{"id":"arxiv:2601.19897","title":"Self-Distillation Enables Continual Learning","date":"2026-01-27","section":"Post-training","quality":-0.07961,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/","source_url":"https://arxiv.org/abs/2601.19897","text_available":true,"authors":["Idan Shenfeld","Mehul Damani","Jonas Hübotter","Pulkit Agrawal"],"outgoing":[{"source":"arxiv:2601.19897","target":"arxiv:2402.13669","evidence":"id","id":"arxiv:2402.13669","title":"Self-Distillation Bridges Distribution Gap in Language Model Fine-Tuning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/1a8d11d48a58f35055e7f88f699cf42688a3698c63bd7026d5789f72a80581bf/"},{"source":"arxiv:2601.19897","target":"arxiv:2504.13837","evidence":"id","id":"arxiv:2504.13837","title":"Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b90f70269aeb02311d74ffaadf47e139683424a47118dc18619000ba72174d17/"},{"source":"arxiv:2601.19897","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:2601.19897","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:2601.19897","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:2601.19897","target":"arxiv:2601.20802","evidence":"id","id":"arxiv:2601.20802","title":"Reinforcement Learning via Self-Distillation","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/6073ec3d17b9a932a68e79ae078928f4635e40fc79cfcbc31709715d1f1e1ffa/"},{"source":"arxiv:2601.19897","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":[{"source":"arxiv:2601.18734","target":"arxiv:2601.19897","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:2601.20802","target":"arxiv:2601.19897","evidence":"id","id":"arxiv:2601.20802","title":"Reinforcement Learning via Self-Distillation","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/6073ec3d17b9a932a68e79ae078928f4635e40fc79cfcbc31709715d1f1e1ffa/"},{"source":"arxiv:2604.02288","target":"arxiv:2601.19897","evidence":"id","id":"arxiv:2604.02288","title":"Unifying Group-Relative and Self-Distillation Policy Optimization via Sample Routing","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/"},{"source":"arxiv:2604.13016","target":"arxiv:2601.19897","evidence":"id","id":"arxiv:2604.13016","title":"Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipe","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d6d7ff1e2e7b3b9c528f19f7144098920bf573f2cb411fcac5879d5d2cac93e0/"},{"source":"arxiv:2606.06021","target":"arxiv:2601.19897","evidence":"id","id":"arxiv:2606.06021","title":"OPRD: On-Policy Representation Distillation","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/"},{"source":"arxiv:2606.18810","target":"arxiv:2601.19897","evidence":"id","id":"arxiv:2606.18810","title":"Learning from Own Solutions: Self-Conditioned Credit Assignment for Reinforcement Learning with Verifiable Rewards","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/9c11d77c6d08392f24aab79c8b45a3a3b0f4b743dd35e150a83ee7f6fa315248/"},{"source":"arxiv:2607.07847","target":"arxiv:2601.19897","evidence":"id","id":"arxiv:2607.07847","title":"When Does Continual Learning Require Learning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/86e88c6cf21a66879eab5c156913fc85d14eb69c394c78c33c0e81d144e528eb/"},{"source":"arxiv:2608.13040","target":"arxiv:2601.19897","evidence":"id","id":"arxiv:2608.13040","title":"Latent On-Policy Self-Distillation","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/"},{"source":"arxiv:2610.02140","target":"arxiv:2601.19897","evidence":"id","id":"arxiv:2610.02140","title":"Finetuning with Sampling: SFT Learns Better Than You Think","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/60f5edfae92f505bf86e8890135a44b4129825b0525b5ad8d242e8ba04182605/"}],"text_parts":[{"start":0,"end":15314,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-1.html"},{"start":15314,"end":30962,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-2.html"},{"start":30962,"end":46658,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-3.html"},{"start":46658,"end":54154,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-4.html"}],"contents":[{"title":"Self-Distillation Enables Continual Learning","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-1.html#text"},{"title":"Abstract","offset":48,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-1.html#text"},{"title":"1 Introduction","offset":1309,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-1.html#text"},{"title":"2 Related Work","offset":5706,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-1.html#text"},{"title":"Off-policy versus On-policy Learning.","offset":5725,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-1.html#text"},{"title":"Inverse Reinforcement Learning.","offset":6988,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-1.html#text"},{"title":"Context Distillation.","offset":8465,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-1.html#text"},{"title":"Self-Distillation.","offset":9690,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-1.html#text"},{"title":"3 Self-Distillation