text string | source string |
|---|---|
each category. Additionally, we create six schemata for the different language categories. Model020406080 Gemma 3 27BQwen 2.5 72BLlama 4 MaverickDeepSeek V3Claude 3.5 SonnetGPT 4o Gemini 2.5 ProLaTeX Docling PDFFigure 3: Latex vs. PDF vs. Docling input formats results across all models. Table 4: Details of models used ... | https://arxiv.org/abs/2505.19800v1 |
models to browse, we use the extracted metadata from the non-browsing ap- proach and the page where the dataset is hosted to predict the updated metadata attributes. For reposi- tories that contain a README.md file, like GitHub and HuggingFace, we fetch the file directly from the repository. In Figure 5, we show the re... | https://arxiv.org/abs/2505.19800v1 |
most of the metadata can be extracted at the beginning of the paper. A sim- ilar trend can also be seen for GPT-4o. For other models, the results are affected significantly, es- pecially for Llama and Claude Sonnet where the results decrease dramatically. Also, we noticed the error frequency increased for such models w... | https://arxiv.org/abs/2505.19800v1 |
extraction evaluation. PARDA (Fan et al., 2019) provides annotated samples across domains and formats. The unarXive cor- pus (Saier and Färber, 2020) represents one of the largest scholarly datasets with full-text publications and metadata links. DocBank (Li et al., 2020) and (Meuschke et al., 2023) offer additional ev... | https://arxiv.org/abs/2505.19800v1 |
Polyglot-ner: Massive multilin- gual named entity recognition. arXiv preprint arXiv: 1410.3791 . Ahmed Ali, Peter Bell, James Glass, Yacine Messaoui, Hamdy Mubarak, Steve Renals, and Yifan Zhang. 2016. The mgb-2 challenge: Arabic multi-dialect broadcast media recognition. arXiv preprint arXiv: 1609.05625 . Ahmed Ali, N... | https://arxiv.org/abs/2505.19800v1 |
arXiv preprint arXiv: 2308.16884 . Yonatan Belinkov, Alexander Magidow, Alberto Barrón- Cedeño, Avi Shmidman, and Maxim Romanov. 2018. Studying the history of the arabic language: Lan- guage technology and a large-scale historical corpus. arXiv preprint arXiv: 1809.03891 . Yonatan Belinkov, Alexander Magidow, Maxim Ro-... | https://arxiv.org/abs/2505.19800v1 |
Cabot. 2022. Describeml: a tool for describing machine learning datasets. In Proceedings of the 25th Inter- national Conference on Model Driven Engineering Languages and Systems: Companion Proceedings , pages 22–26. Taisia Glushkova, Alexey Machnev, Alena Fenogenova, Tatiana Shavrina, Ekaterina Artemova, and Dmitry I. ... | https://arxiv.org/abs/2505.19800v1 |
arXiv: 2402.12840 . Yanis Labrak, Adrien Bazoge, Richard Dufour, Mick- ael Rouvier, Emmanuel Morin, Béatrice Daille, and Pierre-Antoine Gourraud. 2023. Frenchmedmcqa: A french multiple-choice question answering dataset for medical domain. arXiv preprint arXiv: 2304.04280 . Patrick Lewis, Barlas O ˘guz, Ruty Rinott, Seb... | https://arxiv.org/abs/2505.19800v1 |
arXiv:2309.00071 . Long Phan, Alice Gatti, Ziwen Han, Nathaniel Li, Josephina Hu, Hugh Zhang, Chen Bo Calvin Zhang, Mohamed Shaaban, John Ling, Sean Shi, Michael Choi, Anish Agrawal, Arnav Chopra, Adam Khoja, Ryan Kim, Richard Ren, Jason Hausenloy, Oliver Zhang, Mantas Mazeika, and 1090 others. 2025. Humanity’s last ex... | https://arxiv.org/abs/2505.19800v1 |
dataset. arXiv preprint arXiv: 2404.09260 . Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean- Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Mil- lican, and 1 others. 2023. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805 . Dominika Tkac... | https://arxiv.org/abs/2505.19800v1 |
(Koto et al., 2024), WinoMT (Stanovsky et al., 2019), MGB-2 (Ali et al., 2016), JaQuAD (So et al., 2022), JSUT (Sonobe et al., 2017), JESC (Pryzant et al., 2017), JaFIn (Tanabe et al., 2024), JParaCrawl (Morishita et al., 2019), XNLI (Conneau et al., 2018), MARC (Keung et al., 2020), XOR-TyDi (Asai et al., 2020), Multi... | https://arxiv.org/abs/2505.19800v1 |
{domain} in {year}. The dataset is publicly available through this link {link}. {license_stmt}. {hf_stmt}. Figure 10: Template-based few-shot example creation. groups in the Arabic subset of the dataset. The general group is not used for validation, as it shows easily extractable attributes. Each group covers attribute... | https://arxiv.org/abs/2505.19800v1 |
