text string | source string |
|---|---|
reinforcement learning, guided by an evaluator providing multidimensional clinical reward signals. Experimental results demonstrate that DoctorAgent-RL significantly outperforms traditional methods in diagnostic accuracy. These findings not only mark a paradigm shift in medical AI from static Q&A to dynamic reasoning, ... | https://arxiv.org/abs/2505.19630v1 |
Carroll, Charles Lau, Ryutaro Tanno, Ira Ktena, et al. Towards generalist biomedical ai. Nejm Ai , 1(3):AIoa2300138, 2024. [11] Tianyu Han, Lisa C Adams, Jens-Michalis Papaioannou, Paul Grundmann, Tom Oberhauser, Alexander Löser, Daniel Truhn, and Keno K Bressem. Medalpaca–an open-source collection of medical conversat... | https://arxiv.org/abs/2505.19630v1 |
, pages 1–9, 2025. [25] Xi Chen, Huahui Yi, Mingke You, WeiZhi Liu, Li Wang, Hairui Li, Xue Zhang, Yingman Guo, Lei Fan, Gang Chen, et al. Enhancing diagnostic capability with multi-agents conversational large language models. NPJ digital medicine , 8(1):159, 2025. [26] Junkai Li, Yunghwei Lai, Weitao Li, Jingyi Ren, M... | https://arxiv.org/abs/2505.19630v1 |
Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 , 2017. [40] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv prepr... | https://arxiv.org/abs/2505.19630v1 |
actual condition was hypothyroidism combined with iron-deficiency anemia, but the system failed to correlate the combined effects of heavy menstrual flow + cold extremities + fatigue, and misdiagnosed due to a lack of thyroid-related cases in the training dataset. Please check if your eyes appear more prominent than us... | https://arxiv.org/abs/2505.19630v1 |
Doctor Agent This section elaborates on the prompt design for the doctor agent during both training and inference phases. The complete prompt structure is illustrated in Figure 7. E Prompt for Patient Agent This section first details the patient agent’s implicit disease knowledge generation mechanism, with the complete... | https://arxiv.org/abs/2505.19630v1 |
you shouldn’t ask similar questions. 2. Comprehensively analyze the patient’s condition to provide an accurate diagno- sis and appropriate treatment recommendations. Rules: 1. You can only choose one of the options to respond, you cannot both answer questions and provide a diagnosis simultaneously. 2. Absolutely do not... | https://arxiv.org/abs/2505.19630v1 |
reasoning]</think><answer>[Your response]</answer> Figure 9: Prompt for Patient Agent Training and Inference. The diagnostic process unfolds through multiple rounds of interaction between the two agents. The patient’s self-report serves as the starting point for the first round of dialogue between the doctor and patien... | https://arxiv.org/abs/2505.19630v1 |
& Software. All training and evaluation were performed on a system featuring eight NVIDIA A100 80GB PCIe GPUs, an Intel Xeon Platinum 8369B 32-Core Processor, and 1.0 TB of RAM. For supervised fine-tuning, we utilized the LLaMA-Factory framework [ 41] to fine-tune with LoRA [ 42]. Our DoctorAgent-RL, is built upon the ... | https://arxiv.org/abs/2505.19630v1 |
arXiv:2505.19631v1 [cs.CL] 26 May 2025Segment First or Comprehend First? Explore the Limit of Unsupervised Word Segmentation with Large Language Models Zihong Zhang1, Liqi He2, Zuchao Li1,*, Lefei Zhang2, Hai Zhao3,Bo Du2 1School of Artificial Intelligence, Wuhan University, Wuhan, China 2School of Computer Science, Wu... | https://arxiv.org/abs/2505.19631v1 |
in NLP, that is, to test the understand- ing capabilities of a language model. Large Language Models (LLMs) are the best lin- guists. The emergence of LLMs has marked signif- icant advancements in NLU. And the capabilities of LLMs are no longer based on word segmenta- tion. However, word segmentation can serve as an in... | https://arxiv.org/abs/2505.19631v1 |
represents a new era in the development of word segmentation meth- ods. Based on LLMs trained on massive corpora, our framework named LLM-WS can explore the limit of unsupervised word segmentation. Specifically, in LLM-WS, we propose an un- supervised word segmentation method named Figure 2: Evaluations on datasets MSR... | https://arxiv.org/abs/2505.19631v1 |
an iterative process with local maximum strat- egy (Wang et al., 2011). nVBE relies on Varia- tion of Branching Entropy, enhancing performance through normalization and Viterbi decoding while simplifying the model by reducing parameters and thresholds (Magistry and Sagot, 2012). One of the disadvantages of goodness mea... | https://arxiv.org/abs/2505.19631v1 |
