{"id":"arxiv:2505.21444","title":"Can Large Reasoning Models Self-Train?","date":"2025-05-27","section":"Reasoning and the \"physics\" of language models","quality":-0.120113,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/","source_url":"https://arxiv.org/abs/2505.21444","text_available":true,"authors":["Sheikh Shafayat","Fahim Tajwar","Ruslan Salakhutdinov"],"outgoing":[{"source":"arxiv:2505.21444","target":"arxiv:2203.11171","evidence":"title","id":"arxiv:2203.11171","title":"Self-Consistency Improves Chain of Thought Reasoning in Language Models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/02a15b098d96a94d463bb470f59a34723abc34e59196f35c539f9796d8f3c3df/"},{"source":"arxiv:2505.21444","target":"arxiv:2303.17651","evidence":"title","id":"arxiv:2303.17651","title":"Self-Refine: Iterative Refinement with Self-Feedback","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/accef2157d87cf33be93cb3a9af246528b0a3742a9e2e700a194956fac8f5cf4/"},{"source":"arxiv:2505.21444","target":"arxiv:2402.03300","evidence":"id","id":"arxiv:2402.03300","title":"DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/636fb871904bc06f742cfdea069796f9e9665f95bd58dc63260eaff5154ef985/"},{"source":"arxiv:2505.21444","target":"arxiv:2406.11741","evidence":"title","id":"arxiv:2406.11741","title":"Transcendence: Generative Models Can Outperform The Experts That Train Them","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/009cee51dbce762b177f0413e99ef4c30535501f77f54ed03aeb95de376319fd/"},{"source":"arxiv:2505.21444","target":"arxiv:2407.21783","evidence":"id","id":"arxiv:2407.21783","title":"The Llama 3 Herd of Models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/de3cebb4c28bc5d41a843279f219dca14476284991ea5c7c0aeab7068a40e306/"},{"source":"arxiv:2505.21444","target":"arxiv:2503.14476","evidence":"id","id":"arxiv:2503.14476","title":"DAPO: An Open-Source LLM Reinforcement Learning System at Scale","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/548e934b912ef84691ace63bf1cfd670e7b0257d9d6986dde926f569e7b2f8b7/"},{"source":"arxiv:2505.21444","target":"arxiv:2503.20783","evidence":"id","id":"arxiv:2503.20783","title":"Understanding R1-Zero-Like Training: A Critical Perspective","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b49a056f3ce8e269dd65e3b11605927f3c4d0ef074bc68e0b1d7f84e6c9c6c6a/"},{"source":"arxiv:2505.21444","target":"arxiv:2504.16084","evidence":"id","id":"arxiv:2504.16084","title":"TTRL: Test-Time Reinforcement Learning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/ad54dd58d24afdfefc8aa5087d07c4aab2143e90ef06f6f7db8d3f3333452c48/"},{"source":"arxiv:2505.21444","target":"arxiv:2504.20571","evidence":"id","id":"arxiv:2504.20571","title":"Reinforcement Learning for Reasoning in Large Language Models with One Training Example","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b2b21c3df49619588c371411dba9467a41f81b3fd4da5441898b5145bb0de120/"},{"source":"arxiv:2505.21444","target":"arxiv:2505.19590","evidence":"id","id":"arxiv:2505.19590","title":"Learning to Reason without External Rewards","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d8df33b706906a7abf4e660f54c75a81ffe2caaec8b0a10d1a6ad6ec63060832/"},{"source":"arxiv:2505.21444","target":"arxiv:2506.10947","evidence":"id","id":"arxiv:2506.10947","title":"Spurious Rewards: Rethinking Training Signals in RLVR","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f48e13a9656d8eff1882eb80988e07a820c4799401298947deac735c004e8592/"},{"source":"arxiv:2505.21444","target":"arxiv:2507.18071","evidence":"id","id":"arxiv:2507.18071","title":"Group Sequence Policy Optimization","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/69430db904efc7978803673806cc9338004ffd8cf92199f3349dfbfa8a44e440/"}],"incoming":[{"source":"arxiv:2505.19590","target":"arxiv:2505.21444","evidence":"id","id":"arxiv:2505.19590","title":"Learning to Reason without External Rewards","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d8df33b706906a7abf4e660f54c75a81ffe2caaec8b0a10d1a6ad6ec63060832/"},{"source":"arxiv:2506.10947","target":"arxiv:2505.21444","evidence":"id","id":"arxiv:2506.10947","title":"Spurious Rewards: Rethinking Training Signals in RLVR","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/f48e13a9656d8eff1882eb80988e07a820c4799401298947deac735c004e8592/"},{"source":"arxiv:2604.19341","target":"arxiv:2505.21444","evidence":"id","id":"arxiv:2604.19341","title":"Structured Scaling of AI Discovery Across Diverse Scientific Domains","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/33ad3375247f86197210840a19064548b997e8617475e606d7a7ac2cda64740e/"},{"source":"arxiv:2607.13104","target":"arxiv:2505.21444","evidence":"id","id":"arxiv:2607.13104","title":"Self-Improvements