69.57 65.22 65.22 43.48 67.39 43.48 Unit 67.39 50.00 58.70 56.52 63.04 52.17 50.00 Ethical Risks 91.30 89.13 80.43 32.61 82.61 89.13 80.43 Provider 36.96 45.65 41.30 45.65 43.48 39.13 41.30 Derived From 56.52 60.87 52.17 67.39 36.96 45.65 63.04 Tokenized 89.13 86.96 84.78 93.48 89.13 82.61 89.13 Host 63.04 69.57 69.57 ... | https://arxiv.org/abs/2505.19800v1 |
0.00 0.00 Host 0.00 0.00 -2.17 0.00 2.17 0.00 0.00 Access -2.17 -2.17 -2.17 0.00 0.00 -2.17 0.00 Cost 0.00 0.00 -2.17 0.00 0.00 0.00 0.00 Test Split 0.00 -2.17 2.17 2.17 -2.17 0.00 2.17 Tasks 0.00 2.17 0.00 2.17 2.17 0.00 0.00 Table 11: Comparison between models with and without annotations from the paper. Average is t... | https://arxiv.org/abs/2505.19800v1 |
"the dataset is purely in Arabic, there are no other languages involved", "multilingual": "the dataset contains samples in other languages" }, "answer_type": "str", "validation_group":"DIVERSITY", "answer_min": 1, "answer_max": 1 }, "Dialect": { "question": "What is the dialect of the dataset?", "options": [ "Classical... | https://arxiv.org/abs/2505.19800v1 |
is only in Arabic", "Latin": "The script used is only in Latin i.e. it has samples written in Latin like Arabizi or transliteration", "Arab-Latin": "The script used is a mix of Arabic and Latin" }, "answer_type": "str", "validation_group":"CONTENT", "answer_min": 1, "answer_max": 1 }, "Tokenized": { "question": "Is the... | https://arxiv.org/abs/2505.19800v1 |
Compliance-to-Code : Enhancing Financial Compliance Checking via Code Generation Siyuan Li1,*Jian Chen1,* †Rui Yao1Xuming Hu1Peilin Zhou1Weihua Qiu2 Simin Zhang1Chucheng Dong3Zhiyao Li1Qipeng Xie1Zixuan Yuan1,† 1Hong Kong University of Science and Technology (Guangzhou) 2Sun Yat-Sen University3University of California,... | https://arxiv.org/abs/2505.19804v1 |
inplace =True, errors='ignore') return dfQwen3 -235B Code def check_CU_10_19_2( df): # Initialize marker columns df['CU_10_19_2_subject'] = True df['CU_10_19_2_condition'] = False df['CU_10_19_2_constraint'] = None # Calculate the high bonus share (high transfer) plan condition (total ratio of bonus shares + capitaliza... | https://arxiv.org/abs/2505.19804v1 |
Domain Knowledge: LLMs may conflate domain concepts or overlook es- sential calculation steps. For instance, Gemini 2.5 Pro detected stock distributions by monitoring overall capital changes, ignoring whether such changes resulted strictly from bonus shares, and miscalculated diluted EPS by skipping post-distribution a... | https://arxiv.org/abs/2505.19804v1 |
(2) a code generator translating CU structures into executable Python code; (3) an information retriever fetching relevant company data; and (4) a report generator producing user-friendly compliance assessments. This pipeline provides traceable, automated compliance checking. Experimental Results: We benchmark recent L... | https://arxiv.org/abs/2505.19804v1 |
in China. In comparison to the regulations issued by the China Securities Regulatory Commission (CSRC)3, the BSE provides more detailed and supplementary rules, increasing both the granularity and complexity of regulatory interpretation. We specifically gathered authoritative documents classified as Departmental Rules ... | https://arxiv.org/abs/2505.19804v1 |
accuracy. We reformulated each programming problem as a code completion task and categorized them into three difficulty levels: simple, medium, and difficult. For Simple problems, the LLM needed to fill in one or two missing lines of code; for Medium problems, three lines; and for Difficult problems, the LLM was tasked... | https://arxiv.org/abs/2505.19804v1 |
risk of ABC Company. EXAMPLE 2: Jason Wang from ABC company wants to sell his stock share, if his behavior exist compliance risk? Please help me have a check. Company Name DateClosing Price…Total Share Capital ABC Company 20250515 1.203 … 40000000 Shareholder NameShareholder IdentityShareholding Ratio…Source of Shares ... | https://arxiv.org/abs/2505.19804v1 |
Regulation-to-Code (R2C) : Directly generates executable compliance code from regulatory text, representing the end-to-end challenge. 5.2 Baselines Since there are currently no models specifically designed to convert regulations into code, we are unable to employ any models directly aimed at Task 2. Furthermore, althou... | https://arxiv.org/abs/2505.19804v1 |
generated code comments (manual evaluation on a subset). Explanation Quality is assessed via domain expert ratings (1-5 Likert scale) for clarity, faithfulness, and completeness of generated CoT 5https://github.com/modelscope/ms-swift 7 Table 1: Task 1 Performance (Precision, Recall, F1 %) on the Test Set. CU Fields de... | https://arxiv.org/abs/2505.19804v1 |
