image
stringlengths 42
218
| text
stringlengths 100
1k
| paper_id
stringlengths 12
12
| figure_idx
int64 1
312
|
|---|---|---|---|
Figure 1:Comparison between baseline and SMART-OL.Existing topology reasoning methods suffer from limited sensor data. SMART augments online topology reasoning with robust map priors learned from scalable SD and satellite maps, substantially improving lane perception and topology reasoning. | 2502.04329v1 | 1 |
|
Figure 2:Outline of the proposed approach.In the first stage (bottom row), SMART is trained at scale using SD and satellite maps for lane graph prediction, supervised by large-scale geo-referenced HD maps. In the second stage (top row), the robust map priors learned by SMART are seamlessly integrated into any online driving topology reasoning models, significantly enhancing lane perception and topology reasoning. | 2502.04329v1 | 2 |
|
Figure 3:Qualitative comparison of SMART-OL to baselines.The top-left shows the SD map plotted on top of the satellite image. Our method improves baselines consistently, producing more complete lane graphs. | 2502.04329v1 | 3 |
|
Figure 4:Impact of varying sensor data availability.With only 40% of sensor data, SMART-OL achieves performance comparable to using the full sensor data, demonstrating its robustness with reduced sensor data. | 2502.04329v1 | 4 |
|
Figure 2:OlaArchitecture.Olasupports omni-modal inputs including text, image, video, and audio, capable of processing the inputs simultaneously with competitive performance on understanding tasks for all these modalities. Meanwhile,Olasupports user-friendly real-time streaming decoding for texts and speeches thanks to the text detokenizer and the speech decoder. | 2502.04328v1 | 2 |
|
Figure 3:Progressive modality alignment helps to learn better omni-modal models.We compare our progressive alignment strategy with two baseline training pipelines on Image QA(MMBench[40]), Video QA(VideoMME[21]), and ASR(LibriSpeech[54]): 1) direct mixing where all instruction tuning data is merged and trained in a single stage, and 2) balanced sampling where we upsample certain sources to make the training data more balanced among modalities. The experiment is conducted on a subsampled training set for efficiency and we train models for the same number of steps for fair comparisons. The score is normalized based on the score of progressive alignment to calculate the relative score and the ASR score is inverted as lower is better for the WER metric. | 2502.04328v1 | 3 |
|
Figure 5:Generative results on speech and visual understanding tasks.We illustrate results on speech and video understanding and show the strong ability of omni-modalOlacompared with conventional vision-language models. | 2502.04328v1 | 5 |
|
Figure 1:Examples in WorldSense.Unlike existing benchmarks, WorldSense emphasizes the tight coupling of audio-visual perception in real-world scenarios, where accurate comprehension relies on integrating both modalities, as neither alone provides sufficient context for correct answers. In thefirstexample, the video shows a man holding a fruit, but visual information alone is insufficient to determine his specific action, while audio alone cannot identify the fruit type. In thesecondexample, both visual cues and audio are necessary to identify the country elements and determine which segment of music is “lively and energetic.” Through this design, WorldSense aims to serve as a platform for evaluating MLLMs’ real-world perception and understanding capabilities using omni-modal information. | 2502.04326v1 | 1 |
|
Figure 2:Distribution of WorldSense.(a) Videos in WorldSense spans 8 primary categories with 67 fine-grained subcategories. (b) QA pairs are structured across 26 tasks. (c) Acoustic signals distribution. Individual videos may contain multiple audio categories, leading to overlapping counts in statistical analysis. Consequently, the cumulative sum of audio instances exceeds the total video count. (d) Video duration distribution. The average duration of videos is 141.1 seconds. | 2502.04326v1 | 2 |
|
Figure 3:Data collection and QA annotation pipelines.(a) Data collection and curation process. (b) QA annotation and quality control pipeline. | 2502.04326v1 | 3 |
|
Figure 4:Breakdown results.(a) Fine-grained results on task category. (b) Fine-grained results on audio signals. | 2502.04326v1 | 4 |
|
Figure 5:Impact of video frames.We compare performance of several Video MLLMs with different frames input. | 2502.04326v1 | 5 |
|