Model-Inspired Aho-Corasick Automaton The Aho-Corasick automaton (AC automaton), as defined by Aho and Corasick (1975), extends the concept of a Trie, a tree-like data structure com- monly used for storing strings where each node Figure 5: A Directed Acyclic Graph (DAG) for sequence segmentation, starting from "BOS" (B... | https://arxiv.org/abs/2505.19631v1 |
adopt a different approach. We define the previous state prev( w)as the closest final state on the path from the root to the current state w. For example, as shown in Figure 3, the previous state of "her" is "he". And the root, representing ϵ, is considered a special final state. Thus the previous state of "she" is ϵ. ... | https://arxiv.org/abs/2505.19631v1 |
patterns, it’s calculated as4 1+2+3+4because all of the patterns share common prefix ϵdenoting the empty string. Ambiguities within words, such as " 武汉市长" (the mayor of Wuhan city), might be wrongly seg- mented into " 武汉市/长" (Wuhan city/long) by simple unigram models like Jieba, where both " 武 汉市" (Wuhan city) and " 长"... | https://arxiv.org/abs/2505.19631v1 |
all other test datasets maintained their origi- nal splits. Our experiments primarily utilized the Qwen1.5 series of LLMs (Bai et al., 2023). We selected this series because it is reputed to excel in multilingual tasks, offering a wide range of param- eters from 0.5B to 110B and easy to employ. The diversity of paramet... | https://arxiv.org/abs/2505.19631v1 |
LLMs possess enhanced capabilities for generalization and comprehension in NLP. This trend indicates that the expansion in model capacity not only aids in handling more complex linguistic features but also improves adaptability to varied linguistic contexts. 4.3 From LLM-WS to LLACA Although LLMs demonstrate capabiliti... | https://arxiv.org/abs/2505.19631v1 |
Qwen1.5-7B) in Table 4. Table 4: Comparison between SLM (Sun and Deng, 2018) and LLACA on OOV Handling for MSR, PKU, and CTB Datasets Training-ModelTest MSR PKU CTB MSR-SLM-3 73.9 69.8 67.4 MSR-LLACA 86.7 75.2 70.0 PKU-SLM-3 70.8 76.6 69.2 PKU-LLACA 78.2 87.7 72.9 CTB-SLM-3 69.7 70.1 76.0 CTB-LLACA 77.0 75.5 88.0 Note ... | https://arxiv.org/abs/2505.19631v1 |
IEEE Transac- tions on Software Engineering , 15(8):1010–1016. Miguel Arevalillo-Herráez, Pablo Arnau-González, and Naeem Ramzan. 2022. On adapting the DIET ar- chitecture and the rasa conversational toolkit for the sentiment analysis task. IEEE Access , 10:107477– 107487. Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Ka... | https://arxiv.org/abs/2505.19631v1 |
Morris, Jr, and Vaughan R Pratt. 1977. Fast pattern matching in strings. SIAM journal on computing , 6(2):323–350. Krit Kosawat, Monthika Boriboon, Patcharika Chootrakool, Ananlada Chotimongkol, Supon Klaithin, Sarawoot Kongyoung, Kanyanut Kriengket, Sitthaa Phaholphinyo, Sumonmas Purodakananda, Tipraporn Thanakulwarap... | https://arxiv.org/abs/2505.19631v1 |
Wenjie Li. 2012. Fast on- line training with frequency-adaptive learning rates for chinese word segmentation and new word detec- tion. In The 50th Annual Meeting of the Association for Computational Linguistics, Proceedings of the Conference, July 8-14, 2012, Jeju Island, Korea - Volume 1: Long Papers , pages 253–262. ... | https://arxiv.org/abs/2505.19631v1 |
for chinese word segmentation. In Neural Information Processing - 23rd International Conference, ICONIP 2016, Kyoto, Japan, October 16-21, 2016, Proceedings, Part IV , volume 9950 of Lecture Notes in Computer Science , pages 345–353. Zhuosheng Zhang, Hai Zhao, Kangwei Ling, Jiangtong Li, Zuchao Li, Shexia He, and Guoho... | https://arxiv.org/abs/2505.19631v1 |
training and testing data. Statistics and Licenses for Datasets Table 5 pro- vides a summary of the datasets used for evaluation, while Table 6 details the licenses associated with each dataset. Except for BEST was a subuset contained 1,000 sentences randomly sampled with random seed 17 bynumpy.random.choice from the B... | https://arxiv.org/abs/2505.19631v1 |
others. Hence, we use F0as a normaliza- tion factor to measure consistency across datasets. 3https://github.com/hankcs/HanLP/tree/master Algorithm 1 Word Segmentation with LLACA 1:Input: Raw text data 2:Output: Segmented text 3:procedure LLACA(text) 4: Randomly shuffle the raw text into batches 5: foreach batch do 6: G... | https://arxiv.org/abs/2505.19631v1 |