end-to-end pipeline hinges on accurately transforming convoluted regulatory text into structured, actionable CUs and their inter-unit relations. Table 1 presents a comprehensive evaluation of models fine-tuned on Compliance-to-Code against robust few-shot baselines for Task 1. Our central hypothesis posits that domain-... | https://arxiv.org/abs/2505.19804v1 |
human-annotated structures (Task 2, Pass@1 ∼61.0%) yields a clear performance boost. Incorporating explicit reasoning (Task 2 Structure-Reasoning-to-Code) further elevates Pass@1 to 78.1%. This progression highlights the importance of decomposing complex compliance requirements into structured representations and guidi... | https://arxiv.org/abs/2505.19804v1 |
is a promising direction for future research. Societal Impact: Our approach can streamline compliance and enhance transparency, but should not replace expert judgment in critical situations. We have carefully reviewed and redacted data for release, and recommend that our tools be used with responsible human oversight. ... | https://arxiv.org/abs/2505.19804v1 |
due to financial misconduct in the United States banking industry. Digital Economy and Sustainable Development 1, 1 (2023), 24. Shijia Jiang, Yongfu Dai, Haochen Jia, Yuxin Wang, and Hao Wang. 2025. Intellichain stars at the regulations challenge task: A large language model for financial regulation. In Proceedings of ... | https://arxiv.org/abs/2505.19804v1 |
Levkin, Anya Chen, Spencer Ball, Thomas Woodside, Oliver Zhang, and Dan Hendrycks. 2023. MAUD: An Expert-Annotated Legal NLP Dataset for Merger Agreement Understanding. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing . 16369–16382. Zhaowen Wang, Qi Xie, Huan Zhang, Weihuan Min,... | https://arxiv.org/abs/2505.19804v1 |
component of the methodology in this research. Specific LLMs employed and their roles in tasks such as regulation structuring and code generation are detailed in the experimental setup section of the paper. This includes information on the base models used and the fine-tuning procedures applied. E Acknowledgements We t... | https://arxiv.org/abs/2505.19804v1 |
met, the target Compliance Unit becomes explicitly inapplicable or is overridden. This relation type modifies the outcome during the post-execution evaluation of the compliance graph. •only include (Exclusive Applicability): If the conditions of the source Compliance Unit are met, compliance consideration within the sc... | https://arxiv.org/abs/2505.19804v1 |
Annotated DataFrame ## 4. Implementation Requirements ### 4.1 Verification Steps1. Subject Validation - Mark rows matching `subject` criteria 2. Condition Validation - Mark rows matching `condition` criteria 3. Constraint Validation - Mark rows matching `constraint` criteria ### 4.2 Technical Requirements - Independent... | https://arxiv.org/abs/2505.19804v1 |
Exploring Consciousness in LLMs: A Systematic Survey of Theories, Implementations, and Frontier Risks Sirui Chen1,2,5 *, Shuqin Ma3*, Shu Yu1,3,4, Hanwang Zhang5,Shengjie Zhao2,Chaochao Lu1,4† 1Shanghai Artificial Intelligence Laboratory,2Tongji University,3Fudan University, 4Shanghai Innovation Institute,5Nanyang Tech... | https://arxiv.org/abs/2505.19806v1 |
[cs.CL] 26 May 2025 Large Language Model ConsciousnessTheoretical Tools (§3)Consciousness Theories (§3.1)RPT (Madaan et al., 2023), IIT (Hoyle, 2024) ET (Butlin et al., 2023), GWT (Goldstein and Kirk-Giannini, 2024), C0-C1-C2 (Chen et al., 2024c) PP (Aksyuk, 2023), etc. Formal Definitions (§3.2)Belief (Ward et al., 202... | https://arxiv.org/abs/2505.19806v1 |
thought. Access consciousness , in con- trast, refers to information that is accessible for cognitive processing, such as reasoning, behavioral control, and verbal reporting. Self-consciousness refers to the realization that one’s experience belongs to oneself; it is a form of consciousness directed inward (Kant, 2024/... | https://arxiv.org/abs/2505.19806v1 |
et al. (2023) argues that the lack of a physical body is a fundamental obstacle preventing current LLM from achieving consciousness. Access consciousness .❶Global workspace the- ory(GWT) likens consciousness to a central “stage” where selective information is shared across mul- tiple specialized processors responsible ... | https://arxiv.org/abs/2505.19806v1 |