Figure 1.Sample Mondrian Tessellations on𝒳=[0,1]2𝒳superscript012\mathcal{X}=[0,1]^{2}caligraphic_X = [ 0 , 1 ] start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPTwithλ=1𝜆1\lambda=1italic_λ = 1. | 2502.04323v1 | 1 |
|
Figure 2.Sample Isotropic Poisson Hyperplane Tessellation and Poisson Manhattan Tessellation on𝒳=[0,1]2𝒳superscript012\mathcal{X}=[0,1]^{2}caligraphic_X = [ 0 , 1 ] start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPTwithλ=1𝜆1\lambda=1italic_λ = 1. | 2502.04323v1 | 2 |
|
Figure 3.Limiting Kernel for the Uniformly Rotated Mondrian Process onℝ2superscriptℝ2\mathbb{R}^{2}blackboard_R start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPTwith lifetimeλ=1𝜆1\lambda=1italic_λ = 1. | 2502.04323v1 | 3 |
|
(a)Convergence of Random Feature Kernel to Limiting Kernel for Uniform Points on[0,1]2superscript012[0,1]^{2}[ 0 , 1 ] start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT. | 2502.04323v1 | 4 |
|
(a)Convergence of Random Feature Kernel to Limiting Kernel for Uniform Points on[0,1]2superscript012[0,1]^{2}[ 0 , 1 ] start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT. | 2502.04323v1 | 5 |
|
Figure 1:Process flow of human evaluation to identify attributes contributing to jailbreak harmfulness. We collect and curate10101010harmful base examples that meet all four attributes and augment each response into16161616variants with different attribute combinations using GPT-4. Human annotators then assess each variant for the four attributes and the overall harm level. | 2502.04322v1 | 1 |
|
Figure 2:Real-world human-LLM interactions fromWildVis. The left example illustrates a multi-step user-LLM interaction with a malicious query and subsequent follow-ups.
In the right example, the multilingual LLM provides step-by-step instructions in response to a malicious query in Spanish. | 2502.04322v1 | 2 |
|
Figure 3:OurSpeak Easyjailbreak framework.
Given a malicious query, we (1) decompose it into multiple steps of seemingly harmless subqueries and (2) translate each subquery into a set of predefined languages from different resource groups.
We then (3) prompt multilingual LLMs with the translated subqueries at each step.
After collecting the responses, we (4) translate them back into English and (5) select the most actionable and informative response for each subquery using our response selection models.
Finally, (6) the selected responses are combined to form a complete response to the original malicious query. | 2502.04322v1 | 3 |
|
Figure 4:Jailbreak performance measured byASRandHarmScorebefore and after integratingSpeak Easyinto the baselines, with the shaded bars highlighting the difference.Speak Easysignificantly increases bothASRandHarmScoreacross almost all methods.
See Table10for full numerical values. | 2502.04322v1 | 4 |
|
Figure 5:Top: Average actionability and informativeness scores; Bottom: Selection rates for each language, both forn=6𝑛6n=6italic_n = 6.
Each color theme represents a language resource level.
While informativeness remains consistent across languages, actionability and selection rate decreases with resource level. | 2502.04322v1 | 5 |
|
Figure 6:Prompts used to instruct GPT-4o to augment responses by removing each attribute from the response.[QUERY]and[RESPONSE]are replaced with the original query-response pairs from Table7, respectively. | 2502.04322v1 | 6 |
|
Figure 7:Annotation questionnaire for assessing the relationship between the four identified attributes and the harm in jailbreak responses. | 2502.04322v1 | 7 |
|
Figure 8:Annotation instructions and example instances for comparing alignment results betweenASRandHarmScore. | 2502.04322v1 | 8 |
|
Figure 9:The prompt used to decompose the harmful query, along with four in-context examples.[NUMBER OF SUBQUERIES]and[HARMFUL QUERY]are replaced with the number of subqueries and the jailbreak query during test time. By default,[NUMBER OF SUBQUERIES]is set to 3. | 2502.04322v1 | 9 |
|
(a)Prompt used to summarize paragraph-length questions in theStack-Exchange-Preferencesdataset into a single sentence. | 2502.04322v1 | 10 |
|
(a)Prompt used to summarize paragraph-length questions in theStack-Exchange-Preferencesdataset into a single sentence. | 2502.04322v1 | 11 |
|
(b)Prompt used to determine whether a question from theHH-RLHForStack-Exchange-Preferencesdatasets can be answered with an actionable response to filter out irrelevant questions. | 2502.04322v1 | 12 |
|
(c)Prompt used to determine whether a question from theHH-RLHForStack-Exchange-Preferencesdatasets can be answered with an informative response to filter out irrelevant questions. | 2502.04322v1 | 13 |
|
Figure 11:Language selection rates forn=3𝑛3n=3italic_n = 3andn=9𝑛9n=9italic_n = 9.
