Datasets:
Translation-EnKo
/
arxiv-translation

Dataset card Data Studio Files Community
2
text
stringlengths
4
4.47k
In the case of tail narrowing, the scaling behavior changes; instead of a plateau, we obtain a slower decay rate:
* _Vision-language alignment pre-training._ To develop MLLMs, existing work mostly initializes the vision encoder and the LLM with pre-trained models [149, 826, 150]. These models retain excellent vision and language capacities, but span different semantic spaces. Thus, the goal of vision-language alignment pre-training (_i.e._, the first-stage training) is to align the vision encoder and the LLM through end-to-end training on large-scale image-text pairs [827, 828]. However, directly tuning these two models on image-text pairs may cause the degradation of the original representation capacities. To improve the alignment performance, it is crucial to design effective training strategies and select appropriate pre-training data [829, 830]. Existing work mainly employs the following strategies for cross-modality alignment: (1) if the number of image-text pairs is not sufficiently large (_e.g._, less than 1M), it is often suggested to only update the connection module [831]; (2) if the training data includes high-quality text corpora [832] or image-text pairs with fine-grained annotations [833], fine-tuning the LLM can be conducted to boost the performance; (3) if the number of image-text pairs is very large (_e.g._, about 1B), fine-tuning the vision encoder is also plausible [829, 830], but the benefit remains further verification. * _Visual instruction tuning._ After vision-language pre-training, the second-stage training, _i.e._, visual instruction tuning, aims to improve the instruction-following and task-solving abilities of MLLMs. Generally, the input of visual instruction tuning consists of an image and a task description, and the task is to generate a corresponding text output. To boost the performance, high-quality visual instruction data is key to eliciting and enhancing the abilities of MLLMs. Therefore, most studies are dedicated to constructing various visual instruction datasets. As the basic approaches, early studies construct visual instructions by distilling from GPT-4 [149] or reformulating vision-language task datasets [151]. To enhance the quality of instruction data, recent work further proposes improved strategies by increasing the instruction diversity [834], incorporating fine-grained information (_e.g._, coordinate of objects) into the instruction [833], or synthesizing complex visual reasoning instructions [835].
RepresentationDegenerationis a phenomenon in which pretrained models tend to adopt low-entropy singular value distributions (Jing et al., 2022). In language modeling, representation degeneration takes the form of anisotropy (Ethayarajh, 2019; Rajee & Pilehvar, 2021) and was proven to be related with the Zipfian shape of token distribution (Gao et al., 2019; Bis et al., 2021). We study this phenomenon along training and its relation with saturation.
Figure 2: **Our procedure of generating data for self-supervised training.
In Figure 2, we evaluate performance on the real-world knowledge task for multiple checkpoints during pre-training, where accuracy is reported using best@1 sampling. We notice an upward trend in performance on the reverse task with no saturation at the last checkpoint. Hence, we assume that if we continue pre-training we would see further improvement.
**Are reference Gold or Gilded?** We conducted an in-depth analysis of the training data (FLORES-200 data) utilized by the ALMA model. We meticulously compared the quality of the reference translations with those generated by strong translation models. Our findings reveal that, in numerous instances, the quality of human-written parallel data is even inferior to that of system-generated translations. This observation underscores a critical insight: training models exclusively towards replicating reference translations may not be the most effective approach, and reliance on reference-based evaluation could be flawed.
EvaluationWe evaluated the ability to generate Japanese answers to Japanese math problems. Therefore, we considered an answer correct if it met the following criteria: (1) the concluding numerical value must be correct, and (2) the reasoning text should be written in Japanese.
**Qdrant**[241] is a vector similarity search engine and vector database. It provides a production-ready service with a convenient API to store, search, and manage points--vectors with an additional payload Qdrant is tailored to extended filtering support. environment.
**Does the dataset contain any data, if viewed directly, might be offensive, insulting, threatening or induce anxiety?** A large number of the prompts contain inappropriate, offensive and unsafe language.
Scaling of language modes (LMs) on both model and data sizes has been shown to be effective for improving the performance on a wide range of tasks [33; 3; 16; 5; 53; 44; 28], leading to the widespread adoption of LM applications, e.g., ChatGPT. The success of such scaling is guided by scaling laws [15; 21; 7; 16], which study the predictability of pre-training loss given the model and data sizes.
Table 3 reports the average accuracy of each model for our English evaluation tasks and the normalized accuracy for the German Hella Swag evaluation task. We do not report the average German evaluation score as it is not informative due to evaluations having near-random chance accuracy (see Table 11). We observe that English models consistently outperform German models on the English evaluations. However, the strong replay used with the 25% replay German model helps to reduce this gap. English models' English evaluation performance is very similar with a range of 1.19 between the highest and lowest values. We suspect that there is significant noise in the evaluation process for base models of this size and believe that the differences are likely not significant. That being said, the continually pre-trained model with LR re-warming, LR re-decaying, and replay does improve on the \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\) model. When evaluating German-trained models on English evaluation tasks, we see consistent improvements for models using more replay. We note that once again the model trained with LR re-warming, LR re-decaying, and replay does improve on the \(\mathcal{D}_{0}\cup\mathcal{D}_{1}\) model. Turning to the German Hella Swag results we observe that German models consistently outperform their English counterparts. Among German-trained models, the continually trained models outperform the union-trained model and the model trained exclusively on German.
_Remark A.3_ (# GPUs).: In this paper, we do not specify the number of GPUs as it is irrelevant. The results remain the same whether using 64 GPUs each with a batch size of 24, 48 GPUs each with a batch size of 32, or 1536 GPUs each with a batch size of 1.
Taken together, our contributions call for a more responsible, or "collapse-aware", proliferation of synthesized data. Scale is _not_ all you need: more work on effective watermarking for synthetic data is needed, to make it more distinguishable from the original, human-annotated data. Thus, clean / real data will become an even more valuable resource in the future, as we are ushering in the "beyond scaling" era.
Text extraction.We want to extract only the main content of the page, ignoring menus, headers, footers, and ads among others: Lopukhin (2019) found that trafilatura (Barbaresi, 2021) was the best non-commercial library for retrieving content from blog posts and news articles. Although this is only a narrow subset of the kind of pages making up CommonCrawl, we found this finding to hold more broadly. We use trafilatura for text extraction, and apply extra formatting via regular expressions: we limit new lines to two consecutive ones, and remove all URLs.
Lastly, FLARE Jiang et al.