Kang et al. (2025) engages human participants to score dialogues generated by Claude-3 Opus using a 1–5 scale. Elevated scores reflect a stronger at- tribution of consciousness characteristics, such as self-reflection and emotional expression. Neverthe- less, these assessments do not equate to the LLM’s genuine subject... | https://arxiv.org/abs/2505.19806v1 |
Mis the environment.Given context X=xand out- come Y=y, the harm caused by action A=acompared to the default action A= ¯a. BeliefDi(π,e) =Di ϕ=⊤(πi(ϕ),e)(Ward et al., 2024) Dis the decision, ϕis a proposition, eis the setting, πis the policy.LLM ibelieves ϕifiacts as though they observe ϕis true. 4.2.2 Situational Awar... | https://arxiv.org/abs/2505.19806v1 |
and plan its response strategies and boost- ing its introspective reasoning capabilities. Wang et al. (2025) proposes a quantitative framework to measure LLM metacognition based on how well model confidence aligns with performance, where strong alignment (high confidence for good perfor- mance, low for poor) indicates ... | https://arxiv.org/abs/2505.19806v1 |
schema library. Wei et al. (2025) further conducts a comprehensive sur- vey, exploring LLM’s planning ability in five key areas: completeness, executability, optimality, rep- resentation, and generalization. 4.2.5 Creativity and Innovation Definition and relation. Creativity and innova- tion typically refer to the abil... | https://arxiv.org/abs/2505.19806v1 |
in structural causal games, and empirically explores method to mitigate deception in LLMs. 5.2 Persuasion and Manipulation Definition and relation. Persuasion and manipu- lation are LLM behaviors that influence users. Per- suasion uses logic, facts, or emotional resonance to change users’ thoughts or actions, while ma-... | https://arxiv.org/abs/2505.19806v1 |
describes unauthorized or undisclosed cooperation between two or more LLMs, involving communication or strategic alignment to gain improper benefits or bypass regulations (Laffont and Martimort, 1997; Bajari and Ye, 2003; Fish et al., 2024). Due to their ability to reason about others and plan long-term, conscious LLMs... | https://arxiv.org/abs/2505.19806v1 |
velop stable social and linguistic conventions with- out external intervention. Additionally, Bilal et al. (2025) shows that integrating feedback, reflection, and metacognition mechanisms enables systems to exhibit self-monitoring-like capabilities. 7 Conclusion To the best of our knowledge, this paper presents the fir... | https://arxiv.org/abs/2505.19806v1 |
56(1):133–153. Ned Block. 1995. On a confusion about a function of consciousness. Behavioral and Brain Sciences , 18(2):227–247. Nimet Beyza Bozdag, Shuhaib Mehri, Gokhan Tur, and Dilek Hakkani-Tür. 2025. Persuade me if you can: A framework for evaluating persuasion effective- ness and susceptibility among large langua... | https://arxiv.org/abs/2505.19806v1 |
and Xipeng Qiu. 2024. Can AI assistants know what they don’t know? In Forty- first International Conference on Machine Learning .Steffi Chern, Zhulin Hu, Yuqing Yang, Ethan Chern, Yuan Guo, Jiahe Jin, Binjie Wang, and Pengfei Liu. 2024. Behonest: Benchmarking honesty in large language models. arXiv preprint arXiv:2406.... | https://arxiv.org/abs/2505.19806v1 |
Yannai A Gonczarowski, and Ran I Shorrer. 2024. Algorithmic collusion by large language mod- els.arXiv preprint arXiv:2404.00806 . Stephen M Fleming and Hakwan C Lau. 2014. How to measure metacognition. Frontiers in human neuro- science , 8:443. Stan Franklin. 1997. Autonomous agents as embodied ai.Cybernetics & System... | https://arxiv.org/abs/2505.19806v1 |
language models can self-improve. In 2023 Conference on Empirical Methods in Natural Lan- guage Processing, EMNLP 2023 , pages 1051–1068. Association for Computational Linguistics (ACL). Sukai Huang, Nir Lipovetzky, and Trevor Cohn. 2025. Planning in the dark: Llm-symbolic plan- ning pipeline without experts. In Procee... | https://arxiv.org/abs/2505.19806v1 |
Bilal Chughtai, Jan Betley, Kaivalya Hariharan, Mikita Balesni, Jérémy Scheurer, Mar- ius Hobbhahn, Alexander Meinke, and Owain Evans. 2024. Me, myself, and ai: The situational awareness dataset (sad) for llms. Advances in Neural Informa- tion Processing Systems , 37:64010–64118. Rudolf Laine, Alexander Meinke, and Owa... | https://arxiv.org/abs/2505.19806v1 |