Each color theme represents a language resource level.
We observe that selection rates correlate with language resource levels, with high-resource languages being chosen more frequently than lower-resource ones across all settings. | 2502.04322v1 | 15 |
|
Figure 1:ConceptAttentionproduces saliency maps that precisely localize the presence of textual concepts in images.We compare Flux raw cross attention, DAAM(Tang et al.,2022)with SDXL, and TextSpan(Gandelsman et al.,2024)for CLIP. | 2502.04320v1 | 1 |
|
Figure 2:ConceptAttentionaugments multi-modal DiTs with a sequence of concept embeddings that can be used to produce saliency maps.(Left) An unmodified multi-modal attention (MMAttn) layer processes bothpromptandimagetokens. (Right)ConceptAttentionaugments these layers without impacting the image appearance to create a set of contextualizedconcepttokens. | 2502.04320v1 | 2 |
|
Figure 3:ConceptAttentioncan generate high-quality saliency maps for multiple concepts simultaneously.Additionally, our approach is not restricted to concepts in the prompt vocabulary. | 2502.04320v1 | 3 |
|
Figure 4:(a)MMAttncombines cross and self attention operations between the prompt and image tokens. (b) OurConceptAttentionallows the concept tokens to incorporate information from other concept tokens and the image tokens, but not the other way around. | 2502.04320v1 | 4 |
|
Figure 5:ConceptAttentionproduces higher fidelity raw scores and saliency maps than alternative methods, sometimes surpassing in quality even the ground truth saliency map provided by the ImageNet-Segmentation task. Top row shows the soft predictions of each method and the bottom shows the binarized predictions. | 2502.04320v1 | 5 |
|
Figure 6:LaterMMAttnlayers encode richer features for zero-shot segmentation.We investigated the impact of using features from variousMMAttnlayers and found that deeper layers lead to better performance on segmentation metrics like pixelwise accuracy, mIoU, and mAP. We also found that combining the information from all layers further improves performance. | 2502.04320v1 | 6 |
|
Figure 7:Optimal segmentation performance requires some noise to be present in the image.We evaluated the performance ofConceptAttentionby encoding samples from a variety of timesteps (determines the amount of noise). Interestingly, we found that the optimal amount of noise was not zero, but in the middle to later stages of the noise schedule. | 2502.04320v1 | 7 |
|
Figure 8:Pseudo-code depicting the (a) multi-modal attention operation used by Flux DiTs and (b) ourConceptAttentionoperation.We leverage the parameters of a multi-modal attention layer to construct a set of contextualized concept embeddings. The concepts query the image tokens (cross-attention) and other concept tokens (self-attention) in an attention operation. The updated concept embeddings are returned in addition to the image and text embeddings. | 2502.04320v1 | 8 |
|
Figure 1:Overview. Given a number of input images, sshELF first reconstructs several virtual views and only then predicts the 3D Gaussian primitives of the scene from which novel views are rendered. The colors of the latent information correspond to different object classes, such as purple for buildings and green for vegetation. | 2502.04318v1 | 1 |
|
Figure 3:Overview of sshELF. Given a few input images, sshELF first encodes them into latent features using a pre-trained DinoV2 (Sec.3.1). As part of thebackbone, the latent features, together with a pre-trained depth head, are used to initialize the virtual views, which are refined using hierarchical ELF blocks consisting of cross- and self-attention layers (Sec.3.2). Reference and virtual views are then fed into thetranslatorpart to predict 3D Gaussian splats (Sec.3.3). Not shown here is the rasterization part used for creating novel views (Sec.3.4). | 2502.04318v1 | 3 |
|
Figure 1:FIGConvNet: ConvNet for drag prediction using FIG convolution blocks. The encoder and decoder consist of a set of FIG convolution blocks and we connect the encoder and decoder with skip connections. The output of the encoder is used for drag prediction and the output of the decoder is used for pressure prediction. | 2502.04317v1 | 1 |
