large Language Models Are Not Robust Multiple Choice Selectors\"], \"questions\": [\"In L303, are the 10,000 cases described a subset of the 600,000 token cases described in L249? If so, how were these cases chosen?\", \"Detecting errors require you to know which tokens are incorrect to begin with. Do you have any potential application of detecting potential QA errors at scale without prior knowledge of the ground truth?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you for your response. I do think the paper would be improved with the proposed edits, in addition to the overall recommendations from the other reviewers. Perhaps with these revisions the paper may offer some interesting findings for an audience focused on interpretability and NLP. However, I will keep my original score, based on some of the factors mentioned in my original review.\"}", "{\"comment\": \"Thank you for the constructive comments on related work and definition.\\n\\n**Q1. [Clarify Definition] The key terms 'semantic information' and 'semantic dependence' is not defined. Semantic dependency seems to be defined later in section 3 at L273, but it seems to be a pretty general definition that makes categorising false dependencies in Section 4 based on a post-hoc 'dependency is false if the answer is wrong'.**\\n\\nWe appreciate the reviewer\\u2019s feedback on the absence of key word definitions. We will add them to the Introduction section of the revised version.\\n\\nA1.1 **Semantic information** refers to the meaningful content that consists of data or representations carrying meaning interpretable in a specific context.\\n\\nA1.2 **Semantic dependency** can be defined as the relationship between words in a sentence where the meaning of one word (predicate) depends on another word (argument) in the sentence.\\n\\nA1.3 In our case, **false semantic dependency** means the meaning of one word is not dependent on another. For example, in the sequence \\u201cblue sky and red apple,\\u201d the semantic dependency between the word \\u201cblue\\u201d and \\u201capple\\u201d is false. In the experiment of Section 4, we view the model\\u2019s wrong output token and the question token as a false dependency for evaluation.\\n\\n**Q2. [Applications or Possible Interventions] There seems to be not much applications or possible interventions based on these findings. For example, we already know that LMs can be distracted by irrelevant contexts or context changes**\\n\\nThank you for raising this concern. We will add the related work and address the applications or possible interventions based on our new findings in the revised paper.\\n\\nA2.1 (Existing Work) Existing work indeed discusses the **phenomenon** that irrelevant context in prompts can lead to a significant decline in model accuracy. Specifically, Shi et al. (2024) demonstrates that the inclusion of irrelevant context in prompts can lead to an erroneous focus on unrelated content, causing a significant decline in model accuracy.\\n\\nA2.2 (Difference With Ours) Our study tries to explain why and **provides an underlying reason** for the decline in model accuracy given irrelevant context in prompts. We found that **the rank of different semantic information (concepts) encoded in a token** changes dramatically with or without irrelevant context in prompts. Specifically, we show that the proportion of affected tokens\\u2019 semantic dependencies is rather high (around 25%\\u201340%) in different tested models.\\n\\nA2.3 (Our Extra Finding) Additionally, we also found that even **order changes of irrelevant context positions** can still affect the rank of different semantic information encoded in tokens while maintaining overall semantic information unchanged.\\n\\nA2.4 (Applications or Possible Interventions) We believe our insight can further help in training or fine-tuning a robust language model by constraining the rank of encoded semantic dependencies within tokens when given irrelevant context in prompts.\\n\\n**Q3. The claim of 'understanding mistakes' is too broad, as they study only question-answering tasks where the answer already exists verbatim in the context.**\\n\\nThank you for raising this concern. We have acknowledged this point as a limitation and future work. We agree that the scope of our study is currently limited to QA tasks where the answer exists verbatim in the context. This setting allows for a controlled analysis of how semantic information propagates across known input tokens. We will further emphasize this point in the Experiment section.\\n\\n**Q4. [Related Works] The paper does not discuss any related works on interpreting how information flow in transformers to answer questions**\\n\\nWe appreciate the suggestion to include related works, particularly those that focus on understanding information flow in transformers for QA tasks.\\n\\n(Our Change in Revision)\\nIn the revised version, we will include a related work section and discuss the paper you mentioned as follows:\\n\\nMor et al. (2024) analyze how factual associations are recalled in auto-regressive language models, highlighting the roles of MLP sublayers in enriching subject representations and attention heads in extracting attributes. Our study addresses a gap by studying how semantic information flows between tokens through attention layers in both non-auto-regressive (BERT) and auto-regressive models (GPT, Llama).\"}", "{\"comment\": \"**Q1. [Clarify Figure] Figure 1 is not illustrative enough. For example, in Figure 1c, it is hard to understand the key idea, why the two arrows are in the opposite direction? Why one sequence has green followings and the other does not? The answers cannot be found until reading line 112-120 (and still confusing).**\\n\\nThank you for highlighting the need to improve Figure 1 and its explanation. We will ensure that the following points are explicitly clarified in the main text near the figure:\\n\\nA1.1 (Opposite Arrows in Figure 1c) The arrows represent the semantic information propagation flow from the **token \\u201crhino\\u201d at layer 0** to a **group of tokens at layer L**. Tokens with darker blue backgrounds at layer L are more strongly related to the token \\u201crhino.\\u201d Red highlights indicate tokens whose semantic dependency rankings relative to \\\"rhino\\\" change when input context changes. For instance, \\u201cwhite\\u201d becomes more strongly associated with \\u201crhino\\u201d than \\u201cgray\\u201d when the irrelevant context \\u201capples are red\\u201d is added. The arrows' opposite directions are designed to visually align the tokens at the final layer for easier comparison. However, we recognize this may be confusing, and in the revised version, we will provide a more detailed caption to explain this aspect clearly.\\n\\nA1.2 (the green followings) The green background highlights represent **irrelevant context** in the sequence. In this example, \\\"apples are red\\\" serves as irrelevant context to \\\"white rhinos are grey.\\\" For the **context-change example** (left of Figure 1c), we compare the token \\u201crhino\\u201d across two scenarios: one where \\u201cwhite rhinos are grey\\u201d appears alone and another where irrelevant context is added. This comparison illustrates how relationships between \\u201crhino\\u201d and other tokens shift when irrelevant context is introduced. For the **order-change example** (right of Figure 1c), we change the irrelevant context position (highlighted with green background) but maintain overall semantic information unchanged to show order change can still affect the token relationship in \\\"white rhinos are grey\\\". \\n\\n**Q2. [clarify our experiment] Line 184: if the perturbation token is sampled randomly, it is possible that the semantic information of the original sentence does change a lot due to the perturbation token, meaning that the semantic information change may due to the dependency between this perturbation token and the other token in the original input. Thus, the authors may need to consider all the tokens in each sequence when calculating the average change of the jth token.**\\n(Experiment Setting) We appreciate your concern and clarify our experimental setup: In our experiments, we do not limit the calculations to a few randomly sampled tokens. Instead, we compute changes for nearly all tokens (95%) in each sequence, excluding special tokens such as [CLS] and [SEP]. This ensures a comprehensive evaluation of semantic dependencies across the input sequence. We will revise this experiment detail to ensure readers understand that the analysis accounts for the full token sequence rather than a subset.\\n\\n**Q3. [clarify our experiment] The experiment results in Table3 also reflects what I have mentioned in point 2.**\\n\\nA3.1 (What Table 3 Show) Thank you for your concern. In Table 3, the results show that tokens are primarily influenced by semantically dependent tokens in different models, further supporting the robustness of our approach.\\n\\nA3.2 (Datasets Preparation) In Experiment 3, we utilize SpaCy to generate a specialized word dependency dataset derived from the GLUE dataset. This dataset includes sentences from the GLUE dataset, where each word (as 1 case) in the sentence is annotated with its semantically dependent word group as standard dependency data. For each sentence in the dataset, SpaCy generates a set of semantically dependent tokens for nearly all tokens (over 90%, except for some special tokens like punctuation tokens).\\n\\nA3.3 (Experiment Setting) Experiment 3 evaluates our method's alignment with the semantic dependencies in the above specialized dataset. Specifically, for each token variation, we calculated an alignment score against the standard dependency data to validate that the token behavior aligns with semantically dependent tokens. The high alignment scores demonstrate that our method reliably captures the influence of semantically dependent tokens.\"}", "{\"title\": \"Thanks for the response.\", \"comment\": \"Thanks the authors for the clarification. I have increased my score correspondingly.\"}", "{\"summary\": \"This paper examines where information is stored in transformer language model tokens, and how it moves through the layers. For example, how \\\"white\\\" gets attached to rhinos in \\\"white rhinos\\\". The authors use minimal pair counterfactuals to judge how changing one token in the input of the model changes the value in a later layer. The authors use this to show that individual tokens contain mostly their own info ('rhinos' contains 'rhinos' information) and when models make mistakes when making predictions about some token, it is because the wrong information from context gets attached to that token.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"The authors outline several interesting research questions related to the movement of information within individual tokens and explain cases of irrelevant information moving into the representations of key tokens that cause incorrect answers. I think the questions are interesting and the problems being addressed have been observed for some time, although there has not been a principled study on movement of semantically dependent information across layers to my knowledge.\\n\\nThis work connects nicely to previous work on \\\"attention is not explanation\\\" (e.g., https://dl.acm.org/doi/10.1145/3447548.3467307 ), but i think it does not expand enough on this body of work\\n\\nThe authors consider a range of models, including encoder-only models like BERT, which I think is a strength for this kind of study\", \"weaknesses\": \"While this work is original, I'm not sure how much the findings weren't already understood to be the case. For example, distracting text causing performance drops is well known. The authors show that this is due to semantic information from other tokens being encoded in the correct token, but the methods that show the extent to which this is true do not explain enough about how this information is stored in the model. If we really understood how this was represented in (and intruding on) a given token, we would be able to remove it. This information belongs to some value vector(s) that copied the information in some attention operation. Can that be localized? I am generally positive about this paper, but I believe it could go a bit deeper in its analysis to better quantify why certain failure modes take hold.\\n\\nThe figures are unclear and sometimes poorly annotated. For example, figure 2 text should be much larger and it should be more apparent what \\\"score\\\" is. Figure 3, it is unclear what the numbers are next to the arrows\", \"typos_and_suggestions\": [\"L15: semantic dependency is redundant. I would suggest rewriting parts of the abstract to be a bit more descriptive\"], \"questions\": \"Can MLPs introduce erroneous semantic information in the early layers? for example, see https://arxiv.org/abs/2304.14767\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"**Q4. [Figures and Annotations] The figures are unclear and sometimes poorly annotated. For example, figure 2 text should be much larger and it should be more apparent what \\\"score\\\" is. Figure 3, it is unclear what the numbers are next to the arrows**\\n\\nThank you for the suggestion. We agree that the figures could be improved for clarity. In the revised version, we will enlarge the text and labels for Figure 2. We will note the \\\"Dependency Alteration Score (DAS)\\\" in Figure 2 and provide context for the numbers in Figure 3 as \\u201csemantic dependency scores from context tokens to question tokens.\\u201d We will add captions that clearly describe the content and purpose of each figure.\\n\\n**Q5. [Abstract Revision] L15: semantic dependency is redundant. I would suggest rewriting parts of the abstract to be a bit more descriptive**\\n\\nThank you for the suggestion. We will revise the abstract and provide a clear definition in the introduction for the term \\\"semantic dependency.\\u201d\\n\\n**Q6. [Question] Can MLPs Introduce Erroneous Semantic Information in Early Layers?**\\n\\nThank you for the question. \\n\\nA6.1 Our current work focuses on **attention layers** as the primary mechanism for token-level information propagation. However, we acknowledge that MLPs could play a role in introducing or amplifying erroneous semantic information.\\n\\nA6.2 (Our Change in Revision)\", \"we_will_add_this_to_the_discussion_and_future_work_in_the_later_version_as_follows\": \"Studies like Mor et al. (2024) (Dissecting Recall of Factual Associations in Auto-Regressive Language Models) have shown that MLPs encode enriched representations that propagate attributes. Such representations may inadvertently amplify irrelevant or erroneous semantic information. We aim to extend our analysis to quantify the contribution of MLPs to semantic dependency in the future.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"comment\": \"**Q3. [third finding overlap] The conclusion that model mistakes can be attributed to incorrect dependency is also observed in Wu et. al. 2024.**\\n\\nA3.1 (Existing work) This work is parallel with our work. The work shows that **some attention heads** contribute largely to model performance, as they retrieve factual information. In other words, some parameters are important and help encode factual information. The absence or malfunctioning of these retrieval heads may lead to model errors.\\n\\nA3.2 (Our finding) Our study explores how token-level **semantic dependency** influences model mistakes. Specifically, wrong tokens are likely to have stronger semantic dependency on question tokens than correct tokens when models make mistakes.\\n\\nA3.3 (importance) Our findings provide another perspective on understanding and correcting model mistakes under specific question-answering cases.\\n\\nA3.4 (our change in revision) Thank you for your comments, we realize that we should add more work on attention head into our related work. Specifically, we revised our related work as follows:\\n\\nExisting works have studied specific roles of attention heads to explain model errors. Wu et al. (2024) identify specific attention heads, termed retrieval heads, which are critical for retrieving factual information from long contexts. The absence or malfunctioning of these retrieval heads may lead to model errors. Gandelsman et al. (2024) show some attention heads in CLIP have property-specific roles (e.g., location or shape), which are important for model performance. Our study addresses another reason by exploring how token-level semantic dependency influences model mistakes, which provides another critical perspective on understanding and correcting model mistakes under specific question-answering cases.\", \"references\": \"Gandelsman et al. 2024. Interpreting CLIP's Image Representation via Text-Based Decomposition\\n\\n**Q4. rewriting**\\n\\nWe appreciate the reviewer\\u2019s feedback on the complexity of notations. In the revised version, we have revised annotations and added a **symbol list in the appendix (page 19)** to make them comprehensible.\"}", "{\"comment\": \"Thank you for the constructive comments regarding the related work. Our work is very different from the existing one. We have added your mentioned work into our related work to make it more comprehensive and avoid confusion.\\n\\n**Q1.\\u00a0[first finding overlap] The conclusion that the model\\u2019s activation does not change much through layers is already observed by a thread of work, based on which there are techniques derived like early exist: Liao et. al. 2021, Schuster et. al. 2022, Elhoushi et. al. 2024.**\\n\\nA1.1 (Existing work) Existing work, including Liao et al. (2021), Schuster et al. (2022), and Elhoushi et al. (2024), focuses on **model activation stability** in the later layers of transformer models. Specifically, later layers may contribute minimally to the refinement of token representations, which enables techniques like **early exit** to accelerate inference.\\n\\nA1.2 (Our finding) Our finding is that whether the $i$-th token of the last layer $L$ mostly contains its original semantic information from the $i$-th token in layer 0. This has not been studied by existing works.\\n\\nA1.3 (importance) The finding that the last-layer token retains its original semantic information is critical for multiple reasons. It ensures that semantic dependency analyses in later layers are meaningful and reliable. It also provides a stable basis for interpretability, enabling attention-based methods to trace token contributions accurately. Additionally, this property could inspire new architectures that further optimize semantic stability.\\n\\nA1.4 (our change in revision) We follow your constructive comments, and we revise the related work to make a clear difference with existing work as follows:\\nExisting work (Liao et al., 2021; Schuster et al., 2022; Elhoushi et al., 2024) has studied model activation stability in **later layers** of transformer models. Specifically, additional layers may contribute minimally to the refinement of token representations, which enables techniques like early exit to accelerate inference. However, whether the $i$-th token of the last layer $L$ mostly contains its original semantic information from the $i$-th token in layer 0 has not been studied. Our finding ensures that semantic dependency analyses are meaningful and reliable. The fact that most tokens still predominantly reflect their initial semantics highlights the model's strong retention property, which is not inherently expected given the iterative aggregation of semantic information across many layers.\\n\\n**Q2. [second finding overlap] The conclusion that language models behavior changes with perturbation by irrelevant context is observed in Shi et. al. 2024.**\\n\\nA2.1 (Existing work) Existing work discusses the **phenomenon** that irrelevant context in prompts can lead to a significant decline in model accuracy. Specifically, Shi et al. (2024) demonstrates that the inclusion of irrelevant context in prompts can lead to an erroneous focus on unrelated content, causing a significant decline in model accuracy.\\n\\nA2.2 (difference with ours) Our study tries to explain why and **provides an underlying reason** for the decline in model accuracy given irrelevant context in prompts. We found that **the rank of different semantic information (concepts) encoded in a token** changes dramatically with or without irrelevant context in prompts. Specifically, we show that the proportion of affected tokens\\u2019 semantic dependency is rather high (around 25%~40%) in different tested models.\\n\\nA2.3 (our extra finding) Additionally, we also found even **order changes of irrelevant context positions** can still affect the rank of different semantic information encoded in tokens while maintaining overall semantic information unchanged.\\n\\nA2.4 (importance) We believe our insight can further help training or fine-tuning a robust language model in which the rank of encoded semantic dependency within tokens is stable when given irrelevant context in prompts.\\n\\nA2.5 (our change in revision) We have also revised our related work accordingly as follows: Existing works (Shi et al., 2024) have demonstrated that the inclusion of irrelevant context in prompts can lead to a significant decline in model accuracy. Our study tries to explain why and provide an underlying reason for such a decline in model accuracy. We found the rank of different semantic information (concepts) encoded in a token changes when adding irrelevant context or simply changing the order of the context sequence.\", \"title\": \"Response by Authors\"}", "{\"comment\": \"Thank you for your feedback.\\n\\n**Q1. The paper would greatly benefit from a deeper investigation into the mechanism. For example, can you do some intervention to correct the mistakes when the model does a QA task wrongly?**\\n\\nWhile we do not yet provide a complete solution for removing intrusive information to correct model behavior, our current analysis offers a key step in this direction by **identifying specific attention heads responsible for the transmission of erroneous semantic information**. We have added this part in **Appendix A.4**. We can also evaluate how much a single attention head contributes to a token-level semantic dependency. \\n\\n**Q2. I find that the tasks only being QA where the answer exists in the context to be extremely limiting for the paper.**\\n\\nThank you for your concern. Though the scope of our study is currently limited to question-answering (QA) tasks where the answer exists in the context, this setting allows for a **controlled analysis** of how semantic information propagates through the model. We have incorporated explanations into the revised version to clarify our task selection rationale and outline potential directions for broader exploration.\\n\\n**A2.1 (Choice of such QA as the primary task)** We chose the question-answering (QA) task where the answer exists in the context because it is particularly well-suited for evaluating the impact of semantic dependency errors at the token level. QA tasks inherently involve understanding and associating tokens in a question with those in the context, making them ideal for testing the model's ability to handle complex dependencies. This directly aligns with the focus of our study, which explores how semantic dependencies lead to errors in aggregation. Knowledge or contexts beyond inputs will be a distraction to our analysis.\\n\\n**A2.2 (Need for Ground Truth Datasets)** To validate our findings, it is crucial to have ground-truth datasets that clearly present correct and incorrect dependencies. Such QA tasks provide such datasets, where the answers are explicitly tied to certain context tokens. These datasets enable us to systematically evaluate how dependency errors between question and context tokens contribute to prediction errors.\\n\\n**A2.3 (Applicability to Other Tasks)** We acknowledge that wrongly aggregated dependencies may affect other NLP tasks, such as reasoning and in-context learning. In future work, we plan to investigate tasks where dependency relationships are less explicit, such as natural language inference and commonsense reasoning.\\n\\n**Q3. Does this paper show some surprising result about how transformers make mistakes? No, the failure modes are well understood externally, though the authors do show some internal mechanism that correlates with the failure modes.**\\n\\nA3.1 Regarding your point about the external understanding of failure modes, **we would appreciate it if you could provide studies that well understand externally the specific examples of failure modes**. We believe if we are unaware of how internal structures work, it would be difficult to optimize and understand the model effectively. \\n\\nA3.2 For example, in the case of a question like, \\u201cWhere do A live?\\u201d with context stating, \\u201cA lives on an island. B lives in a mountain.\\u201d the model may incorrectly output \\u201cmountain\\u201d instead of \\u201cisland.\\u201d Knowing the existence of irrelevant context is not enough to explain why it happens. In our work, we can examine the relationship between \\u201cmountain\\u201d and question tokens to see if they are falsely dependent. The **prior study usually works by analyzing model performance** on various types of adversarial examples and attributing the decline to broader issues, such as the model's tendency to rely on surface-level features like word overlap and positional cues. However, we offer a **finer-grained explanation** for model errors, which provides an evaluation method for various language models and insights into fixing these errors.\\n\\n**Q4. If not, does the paper then convincingly show that the mechanism exists in a meaningful way and is causally responsible for certain behaviors of the model? Unfortunately, I don't think the paper has hit this bar. I would encourage the authors to do further work into understanding how this mechanism work and how significant this mechanism is in the model's functioning.**\\n\\nA4.1 In experiments across a large number of QA cases and different models, we observed a very high percentage indicating a strong correlation between false semantic dependencies and model errors. \\n\\nA4.2 Our further exploration in Appendix A.4 locates attention heads responsible for any token-level semantic dependency, indicating semantic dependency is mutually contributed by a small group of attention heads. We also found that the model\\u2019s attention head performance for semantic information storage differs in various QA cases.\"}", "{\"comment\": \"**1. Overall weaknesses**\\n\\n**Q1.1 [clarify presentation] The paper is structured as a sequence of loosely related experiments and analysis. It was a bit difficult to understand which findings were new and interesting. I think the paper could be improved by\\u00a0*removing*\\u00a0some content or deferring it to the appendix in order to focus and expand on the most interesting results.**\\n\\nThank you for this suggestion. The four findings are progressively related. In the revision, we refine the importance of the four findings to make them more clear and add a brief introduction to their relationships in the Introduction. We also focus on the most impactful results by moving some trivial parts into the appendix.\\n\\nA1.1.1 [**clarify finding 1 importance:** Most tokens primarily retain their original semantic information, even as they pass through multiple layers.] Our finding focuses on whether the **$i$-th token of the last layer L** mostly contains its original semantic information from the **$i$-th token in Layer 0**. This has not been studied by existing works. This is a fundamental prerequisite for studying token-level semantic dependencies. If the semantic information of the $i$-th token in layer 0 is very different from the $i$-th token in the last layer, our later study on token semantic dependencies would be invalid. These results directly support the feasibility of studying how semantic information propagates between tokens in transformer architectures.\\n\\nA1.1.2[**clarify finding 2 importance**: Semantically dependent information is usually encoded together within a token.] Here, we focus on investigating **truthful semantic dependency** in tokens. The experiment for the second finding validates that tokens are primarily influenced by semantically dependent tokens, showing our perturbation-based method captures semantic dependencies, reinforcing the robustness and reliability of our approach. \\n\\nA1.1.3 [**clarify finding 3 importance**: The semantic dependency within a token is sensitive to even irrelevant changes in context and order of prompts.] The third finding reveals how the model\\u2019s performance is affected by irrelevant context at the token level, evidenced by changes in **the rank of tokens\\u2019 semantic dependencies**. Furthermore, we observed that even altering the position of irrelevant context can impact token relationships within the sequence while preserving the overall semantic meaning. This highlights how token-level analysis uncovers additional insights beyond performance metrics, providing alternative measures for evaluating model behavior.\\n\\nThe above three findings are all prerequisites of studying how token-level semantic dependency influences model mistakes, which provide a critical perspective on understanding and correcting model mistakes. \\n\\n**Q1.2 [prior work] Contextualizing findings in prior work: It would have been nice to contextualize the methods and findings more in prior work, e.g. probing studies (e.g. summarized in Rogers et al. 2020) and studies related to robustness to adversarial changes in context (e.g. Jia & Liang 2017). It would have been nice to have included discussion of in what cases the authors' findings are validating those of prior work vs. contradicting.**\\n\\nThank you for this suggestion. We will enhance the discussion in related works by linking our findings to prior studies, such as probing (e.g., Rogers et al., 2020) and robustness studies (e.g., Jia & Liang, 2017). \\n\\n**A1.2.1 Probing Study**\\n\\n(Existing Work) Probing tasks (Rogers et al., 2020). have revealed how BERT encodes syntactic and semantic features across its layers, showing that lower layers focus on syntactic properties, while higher layers capture semantic dependencies. These studies often use lightweight classifiers or attention pattern analysis to uncover linguistic structures within BERT\\u2019s representations. \\n\\n(Difference with Our Work) However, probing tasks typically analyze static, pre-trained representations. In contrast, our work dynamically tests token-level semantic dependencies by introducing perturbations and measuring their effects on token semantic dependencies. This approach enables us to identify how falsely encoded semantic dependency lead to model mistakes. Moreover, we extend the scope of probing studies by examining the influence of irrelevant context specifically on the rank of token semantic dependenciy strength.\"}", "{\"metareview\": \"This paper studies how semantic dependency is captured in pretrained LMs. They find that many tokens do not change substantially over layers, and that analyzing this phenomena can be diagnostic of some of the errors made by models.\\n\\nDespite some interesting analyses, there are major weaknesses in the paper with regard to novelty (many of the findings are already known), as well as clarity/writing. I am therefore recommending that this paper be rejected.\", \"additional_comments_on_reviewer_discussion\": \"Many reviewers found the results not so novel in light of the large body of existing work in this area. Moreover, there was general consensus that the writing was unclear (e.g., no definition of \\\"semantic dependency\\\"), and the paper could be structured better. While the authors responded to such points in the rebuttal, the rebuttal did not change my opinion.\"}", "{\"comment\": \"**Q1. [Compare to Existing Work] While this work is original, I'm not sure how much the findings weren't already understood to be the case. For example, distracting text causing performance drops is well known.**\\n\\nThank you for addressing this concern. We will add related works and clarify our new findings.\\n\\nA1.1 (Existing Work) Existing work indeed discusses the phenomenon that irrelevant context in prompts can lead to a significant decline in model accuracy. For example, Shi et al. (2024) demonstrates that the inclusion of irrelevant context in prompts can lead to erroneous focus on unrelated content, causing a significant decline in model accuracy.\\n\\nA1.2 (Difference with Ours) Our study tries to explain why and provide an underlying reason for the decline in model accuracy given irrelevant context in prompts. We found that the rank of different semantic information (concepts) encoded in a token changes dramatically with or without irrelevant context in prompts. Specifically, we show that the proportion of affected tokens\\u2019 semantic dependency is rather high (around 25%\\u201340%) in different tested models.\\n\\nA1.3 (Our Extra Finding) Additionally, we also found that even an **order change of irrelevant context position** can still affect the rank of different semantic information encoded in tokens while maintaining overall semantic information unchanged.\\n\\nA1.4 (Importance) Our insight can further help in training or fine-tuning a robust language model in which the rank of encoded semantic dependency within tokens is stable when given irrelevant context in prompts.\\n\\nA1.5 (Our Change in Revision) We have also revised our related work accordingly as follows: Existing works (Shi et al., 2024) have demonstrated that the inclusion of irrelevant context in prompts can lead to a significant decline in model accuracy. Our study tries to explain why and provide an underlying reason for such a decline in model accuracy. We found that the rank of different semantic information (concepts) encoded in a token changes when adding irrelevant context or simply changing the order of the context sequence.\\n\\n**Q2. [Clarify Our Method] The authors show that this is due to semantic information from other tokens being encoded in the correct token, but the methods that show the extent to which this is true do not explain enough about how this information is stored in the model.** \\n\\nThank you for addressing this concern. \\n\\nA2.1 The token perturbation method can show how much semantic information from each token in the input sequence is stored in each token. Our experiment in Section 4 shows that question tokens are likely encoded with more semantic information from wrong answer tokens than correct tokens when the model makes mistakes.\\n\\nA2.2 For a token\\u2019s semantic dependency, we also designed a method to localize how the model\\u2019s attention heads contribute to it. We will introduce this method and corresponding results in the appendix and continue studying precise information storage in our future work.\\n\\n**Q3. [Potential Applications for Removing Erroneous Information] If we really understood how this was represented in (and intruding on) a given token, we would be able to remove it. This information belongs to some value vector(s) that copied the information in some attention operation. Can that be localized? I am generally positive about this paper, but I believe it could go a bit deeper in its analysis to better quantify why certain failure modes take hold.**\\n\\nWe appreciate the reviewer\\u2019s insightful comment on the potential to localize and remove erroneous information during attention operations. While we do not yet provide a complete solution for removing intrusive information, our current analysis offers a key step in this direction by **identifying specific attention heads responsible for the transmission of erroneous semantic information**. Our study is as follows:\\n\\nA3.1 **Semantic dependency can be localized**. We have designed a method based on token perturbation to localize the attention head groups that are responsible for the semantic dependency between tokens. We can also evaluate how much a single attention head contributes to a token-level semantic dependency.\\n\\nA3.2 However, we also found that the model\\u2019s attention head performance for semantic information storage differs in various QA cases, making it impossible to unify a group of specific heads for general mistakes. We will introduce this method and corresponding results in the appendix and continue studying precise information storage in our future work.\"}" ] }

[ "Zhen Liu", "Tim Z. Xiao", "Weiyang Liu", "Yoshua Bengio", "Dinghuai Zhang" ]

[ "Reward Finetuning", "Diffusion Models", "Alignment", "GFlowNet", "Generative Models" ]

{ "note_id": [ "zW2PCTjTfe", "xRoqJKjfG5", "xMotw4IEDc", "wk2sg1ATWu", "w8DF4fQthv", "slTxSOAXfV", "sPHGJfNZbO", "puDI282thk", "kjo7fzGIJu", "igHptFJpsC", "fu4spbPMf5", "fok4sWHWdA", "fhJfmJlSQ8", "fNr2sAY9Bo", "etmlpQLpRQ", "ZoJGXLJJJ9", "Ui79pKKhHH", "S9XzNTbyMM", "OY73Orbsin", "O2A1Iuf35a", "NtZkE9RCl1", "MjM9LqIFuF", "M3qkzcYUud", "L9UfYyc5fw", "Iskh93fAXI", "Fv58JFGDsu", "B0j5CoeHgV", "4w9EYXG1BF", "2T7149GaTr" ], "note_type": [ "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "decision", "official_comment", "meta_review", "official_comment", "official_review" ], "note_created": [ 1732499857098, 1730593399068, 1732190423909, 1733151004844, 1732189724813, 1730617055921, 1732191115251, 1730962572904, 1733206380279, 1732606172267, 1732191540952, 1733240969361, 1732363675583, 1730467043352, 1732504603794, 1733241144995, 1732487672212, 1732587162109, 1732986685687, 1732191855421, 1732192533251, 1733210152736, 1733074156204, 1732573977411, 1737523911491, 1732986404092, 1734674760589, 1732426207688, 1730721336765 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission8475/Authors" ], [ "ICLR.cc/2025/Conference/Submission8475/Reviewer_5N8Z" ], [ "ICLR.cc/2025/Conference/Submission8475/Authors" ], [ "ICLR.cc/2025/Conference/Submission8475/Authors" ], [ "ICLR.cc/2025/Conference/Submission8475/Authors" ], [ "ICLR.cc/2025/Conference/Submission8475/Reviewer_y62f" ], [ "ICLR.cc/2025/Conference/Submission8475/Authors" ], [ "ICLR.cc/2025/Conference/Submission8475/Reviewer_ZKLE" ], [ "ICLR.cc/2025/Conference/Submission8475/Reviewer_BEKY" ], [ "ICLR.cc/2025/Conference/Submission8475/Reviewer_5qSt" ], [ "ICLR.cc/2025/Conference/Submission8475/Authors" ], [ "ICLR.cc/2025/Conference/Submission8475/Authors" ], [ "ICLR.cc/2025/Conference/Submission8475/Authors" ], [ "ICLR.cc/2025/Conference/Submission8475/Reviewer_5qSt" ], [ "ICLR.cc/2025/Conference/Submission8475/Authors" ], [ "ICLR.cc/2025/Conference/Submission8475/Authors" ], [ "ICLR.cc/2025/Conference/Submission8475/Reviewer_ZKLE" ], [ "ICLR.cc/2025/Conference/Submission8475/Reviewer_y62f" ], [ "ICLR.cc/2025/Conference/Submission8475/Authors" ], [ "ICLR.cc/2025/Conference/Submission8475/Authors" ], [ "ICLR.cc/2025/Conference/Submission8475/Authors" ], [ "ICLR.cc/2025/Conference/Submission8475/Authors" ], [ "ICLR.cc/2025/Conference/Submission8475/Reviewer_5N8Z" ], [ "ICLR.cc/2025/Conference/Submission8475/Reviewer_5N8Z" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission8475/Authors" ], [ "ICLR.cc/2025/Conference/Submission8475/Area_Chair_q9LY" ], [ "ICLR.cc/2025/Conference/Submission8475/Authors" ], [ "ICLR.cc/2025/Conference/Submission8475/Reviewer_BEKY" ] ], "structured_content_str": [ "{\"comment\": \"We sincerely appreciate the reviewer's time and efforts in reviewing our response, and their further acknowledgement on our contribution. We are glad that we addressed your concerns.\\n\\nAs suggested by the reviewer, we have included the weighting scheme at the end of Section 3 in the main text.\\n\\nIn the meantime, we would also like to take this opportunity to further clarify and highlight the novelty and technical contributions of our work from two perspectives:\\n\\n**From the perspective of reward finetuning.** \\n\\nOur proposed method is a *practical, scalable and provable* method that\\n\\n- works and scales well on **large diffusion models**;\\n- has **theoretical guarantees** of unbiased optimal solution to the target distribution; and,\\n- achieves **Pareto improvements** on reward convergence and diversity & prior preservation on **all reward functions** we experimented with.\\n\\nWhile the derivation naturally follows GFlowNet principles, we believe that not only the novel combination of these principles, with the scalability to large models, is a significant step. Plus, we believe the introduction and application of this new GFlowNet method to reward finetuning of diffusion models builds significant, and perhaps previously overlooked, connections between fields.\\n\\n**From the perspective of GFlowNets.**\\n\\n- We propose the **first** GFlowNet objective that leverages **gradient signals** in training GFlowNets, while all other GFlowNet methods only use zeroth-order signals.\\n- Our method demonstrates scalability to large diffusion models, such as StableDiffusion with a substantial number of sampling steps\\u2014something that few existing GFlowNet approaches have achieved.\\n\\nWe believe these aspects constitute a novel contribution to the GFlowNet literature, which we regret not having fully explained in our paper or earlier responses. We sincerely hope this perspective provides further context for evaluating our work.\"}", "{\"summary\": \"The authors propose a method, called $\\\\nabla$-GFlowNet which which is modification of GFLowNets. They define a new objective $\\\\nabla$-DB, which is a gradient informed version the Detailed Balance objective. The paper then goes on to propose a residual version of this loss which at optimality samples proportionally to the argumented distribution $r(x_T)p^\\\\sharp (x_T)$, which maintains diversity of generations. The paper then presents experiments which shows a good pareto frontier of reward vs diversity on two tasks.\", \"soundness\": \"3\", \"presentation\": \"4\", \"contribution\": \"3\", \"strengths\": [\"This paper offers a good way at maintaining diversity in generations while aligning generations to a reward model.\", \"The paper is theoretically founded and shows that their new training objective maintains the validity of GFlowNets while taking into account gradient information.\", \"The work has a number of good ablations and experiments to show diversity and reward tradeoff.\"], \"weaknesses\": [\"The problem of bias is well known in optimizations for diffusion models (elaborated in question section). Some treatment of this problem problem would be desirable since it seems that this method may be optimizing for the biased distribution\"], \"questions\": \"- It is known that optimizing the KL constrained optimization problem, with the closed form solution of the augmented distribution, can lead to a biased result [0, 1]. Does this method have this bias problem? If not, how do you get around it?\\n\\n[0] https://arxiv.org/abs/2409.08861\\n[1] https://arxiv.org/abs/2402.15194\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"We appreciate the reviewer for acknowledging our contribution and raising their concerns.\\n\\n> I think the \\\"predicted reward\\\" estimation in Eq. 15 can be severely unreliable, especially for the high-noise time-steps of the diffusion model.\\n\\nThanks for your insightful comment. It is undoubtedly true that the predicted reward is never accurate. Indeed, this is the core reason why we need to learn the residual flow function to correct this error, which differentiates our method from others like ReFL (which also computes the predicted reward).\\n\\nBy design, our method does not rely on the predicted reward \\u2014 it is only a technique to set a good initialization for the flow function. If we are allowed to parameterize the flow function with a sufficiently large neural net and to spend much more compute (e.g., very large batch size) to optimize the flow function, we do not need this predicted reward because the detailed-balance (DB) consistency losses will propagate the reward signal on the terminal state back to all previous ones.\\n\\nIn practice, as one of our objectives is to achieve fast convergence for reward finetuning, we are not allowed to update the policy for too many steps with small learning rate / very large batch size. And indeed as what you may have been concerned about, the inaccurate reward prediction, under high learning rate, can lead to worse performance. Both to answer your question and to address this issue, we experimented to attenuate the scale of the predicted reward according to the diffusion time step \\u2014 the further away the diffusion time step is from the data distribution, the less weight we place on the predicted reward. We found that not only the performance (in terms of reward vs diversity trade-off) significantly improves with convergence speed slightly slower. To better demonstrate the effectiveness of our method, in the main text we show the results of our method with reward scaling, but show the comparison between w/ and w/o attenuation in Figure 20 and 21 in the appendix.\\n\\n> The parameter $\\\\lambda$ and the output regularization described in Page 7 seems to be crucial to the model's performance, but they are not the paper's contribution.\\n\\nThe regularization term is a common technique in the literature of reinforcement learning (especially for on-policy settings where the training trajectories are sampled from the policy currently being finetuned) to avoid too abrupt change in the policy during training, as it may lead to collapse of training. While it may be more common to find papers using KL divergence to do this (for instance, in the very popular methods of TRPO [1] and PPO [2]), we use Fisher divergence (therefore the L2 loss between diffusion policy outputs \\u2014- the gradients of the log probabilities). Indeed, the regularization term is used in one of our baselines DAG-DB [3]. Empirically, as long as this divergence is sufficiently large to prevent divergence in training loss, the results will be good. \\n\\n> I think a qualitative comparison on the HPSv2 can be more valuable and insightful about the model's performance.\\n\\nWe thank the reviewer for pointing this out. We have included more qualitative results on HPSv2 (with improved performance due to our modifications) in Figure 4 in the main text and Figure 25, 26 in the appendix. In addition, we now include the results on another reward function, ImageReward [4], to demonstrate the capability (Figure 5). \\n\\n---\\n\\n[1] Trust Region Policy Optimization. John Schulman, Sergey Levine, Philipp Moritz, Michael I. Jordan, Pieter Abbeel. ICML 2015\\n\\n[2] Proximal Policy Optimization Algorithms. John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, Oleg Klimov. https://arxiv.org/abs/1707.06347\\n\\n[3] Improving GFlowNets for Text-to-Image Diffusion Alignment. Dinghuai Zhang, Yizhe Zhang, Jiatao Gu, Ruixiang Zhang, Josh Susskind, Navdeep Jaitly, Shuangfei Zhai. https://arxiv.org/abs/2406.00633\\n\\n[4] ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation. Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, Yuxiao Dong. NeurIPS 2023\"}", "{\"comment\": \"We appreciate the reviewer's constructive comments.\\n\\nWe agree with the reviewer that the term FID is a bit misleading and we will term it instead as \\\"prompt-averaged FID\\\" (this practice follows the way that the adjoint matching paper does with diversity scores).\", \"regarding_limitations\": \"we will mention the limitations in our next draft, as we agree that flow matching is one of the mainstream models these days. In the meantime, we would like to point out that (which we are inspired by the reviewer to check), since the typical flow matching models like SD3 are those with *Gaussian conditional probability paths*, we have a very simple correspondence between the learned vector field $v(x,t)$ and the score function $\\\\nabla_{x_t} \\\\log p_t(x_t)$. For instance, for linear flow:\\n\\n$$\\\\nabla_{x_t} \\\\log p_t(x_t) = \\\\frac{t-1}{t}v(x_t, t) - \\\\frac{x}{t}$$\\n\\nSuch a relationship indeed induces a corresponding SDE in this flow matching setting. As a result, we are able to use SD3 as an initialization and use $\\\\nabla$-GFlowNet to obtained a finetuned diffusion model out of it. This is slightly different from directly obtaining a flow matching model, but arguably we may always compute the probability flow ODE or use DDIM-like solvers if ODE-style inference is preferred.\\n\\nDue to the extremely limited amount of time, we are less confident to show the reviewer empirical results before the end of the rebuttal period. Nevertheless, we are trying our best to produce the results, and at the same time hope our explanations above can alleviate some of your concerns.\"}", "{\"title\": \"General response to the reviewers\", \"comment\": [\"We sincerely thank the reviewers for their time and effort in reviewing our paper and making valuable suggestions. We have uploaded a revised draft of our paper. The major differences are:\", \"We have 1) fixed the numerical issue during training, 2) introduced **time-dependent weighting on the predicted rewards** of the residual $\\\\nabla$-DB loss. As a result, our proposed method is now able to achieves a **much better trade-off between reward, diversity and prior-following on all reward models**, shown in Table 1 and Figure 6, 7, 12 & 13 for the main experiments in the revised draft. We have therefore also revised our Figure 2 on the results using Aesthetic Score.\", \"To further demonstrate the effectiveness of our methods and in response to Reviewer zWHc and 5qSt, we include more qualitative results in Figure 3, 4, 5 as well as the last three pages in the appendix.\", \"We now introduce a **new hyperparameter to control the strength of the pretrained prior**, for which the ablation results are shown in Figure Figure 16 and 17.\", \"We now include a **new metric of prior preservation** by measuring the FID score (in general, lower the better) between samples generated by the finetuned model and the pretrained model (as our objective is $P_\\\\text{pretrained}(x)^\\\\eta * R(x)^\\\\beta$). For a fair comparison, we draw the Pareto frontier to show the FID scores of models at roughly the same reward. Figure 7 shows that our method is significantly better in this dimension.\", \"We now include **another reward function ImageReward** [1], a general-purpose text-to-image human preference reward model.\", \"In response to Reviewer y26f and 5qSt, we include some more details in the experiment section to avoid confusion.\", \"In Figure 2, we used to show the final results (trained with 200 update steps) of all methods. While it is indeed the case that our method is more robust, it might be a bit unfair as one may perform early stopping when finetuning models. Therefore, we now visualize the samples from the best model checkpoint (throughout the finetuning process) for each method with an average reward of the visualized set of images, while we qualitatively show the training collapse issue of the baseline methods in Figure 3.\", \"Due to the limited space in the paper, we moved some figures like the one for ablation study on trajectory subsampling rate to the appendix.\", \"In response to Reviewer y62f, we have experimented with a different scheduler. Specifically, we pick **SDE-DPM-Solver++** and construct an MDP. In Figure 18 and 19, we show that our method is still able to perform well.\", \"For every experiment we now show the standard deviation of metrics with 3 random seeds.\", \"We polished the figures, the captions the related work section for better flow.\"]}", "{\"summary\": \"The authors propose Nabla-GFlowNet (\\u2207-GFlowNet) to efficiently finetune pretrained diffusion models. This approach addresses issues of limited sample diversity and slow convergence in existing methods by leveraging reward gradients through \\u2207-DB and its variant, residual \\u2207-DB. Empirical results show that residual \\u2207-DB enables fast, diversity-preserving finetuning of models like StableDiffusion on various realistic reward functions.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"3\", \"strengths\": [\"The paper offers a comprehensive theoretical deduction of the proposed method, thoroughly explaining how the objectives nabla-DB and residual nabla-DB are derived.\", \"By introducing residual \\u2207-DB, the authors extend the applicability of their work to pretrained large-scale models, which is crucial.\", \"The paper enhances the quantitative evaluation of diversity in generated samples. By employing a broader range of metrics and more extensive comparisons.\"], \"weaknesses\": [\"The current experimental setting appears somewhat outdated. To enhance the study's relevance, please consider using more recent schedulers and pre-trained models instead of DDPM or Stable Diffusion 1.5.\", \"The qualitative results shown in Figure 2 are confusing. Additional explanation is needed to clearly demonstrate the superiority of \\u2207-DB, as DDPO and DAG-DB also exhibit strong performance.\", \"A user study would be helpful for evaluating diversity.\"], \"questions\": \"Please see weaknesses.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"We appreciate the review's acknowledge on our contribution and their valuable comments. We have made some further refinements on the paper presentation and the experiments (detailed in the general response).\\n\\n> The function $g_\\\\phi(x_t)$ is an interesting and reasonable choice for achieving the fitness task; however, it results in approximately zero vectors, with a terminal constraint of $g_\\\\phi(x_T)=0$. It remains unclear whether Unet is a suitable option for this purpose. \\n\\nGood point! While the $g_\\\\phi(x)$ function at the terminal state is zero, in general it is moderately far from zero for states far from the terminal state. Indeed, this is the reason why we resort to GFlowNet to solve this issue. We therefore argue that it is necessary to predict a vector field from the input, for which U-Net is a good choice. Indeed, the results with $w_B = 0$ in our paper shows that when the detailed balance constraints are not well obeyed, the model suffers from worse diversity and worse prior following, which partially demonstrates the importance of learning a good $g_\\\\phi(x_t)$.\\n\\n> The regularization term appears significant, with $\\\\lambda=1000$ in the Aesthetic Score experiments and $\\\\lambda=100$ in the HPSv2 experiments. However, Section 3.2 states that it \\\"may eventually over-optimize the reward and thus neglect the pretrained prior.\\\" Is Section 3.2 more helpful for preventing over-optimization than the regularization term?\\n\\nThe regularization is indeed more concerned about training stability, as the training examples are generated by the policy currently being finetuned (i.e., the so-called on-policy training in reinforcement learning literature). RL Algorithms like TRPO [1] and PPO [2] have similar regularization to avoid sudden collapse of the policy.\\n\\nWe stated the overfitting issue because we set a relatively high reward-to-prior ratio (i.e., the $\\\\beta$ parameter) for faster convergence. As a result, if we train sufficiently long (say several days), it is possible for the policy to completely ignore the prior.\\n\\n> The method is interesting and useful, but the paper's claim of \\\"Diversity-Preserving\\\" in the title creates a gap. Does this imply that the method theoretically ensures better diversity, or is it based solely on observational results? From my perspective, this method prioritizes improving information backpropagation for fine-tuning multistep sampling rather than enhancing diversity. \\n\\nThanks for raising your concern and we apologize for our unclear previous descriptions. \\n\\nBy \\u201cdiversity-preserving\\u201d, we mean that our method aims to **sample** from the target probability distribution, versus to **maximize** the probability distribution (as in DDPO and ReFL). Here we give a simple example: suppose that the policy is a single 1D Gaussian distribution (i.e., running diffusion for only one step) with both the mean and variance parameters learnable, and the target distribution is a mixture of two Gaussians of different variance and moderately far from each other. With the **reward maximization** objective, the optimal solution is to have the policy to always (i.e., deterministically) output the point with the maximum probability (i.e., a Dirac delta distribution), while with the **sampling objective** one aims to minimize the distribution distance between the policy and the target probability, which in general leads to non-zero overlap between distributions in the visualization of both distributions.\\n\\nTherefore, we argue that these non-sampling baseline methods (DDPO, ReFL and DRaFT) are more prone to mode collapse (since if they find a mode in the distribution, there is less incentive for these models to discover other modes). \\n\\n---\\n[1] Trust Region Policy Optimization. John Schulman, Sergey Levine, Philipp Moritz, Michael I. Jordan, Pieter Abbeel. ICML 2015\\n\\n[2] Proximal Policy Optimization Algorithms. John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, Oleg Klimov. https://arxiv.org/abs/1707.06347\"}", "{\"summary\": \"The paper addresses the problem of fine-tuning the pretrained diffusion models on a target reward while aiming for 1) preserving diversity of generated images and 2) fast convergence. It proposes Nabla-GFlowNet to do so, inspired by generative Flow Nets (GFlow-Nets) that sample with unnormalized density of the reward function. Experiments on different benchmarks show that the proposed method generally achieves the best diversity vs reward trade-off compared to baselines.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": [\"The proposed idea is based on the generative flow nets, which makes it intuitive and straightforward.\", \"The Nabla-GFlowNet can leverage the first order information of the reward function (gradient) while the baselines only use the zero-order information.\", \"The experimental results show that the proposed method can generally achieve the best diversity vs. reward trade-off frontiers.\"], \"weaknesses\": [\"I think the \\\"predicted reward\\\" estimation in Eq. 15 can be severely unreliable, especially for the high-noise time-steps of the diffusion model. The predicted clean image will be noisy, and if the reward function is calculated by a model that has been trained on not noisy images, the predicted reward will be inaccurate.\", \"The parameter \\\\lambda and the output regularization described in Page 7 seems to be crucial to the model's performance, but they are not the paper's contribution.\", \"The qualitative samples that are compared with the baselines are only for the aesthetic score. I think a qualitative comparison on the HPSv2 can be more valuable and insightful about the model's performance.\"], \"questions\": \"- I suggest that the authors include the qualitative comparative results for the HPS-v2 reward in the paper.\\n\\nIf the authors address my concerns, I am willing to increase my score.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"The author partially addressed my concerns, and considering everything, I agree to accept this paper.\\n1. I suggest that the author should include these points in the latest version of the paper, especially regarding diversity, as it was mentioned in the title.\\n2. I have an additional question; given the deadline, the author does not need to respond but I recommend considering it for future work. I'd like to ask if a good reward could completely eliminate the need for a regularization term or whether the existence of a regularization term is merely a patch for currently unexplained issues within rewards.\"}", "{\"comment\": \"Thank you for your detailed response. I appreciate the clarifications and the efforts you've made to address most of my concerns. I have decided to raise my rating to 6.\"}", "{\"comment\": \"We thank the review for the valuable comments and suggestions. We have made some further refinements on the paper presentation and the experiments (detailed in the general response).\\n\\n> The current experimental setting appears somewhat outdated. To enhance the study's relevance, please consider using more recent schedulers and pre-trained models instead of DDPM or Stable Diffusion 1.5. \\n\\nThank you for raising this concern. We would like to first state that our paper follows the common practice in evaluating reward finetuning methods, for instance [1,2,3], almost all of which finetune with DDPM on StableDiffusion v1.5.\\n\\nThat\\u2019s said, it is definitely interesting to see how our method may generalize to some other MDPs constructed by different diffusion SDE solvers. For this purpose, we constructed another MDP with the schedule of SDE-DPM-Solver++ [4] and showed in the appendix (Figure 18 and 19) that our method still works well in this setting.\\n\\nRegarding pretrained models, SD v1.5 is probably the best open-sourced large diffusion models for verifying different methods. SDXL can be too large and require more compute that a typical academic lab does not afford. SD3 is a flow matching model (despite the word \\u201cdiffusion\\u201d in the name \\u201cStableDiffusion\\u201d), of which the sampling process is deterministic (given an initial Gaussian noise) by solving an ordinary differential equation (ODE). As GFlowNets are probabilistic models, they by design do not model deterministic processes and therefore we do not consider SD3 as a suitable model to benchmark methods for diffusion finetuning.\\n\\nAnd we would also like to share our insight that, since our method is mostly algorithm- and model-agnostic (as it is indeed derived from a general probabilistic model perspective) for diffusion models, we confidently believe that there is not much gap between the results on different pretrained diffusion models and different MDP.\\n\\n> The qualitative results shown in Figure 2 are confusing. Additional explanation is needed to clearly demonstrate the superiority of \\u2207-DB, as DDPO and DAG-DB also exhibit strong performance. \\n\\nWe apologize for our unclear presentation on this result. We have revised the figure and the captions.\\n\\nIn the revised draft (Table 1 and Fig 12 & 13), our method achieves the best performance (in the sense that it has comparable speed to ReFL while maintaining good diversity & prior preservation) with our latest modifications: \\n\\n- Introduction of attenuating scaling on the predicted rewards (so that the less reliable predicted rewards at time steps far from the generated samples are down-weighted)\\n- Fix on the numerical issues in mixed-precision training\\n\\n> A user study would be helpful for evaluating diversity.\\n\\nWe appreciate the reviewer\\u2019s suggestion \\u2014- we are currently working on it and hopefully we can report the results very soon.\\n\\n---\\n[1] Training Diffusion Models with Reinforcement Learning. Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, Sergey Levine. ICLR 2024\\n\\n[2] Directly Fine-Tuning Diffusion Models on Differentiable Rewards. Kevin Clark, Paul Vicol, Kevin Swersky, David J. Fleet. ICLR 2024\\n\\n[3] Improving GFlowNets for Text-to-Image Diffusion Alignment. Dinghuai Zhang, Yizhe Zhang, Jiatao Gu, Ruixiang Zhang, Josh Susskind, Navdeep Jaitly, Shuangfei Zhai. https://arxiv.org/abs/2406.00633\\n\\n[4] DPM-Solver++: Fast Solver for Guided Sampling of Diffusion Probabilistic Models. Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, Jun Zhu. ICLR 2023\"}", "{\"comment\": \"As it is coming to the end of the rebuttal period, we would like to thank the reviewer for their time and efforts and their continued overall positive rating, and we are glad that we addressed your concerns.\"}", "{\"comment\": \"We would like to thank the reviewer for their patience. Here are the results of a simple user study, where we randomly picked 5 categories in animal prompts and for each generated 16 images to test on our models finetuned on Aesthetic Score with (residual $\\\\nabla$-DB) and the baselines (Residual DB, DDPO, ReFL and DRaFT-LV). For each category, we present the corresponding images generated for all 5 methods and ask 3 questions:\\n1. Which set is the most aesthetic one?\\n2. Which set is the most diverse one?\\n3. Only consider those images you feel aesthetic. Now, which one is the most diverse?\", \"we_collected_27_responses_and_computed_the_win_rate_averaged_on_5_prompts\": \"| Method | Q1 | Q2 | Q3 |\\n|------------------|------------------|------------------|------------------|\\n| Residual $\\\\nabla$-DB | 63.7% | 52.6% | 61.7% |\\n| Residual DB | 7.4% | 0% | 0% |\\n| DDPO | 11.9% | 10.4% | 8.9% |\\n| ReFL | 5.9% | 18.5% | 13.3% |\\n| DRaFT-LV | 11.1% | 18.5% | 17.0% |\\n\\n\\nThe results of this user study are largely consistent with our quantitative results with the ground truth reward function (Aesthetic Score) and the proxy diversity score (DreamSim diversity).\"}", "{\"summary\": \"This paper introduces Nabla-GFlowNet with a new objective, $ \\\\nabla$-DB, and its variant, residual $ \\\\nabla$-DB, for finetuning pretrained diffusion models. The method aims to enhance sample diversity and finetuning efficiency by utilizing reward gradients. Experiments on two reward functions, Aesthetic Score and Human Preference Score (HPSv2), demonstrate improved performance.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": [\"The paper has a clear motivation, addressing the challenge of preserving diversity and improving efficiency in finetuning diffusion models, which is crucial for real-world applications.\", \"The paper introduces a unique application of GFlowNet principles to diffusion model finetuning, specifically focusing on the reward gradient to preserve sample diversity.\"], \"weaknesses\": [\"In the quantitative evaluation, the proposed method performs well on the smaller dataset of Aesthetic Score, nearly doubling the baseline in the DreamSim metric. However, on the larger dataset of HPSv2, its performance is similar to baselines, showing no clear advantage. The effectiveness of the method remains inconclusive due to the differences in dataset size.\", \"While the authors mention \\u201cfast and efficient finetuning\\u201d as a contribution, only Figure 4 shows comparable convergence speed to the baseline. It would be helpful to include details on training resource consumption, such as GPU usage and computational cost, to substantiate this claim.\", \"Figure 3 only compares the results of the pretrained model and the proposed method, lacking visual comparisons with other baselines.\", \"Figure 5 lacks sufficient information in the title and annotations to clarify what each point represents (e.g., training iteration). To improve clarity, consider adding an explanation in the figure caption or in Section 4.4 to specify the meaning of each point.\", \"The conclusion section is missing, which may limit the clarity of the paper\\u2019s overall findings and contributions.\"], \"questions\": \"Please see weaknesses.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Summary of our contributions\", \"comment\": [\"We would like to copy and paste parts of our supplementary response to Reviewer ZKLE, to give a better summarization and explanation on our contributions (especially from the perspective of GFlowNets):\", \"**From the perspective of reward finetuning.**\", \"Our proposed method is a *practical, scalable and provable* method that\", \"works and scales well on **large diffusion models**;\", \"has **theoretical guarantees** of unbiased optimal solution to the target distribution; and,\", \"achieves **Pareto improvements** on reward convergence and diversity & prior preservation on **all reward functions** we experimented with.\", \"**From the perspective of GFlowNets.**\", \"We propose the **first** GFlowNet objective that leverages **gradient signals** in training GFlowNets, while all other GFlowNet methods only use zeroth-order signals.\", \"Our method demonstrates scalability to large diffusion models, such as StableDiffusion with a substantial number of sampling steps\\u2014something that few existing GFlowNet approaches have achieved.\"]}", "{\"comment\": \"As it is coming to the end of the rebuttal period, we would like to deeply appreciate the reviewer's time and efforts in positively re-evaluating our paper, and we are happy to see that we addressed most of your concerns.\"}", "{\"title\": \"Response to Authors' Rebuttal\", \"comment\": \"I thank the authors for their efforts for the rebuttal. The rebuttal addressed most of my concerns, and I raise my score to 6. Specially, I appreciate the experiments when attenuating the scale of reward value for noisier time-steps, and I recommend that the authors include the weighting scheme that they used for this experiment in the supplementary. Yet, I didn't give a higher score because I believe the novelty and technical contribution of the paper is relatively limited.\"}", "{\"comment\": \"Thanks for your reply. The authors address most of my concerns. I decide to maintain my initial rating 6.\"}", "{\"comment\": \"Dear Reviewer BEKY,\\n\\nWe wanted to follow up and check if you have any concerns or questions regarding our previous response. If so, we would be more than happy to provide further clarification or address them in detail. We would also like to share with you that we have included additional experiments in response to some other reviewers.\\n\\nWe greatly appreciate your time and effort in reviewing our response and look forward to your feedback.\\n\\nBest,\\nPaper 8475 Authors\"}", "{\"comment\": \"We thank the reviewer for their acknowledge on our paper's contribution and sharing their concerns. We have made some further refinements on the paper presentation and the experiments (detailed in the general response).\\n\\n> The problem of bias is well known in optimizations for diffusion models (elaborated in question section). Some treatment of this problem problem would be desirable since it seems that this method may be optimizing for the biased distribution\\n\\nThank you for raising this interesting point, which indeed demonstrates the advantage of our method. \\n\\nThe source of the bias in the papers cited above lies in their **stochastic optimal control** (SOC) formulation. As shown in Equation 19 in the adjoint matching paper [0], the \\u201creward\\u201d in the equivalent RL formulation is $-\\\\int_0^1 f(x_t, t)dt - g(x_1)$ on the terminal state $x_1$, which is different from $R(x_1)$ (if it were some RL method). It therefore inevitably depends on the initial value $V(x_0)$, as shown in Equation 20 and 23 in the paper. We note that this is different from the standard RL setting where one optimizes $R(x_T)$ which does not have this bias issue.\\n\\nIn contrast, GFlowNets are proposed [2] in the first place to sample from non-negative rewards (i.e., unnormalized density functions) **in an unbiased way**. By directly optimizing the detailed balance (DB) condition (or the gradient-based version), the log-flow function (or the gradient of that) learns to correct any deviation from the target distribution. Therefore, instead of saying that we \\u201cget around\\u201d the bias issue, we would prefer to say that our framework is theoretically bias-free by design.\\n\\n---\\n[2] GFlowNet Foundations. Yoshua Bengio, Salem Lahlou, Tristan Deleu, Edward J. Hu, Mo Tiwari, Emmanuel Bengio. https://arxiv.org/abs/2111.09266\"}", "{\"comment\": \"We thank the reviewer for their time and efforts and their constructive comments. We have made some further refinements on the paper presentation and the experiments (detailed in the general response).\\n\\n> In the quantitative evaluation, the proposed method performs well on the smaller dataset of Aesthetic Score, nearly doubling the baseline in the DreamSim metric. However, on the larger dataset of HPSv2, its performance is similar to baselines, showing no clear advantage. The effectiveness of the method remains inconclusive due to the differences in dataset size. \\n\\nThanks for raising this concern. In our revised draft, we show that on HPSv2 our method achieves the best performance (comparable speed to ReFL, yet still with good diversity and prior preservation), after we 1) introduce the time-dependent scaling of the predicted reward and 2) fix some numerical issues during mixed-precision training (Table 1 and Figure 12 & 13).\\n\\nIn the meantime, we would like to point out that the evaluation results in the presented table is probably misleading. A method can converge really fast (in terms of reward) but fail to output diverse outputs or fail to follow the pretrained prior (in the worst case, producing nonsensical outputs, as illustrated in Fig 3 in the revised draft). Therefore, we believe that the best way to evaluate these methods is to look at the \\u201cPareto frontiers\\u201d (quite similar to the precision-recall curve), where one can see the performance of different checkpoints of the same model and directly compare both diversity- and prior-preservation capability of model checkpoints with similar rewards.\\n\\nOn the dataset/reward model concern: the reward model of Aesthetic Score is indeed trained on a relatively large dataset (LAION-aesthetic, a subset with ~238k images from the LAION dataset [1]) whereas HPSv2 is trained on HPDv2 dataset [2] of ~433k pairs of images. The major differences between these reward models are more related to the specific objectives. Specifically, Aesthetic Score cares more about the style of images (which is less correlated with text prompts), while HPSv2 demands more on alignment between text prompts and generated images.\\n\\n> While the authors mention \\u201cfast and efficient finetuning\\u201d as a contribution, only Figure 4 shows comparable convergence speed to the baseline. It would be helpful to include details on training resource consumption, such as GPU usage and computational cost, to substantiate this claim. \\n\\nGood point! We now include the convergence plots (Figure 9) with the x-axis being the relative wall time to quantitatively evaluate how costly (in time) each method can be, with the caption stating the amount of compute used.\\n\\n> Figure 3 only compares the results of the pretrained model and the proposed method, lacking visual comparisons with other baselines. \\n\\nThank you for pointing this issue out. We have included more qualitative comparisons based on your suggestions, in both the experiment section (Figure 4 and 5 in the revised draft) and the last two pages of the appendix (Figure 24, 25 & 26).\\n\\n> Figure 5 lacks sufficient information in the title and annotations to clarify what each point represents (e.g., training iteration). To improve clarity, consider adding an explanation in the figure caption or in Section 4.4 to specify the meaning of each point. \\n\\nThank you for reminding us of our overlook on explaining the Pareto frontier figure. We have revised the caption (Figure 7 in the revised draft) and hopefully it is clearer now.\\n\\n> The conclusion section is missing, which may limit the clarity of the paper\\u2019s overall findings and contributions.\\n\\nThanks for your suggestion. We now include a conclusion section that summarizes the contributions.\\n\\n---\\n[1] https://laion.ai/blog/laion-aesthetics/\\n\\n[2] Human Preference Score v2: A Solid Benchmark for Evaluating Human Preferences of Text-to-Image Synthesis. Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, Hongsheng Li. https://arxiv.org/abs/2306.09341\"}", "{\"comment\": \"We appreciate the reviewer's acknowledgement on our contribution and suggestions. We will make the diversity claim clearer in our next draft.\\n\\n**Regarding the regularization term.** The issue is rooted in the high *variance* nature of exploration in policy optimization (even though the objective is unbiased) plus the highly complex optimization landscape, especially when we are dealing with high-dimensional action spaces (in our case, the action is some denoised image). RL methods, which are very akin to our method, are shown in the literature of RL, control theory and optimization that the underlying optimization problem is hard. For instance, the famous maximization bias problem shows that even for simple unbiased reward functions in some tiny MDP of few states, a model can take a long time to overcome overestimations caused by early learning dynamics [1]. For another example: it is shown that the optimization with the so-called temporal difference (one of the core concepts in RL) can be unstable if certain conditions are violated [2].\\n\\nEmpirically speaking, even in relatively small-scale control problems (compared to an action space of images), like the classical benchmark of Mujoco tasks (e.g., controlling a 27-degree-of-freedom humanoid to walk in simulated environments), naive RL algorithms hardly work well without regularization. There are papers to investigate the importance of regularization of RL algorithms, for example [3]. Due to the similarity between our method and RL ones, we expect regularization as an important component to alleviate the burden in tuning optimization hyperparamters.\\n\\nThat's said, the field of deep RL has been developed for many years and therefore many standard and widely accepted and applicable engineering tricks besides KL regularization, just as techniques like gradient clipping, Xavier initialization (for ResNets), orthogonal initialization (for LSTM) and batch normalization in training deep neural nets for classification. Many of these RL techniques are used in large-scale systems like AlphaGo and AlphaStar. Therefore, we are relatively confident to claim that the use of RL regularization techniques in GFlowNet-related methods (including ours) is not a big issue. \\n\\n**Regarding picking reward functions.** As we aim to accommodate different reward functions (*e.g.*, a large transformer-based neural reward model learned from noisy human preference data, or rewards from accurate physical simulators), we do not make specific assumptions on what the rewards are, so as to find methods that are generalizable to different reward functions.\\n\\nWe hope our explanations may resolve some of your concerns.\\n\\n---\\n\\n[1] https://web.stanford.edu/class/cs234/CS234Win2023/slides/lecture6post.pdf (Slides from the RL course by Emma Brunskill)\\n\\n[2] Simplifying Deep Temporal Difference Learning. ICLR 2025 Conference Submission7833. https://openreview.net/forum?id=7IzeL0kflu\\n\\n[3] Regularization Matters in Policy Optimization - An Empirical Study on Continuous Control. Zhuang Liu, Xuanlin Li, Bingyi Kang, Trevor Darrell. ICLR 2021. https://openreview.net/forum?id=yr1mzrH3IC\"}", "{\"comment\": \"I thank the authors for their effort in implementing adjoint matching and comparison to elegant.\\n\\nThese results seem to show that there is a qualitative benefit for $\\\\nabla$-DB method in comparison to adjoint matching. One concern that I have is that the definition of FID score. As mentioned in the manuscript, this is computed by the FID score \\\"between images generated from the pre-trained model and from the fine-tuned model and take the average FID score over all evaluation prompts\\\". This is not the traditional use of FID score and I think that this should be renamed to reflect this deviation from the norm.\\n\\nI think there should be some mention of limitations to this method. As mentioned in another review, this only works with SDE diffusion models, not with flow matching which has become used more frequently. Although the authors mentioned that this is not a significant limitation, I think that this is quite important. I hope that the authors can make further modifications to method in this work to show that this is possible for flow matching.\"}", "{\"title\": \"Response to Official Comment\", \"comment\": \"Thank you for the reply. I think this is correct (although not because they are optimizing a different function as mentioned, but because they add a second term to the drift coefficient). I would like to see some comparison to these SOC methods.\\n\\nIn the absence of this, I will maintain my score since I believe that this is a strong paper, but still with limited impact.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}", "{\"comment\": \"We sincerely thank the reviewer for their response and acknowledgement that our paper is a strong one.\", \"on_your_comment_on_the_root_of_bias\": \"yes, you are absolutely correct. Equation 20 and 23 in their paper are merely trying to provide equivalent RL objectives to make direct comparisons between their method and RL.\\n\\n---\\nWe greatly appreciate that the reviewer pointed out SOC methods as baselines, which are indeed important and inspiring in the field of reward finetuning for diffusion models. And we would like to further share our opinions and results regarding your comments on comparison with SOC methods:\\n\\n**ELEGANT.** \\n\\nDirectly comparison against ELEGANT is tricky, as ELEGANT essentially employs a 3-stage training procedure in which only the first one queries reward functions. Here we directly compare the final generated images with ELEGANT (directly copied and pasted from their paper) and our $\\\\nabla$-GFlowNet in this anonymous link: [https://drive.google.com/file/d/1FLI1rypZCIg2qerVaa8dQNqq-3vNeiB7/view?usp=sharing](https://drive.google.com/file/d/1FLI1rypZCIg2qerVaa8dQNqq-3vNeiB7/view?usp=sharing). We may obverse that the generated images of ELEGANT suffer from collapse in generation, illustrated by their consistent production of distorted shapes of animals and a very consistent style of non-realistic images. In contrast, the generated images with our method show much better aesthetics (colorful images), diversity (as the subjects, animal poses and image styles vary) and prior preservation (as the images look much more realistic).\\n\\n**Adjoint matching.** \\n\\n- It is a really new paper, of which the first arXiv version is released on Sept 13, only few days before the ICLR submission deadline. Plus, till now the authors have not released their codes nor their finetuned weights.\\n\\n- Adjoint matching is proposed for flow matching, although some very special cases of diffusion models with appropriate noise schedule satisfy the so-called \\\"memoryless\\\" property, to which adjoint matching happens to be applicable if we may treat the trained diffusion model as a continuous process (i.e., a SDE). Our method, in comparison, can be applied to any MDP constructed by a diffusion model (for instance, we show that our method works with the MDP constructed with SDE-DPM-solver++ in Figure 18 and 19). What's more, our $\\\\nabla$-DB objective can work for many other MDPs as it is derived under a non-diffusion-specific framework.\\n\\nWe totally agree that it is definitely worth seeing the comparison between our method and adjoint matching despite their different applicability, and therefore we tried our best to implement adjoint matching by ourselves (as it is not open-sourced yet), do the experiments that are possible in this short period of time, and show the results in these anonymous links:\\n\\n- Comparison on reward convergence: [https://drive.google.com/file/d/1JpvOoKhAxni5-Z5NadhmzdDDQWEClZh8/view?usp=sharing](https://drive.google.com/file/d/1JpvOoKhAxni5-Z5NadhmzdDDQWEClZh8/view?usp=sharing)\\n- Comparison on diversity evolution: [https://drive.google.com/file/d/1PYpn1h_O5_7APGRkucJFZmHtjpOJnyiV/view?usp=sharing](https://drive.google.com/file/d/1PYpn1h_O5_7APGRkucJFZmHtjpOJnyiV/view?usp=sharing)\\n- Comparison on FID evolution: [https://drive.google.com/file/d/1Zb0kW_ySMm1b5aqkqFu0jADWXUMO5pne/view?usp=sharing](https://drive.google.com/file/d/1Zb0kW_ySMm1b5aqkqFu0jADWXUMO5pne/view?usp=sharing)\\n- Comparion on generated images: [https://drive.google.com/file/d/1H4ulhKo4MKrV04bYKWTN4lfmebIofsZP/view?usp=sharing](https://drive.google.com/file/d/1H4ulhKo4MKrV04bYKWTN4lfmebIofsZP/view?usp=sharing)\\n\\nBoth the quantitative and the qualitative results show that adjoint matching, compared to our method, generates images with less diversity and less prior preservation.\\n\\n---\\nWe thank the reviewer in advance for the reviewer's time in reading this response, and sincerely hope that the reviewer may take our new experiments and explanations into consideration.\"}", "{\"metareview\": \"This paper proposes a novel GFlowNet method dubbed Nabla-GFlowNet (abbreviated as $\\\\nabla$-GFlowNet), together with an objective called $\\\\nabla$-DB, plus its variant residual $\\\\nabla$-DB for finetuning pretrained diffusion models, to achieve fast yet diverse sampling.\\n\\nAll 5 reviewers give a final rating of 6. I am on board with them and recommend accept.\", \"additional_comments_on_reviewer_discussion\": \"Initial concerns include: the reliability of the \\\"predicted reward\\\" estimation in Eq. 15; the claim of diversity preserving is untenable; the current experimental setting appears somewhat outdated; this paper may be optimizing for a bias distribution while preserving diversity, etc.\\n\\nMost of these concerns have been well addressed after the rebuttal.\"}", "{\"title\": \"Following-up on reviewers' concerns\", \"comment\": \"Dear Reviewers and AC,\\n\\nWe thank again for your time and efforts in reviewing our paper and going through the possibly long responses to your questions. As it is coming to the end of the rebuttal period, we would like to learn if we have addressed your concerns or the reviewers have any other questions. If any, we are more than happy to answer them.\\n\\nMany Thanks,\\n\\nPaper 8475 Authors\"}", "{\"summary\": \"This paper introduces an interesting method to address the challenges of fine-tuning multistep sampling in diffusion models. It employs GFlowNets to incorporate a middle term, $F(x_t)$ or $g_\\\\phi(x_t)$, which allows the reward score to effectively influence different timesteps. This method successfully eliminates the need to train a reward model that accepts noisy input. This paper implements their idea in both theoretical and practical contexts. Section 3.1 covers the theoretical aspect, while sections 3.2 and 3.3 address the practical application. Experiments also show that this method can enhance reward tuning in diffusion models.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"4\", \"strengths\": \"1. This paper presents a new method for addressing the challenges of fine-tuning multistep sampling in diffusion models using GFlowNets. This method effectively eliminates the need to train a reward model that processes noisy input.\\n2. This paper implements their idea in both theoretical and practical contexts. Section 3.1 covers the theoretical aspect, while sections 3.2 and 3.3 address the practical application.\", \"weaknesses\": \"The main weakness is in the experiment part.\\n1. The function $g_\\\\phi(x_t)$ is an interesting and reasonable choice for achieving the fitness task; however, it results in approximately zero vectors, with a terminal constraint of $g_\\\\phi(x_T) = 0$. It remains unclear whether Unet is a suitable option for this purpose.\\n2. The regularization term appears significant, with $\\\\lambda=1000$ in the Aesthetic Score experiments and $\\\\lambda=100$ in the HPSv2 experiments. However, Section 3.2 states that it \\\"may eventually over-optimize the reward and thus neglect the pretrained prior.\\\" Is Section 3.2 more helpful for preventing over-optimization than the regularization term?\\n3. The method is interesting and useful, but the paper's claim of \\\"Diversity-Preserving\\\" in the title creates a gap. Does this imply that the method theoretically ensures better diversity, or is it based solely on observational results? From my perspective, this method prioritizes improving information backpropagation for fine-tuning multistep sampling rather than enhancing diversity.\\n\\nIf the authors can resolve my issue, I will contemplate raising my score.\", \"questions\": \"See the weaknesses.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"details_of_ethics_concerns\": \"None\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}" ] }