at the state "she" and the next input is "r" without a corresponding edge, the automaton utilizes failure links to back- track until a valid node with the "r" edge is found or until it returns to the root. When the automaton reaches the state "her", not only is the pattern "her" itself recognized but the state of its f... | https://arxiv.org/abs/2505.19631v1 |
GitHub. Details of them are below: •Jieba4: Jieba is a highly popular Chinese text segmentation tool known for its ease of use and versatility. Jieba also allows for custom dictionary integration, making it adaptable for specific vocabularies or industry terms. •MeCab5: MeCab is a sophisticated morpho- logical analyzer... | https://arxiv.org/abs/2505.19631v1 |
Perplexity (the lower the better) and F- measure (the larger the better) on datasets of different languages. LLM-WS-Uni and LLACA’s construction were both conducted on GPT-4. CITYU MSR Model F↑ ppl.↓ F↑ ppl.↓ LLM-WS-Uni 83.7 121 84.1 94 LLACA 84.1 42 84.2 28 KWDLC BEST Model F↑ ppl.↓ F↑ ppl.↓ LLM-WS-Uni 82.7 51 72.1 14... | https://arxiv.org/abs/2505.19631v1 |
vintoQ 6: end for 7: while notQ.isEmpty() do 8: u←Q.dequeue() 9: fori←possible transitions from udo 10: v←child( u,i) 11: ifv̸=ϵthen 12: fail(v)←child(fail( u),i) ▷Update the failure pointer 13: Enqueue vintoQ 14: else 15: child( u,i)←child(fail( u),i) ▷Set virtual transition 16: end if 17: end for 18: end while 19:end... | https://arxiv.org/abs/2505.19631v1 |
Faster and Better LLMs via Latency-Aware Test-Time Scaling Zili Wang1,2*, Tianyu Zhang3∗, Lei Zhu3,†, Haoli Bai3, Lu Hou3, Shiming Xiang1,2,†, Xianzhi Yu3, Wulong Liu3 1School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China 2Institute of Automation, Chinese Academy of Sciences, Chi... | https://arxiv.org/abs/2505.19634v3 |
210212214 #Output token406080DeepSeek-R1-Distill-Qwen-32B 2324 Second405060Accuracy (%) 232425 Second50556065 2526 Second707580859095 24252627 Second406080Sequential scaling Parallel scalingFigure 2: Latency-aware test-time scaling on MATH-500 with different model types, with sequential scaling in red and parallel scal... | https://arxiv.org/abs/2505.19634v3 |
(TFLOP/GB)050100150200250300Attainable FLOPS (TFLOP/s) Roofline model with s1.1-32B 0 40 80 120 160 200 240 Batch size0500100015002000Throughput (toks/sec) Throughput of s1.1-32B 1 2 4 816 32 64128 256 #Branch020406080100120140Latency (ms)Latency of s1.1-32B sequence length 512 1024Figure 3: System state, latency, and ... | https://arxiv.org/abs/2505.19634v3 |
3.1 Latency-Aware Test-time Scaling The inference of LLMs on modern accelerators is a memory-bound process, constrained by mem- ory bandwidth. Nowadays, LLMs have billions of parameters (Bai et al., 2023; Yang et al., 2024; Team, 2025; Guo et al., 2025; Meta, 2024). In the autoregressive decoding, parameters (weights &... | https://arxiv.org/abs/2505.19634v3 |
One approach is em- ploying multiple concurrent branches Bfor the question, as shown in Figure 4 (b). For instance, when a 2048-token response fails to yield a cor- rect answer, users can infer more branches to gen- erate multiple responses of the same length and determine the final answer through majority vot- ing. Th... | https://arxiv.org/abs/2505.19634v3 |
our test-time scaling baseline. s1.1-7B is used as its draft model for speculative decoding. Also, we conduct rele- vant experiments on LLaMa-3.1-8B-Instruct (Meta, 2024) with Eagle3 (Li et al., 2025) as the draft model. For RL-based thinking model, we employ DeepSeek-R1-Distill-Qwen-32B (Guo et al., 2025) and QwQ-32B ... | https://arxiv.org/abs/2505.19634v3 |
However, the extent of performance improve- ment varies depending on the model type. Specifi- cally, for reasoning models like QwQ-32B, parallel scaling reaches 80.1%accuracy on the MATH-500 dataset within 30seconds, but sequential scaling requires 1.4×more time to achieve the comparable performance. Sequential scaling... | https://arxiv.org/abs/2505.19634v3 |
LLMs hat exhibit improve- ments from branches, increasing the number of branches yields huge performance gains. This dif- ference stems from the different token efficiency. For reasoning models like QwQ, accuracy improve- ments are achieved through long CoT, making bet- ter SD acceleration more advantageous within lim-... | https://arxiv.org/abs/2505.19634v3 |