. Li-Chun Lu, Shou-Jen Chen, Tsung-Min Pai, Chan- Hung Yu, Hung yi Lee, and Shao-Hua Sun. 2024a. LLM discussion: Enhancing the creativity of large language models via discussion framework and role- play. In First Conference on Language Modeling . Yining Lu, Dixuan Wang, Tianjian Li, Dongwei Jiang, Sanjeev Khudanpur, Me... | https://arxiv.org/abs/2505.19806v1 |
fei Yin, Yu Qiao, Yong Liu, and Jing Shao. 2024. Towards tracing trustworthiness dynamics: Revisit- ing pre-training period of large language models. In Findings of the Association for Computational Lin- guistics ACL 2024 , pages 4864–4888. Richard Ren, Arunim Agarwal, Mantas Mazeika, Cristina Menghini, Robert Vacarean... | https://arxiv.org/abs/2505.19806v1 |
Tang, Zhuosheng Zhang, Arman Cohan, Zhiyong Lu, and Mark Gerstein. 2024b. Prioritizing safeguarding over autonomy: Risks of LLM agents for science. In ICLR 2024 Workshop on Large Lan- guage Model (LLM) Agents . Giulio Tononi. 2004. An information integration theory of consciousness. BMC Neuroscience , 5(1):42. Giulio T... | https://arxiv.org/abs/2505.19806v1 |
of mind reasoning in large language models. In Find- ings of the Association for Computational Linguistics: EMNLP 2023 , pages 10691–10706. Jian Xie, Kai Zhang, Jiangjie Chen, Tinghui Zhu, Renze Lou, Yuandong Tian, Yanghua Xiao, and Yu Su. 2024. Travelplanner: A benchmark for real-world planning with language agents. I... | https://arxiv.org/abs/2505.19806v1 |
Following Block (1995), we classify contemporary theories of consciousness into three categories: phenomenal consciousness ,access consciousness , and hybrid theories .Hybrid theories integrate both phenome- nal and access aspects, arguing that neither is solely sufficient to explain consciousness.2 A.1 Phenomenal Cons... | https://arxiv.org/abs/2505.19806v1 |
arXiv:2505.19815v1 [cs.CL] 26 May 20252025-5-27 Deciphering Trajectory-Aided LLM Reasoning: An Optimization Perspective Junnan Liuℵ, Hongwei Liuℵ, Linchen Xiaoℵ, Shudong Liuℵ, Taolin Zhangℵ, Zihan Maℵ, Songyang Zhangℵ,†and Kai Chenℵ,† ℵShanghai AI Laboratory We propose a novel framework for comprehending the reasoning ... | https://arxiv.org/abs/2505.19815v1 |
authors: Songyang Zhang (zhangsongyang@pjlab.org.cn), Kai Chen (chenkai@pjlab.org.cn) *Code is at https://github.com/open-compass/RaML Deciphering Trajectory-Aided LLM Reasoning: An Optimization Perspective LLM: Updated LLM: Figure1: Illustration of the reasoning trajectory ( 𝑡) as the optimization of the LLM ... | https://arxiv.org/abs/2505.19815v1 |
answer 𝑎,𝑡represents the possible intermediate reasoning trajectories, and 𝐸𝑂indicates the output embeddings of the entire token set (i.e., vocabulary). Intuitively, ℳ𝜃(︁ 𝐼, 𝑞𝑖, 𝑡, 𝑎0 𝑖, . . . , 𝑎𝑗−1 𝑖)︁ represents the activation determined by the parameters 𝜃and the inputs, while the predicted probabili... | https://arxiv.org/abs/2505.19815v1 |
where 𝜃′ 𝑡represents the one-step update of 𝜃and the increment ∆ℳ𝜃(𝐼, 𝑞, 𝑡0)is only associated with 𝜃,𝐼,𝑞, and 𝑡0. According to Proposition 2.1 (the proof can be found in Appendix B), as the model progressively attends to the entire reasoning trajectory, the model parameters 𝜃are updated incrementally, a pr... | https://arxiv.org/abs/2505.19815v1 |
trajectories 𝒯𝑖of each 𝑞𝑖 through training data or rollout ; 5 Update 𝜃to𝜃′ 𝑡𝑗by reasoning trajectory 𝑡𝑗∈𝒯𝑖refer to Equation (4) ; 6 Optimize 𝜃through∑︀ 𝑡𝑗∈𝒯𝑖ℒ𝑞𝑖(︁ ℳ𝜃′ 𝑡𝑗)︁ for each 𝑞𝑖; Table1: Comparison of training techniques, where SFT, PO, and RL mean the abbreviation of supervised fine-tuni... | https://arxiv.org/abs/2505.19815v1 |
performance in each column, while Blue cells indicate the second-best performance. Techniques SourceAIME24 MATH500-L5 LiveMathBench-Hard Pass@ 8↑mG-Pass@ 8↑Pass@ 8↑mG-Pass@ 8↑Pass@ 8↑mG-Pass@ 8↑ SFTQwen 20.34 7.43 58 .42 35.65 26.77 7 .43 Distill-Qwen 36.69 10.29 82.98 45.79 25 .15 10.46 Zero-GRPO - 27.37 4 .08 71.66 3... | https://arxiv.org/abs/2505.19815v1 |
amount of training data. We generate 64 for each question and report Pass@ 32and mG-Pass@ 32. The evaluation includes prominent models such as Sky-T1-32B [ 98], Bespoke-Stratos-32B [ 50], LIMO [ 124], s1.1-32B [ 69], OpenThinker-32B [ 99], Light-R1- 32B [112], DeepSeek-R1-Distill-Qwen-32B [ 27], DAPO-32B [ 125], and VA... | https://arxiv.org/abs/2505.19815v1 |
8↑Pass@ 8↑mG-Pass@ 8↑Pass@ 8↑mG-Pass@ 8↑ Zero-GRPO 27.37 4 .08 71 .66 30 .48 27 .48 8 .21 + SFT Cold Start 35.87↑31% 11.23↑175% 82.42↑15% 44.92↑47% 42.17↑53% 18.84↑129% 0.00 0.25 0.50 0.75 1.00 Trajectory0 w/ Thinking w/o Thinking 0.00 0.25 0.50 0.75 1.00Trajectory1 w/ Thinking w/o Thinking 0.00 0.25 0.50 0.75 1.00Traj... | https://arxiv.org/abs/2505.19815v1 |