[ "Ruisi Cai", "Saurav Muralidharan", "Hongxu Yin", "Zhangyang Wang", "Jan Kautz", "Pavlo Molchanov" ]

[ "large language models", "elastic networks", "training efficiency", "inference efficiency" ]

{ "note_id": [ "zOkTNMe3dG", "s6KR9DDrHg", "rSGWOmcJgb", "pn80zReG7l", "oSMfEwEjW3", "lck3Pmq9UF", "lNlAUFj0xC", "kJCePP8UlC", "g6uxXwT0KK", "g2t5C3HQSn", "ch5Aphu6bd", "ca8PvMdJ7v", "bPBXqzKQUv", "VmZ1bLV0cZ", "Vd6SN9pfbi", "ToiormCfwE", "QekIeYDj09", "K3hJ0A6maK", "GkzYXtl9zE", "CssC4803TQ", "6FRikRg2fu", "3jfs35AeoY" ], "note_type": [ "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "decision", "official_comment", "official_review", "official_comment", "official_comment", "meta_review", "official_comment", "official_comment", "official_review", "official_comment", "official_comment" ], "note_created": [ 1732628872412, 1732867184117, 1733148166253, 1730705581318, 1733167634280, 1732868288133, 1730705378172, 1732266145094, 1732267312509, 1732868327261, 1732268248349, 1737523439571, 1732268482727, 1730694338671, 1732605056736, 1732957568682, 1733659119306, 1732267804854, 1732867029431, 1730624865881, 1732958848264, 1732771652223 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission1194/Reviewer_uPJk" ], [ "ICLR.cc/2025/Conference/Submission1194/Authors" ], [ "ICLR.cc/2025/Conference/Submission1194/Reviewer_gENy" ], [ "ICLR.cc/2025/Conference/Submission1194/Reviewer_7Wyq" ], [ "ICLR.cc/2025/Conference/Submission1194/Authors" ], [ "ICLR.cc/2025/Conference/Submission1194/Authors" ], [ "ICLR.cc/2025/Conference/Submission1194/Reviewer_uPJk" ], [ "ICLR.cc/2025/Conference/Submission1194/Authors" ], [ "ICLR.cc/2025/Conference/Submission1194/Authors" ], [ "ICLR.cc/2025/Conference/Submission1194/Authors" ], [ "ICLR.cc/2025/Conference/Submission1194/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission1194/Authors" ], [ "ICLR.cc/2025/Conference/Submission1194/Reviewer_XXXs" ], [ "ICLR.cc/2025/Conference/Submission1194/Reviewer_7Wyq" ], [ "ICLR.cc/2025/Conference/Submission1194/Area_Chair_gXjG" ], [ "ICLR.cc/2025/Conference/Submission1194/Area_Chair_gXjG" ], [ "ICLR.cc/2025/Conference/Submission1194/Authors" ], [ "ICLR.cc/2025/Conference/Submission1194/Authors" ], [ "ICLR.cc/2025/Conference/Submission1194/Reviewer_gENy" ], [ "ICLR.cc/2025/Conference/Submission1194/Area_Chair_gXjG" ], [ "ICLR.cc/2025/Conference/Submission1194/Area_Chair_gXjG" ] ], "structured_content_str": [ "{\"title\": \"Comment\", \"comment\": \"Thank you for your feedback, and I recommend that you organize your newly provided experimental results in the main text or in an appendix. Methods such as Flextron and Sheared-llama that were used as comparisons should still be re-used with Llama-3.1 as the base model to ensure fair comparisons (you need not provide these comparisons at the rebuttal stage due to time limitations). Your reply solved my problem to a considerable extent and I would like to keep my score the same, thanks.\"}", "{\"comment\": \"Thank you for your thoughtful comments and suggestions. Due to the limited rebuttal period, we were unable to complete the full comparison. However, we greatly appreciate your support and will incorporate all your feedback into the revised version.\"}", "{\"comment\": \"Dear authors, thank you for your thorough responses, which addressed my concerns, therefore I raise the score to 6.\"}", "{\"summary\": \"This paper introduces a novel LLM compression framework named LLaMaFlex. With around ~60B tokens for elastic training, the network can be resized both in the depth and the width dimension, producing better results than previous pruning or distillation-based methods. The method composes a Gumbel-softmax router to select the sub-network after the elastic training. For the training, the model is first arranged by the importance of sub-networks and a policy-aware modulation is proposed. Experiments show that it can achieve much better results than previous compression methods.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"1. The idea of \\u201ctrain once, generate many\\u201d for LLM is interesting for compressing large language models. While this paradigm isn\\u2019t entirely new, this paper achieves impressive results to apply this on LLM.\\n2. The proposed method is both well-founded and novel. A particularly interesting discovery is that the learned router can interpolate, enhancing the method\\u2019s generalizability.\\n3. The experimental results demonstrate strong performance.\", \"weaknesses\": \"I did not find significant weaknesses for this paper\", \"questions\": \"1. How does the performance compare if elastic training is not applied to retrain the LLM? Specifically, if only the Gumbel router is learned on the pre-trained LLM, while retaining the pre-trained weights of LLaMA-3.1-8B.\\n2. Have you conducted experiments on other LLMs to demonstrate that this method can be extended across different types of language models?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Many thanks for your thoughtful comments and suggestions. We value your support and will be incorporating all your feedback in our revised version.\"}", "{\"comment\": \"Thank you very much for your thorough review and for the thoughtful comments you provided on our work.\\n\\nWe have made the necessary revisions in response to your comments, and we hope that our revisions meet your expectations. If there are any remaining concerns or if you have any further suggestions that could help us improve the quality of our work, please do not hesitate to let us know. We would greatly appreciate any further guidance you can provide.\\n\\nOnce again, we are truly grateful for your support. \\n\\nAuthors\"}", "{\"summary\": \"LLAMAFLEX introduces a new approach to efficiently and effectively generate high-accuracy SLMs. LLAMAFLEX employs a nested weight-shared architecture. Starting from a pre-trained model, it uses only 60 billion tokens for a single training phase, enabling zero-shot pruning across both width and depth and a \\u201ctrain once, deploy many\\u201d pruning paradigm.\", \"soundness\": \"3\", \"presentation\": \"1\", \"contribution\": \"2\", \"strengths\": [\"Introduces a gumbel softmax-based end-to-end learnable router to train the candidate models.\", \"Enhances generality by incorporating a policy-aware modulation technique.\", \"Enables interpolation-based extrapolation of pruning rates for flexible model compression.\", \"Achieves accuracy comparable to state-of-the-art SLMs like minitron.\"], \"weaknesses\": [\"The training process and methodology flow are difficult to understand and a figure like Figure 3 in Flextron paper should be provided.\", \"The presentation in the main text is somewhat weaker than in the abstract, e.g. the explanation of Equation 3 (L170-L180), and the textual ordering needs to be organized. For another example, paragraphs L199-L210 are intended to be placed in the heading \\u201cRouter Design\\u201d rather than under \\u201cProblem Formulation\\u201d. Apart from that, parts of 2.3 seem to fit into the experimental chapter. More, the description of the amount of data on line 325 should be placed on line 309 and the source of the data set should be described.\", \"In Eq. 2, if \\\\lambda_i^j=0, it would be misleading to think that all previous layers have been pruned, and the formula should be reshaped to alleviate the ambiguity. The upper limit of j is not defined, and if j can be taken to infinity, this should be stated.\", \"Unfair comparison with Flextron, you should reproduce Flextron-Llama3.1-8B to provide a fair comparison. Unfair comparison with structured pruning methods: 1) All of these methods need to be reproduced with llama-3.1-8B as the base model. 2) Structured pruning methods almost always support fine-tuning after pruning, and you should implement training using the same amount of data as your work to achieve a fair comparison.\", \"While LLAMAFLEX can incubate SLMs similar to Minitron accuracy, the layer-pruning-only version of Minitron-LLaMA-3.1 will embody far more efficiency than an SLM of the same size, and you need to make a full comparison.\"], \"questions\": [\"L094-L096: The value of a heterogeneous architecture should not be dismissed because of unfriendly backend support. I believe that loosening the homogeneous restrictions in LLAMAFLEX will boost the potential to produce stronger models at the same scale. Can you provide some relevant results (or insights if it is resource-intensive to build this experiment)?\", \"Are open-source SLMs in the Phi series and the LLaMA 3.2 series comparable in effectiveness and worth adding to the evaluation?\", \"How much computational and time resources were used for this work?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"**[Q1: Learn routers while retaining pre-trained model weights]**\\nThank you for your insightful suggestion. Our method remains effective when retaining model weights, and the routers are able to search the best configurations of the sub-networks. We conduct an experiment where we fix the backbone weights and enable only the learnable routers to search for the optimal sub-network configurations under two parameter budgets: 25% and 50%. The results, compared against heuristic methods, are presented in Table A1 below.\", \"table_a1\": \"Comparison of learnable searching framework (ours) with heuristic based methods, on elastic training while retaining the original weights.\\n| Method | Valid Loss (50% #param) | Valid Loss (25% #param) |\\n|:-:|:-:|:-:|\\n| MLP pruning-only | 4.35| 7.14 |\\n| Hidden Dim. pruning-only | 5.41 | 8.99 |\\n| Layer skipping-only | 9.13 | 10.48 |\\n| Learnable (ours) | 4.12| 6.28 |\\n\\n\\n**[Q2: Extend to other Pre-trained Model Type]**\\nThank you for the suggestion. We validated our method on Minitron-4B [1], a representative efficient large language model of relatively small scale. Using the same hyperparameter settings as in our original experiments on Llama-3.1-8B, we trained the model for 10k iterations and evaluated the perplexity (PPL) on four resultant sub-networks. Table A2 below provides the results for this experiment.\", \"table_a2\": \"Evaluation of our method using Minitron-4B as the pre-training model. We report the anchor model (the sub-models directly received gradient during training), as well as the interpolated model (the sub-models generated by router interpolation).\\n| Model size | 25% (anchor) | 37.5% | 50% (anchor) | 62.5% | 75% (anchor) | 87.5% | 100% (anchor) |\\n|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|\\n| PPL | 25.25 | 22.89 | 21.22 | 20.03 | 17.42 | 14.89 | 11.78 |\\n\\n[1] Compact language models via pruning and knowledge distillation\"}", "{\"comment\": \"**[W1: Extend to other Pre-trained Model Type]**\\nThank you for the suggestion. We validated our method on Minitron-4B [1], a representative efficient large language model of relatively small scale. Using the same hyperparameter settings as in our original experiments on Llama-3.1-8B, we trained the model for 10k iterations and evaluated the perplexity (PPL) on four resultant sub-networks.\", \"table_d1\": \"Evaluation of our method using Minitron-4B as the pre-training model. We report the anchor model (the sub-models directly received gradient during training), as well as the interpolated model (the sub-models generated by router interpolation).\\n| Model size | 25% (anchor) | 37.5% | 50% (anchor) | 62.5% | 75% (anchor) | 87.5% | 100% (anchor) |\\n|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|\\n| PPL | 25.25 | 22.89 | 21.22 | 20.03 | 17.42 | 14.89 | 11.78 |\\n\\n**[W2: Clarify our approach, and Metric Ablation]**\\nThank you for the suggestion. Using the proposed importance metric (cumulative activated values), we rank the channels and permute the weight matrix to reorder the channels/heads in decreasing order of importance. To ablate the metric, we compare this ranking with permutation to a setup without permutation. After training the model for 2k iterations, we observe a significant improvement in validation loss when incorporating the importance-based metric.\", \"table_d2\": \"Ablation experiments on our proposed importance metric.\\n| Model size | 25% | 50% | 75% | 100% |\\n|:-:|:-:|:-:|:-:|:-:|\\n| wo Ranking & Permutation | 2.89 | 2.53 | 2.25 | 2.01 |\\n| w/ Ranking & Permutation | 2.58 | 2.37 | 2.18 | 2.01 |\\n\\nIt is worth noting that our importance metric builds upon previous work. Specifically, both the Minitron[1] and Flextron[2] papers include detailed ablations comparing various metrics for importance estimation (see Sections 4.2 and 5.5 in the Minitron and Flextron papers, respectively). We use the best metrics proposed in these works.\\n\\n**[W3: sampled network]**\\nThank you for the question. As detailed in Equation 4, our overall training objective is to minimize the loss of the sampled sub-networks while ensuring they satisfy the parameter budget constraints. This is achieved by incorporating an additional loss term that is activated only when the actual number of activated parameters exceeds the budget. This approach is standard and has been adopted in existing works [2]. Further details are provided in Section 2.3. In terms of joint distribution of Q, we independently model the categorical distribution of each architectural choice by a MLP and sample one category with gumbel softmax. Specifically, $Q(D^j, N_A^j, H^j, N^j, \\\\boldsymbol{\\\\lambda}^j) = P(D^j) P(N_A^j) P(H^j) P(N^j) \\\\prod_i P(\\\\lambda^j_i)$.\\n\\n**[W4: Limitation discussion]**\\nThanks for your suggestion. We have included a limitations section in the revised version (marked in blue) as below:\\n\\nOur method is not training free, as it requires router tuning and necessitates modification to backbone weights to convert the pre-trained model into an elastic one. There is a potential application of training LlamaFlex from scratch; we leave this to future work. \\n\\n**[Q1: Router interpolation]** Thank you for the interesting question. The router outputs cannot be interpolated because they are designed to model the categorical distribution of each architectural choice. The model samples one architecture choice from this distribution, and as training progresses, the sampling process becomes increasingly deterministic, consistently selecting the optimal choice. This results in the router outputs converging to a nearly one-hot distribution. Interpolating the router outputs would disrupt this one-hot distribution, undermining the intended selection mechanism.\\n\\n**[Q2: Configuration details]**\\nThanks for the question. We provide the choices of \\\\mathcal{D}, \\\\mathcal{N_A}, \\\\mathcal{N}, \\\\mathcal{H} in Table 1 in Section 3.1. Specifically:\\n- \\\\mathcal{N}_A = \\\\{25\\\\%, 50\\\\%, 75\\\\%, 100\\\\%\\\\}\\n- \\\\mathcal{D} = \\\\{25\\\\%, 37.5\\\\%, 50\\\\%, 62.5\\\\%, 75\\\\%, 87.5\\\\% , 100\\\\%\\\\}\\n- \\\\mathcal{H} = \\\\{50\\\\%, 62.5\\\\%, 75\\\\%, 87.5\\\\% , 100\\\\%\\\\} \\n\\nWe observe that the final results are largely independent of specific choices within these sets; in other words, the selections can be made flexibly, as long as there exist combinations of these sets that satisfy the parameter budget constraints and the elements within each set are reasonably spaced.\\n\\n[1] Compact language models via pruning and knowledge distillation\\n\\n[2] Flextron: Many-in-One Flexible Large Language Model\"}", "{\"comment\": \"Thank you very much for your thorough review and for the thoughtful comments you provided on our work.\\n\\nWe have made the necessary revisions in response to your comments, and we hope that our revisions meet your expectations. If there are any remaining concerns or if you have any further suggestions that could help us improve the quality of our work, please do not hesitate to let us know. We would greatly appreciate any further guidance you can provide.\\n\\nOnce again, we are truly grateful for your support. \\n\\nAuthors\"}", "{\"comment\": \"**[W3: Comparison with Distillation]**\\nThank you for your insightful question. First, we would like to emphasize that elastic frameworks, including LlamaFlex, are orthogonal to traditional fixed-size compression methods. While our primary focus is on the elastic component, our framework is fully compatible with various fixed-size compression approaches and can potentially benefit from advanced methods in this domain. To validate this compatibility, we incorporated knowledge distillation into our framework by adding an additional knowledge distillation loss. We trained the model for 2k iterations and reported the validation loss in Table C2 below. This experiment demonstrates the flexibility of our approach in integrating with other compression techniques.\", \"table_c2\": \"Compatibility of the proposed framework with existing fix-sized compression methods.\\n| Model size| 25% | 50% | 75% | 100%|\\n|:-:|:-:|:-:|:-:|:-:|\\n| wo distillation | 2.58 | 2.35 | 2.16 | 1.97 |\\n| w/ distillation | 2.49 | 2.26 | 2.15 | 1.99 |\\n\\nSecond, we compared our method with representative fixed size compression methods - Minitron [1]. To ensure fair comparison, we use the same pre-trained model and similar data sources as Minitron, and compare the downstream performance in Table C3.\", \"table_c3\": \"Downstream comparison with Minitron method (summarized from Table 2).\\n| Model | #Params | ARC-E | LAMB. | PIQA | Wino. | Hell. | MMLU | Avg.|\\n|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|\\n| LlamaFlex-50% | 3.5B | 76.1% | 66.3% | 77.0% | 67.9% | 76.0% | 60.3% | 70.6% |\\n| Minitron-Depth | 3.7 B | 74.7% | 61.5% | 74.9% | 72.5% | 73.2% | 58.7% | 69.3% |\\n| Minitron-Width | 3.7 B| 75.1%| 63.0% | 75.0%| 73.1%| 76.1%| 60.5%| 70.5%|\\n\\n**[W4: Discussion on optimal architectures]**\\nThank you for the suggestion! In our framework, a lightweight sub-network can be derived by reducing the hidden dimension ($H$), MLP intermediate dimension ($D$), number of attention heads ($N_A$), and the number of remaining layers ($N$). This framework offers a wide range of possible combinations to achieve approximately the same parameter budget (e.g. 50% remaining parameters). However, different configurations would yield different performance. Thus we apply the learnable framework that aims to search for the sub-network with optimal configurations.\", \"table_c4\": \"The learnable framework yields optimal sub-network configurations. With a 50% remaining parameter budget, the sub-network can adopt diverse configurations. For example, we can prune the MLP by retaining only half of its intermediate dimension, prune the hidden dimension by keeping only half of the hidden feature size, or skip half of the layers to meet the 50% target. However, our learnable framework identifies the optimal configurations that combine multiple strategies. All models are fine-tuned for 2k iterations in these experiments.\\n\\n|Method | $H$ | $D$ | $N_A$ | $N$ | Budget (#param) | Valid Loss |\\n|:-:|:-:|:-:|:-:|:-:|:-:|:-:|\\n| MLP pruning-only | 4096 | 7168 (50%) | 32 | 32 | 4.16 B (59%) | 2.16 |\\n| Hidden-dim pruning-only | 2048 (50%) | 14336 | 32 | 32 | 3.49 B (50%) | 2.22 |\\n| Layer pruning-only | 4096 | 14336 | 32 | 16 (50%) | 3.49 B (50%) | 2.31 |\\n| Learnable (ours) | 2560 (62.5%) | 12544 (87.5%) | 24 (75%) | 30 (93.75%) | 3.48 B (50%) | 2.12 |\\n\\n[1] Compact language models via pruning and knowledge distillation\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}", "{\"comment\": \"**[W1: Pipeline Figure]**\\nThank you for the suggestion. We have updated Figure 3 to explicitly illustrate how the routers determine the sub-network configurations. Additionally, we have included a new Figure 4 to provide a clearer depiction of the modulation process.\\n\\n**[W2: Text and subtitles]**\\nThank you for your valuable suggestions. We have updated the revised PDF and marked the changes in **blue** based on your suggestions. Please let us know if there are any further questions or points of confusion. Regarding data resources, we utilize in-house data but are unfortunately unable to disclose specific details at this stage to avoid revealing author identities, in compliance with the double-blind review policy. More detailed information will be provided upon paper acceptance.\\n\\n**[W3: Confusion on Equation 2]**\\nThanks for pointing this out and our apologies for the confusion. We have updated Equations 1 and Equation 2 in the revised PDF, and have added the skip connection term in the second equation. In this case, the layer will be skipped and only the skip connection is enabled if $\\\\lambda_i$ equals 0. \\n\\n**[W4: Unfair comparison, and the comparison with layer-pruning-only Minitron]**\\nThank you for pointing this out. Our method primarily compares with Minitron, a representative compression approach that integrates pruning and distillation. To ensure a fair comparison, we used Llama-3.1-8B as the pre-trained base model and the similar data distribution (confirmed with the Minitron authors), aligning with the setup of Minitron-4B to the best of our knowledge.\\n\\nAs shown in Table B1 below, our method outperforms the Minitron models (both the layer-pruning and width-pruning versions). Notably, our approach requires only 60 billion tokens, compared to Minitron's 200 billion tokens (100 billion tokens per model version, with computations repeated). Despite using significantly fewer tokens, we notice that our approach achieves better performance, underscoring its effectiveness.\", \"table_b1\": \"Downstream comparison with Minitron method (summarized from Table 2).\\n| Model | #Params | ARC-E | LAMB. | PIQA | Wino. | Hell. | MMLU | Avg.|\\n|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|:-:|\\n| LlamaFlex-50% | 3.5B | 76.1% | 66.3% | 77.0% | 67.9% | 76.0% | 60.3% | 70.6% |\\n| Minitron-Depth | 3.7 B | 74.7% | 61.5% | 74.9% | 72.5% | 73.2% | 58.7% | 69.3% |\\n| Minitron-Width | 3.7 B| 75.1%| 63.0% | 75.0%| 73.1%| 76.1%| 60.5%| 70.5%|\\n\\nWe acknowledge that the claim \\\"LlamaFlex can incubate SLMs similar to Minitron accuracy\\\" may not fully capture the nuances of the comparison. As shown in Table B1, LlamaFlex-50% achieves higher average accuracy (70.6%) compared to Minitron-Depth (69.3%) and is on par with Minitron-Width (70.5%). Moreover, LlamaFlex requires significantly fewer tokens for fine-tuning.\\n\\n**[Q1: Heterogeneous Layer]**\\nThank you for your insightful question. While our primary focus is on a system-friendly elastic framework, we observed the trade-off between performance and system efficiency. Loosening the homogeneous restrictions in LlamaFlex could enhance its potential to produce stronger models at the same scale. Specifically, we explored using layer-wise routers instead of a single router for all layers, and this approach demonstrated improved performance while maintaining the same model size.\\n\\n**[Q2: Comparison with Phi]**\\nThank you for the comments. Our model cannot be directly compared with the Phi series or Llama 3.2 due to differences in data resources. To ensure a fair comparison, we focus on benchmarking against the Minitron method, where our model demonstrates exceptional performance while utilizing fewer training resources.\\n\\n**[Q3: Training details]** Thank you for the question. We used a total of 60.4 billion training tokens and approximately 12,288 GPU hours. Specifically, we set the sequence length to 4096, the batch size to 128, and fine-tuned the model for 28,800 iterations.\"}", "{\"summary\": \"This paper presents LLaMaFlex, a novel approach to create elastic Large Language Models (LLMs) that can be dynamically resized without additional fine-tuning. The key innovation is a nested weight-shared architecture combined with a Gumbel Softmax-based router that allows zero-shot pruning across both width (attention heads, hidden dimensions, MLP intermediate size) and depth (layers) dimensions. Unlike previous approaches that require separate training or distillation for each model size, LLaMaFlex needs only a single continued training phase of ~60B tokens to enable \\\"train once, deploy many\\\" capabilities. The authors also introduce a policy-aware modulation technique inspired by diffusion models to enhance the expressiveness of nested architectures. When applied to Llama 3.1 8B, LLaMaFlex produces compressed models that outperform state-of-the-art pruned models, elastic frameworks, and models trained from scratch. The router can interpolate smoothly between different model sizes, allowing deployment flexibility without accuracy compromises.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": [\"The paper introduces several innovative technical components, including an end-to-end learnable router with Gumbel Softmax for architecture selection, policy-aware modulation for enhanced expressiveness, and support for both width and depth pruning dimensions. This comprehensive approach advances the state-of-the-art in elastic LLM architectures.\", \"The method demonstrates impressive performance across multiple downstream tasks (ARC-E, LAMBADA, PIQA, WinoGrande, HellaSwag, MMLU), consistently outperforming existing approaches including Flextron, Minitron, and models trained from scratch. The results are particularly notable given that LLaMaFlex requires only 60B tokens for training compared to hundreds of billions for traditional approaches.\", \"LLaMaFlex produces uniform architectures that are compatible with common deployment frameworks like TensorRT-LLM and llama.cpp, addressing a significant limitation of previous elastic approaches that generated heterogeneous architectures. The ability to interpolate between model sizes without additional training also offers practical deployment flexibility.\"], \"weaknesses\": \"- While the paper presents strong technical contributions, there are opportunities for additional context and analysis that could further strengthen the work. The authors could enrich their discussion by connecting their approach to the broader context of supernet-based neural architecture search. For instance, works like HAT and HELP have previously explored supernet-based approaches for generating target-budget subnets in the field of small language models and Meta-NAS or DARTS-EGS have handled Gumbel-Softmax-based supernet. Additionally, a more thorough discussion and comparison of the paper's relationship to MatFormer's subnet extraction approach for Transformer architectures (applied to Llama or ViT) would be valuable.\\n- The paper would benefit from additional analysis of training dynamics typically associated with supernet approaches. In supernet-based methods, there are well-known challenges to address. First, training imbalances between larger and smaller subnets often occur, where larger architectures receive less training attention and may not fully converge. Second, weight sharing interference can arise when different architectural configurations compete for optimal parameter values. While the authors present strong results, providing insights into how LLaMaFlex addresses these potential training imbalances would be informative. Similarly, a discussion of how weight sharing affects different architectural configurations could offer valuable implementation insights for practitioners.\\n- The comparative analysis could be more comprehensive, particularly regarding scenarios with fixed target sizes. A direct comparison with knowledge distillation approaches would be valuable, considering factors such as training computational costs, final model performance, and the associated trade-offs. Such analysis would help practitioners better understand when to choose LLaMaFlex over traditional approaches.\\n- The process for selecting optimal architectures when multiple configurations achieve similar parameter counts needs more detailed explanation.\\n\\n[HAT] HAT: Hardware-Aware Transformers for Efficient Natural Language Processing\\n[HELP] HELP: Hardware-Adaptive Efficient Latency Predictor for NAS via Meta-Learning\\n[MatFormer] MatFormer: Nested Transformer for Elastic Inference\\n[MetaNAS] Meta-Learning of Neural Architectures for Few-Shot Learning\\n[DARTS-EGS] Differentiable Architecture Search with Ensemble Gumbel-Softmax\", \"questions\": \"Please address concerns in Weaknesses.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thanks for your detailed response. I maintain my score.\"}", "{\"comment\": \"Dear Reviewer,\\n\\nCould you kindly respond and indicate whether authors have addressed your concerns? \\n\\nThanks, AC\"}", "{\"metareview\": \"(a) Summary: introduces LLaMaFlex, a new weight-shared training architecture with routers, for LLMs. It enables zero-shot pruning for different widths and depths.\\n\\n(b) Strengths: being able to \\\"train once, deploy many\\\"; strong performance; easy deployment with uniform architectures, etc.\\n\\n(c) Weaknesses: limited evaluation (LLaMA 3.1 8B); clarity issues in writing/organization; more fair comparisons needed.\\n\\n(d) Reasons for decision: reviewers' unanimous support; \\\"train once, deploy many\\\" is highly demanded for LLMs.\", \"additional_comments_on_reviewer_discussion\": \"Authors added experiments on Minitron-4B, added ablations on importance metrics, clarified architectural constraints, reorganized sections, and included a limitations discussion. Reviewers acknowledged these, with one reviewer upgrading the rating.\"}", "{\"comment\": [\"**[W1: Comparison with Supernet-based NAS]**\", \"Thanks for the insightful suggestion. Our work primarily focuses on LLMs, which sets it apart from previous studies that primarily target relatively small models. Consequently, our work offers unique contributions specifically designed to address the challenges associated with LLMs:\", \"LLMs demand substantial computational resources and extensive datasets for training. To mitigate these high costs, our framework focuses on **transforming a pre-trained model** into an elastic framework, enabling efficient training and adaptation. In contrast, supernet-based methods such as HAT [1], HELP [2], DARTS-EGS [3] and MetaNAS [4], as well as recent work like Matformer [5], primarily rely on training models from scratch. Additionally, our method enables zero-shot generation of sub-networks for any parameter budget through router interpolation, a capability not achieved by previous supernet-based methods.\", \"LLMs typically consist of billions of parameters, making GPU memory usage a critical bottleneck during training. To address this, we adopt weight sharing, ensuring the number of training parameters remains comparable to that of a dense pre-trained model. In contrast, methods like HELP [2], DARTS-EGS [3], and MetaNAS [4] train a supernet with multiple branches. While HAT [1] also employs weight sharing, its use of heterogeneous layers complicates system deployment.\", \"Matformer [5] relies on manual mix-and-match to extract sub-models, requiring manual selection of optimal configurations. In contrast, we present a fully learnable router that can do this sub-model extraction automatically.\"], \"we_have_added_the_following_paragraph_to_the_related_work_section_in_the_revised_pdf\": \"HAT[1] and HELP[2] have explored supernet-based approaches for generating sub-networks, primarily targeting relatively small-sized models, while Meta-NAS[4] and DARTS-EGS[3] have utilized Gumbel-Softmax-based supernet techniques. Our work focuses on addressing the unique challenges of LLMs, which differ significantly from prior studies targeting smaller models. To avoid costly training processes and huge GPU memory usage, we transform a pre-trained model into an elastic framework, and employ weight-sharing, unlike previous attempts. To avoid repeated computation, we enable zero-shot sub-network generation for any parameter budget through router interpolation - capabilities not achieved by supernet-based methods.\\n\\n\\n**[W2: Training dynamics]**\\nOur framework leverages **pre-trained LLMs** and applies weight matrix permutation based on an importance metric, ensuring that the largest model variant (i.e., the full pre-trained model) is already converged while other sub-models are reasonably initialized. Furthermore, the algorithm inherently avoids selecting the converged large model variant, as it is penalized by the parameter loss for exceeding the budget requirement.\\n\\nTo mitigate the interference caused by weight sharing, we propose the **modulation** technique (detailed in Sec 2.4). Specifically, the nesting constraint can potentially limit the representational capacity of elastic networks, as the same set of weights must accommodate inputs across all possible sub-networks. To address this limitation, we propose the use of policy-aware modulation, a technique inspired by methods in the literature on diffusion models. We introduce lightweight non-linear modulation heads, which are applied after the elastic components (i.e., elastic MLP/elastic MHA). These heads modulate the outputs of elastic operations based on the elastic choice. We ablate the modulation technique in Section 4 and provide a detailed comparison in Table C1 below.\", \"table_c1\": \"Results of our ablation study on policy-aware modulation. We run LLAMAFLEX for 800 iterations and report the validation loss. All sub-networks show improved performance when modulation is enabled, reducing validation loss by 0.08 on average.\\n|$b_j$|25%|50%|75%|100%|avg.|\\n|:-:|:-:|:-:|:-:|:-:|:-:|\\n|w/ Modulation|2.53 ($\\\\downarrow$ 0.08)|2.41 ($\\\\downarrow$ 0.13)|2.08 ($\\\\downarrow$ 0.09)|1.88|2.22 ($\\\\downarrow$ 0.08)|\\n|wo Modulation|2.61|2.54|2.17|1.88|2.30|\\n\\n[1] HAT: Hardware-Aware Transformers for Efficient Natural Language Processing \\n\\n[2] HELP: Hardware-Adaptive Efficient Latency Predictor for NAS via Meta-Learning \\n\\n[3] Differentiable Architecture Search with Ensemble Gumbel-Softmax\\n\\n[4] Meta-Learning of Neural Architectures for Few-Shot Learning \\n\\n[5] MatFormer: Nested Transformer for Elastic Inference\"}", "{\"comment\": \"Many thanks for your thoughtful comments and suggestions. We value your support and will be incorporating all your feedback in our revised version.\"}", "{\"summary\": \"The paper presents LLaMaFlex, an elastic architecture for large language models that enables zero-shot resizing across both width and depth dimensions, allowing for the instant generation of multiple compressed models without additional fine-tuning. LLaMaFlex utilizes a nested weight-shared network architecture and a Gumbel Softmax-based router for smooth interpolation between model sizes, achieving a \\\"train once, deploy many\\\" paradigm. It introduces policy-aware modulation to enhance the expressivity of nested architectures and produces a family of compressed models from Llama 3.1 8B that outperform state-of-the-art compressed, elastic/flexible, and trained-from-scratch models in accuracy. The uniform architectures generated by LLaMaFlex are also easier to deploy using existing LLM frameworks, offering a significant advancement in training efficiency and deployment flexibility for large language models.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"3\", \"strengths\": \"1. The authors introduce a novel elastic architecture for large language models that enables zero-shot resizing across both width and depth, addressing a key challenge for efficient model deployment.\\n\\n2. The end-to-end trained router could dynamically determine the optimal network architecture under a given parameter budget.\\n\\n3. It achieves a promising performance of the compressed model in various of size, which outperform the same size model trained from scratch.\", \"weaknesses\": \"1. The authors validate their proposed method in only one model, Llama 3.1 8B. The serviceability to other models and architecture should be considered.\\n\\n2. I understand that the router is used to determine the optimal architecture, that is the hyperparams of the network architecture, under a given parameter budget. Once the hyperparams of the network architecture are determined, the algorithm will select the top-k importance submodule, which is ranked by a pre-defined metric, e.g. the accumulated magnitude of activations as stated in Section 2.3. There is no ablation on the pre-defined metric, which might limit the rigor and integrity of this work.\\n\\n3. How to guarantee that the sampled submodel (i.e., the combination of N, H, D, N_A, lambda) satisfies the parameter budget (i.e., the second equation in Eq. 4), given that N, H, D, N_A, and lambda are modeled independently in the routers (i.e., Eq. 5)? What is the explicit joint distribution of Q in Eq. 6?\\n\\n4. The authors did not explicitly discuss the limitations of the proposed method.\", \"questions\": \"1. Instead of interpolating the router input h_k in Eq. 7, is it possible to interpolate the router output in a similar way given h_{n+1} and h_{n}?\\n\\n2. In Line 196, the author indicates that \\\\mathcal{D}, \\\\mathcal{N_A}, \\\\mathcal{N}, \\\\mathcal{H} are predefined, how to define them?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Dear Reviewer,\\n\\nCould you kindly respond and indicate whether authors have addressed your concerns?\\n\\nThanks, AC\"}", "{\"title\": \"Reviewers, please kindly respond\", \"comment\": \"Dear Reviewers,\\n\\nIf you have not responded to author's rebuttal, please kindly do so as soon as possible. The deadline is Dec 2, but the authors can potentially further clarify questions if you respond earlier. Thanks!\\n\\nBest, AC\"}" ] }

[ "Olivier Lepel", "Anas Barakat" ]

[ "Cumulative prospect theory", "policy gradient", "policy optimization", "reinforcement learning" ]

