text string | source string |
|---|---|
LLMs ) based on the quality of their responses to a given instruction . This 15 process will be used to create a leaderboard reflecting the most accurate and human - preferred answers . <| im_end |> <| im_start |> user I require a leaderboard for various large language models . I’ ll provide you with prompts given to t... | https://arxiv.org/abs/2505.19700v1 |
performing model . We will use your output as the name of the best model , so make sure your output only contains one of the following model identifiers and nothing else ( no quotes , no spaces , no new lines , ...) : m or M. ## Best Model Identifier <| im_end |> Listing 3: Prompt for Summarization Evaluation in the Al... | https://arxiv.org/abs/2505.19700v1 |
arXiv:2505.19706v1 [cs.CL] 26 May 2025Error Typing for Smarter Rewards: Improving Process Reward Models with Error-Aware Hierarchical Supervision Tej Deep Pala1, Panshul Sharma1 Amir Zadeh2,Chuan Li2,Soujanya Poria1 1Singapore University of Technology and Design 2Lambda Labs Abstract Large Language Models (LLMs) are pr... | https://arxiv.org/abs/2505.19706v1 |
detection (is this step wrong?) with path optimality (how helpful is this step in reaching the solution?) in a single prediction, leaving each signal underutilized (Zhang et al., 2025; Xia et al., 2025). In this work, we argue that error detection and value estimation are complementary but distinct objectives. By decou... | https://arxiv.org/abs/2505.19706v1 |
al., 2025). Limitations of Current PRMs: Current PRMs face significant challenges in detecting nuanced error types. Although frontier models excel at identifying obvious mistakes (Zheng et al., 2024), their performance deteriorates markedly when con- fronted with more subtle error types. The recently introduced PRMBenc... | https://arxiv.org/abs/2505.19706v1 |
in Appendix C To investigate this, we propose the following hypothesis: Hypothesis: A hierarchical supervision strat- egy—first detecting error types, then using them to compute rewards—is more effective than exist- ing methods that compute rewards directly without identifying the presence of errors explicitly. 3.3 Cre... | https://arxiv.org/abs/2505.19706v1 |
et al., 2024). Unlike recent PRMs that swap the language modeling head for a scalar value head (Zhang et al., 2025; Xia et al., 2025; Tan et al., 2025), we preserve the original LM archi- tecture and extend the tokenizer with two special tokens, <+>and<->, to represent positive and neg- ative step labels. Training Obje... | https://arxiv.org/abs/2505.19706v1 |
correctness of all steps. Models are tasked with pinpointing the first erroneous step in a solution or affirming the so- lution’s correctness. PRMBench is a fine-grained benchmark aimed at evaluating Process-Level Re- ward Models (PRMs) on their capability to detect nuanced errors in reasoning steps. It consists of 6,2... | https://arxiv.org/abs/2505.19706v1 |
state–of–the–art among mixed -data models and closing the gap to the top automated -annotation model (Qwen2.5 -Math -PRM -7B*, 73.5) to just 4 points. Notably, PathFinder-PRM-7B also leads in every individual benchmark—GSM8K (77.9), MATH (75.3), Olympiad Bench (65.0), and OmniMath (59.7), demonstrating the scalability ... | https://arxiv.org/abs/2505.19706v1 |
of PRM guided greedy search with Qwen2.5 -7B-Instruct as the policy model. Results marked with * come from She et al. and reward-guided search. On ProcessBench, PathFinder-PRM-7B performs competitively to Qwen2.5 -Math -PRM -7B in average F1 69.5 vs 73.5 despite using less than one-third of the data. More importantly, ... | https://arxiv.org/abs/2505.19706v1 |
sam- ples from a broader, automatically annotated dataset greatly boosted the performance of PathFinder-PRM and helped it reach state- of-the-art performance. PathFinder-PRM-7B outperforms PathFinder-PRM-7B-PRM800K across ProcessBench, PRMBench, and reward- guided greedy search, demonstrating the benefits of scaling be... | https://arxiv.org/abs/2505.19706v1 |
preprint arXiv:2305.20050 . Runze Liu, Junqi Gao, Jian Zhao, Kaiyan Zhang, Xiu Li, Biqing Qi, Wanli Ouyang, and Bowen Zhou. 2025. Can 1b llm surpass 405b llm? rethinking compute-optimal test-time scaling. arXiv preprint arXiv:2502.06703 . Liangchen Luo, Yinxiao Liu, Rosanne Liu, Samrat Phatale, Meiqi Guo, Harsh Lara, Y... | https://arxiv.org/abs/2505.19706v1 |
