{"id":"arxiv:2507.12856","title":"Supervised Fine Tuning on Curated Data is Reinforcement Learning (and can be improved)","date":"2025-07-17","section":"Data, training, optimization","quality":-0.231891,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/","source_url":"https://arxiv.org/abs/2507.12856","text_available":true,"authors":["Chongli Qin","Jost Tobias Springenberg"],"outgoing":[{"source":"arxiv:2507.12856","target":"arxiv:2106.06860","evidence":"title","id":"arxiv:2106.06860","title":"A Minimalist Approach to Offline Reinforcement Learning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/00a35c7a58d8775f23035efbd1cb0a5f540440d47b8a996be8a4fa0da2da3e0e/"},{"source":"arxiv:2507.12856","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:2507.12856","target":"arxiv:2501.19393","evidence":"id","id":"arxiv:2501.19393","title":"s1: Simple test-time scaling","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/2a88818f7605b71cd0160831faebd51b3274c24ffd29626b4d519390b4bffffe/"},{"source":"arxiv:2507.12856","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:2507.12856","target":"url:rlhfbook.com","evidence":"title","id":"url:rlhfbook.com","title":"Reinforcement Learning from Human Feedback","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/20e354559fad09ad4d6338d104c9c3f822a48e7fbf86cd8afa1ddf951b223658/"}],"incoming":[{"source":"arxiv:2508.05629","target":"arxiv:2507.12856","evidence":"id","id":"arxiv:2508.05629","title":"On the Generalization of SFT: A Reinforcement Learning Perspective with Reward Rectification","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/0a17fccfeffa8a04070bb16903a7217d8adee804cf42e0cfc8dd128d3390474f/"},{"source":"arxiv:2508.11408","target":"arxiv:2507.12856","evidence":"id","id":"arxiv:2508.11408","title":"On-Policy RL Meets Off-Policy Experts: Harmonizing Supervised Fine-Tuning and Reinforcement Learning via Dynamic Weighting","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/48594da2e3f2cdfeb17195dd033d047235fc4b4ffb65030800f64c6faedc4fbe/"}],"text_parts":[{"start":0,"end":15691,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-1.html"},{"start":15691,"end":31563,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-2.html"},{"start":31563,"end":46967,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-3.html"},{"start":46967,"end":52204,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-4.html"}],"contents":[{"title":"Supervised Fine Tuning on Curated Data is Reinforcement Learning (and can be improved)","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-1.html#text"},{"title":"Abstract","offset":90,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-1.html#text"},{"title":"1 Introduction","offset":1321,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-1.html#text"},{"title":"2 Related Work","offset":5160,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-1.html#text"},{"title":"3 SFT on Curated Data as Reinforcement Learning","offset":9487,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-1.html#text"},{"title":"3.1 Notation and Background","offset":10825,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-1.html#text"},{"title":"Notation","offset":11156,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-1.html#text"},{"title":"Reinforcement learning","offset":12636,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-1.html#text"},{"title":"3.2 SFT lower bounds RL","offset":15662,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-1.html#text"},{"title":"3.3 Improving SFT from the perspective of RL","offset":20383,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-2.html#text"},{"title":"3.4 The importance of choosing q ⁡ ( τ ) q(\\tau)","offset":24796,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-2.html#text"},{"title":"Clipping per step importance weights","offset":26733,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-2.html#text"},{"title":"Smoothing importance weights","offset":27607,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-2.html#text"},{"title":"3.5 A Toy Example to build Intuition for SFT on curated/filtered data","offset":29103,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-2.html#text"},{"title":"4 Experiments","offset":32846,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-3.html#text"},{"title":"4.1 Training LLMs for reasoning via SFT on curated data","offset":33046,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-3.html#text"},{"title":"4.2 Offline RL for Continuous Control","offset":36548,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-3.html#text"},{"title":"4.3 Fine-tuning control policies with little data and noisy quality scores","offset":39569,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-3.html#text"},{"title":"5 Conclusion and Discussion","offset":40734,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-3.html#text"},{"title":"Appendix A Derivation for the iw-SFT objective","offset":42356,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-3.html#text"},{"title":"Appendix B Detailed algorithm","offset":46812,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-3.html#text"},{"title":"Appendix C Hyperparameters for LLM experiments","offset":46916,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-3.html#text"},{"title":"Appendix D Compute resources and infrastructure","offset":48282,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-4.html#text"},{"title":"LLM training","offset":48334,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-4.html#text"},{"title":"Continuous control experiments","offset":48619,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-4.html#text"},{"title":"Appendix E Dataset curation for control tasks","offset":49069,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-4.html#text"},{"title":"Appendix F Additional details for training on D4RL control tasks","offset":49183,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-4.html#text"},{"title":"Appendix G Ablations with Budget Forcing","offset":49967,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-4.html#text"},{"title":"Appendix H Code","offset":50712,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-4.html#text"},{"title":"H.1 Expanded D4RL results","offset":51080,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-4.html#text"},{"title":"Instructions for reporting errors","offset":51240,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/part-4.html#text"}],"text_code_points":52204,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5f6c1906eaa67369d9ff15b974756b89141d1b4cf9d7c0f6554e0d56f26e79be/source.md"}