{ "note_id": [ "yNAF64WSKP", "xS39L21Fs9", "xAVlUPHXmF", "vjLTZUMrrM", "uJI4nHKyQR", "omknRkIV8p", "nR7oNsLSRl", "k1Wa9YXWrr", "jRgfBcrZvB", "igoZr1oEys", "hX5urtYp6s", "Zbw7jLc7dJ", "VioUkhQDRY", "U15yJSlAwR", "TFFLfxKExZ", "PgSfR1DxOM", "NzrteCaibJ", "NrjNpeICFp", "LAhBzhuMn5", "HomOOcvG8x", "HkSKhbM6as", "897YZLUdhx", "66YnAZyPUk", "5RRRcbHrvI", "3Rj4cqh61w", "2ipEj4zSoA" ], "note_type": [ "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "decision", "meta_review" ], "note_created": [ 1732215870507, 1732214701633, 1733209823431, 1732341666140, 1730666717923, 1733208506097, 1730498204601, 1733216834310, 1732215323051, 1732796205415, 1733211274851, 1732213487930, 1733179057859, 1732214990922, 1732214116829, 1732796286203, 1732215636874, 1733208710519, 1733184946075, 1732216124892, 1732882215471, 1732212470977, 1730151286470, 1732216380155, 1737523422252, 1734643213327 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission901/Authors" ], [ "ICLR.cc/2025/Conference/Submission901/Authors" ], [ "ICLR.cc/2025/Conference/Submission901/Reviewer_aLSN" ], [ "ICLR.cc/2025/Conference/Submission901/Reviewer_hWfL" ], [ "ICLR.cc/2025/Conference/Submission901/Reviewer_Ydtj" ], [ "ICLR.cc/2025/Conference/Submission901/Authors" ], [ "ICLR.cc/2025/Conference/Submission901/Reviewer_hWfL" ], [ "ICLR.cc/2025/Conference/Submission901/Authors" ], [ "ICLR.cc/2025/Conference/Submission901/Authors" ], [ "ICLR.cc/2025/Conference/Submission901/Authors" ], [ "ICLR.cc/2025/Conference/Submission901/Authors" ], [ "ICLR.cc/2025/Conference/Submission901/Authors" ], [ "ICLR.cc/2025/Conference/Submission901/Reviewer_Ydtj" ], [ "ICLR.cc/2025/Conference/Submission901/Authors" ], [ "ICLR.cc/2025/Conference/Submission901/Authors" ], [ "ICLR.cc/2025/Conference/Submission901/Authors" ], [ "ICLR.cc/2025/Conference/Submission901/Authors" ], [ "ICLR.cc/2025/Conference/Submission901/Authors" ], [ "ICLR.cc/2025/Conference/Submission901/Reviewer_aLSN" ], [ "ICLR.cc/2025/Conference/Submission901/Authors" ], [ "ICLR.cc/2025/Conference/Submission901/Authors" ], [ "ICLR.cc/2025/Conference/Submission901/Authors" ], [ "ICLR.cc/2025/Conference/Submission901/Reviewer_aLSN" ], [ "ICLR.cc/2025/Conference/Submission901/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission901/Area_Chair_UHSq" ] ], "structured_content_str": [ "{\"title\": \"Response to reviewer aLSN (continued)\", \"comment\": \"- Regarding the structure of the final reward and the metric learning you mention, this is a fair point and we agree that our work requires so far access to utility and weight functions. However, let us mention a few points:\\n1. These can be readily available in specific applications (for risk modeling or even chosen at will by the users themselves); \\n2. CPT relies on a predefined model, this can be beneficial in applications such as portfolio optimization or medical treatment where trade-offs have to be made and models might be readily available; \\n3. Furthermore, we argue that having such a model allows it to be more explainable compared to a model entirely relying on human feedback and fine tuning, let alone the discussion about the cost of collecting human feedback. We also note that some of the most widely used algorithms in RLHF (e.g. DPO) do rely on the fact that the reward follows a Bradley-Terry model for instance (either for learning the reward or at least to derive the algorithm to bypass reward learning); \\n4. Let us mention that one can also learn the utility and weight functions. We mentioned this promising possibility in our conclusion although we did not pursue this direction in this work. One can for instance represent the utility and weight functions by neural networks and train models to learn them using available data with relevant losses, jointly with the policy optimization task. One can also simply fit the predefined functions (say e.g. Tversky and Kahneman\\u2019s function) to the data by estimating the parameters of these functions (see $\\\\eta$ in l. 162 with our notations and exponents of the utility function in table 1 p. 17 for the CPT row). This last approach is already commonly used in practice, see e.g. Rieger et al. 2017 (reference 1/ provided below in the discussion). \\n\\n- **CPT vs RLHF: General comparison.** CPT has been particularly useful when modeling specific biases in decision making under risk to account for biased probability perceptions. It allows to **explicitly** model cognitive biases. In contrast, RLHF has been successful in training LLMs which are aligned with human preferences where these are complex and potentially evolving and where biases cannot be explicitly and reasonably modeled. RLHF has been rather focused on learning **implicit** human preferences through interaction (e.g. using rankings and/or pairwise comparisons). Overall, CPT can be useful for tasks where risk modeling is essential and critical whereas RLHF can be useful for general preference alignment although RLHF can also be adapted to model risk if human preferences are observable and abundantly available at a reasonable cost. This might not be the case in healthcare applications for instance, where one can be satisfied with a tunable risk model. On the other hand, so far CPT does not have this ability to adapt to evolving preferences over time unlike RLHF which can do so via feedback. \\n\\n- **CPT and RLHF: Pros and cons.** To summarize the pros and cons of both approaches, we provide the following elements. As for the pros, CPT directly models psychological human biases in decision making via a structured framework which is particularly effective for risk preferences. RLHF can generalize to different scenarios with sufficient feedback and handle complex preferences via learning from diverse human interactions, it is particularly useful in settings where preferences are not explicitly defined such as for LLMs for aligning the systems with human preferences and values. As for the cons, CPT is a static framework since the utility and probability weight functions are fixed, it is hence less adaptive to changing preferences. It uses a predefined model of human behavior which is not directly using feedback. It also requires to estimate model parameters precisely, often for specific domains. As for RLHF on the other hand, the quality and the quantity of the human feedback is essential and this dependence on the feedback clearly impacts performance. This dependence can also cause undesirable bias amplification which is present in the human feedback. We also note that training such models is computationally expensive in large scale applications.\"}", "{\"title\": \"Rebuttal: response to reviewer hWfL\", \"comment\": \"We thank the reviewer for their time and feedback. We answer their questions and reply to their comments in the following.\\n\\n> The algorithm is only validated in a few small domains, making it difficult to assess the performance of this policy gradient in more complex tasks.\\n\\nBesides our small scale grid world experiments, please note that we did also test the performance of our PG algorithm on a continuous state action space setting with our electricity management scenario. In practice, prior work in CPT-RL (L.A. et al. 2016, ICML) did only consider a SPSA (zeroth order) algorithm (which we also compared to) in a traffic signal control application on a 2x2-grid network. Our PG algorithm has clearly an advantage compared to a zeroth order algorithm which only relies on function values and which is harder to scale to larger state action spaces. **We have now performed additional simulations during the rebuttal phase for a financial trading application** (see response to reviewer Ydtj, results and precise description in the last two pages of our revised manuscript, all modifications in blue). \\n\\n> The computation of the integral for the CPT value appears complex and time-consuming due to the use of Monte Carlo sampling.\\n\\nMonte Carlo sampling is used even for the vanilla policy gradient algorithm (say Reinforce) for standard expected return RL. We use a similar procedure here which is simple. Our additional required quantile estimation procedure requires a mild sorting step which can be executed in $O(n \\\\log n)$ running time (without even invoking parallel implementations) where $n$ is the length of the rewards to be sorted. \\n\\n> Suggestion: The authors may consider demonstrating the benefits of using CPT over standard risk-neutral RL in a concrete reinforcement learning problem at the beginning of the paper to help readers better understand CPT.\\n\\nWe have provided concrete simple examples of CPT in the introduction to give the reader some intuition without introducing any notation nor background. For clarity, we prefer to defer a detailed exposition of examples of CPT-RL to later in the paper (simulation section) once the problem formulation is precisely exposed and explained. Thank you for the suggestion, we will add a pointer to the introduction to directly refer the reader to the relevant section for an example.\\n\\n> This paper considers the problem in a non-discounted setting, i.e., \\u03b3=1 as shown in Line 181. The claim in line 261 does not apply to the discounted case for exponential utilities. Discussion may be needed. see [Entropic Risk Optimization in Discounted MDPs, AISTATS 2023]. \\n\\n- We thank the reviewer for their comment and for the reference. Indeed, we focus on the undiscounted setting in our work. This is precisely because some of the issues discussed in the paper you mention arise when considering discounted rewards (e.g. lack of positive homogeneity). The entropic risk measure (ERM) mentioned in the reference you provide seems to be actually rather consistent with our result (Theorem 4), up to the monotone log function applied to the objective (which does not fundamentally change the policy optimization problem). Comparing (5) and (8) in the reference with our (EUT-PO) problem formulation, ERM is rather consistent with the exponential form we provide in Theorem 4 (up to the fact that our theorem does not apply to the discounted setting as we currently state it). As for the entropic value-at-risk (which builds on the entropic risk measure), it seems that it is not exactly an instance of our problem formulation (EUT-PO) because of the supremum over $\\\\beta$ transformation coupling ERM with the additive log term (besides the discounted setting). \\n\\n\\n- Although discounting is widely adopted in the RL community, especially for infinite horizon settings, we do not find the finite horizon to be a restrictive setting in practice in applications. As for the extension to the infinite horizon, we find the use of positive quasi-homogeneity to adapt the proof and show the existence of an optimal deterministic Markov policy (which is crucially non-stationary) for the special case of Entropic Risk Measure (Theorem 3.3 therein) interesting. This observation might actually give some hope to extend our result to the discounted setting using similar techniques. \\nWe will add a discussion regarding this interesting point, thanks again for your comment and the useful reference that we were not aware of.\"}", "{\"comment\": \"I see. In this case, I feel that it would be better by not restricting to continuous weight functions when you define CPT values. Then you can say that CPT values encompass CVar, Var, etc. Then when you introduce the method, you could say that the method only works for continuous weight functions. But discontinuous functions are probably not that important to consider. Plus, our method has advantages A, B, and C.\\n\\nI don't have more questions.\"}", "{\"comment\": \"Thanks for the detailed response! I have the same feeling as reviewer aLSN that I did not capture the importance of CPT-based RL after reading the paper. The advantage of CPT is that it is a generalization, while the drawback is that it is not intuitive to interpret. The paper provides a very brief example in the introduction, which does not adequately motivate the use of CPT. The authors may consider addressing questions such as how to choose the utility function and what the CPT represents when choosing a specific utility, to help readers better understand the CPT-RL problem.\\n\\nIn addition, the experimental domains are relatively simple. Many risk RL papers conduct experiments in Mujoco (e.g., [1, 2, 3, 4]). The authors may consider designing more complex experiments to validate the proposed algorithm.\\n\\n[1] Risk-Averse Offline Reinforcement Learning, ICLR 2021\\n[2] Mean-variance policy iteration for risk-averse reinforcement learning, AAAI 2021\\n[3] An alternative to variance: Gini deviation for risk-averse policy gradient, NeurIPS 2023\\n[4] One risk to rule them all: A risk-sensitive perspective on model-based offline reinforcement learning, NeurIPS 2023\"}", "{\"summary\": \"The paper introduces a novel approach to reinforcement learning (RL) by leveraging Cumulative Prospect Theory (CPT) to account for human decision-making biases, moving beyond the traditional expected utility framework. It focuses on the policy optimization problem, where the objective is the CPT value of the random variable recording the cumulative discounted rewards (CPT-PO). Additionally, it considers the particular case of the expected utility objective, where only returns are distorted by the utility function (EUT-PO). This work derives a policy gradient (PG) theorem for CPT-based RL, presenting a model-free policy gradient algorithm for optimizing a CPT policy.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"This paper addresses an objective that encompasses a broad class, including Conditional Value at Risk (CVaR), distortion risk measures, and expected utility.\\n\\nThe paper effectively explains various policy classes associated with different objectives, helping to clarify how these classes relate to CPT-RL versus standard RL approaches.\\n\\nThe derivation of a policy gradient theorem for CPT objectives is a significant extension of the traditional PG results in reinforcement learning, broadening its applicability in human-centered decision contexts.\", \"weaknesses\": \"The experimental section provides limited insight due to its use of relatively straightforward examples (traffic control and electricity management), which may not fully illustrate the complexities that arise in more realistic, high-stakes environments, such as finance or healthcare.\\n\\nIn Section 5(a), the observation that different weight functions lead to different policies in the grid environment could be strengthened by assessing these policies against risk measures (such as CVaR) beyond expected return. Without this, it is unclear if the CPT-RL-PO policy outperforms standard RL-PO under any specific risk-sensitive criteria.\\n\\nThe grid environment results are overly simplified and provide little substantive information. It would be beneficial to evaluate whether the derived policies' performance aligns meaningfully with their respective objectives.\", \"limited_applicability\": \"The experiments did not demonstrate the advantages of using the proposed algorithm. It is unclear what the benefits of using CPT learning are over risk-sensitive RL or distributional RL. Although CPT is theoretically valuable, the empirical advantages of CPT over other risk-sensitive measures remain ambiguous in the presented results. The authors may consider including direct comparisons with risk-sensitive RL or distributional RL methods on the same tasks.\", \"questions\": \"Minor comment: On page 4, under Problem formulation: CPT-RL, the cumulative discounted rewards variable $X$ is referenced, but the definition provided does not discount the rewards. It would be clearer to either update the variable's definition to include discounting or adjust the notation accordingly to prevent confusion.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Thank you for your feedback, our response to your follow-up questions\", \"comment\": [\"We sincerely thank the reviewer for their valuable feedback and suggestions which have helped improve our work. We reply to their comments and follow-up questions in details in the following.\", \"1. **About our healthcare example.** Thank you for the useful suggestion which we will follow, we agree that it will make the example even more compelling. There are indeed several behavioral studies in healthcare works using CPT and studying what you suggested. We will add them to support the example. We have found a number of recent such studies from which we cite a few together with some quotes to support our point:\", \"**Mkrtchian, A., Valton, V., & Roiser, J. P. (2023). Reliability of decision-making and reinforcement learning computational parameters. Computational Psychiatry, 7(1), 30.** This is a psychological study conducted on 50 participants recruited from the UCL Institute of Cognitive Neuroscience Subject Database and supported by the (British) National Institute for Health Research (NIHR). Here is a quote from the paper: \\u2018[...]risk aversion and loss aversion parameters from a prospect theory model exhibited good and excellent reliability, respectively. [...] These results suggest that reinforcement learning, and particularly prospect theory parameters, as derived from a restless four-armed bandit and a calibrated gambling task, can be measured reliably to assess learning and decision-making mechanisms. Overall, these findings indicate the translational potential of clinically-relevant computational parameters for precision psychiatry.\\u2019 The authors further add in the conclusion: \\u2018These models can further be used to predict future behaviour in the same individuals, especially PT model parameters, indicating that the decision-making processes assessed in these tasks represent relatively consistent and unique characteristics of an individual. These findings take us one step closer to translating computational measures of behaviour into clinical application.\\u2019\", \"**George, S. A., Sheynin, J., Gonzalez, R., Liberzon, I., & Abelson, J. L. (2019). Diminished value discrimination in obsessive-compulsive disorder: A prospect theory model of decision-making under risk. Frontiers in Psychiatry, 10, 469.** This work studied decision-making in a clinical cohort consisting of \\u2018patients diagnosed with OCD (n = 10), generalized anxiety disorder (n = 15), social anxiety disorder (n = 14), and healthy controls (n = 20) [which] were given a decision-making task and choices were modeled using a cumulative prospect theory framework.\\u2019 Here is what authors mention in their introduction: \\u2018Recently, behavioral neuroeconomic tools have been touted as having high potential utility in assessing the decision-making characteristics of people with psychiatric disorders (1, 8, 9). Such an approach computes \\u201coptimal\\u201d or normative behavior on a variety of dimensions, thus allowing for precise quantification of deviation from these norms. The traditional view conceptualizes decision-making as a rational process involving simple comparisons of expected values or expected utilities. However, because human behavior routinely deviates from purely \\u201crational\\u201d choice, cumulative prospect theory offers empirically validated mathematical formulations of psychological effects in decision-making, such as loss aversion, and the circumstances when risk-seeking or risk-averse behaviors are likely to occur.\\u2019 Please refer to e.g. Figure 3 in their paper for estimated parameters of the CPT probability weighting and utility functions as well as further details regarding the results.\", \"**Sip, K. E., Gonzalez, R., Taylor, S. F., & Stern, E. R. (2018). Increased loss aversion in unmedicated patients with obsessive\\u2013compulsive disorder. Frontiers in Psychiatry, 8, 309.** This is a study conducted on 43 obsessive\\u2013compulsive disorder patients across two sites (Icahn School of Medicine at Mount Sinai in New York and University of Michigan) supported by the US National Institutes of Health (NIH). \\u2018Obsessive\\u2013compulsive disorder (OCD) patients show abnormalities in decision-making and, clinically, appear to show heightened sensitivity to potential negative outcomes. Despite the importance of these cognitive processes in OCD, few studies have examined the disorder within an economic decision-making framework. Here, we investigated loss aversion, a key construct in the prospect theory that describes the tendency for individuals to be more sensitive to potential losses than gains when making decisions.\\u2019 The authors add: \\u2018These data identify abnormalities of decision-making in a subgroup of OCD patients not taking psychotropic medication. The findings help elucidate the cognitive mechanisms of the disorder and suggest that future treatments could aim to target abnormalities of loss/gain processing during decision-making in this population.\\u2019\"]}", "{\"summary\": \"This paper is a follow-up work to [L.A. et al., 2016] on Cumulative Prospect Theory (CPT) value optimization in reinforcement learning problems. The original paper utilized the simultaneous perturbation stochastic approximation (SPSA) method to update the policy, while this paper provides a policy gradient method. The policy gradient is evaluated and compared with the SPSA method in several domains.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": [\"Discussion on the optimal policies under CPT, showing that optimal policies are generally stochastic and non-Markovian when CPT is applied.\", \"Policy gradient algorithm for CPT value optimization, compared to SPSA methods in the original paper.\"], \"weaknesses\": [\"The algorithm is only validated in a few small domains, making it difficult to assess the performance of this policy gradient in more complex tasks.\", \"The computation of the integral for the CPT value appears complex and time-consuming due to the use of Monte Carlo sampling.\"], \"suggestions\": [\"The authors may consider demonstrating the benefits of using CPT over standard risk-neutral RL in a concrete reinforcement learning problem at the beginning of the paper to help readers better understand CPT.\", \"This paper considers the problem in a non-discounted setting, i.e., $\\\\gamma=1$ as shown in Line 181. The claim in line 261 does not apply to the discounted case for exponential utilities. Discussion may be needed. see [Entropic Risk Optimization in Discounted MDPs, AISTATS 2023]\", \"As mentioned in line 309, Proposition 7 in the paper is different from the Proposition 6 in [L.A. et.al. 2016], authors may need to provide a comparison and justification.\"], \"questions\": [\"$\\\\phi(R(\\\\tau))$ in line 295, 303, 305 should be $\\\\varphi(R(\\\\tau))$ in Theorem 6?\", \"How to choose the proper utility function for different problems?\", \"The policy gradient calculation depends on the property that the function $u$ is non-decreasing. Can we work with other $u$ where non-decreasing is not guaranteed?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response to the reviewer's comments\", \"comment\": [\"Thank you for your feedback.\", \"**Our motivation and goal.** Please note that our goal is not to design an algorithm to compare to existing special risk measures which are studied in the literature. CPT-RL is our problem formulation and it has many interesting features that are not captured by other existing distortion risk measures (as we have discussed in our previous response) which also have their own merits depending on the application and the goal the agent pursues. Our main motivation is to address the CPT-RL problem and showcase its importance. We do not believe that comparing to existing risk-sensitive RL algorithms for specific risk-sensitive objectives brings an added value in that it \\u2018will help ensure that the learned CPT policy\\u2019s performance aligns meaningfully with their respective objectives\\u2019 as the reviewer suggests. Comparing to other algorithms for special cases of our problem (which do not emphasize or even consider the motivation of CPT-RL and its features) does not enhance our main motivation in this work.\", \"**About our experiments.** We have tested our algorithm in a number of applications, going far beyond prior work in CPT-RL which only considered solving a single traffic control application on a 2x2 grid using a zeroth order algorithm. We have investigated their alignment with the objective we pursue, investigating even the policies we obtain. Our additional experiments explore the sensitivity to different reference points as well as different parameters of the CPT model to show how CPT return is affected, including in risk-sensitive variants of the problem. **In the financial training and MuJoCo experiments we considered, existing risk-sensitive RL algorithms cannot be applied to our setting in which we consider both probability weight and utility functions (which are not both the identity, nor they are any step function such as the ones needed to model Var or CVar) such as the KT model.**\", \"**Instantiation of our PG algorithm for exponential criteria.** Particularizing our problem and algorithm to the special case of the risk-sensitive setting with exponential criteria essentially results in an algorithm which is the PG algorithm proposed in Noorani et al. 2023 and we refer the reader to this paper for comparisons to other risk-sensitive algorithms for that particular risk-sensitive task. Our work focuses on showcasing the benefits of CPT-RL objectives which were not previously studied in practice.\", \"**Comparison to SPSA which was also investigated in risk-sensitive RL.** We also note that prior work in risk-sensitive RL (e.g. Vijayan and L. A. AISTATS 2023 that we mention in related works) has also used similar SPSA algorithms as the one we compare to in section 5. (a) (for CPT-RL) and we did compare to such an algorithm to show the advantage of our method.\", \"> \\u2018Also, the paper is difficult to follow. The writing and presentation of the manuscript need improvement to make the content clear and accessible to readers. I will maintain the current score.\\u2019\", \"**Presentation.** Regarding the presentation, we would like to highlight that reviewer aLSN (in its original review) finds that **\\u2018the paper\\u2019s presentation is clear\\u2019**, \\u2018the theoretical and algorithmic contributions are original and **clear**. If one believes that CPT-based RL problems are important, then these contributions are also significant' and that \\u2018the paper did a very nice job relating its studied problem with other related problem settings studied previously, **making it easy for readers to understand the position of this paper in the literature**.\\u2019 Please let us know if you have any concrete and constructive comment to further improve our presentation and we will be happy to take it into account. During the rebuttal, we have also made substantial efforts to revise the manuscript and strengthen it, please see our general comment for a summary of our revisions where we also highlight our contributions and their importance.\"]}", "{\"title\": \"Response to 3rd question\", \"comment\": [\"> The policy gradient calculation depends on the property that the function u is non-decreasing. Can we work with other u where non-decreasing is not guaranteed?\", \"Strict monotonicity is needed in some of the proofs of our results, mainly Theorem 4, see e.g. proof of 1 implies 2 in p. 24 and proof of 5 implies 2 in p. 22. However, it is not formally required to derive our policy gradient theorem. Note also that even Proposition 7 for estimating the integrals does not require it as soon as $\\\\xi_{i/n}^{+,-}$ denote the right quantiles as we defined them in the proposition. See point 3 below for further comments.\", \"We focus on the monotone setting because typical human behavior tends to prefer better outcomes over worst ones, e.g. higher gains over smaller ones and smaller losses over larger losses. This is captured by the monotonicity assumption on the utility function: utility increases with increasing gains and decreases with increasing losses. This is a fundamental assumption in economics and decision theory which is consistent with how humans evaluate outcomes. Nevertheless, there are cases where this assumption might not hold if this is what the reviewer is referring to.\", \"Mere monotonicity is mainly useful in our algorithm to guarantee that the utility values are sorted in the same order as the returns obtained from the sampled trajectories (see step 6 of the algorithm for quantile estimation computation). This leads to a simple computation of the quantiles of the utilities: Once the returns are sorted, the utilities are also sorted in the same way and quantiles can be read from the sorted list. If monotonicity does not hold anymore, one has to be more careful about this computation and adapt step 6 of the algorithm accordingly by simply computing the quantiles of the utilities using the right sorting of these (which would be different from the sorting of the returns). In that case, sorting the returns in step 5 is not needed and one only needs to sort the utilities (utility function applied to the returns). Apart from this technical detail, we do not foresee any major impediment to using the algorithm without the strict monotonicity assumption.\", \"We thank the reviewer for this question, we will add a remark to the paper accordingly.\", \"Thank you for your review, please let us know if you have any further concern or questions.\"]}", "{\"title\": \"Additional example and experiment\", \"comment\": \"We thank the reviewer for their feedback. We address their concerns in the following. Regarding the importance of CPT-RL, we refer the reviewer to our detailed response to reviewer aLSN and appendix C p. 17 which we added to the manuscript. We have also augmented the introduction with l. 63-68 to highlight applications. Concerning examples motivating CPT-RL, we provide another example below and we have now added it to the main part of the paper in section 2 p. 4-5 to illustrate our CPT-RL problem formulation as suggested by the reviewer (deferring related work to the appendix due to space constraints). As for the utility function choice, we will add a discussion following our response to your question above.\\n\\n**Additional concrete example of CPT-RL in healthcare: Personalized Treatment for Pain Management.** CPT is not just a generalization, its features are important in applications, especially when human perception and behavior matter. Existing traditional RL approaches often lack a behavioral perspective and CPT-RL allows for more empathetic and realistic sequential decision making. Here is another concrete example to illustrate the importance of CPT-RL and its differences compared to risk sensitive RL to provide more intuition to the reader. \\n\\n-**Scenario**: The goal is to help a physician manage a patient's chronic pain by suggesting a personalized treatment plan over time. The challenge here is to balance pain relief and the risk of opioid dependency or other side effects that might be due to the treatment, i.e. short-term relief and longer term risks. \\n\\n-**Our approach**: We propose to train a CPT-RL agent to help the physician. \\n\\n**Why sequential decision making?** \\n\\n(a) The physician needs to adjust treatment at each time step depending on the patient\\u2019s reported pain level as well as the observed side effects. Note here that this is relevant to dynamic treatment regimes in general (such as for chronic diseases) in which considering delayed effects of treatments is also important (and RL does account for such effects). We refer the reader to section IV of Yu and Liu 2020, \\u2018Reinforcement Learning in Healthcare: A Survey.\\u2019 ACM Computing Surveys.\\n\\n(b) Decisions clearly impact the patient's immediate pain relief, dependency risks in the future and their overall health condition.\\n\\n**Why CPT?** Patients and clinicians make decisions influenced by psychological biases. We illustrate the importance of each one of the three features of CPT as introduced in our paper in section 2 (reference point, utility and probability distortion weight functions) via this example: \\n\\n (a) *Reference points*: Patients assess and report pain levels according to their subjective (psychologically biased) baseline. Incorporating reference point dependence leads to a more realistic model of human decision-making as this allows for capturing e.g. expectations, past experience as well as their desired outcomes to define their perceived gains and losses. In our example, reducing pain from a level of 7 to 5 is not perceived the same way if the reference point of the patient is 3 of it is 5. In contrast, risk-sensitive RL treats every pain reduction as a uniform gain, regardless of the patient\\u2019s starting reference pain level. \\n\\n(b) *Utility transformation*: Patients might often show a loss averse behavior, i.e., they might perceive pain increase or withdrawal symptoms as worse than equivalent gains in pain relief. **Note here that loss aversion should not be confused with risk aversion** (see definition and discussion in Schmidt and Zank 2005. \\u2018What is loss aversion?\\u2019 The Journal of Risk and Uncertainty.) In short, loss aversion can be defined as a cognitive bias in which the emotional impact of a loss is more intense than the satisfaction derived from an equivalent gain. For instance, in our example, a 2-point increase in pain might be seen as much worse than a 2-point reduction even if the change is the same in absolute value. This loss aversion concept is a cornerstone of Kahneman and Tversky\\u2019s theory. In contrast, risk aversion rather refers to the **rational** behavior of undervaluing an uncertain outcome compared to its expected value. Risk sensitive approaches might be less adaptive to a patient\\u2019s subjective preferences if they deviate from objective risk assessments.\\n\\n(c) *Probability weighting*: Low probability events such as severe side effects (e.g., opioid overdose or dependency) might be overweighted or underweighted based on the patient's psychology.\\n\\n**Environment and transitions.** A state is a vector of three coordinates (current pain level, dependency risk, side effect severity). Actions are treatments, e.g. no treatment, alternative treatment or opioid treatment. An episode ends if the patient develops full dependency or if pain is effectively managed.\"}", "{\"title\": \"Thank you\", \"comment\": \"Thank you for your response. Let us mention that we have clearly indicated the continuity and differentiability assumptions in the statement of our policy gradient theorem (Theorem 6). We will surely make this clearer as for the special cases like CVar and Var, thank you for the suggestions regarding this that we will follow.\"}", "{\"title\": \"About applicability\", \"comment\": \"> Limited applicability: The experiments did not demonstrate the advantages of using the proposed algorithm. It is unclear what the benefits of using CPT learning are over risk-sensitive RL or distributional RL. Although CPT is theoretically valuable, the empirical advantages of CPT over other risk-sensitive measures remain ambiguous in the presented results. The authors may consider including direct comparisons with risk-sensitive RL or distributional RL methods on the same tasks.\\n\\n**Our general PG theorem result and our novel proposed algorithm** expand the applicability of CPT-RL in practice besides being theoretically grounded. We would like to highlight that our PG theorem unifies several settings including standard RL, risk sensitive, risk seeking, probability distorted settings and beyond under the umbrella of CPT-RL. Our algorithm which stems from such a general result is therefore general as for the different problem settings it can address. We refer the reader to p. 17 for a diagram representing our framework in the literature and Table 1 for different special case examples. \\n\\n**Advantage and applicability.** We devoted a specific part of our experiments to clearly illustrate the scalability advantage of our PG algorithm compared to the zeroth order algorithm (please see sec. 5. (b) and fig. 3, the performance of the 0th order algorithm gets even worse with an even larger scale). We also demonstrated its applicability to different settings including the case of continuous state action settings, different utility functions including risk seeking and risk averse ones, and different probability weight functions as well. \\n\\n**Comparison to Risk sensitive RL.** Risk sensitive RL is a particular case of our framework and our goal is not to outperform existing algorithms for specific tasks using our general algorithm. We also provided several examples where we use risk-sensitive measures though. We also note that prior work in risk-sensitive RL (e.g. Vijayan and L. A. AISTATS 2023 that we mention in related works) has also used similar SPSA algorithms as the one we compare to in section 5. (a). Moreover, particularizing our problem and algorithm to the special case of the risk-sensitive setting with exponential criteria results in an algorithm that bears similarity to existing work and PG algorithms such as the work of Noorani et al. 2023 and we refer the reader to this paper for comparisons to other risk-sensitive algorithms for that particular task. Distributional RL focuses on modeling and computing the entire distribution of returns. While distributional RL allows handling risk explicitly by approximating the distribution of returns, it is in general computationally demanding and it does not model subjective risk handling via utility and probability distortions like CPT.\\n\\n**Comparison to CPT.** \\n- CVar or VaR in risk sensitive RL do not take into account transformed utilities (modeling perceived/subjective gain and loss values differently and wrt reference points) whereas exponential risk sensitive RL does not account for probability weighting (which models low and high probability events subjective perception). To appreciate this, please see Table 1 in p.17 and our experiments where we compare the different examples (please see figure 2). \\n\\n- When an agent perceives a return differently from the raw return, then this is also modeled in our CPT framework and it is clear that CPT is more suitable than existing risk sensitive approaches. This is the case when (a) it has a different reference point regarding their perceived value, (b) its utility is not linear in the return, it is rather concave for gains and convex for losses (wrt their reference). \\n- Most importantly, we stress that our CPT framework captures several existing risk sensitive measures as particular cases and offers the flexibility to the decision maker to design their own. In particular, CPT can at the same time overweight low probability events and underweight highly probable ones with different magnitudes. Therefore, CPT allows to handle losses and gains separately. A similar flexibility is offered for utility functions. See figure 6 p. 18 for illustrations.\"}", "{\"comment\": \"Thank you to the authors for the responses and the additional experiments.\\n\\nSome concerns remain. The authors stated that \\u201cseveral risk measures are also particular cases of CPT values: Variance, Conditional Value at Risk (CVaR), distortion risk measures, to name a few.\\u201d It would be worthwhile to demonstrate that the proposed algorithm is at least comparable to existing risk-sensitive RL algorithms in some particular cases, such as under the popular risk metric CVaR. This comparison will help ensure that the learned CPT policy\\u2019s performance aligns meaningfully with their respective objectives. For example, in the financial trading and MuJoCo experiments, it is unclear how the proposed algorithm performs compared to existing risk-sensitive RL algorithms.\\n\\nAlso, the paper is difficult to follow. The writing and presentation of the manuscript need improvement to make the content clear and accessible to readers. I will maintain the current score.\"}", "{\"title\": \"Response to remaining comments and questions\", \"comment\": \"> As mentioned in line 309, Proposition 7 in the paper is different from the Proposition 6 in [L.A. et.al. 2016], authors may need to provide a comparison and justification.\\n\\nWe provided a brief comment about this difference in l. 343-346. As we mention it, the reason for this is that the policy gradient theorem (Theorem 6) involves the exact same integral term (l. 288) that we approximate in Proposition 7 (l. 319) and this term features a first order derivative of the weight function since we are considering gradients. L. A. et al. 2016 do not require such derivative in the integral because they only consider zeroth order estimates of the policy gradient. Overall, the approximation result is fundamentally the same as the Riemann scheme approximation of the integral using simple staircase functions does not crucially depend on the integrand ($w\\u2019$ in our case, $w$ in their case) which we replace for our purpose. \\n\\n> $\\\\phi(R(\\\\tau))$ in line 295, 303, 305 should be $\\\\varphi(R(\\\\tau))$ in Theorem 6? \\n\\nThank you for spotting the typo. It is corrected now. \\n\\n> How to choose the proper utility function for different problems?\\n\\n- The problems themselves might dictate to the user or decision maker the utility function to be used. The user might also design their own according to their own beliefs, behaviors and objectives, based on the goal to be achieved (e.g. risk-seeking, risk-neutral, risk-averse). Specific applications might also suggest specific utility functions such as specific risk measures like in risk sensitive RL for instance. We have provided in table 1 p. 17 a list of different examples one might consider. Learning the utility function is also an interesting direction to investigate as we mention in the conclusion. In practice, it is rather common to use the example we provide in table 1 p. 17 (CPT row) with exponent parameters which are estimated using data. \\n\\n- We provide a few concrete examples in the following. For instance, Rieger et al. 2017 (see reference below in response to reviewer aLSN) adopt such an approach (see sections 3.1, 3.2 and 3.3 therein for a detailed discussion about parameter estimation). Ebrahimigharehbaghi et al. 2022 (see reference below) choose some similar variation of this utility (see eqs. 2-3 therein) while still using KT\\u2019s probability weighting functions. Gao et al. 2023 compare different functions for different similar power utility functions with fitted parameters (see Tables 1, 2 and 3 therein p. 3, 4, 6 for extensive comparisons with the existing literature). Similar investigations were conducted in Yan et al. 2020. Dorahaki et al. 2022 consider psychological time discounted utility functions (variations of the same power functions) in their model with additional relevant hyperparameters, motivated by (domain-specific) psychological studies (see eq. (4) therein). It is worth noting that all these examples are only in the static stateless setting.\\n\\nWe thank the reviewer for the question, we will add a remark along these lines to our paper about this point.\"}", "{\"title\": \"About our theoretical contributions and response to minor point\", \"comment\": \"**About our theoretical contributions.** Besides our algorithmic contributions and experiments, we would like to draw the attention of the reviewer to our theoretical contributions. We are not aware of any work which establishes a PG theorem in this level of generality. This result on its own is conducive to a number of possible algorithmic schemes from which we propose a vanilla PG algorithm (one can think about actor-critic methods, n-step methods, variance reduced schemes and many others). We investigate the nature of optimal policies on our problem and show the differences with respect to the standard RL setting. In particular, we characterize the utility functions allowing for Markovian policies (which are sufficient policies in standard RL but not necessarily in our setting) and we show that they reduce to the class of affine and exponential utilities. We are not aware of any such result in the literature.\\n\\n> **Minor comment:** On page 4, under Problem formulation: CPT-RL, the cumulative discounted rewards variable X is referenced, but the definition provided does not discount the rewards. It would be clearer to either update the variable's definition to include discounting or adjust the notation accordingly to prevent confusion.\\n\\nThank you for spotting this. Throughout the paper, we focus on the finite horizon undiscounted setting corresponding to the formulation of page 4 and the definition of X in the same page. We comment on the extensions to the discounted and infinite horizon settings in remarks 8 (l. 887) and 13 (l. 1373). This is fixed now on page 4. \\n\\nWe thank the reviewer again for their time and feedback. Please let us know if you have any further comment or questions, we will be happy to address them.\"}", "{\"comment\": \"**CPT-RL vs Risk-averse RL: comparison.**\\nIn terms of policies, risk-averse RL would favor non-opioid treatments unless extreme pain levels make opioids justifiable. In contrast, CPT-RL policies would prescribe opioids if pain significantly exceeds the patient\\u2019s reference point. As dependency risk increases, CPT-RL policies would transition to non-opioid treatments as a consequence of overweighting the probability of rare catastrophic outcomes. Notably, CPT-RL policies can oscillate between risk-seeking (to address high pain) and risk-averse (to avoid severe side effects). In contrast, a risk-sensitive agent focuses on minimizing variability in health states and dependency risks and would likely avoid opioids in most cases unless pain levels become extreme. Such risk-sensitive policies favor stable strategies (e.g., consistent non-opioid use), prioritizing low variance in patient outcomes.\\n\\n**Experiments.** In our simulations, we have focused on examples emphasizing the behavioral economics motivation of CPT-RL. As for Mujoco, we have **now also added an example on the inverted pendulum environment to demonstrate that our PG algorithm can be readily used in such environments as well (see page 38 in the appendix).**\"}", "{\"title\": \"Response to reviewer aLSN\", \"comment\": \"We thank the reviewer for their time, for their valuable feedback and thoughtful comments. We appreciate that the reviewer is faithfully and accurately reporting our contributions. We provide a detailed discussion below regarding the reviewer\\u2019s concerns about the importance of the problem and its relevance in comparison to other approaches such as trajectory-based RL and RLHF in particular.\\n\\n> While I appreciate the contributions the authors made to CPT-based RL problems, I am not convinced of the importance of CPT-based RL problems by reading this paper. The paper cites another paper \\\"... this is particularly important in applications directly involving humans in the loop such as e-commerce, crowdsourcing and recommendation to name a few\\\" to argue that this class of problems has important applications.\\n\\nCPT is a popular model in behavioral economics originating from economics and psychology that was recognized by a Nobel prize attributed to Daniel Kahneman in 2002. We believe that developing a paradigm extending CPT to sequential decision making is an important and natural extension as it allows to broaden the scope of applications and factor in its main features by taking into account the subjective valuation of outcomes (utility) and the subjective weighting of probabilities. CPT-based RL unifies different settings including in particular risk-sensitive RL that has by now endless applications in RL.\\nPlease see our response below to your last comment for a detailed discussion regarding the importance of our problem and existing applications. We thank the reviewer for their comment, we will certainly add further motivation in the introduction to further support the importance of our problem. We have now added section C in the appendix p. 17 (due to space constraints) to discuss applications along the lines of the last comment below. \\n\\n> However, it is not clear from the cited paper whether other formulations that take into account human behavior can also handle these applications well. In fact, compared with CPT, I feel that a more principled way that considers human behavior is trajectory-based reward RL problems (one reward for the entire trajectory) with human preferences, like RLHF in large language models. While CPT assumes a structure of the final reward (the final metric to optimize involves a weight function, a utility function, and a summation of rewards), these weight functions are unknown and must be learned from data. In contrast, in trajectory-based reward RL problems, there is no assumption on the structure of the metric being optimized. And the metric is learned with human preference data. I wonder how authors think about the pros and cons of CPT in terms of its applications compared to trajectory-based reward RL problem settings.\\n\\nFirst of all, we do not exclude that other approaches might also be useful for modeling human behavior (and there are others as you mention), this is an open and promising research area that we believe has yet many interesting research directions to offer. We also do not claim that our approach is the unique best way to tackle the problem. Nevertheless, we do believe it is a principled approach rooted in an established literature in behavioral economics that would gain to be developed for sequential decision making. Our paper contributes to this effort by proposing a practical PG algorithm. While the CPT approach has not yet been well-established in the machine learning community, we believe it has interesting features to offer and it is already pervasive in RL through its particular cases risk-sensitive and safe RL. We provide a more detailed discussion below regarding these aspects and comment on your interesting question regarding the comparison with other paradigms such as RLHF/trajectory-based reward RL and the pros and cons of CPT. \\n\\nPlease see the rest of our response to your above comment in what follows.\"}", "{\"title\": \"response (continued)\", \"comment\": \"Here a few additional relevant studies:\\n- *Zhao, M., Wang, Y., Meng, X., & Liao, H. (2023). A three-way decision method based on cumulative prospect theory for the hierarchical diagnosis and treatment system of chronic diseases. Applied Soft Computing, 149, 110960.* \\n- *Sun, J., Zhou, X., Zhang, J., Xiang, K., Zhang, X., & Li, L. (2022). A cumulative prospect theory-based method for group medical emergency decision-making with interval uncertainty. BMC Medical Informatics and Decision Making, 22(1), 124.* \\n\\n2. **Question 1.** One has to be careful about such extensions because the nonlinearity of both the utility and probability weighting functions prevents from leveraging dynamic programming, especially under our (natural) formulation of the CPT-RL problem. There are a few works discussing value-based methods for CPT (e.g. Q-learning) that we mentioned in our related works section (e.g. Borkar and Chandak 2021, Ramasubramanian et al. 2021). However, as we discuss it in related works, \\u2018these works are concerned with maximizing a sum of CPT value period costs which is amenable to dynamic programming. In contrast to their accumulated CPT-based cost (see their remark 1), our CPT policy optimization problem formulation is different: we maximize the CPT value of the return of a policy (see (CPT-PO)). In particular, this objective does not enjoy an additive structure and hence does not satisfy a Bellman equation.\\u2019\\n\\n**Question 2.** CVar, Var and distortion risk measures are indeed CPT values **if we allow for discontinuous probability weighting functions**, we provided a short proof for our claim in appendix E.3 for completeness. It is based on distortion risk measures from which VaR and CVaR have been shown to be special cases (Wirch & Hardy, 2001). To see this, for Var, we use a step function that focuses on the quantile. For CVaR, we use a piecewise uniform weighting function for the tail. The idea is that considering discontinuous weighting functions allows CPT to focus only on specific parts of the probability distribution which are relevant to Var and CVar. In general, probability weights and utility functions are supposed to be continuous as people might less frequently exhibit a behavior corresponding to hard thresholding probabilities. That was the point of the footnote. VaR and CVaR are rather rooted in the financial risk management literature and rely on objective probability distributions whereas CPT originates from behavioral economics and focuses on subjective risk perception and decision-making.\\n\\nThe footnote in the caption is a bit misleading, we will update the formulation, the point of that footnote is to mention for completeness and to be completely rigorous that in those cases the probability weight functions are not continuous and this is a caveat. The point of our remark in the main part of the paper is to illustrate the modeling ability of the CPT framework in general. \\n\\nWe suppose differentiability of the probability weight functions in our paper to be able to compute gradients everywhere for simplicity. Note though that non-differentiability of the weight functions is actually mild in those cases as it only occurs at a single point for Var and CVar, please see figure 6 in appendix E.2 (p. 19) where we represent these weight functions for Var and CVar for illustration purposes. To apply our first-order algorithm to non-smooth settings without worrying about this single point issue for example, one can consider smooth approximations of these distortion functions (we did discuss and test such an alternative, see appendix G.2 for details). One can also further consider Clark subdifferentials to address the nonsmooth setting, we stick to the differentiable case which already allows us to consider a wide range of utility and weighting functions for simplicity and clarity of the paper. We note though that there are better approaches to address the specific case of CVar in terms of optimization (by fully leveraging its structure, see e.g. Meng, S. Y., & Gower, R. M. A model-based method for minimizing CVaR and beyond. ICML 2023, for an approach based on subgradients as we highlight and a more sophisticated one using a stochastic prox linear method, see also Duchi and Ruan 2018 or Davis and Drusvyatskiy 2019). We highlight that our main goal is to make use of the features of general CPT and not just to generalize prior risk measures which do not capture all the important features of CPT that we have expanded on. As a side note, we also do not claim that our method should be uniformly the best algorithm to use for all settings (as one might often exploit specific structures of special problems). We hope this clarifies your confusion, we will further clarify this in the paper. \\n\\nWe hope this answers your questions precisely. Please let us know if you have any further questions, we will be happy to address them.\"}", "{\"title\": \"My feedback and two follow-up questions\", \"comment\": \"Your feedback on the usefulness of CPT-RL is appreciated. One suggestion is to make the example in your paper more concrete. Specifically, in the patient treatment example, you explained the potential benefits of weight functions, utility functions, and reference points. However, it would be more convincing if you could reference a study, if available, that models these functions and points using real patient treatment data.\\n\\nI have another question regarding the extension of algorithms beyond Policy Gradient (PG) to the CPT setting. Specifically, when it comes to learning a value function, do you think there are straightforward ways to adapt RL algorithms like TD learning or Q-learning to the CPT framework?\\n\\nLastly, I'm a bit confused about your claim that \\\"several risk measures are also particular cases of CPT values: Variance, Conditional Value at Risk (CVaR), distortion risk measures, to name a few.\\\" You also mentioned in the caption of Table 1 that \\\"w+ and w\\u2212 are often required to be continuous, which would exclude VaR and CVaR.\\\" These two statements seem conflicting\\u2014one suggests that CPT encompasses CVaR, while the other implies it does not. Could you clarify this?\"}", "{\"title\": \"Response continued\", \"comment\": \"- **CPT and RLHF are not mutually exclusive.** While CPT and trajectory-based RL (say e.g. RLHF) both offer frameworks for incorporating human preferences into decision making, we would like to highlight that CPT and RLHF are not mutually exclusive. We can for instance use CPT to design an initial reward structure reflecting human biases, then refine it with RLHF. We can also consider to further relax the requirement of sum of rewards (which already has several applications on its own) and think about incorporating CPT features to RLHF. Some recent efforts in the literature in this direction that we mentioned in our paper include the work of Ethayarajh et al. (ICML 2024) \\u2018Model alignment as prospect theoretic optimization\\u2019 which combines prospect theory with RLHF (without probability weight distortion though, which limits its power). Note that the ideas of utility transformation and probability weighting are not crucially dependent on the sum of rewards structure and can also be applied to trajectory-based rewards or trajectory frequencies for instance. We believe this direction deserves further research, one interesting point would be how to incorporate risk awareness from human behavior to such RLHF models using ideas from CPT.\\n\\nThank you for the interesting question. We have now incorporated this discussion to appendix D (p. 18, in the revised manuscript) due to space constraints. \\n\\n\\n> Another weakness is related to the empirical study of the paper. The experiments only demonstrate the effectiveness of the proposed algorithm but do not include those justifying the necessity of the CPT problem. It would be much better if the authors could demonstrate that in some applications, the data shows that human behavior can be accurately predicted with CPT but not other theories.\\n\\n**About the importance of our problem formulation and its empirical relevance.** While our empirical results do not focus on the importance of our CPT problem formulation and mainly serve an illustrative purpose given our theoretical and algorithmic contributions to CPT-RL, we would like to stress that CPT has been tested and effectively used in a large number of compelling behavioral studies that we cannot hope to give justice to here. It is far from being an intellectual theoretical curiosity disconnected from practice. Besides the initial findings of Tversky and Kahneman for which the latter won the Nobel Prize in economics in 2002, please see a few recent references below for a broad spectrum of real-world applications ranging from economics to transport, security and energy, mostly in the stateless (static) setting, also reflecting how active and impactful this research is in other fields. We quote some of the results for each paper for the convenience of the reader to reply to the reviewer\\u2019s concern. We hope these examples convince the reviewer of the importance of the problem we address which is an extension to the dynamical setting via our RL formulation for sequential decision making. \\n\\n1. *Rieger, M. O., Wang, M., & Hens, T. (2017). Estimating cumulative prospect theory parameters from an international survey. Theory and Decision, 82, 567-596.*\\n\\u2018We conduct a standardized survey on risk preferences in 53 countries worldwide and estimate cumulative prospect theory parameters from the data. The parameter estimates show that significant differences on the cross-country level are to some extent robust and related to economic and cultural differences. In particular, a closer look on probability weighting underlines gender differences, economic effects, and cultural impact on probability weighting.\\u2019 \\nNote here the explainability feature of CPT that we highlighted above in our discussion. \\n\\n2. *Ebrahimigharehbaghi, S., Qian, Q. K., de Vries, G., & Visscher, H. J. (2022). Application of cumulative prospect theory in understanding energy retrofit decision: A study of homeowners in the Netherlands. Energy and Buildings, 261, 111958.* \\u2018CPT correctly predicts the decisions of 86% of homeowners to renovate their homes to be energy efficient or not. EUT, on the other hand, overestimates the number of decisions to renovate: it incorrectly predicts retrofit for 52% of homeowners who did not renovate for energy efficiency reasons. Using the estimated parameters of CPT, the cognitive biases of reference dependence, loss aversion, diminishing sensitivity, and probability weighting can be clearly seen for different target groups.\\u2019\"}", "{\"title\": \"Revision and contributions\", \"comment\": \"We would like to thank the reviewers again for their feedback and their valuable suggestions.\\n\\n**Revision.** In addition to our detailed responses to each one of the reviewers, we have uploaded a revised version of our manuscript containing the following main modifications (highlighted in blue) to improve our work based on the reviewers\\u2019 feedback which we are grateful for: \\n- Regarding the importance of CPT-RL (highlighted by reviewers aLSN and hWfL), we have updated the introduction with l. 63-68 to highlight applications of CPT in general and we have incorporated a new extended discussion in appendix C p. 17. \\n- Concerning examples to clarify the features of CPT-RL and its comparison to risk-sensitive RL, we have added another concrete example in healthcare (personalized treatment for pain management) to section 2 right after the problem formulation to illustrate it and highlight why we consider CPT-RL instead of prior existing risk-sensitive approaches. We will further update the manuscript to incorporate more details in the lines of our response to reviewer hWfL. \\n- We have added appendix D p. 18-19 to compare CPT-RL and trajectory-based reward RL as preference based learning paradigms to address the questions and concerns of reviewer aLSN. \\n- We deferred the related work section to appendix B due to space constraints. \\n- **Experiments:** In addition to our simulations in grid worlds, simple traffic control settings as well as our electricity management application using real-world data, we have added simulations for trading in financial markets (appendix H. 7, as suggested by reviewer Ydtj) and on the Mujoco inverted pendulum environment (appendix H. 8, as suggested by reviewer hWfL) to show the applicability of our PG algorithm to other settings. \\n\\n**Our key contributions.** Our main goal in this work is to study and solve a policy optimization problem accounting for human perception and behavior via the use of the well-established cumulative prospect theory from behavioral economics. We provide further insights on the optimal policies and their nature in such problems compared to standard RL policy optimization for maximizing expected returns. We would like to highlight that one of our main contributions is our PG theorem and algorithm to solve this problem. We believe that our practical PG algorithm is widely applicable. We have tested it in several applications, **going far beyond prior work in CPT-RL which only considered solving a single traffic control application on a 2x2 grid using a zeroth order algorithm**. Testing it in other applications and even larger scale problems would certainly be interesting. Our work also opens up the way to several interesting future work opportunities towards more realistic human-based sequential decision making both from the algorithmic and methodological viewpoints (such as learning reference points, utilities and probability weight functions as we highlight in our paper and in our responses to reviewers). \\n\\n**Potential.** We are excited by the potential that CPT-RL has to offer to better model human sequential decision making and incorporate cognitive and psychological biases which are of most importance in several high-stakes applications. Only few works have been devoted to this endeavour in RL and our work contributes to this goal. We have only scratched the surface in this regard by showcasing the potential of solving such a CPT-RL problem formulation in simple settings for healthcare decision making, energy, finance and investment. Many other meaningful applications such as legal and ethical decision making, cybersecurity and human-robot interaction are yet to be explored. \\n\\nWe would be happy to address any further questions or concerns during the remainder of the discussion period.\"}", "{\"title\": \"Rebuttal: response to Reviewer Ydtj\", \"comment\": \"We thank the reviewer for their feedback. Please find below a point by point reply to your comments and concerns.\\n\\n> 'The experimental section provides limited insight due to its use of relatively straightforward examples (traffic control and electricity management), which may not fully illustrate the complexities that arise in more realistic, high-stakes environments, such as finance or healthcare.'\\n\\n**Traffic control and electricity management** are major high-stakes and complex applications which are extremely active research areas. Such applications need no motivation regarding their complexity and relevance in the real world. As a matter of fact a quick keyword search on app.dimensions.ai shows more than 200 000 (resp. more than 100 000) publications around traffic control (resp. electricity management) in 2023. \\n\\n**Our simulations.** While our simulations rely on simplified models that cannot capture all the real-world intricacies of such complex applications, our goal is to illustrate how our problem formulation can be useful to solve concrete problems by giving a flavor about the scenarios one might consider in practice. We purposefully designed them to go beyond the standard vanilla gridworld that we have also considered. In particular, please note that we are using real-world public data available online for our simple electricity management application (please see appendix F. 6, Figure 19). We report electricity prices on a typical day on the market together with the total electricity production. \\t\\n\\nBesides our theoretical contributions which we would like to highlight, please let us briefly recall the purpose of each one of the applications we consider and the insights we provide: (a) in our traffic control application, we show the influence of the probability distortion function, (b) we illustrate the scalability of our PG algorithm to larger state spaces compared to the existing zeroth order algorithm in a grid environment with increasing state space size and (c) we test our algorithm in a continuous state-action space setting in our electricity management application. \\n\\n**Additional simulations.** Finance and healthcare are also important applications. We thank the reviewer for mentioning these. Exploring all these applications in depth is beyond the scope of this paper. Nevertheless, **during the rebuttal period we have now performed additional simulations in a financial application as requested by the reviewer**. The goal is to train RL trading agents using our general PG algorithm in the setting of our CPT-RL framework. We use a Gym Trading environment available online using data from the Bitcoin USD (BTC-USD) market between May 15th 2018 and March 1st 2022. One can easily use any other real-world publicly available stock market data by providing a dataframe as input to build a corresponding RL environment. Our BTC-USD data is of the same kind as any historical data which can be found online (see e.g. https://finance.yahoo.com/quote/BTC-USD/history). The RL trading agent can take three classical positions (SHORT, OUT and LONG). We refer the reader to the last pages of our revised manuscript (modifications all in blue) for the figure as well as more details regarding the experiment. We hope these additional simulations further convince the reviewer about the practicality of our algorithm. \\n\\n> In Section 5(a), the observation that different weight functions lead to different policies in the grid environment could be strengthened by assessing these policies against risk measures (such as CVaR) beyond expected return. Without this, it is unclear if the CPT-RL-PO policy outperforms standard RL-PO under any specific risk-sensitive criteria.\\n\\n Please note that we do consider a risk averse probability weight function to compare to the risk neutral case. We refer the reader to Table 2 p. 28, appendix F. 5 p. 31 for an illustration of the probability weight function w^+ used and Figure 18 for representations of the policies obtained that illustrate that the output policies solving our PO problem are meaningful compared to standard RL-PO. \\n\\n> The grid environment results are overly simplified and provide little substantive information. It would be beneficial to evaluate whether the derived policies' performance aligns meaningfully with their respective objectives.\\n\\n- Our experiment with a varying grid world state space size illustrates the better scalability of our PG algorithm compared to the previously known zeroth order algorithm (see Figure 3). \\n- We did investigate the policies that we obtain by solving our problem using our novel PG algorithm, see appendix F.4, figure 16 p. 31 for representations of the policies we obtain and how they are consistent with the objectives set. See also figure 18 p. 32. \\n- We have now performed additional simulations to apply our methodology and use our PG algorithm in financial trading (see last part of the revised manuscript for a detailed exposition).\"}", "{\"summary\": \"This paper makes three key contributions to the field of cumulative-prospect-theory (CPT)-based reinforcement learning, a framework where the agent aims to maximize not the raw cumulative reward but a transformed version of it. This transformation involves both a utility function and a probability weighting function, introduced to better reflect human value assessments of rewards (e.g., monetary amounts converted into perceived value). The objective is to find a policy that maximizes this transformed cumulative reward.\\n\\nCPT-based reinforcement learning generalizes various prior approaches, including: (1) the standard RL framework, (2) risk-sensitive RL with an exponential utility function (e.g., as in Noorani et al., 2022), (3) optimization based on VaR/CVaR of the total reward, and (4) the case studied by Kahneman and Tversky (1992).\\n\\nThe first contribution presents new insights into the characteristics of optimal policies within CPT-based reinforcement learning. These findings include: (1) deterministic, history-dependent optimal policies may not always exist; (2) in the absence of weighting functions, a deterministic, non-stationary optimal policy exists that depends on the current state and past cumulative reward; and (3) without weighting functions, a Markovian, non-stationary optimal policy exists if and only if the utility function is either affine or exponential.\\n\\nThe second contribution introduces a policy gradient theorem and a novel algorithm for CPT-based reinforcement learning.\\n\\nThe third contribution is an empirical evaluation of the proposed algorithm. This study includes (1) a performance comparison with CPT-SPSA-G, an existing RL algorithm for CPT problems, across various grid-world environments, demonstrating the proposed algorithm's advantages, particularly in larger environments; and (2) an exploration of different behaviors under varying utility and weighting functions.\", \"soundness\": \"4\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"The theoretical and algorithmic contributions are original and clear. If one believes that CPT-based RL problems are important, then these contributions are also significant.\\n\\nThe paper did a very nice job relating its studied problem with other related problem settings studied previously, making it easy for readers to understand the position of this paper in the literature.\\n\\nThe paper's presentation is clear.\", \"weaknesses\": \"While I appreciate the contributions the authors made to CPT-based RL problems, I am not convinced of the importance of CPT-based RL problems by reading this paper. The paper cites another paper \\\"... this is particularly important in applications directly involving humans\\nin the loop such as e-commerce, crowdsourcing and recommendation to name a few\\\" to argue that this class of problems has important applications. However, it is not clear from the cited paper whether other formulations that take into account human behavior can also handle these applications well. In fact, compared with CPT, I feel that a more principled way that considers human behavior is trajectory-based reward RL problems (one reward for the entire trajectory) with human preferences, like RLHF in large language models. While CPT assumes a structure of the final reward (the final metric to optimize involves a weight function, a utility function, and a summation of rewards), these weight functions are unknown and must be learned from data. In contrast, in trajectory-based reward RL problems, there is no assumption on the structure of the metric being optimized. And the metric is learned with human preference data. I wonder how authors think about the pros and cons of CPT in terms of its applications compared to trajectory-based reward RL problem settings.\\n\\nAnother weakness is related to the empirical study of the paper. The experiments only demonstrate the effectiveness of the proposed algorithm but do not include those justifying the necessity of the CPT problem. It would be much better if the authors could demonstrate that in some applications, the data shows that human behavior can be accurately predicted with CPT but not other theories.\", \"questions\": \"See weaknesses.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Further examples\", \"comment\": \"3. *Gao, D., Xie, W., Cao, R., Weng, J., & Lee, E. W. M. (2023). The performance of cumulative prospect theory's functional forms in decision-making behavior during building evacuation. International Journal of Disaster Risk Reduction, 104132.*\\n\\u2018Understanding the performance of decision-making behavior in building evacuation is essential for predicting pedestrian dynamics, designing appropriate facility safety management, optimizing emergency management strategies, and reducing the impact of disasters. While many pedestrian movement models have been developed based on the hypothesis of rational and strategic decision-making, only a limited number of works consider individual psychology and irrational behavior. To address this issue, we have successfully integrated the cumulative prospect theory (CPT) into modeling evacuations.\\u2019\\n\\n4. *Yan, Q., Feng, T., & Timmermans, H. (2020). Investigating private parking space owners\\u2019 propensity to engage in shared parking schemes under conditions of uncertainty using a hybrid random-parameter logit-cumulative prospect theoretic model. Transportation Research Part C: Emerging Technologies, 120, 102776.* \\n\\u2018Results show that socio-demographic characteristics, context variables, revenues and psychological concerns are all important factors in explaining parking space owners\\u2019 propensity to engage in platform-based shared parking schemes. [...] Understanding parking space owners\\u2019 propensity to share their parking spaces in relation to their psychological concerns and uncertain conditions is critical to improve shared parking policies. The results of this paper may help designers and planners in the delivery of shared parking services and promote the success and future growth of the shared parking industry.\\u2019\\n\\n5. *Dorahaki, S., Rashidinejad, M., Ardestani, S. F. F., Abdollahi, A., & Salehizadeh, M. R. (2022). A home energy management model considering energy storage and smart flexible appliances: A modified time-driven prospect theory approach. Journal of Energy Storage, 48, 104049.* \\u2018Smart home is a small but an important energy segment that has a significant potential to implement authentic energy policies, where human is a major decision-maker in the home energy management dilemma. Therefore, humans\\u2019 emotions and tendencies plays a vital role as the End-User's daily decisions. In this paper, we develop a behavioral home energy management model based on time-driven prospect theory incorporating energy storage devices, distributed energy resources, and smart flexible home appliances. [...] The results of the simulation studies show that the End-User's satisfaction in the proposed home energy management behavioral structure will be increased substantially compared with the conventional monetary home energy management models.\\u2019\\n\\n6. *Ladr\\u00f3n de Guevara Cort\\u00e9s, R., Tolosa, L. E., & Rojo, M. P. (2023). Prospect theory in the financial decision-making process: An empirical study of two Argentine universities. Journal of Economics, Finance and Administrative Science, 28(55), 116-133.* \\u2018This paper aims to provide empirical evidence for using the prospect theory (PT) basic assumptions in the Argentine context. Mainly, this study analysed the financial decision-making process in students of the economic-administrative academic area of two universities, one public and one private, in C\\u00f3rdoba. [...] The empirical results provided evidence on the effects of certainty, reflection and isolation in both universities, concluding that the study participants make financial decisions in situations of uncertainty based more on PT than on expected utility theory. This study contributes to the empirical evidence in a different Latin-American context, confirming that individuals make financial decisions based on the PT [...].\\u2019 \\n\\nWe thank the reviewer for their comment. We have now added a section in the appendix (see appendix C p. 17 in the revised manuscript, modifications all in blue) to further support the relevance of CPT compared to expected utility theory to model human behavior as suggested. We will also briefly update the introduction with the additional references and refer to the appendix for further details (due to space constraints). CPT provides a more realistic framework for modeling decision-making in uncertain environments than expected utility theory and this has been supported by extensive empirical studies. We believe that the machine learning community has a lot more to offer to address these questions given all the modern developments.\\n \\nThanks again for your thoughtful review, please let us know if our response addresses your concerns. If you have any further questions or comments, we will be happy to address them.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"metareview\": \"This paper is on policy optimization when the objective is Cumulative Prospect Theory (CPT) value of the cumulative reward random variable. It proves a policy gradient theorem for CPT-RL and derives a policy gradient algorithm for optimization in CPT-RL. Finally, it empirically shows the efficacy of the proposed algorithm and compares it with an existing method for optimizing CPT-RL that is based on simultaneous perturbation stochastic approximation (SPSA).\\n\\n(+) The policy gradient theorem, the corresponding algorithm, and experimenting with several problems are all positive aspects of this work. \\n\\n(-) Given that CPT-RL was previously studied by L.A. Prashanth et al. (2016), the reviewers expected that this paper provides a better comparison, in terms of both theoretical and empirical results, with the existing work. \\n(-) The reviewers seem to have concerns about the motivation behind the CPT formulation in RL. I believe the authors need to better justify the importance of this formulation in RL. \\n(-) There are several questions about the connections between the CPT-RL formulation and the proposed policy gradient algorithm, and the variety of risk-sensitive RL formulations and algorithms (many of them being policy gradient) that exist in the literature. I believe addressing this issue can significantly improve the overall quality of the paper.\", \"additional_comments_on_reviewer_discussion\": \"The authors tried to address the issues raised by the reviewers and answer their questions. Although some of the questions were addressed, the reviewers still had some concerns, especially on topics that were summarized in the meta-review. I believe the paper has potential and I strongly recommend that the authors take the reviewers' comments into account, improve the quality of their work, and submit it to an upcoming venue.\"}" ] }