memory access overhead. This eliminates the benefits of parallelism, leading to increased la- tency and sometimes even worse than the baseline. This suggests that a latency-optimal TTS strategy requires extremely fine-grained parameter tuning. 6 Conclusion In this paper, we propose to rethink test-time scal- ing in lat... | https://arxiv.org/abs/2505.19634v3 |
Lespiau, Laurent Sifre, and John Jumper. 2023. Accelerating large language model decoding with speculative sampling. arXiv preprint arXiv:2302.01318 . Xinyu Guan, Li Lyna Zhang, Yifei Liu, Ning Shang, Youran Sun, Yi Zhu, Fan Yang, and Mao Yang. 2025. rstar-math: Small llms can master math reason- ing with self-evolved ... | https://arxiv.org/abs/2505.19634v3 |
Lionel M Ni, and 1 others. 2024a. Openr: An open source framework for advanced reasoning with large language models. arXiv preprint arXiv:2410.09671 . Junlin Wang, Jue Wang, Ben Athiwaratkun, Ce Zhang, and James Zou. 2024b. Mixture-of-agents enhances large language model capabilities. arXiv preprint arXiv:2406.04692 . ... | https://arxiv.org/abs/2505.19634v3 |
γ=5. Figure 9: MATH-500: B=16, γ=5. GPQA-Diamond: B=32, γ=5. AIME24:B=32, γ=4. AIME25: B=32, γ=5. Figure 10: MATH-500: B=32, γ=5. GPQA-Diamond: B=16, γ=5. AIME24: B=8, γ=5. AIME25: B=16, γ=5. 27282975.077.580.082.585.087.590.092.5MATH500s1.1-32B 242526271020304050607080DeepSeek-R1-Distill-Qwen-32B 2526277580859095QwQ-3... | https://arxiv.org/abs/2505.19634v3 |
Interleaved Reasoning for Large Language Models via Reinforcement Learning Roy Xie†‡David Qiu†Deepak Gopinath†Dong Lin†Yanchao Sun† Chong Wang†Saloni Potdar†Bhuwan Dhingra†‡ †Apple‡Duke University Abstract Long chain-of-thought (CoT) significantly enhances large language models’ (LLM) reasoning capabilities. However, t... | https://arxiv.org/abs/2505.19640v1 |
the full reasoning trace before generating answers, introduces two critical limitations. First, it significantly increases time-to-first-token (TTFT), taking seconds or minutes for answer generation. This breaks the interaction flow in real-time AI applications such as conversational assistants, resulting poor user exp... | https://arxiv.org/abs/2505.19640v1 |
GPQA [35], and MMLU [14]. We summarize our key contributions as follows: •We propose a novel RL training paradigm that trains LLMs to alternate thinking and answering, inherently reducing Time-to-First-Token (TTFT) by over 80% on average. •We introduce a rule-based reward that provides consistent, dense feedback for in... | https://arxiv.org/abs/2505.19640v1 |
Concurrent work mainly focuses on leveraging external tools such as search engine [ 19,4,41,26] during the reasoning process. In contrast, we focus on model’s internal ability of generating verifiable intermediate answers, which can be later used as additional reward signal for training. 3 Training LLMs for Interleaved... | https://arxiv.org/abs/2505.19640v1 |
reason through problems step by step before providing an answer. You conduct your reasoning within <think></think> and share partial answers within <answer></answer> as soon as you become confident about the intermediate results. You continue this pattern of <think></think><answer></answer><think></think><answer></answ... | https://arxiv.org/abs/2505.19640v1 |
TTFT, we design a conditional reward strategy that incentivizes the model to generate correct intermediate answers early, in order to guide the reasoning toward the correct final answer. We apply a conditional reward scheme where intermediate rewards are only invoked when the model demonstrates foundational competence ... | https://arxiv.org/abs/2505.19640v1 |
us to comprehensively evaluate the robustness of our approach. More details about the datasets are provided in Appendix D. Models and Baselines. We conduct experiments using Qwen2.5 instruct models with 1.5B and 7B parameters. To comprehensively evaluate the effectiveness of our approach, we compare it against verious ... | https://arxiv.org/abs/2505.19640v1 |
Section 5). All experiments are conducted on eight H100 GPU with 80GB memory. More training details are provided in Appendix B. Main Results. The results in Table 2 demonstrate the benefits of interleaved reasoning. Our base interleaved approach (Interleave), without using intermediate rewards, maintains Pass@1 accurac... | https://arxiv.org/abs/2505.19640v1 |
in practical applications. We present a qualitative analysis of interleaved reasoning in Appendix F and examples in Appendix H. 5 Analysis and Discussions Impact of Intermediate Answers. Using the Qwen2.5-1.5B-Instruct model, we investigate how intermediate answers influence model performance and training dynamics. Fir... | https://arxiv.org/abs/2505.19640v1 |