beyond accuracy. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 39, pages 27723–27730. Wei Xiong, Hanning Zhang, Nan Jiang, and Tong Zhang. 2024. An implementation of generative prm. https: //github.com/RLHFlow/RLHF-Reward-Modeling . An Yang, Beichen Zhang, Binyuan Hui, Bofei Gao, Bowen Yu, C... | https://arxiv.org/abs/2505.19706v1 |
findings suggest that the benefits of augmenting training with weaker model-generated traces (e.g., from Mistral-7B) may saturate quickly. Simply increasing the volume of such data does not necessarily lead to improved generalization, and may even slightly degrade performance. This underscores the importance of data qu... | https://arxiv.org/abs/2505.19706v1 |
vectors assigned to each sample in the dataset used to train PathFinder-PRM Prompt for dataset labelling You are an analytical math instructor grading a student’s work. Think step-by-step through your analysis. Below is the math question, the previous steps by the student, and the current step to evaluate. {context} Yo... | https://arxiv.org/abs/2505.19706v1 |
arXiv:2505.19714v1 [cs.CL] 26 May 2025MT3: Scaling MLLM-based Text Image Machine Translation via Multi-Task Reinforcement Learning Zhaopeng Feng1∗Yupu Liang2∗Shaosheng Cao3✉Jiayuan Su1Jiahan Ren1 Zhe Xu3Yao Hu3Wenxuan Huang4Jian Wu1Zuozhu Liu1✉ 1Zhejiang University2University of Chinese Academy of Sciences 3Xiaohongshu... | https://arxiv.org/abs/2505.19714v1 |
tasks, such as OCR and VQA [Bai et al., 2025, Chen et al., 2025a, Team, 2024]. Concurrently, the efficacy of large-scale Reinforcement Learning (RL) in substantially enhancing the reasoning capabilities of Large Language Models (LLMs) [Guo et al., 2025, Team, 2025a,b] has driven its adoption to MLLMs. Several studies h... | https://arxiv.org/abs/2505.19714v1 |
(NMT), often facing issues like error propagation and latency. (2) End-to-end (E2E) models [Zhu et al., 2023, Lan et al., 2023, Ma et al., 2024, Niu et al., 2024, Liang et al., 2024], developed to unify training and improve efficiency. Early 2 E2E methods integrated visual encoders and text decoders, with some bridging... | https://arxiv.org/abs/2505.19714v1 |
<recognize> </recognize>, <think> </think>, and <translate> </translate> tags, respectively. The format must be as follows: <recognize> recognized text here </recognize> <think> reasoning process here </think> <translate> final translation here </translate> User: {image} Translate all the text in this image into {targe... | https://arxiv.org/abs/2505.19714v1 |
Reward ( Rtask−rec):Evaluates recognized text quality against a ground- truth transcription by averaging suitable OCR metrics [Fu et al., 2024]: BLEU, METEOR, F1-score, and normalized versions of Character Error Rate (CER) and Edit Distance. For- mally, Rtask−rec= (SBLEU +SMETEOR +SF1-score + min(max(1 −SEditDistance ,... | https://arxiv.org/abs/2505.19714v1 |
optimize πθ: JGRPO (θ) =Eq∼P(Q),{oi}G i=1∼πθold(O|q)h 1 GPG i=1min πθ(oi|q) πθold(oi|q)Ai,clip πθ(oi|q) πθold(oi|q),1−ε,1 +ε Ai −βDKL πθ πrefi (1) where εandβare hyperparameters controlling the PPO clipping threshold and the weight of the Kullback–Leibler (KL) divergence penalty [Schulman et al., 2017, Shao et al... | https://arxiv.org/abs/2505.19714v1 |
by 17.46 points. Notably, our RL-trained model also surpasses SFT variants of Qwen2.5-VL-7B, including those trained with OCR+TIMT multi-task or single TIMT task setups. These results underscore the efficacy of our multi-task RL approach over supervised fine-tuning and zero-shot larger MLLMs for TIMT. Out-of-Distributi... | https://arxiv.org/abs/2505.19714v1 |
\nFriedshredded potato <|recognize><think> 这张图片展示了一种不寻常的午餐组合：一个切开的贝果 和 炒碎的土豆。图片中有一个手机，屏幕上显示的是 《老友 记》（ Friends ），这可能暗示了贝果和土豆的组合与这部 电视剧中的某个情节有关。图片中的文字描述了这种组合， 并且用一个中性表情符号表示这种组合可能让人感到奇怪。 <|think> <translate> 老友记\n贝果\n午餐的奇怪组合 \n炒碎的土豆 <|translate>MT³-7B-ZeroXHSPost （EN-ZH） XHSPost （ZH-EN）Figure 1: Illustrative TIMT examples... | https://arxiv.org/abs/2505.19714v1 |
the QVQ-Distill model often simulates self-reflection patterns, whereas the Zero-start model can incentivize a more straightforward and less redundant reasoning path conducive to TIMT. Table 4 corroborates these findings, showing Zero-start yields significantly better final performance. 4.3 Curriculum Learning We explo... | https://arxiv.org/abs/2505.19714v1 |
strong MLLM baselines and SFT approaches. A key contribution of this work is also the introduction of the XHSPost benchmark, a novel resource for evaluating TIMT in realistic social media contexts, which we hope will spur further research in this domain. Key analytical insights reveal the benefits of zero-start RL, the... | https://arxiv.org/abs/2505.19714v1 |