Fine-Tuning","offset":9983,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-1.html#text"},{"title":"3.1 Self-Distillation as Inverse RL","offset":13808,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-1.html#text"},{"title":"3.2 Validating the ICL Assumption","offset":17821,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-2.html#text"},{"title":"4 Experiments","offset":21536,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-2.html#text"},{"title":"4.1 Experimental Setting","offset":21554,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-2.html#text"},{"title":"Evaluation.","offset":23338,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-2.html#text"},{"title":"Baselines.","offset":24537,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-2.html#text"},{"title":"4.2 On-policy learning leads to better generalization","offset":25723,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-2.html#text"},{"title":"Results.","offset":26215,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-2.html#text"},{"title":"4.3 Learning without forgetting","offset":28151,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-2.html#text"},{"title":"Results.","offset":28931,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-2.html#text"},{"title":"4.4 Effect of model size","offset":30305,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-2.html#text"},{"title":"4.5 Training reasoning models without reasoning data","offset":31652,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-3.html#text"},{"title":"Results.","offset":33592,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-3.html#text"},{"title":"4.6 What drives the improvement in performance?","offset":34024,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-3.html#text"},{"title":"Results.","offset":35021,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-3.html#text"},{"title":"5 Discussion and Limitations","offset":35418,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-3.html#text"},{"title":"Relationship to on-policy RL.","offset":35451,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-3.html#text"},{"title":"Computational Costs.","offset":36703,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-3.html#text"},{"title":"Learned Artifacts.","offset":37430,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-3.html#text"},{"title":"Requirements for Model Capability.","offset":38207,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-3.html#text"},{"title":"Future Work.","offset":39068,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-3.html#text"},{"title":"Author Contributions","offset":39948,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-3.html#text"},{"title":"Acknowledgment","offset":40505,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-3.html#text"},{"title":"Appendix A Additional Ablations","offset":42211,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-3.html#text"},{"title":"A.1 Estimating the KL gradient","offset":42247,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-3.html#text"},{"title":"Token-level (partial) estimator","offset":43135,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-3.html#text"},{"title":"Full analytic per-token estimator.","offset":43816,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-3.html#text"},{"title":"Rao-Blackwellized estimator.","offset":44613,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-3.html#text"},{"title":"A.2 The Importance of Demonstration-Conditioned Context","offset":46597,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-3.html#text"},{"title":"Results.","offset":47517,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-4.html#text"},{"title":"A.3 Choice of teacher model","offset":48106,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-4.html#text"},{"title":"Appendix B Training and Evaluation details","offset":49078,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-4.html#text"},{"title":"B.1 Training Details","offset":49125,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-4.html#text"},{"title":"B.2 Evaluation Details","offset":50223,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-4.html#text"},{"title":"B.3 Dataset Details","offset":51051,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-4.html#text"},{"title":"Science Q&A.","offset":51076,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-4.html#text"},{"title":"Tool Use.","offset":51627,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-4.html#text"},{"title":"Medical.","offset":51952,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-4.html#text"},{"title":"Knowledge Acquisition.","offset":52493,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-4.html#text"},{"title":"Instructions for reporting errors","offset":53190,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/part-4.html#text"}],"text_code_points":54154,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/cc146df3e9ca40c6d111a5264e08710100c1b34bd0eabd9052e8d55d13589efe/source.md"}