one; encouraging the model to articulate and produce correct intermediate steps keeps the deductive chain intact and makes a correct final conclusion more likely. This trend indicates that interleaved reasoning not only offers practical speedups on TTFT but also improves overall reasoning, especially for harder multi-h... | https://arxiv.org/abs/2505.19640v1 |
steps and ulti- mately improve final accuracy. The conditional reward strategy thus works as intended: a fre- quent, always-on intermediate reward is not nec- essary – a targeted, conditional approach is suffi- cient to guide the model. Figure 4: Visualization of intermediate reward application rate during training. Th... | https://arxiv.org/abs/2505.19640v1 |
Melanie Subbiah, J. Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-V oss, Gretchen Krueger, T. Henighan, R. Child, A. Ramesh, Daniel M. Ziegler, Jeff Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Ma teusz Litwin, Scott Gray, Be... | https://arxiv.org/abs/2505.19640v1 |
learning. arXiv preprint arXiv:2501.12948 , 2025. [14] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Xi- aodong Song, and Jacob Steinhardt. Measuring massive multitask language understand- ing. ArXiv , abs/2009.03300, 2020. URL https://api.semanticscholar.org/CorpusID: 221516475 . [15] Dan ... | https://arxiv.org/abs/2505.19640v1 |
URL https://openreview.net/forum? id=v8L0pN6EOi . [29] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys , 55:1 – 35, 2021. URL http://dl.acm.org/citation... | https://arxiv.org/abs/2505.19640v1 |
language models. arXiv preprint arXiv:2503.16419 , 2025. [43] Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang Du, Chonghua Liao, Chuning Tang, Congcong Wang, Dehao Zhang, Enming Yuan, Enzhe Lu, Fengxiang Tang, Flood Sung, Guangda Wei, Guokun Lai, Haiqing Guo, ... | https://arxiv.org/abs/2505.19640v1 |
Neural Information Processing Systems , volume 35, pages 24824–24837. Curran Associates, Inc., 2022. [51] Chulin Xie, Yangsibo Huang, Chiyuan Zhang, Da Yu, Xinyun Chen, Bill Yuchen Lin, Bo Li, Badih Ghazi, and Ravi Kumar. On memorization of large language models in logical reasoning. https://arxiv.org/abs/2410.23123 , ... | https://arxiv.org/abs/2505.19640v1 |
the generated response. It checks whether the model properly alternates between thinking and answering phases using the designated tags ( <think></think> and<answer></answer> ). The reward is calculated as: rformat(y) =λf·1.0 if format is correct −1.0if format is incorrect(6) where “correct” format means all tags are ... | https://arxiv.org/abs/2505.19640v1 |
if one intermediate answer is incorrect. This design choice was intentionally made to strongly incentivize the model to generate all correct intermediate steps, rather than being satisfied with partial correctness. 16 Algorithm 1 Intermediate Reward Calculation 1:Input: Generated sequence y, ground truth intermediate a... | https://arxiv.org/abs/2505.19640v1 |
from [ 12]. To match with GPQA, we select a subset of 198 data points from domains requiring formal reasoning: college computer science, college mathematics, abstract algebra, formal logic, college physics, and machine learning. MATH. We use 198 data points from the level 5 subset of MATH [ 15], which are the most chal... | https://arxiv.org/abs/2505.19640v1 |
and informativeness of feedback, and (3) overall user experience. The LLM was instructed to mimic a human evaluator and assign scores for each criterion and to select a winner between the two methods for each example. The evaluation prompt is shown in Appendix F.1. We calculate the win rates for each method, as shown i... | https://arxiv.org/abs/2505.19640v1 |
tokens, it matches or exceeds Think-answer on every dataset (see main results in Table 2). Additionally, for every training recipe and model size we evaluated, correct solutions are consistently shorter than incorrect ones. This suggests that once the model has found a reliable path it stops generating unnecessary thou... | https://arxiv.org/abs/2505.19640v1 |
step, and arrives at the correct final answer. Question : A very special island is inhabited only by knights and knaves. Knights always tell the truth, and knaves always lie. You meet 5 inhabitants: Victoria, Mia, Sebastian, Ethan, and Olivia. "Mia is a knave and Ethan is a knave" - Victoria. Mia said that If Mia is a ... | https://arxiv.org/abs/2505.19640v1 |