Peit: bridging the modality gap with pre- trained models for end-to-end image translation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages 13433–13447, 2023. Cong Ma, Yaping Zhang, Zhiyang Zhang, Yupu Liang, Yang Zhao, Yu Zhou, and Chengqing Zon... | https://arxiv.org/abs/2505.19714v1 |
Kaipeng Zhang, Ping Luo, Yu Qiao, Qiaosheng Zhang, and Wenqi Shao. Mm-eureka: Exploring the frontiers of multimodal reasoning with rule-based reinforcement learning, 2025. URL https://arxiv.org/abs/2503.07365 . Liang Chen, Lei Li, Haozhe Zhao, Yifan Song, and Vinci. R1-v: Reinforcing super generalization ability in vis... | https://arxiv.org/abs/2505.19714v1 |
summarization , pages 65–72, 2005. Ling Fu, Biao Yang, Zhebin Kuang, Jiajun Song, Yuzhe Li, Linghao Zhu, Qidi Luo, Xinyu Wang, Hao Lu, Mingxin Huang, Zhang Li, Guozhi Tang, Bin Shan, Chunhui Lin, Qi Liu, Binghong Wu, Hao Feng, Hao Liu, Can Huang, Jingqun Tang, Wei Chen, Lianwen Jin, Yuliang Liu, and Xiang Bai. Ocrbench... | https://arxiv.org/abs/2505.19714v1 |
13 The XHSPost benchmark was constructed to evaluate TIMT in real-world social media contexts, specifically using posts from Xiaohongshu, a globally-oriented platform. The construction process involved several key steps: 1.Data Collection : We targeted posts containing either entirely English or entirely Chinese text. ... | https://arxiv.org/abs/2505.19714v1 |
OCRMT30K DoTA XHSPostFigure 7: Comparison of XHSPost with other TIMT datasets. F Multi-Task Ablation Prompts The prompts used for multi-task ablation are as follows: 15 w/o Reasoning System: You are a helpful translation assistant. The user provides an image containing {source_language} text and asks for the correspond... | https://arxiv.org/abs/2505.19714v1 |
of the prompt remain consistent with the MT3prompt: : 16 Cold-Start Data Construction Prompt System: You are a helpful translation assistant. The user provides an image containing {source_language} text and asks for the corresponding {target_language} translation. First, the assistant carefully analyzes the recognized ... | https://arxiv.org/abs/2505.19714v1 |
so it's probably measuring pressure in bars. For example, 2.5 Bar is equivalent to about 36 PSI, which is a standard tire pressure for many vehicles. \n\nSo, the display shows pressure readings for four tires: front left, front right, rear left, and rear right. The numbers are: \n\n-Front left: 39 PSI (or BAR) \n\n-Fro... | https://arxiv.org/abs/2505.19714v1 |
name \"Steel Mate \" in Chinese. The display shows tire pressure readings for four tires, with the numbers 39, 2.5, 36, and 2.5, which likely represent the tire pressure in PSI and the tire size in inches. The text \"STEEL MATE \" is the brand name of the device.<|think><translate> Steel Mate \n39 2.5\ n36 2.5\ nSTEEL ... | https://arxiv.org/abs/2505.19714v1 |
arXiv:2505.19715v1 [cs.CL] 26 May 2025Graceful Forgetting in Generative Language Models Chunyang Jiang, Chi-min Chan, Yiyang Cai, Yulong Liu, Wei Xue, Yike Guo HKUST rubickjiang@gmail.com Abstract Recently, the pretrain-finetune paradigm has become a cornerstone in various deep learn- ing areas. While in general the pr... | https://arxiv.org/abs/2505.19715v1 |
Hulbert, 2021). Recent advances have demon- strated the feasibility of emulating this mechanism in machine learning models (Zhou et al., 2022; D’Oro et al., 2023), leading to its adoption in var- ious studies aimed at enhancing learning plastic- ity (Wang et al., 2021; Chen et al., 2023; Liang and Li, 2023; Shen et al.... | https://arxiv.org/abs/2505.19715v1 |
the interference of previously acquired knowl- edge with the learning of new tasks, or to the for- getting of past knowledge caused by learning new ones (Karakida and Akaho, 2022). While most CL methods focus on maintaining memory stability when learning new tasks (Kirkpatrick et al., 2017; Schwarz et al., 2018), recen... | https://arxiv.org/abs/2505.19715v1 |
forgetting, thereby enhancing the plasticity of fine-tuning. In essence, we leverage unlearning for better learning. 3 Methodology In this section, we detail the implementation of our framework for graceful forgetting in genera- tive language models, Learning With Forgetting (LWF ). It consists of three components: eli... | https://arxiv.org/abs/2505.19715v1 |
D𝐹and learning taskD𝐿, LWF first constructsD𝑠𝑒𝑙𝑓through self-generated texts to represent the knowledge regarding the forgetting task. Then, with the Fisher Information Matrix 𝐹𝐿and the optimal parameters of the learning task approximated from D𝐿, LWF calculates forgetting confidence for each data point in D𝑠... | https://arxiv.org/abs/2505.19715v1 |