|
Figure 3:Factorized Implicit Global Convolution 3D:
The FIG convolution first creates a set of voxel grids that factorizes the domain. This allows representing a high resolution voxel grid domain implicitly that can be computationally prohibitive to save explicitly. Then, a set of global convolution operations are applied in parallel to these voxel grids to capture the global context. Finally, the voxel grids are aggregated to predict the output. | 2502.04317v1 | 3 |
|
Figure 4:Point Convolution: The features from source and target nodes as well as offset are fed into an MLP to lift the features, which are then aggregated and projected back to the original feature space using an MLP. | 2502.04317v1 | 4 |
|
(a)Number of Sample Points on Drag Prediction: The networks are robust to the number of sample points used for drag prediction. | 2502.04317v1 | 5 |
|
(a)Number of Sample Points on Drag Prediction: The networks are robust to the number of sample points used for drag prediction. | 2502.04317v1 | 6 |
|
(b)Drag prediction vs. Ground truth drag on DrivAerNet. The drag prediction closely matches the drag ground truth withR2superscript𝑅2R^{2}italic_R start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPTof 0.95. | 2502.04317v1 | 7 |
|
Figure 6:Normalized Pressure Prediction and Error Visualization on DrivAerNet.Our network predicts both drag coefficients and per vertex pressure. We visualize the ground truth pressure and prediction along with the absolute error of the pressure. Note that the pressures are normalized to highlight the errors clearly. | 2502.04317v1 | 8 |
|
Figure 1:Our Main Contributions. We develop a novel probabilistic metric for model similarity, CAPA (κpsubscript𝜅𝑝\kappa_{p}italic_κ start_POSTSUBSCRIPT italic_p end_POSTSUBSCRIPT), which adjusts for chance agreement due to accuracy. Using this, we find (1) LLM-as-a-judge scores are biased towards more similar models controlling for the model’s capability (2) Gain from training strong models on annotations of weak supervisors (weak-to-strong generalization) is higher when the two models are more different, (3) Concerningly, model errors are getting more correlated as capabilities increase. | 2502.04313v1 | 1 |
|
Figure 2:Judgment Score relation with Model Similarity.Each line is a regression model fit between judgment and similarity scores as computed between model and judge pairs. Each point represents a single pair, and⋄⋄\diamond⋄indicates that both, the model and the judge, come from the same model family. We report for each fit the corresponding Pearson correlation values,r𝑟ritalic_r. We found significant positive correlation between judgment scores and similarity across all judges,∗∗**∗ ∗indicatesp<0.01𝑝0.01p<0.01italic_p < 0.01. | 2502.04313v1 | 2 |
|
Figure 3:Similarity vs Gain from Weak-to-Strong Training.Across 12 model pairs, the strong student gains more from weak-to-strong training on tasks where it is more different from the weak supervisor (p<0.01𝑝0.01p<0.01italic_p < 0.01). | 2502.04313v1 | 3 |
|
Figure 4:Role of Complementary Knowledge and Elicitation in Weak-to-Strong Generalization. We decompose the accuracy of the weak-to-strong trained model on four parts of the test data distribution, based on the correctness of the weak supervisor and an oracle strong elicited model which uses ground-truth annotations. Sub-rectangles represent weak, strong model pairs. Results are averaged across 15 tasks. Complementary knowledge transfer explains weak-to-strong model accuracy beyond elicitation. | 2502.04313v1 | 4 |
|
Figure 5:Average Similarity (κpsubscript𝜅𝑝\kappa_{p}italic_κ start_POSTSUBSCRIPT italic_p end_POSTSUBSCRIPT) vs Model Capability. We split 130 LMs into 5 buckets based on their accuracy percentile. For each LM we compute its mean similarity within the bucket (across models from different developers), and plot it against model accuracy. The size of the scatter points is proportional to model size. Asκpsubscript𝜅𝑝\kappa_{p}italic_κ start_POSTSUBSCRIPT italic_p end_POSTSUBSCRIPTmeasures overlap in mistakes, the positive correlation indicates LM mistakes are getting more correlated with increasing capabilities. | 2502.04313v1 | 5 |