in population among the cities of the United Kingdom. </think> <answer> 30th </answer> <think> Fifth, state the final answer. </think> <answer> 30th </answer> MMLU The model translates logic formulas stepwise, mapping each part to natural language and selecting the correct answer. Question : Use the following key to tr... | https://arxiv.org/abs/2505.19640v1 |
successfully applies interleaved reasoning to derive and combine gradients step by step, leading to the correct answer. Question : Suppose we have the following objective function: arg minw1 2∥Xw−y∥2 2+1 2γ∥w∥2 2. What is the gradient of1 2∥Xw−y∥2 2+1 2λ∥w∥2 2with respect to w? A)∇wf(w) = (XTX+λI)w−XTy+λw B)∇wf(w) =XTX... | https://arxiv.org/abs/2505.19640v1 |
<answer> The IR spectrum strongly suggests the presence of a carboxylic acid group (-COOH) due to the broad O-H stretch (3500-2700 cm−1) and the C=O stretch (1720 cm−1).</answer> <think> Third, analyze the1H NMR Spectrum. The signal at 11.0 ppm (singlet, 1H) is highly deshielded and is characteristic of a carboxylic ac... | https://arxiv.org/abs/2505.19640v1 |
shorter thinking and longer answers than necessary. Question : Reaction of solid A with 8 equivalents of gas B forms bright red product C. When C reacts with 2 equivalents of gas D, it produces the extremely hazardous product E. C reacts with water to reform A plus two different acids F and G. F is a strong acid while ... | https://arxiv.org/abs/2505.19640v1 |
arXiv:2505.19641v3 [cs.AI] 28 May 2025 S y n L o g i cS ynthesizing V erifiable R easoning Data at Sc alef or L earning L ogic al R easoning and Be y ondJunteng Liu1,2Yuanxiang Fan2Zhuo Jiang2Han Ding2Yongyi Hu2Chi Zhang2Yiqi Shi2 Shitong Weng2Aili Chen2Shiqi Chen3Yunan Huang2Mozhi Zhang2Pengyu Zhao2 Junjie Yan2Junxian... | https://arxiv.org/abs/2505.19641v3 |
number that needs to be filled in. Please solve it.Logic Instance GenerationDifficulty ControlPrompt FormalizationTask Selection & Parameter IdentificationBag of Logical Tasks Our Working Flow Cipher Grid 7x7 Figure 1 |The framework of logic data synthesis. The process begins with the selection of suitable tasks and th... | https://arxiv.org/abs/2505.19641v3 |
BBH (Suzgun et al., 2022) 23 ✗ ✗ Zebra Logic (Lin et al., 2024) 1 ✗ ✓ KOR-Bench (Ma et al., 2024) 5*✗ ✗ K&K (Xie et al., 2025a) 1 ✓ ✓ BBEH (Kazemi et al., 2025) 23 ✗ ✗ SynLogic 35 ✓ ✓ training. Surprisingly, conducting the mixed training on Qwen2.5-7B-Base model (Yang et al., 2024), incorporating SynLogic data improves... | https://arxiv.org/abs/2505.19641v3 |
35 tasks can be found in the Appendix A.1. 2.Parameter Identification For each task, we identify key parameters that control difficulty (e.g., grid size in Sudoku, or missing numbers in Math Path). These parameters form the basis for scalable and adjustable difficulty data synthesis. 3.Logic Instance Generation We form... | https://arxiv.org/abs/2505.19641v3 |
(a) 7B models on SynLogic -Easy Qwen2.5-32B-Instruct DeepSeek-R1-Distill-32B Models0.00.20.40.60.8Performance 0.120.33 0.280.66avg@8 pass@8 (b) 32B models on SynLogic -Hard Figure 2 |Evaluation of task difficulty across our dataset versions. (a) Shows the performance of 7B-scale models on the SynLogic -Easy dataset, wh... | https://arxiv.org/abs/2505.19641v3 |
19.6 82.0 57.5 on applying reinforcement learning techniques to the SynLogic dataset, leveraging the verifiable nature of logical reasoning tasks. In this section, we validate the effectiveness of reinforcement learning training on the SynLogic dataset using Qwen2.5-7B-Base and Qwen2.5-32B-Base models. 3.1. Setup Detai... | https://arxiv.org/abs/2505.19641v3 |
BBH (Suzgun et al., 2022), and the substantially more challenging BBEH (Kazemi et al., 2025). To investigate cross- domain generalization effects, we incorporate mathematics evaluations on MATH 500 (Hendrycks et al., 2021), AMC 2023, and AIME 2024. All evaluations are conducted in a zero-shot setting, with avg@8 metric... | https://arxiv.org/abs/2505.19641v3 |
code and mathematical data influences training efficiency on 7B models and enhances the Zero-RL capabilities for 32B models. 4.1. Setup Details Formathematicaltrainingdata, wedirectlyutilizethe17ksamplesprovidedinDAPO(Yuetal.,2025). For coding data, we assembled approximately 9k samples from various online coding platf... | https://arxiv.org/abs/2505.19641v3 |