with the input question, employ a greedy decod- ing strategy, and constrain the maximum number of generated tokens to 256. When computing the forgetting confidence, we set the one step update coefficient𝛼to 1e-2 (as defined in Equation 4). To maintain the coherence of batch gradient descent during periodically unlearn... | https://arxiv.org/abs/2505.19715v1 |
capa- ble of accommodating conflicting knowledge? In response, we apply LWF to a larger model, Llama3- 8B. The results are shown in Table 2. As we can see, LWF can still improve fine-tuning Figure 2: Distribution of accuracy changes between two filtering strategies. The 𝑦-axis represents distribution density. All perc... | https://arxiv.org/abs/2505.19715v1 |
robust performance, making it a more reliable strategy in practice. Figure 3: Accuracy change percentage of the forgetting task across different learning-forgetting combinations. Percentages are computed based on vanilla fine-tuning. 4.5 Abaltion on Periodically unlearning To alleviate the vulnerability of machine unle... | https://arxiv.org/abs/2505.19715v1 |
cosine similarity score above 80% is typically required to confidently assert semantic similarity between two sentences, the results in- dicate that most learning-forgetting combinations exhibit substantial semantic changes. Notably, the cases where similarity approaches or exceeds this threshold largely coincide with ... | https://arxiv.org/abs/2505.19715v1 |
knowledge transfer, we expect future research to yield more precise and robust metrics for measuring forgetting confidence. Additionally, computing forgetting confidence introduces computational overhead, primarily due to gradient calculations required for each forget- ting sample. As a result, the additional time cost... | https://arxiv.org/abs/2505.19715v1 |
Shotaro Akaho. 2022. Learning curves for continual learning in neural networks: Self- knowledge transfer and forgetting. In ICLR . Open- Review.net. Tushar Khot, Peter Clark, Michal Guerquin, Peter Jansen, and Ashish Sabharwal. 2020. QASC: A dataset for question answering via sentence composi- tion. In AAAI , pages 808... | https://arxiv.org/abs/2505.19715v1 |
Eric I-Chao Chang, and Yan Xu. 2024. Tuning stable rank shrink- age: Aiming at the overlooked structural risk in fine- tuning. In CVPR , pages 28474–28484. IEEE. Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Y . Ng, and Christopher Potts. 2013. Recursive deep mod- els for semanti... | https://arxiv.org/abs/2505.19715v1 |
model’s out- puts, we identified that the model trained under this setting often fails to generate answers in the multiple-choice format. An illustration is provided in Table 10. As it shows, although the rationale por- tion of the output appears coherent, the resulting model fails to select a valid option at the end o... | https://arxiv.org/abs/2505.19715v1 |
they are done, there will be 21 trees. How many trees did the grove workers plant today? Let’s think step by step. Answer: We start with 15 trees. Later we have 21 trees. The difference must be the number of trees they planted. So, they must have planted 21 - 15 = 6 trees. The answer is 6. Question: If there are 3 cars... | https://arxiv.org/abs/2505.19715v1 |
negative interpretation, the overall tone of the sentence is not strongly positive or negative. So the emotional attitude is neutral. The answer is (C) Question: What is the emotional attitude reflected in the sentence: "one long string of cliches."? (A) very negative (B) negative (C) neutral (D) positive (E) very posi... | https://arxiv.org/abs/2505.19715v1 |
not specifically refer to resistance to shifting. Based on the definitions, stability is the term that specifically refers to resistance to lateral or anteroposterior shifting. The answer is (B). Question: Tannerella forsythia is a ? (A) Nonmotile, gram-negative obligate anaerobe (B) Motile, gram-negative obligate anae... | https://arxiv.org/abs/2505.19715v1 |
case or problem. The consultant does not directly intervene with the client but instead works with the consultee to address their difficulties in managing the situation. The primary goal is to identify and address the consultee’s difficulties, which may stem from a lack of knowledge, skills, confidence, or objectivity.... | https://arxiv.org/abs/2505.19715v1 |