[ "Claudio Battiloro", "Ege Karaismailoglu", "Mauricio Tec", "George Dasoulas", "Michelle Audirac", "Francesca Dominici" ]

[ "Topological Deep Learning", "Equivariance", "Equivariant Neural Networks", "Geometric Deep Learning", "Geospatial data", "Air Pollution Prediction", "Molecular Property Prediction" ]

{ "note_id": [ "ydZUvBNhTf", "wvtl1TSYnX", "v88eWFkxXX", "slfQAZYSSU", "qUvX6BQR7j", "qK9k2fFaMV", "ozw7lX54Tg", "mtjEK8tv2N", "kupIlx1BMm", "k9hioMimVG", "jbfCPcNnnw", "iSlmX5RRiI", "h2gY6GJqsB", "dzp2j9EsR9", "duods1ybT5", "dRXSoXw0Ss", "ZLq8szkOoY", "XF2wsVRtCr", "Vl3ZzQEzLh", "RVGWpQLAL4", "RIaBJhXHiQ", "Q7L1jhXEF3", "PN1JABgCvg", "Nu01XEbWNY", "MdO055yRe2", "MWQqN6Ngv6", "MCsUohADee", "L4EIzA9TGo", "I1RyStGqOi", "Hb8vH5h5e8", "HEvSnP3YRl", "H4T6upCkas", "GYhxf7amyB", "FEipEjLGAT", "Ell1hRhIFM", "E940uXyZNn", "DtKdKE4T0a", "DXZ0CKri20", "DLaU5ZhREV", "Bjdxkg4uV3", "BA7WRUczVB", "AAfzaUUUrL", "7fndy1wsfr", "6jqaNcobEZ", "4IngJZx19B", "37xlvOAMeB", "2wN616llIH", "1sxIbjE8ZE", "0G4uQJ49gD" ], "note_type": [ "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "meta_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "decision" ], "note_created": [ 1732808827670, 1731719425964, 1731721004469, 1732559264161, 1733239889702, 1732808852604, 1731716372444, 1732308208829, 1732593410084, 1732467650038, 1732322765503, 1732158318468, 1732808726664, 1730598807523, 1731721891874, 1731968263005, 1732227046230, 1732209336583, 1733193484304, 1732582787400, 1731923608308, 1732593951872, 1731720724298, 1732116772137, 1730372385653, 1730658626723, 1732489657798, 1732338612483, 1731715943071, 1731717942790, 1734836268754, 1732322467413, 1731956071126, 1732297045856, 1731721857733, 1730306459317, 1732808877762, 1732296355402, 1732296603300, 1732139112956, 1731719584688, 1732158124622, 1732487052491, 1731945309249, 1732446534696, 1731716608402, 1732227165108, 1731717670928, 1737523907045 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Reviewer_h23Z" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Reviewer_h23Z" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Reviewer_h23Z" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Reviewer_NQcw" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Reviewer_wwvW" ], [ "ICLR.cc/2025/Conference/Submission8415/Reviewer_h23Z" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Reviewer_wwvW" ], [ "ICLR.cc/2025/Conference/Submission8415/Reviewer_NQcw" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Reviewer_wwvW" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Area_Chair_z9ho" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Reviewer_J6CW" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Reviewer_J6CW" ], [ "ICLR.cc/2025/Conference/Submission8415/Reviewer_J6CW" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Submission8415/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ] ], "structured_content_str": [ "{\"title\": \"Follow-up after deadline extension\", \"comment\": \"In light of the extension of the discussion deadline, we would like to thank the reviewer again for the already positive assessment of our work and for their engagement, and ask if there is anything else we can do to further improve the manuscript and its score.\\n\\nWe also kindly invite the reviewer to parse the discussions we had with the other reviewers, as we believe that the overall engagement and the individual feedback have clarified or improved several aspects of the paper.\\n\\nThanks a lot!\"}", "{\"title\": \"Addressing Reviewer's comments\", \"comment\": \"We would like to thank the Reviewer for both pointing out the paper\\u2019s strengths, as well as their raised concerns regarding novelty, expressivity and the comparison with the CCMPNs. Next, we address these points.\\n\\n## Weaknesses.\\n**[W1] Novelty**\\n\\nOn the one hand, it is true that ETNNs are scalarization-based architectures, as EGNNs [4], on higher-order combinatorial domains, as EMPSNs [2]. As such, they inevitably resound with each other. \\n\\nHowever, we kindly disagree with the reviewer\\u2019s comment since ETNNs are a formal and more expressive generalization of both EMPSNs and EGNNs. For this reason, ETNNs unlock several features (e.g. arbitrary modeling of heterogeneous hierarchical higher-order interactions, tunable expressive power, and general learnable geometric invariants) that the constrained graph or simplicial structures of EGNNs and EMPSNs cannot accommodate (see **Appendix C** and **F**). As such, our framework can be used to design arbitrary E(n) equivariant TDL models. No other framework has the same power. \\n\\nMoreover, although we believe that the generality of our framework is fundamental, we also recognized the necessity for applications whose exhaustive modeling (in a combinatorial/topological sense) is possible only via combinatorial complexes, and we introduced MolecularCCs and GeospatialCCs to tackle this problem (no other complex nor graph can model molecular and multi-resolution irregular geospatial data as straightforwardly as CCs). As a consequence, we achieved SotA results or matched the existing SotA results among the Equivariant TDL models with a huge reduction in computational complexity. \\nAs a byproduct, as reviewer NQcw noticed too, our air pollution downscaling task represents a novel benchmark for the TDL community, which can be proven valuable to its needs, as noticed in the position paper [5]. \\n\\nFinally, we believe that the expressivity analysis is novel. In particular, our approach is, to the best of our knowledge, currently the most exhaustive one for scalarization-based equivariant TDL models, as it could be applied to analyze the (geometric WL) expressivity of any scalarization-based equivariant TDL model without the need for domain-specific coloring procedures. \\n\\n**[W2] Expressivity and Geometric Invariance**\\n\\nOur expressivity analysis in **Appendix F** assumes that the employed geometric invariant is the sum of pairwise distances. We believe that this is already the simplest and most computationally efficient geometric invariant. Moreover, as discussed in **Remark 5**, the assumptions of Proposition 1 are required for a clean theoretical treatment, but we empirically observed that different aggregations, as well as different message functions and geometric invariants, do not affect the expressiveness. This makes sense, as the assumptions in **Proposition 1** lead to the simplest possible architectural setting for an ETNN. About what geometric invariant functions graph models are unable to compute/learn, we are not sure we understood what the reviewer meant. The geometric invariants take as input all the geometric features of the nodes belonging to a certain cell. As such, a graph cannot compute/learn anything that takes as input more than two node geometric features (i.e., what is induced by the presence of an edge), as it cannot jointly access higher-order information (i.e., geometric features of more than two nodes).\\n\\n**[W3] ETNN vs EGNN expressivity**\\n\\nThe reviewer is correct that we need to refine the statement to make it clearer. In the revised manuscript, we clearly stated the condition of improved expressivity induced by **Proposition 2** in the new **Corollary 1**. In particular, an ETNN is strictly more powerful than an EGNN in distinguishing $k$-hop distinct graphs if the skeleton-preserving lifting is such that the number of ETNN layers required to have at least one cell in $X_{G_1}$/$X_{G_2}$ whose receptive field is the whole set of nodes in $G_1$/$G_2$ is smaller than the number of EGNN layers required to have at least one node in $G_1$/$G_2$ whose receptive field is the whole set of nodes in $G_1$/$G_2$. This statement is exhaustively empirically confirmed in **Table 3**. Regarding the plot, including it in the main body of the paper is challenging due to space constraints. However, we will make an effort to incorporate it in the final camera-ready version of the paper.\\n\\n*Continuing response in the thread (References are included in the last part of the thread)*\"}", "{\"comment\": \"**[W4] ETNN vs EMPSN expressivity**\\n\\nAs a consequence of what we wrote in our reply to **[W1](iii)**, a comparison is not required, because CCs generalize Cellular/Simplicial Complexes (SCs) and our approach based on the novel notion of *Geometric Augmented Hasse Graph* can thus be directly applied to SCs. As we carefully explained in the paper, an expressivity comparison (in the geometric WL sense) is fair only when the lift is skeleton-preserving, i.e. when the underlying graph connectivity is preserved; in [2], they do not lift the graphs (and they do not provide any expressivity analysis), but rather they replace the connectivity with a Vietori-Rips SC. Although one can indeed lift a graph into a simplicial complex preserving the graph connectivity by taking triangles or cliques as higher-order simplices, the expressivity would be characterized exactly as for **Proposition 2** (and as in [W3] to compare with GNNs).\\n\\n**[W5] Statistical Significance on QM9**\\n\\nAt the beginning of **Appendix J**, we exhaustively explain why we don\\u2019t run the model on multiple splits (each one of the baseline models uses different splits) and, in general, our meticulous care for reproducibility and fairness. Practical examples of this are also the fact that, despite its irreproducibility, we cared a lot about a fair comparison with EMPSN, or with an EGNN having our same parameter budget. Overall, for the QM9 benchmark, we followed the established experimentation setup of this specific dataset, following prior works, in which authors reported their best performance results and we were completely transparent about it (see **Appendix J** and **I**). This can be resonated by the large size of the molecular dataset, which makes multiple iterations of experiments challenging. In this sense, ETNN sets an unprecedented standard in the TDL community regarding exhaustiveness and transparency in the presentation of the results.\\n\\n**[W6] Statistical Significance on Air Pollution Downscaling**\\n\\nWe originally did not report variances since our experiments are averaged over 3 seeds, which makes variance calculations unreliable. We are now working on running with additional seeds and will report the results when the experiments are complete.\\n\\n## Questions.\\n**[Q1] K-hop distinct graphs**\\n\\nPlease see **[W2]**.\\n\\n**[Q2] ETNN vs EMPSN expressivity**\\n\\nPlease see **[W4]**.\\n\\n\\n**[Q3] Statement of Proposition 2**\\n\\nPlease see **[W3]**\\n\\n**[Q4] Statistical Significance**\\n\\nPlease see **[W5]-[W6]**.\\n\\n## References\\n\\n[1] V\\u00edctor Garcia Satorras, Emiel Hoogeboom, and Max Welling. \\\"E(n) Equivariant Graph Neural Networks.\\\" In *International Conference on Machine Learning*, pages 9323\\u20139332. PMLR, 2021.\\n\\n[2] Floor Eijkelboom, Rob Hesselink, and Erik J Bekkers. \\\"E(n) Equivariant Message Passing Simplicial Networks.\\\" In *International Conference on Machine Learning*, pages 9071\\u20139081. PMLR, 2023.\\n\\n[3] Papamarkou, Theodore, et al. \\\"Position: Topological Deep Learning is the New Frontier for Relational Learning.\\\" *Forty-first International Conference on Machine Learning*, 2024.\\n\\n[4] Liu, Cong, et al. \\\"Clifford group equivariant simplicial message passing networks.\\\" ICLR 2024.\"}", "{\"title\": \"Response to the Authors\", \"comment\": \"I would like to thank the authors for the detailed response.\\n\\n**[W1. Q6]** I apologize if my previous comment on simplicity of the model sounds impolite. I agree with the authors on the scope of the project, whose novelties focus on a flexible and a general pipeline for TDL. However, my main point still remains the same; it seems to me that your experiments cannot fully support the pipeline. However, I think it is worth a future research direction on defining geometric features for non-standard collection of neighborhoods. Given that your paper already brought many interesting insights and a novel benchmark, and my expectation may not be reasonable for an already informatively dense paper, I will consider to raise my score after the following concerns are resolved.\\n\\n> However, E(n) Equivariant versions of architectures, e.g., [1], leveraging powerful concepts coming from algebraic topology, e.g., homology groups, stratifications, filtrations, etc., arise when ETNNs and CCs are particularized to specific combinatorial topological spaces, e.g., cell complexes. \\n\\nCould you reference relevant works on E(n) architectures that bridge the gap between Topological Deep Learning (TDL) and Traditional Topological Machine Learning methods that rely on homology groups, stratifications, filtrations, ... I checked [1], and it didn't mention anything about TDL and the traditional topological machine learning methods. Please kindly correct me if I am wrong. I think there is one concurrent work [2] trying to bridge the literature gap, but it seems the goal of your paper and this paper is very different. Because it is a concurrent work, so I just put it here as a reference and it doesn't affect the the paper score.\\n\\n> Regarding the geospatial task, we use the coordinates of the points (0-cells) as geometric features.\\n\\nThis is my main concern regarding your response. Even though coordinates play as an important geometric features for molecular graphs, I don't think it is the same for the novel benchmark. In many real world applications, if you compute geometric invariant features based on coordinates, it doesn't make a lot of sense. For example, housing price in big cities will be different from that in rural area; however, under the E(n) architecture, you are computing geometric invariant features and eliminate such a useful information. Can you elaborate more how coordinates of the points are appropriate geometric features for the benchmark?\\n\\n**[Q2]** I completely agree with your argument; CC is the most general form of complex where one can define relations flexibly. I just meant that you should be more explicit on how you define Lift 4b to avoid confusion, as there are many complexes defined on paths.\\n\\n[1] Battiloro, Claudio, et al. \\\"Generalized simplicial attention neural networks.\\\" IEEE Transactions on Signal and Information Processing over Networks (2024).\\n\\n[2] Topological Blindspots: Understanding and Extending Topological Deep Learning Through the Lens of Expressivity. Under review at ICLR'25.\"}", "{\"title\": \"Final Global Response\", \"comment\": \"We thank the reviewers for their valuable feedback and for their constant engagement in the rebuttal period. **The quality of the reviews was particularly high and our manuscript surely benefited from this and from the exhaustive and constructive communication we had the possibility to constantly have with the reviewers during the discussion period**. Thank you.\\n\\nOverall, **all** the reviewers agreed that this work is a significant contribution, **deserves to be published** (NCQw, wwvW, J6CW) and it is a **critical** work for the TDL field (h23Z).\\n \\nAlthough novelty seems to *partly* remain somehow a concern, we would like to stress that *architectural novelty* is not representative of the novelty of our work as a whole. **Most importantly**, it is undoubtedly true that ETNN is a simple equivariant generalization of CCMPNs [1], being a scalarization-based equivariant model, similar to EMPSN [2] being a simple generalization of MPSNs [3], and EGNN [4] being a simple generalization of MPNs [5]. However, beyond the technical differences we already described, overall we showed that ETNNs are **more general and flexible**, **more expressive**, and **similar or better performing while being hugely more scalable** than SotA methods in the same class. These facts, together with the **benchmark**, **software**, and **outreaching** contribution of our work, should motivate its publication. This said, we already wrote in the future directions section that beyond scalarization-based architectures are of great interest. However, given that **a gap in the literature was still present for scalarization-based models** too, it made sense to us to **first comprehensively fill it** by working in the direction of ETNN.\\n\\n\\nFinally, elaborating more to strengthen our previous global response, ETNN has:\\n\\n(a) **Architectural** novelty: ETNN improves on the two most important components of a scalarization-based equivariant architecture, i.e., feature updates and geometric invariants. Regarding features update, graphs and SCs induce specific neighborhood functions (node adjacency for the former, boundary, coboundary, up/down adjacencies for the latter), thus if a CC comes with a non-standard collection of neighborhoods, neither EGNN nor EMPSN can handle it, both for non-geometric and geometric features updates. Regarding geometric invariants, EGNN/EMPSN uses tailored geometric invariants for graphs/SCs. Imagine applying EMPSN on a CC. Then, computing the volume of a non-simplex makes no sense. Similarly, pruning the pairwise distances based on the simplicial structure or using the angles of the planes induced by the dimension of the simplices makes no sense. On the other hand, ETNN can handle arbitrary neighborhoods, and its geometric invariants work on arbitrary CTSs (and formally generalize some of the invariants of EMPSN and EGNN). We also enriched the characterization of the geometric invariants with our ablation presented in the reply to reviewer J6CW, that will be extended and integrated in the camera-ready version of the paper.\\n\\n(b) **Experimental** novelty: the introduction of the new real-world geospatial benchmark and our novel approach for molecular modeling represent a significant effort to tackle some of the **most relevant open problems of the TDL field** (as described in [1] and recognized by multiple reviewers) and are a significant source of novelty (and an important resource for the community) as well.\\n\\n(c ) **Theoretical** novelty: ETNN is the only available framework for designing *arbitrary* E(n) equivariant scalarization-based TDL models, and the expressivity proof leveraging the novel notion of geometric augmented Hasse graph has not appeared before in the literature of equivariant TDL.\\n\\n**References**\\n\\n- [1] Hajij et al. (2022) *Topological deep learning: Going beyond graph data*. arXiv:2206.00606.\\n- [2] Eijkelboom et al. (2023). *E(n) Equivariant Message Passing Simplicial Networks*. ICML.\\n- [3] Bodnar et al. (2021). *Weisfeiler and lehman go topological: Message passing simplicial networks*. ICML.\\n- [4] Satorras et al. (2021). *E(n) equivariant graph neural networks*. ICML.\\n- [5] Gilmer et al. (2017). *Neural message passing for Quantum chemistry*. ICML.\"}", "{\"title\": \"Follow-up after deadline extension\", \"comment\": \"In light of the extension of the discussion deadline, we would like to thank the reviewer again for the already positive assessment of our work and for their engagement, and ask if there is anything else we can do to further improve the manuscript and its score.\\n\\nWe also kindly invite the reviewer to parse the discussions we had with the other reviewers, as we believe that the overall engagement and the individual feedback have clarified or improved several aspects of the paper.\\n\\nThanks a lot!\"}", "{\"comment\": \"**[W4] QM9 Testing Unfairness**\\n\\nWe kindly disagree with the reviewer, because we put most of our efforts into the fairness of the experiments. At the beginning of Appendix J (and in Appendix I, as the reviewer mentioned), we exhaustively explain why we don\\u2019t run the model on multiple splits (each one of the baseline models uses different splits) and, in general, our meticulous care for reproducibility and fairness. Practical examples of this are also the fact that, despite its partial irreproducibility, we cared a lot about a fair comparison with EMPSN, or with an EGNN having our same parameter budget (ETNN-graph-W). Moreover, the reported results from the other baselines are indeed their strongest performances, occurring in their own experimentation setup. This means that the reported results of the other models have already undergone a hyperparameter tuning process in the corresponding papers. Overall, we believe that ETNN sets an unprecedented standard in the TDL community regarding exhaustiveness and transparency in the presentation of the results.\\n\\n\\n**[W5] QM9 ETNN vs EGNN**\\n\\nWe report the improvement of ETNN over EGNN because our main focus is not showing overall SotA results (that, by the way, we mostly achieve in our class of interest Equiv+TDL) but rather the advantages of using CCs together with E(n) equi/invariance. We report the results of Equiformer exactly to be transparent and show what the current SotA for molecular graphs is (indeed it is already marked as SotA). As a side note, please notice that Equiformer and DimeNet are both tailored for molecules, while we show ETNN can be used in very different domains with very good performance. To make things clearer, we highlighted in bold the best-performing model in **Table 1** per each property. However, we believe that the improvement over EGNN should be kept (as the name of the row is not misleading).\\n\\n## Questions.\\n**[Q1] Functional Groups**\\n\\nWe extracted functional groups through RDKit\\u2019s substructure matching library. This can be seen in our shared repo, in the functional group lifting module. After the identification of the node subsets corresponding to functional groups, we filtered each group\\u2019s features through the following categorization:\\n| Functional Group | SMILES Pattern |\\n|------------------|-------------------------|\\n| carboxyl | C(=O)O |\\n| nitro | [N+](=O)[O-] |\\n| ketone | [CX3](=O)[C] |\\n| ester | [CX3](=O)[OX2H0][#6] |\\n| ether | [OD2]([#6])[#6] |\\n| amide | [NX3][CX3](=[OX1])[#6] |\\n| benzene | c1ccccc1 |\\n| aniline | Nc1ccccc1 |\\n| phenol | Oc1ccccc1 |\\n| carbamate | [NX3][CX3](=[OX1])[OX2H0] |\\n\\nFurther details can be found in the functional group lifting module in the shared repo. \\n\\n**[Q2] Contribution of Higher-Order Cells**\\n\\nWe kindly disagree with the reviewer. In **Tables 10-11**, on 9 out of 11 properties, the variants of ETNN matching the best results use higher-order features. Moreover, please also notice that even when only bonds are used, we are still in a higher-order scenario, because ETNN leverages multiple adjacencies among edges that could not be leveraged with a GNN.\\n\\n**[Q2] Novelty: TDL vs GDL**\\n\\nOur comments in **[W1]-[W2]-[W3]** should jointly provide an exhaustive answer to this question. Of course, we are more than happy to elaborate more if needed during the discussion phase.\\n\\n**[Q3] Virtual Cell and Heterogeneity**\\n\\nThe virtual cell increases performance in general, and this is expected, as the virtual node [7] had a similar effect on graph-based models. Defining a family of fully connected EGNNs over the different orders is computationally similar (but not equivalent, as in the case proposed by the reviewer is not clear how to handle geometric features) to every configuration in **Tables 10-11** that has only \\u201cmax\\u201d in the Adjacencies column and \\u201c0\\u201d in the incidence column. As the reviewer can notice, only on 2 properties out of 11 these architectures perform better. About the heterogeneity, the reviewer is right, we cannot be certain that the two (or more) types of propagation add to performance, but this is just an additional amenable property of ETNN, and the user has full control over the architecture configuration, thus on the choice of including multiple types of relations or not.\\n\\n\\n*Continuing response in the thread (References are included in the last part of the thread)*\"}", "{\"title\": \"Response to the Authors regarding Weaknesses and Questions\", \"comment\": \"First of all, I apologize for the late response. Below are my detailed responses regarding your rebuttal.\\n\\n**[W1, Q6] Novelty and Higher-order Geometric Features**\\n\\nI kindly disagree with the authors on the novelty of the paper. I acknowledge that the work is critical to the field as there is limited existing work attempts to formally generalize the Topological Deep Learning (TDL) framework to geometric spaces; however, the generalization is trivial as similar to how EGNNs generalize to GNNs. As higher-order relations are just a generalization of adjacency matrices in graphs, it is trivial to insert any geometric invariant features during message passing with higher-order relations to make the TDL framework equivariant. \\n\\nWhat I am interested more is how the authors construct geometric features for higher-order cells as stated in Q6. As the authors clarify, it is possible that higher-order cells are not physical entities so they may not have higher-order geometric features. Therefore, the geometric information is only inserted for cells that contains this information, and we rely on message passing to pass this geometric information to subsequent higher-order cells. In other words, it seems that there is no geometric features for higher-order cells. While the proposed framework allows TDL to be equivariant for every order, the experiment for QM9 only leverages the geometric invariant features for 0-cells only. It makes the experiments may not support the framework fully.\\n\\nAlso, I still have questions regarding the Geospatial CCs. Which geometric information is provided for different cells? I checked the Appendix and I only saw non-geometric features; I may be wrong, so please kindly point out which features are geometric. Or do you mean that distance to nearest is geometric invariant feature in this benchmark?\\n\\n**[Q2] Path Complexes**\\nPath Complex in [2] strictly follows the boundary operation defined in [1], where a k-path contains (k-1)-paths that exist in the graph. With that being said, it is possible that given a triangle, there can be 3 distinct 2-paths, but each 2-path shares the same set of edges on its boundary, even when an edge may not be a subset of the 2-path. I think your ablation studies treat 3-paths as 2-cells and consider edges belong to the 3-paths as their boundaries. This notion is different from what is proposed in [2]. Even though this part may not directly affect your work, I strongly encourage the authors to explain more on \\\"Lift 4b\\\" instead of just stating this is equivalent to the Path complex, as it may create confusion for future readers.\\n\\n**[Q5] Notation**\\nThis is just my personal preference to improve readability, so it doesn't affect the score. I meant Eq. 6 and Eq. 7 on page 5, where $\\\\mathcal{N} \\\\in \\\\mathcal{CN}$ and $x \\\\in \\\\mathcal{X}$ both use $\\\\in$ to describe two different notions. \\n\\n**[W2] Scalability**\\nWhat I meant is that if there are any large graphs with geometric features (maybe millions of nodes for example) to see how the framework performs with respect to GNNs. However, I acknowledge this part may be out of this paper's scope, and as the authors demonstrated the framework superiority with respect to EMPSN, this concern is addressed.\\n\\n[1] Grigor\\u2019yan, A.A., Lin, Y., Muranov, Y.V., et al. \\\"Path Complexes and their Homologies.\\\" Journal of Mathematical Sciences, 248, 564\\u2013599 (2020).\\n\\n[2] Truong, Q., & Chin, P. \\\"Weisfeiler and Lehman Go Paths: Learning Topological Features via Path Complexes.\\\" Proceedings of the AAAI Conference on Artificial Intelligence, 38(14) (2024).\"}", "{\"comment\": \"**[W1. Q6] TDL and TML**\\nWe thank the reviewer for bringing the reference to our attention. The goal of our previous reply was not to reference methods that bridge TDL with classical TML though, as reviewer\\u2019s comment **was not** about this. We wanted to stress that higher-order interactions are not just encoded as generalization of graph adjacencies, but can rather be leveraged using algebraic topology tools as well. In particular, the work in [1] we refer to as an example, (i) explicitly and formally uses (weighted) Hodge and Dirac theories, (ii) explicitly uses their corresponding spectral theories [2,3,4] (that can be leveraged in combination with filtrations [5]), and (iii) implicitly uses homology theory because the kernels of the Hodge Laplacian (and, as a consequence, the kernel of the Dirac operator) contain homology information as their dimensions are the Betti numbers of the complex. To link to our previous comment about this, E(n) Equivariant versions of architectures like [1] can be readily obtained from our ETNN framework. \\n\\nIn terms of macro-areas, works like [1] help to bridge TDL with algebraic topology using more of a Topological Signal Processing (TSP) [6] perspective, while we agree that works like [7] help to bridge TDL with algebraic topology using more of a TML perspective. Furthermore, works like [8] help to bridge TDL with equivariant DL using more of a TML perspective, while a general work like ETNN helps to bridge TDL with equivariant DL implicitly using more of a TSP (and traditional relational DL) perspective. Overall, the three mentioned disciplines (TSP, TDL, TML) are surely overlapping, fall under the same umbrella, and many links among them are yet to be studied (e.g., further insight on if and how results from [7] applies to proper convolutional models like [1] that explicitly take some topological invariants into account). However, this surely is an exciting motivation for being in this field at this time. Finally, we will also cite [7] in our paper as it is surely relevant to our work.\\n\\n**[W1. Q6] Geospatial task coordinates as geometric features**\\nWe would like to clarify some potential misunderstandings. Most coordinate systems are, by construction, artifacts built with notions of symmetry and equivariance, as they are often derived from projections. Specifically, we use the Mercator projection, which uses meters as the unit scale. This projection has the key property that Euclidean distance in coordinate space corresponds to geodesic distance. Thus, E(n)-equivariance is implicitly essential to the purpose of the Mercator projection. The exact coordinate values are irrelevant; what is important for our task is the projection's usefulness in expressing geodesic distances. Consequently, it is never desirable for the model to behave erratically with respect to a Euclidean transformation of the Mercator coordinates.\\n\\nIn light of the above remark, it is not true that geometric invariant features eliminate useful information. It is known that even just pairwise distances are sufficient to recover angles in Euclidean space (cf., [9; Appendix E]), and therefore a coordinate system modulo the Euclidean group. The same logic applies to our study: by using geometric invariants in geospatial coordinate systems, we allow the model to use the coordinates modulo Euclidean transformations, which is the original intention of the Mercator projection. This implies that the model can still capture location-specific random effects in node-level tasks using only invariants, which is the reviewer's main concern.\\n\\nWe recognize that this is a confusing point and central to the paper. Therefore, we have added a summarized version of our response to the revised text.\\n\\n[1] Battiloro, et al., 2024. \\\"Generalized simplicial attention neural networks.\\\" IEEE Transactions on Signal and Information Processing over Networks.\\n\\n[2] Calmon et al. 2023. \\\"Dirac signal processing of higher-order topological signals.\\\" New Journal of Physics 25.9.\\n\\n[3] Hansen & Ghrist, 2019. Toward a spectral theory of cellular sheaves. Journal of Applied and Computational Topology.\\n\\n[4] Yang, et al. 2022. \\\"Simplicial convolutional filters.\\\" IEEE Transactions on Signal Processing.\\n\\n[5] Grande & Schaub, 2024. \\\"Disentangling the Spectral Properties of the Hodge Laplacian: not all small Eigenvalues are Equal.\\\" ICASSP.\\n\\n[6] Barbarossa & Sardellitti 2020. \\\"Topological signal processing over simplicial complexes.\\\" IEEE Transactions on Signal Processing.\\n\\n[7] Eitan et al., 2025. \\\"Topological blind spots: Understanding and extending topological deep learning through the lens of expressivity.\\\", under review at ICLR'25.\\n\\n[8] Verma et al., 2024. \\\"Topological Neural Networks go Persistent, Equivariant, and Continuous.\\\" ICML.\\n\\n[9] Satorras et al., 2021. \\\"E(n) Equivariant Graph Neural Networks.\\\" ICML.\"}", "{\"title\": \"Did our response address the Reviewer's questions?\", \"comment\": \"We\\u2019d like to ask the Reviewer whether our previous response addresses their latest questions. We\\u2019re more than happy to elaborate more on any raised concern from the Reviewer\\u2019s side.\\nOnce again, we'd like to thank them for the engaging discussion so far!\"}", "{\"comment\": \"**[W1, Q6] Cont.** Regarding the geometric features for higher-order cells, we apologize if our previous reply was not clear enough. It is not true that \\u201cthe geometric information is only inserted for cells that contain this information, and we rely on message passing to pass this geometric information to subsequent higher-order cells\\u201d. This is because geometric information can be leveraged either in the form of geometric features and/or in the form of geometric invariants. It is true that the **geometric features** are attacched only to cells that come with them, but the **geometric invariants** are used to update the subsequent (possibly) higher-order cells without any reliance on message passing, i.e. the geometric information contained in the geometric invariants is directly leveraged from the subsequent cells even if they do not come with geometric features. In this sense, what message-passing enables is the exchange of the (already attached) \\u201cgeometric information\\u201d of all the cells, realized as geometric features and/or geometric invariants. This said, we totally agree that tailoring ETNN for applications in which higher-order cells have a (possibly) physical meaning with attached geometric features is intriguing, but it is an interesting future direction. Here, geometric invariants already represent a source of geometric higher-order information.\\n\\nRegarding the geospatial task, we use the coordinates of the points (0-cells) as geometric features. We are really sorry we forgot to write it in Appendix I, it was not on purpose. We will add this specification in the next revision of the manuscript we are planning to submit in the next few days.\\n\\n\\n**[Q2] Path complexes** \\nThe reviewer is right about Lift 4b. It is not equivalent to a path complex, we will delete the sentence in the next revision to avoid adding further complexity. However, the reviewer's previous comment was about the possibility of rewriting a path complex as a CC. Regarding this, our answer remains valid: a CC where the cells (i.e., the elementary paths) are correctly specified and the boundary operation is used to define the neighborhood functions is equivalent to a path complex as in [2].\\n\\n**[Q5] Notation** We believe they both correctly indicate containment. A cell $x$ is an element of the set of cells $\\\\mathcal{X}$, a neighborhood function $\\\\mathcal{N}$ is an element of the set of the neighborhood functions $\\\\mathcal{CN}$ (i.e. the collection of neighborhoods). However, in the revision, we will make it clearer that $\\\\mathcal{CN}$ is a set when we first introduce it before Equation 3.\\n\\n## References.\\n\\n[1] Grigor\\u2019yan, A.A., Lin, Y., Muranov, Y.V., et al. \\\"Path Complexes and their Homologies.\\\" Journal of Mathematical Sciences, 248, 564\\u2013599 (2020).\\n\\n[2] Truong, Q., & Chin, P. \\\"Weisfeiler and Lehman Go Paths: Learning Topological Features via Path Complexes.\\\" Proceedings of the AAAI Conference on Artificial Intelligence, 38(14) (2024).\\n\\n[3] Papamarkou, Theodore, et al. \\\"Position: Topological Deep Learning is the New Frontier for Relational Learning.\\\" Forty-first International Conference on Machine Learning. 2024.\\n\\n[4] Battiloro, Claudio, et al. \\\"Generalized simplicial attention neural networks.\\\" IEEE Transactions on Signal and Information Processing over Networks (2024).\\n\\n[5] Bai, Song, Feihu Zhang, and Philip HS Torr. \\\"Hypergraph convolution and hypergraph attention.\\\" Pattern Recognition 110 (2021): 107637.\"}", "{\"comment\": \"\\u200b\\u200b**Standard errors for the geospatial task and additional seeds**\\n\\nWe have conducted multiple additional runs for the geospatial task based on your suggestion. We ran 10x more seeds (30 in total). Remarkedly, the results are very similar to those reported in the paper despite the high variance reflected by the standard error, \\n\\n*Updated tables*\\n\\n| Baseline | Av. Explained Variance ($R^2$) | Std. Err ($R^2$) | **MSE** |\\n| -------- | ------------------------------ | ---------------- | ------- |\\n| ETNN | 9.34% | 2.05% | 0.935 |\\n| GNN | 2.44% | 0.99% | 0.987 |\\n| MLP | 2.35% | 1.61% | 1.022 |\\n| EGNN | 1.43% | 1.2% | 1.041 |\\n| Linear | 0.51% | 0.95% | 1.106 |\\n**Table 2a**. Baseline model comparison\\n\\n| Baseline | Diff. Explained Variance ($R^2$) | Std. Err ($R^2$) | **MSE** |\\n| ----------------------------- | -------------------------------- | ---------------- | ------- |\\n| no geometric features (CCMPN) | -1.08% | 2.05% | 0.946 |\\n| no position update (invariant ETNN) | -1.37% | 2.52% | 0.956 |\\n| no virtual node | -1.80% | 2.32% | 0.957 |\\n| | | | |\\n**Table 2b**. Ablation study\"}", "{\"title\": \"Follow-up after deadline extension\", \"comment\": \"In light of the extension of the discussion deadline, we would like to thank the reviewer again for the already positive assessment of our work and for their engagement, and ask if there is anything else we can do to further improve the manuscript and its score.\\n\\nWe also kindly invite the reviewer to parse the discussions we had with the other reviewers, as we believe that the overall engagement and the individual feedback have clarified or improved several aspects of the paper.\\n\\nThanks a lot!\"}", "{\"summary\": \"The authors propose an equivariant Topological Deep Learning framework that deals with geometric node features. The frameworks can be generalized to many topological domains including simplicial, cell, combinatorial, and path complexes. The authors provide theoretical analysis regarding his design choice as well as expressiveness of the proposed method. Lastly, the authors support his arguments via real-world datasets QM9 and his proposed benchmark - air pollution downscaling benchmark.\", \"soundness\": \"2\", \"presentation\": \"4\", \"contribution\": \"3\", \"strengths\": \"The authors add an important piece of work for the Topological Deep Learning (TDL) community as there is not much literature on Equivariant TDL. The work is well-formulated with clear motivations. The theoretical contributions are well-written. The paper is also self-contained and easy to follow, given the substantial explanations from related prior literature. The ablation studies are well-conducted via many synthetic graphs and additional information (hyperparameters, data statistics, etc.) are provided. The novel benchmark based on geospatial information is novel.\", \"weaknesses\": \"Novelty is the key disadvantage of the paper. It seems that the work just extends prior works on graphs to TDL. Even though the theoretical insights are important, yet they are mostly an extension from graphs. Another important weakness is scalability and practicability of the problem. There are only two real-world datasets evaluated, and in both cases, graph sizes are small. Furthermore, the performance isn\\u2019t convincing given there are only minor improvements over the graph counterparts. The performance could be due to extra parameters for higher-order filters. Perhaps an experiment on model comparison with constraints on parameter budgets and an ablation study on higher-order features masked out are needed to prove your arguments. Lastly, even when the framework makes sense, it is unclear how we can obtain geometric features for higher-order cells. Please refer to question 6 for my concern.\", \"questions\": \"1. If I understand correctly, your argument on \\u201cheterogeneous interactions\\u201d focuses on different relationships between cells. Meanwhile, this property was actually mentioned in prior literature ([3] to name a few). I don\\u2019t think it is fair to claim that ETNNs are set up for this characteristic, but more like TDL in general already possesses this property.\\n2. The paper mentioned that Combinatorial Complex subsumes Path Complex. I believe there are two distinct lines of work regarding complexes arising from paths. One is path-based Combinatorial Complex [3], and another work is a simplified path complex [2] based on path complex [1]. [2] can\\u2019t be derived directly from [3] because [2] reserves the sequential information of paths. From my understanding, your work focuses more on path-based combinatorial complexes.\\n3. A relevant work [4] is not discussed in the paper.\\n4. It seems to me that $\\\\mathcal{N}_{A, \\\\text{max}}$ isn\\u2019t supported by any experiments. It would be better if the authors provide a set of experiments to support this design choice.\\n5. A minor subjective feedback on writing. I think it is better to have different notations for \\u201ccontainment\\u201d and \\u201cis a subset of\\u201d operation for clarity (Equation 6 and 7).\\n6. Regarding section I.1.2, it seems that 2-cell features do not encode any geometric features (velocity for example) but only invariant features. I think it is a missing piece to convince the audience that your framework can work with geometric features. Also, suppose that even when we have geometric features at node levels, it is unclear how to lift these features into higher-order spaces.\\n\\n[1] Grigor\\u2019yan, A.A., Lin, Y., Muranov, Y.V. et al. Path Complexes and their Homologies. J Math Sci 248, 564\\u2013599 (2020).\\n\\n[2] Truong, Q., & Chin, P. (2024). Weisfeiler and Lehman Go Paths: Learning Topological Features via Path Complexes. Proceedings of the AAAI Conference on Artificial Intelligence, 38(14).\\n\\n[3] Hajij, M. et al. Topological Deep Learning: Going Beyond Graph Data (2023).\\n\\n[4] Li, L. et al. Path Complex Neural Network for Molecular Property Prediction. ICML 2024 Workshop GRaM (2024).\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"## References.\\n[1] Bronstein, Michael M., et al. \\\"Geometric deep learning: Grids, groups, graphs, geodesics, and gauges.\\\" arXiv preprint arXiv:2104.13478 (2021).\\n\\n[2] Bodnar, Christian, \\u201cTopological Deep Learning: Graphs, Complexes, Sheaves\\u201d, PhD Thesis, https://www.repository.cam.ac.uk/items/06b0b8e5-57d1-4120-8fad-643ce4d40eda (2022).\\n\\n[3] Battiloro, Claudio, \\u201cSignal Processing and Learning over Topological Spaces\\u201d, PhD Thesis, https://theses.eurasip.org/theses/974/signal-processing-and-learning-over-topological/ (2024).\\n\\n[4] Papamarkou, Theodore, et al. \\\"Position: Topological Deep Learning is the New Frontier for Relational Learning.\\\" Forty-first International Conference on Machine Learning. 2024.\"}", "{\"comment\": \"We would like to thank the reviewer for the swift response, allowing us to have an engaging discussion. Next, we address the four points they raised:\\n\\n**[W1]** Architectural novelty: ETNNs are a direct generalization of EMPSNs.\\n\\nIn mere terms of forward pass (which, although an important component, represents just one aspect of our more comprehensive work), EMPSN forward cannot be used to process signals on a CC. The main reasons concern the two most critical components of a scalarization-based equivariant architecture, i.e., (1) Feature updates and (2) Geometric invariants.\\n\\nRegarding (1), SCs induce specific neighborhood functions (Boundary, Coboundary, Up/Down Adjacencies), thus if a CC comes with a non-standard collection of neighborhoods, EMPSN forward cannot handle it, both for non-geometric and geometric features updates.\\n\\nRegarding (2), EMPSN forward uses tailored geometric invariants for SCs. Computing the volume of a non-simplex makes no sense. Similarly, pruning the pairwise distances based on the simplicial structure or using the angles of the planes induced by the dimension of the simplices makes no sense in the CC case.\\n\\nPlease notice that an ETNN forward can be used to process signals on an SC, as all the geometric invariants tailored for SCs and the neighborhoods induced by an SC can be directly used in our framework. \\n\\n\\n**[W2]** Expressivity on CTSs.\\n\\nThanks for clarifying the question. In fact, we have already discussed an example where the lifting of $k$-chain graphs would result in two undistinguishable (geometric) CCs for ETNN. In particular, in **Remark 6**, we explain that in Lift 3a, using the up adjacency from (2) without including the edges as cells, would prevent solving the task. This is because it would result in a disconnected geometric augmented Hasse graph, since $p_1$/$p_2$ and $o_1$/$o_2$ would not have been linked. We are more than happy to enhance this part more in the revised manuscript, highlighting that failing to solve the task is equivalent to proving that there exists a collection of neighborhoods and a pair of geometric CCs that are undistinguishable by ETNN.\\n\\n**[W3]** Expressivity comparison to equivariant TDL methods. \\n\\nYes, lifting to a CC and using ETNN is indeed better than lifting to a SC and using EMPSN. We already showed this indirectly in the paper. In particular, on both the numerical expressivity tasks that we show, the $k$-chain graphs (**Table 3**) and the counterexamples (**Table 4**), EMPSN with any skeleton preserving lifting would be as expressive as an EGNN (and, thus, way less expressive than an ETNN), as there are no triangles nor cliques. We will add a remark in the paper to make this clearer.\\n\\n**[W4]** Statistical significance of experimental results.\\n\\nWe thank the reviewer for their understanding regarding QM9, and we hope that our particular focus on transparency and honesty in the description of our experimental setup (**Appendices I** and **J**) can compensate for this partially problematic habit of the community. \\n\\nWe are currently running the geospatial experiment on a significantly higher number of seeds and we will post the results as soon as the runs are done.\"}", "{\"comment\": \"We would like to thank the reviewer for the prompt feedback, their reconsideration of the score, and their opinion about the fact that our work deserves to be published. Overall, their comments have certainly improved the quality of our manuscripts. Here, we address their latest questions and comments. We are happy to work further to improve the already positive opinion the reviewer has of our work.\\n\\n**[W1]** We are happy to discuss this further. About GDL and TDL, the reviewer is totally right on everything, and the immediate main difference remains *the focus on higher-order interactions*. \\n\\nOur point was more subtle and related to the approach that the two disciplines *historically* followed during their development, keeping again in mind that the intersection is not empty and the object of study (non-Euclidean spaces) is loosely the same. \\n\\nOn a more technical side, GDL uses mainly geometric arguments (norms, groups, metrics, and deformations are at the core of the framework) while TDL uses mainly topological arguments (again, our ETNN framework kind of mixes them). \\n\\nGiven the scope and goals of this work and for the sake of clarity, we did not overcomplicate the definition of the signals. However, in general, in TDL (and Topological Signal Processing), a more exhaustive definition of signals is the vector representation (due to natural isomorphism) of the cochain spaces with coefficients in $\\\\mathbb{R}$ associated with the simplicial/cell complex [1][2] (a more general but less theoretically rich notion of cochain space has been given for combinatorial complexes too [3], however, we believe the reviewer will resonate more with the standard notion). In this case, the cochain spaces are Hilbert spaces as well, as the reviewer correctly pointed out. However, to retain the whole machinery and results of the algebraic topology machinery (and, more in general, to retain the relational structure induced by the space), a cochain space is solely required to have an abelian group structure. This enables the usage of more exotic data structures, e.g. lattices [4]. Unfortunately, it is impossible to exhaustively treat this topic in this rebuttal, and it is more than sufficient to highlight the crucial importance of higher-order interactions. But, to wrap up, an intuitive but *absolutely not exhaustive* way of looking at this comparison, is mathematically looking at GDL as more of a top-down approach while TDL is a bottom-up. This *loose* difference is the main drive for modeling higher-order interactions (that in GDL-related spaces, e.g. manifolds, grids, or graphs, are harder to derive). \\n\\nAbout the geospatial benchmark, we only partially agree on the specificity for the TDL community. We shall note that this is the first, to our knowledge, TDL work that utilizes a geospatial task as a benchmark, going beyond the more standard molecular tasks, which are widespread in this line of research, and, also, moving away from how classic biostatistics methods treat the downscaling problem. This is exactly what we find exciting in this application domain, and we built this task to show it. So, on one hand, it is true that the specific task is tailored to be a benchmark for TDL models (and this is good for the community as a whole) but on the other hand, we believe it will be useful for *general-purpose ML applications*. To provide the reviewer with a sense of it, we refer to the geospatial foundation model in [4], which appeared online only some days ago, that uses similar (but muchsimpler) arguments to ours for irregular multi-resolution geospatial data employing a heterogeneous graph jargon rather than a TDL one.\\n\\n**[W2]** We believe that a way to look at it is the following: ETNNs are very general architectures; to prove WL expressivity what really matters is the relational structure exploited by the network rather than its specific architectural instance; it then makes sense to use the simplest instance of ETNN (the one respecting the assumptions 1-3 in Proposition 1); this specific architecture results in a non-standard EGNN on the geometric augmented Hasse graph and it is sufficient to elegantly prove the WL expressiveness of ETNN.\\n\\nTherefore, the non-standard EGNN running on the geometric augmented Hasse graph is just a specific instance of ETNN. Indeed, this ETNN resulting from Assumptions 1-3, does not take into account either the different ranks (because of Assumptions 1-2), or the different geometric invariants (because of Assumption 3), or the different neighborhoods (because, from (27), the graph is built from the union over the neighborhoods of the complex). This collapses many relations induced by the topological domain and applies the same set of weights to the connections that survive the collapse.\\n\\n*Continuing the response in the next message. References are included, also, in the next message.*\"}", "{\"comment\": \"I thank the authors for their detailled answers!\\n\\n[W1]\\n\\nI must admit I\\u2019m a little confused about this point. Is the Hilbert space structure due to the feature vectors above each node on the graph? If that\\u2019s the case, even in TDL we still have (Euclidean) feature vectors sitting above each node in the augmented hasse diagram etc. My understanding of the group-theoretic arguments was that they relate to permutation equivariance of basis vectors in the feature space, which equally applies to TDL given that neighbourhoods need Euclidean \\ufefffeature vectors. I take your point about the emphasis on the higher order interactions, and accept that this is indeed a different emphasis.\\n(ii) I totally agree that TDL needs more empirical justification, and that this paper is a good foundation for it. I do worry a little (as one of the other reveiwers) that something like the geospatial benchmark \\ufeffis perhaps a little over-designed specifically TDL methods to be of interest outside of the TDL community --- apart from that I\\u2019m on board.\\n\\n[W2]\\nI take your point. I think our only disgreement is the extent to which a \\u201cnon-standard EGNN over a Geometric Augmented Hasse Graph\\u201d is a significantly novel deviation from a standard EGNN. This still feels (to me at least) well within the vicinity of message passing over a graph \\u2014 albeit with non-uniform graded feature spaces. Again though, I understand your point that the novelty is that the actual Hasse graph emphasises higher-order combinatorial structure. \\n\\n[W3] \\nFair enough.\\n\\n[W4]\\nOK good explanation \\u2014 thanks for clarifying, in particular about the hyperparameters. \\n\\n[W5]\\nI appreciate the highlighting of the benchmarks! \\n\\ufeff\\n\\ufeffRelating to my previous point, I think that beating EGNNs will mainly be of interest to researchers within the TDL community rather than researchers more broadly. With that \\ufeffin mind, the bottom line still looks strange to me. It\\u2019s not a massive deal so I\\u2019m happy for it to be left in if that\\u2019s what the authors want. \\n\\n\\n[Q1]\\nThanks for pointing out where to find this.\\n\\n[Q2]\\nGood explanation!\\n\\n[Q3]\\nAh I see, I think that actually clarifies a misunderstanding I had. So the bonds still inherit the adacencies of neighbouring higher order cells even when we are not using the feature vectors of the higher order cells in the learning? Have I understood that correctly?\\n\\n[Q4]\\nUpdated definition looks good to me!\\n\\n[Q5]\\nThanks for adding references. \\n\\n[Q6]\\nThanks for rerunning the experiments! Would it be possible to include the variances in the paper? I think it\\u2019s useful for the reader.\\n\\n\\nBased on the responses I'll be upgrading my score to a 6. Despite being a very well written and tested paper, the main issue I have is the extent of contribution and novelty, as I outline in my answers above. I think the paper deserves to be published in any case.\"}", "{\"title\": \"Extension of the ablation study on the geometric invariants\", \"comment\": \"As we mentioned in our previous message about the ablation study over the impact of the geometric invariant choice, we follow up with an extended study over multiple targets, and multiple experiment configurations (as these are reported in **Appendix J**. Below, we present the results of this study. The numbers of the experiment configurations correspond to the Combinatorial complex configuration, as presented in **Table 11**.\\n\\n**Remark:** We note that due to time constraints we run the next configurations for a total of 350 epochs (instead of 1000 epochs of our original results in Table 11).\", \"experiment_configuration\": \"4\\n| Invariant Choice | $\\\\alpha$| $\\\\Delta\\\\epsilon$ | $\\\\epsilon_\\\\text{HOMO}$ | $\\\\epsilon_\\\\text{LUMO}$ | $mu$ |\\n|:--------------------|-----------:|------------:|------------:|------------:|-----------:|\\n| Centroid | 0.123991 | 0.0591344 | 0.0320109 | 0.027891 | 0.320119 |\\n| No invariants | 0.180012 | 0.0712032 | 0.0371132 | 0.0314019 | 0.037239 |\\n| Centroid + Haudsorff| 0.105327 | 0.0587845 | 0.0355854 | 0.0264562 | 0.031021 |\\n| Hausdorff | 0.112335 | 0.0595528 | 0.0347739 | 0.0292914 | 0.034909 |\\n\\n\\nFollowing our preliminary results (see previous response), we validate again that:\\n1. Not using Centroids distance nor Hausdorff distance exhibits the **worst performance**.\\n2. In the majority of the runs, Hausdorff distance seems to perform **slightly better** than the Centroids distance. \\n3. Using both invariants usually yields the **best performance**. \\n\\nWe\\u2019d like to thank once again the reviewer for the constructive feedback, and we plan to include such analysis in the camera-ready version.\"}", "{\"title\": \"Ablation study on geometric invariants\", \"comment\": \"**[W2]** Although, as the reviewer agreed, it is unfeasible but interesting to add further theoretical results about this matter to this already extensive work, we, as we said, agree and believe that the theoretical and empirical impact of geometric invariants matters. For this reason, we conducted a small ablation study to analyze the impact of different geometric invariants on model performance. In this experiment, we used the property $\\\\alpha$ of QM9 as the target variable. For the Combinatorial Complex configuration, we employed the first configuration outlined in Table 11 of Appendix J (which consists of atom, bond, and virtual cell with incidences and max adjacency). Regarding the choice of invariants, we utilized the following:\\n\\n- Centroids Distance: This metric represents the Euclidean distance between the centroids of the two participating cells. It belongs to the second family of measures outlined in the manuscript, specifically the \\u201cDistances of permutation-invariant functions\\u201d.\\n\\n- Hausdorff Distance: AS described in the paper. This measures the largest minimum distance between nodes in the two cells.\\n\\nWe indicate whether an invariant is active or inactive by setting it to 1 or 0, respectively. \\n\\n\\n| Run | Centroids Distance | Hausdorff Distance | Test MAE |\\n|------:|---------------------:|---------------------:|-----------:|\\n| 1 | 1 | 0 | 0.0834196 |\\n| 2 | 0 | 1 | 0.0747732 |\\n| 3 | 1 | 1 | 0.0733149 |\\n| 4 | 0 | 0 | 0.416896 |\\n\\nBased on the results in the table, we make the following preliminary observations:\\n\\n- Using both geometric invariants yields the best performance (Run 3).\\n\\n- Excluding both Centroids Distance and Hausdorff Distance (no geometric information, CCMPN-like) significantly harms performance (Run 4).\\n\\n- Hausdorff Distance appears to be a slightly better choice than Centroids Distance as a geometric invariant (Run 2 vs. Run 1).\\n\\nThese results show an interesting interplay among the invariants, giving hints on the impact of each one. As today is the last day to submit a revision of the paper, we plan and commit to extending this ablation study to include additional configurations and target variables for the camera-ready version of the paper.\"}", "{\"title\": \"Response:\", \"comment\": \"I appreciate the authors' detailed response, which I will now address:\\n\\n[W1] I recognize that ETNNs can process combinatorial complexes, whereas EMPSNs are limited to simplicial complexes, making ETNNs more versatile. I also agree that the inclusion of two novel combinatorial complex benchmarks is a valuable contribution, highlighting the strengths of ETNNs. However, I believe the main advantage of ETNNs over EMPSNs is primarily technical, as both rely on scalarization combined with neighborhood-dependent updates. To my understanding, there is nothing inherently restrictive in EMPSNs update rule that prevents them from handling combinatorial complexes, even though they have been framed as a simplicial complex-focused architecture. Despite this, I agree that the ETNN update rule is more generic and flexible than EMPSNs. \\n\\nRegarding the expressivity analysis, I agree that naturally extending the augmented geometric WL test is relevant and provides a non-domain-specific framework for analyzing geometric WL expressivity. However, the reduction of combinatorial complexes to Hasse graphs has been explored on multiple occasions and, in my view, applying a linear permutation-invariant map to the nodes of the cells appears to be a straightforward extension. Still the additional experiments testing ETNN on the counterexample structures from previous papers are a nice addition and the experimental section of the paper in general is strong.\\n\\n\\n\\n[W2] I agree that the sum of pairwise distances is a natural and straightforward invariant to consider, and I appreciate the empirical demonstration that different aggregation methods do not impact expressiveness. However, from a theoretical perspective, I would have liked to see formal proofs for statements such as: \\u201cETNN with permutation-invariant functions of pairwise distances can implement any function that ETNN with Hausdorff distance or convex hull volume can\\u201d. This would theoretically validate the empirical claim that different aggregation methods do not impact expressiveness. \\n\\nAdditionally, I believe that a graph could theoretically learn geometric node features for a node x, such as \\\"the volume of the convex hull of the k-neighborhood of x\\\" or \\\"the distance of \\nx from the barycenter of the graph.\\\" However, it is unclear whether either ETNN with permutation-invariant functions of pairwise distances or EGNN can effectively learn these types of node features.\\n\\n[W3] I thank the authors for the clarification.\\n\\n[W4] The authors note that \\u201cone can indeed lift a graph into a simplicial complex preserving the graph connectivity by taking triangles or cliques as higher-order simplices\\u201d. Given the paper's claim that using combinatorial complexes is more beneficial than simplicial complexes, a compelling theoretical result would be to demonstrate that ETNNs using lifts that result in non-simplicial complexes are more expressive than ETNNs using lifts that result in simplicial complexes. This might be a generalization of proposition 2, but I think it is an important one.\\n\\n[W5] I thank the authors for their commitment to adding the requested experiments.\"}", "{\"comment\": \"Thank you very much for the clarification. I really appreciate your dedicated efforts. As all concerns are well answered, I am pleased to raise the score. I hope this self-contained and informative piece of work will be an important literature for TDL in general.\"}", "{\"title\": \"Addressing Reviewer's comments and questions\", \"comment\": \"We\\u2019d like to thank the Reviewer for their feedback, as well as the raised concerns and questions. Also, we\\u2019d like to thank them for identifying as strengths of the paper both the framework and the theoretical contributions, the empirical results, as well as the applications and the novel dataset. Next, we address each raised weakness in detail, and also point them out as answers to their questions.\\n\\n## Weaknesses.\\n**[W1] Novelty**\\n\\nOn the one hand, it is true that ETNNs are scalarization-based architectures, as EGNNs [1], on higher-order combinatorial domains, as EMPSNs [2]. As such, they inevitably resound with each other. \\nHowever, we kindly disagree with the reviewer\\u2019s comment since ETNNs are a formal and more expressive generalization of both EMPSNs and EGNNs. For this reason, ETNNs unlock several features (e.g. arbitrary modeling of heterogeneous hierarchical higher-order interactions, tunable expressive power, and general learnable geometric invariants) that the constrained graph or simplicial structures of EGNNs and EMPSNs cannot accommodate (see Appendix C and F). As such, our framework can be used to design arbitrary E(n) equivariant TDL models. No other framework has the same power. \\n\\nMoreover, although we believe that the generality of our framework is fundamental, we also recognized the necessity for applications whose exhaustive modeling (in a combinatorial/topological sense) is possible only via combinatorial complexes, and we introduced MolecularCCs and GeospatialCCs to tackle this problem (no other complex nor graph can model molecular and multi-resolution irregular geospatial data as straightforwardly as CCs). As a consequence, we achieved or matched SotA results among the Equivariant TDL models with a huge reduction in computational complexity. As a byproduct, as reviewer NQcw noticed too, our air pollution downscaling task represents a novel benchmark for the TDL community, addressing a need highlighted in the recent position paper [3]. \\n\\nFinally, we believe that the expressivity analysis is novel. In particular, our approach is, to the best of our knowledge, currently the most exhaustive one for scalarization-based equivariant TDL models, as it could be applied to analyze the (geometric WL) expressivity of any scalarization-based equivariant TDL model without the need for domain-specific coloring procedures.\\n\\n**[W2] Expressivity on CTSs**\\n\\nOur choice of studying expressivity in terms of distinguishing geometric graphs (empirically, we have the counterexamples experiment too, not only the $k$-hop distinct graphs experiment) is due to 3 main reasons: \\n\\n(i) Equivariant TDL is an extremely recent (sub-)field, thus it made sense to us to first show that ETNNs are more expressive than scalarization-based graph models (i.e., EGNNs). This is something already new, as neither [2] nor [4] show any expressivity result. \\n\\n(ii) To achieve (i), we wanted to rely on strong and recognized theoretical tools, and the Geometric WL (and related tasks, e.g., $k$-hop distinct graphs and counterexamples) looked like a natural choice.\\n\\n(iii) The approach we adopt to port the Geometric WL machinery in our setting is itself completely novel and represents a versatile and general approach for studying the (geometric WL) expressivity of any scalarization-based equivariant TDL model that can be derived from the ETNN framework. \\n\\nThis said, we undoubtedly agree with the reviewer's comment. Having a higher-order geometric WL (and related tasks, e.g., $k$-hop distinct complexes) to distinguish among geometric realizations of CTSs is one of the most interesting directions. However, adapting the entire topological-/k-WL machinery to the geometric setting is a highly non-trivial task, and situating it within the context of existing literature adds further complexity. We believe that addressing these challenges fully would necessitate a separate dedicated paper.\\n\\n\\n**[W3] Statement of Proposition 2**\\n\\nThe reviewer is correct that we need to refine the statement to make it clearer. In the revised manuscript, we clearly state the condition of improved expressivity induced by **Proposition 2** in the updated **Corollary 1**. In particular, an ETNN is strictly more powerful than an EGNN in distinguishing $k$-hop distinct graphs if the skeleton-preserving lifting is such that the number of ETNN layers required to have at least one cell in $X_{G_1}$/$X_{G_2}$ whose receptive field is the whole set of nodes in $G_1$/$G_2$ is smaller than the number of EGNN layers required to have at least one node in $G_1$/$G_2$ whose receptive field is the whole set of nodes in $G_1$/$G_2$.\\n\\n*Continuing response in the thread (References are included in the last part of the thread)*\"}", "{\"title\": \"Follow-Up on Reviewer Feedback\", \"comment\": \"We would like to follow up to ask if our response addresses the reviewer\\u2019s concerns, weaknesses, and questions. We would greatly appreciate a prompt feedback, as it would allow us to clarify any remaining issues and further improve the quality of our manuscript. In case the Reviewer is satisfied with our response and the clarifications, we would kindly ask them to reconsider the rating of our submission.\"}", "{\"summary\": \"This paper extends message passing supported on geometric graphs [3] and geometric simplicial complexes [2] to the more general setting of geometric combinatorial complexes. Theoretically, similar to how [1] establishes that higher-order message passing is more expressive than standard graph message passing, this paper demonstrates that the same holds true in the geometric setting. The paper also highlights the effectiveness of the proposed architecture through two novel real-world applications: (1) property prediction over geometric graphs representing molecules, where higher-order cells consist of rings and active groups, and (2) a regression task over geospatial combinatorial complexes. The proposed architecture outperforms previous geometric graph architectures on these tasks.\\n\\n\\n[1] Cristian Bodnar, Fabrizio Frasca, Nina Otter, Yuguang Wang, Pietro Lio, Guido F Montufar, and Michael Bronstein. Weisfeiler and lehman go cellular: Cw networks. Advances in neural information processing systems, 34:2625\\u20132640, 2021.