|
Figure 6:Metric comparison when models tend towards agreement. We compare different metric values for two models in a binary setting. For first model we set 90% accuracy and calibration to 0.99 (meaning the model is highly confident in its answers). For the second model, we increase it’s calibration from 0.01 to 0.99 to approach the same distribution as the first model. On y-axis we are plotting metric value on x-axis we are reportingp¯¯𝑝\overline{p}over¯ start_ARG italic_p end_ARGfor the second model which as the model becomes more calibrated approaches accuracy of the first model. | 2502.04313v1 | 6 |
|
Figure 7:Metric comparison when models tend towards disagreement (Read plot from right to left). We compare different metric values for two models in a binary setting. For the first model, we set accuracy to 90% and calibration to 0.99 (the model is highly confident in its answers). For the second model, we incrementally increase its disagreement with model one by pushing its probability mass to the second option and increasing its calibration to 0.99. | 2502.04313v1 | 7 |
|
Figure 7:Metric comparison when models tend towards disagreement (Read plot from right to left). We compare different metric values for two models in a binary setting. For the first model, we set accuracy to 90% and calibration to 0.99 (the model is highly confident in its answers). For the second model, we incrementally increase its disagreement with model one by pushing its probability mass to the second option and increasing its calibration to 0.99. | 2502.04313v1 | 8 |
|
Figure 8:Metric comparison when models tend towards disagreement with adjustedκpsubscript𝜅𝑝\kappa_{p}italic_κ start_POSTSUBSCRIPT italic_p end_POSTSUBSCRIPT. Replication of fig.8but with adjustedκpsubscript𝜅𝑝\kappa_{p}italic_κ start_POSTSUBSCRIPT italic_p end_POSTSUBSCRIPTas asκp^^subscript𝜅𝑝\hat{\kappa_{p}}over^ start_ARG italic_κ start_POSTSUBSCRIPT italic_p end_POSTSUBSCRIPT end_ARG, computation following eq.18. | 2502.04313v1 | 9 |
|
Figure 10:Accuracy of free-form responses compared with multiple-choice accuracy on MMLU-Pro.The free-form responses were rated using an ensemble of five capable LM judges. Each judge was given access to the original MMLU-Pro reference answers and their decisions whether a given response is correct or not were aggregated using majority voting. | 2502.04313v1 | 11 |
|
Figure 11:Judgment Scores compared with the ensemble judgment accuracy given access to reference answers.We compare the judgment scores of each judge using only their own knowledge and capabilities to the rating of a judge ensemble that has access to the ground-truth options. The latter is a good proxy of the real correctness of responses. | 2502.04313v1 | 12 |
|
Figure 12:We decompose the accuracy of the weak to strong trained model on four parts of the train data distribution based on whether the weak supervisor and an oracle strong elicited model (using ground-truth annotations) are correct or wrong. All results are averaged over 15 datasets. Sub-rectangles represent weak, strong model pairs. On the train dataset, complementary knowledge transfer (mean accuracy0.590.590.590.59) plays an equal role as elicitation (mean accuracy0.560.560.560.56). | 2502.04313v1 | 13 |
|
Figure 13:Various Similarity Metrics vs Weak-to-Strong gain. The highest correlation is seen for CAPAκpsubscript𝜅𝑝\kappa_{p}italic_κ start_POSTSUBSCRIPT italic_p end_POSTSUBSCRIPT, though in the binary classification setup of weak-to-strong generalization the probabilistic information does not add much value compared to error consistency.1−JSD1𝐽𝑆𝐷1-JSD1 - italic_J italic_S italic_Dgives a more noisy scatter plot, with lower correlation (r𝑟ritalic_r). | 2502.04313v1 | 14 |
|