in Modern Agentic Systems: A Survey","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1452caf7b16ffdb8c2efd5093f46dc4df0f89c3d45a5566cd5fdafd69385b2a7/"},{"source":"arxiv:2608.19197","target":"arxiv:2505.21444","evidence":"id","id":"arxiv:2608.19197","title":"SPADE: Self-Play in Adaptive Synthetic Executable Environments","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c524376d6034b5d79661d6f1a444591404f4a250bcb0790d802c05e02f8249df/"}],"text_parts":[{"start":0,"end":14938,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-1.html"},{"start":14938,"end":30775,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-2.html"},{"start":30775,"end":46755,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-3.html"},{"start":46755,"end":61982,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-4.html"},{"start":61982,"end":69102,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-5.html"}],"contents":[{"title":"Can Large Reasoning Models Self-Train?","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-1.html#text"},{"title":"Abstract","offset":42,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-1.html#text"},{"title":"1 Introduction","offset":1075,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-1.html#text"},{"title":"2 Preliminaries","offset":6391,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-1.html#text"},{"title":"3 Self-Rewarded Training","offset":9265,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-1.html#text"},{"title":"Self-supervision via majority voting.","offset":10328,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-1.html#text"},{"title":"Self-Rewarded Training (SRT).","offset":11592,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-1.html#text"},{"title":"4 Experiments and Analysis","offset":13487,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-1.html#text"},{"title":"4.1 Can SRT go beyond the base model’s capabilities?","offset":14279,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-1.html#text"},{"title":"Experiments on synthetic reasoning tasks.","offset":14890,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-1.html#text"},{"title":"Results.","offset":16736,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-2.html#text"},{"title":"4.2 Can Self-Improvement from SRT be Sustained Indefinitely?","offset":19847,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-2.html#text"},{"title":"Multi-level self-improvement on synthetic tasks using curriculum.","offset":20295,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-2.html#text"},{"title":"Extended SRT -training on math problems.","offset":21754,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-2.html#text"},{"title":"What happens after SRT peaks in performance?","offset":22565,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-2.html#text"},{"title":"Additional experiments and ablations.","offset":24571,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-2.html#text"},{"title":"5 Related Works","offset":25870,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-2.html#text"},{"title":"Self Improving LLM.","offset":25890,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-2.html#text"},{"title":"Online RLVR and Easy to Hard Generalization.","offset":28325,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-2.html#text"},{"title":"Model Collapse and Reward Hacking.","offset":29380,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-2.html#text"},{"title":"Data Efficient RLVR.","offset":30748,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-2.html#text"},{"title":"6 Limitations and Conclusion","offset":31257,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-3.html#text"},{"title":"Acknowledgements","offset":32476,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-3.html#text"},{"title":"Appendix A Definition of Generation-Verification Gap","offset":34123,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-3.html#text"},{"title":"A.1 Generation Verification Gap Through Majority Voting","offset":35287,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-3.html#text"},{"title":"Appendix B Details on Implementation and Training","offset":35348,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-3.html#text"},{"title":"B.1 RL Algorithms","offset":35402,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-3.html#text"},{"title":"Sampling.","offset":38982,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-3.html#text"},{"title":"B.2 GPU Infrastructure","offset":39399,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-3.html#text"},{"title":"B.3 Details on Training Settings","offset":39947,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-3.html#text"},{"title":"Default training