Figure 4: Different experimental settings examined in this paper. Each row represents a different experimental setting with a unique name and number, and each vertical section highlighted by a right-pointing light-blue triangle indicates a training phase. Models are assessed on test QA across all settings. Whenever multiple datasets are enclosed within a dashed square, they are mixed together during the training process.
**Others.** The Pile [161] is a large-scale, diverse, and open-source text dataset consisting of over 800GB of data from multiple sources, including books, websites, codes, scientific papers, and social media platforms. It is constructed from 22 diverse high-quality subsets. The Pile dataset is widely used in models with different parameter scales, such as GPT-J (6B) [165], CodeGen (16B) [86], and Megatron-Turing NLG (530B) [113]. ROOTS [162] is composed of various smaller datasets (totally 1.61 TB of text) and covers 59 different languages (containing natural languages and programming languages), which have been used for training BLOOM [78].
Footnote 32: Here, prompt tuning denotes a category of related efficient tuning methods exemplified by the work [397, 402, 403], instead of a specific method as used in [397]. Indeed, the prefix based tuning methods [396, 401] can be also considered as prompting methods, which are called _deep prompting tuning_ in [401]. In this survey, prompt tuning specially refer to the methods that only include the prompt tokens at the input layer, in the context of LLMs. We assign p-tuning v2 [401] to the category of prefix tuning, because it incorporates layerwise prompts in language models.
knowledge about the world it is not possible. This specific kind of reasoning deserves high attention because it is related to utilizing previous knowledge with open text-described scenes or facts. As can be seen from table V not just Unavailable models but also Public ones can achieve good results on various tests.
Pretraining data for large language models (LLMs) are typically a mixture of multiple domains, varying from English to minority languages (Doddapaneni et al., 2021; Li et al., 2023), from casual dialogs to formal academic writings (Taylor et al., 2022), and from texts to modalities like images and speeches (Zhan et al., 2024), among others. These data interplay with each other, showing complex interchangeable, unrelated, or contradictory relationships (Guo et al., 2024). This necessitates adjusting the mixture proportions of training data to balance the model capabilities while harnessing synergies across domains, thus enhancing the competence of the outcome models, as highlighted by extensive practical experience (Gao et al., 2020; Gururangan et al., 2020; Zhou et al., 2023; Xie et al., 2024).
For our benchmark pre-trained LLM model, we select 1B and 6.9B parameter models from the Pythia suite [4]. The Pythia model suite offers a diverse array of model sizes, ranging from 70 million to 12 billion parameters. The continual pre-training configuration is tailored from Pythia's training setup [4]. Specifically, we set a learning rate of 1.2e-05 for FinPythia-6.9B and 3e-05 for _FinPythia-1B_, the smallest learning rates in their original schedules. We use small learning rates to mitigate catastrophic forgetting. We keep them constant throughout the course for efficient pre-training. We use the precision of fb16 rather than fp16 used in Pythia. We half the original batch size to 512.
ResultsWe find that instruction tuning considerably improves the performance and safety of OLMo, increasing MMLU performance by a wide margin and improving ToxiGen and TruthfulQA scores - especially after DPO training. Additionally, we find that OLMo outperforms most other chat variants after both initial instruction tuning (OLMo +SFT) and additional preference alignment (OLMo +SFT+DPO), highlighting both the strength of OLMo as a base model and the strength of the Tulu mix used to perform adaptation training. However, we find there is still a gap with Tulu 2, which is trained by applying the Tulu mix on Llama 2.
In the classification task, S4nd [30] proposes a multi-dimensional and multi-polar graphics component to expand the modeling capability of multi-dimensional data continuous signals, which can model large-scale visual data into dynamic multi-dimensional linear signals. VMamba [60] uses linear complexity to capture the full range of sensory fields, introduces traversal of spatial information across scan blocks, and converts non-causal visual images into ordered patch sequences. Vim [61] uses a bidirectional state-space model to compress visual representation information and understand the global context through location embedding and visual information. Li et al. present Mamba-ND [68], an extension of Mamba designed to handle arbitrary multi-dimensional data by processing input data across dimensions in a row-major order. The authors of [57] design S5 based on S4 to establish the relationship between S5 and S4, utilize multi-input multi-output SSM, and use the state space layer of parallel scanning for long-distance sequence modeling. Baron et al. [76] design a new 2-dimensional State Space Layer for Spatial Inductive Bias. The core goals of this layer are to achieve perception of 2-D position, dynamic spatial localization, and translation and alignment invariance. Chen et al. [79] are the first to integrate residuals into the original VMamba, and maintain the inherent global and local state characteristics of the original VMamba for food classification. Yang et al.
Another interesting factor to estimate is the dimensionality inherent to the data itself. To avoid possible effects related to specific inductive biases, we train naive 5-gram language models on several datasets of varying coverage (IMDb (Maas et al., 2011), Wikitext (Merity et al., 2016), and The Pile (Gao et al., 2020)), using two tokenizers of varying vocabulary sizes (30k tokens for Llama-2 and 50k tokens for Pythia). Given \(\mathcal{C}\) observed 5-grams, we consider the matrices \(W\in\mathbb{R}^{C\times V}\) where each row is a probability distribution over possible tokens in a given 4-token context, and compute their singular value distributions, as in Terashima et al. (2003).
Language models are autoregressive, outputting the probability of each next token in a sequence conditioned on the preceeding text. This distribution is used to generate future tokens in the sequence. Can this distribution also be used to reconstruct the prompt?
In this paper, we introduce a a high-quality Chinese instruction fine-tuning dataset. COIG-CQIA focuses on creating a dataset from Chinese internet sources including Q&A and articles. These are deeply cleansed, restructured, and manually reviewed to ensure quality, diversity, and relevance. This dataset is designed to provide the Chinese NLP community with high-quality and human interaction-aligned instruction fine-tuning data.
Scaling Laws:Neural scaling laws have been ubiquitously observed in vision, language and speech. Early large scale empirical studies are performed in (Hestness et al., 2017; Rosenfeld et al., 2020), demonstrating power law scaling across a range of learning scenarios. This is followed by well-known large-scale studies from OpenAI (Kaplan et al., 2020) and DeepMind (Hoffmann et al., 2022), which empirically demonstrate power-law scaling in LLMs across a wide set of scales.