3 |Performance comparison across multiple benchmarks. The evaluation metrics vary by dataset: BBEH (Kazemi et al., 2025) uses pass@1, while KOR-Bench (Ma et al., 2024), LiveCodeBench (LCB)(Jain et al., 2025), and GPQA-Diamond(Rein et al., 2024) use avg@4. AIME 2024 is evaluated using avg@8. All training configurations ... | https://arxiv.org/abs/2505.19641v3 |
Mix-2(Math+Coding) experiment, Zero-Mix-3 consistently delivers higher performance across all benchmarks. Specifically, Zero-Mix-3 shows a significant improvement of over 10 points on BBEH, 6 points on KOR-Bench, and over 2 points on the out-of-domain benchmark GPQA Diamond. These resultsstronglyvalidatethesignificantg... | https://arxiv.org/abs/2505.19641v3 |
Yuxiang Luo, Yuxiang You, Yuxuan Liu, Yuyang Zhou, Y. X. Zhu, Yanhong Xu, Yanping Huang, Yaohui Li, Yi Zheng, Yuchen Zhu, Yunxian Ma, Ying Tang, Yukun Zha, Yuting Yan, Z. Z. Ren, Zehui Ren, Zhangli Sha, Zhe Fu, Zhean Xu, Zhenda Xie, Zhengyan Zhang, Zhewen Hao, Zhicheng Ma, Zhigang Yan, Zhiyu Wu, Zihui Gu, Zijia Zhu, Zi... | https://arxiv.org/abs/2505.19641v3 |
Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R. Bowman. GPQA: A graduate-level google-proof q&a benchmark. InFirst Conference on Language Modeling , 2024. URL https://openreview.net/forum?id= Ti67584b98 . ByteDance Seed, Yufeng Yuan, Yu Yue, Mingxuan Wang, Xiaochen Zuo, Jiaze Chen, Lin... | https://arxiv.org/abs/2505.19641v3 |
SynLogic -Easy TheSynLogic -Hard dataset encompasses all 35 tasks, representing a challenging upper bound calibrated to the solvability thresholds of DeepSeek R1 and OpenAI-o3-mini. During our experiments with Qwen2.5-32B-Base, we observed consistent training accuracy gains across this comprehensive task set. However, ... | https://arxiv.org/abs/2505.19641v3 |
Training Prompt Template You are a helpful assistant. You always first think about the reasoning process in the mind and then provides the user with the answer.\nThe reasoning process and answer are enclosed within ‘<think>’ ‘</think>’ and ‘<answer>’ ‘</answer>’ tags, respectively, e.g.,\n<think>\nA detailed reasoning ... | https://arxiv.org/abs/2505.19641v3 |
AMC 2023. All evaluations of the figure use avg@8 scoring. B.2.1. Performance Analysis of Mixed Training with Math We analyze the training dynamics when combining our logical reasoning dataset with mathematical content. Figure 10 presents performance results across three mathematical benchmarks: MATH 500, AIME 2024, an... | https://arxiv.org/abs/2505.19641v3 |
arXiv:2505.19647v1 [cs.CL] 26 May 2025Select, Read, and Write: A Multi-Agent Framework of Full-Text-based Related Work Generation Xiaochuan Liu1,2,3, Ruihua Song1,2,3 *, Xiting Wang1,2,3 *, Xu Chen1,2,3 1Gaoling School of Artificial Intelligence, Renmin University of China, Beijing, China 2Engineering Research Center o... | https://arxiv.org/abs/2505.19647v1 |
Ge et al., 2021), or retrieved text spans (Li et al., 2023; Li and Ouyang, 2025), rather than leveraging the full texts. The lack of rich full-text information often prevents models from fully capturing the content and relationships among references, leading to fre- quent misinterpretations and hallucinations (Xu et al... | https://arxiv.org/abs/2505.19647v1 |
be categorized into two types: extractive and abstractive meth- ods. Extractive methods focus on selecting key sentences from cited papers and concatenating them to form the related work section (Hoang and Kan, 2010; Hu and Wan, 2014; Wang et al., 2018; Chen and Zhuge, 2019; Wang et al., 2019). Re- cent RWG models pred... | https://arxiv.org/abs/2505.19647v1 |
2022; Chen et al., 2023a; Wang et al., 2024b). We adopt an agent-based framework, however, existing agent-based methods primarily focus on question answering (QA) tasks, where the agent only needs to locate an answer and a single agent suffices. In contrast, in RWG tasks, the reading order can im- pact the model’s unde... | https://arxiv.org/abs/2505.19647v1 |
a (c) co-occurrence graph. and section name to be read at step t: (Rt, st) =Selector ((s0,1, . . . , s N,1), Mt−1, Ht−1), (1) Here, s0,1tosN,1denote the abstracts of all papers. Importantly, the selected section (Rt, st)must not already exist in Ht−1. When the selector deter- mines that no further reading is necessary,... | https://arxiv.org/abs/2505.19647v1 |
is crucial for RWG tasks. The graph structure is an intuitive way to de- scribe relationships. Therefore, we propose a novel graph-aware selector, which constrains the read- ing order within the graph, enabling the selector to capture the relationships among papers. Specifi- cally, we propose building two types of grap... | https://arxiv.org/abs/2505.19647v1 |