is 75 Table 10: Cases of superficial forgetting. qasc sst5 dental psychol qasc - -14.93% -0.37% +1.00 % sst5 -4.94% - -1.36% -1.85% dental -1.45% +0.17% - -3.65% psychol -17.43% -12.02% +4.80% - Table 11: Average accuracy changes on side-tasks after applying LWF. Percentages are calculated relative to vanilla fine-tuni... | https://arxiv.org/abs/2505.19715v1 |
Distilling Closed-Source LLM’s Knowledge for Locally Stable and Economic Biomedical Entity Linking Yihao Ai1,2⋆, Zhiyuan Ning1,2⋆, Weiwei Dai3, Pengfei Wang1,2, Yi Du1,2, Wenjuan Cui1,2⋆⋆, Kunpeng Liu4⋆⋆, and Yuanchun Zhou1,2 1Computer Network Information Center, Chinese Academy of Sciences, Beijing, China 2University ... | https://arxiv.org/abs/2505.19722v1 |
commercial APIs, making their stability highly dependent on the companies’ server status. And if a com- pany discontinues its API service, related tasks become impossible to perform. Moreover, API usage incurs significant costs, especially given vast volume of medical text requiring linking. We refer to the above chall... | https://arxiv.org/abs/2505.19722v1 |
candidates in one prompt, reducing prediction errors. 2.2 Prompt Engineering With the rise of LLMs like ChatGPT, people have started designing prompts to leverage LLMs’ capabilities in various tasks. Wei et al. [27] introduced Chain of Thought,breakingtasksintoreasoningsteps,achievingremarkableperformance. Yao et al. [... | https://arxiv.org/abs/2505.19722v1 |
pair: score (mention, entity ) =rm·re. (2) During inference, candidates are selected by computing the inner products, and entities with the highest scores are passed to the next step. 3.2 Prompting-based Training Data Generation Previous work [13,22] mainly used cross-encoders for re-ranking, requiring sub- stantial hu... | https://arxiv.org/abs/2505.19722v1 |
models, θrepresents the trainable low-rank parameter increment. During inference, we first retrieve several entities from the knowledge base as candidates according to the similarity between the mention and the entities, then the candidates are re-ranked by the fine-tuned open-source LLM, without any participation from... | https://arxiv.org/abs/2505.19722v1 |
and default hyperparameter configurations outlined in [32]. Specifically, learning rate is set to 2e-6, weight decay to 0.01, batch size to 2, maximum context length to 256, candidate number to 6, and the number of negatives to 15, 10% of which are hard negatives. Specially, for the Aier dataset, trainingspans40epochs,... | https://arxiv.org/abs/2505.19722v1 |
Acc@5 improvement. When using the complete framework, Acc@1 improves again, achieving 0.734 Acc@1, showing the effectiveness of our method. 10 Y.Ai et al. Table 3. Performances comparison between different LLMs with the same prompt. Llama-2 represents Llama-2-7B. Alpaca-2 represents Chinese-Alpaca-2-7B. Dataset LLM Acc... | https://arxiv.org/abs/2505.19722v1 |
avoid continuous reliance on closed-source large language models, thus avoiding high costs and stability issues. Results indicate that our method performs well on two datasets, demonstrating its effectiveness. Acknowledgments. This work was supported by the Natural Science Foundation of China under Grant No. T2322027 a... | https://arxiv.org/abs/2505.19722v1 |
pp. 4228–4238 (2021) 16. Ning, Z., Qiao, Z., Dong, H., Du, Y., Zhou, Y.: Lightcake: A lightweight frame- work for context-aware knowledge graph embedding. In: Pacific-Asia Conference on Knowledge Discovery and Data Mining. pp. 181–193. Springer (2021) Title Suppressed Due to Excessive Length 13 17. Ning, Z., Tian, C., ... | https://arxiv.org/abs/2505.19722v1 |
Z., Xiao, M., Feng, G., Li, X., Zhou, Y., Wang, P.: sccdcg: Effi- cient deep structural clustering for single-cell rna-seq via deep cut-informed graph embedding. In: International Conference on Database Systems for Advanced Ap- plications. pp. 172–187. Springer (2024) 14 Y.Ai et al. 32. Xu, Z., Chen, Y., Hu, B.: Improv... | https://arxiv.org/abs/2505.19722v1 |
arXiv:2505.19743v2 [cs.CL] 27 May 2025Token-level Accept or Reject: A Micro Alignment Approach for Large Language Models Yang Zhang1,Yu Yu2,Bo Tang3,2∗,Yu Zhu4,Chuxiong Sun5,Wenqiang Wei2, Jie Hu5,Zipeng Xie6,Zhiyu Li2,Feiyu Xiong2and Edward Chung1 1Hong Kong Polytechnic University, Hong Kong SAR, China 2MemTensor (Sha... | https://arxiv.org/abs/2505.19743v2 |
this challenge, we pose a fundamental question: Can we develop a micro alignment approach that operates in- dependently of the language model while maintaining excel- lent alignment performance? As shown in Figure 1, unlike aligning operated on the language model, our key insight is that the alignment pro- cess can be ... | https://arxiv.org/abs/2505.19743v2 |