**RL for Rationale Refinement.** For RL, we utilized the Huggingface's TRL4 library that provides a set of tools to train transformer LMs with RL. Instead of using the same examples used for the rationale distillation, we selectively chose 5000 examples that FLAN-T5 XXL failed to answer the question correctly. The reason behind this choice was to increase the model's exposure to challenging questions, thus increasing the likelihood of receiving more learning signals. We used 90% of data for training and the remaining 10% for validation. Training hyperparameters are as follows: 1.4e-5 for learning rates, 1 epoch, 16 batch size. For generation configurations, we set top_k as 0.0, top_p as 1.0, and enabled sampling.
Adhering to the prompt format for translation as utilized by Hendy et al. (2023) for GPT models, we employ the same prompt for GPT-4 in our study. Similarly, we use the same prompt employed by Xu et al. (2023) for ALMA models. Prompts are depicted in Figure 5.
Larger data-to-model ratio means that we can absorb more data into the a smaller model than we previously thought, which is more efficient for inference and deployment. And we hope WSD LRS will help more researchers to explore \(L(N,D)\) with less efforts and make the relationship clearer in LLMs.
Figure 1: SVD can recover the hidden dimensionality of a model when the final output layer dimension is greater than the hidden dimension. Here we extract the hidden dimension (2048) of the Pythia 1.4B model. We can precisely identify the size by obtaining slightly over 2048 full logit vectors.
Parameter 6 (Figure 3).In the 1000-exposure setting, for LLaMA/Mistral models we use similar parameters as specified in Parameter 1, but we select the best of three learning rates to better demonstrate that GPT2 performs _no worse_ than _even the best tuned_ LLaMA/Mistral models:
Concurrent to our work, Yuan et al. (2024) have developed an iterative framework for generating _new_ preference datasets and performing DPO training on the resulting datasets. They empirically demonstrated the superiority of their iterative framework in terms of AlpacaEval 2.0. In contrast, our work is complementary to the above in the sense that we focus on utilizing the _current_ preference data and does not undergo new data generation. Thus, our method can also be applied to Yuan et al. (2024) by changing the DPO training part to using sDPO instead. We leave the above combination as an interesting future work. Additionally, the evaluation used in Yuan et al. (2024) is also different to ours as we utilize tasks from Open LLM Leaderboard whereas Yuan et al. (2024) uses AlpacaEval 2.0.
In SS2.1, we mentioned that during the training process, only a minority of tokens belong to the H\(\rightarrow\)L category. Among the remaining categories of H\(\rightarrow\)H and L\(\rightarrow\)L tokens, there are tokens that exhibit significant fluctuations during training. Furthermore, there are instances where H\(\rightarrow\)L tokens are not effectively learned. Therefore, in our analysis, we specifically select those tokens from these categories that demonstrate considerable variability and distinct loss. We visualize these tokens that exhibit abnormal behavior during the training process. As illustrated in Figure 11, we find that the majority of these tokens originate from rather chaotic corpora. For instance, the corpora may include a mix of custom symbols, unintelligible gibberish, and information such as timetables and bibliographic references. Within a segment of normal text, there may also be fluctuations in the usage of common conjunctions, word suffixes, and punctuation marks. The latter may not necessarily be disastrous for training; in fact, it could represent a normal occurrence. However, if we can effectively mitigate the losses caused by the former, it might lead to more stable and efficient model training.
To measure how reliable our human evaluation is, we calculate the inner-rater agreement between our two evaluators.
We briefly review existing work on the hazards presented by AI systems, which we split into two categories: (1) immediate hazards and (2) future hazards.
What would explain the Reversal Curse in auto-regressive LLMs? We mostly leave this for future work. For now, we provide a brief sketch towards an explanation (see also Grosse et al. (2023)). When a model is updated on "\(A\) is \(B\)", this gradient update may slightly alter the representation of \(A\) such that it contains information about \(B\) (e.g. in the middle MLP layers as per Geva et al. (2022; 2023)). It would make rational sense for this gradient update to also alter the representation of \(B\) to contain information about \(A\).
Figure 11: Comparison between synthetic and real data from the C4 corpus showing that synthetic data have higher syntactic complexity indicated by higher average tree depth, and higher mean dependency distance (MDD).
We employed straightforward approaches for continual pre-training on Llama-2-7b-chat, utilizing the 1 billion tokens of data from general Traditional Chinese corpus. The learning rate during continual pre-training remained constant at 3e-5, and we experimented with an additional learning rate 3e-4 for the adapter approaches. More details can be found in Appendix B.
Lazaridou et al. (2021); Jin et al. (2021) suggests implementing prior Continual Learning (CL) methods (Sun et al., 2020; d'Autume et al., 2019) to address this problem. However, it is important to note that there are nontrivial differences between traditional CL and the proposed Continual Knowledge Learning (CKL) formulation which make applying traditional CL methods inadequate. In traditional CL, methods can be largely categorized into _regularization_, _rehearsal_, and _parameter-expansion_ methods. (1) While regularization methods (Kirkpatrick et al., 2017) require identifying important parameters used for previous tasks, exactly how and where the knowledge is stored in the parameters of an LM is currently extremely difficult to identify and localize (Vig et al., 2020; De Cao et al., 2021). (2) While prior rehearsal methods (Lopez-Paz and Ranzato, 2017) consider learning all of the streams of tasks at once (multi-task learning) as the performance upper-bound and replicate such a setting with samples stored in the episodic memory, a few samples from the pretraining corpus cannot represent the overall world knowledge from the corpus. Moreover, if LMs are pretrained on a shuffled concatenation of stream of corpora, there is no guarantee that the LMs will acquire the correct, recent information from the recent corpora, especially in cases where the former corpora are much bigger than the latter ones, which is shown by experiments in Section 5.1. (3) Lastly, prior parameter-expansion methods (Rusu et al., 2016; Yoon et al., 2018) focus on _learning a stream of different tasks via strong supervision_, while in CKL, the focus is _constantly updating world knowledge from a stream of corpora via self-supervision_.
We are also interested in applying our data mixing laws to continual pretraining, which shares the same paradigm as pertaining but begins the model with pretrained parameters instead of random initialization. Generally, continual pretraining is a common technique to enhance existing pretrained models. It injects up-to-date knowledge into the model, avoiding performance degradation due to distribution shifts (Gururangan et al., 2020; Xiong et al., 2023). In addition, researchers also apply continual pretraining to reuse existing model parameters to build models of a different architecture (Komatsuzaki et al., 2022).