Figure 14:Test Accuracies for various models and ceiling estimates in Weak-to-Strong training. The accuracies are averaged over 12 model pairs. The initial strong student model has consistently lower accuracy than the weak supervisor consistent withBurns et al. (2024); Scherlis et al. (2024). The weak-to-strong trained student surpasses the weak-supervisor across datasets. However it has lower accuracy than the elicitation ceiling which trains the strong student on ground-truth annotations. Finally, our new estimated ceiling which incorporates the complementary knowledge of the weak supervisor has even higher accuracies, showing even more scope for improvements. | 2502.04313v1 | 19 |
|
(a)LM Similarity (κpsubscript𝜅𝑝\kappa_{p}italic_κ start_POSTSUBSCRIPT italic_p end_POSTSUBSCRIPTforM>2𝑀2M>2italic_M > 2) vs Average Accuracy of Model Pairs in each Capability bin | 2502.04313v1 | 23 |
|
(a)LM Similarity (κpsubscript𝜅𝑝\kappa_{p}italic_κ start_POSTSUBSCRIPT italic_p end_POSTSUBSCRIPTforM>2𝑀2M>2italic_M > 2) vs Average Accuracy of Model Pairs in each Capability bin | 2502.04313v1 | 24 |
|
(b)LM Similarity (Discreteκpsubscript𝜅𝑝\kappa_{p}italic_κ start_POSTSUBSCRIPT italic_p end_POSTSUBSCRIPT) vs Average Accuracy of Model Pairs in each Capability bin | 2502.04313v1 | 25 |
|
(a)When choosingεp,εw,εnsubscript𝜀psubscript𝜀wsubscript𝜀n\varepsilon_{\rm p},\varepsilon_{\rm w},\varepsilon_{\rm n}italic_ε start_POSTSUBSCRIPT roman_p end_POSTSUBSCRIPT , italic_ε start_POSTSUBSCRIPT roman_w end_POSTSUBSCRIPT , italic_ε start_POSTSUBSCRIPT roman_n end_POSTSUBSCRIPTaccording to Table1withτ=2𝜏2\tau=2italic_τ = 2,η=1𝜂1\eta=1italic_η = 1. | 2502.04312v1 | 1 |
|
(a)When choosingεp,εw,εnsubscript𝜀psubscript𝜀wsubscript𝜀n\varepsilon_{\rm p},\varepsilon_{\rm w},\varepsilon_{\rm n}italic_ε start_POSTSUBSCRIPT roman_p end_POSTSUBSCRIPT , italic_ε start_POSTSUBSCRIPT roman_w end_POSTSUBSCRIPT , italic_ε start_POSTSUBSCRIPT roman_n end_POSTSUBSCRIPTaccording to Table1withτ=2𝜏2\tau=2italic_τ = 2,η=1𝜂1\eta=1italic_η = 1. | 2502.04312v1 | 2 |
|
(b)When choosingεp=εw=εsubscript𝜀psubscript𝜀w𝜀\varepsilon_{\rm p}=\varepsilon_{\rm w}=\varepsilonitalic_ε start_POSTSUBSCRIPT roman_p end_POSTSUBSCRIPT = italic_ε start_POSTSUBSCRIPT roman_w end_POSTSUBSCRIPT = italic_ε. | 2502.04312v1 | 3 |
|
Figure 2:A two-dimensional sphere, one of its tangent planes, and an isomorphic copy of the latter as the Euclidean spaceℝ2superscriptℝ2\mathbb{R}^{2}blackboard_R start_POSTSUPERSCRIPT 2 end_POSTSUPERSCRIPT | 2502.04312v1 | 4 |
|
(a)A two-dimensional sphere inℝ3superscriptℝ3\mathbb{R}^{3}blackboard_R start_POSTSUPERSCRIPT 3 end_POSTSUPERSCRIPTand one of its tangent planes colored in pink | 2502.04312v1 | 5 |
|
(a)A two-dimensional sphere inℝ3superscriptℝ3\mathbb{R}^{3}blackboard_R start_POSTSUPERSCRIPT 3 end_POSTSUPERSCRIPTand one of its tangent planes colored in pink | 2502.04312v1 | 8 |
|
Figure 1:AD using idealised Finding Pegasus dataset illustrating the distinction between on- and off-manifold anomalies. (Horse images created using Canva.com) | 2502.04310v1 | 1 |
|
Figure 2:Schematic of formal framework for AD using DR. Detected off- and on-manifold anomalies are shown by the shaded regions. | 2502.04310v1 | 2 |
|
Figure 3:Schematic showing relationship between on- and off-manifold anomalies when performing AD after dimensionality reduction. | 2502.04310v1 | 3 |
|
Figure 5:7777s ‘anomalies’ from MNIST handwritten digits dataset representing extreme forms of the bulk topology -EohippusandSampsonpoints. | 2502.04310v1 | 5 |
|
Figure 6:8888s ‘anomalies’ from MNIST handwritten digits dataset representing a radically different topology -Pegasuspoints. | 2502.04310v1 | 6 |
|
Figure 7:t-SNE 2D representation of distribution of dataset points upon the PCA manifoldM=84𝑀84M=84italic_M = 84with1111s shown in red,7777s in blue and8888s in yellow. We can see clearly the two “islands” where most of the7777s and8888s reside. This is shown zoomed in in the figure on the right. | 2502.04310v1 | 7 |
|
Figure 8:t-SNE 2D representation of distribution of dataset points upon the PCA manifoldM=84𝑀84M=84italic_M = 84with points detected as anomalies by off-manifold reconstruction error highlighted. Detected7777s are shown in magenta, detected8888s in cyan, and the detected1111s (false positive anomalies) are shown in black. The subfigure to the right shows a zoomed-in view of this.