\\n\\n[2] Floor Eijkelboom, Rob Hesselink, and Erik J Bekkers. E (n) equivariant message passing simplicial networks. In International Conference on Machine Learning, pages 9071\\u20139081. PMLR, 2023.\\n\\n[3] V\\u0131ctor Garcia Satorras, Emiel Hoogeboom, and Max Welling. E (n) equivariant graph neural networks. In International conference on machine learning, pages 9323\\u20139332. PMLR, 2021.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. The paper offers a simple and straightforward way to adapt higher order message passing to respect $O(d)$ symmetries.\\n\\n2. The experimental section effectively demonstrates the architecture's performance and introduces two new, interesting real-world TDL benchmarks, addressing a need highlighted in a recent position paper [1]. \\n\\n\\n\\n[1] Theodore Papamarkou, Tolga Birdal, Michael M Bronstein, Gunnar E Carlsson, Justin Curry, Yue Gao, Mustafa Hajij, Roland Kwitt, Pietro Lio, Paolo Di Lorenzo, et al. Position: Topological deep learning is the new frontier for relational learning. In Forty-first International Conference on Machine Learning, 2024.\", \"weaknesses\": \"1. The novelty of the proposed architecture and the theoretical section is somewhat limited. The architecture closely resembles those presented in [4] and [2]. Additionally, the theoretical contributions feel somewhat straightforward, offering limited new insights.\\n\\n2. Building on the previous comment about the theoretical section, the paper lacks an analysis of how the choice of invariant function (see Equation (6)) affects the architecture's expressivity. It would be valuable to examine whether simpler, computationally efficient invariant functions could result in architectures that are as expressive as architectures which use more complex alternatives, thereby guiding the choice of which invariant function to use in practice. Additionally, it would have been insightful to analyze whether there exist any natural geometric invariant functions that geometric graph models are unable to compute while the proposed model succeeds in doing so.\\n\\n3. The end of Proposition 2 states \\\"In most of the cases, an ETNN is strictly more powerful than an EGNN\\\". I tried finding the proof to this in appendix F and had a hard time. I think the authors refer to proposition 2 in appendix F but I'm not sure. A clearer framing of this result in the appendix, and perhaps an illustrative example or plot in the main body, would improve readability and support this claim.\\n\\n4. The paper refers to the architecture proposed in [2] for geometric simplicial complexes. A comparison of the expressive power of ETNN using different lifts compared to this architecture would be interesting.\\n\\n5. The paper [3] benchmarks higher-order message passing on several geometric benchmarks, using data augmentation to address $O(d)$ symmetries. An empirical comparison to this approach would provide valuable insight.\\n\\n[1] Cristian Bodnar, Fabrizio Frasca, Nina Otter, Yuguang Wang, Pietro Lio, Guido F Montufar, and Michael Bronstein. Weisfeiler and lehman go cellular: Cw networks. Advances in neural information processing systems, 34:2625\\u20132640, 2021.\\n\\n[2] Floor Eijkelboom, Rob Hesselink, and Erik J Bekkers. E (n) equivariant message passing simplicial networks. In International Conference on Machine Learning, pages 9071\\u20139081. PMLR, 2023.\\n\\n[3] Mustafa Hajij, Ghada Zamzmi, Theodore Papamarkou, Nina Miolane, Aldo Guzm\\u00b4an-S\\u00b4aenz, Karthikeyan Natesan Ramamurthy, Tolga Birdal, Tamal K Dey, Soham Mukherjee, Shreyas N Samaga, et al. Topological deep learning: Going beyond graph data. arXiv preprint arXiv:2206.00606, 2022.\\n\\n[4] V\\u0131ctor Garcia Satorras, Emiel Hoogeboom, and Max Welling. E (n) equivariant graph neural networks. In International conference on machine learning, pages 9323\\u20139332. PMLR, 2021.\", \"questions\": \"see weaknesses.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper introduces an equivariant model within the framework of topological deep learning. The architecture generalizaes the equivariant graph neural network architecture from Santorras at al. from the setting of graphs to message passing over combinatorial complexes. Notably, this architecture allows for message passing with cells that have heterogeneous node feature over differening ranks.\\n\\nThe authors first introduce the relevant theory of combinatorial complexes, topological deep learning and important noitions of equivariance. In section 3, they introduce their architecture using the formalism from the previous section. The authors then discuss important theoretical aspects of their model; proving equivariance, comparing expressiveness with traditional equivariant, scalarization-based techniques and discussing computational complexity of their design.\", \"the_authors_discuss_in_section_4_two_complementary_examples_of_how_to_attain_combinatorial_complexes\": \"firstly in molecular data and secondly in geospatial data. In section 5, the authors use said combinatorial complexes to benchmark their method against common architectures on the QM9 dataset. Secondly, the authors introduce a novel benchmark for predicting airpollution from annoted geospatial data.\", \"soundness\": \"3\", \"presentation\": \"4\", \"contribution\": \"2\", \"strengths\": \"Presentation, clarity and experimental transparency\\n\\nI found the paper to be written clearly, and the mathematical statements and proofs were precise, well formulated and easy to follow. Further, the authors included a lot of helpful background about topological deep learning, and explained concepts clearly. \\n\\nOne of the biggest strengths of this paper is the much appreciated transparency around testing in the appendix. This allowed a confident and clear understanding of what the authors actually did to obtain their results. I really appreciated the detailled description of cell features in section I.1.2 and I.2.2. The detailled ablation studies on the were also super helpful to understand and evaluate overall performance.\\n\\nNovel geospatial task\\n\\nI really like that the authors introduced a novel geometric prediction task into the literature. In reading the literature for this review, it struck me that many of the benchmarks for TDL were somewhat old and outdated. It seemed appropriate that the task featured integration of data over different dimensional regions (points, lines, cells) in a way that showcased the central feature of the paper \\u2014 reconciling data with features on subspaces of differing dimension i.e. the \\u2018heterogeneous interactions\\u2019 promised in the abstract. \\n\\nComputational\\n\\nI think one of the main strengths of the paper is the computational benefits, and I would personally focus on this more in the introduction. The tailored lifting of molecular graphs to the higher order CCs dramatically decreases the number of higher order cell, which is a problem that plagues many architectures in learning on simplicial/cellular complexes. The section on computational complexity also demonstrates this in a robust mathematical setting.\", \"weaknesses\": \"TDL vs GDL: novelty as a conceptual framework\\n\\nFor me personally I find it hard to understand the framing of this as a part of an entirely new conceptual field of topological deep learning beyond GNNs, and question the genuine novelty of papers like this. This is a concern I have of the field of TDL more generally, but I hope that the authors may be able to help clarify given their excellent communication skills demonstrated in the paper. \\n\\nUnless I\\u2019m mistaken, the basic content of proposition (1) is that an ETNN can be reformulated as an EGNN. This means that the main novelty is the clever choice of \\u2018lifting' the data into a certain graph and some delination of the learning based on \\u2018rank\\u2019. Indeed, even the proof of theorem (1) is basically a straightforward adaption of the corresponding result for EGNNs. I think the need to reformulate everything as an ETNN then show it\\u2019s equivalent to some specific EGNN needs more justification.\\n\\nI still think that the experiments, results and set-up of the paper are interesting, but I personally get the sense that the \\u2019topology\\u2019 part \\u2014 and hence the novelty of these kinds of architectures \\u2014 is overplayed a little. One gets the impression from reading the introduction that there is some topological thing deep down somewhere that is making the difference in performance, whereas my feeling is that the inclusion of domain-specific data \\u2014 functional groups, rings, etc. \\u2014 along with the design of the graph is doing most of the work. \\n\\nOn a similar theme, I don\\u2019t personally see a strong connection with the \\u2018lifted\\u2019 combinatorial complexes used in this paper and topology in the classical sense. It\\u2019s true that cell complexes and simplicial complexes also have the structure of a combinatorial complex (as in Appendix C), but these classical objects have additional connections to topology \\u2014 i.e. they are stratifications of a genuine topological space, posses a homology theory, etc. \\n\\nQM7 Benchmarking comments\\n\\nI find the approach to testing for this benchmark slightly unfair on other methods. Searching through so many iterations of hyperparameters, it seems inevitable that eventually some parameter set will outperform current methods at least once just on the balance of probabilities rather than model strength. \\ufeffSurely a fairer test would be run the competing methods along a similarly vast set of hyperparameters? At minimum, this experiment should be run multiple times, with the variance on results included, to test the robustness of these results.\\n\\nCould the authors elaborate on why they chose to focus only on comparing the ETNN only with EGNN in the bottom line of Table 1? Upon first look, I assumed that ETNN outperformed SotA, but then on closer look saw that Equiformer outperformed ETNN in (almost)( every category. I would recommend highlighting the best performing method in bold in each column so that it\\u2019s more immediately clear that the while ETNN is improving on EGNN, it is still behind the SotA methods like DimeNet++ and Equiformer. My personal opinion is that the bottom line should be removed \\u2014 the fact that it outperforms EGNNs is not surprising considering the fact that EGNNs are somehow a sub-architecture given the results of the expressiveness appendix.\", \"questions\": \"Experiments\\n\\nCould you provide more details on how functional groups were defined or identified in the QM7 dataset? Were these pre-annotated in the data or determined through some other process?\\n\\nIt also seems like many of the best results are given in hyper parameter configurations without the use of higher order cells. Can the authors comment on this? Does this detract from the main thrust of the paper, which is the inclusion of higher order cells?\", \"novelty\": \"TDL vs GDL\", \"turning_my_comments_into_the_weaknesses_section_in_a_specific_question\": \"what do we gain from introducing the notion of a combinatorial complex? As above, (1) the procedures for turning graphs into combinatorial complexes in this paper are not objects commonly studied in topology \\u2014 i.e. do not correspond to a simplicial or cellular complex \\u2014 so do not have access to additional mathematical theory and (2) we could seemingly equally well-formulate this paper as an EGNN over a specific procedure for turning the data into a graph\\ufeff as per Prop. 1. Why not just say this is a specific example of a EGNN using the construction in Prop 1?\\n\\nI am very open to hear the author\\u2019s thoughts/be corrected on this!\\n\\nVirtual cells\\n\\nThe virtual cell seems like to dramatically increase performance. However, my understanding is that this totally change the connectivity of the underlying data, meaning that\\neverything is now connected (as per the comment on 389). \\ufeffWhat does this say about the actual importance of the graph structure? \\n\\nIt seems to suggest the \\u2019topology\\u2019 of the underlying molecular graph is basically ignored once the virtual cell is included. Why not just define the architecture as a family of fully connected EGNN over the nodes, edges and potentially the faces, given that the hyperparameters including the virtual cell almost always are superior?\", \"on_a_related_note\": \"\\\"it is clear how ETNN naturally handles heterogeneity, e.g., the same pair of bonds could be connected because part of the same functional group (up adjacency) and the virtual cell (max adjacency), but the two messages will be different across the neighborhood\\u201d. Unless I\\u2019ve missed something, how are we certain that the two types of propagation add to performance, rather than just overcomplicating the situtation?\\n\\nMiscellaneous\", \"line_354\": \"In the defintiion of geospatial CC, the mapping function s : X \\\\to T takes cells to points in T, but the geographic space s(X) is now a subset of T rather than an element. Should s be mapping to the powerset of T, rather than T, if indeed it takes each cell to a subspace of T?\\n\\nIn general, it would better to have a more concrete definition of what is meant by a \\u2018geospatial combinatorial complex\\u2019. I\\u2019m also unfamilar with the term polyline.\", \"line_375\": \"\\\"This work is the first to explore combinatorial topological modeling of multi-resolution irregular geospatial data..\\u201d The claim that this is the first work to explore combinatorial topological modeling of multi-resolution irregular geospatial data seems overstated given existing literature in Topological Data Analysis applied to geospatial data. Could you clarify how your approach differs from or advances beyond these existing works?\\\" https://arxiv.org/pdf/2104.00720, https://www.researchgate.net/publication/366891451_Topological_data_analysis_for_geographical_information_science_using_persistent_homology, https://pubmed.ncbi.nlm.nih.gov/26353267/\", \"table_2\": \"It would be helpful to have the variance included over the multiple runs\", \"line_242\": \"\\\"geometric invariants should make use of the underlying topological structure.\\u201d The pairwise distance is more of a geometric structure than a topological one.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"We thank the reviewer for their engagement, further feedback, and for agreeing that our paper deserves to be published. We deeply appreciate the time the reviewer dedicated to our work. Here are some further comments.\\n\\n\\n**[W1]** We thank the reviewer for clarifying their point. Overall, it is undoubtedly true that ETNN is a simple equivariant generalization of CCMPNs, similar to EMPSN being a simple generalization of MPSNs, and EGNN being a simple generalization of MPNs. Moreover, it is also true that EMPSN resonates with ETNN being both scalarization-based equivariant TDL models. However, we showed that ETNNs are **more general and flexible**, **more expressive**, and **similar or better performing while being hugely more scalable** than SotA methods in the same class. These facts, together with the **benchmark**, **software**, and **outreaching** contribution of our work, should motivate its publication, as the reviewer agrees. This said, we already wrote in the future directions section that beyond scalarization-based architectures are of great interest. However, given that **a gap in the literature was still present for scalarization-based models** too, it made sense to us to **first comprehensively fill it** by working in the direction of ETNN.\\n\\n\\n**[W2]** We will highlight it in the next revision of the paper we will upload in the next couple of days. Moreover, we agree with the reviewer, that carrying out an in-depth study of which are the distinguishable classes of CCs beyond showing the existence of undistinguishable pairs of CCs is definitely an interesting direction. However, given its complexity and relevance, it would require a standalone work (that builds on top of our expressivity analysis).\\n\\n\\n**[W3]** We agree with the reviewer (and this is somehow also linked to [W2]), although we believe that the generality of Proposition 2 is already a valuable theoretical tool to compare the two models.\\n\\nThanks again!\"}", "{\"title\": \"Response\", \"comment\": \"I thank the authors for their detailed response and address it below.\\n\\n[W1] I want to emphasize that I don\\u2019t believe the paper lacks novelty entirely; rather, its novelty is somewhat limited. Though I agree that the generalization from simplical complexes to combinatorial complexes has value, and that including geometric invariants in the update is a good addition to previous methods, I still find the novelty of these contributions to be somewhat modest.\\n\\n[W2] Though I agree that \\u201cETNN is natively ready to be fed directly with precomputed Hausdorff distances, the volume convex hulls or, in general, higher-order geometric invariants\\u201d I feel like answering these type of questions would make for a more comprehensive theoretical section. While I don\\u2019t expect the authors to answer all of the theoretical questions I posed in my review, I gave these as examples to demonstrate that many theoretical aspects of the proposed architectures remain unexplored.\\n\\n[W4] I thank the reviewer for this answer and agree with it.\\n\\nIn general, I agree with other reviewers saying \\\"Despite being a very well written and tested paper, the main issue I have is the extent of contribution and novelty, as I outline in my answers above. I think the paper deserves to be published in any case\\\" and will keep my score as it is.\"}", "{\"title\": \"Addressing Reviewer's comments and questions\", \"comment\": \"We appreciate the thorough reviewer\\u2019s constructive comments and the valuable questions. We would also like to thank the reviewer for pointing out as strengths the transparency of our experiments, the text clarity, the novelty of the geospatial task, and the computational efficiency of ETNN. Next, we address each point raised as weakness and answer all of the reviewer\\u2019s questions.\\n\\n## Weaknesses.\\n**[W1] TDL vs GDL: novelty as a conceptual framework**\\n\\nAs long-time practitioners of TDL (and GDL) we the authors completely understand the point raised by the reviewer. However, there are multiple answers to why TDL is a motivated framework, spanning from (i) more conceptual (about the framing of TDL and GDL as fields in the ML landscape) to (ii) more pragmatic (about the usage of TDL models in real-world applications). \\n\\nAbout (i), the quick (but not exhaustive) answer is that GDL and TDL stem from two formally and inherently different perspectives on the same objects, \\u201cnon-Euclidean spaces\\u201d. GDL (in the sense of [1]) is built on group-theoretic arguments along with the frequent usage of Hilbert Spaces (strictly related to manifold learning and, in general, to metric spaces), while TDL is solely built on the modeling assumption of data living on the neighborhoods of a combinatorial and/or topological space and having a relational structure induced by the neighborhoods\\u2019 overlap. As such, TDL works tend to put more emphasis on what graph (or manifold) based models struggle to capture, e.g., higher-order combinatorial interactions. Overall, the intersection of GDL and TDL is clearly not empty, but we believe that both fields have a well-framed conceptual motivation. Further insights can be gained from the thesis in [2]-[3].\\n\\nAbout (ii), we firmly believe that TDL as a field needs stronger empirical evidence of its effectiveness. For this reason, in this work, we focused on proposing applications in which higher-order interactions (a) matter and (b) can be captured via ETNN (and some other TDL models in general). Indeed, and this is also a partial answer to **[Q2]** below, the aim of the proposed molecular CC and the novel geospatial benchmark is exactly to prove (a)-(b) true and, more, to also show that there are situations in which neither graphs nor higher-order combinatorial topological spaces (SC or CW complexes) can leverage the available information in a jointly exhaustive and computationally efficient way as CCs do. We thus prove that the modeling of higher-order interactions through CCs indeed offers benefits, especially for structured hierarchical data (e.g., molecular and geospatial data). \\n\\nFinally and most importantly, to the best of our knowledge, ETNN is currently the most exhaustive framework for merging scalarization-based equivariance arguments from GDL with pure TDL arguments.\\n\\n**[W2] ETNN vs EGNN**\\n\\nWe kindly disagree with this observation since the aim of **Proposition 1** is not to show that ETNN is equivalent to some specific EGNN, but rather the opposite: to show that some specific ETNN is computationally equivalent to a (non-standard) EGNN over a Geometric Augmented Hasse Graph. As such, the considered subclass (that is the easiest possible) is not representative at all of the whole class of architectures that can be derived from the ETNN framework. EGNN is formally a particular case of ETNN. The \\u201cgraph-based\\u201d method we introduce is only useful to easily and elegantly prove expressiveness in the geometric WL sense, which is dependent only on the relational structure and the geometric features induced by the CC, not on the specific ETNN architectural choice. This is an intuition already proved successful in works like [4] for non-geometric settings.\\n\\n**[W3] Classical Topology in TDL**\\n\\nThis is a key comment (for this work, its novelty, and TDL in general). We agree that the inclusion of domain-specific data along with the design of a CC (not the graph) that can directly exploit them in a principled way, is doing most of the work. However, given our reply to **[W1]**, we do not believe that this should not be considered topological (in a non-classical sense), as it is the higher-order relational structure induced by the underlying space that contributes to the effectiveness of ETNN. That said, the formal concepts coming from classic algebraic topology, e.g., homology groups, stratifications, filtrations, etc., arise when CCs are particularized to specific combinatorial topological spaces. Several works, e.g., [5]-[6], have shown improvements related to classical topological arguments. E(n) Equivariant versions of these architectures can be directly derived from the ETNN framework.\\n\\n*Continuing response in the thread (References are included in the last part of the thread)*\"}", "{\"comment\": \"**[Q4] Max Adjacency**\\n\\nWe exhaustively study how the max adjacency (and the virtual cell) impacts the performance. In **Tables 10-11**, the \\u201cAdjacencies\\u201d column tells exactly what adjacencies are used. In general, the max adjacency and the virtual cell increase performance, and this is expected, as the virtual node [7] had a similar effect on graph-based models.\\n\\n**[Q5] Notation**\\n\\nThe equations only require \\u201ccontainment\\u201d notation since $\\\\mathcal{CN}$ contains neighborhood systems, $\\\\mathcal{N}(x)$ contains neighboring cells, $x$ and $y$ are cells that contain atoms $z$. \\n\\n***Could the Reviewer please clarify which components of **Equations 6** and **7** are they referring to?***\\n\\n**[Q6] Higher-order geometric features**\\n\\nThe setting that we tackle, which is the same as EGNN and EMPSN and is clearly stated at the beginning of **Section 3**, is the setting in which nodes (0-cells) are embedded in some Euclidean space, i.e., they come with both non-geometric and geometric features. This makes sense as higher-order cells are not necessarily physical entities. That said, in **Remark 1** we also explain how to modify ETNN to integrate higher-order cells with geometric features, if available. Further, **Appendix D** discusses velocity-type inputs. However, in both cases, it is not clear to us what the reviewer means when they write about lifting (node) geometric features. The geometric features, as it is written in **Equation 7**, are updated only for the cells (say the nodes) that come with geometric features. On the other hand, the non-geometric features are updated for all the cells (**Equation 6**), and geometric invariants take into account the geometric features of all the nodes that are a part of a higher-order cell. Because of the message-passing operations between the different ranks, subsequent layers of message-passing at higher-order ranks will use the updated information from the geometric features at the node level. \\n\\n## References\\n\\n[1] Grigor\\u2019yan, A.A., Lin, Y., Muranov, Y.V., et al. \\\"Path Complexes and their Homologies.\\\" *Journal of Mathematical Sciences*, 248, 564\\u2013599 (2020).\\n\\n[2] Truong, Q., & Chin, P. \\\"Weisfeiler and Lehman Go Paths: Learning Topological Features via Path Complexes.\\\" *Proceedings of the AAAI Conference on Artificial Intelligence*, 38(14) (2024).\\n\\n[3] Hajij, M., et al. \\\"Topological Deep Learning: Going Beyond Graph Data.\\\" (2023).\\n\\n[4] Li, L., et al. \\\"Path Complex Neural Network for Molecular Property Prediction.\\\" *ICML 2024 Workshop GRaM* (2024).\\n\\n[5] Lecha, Manuel, et al. \\\"Higher-Order Topological Directionality and Directed Simplicial Neural Networks.\\\" *arXiv preprint* arXiv:2409.08389 (2024).\\n\\n[6] Papamarkou, Theodore, et al. \\\"Position: Topological Deep Learning is the New Frontier for Relational Learning.\\\" *Forty-first International Conference on Machine Learning* (2024).\\n\\n[7] Sestak, Florian, et al. \\\"VN-EGNN: E (3)-Equivariant Graph Neural Networks with Virtual Nodes Enhance Protein Binding Site Identification.\\\" *arXiv preprint* arXiv:2404.07194 (2024).\"}", "{\"metareview\": \"This paper extends TNN to ETNN, similar to the extension of GNN to EGNN. There have been very intensive discussions among the authors and reviewers, and the authors put a lot of efforts during the process. Eventually, I believe all the major issues have been resolved. I agree with some of the reviewers that the work is not entirely novel given the similarity with extension of GNN to EGNN. In this sense, I tend to think this is a borderline paper. On the other hand, given the many empirical comparisons and insights of add equivariance to existing methods, I tend to recommend an accept.\", \"additional_comments_on_reviewer_discussion\": \"There have been extensive discussions and all the major issues have been resolved.\"}", "{\"comment\": \"We thank the reviewer for the detailed response and for agreeing that our work is critical to the TDL field and starts filling an important gap in the literature. In the following, we address their follow-up comments. We hope that these further replies along with the overall positive feelings the reviewer showed for our work, can convince them to raise their score.\\n\\n**[W1, Q6] Novelty and Higher-order Geometric Features**\\n\\nElaborating more about what we wrote in the previous reply, we would like to point out that: \\n\\n(i) The *architectural* novelty is not representative of the novelty of our work as a whole. In this specific case, the novelty is given by:\\n\\n- (a) **Architectural** novelty: ETNN improves on the two most important components of a scalarization-based equivariant architecture, i.e., feature updates and geometric invariants. Regarding features update, graphs and SCs induce specific neighborhood functions (node adjacency for the former, boundary, coboundary, up/down adjacencies for the latter), thus if a CC comes with a non-standard collection of neighborhoods, neither (ofc) EGNN nor EMPSN can handle it, both for non-geometric and geometric features updates. Regarding geometric invariants, EGNN/EMPSN uses tailored geometric invariants for graphs/SCs. Imagine applying EMPSN on a CC. Then, computing the volume of a non-simplex makes no sense. Similarly, pruning the pairwise distances based on the simplicial structure or using the angles of the planes induced by the dimension of the simplices makes no sense. On the other hand, ETNN can handle arbitrary neighborhoods, and its geometric invariants work on arbitrary CTSs (and formally generalize some of the invariants of EMPSN and EGNN).\\n\\n- (b) **Experimental** novelty: the introduction of the new geospatial benchmark and our novel approach for molecular modeling represent a significant effort to tackle some of the important open problems of the field [3] and are a significant source of novelty (and an important resource for the community) as well.\\n\\n- (c) **Theoretical** novelty: ETNN is the only available framework for designing arbitrary E(n) equivariant scalarization-based TDL models and the expressivity proof leveraging the novel notion of geometric augmented Hasse graph has not appeared before in the literature of equivariant TDL.\\n\\n(ii) We believe that our architecture should be considered **simple** rather than *trivial*. If a model is designed to be general and flexible, and it is shown to be more expressive, scalable, and better performing than the SotA of its class, defining it trivial just because of the simplicity of its architectural definition sounds reductive (and slightly impolite) to us. This said, we explicitly wrote in the future directions section that beyond scalarization-based architectures are of great interest. However, given that the gap in the literature was present for scalarization-based models too, it made sense to us to comprehensively work in the direction of ETNN.\\n\\n(iii) Higher-order relations are not just a generalization of adjacency matrices in graphs. In the context of ETNN, it is true that it is a widely general class of models on a broadly general space. However, E(n) Equivariant versions of architectures, e.g., [4], leveraging powerful concepts coming from algebraic topology, e.g., homology groups, stratifications, filtrations, etc., arise when ETNNs and CCs are particularized to specific combinatorial topological spaces, e.g., cell complexes. Similarly, E(n) Equivariant versions of architectures using specific insights on hypergraphs, e.g. [5], arise when ETNNs and CCs are particularized to hypergraphs. ETNN is the only framework having this joint feature, to the best of our knowledge. \\n\\n*Continuing response in the thread (references are at the end of the thread)*\"}", "{\"comment\": \"**[W1]** We are glad the reviewer appreciated the increased versatility of ETNN and the strength of our experimental section, recognizing the contributions of our work. We also thank them for their prompt response, as usually in these cases, it is vital to have a constant and fruitful discussion.\\n\\nHowever, we would like to point out that the sentence *\\u201cTo my understanding, there is nothing inherently restrictive in EMPSNs update rule that prevents them from handling combinatorial complexes, even though they have been framed as a simplicial complex-focused architecture.\\u201d* is slightly tendentious: the fact that an architecture can be directly generalized does not imply that the generalization has no value. Especially when the treatment, as in our ETNN case, is clearly broader, more comprehensive, and supported by novel and non-trivial experiments. EMPSN is a simplicial neural network and should be considered so. It could be generalized, and we did it by clearly motivating it both in methodological (i.e., expressivity and geometric invariants), experimental (i.e., a novel benchmark), and computational (i.e., less than half of EMPSN memory usage and almost half of EMPSN runtime with same or better performance) terms. In mere terms of forward pass (that obviously is just a piece of our work), while ETNN forward can process data on a SC, EMPSN forward cannot be used to process data on a CC. The main reasons regard the two most important components of a scalarization-based equivariant architecture, i.e., (1) Feature updates and (2) Geometric invariants.\\n\\nRegarding (1), SCs induce specific neighborhood functions (Boundary, Coboundary, Up/Down Adjacencies), thus if a CC comes with a non-standard collection of neighborhoods, EMPSN forward cannot handle it, both for non-geometric and geometric features updates.\\n\\nRegarding (2), EMPSN forward uses tailored geometric invariants for SCs. Computing the volume of a non-simplex makes no sense. Similarly, pruning the pairwise distances based on the simplicial structure or using the angles of the planes induced by the dimension of the simplices makes no sense in the CC case.\\n\\nRegarding expressivity, although the reduction of combinatorial complexes to Hasse graphs has been explored in a couple of works, the only benchmark study using it for expressivity purposes is [1]. As such, we still believe that our expressivity analysis, being the first one to (partially) generalize the arguments from [1] to the geometric setting, relating it to the geometric WL, is a significant contribution.\\n\\n**[W2]** We agree with the reviewer that questions like *\\u201ccan ETNN with permutation-invariant functions of pairwise distances implement any function that ETNN with Hausdorff distance can?\\u201d* would be interesting for (mainly) computational aspects. However, formally proving general statements on this topic would require an entire paper (maybe a paper also related to the generalization capabilities of ETNN?). The simplistic answer to questions like the one above is yes. In the specific case of the Hausdorff distance, for example, it would suffice to use a learnable weighted sum of pairwise distances and the network should learn to set to zero all the weights of the sum not corresponding to the pair of points in the two cells that are maximally distant (in the set sense of the Hausdorff distance).\\n\\nRegarding graph models, could the reviewer point us to some work showing what are the geometric invariants that geometric graph models can learn starting from pairwise distances? Intuitively, given our answer above, ETNN should be able to do it more straightforwardly anyway.\\n\\nOverall and most importantly, we would like to stress that, although an interesting problem, there is no need to focus too much on which kind of invariants can be learned from pairwise distances because ETNN is natively ready to be fed directly with precomputed Hausdorff distances, the volume convex hulls or, in general, higher-order geometric invariants. This is actually one of the main features of Equivariant TDL models in general. \\n\\n**[W4]** We believe that the notion itself of a theoretical expressivity result about (geometric) CCs vs SCs is ill-conditioned. The theoretical reason is that the statement would be equivalent to the new **Corollary 1** but replacing EGNN and graphs with EMPSN and simplicial complexes. More importantly, the practical reason is that on both the numerical expressivity tasks that we show, the $k$-chain graphs (**Table 3**) and the counterexamples (**Table 4**), EMPSN with any skeleton preserving lifting would be as expressive as an EGNN (and, thus, way less expressive than an ETNN), as there are no triangles nor cliques. \\n\\n## References.\\n[1] Jogl, Fabian, Maximilian Thiessen, and Thomas G\\u00e4rtner. \\\"Expressivity-preserving GNN simulation.\\\" Advances in Neural Information Processing Systems 36 (2024).\"}", "{\"title\": \"Follow-up: Did our responses address the Reviewer's questions?\", \"comment\": \"We\\u2019d like to follow up to ask if our responses sufficiently addressed the Reviewer\\u2019s comments.\\n\\nAlso, we'd like to ask them if the extra experiments we conducted to provide the statistical measures (as shown in the tables above) meet expectations or if further experiments are needed.\\nOtherwise, in case the Reviewer is satisfied with our response and the clarifications, we'd greatly appreciate it if they could reconsider the rating of our submission.\"}", "{\"title\": \"Global Response\", \"comment\": \"We would first like to thank the reviewers for their detailed and constructive feedback. The paper is noted for its clarity (**Reviewers NQcw, h23Z**), rigor and elegance (**Reviewers NQcw, J6CW**), self-containment (**Reviewer h23Z**), computational efficiency (**Reviewer NQcw**), and transparency (**Reviewer NQcw, h23Z**). The expressivity analysis is overall found thorough. Moreover, we are really glad all the reviewers agreed on the importance to the TDL community of the two real-world applications and the novel geospatial benchmark. Overall, the paper is seen as addressing important gaps in the TDL literature with theoretical and practical contributions.\\n\\nA point-by-point reply to all the comments is individually given in the sequel. We reply to comments in the same order they were presented by the reviewers, and we also group them by \\\"Weaknesses\\\" and \\\"Questions\\\". All the changes in the **revised manuscript** appear in blue color, to facilitate checking. We also summarize here our discussion about the novelty of our work.\\n\\n**Novelty.** \\n\\nThe reviewers indicated novelty concerns about our work. However, we believe that this work is conceptually, technically, and practically novel.\\n\\nETNN (and TDL in general) is **conceptually** novel and motivated because, Geometric Deep Learning (GDL) and TDL stem from two formally and inherently different perspectives on the same objects, \\u201cnon-Euclidean spaces\\u201d. GDL (in the sense of [1]) is built on group-theoretic arguments along with the frequent usage of Hilbert Spaces (strictly related to manifold learning and, in general, to metric spaces), while TDL is solely built on the modeling assumption of data living on the neighborhoods of a combinatorial and/or topological space and having a relational structure induced by the neighborhoods\\u2019 overlap. As such, TDL works tend to put more emphasis on what graph (or manifold) based models struggle to capture, e.g., higher-order combinatorial interactions. Overall, the intersection of GDL and TDL is clearly not empty, but we believe that both fields have a well-framed conceptual motivation. Further insights can be gained from the thesis in [2]-[3]. Overall, to the best of our knowledge, ETNN is currently the most exhaustive framework for merging scalarization-based equivariance arguments from GDL with pure TDL arguments.\\n\\nOn the one hand, it is true that ETNNs are scalarization-based architectures, as EGNNs, on higher-order combinatorial domains, as EMPSNs. As such, they inevitably resound with each other. \\n\\nHowever, ETNN is **technically** novel as it is a formal and more expressive generalization of both EMPSNs and EGNNs. For this reason, ETNNs unlock several features (e.g. arbitrary modeling of heterogeneous hierarchical higher-order interactions, tunable expressive power, and general learnable geometric invariants) that the constrained graph or simplicial structures of EGNNs and EMPSNs cannot accommodate (see Appendix C and F). As such, our framework can be used to design arbitrary E(n) equivariant TDL models. No other framework has the same power. The expressivity analysis is novel as well. In particular, our approach is, to the best of our knowledge, currently the most exhaustive one for scalarization-based equivariant TDL models, as it could be applied to analyze the (geometric WL) expressivity of any scalarization-based equivariant TDL model without the need for domain-specific coloring procedures. \\n\\nETNN is **practically** novel because it tackles the fundamental problem of TDL as a field needing stronger empirical evidence of its effectiveness, as described in [4]. In this work, we focused on proposing applications of vastly different scales in which higher-order interactions (a) matter and (b) can be captured via ETNN (and some other TDL models in general). Indeed, the aim of the proposed molecular CC and the novel geospatial benchmark is exactly to prove (a)-(b) true and, more, to also show that there are situations in which neither graphs nor higher-order combinatorial topological spaces (SC or CW complexes) can leverage the available information in a jointly exhaustive and computationally efficient way as CCs do. Moreover, the experiments also show the versatility of ETNN in tackling very different problems. We thus prove that the modeling of higher-order interactions through CCs indeed offers benefits, especially for structured hierarchical data (e.g., molecular and geospatial data). As a consequence, we achieved or matched SotA results among the Equivariant TDL models with a huge reduction in computational complexity. As a byproduct, as **Reviewer NQcw** noticed too, our air pollution downscaling task represents a novel benchmark for the TDL community.\\n\\nFinally, as a side minor note, we'd like to highlight the outreaching scope of our work. In its current form, it is sufficiently self-contained to be accessible to any ML practitioner, expert, or non-expert.\\n\\n*The references are included in the next thread.*\"}", "{\"summary\": \"**Method:** The paper introduces a framework, termed ETNN, for processing combinatorial complexes with geometric cell features in an $\\\\mathrm{E}(n)$-equivariant way (explicitly introduces a method for $0$-cell geometric features and explains how to generalize to geometric features of cells of arbitrary rank). ETNNs are equivariant to the action of both $\\\\mathrm{Sym}(\\\\mathcal{X})$ (cell renaming) and $\\\\mathrm{E}(n)$ isometries acting on the geometric features. The architecture builds on higher-order message-passing over CCs and has both invariant and equivariant versions. ETNNs generalizes previous work on $\\\\mathrm{E}(n)$-equivariant graph neural networks, and $\\\\mathrm{E}(n)$-equivariant cellular complex networks.\\n\\n**Theory:** The expressivity of the proposed methods is evaluated based on its ability to distinguish $k$-hop distinct geometric graphs. The paper proves that in *most cases* ETNN is strictly more expressive than baseline geometric graph methods (i.e. message-passing architectures on geometric graphs that do not operate on higher-order cells).\\n\\n**Experiments:** \\n- **QM9:** The first application considered in the paper is QM9 molecular property prediction, where molecules are represented as combinatorial complexes with geometric features. ETNN variants demonstrate clear performance increase over standard geometric graph methods.\\n- **Air Pollution Benchmark:** The second application introduces a new benchmark for air pollution prediction. CCs are constructed from point measurements (0-cells), road measurements (1-cells), and census data (2-cells). \\n\\n**Notes.**\\n- Typo in equation (13); should be \\\"$x \\\\subset z$, and $y \\\\subset z$\\\".\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"3\", \"strengths\": [\"**Framework and Theoretical Contributions:**\", \"The ETNN architecture introduced in the paper is a novel framework for $\\\\mathrm{E}(n)$-equivariant processing of a wide class of topological data objects.\", \"The framework elegantly unifies and generalizes existing architectures -- e.g. EGNN [1] and EMPSN [2].\", \"The paper theoretically proves expressivity improvements over $\\\\mathrm{E}(n)$-equivariant graph methods.\", \"**Empirical Results:**\", \"ETNN variants achieve clear performance gains over standard $\\\\mathrm{E}(n)$-equivariant graph methods for QM9 molecular property prediction tasks; ETNNs also improves upon EMPSNs while using less memory and having a faster runtime.\", \"**Applications and Dataset Contributions:**\", \"The paper introduces a principled approach to modeling irregular multi-resolution geospatial data using combinatorial complexes.\", \"The *air pollution downscaling* benchmark introduced in the paper is a new dataset for benchmarking TDL architectures. Additionally, the construction and analysis of geospatial combinatorial complexes is in itself an interesting contribution of the paper.\", \"[1] Satorras et al. \\\"E(n) equivariant graph neural networks\\\", ICML 2021.\", \"[2] Eijkelboom et al. \\\"E(n) equivariant message passing simplicial networks\\\", ICML 2023.\"], \"weaknesses\": [\"**Novelty:** While the proposed method generalizes previous work to general combinatorial complexes, the architectural changes are incremental.\", \"**Theoretical analysis:**\", \"The expressivity analysis focuses solely on distinguishing geometric *graphs*, despite defining expressive power in terms of separating more general non-isomorphic *CTSs*.\", \"Proposition 2's statement is imprecise: the claim of improved expressivity \\\"in most cases\\\" lacks formal qualification, and the relationship between the expressivity gap and choice of lifting method needs clearer formulation and concrete characterization. A clearer restatement would be helpful.\", \"No formal comparison to expressivity of other TDL methods (e.g. equivariant simplicial networks).\", \"**Empirical evaluation:**\", \"Results lack statistical significance analysis (no standard deviations reported).\", \"On the air pollution benchmark, improvements over the MLP baseline are modest (~1.5% RMSE reduction) and their statistical significance cannot be assessed due to unreported standard deviations.\"], \"questions\": [\"Can the authors clarify the motivation for using \\\"k-hop distinct\\\" graphs as the primary measure of expressivity? Is it possible extend the expressivity results to CCs? E.g. is it possible to define a notion of \\\"k-hop distinctness\\\" for geometric combinatorial complexes and analyze ETNN's ability to distinguish between such complexes?\", \"How does ETNN's expressivity compare to that of simplicial complex networks in distinguishing non-isomorphic geometric graphs/complexes?\", \"In proposition 2, can you specify the conditions under which ETNNs are provably more expressive than EGNNs?\", \"Could you include standard deviations for the air pollution benchmark and QM9 results to assess statistical significance?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Follow-up after deadline extension\", \"comment\": \"In light of the extension of the discussion deadline, we would like to thank the reviewer again for the already positive assessment of our work and for their engagement, and ask if there is anything else we can do to further improve the manuscript and its score.\\n\\nWe also kindly invite the reviewer to parse the discussions we had with the other reviewers, as we believe that the overall engagement and the individual feedback have clarified or improved several aspects of the paper.\\n\\nThanks a lot!\"}", "{\"title\": \"2nd Follow-Up: Are there any outstanding questions or concerns?\", \"comment\": \"We would like to ask the reviewer if there are any outstanding questions or concerns, that were not addressed in our previous responses. We would greatly appreciate prompt feedback to address any points and ensure our manuscript meets expectations.\"}", "{\"title\": \"2nd Follow-up: Are there any outstanding points?\", \"comment\": \"We'd like to thank once again the Reviewer for the constructive review. We want to follow up to ask whether our responses addressed their questions so far. Otherwise, if the Reviewer is satisfied with our response and the clarifications, we would kindly ask them to reconsider the rating of our submission.\"}", "{\"title\": \"Follow-up on Reviewer\\u2019s feedback\", \"comment\": \"We would like to ask you whether our response addressed your concerns, weaknesses, and questions so far. Also, we\\u2019d like to know whether you have any other questions?\\n\\nWe would greatly appreciate a prompt feedback, as it would allow us to clarify any remaining issues and further improve the quality of our manuscript.\"}", "{\"comment\": \"**[W4] ETNN vs EMPSN expressivity**\\n\\nAs a consequence of what we wrote in our reply to **[W1]**, a comparison is not required, because CCs generalize Simplicial Complexes (SCs) and our approach based on the novel notion of *Geometric Augmented Hasse Graph* can thus be directly applied to SCs. As we carefully explained in the paper, an expressivity comparison (in the geometric WL sense) is fair only when the lift is skeleton-preserving, i.e. when the underlying graph connectivity is preserved; in [2], they do not lift the graphs (and they do not provide any expressivity analysis), but rather they replace the connectivity with a Vietori-Rips SC. Although one can indeed lift a graph into a simplicial complex preserving the graph connectivity by taking triangles or cliques as higher-order simplices, the expressivity would be characterized exactly as for **Proposition 2** (and as in **[W3]** to compare with EGNNs).\\n\\n**[W5] Comparison with CCMPNs**\\n\\nWe believe this is an interesting point, as there is an ongoing debate on where and when either hard equivariance inductive bias or a massive data augmentation is required. However, implementing the data augmentations and rerunning all the experiments is beyond what can be accomplished within the given 1-week timeframe of the rebuttal. Having said that, we are committed to conducting these experiments and incorporating the results in the final camera-ready version of the paper. Nevertheless, we exhaustively showed the superior performance of ETNNs against CCMPNs with no data augmentation (configurations. 16-17-25-28-37-38-39 of **Tables 10-11**).\\n\\n## References\\n\\n[1] Cristian Bodnar, Fabrizio Frasca, Nina Otter, Yuguang Wang, Pietro Lio, Guido F Montufar, and Michael Bronstein. \\\"Weisfeiler and Lehman Go Cellular: CW Networks.\\\" *Advances in Neural Information Processing Systems*, 34:2625\\u20132640, 2021.\\n\\n[2] Floor Eijkelboom, Rob Hesselink, and Erik J Bekkers. \\\"E(n) Equivariant Message Passing Simplicial Networks.\\\" In *International Conference on Machine Learning*, pages 9071\\u20139081. PMLR, 2023.\\n\\n[3] Mustafa Hajij, Ghada Zamzmi, Theodore Papamarkou, Nina Miolane, Aldo Guzm\\u00e1n-S\\u00e1enz, Karthikeyan Natesan Ramamurthy, Tolga Birdal, Tamal K Dey, Soham Mukherjee, Shreyas N Samaga, et al. \\\"Topological Deep Learning: Going Beyond Graph Data.\\\" *arXiv preprint* arXiv:2206.00606, 2022.\\n\\n[4] V\\u00edctor Garcia Satorras, Emiel Hoogeboom, and Max Welling. \\\"E(n) Equivariant Graph Neural Networks.\\\" In *International Conference on Machine Learning*, pages 9323\\u20139332. PMLR, 2021.\\n\\n[5] Papamarkou, Theodore, et al. \\\"Position: Topological Deep Learning is the New Frontier for Relational Learning.\\\" *Forty-first International Conference on Machine Learning*, 2024.\"}", "{\"comment\": \"\\u200b\\u200b**Standard errors for the geospatial task and additional seeds**\\n\\nDear reviewer, we have conducted multiple additional runs for the geospatial task and computed standard errors based on your suggestion. We ran 10x more seeds (30 in total). Despite the high variance, the results are very similar to those reported in the paper. Standard errors are computed with the formula $\\\\sigma/\\\\sqrt{n}$, which is recommended when having at least sample size 30.\\n\\n| Baseline | Av. Explained Variance ($R^2$) | Std. Err ($R^2$) | **MSE** |\\n| -------- | ------------------------------ | ---------------- | ------- |\\n| ETNN | 9.34% | 2.05% | 0.935 |\\n| GNN | 2.44% | 0.99% | 0.987 |\\n| MLP | 2.35% | 1.61% | 1.022 |\\n| EGNN | 1.43% | 1.2% | 1.041 |\\n| Linear | 0.51% | 0.95% | 1.106 |\\n**Updated Table 2a of the paper**. Baseline model comparison\\n\\n| Baseline | Diff. Explained Variance ($R^2$) | Std. Err ($R^2$) | **MSE** |\\n| ----------------------------- | -------------------------------- | ---------------- | ------- |\\n| no geometric features (CCMPN) | -1.08% | 2.05% | 0.946 |\\n| no position update (invariant ETNN) | -1.37% | 2.52% | 0.956 |\\n| no virtual node | -1.80% | 2.32% | 0.957 |\\n| | | | |\\n**Updated Table 2b of the paper**. Ablation study\"}", "{\"comment\": \"We thank the reviewer for their engagement, further feedback, and for stating that our paper deserves to be published. We sincerely and deeply appreciate the time the reviewer dedicated to our work. Here are some further comments.\\n\\n\\n**[W1]** It is undoubtedly true that ETNN is a simple equivariant generalization of CCMPNs, being a scalarization-based equivariant model, similar to EMPSN being a simple generalization of MPSNs, and EGNN being a simple generalization of MPNs. However, beyond the technical differences we already described, overall we showed that ETNNs are **more general and flexible**, **more expressive**, and **similar or better performing while being hugely more scalable** than SotA methods in the same class. These facts, together with the **benchmark**, **software**, and **outreaching** contribution of our work, should motivate its publication, as the reviewer agrees. This said, we already wrote in the future directions section that beyond scalarization-based architectures are of great interest. However, given that **a gap in the literature was still present for scalarization-based models** too, it made sense to us to **first comprehensively fill it** by working in the direction of ETNN.\\n\\n\\n\\n**[W2]** We agree with the reviewer that open questions remain despite our extensive analysis and treatment. However, we believe this is almost always the case for any new architecture, especially in the case of a very young field such as Equivariant TDL (indeed, the same questions could be rightfully asked for the other architectures in the field). In this sense, ETNN, in our opinion, already represents the most comprehensive work in its related literature, and we believe it can spark interest to be further studied, by the community and by us as well.\\n\\nThanks again!\"}", "{\"comment\": [\"I Thank the authors for their thoughtful and detailed response. A few followup questions/comments:\", \"**Re: architectural novelty:** ETNNs are a direct generalization of EMPSNs. Apart from the fact that the input to an ETNN model is a signal defined over a combinatorial complex (CC) and the input to an EMPSN model is a signal defined over a simplicial complex (SC), what is the difference between their forward passes? That is, could I have used an EMPSN model to process signals on a CC?\", \"While I understand that applying this architecture to a more general data object is interesting and allows for experimental flexibility (and that is a contribution of the paper), I\\u2019m still trying to understand the architectural modifications needed in order to do that.\", \"**Re: Expressivity on CTSs:** I recognize the difficulty of analyzing expressivity in a new (sub)field, but I believe this point is important and worth further investigation. Since the premise of the paper is that operating on CCs with geometric features is useful, it\\u2019s important to understand the expressive power *in this domain specifically*. Even without a higher order geometric WL hierarchy, a concrete example of CCs with geometric features that ETNNs can/cannot distinguish (provably) is valuable.\", \"**Re: expressivity comparison to equivariant TDL methods:** While ETNNs and e.g. EMPSNs operate on different objects, both objects (CCs and SCs) can be lifted from (geometric) graphs. In the graph expressivity sense, is lifting to a CC and using ETNN better than lifting to a SN and using EMPSN?\", \"**Re: statistical significance of experimental results:** I see that other baselines don\\u2019t report std results on QM9 either. While I still believe this is problematic I accept the authors\\u2019 answer. As to the air pollution benchmark, standard deviation results (even on 3 seeds, but ideally more) would be appreciated.\"]}", "{\"title\": \"Addressing authors' clarifications\", \"comment\": \"I thank the reviewers for thorough responses, which have strengthened my understanding of this work. The paper presents good contributions that merit publication, and I'd like to address the clarifications while explaining why I maintain my current score.\\n\\n**[W1]** The authors' response clarifies the key differences between ETNNs and EMPSNs, particularly regarding neighborhood functions and geometric invariants. While I now have a clearer technical understanding of these differences, I view them as adaptations for handling CCs rather than fundamental architectural innovations. As I noted in my review, this contribution is still valuable \\u2013 I simply maintain this as a non-detrimental weakness.\\n\\n**[W2]** The authors' clarification about Remark 6 and the example of undistinguishable geometric CCs is helpful, and I think the paper would benefit from highlighting it in the main text. However, two points remain: (a) while instructive, this specific example doesn't easily generalize, and it\\u2019s unclear how we can establish a more comprehensive expressivity hierarchy, and (b) it demonstrates one direction (indistinguishable geometric CCs) without fully characterizing which geometric CC classes can be distinguished. I again want to stress that this is a non-detrimental weakness and that, as I pointed out in my review, the paper's theoretical contributions are valuable.\\n\\n**[W3]** I understand the authors' argument re: the expressivity comparison (for $k$-chains and counterexamples) through Tables 3 and 4 + the fact that EMPSN with skeleton-preserving lifting would be limited to EGNN expressivity in those cases. This is a subtle argument that mixes empirical and theoretical observations. A formal theoretical characterization would be valuable in future work.\\n\\n**[W4]** The additional seeds for the geospatial experiment strengthen the empirical validation and are a good precedent for the benchmark.\\n\\nIn conclusion, I maintain my positive assessment of the paper. The paper makes valuable contributions to equivariant topological deep learning, particularly through its practical innovations and empirical results. While the noted limitations prevent me from increasing my score, they don't detract from the paper's value and publishability.\"}", "{\"comment\": \"**[Q4] Definition of Geospatial CC**\\n\\nThank you for your careful reading and for bringing this omission to our attention. The corrected notation should indeed be $s: X \\\\to \\\\mathcal{P}(T)$. \\n\\nRegarding your second comment, we have revised the manuscript to include a more concrete definition of a geospatial CC (GCC) at the beginning of the relevant section. GCCs are now explicitly defined as CCs in which all cells are subsets of a geospatial domain. However, we believe that the mapping $s$ remains necessary for a more formal definition and to explain the construction of neighborhood systems. \\n\\nWe have also clarified the definition of a polyline in the revised manuscript. It is now stated that a polyline is simply a sequence of connected lines, and we have included a citation to an introduction to common data formats in geographic information systems.\\nWe hope that these improvements to the wording highlighted in the revised manuscript provide greater clarity. \\n\\n**[Q5] Combinatorial topological modeling of multi-resolution irregular geospatial data**\\n\\nAs we explain in Appendix H.2, TDA methods have been applied to geospatial data (we added the missing references among the ones indicated by the reviewer). However, TDA is different from TDL (see again Appendix H.2), and in all the referenced works there is no combinatorial characterization induced by, e.g., political partition and no multi-resolution. They are mostly all variants of PH-related arguments applied to geospatial data.\\n\\n**[Q6] Variances in Table 2.**\\n\\nWe originally did not report variances since our experiments are averaged over 3 seeds, which makes variance calculations unreliable. We are now working on running with additional seeds and will report the results in the next few days when the experiments are complete.\\n\\n**[Q7] Line 242**\\n\\nAll the geometric invariants are geometric objects. However, their definition is totally dependent on the underlying topological combinatorial space. E.g., which are the pairwise distances to be summed are given by the relational structure of the space and its neighborhoods.\\n\\n## References\\n\\n[1] Bronstein, Michael M., et al. \\\"Geometric deep learning: Grids, groups, graphs, geodesics, and gauges.\\\" *arXiv preprint* arXiv:2104.13478 (2021).\\n\\n[2] Bodnar, Christian, \\u201cTopological Deep Learning: Graphs, Complexes, Sheaves\\u201d, *PhD Thesis*, [https://www.repository.cam.ac.uk/items/06b0b8e5-57d1-4120-8fad-643ce4d40eda](https://www.repository.cam.ac.uk/items/06b0b8e5-57d1-4120-8fad-643ce4d40eda) (2022).\\n\\n[3] Battiloro, Claudio, \\u201cSignal Processing and Learning over Topological Spaces\\u201d, *PhD Thesis*, [https://theses.eurasip.org/theses/974/signal-processing-and-learning-over-topological/](https://theses.eurasip.org/theses/974/signal-processing-and-learning-over-topological/) (2024).\\n\\n[4] Jogl, Fabian, Maximilian Thiessen, and Thomas G\\u00e4rtner. \\\"Expressivity-preserving GNN simulation.\\\" *Advances in Neural Information Processing Systems* 36 (2024).\\n\\n[5] Yang, Maosheng, Elvin Isufi, and Geert Leus. \\\"Simplicial convolutional neural networks.\\\" *ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)*. IEEE, 2022.\\n\\n[6] Battiloro, Claudio, et al. \\\"Generalized simplicial attention neural networks.\\\" *IEEE Transactions on Signal and Information Processing over Networks* (2024).\\n\\n[7] Sestak, Florian, et al. \\\"VN-EGNN: E (3)-Equivariant Graph Neural Networks with Virtual Nodes Enhance Protein Binding Site Identification.\\\" *arXiv preprint* arXiv:2404.07194 (2024).\"}", "{\"comment\": \"*Continuing response*\", \"take_again_the_molcc_as_an_example\": \"bonds are modeled as edges (1-cells) and rings such as carbon rings are modeled as faces (2-cells). Two bonds can simultaneously share multiple neighborhoods. For instance, they could be lower adjacent because they have a common atom (0-cell) and, at the same time, also be upper adjacent because they are part of the same molecular ring (2-cell). Despite their different chemical meaning, the whole geometric augmented Hasse graph would collapse these two relations (upper adjacent, lower adjacent) into one. Moreover, the resulting non-standard EGNN would not be able to distinguish anymore which node of the geometric augmented Hasse graph was an atom, a bond, or a ring in the original molecule, and would process all the connections with the same set of weights. However, and this motivates our approach and validates its elegance, the non-standard EGNN resulting from Assumptions 1-3 is as expressive as any general ETNN, as it is the most specific and constrained instance of it.\\n\\n[Q3] Exactly, as long as the adjacencies are in the collection of neighbrhoods, even if the higher-order cell features are not used.\\n\\n[Q6] Sure, we are collecting some more suggestions and we will update a further revision in the next few days.\\n\\n## References.\\n[1] Battiloro, Claudio, et al. \\\"Topological signal processing over weighted simplicial complexes.\\\" ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2023.\\n\\n[2] Roddenberry, T. Mitchell, Michael T. Schaub, and Mustafa Hajij. \\\"Signal processing on cell complexes.\\\" ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2022.\\n\\n[3] Hajij, Mustafa, et al. \\\"Topological deep learning: Going beyond graph data.\\\" arXiv preprint arXiv:2206.00606 (2022).\\n\\n[4] Ghrist, Robert, and Hans Riess. \\\"Cellular sheaves of lattices and the tarski laplacian.\\\" arXiv preprint arXiv:2007.04099 (2020).\"}", "{\"comment\": \"We sincerely thank the reviewer for their valuable feedback and suggestions. We would also like to thank the reviewer for pointing out as strengths the importance of this work for the TDL community, the text clarity, its clarity, completeness and self-containment. Next, we address each point raised as weakness and answer all of the reviewer\\u2019s questions.\\n\\n## Weaknesses.\\n**[W1] Novelty**\\n\\nOn the one hand, it is true that ETNNs are scalarization-based architectures, as EGNNs, on higher-order combinatorial domains, as EMPSNs. As such, they inevitably resound with each other. \\n\\nHowever, we kindly disagree with the reviewer\\u2019s comment since ETNNs are a formal and more expressive generalization of both EMPSNs and EGNNs. For this reason, ETNNs unlock several features (e.g. arbitrary modeling of heterogeneous hierarchical higher-order interactions, tunable expressive power, and general learnable geometric invariants) that the constrained graph or simplicial structures of EGNNs and EMPSNs cannot accommodate (see **Appendix C** and **F**). As such, our framework can be used to design arbitrary E(n) equivariant TDL models. No other framework has the same power. \\n\\nMoreover, although we believe that the generality of our framework is fundamental, we also recognized the necessity for applications whose exhaustive modeling (in a combinatorial/topological sense) is possible only via combinatorial complexes, and we introduced MolecularCCs and GeospatialCCs to tackle this problem (no other complex nor graph can model molecular and multi-resolution irregular geospatial data as straightforwardly as CCs). As a consequence, we achieved or matched SotA results among the Equivariant TDL models with a huge reduction in computational complexity. \\n\\nAs a byproduct, as **Reviewer NQcw** noticed too, our air pollution downscaling task represents a novel benchmark for the TDL community, addressing a need highlighted in the recent position paper [6]. \\nFinally, we believe that the expressivity analysis is novel. In particular, our approach is, to the best of our knowledge, currently the most exhaustive one for scalarization-based equivariant TDL models, as it could be applied to analyze the (geometric WL) expressivity of any scalarization-based equivariant TDL model without the need for domain-specific coloring procedures. \\n\\n**[W2] Scalability**\\n\\nWe kindly disagree with the reviewer since we believe that scalability is one of ETNN\\u2019s strengths (as pointed out by **Reviewer NQcw** too). In particular, as described in **Section 5** and **Appendix G**, we showed that ETNN achieves or matches SotA results among E(n) Equivariant TDL models with considerable gains in memory, time, and, overall, scalability. These facts make ETNN amenable to be used on larger datasets in tailored, future works. The scope of this work is to introduce the framework, analyze it in detail, make it accessible, and show how it can be easily used for problems of vastly different natures. About the ablation, EGNN-graph-W has been introduced exactly to show how a graph model with the same parameter budget would perform, while the use or not of features has been extensively studied in **Appendix J**, **Tables 10-11** (HO Features column). Finally, we do not consider an average improvement of more than 11% incremental.\\n\\n## Questions.\\n\\n**[Q1] Heterogeneous Interactions**\\n\\nThe reviewer is right. Indeed, we state that CCMPNs can handle heterogeneity before we even introduce ETNN, and we give the right credit to [3]. However, it remains true that ETNN are set up for heterogeneity (in the geometric setting too). To improve the fairness of our presentation, in the revised manuscript we made clear that TDL models have this feature in the abstract.\\n\\n**[Q2] Path Complexes as CCs**\\nWe believe that, as long as the considered paths are undirected (and this seems to be the case in [2], [3], [4] but not in [1]), a path complex can be easily cast as a CC in which the cells are the paths and the incidences and adjacencies are defined via the boundary relation of the complex (that, of course, takes into account the sequential nature of paths by definition) and not the set inclusion, as it happens in cell complexes (that indeed generalize path complexes). In the case of directed paths in the complex, then CCs cannot model it as the neighborhood structure would be possibly asymmetric. In this case, a different approach as the one in [5] should be employed. We also added [4] in our related works section.\\n\\n\\n*Continuing response in the thread (References are included in the last part of the thread)*\", \"title\": \"Addressing Reviewer's comments and questions\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}" ] }