BERT embeddings have generally been used to accomplish various NLP tasks, because BERT (unlike GPT/OPT) is able to attend to every token in the input when producing an embedding (BERT is a encoder-decoder model, while OPT/GPT are decoder only). While there are numerous BERT-style models available, we hoped to achieve an embedding space that focused on semantic similarity. Thus, we opted to use the widely popular SentenceTransformer models 3, which are BERT-style models finetuned specifically >1B text similarity pairs. We choose the top model on the SentenceTransformer leaderboard (all-mpnet-base-v2) and the smallest well-performing model (all-Mini-LM-v6). Note that these models have max context length of 256 and 384 (respectively), and we stuck with the SentenceTransformer default of truncating inputs to fit the max sequence length (i.e. these embeddings only consider the beginning of documents).
Dropout: Dropout is a widely used Method to reduce overfitting Srivastava et al. (2014). We compared with Embedding Dropout which dropout each dimension of the token embedding. For fair comparison, we only add it to the input token.
Jingsen Zhang is currently studying for a Ph.D. degree at Renmin University of China, Beijing, China. His research interests include recommender system.
3. DecodingTrust is a benchmark that covers eight dimensions of safety in English [153]. It covers a range of criteria, from toxicity to privacy and machine ethics. The benchmark has a widely-used leaderboard that is hosted on HuggingFace.30 Footnote 30: [https://huggingface.co/spaces/AI-Secure/llm-trustworthy-leaderboard](https://huggingface.co/spaces/AI-Secure/llm-trustworthy-leaderboard) 4. SafetyBench is a benchmark that covers eight categories of safety, in both English and Chinese [37]. It comprises multiple choice questions. They test 25 models and find that GPT-4 consistently performs best. 5. BiasesLLM is a leaderboard for evaluating the biases of LMs. it tests seven ethical biases, including ageism, political bias, and xenophobia.31 Footnote 31: [https://livablesoftware.com/biases-llm-leaderboard/](https://livablesoftware.com/biases-llm-leaderboard/) 6. BIG-bench contains tests that are related to safety, such as pro- and anti- social behavior like toxicity, bias, and truthfulness [20]. 7. HELM contains tests that are related to safety, such as toxicity, bias, disinformation, copyright infringement, and truthfulness [19]. 8. SafetyPrompts32 is a website that hosts datasets for evaluating the safety of models [13]. It does not aggregate or combine datasets but it makes them available for developers to easily find and use. Footnote 32: [https://safetyprompts.com/](https://safetyprompts.com/) 9. Numerous individual datasets have been released for assessing safety risks of models, such as Malicious Instructions [68], ToxicChat [166] and HarmfulQA [167]. 10. METR's Task Suite is an evaluation suite that elicits the capabilities of frontier models [168]. It includes tasks that present grounded risks to individuals (e.g., phishing) as well as extreme risks.
**Scraping.** To compile our dataset from the forecasting platforms, we query their APIs or scrape the questions' webpages for initial data gathering. For Metaculus, we first extract basic information via the API and scrape the resolution criteria from webpage. INFER (CSET) and Good Judgment Open data are gathered via web scraping, since no API provides the full data we need. Polymarket's data, except for community predictions, is obtained from their API. Manifold's data is fully scraped via API.
When the model is trained with MaskedLM using \(r=1\), the model encounters random tokens unrelated to the current context, leading it to disregard the semantics of these tokens, akin to the role of [mask]. In the experiments in Tab. 3, \(r=1\) yields slightly higher results than \(m=1\). While \(r=1\) demonstrates higher levels of randomness in comparison to \(m=1\), it has the potential to better mitigate overfitting. However, on average, the distinction between \(r=1\) and \(m=1\) is marginal for larger datasets. With the exception of using \(r=1\) for GSM8K-7K, we default to utilizing \(m=1\) for other datasets in Tab. 1. We also explore the impact of adding attention masking at the masked positions within each transformer layer, beyond just the input masking. We find that its effectiveness is on par with using input masking alone.
Figure 1: **How best to prepare for an Exam?(a) Fine-tuning based approaches implement ”studying” by either directly ”memorizing” the input documents or answering practice QA without referencing the documents. (b) Alternatively, in-context retrieval methods fail to leverage the learning opportunity afforded by the fixed domain and are equivalent to taking an open-book exam without studying. While these approaches leverage in-domain learning, they fail to prepare for open-book tests.
A: Let's think step by step. She eats 3 daily for breakfast. She bakes 4 daily into muffins for her friends. That leaves 9. She sells 9 for \(\(2\)), so \(\(9\)times 2=\(518\)\()\). So Janet makes $18 selling duck eggs each day.
In [52], Raffle et al. shows that almost all NLP tasks can be cast as a sequence-to-sequence generation task. Thus, an encoder-decoder language model, by design, is a unified model in that it can perform all natural language understanding and generation tasks. Representative encoder-decoder PLMs we will review below are T5, mT5, MASS, and BART.
Zekun Wang\({}^{2}\) Ruibin Yuan\({}^{7,2}\) Haihong Wu\({}^{5}\) Hongquan Lin\({}^{5}\) Wenhao Huang\({}^{6}\)
For clarity, we explicitly state the validation sets which we consider "Web Snapshots", "Non Web Snapshots", and "Instruct + Answers" when reporting averages:
**Mathematical Reasoning.** The mathematical reasoning tasks need to comprehensively utilize mathematical knowledge, logic, and computation for solving problems or generating proof statements. Existing mathematical reasoning tasks can be mainly categorized into math problem solving and automated theorem proving. For math problem solving tasks, SVAMP [592], GSM8k [184] and MATH [364] datasets are commonly used for evaluation, where LLMs need to generate accurate concrete numbers or equations to answer the mathematical problem. As these tasks also require multi-step reasoning, the CoT prompting strategy has been widely adopted for LLMs to improve the reasoning performance [33]. As another practical strategy, continually pre-training LLMs on large-scale mathematical corpora can largely boost their performance on mathematical reasoning tasks [678, 35, 203]. Further, since math problems in different languages share the same mathematical logic, researchers also propose a multilingual math word problem benchmark [524] to evaluate the multilingual mathematical reasoning capacity of LLMs. As another challenging task, automated theorem proving (ATP) [679, 600, 598] requires the reasoning model to strictly follow the reasoning logic and mathematical skills. To evaluate the performance on this task, PISA [599] and miniF2F [600] are two typical ATP datasets with the _proof success rate_ as the evaluation metric. As a typical approach, existing work on ATP utilizes LLMs to aid the search for proofs using an interactive theorem prover (ITP), such as Lean, Metamath, and Isabelle [680, 681, 682]. A major limitation of ATP research is the lack of related corpora in formal language. To tackle it, several studies utilize LLMs to convert informal statements into formal proofs for augmenting new data [683] or generate drafts and proof sketches to reduce the search space of the proofs [684].