We can see immediately that the reconstruction error approach is doing a decent job of identifying the “islands” of both7777s and8888s. | 2502.04310v1 | 8 |
|
Figure 9:‘Anomalies’ detected from MNIST handwritten digits dataset using reconstruction error from PCA manifold atM=84𝑀84M=84italic_M = 84 | 2502.04310v1 | 9 |
|
Figure 10:Zoomed-in view of TSNE 2D representation of distribution of dataset points upon the PCA manifoldM=84𝑀84M=84italic_M = 84showing those detected by PCA RE and those by all the on-manifold models combined. The points in cyan are anomalies that are only detected by the off-manifold method. | 2502.04310v1 | 10 |
|
Figure 11:MSE of reconstruction error between PCA manifold and original data representations. Anomalies are defined as the most outlying 240 points | 2502.04310v1 | 11 |
|
Figure 12:Chart showing distribution of different underlying data classes on t-SNE 2D representation of the AEM=30𝑀30M=30italic_M = 30manifold | 2502.04310v1 | 12 |
|
Figure 13:Chart showing distribution of anomalies detected through AE reconstruction error from the AEM=30𝑀30M=30italic_M = 30manifold | 2502.04310v1 | 13 |
|
Figure 1.Plots for estimating demographic parity in simulation Setting 1. Each point represents an estimate from a simulated dataset. In both plots, the X-axis is sample size and the Y-axis is the estimated demographic parity. The dotted line represents the true value of demographic parity for the distribution. For the lower plot, the bands around each point represent the 95% confidence interval. | 2502.04309v1 | 1 |
|
Figure 2.Heatmap demonstrating coverage in different scenarios and as sample size varies. For each cell in the heatmap, coverage is calculated over 100 simulations. | 2502.04309v1 | 2 |
|
Figure 3.Line plot comparing targeted learning estimate of variance to a t-test estimate of variance. | 2502.04309v1 | 3 |
|
Figure 4.Line plot and heatmap demonstrating error and coverage results for CMI. In the line plots, the X-axis is the value of c and the Y-axis is the error. The solid horizontal line represents and error of 0. For every combination of (c,estimator type, sample size), we perform 100 simulations. The heatmap shows coverage for the TL estimator as c and sample size varies. | 2502.04309v1 | 4 |
|
Figure 5.Feature importance to fairness scores for both the Adult and Law school datasets. The first row contains feature importances for Adult, and the second row contains feature importances for Law school. | 2502.04309v1 | 5 |
|
Figure 1:Overivew of HOG-Diff.The dashed line above illustrates the classical generation process, where graphs quickly degrade into random structures with uniformly distributed entries. In contrast, as shown in the coloured region below, HOG-Diff adopts a coarse-to-fine generation curriculum based on the diffusion bridge, explicitly learning higher-order structures during intermediate steps with theoretically guaranteed performance. | 2502.04308v1 | 1 |
|
Figure 2:Visualization of Cell Complex Transformations.(a) An example graph. (b) Cell complex representation with corresponding homeomorphisms to closed balls for three example cells. (c) Black elements represent high-order structures extracted from the original graph through 2-cell filtering, while those in grey denote corresponding peripheral structures. | 2502.04308v1 | 2 |
|
Figure 2:Overview of proposed framework.Left (Training):We collect a large multi-task dexterous in-hand manipulation dataset in simulation to pretrain a generative model that can generate diverse actions conditioned on the current state. The pretrained generative model can produce useful actions including rotation, translation, and more intricated behaviors.Right (Inference):During inference, we can project dangerous motion produced by teleoperation or policy back to a high-likelihood action with guided sampling. This makes DexGen capable of assisting a coarse high-level policy to perform complex object manipulations. | 2502.04307v1 | 2 |
|
Figure 3:Dataset:The Anygrasp-to-Anygrasp dataset generation pipeline is designed for the generative pretraining of DexGen. For a wide variety of objects, we extensively search for potential grasp configurations, using these as both the initial and goal states for RL policies. To ensure our diffusion model can manage diverse scenarios, we incorporate varied wrist poses, movements, and domain randomization during RL training and data collection. | 2502.04307v1 | 3 |
|