tent performance improvements across models with varying capabilities, we can validate the generaliz- ability of our framework. While the use of closed commercial LLMs is common in NLP research, it poses challenges for reproducibility. To address this concern, we ensure that all experiments con- ducted with Llama3-8B a... | https://arxiv.org/abs/2505.19647v1 |
with Extended Context Windows. Certain advanced LLMs are equipped with long input windows, en- abling them to process the full-text of all references simultaneously. We choose Claude-3-Haiku (200K-token) and GPT-4o (128K-token) as baselines. 5.2 Metrics To avoid the poor correlation with human judg- ments in traditiona... | https://arxiv.org/abs/2505.19647v1 |
tasks. (2) Feeding all textual data in a single pass is not op- timal. While many advanced LLMs claim to han- dle long inputs, their methods for extending input windows often come at the cost of performance. As shown in Table 1, for GPT-4o and Claude-3-Haiku, feeding all the content does not perform as well as providin... | https://arxiv.org/abs/2505.19647v1 |
C. 5.6 Different Input Configurations In addition to the abstracts, existing works also uti- lize introduction and conclusion (Chen and Zhuge, 2019; Deng et al., 2021) or RWS (Xing et al., 2020;Input ModelAvg. Overall Edges LLM (Graph) -based Llama3-8B 1.063 2.93 Intro. Llama3-8B Graph-Ci 1.163 3.29 &Claude-3-Haiku 1.4... | https://arxiv.org/abs/2505.19647v1 |
a target domain with limited labels. A prevalent strategy involves learning domain-invariant feature representations [6, 8, 9, 18], where methods like Domain-Adversarial Neural Networks (DANN) [9] and Deep CORAL [18] align feature distributions to minimize domain shift. However, these approaches often overlook domain-s... | https://arxiv.org/abs/2505.19647v1 |
may be overly elaborated by the model, leading to lower coherence and relevance in the generated re- lated work. Addressing this issue is a key focus for our future work. Our framework also requires that users provide a set of references in advance. Significant effort still needs to be spent on manually retrieving and ... | https://arxiv.org/abs/2505.19647v1 |
Zhang. 2024. Rethinking scientific sum- marization evaluation: Grounding explainable met- rics on facet-aware benchmark. arXiv preprint arXiv:2402.14359 . Alexis Chevalier, Alexander Wettig, Anirudh Ajith, and Danqi Chen. 2023. Adapting language models to compress contexts. In Proceedings of the 2023 Con- ference on Em... | https://arxiv.org/abs/2505.19647v1 |
document summarization of scientific articles. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages 8068–8074. Biswadip Mandal, Xiangci Li, and Jessica Ouyang. 2024. Contextualizing generated citation texts. In Proceedings of the 2024 Joint International Con- ference... | https://arxiv.org/abs/2505.19647v1 |
2020. Automatic generation of citation texts in scholarly papers: A pilot study. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , pages 6181–6190. Ziwei Xu, Sanjay Jain, and Mohan Kankanhalli. 2024. Hallucination is inevitable: An innate lim- itation of large language models.... | https://arxiv.org/abs/2505.19647v1 |
these guidelines, are generally able to operate with few errors. We analyze the final working memory contents and compute the retention ratio—the proportion of cited papers that remain in working memory by the end of the process. As shown in Table 6, in most cases, LLMs successfully retain the majority of cited papers ... | https://arxiv.org/abs/2505.19647v1 |
cate- gories: Introduction, Related Work, Methodology, Experiments, and Conclusion. We then calculate the proportion of these sections that are selected for reading, as shown in Figure 5. We do not include the abstracts because we provide the abstracts of all papers for the model. The results reveal that the selector m... | https://arxiv.org/abs/2505.19647v1 |
input configurations. Our proposed framework consistently improves the performance of all three base models across both settings. Prompt for Selector-Vanilla SystemPrompt: You are aresearch worker with excellent paperreading skills. UserPrompt: I am writing a scientific paper. Now I need to cite some reference papers a... | https://arxiv.org/abs/2505.19647v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.