[ "Brandon Zhao", "Aviad Levis", "Liam Connor", "Pratul P. Srinivasan", "Katherine Bouman" ]

[ "computational imaging", "signal processing", "inverse problems", "astrophysics", "cosmology", "neural fields", "machine learning for physical sciences" ]

{ "note_id": [ "qaGiUB5HgN", "q91RWhKe1q", "pONvtVr8fW", "oPZrjsBP55", "lbHbG4QHKr", "lEm9K1Q883", "k7NIlXnEsR", "gy3dID3NH7", "e0trP7PN38", "ZWtJSI3ovv", "Wt1iyxjciQ", "L1HHQ4qjOR", "KWKwFclFtF", "Dd56VPd4jO", "B9PLerKUyq", "9U0i0dSQAd", "8ClHF6zbVd", "6wv5IlMdDF", "55x9No71eZ", "2Cgjz0N57G" ], "note_type": [ "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "decision", "official_review", "official_comment", "official_comment", "official_comment", "official_review", "meta_review", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_comment" ], "note_created": [ 1732673177956, 1731729161436, 1731729205969, 1732706629644, 1732673201442, 1732557845163, 1730285952415, 1737523874078, 1731469457487, 1731729278041, 1732547227256, 1731901098242, 1730267744690, 1734576375031, 1731729326068, 1732328312836, 1733110286215, 1730700882731, 1732598419352, 1731729101221 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission7909/Area_Chair_6q1K" ], [ "ICLR.cc/2025/Conference/Submission7909/Authors" ], [ "ICLR.cc/2025/Conference/Submission7909/Authors" ], [ "ICLR.cc/2025/Conference/Submission7909/Reviewer_p3hC" ], [ "ICLR.cc/2025/Conference/Submission7909/Area_Chair_6q1K" ], [ "ICLR.cc/2025/Conference/Submission7909/Authors" ], [ "ICLR.cc/2025/Conference/Submission7909/Reviewer_gGfb" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission7909/Reviewer_p3hC" ], [ "ICLR.cc/2025/Conference/Submission7909/Authors" ], [ "ICLR.cc/2025/Conference/Submission7909/Reviewer_gGfb" ], [ "ICLR.cc/2025/Conference/Submission7909/Reviewer_pR5T" ], [ "ICLR.cc/2025/Conference/Submission7909/Reviewer_pR5T" ], [ "ICLR.cc/2025/Conference/Submission7909/Area_Chair_6q1K" ], [ "ICLR.cc/2025/Conference/Submission7909/Authors" ], [ "ICLR.cc/2025/Conference/Submission7909/Authors" ], [ "ICLR.cc/2025/Conference/Submission7909/Reviewer_x8ZD" ], [ "ICLR.cc/2025/Conference/Submission7909/Reviewer_x8ZD" ], [ "ICLR.cc/2025/Conference/Submission7909/Authors" ], [ "ICLR.cc/2025/Conference/Submission7909/Authors" ] ], "structured_content_str": [ "{\"title\": \"Response\", \"comment\": \"Dear Reviewer,\\nWhile the review was positive and response brief, it would still be helpful to acknowledge and note any other reviews or response influence your score. Thank you.\"}", "{\"comment\": \"We train our network through the Adam optimizer with an exponential learning rate decay schedule for a fixed number of iterations. We are able to perform full gradient descent without batching on a single card due to our use of the Fast Fourier Transform in the forward model, so there is no stochasticity after initialization. We have not observed significant instabilities when training our model for experiments. We will make all code and data publicly available for reproducibility.\"}", "{\"comment\": [\"Training a single neural field model requires 2.3GB of memory and takes 15 minutes for 100k iterations on an A100 GPU. Inference for a single model is almost instantaneous (order of hundredths of a second).\", \"3D dark matter maps are particularly useful for many downstream applications. As highlighted in our introduction, these maps can help to answer questions about the fundamental nature of our universe: Can Dark Matter exhibit trace amounts of gamma ray radiation? How do dark matter structures evolve over time? How can we characterize the primordial distribution of the dark matter field at times close to the Big Bang? Most of these applications specifically target non-Gaussian features of interest, such as galaxy clusters, which makes our method particularly well-suited for downstream science.\"]}", "{\"title\": \"Thanks\", \"comment\": \"Thanks! I did follow the other reviews and responses. I am happy to keep my score as clear accept.\"}", "{\"title\": \"Response\", \"comment\": \"Dear Reviewer,\\nDo you mind letting the authors know if their rebuttal has addressed your concerns and questions? Thanks!\\n-AC\"}", "{\"comment\": \"We have written a couple of paragraphs in our global comment which explain why we believe our work constitutes a significant contribution to the machine learning community. As the end of the discussion period approaches, we would greatly appreciate it if you could confirm whether our response has adequately addressed your concerns. If you have any remaining questions, please let us know, and we will do our best to respond within the remaining time.\\n\\nIf your concerns have been addressed, we kindly ask you to consider raising your rating. Thank you again for your time and efforts in reviewing our manuscript.\"}", "{\"summary\": \"The article considers the task of reconstructing a 3D map of the universe from weak gravitational lensing obtained in 2D signals. The methods consider first to encode the overdensities using a neural field based on the spherical coordinate and its Fourier decomposition. The target density field is used for two purposes: (i) to match the power spectrum of the dataset under consideration and (ii) to match the shear observed (after using the transformation from overdensity to shear) in the data. The method is tested on simulated data mimicking the instrumental observations from an N-body simulation. The results show that the method seems better than using Wiener filter for the same task.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"The authors apply ML techniques to a dataset that is different from the usual benchmarks.\", \"weaknesses\": \"My overall impression is that we have just an application of ML methods to a cosmological problem, which neither\\n- make relevant improvement on the side of ML (unless I'm mistaking, it takes out-of-the-box methods)\\n- make a strong advance in the considered field\\nI would like the authors to comment on how their work is particularly timely. I'm also thinking that the Wiener filter is not very modern. By quickly looking on the web, we can find more elaborated techniques:\\n * sparsity prior: https://www.aanda.org/articles/aa/pdf/2021/05/aa39451-20.pdf, https://arxiv.org/pdf/1801.08945\\n * wavelet: https://www.aanda.org/articles/aa/pdf/2006/21/aa2997-05.pdf\\nAnd I can imagine there are many others. This work does not provide any comparison to these methods.\\n\\nThe last section that investigate non-Gaussian structure is a bit of a mystery to me. Does the application of the method to MNIST (whatever it means) should be considered of a test of something precise ? In addition,\", \"questions\": [\"it might be useful to related the various red-shift to a number of years, it is done for z=0 and z=2 but not for z=1.\", \"The caption of Fig. 3 says: \\\"results includes 12 lens planes from redshift z=0 to z=1\\\", can the authors specify the precise value of the considered z ?\", \"In addition, the authors say that they average over 3 lens planes. Why ? to which values of z does that correspond ?\", \"The ground truth images are blurred with a factor that maximize the cross-correlation with the reconstructed picture. It seems a bit weird to do so, it might bias the reader to believe that a reconstruction is better than what it should be.\", \"The cross-correlation between the reconstructed signal and the true one seems quite small... while we can agree that it is even worse using the Wiener filter, I have the impression that it is working very-well.\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}", "{\"summary\": \"***Late review: Apologies my review is in one day late. I will aim to engage as quickly as needed to makeup.***\\n\\nThis very interesting paper considers a very important application, although using some well-established tools in machine learning and computer vision. The paper focuses on the difficult weak-lensing regime, where deflections are too weak to be interpretable reliably for a single galaxy. The idea is that attempting to understand how weak lensing alters (or shears) images of galaxies in dense fields, one could reconstruct a density map of the underlying matter. This is a difficult problem as the matter density is usually analyzed in 2D: the measured shear map is used to recover a projected 2D estimate. The paper aims to solve the inverse problem i.e. to obtain a 3D reconstruction of the dark matter field from 2D images. From a modeling perspective, this becomes challenging as galaxies are observed from a single view. Further, unlensed shapes of known visible sources are themselves not fully understood, introducing a large amount of uncertainty. The main idea of the paper is to use the underlying physics of gravitational lensing to recover a continuous 3D field, with a focus on also capturing non-Gaussian features, which are not easy with traditional methods, which use strong priors. The idea is to use coordinate-based neural fields augmented with the underlying physics (clearly described in figures 1 and 2), with the total loss function described in equation 7. The experiments, over simulated data, are able to show that the method has promise in both reconstructing the 3D matter field and also non-Gaussian features.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"-- The paper attacks a problem of quite some importance. While the machine learning used is standard, its combination with the underlying physics allows the authors to obtain a promising solution, although it is only tested on simulations (standard for this area of physics at the moment).\\n\\n-- The paper is well-written, well-structured and easy to follow. While I am not an expert in this specific application, I have worked in some adjacent applications. I think the results support the underlying proposal and motivation well, and are creative in combining standard ML tools with physics-based models.\", \"weaknesses\": \"-- More of a question: Could the authors elaborate in the paper why are strong Gaussian priors usually made? This will not be obvious to a standard ML audience. I see that it could make sense from the perspective of filtering. But I am not sure I understand why this should drop out from the physics models. I would expect the features to be highly non-Gaussian. I find this confusing whenever I venture to look at papers in this area.\", \"questions\": \"-- Could the authors elaborate on the training protocols and how difficult it was to get stabilized training? From the writeup, it is not clear how easy it is to get the model to work.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"8\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": [\"We will add a table relating redshifts of each lens plane to light years in the appendix. The 12 lens planes are equally spaced in comoving distance, which is standard in works of this type. We average over 3 adjacent lens planes for visual brevity, but we can also add a figure with all 12 planes to the appendix.\", \"Because of the nature of noisy single-view backprojection, reconstructions typically suffer from some amount of blurring in the z direction which is not trivial to determine. By finding the optimal blur kernel we are able to estimate this effective blur amount which tells us the resolution at which we can reliably recover structural information.\", \"The 3D cross correlation is commonly used in astrophysics for comparing the similarities of 3D fields (see, for example, [2]). It harshly punishes the prediction of overdensities of the wrong sign.\", \"We find that a Gaussian assumption can be limiting for 3D dark matter mapping. The purpose of the MNIST experiment was to highlight this point on a toy example for which a Gaussian prior does not match the true field well. Our method performs well in both situations that have Gaussian and non-Gaussian structure. Since non-Gaussian structures, such as galaxy filaments and clusters, are of great scientific interest, we believe our method is better suited for their downstream analysis.\", \"[2] https://academic.oup.com/pasj/article/70/SP1/S26/4097646\"]}", "{\"comment\": \"I thank the authors for their careful answer. It is still hard to me to consider this an application of ML that should be published in this venue, I agree this is somewhat subjective and this is confirmed by my confidence score.\\n\\nI recognize the effort on the answer, I would put the vote 4, but it is not possible anymore....\"}", "{\"comment\": \"I thank the authors for considering my comments. I still believe that the premise if the study is too empirical for the lensing application -- this is unfortunately a somewhat general concern with application of modern ML methods (which tend to be highly empirical) to the physical sciences (which require a high degree of calibration, control, and robustness, with the current target application being a good example). For example, the role of the implicit induced prior here makes it quite difficult to understand how much of the reconstruction is \\\"real\\\" dark matter, and how much is attributed to the induced prior. While this is not particularly an issue in more traditional application of neural fields to e.g. scene reconstruction, for cosmological applications it is critical. Ablation studies varying over $L$ would be helpful if included, but this is not particularly generalizable -- the effect of bandwidth will be unpredictable when applying the method to a different setting (different survey configuration, survey volume, forward model etc). While shown to be more generalizable than the Gaussian prior of previous approaches, as mentioned in the original review there are other methods that leverage physical priors that are probably more appropriate comparison points. Finally, I appreciate the added study comparing the ensemble error to per-point reconstruction error -- while the ensembling approach seems a bit hacky compared to an in-site probabilistic one, it does give some idea of calibration. Overall, I'm happy to raise my score a notch. Applications to physics problems are challenging, and this is a good effort which I think would be motivating to the conference audience.\"}", "{\"summary\": \"The paper presents a method for reconstructing the underlying 3D dark matter distribution from weak gravitational lensing observations (specifically, shape distortions of galaxies). This is an important problem in cosmology, and is of contemporary relevance given the number of astronomical surveys which will be measuring galaxy shapes. The overall approach is to model the dark matter distribution implicitly through a neural network (NeRF-like or implicit neural representation), and use a differentiable forward model that models the observed shear field. The authors compare the method to a few traditional approaches for mass mapping, finding it to especially perform well in reconstructing non-Gaussian structures.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": [\"**Novel approach and good technical execution:** Using an implicit representation to model the underlying dark matter field along with a differentiable forward model is a very neat but challenging idea, and the authors set it up and executed it very well.\", \"**Paper straddling AI and fundamental physics:** This is a bit of a meta point, but AI + physics (in particular a sub-field like cosmology) is not particularly mainstream at AI conference so far, in contrast to e.g. AI + bio. This makes presenting a paper like this while being faithful to the domain science particularly challenging, and the authors do a good job in this respect.\", \"**Simulation realism:** The authors use a realistic set of tools from the cosmology literature, e.g. JaxPM in the differentiable forward model, making it a contribution beyond a simple proof-of-principle.\", \"**Baseline comparisons:** Comparisons with baselines are well-presented, and highlight the advantage of the method in particular regimes (e.g. in recovering non-Gaussian structures).\"], \"weaknesses\": [\"**Context in current literature and state of the field:** There exists substantial literature in neural approaches to mass mapping; e.g., https://arxiv.org/abs/2201.05561. While the authors compare to traditional approaches like Kaiser-Squires (KS) and Wiener filter, they do not contextualize their work in the more recent ML-based approaches to weak lensing mass mapping. Another example is https://arxiv.org/abs/2206.14820, which uses an implicit neural representation to perform *strong* lensing reconstruction; high-level comparison with existing literature could be improved.\", \"**Hyperparameter choices:** The authors mention specific hyperparameter choices, e.g. bandwidth of positional encodings L=2 and 5, without further justification. This should be expected to have a significant impact on the results, as it controls the spectral biases of the implicit representation, and should be further expanded upon, possibly including ablations.\", \"**Role of induced prior:** Traditional approaches typically have a regularization mechanism, often explicit, which allows for mass reconstruction (since the inverse problem is fundamentally ill-posed). In this study, the authors mention that \\\"neural fields have been shown to provide a good implicit prior\\\"; while true in the real-world setting of scene reconstruction, scientific problems are inherently different in nature, and it is not clear a-priori whether the implicit prior induced by the neural field is a good one. The authors approach this empirically, however further understanding of the role of induced priors is necessary for downstream science from the mass maps.\"], \"questions\": [\"The authors use an ensembling approach, taking the median to get a point estimate on the reconstructed mass map. Much of the ongoing literature/effort is focused on _probabilistic_ approaches to mass mapping. The authors mention in a footnote that ensembling can lead to good predictive uncertainties -- can it be expected that this can be used to yield a calibrated distribution over plausible mass maps? If not, are there alternative approaches to this incorporating neural fields, e.g. variational or Bayesian approaches?\", \"What is the role of bandwidth $L$ and the induced spectral bias in the study? How does the induced prior compare with that assumed in traditional and model ML-based approaches?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"metareview\": \"This paper studies the problem of reconstructing the 3D dark matter field from 2D observations. The method uses a coordinate-based neural field with positional encodings to represent the matter field which is then passed through a physics-based differentiable forward model to obtain a shear field which can be compared to observations.\", \"strengths\": \"This paper addresses an important problem in astronomy. The technique is a novel one in weak lensing, although the underlying neural field methods are not new, and is an interesting combination of ML methods with a physics-based simulator. The evaluations are promising and show good results against reasonable baselines on simulated data. The paper is well-written and contains a good background on the application area for a general audience.\", \"weaknesses\": \"The main the noted weakness of the paper is limited innovation the ML methodology which uses standard techniques, although they have not been applied to this physics problem before. Another noted limitation of the method is limited ability to determine the impact of implicit priors in the reconstruction.\", \"conclusion\": \"Despite the range of scores, reviewers broadly agreed on the merits and weaknesses of the paper. All reviewers actually agree that the paper is of high quality, but question whether ICLR is the right venue for publication since the paper does not make any new contributions to core ML methodology and is largely an application of ML to a new domain. There is not a clear cut consensus here, and I appreciate the reviewers openness and flexibility during the discussion phase. In the end, I tend to believe that scientific applications papers are important to the field and figuring out how to deploy even existing methods in new domains is worthwhile to our community.\", \"additional_comments_on_reviewer_discussion\": \"Reviews for the paper were quite split with two reviewers on the side of rejecting (x8ZD,gGfb) and two on the side of acceptance (pR5T, p3hC). Despite the range of scores, reviewers broadly agreed on the merits and weaknesses of the paper. From the discussion phase during which all reviewers contributed, all reviewers actually agree that the paper is of high quality, but question whether ICLR is the right venue for publication since the paper does not make any new contributions to core ML methodology and is largely an application of ML to a new domain. Both x8ZD and gGfb indicated that despite their negative score, they are not dead set against the paper being accepted. pR5T who raised their score from 5 to 6, pointed out the contributions to astronomy probably do not yet meet the bar for publication in that field. I cannot evaluate that point. While there is not a clear cut consensus here, I appreciate the reviewers openness and flexibility. In the end, I tend to believe that scientific applications papers are important to the field and figuring out how to deploy even existing methods in new domains is worthwhile to our community.\"}", "{\"comment\": [\"Thank you for bringing these works applying neural representations to other lensing paradigms; we will be sure to expand our discussion to include these in related works.\", \"While we chose the hyperparameter settings empirically, we expect the reconstructions to have a degree of robustness in terms of spectral bias because we are explicitly regularizing the power spectrum of our reconstruction. However, we agree that an ablation on the positional encoding degree would be appropriate and will add one to the appendix.\", \"Previous works have been proposed for understanding the implicit prior induced by positional encoding in the context of Neural Tangent Kernel (NTK) theory [1]. Intuitively increasing the bandwidth L should result in an implicit prior with a wider spectrum, allowing the network to fit to higher frequencies. However this is a theoretical result for the infinite-width limit and does not take into consideration our spectral regularization; we are very interested in studying the implicit prior of our method and tuning it appropriately for future work. Empirically, the induced prior of our method appears to be more generalizable than the Gaussian prior of previous approaches.\", \"Although our current ensemble representation doesn\\u2019t mathematically correspond to a Bayesian posterior, we find that our current method already provides meaningful uncertainty estimates. In our appendix we have added a calibration plot for the experiment shown in Fig. 3. There is a strong correlation between the ensemble standard deviation and reconstruction error, except for a few outlier regions. In regions with high ensemble variance we find our model is underconfident. In regions with low ensemble variance our uncertainty estimates are slightly overconfident, but the error is still well within 2 standard deviations.\", \"[1] https://arxiv.org/abs/2007.05864\"]}", "{\"comment\": \"We have now added a table with distance measures for each reconstructed lens plane, as well as a figure visualizing all 12 reconstructed lens planes (without averaging) to the appendix. As the end of the discussion period approaches, we would greatly appreciate it if you could confirm whether our responses have adequately addressed your concerns. If so, we encourage you to consider updating your rating to reflect the improvements made based on your valuable feedback.\"}", "{\"title\": \"Re: Official Comment by Authors\", \"comment\": \"Dear Authors,\\n\\nThanks for your response to my concerns. \\n\\nI still believe that this paper does not have enough new ML contributions. In my opinion, the most important scientific contributions of the paper are in astrophysics and not in machine learning. I do agree, however, that the problem itself can be an important ML application domain. \\n\\nTherefore, I kept my original rating score, but I also lowered my confidence score to indicate that it would be fine with me if the paper gets accepted.\"}", "{\"summary\": \"The paper provides a new method for recovering the 3D cosmic web from galaxy weak lensing signal.\\nThe proposed approach uses a coordinate-based neural field method with positional encoding. (Equation 6).\\nThe reconstruction results are demonstrated in a series of experiments.\", \"soundness\": \"3\", \"presentation\": \"4\", \"contribution\": \"2\", \"strengths\": [\"The paper proposes a new method for an important physics problem.\", \"The paper is well-written; I enjoyed reading it.\", \"The physics background is exceptionally well-presented, and even non-experts can understand the main scientific ideas.\", \"The proposed method works better than the leading baseline.\"], \"weaknesses\": \"I think this is a great paper, but in my opinion, ICLR might not be the best venue for this.\", \"the_new_machine_learning_contributions_are_quite_limited\": \"neural field models and positional encoding have been around for some time, and I didn't find significantly new machine learning contributions in the paper.\\n\\nNonetheless, the proposed algorithm works well, and the physics application, estimating the 3D cosmic web, is a very important problem.\\nSince the most significant contributions of the paper are in astrophysics, I think an astrophysics or cosmology journal would be a better venue for this paper.\", \"questions\": [\"I would like to know some more details about the number of CPUs/GPUs used for training, memory requirements, and the training and inference time of the proposed method.\", \"I'm also curious about the next steps. After the 3D cosmic web estimation step is done, what are the next science questions that can be answered with the new results?\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Thank you for the score revision! If you get a chance, would you mind trying again to update your official score in openreview? Reviewers are still supposed to be able to update it at the moment, and we know of other reviewers who updated their scores today.\"}", "{\"comment\": \"Thank you for your helpful comments and questions. We present a \\u201ccreative\\u201d (p3hC) and \\u201cnovel approach\\u201d (pR5T) that attacks \\u201can important physics problem\\u201d (x8ZD) with comparisons to baselines that \\u201care well-presented, and highlight the advantage of the method in particular regimes\\u201d (pR5T). Reviewer x8ZD mentions that it is a \\u201cgreat paper\\u201d that is \\u201cexceptionally well-presented\\u201d so that \\u201c even non-experts can understand the main scientific ideas.\\u201d As highlighted by reviewer pR5T, \\u201cpresenting a paper like this while being faithful to the domain science [is] particularly challenging, and the authors do a good job in this respect.\\u201d\\n\\n**Machine Learning Contributions (x8ZD, gGfb)**\\n\\nWe firmly believe that our work represents a significant contribution to the ML community. Below, we outline several reasons that we will clarify in the updated paper.\\n\\n**Neural Fields for Science:** While neural field models with positional encoding have previously been applied to various problems, our work demonstrates how the framework can be incorporated with the underlying physics specific to our problem. We work with noisy measurements from a single viewpoint, and the cosmic web volumes we study have different statistical properties compared to natural scenes. As highlighted by pR5T, this unique context raises important ML questions: How does the implicit regularization in neural fields help estimate solutions to severely ill-posed, underconstrained inverse problems? And how well does positional encoding, effective in natural scenes, generalize to these new scientific contexts? Applications of neural fields was the topic of 2023 ICLR (https://sites.google.com/view/neural-fields) and 2024 ECCV (https://neural-fields-beyond-cams.github.io/) workshops, featuring invited talks that apply neural fields with positional encoding to different 3D reconstruction tasks. We believe that adapting neural fields to a new context with completely different physics represents a significant contribution, and gives value to the ML community.\\n\\n**Introducing a New Problem to ML:** Additionally, one of the main challenges of interdisciplinary work is effective communication between fields, and we are glad to have expressed our results in a way that is understandable to the ML community (x8ZD, p3hC, pR5T), as this important problem could greatly benefit from their insights and contributions. We also wish to highlight that, to our knowledge, this is the first time that a comparison of 3D mass mapping methods has been done. We will make all code and datasets available so that this problem can be further studied by the ML community. \\n\\n**A Stepping Stone to Future Work:** We believe our work is a necessary stepping stone to future work that would take advantage of the neural representation we\\u2019ve presented. One direction we are actively pursuing is leveraging the neural representation to obtain efficient uncertainty quantification, which, as pR5T notes, is important for downstream science. While probabilistic approaches to 2D mass mapping are an active field of study, to our knowledge none have yet been proposed for 3D mass mapping, possibly due to computational intractability arising from the extra dimension. Even in the simple case of the Wiener filter, which has an analytic posterior, computing this posterior with a full covariance becomes intractable in 3D. In future work, we plan to build on the theory from [1], which demonstrates how neural representations can play an essential role in enabling efficient posterior estimation. The method we present serves as a solid foundation for further progress in this important direction.\\n\\n**Wiener Filter Baseline (p3hC, gGfb)**\\n\\nAs gGfb highlights through various references, there have been recent developments in the field of 2D mass mapping. However, these methods are not directly applicable to 3D. Currently, there are no publicly available 3D mass mapping codes in working condition. While one code exists, it is non-functional, and the authors were unable to provide a working version. To the best of the authors' knowledge, the Wiener filter remains the only 3D method that has been applied to real cosmic shear data, with its most recent usage documented in 2018 [2]. Although code for this method was also not publicly available, we implemented our own transverse Wiener filter to establish a baseline.\\n\\nIn regards to p3hC\\u2019s question about Gaussian priors: On scales much larger than those we are targeting, the overdensity field of the universe can be described as Gaussian [3]. This coupled with the simplicity of the Gaussian prior makes it a popular choice for reconstructions of this type. However, non-Gaussianity present on small cosmic scales motivates our more general ML-based approach.\\n\\n[1] https://arxiv.org/abs/2007.05864\\n\\n[2] https://academic.oup.com/pasj/article/70/SP1/S26/4097646\\n\\n[3] https://ned.ipac.caltech.edu/level5/March01/Coles/Coles4.html\"}" ] }

[ "Bishwajit Saha", "Dmitry Krotov", "Mohammed J Zaki", "Parikshit Ram" ]

[ "deep clustering", "associative memories", "representation learning", "Hopfield networks" ]