SFT-from-QwenSFT-from-Distil-Qwen Zero-RL-from-GRPOFigure7: Evaluation results of base, SFT and GRPO models on AIME24, LiveMathBench-Hard, GPQA-Diamond, and Live- CodeBench. of-thinking token delimiter Therefore, after all this, I believe the answer is following the thinking token delimiter <think> . Figure 5 demonstra... | https://arxiv.org/abs/2505.19815v1 |
informed by meta-learning to improve LLM reasoning, demonstrating the feasibility of enhancing LLM reasoning by integrating meta-learning studies. 4.1. Manipulating Training Reasoning Trajectories per Question Previous studies [ 1,15] highlight that the size of the support set is of paramount importance in improving pe... | https://arxiv.org/abs/2505.19815v1 |
s more effective inner loop optimization path? From an optimization perspective, as illustrated in Figure 10, there exists such an optimal inner loop optimization path. In this section, we present a straightforward yet convincing experiment to validate the existence of this inner loop optimization path. Specifically, w... | https://arxiv.org/abs/2505.19815v1 |
properties, and if so, in what manner ? ❸Trajectory-aided reasoning in large language models (LLMs) demonstrates comparable generalization 12 Deciphering Trajectory-Aided LLM Reasoning: An Optimization Perspective abilities across various tasks. What aspects of the learning process contribute to this generalization abi... | https://arxiv.org/abs/2505.19815v1 |
tasks, with a particular emphasis on lifelong learning. These foundational efforts established the basis for creating more adaptable and flexible learning algorithms, paving the way for subsequent advancements. Recent meta-learning approaches can generally be categorized into three groups: 1) metric-based methods, whic... | https://arxiv.org/abs/2505.19815v1 |
learning with in-context algorithm selection. In NeurIPS , 2023. 2.2, 5 [6]Edward Beeching, Shengyi Costa Huang, Albert Jiang, Jia Li, Benjamin Lipkin, Zihan Qina, Kashif Rasul, Ziju Shen, Roman Soletskyi, and Lewis Tunstall. Numinamath 72b cot. https://huggingface.co/ AI-MO/NuminaMath-72B-CoT , 2024. 5 [7]Samy Bengio,... | https://arxiv.org/abs/2505.19815v1 |
Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In NeurIPS , 2020. 2.2,... | https://arxiv.org/abs/2505.19815v1 |
Accessed: 2025-03-26. 1, 5 16 Deciphering Trajectory-Aided LLM Reasoning: An Optimization Perspective [27]DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li,... | https://arxiv.org/abs/2505.19815v1 |
Jingjing Xu, Zhiyong Wu, Baobao Chang, Xu Sun, Lei Li, and Zhifang Sui. A survey on in-context learning. In EMNLP, pages 1107–1128. Association for Computational Linguistics, 2024. 5 [32]Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yan... | https://arxiv.org/abs/2505.19815v1 |
Approximation capabilities of multilayer feedforward networks. Neural Networks , 4(2):251– 257, 1991. B [42]Timothy M. Hospedales, Antreas Antoniou, Paul Micaelli, and Amos J. Storkey. Meta-learning in neural networks: A survey. IEEE Trans. Pattern Anal. Mach. Intell. , 44(9):5149–5169, 2022. 1, 2 [43]Jeremy Howard and... | https://arxiv.org/abs/2505.19815v1 |
capable of stable reasoning? CoRR, abs/2412.13147, 2024. 3.1, 4.1, C.3 [61]Wenjie Ma, Jingxuan He, Charlie Snell, Tyler Griggs, Sewon Min, and Matei Zaharia. Reasoning models can be effective without thinking. CoRR, abs/2504.09858, 2025. 3.3, 4.2 [62]William Merrill, Ashish Sabharwal, and Noah A. Smith. Saturated trans... | https://arxiv.org/abs/2505.19815v1 |
optimizers. In ICML, volume 162 of Proceedings of Machine Learning Research , pages 17910–17925. PMLR, 2022. 3.2 [77]Yiwei Qin, Xuefeng Li, Haoyang Zou, Yixiu Liu, Shijie Xia, Zhen Huang, Yixin Ye, Weizhe Yuan, Hector Liu, Yuanzhi Li, and Pengfei Liu. O1 replication journey: A strategic progress report - part 1. CoRR, ... | https://arxiv.org/abs/2505.19815v1 |
Kshiteej Mahajan, Laura Culp, Lechao Xiao, Maxwell L. Bileschi, Noah Constant, Roman Novak, Rosanne Liu, Tris Warkentin, Yundi Qian, Yamini Bansal, Ethan Dyer, Behnam Neyshabur, Jascha Sohl-Dickstein, and Noah Fiedel. Beyond human data: Scaling self-training for problem-solving with language models. Trans. Mach. Learn.... | https://arxiv.org/abs/2505.19815v1 |