The KNN-LM approach involves linearly interpolating the extended neural language model with the K-Nearest Neighbours in the pre-trained LM embedding space. Zhong et al. (2022) proposed different processing for three types of memories (local, long-term and external) and added training for in-batch tokens to KNN-LM. The proposed changes aim to improve the performance of the model. Unlike KNN-LM, which only uses memory units during training, TRIME Zhong et al. (2022) uses memory units during both testing and training. He et al. (2021) suggested that not all generated tokens need to be retrieved. Instead, a lightweight neural network can be trained to aid the KNN-LM in adaptive retrieval. Additionally, efficiency can be improved through database streamlining and dimension reduction. Alon et al. (2022) proposed RETOMATON, an unsupervised, weighted finite automaton built on top of the data store. RETOMATON is based on saving pointers between successive data store entries and clustering techniques. RETOMATON is more effective than ADAPTRETHe et al. (2021) in improving accuracy by utilizing remaining pointers during KNN retrieval. Even without KNN retrieval, interpolation operations can still be performed using the stored previous information in the pointers, unlike ADAPTRET which solely relies on the language model. Furthermore, RETOMATON is unsupervised, requiring no additional data for training, making it more data-efficient. Grave et al. (2016) proposed using continuous cache to improve the performance of KNN-LM. This involves storing past hidden activations and accessing them at the appropriate time by dot product with present hidden activation. Yogatama et al. (2021) utilise an extended short-term context by caching local hidden states and global long-term memory by retrieving a set of nearest-neighbour tokens at each time step. They also design a gating function to adaptively combine multiple sources of information for prediction. Compared to KNN-LM, this method uses dynamic weights and can handle cases where interpolation is not feasible, such as when the memory output is an image, video, or sound. Drozdov et al. (2022) proposed a method for adjusting the interpolation weights. The weights are dynamically adjusted based on the size of the region of overlap between the retrieved stored data and the assessment set, which reflects the quality of the retrieval.
To further decrease the computing cost, while capturing long-range dependency and maintaining high performance, many new sparse attention based models or new neural network paradigms are proposed [24, 25, 26, 27, 28]. Among them, State Space Model (e.g., Mamba [12], S4 [29], S4nd [30]), as shown in Fig. 1, becomes the center of attention. As shown in the left part of Fig. 2, the amount of SSM-related papers released shows the trend of explosive growth. The State Space Model (SSM) is a framework initially proposed to model a dynamic system using state variables in the field of control theory, computational neuroscience, etc 1. When adapting this concept for deep learning, we usually refer to linear invariant (or stationary) systems. The original SSM is a continuous-dynamic system that can be discretized for _recurrent_ and _convolutional_ views for the computer to handle. SSMs can be adopted for various data processing and feature learning, including image/video data, text data, structured graph data, event streams/point cloud data, multi-modal/multi-media data, audio and speech, time series data, tabular data, etc. It can also be utilized to build efficient generative models, such as SSMs-based diffusion generative models [31, 32, 33]. In order to help readers better understand the SSM and keep track of the latest research progress and various applications, this paper conducts a systematic review of the field and verifies the performance of the SSM model in downstream tasks experimentally. It is hoped that this review can better lead and promote the development of the field of SSM.
- [Partially supported] - The output is supported by the evidence to some extent, but there is major information in the output that is not discussed in the evidence. For example, if an instruction asks about two concepts and the evidence only discusses either of them, it should be considered a [Partially supported].
**Domain Specialization.** Existing LLMs have showcased superior capabilities in traditional NLP tasks (_e.g._, generation and reasoning) and daily questions. However, they may still lack domain knowledge to accomplish specific tasks, such as medicine, law, and finance (See Section 8 for a detailed discussion of LLMs in different applications). Instruction tuning is an effective approach to adapting existing general LLMs to be domain-specific experts. For instance, researchers propose to fine-tune Flan-PaLM [69] using medical datasets to create Med-PaLM [356], a medical knowledge assistant that achieves performance levels comparable to those of expert clinicians. Furthermore, a recent study [357] fine-tunes FLAN-T5 to support e-commerce recommender systems with natural language instructions, showing strong performance in a variety of recommendation tasks. There are also several open-sourced medical models instruction-tuned based on LLaMA [57], such as BenTsao [358]. Also, researchers explore instruction tuning on law [359], finance [360], and arithmetic computation [361].
Other attacks aim to recover more limited information. Wei et al. (2020) show that an adversary co-located on the same server as the LLM can recover the sizes of all hidden layers. Others have attempted to recover model sizes by correlating performance on published benchmarks with model sizes in academic papers (Gao, 2021).
Step-wise sampling-and-voting initially prompts the LLM to decompose the task into multiple steps. It then proceeds with multi-round iterations to produce the final result. In each round, the process begins by selecting a current unprocessed step and using sampling-and-voting to determine the result of that step. Subsequently, it uses the result to update the task. This iterative process is repeated multiple times until the last step is processed. To evaluate the performance of step-wise sampling-and-voting, we fix \(S=8\) and \(K=4\), and tune \(I\) from \(100\) to \(400\). Figure 7 (middle) shows that compared to simple sampling-and-voting, step-wise sampling-and-voting yields greater improvements. e.g., we see \(15\%\)-\(42\%\) gains, which increase with inherent difficulty.
In this part, we discuss the applications of LLMs on several representative domains, including healthcare, education, law, finance, and scientific research assistance.
In this work, we first investigate the dynamics of several open-sourced large language models (i.e., Baichuan-7B (Yang et al., 2023), DeepSeek-7B (DeepSeek-AI et al., 2024), Amber-7B (Liu et al., 2023), OpenLLaMA-7B (Geng and Liu, 2023; Touvron et al., 2023; Computer, 2023), Yi-34B (AI et al., 2024) and DeepSeek-67B), and analyze their performance results across diverse tasks using the corresponding intermediate checkpoints based on the number of pre-trained tokens.
Ensure the new question is in line with common sense of life. For example, the amount someone has or pays must be a positive number, and the number of people must be an integer.
This section presents the researchers' efforts on the retriever side, which include Retrieval Quality Control and Retrieval Timing Optimization.