{ "note_id": [ "zFRpWqtA2I", "y7rIGY22iy", "teqpjiR56W", "q77ZFDaKHz", "obYsucDvn4", "oY9Q0ItwHU", "ebazIWUcNv", "dF0B4gr0fw", "cs6uLwJvJB", "c0VelNjAMy", "YC9R9WeaY0", "XGDqR3uwx7", "WizqLNHFkY", "Vk3LrKmg4M", "S7MK6Oun7a", "QKRppYDvUf", "Oie0uqqh6j", "MvCoOtFHqV", "J7GWg2dRZb", "Hhv2IzSGsN", "GwUOZbcsr5", "GjfOcGC0Ff", "Gg0IV5HQqO", "9SgaPWjhyl", "7XC8xEXKbV", "5vhVoFAHzR", "3xQIwX5wYp" ], "note_type": [ "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_review", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "decision", "official_comment", "official_comment", "official_comment", "meta_review" ], "note_created": [ 1732321241991, 1732320723149, 1730698472269, 1732321463517, 1732923792328, 1732925041113, 1732321612803, 1733168337259, 1732319733065, 1733153644198, 1733288506666, 1730659788563, 1730879605170, 1730661789040, 1732464053281, 1733172014855, 1733159775121, 1733171341501, 1732320381343, 1732924696926, 1732924550536, 1732321165977, 1737523713302, 1732599645681, 1732322044146, 1732598767834, 1734716983778 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission5548/Authors" ], [ "ICLR.cc/2025/Conference/Submission5548/Authors" ], [ "ICLR.cc/2025/Conference/Submission5548/Reviewer_AgR2" ], [ "ICLR.cc/2025/Conference/Submission5548/Authors" ], [ "ICLR.cc/2025/Conference/Submission5548/Authors" ], [ "ICLR.cc/2025/Conference/Submission5548/Authors" ], [ "ICLR.cc/2025/Conference/Submission5548/Authors" ], [ "ICLR.cc/2025/Conference/Submission5548/Reviewer_AgR2" ], [ "ICLR.cc/2025/Conference/Submission5548/Authors" ], [ "ICLR.cc/2025/Conference/Submission5548/Reviewer_whZj" ], [ "ICLR.cc/2025/Conference/Submission5548/Authors" ], [ "ICLR.cc/2025/Conference/Submission5548/Reviewer_Z3Ff" ], [ "ICLR.cc/2025/Conference/Submission5548/Reviewer_whZj" ], [ "ICLR.cc/2025/Conference/Submission5548/Reviewer_ZsFx" ], [ "ICLR.cc/2025/Conference/Submission5548/Reviewer_whZj" ], [ "ICLR.cc/2025/Conference/Submission5548/Authors" ], [ "ICLR.cc/2025/Conference/Submission5548/Authors" ], [ "ICLR.cc/2025/Conference/Submission5548/Authors" ], [ "ICLR.cc/2025/Conference/Submission5548/Authors" ], [ "ICLR.cc/2025/Conference/Submission5548/Authors" ], [ "ICLR.cc/2025/Conference/Submission5548/Authors" ], [ "ICLR.cc/2025/Conference/Submission5548/Authors" ], [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission5548/Authors" ], [ "ICLR.cc/2025/Conference/Submission5548/Authors" ], [ "ICLR.cc/2025/Conference/Submission5548/Authors" ], [ "ICLR.cc/2025/Conference/Submission5548/Area_Chair_BseA" ] ], "structured_content_str": [ "{\"title\": \"Response (Part 2)\", \"comment\": \"> In contemporary literature standards, both the neural network trained in the work and the dataset are small. It is unclear whether the proposed method would generalize to more realistic network sizes and datasets.\\n\\nWe have experimented with a diverse set of standard benchmark datasets ranging from 2000 to 60000 data points, and from 10 to 200 clusters, that cover most of the commonly used datasets in deep clustering literature. \\n\\nAs such, our method should scale to much larger datasets, since the Train subroutine of $\\\\texttt{DClAM}$ in algorithm 1 has a linear complexity of $O(dkT N|S|)$, where $d$ is the latent dimension, $k$ is the number of clusters, $T$ is the number of AM steps, $N$ is the number of epochs and $|S|$ is the dataset size. Also, $\\\\texttt{DClAM}$ leverages SGD and works on batches of training data, so we anticipate that there are no barriers to scaling to larger datasets. Another thing we want to mention regarding the network size is that $\\\\texttt{DClAM}$ works seamless with any choice of encoder and decoder (for example, convolutional or residual networks for images or fully-connected feed-forward networks for text or tabular data, and so on).\\n\\n> It seems like the entire (vision) transformer literature is missing from the analysis. This is not a critical concern, but it would be interesting to evaluate DClAM's performance on transformers across image and text tasks.\\n\\nWe thank the reviewer for suggesting this. As we note, our framework can easily utilize any encoder/decoder framework. In the future we will also explore transformer based AE approaches.\"}", "{\"title\": \"Response (Part 3)\", \"comment\": \"> Continuing from part 2 ...\\n\\nThe above implies the existence of a $k$-dimensional space where the points (coming from $k$ ground-truth\\nclusters) are well-separated into $k$ clusters. Thus, a latent space of $k$ dimensions is necessary to obtain\\nwell-separated clusters. As Euclidean clustering becomes more challenging with increasing representation dimensionality\\n(the representation in which the clustering is happening), the motivation is to keep the latent space dimension as low\\nas possible as long as we have enough dimensions to separate the clusters. For this reason, the latent space\\ndimensionality is most deep clustering usually matches the desired number of clusters (as long as the Euclidean\\nclustering is happening with the latent representations). A higher latent dimensionality will definitely help with the\\nreconstruction but can potentially hurt Euclidean clustering; a lower latent dimensionality would not be sufficient to\\nobtain $k$ well-separated clusters.\\nFortunately, given the extremely expressive modern deep learning encoder and decoders, we are able to still get quite low reconstruction loss with a $m=k$ dimensional latent space.\\n\\n[1] Xifeng Guo, Xinwang Liu, En Zhu, and Jianping Yin. Deep clustering with convolutional autoencoders. In International conference on neural information processing, pp. 373\\u2013382. Springer, 2017b.\\n\\n[2] Junyuan Xie, Ross Girshick, and Ali Farhadi. Unsupervised deep embedding for clustering analysis. In International conference on machine learning, pp. 478\\u2013487. PMLR, 2016.\\n\\n[3] Wengang Guo, Kaiyan Lin, and Wei Ye. Deep embedded k-means clustering. In 2021 International\\nConference on Data Mining Workshops (ICDMW), pp. 686\\u2013694. IEEE, 2021.\\n\\n[4] Amin G. Oskouei, Mohammad A. Balafar, and Cina Motamed. Edcwrn: efficient deep clustering with\\nthe weight of representations and the help of neighbors. Applied Intelligence, 53(5):5845\\u20135867,\\n2023.\\n\\n[5] Von Luxburg, Ulrike. A tutorial on spectral clustering. Statistics and computing 17 (2007): 395-416.\\n\\n> How is the number of centers (clusters) chosen? It\\u2019s unclear whether this is a fixed number or if it corresponds to the number of classes in the dataset.\\n\\nWe apologize for any confusion. The number of centers (clusters) is not a hyperparameter; they are always chosen as number of true classes in each dataset, i.e., $m=k$. We did mentioned this in Table 5 in Appendix (line 770), but we will reemphasize this in the main text.\\n\\n> Why is reconstruction considered significant in this context? It seems that the assumption of minimal information loss in the latent space is a prerequisite for effective clustering. To my understanding, a low reconstruction loss simply indicates that good clustering may be achievable, rather than serving as a goal in itself.\\n\\nWe thank the reviewer for this insightful question. We agree that the assumption of minimal reconstruction loss is a prerequisite for a good clustering in latent space.\\nHowever, there are mainly two options while clustering in latent space.\\n\\nFirst option is to pretrain the autoencoder (AE) to minimize the reconstruction loss (RL), and then fix this latent space. Next, apply some clustering scheme to group the points. In this case, we do not have to consider reconstruction loss while clustering since the AE is fixed.\\nHowever, as we demonstrate experimentally, keeping the autoencoder fixed does not lead to the best clusterings. See for example the baseline comparison in Tables 2 and 3 that employs a CAE autoencoder, followed by K-means or ClAM in latent space; $\\\\texttt{DClAM}$ outperforms all these baselines in terms of clustering quality; and in several cases it also improves on the RL compared to the fixed autoencoder!\\n\\nThe second option is to fine-tune the autoencoder (i.e., both the encoder and decoder) so that it modifies the latent space along with the task of finding clusters.\\nIn fact, if we do not consider reconstruction loss, then it means that there is effectively no decoder, since there is no RL constraint, and this can lead to trivial clusterings. As an extreme example, the encoder can map each point to its corresponding center, which would lead to a trivial clustering, but one that has a sum of squared errors loss as zero (or if we consider Silhouette Coefficient (SC), this trivial clustering will have a SC value of 1).\\n\\nIn other words, if we do want to fine-tune both the encoder and decoder, to allow more flexibility to learn a better ``clustering-guided'' latent space, then considering RL is a must; it acts as a regularization constraint on the latent space. The key insight and contribution of $\\\\texttt{DClAM}$ is that we seamlessly combine the clustering and RL objectives into one expression that tackles the task of clustering-guided latent representations, whereas previous deep clustering methods considered these separately.\"}", "{\"summary\": \"The paper applies ClAM (Saha et al. 2023) to various neural architectures (CNN, MLP) and achieves competitive performances on various text and image clustering benchmarks. The main metric of concern is the Silhouette Coefficient (SC), commonly used in clustering literature for unsupervised clustering quality.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": [\"Paper is well written and easy to follow\", \"Very clear motivation, theoretical analysis appear to be correct, core algorithm clearly explained, and experiments are presented thoroughly\"], \"weaknesses\": [\"The work builds up on ClAM (Saha et al., 2023), which may limit its novelty.\", \"Arguably, since the work is concerned with deep clustering, baseline methods should include traditionally metric learning-based approaches.\", \"In contemporary literature standards, both the neural network trained in the work and the dataset are small. It is unclear whether the proposed method would generalize to more realistic network sizes and datasets.\"], \"questions\": [\"It seems like the entire (vision) transformer literature is missing from the analysis. This is not a critical concern, but it would be interesting to evaluate DClAM's performance on transformers across image and text tasks.\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"2\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Response (Part 1)\", \"comment\": \"We thank reviewer ZsFx for their thoughtful comments and questions. Below, we aim to address each of the comments provided. We trust that our responses will assist the reviewer in reassessing their evaluation of our paper.\\n\\n> The evaluation process only included SC although the ground truth labels are available for the used datasets. The SC alone can\\u2019t be enough especially with its underlying assumptions about the cluster\\u2019s distributions. The reported SCs for all the experiments are constrained by 10\\\\% range of change in the reconstruction loss. This is limiting the space for clustering improvement during training. Which is not fair for clustering algorithms, especially those that don\\u2019t use the decoder anymore in the clustering process. It is mentioned in the paper that \\u201capplying CLAM in a latent space learned by a pretrained autoencoder is not an effective strategy\\u201d. However, the results reported don\\u2019t test this hypothesis. The proposed method requires a choice of three learning rates which are difficult to select over different applications. The belief mentioned in the evaluation section of using unsupervised metrics contradicts with using the information of number of clusters which is unknown in many unsupervised applications.\\n\\nWe thank the reviewer for valuable feedback. We address each point in detail below.\\n\\n1) We want to mention at the outset that we also evaluated the methods on the Normalized Mutual Information (NMI) metric, which measures how well the ground truth labels correspond to the clusters. We presented the best NMI (and other associated metrics) across all datasets and methods in Table 10 in the Appendix (and also mentioned this in line 414-416 in the main text). Table 10 highlights that $\\\\texttt{DClAM}$ exhibits strong performance not only in terms of unsupervised SC and RL, but also when compared to the ground truth labels via NMI. In fact, for NMI, $\\\\texttt{DClAM}$ has the best value in 5 out of the 8 datasets.\\n\\n\\n2) Regarding the comment about fairness of reporting the best SC within 10\\\\% of the RL, please note that in general there are two main approaches while clustering in latent space.\\nThe first option is to pretrain the autoencoder (AE) to minimize the reconstruction loss (RL), and then fix this latent space. Next, apply some clustering scheme to group the points. In this case, we do not have to consider reconstruction loss while clustering since the AE is fixed.\\nHowever, as we demonstrate experimentally, keeping the autoencoder fixed does not lead to the best clustering. This is shown in the baseline comparison in Tables 2 and 3 that employ a CAE autoencoder, followed by K-means or ClAM in latent space; $\\\\texttt{DClAM}$ outperforms all these baselines in terms of clustering quality; and in several cases it also improves on the RL compared to the fixed autoencoder! \\n\\nSpecifically, regarding ClAM in latent space, we did test this hypothesis and reported all the results in Tables 2, 3, 8, 9 \\\\& 10. We evaluate ClAM both in the ambient space (denoted as NAE for No AE) and in the latent space obtained through a pretrained Convolutional Autoencoder (CAE). And we found that running ClAM in the (pretrained) latent space is not an effective clustering strategy while $\\\\texttt{DClAM}$ outperforms ClAM with a very high margin utilizing its intrinsic clustering-friendly architecture. \\n\\nReturning to the second option for clustering in latent space, instead of fixing the encoding, we can fine-tune the autoencoder (i.e., both the encoder and decoder) so that is modifies the latent space along with the task of finding clusters. Here, if we do not consider reconstruction loss, then it means that there is effectively no decoder, since there is no RL constraint, and this can lead to trivial clusterings. As an extreme example, the encoder can map each point to its corresponding center, which would lead to a trivial clustering, but one that has a sum of squared errors loss as zero (or if we consider Silhouette Coefficient (SC), this trivial clustering will have a SC value of 1). In other words, if we do want to fine-tune both the encoder and decoder, to allow more flexibility to learn a better clustering-guided latent space, then considering RL is a must; it acts as a regularization constraint on the latent space. The key insight and contribution of $\\\\texttt{DClAM}$ is that we seamlessly combine the clustering and RL objectives into one expression that tackles the task of clustering-guided latent representations, whereas previous deep clustering methods considered these separately.\"}", "{\"comment\": \"Dear Reviewer ZsFx,\\n\\nWe would like to once again thank you for taking the time to review our work and for your help in improving it. We have addressed all your concerns in the revised version of our paper. We would greatly appreciate it if you could consider increasing your score for our submission.\"}", "{\"comment\": \"Dear Reviewer Z3Ff,\\n\\nWe would like to once again thank you for taking the time to review our work and your help in improving it. We have addressed all your concerns and included the following in the revised version of our paper.\\n\\n1. We have included a discussion of VQ-VAEs in the related work section. \\n2. We have also included ablation results on the effect of varying the number of AM steps $T$ in Appendix C.4 and Figure 11, for both USPS and FMNIST datasets.\\n\\nWe would greatly appreciate it if you could consider increasing your score for our submission.\"}", "{\"title\": \"Response (Part 2)\", \"comment\": \"> Continuing from part 1 ...\\n\\nFor fairness sake, we argue the opposite. Since it is not immediately clear whether SC or RL should have more weight, we report in Table 2 the best SC value obtained when ensuring that RL does not degrade by more than 10\\\\% of the initial pretrained AE. We also report in Table 3, the best RL for the methods when SC is within 10\\\\% of the best SC obtained by the method. Table 8 and 9 give more detailed results for these two cases, respectively. What we observe is that in either of these two approaches $\\\\texttt{DClAM}$ has the best results for the SC, and best RL in most of the cases.\\n\\n3) We appreciate the reviewer's concern regarding the three learning rates of $\\\\texttt{DClAM}$. We would like to emphasize that there are two main hyperparameters for $\\\\texttt{DClAM}$, namely inverse temperature ($\\\\beta$), and number of steps of AM ($T$) (line 760-761). The other hyperparameters listed in Table 5 in Appendix (762-221), such as batch size, learning rate, patience, and so on are very common for almost all deep clustering/learning schemes. Furthermore, it is clear from Table 6 in the appendix that most of these hyperparameter values can be quite stable across the datasets. For example, there is little variation in the encoder, decoder or AM learning rates. The same is true for batch size, and the number of AM steps ($T$) is usually 10-15. The only parameter that requires some tuning is $\\\\beta$, whose optimal value ranges from 0.00015 to 10. In general, the convolutional autoencoder (CAE) is less sensitive to $\\\\beta$, whereas resnet autoencoder (RAE) is more sensitive, with typically smaller values of $\\\\beta$ for larger latent dimensions ($m$), which corresponds to the number of true classes in the dataset ($k$).\\n\\n4) Finally, we agree with the reviewer that in the pure unsupervised process, the number of clusters is unknown. However, we are using this information only for how many clusters we want to find, which is also the user input required for virtually all clustering methods. However, in future work, we do plan to explore how we can leverage our approach to automatically find the true number of clusters (we also mentioned this limitation in line 485).\\n\\n> Describing the method to be agnostic to architecture can be easily misunderstood. Because architecture is important according to the type of dataset. What could be deduced is that the method itself is flexible and can be integrated with different autoencoders architectures. Evaluating the proposed method using ground truth-dependent metrics would make the results more reliable. Mitigating data leakage between training and evaluation can still be done without completely ignoring the ground truth information. The paper lacks a solid justification why balancing the SC and reconstruction loss is important.\\n\\nThanks for suggesting the point about ``agnostic''. We agree that is better to state that our approach can flexibly be combined with different AE architectures. We will rephrase this statement. \\n\\nRegarding the ground-truth-dependent metrics, we responded to this in point 2 above. If we employ autoencoder pretraining, then we can optimize for the clustering quality (such as SC) while ensuring that the reconstruction loss is within some margin (say 10\\\\%) of the reconstruction loss of the pretrained autoencoder. In our response 2) above we provide the justification why it is critical to balance the clustering quality with reconstruction loss when allowing both the encoder and decoder to evolve while clustering.\\n\\nFinally, we believe that the hyperparameters for clustering and clustering evaluation should ideally be based on unsupervised metrics, without utilizing ground-truth label information. Nevertheless, we do compare the methods on NMI too and show that $\\\\texttt{DClAM}$ retains its superiority even on this supervised metric.\"}", "{\"comment\": \"Thank you; I'm satisfied with the updated version and have adjusted my score accordingly.\"}", "{\"comment\": \"We thank reviewer whZj for their thoughtful comments and questions. In what follows, we try to address each of the comments. We hope that these help the reviewer reconsider their evaluation of our paper.\\n\\n> First and foremost, this paper resembles the CLAM paper strongly, particularly in the introduction section, which feels almost identical. Given that this work relies heavily on CLAM, it is essential to revise and reframe the introduction to clearly distinguish this approach as more than a direct application of CLAM with an autoencoder. Establishing this work's unique contributions will help clarify its originality and value.\\n\\nWe want to thank the reviewer for this valuable feedback. \\n\\nThere are fundamental differences between ClAM and $\\\\texttt{DClAM}$. The former proposes a differential clustering scheme in the input space utilizing associative memories (AM), whereas our $\\\\texttt{DClAM}$ approach focuses on clustering in latent space, utilizing AM to find good clusters and yet retaining good reconstruction (which reflects the quality of the latent representations). In fact, we do mention the key differences in detail in our submission, in lines 293-300 and 280-290.\\n\\nThe discussion in lines 293-300 points out that in ClAM, AM is utilized as a differentiable argmin solver for the $k$-means objective, whereas in $\\\\texttt{DClAM}$, which involves representation learning, AM recursion actually has a more elaborate effect. The AM augmented encoder ($A_{\\\\mathbf{\\\\rho}}^T \\\\circ \\\\mathbf{e}$) explicitly creates basins of attraction in the latent space, and moves/pushes the latent representations of the points into these basins, thereby explicitly inducing a clustered data distribution in the latent space. While the encoder is moving points into basins of attraction, the $\\\\texttt{DClAM}$ loss tries to minimize the information loss in the latent representations by having the decoder reconstruct these relocated latent representations.\\n\\nThe discussion in lines 280-290 lists the various advantages of $\\\\texttt{DClAM}$, namely (i) First, it does not involve any balancing hyperparameter $\\\\gamma$ since the loss involves all parameters in a single term in the per-sample loss $\\\\bar \\\\ell(x, \\\\mathbf{e}, \\\\mathbf{d}, {\\\\mathbf{\\\\rho}})$. (ii) Second, the updates for all the parameters in the pipeline are more explicitly tied together with the $\\\\mathbf{d} \\\\circ A_{\\\\mathbf{\\\\rho}}^T \\\\circ \\\\mathbf{e}$ composition in the $\\\\mathbf{d}( A_{\\\\mathbf{\\\\rho}}^T( \\\\mathbf{e}( x ) ) )$ term. This ties the representation learning and clustering objectives more intricately. (iii) Third, it continues to have all the advantages of \\ntraditional deep clustering, being end-to-end differentiable since all operators in the above composition are differentiable, and performing a discrete cluster center assignment with $T$ recursions of the attractor dynamics operator $A_{\\\\mathbf{\\\\rho}}$. (iv) Forth, this deep clustering can be combined with different auto-encoder architectures -- we can select a problem dependent encoder and decoder (for example, convolutional or residual networks for images or fully-connected feed-forward networks for text or tabular data).\\n(v) Fifth, it does not involve any additional entropy regularization based hyperparameters as with existing deep clustering algorithms. \\n\\nHowever, to highlight these key contributions and differences, we will certainly restructure our introduction section emphasizing these unique aspects of $\\\\texttt{DClAM}$.\", \"title\": \"Response (Part 1)\"}", "{\"comment\": \"Dear Authors,\\n\\nI am satisfied with the revised version and have adjusted my score accordingly.\"}", "{\"title\": \"Summary of Review and Response\", \"comment\": [\"We express our sincere gratitude to the Reviewers and Area Chairs for their dedicated time and constructive feedback. Here is a concise summary of the review and our responses for easy reference.\", \"**Reviewer Acknowledgments:**\", \"Our method, $\\\\texttt{DClAM}$, has been recognized for its **Novelty** and **Superior Results**. Key highlights include:\", \"**Clear Motivation and Novelty:** $\\\\texttt{DClAM}$ comes up with **very clear motivation** with precise **theoretical analysis** and explanation of the core algorithm [AgR2]. The core idea of integrating AMs into the (deep) autoencoder pipeline via a differentiable loss is **novel** and **interesting** [Z3Ff].\", \"**Superior Results, Generalizability and Reproducibility:** $\\\\texttt{DClAM}$ achieves superior results across various benchmarks with diverse domains and architectural types, effectively showcases the **generalizability** of the method [whZj]. It describes the technical details well which make it **reproducible** [ZsFx]. Moreover, $\\\\texttt{DClAM}$ is **agnostic** to architectural choices on the encoder and decoder that makes it easy to incorporate with any architecture [Z3Ff].\", \"**Quality of Writing:** The manuscript is praised as **\\\"well written\\\"** and **\\\"easy to follow\\\"** [AgR2], making it easily **accessible** to the field [whZj].\", \"**Addressing Weaknesses:**\", \"The reviewers raised concerns regarding the clarification of unique contributions of $\\\\texttt{DClAM}$ [whZj, AgR2], hyperparameter tuning [whZj, ZsFx], more ablation studies [whZj, Z3Ff], generalization [AgR2] and evaluation metrics [ZsFx]. Our responses include:\", \"*Unique contributions of $\\\\texttt{DClAM}$:* We clarified in this [comment](https://openreview.net/forum?id=Awsb8jhEx3&noteId=cs6uLwJvJB) the unique contributions of $\\\\texttt{DClAM}$ and fundamental differences from $\\\\texttt{ClAM}$ and other deep clustering methods.\", \"*Ablation studies:* Following suggestions from Reviewers whZj and Z3Ff, we have added the ablation studies by varying the **latent dimension $\\\\textbf{m}$** for USPS and varying the number of **AM steps $\\\\textbf{T}$** for both USPS and FMNIST datasets.\", \"*HP tuning and generalization:* We clarified [here](https://openreview.net/forum?id=Awsb8jhEx3&noteId=J7GWg2dRZb) that there are two main hyperparameters for $\\\\texttt{DClAM}$, namely inverse temperature ($\\\\beta$), and number of steps of AM ($T$). The number of AM steps ($T$) is usually 10-15. While $\\\\texttt{DClAM}$ requires tuning $\\\\beta$, it is a completely new strategy for deep clustering using associative memories that **outperforms** all the related baseline schemes. We clarified [here](https://openreview.net/forum?id=Awsb8jhEx3&noteId=zFRpWqtA2I) about the generalization of $\\\\texttt{DClAM}$ to any network and dataset size.\", \"*evaluation metrics:* We clarified [here](https://openreview.net/forum?id=Awsb8jhEx3&noteId=q77ZFDaKHz) and [here](https://openreview.net/forum?id=Awsb8jhEx3&noteId=ebazIWUcNv) that we also evaluated $\\\\texttt{DClAM}$ on the supervised Normalized Mutual Information (NMI) metric where $\\\\texttt{DClAM}$ **wins** for **5** out of the 8 datasets. We have also discussed why unsupervised metric like Silhouette Coefficient (SC) and reconstruction loss (RL) are necessary to evaluate unsupervised clustering methods and added new experiments on the Pareto frontier for all the hyperparameter configurations for different vision-based datasets to show how we selected best SC and RL across all datasets.\", \"The responses were **acknowledged positively** by reviewer **[whZj](https://openreview.net/forum?id=Awsb8jhEx3&noteId=c0VelNjAMy)** and reviewer **[AgR2](https://openreview.net/forum?id=Awsb8jhEx3&noteId=dF0B4gr0fw)** and they have **increased** their score to **6**. We have not received any responses from reviewer ZsFx and Z3Ff while we have addressed all of their concerns.\", \"**Revision Overview:**\", \"To enhance clarity, we have revised our introduction and related work sections, to outline the key differences between $\\\\texttt{DClAM}$ and $\\\\texttt{ClAM}$ and other deep clustering works.\", \"We have added the ablation study by varying the latent dimension $m$ for USPS in Appendix C.4, Fig 10. We have also clarified that the number of clusters $k$ is not a hyperparameter (end of sec A.2).\", \"We have added new experiments on the Pareto frontier for all the hyperparameter configurations for different vision-based datasets in Appendix C.2, and Figures 5,6,7,8,9.\", \"We have also included ablation results on the effect of varying the number of AM steps $T$ in Appendix C.4 and Figure 11, for both USPS and FMNIST datasets.\", \"We are thankful for the valuable feedback from the reviewers, and we believe we have thoroughly addressed all the concerns raised by the reviewers in our responses and revised manuscript.\"]}", "{\"summary\": \"This paper proposes DClAM, a deep clustering method that integrates the associate memory idea of ClAM into the deep learning pipeline. Their method involves fine-tuning an autoencoder model by taking multiple associative memory (AM) steps on encoded embeddings before decoding, which enables joint optimization of representation learning and clustering. The paper shows that DClAM outperforms existing deep clustering methods on eight datasets spanning image and text with multiple architectures.\", \"soundness\": \"3\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": [\"I think the core idea of integrating AMs into the (deep) autoencoder pipeline via a differentiable loss is novel and interesting. The method is agnostic to architectural choices on the encoder and decoder.\", \"The paper comprehensively evaluates on diverse datasets, comparing against a range of clustering and deep clustering baselines.\"], \"weaknesses\": [\"There's a large body of work on vector-quantized VAEs [1], which seems related to DClAM in that they're both learning an autoencoder with a latent space with a discrete structure. One obvious difference is that this model isn't a VAE (doesn't consider a prior over latents). Still, I think the learned centers play a similar role to the learned vectors in a VQ-VAE; their objective is essentially yours, with only one AM step. I think the paper should discuss the connection in more detail.\", \"The paper could benefit from more detailed ablation studies on key hyperparameters. I think reporting sensitivity to (1) the number of AM steps and (2) the number of clusters would give a better sense of the method's tuning requirements. Also, what is the number of AM steps you use? I can't find it in Table 5.\", \"What does the \\\"restart\\\" hyperparameter mean? Does it mean you ran the entire pipeline with five random seeds and reported the best performance?\", \"How does this method compare to other deep clustering methods in terms of computational efficiency? Mainly, I'd like to better understand how computationally expensive the AM steps are.\", \"[1] Van Den Oord, Aaron, and Oriol Vinyals. \\\"Neural discrete representation learning.\\\" Advances in neural information processing systems 30 (2017).\", \"Minor\", \"I think it's generally good to have the proposed method's explanation (sec 4) self-contained. A core component (eq 7) is explained separately in section 3; you might consider briefly re-stating this operator in section 4.\"], \"questions\": \"Please see questions in weaknesses section above.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper presents DClAM, a deep clustering method that builds on the previous ClAM approach by integrating it with deep clustering techniques, specifically through a deep autoencoder (AE). DClAM is architecture-agnostic, demonstrating strong performance across different autoencoder designs and dataset types. It achieves high clustering quality and low reconstruction loss in both ambient and latent spaces, surpassing previous methods.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"The paper demonstrates superior results across various benchmarks, highlighting the robustness of the proposed approach. Additionally, the authors conduct experiments across diverse domains and architectural types, effectively showcasing the generalizability of their method. The preliminary section provides a thorough overview of foundational concepts, enhancing the accessibility of the paper. Furthermore, the appendix includes detailed hyperparameter information, essential for validating experiments in this hyperparameter-sensitive area and supporting the reproducibility of the study.\", \"weaknesses\": \"First and foremost, this paper resembles the CLAM paper strongly, particularly in the introduction section, which feels almost identical. Given that this work relies heavily on CLAM, it is essential to revise and reframe the introduction to clearly distinguish this approach as more than a direct application of CLAM with an autoencoder. Establishing this work's unique contributions will help clarify its originality and value.\\n\\nWhile the authors claim that the method does not require $\\\\gamma$ tuning, it does not appear to resolve the underlying parameter sensitivity issue. In fact, it introduces additional hyperparameters that need careful adjustment. The chosen values suggest that previous tuning decisions are not reusable across different datasets and models, requiring high tuning complexity. This is particularly challenging, given that model training for reconstruction must occur simultaneously with the tuning process.\\n\\n An essential missing experiment is an ablation study on the latent dimension \\\\( m \\\\). Evaluating this parameter, even for one of the models used, would provide valuable insights into the impact of latent dimensionality on performance.\", \"questions\": \"How is the number of centers (clusters) chosen? It\\u2019s unclear whether this is a fixed number or if it corresponds to the number of classes in the dataset.\\n\\nWhy is reconstruction considered significant in this context? It seems that the assumption of minimal information loss in the latent space is a prerequisite for effective clustering. To my understanding, a low reconstruction loss simply indicates that good clustering may be achievable, rather than serving as a goal in itself.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This study presents DCLAM a deep clustering algorithm that internally uses Clustering with Associative Memory (CLAM) algorithm. The proposed method uses encoder-decoder architecture which is optimized by an associative memory-inspired loss function. Finally, the method is evaluated in comparison with existing methods using silhouette coefficient (SC) and reconstruction loss over image and text datasets.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"The research integrates autoencoder and pretraining while optimizing the data in the latent space using the CLAM algorithm. Which is a good extension to CLAM.\\nThe approach tries to maintain a minimal reconstruction loss during the process of finding clusters.\\nThe paper describes the technical details well which make it reproducible.\", \"weaknesses\": \"The evaluation process only included SC although the ground truth labels are available for the used datasets. The SC alone can\\u2019t be enough especially with its underlying assumptions about the cluster\\u2019s distributions.\\nThe reported SCs for all the experiments are constrained by 10% range of change in the reconstruction loss. This is limiting the space for clustering improvement during training. Which is not fair for clustering algorithms, especially those that don\\u2019t use the decoder anymore in the clustering process.\\nIt is mentioned in the paper that \\u201capplying CLAM in a latent space learned by a pretrained autoencoder is not an effective strategy\\u201d. However, the results reported don\\u2019t test this hypothesis.\\nThe proposed method requires a choice of three learning rates which are difficult to select over different applications.\\nThe belief mentioned in the evaluation section of using unsupervised metrics contradicts with using the information of number of clusters which is unknown in many unsupervised applications.\", \"questions\": \"Describing the method to be agnostic to architecture can be easily misunderstood. Because architecture is important according to the type of dataset. What could be deduced is that the method itself is flexible and can be integrated with different autoencoders architectures.\\nEvaluating the proposed method using ground truth-dependent metrics would make the results more reliable. Mitigating data leakage between training and evaluation can still be done without completely ignoring the ground truth information.\\nThe paper lacks a solid justification why balancing the SC and reconstruction loss is important.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"details_of_ethics_concerns\": \"No ethics review needed.\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"Dear Authors,\\n\\nThank you for addressing my concerns in your responses. While I have no further questions, I would like to review a revised version of the paper to reconsider my score.\"}", "{\"comment\": \"Dear Reviewers, Dear Area Chair,\\n\\nThanks a lot for taking the time to read our revised manuscript! We are very happy to hear that reviewers whZj and AgR2 are satisfied with our response and have updated their scores. \\n\\nThere is still one reject score (3) from reviewer ZsFx, and we have not received any response from them regarding our rebuttal and revised manuscript. Since there is less than 24 hours left for us to be able to communicate with reviewers, we kindly ask reviewer ZsFx to consider increasing the numerical score for our submission.\\n\\nWe are thankful to reviewers whZj and AgR2 for raising their scores to 6. With the decision-making process soon underway, if whZj or AgR2 believe our paper is worthy of acceptance, we kindly ask considering a further increase of the score to a \\\"full accept\\\" (8). We remain available to resolve any outstanding issues that would bump our paper from a borderline to an accept in the remaining time dedicated for discussion.\\n\\nSincerely, Authors\"}", "{\"comment\": \"Dear Reviewer whZj,\\n\\nWe sincerely appreciate your support of our paper.\"}", "{\"title\": \"Thank you!\", \"comment\": \"Thank you very much! We are very happy to hear that you are satisfied with our responses. Given that the decision-making process will soon be underway, if you believe that our paper is worthy of acceptance, we kindly ask considering a further increase of the score to a \\\"full accept\\\" (8). We remain available to resolve any outstanding issues that would bump our paper from a borderline to an accept in the remaining time dedicated for discussion.\"}", "{\"comment\": \"> While the authors claim that the method does not require $\\\\gamma$\\n tuning, it does not appear to resolve the underlying parameter sensitivity issue. In fact, it introduces additional hyperparameters that need careful adjustment. The chosen values suggest that previous tuning decisions are not reusable across different datasets and models, requiring high tuning complexity. This is particularly challenging, given that model training for reconstruction must occur simultaneously with the tuning process.\\n\\nWe appreciate the reviewer's concern regarding the hyperparameter sensitivity of $\\\\texttt{DClAM}$. We would like to emphasize that there are two main hyperparameters for $\\\\texttt{DClAM}$, namely inverse temperature ($\\\\beta$), and number of steps of AM ($T$) (line 760-761). The other hyperparameters listed in Table 5 in Appendix (762-221), such as batch size, learning rate, patience, and so on are very common for almost all deep clustering/learning schemes. \\n\\nFurthermore, it is clear from Table 6 in the appendix that most of these hyperparameter values can be quite stable across the datasets. For example, there is little variation in the encoder, decoder or AM learning rates. The same is true for batch size, and the number of AM steps ($T$) is usually 10-15. The only parameter that requires some tuning is $\\\\beta$, whose optimal value ranges from 0.00015 to 10. In general, the convolutional autoencoder (CAE) is less sensitive to $\\\\beta$, whereas resnet autoencoder (RAE) is more sensitive, with typically smaller values of $\\\\beta$ for larger latent dimensions ($m$), which corresponds to the number of true classes in the dataset ($k$).\\nIt is worth noting that while $\\\\texttt{DClAM}$ requires tuning $\\\\beta$, it is a completely new strategy for deep clustering using associative memories that outperforms all the related baseline schemes.\\n\\n> An essential missing experiment is an ablation study on the latent dimension (m). Evaluating this parameter, even for one of the models used, would provide valuable insights into the impact of latent dimensionality on performance.\\n\\nIt is important to clarify that in $\\\\texttt{DClAM}$ the latent dimensions $m$ is always set as the true number of classes per dataset, i.e., $m=k$. Indeed, most deep clustering schemes in the literature such as DCEC[1], DEC [2], DEKM[3], and EDCWRN [4] either follow this strategy or fix this to a specific number (e.g., 10), since latent representations are not only just good representations of the data points, they also represent the clusters. In $\\\\texttt{DClAM}$, we always set this latent dimension as the number of clusters. By setting this exactly the same as the number of clusters, each latent dimension should ideally represent one specific cluster. If the latent dimensions are larger than the number of clusters, some dimensions might not align with any specific cluster, or multiple dimensions could end up representing the same cluster. This can introduce redundancy and result in a less efficient representation. On the other hand, if the latent dimensions are smaller than the number of clusters, some clusters may not be adequately represented. This forces multiple clusters to share the same dimension, making it challenging for the model to distinguish between them accurately.\\n\\nWe can also consider a spectral argument [5] for setting $m=k$ .\\nGiven a set of $n$ points (in any representation) from $k$ ground-truth clusters, consider a graph (weighted or\\nunweighted and undirected) with each point as a node, and edges between points belonging to the same cluster, and no\\ninter-cluster edges. This graph would have $k$ connected components, and the Laplacian $L \\\\in \\\\mathbb{R}^{n \\\\times n}$ of\\nthis graph will have $k$ zero eigenvalues (for example, see Von Luxburg [5, Proposition 2]). Now consider the first $k$\\neigenvectors $u_1, \\\\ldots, u_k \\\\in \\\\mathbb{R}^n$ forming the columns of the matrix $U \\\\in \\\\mathbb{R}^{n \\\\times k}$. Then\\neach row $z_i \\\\in \\\\mathbb{R}^k$ of $U$ can serve as a representation of the point $i$, and the points will be\\nwell-separated into $k$ clusters in this representation. This is the intuition that forms the basis of various spectral clustering algorithms.\", \"title\": \"Response (Part 2)\"}", "{\"comment\": \"Dear Reviewer whZj,\\n\\nWe would like to once again thank you for taking the time to review our work and your help in improving it. We have addressed all your concerns in the revised version of our paper. We would greatly appreciate it if you could consider increasing your score for our submission.\"}", "{\"comment\": \"Dear Reviewer AgR2,\\n\\nWe would like to once again thank you for taking the time to review our work and your help in improving it. We have addressed all your concerns and included the fundamental differences between DClAM and ClAM, in the revised introduction and related works section of our paper. We have added the relation to metric learning too.\\n\\nWe would greatly appreciate it if you could consider increasing your score for our submission.\"}", "{\"title\": \"Response (Part 1)\", \"comment\": \"We thank reviewer AgR2 for their review and kind words of support for our paper, acknowledging its \\\"well-written\\\" and \\\"very clear motivation\\\". We address their specific comments in the sections below and wish the reviewer to reconsider their evaluation of our paper towards acceptance.\\n\\n> The work builds up on ClAM (Saha et al., 2023), which may limit its novelty.\\n\\nWe would like to emphasize that there are fundamental differences between ClAM and $\\\\texttt{DClAM}$. The former proposes a differential clustering scheme in the input space utilizing associative memories (AM), whereas our $\\\\texttt{DClAM}$ approach focuses on clustering in latent space, utilizing AM to find good clusters and yet retaining good reconstruction (which reflects the quality of the latent representations). Achieving the joint goal of clustering-guided latent representation is a novel task.\\n\\nFurthermore, we did elucidate the key differences between them in our submission. The discussion in lines 293-300 points out that ClAM, AM is utilized as a differentiable argmin solver for the $k$-means objective, whereas in $\\\\texttt{DClAM}$, which involves representation learning, AM recursion actually has a more elaborate effect. The AM augmented encoder ($A_{\\\\mathbf{\\\\rho}}^T \\\\circ \\\\mathbf{e}$) explicitly creates basins of attraction in the latent space, and moves/pushes the latent representations of the points into these basins, thereby explicitly inducing a clustered data distribution in the latent space. While the encoder is moving points into basins of attraction, the $\\\\texttt{DClAM}$ loss tries to minimize the information loss in the latent representations by having the decoder reconstruct these relocated latent representations.\\n\\nThe discussion in lines 280-290 lists the various advantages of $\\\\texttt{DClAM}$, namely (i) First, it does not involve any balancing hyperparameter $\\\\gamma$ since the loss involves all parameters in a single term in the per-sample loss $\\\\bar \\\\ell(x, \\\\mathbf{e}, \\\\mathbf{d}, {\\\\mathbf{\\\\rho}})$. (ii) Second, the updates for all the parameters in the pipeline are more explicitly tied together with the $\\\\mathbf{d} \\\\circ A_{\\\\mathbf{\\\\rho}}^T \\\\circ \\\\mathbf{e}$ composition in the $\\\\mathbf{d}( A_{\\\\mathbf{\\\\rho}}^T( \\\\mathbf{e}( x ) ) )$ term. This ties the representation learning and clustering objectives more intricately. (iii) Third, it continues to have all the advantages of \\ntraditional deep clustering, being end-to-end differentiable since all operators in the above composition are differentiable, and performing a discrete cluster center assignment with $T$ recursions of the attractor dynamics operator $A_{\\\\mathbf{\\\\rho}}$. (iv) Forth, this deep clustering can be combined with different auto-encoder architectures -- we can select a problem dependent encoder and decoder (for example, convolutional or residual networks for images or fully-connected feed-forward networks for text or tabular data).\\n(v) Fifth, it does not involve any additional entropy regularization based hyperparameters as with existing deep clustering algorithms. \\n\\nTo highlight these key contributions and differences, we will certainly restructure our introduction section emphasizing these unique aspects of $\\\\texttt{DClAM}$.\\n\\n> Arguably, since the work is concerned with deep clustering, baseline methods should include traditionally metric learning-based approaches.\\n\\nWe want to mention that in our study, we prioritized baseline methods that align closely with the primary focus of our work, namely deep clustering models that jointly learn representations and cluster assignments. Metric learning\\nrefers to the task of learning the ``similarity'' between pairs of points. In an unsupervised setting the complexity of this task is at least $O(|S|^2)$ where $|S|$ is the dataset size (number of points). Therefore, metric learning is typically performed either in a supervised setting where the label is known and we learn a distance metric that maximizes similarity between points in the same class and minimize it between points across classes. Alternatively, it can be done in a weakly supervised setting, where pairs of close and far away points are given as positive and negative sets, with the goal of learning a distance metric that puts positive pairs close together and negative pairs far away.\\n\\nAs such metric learning task is not directly comparable to the objectives of unsupervised deep clustering, with the aim of learning a latent representation and clustering at the same time.\\nIf the reviewer has any specific metric learning approach in mind, we will be happy to discuss that.\"}", "{\"title\": \"Paper Decision\", \"decision\": \"Reject\"}", "{\"comment\": \"Dear Reviewer,\\n\\nWe have submitted a revised version of our paper, and we cordially request you to take a review of this revised version to reconsider our paper toward acceptance. Thank you for your kind support of our paper.\"}", "{\"title\": \"Response\", \"comment\": \"We want to thank Reviewer Z3Ff for their support of our paper, highlighting how our method is novel and interesting. We respond to specific comments in the following:\\n\\n> There's a large body of work on vector-quantized VAEs [1], which seems related to DClAM in that they're both learning an autoencoder with a latent space with a discrete structure. One obvious difference is that this model isn't a VAE (doesn't consider a prior over latents). Still, I think the learned centers play a similar role to the learned vectors in a VQ-VAE; their objective is essentially yours, with only one AM step. I think the paper should discuss the connection in more detail.\\n\\nWe thank the reviewer for this insightful comment. There is indeed a connection between $\\\\texttt{DClAM}$ and VQ-VAE in that both of these models learn distinct embedding vectors -- K embedding vectors as discrete latent space for VQ-VAE, and k memories for $\\\\texttt{DClAM}$. However, there are fundamental differences between them. First, in VQ-VAE, every data point is mapped to one of the stored K (typically 512; and much larger than the number of clusters k) discrete embedding vectors (dictionary mapping), and then this mapped embedding vector is passed through the decoder to construct the original data point to calculate reconstruction loss. In contrast, in $\\\\texttt{DClAM}$, the original data point is not dictionary mapped to one of the stored memories, but rather it moves towards the closest memories after $T$ steps of AM recursion and then it is passed through the decoder. Second, in VQ-VAE, it is not possible to train the original architecture with standard backpropagation as the gradient would not follow through the argmin operation. To overcome this issue, they try to approximate the gradient using straight-through estimator approach (i.e., copying it from the decoder input to encoder output), which loses the ability to minimize the expected loss function. In contrast, in $\\\\texttt{DClAM}$ the AM steps are fully differentiable and can optimize the loss function with full control. Third, VQ-VAE consists of three loss terms (as it loses control of the end-to-end differentiability) whereas $\\\\texttt{DClAM}$ has only one unified loss term. Finally, being a generative model, VQ-VAE focuses on the generation of new data points, whereas $\\\\texttt{DClAM}$ focuses on learning a clustering-friendly autoencoder. As suggested by the reviewer, we will add this discussion in the related work section in our revision. \\n\\n[1] Van Den Oord, Aaron, and Oriol Vinyals. \\\"Neural discrete representation learning.\\\" Advances in neural information processing systems 30 (2017).\\n\\n> The paper could benefit from more detailed ablation studies on key hyperparameters. I think reporting sensitivity to (1) the number of AM steps and (2) the number of clusters would give a better sense of the method's tuning requirements. Also, what is the number of AM steps you use? I can't find it in Table 5.\\n\\nWe thank the reviewer for this valuable feedback. For the number of AM steps, they are reported in Table 6 across all the datasets and architectures and are usually between 10-15. We plan to include ablation results on varying the AM steps in our revision.\\n\\nRegarding the number of clusters, it is not a hyperparameter. The number of clusters is chosen as the number of true classes as per dataset. We mentioned it in Table 5 in Appendix (line 770), however, we will certainly include this in the main text to remove the confusion.\\n\\n> What does the \\\"restart\\\" hyperparameter mean? Does it mean you ran the entire pipeline with five random seeds and reported the best performance?\\n\\nYes, the reviewer is correct.\\n\\n> How does this method compare to other deep clustering methods in terms of computational efficiency? Mainly, I'd like to better understand how computationally expensive the AM steps are.\\n\\n$\\\\texttt{DClAM}$ (Algorithm 1) has a complexity of $O(dkT N|S|)$, where $d$ is the latent dimension, $k$ is the number of clusters, $T$ is the number of AM steps, $N$ is the number of epochs and $|S|$ is the size of the dataset. Thus, the runtime complexity is linear in the number of AM steps $T$, and it is usually small (10-15 in Table 6 in Appendix). \\n\\n> I think it's generally good to have the proposed method's explanation (sec 4) self-contained. A core component (eq 7) is explained separately in section 3; you might consider briefly re-stating this operator in section 4.\\n\\nThanks for the suggestion. We will incorporate this in our revision.\"}", "{\"title\": \"General response (with revised version)\", \"comment\": \"We sincerely thank all reviewers for their valuable comments, which we\\nhave thoroughly addressed in our responses below, as well as in the revised\\nmanuscript that has been updated on openreview. All changes have been\\nhighlighted in blue in the main text and the Appendix. We summarize our main\\nchanges below.\\n\\nAs requested by reviewer whZj we have revised the introduction and related\\nwork sections, to outline the key differences between DClAM and ClAM and\\nother deep clustering works. Likewise, we have added the ablation study by\\nvarying the latent dimension $m$ for USPS in Appendix C.4, Fig 10. We have\\nalso clarified that the number of clusters $k$ is not a hyperparameter (end of\\nsec A.2).\\n\\nFor reviewer AgR2, we note the fundamental differences between DClAM and\\nClAM, as now done in the revised introduction and related works section. We\\nhave added the relation to metric learning too.\\n\\nFor reviewer ZsFx, we have added new experiments on the Pareto frontier for\\nall the hyperparameter configurations for different vision-based datasets in\\nAppendix C.2, and Figures 5,6,7,8,9. We also show the Pareto optimal\\nparameters reported for the methods for the best SC and RL, and within the\\n10% thresholds, as reported in Tables 2,3,4 and Tables 8,9. These results\\nhighlight how we thoroughly optimize the hyperparameters, and how we select\\nthe final Pareto optimal performance values from the Pareto front to be\\nconsistent and fair across all methods. The results clearly indicate that\\nDClAM offers the best clustering performance in terms of SC, as well as\\nhaving low reconstruction loss. It also performs very well on the\\nsupervised NMI metric. In fact, for NMI, DClAM has the best value in 5 out\\nof the 8 datasets.\\n\\nFinally, for reviewer Z3Ff, we include a discussion of VQ-VAEs in the related\\nwork section. We also include ablation results on the effect of varying\\nthe number of AM steps $T$ in Appendix C.4 and Figure 11, for both USPS and\\nFMNIST datasets.\\n\\nWe request the reviewers to reconsider their evaluation of our paper in\\nlight of these revisions.\"}", "{\"metareview\": \"This paper proposes a method for deep clustering based on an autoencoder model which applies the gradient descent step of an associative memory clustering method (ClAM) in the latent space of the encoder. The reviewers are largely borderline, along with one more negative review. Many reviewers note that the method appears to be largely similar to ClAM with the main distinction being incorporating the ClAM gradient descent operator in the latent space of an autoencoder. The high level approach of incorporating a clustering metric on the latent space of an autoencoder is a well-known strategy in the deep clustering literature (though it also can often result in ill-posed formulations as well if not done carefully). Here the authors argue that their method enjoys the advantage of not requiring a balancing between the sum of an autoencoder reconstruction loss and a clustering loss in the latent space as their method instead directly employs the gradient descent operator of ClAM in the latent space to drive latent points towards cluster centers. However, while this does eliminate the need for balancing a weighting hyperparameters between two losses, one is still required to choose a hyperparameter for the number of time steps (T) used in applying the ClAM gradient descent steps. The authors also provide an experimental evaluation of their approach which gives reasonable results, but some reviewers note that the datasets used are relatively small by modern standards.\\n\\nOverall, the reviewers are largely lukewarm on the paper, and I am unfortunately inclined to agree with the critiques of the reviewers and recommend rejection. Using an autoencoder with a clustering promoting loss or operator in the latent space is well established in the deep clustering literature. To meet the bar for publication I would expect to see either a well motivated formulation along with strong theoretical analysis as to the merits of the formulation or convincing experiments on large scale datasets. Here the authors have taken steps in both directions, but neither has led to the reviewers being particularly convinced about the approach. I would encourage the authors to consider developing more extensive theoretical motivations for their proposed approach or evaluating the approach on larger scale datasets and look forward seeing an improved version of the work in future meetings.\", \"additional_comments_on_reviewer_discussion\": \"The authors were largely responsive to the critiques of the reviewers in their rebuttal, leading to the reviewers who did respond to raise the scores, but not of the reviewers raised their scores to the point of arguing strongly for acceptance.\"}" ] }

[ "Shu Zou", "Xinyu Tian", "Qinyu Zhao", "Zhaoyuan Yang", "Jing Zhang" ]

[ "Out-of-distribution detection", "Vision-Language Models" ]

{ "note_id": [ "zcmq24nCnu", "yQr14A4amg", "vvtAwZLbDX", "pVpvK3A1mP", "X4325FMbhw", "TxvuBYYHnb", "7XjxQjvqVx" ], "note_type": [ "official_review", "official_review", "official_review", "official_review", "comment", "official_review", "official_review" ], "note_created": [ 1730353938230, 1730100119891, 1730631516312, 1730291049521, 1731434300410, 1730277646063, 1730744578026 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission6402/Reviewer_b5Cd" ], [ "ICLR.cc/2025/Conference/Submission6402/Reviewer_Dumo" ], [ "ICLR.cc/2025/Conference/Submission6402/Reviewer_tfLP" ], [ "ICLR.cc/2025/Conference/Submission6402/Reviewer_Yowj" ], [ "ICLR.cc/2025/Conference/Submission6402/Authors" ], [ "ICLR.cc/2025/Conference/Submission6402/Reviewer_FJor" ], [ "ICLR.cc/2025/Conference/Submission6402/Reviewer_uWWf" ] ], "structured_content_str": [ "{\"summary\": \"This paper investigates the ability of image-text comprehension among different semantic-related ID labels in VLMs and proposes a novel post-hoc strategy called SimLabel.\\n\\nSimLabel enhances the separability between ID and OOD samples by establishing a more robust image-class similarity metric that considers consistency over a set of similar class labels.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. The introduced different strategies for generating similar labels is interesting and the authors provide comprehensive insights.\\n\\n2. The paper learns a robust and discriminative image-class matching score, potentially improving visual classification ability.\", \"weaknesses\": \"1. Too strong assumption: the work is based on the assumption that ID sample should consistently have high similarity scores across similar ID classes, which is not often the case, and even worse when encounted datasets with domain gaps.\\n\\n2. Lack of comparisons and baselines. The comparion of baseline is scarace and the datasets are simple: mcm is not the only post-hoc method and no comparsion with complex OOD setting are displayed.\\n\\n3. Lack of comparisons and baselines. The comparion of baseline is scarace and the datasets are simple: mcm is not the only post-hoc method and no comparsion with complex OOD setting are displayed.\\n\\n4. Efficiency: since the method involves generation procedure, e,g. using the external large language models/world knowledge in generating similar classes and so on ( line 156-161), too many stored knowledge would slow down the efficiecny. I'm curious about the efficiency (real-time throughput and gpu memory consumption) of the method compared with existing method.\", \"questions\": \"1. Even MCM serves as the common baseline for zero-shot OOD detection, to my humble understanding, many existing few-shot methods with designed OOD score also suit for post-hoc framework[1, 2, 3], which should also be included for comprehensive coparison.\\n\\n2. Datasets is not diversity enough. The assumption proposed in line78-79 is not often the case, and even worse when encounted datasets with domain gaps, where OOD datasets are also similar to ID classes. Incorporating more OOD datasets with domian gaps including but not limited to ImageNet-R would strengthen the assumption.\\n\\n[1] CLIPN for Zero-Shot OOD Detection: Teaching CLIP to Say No. ICCV2023.\\n\\n[2] Enhancing Outlier Knowledge for Few-Shot Out-of-Distribution Detection with Extensible Local Prompts.\\n\\n[3] Learning Transferable Negative Prompts for Out-of-Distribution Detection CVPR2024.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper addresses the challenge of OOD detection, critical for preventing errors in machine learning applications, particularly in safety-sensitive areas. Existing methods using VLMs improve OOD detection by leveraging class-specific textual information. However, they focus mainly on the similarity between a new sample and each ID class, often neglecting the broader context of similar classes. To address this, the authors propose a novel approach, SimLabel, which enhances OOD detection by utilizing a robust image-class similarity metric that consistently evaluates similarity across related classes. Experimental results highlight SimLabel\\u2019s superior performance in zero-shot OOD benchmarks, underscoring its effectiveness in enhancing the separability of ID and OOD samples and promoting more reliable OOD detection.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. The paper is well-written and easy to understand. The framework diagram and visualization results are helpful for understanding the method.\\n2. The paper proposes multiple feasible concrete solutions based on a single concept, providing a comprehensive methodology section.\", \"weaknesses\": \"1. The performance reported in this paper is poor, showing a significant gap from current state-of-the-art (SOTA) methods. Specifically, its FPR95 metric on the ImageNet-1k benchmark is only 36.46%, whereas existing methods, such as CLIPN and LSN, have achieved around 30%, and NegLabel has reached 25%. Although the proposed method shows improvements over older baselines, this does not sufficiently demonstrate its effectiveness or make a meaningful contribution to the field. While the authors emphasize that NegLabel uses additional textual information, utilizing a public vocabulary is a standard approach for zero-shot tasks. My personal recommendation is that the authors should validate the method's effectiveness on more recent approaches before resubmitting. In its current form, the paper does not meet the standards for acceptance at ICLR.\\n2. There appears to be a citation error in Line 519 of the paper; CLIPEN (which might be intended as CLIPN) is not from Dai et al., 2023.\", \"questions\": \"See Weakness.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper presents a post-hoc framework for OOD detection with CLIP, SimLabel, which is based on a basic assumption that is an ID sample should consistently have high similarity scores across similar ID classes. The authors introduce three strategies for selecting similar labels. Experiments on several benchmarks show improvement over the baselines.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"Overall, the motivation and contribution of this work is clearly presented, understandable and consistent with the intuition. Experiment comparison is comprehensive.\", \"weaknesses\": \"1) In Figure 1 (b), If the model predicts a low probability on other similar categories, then why is the probability high only on the specific category (Egyptian Cat e.g.), and what is the logic behind this?\\n2) Experimental analysis needs to be added to the experimental results in Table 1, e.g., explanation for the inferior results of this paper's method on the SUN and Textures datasets, and explanation for the difference in results between the three variants.\\n3) In Figure 4., Why does the FPP increase first when K increases on Places and SUN, while the rest decreases directly?\", \"questions\": \"Please refer to the above section.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper explores the capacity of VLMs for image-text comprehension among closely related ID labels and introduces a novel post-hoc approach called SimLabel. SimLabel strengthens the distinction between ID and OOD samples by establishing a more resilient image-class similarity metric that takes into account the consistency across similar class labels. Extensive experiments validate SimLabel\\u2019s superior performance on various zero-shot OOD detection benchmarks, highlighting its effectiveness in achieving robust OOD detection.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"1. The motivation and paper writing are clear.\\n2. The experimental analysis is sufficient. \\n3. The method is simple and easy to implement.\", \"weaknesses\": \"1. The motivation does not convince me: the ood samples always have the similar appearance with the ID samples. Therefore, their affinity to similar classes should be the approximately the same to ID samples.\\n2. The OOD example shown in Figure 2 is not the typical sample in current stage, which should be visually similar the ID classes. \\n3. The ID dataset in experiment section is limited (only imagenet-1k). \\n4. The ID accuracy is not provided.\", \"questions\": \"My main concern is the motivation: See the weakness 1, Why the affinity to similar classes can be helpful to separate the OOD from ID ? I think it can only be helpful to the easy case but not the hard case in OOD detection (visually similar case) ?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"withdrawal_confirmation\": \"I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors.\", \"comment\": \"Thanks for the reviewers and chair.\"}", "{\"summary\": \"This paper introduces SimLabel, a novel post-hoc strategy that enhances the separability of ID and OOD samples by developing a more robust similarity metric that accounts for consistency across semantically related ID classes. By investigating the image-text comprehension capabilities of VLMs, SimLabel significantly improves OOD detection performance. Extensive experiments across various zero-shot benchmarks highlight its effectiveness, showcasing a promising advancement in achieving robust OOD detection.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"Strengths\\uff1a\\n1. The research contributes to the advancement of zero-shot learning techniques, showcasing how they can be applied effectively to OOD detection scenarios. Besides, it explores the robustness of zero-shot approaches in varied settings, providing insights into how these methods can maintain performance across different data distributions.\\n2. Despite the performance limitations, the results are presented in a clear and organized manner, allowing readers to understand the implications and significance of the findings within the broader context of the field.\\n3. The article includes a comprehensive experimental analysis, allowing for a better understanding of the strengths and weaknesses of the proposed method compared to traditional approaches.\", \"weaknesses\": \"1. The lack of comparison with several state-of-the-art methods that demonstrate superior performance, such as CLIPN [1], LSN [2], NegLabel [3], and CSP [4], represents a significant shortcoming of this study. The relatively low performance is a primary flaw of the article.\\n\\n[1] CLIPN for zero-shot OOD detection: Teaching CLIP to say no. ICCV, 2023.\\n[2] Out-of-distribution detection with negative prompts. ICLR, 2024.\\n[3] Negative label guided OOD detection with pretrained vision-language models. ICLR, 2024.\\n[4] Conjugated semantic pool improves OOD detection with pretrained vision-language models. NeurIPS, 2024.\\n\\n2. It is recommended to include the performance of the proposed method under different CLIP models to achieve a more comprehensive evaluation of its effectiveness.\\n\\n3. It seems that SimLabel-H and SimLabel-L may not ensure that the similar labels obtained for the ID samples are not from the OOD category. I am unsure if my understanding is correct. If this is indeed the case, I would appreciate it if the authors could provide some clarification regarding the rationale behind the method.\", \"questions\": \"Please refer to the weakness part.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"This paper introduces a novel post-processing approach called SimLabel, which leverages pre-trained visual-language models (VLMs) for robust anomaly detection. SimLabel utilizes the image-text understanding capabilities of VLMs and emphasizes the consistency of semantic similarity among class labels in the training set, rather than solely relying on predicted similarities for these labels. This approach allows for effective differentiation between in-distribution (ID) and out-of-distribution (OOD) samples. The authors propose three strategies for selecting high-quality, semantically similar class labels: utilizing hierarchical text structures, prompt learning with large language models, and pseudo image-text pairing. Extensive experiments show that SimLabel achieves impressive results across various zero-shot OOD detection benchmarks.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": [\"SimLabel harnesses the image-text understanding abilities of pre-trained VLMs, focusing on the semantic consistency among training set class labels rather than relying exclusively on single-class label predictions. This novel approach of leveraging inter-class semantic similarity sets it apart from previous methods, enhancing its capacity to distinguish ID and OOD samples.\", \"The paper presents three distinct strategies for selecting semantically relevant class labels, including hierarchical text structure utilization, prompt-based learning with large language models, and pseudo image-text pairing. These strategies facilitate the selection of high-quality labels from various perspectives, which strengthens the performance of the SimLabel method.\", \"The paper clearly explains the underlying principles and implementation details of SimLabel and thoroughly validates its effectiveness through extensive experiments. Results across multiple zero-shot OOD detection benchmarks highlight SimLabel's excellent performance, underscoring the method's advantages.\"], \"weaknesses\": [\"SimLabel\\u2019s performance may suffer in long-tail distribution scenarios, where selecting effective similar classes for tail classes becomes challenging. This limitation could reduce its effectiveness in zero-shot OOD detection tasks. One possible improvement could involve expanding tail classes by introducing additional sibling or subclass labels, although careful design would be needed to avoid introducing noise.\", \"The current implementation of SimLabel assumes equal contribution weights for each similar class when calculating the similarity between images and class labels. In practice, however, the semantic similarity between different classes varies, so a uniform weighting may not capture the full depth of semantic information. Future work could explore refined measures of similarity between classes to improve performance.\"], \"questions\": \"Please refer to the Weaknesses section.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}" ] }

[ "Yuki Takezawa", "Sebastian U Stich" ]

[ "decentralized learning" ]