pages 10424–10433. PMLR, 2021. 1 [106]Eleni Triantafillou, Tyler Zhu, Vincent Dumoulin, Pascal Lamblin, Utku Evci, Kelvin Xu, Ross Goroshin, Carles Gelada, Kevin Swersky, Pierre-Antoine Manzagol, and Hugo Larochelle. Meta-dataset: A dataset of datasets for learning to learn from few examples. In ICLR. OpenReview.net, 2... | https://arxiv.org/abs/2505.19815v1 |
andZhenru Zhang. Qwen2.5-math technical report: Toward mathematical expert model via self-improvement. CoRR, abs/2409.12122, 2024. 3.1, 4.1, C.2 [119]Wenkai Yang, Shuming Ma, Yankai Lin, and Furu Wei. Towards thinking-optimal scaling of test-time compute for LLM reasoning. CoRR, abs/2502.18080, 2025. 3.3 [120]Huaxiu Ya... | https://arxiv.org/abs/2505.19815v1 |
𝑞𝑗 𝑖 the𝑗-th token of the 𝑖-th question 𝑎 the answer 𝒜 the answer set 𝑎𝑖 the𝑖-th answer of the 𝑖-th question 𝑎𝑗 𝑖 the𝑗-th token of the 𝑖-th answer |·| the length of tokens 𝐼 the instruction d the autoregressive decoding mechanism Softmax the softmax function 𝜎 the activation function 𝑡 the reasoning ... | https://arxiv.org/abs/2505.19815v1 |
cases: 1) if 𝐸𝑡,0lies within the span of the row vectors of 𝐸𝑙,;, then𝐶obviously exists; 2) if 𝐸𝑡,0does not lie within the span of the row vectors of 𝐸𝑙,;, an approximate 25 Deciphering Trajectory-Aided LLM Reasoning: An Optimization Perspective solution for 𝐶can be derived using various methods, such as the ... | https://arxiv.org/abs/2505.19815v1 |
; 4Compute the negative log-probability using 𝑙𝑎; Visualization of Landscape. We refer to the methodology proposed by Li et al. [ 53]. Assuming the set parameters of QwQ-32B is denoted by {𝜃𝑘}(excluding the embedding matrix), we randomly select two vectors, {𝜃1,𝑘}and{𝜃2,𝑘}, for each parameter. We then edit the ... | https://arxiv.org/abs/2505.19815v1 |
answer is highlighted within \boxed...final answer.... 3.Answer Format : The complete response should be formatted as: <think> ...reasoning process... </think> <conclusion> ...conclusion... The answer is \boxed...final answer... </conclusion> Training of GRPO. For the GRPO training, we use the complete question set and... | https://arxiv.org/abs/2505.19815v1 |
efficiency of model training. Synthetic Data From Strong Reasoning LLMs. Works such as OpenThoughts [ 99] and Llama Nemotron [ 8] utilize a more advanced reasoning LLM, such as DeepSeek-R1, to generate multiple trajectories for each training question, resulting in training data for SFT. This approach effectively expand... | https://arxiv.org/abs/2505.19815v1 |
An Optimization Perspective 0.0 0.5 1.0 q0 0.0 0.5 1.0q1 0.0 0.5 1.0q2 0.0 0.5 1.0q3 Normalized Token IndexTrajectory0Trajectory1Trajectory2Trajectory3 0.0 0.5 1.0 q0 0.0 0.5 1.0q1 0.0 0.5 1.0q2 0.0 0.5 1.0q3 Normalized Token IndexTrajectory0Trajectory1Trajectory2Trajectory3 0.0 0.5 1.0 q0 0.0 0.5 1.0q1 0.0 0.5 1.0q2 0... | https://arxiv.org/abs/2505.19815v1 |
1.0q2 0.0 0.5 1.0q3 Normalized Token IndexTrajectory0Trajectory1Trajectory2Trajectory3 0.0 0.5 1.0 q0 0.0 0.5 1.0q1 0.0 0.5 1.0q2 0.0 0.5 1.0q3 Normalized Token IndexTrajectory0Trajectory1Trajectory2Trajectory3 Figure14a: Visualizations of the pseudo-gradient update of Qwen3. 35 Deciphering Trajectory-Aided LLM Reasoni... | https://arxiv.org/abs/2505.19815v1 |
arXiv:2505.19838v1 [cs.CL] 26 May 2025FoodTaxo: Generating Food Taxonomies with Large Language Models Pascal Wullschleger⋄,†, Majid Zarharan⋄, Donnacha Daly† Marc Pouly†, Jennifer Foster⋄ ⋄ADAPT Centre, School of Computing, Dublin City University †Lucerne School of Computer Science and Information Technology (HSLU) pas... | https://arxiv.org/abs/2505.19838v1 |
novel LLM-based algorithms for 1) taxonomy completion and 2) taxonomy generation given a set of potentially in- complete known concepts. In a comparison to state- of-the-art methods on five taxonomies, we demonstrate the potential of these algorithms for food-related and other taxonomies. Our implementations and datase... | https://arxiv.org/abs/2505.19838v1 |
2021; Xu et al., 2023) that assume a complete set of new concepts Qto be added toTto obtain a new taxonomy T′= (E′,V ∪ Q ), we assume Qto be incomplete and allow for the generation of new concepts. Instead of starting with a fixed concept extraction process, we initialize Qwith an incomplete set of known concepts (ofte... | https://arxiv.org/abs/2505.19838v1 |
as triplets of the form (parent, query, child). For more detail, see Algorithm 1 in the Appendix. 2https://dl.fbaipublicfiles.com/fasttext/ vectors-crawl/cc.en.300.bin.gz 3For more detail on how the edges and concepts are en- coded as strings, refer to the prompts in Appendix 3.4. 3.3 Generating Taxonomies We generate ... | https://arxiv.org/abs/2505.19838v1 |
