{"id":"arxiv:2604.02288","title":"Unifying Group-Relative and Self-Distillation Policy Optimization via Sample Routing","date":"2026-04-02","section":"Post-training","quality":0.005794,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/","source_url":"https://arxiv.org/abs/2604.02288","text_available":true,"authors":["Gengsheng Li","Tianyu Yang","Junfeng Fang","Mingyang Song"],"outgoing":[{"source":"arxiv:2604.02288","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:2604.02288","target":"arxiv:2503.14476","evidence":"id","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:2604.02288","target":"arxiv:2503.20783","evidence":"title","id":"arxiv:2503.20783","title":"Understanding R1-Zero-Like Training: A Critical Perspective","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b49a056f3ce8e269dd65e3b11605927f3c4d0ef074bc68e0b1d7f84e6c9c6c6a/"},{"source":"arxiv:2604.02288","target":"arxiv:2507.18071","evidence":"id","id":"arxiv:2507.18071","title":"Group Sequence Policy Optimization","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/69430db904efc7978803673806cc9338004ffd8cf92199f3349dfbfa8a44e440/"},{"source":"arxiv:2604.02288","target":"arxiv:2510.13786","evidence":"id","id":"arxiv:2510.13786","title":"The Art of Scaling Reinforcement Learning Compute for LLMs","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/72580bcba3d9b32b2f3e83d074f626234b26dc18126edd22a5eacef927cf87da/"},{"source":"arxiv:2604.02288","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:2604.02288","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:2604.02288","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:2604.02288","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:2604.02288","target":"arxiv:2603.24472","evidence":"id","id":"arxiv:2603.24472","title":"Why Does Self-Distillation (Sometimes) Degrade the Reasoning Capability of LLMs?","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/d99edb6fac2f4ad6d63e98ec0a2952187ed2ae190604fbed3507b3ed56d57315/"}],"incoming":[{"source":"arxiv:2604.13016","target":"arxiv:2604.02288","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:2604.02288","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:2608.27448","target":"arxiv:2604.02288","evidence":"id","id":"arxiv:2608.27448","title":"TTPO: Test-Time Policy Optimization","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3473ad1e90105976912e40eabe716441e830da401749e9d85b5162e1c7bbd964/"}],"text_parts":[{"start":0,"end":15705,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-1.html"},{"start":15705,"end":31182,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-2.html"},{"start":31182,"end":42631,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-3.html"}],"contents":[{"title":"Unifying Group-Relative and Self-Distillation Policy Optimization via Sample Routing","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-1.html#text"},{"title":"Abstract","offset":88,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-1.html#text"},{"title":"1 Introduction","offset":1678,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-1.html#text"},{"title":"2 Preliminaries","offset":7723,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-1.html#text"},{"title":"2.1 Group Relative Policy Optimization","offset":8034,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-1.html#text"},{"title":"2.2 Self-Distillation Policy Optimization","offset":9600,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-1.html#text"},{"title":"3 The SRPO","offset":11841,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-1.html#text"},{"title":"3.1 Sample-Level Routing","offset":12400,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-1.html#text"},{"title":"3.2 Dynamic-Weighted SDPO","offset":14465,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-1.html#text"},{"title":"3.3 Training Objective","offset":16230,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-2.html#text"},{"title":"4 Experiments","offset":17529,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-2.html#text"},{"title":"4.1 Experimental Setup","offset":17547,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-2.html#text"},{"title":"4.2 Main Results","offset":20109,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-2.html#text"},{"title":"4.3 Ablation Study","offset":23134,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-2.html#text"},{"title":"4.4 Response Length and Compute Time","offset":25453,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-2.html#text"},{"title":"5 Conclusion","offset":27221,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-2.html#text"},{"title":"Ethics Statement","offset":28238,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-2.html#text"},{"title":"Appendix A Related Work","offset":29435,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-2.html#text"},{"title":"A.1 Reinforcement Learning with Verifiable Rewards","offset":29463,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-2.html#text"},{"title":"A.2 On-Policy Distillation and Self-Distillation","offset":31128,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-2.html#text"},{"title":"Appendix B Experimental Details","offset":33539,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-3.html#text"},{"title":"B.1 Technical Setup","offset":33575,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-3.html#text"},{"title":"B.2 Hyperparameters","offset":34478,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-3.html#text"},{"title":"B.3 Prompt Templates","offset":35351,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-3.html#text"},{"title":"B.4 Benchmark Details","offset":35715,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-3.html#text"},{"title":"B.5 Teacher Information Construction","offset":36620,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-3.html#text"},{"title":"Construction procedure.","offset":37249,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-3.html#text"},{"title":"Illustrative example.","offset":38238,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-3.html#text"},{"title":"Fallback to GRPO when no teacher information is available.","offset":38517,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-3.html#text"},{"title":"Appendix C Routing Statistics Over Training","offset":39415,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-3.html#text"},{"title":"Instructions for reporting errors","offset":41667,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/part-3.html#text"}],"text_code_points":42631,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/3d1160029d321f2069016f7702d4ebebee0de6cd31a8f296ee220abd6f69c99b/source.md"}