Figure 4:Model:Architecture of the DexGen controller. The whole system takes robot state, external motion conditioning, and mode conditioning as input. A diffusion model first generates the motion as the intermediate action representation. The motion conditioning is not fed into the diffusion model directly but as the gradient guidance during the diffusion sampling. Then, another inverse dynamics model will translate the generated motion to executable robot action. We implement our diffusion model as a UNet in this paper. The inverse dynamics model is a residual multilayer perceptron. | 2502.04307v1 | 4 |
|
Figure 5:Our large-scale, multi-task pretraining dataset covers diverse grasp to grasp transitions (arrows). DexGen controller learns the dataset action distribution (purple shaded area) at each state, and we can use sequential motion prompting (purple triangle) to perform a useful long-horizon skill, connecting two distance states. | 2502.04307v1 | 5 |
|
Figure 6:Real world experimental setup based on Allegro Hand with a Franka Panda Arm (Left). We use human teleoperation (Right) as a proxy for high-level policy. | 2502.04307v1 | 6 |
|
Figure 7:Part of our real world testing objects, which are not present in our pretraining dataset. We include objects of different sizes, masses, and aspect ratios. | 2502.04307v1 | 7 |
|
Figure 8:Results of simulation evaluation. We use DexGen to correct several noise-corrupted expert policies. Note that each dimension of action space is bounded by [-1, 1] and these noises ruin the expert action most of the time. We measure the average duration (in seconds) and number of achieved goals per trial over a 20-minute simulated experiment. As shown in the figure, DexGen can successfully improve the performance of these policies. Across the experiments, DexGen can boost the duration by 10-100x and even help an extremely perturbed policy to achieve success where the baseline fails. | 2502.04307v1 | 8 |
|
Figure 9:DexGen can maximally preserve input action while correcting dangerous actions. DexGen can reject users’ behavior (open up the palm) and keep holding the object. | 2502.04307v1 | 9 |
|
Figure 10:Diffusion Model in DexGen Controller. We use a standard U-Net based diffusion model with FiLM conditioning. | 2502.04307v1 | 10 |
|
Figure 2:Overview ofMotionCanvas. Given an input image and high-level scene-space motion intent, MotionCanvas decomposes and translates the motion (camera and object motion with their timing) into screen space by leveraging the depth-based synthesis and hierarchical transformation with the Motion Signal Translation module. These screen-space motion signals are subsequently passed to a video generation model to produce the final cinematic shots. | 2502.04299v1 | 2 |
End of preview. Expand
in Data Studio
ArXiv AI Figure-Caption Dataset
This dataset contains figures and their captions extracted from recent AI-related papers on arXiv.
Dataset Information
- Total figures: 17094
- Source: arXiv papers from 2023.12 onwards
- Categories: CS.AI, CS.LG, CS.CV, CS.CL, CS.NE, STAT.ML
Content Format
Each entry contains:
image: URL of the figure imagetext: Caption textpaper_id: arXiv paper IDfigure_idx: Index of the figure in the paper
Data Quality
- All captions have passed quality filters for length and content
- Images are verified for appropriate dimensions and format
License
This dataset is released under CC BY-SA 4.0 License. This means you are free to:
- Share: Copy and redistribute the material in any medium or format
- Adapt: Remix, transform, and build upon the material for any purpose
Under the following terms:
- Attribution: You must give appropriate credit, provide a link to the license, and indicate if changes were made
- ShareAlike: If you remix, transform, or build upon the material, you must distribute your contributions under the same license
Citation
If you use this dataset in your research, please cite it as:
@dataset{arxiv_figures,
author = {Yiming Liu},
title = {ArXiv AI Figure-Caption Dataset},
year = {2025},
publisher = {HuggingFace},
url = {https://huggingface.co/datasets/Leonardo6/ArXiv-AI-Figure-Caption}
}
Acknowledgments
All figures and captions are sourced from papers on arXiv.org. We acknowledge the authors of all papers included in this dataset.
- Downloads last month
- 43
Size of downloaded dataset files:
3.05 MB
Size of the auto-converted Parquet files:
3.05 MB
Number of rows:
17,094
Data Sourcing report
powered
by
Spawning.aiNo elements in this dataset have been identified as either opted-out, or opted-in, by their creator.