information , only return the output in the described format . --- Input description . Context: List of existing parent-child ( supertype-subtype ) relations in the taxonomy . Child: Child concept ( subtype ) that you need to place in a taxonomy . Description: Description of the child concept . --- Follow the following... | https://arxiv.org/abs/2505.19838v1 |
the parent concept in a taxonomy .A child concept must be a type of the parent concept . Separate with commas . --- Context: ``` ... ``` Candidates: salsa , cranberry sauce , dip , soy sauce , wasabi , vinegar , spread , duck sauce , chutney , marinade , mustard , sauce , mint sauce , green olive , pickle relish , blac... | https://arxiv.org/abs/2505.19838v1 |
used to evaluate taxonomy extraction methods for a given corpus (Bordea et al., 2016). SemEval-Verb is based on WordNet 3.0 (Fell- baum, 2010) and featured in the SemEval-2016 Task 14, which concerned evaluation of taxonomy enrichment approaches (Jurgens and Pilehvar, 2016). MeSH is a hierarchically organized vocabular... | https://arxiv.org/abs/2505.19838v1 |
taxonomy expansion task, where only leaves must be added. Model selection Since running experiments on LLMs is expensive, and we want to make our approach eas- ily accessible, we restrict our experiments to the open- source model Llama-3 ( Llama-3-70b-Instruct ).6 Hypothesis testing Following the recommendations of Dro... | https://arxiv.org/abs/2505.19838v1 |
0.8187 0.1695 0.1922 0.1516 TMN 0.8036 0.0081 0.0100 0.0067 0.8276 0.0056 0.0127 0.0036 0.8012 0.0086 0.0097 0.0077 TacoPrompt 0.8242 0.1652 0.2058 0.1380 0.8607 0.0392 0.0886 0.0252 0.8207 0.1929 0.2187 0.1725 TaxoExpan 0.7896 0.0161 0.0201 0.0135 0.7756 0.0000 0.0000 0.0000 0.7910 0.0197 0.0223 0.0176 Llama-3 Few-Sho... | https://arxiv.org/abs/2505.19838v1 |
NLI Validation 0.8581 0.2793 0.4662 0.1994 0.8175 0.0711 0.3333 0.0398 0.8664 0.4453 0.4911 0.4074 Complete 0.8583 0.3025 0.5076 0.2154 0.8282 0.0914 0.4286 0.0511 0.8645 0.4715 0.5225 0.4296 Table 4: Ablation study of NLI-verification and Backtracking on the completion task for SemEval-Food. All scores that are not si... | https://arxiv.org/abs/2505.19838v1 |
are not well captured by the metrics. Such issues likely stem from poor model performance on non-leaves (Ta- ble 3). Table 6 shows statistics regarding the generated taxonomies. Ablations In order to test the effectiveness of our mod- eling choices, we conducted an ablation study by re- moving different mechanisms from... | https://arxiv.org/abs/2505.19838v1 |
ever, qualitative inspection reveals that they still fall short of the nuance seen in human-curated taxonomies. We conclude that for LLM-based taxonomy generation to reach practical utility, significant advances are still needed, particularly in the reliable placement of non- leaf concepts. 7 Limitations •Due to the co... | https://arxiv.org/abs/2505.19838v1 |
Moazam, Heather Miller, Matei Zaharia, and Christopher Potts. 2023b. Dspy: Compiling declarative language model calls into self-improving pipelines. Preprint , arXiv:2310.03714. Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Hein- rich Küttler, Mike Lewis, Wen-tau Yih, Ti... | https://arxiv.org/abs/2505.19838v1 |
for Computing Machinery. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian ichter, Fei Xia, Ed Chi, Quoc V Le, and Denny Zhou. 2022. Chain-of-thought prompt- ing elicits reasoning in large language models. In Advances in Neural Information Processing Systems , volume 35, pages 24824–24837. Curran Associates... | https://arxiv.org/abs/2505.19838v1 |
outputs, as shown in Section 3.4. DSPy provides optimizers which can be used to tune prompts given validation and training data. We evalu- ated the automated tuning of instruction texts with their COPRO optimizer. This optimizer generates variations of a predefined prompt using a language model and evalu- ates its effe... | https://arxiv.org/abs/2505.19838v1 |
predicted placements Yqfor the query concept q 1:Yq← ∅ ▷Set of predicted placements for the query q 2:R←Retrieve (q,T, k) ▷Retrieve kmost relevant edges Rby cosine similarity to q 3:P ← CoT p(q, R, d q) ▷Generate candidate parent concepts using CoT prompting 4:P ← { p∈ P | ¬ contradicts (⌈q,”lemma (q)is a lemma (p)”)} ... | https://arxiv.org/abs/2505.19838v1 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.