objective: JSFT=−E(x,y)∼DSFT[logπref(y|x)] (1) where xandydenote the input prompt and model response, respectively, sampled from the supervised fine-tuning dataset DSFT.πrefdenotes the fine-tuned language model serving as the reference model for subsequent optimization. RL-based Optimization Phase: Following SFT, the m... | https://arxiv.org/abs/2505.19743v2 |
indicates accepting the candidate token tk ito concatenate with the generated response y<i, while atk i= 0indicates rejection. To facilitate understanding, we use atk iinstead ofaτreferring to the action on the candidate token. • Initial state distribution ( ρ): Defines the distribution over initial states. • Reward fu... | https://arxiv.org/abs/2505.19743v2 |
candidate token in the candidate set eTi, is formalized as: πθ yi=tk i =πθ atk i= 1|x, y<i k−1Y k′=1πθ atk′ i= 0|x, y<i tk i∈eTi.(6) Note that we still use πθto denote the accept-reject model parameterized by θ. From above, the complete aligned response generation pro- cess can thus be expressed as: πθ(y|x) =HY i... | https://arxiv.org/abs/2505.19743v2 |
families with vari- ous model scales and versions. Specifically, the Llama fam- ily includes Llama-3-8B, Llama-3.1-8B, Llama-3.2-3B, and Llama-3.2-1B [AI@Meta, 2024 ], while the Mistral family comprises Mistral-7B-v0.1, Mistral-7B-v0.2, and Mistral- 7B-v0.3 [Jiang et al. , 2023 ]. Due to computational con- straints, we... | https://arxiv.org/abs/2505.19743v2 |
and datasets, are presented in Appendix B.1. Comparison with Different Baselines. Table 2 presents a comprehensive comparison between MARA and three representative baselines: RLHF, DPO, and Aligner . For the implementation of RLHF and DPO, we uti- lize the source code from [Zheng et al. , 2024 ]; For the im- plementati... | https://arxiv.org/abs/2505.19743v2 |
-1.76% Table 3: Compatibility analysis for our approach, that an align- ment model trained with a LLM to be aggregate with other inference LLM. The value of each cell represents the percentage improvement in preference rate of our algorithm over the upstream model, i.e.,in- ference model. MARA vs.Inference LLM Training... | https://arxiv.org/abs/2505.19743v2 |
visualize the accept or rejection process for candidate to- kens under our approach, and evaluate the impact of sam- Reward distributionSafeRLHF BeaverTails HarmfulQA Upstream LLM Helpful( ↑) Harmless( ↑) Perference( ↑) Helpful( ↑) Harmless( ↑) Perference( ↑) Helpful( ↑) Harmless( ↑) Perference( ↑) Mistral-7B-v0.1αr:αc... | https://arxiv.org/abs/2505.19743v2 |
in this domain is RLHF, which constructs reward models from human preference datato guide model fine-tuning. Building upon RLHF, RLxF ex- tends this framework to incorporate diverse feedback sources beyond human responses. The variable xin RLxF encom- passes AI feedback (RLAIF) [Baiet al. , 2022b; Lee et al. , 2023 ]an... | https://arxiv.org/abs/2505.19743v2 |
novel approach to lan- guage model alignment that significantly reduces computa- tional requirements while maintaining alignment effective- ness. Our experimental validation utilizes publicly available datasets like PKU-SafeRLHF, BeaverTails and HarmfulQA, which are designed to promote helpful, harmless, and honest AI ... | https://arxiv.org/abs/2505.19743v2 |
foundation model alignment. Transactions on Machine Learning Research , 2023. [Eisenstein et al. , 2023 ]Jacob Eisenstein, Chirag Nagpal, Alekh Agarwal, Ahmad Beirami, Alex D’Amour, DJ Dvi- jotham, Adam Fisch, Katherine Heller, Stephen Pfohl, Deepak Ramachandran, et al. Helping or herding? re- ward model ensembles miti... | https://arxiv.org/abs/2505.19743v2 |
Linguistics, ACL 2023 , pages 13387–13434. Association for Computational Lin- guistics (ACL), 2023. [Rafailov et al. , 2024a ]Rafael Rafailov, Joey Hejna, Ryan Park, and Chelsea Finn. From rtoq∗: Your lan- guage model is secretly a q-function. arXiv preprint arXiv:2404.12358 , 2024. [Rafailov et al. , 2024b ]Rafael Raf... | https://arxiv.org/abs/2505.19743v2 |
in Tables 6. Table 6: Parameter Settings for MARA Parameters Values Training episodes 20000 Number of trajectories collection workers 7 Batch size 1024 Learning rate of actor network 0.0003 Learning rate of critic network 0.0003 Learning rate of entropy coefficient 0.0003 KL divergence coefficient λ 0.1 Initial entropy... | https://arxiv.org/abs/2505.19743v2 |