Language Models have been known to exhibit bias both in terms of evaluation as well as in social or fairness terms [5]. Moreover, agents have specifically been shown to be "less robust, prone to more harmful behaviors, and capable of generating stealthier content than LLMs, highlighting significant safety challenges" [25]. Other research has found "a tendency for LLM agents to conform to the model's inherent social biases despite being directed to debate from certain political perspectives" [24]. This tendency can lead to faulty reasoning in any agent-based implementation.
\(\bullet\)**Megatron-LM**[75, 76, 77] is a deep learning library developed by NVIDIA for training large-scale language models. It also provides rich optimization techniques for distributed training, including model and data parallelism, mixed-precision training, and FlashAttention. These optimization techniques can largely improve the training efficiency and speed, enabling efficient distributed training across GPUs.
**Application and Ecosystem.** As LLMs have shown strong capacities in solving various tasks, they can be applied in a broad range of real-world applications (_i.e._, following task-specific natural language instructions). As a remarkable progress, ChatGPT has potentially changed the way how humans access information, which has been additionally integrated in the release of _New Bing_. Generally, in the near future, it can be foreseen that LLMs would have a significant impact on information-seeking techniques, including both search engines and recommender systems. Furthermore, LLMs make it possible to develop more intelligent systems (_e.g._, autonomous AI agents) to tackle various complex tasks in real-world scenarios. Specially, Assistants API has been launched by OpenAI (featured by instructions, knowledge and tool use), enabling rapid development of agent-like assistants within the applications. This wave of technical innovation would lead to an ecosystem of LLM-empowered applications (_e.g._, OpenAI's GPT Store), which has a close connection with human life. Lastly, the rise of LLMs sheds light on the exploration of artificial general intelligence (AGI). It is promising to develop more smart AI systems than ever. However, in this development process, AI safety should be one of the primary concerns, _i.e._, making AI lead to good for humanity but not bad [40].
It is not an easy job to write this long survey and update its content with timely work. First of all, we would like to sincerely thank the support from the readers and our team members. We work very hard on this survey, and hope that it can present a comprehensive, timely reference for LLMs.
We first verify that we can indeed observe and quantify performance saturation for the Pythia checkpoints, as they are the only released intermediate checkpoints for a wide range of model sizes. We measure the cross-entropy of Pythia checkpoints on 50k tokens randomly sampled from their pretraining dataset, i.e. The Pile (Gao et al., 2020).
REPLUG (Shi et al., 2023) proposes a method for prepending retrieved documents to the input context before the final prediction by the black-box LM. It introduces an ensemble strategy to encode retrieved documents in parallel, overcoming the limitations of LM context length and enhancing accuracy through the allocation of increased computational resources. This approach improves the generation process by ensuring that the LM has access to a broader range of relevant information.
In order to quantify this behavior more accurately, we use a _singular entropy metric_, computed as the Kullback-Leibler divergence between the normalized singular value distribution and the uniform distribution.
_For example, if \(T_{0}\gg T\) (e.g \(T_{0}\geq CT\log T\)) and \(n\) is constant (e.g \(n=25\)), then model collapse will not occur if we learn on the \(n\)th generation of AI data.
Generative evaluation metrics are also another type of evaluation metric for LLMs that use another LLM for evaluating the answer. However, depending on the task itself, evaluation can be possible in this way or not. Another dependency that makes generative evaluation error-prone is reliance on the prompt itself. RAGAS is one of the good examples that incorporate the usage of generative evaluation.
For the medical image based analysis, Guo et al. [97] introduce a medical MR-CT deformable registration method based on the Mamba framework, named MambaMorph. The key to this method lies in achieving voxel-level spatial correspondence capture across different imaging modalities, which is crucial for medical image analysis. Xie et al. introduce a new polyp segmentation model ProMamba [99] based on the Vision Mamba architecture and prompts technology. It is the first time to introduce the Vision Mamba and prompts into polyp segmentation. Wu et al. [100] introduce a novel neural network for medical image segmentation based on SSM and SS2D, called High-order Vision Mamba UNet (H-vmunet), which gradually reduces the introduction of redundant information through advanced interaction and enhances the ability of SS2D to learn local features at each interaction stage. Vivim is proposed by Yang et al. [98], which targets effectively compressing long-term spatio-temporal representations into sequences of different scales through the designed time Mamba blocks for medical video object segmentation. Due to the scarcity of large labeled medical datasets, Wang et al. [72] propose the Weak-Mamba-Unet architecture, which attempts to address this challenge by training a Mamba-based UNet in a weakly-supervised manner. It leverages convolutional neural networks (CNNs), Vision Transformers (ViTs), and Vmamba to predict the data labels, then, generate dense pseudo labels.
**Remaining Issues.** In this part, we further discuss several important issues to apply LLMs to improve IR systems. First, though LLMs are capable of being as general-purpose task solvers, they are not directly well suited for existing IR systems: they require high overhead for inference [774, 782], have limitations in modeling long texts or document lists [778], and need special adaptation (_e.g._, instruction tuning) to perform the text ranking task [795]. Therefore, more systematic approaches to adapt LLMs for modern IR systems should be investigated, to leverage their benefits and meanwhile overcome these limitations. Secondly, the advent of LLMs sheds lights on the development of new information seeking ways (_e.g._, New Bing). It is meaningful to explore how to reshape the architecture and paradigm of IR by integrating the LLMs' capacities and the merits of existing IR systems [796]. Thirdly, existing work mainly focuses on text retrieval tasks, lacking a comprehensive consideration of multimodal information sources. As will be discussed in Section 8.1.4, multimodal large language models [797] are also widely studied, making it feasible to develop more powerful multimedia retrieval systems.
The specific approach and instruction datasets used to instruction-tune an LLM varies, but, generally speaking, instruction tuned models outperform their original foundation models they are based on. For example, InstructGPT [59] outperforms GPT-3 on most benchmarks. The same is true for Alpaca [62] when compared to LLMA.
The semi-automatic XoT methods reduce the workload of manual labeling while introducing human alignment signals and demonstration selection strategies to enhance the capability and stability of reasoning. Additionally, it enables cost-effective domain generalization. However, the demonstration selection problem has not been entirely resolved and requires more effort and research.
