{"id":"arxiv:2610.02140","title":"Finetuning with Sampling: SFT Learns Better Than You Think","date":"2026-10-01","section":"Post-training","quality":null,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/","source_url":"https://arxiv.org/abs/2610.02140","text_available":true,"authors":["Aayush Karan","Sitan Chen","Yilun Du"],"outgoing":[{"source":"arxiv:2610.02140","target":"arxiv:2508.05629","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:2610.02140","target":"arxiv:2510.14901","evidence":"id","id":"arxiv:2510.14901","title":"Reasoning with Sampling: Your Base Model is Smarter Than You Think","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/46373e6f386387299bbade0a302219a611c3497c5e5b5970e3502f7a5f189687/"},{"source":"arxiv:2610.02140","target":"arxiv:2512.13961","evidence":"id","id":"arxiv:2512.13961","title":"Olmo 3","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/"},{"source":"arxiv:2610.02140","target":"arxiv:2601.18734","evidence":"id","id":"arxiv:2601.18734","title":"Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8a352e37fa79840a4f35f86ebc39661cfc34fbfc4391cd020a69af158c63485a/"},{"source":"arxiv:2610.02140","target":"arxiv:2601.19897","evidence":"id","id":"arxiv:2601.19897","title":"Self-Distillation Enables Continual Learning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/"},{"source":"arxiv:2610.02140","target":"doi:10.1073/pnas.1611835114","evidence":"title","id":"doi:10.1073/pnas.1611835114","title":"Overcoming catastrophic forgetting in neural networks","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/849082368a44d06b93ef3ba95bcfddc6672b7a5e8064ad2b2d19601639266e8b/"}],"incoming":[],"text_parts":[{"start":0,"end":15641,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-1.html"},{"start":15641,"end":31238,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-2.html"},{"start":31238,"end":44614,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-3.html"}],"contents":[{"title":"Finetuning with Sampling: SFT Learns Better Than You Think","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-1.html#text"},{"title":"Abstract","offset":62,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-1.html#text"},{"title":"1 Introduction","offset":1703,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-1.html#text"},{"title":"2 Related Works","offset":6026,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-1.html#text"},{"title":"Self-Distillation.","offset":7977,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-1.html#text"},{"title":"MCMC Sampling with LLMs.","offset":8688,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-1.html#text"},{"title":"3 Preliminaries","offset":9411,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-1.html#text"},{"title":"Supervised Finetuning.","offset":10678,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-1.html#text"},{"title":"4 Boosting Off-Policy Data with MCMC Sampling","offset":11211,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-1.html#text"},{"title":"4.1 Targeting the Information Projection","offset":12328,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-1.html#text"},{"title":"Definition 1 .","offset":13620,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-1.html#text"},{"title":"Definition 2 .","offset":14743,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-1.html#text"},{"title":"Proposition 1 .","offset":15020,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-1.html#text"},{"title":"Proof.","offset":15626,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-1.html#text"},{"title":"4.2 The Metropolis-Hastings Algorithm","offset":18374,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-2.html#text"},{"title":"Definition 3 .","offset":19539,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-2.html#text"},{"title":"Proposition 2 .","offset":21304,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-2.html#text"},{"title":"Proposition 3 .","offset":22016,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-2.html#text"},{"title":"Proof.","offset":22351,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-2.html#text"},{"title":"4.3 Projection Sampling for LLMs","offset":23594,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-2.html#text"},{"title":"5 Experiments","offset":27113,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-2.html#text"},{"title":"5.1 Experimental Setup","offset":27350,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-2.html#text"},{"title":"Tasks.","offset":27378,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-2.html#text"},{"title":"Models.","offset":29485,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-2.html#text"},{"title":"Evaluation.","offset":30027,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-2.html#text"},{"title":"Sampling.","offset":32032,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-3.html#text"},{"title":"Training.","offset":32196,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-3.html#text"},{"title":"5.2 Main Results","offset":32894,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-3.html#text"},{"title":"5.3 Analysis","offset":35713,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-3.html#text"},{"title":"6 Conclusion","offset":39391,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-3.html#text"},{"title":"7 Acknowledgements","offset":40903,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-3.html#text"},{"title":"Appendix A Examples of Boosted Off-Policy Data","offset":41174,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-3.html#text"},{"title":"Appendix B Additional Results","offset":41225,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-3.html#text"},{"title":"Appendix C Further Experimental Details","offset":42164,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-3.html#text"},{"title":"C.1 Information-Preservation","offset":42208,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-3.html#text"},{"title":"C.2 Sampling","offset":42605,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-3.html#text"},{"title":"C.3 Proposal and Grading Prompts","offset":43373,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-3.html#text"},{"title":"Instructions for reporting errors","offset":43650,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/part-3.html#text"}],"text_code_points":44614,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/source.md"}