hyperparameters for Reasoning Gym tasks.","offset":41673,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-3.html#text"},{"title":"Default training hyperparameters for Math Datasets.","offset":42001,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-3.html#text"},{"title":"B.4 Code and Dataset Release","offset":42979,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-3.html#text"},{"title":"Appendix C Additional Experimental Results","offset":44406,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-3.html#text"},{"title":"C.1 Qwen2.5-Math-7B","offset":44453,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-3.html#text"},{"title":"C.2 Qwen3-14B-Base","offset":45071,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-3.html#text"},{"title":"Appendix D Effect of Decoding Temperature (Qwen2.5-Math-7B)","offset":45814,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-3.html#text"},{"title":"Appendix E Additional Self-Training Metrics","offset":45878,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-3.html#text"},{"title":"Appendix F More Details on Baselines","offset":46683,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-3.html#text"},{"title":"Baseline Implementation.","offset":46724,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-3.html#text"},{"title":"Dataset Curation","offset":47680,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-4.html#text"},{"title":"Appendix G Detailed Experiment Results on Different Training Settings","offset":48006,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-4.html#text"},{"title":"G.1 GRPO vs RLOO","offset":48080,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-4.html#text"},{"title":"G.2 Different KL Penalty Coefficients","offset":48464,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-4.html#text"},{"title":"G.3 Learning Rate","offset":48873,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-4.html#text"},{"title":"G.4 Different Number of generations per prompt","offset":49670,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-4.html#text"},{"title":"G.5 Entropy Coefficient","offset":50601,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-4.html#text"},{"title":"Appendix H Test-Time Self-Improvement","offset":51998,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-4.html#text"},{"title":"H.1 SRT can be used for test-time training","offset":52040,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-4.html#text"},{"title":"H.2 Why Doesn’t the Performance Collapse during Test-Time-Training?","offset":53078,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-4.html#text"},{"title":"Appendix I More on Model Collapse Resulting from Self-Training","offset":54473,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-4.html#text"},{"title":"I.1 Early Stopping","offset":55194,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-4.html#text"},{"title":"I.2 Self-Training with Offline-Generated Labels","offset":55808,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-4.html#text"},{"title":"I.3 Self-Training with Curriculum Learning","offset":57078,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-4.html#text"},{"title":"I.4 Training Dynamics of SRT (Qwen2.5-Math-7B) on the Easy DAPO Subset","offset":58687,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-4.html#text"},{"title":"I.5 Training Qwen3-14B-Base on the Easy DAPO Subset","offset":60274,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-4.html#text"},{"title":"I.6 Generating Easy DAPO Subset using Qwen2.5-Math-1.5B","offset":61075,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-4.html#text"},{"title":"I.7 SRT training with online curriculum","offset":61937,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-4.html#text"},{"title":"Appendix J Detailed Experiment Results using non-Qwen models","offset":63849,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-5.html#text"},{"title":"J.1 Deepseek-Math-7B-Instruct","offset":64194,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-5.html#text"},{"title":"J.2 Llama-3.1-8B-Instruct","offset":64997,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-5.html#text"},{"title":"Appendix K Example Tasks from Reasoning Gym","offset":67127,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-5.html#text"},{"title":"Appendix L Behavior of a Model after Performance Collapse","offset":67571,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-5.html#text"},{"title":"Instructions for reporting errors","offset":68138,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/part-5.html#text"}],"text_code_points":69102,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/206224ca50098e0bdc230140722feec107cfc90c1cef2ed8d3ed52355ef56bca/source.md"}