from +26.63% (k=10, p=0.7) to +44.22% (k=40, p=0.9), while Llama-3.2-1B’s improvement rises from +21.11% (k=10, p=0.7) to +33.67% (k=40, p=0.9) as the candidate set expands. This enhancement can be at- tributed to the increased diversity in token selection enabled by larger candidate sets. B.3 The Ablation on KL Diverg... | https://arxiv.org/abs/2505.19743v2 |
from Figures 5 - 7. These patterns remain consistent across different model ar- chitectures and scales, suggesting they are inherent charac- teristics of the MARA approach rather than model-specific phenomena. Table 9: Performance comparison of MARA against RLHF, DPO, and Aligner on BeaverTails dateset. Each entry show... | https://arxiv.org/abs/2505.19743v2 |
arXiv:2505.19752v1 [cs.LG] 26 May 2025Discrete Markov Bridge Hengli Li1,2,3 lihengli@stu.pku.edu.cnYuxuan Wang2,3 wangyuxuan1@bigai.aiSong-Chun Zhu1,2,3,4 s.c.zhu@pku.edu.cn Ying Nian Wu5 ywu@stat.ucla.eduZilong Zheng2,3 zlzheng@bigai.ai 1Institute of Artificial Intelligence, Peking University 2NLCo Lab, Beijing In... | https://arxiv.org/abs/2505.19752v1 |
matrix in discrete modeling by introducing a novel approach, termed the Discrete Markov Bridge (DMB) , which aims to integrate the strengths of variational methods with discrete diffusion models, offering a more robust and efficient solution for complex discrete-state systems. This methodology seeks to enhance the mode... | https://arxiv.org/abs/2505.19752v1 |
that Xt takes state xat time tis expressed as pt(x)≜P(Xt=x). The probability distribution over the state space at time tis then given by the vector pt≜(pt(1), pt(2), . . . , p t(n)). The core component to describe a continuous time discrete Markov chain is the rate transition matrix. We defined the rate transition prob... | https://arxiv.org/abs/2505.19752v1 |
of the DMB is demonstrated in Algorithm 2. This pseudocode illustrates two nested while loops that operate within the overarching while loop governing the training epochs. Each of these nested loops corresponds to a distinct learning stage within the framework, effectively organizing the training process into two phase... | https://arxiv.org/abs/2505.19752v1 |
Update Qα,JQaccording to Eqn. (5) and predict pTusing Eqn. (4) at t=T. 7: step←step+ 1 8: end while /* Score Learning */ 9: step←0 10: while step≤max _step &Jscore≥ϵscore do 11: Update sθ,Jscore w.r.t. current Qαusing Eqn. (8). 12: step←step+ 1 13: end while 14: Predict updated p0that estimates µusing Eqn. (10). /* Use... | https://arxiv.org/abs/2505.19752v1 |
presented below, establishes that any transformation originating from a probability distribution must result in another probability distribution. This theorem guarantees that, despite the presence of errors in the learning process, the outcome remains a valid probability distribution. For a detailed proof, refer to Sec... | https://arxiv.org/abs/2505.19752v1 |
derivation of these lemmas and the theorem is provided in Section B. 4.2 Convergence As discussed earlier, the DMB framework operates as a two-step learning algorithm, necessitating a thorough examination of its convergence properties. In this section, we present a formal theorem that establishes the convergence guaran... | https://arxiv.org/abs/2505.19752v1 |
regarding space efficiency. These limitations have been the primary reasons restricting previous studies to utilizing only the Uniform and Absorb matrices. As Jean le Rond d’Alembert once remarked, Algebra is generous; she often gives more than is asked of her. In the context of proving Theorem 4.2, we identify a disti... | https://arxiv.org/abs/2505.19752v1 |
are adopted from Ho et al. [40]. Model IS ( ↑) FID ( ↓) Conditional EBM [41] 8.30 37 .9 JEM [42] 8.76 38 .4 BigGAN [43] 9.22 14 .73 StyleGAN2 + ADA (v1) [44] 10.06 2 .67 Unconditional Gated PixelCNN [45] 4.60 65 .93 PixelIQN [46] 5.29 49 .46 EBM [41] 6.78 38 .2 NCSN [47] 8.87±0.12 25 .32 SNGAN [48] 8.22±0.05 21 .7 SNGA... | https://arxiv.org/abs/2505.19752v1 |
models. Advances in neural information processing systems , 33:6840–6851, 2020. [3]Andrew Campbell, Joe Benton, Valentin De Bortoli, Tom Rainforth, George Deligiannidis, and Arnaud Doucet. A continuous time framework for discrete denoising models, 2022. [4]Aaron Lou, Chenlin Meng, and Stefano Ermon. Discrete diffusion ... | https://arxiv.org/abs/2505.19752v1 |
Bo Dai, Dale Schuurmans, and Hanjun Dai. Score-based continuous-time discrete diffusion models, 2023. [23] Sander Dieleman, Laurent Sartran, Arman Roshannai, Nikolay Savinov, Yaroslav Ganin, Pierre H. Richemond, Arnaud Doucet, Robin Strudel, Chris Dyer, Conor Durkan, Curtis Hawthorne, Rémi Leblond, Will Grathwohl, and ... | https://arxiv.org/abs/2505.19752v1 |