{ "note_id": [ "w8BI0HZWw1", "tDMhBgwZl5", "oNF5Q3EkMs", "nAMoMSyAon", "l0rDFkerZc", "ilNH3G2sNU", "iXW2s9NnoY", "hzDOk2EDm6", "eyCH0egM1p", "cocEV9e8cH", "UCD00gRTd7", "M5WZA4onf0", "HWA8OVC6IG", "FyugfG5kEe", "Cgv39v4Ee0", "BWJp9uhxF7", "BMdNalsIu6", "8r5Lc9qI9n", "5nV4aXcax4", "50jTYEX244", "33iZdD84xb" ], "note_type": [ "decision", "official_comment", "official_comment", "official_comment", "official_review", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "meta_review", "official_comment", "official_comment", "official_review", "official_comment", "official_comment", "official_comment", "official_comment", "official_comment", "official_review", "official_comment" ], "note_created": [ 1737523593413, 1732310636541, 1732528358569, 1732896713205, 1730697856687, 1732117536854, 1730698810980, 1732285348449, 1733019420541, 1732025552360, 1735185196263, 1732009927119, 1732285415691, 1730872287770, 1732292549701, 1732291475118, 1732025594052, 1732466980212, 1732540240462, 1730306524374, 1732117495830 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Program_Chairs" ], [ "ICLR.cc/2025/Conference/Submission3738/Reviewer_bkGn" ], [ "ICLR.cc/2025/Conference/Submission3738/Authors" ], [ "ICLR.cc/2025/Conference/Submission3738/Authors" ], [ "ICLR.cc/2025/Conference/Submission3738/Reviewer_bkGn" ], [ "ICLR.cc/2025/Conference/Submission3738/Authors" ], [ "ICLR.cc/2025/Conference/Submission3738/Reviewer_AkDU" ], [ "ICLR.cc/2025/Conference/Submission3738/Authors" ], [ "ICLR.cc/2025/Conference/Submission3738/Reviewer_Neby" ], [ "ICLR.cc/2025/Conference/Submission3738/Authors" ], [ "ICLR.cc/2025/Conference/Submission3738/Area_Chair_BA24" ], [ "ICLR.cc/2025/Conference/Submission3738/Authors" ], [ "ICLR.cc/2025/Conference/Submission3738/Authors" ], [ "ICLR.cc/2025/Conference/Submission3738/Reviewer_Neby" ], [ "ICLR.cc/2025/Conference/Submission3738/Authors" ], [ "ICLR.cc/2025/Conference/Submission3738/Reviewer_AkDU" ], [ "ICLR.cc/2025/Conference/Submission3738/Authors" ], [ "ICLR.cc/2025/Conference/Submission3738/Reviewer_67Bz" ], [ "ICLR.cc/2025/Conference/Submission3738/Authors" ], [ "ICLR.cc/2025/Conference/Submission3738/Reviewer_67Bz" ], [ "ICLR.cc/2025/Conference/Submission3738/Authors" ] ], "structured_content_str": [ "{\"title\": \"Paper Decision\", \"decision\": \"Accept (Poster)\"}", "{\"title\": \"Thanks for the rebuttal.\", \"comment\": \"Thanks for the responses. Most of my questions/concerns have been addressed so I increase my point to 6.\\n\\nHere are some further thoughts.\\n\\n1. I understand that in the convergence rates, the first term decreases as $n$ increases, while the remaining terms would increase if $n$ is further increased. My point is that not every value of $n$ will decrease the convergence rate. Within a certain range (when the term $1/\\\\sqrt{nT}$ dominates), a larger $n$ might actually accelerate the convergence rate, leading to a linear speedup.\\n\\n2. I suggest that the authors add a discussion on the lower bound in the revised paper.\\n\\n3. It would also be beneficial to expand on why TELEPORTATION can alleviate this degradation in the discussion that follows.\"}", "{\"comment\": \"We thank the reviewer for his/her positive feedback.\\nWe promise to include the above discussion in the revised manuscript to make this paper more intuitive for the reader.\"}", "{\"comment\": \"We appreciate your efforts in reviewing our paper.\\nThe rebuttal deadline is approaching. We would like to revise our manuscript to ensure that our claims are accurately presented and that readers do not feel that our claim is overstated. We would greatly appreciate it if you could kindly provide feedback at your convenience.\\nIf there are any further questions, we would be happy to address them.\"}", "{\"summary\": \"This paper introduces a method called TELEPORTATION aimed at improving decentralized learning, particularly decentralized stochastic gradient descent (SGD).\\n\\nTELEPORTATION is have a better dependence on the spectral gap of the communication graph in the convergence rate. So, the authors claim that their method can alleviate the convergence rate degradation while maintaining communication efficiency. \\n\\nThey also propose a method to optimize the number of active nodes so as to fasten the convergence rate. \\n\\nThe experiments suggest that TELEPORTATION outperforms decentralized SGD in both convergence speed and stability, especially when data distribution is uneven across nodes.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1. **Reduced Communication Overhead**:\\n\\nThe paper proposes a practical solution to reduce the communication costs in decentralized learning by activating fewer nodes. The methods developed to tune the hyperparameters would inspire future work.\\n \\n2. **Clear Theoretical Analysis**: \\n\\nThe paper provides solid theoretical support for the claims, including convergence rate bounds and proofs for the proposed method. \\n\\n3. The paper is mostly well-written and easy to follow.\", \"weaknesses\": \"1. **Lack of Clarity in Problem Formulation**:\\n The paper could benefit from a clearer and more formal description of the problems posed by decentralized learning, particularly why more nodes lead to degraded performance. Without this clarification, it becomes difficult to fully appreciate the novelty and impact of TELEPORTATION.\\n\\n2. **Loose Theoretical Guarantees**: \\n\\n Beyond Proposition 1, the paper could be more rigorous in establishing the tightness of the bounds used to justify the method's benefits. A deeper exploration of whether these bounds hold tightly in practical cases (e.g., through empirical validation or consider quadratic cases) would enhance confidence in the theoretical contributions.\\n\\n3. **Limited Practical Applicability**:\\n\\n The assumption that any two nodes can communicate directly (as mentioned in the abstract and discussions) is overly idealized. In many real-world applications, such as wireless sensor networks or geographically distributed systems, this assumption may not hold.\", \"questions\": \"1. **Convergence and the Number of Nodes**:\\n I am unsure why decentralized learning methods should necessarily converge more slowly as the number of nodes increases (line 53-line 60). In fact, in fully synchronized systems, having more nodes typically leads to a linear speedup in convergence. It would be helpful if the authors could clearly formulate why this slowdown happens in decentralized settings and explicitly highlight the limitations of existing methods. A more specific problem formulation, accompanied by examples illustrating why current approaches fail to address this, would make the contribution clearer. (e.g., using a table summarizing existing results).\\n\\n2. **Discussion on Proposition 1**: \\n\\n The discussion in lines 140-148 seems to rely on the upper bound from Proposition 1. However, this argument feels weak, as the tightness of the upper bound is never discussed. If the bound is loose due to flaws in the analysis, the subsequent discussion would be rendered meaningless. Is there any way to demonstrate that the bound is tight? For instance, you could explore the behavior of a simple quadratic function to provide more concrete support for this bound.\\n\\n3. **Explanation of Theoretical Improvements**: \\n\\n It appears that the theoretical improvement primarily stems from the dependence on $p_n$. Could the authors clarify the source of this improvement? Is it driven by a more refined analysis, or is it due to a novel aspect of the algorithm itself? Providing a brief but clear explanation of what underlies this theoretical gain would strengthen the paper.\\n\\n4. **Unclear Impact of Hyperparameter Tuning**:\\n\\n The proposed hyperparameter-tuning method for selecting the number of active nodes adds another layer of complexity to the method. However, the impact of this additional tuning on overall training time and resource consumption is not fully explored. In practice, the benefits of communication reduction could be outweighed by the cost of hyperparameter search. A detailed examination of this trade-off would provide more practical insight.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Official Comment by Authors (2/2)\", \"comment\": \"> Will it make sense to change the number of activated nodes in every iteration?\\n\\nTo simplify our analysis, we considered the setting where the number of active nodes is constant.\\nIf we consider a setting where the number of active nodes can change in every iteration, we might further improve the convergence rate of TELEPORTATION.\\nWe believe that it is one of the most promising directions for future research.\\n\\n> In the current proposed algorithm, data heterogeneity is not taken into account. Some discussions on how the algorithm might need to be modified in the presence of data heterogeneity would be appreciated. [...]\\n\\nMany existing papers tried to make DSGD robust to data heterogeneity by using variance reduction,\\nproposing decentralized learning methods whose convergence rates are independent of data heterogeneity $\\\\zeta$ [1,2,3].\\nSimilarly, it may be possible to make the convergence rate of TELEPORTATION independent of data heterogeneity by leveraging variance reduction. \\nHowever, we believe that making TELEPORTATION robust to data heterogeneity is out of scope and leave it for future research.\\n\\n\\n> Although less related, in light of the popularity of federated learning literature, the privacy of activated nodes might be an important criteria to consider. Since each node knows which node participated in the previous iteration, this can lead to some degree of privacy leakage, that needs to be quantified. [...]\\n\\nDifferential privacy is the most common technique for ensuring privacy in federated learning literature.\\nIn the following, we briefly discuss the noise to be added to implement differential privacy.\\n\\nAs the reviewer pointed out, TELEPORTATION may necessitate adding substantial noise since only a few nodes are activated.\\nHowever, since TELEPORTATION can converge faster than DSGD, it may allow us to reduce the required noise.\\nDifferential privacy requires significant noise as the number of iterations increases [4]. \\nTherefore, it is not trivial to draw a conclusion here, and we would like to leave it for future research.\\n\\n\\n## Reference\\n[1] Pu et. al., Distributed stochastic gradient tracking methods, In Mathematical Programming 2021\\n\\n[2] Yuan et. al., Exact diffusion for distributed optimization and learning\\u2014Part I: Algorithm development, In IEEE Transactions on Signal Processing 2018\\n\\n[3] Tang et. al., $D^ 2$: Decentralized training over decentralized data, In International Conference on Machine Learning 2018\\n\\n[4] Abadi et. al., Deep learning with differential privacy, In ACM Conference on Computer and Communications Security 2016\"}", "{\"summary\": \"This paper proposes a novel decentralized algorithm designed to perform well when the number of nodes $n$ is very large. The proposed algorithm synchronizes a smaller set of local states $\\\\\\\\{z_i\\\\\\\\}_{i \\\\in \\\\\\\\{1,...,k\\\\\\\\}}$ for $k < n$ by running a single step of Decentralized SGD (DSGD) [Lian et al., 2017] on a subgraph of $k$ nodes, and then transfer the local states to another set of $k$ nodes for computing the gradients of other local objectives. Such algorithm is suitable for fully-connected distributed system. The proposed theorem suggests that this algorithm enjoys linear speedup by the total number of nodes $n$, at the cost of consuming $k$ local gradient computations only.\", \"soundness\": \"2\", \"presentation\": \"4\", \"contribution\": \"2\", \"strengths\": [\"The idea of sacrafising gradient updates of the whole network in favor of maintaning lower consensus error is novel.\"], \"weaknesses\": \"- In line 268, $k$ can be simplified as\\n$$\\nk = \\\\max\\\\left\\\\\\\\{ 1, \\\\min \\\\left\\\\\\\\{\\\\left\\\\lceil \\\\left(\\\\frac{T(\\\\sigma^2 + \\\\zeta^2)}{Lr_0} \\\\right)^{1/7} \\\\right\\\\rceil, n \\\\right\\\\\\\\} \\\\right\\\\\\\\}\\n$$\\nbecause $a^{1/7} < a^{1/5}$ for any $a > 1$.\\n- The main claim of Theorem 2 might be theoretically flawed and deserves more attention, which relates to the $n$ linear speedup results. More discussion is provided in the section Questions: (**About the proof of Lemma 9**).\", \"questions\": [\"(**About experiment**)\", \"I suggest the authors to plot the convergence of consensus error in the experiments as well. This can give a clearer picture to whether the $k$ nodes interaction in the proposed algorithm achieve faster convergence in consensus error than the $n$ nodes interaction in Decentralized SGD.\", \"(**About the proof of Lemma 9**)\", \"Let's make it clear when does the dominance order changes in (14) for different values of $k$ by the following statements:\", \"$$ \\\\sqrt{A/k} \\\\leq Bk^{2/3} \\\\Leftrightarrow k \\\\ge (A^3 / B^6)^{1/7} \\\\quad \\\\cdots (*)$$\", \"$$ \\\\sqrt{A/k} \\\\leq Ck^{2} \\\\Leftrightarrow k \\\\ge (A / C^2)^{1/5} \\\\quad \\\\cdots (**)$$\", \"Then, by the choice of $k$ in line 1099,\", \"when $k = \\\\lceil (A^3 / B^6)^{1/7}\\\\rceil \\\\leq \\\\lceil (A / C^2)^{1/5}\\\\rceil$, $(*)$ is true while $(**)$ is false, i.e., $\\\\sqrt{A/k} = \\\\mathcal{O}(Bk^{2/3}) = \\\\mathcal{O}(A^{2/7}B^{3/7})$ and $Ck^2 = C A^{6/7} B^{-12/7}$. Then line 1106 should be $\\\\mathcal{O}(A^{2/7}B^{3/7} + C A^{6/7} B^{-12/7})$ instead.\", \"when $k = \\\\lceil (A / C^2)^{1/5}\\\\rceil \\\\leq \\\\lceil (A^3 / B^6)^{1/7}\\\\rceil$, $(**)$ is true while $(*)$ is false, so that by the similar arguement line 1112 should be $\\\\mathcal{O}(A^{2/15} B C^{-4/15} + A^{2/5}C^{1/5})$.\", \"when $k = n \\\\leq \\\\min\\\\{ \\\\lceil (A^3 / B^6)^{1/7}\\\\rceil, \\\\lceil (A / C^2)^{1/5}\\\\rceil \\\\}$, both $(*)$ and $(**)$ are false and line 1117 should be $\\\\mathcal{O}(\\\\sqrt{A/n} + Bn^{2/3} + Cn^2)$.\", \"Therefore, none of the above cases show a convergence bound that is consistent with the main claim in Theorem 2, i.e., a linear speedup with total number of nodes $n$. May I request the authors to address this issue, and point out my mistakes if I am wrong in the above calculation.\", \"The similar argument applies to the proof of Lemma 10.\", \"(**Connection to DSGD**)\", \"We can equivalently interpret the proposed algorithm as a DSGD [Lian et. al., 2017] algorithm of $k$ nodes and each node has access to all local objective $f_i, ~i\\\\in \\\\\\\\{1, ..., n\\\\\\\\}$. Each iteration of the proposed algorithm with an active node set $V^{(t)}$ corresponds to an iteration of the above-mentioned DSGD algorithm where $\\\\\\\\{\\\\nabla f_j(z^{(t)}_{{\\\\rm token\\\\\\\\_id}_j^{(t)}}, \\\\xi_j^{(t)}): v_j \\\\in V^{(t)}\\\\\\\\}$ are the sampled local gradients. Therefore theoretically, by the results of [Lian et. al., 2017], I expect that the proposed algorithm to only achieve the linear speedup of $\\\\mathcal{O}(1/\\\\sqrt{kT})$.\", \"Also, by the above equivalence, the proposed algorithm is only serving as a new implementation of DSGD on large graph without data heterogeneity, i.e., $\\\\varsigma = 0$ in Assumption 1.3 of [Lian et. al, 2017].\", \"I suggest the authors to consider along this argument and compare with DSGD in the main text in terms of such equivalence. Also, since this algorithm is potentially only contributing an efficient large graph implementation of DSGD, I suggest the authors to expand the experiment section with larger scale experiments such as larger dataset and larger number of nodes $n> 100$.\"], \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"6\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Official Comment by Authors (1/2)\", \"comment\": \"We thank the reviewer for your constructive feedback.\\n\\n> In line 268, $k$ can be simplified as [...]\\n\\nThank you for the suggestion.\\nWe have not revised the statement of our theorems in the main paper yet because revising the theorem would complicate my response to your other comments.\\nWe will revise it as you suggested in the camera-ready version.\\n\\n> About the experiments: [...]\\n\\nWe are currently conducting these experiments, but we do not have enough time left to show you the results in this rebuttal period.\\nWe promise to add these experiments in the camera-ready version.\\n\\n> About the proof of Lemma 9 [...]\\n\\nTo derive the statement of Lemma 9, further expansion of the equations is necessary.\\n\\n* When $k = \\\\lceil (A^3 / B^6)^{1/7} \\\\rceil \\\\leq \\\\min \\\\\\\\{ \\\\lceil (A / C^2)^{1/5} \\\\rceil, n \\\\\\\\}$, we get \\n\\\\begin{align*}\\n\\\\sqrt{\\\\frac{A}{k}} = \\\\mathcal{O} (A^{\\\\frac{2}{7}} B^{\\\\frac{3}{7}}),\\n\\\\quad B k^{\\\\frac{2}{3}} = \\\\mathcal{O} (A^{\\\\frac{2}{7}} B^{\\\\frac{3}{7}}), \\n\\\\quad C k^2 \\\\leq \\\\mathcal{O} (A^{\\\\frac{2}{5}} C^{\\\\frac{1}{5}}). \\n\\\\end{align*}\\nIn the last inequality, we use $k \\\\leq \\\\lceil (A / C^2)^{1/5} \\\\rceil$. Note that we need to use $k \\\\leq \\\\lceil (A / C^2)^{1/5} \\\\rceil$ instead of $k = \\\\lceil (A^3 / B^6)^{1/7} \\\\rceil$ to get the statement of Lemma 9.\\nThen, using the above inequalities, we get the following convergence rate:\\n\\\\begin{align*}\\n \\\\mathcal{O} \\\\left( A^{\\\\frac{2}{7}} B^{\\\\frac{3}{7}} + A^{\\\\frac{2}{5}} C^{\\\\frac{1}{5}} \\\\right).\\n\\\\end{align*}\\n\\n* When $k = \\\\lceil (A/C^2)^{1/5} \\\\rceil \\\\leq \\\\min \\\\\\\\{ \\\\lceil (A^3 / B^6)^{1/7} \\\\rceil, n \\\\\\\\}$, we get\\n\\\\begin{align*}\\n\\\\sqrt{\\\\frac{A}{k}} = \\\\mathcal{O} (A^{\\\\frac{2}{5}} C^{\\\\frac{1}{5}}),\\n\\\\quad B k^{\\\\frac{2}{3}} \\\\leq \\\\mathcal{O} (A^{\\\\frac{2}{7}} B^{\\\\frac{3}{7}}),\\n\\\\quad C k^2 = \\\\mathcal{O} (A^{\\\\frac{2}{5}} C^{\\\\frac{1}{5}}). \\n\\\\end{align*}\\nNote that we use $k \\\\leq \\\\lceil (A^3 / B^6)^{1/7} \\\\rceil$ to obtain $B k^{2/3} \\\\leq \\\\mathcal{O} (A^{2/7} B^{3/7})$.\\nThen, using the above inequalities, we get the following convergence rate:\\n\\\\begin{align*}\\n \\\\mathcal{O} \\\\left( A^{\\\\frac{2}{7}} B^{\\\\frac{3}{7}} + A^{\\\\frac{2}{5}} C^{\\\\frac{1}{5}} \\\\right).\\n\\\\end{align*}\\n\\n\\n* When $k=n \\\\leq \\\\min \\\\\\\\{ \\\\lceil (A^3 / B^6)^{1/7} \\\\rceil, \\\\lceil (A/C^2)^{1/5} \\\\rceil \\\\\\\\}$, we get\\n\\\\begin{align*}\\n\\\\sqrt{\\\\frac{L r_0 (\\\\sigma^2 + (1 - \\\\frac{k-1}{n-1})\\\\zeta^2)}{k T}} = \\\\sqrt{\\\\frac{L r_0 \\\\sigma^2}{n T}}, \\n\\\\quad B k^{\\\\frac{2}{3}} \\\\leq \\\\mathcal{O} (A^{\\\\frac{2}{7}} B^{\\\\frac{3}{7}}),\\n\\\\quad C k^2 \\\\leq \\\\mathcal{O} (A^{\\\\frac{2}{5}} C^{\\\\frac{1}{5}}), \\n\\\\end{align*}\\nwhere we use $k \\\\leq \\\\lceil (A^3 / B^6)^{1/7} \\\\rceil$ and $k \\\\leq \\\\lceil (A/C^2)^{1/5} \\\\rceil$.\\nThen, we get the following rate:\\n\\\\begin{align*}\\n \\\\mathcal{O} \\\\left(\\\\sqrt{\\\\frac{L r_0 \\\\sigma^2}{n T}} + A^{\\\\frac{2}{7}} B^{\\\\frac{3}{7}} + A^{\\\\frac{2}{5}} C^{\\\\frac{1}{5}} \\\\right).\\n\\\\end{align*}\\n\\n\\nBy summarizing the above inequalities, we obtain the statement of Lemma 9.\\nWe have clarified the explanation above in the revised manuscript, with the revisions highlighted in blue.\\n\\n\\n\\n> [...] I expect that the proposed algorithm to only achieve the linear speedup of $\\\\mathcal{O}(1 / \\\\sqrt{kT})$.\\n\\nIf we carefully tune $k$, the convergence rate of TELEPORTATION can be $\\\\mathcal{O}(\\\\frac{1}{\\\\sqrt{n T}})$ as shown in Theorem 2.\\n\\nHere, we intuitively explain why setting $k$ as in Theorem 2 achieves the linear speedup $\\\\mathcal{O}(\\\\frac{1}{\\\\sqrt{n T}})$.\\nSee the proof in Lemma 9 and the above response for the precise discussion.\\n\\n* When $T$ is small, the second and third terms $\\\\mathcal{O}((\\\\frac{(1- p_k)}{T^2 p_k})^{1/3} + \\\\frac{1}{T p_k})$ dominate the convergence rate in Theorem 1. \\nIn this case, we would like to use a small $k$.\\n* When $T$ is large, the first term $\\\\mathcal{O}(\\\\frac{1}{\\\\sqrt{k T}})$ dominates the convergence rate in Theorem 1.\\nIn this case, we would like to set $k=n$.\\n\\nThus, if we carefully set $k$ so that $k$ increases to $n$ as $T$ increases, we can balance $\\\\mathcal{O}(\\\\frac{1}{\\\\sqrt{k T}})$ and $\\\\mathcal{O}((\\\\frac{(1- p_k)}{T^2 p_k})^{1/3} + \\\\frac{1}{T p_k})$. Then, TELEPORTATION can achieve the linear speedup $\\\\mathcal{O}(\\\\frac{1}{\\\\sqrt{nT}})$ and can totally eliminate the degradation caused by large $n$.\"}", "{\"title\": \"Ack\", \"comment\": \"Thank you for your response.\\n\\nI will be keeping my score, as I believe this paper is good and leaning towards acceptance.\", \"justification_for_not_higher_score\": \"Lots of questions as discussed above are open, and there is a huge score for improvement. I'd increase my score to 7 if ICLR provisioned that.\"}", "{\"title\": \"Official Comment by Authors (1/2)\", \"comment\": \"We thank the reviewer for our comments.\\n\\n> I am unsure why decentralized learning methods should necessarily converge more slowly as the number of nodes increases [...]\\n\\nWe would like to briefly explain why the convergence rate of DSGD can be degraded as $n$ is substantially large.\", \"dsgd_satisfies_the_following\": \"\\\\begin{align}\\n \\\\bar{\\\\mathbf{x}}^{(t+1)} = \\\\bar{\\\\mathbf{x}}^{(t)} - \\\\frac{\\\\eta}{n} \\\\sum_{i=1}^n \\\\nabla F_i (\\\\mathbf{x}_i ; \\\\xi_i^{(t)}),\\n\\\\end{align}\\n\\nwhere $\\\\bar{\\\\mathbf{x}} \\\\coloneqq \\\\frac{1}{n} \\\\sum_{i=1}^n \\\\mathbf{x}_i$ (see the proof of Lemma 11 in [1]).\\n$\\\\nabla F_i$ is calculated at different parameters $\\\\mathbf{x}_i$, and the convergence rate degrades when $\\\\mathbf{x}_i$ is far from $\\\\bar{\\\\mathbf{x}}$.\\nIn decentralized learning, each node communicates with a few nodes for communication efficiency.\\nFor instance, if we use a ring as the underlying topology, each node communicates with two neighboring nodes.\\nThus, $\\\\mathbf{x}_i$ comes to be far from $\\\\bar{\\\\mathbf{x}}$ as $n$ increases, and the convergence rate of DSGD degrades when $n$ is substantially large.\\n\\nThe following table summarizes the convergence rate of DSGD with various topologies.\\nFor all topology, $n$ appears in the numerator in the convergence rate, and the rate degrades when $n$ is substantially large.\\nWe show this table in Sec. D.\\n\\n| Topology | Convergence Rate | \\n| -------- | -------- | \\n| Ring | $\\\\mathcal{O} \\\\left( \\\\sqrt{\\\\frac{L r_0 \\\\sigma^2}{n T}} + \\\\left( \\\\frac{L^2 r_0^2 n^2 (\\\\sigma^2 + n^2 \\\\zeta^2)}{T^2} \\\\left( 1 - \\\\frac{1}{n^2} \\\\right) \\\\right)^\\\\frac{1}{3} + \\\\frac{L r_0 n^2}{T} \\\\right)$ | \\n| Torus | $\\\\mathcal{O} \\\\left( \\\\sqrt{\\\\frac{L r_0 \\\\sigma^2}{n T}} + \\\\left( \\\\frac{L^2 r_0^2 n (\\\\sigma^2 + n \\\\zeta^2)}{T^2} \\\\left( 1 - \\\\frac{1}{n} \\\\right)\\\\right)^\\\\frac{1}{3} + \\\\frac{L r_0 n}{T} \\\\right)$ |\\n| Exponential Graph | $\\\\mathcal{O} \\\\left( \\\\sqrt{\\\\frac{L r_0 \\\\sigma^2}{n T}} + \\\\left( \\\\frac{L^2 r_0^2 \\\\log_3 (n) (\\\\sigma^2 + \\\\log_2 (n) \\\\zeta^2)}{T^2} \\\\left( 1 - \\\\frac{1}{\\\\log_2 (n)} \\\\right) \\\\right)^\\\\frac{1}{3} + \\\\frac{L r_0 \\\\log_2 (n)}{T} \\\\right)$ | \\n| Base-2 Graph | $\\\\mathcal{O} \\\\left( \\\\sqrt{\\\\frac{L r_0 \\\\sigma^2}{n T}} + \\\\left( \\\\frac{L^2 r_0^2 \\\\log_2 (n) (\\\\sigma^2 + \\\\log_2 (n) \\\\zeta^2)}{T^2} \\\\right)^\\\\frac{1}{3} + \\\\frac{L r_0 \\\\log_2 (n)}{T} \\\\right)$ | \\n\\n\\n> The discussion in lines 140-148 seems to rely on the upper bound from Proposition 1. However, this argument feels weak, as the tightness of the upper bound is never discussed. [...] Is there any way to demonstrate that the bound is tight? [...]\\n\\nThank you for the comment.\\n**[1] analyzed the lower bound of the convergence rate of SGD and showed that it is inevitable that the convergence rate of DSGD degrades when $n$ is substantially large.**\\nWe will add the following discussion in the revised manuscript.\\n\\nSpecifically, Theorem 3 in [1] shows the lower bound of the convergence rate of DSGD when $f_i$ is $\\\\mu$-strongly convex and $L$-smooth with $\\\\mu = L=1$ and $\\\\sigma=0$.\\nUnder these assumptions, [1] showed that it requires\\n\\\\begin{align}\\n \\\\tilde{\\\\Omega} (\\\\frac{\\\\zeta (1-p_n)}{\\\\sqrt{\\\\epsilon} p_n})\\n\\\\end{align}\\niterations to converge to accuracy $\\\\epsilon$.\\nSee Theorem 3 in [1] for the precise statement.\\nThus, the convergence rate of DSGD must depend on $p_n$, and it is inevitable that the convergence rate of DSGD degrades when $n$ is substantially large since $p_n$ reaches zero as $n$ increases.\\n\\n> It appears that the theoretical improvement primarily stems from the dependence on $p_n$. Could the authors clarify the source of this improvement? Is it driven by a more refined analysis, or is it due to a novel aspect of the algorithm itself? [...]\\n\\nAs we mentioned above, it is inevitable that the convergence rate of DSGD degrades when $n$ is substantially large.\\n**Thus, the property that TELEPORTATION can alleviate the degradation caused by large $n$ is not obtained by the refined analysis but by its methodology.**\\n\\nWe would like to briefly explain why TELEPORTATION can alleviate this degradation in the following.\\n\\nThe convergence rate of DSGD degrades when $n$ is substantially large since $\\\\mathbf{x}_i$ comes to be far from $\\\\bar{\\\\mathbf{x}}$ due to the sparse communication characteristic.\\nTo prevent $\\\\mathbf{x}_i$ from being far from $\\\\bar{\\\\mathbf{x}}$, TELEPORTATION activates only $k$ nodes.\\nIf we use a small $k$, the gossip averaging is performed on a small topology, and the convergence rate does not depend on $p_n$, as shown in Theorem 1.\\nComparing Proposition 1 and Theorem 1, the second and third terms in the convergence rate in Theorem 1 are better since $p_k \\\\geq p_n$, but the first term is worse.\\nFinally, by carefully tuning $k$, we can balance these terms, obtaining the statement of Theorem 2, and TELEPORTATION can totally alleviate the degradation caused by large $n$.\"}", "{\"metareview\": \"This paper presents TELEPORTATION, a decentralized optimization algorithm designed to mitigate the common issue of deteriorating convergence rates in decentralized SGD as the number of nodes increases. The key insight is that larger spectral gaps in the communication graph typically require more iterations for convergence. TELEPORTATION addresses this by randomly activating a small subset of nodes at each iteration, reducing communication costs while maintaining effective learning. These active nodes update the model, perform a descent step, and achieve local consensus with other active nodes, repeating this process over multiple iterations. The paper provides a theoretical analysis of the algorithm\\u2019s convergence and an efficient method for tuning the number of active nodes.\\n\\nThere were a few concern raised by reviewers that some of them are partially resolved by authors, but the overall consensus was on accepting the paper; noting that the paper needs a through revision to incorporate suggested modifications.\", \"additional_comments_on_reviewer_discussion\": \"Presentation of claimed results needs further clarification (please consult suggestions from reviewers)\"}", "{\"comment\": \"We thank the reviewer for your comments.\\n\\n> Even though the distinction of TELEPORTATION from client sampling has been discussed, it still looks to me a follow-up variant in this category (McMahan 2017) [...] \\\"completely alleviate the convergence rate degradation\\\" is not that interesting any more.\\n\\n**We believe that the similarity between TELEPORTATION and client sampling does not mean that our paper is less novel.**\\nMitigating the degradation caused by large $n$ is an essential topic in decentralized learning, and our paper successfully proposed a method to solve this degradation.\\n\\nScaling decentralized learning to a large number of nodes is a challenging and important research topic, since the convergence rate of DSGD degrades when $n$ is significantly large.\\nExisting papers have attempted to alleviate this problem by proposing topologies with a large spectral gap [1,2,3].\\nHowever, no existing work can completely eliminate the convergence rate degradation caused by large $n$ (see Table 2 in Sec. D).\\nIt has been an open question of how to eliminate this degradation completely.\\n\\nIn this study, we solved this issue by activating only the appropriate number of nodes instead of designing a graph with a large spectral gap.\\nThen, we found that TELEPORTATION can **completely** eliminate this degradation without sacrificing the linear speedup property $\\\\mathcal{O}(1 / \\\\sqrt{n T})$.\\nWe believe that it is surprising that such a simple idea can totally alleviate this degradation.\\n\\nAs the reviewer mentioned, client sampling is widely studied in the federated learning literature to reduce the communication costs between the central server and nodes.\\nHowever, it is novel to use the idea of activating only a few nodes to mitigate the degradation caused by large $n$ in decentralized learning,\\nand it is not trivial that this idea can totally eliminate the degradation caused by $n$.\\nWe believe that the fact that the proposed method is similar to client sampling does not diminish our contribution.\\n\\n\\n> What are the Base-2 graph and what do you mean by superiority of this graph?\\n\\nBase-2 Graph is a topology that is proposed to alleviate the convergence rate degradation caused by large $n$, and it can achieve a reasonable balance between the convergence rate and communication efficiency.\\n[1] demonstrated that Base-2 Graph enables DSGD to more successfully reconcile accuracy and communication efficiency than the other existing topologies in their experiments.\\nThus, we compared TELEPORTATION and DSGD with Base-2 Graph in our experiments.\\n\\n\\n## Reference\\n\\n[1] Takezawa et. al., Beyond Exponential Graph: Communication-efficient Topologies for Decentralized Learning via Finite-time Convergence, In NeurIPS 2023\\n\\n[2] Ying et. al., Exponential Graph is Provably Efficient for Decentralized Deep Training, In NeurIPS 2021\\n\\n[3] Ding et. al., . DSGD-CECA: Decentralized SGD with Communication-optimal Exact Consensus Algorithm, In ICML 2023\"}", "{\"title\": \"Official Comment by Authors (2/2)\", \"comment\": \"> Also, by the above equivalence, the proposed algorithm is only serving as a new implementation of DSGD on large graph without data heterogeneity [...]\\n\\nWe respectfully disagree with this reviewer's comment that TELEPORTATION is equivalent to DSGD in the homogeneous case. Even in the homogeneous case, DSGD suffers from a degradation in convergence rate when $n$ is substantially large because stochastic gradient noise causes the parameters held by each node to drift away.\\nActivating only an appropriate number of nodes in TELEPORTATION is critical to mitigating this issue in both homogeneous and heterogeneous cases.\\n\\nSpecifically, TELEPORTATION with ring and DSGD achieve the following convergence rate when $\\\\zeta=0$, respectively:\\n\\\\begin{align*}\\n &\\\\textbf{TELEPORTATION}: \\\\\\\\;\\\\\\\\; \\\\mathcal{O} \\\\left( \\\\sqrt{\\\\frac{L r_0 \\\\sigma^2}{n T}} + \\\\left( \\\\frac{L r_0 \\\\sigma^{\\\\frac{3}{2}}}{T }\\\\right)^\\\\frac{4}{7} + \\\\left( \\\\frac{L r_0 \\\\sigma^\\\\frac{4}{3}}{T} \\\\right)^\\\\frac{3}{5} + \\\\frac{L r_0}{T} \\\\right), \\\\\\\\\\\\\\\\\\n &\\\\textbf{DSGD}: \\\\\\\\;\\\\\\\\; \\\\mathcal{O} \\\\left( \\\\sqrt{\\\\frac{L r_0 \\\\sigma^2}{n T}} + \\\\left( \\\\frac{L^2 r_0^2 \\\\sigma^2 (1 - p_n)}{T^2 p_n} \\\\right)^\\\\frac{1}{3} + \\\\frac{L r_0}{T p_n} \\\\right).\\n\\\\end{align*}\\nThe convergence rate of DSGD degrades when $n$ is substantially large, while the convergence rate of TELEPORTATION is consistently improved as $n$ increases.\\nOur experiments in Figure 2 can also show the superiority of TELEPORTATION, demonstrating that TELEPORTATION can converge faster than DSGD in both homogeneous and heterogeneous cases.\\n\\n> I suggest the authors to consider along this argument and compare with DSGD in the main text in terms of such equivalence.\\n\\nThank you for your suggestion.\\nWe will clarify this point in the revised manuscript.\\n\\n> I suggest the authors to expand the experiment section with larger scale experiments such as larger dataset and larger number of nodes $n > 100$.\\n\\nSince there is not enough time left until the end of the rebuttal, we cannot show the results with a larger number of nodes in this rebuttal period. We will add the results with a larger number of nodes in the camera-ready version.\"}", "{\"summary\": \"This work studies a decentralized optimization algorithm whose convergence rate does not deteriorate with increasing number of nodes. Usually, when number of nodes in a network increases, the spectral gap, i.e., $(1 - p_n)$ (according to the notation of the paper), also increases. Larger spectral gap means that decentralized SGD requires a larger number of iterations to converge.\\n\\nIn order to make decentralized SGD scalable for large networks, the proposed algorithm randomly activates a small subset of nodes by randomly sampling them. These nodes get the updated models from nodes activated in the last iteration, does a descent step, and subsequently communicate with other nodes active in the current iteration to do a local consensus step (only amongst the currently activated node). this process is repeated over multiple iterations.\\n\\nThe paper analyzes the convergence of this algorithm, and also provides an efficient algorithm to tune the number of active nodes.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"3\", \"strengths\": \"The paper is generally well-written. The problem statement being addressed is clear, and the proposed algorithm is also quite simple. The paper does a commendable job analyzing their proposed algorithm, in addition to validating numerically. The assumptions are also clearly stated, which are quite standard in decentralized optimization.\\n\\nThe work also compares their algorithm with client sampling, which is a natural question that came to my mind while reading the paper, and I appreciate that it was answered. I have a few concerns and would be glad if these are addressed. Despite my concerns, I strongly believe that within the current scope of the problem statement considered in the paper, it is adequately addressed.\", \"weaknesses\": \"I would appreciate it if the authors would answer some of my concerns.\\n\\n1. How are the consensus weights $\\\\mathbf{W}$ chosen, and how is it assumed that the randomly activated nodes at any iteration know these weights? \\n\\n2. How are the nodes sampled? Are they sampled uniformly at random? Would it help if the nodes were sampled taking some other criteria in mind, such as activating nodes that are more likely to have an edge between them? More generally, how do the authors expect the sampling to be implemented in practice in a fully decentralized topology? More generally, can the authors highlight a practical application scenario?\\n\\n3. The paper would benefit with some explicit discussion on the dependence of the node-selection strategy, and/or the required number of active nodes on the network topology? For instance, it seems like if the network is sparse, perhaps more nodes need to be activated at every iteration?\\n\\n4. Will it make sense to change the number of activated nodes in every iteration?\\n\\nI must acknowledge that I am not completely familiar with the current literature in topology selection for decentralized optimization. So I cannot definitively comment on the novelty of the paper, but based on how it is placed in the writing itself, the contribution seems non-trivial (despite the weaknesses mentioned here).\", \"questions\": \"I have some more questions which are not very critical, but I believe that the paper will benefit from:\\n\\n1. In the current proposed algorithm, data heterogeneity is not taken into account. Some discussions on how the algorithm might need to be modified in the presence of data heterogeneity would be appreciated. Perhaps optimizing the weight matrix $\\\\mathbf{W}$ would be helpful?\\n\\n2. Although less related, in light of the popularity of federated learning literature, the privacy of activated nodes might be an important criteria to consider. Since each node knows which node participated in the previous iteration, this can lead to some degree of privacy leakage, that needs to be quantified. (Please note that I do not see this as a major drawback of the paper -- it is just a suggestion). It ties more with my previous question of practical application scenario, where the network has a large number of nodes, so privacy becomes important.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"details_of_ethics_concerns\": \"None needed.\", \"rating\": \"6\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"comment\": \"We thank the reviewer again for carefully reviewing our theorem and response.\\nWe promise to include our discussion in the camera-ready version and revise the paper to make our methods more intuitive to the reader.\"}", "{\"title\": \"Response to Authors Comment\", \"comment\": \"Thank you the authors for the clear response. I am convinced by the response and will raise the rating to 6.\\nI suggest the authors to include the above discussion around different regime of $T$ (e.g., $T \\\\leq n^7 \\\\Rightarrow k = \\\\mathcal{O}(T^{1/7})$ vs $T > n^7 \\\\Rightarrow k = n$ for ring graph) and the corresponding choices of $k$ in the main text after the paragraph of Theorem 2, so that readers can make a clear interpretation of the results.\"}", "{\"title\": \"Official Comment by Authors (2/2)\", \"comment\": \"> The proposed hyperparameter-tuning method for selecting the number of active nodes adds another layer of complexity to the method. However, the impact of this additional tuning on overall training time and resource consumption is not fully explored. [...]\\n\\nWe will add the following discussion in the revised manuscript.\\n\\nBy comparing with DSGD, TELEPORTATION has an additional hyperparameter $k$.\\n**However, thanks to Algorithm 2, its hyperparameter-tuning requires only $2T$ iteration in total, which is not very expensive.**\\n\\nFrom the viewpoint of the convergence rate, TELEPORTATION is superior to DSGD, even considering the cost of hyperparameter-tuning with Algorithm 2 since the cost of hyperparameter-tuning with Algorithm 2 is constant, which is negligible compared to the degradation of the convergence rate of DSGD caused by large $n$.\\n\\nFrom an experimental viewpoint, TELEPORTATION can be superior to DSGD in the heterogeneous case.\\nFigure 3 shows that TELEPORTATION can train the neural network more stably than DSGD in the heterogeneous case.\\nThus, even considering the cost of hyperparameter-tuning with Algorithm 2, TELEPORTATION can be superior to DSGD.\\nIn the homogeneous case, DSGD and TELEPORTATION achieved comparable performance in Figure 3.\\nThus, in the homogeneous case, DSGD is superior to TELEPORTATION.\\n\\n\\n> The assumption that any two nodes can communicate directly (as mentioned in the abstract and discussions) is overly idealized. [...]\\n\\nAs the reviewer mentioned, there are cases, such as wireless networks, where there are pairs of nodes that cannot communicate. However, there are also many settings where this condition is satisfied. \\nFor example, in a data center or in a setting where nodes are connected to the Internet, any two nodes can exchange parameters. \\nTELEPORTATION can be used in these cases to scale decentralized learning to a large number of nodes.\\n\\nFurthermore, we would like to emphasize that even in this setting, it is not trivial to eliminate the convergence rate degradation caused by large $n$.\\nIn this setting, many existing papers have also tried to alleviate this degradation by designing a topology with a large spectral gap [2-4], while none of the existing papers can completely eliminate this degradation (see the discussion in Section 2.2).\\nTELEPORTATION is the first decentralized learning method that can completely alleviate the degradation caused by large $n$.\\n\\nIt is a limitation that TELEPORTATION can only work when any two nodes can communicate.\\nHowever, we believe that our work has achieved the important steps for scaling decentralized learning to a large number of nodes in the general case where there are pairs of nodes that cannot communicate.\\n\\n\\n## Reference\\n[1] Koloskova et. al., A Unified Theory of Decentralized SGD with Changing Topology and Local Updates, In ICML 2019\\n\\n[2] Takezawa et. al., Beyond Exponential Graph: Communication-efficient Topologies for Decentralized Learning via Finite-time Convergence, In NeurIPS 2023\\n\\n[3] Ying et. al., Exponential Graph is Provably Efficient for Decentralized Deep Training, In NeurIPS 2021\\n\\n[4] Ding et. al., . DSGD-CECA: Decentralized SGD with Communication-optimal Exact Consensus Algorithm, In ICML 2023\"}", "{\"comment\": \"I appreciate the contributions of the paper; however, I feel that the presentation may overstate the significance of the results. The analysis by Koloskova et al. establishes a connection between the convergence rate and parameters $p$ and $n$, showing that the required convergence time decreases as $n$ increases and increases as $p$ decreases. In scenarios where\\n$p$ depends on $\\ud835\\udc5b$, as seen in all the cases presented in Table 2 of this paper, their bounds have the claimed degradation in terms of $n$. \\n\\nIn the context where $p$ depends on $n$, denoted here as $p_n$, the primary contribution of the present paper is to replace $p_n$ in Koloskova et al.'s bounds with $p_k$. The authors then claim to have completely removed the degradation with respect to $n$. However, this claim appears to be somewhat overstated for the following reasons:\\n1. Dependency of $p_k$ on $n$: While $p_k$ is introduced as distinct from $p_n$, it is not entirely independent of $n$. For instance, in lines 277-282 regarding the Exp. Graph, $k$ is defined as $\\\\max (1,n, \\\\dots )$. Consequently, $p_k$ inherits a dependency on $n$ through $k$.\\n2. Addition of a new parameter: Even if $p_k$ does not depend on $n$, the introduction of $p_k$ adds an additional parameter to the analysis.. Improvements of this nature could also be achieved using other techniques (e.g. client activation, graph sparsification) that similarly parameterize $p$ with a variable independent of $n$.\\n3. Limited impact in certain cases: The paper does not demonstrate any improvement in cases where the spectral gap does not scale with $n$.\"}", "{\"comment\": \"We thank the reviewer for clarifying the concerns.\\n\\n\\nWe suspect the reviewer feels that our claim that TELEPORTATION can \\\"completely\\\" mitigate the degradation caused by $n$ is overstated.\\nFor example, we could remove this statement and modify this statement to be more accurate by claiming that \\u201cthe convergence rate of TELEPORTATION consistently improves convergence rate as $n$ increases.\\u201d\\n\\nWould this modification address the reviewer's concerns?\\nIf the reviewer's concerns are still unresolved, please let us know. We will gladly address your concerns.\\n\\nSee below for more detailed comments.\\n\\n> While $p_k$ is introduced as distinct from, it is not entirely independent of $n$. For instance, in lines 277-282 regarding the Exp. Graph, $p_k$ is defined as $\\\\max (1, n, \\\\dots)$. Consequently, $p_k$ inherits a dependency on $n$ through $k$.\\n\\nWe agree with the reviewer that $p_k$ is not entirely independent of $n$ since the choice of $k$ depends on $n$.\\nHowever, the dependence of $n$ is no longer harmful in TELEPORTATION, as shown in Theorem 2, where the convergence rate of TELEPORTATION consistently improves as $n$ increases.\\nFor this reason, we claimed that TELEPORTATION can completely eliminate the convergence rate degradation caused by large $n$ in the current manuscript, but if the reviewer feels that this claim is an overstatement, we would be happy to revise it as mentioned above.\\n\\n> Even if $p_k$ does not depend on $n$, the introduction of $p_k$ adds an additional parameter to the analysis.. Improvements of this nature could also be achieved using other techniques (e.g. client activation, graph sparsification) that similarly parameterize $p$ with a variable independent of $n$.\\n\\nAs the reviewer mentioned, we need to make $p$ independent of $n$ to alleviate the convergence rate degradation caused by large $n$.\\n\\nHowever, we would like to emphasize again that no existing paper has succeeded in eliminating the degradation associated with large $n$ (see Table 2 in Sec. D), and our paper is the first to propose a decentralized learning method whose convergence rate can consistently improve as $n$ increases.\\n\\nThe reviewer mentioned that other techniques might achieve results similar to TELEPORTATION, but we believe the possibility of unpublished alternative methods does not diminish our contribution.\\n\\n> The paper does not demonstrate any improvement in cases where the spectral gap does not scale with $n$.\\n\\nIn decentralized learning, it is essential to reconcile the communication efficiency and convergence rate, e.g., $p_n$.\\nAs we discussed in our paper, no existing paper has succeeded in proposing a topology that has a spectral gap that does not decrease as $n$ increases without sacrificing the communication efficiency (see Table 2 in Sec. D).\\nTELEPORTATION is the first method that can eliminate the convergence rate degradation caused by large $n$ without sacrificing the communication efficiency.\"}", "{\"summary\": \"The paper addresses degradation in decentralized learning arising from fixed communication patterns. Given the flexibility to select both a communication pattern and a subset of active nodes, the authors propose the TELEPORTATION algorithm. This algorithm activates k nodes per iteration and applies a specific topology to these nodes, facilitating adaptable communication.\\n\\nThe convergence analysis for TELEPORTATION follows the approach used by Koloskova (2020b), extending it to any number of active nodes k. Additionally, the authors provide an efficient method for tuning k, including determining optimal values for specific graph types, such as rings and exponential graphs, where they establish the theoretically best k and its associated convergence rate.\\n\\nExperimental results demonstrate that TELEPORTATION consistently outperforms decentralized SGD across various settings and graph structures, reinforcing its potential advantages in decentralized learning.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"The primary strength of this paper lies in its proof of Theorem 1, which effectively addresses a key limitation in previous analyses of similar algorithms. Prior approaches encountered challenges due to the disconnectivity of graphs under certain node activation conditions (e.g., when p=0), resulting in an infinite bound on the number of iterations required for convergence. Remarkably, this paper recovers the bound obtained by Kolosokova (2020b) under a generalized node selective activation scheme, which could prove potential for future research in this domain.\\n\\nIn addition to the theoretical insights, the paper is also notable for its clear presentation of contributions. The experimental section is both comprehensive and persuasive, further validating the authors' approach and supporting the potential utility of their bound.\", \"weaknesses\": \"Even though the distinction of TELEPORTATION from client sampling has been discussed, it still looks to me a follow-up variant in this category (McMahan 2017), and client sampling has been a heavily discussed approach. In the special case where any node is allowed to connect to any node, decentralized learning with client sampling is a very natural idea. Admittedly the proof of this paper could be something new, but the approach they proposed, at least in practice, might have been discussed/implemented allready. In particular, it is expected when you select a fixed graph of size k, where its spectral gap is known for example in ring and exp. graph, the spectral gap does not show up in the bound. In this sense, \\\"completely alleviate the convergence rate degradation\\\" is not that interesting any more.\", \"minor\": \"there are no black nodes in Figure 1.\", \"questions\": \"What are the Base-2 graph and what do you mean by superiority of this graph?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"title\": \"Official Comment by Authors (1/2)\", \"comment\": \"We thank the reviewer for your positive feedback.\\n\\n> How are the consensus weights $\\\\mathbf{W}$ chosen, and how is it assumed that the randomly activated nodes at any iteration know these weights?\\n\\nWe explain how the sampling of active nodes can be implemented in more detail.\\n\\nAll nodes communicate before starting the training and have the same seed value, as we mentioned in line 215.\\nThen, we sample the active nodes $V_\\\\text{active}$ and assign $\\\\\\\\{1, 2, \\\\cdots, k \\\\\\\\}$ to variables $\\\\\\\\{ \\\\\\\\text{token\\\\\\\\_id}_i \\\\\\\\}$ by using this seed value for each iteration in line 3 in Algorithm 1.\\nSince all nodes have the same seed value, all nodes can know which nodes are active and can obtain the same variables $\\\\\\\\{ \\\\text{token\\\\\\\\_id}_i \\\\\\\\}$ without communicating with other nodes.\\nUsing these variables, active nodes exchange parameters with their neighbors and compute the weighted average in lines 11-12 in Algorithm 1.\\n\\n> Would it help if the nodes were sampled taking some other criteria in mind, such as activating nodes that are more likely to have an edge between them?\\n\\nIn our current proposed method, we assume that the active nodes are sampled uniformly.\\nIn the federated learning literature, many client sampling strategies have been proposed, as we summarized in Sec. 4.\\nThus, there might be a better scheme for selecting active nodes than random sampling.\\nHowever, the primary objective of our work is to address the convergence rate degradation associated with large $n$ and to make decentralized learning scalable to a large number of nodes.\\nWe have shown that random sampling of active nodes successfully achieves this goal.\\nWe believe that proposing a better node sampling scheme is beyond the scope of our paper, and we leave it for future research.\\n\\n\\n> More generally, how do the authors expect the sampling to be implemented in practice in a fully decentralized topology? More generally, can the authors highlight a practical application scenario?\\n\\nSampling of the active nodes can be performed in a decentralized manner.\\nSee our first response for more details.\\n\\n> The paper would benefit with some explicit discussion on the dependence of the node-selection strategy, and/or the required number of active nodes on the network topology? For instance, it seems like if the network is sparse, perhaps more nodes need to be activated at every iteration?\\n\\nIn Theorem 2, we show the optimal number of active nodes when we use a ring and exponential graph as a topology to connect active nodes.\\nAn exponential graph has a larger spectral gap than a ring.\\nComparing the results with a ring and an exponential graph, more nodes are activated when we use an exponential graph.\\nThis is because the parameters that each node has are more likely to drift away when the number of active nodes increases when a ring is used than when an exponential graph is used.\\n\\nSee the above response for the discussion of the node-selection strategy.\"}" ] }

[ "Teng LIN", "Yizhang Zhu", "Yuyu Luo", "Nan Tang" ]

[ "Multi-Entity QA", "Wikipedia Graph", "Structured QA", "RAG" ]

{ "note_id": [ "l8LjSfXI49", "ga0LbtqXBZ", "YUi8L57xBa", "OyZGRPajHD", "Cqs09nHLQo", "3jiWAGcJze" ], "note_type": [ "official_review", "official_review", "comment", "official_review", "official_review", "official_review" ], "note_created": [ 1730606190121, 1730558067983, 1732154343713, 1730401510401, 1730691783643, 1730538510180 ], "note_signatures": [ [ "ICLR.cc/2025/Conference/Submission6515/Reviewer_o1jz" ], [ "ICLR.cc/2025/Conference/Submission6515/Reviewer_LFcr" ], [ "ICLR.cc/2025/Conference/Submission6515/Authors" ], [ "ICLR.cc/2025/Conference/Submission6515/Reviewer_tgBr" ], [ "ICLR.cc/2025/Conference/Submission6515/Reviewer_7eqN" ], [ "ICLR.cc/2025/Conference/Submission6515/Reviewer_sfh1" ] ], "structured_content_str": [ "{\"summary\": \"This paper introduces MEgraph QA, a benchmark targeting multi-entity question answering, and proposes the Structured QA framework to address this challenge. The framework\\u2019s core innovation is converting long texts into tables, effectively compressing input length and filtering irrelevant information. Structured QA demonstrates optimal performance on the MEgraph QA benchmark.\", \"soundness\": \"2\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1.\\tThe benchmark construction process is clear. The author describes the benchmark construction process in detail, including question templates. The dataset is also analyzed in detail, including the types of questions, entity numbers, etc.\\n2.\\tThe motivation is straightforward. Excessive retrieved text often introduces irrelevant information, yet the authors effectively improve answer accuracy by extracting key information into a table to filter the noise.\", \"weaknesses\": \"1.\\tThere is a weak correlation between the problem and the solution. While the authors focus on multi-entity question answering, the method primarily addresses the simplification of long retrieval documents. In this sense, the proposed method does not need to focus on entities but can be applied in any scenario where the retrieval text is too long, but the authors do not experimentally validate it.\\n2.\\tThe significance of the benchmark remains unverified. In the main experiments, the authors evaluate only the base model, RAG, and Structure QA, without exploring alternative RAG methods [1] or other graph-based approaches[2,3]. We do not know if the benchmark is necessary for further research.\\n3.\\tThe effectiveness of the proposed method is not convincing. As noted in W1 and W2, SQA in the main experiment is only compared with basic methods, Specifically, the experiment employs the highly capable GPT-4 as the main model of SQA (refer to lines 212 and 252), resulting in a potentially imbalanced comparison with less advanced models such as GPT-3.5 and Llama. We hope that the authors can add more baseline[1,2,3].\\n4.\\tThe authors should consider conducting ablation experiments to verify the necessity of SQL in their approach. A comparative analysis between the SQL query-based table method and simpler techniques, such as keyword extraction and summarization, would clarify the relative advantages of SQL in this context.\\n\\n[1] Wang H, Li R, Jiang H, et al. Blendfilter: Advancing retrieval-augmented large language models via query generation blending and knowledge filtering[J]. arXiv preprint arXiv:2402.11129, 2024.\\n[2] Wu Y, Huang Y, Hu N, et al. CoTKR: Chain-of-Thought Enhanced Knowledge Rewriting for Complex Knowledge Graph Question Answering[J]. arXiv preprint arXiv:2409.19753, 2024.\\n[3] Sun J, Xu C, Tang L, et al. Think-on-Graph: Deep and Responsible Reasoning of Large Language Model on Knowledge Graph[C]//The Twelfth International Conference on Learning Representations.\", \"questions\": \"1.\\tCan you describe in detail how to use the entities and relations in the text to query the graph in Figure 1-a3? Is it query by string matching?\\n2.\\tFor the executor in Section 4.3, how do LLM and SQL interact with each other? Does LLM write the SQL query and then execute it?\\n3.\\tFor Section 5.3, why is quality control necessary? In practice, the semantics of the questions are usually unclear.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"5\", \"confidence\": \"5\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"Wikipedia is widely used for QA, but traditional methods struggle with multi-page information. This paper proposes a Structured QA (SQA) approach, organizing entities into relational tables and using table-based QA, improving accuracy by 29.6% over current methods.\", \"soundness\": \"3\", \"presentation\": \"3\", \"contribution\": \"2\", \"strengths\": \"1, The structure of the presentation is clear, covering the whole process from classified query and graph search to table generation and execution.\\n2. This paper constructed a specialized benchmark for Multi-Entity QA. However, no access link is provided, so it is not possible to judge its quality.\\n3. The SQA approach, using a relational table to structure multi-entity data, aims to enhancethe retrieval accuracy and scalability of question answering over Wikipedia, which is reasonable.\", \"weaknesses\": \"1. The core idead in this paper, i.e., using a relational table to structure multi-entity data, has been studied in work \\\"Semantic Table Retrieval Using Keyword and Table Queries\\\", \\\"Web Table Extraction, Retrieval, and Augmentation: A Survey\\\", \\\"Novel Entity Discovery from Web Tables\\\" and \\\"Web Table Extraction, Retrieval and Augmentation\\\".\\n2. The question and answer is based on the contents of the relational table. How to ensure the accuracy of the entity relations extracted by GPT-4? it has not been explained whether there is a more efficient way to build structured entity relations, and whether the additional time and space complexity introduced by the construction table.\\n3. The performance achieved is heavily dependent on GPT-4, and thus the availability is relatively limited. Therefore, it is recommended that the author utilize other open-source large models as tools to replace GPT-4 for comparison with existing methods, in order to verify the validity of the core idea.\", \"questions\": \"1. The core idead in this paper, i.e., using a relational table to structure multi-entity data, has been studied in work \\\"Semantic Table Retrieval Using Keyword and Table Queries\\\", \\\"Web Table Extraction, Retrieval, and Augmentation: A Survey\\\", \\\"Novel Entity Discovery from Web Tables\\\" and \\\"Web Table Extraction, Retrieval and Augmentation\\\".\\n2. The question and answer is based on the contents of the relational table. How to ensure the accuracy of the entity relations extracted by GPT-4? it has not been explained whether there is a more efficient way to build structured entity relations, and whether the additional time and space complexity introduced by the construction table.\\n3. The performance achieved is heavily dependent on GPT-4, and thus the availability is relatively limited. Therefore, it is recommended that the author utilize other open-source large models as tools to replace GPT-4 for comparison with existing methods, in order to verify the validity of the core idea.\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"withdrawal_confirmation\": \"I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors.\"}", "{\"summary\": \"The paper introduces a dataset of about 4780 question answers based on Wikipedia. The questions involve multiple entities and can be classified into three broad categories: Comparison, Statistical, and Relationship questions. The paper also proposes an approach called SQA to answer such questions. The key idea is to create an intermediate table that summarizes the information needed to answer the question.\", \"soundness\": \"1\", \"presentation\": \"1\", \"contribution\": \"1\", \"strengths\": \"The paper proposes a method for solving the difficult task of answering non-trivial questions involving multiple entities. However, due to the shortcomings mentioned in the weakness section, it is difficult to assess the paper's significance based on the current draft.\", \"weaknesses\": \"**Clarity**: The paper lacks clarity and purpose. Therefore, in its current form, it may not be useful to the community.\\n* There are several places where sufficient information is not provided to get a point through. For example, in section 2.3, the authors cite a couple of papers stating that those papers propose GraphQA benchmarks. The authors then simply go on to say that those benchmarks are insufficient without providing any specific reason. In section 4.1, the authors mention that they use named entity recognition and relation extraction models but fail to provide any details of the model or a reference to a pre-existing model they use for the purpose. Another example is section 5.2, which tries to describe the process of generating QA pairs to create the dataset. From the two paragraphs in the section (lines 306 to 320), I was unable to understand how they created the QA pairs. \\n\\n**Quality**: The experiments are insufficient to establish the usefulness of the dataset as well as the real efficacy of the proposed model.\", \"questions\": \"1. Around line 403, you mention that for some subcategories of questions under the Statistics, you do not evaluate the accuracy of the model on the numerical answers but instead check whether the SQA model can identify the relevant columns in the table that it creates. Then, for these categories, how do we evaluate the baselines? As I understand, the baselines do not create the intermediate table before answering the queries.\\n\\n2. Could you please provide the details of the name entity recognition and relation extraction models you use?\\n\\n3. Since the paper is highly empirical, could you please include the reproducibility statement as described in the [ICLR author instructions](https://iclr.cc/Conferences/2025/AuthorGuide).\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"3\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The paper addresses the challenge of multi-entity question answering (QA) on Wikipedia, where queries often involve complex relationships across multiple entities, like \\u201cHow many Turing Award winners are Canadian?\\u201d Traditional approaches, such as using retrieval-augmented generation (RAG), struggle with aggregating insights from multiple pages effectively. To overcome this, the authors propose a Structured Question Answering (SQA) approach that organizes information into relational tables tailored to each query. This structure enables advanced table-based methods, like TableQA or NL2SQL, for more accurate responses.\", \"soundness\": \"2\", \"presentation\": \"2\", \"contribution\": \"2\", \"strengths\": \"1. The development of MEGraphQA as a benchmark specifically for multi-entity QA over Wikipedia is a notable contribution.\\n2. The paper explores a practical area of application by focusing on Wikipedia, a widely used knowledge base, for QA. This emphasis on a real-world source may make the research outcomes directly relevant to practical implementations of QA systems.\", \"weaknesses\": \"1. Many KBQA systems already handle multi-entity questions effectively, utilizing pre-existing data structures within the knowledge base to support relational queries. The paper does not convincingly establish why a dedicated approach, such as SQA, is necessary.\\n2. The idea of converting entity data into structured tables is well-established in database query systems, and approaches like NL2SQL have long enabled structured QA without the need for a specialized framework. Thus, SQA lacks originality and does not present a significant innovation in the context of established KBQA solutions.\\n3. The description of the SQA approach lacks depth in explaining how the schema is derived or validated, how the table generation process addresses entity disambiguation, and the specific steps of entity extraction. A more detailed breakdown of each process would help clarify the novelty and complexity of SQA.\\n4. The paper primarily discusses simple schema extraction for relatively straightforward queries. However, in real-world applications, queries can vary widely in complexity, requiring more sophisticated relational or semantic inference. The paper could be strengthened by addressing how SQA would handle more complex or ambiguous queries.\", \"questions\": \"1. How does the proposed system ensure the correctness of the inferred schema, especially in complex or ambiguous queries? What mechanisms are in place to handle errors in schema derivation or entity disambiguation?\\n2. Does SQA encounter any specific failure modes or limitations in terms of entity coverage, complex relationship resolution, or accuracy? How does it handle such limitations, and are there planned mitigations?\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}", "{\"summary\": \"The paper addresses multi-entity question answering (QA) using Wikipedia as a knowledge source. It argues that modern methods, particularly those involving retrieval-augmented generation (RAG) with large language models (LLMs), struggle to synthesize information from multiple Wikipedia pages effectively. Instead, it suggests an approach that constructs a table as a summary to answer multi-entity questions. The paper proposes a Structured Question Answering (SQA) approach that works in three main steps: (1) graph search to identify entities from the question, (2) table generation to extract relevant information from graphs, and (3) execution of SQL to get the answer. In order to test the proposed approach, the paper introduces MEGraphQA, a new benchmark for evaluating multi-entity QA on Wikipedia. In experiments, the SQA approach significantly outperforms existing methods.\", \"soundness\": \"2\", \"presentation\": \"1\", \"contribution\": \"2\", \"strengths\": \"1. The paper introduces using a table as an intermediate object for context construction.\\n2. The paper introduces another dataset for multi-hop reasoning question answering.\", \"weaknesses\": \"1. While the overall idea of the SQA approach is straightforward and original, its description is rather vague.\\n 1. In Section 4.1, the paper describes the first step to parse the question into entities and relations. The paper mentions the use of GPT-4, NER, and RE. However, no details are given on how these are achieved.\\n 2. In Section 4.2, the paper describes the schema generation prompts in the appendix but does not state how Mistral-7B populates the table.\\n 3. In Section 4.3, there is no detail of how a question is converted to SQL.\\n2. The paper should discuss many existing multi-hop reasoning datasets to motivate another benchmark. For example, this article reviews seven multi-hop QA datasets ([ref](https://arxiv.org/pdf/2204.09140#page=27.11)). In addition, the paper mentions a quality control step in Section 5.3 but does not provide any detail or evidence that it is done thoroughly. \\n3. The comparison in Table 3 shows that SQA is better than RAG. However, some important, missing baselines, such as [StructQA](https://arxiv.org/abs/2311.03734) or [HOMLES](https://aclanthology.org/2024.acl-long.717/), might be missing.\\n4. The paper could benefit from writing improvement. Several paragraphs do not convey important reasoning or evidence but rather an opinionated advantage of the approach, such as Section 4.4, Lines 103 - 107, and Lines 358 - 362. Some minor comments about the notations include notations not being used, unclear what $P$ and $t$ are, duplicate duplicates of $V$, and lack of definition of $P_v$.\", \"questions\": \"-\", \"flag_for_ethics_review\": \"['No ethics review needed.']\", \"rating\": \"3\", \"confidence\": \"4\", \"code_of_conduct\": \"Yes\"}" ] }