{"id":"arxiv:2608.13040","title":"Latent On-Policy Self-Distillation","date":"2026-08-13","section":"Post-training","quality":0.138951,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/","source_url":"https://arxiv.org/abs/2608.13040","text_available":true,"authors":["Guibin Zhang","Jiayang Lyu","Ran Sun","Xinlei Yu"],"outgoing":[{"source":"arxiv:2608.13040","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:2608.13040","target":"arxiv:2412.06769","evidence":"id","id":"arxiv:2412.06769","title":"Training Large Language Models to Reason in a Continuous Latent Space","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/68c71dc1a042dedd5e3c4d9c1ad32c9d046fd31b8dc79080bd6dff1c9800abc7/"},{"source":"arxiv:2608.13040","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:2608.13040","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:2608.13040","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:2608.13040","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:2608.13040","target":"arxiv:2603.25562","evidence":"id","id":"arxiv:2603.25562","title":"Revisiting On-Policy Distillation: Empirical Failure Modes and Simple Fixes","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a66707a05c8aa09328e4a1f322a57800a3e12262a85a9a1cf195b41dc1926b32/"},{"source":"arxiv:2608.13040","target":"arxiv:2604.13016","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/"}],"incoming":[],"text_parts":[{"start":0,"end":14206,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-1.html"},{"start":14206,"end":30189,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-2.html"},{"start":30189,"end":46172,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-3.html"},{"start":46172,"end":56618,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-4.html"}],"contents":[{"title":"Latent On-Policy Self-Distillation","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-1.html#text"},{"title":"Abstract","offset":38,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-1.html#text"},{"title":"1 Introduction","offset":1793,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-1.html#text"},{"title":"2 Related Work","offset":7295,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-1.html#text"},{"title":"On-Policy (Self-)Distillation.","offset":7314,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-1.html#text"},{"title":"Latent Computation.","offset":9254,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-1.html#text"},{"title":"3 Method","offset":10439,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-1.html#text"},{"title":"3.1 Problem Setup: Privileged Context in OPSD","offset":10452,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-1.html#text"},{"title":"3.2 Learnable Latent Privileged Context","offset":13564,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-1.html#text"},{"title":"3.3 Latent On-Policy Self-Distillation","offset":16818,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-2.html#text"},{"title":"Privileged-Margin Constraint.","offset":20334,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-2.html#text"},{"title":"Algorithm Summary.","offset":23372,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-2.html#text"},{"title":"4 Experiments","offset":23835,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-2.html#text"},{"title":"4.1 Experiment Setup","offset":23853,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-2.html#text"},{"title":"Training.","offset":23879,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-2.html#text"},{"title":"Evaluation.","offset":24376,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-2.html#text"},{"title":"Baselines.","offset":24976,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-2.html#text"},{"title":"Configurations.","offset":25574,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-2.html#text"},{"title":"4.2 Main Results","offset":26624,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-2.html#text"},{"title":"Performance.","offset":26793,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-2.html#text"},{"title":"No Hand-Crafted Context Is Universally Optimal.","offset":27719,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-2.html#text"},{"title":"Learnable Context Matters.","offset":29233,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-2.html#text"},{"title":"4.3 Framework Analysis","offset":30189,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-3.html#text"},{"title":"Effect of Joint Optimization.","offset":30217,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-3.html#text"},{"title":"Training Dynamics and Sample Efficiency.","offset":31526,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-3.html#text"},{"title":"Sensitivity Analysis.","offset":32327,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-3.html#text"},{"title":"Behavioral Internalization.","offset":33765,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-3.html#text"},{"title":"4.4 Case Study","offset":34795,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-3.html#text"},{"title":"5 Conclusion","offset":35846,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-3.html#text"},{"title":"Appendix A Implementation Details","offset":37393,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-3.html#text"},{"title":"A.1 Composer Architecture","offset":37431,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-3.html#text"},{"title":"A.2 Cold-Start Initialization","offset":39144,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-3.html#text"},{"title":"A.3 Retrieval Configuration","offset":40938,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-3.html#text"},{"title":"A.4 Latent Injection Mechanism","offset":43526,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-3.html#text"},{"title":"Placeholder strategy.","offset":43930,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-3.html#text"},{"title":"Embedding replacement.","offset":44893,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-3.html#text"},{"title":"Engine-specific implementations.","offset":45682,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-3.html#text"},{"title":"Equivalence verification.","offset":46172,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-4.html#text"},{"title":"Encoder LoRA isolation.","offset":46635,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-4.html#text"},{"title":"A.5 Prompt and Interaction Format","offset":47130,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-4.html#text"},{"title":"Agentic system prompt.","offset":47169,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-4.html#text"},{"title":"Coding system prompt.","offset":47343,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-4.html#text"},{"title":"Latent-context framing text.","offset":47603,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-4.html#text"},{"title":"Encoder input format.","offset":47998,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-4.html#text"},{"title":"Appendix B Experiment Details","offset":48430,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-4.html#text"},{"title":"B.1 Baseline Setup","offset":48464,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-4.html#text"},{"title":"Vanilla.","offset":49564,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-4.html#text"},{"title":"GRPO.","offset":49647,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-4.html#text"},{"title":"SDFT.","offset":50040,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-4.html#text"},{"title":"OPSD.","offset":50466,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-4.html#text"},{"title":"SDPO.","offset":50789,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-4.html#text"},{"title":"Skill-SD.","offset":51284,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-4.html#text"},{"title":"B.2 Training Data and Domain Setup","offset":51811,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-4.html#text"},{"title":"B.3 Evaluation Protocols","offset":52775,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-4.html#text"},{"title":"EnvScaler.","offset":52915,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-4.html#text"},{"title":"BFCL-v3.","offset":53066,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-4.html#text"},{"title":"ACEBench.","offset":53333,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-4.html#text"},{"title":"LiveCodeBench.","offset":53529,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-4.html#text"},{"title":"EvalPlus.","offset":53726,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-4.html#text"},{"title":"B.4 Ablation and Analysis Setup","offset":53857,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-4.html#text"},{"title":"Effect of Joint Optimization ( Figure 3 ).","offset":53979,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-4.html#text"},{"title":"Training Dynamics ( Figure 4 ).","offset":54452,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-4.html#text"},{"title":"Sensitivity Analysis ( Figure 5 ).","offset":54592,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-4.html#text"},{"title":"Behavioral Internalization ( Table 3 ).","offset":55172,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-4.html#text"},{"title":"Instructions for reporting errors","offset":55654,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/part-4.html#text"}],"text_code_points":56618,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/c2cfd504d6ca36ba88d75aba018620867d5fafd7433b03f3b2c9f51408422a02/source.md"}