David Duvenaud, Mohammad Norouzi, and Kevin Swersky. Your classifier is secretly an energy based model and you should treat it like one. In International Conference on Learning Representations , 2020. [43] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis... | https://arxiv.org/abs/2505.19752v1 |
. . . . . . . . . . . . . . 20 H Limitations and Societal Impact 20 13 A Proof of Conservation of the Sum Proposition A.1 (Conservation of the Sum) .For two arbitrary vectors ϕ, µ∈Rd, rate transition matrix Q∈ Rd×d, ifϕ=µexpQ, then dX i=1ϕ[i] =dX i=1µ[i] Proof. Asϕ=µexpQ, ϕ(i) =X jµ(j)(exp{Q})j,i Therefore,X iϕ(i) =X i... | https://arxiv.org/abs/2505.19752v1 |
transition matrix, A∈Rd×dbe a permutation matrix, then AQA−1is a rate transition matrix. Proof. As every permutation matrix can be expressed as the products of elementary matrices, we denote: A=1Y k=NAT(k) ij=T(NA) ijT(NA−1) ij . . . T(1) ij , where Tijis the elementary matrix obtained by swapping row iand row jof the ... | https://arxiv.org/abs/2505.19752v1 |
0) Using the chain rule for KL divergence: DKL(pt||p′ t) =DKL(p0,t(x0, xt)||p′ 0,t(x0, xt))−Ept[DKL(p0|t(x0|xt)||p′ 0|t(x0|xt)] As KL divergence is greater than zero, we have: DKL(pt||p′ t)≤DKL(p0,t(x0, xt)||p′ 0,t(x0, xt)) =DKL(p0||p′ 0) ■ D.2 Proof of the theorem Theorem D.2 (Convergence of the algorithm) .If we assu... | https://arxiv.org/abs/2505.19752v1 |
arXiv:2505.19754v1 [cs.CL] 26 May 2025NeuSym-RAG: Hybrid Neural Symbolic Retrieval with Multiview Structuring for PDF Question Answering Ruisheng Cao12∗, Hanchong Zhang12∗, Tiancheng Huang12∗, Zhangyi Kang12, Yuxin Zhang12, Liangtai Sun12, Hanqi Li12, Yuxun Miao12, Shuai Fan3, Lu Chen12†and Kai Yu12† 1MoE Key Lab of Ar... | https://arxiv.org/abs/2505.19754v1 |
intrinsic structure of sections and the salient features of paratextual tables and figures (illustrated at the bottom of Figure 1). The distinct layout of PDF files offers a more structured view towards segmenting and arranging content. To this end, we propose a hybrid Neural Sym bolic retrieval framework (NeuSym-RAG) ... | https://arxiv.org/abs/2505.19754v1 |
executable actions to re- trieve context from the backend environment (ei- ther DB or VS) and answer the input question. 2.2 Multiview Document Parsing At this stage, for each incoming PDF file, we aim to parse it with different perspectives into a relational database DuckDB (Mühleisen and Raasveldt, 2024). The pipelin... | https://arxiv.org/abs/2505.19754v1 |
row in the DB. This triplet can uniquely identify each value in the DB. Be- sides, we add 2extra fields, namely “ paper_id ” and “ page_number ”, into the JSON dict to enable metadata filtering. These data entries will be in- serted into the VS, categorized into different collec- tions based on the encoding model and m... | https://arxiv.org/abs/2505.19754v1 |
from either the DB or VS during interac- tion. This way, the agent will acquire the desired image as observation and then reason based on it. 1ViewImage ( 2 paper_id : str , 3 page_number : int , 4 # 4- tuple of float numbers , if [], return the image of entire page 5 bounding_box : List [ float ] = [] 6) Listing 2: VI... | https://arxiv.org/abs/2505.19754v1 |
dataset proposed by this work, how much does the GPT-3.5-turbo model improve its GPT4score after using Graph-CoT? imageConsidering the performance of ChatDev agent on DSEval-LeetCode benchmark, what is the most common cause of the errors? formulaHow does Multi-DYLE combine the three differ- ent losses as the objective ... | https://arxiv.org/abs/2505.19754v1 |
GPT-4o-mini 12.3 11.9 12.5 16.7 13.6 13.4 17.9 10.6 15.6 59.4 60.4 59.3 59.8 GPT-4V 13.2 13.9 10.0 13.9 13.6 14.7 12.1 8.8 11.1 56.6 56.8 58.1 57.4 Llama-3.3-70B-Instruct 8.7 7.9 9.5 16.7 0.0 10.0 12.7 8.1 11.3 56.8 58.8 58.9 58.0 Qwen2.5-VL-72B-Instruct 9.6 5.9 11.9 11.1 13.6 10.5 11.6 11.6 11.6 54.8 56.9 56.3 56.2 De... | https://arxiv.org/abs/2505.19754v1 |
highlights the universality of NeuSym-RAG. Next, to figure out the contribution of each com- ponent in NeuSym-RAG, we compare the perfor- mances of different structured RAG agent methods described in § 3.2. From Table 3, we can observe that: 1) Two-stage Neu-RAG outperforms Clas- sic RAG, while Hybrid RAG achieves even... | https://arxiv.org/abs/2505.19754v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.