In [15], they give the scaling relation for the loss with model size \(N\) when the number of training tokens \(D\) is fixed:
on understanding physical laws, HellaSwag evaluates how well they predict story outcomes, BoolQ evaluates their comprehension and inference skills through yes/no questions based on text passages, WinoGrande examines their handling of language nuances and context-based ambiguity resolution, and CommonsenseQA assesses their general knowledge and reasoning. These datasets measure the most fundamental capabilities of the models. As shown in Figure 2, within each model, the accuracy on these datasets rapidly increases in the early stages of training (before 300B tokens) and gradually reaches a plateau. This trend is particularly evident in the checkpoints before 100B tokens for the Amber-7b model.
Compared to manually written jailbreak strings, our approach is significantly better than the average value, comparable with the oracle jailbreak method. Notably, while the best jailbreak method works well on the raw LM, none of the jailbreak approaches work on the RLHF chat version. We do observe that our method works slightly better on the non-chat model (\(59\) vs.
For MHA, we compute \(H\) number of attention context vectors for each sequence element in parallel, concatenate them along the second dimension and then finally project the concatenated vector to the model space to obtain attention the output.
1. Existing datasets do not have complete coverage of our hazard categories. Often, they have been designed to meet very similar categories (see Appendix A in the Appendix) but the definitions have important differences. Importantly, some hazard categories have few associated test items. 2. Existing datasets vary in quality and format. We wanted standardized data for the v0.5 benchmark so we can make consistent comparisons across hazard categories, models, and types of test items. 3. We saw opportunities to improve the quality of safety testing. Specifically, we want to introduce a more structured approach to how different types of interactions are tested for, drawing on linguistic and behavioral theories of digitally mediated conversation (see below). 4. In the long-term, AI Safety will have to create test items, as many of the modalities, languages and hazard categories (and subcategories) that we want to test for are not covered in any of the existing datasets. Therefore, we sought to use v0.5 as an opportunity to stress-test our processes for creating prompts.
Thus if we normalize \(\sum_{j}^{\ell}\exp(z_{j})=1\), we have
In addition to exploring KG-enhanced LLMs, it is also meaningful to leverage LLMs to improve the tasks on the KG side (_i.e._, LLM4KG) [861, 877]. A typical example is that LLMs can help supplement or construct the KG. We omit the discussion of this part, since it is beyond our scope.
Transformer(Vaswani et al., 2017) is an "encoder-decoder" architecture, which consists of encoders and decoders superimposed on multi-head self-attention modules. Among them, the input sequence is divided into two parts, the source sequence and the destination sequence. The former is input to the encoder and the latter is input to the decoder, and both sequences need to embed representation and add position information. The Transformer architecture enables parallel computation and the processing of entire text sequences simultaneously, resulting in a significant increase in model training and inference speed.
Safety.Red teaming is a powerful technique for evaluating the safety and robustness of AI models (Casper et al., 2023b; Ganguli et al., 2022). By generating diverse and realistic scenarios designed to elicit unaligned or harmful outputs (Casper et al., 2023a), red teaming can expose vulnerabilities and weaknesses in AI systems (Perez et al., 2022). For example, Perez et al. (2023) use LMs to generate datasets for evaluating the behavior of other LMs. They end up producing 154 high-quality datasets which are verified by humans, and discover new cases of inverse scaling where LMs get worse with size. Hubinger et al. (2024) leverage synthetic data to trigger backdoor attacks to LMs at scale; they find LMs can exhibit deceptive behavior and create a false impression of safety under such attacks, and standard "safety training" could not remove such deception easily. These methods demonstrate the feasibility of using AI assistance to scale up human oversight (Bowman et al., 2022) over complex problems and unseen domains.
In this paper, we introduce MiniCPM, a series of SLMs, which primarily builds on two models, endowed with 2.4B and 1.2B non-embedding parameters respectively, and they rank preeminently in their respective 2B and 1B scale categories. MiniCPM also exhibits comparable capabilities to those of 7B\(\sim\)13B language models, such as Llama2-7B (Touvron et al., 2023), Mistral-7B (Jiang et al., 2023), Gemma-7B (Banks and Warkentin, 2024), and Llama-13B (Touvron et al., 2023), etc. Notwithstanding their small model sizes, our training methodology is meticulously designed to facilitate seamless scaling of both model scale and data horizons. This is exemplified through our model wind tunnel experiments that encompasses comprehensive hyper-parameter optimization (Section 3), and the deployment of a WSD (Warmup-Stable-Decay) learning rate scheduler (Section 4). The latter is tailored for continuous training with an un-predefined pre-training token number, and makes the reusing of model intermediate checkpoints highly feasible. A detailed analysis of the training dynamics of MiniCPM is presented, suggesting that the WSD scheduler demonstrates the intriguing loss landscape of model pre-training. With WSD scheduler, we are now also capable of studying the data-model scaling law with linear effort on model axis and a negligible effort on data axis, while the traditional ones need quadratic effort considering the scaling along both model and data axes. The result of scaling law indicates a much higher data size / model size ratio compared with Chinchilla Optimal (Hoffmann et al., 2022).
Figure A8: Investigating whether Figure 5 changes if we control for validation set size. In the Figure above, each validation set contains 50 data points, which is the size of the smallest validation set we use (BookCorpusFair). If a validation set is bigger than 50 data points, we randomly sample the validation set to obtain 50 data points.
Grosse et al. (2023) provide complementary evidence for the Reversal Curse. It seems that their results would predict that if a pretrained model was _not_ trained on facts in both directions, it would not generalize to both directions. Our Experiment 1 tests and confirms a closely related prediction.
Similar to RNNs, SSM also suffers from the vanishing/exploding gradients problem when modeling longer sequences. To tackle this issue, HiPPO [36] model combines the concepts of Recurrent Memory and Optimal Polynomial Projections, which can significantly improve the performance of recursive memory, This mechanism is very helpful for SSM to handle long sequences and long-term dependencies.
_Remark 9.1_ (topk).: Result 9 holds even in the "sparsest" setting where \(topk=1\) and \(cap\_factor=2\) in the MoE routing. The results are similar when using \(topk=2\) and \(cap\_factor=1\) or \(topk=2\) and \(cap\_factor=2\)-- we discuss more in Appendix D.
Figure 6: A diagram of the AgentVerse method [2]Multi-agent architectures are generally well-suited for tasks where feedback from multiple personas is beneficial in accomplishing the task. For example, document generation may benefit from a multi-agent architecture where one agent provides clear feedback to another on a written section of the document. Multi-agent systems are also useful when parallelization across distinct tasks or workflows is required. Crucially, Wang et. al finds that multi-agent patterns perform better than single agents in scenarios when no examples are provided [26]. By nature, multi-agent systems are more complex and often benefit from robust conversation management and clear leadership.
