{"id":"arxiv:2603.24472","title":"Why Does Self-Distillation (Sometimes) Degrade the Reasoning Capability of LLMs?","date":"2026-03-25","section":"Post-training","quality":-0.15648,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/","source_url":"https://arxiv.org/abs/2603.24472","text_available":true,"authors":["Jeonghye Kim","Xufang Luo","Minbeom Kim","Sangmook Lee"],"outgoing":[{"source":"arxiv:2603.24472","target":"arxiv:2503.14476","evidence":"title","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:2603.24472","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:2603.24472","target":"arxiv:2601.18734","evidence":"title","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:2603.24472","target":"arxiv:2601.20802","evidence":"title","id":"arxiv:2601.20802","title":"Reinforcement Learning via Self-Distillation","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/6073ec3d17b9a932a68e79ae078928f4635e40fc79cfcbc31709715d1f1e1ffa/"}],"incoming":[{"source":"arxiv:2604.02288","target":"arxiv:2603.24472","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:2603.24472","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:2603.24472","evidence":"id","id":"arxiv:2606.06021","title":"OPRD: On-Policy Representation Distillation","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a16a9d7ea78c5748acc7f651c9a96663e0200d72e51d6a52c9528e1739a27d8d/"}],"text_parts":[{"start":0,"end":15647,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-1.html"},{"start":15647,"end":31527,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-2.html"},{"start":31527,"end":47444,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-3.html"},{"start":47444,"end":59358,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-4.html"}],"contents":[{"title":"Why Does Self-Distillation (Sometimes) Degrade the Reasoning Capability of LLMs?","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-1.html#text"},{"title":"Abstract","offset":84,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-1.html#text"},{"title":"1 Introduction","offset":1171,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-1.html#text"},{"title":"2 Preliminaries","offset":5288,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-1.html#text"},{"title":"Self-Distillation","offset":5308,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-1.html#text"},{"title":"Key Characteristics of Math Reasoning","offset":6684,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-1.html#text"},{"title":"3 LLM Reasoning Behavior Under Richer Information","offset":7740,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-1.html#text"},{"title":"Prompts","offset":11093,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-1.html#text"},{"title":"Epistemic Tokens","offset":11340,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-1.html#text"},{"title":"Results","offset":12146,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-1.html#text"},{"title":"4 Supervised Finetuning with Self-Distillation","offset":13929,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-1.html#text"},{"title":"5 On-Policy Self-Distillation","offset":15613,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-1.html#text"},{"title":"5.1 DeepSeek-R1-Distill-Qwen-7B","offset":16844,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-2.html#text"},{"title":"Training Performance","offset":17150,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-2.html#text"},{"title":"OOD Evaluation - AIME24, AMC23","offset":17818,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-2.html#text"},{"title":"Reasoning Pattern","offset":18335,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-2.html#text"},{"title":"5.2 Qwen3-8B (Thinking Mode: ON)","offset":19056,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-2.html#text"},{"title":"Training Performance","offset":19454,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-2.html#text"},{"title":"OOD Evaluation - AIME24, AMC23","offset":20098,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-2.html#text"},{"title":"Reasoning Pattern","offset":20696,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-2.html#text"},{"title":"5.3 Qwen3-8B (Thinking Mode: OFF)","offset":21102,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-2.html#text"},{"title":"5.4 Causal Study: Epistemic Token Suppression during GRPO","offset":21824,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-2.html#text"},{"title":"5.5 Ablation Study: Fixed vs. Moving Target Teacher","offset":23287,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-2.html#text"},{"title":"6 Relationship Between Task Coverage, Epistemic Verbalization and Generalization Ability","offset":24432,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-2.html#text"},{"title":"6.1 Comparison of Task Coverage","offset":25450,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-2.html#text"},{"title":"6.2 Relationship Between Task Coverage and Learning Performance","offset":26217,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-2.html#text"},{"title":"Training Logs","offset":26581,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-2.html#text"},{"title":"OOD Evaluation - AIME24, MATH500","offset":27625,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-2.html#text"},{"title":"7 Conclusion","offset":28329,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-2.html#text"},{"title":"Acknowledgments","offset":29593,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-2.html#text"},{"title":"Ethics Statement","offset":29862,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-2.html#text"},{"title":"Reproducibility Statement","offset":30924,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-2.html#text"},{"title":"Appendix A Related Works","offset":31455,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-2.html#text"},{"title":"Self-Distillation for LLM Reasoning.","offset":31484,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-2.html#text"},{"title":"Uncertainty Expression in LLM Reasoning.","offset":32811,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-3.html#text"},{"title":"Reasoning Compression.","offset":33608,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-3.html#text"},{"title":"Appendix B Basis for Selecting Epistemic Tokens","offset":34946,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-3.html#text"},{"title":"Appendix C Experimental Details","offset":37215,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-3.html#text"},{"title":"Training","offset":37251,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-3.html#text"},{"title":"Evaluation","offset":38540,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-3.html#text"},{"title":"Chat Templates for Different Model Series","offset":38806,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-3.html#text"},{"title":"Appendix D More Ablation Study","offset":39069,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-3.html#text"},{"title":"D.1 Effect of Maximum Response Length","offset":39104,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-3.html#text"},{"title":"D.2 Top- k k Distillation and Learning Rate","offset":40152,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-3.html#text"},{"title":"Appendix E Additional Analysis of Epistemic Verbalization","offset":40759,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-3.html#text"},{"title":"E.1 Per-Token Analysis of Epistemic Verbalization","offset":40821,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-3.html#text"},{"title":"E.2 Comparison of Epistemic Token Usage Across Models","offset":41564,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-3.html#text"},{"title":"E.3 Analysis of GRPO and SDPO Reasoning Patterns by Task Coverage","offset":42824,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-3.html#text"},{"title":"E.4 Comparative Analysis of Epistemic and Overall Word Usage","offset":43379,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-3.html#text"},{"title":"E.5 Uncertainty Annotation via LLM-as-Judge","offset":45581,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-3.html#text"},{"title":"Appendix F Student–Teacher KL Divergence on Epistemic Tokens","offset":47379,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-3.html#text"},{"title":"Appendix G Analysis of Changes in Model Internal Uncertainty under Self-Distillation","offset":48264,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-4.html#text"},{"title":"Appendix H Can Uncertainty-Encouraging Prompts Mitigate the Suppression?","offset":49656,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-4.html#text"},{"title":"Appendix I Comparison with OPSD from Zhao et al. (2026)","offset":51116,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-4.html#text"},{"title":"Appendix J More On-Policy Self-Distillation Results","offset":53550,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-4.html#text"},{"title":"J.1 Qwen3-8B (Thinking Mode: OFF)","offset":53606,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-4.html#text"},{"title":"J.2 Qwen3-1.7B (Thinking Mode: ON)","offset":54477,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-4.html#text"},{"title":"J.3 Olmo-3-7B-Instruct","offset":55222,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-4.html#text"},{"title":"J.4 Pass@16 Score","offset":55738,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-4.html#text"},{"title":"Appendix K Qualitative Analysis","offset":56103,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-4.html#text"},{"title":"K.1 Unguided vs. Solution-Guided Generation","offset":56139,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-4.html#text"},{"title":"K.2 GRPO vs. SDPO Generation","offset":57119,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-4.html#text"},{"title":"Problem (AIME24).","offset":57153,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-4.html#text"},{"title":"Appendix L Disclosure of LLM Usage","offset":58085,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-4.html#text"},{"title":"Instructions for reporting errors","offset":58394,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/part-4.html#text"}],"text_code_points":59358,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/source.md"}