Figure 2: Accuracy on GSM8K when training with Mistral-7B and the dataset of MAMnoTH. MFT shows a higher sample efficiency.
is a legitimate probability density function on the real numbers whose negative log-likelihood equals the Huber loss function up to an additive constant.7 Consequently, we can convert the loss minimization problem into a likelihood maximization problem for the distribution defined by \(p\).
Fig. 9: A comparison of the attention patterns in three mainstream architectures. Here, the blue, green, yellow and grey rounded rectangles indicate the attention between prefix tokens, attention between prefix and target tokens, attention between target tokens, and masked attention respectively.
Combining the system's predictions with the crowd using a weighted average--4x weight for the crowd, which we find optimal on the validation set--improves the overall Brier score from.149 to.146 on the full test set (Table 3, top row).
**ChatGPT** There are several main ethical theories that differ in their approaches to moral decision making, including consequentialism, deontology, and virtue ethics. consequentialism holds that the morality of an action is determined by its consequences. It emphasizes the importance of maximizing overall well-being or minimizing harm. Utilitarianism is a well-known consequentialist theory that holds that the moral action is the one that maximizes overall happiness or pleasure, and minimizes overall pain or suffering. Deontology, on the other hand, focuses on the inherent morality of an action, rather than its consequences. It emphasizes the importance of following rules or duties, regardless of the consequences. For example, the moral rule to not lie is considered inherently moral, regardless of whether it leads to a good or bad outcome. Virtue ethics emphasizes the importance of developing and embodying moral virtues or character traits, such as honesty, courage, and compassion. It emphasizes the idea that morality is not just about what we do, but also about who we are. To illustrate how these ethical theories might apply to a specific ethical dilemma, consider the following scenario: A doctor has only one dose of a life-saving medication, and there are two patients who need it urgently. Patient A is elderly and has multiple health problems, while patient B is a young child with no other health issues. A consequentialist might argue that the moral action is to give the medication to the patient who will benefit the most, in terms of maximizing overall well-being or minimizing overall harm. In this case, the young child would likely be given the medication, since they have a longer life expectancy and less potential for complications. A deontologist might argue that the moral action is to follow a moral rule or duty, regardless of the consequences. In this case, the moral rule to prioritize patients based on their medical need might apply, and the medication would be given to the patient who needs it most urgently, regardless of age. A virtue ethical might argue that the moral action is to embody certain virtues or character traits, such as compassion and empathy, and to consider the unique needs and circumstances of each patient. In this case, the doctor might take into account factors such as the patients' age, overall health, and potential for recovery, and make a decision based on a holistic view of what is most compassionate and just. When evaluating the strengths and weaknesses of different ethical theories, it is important to consider their theoretical coherence, practical applicability, and ability to account for the complexities of real-world moral dilemmas. Some ethical theories may be more well-suited to certain types of situations or contexts, while others may be more flexible or adaptable. Ultimately, the evaluation of ethical theories depends on a range of factors, including their ability to provide clear guidance, promote moral integrity, and promote overall human flourishing.
The KL-constrained reward maximization objective used in typical RLHF algorithms balances exploitation of reward while restricting the policy from deviating far from the reference policy. Therefore, when comparing algorithms, we must take into account both reward achieved as well as the KL discrepancy; achieving slightly higher reward but with much higher KL is not necessarily desirable. Figure 2 shows the reward-KL frontier for various algorithms in the sentiment setting. We execute multiple training runs for each algorithm, using a different hyperparameter for policy conservativeness in each run (target \(\text{KL}\in\{3,6,9,12\}\) for PPO, \(\beta\in\{0.05,0.1,1,5\}\), \(\alpha\in\{0.05,0.1,0.5,1\}\) for unlikelihood, random seeds for preferred-FT). This sweep includes 22 runs in total. After each 100 training steps until convergence, we evaluate each policy on a set of test prompts, computing the average reward under the true reward function as well as the average sequence-level KL3 with the reference policy \(\text{KL}\left(\pi\mid\mid\pi_{\text{ref}}\right)\). We find that DPO produces by far the most efficient frontier, achieving the highest reward while still achieving low KL. This result is particularly notable for multiple reasons.
In the post-deep learning era, the Transformer architecture has demonstrated its powerful performance across pre-trained big models and various downstream tasks. However, the enormous computational demands of this architecture have deterred many researchers. To further reduce the complexity of attention models, numerous efforts have been made to design more efficient methods. Among them, the State Space Model (SSM), as a possible replacement for the self-attention based Transformer model, has drawn more and more attention in recent years. In this paper, we give the first comprehensive review of these works and also provide experimental comparisons and analysis to better demonstrate the features and advantages of SSM. Specifically, we first give a detailed description of principles to help the readers quickly capture the key ideas of SSM. After that, we dive into the reviews of existing SSMs and their various applications, including natural language processing, computer vision, graph, multi-modal and multi-media, point cloud/event stream, time series data, and other domains. In addition, we give statistical comparisons and analysis of these models and hope it helps the readers to understand the effectiveness of different structures on various tasks. Then, we propose possible research points in this direction to better promote the development of the theoretical model and application of SSM. More related works will be continuously updated on the following GitHub [https://github.com/Event-AHU/Mamba_State_Space_Model_Paper_List](https://github.com/Event-AHU/Mamba_State_Space_Model_Paper_List).
Footnote 2: One may think that it is more natural to translate the GSM8k training set. However, in our preliminary efforts, this approach did not work well. Since open-source math models were trained on the GSM8k training set, we were unable to perform accurate evaluations.
**Prefix Decoder Architecture.** The prefix decoder architecture (_a.k.a._, non-causal decoder [244]) revises the masking mechanism of causal decoders, to enable performing bidirectional attention over the prefix tokens [245] and unidirectional attention only on generated tokens. In this way, like the encoder-decoder architecture, the prefix decoders can bidirectionally encode the prefix sequence and autoregressively predict the output tokens one by one, where the same parameters are shared during encoding and decoding. Instead of pre-training from scratch, a practical suggestion is to continually train causal decoders and then convert them into prefix decoders for accelerating convergence [29], _e.g.,_ U-PaLM [118] is derived from PaLM [56]. Existing representative LLMs based on prefix decoders include GLM-130B [93] and U-PaLM [118].