{"id":"arxiv:2605.22074","title":"From Reasoning Chains to Verifiable Subproblems: Curriculum Reinforcement Learning Enables Credit Assignment for LLM Reasoning","date":"2026-05-21","section":"Post-training","quality":0.202613,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/","source_url":"https://arxiv.org/abs/2605.22074","text_available":true,"authors":["Xitai Jiang","Zihan Tang","Wenze Lin","Yang Yue"],"outgoing":[{"source":"arxiv:2605.22074","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:2605.22074","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:2605.22074","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:2605.22074","target":"arxiv:2506.01939","evidence":"id","id":"arxiv:2506.01939","title":"Beyond the 80/20 Rule: High-Entropy Minority Tokens Drive Effective Reinforcement Learning for LLM Reasoning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e0f4a55bca612bce5895b6b98f89b0efc14cbd8cf45946a7743347cf6daee1a1/"},{"source":"arxiv:2605.22074","target":"arxiv:2506.06632","evidence":"title","id":"arxiv:2506.06632","title":"Curriculum Reinforcement Learning from Easy to Hard Tasks Improves LLM Reasoning","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/79fda9609f9b88bb71f71f531b63ea6b5237d1c889b7ae59e9ca07c12dbe70d1/"}],"incoming":[],"text_parts":[{"start":0,"end":15039,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-1.html"},{"start":15039,"end":30748,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-2.html"},{"start":30748,"end":45874,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-3.html"}],"contents":[{"title":"1 Introduction","offset":892,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-1.html#text"},{"title":"2 Related Work","offset":8143,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-1.html#text"},{"title":"Reinforcement Learning with Verifiable Rewards (RLVR)","offset":8162,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-1.html#text"},{"title":"Curriculum Learning for Reasoning","offset":9817,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-1.html#text"},{"title":"3 Method","offset":11497,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-1.html#text"},{"title":"3.1 Preliminaries: GRPO","offset":12363,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-1.html#text"},{"title":"3.2 SCRL Framework","offset":13558,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-1.html#text"},{"title":"Build subproblems.","offset":13582,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-1.html#text"},{"title":"Curriculum prompt.","offset":13938,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-1.html#text"},{"title":"Response format.","offset":14386,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-1.html#text"},{"title":"3.2.1 Progress-Aware Subproblem Rewards","offset":14965,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-1.html#text"},{"title":"Curriculum progress.","offset":15011,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-1.html#text"},{"title":"Progress-aware correction.","offset":16048,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-2.html#text"},{"title":"3.2.2 SCRL Training Algorithm","offset":16884,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-2.html#text"},{"title":"Subproblem-level normalization.","offset":17163,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-2.html#text"},{"title":"Token-level credit assignment.","offset":17816,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-2.html#text"},{"title":"Mixed-group training.","offset":18505,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-2.html#text"},{"title":"4 Theoretical Analysis","offset":19866,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-2.html#text"},{"title":"Definition 4.1 (Effective and Lifted Gradient Information Matrices) .","offset":20321,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-2.html#text"},{"title":"Theorem 4.2 (Gradient Dead Zone) .","offset":21824,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-2.html#text"},{"title":"Theorem 4.3 (Metric Recovery via Subproblem Decomposition) .","offset":22583,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-2.html#text"},{"title":"5 Experiment","offset":24256,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-2.html#text"},{"title":"5.1 Experimental Setup","offset":24273,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-2.html#text"},{"title":"Models.","offset":24301,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-2.html#text"},{"title":"Training Setup.","offset":24580,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-2.html#text"},{"title":"Benchmark.","offset":25019,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-2.html#text"},{"title":"Baseline Settings.","offset":25200,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-2.html#text"},{"title":"5.2 Main Results and Further Analysis","offset":25457,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-2.html#text"},{"title":"Superior Performance Across All Benchmarks.","offset":25649,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-2.html#text"},{"title":"Curriculum progress transfers to hard-problem solving.","offset":26305,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-2.html#text"},{"title":"SCRL does not rely on highly curated subproblems.","offset":27161,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-2.html#text"},{"title":"Subproblem-level normalization enables better credit assignment.","offset":28394,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-2.html#text"},{"title":"6 Conclusion","offset":29150,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-2.html#text"},{"title":"Appendix A More Ablation Study","offset":29722,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-2.html#text"},{"title":"Number of subproblems.","offset":29757,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-2.html#text"},{"title":"Training data construction.","offset":31171,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-3.html#text"},{"title":"Appendix B Proofs for Section 4","offset":33234,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-3.html#text"},{"title":"B.1 Proof of Theorem 4.2 (Bound on λ min ​ ( 𝑭 x ​ ( θ ) ) \\lambda_{\\min}(\\bm{F}_{x}(\\theta)) )","offset":33270,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-3.html#text"},{"title":"B.2 Proof of Theorem 4.3","offset":34958,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-3.html#text"},{"title":"First claim.","offset":34988,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-3.html#text"},{"title":"Second claim.","offset":38330,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-3.html#text"},{"title":"B.3 Bound on C A ^ C_{\\hat{A}}","offset":38804,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-3.html#text"},{"title":"Appendix C SCRL Training Algorithm","offset":39750,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-3.html#text"},{"title":"Appendix D OOD Task Performance","offset":39789,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-3.html#text"},{"title":"SCRL generalizes to out-of-distribution tasks.","offset":39825,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-3.html#text"},{"title":"Appendix E Detailed Experimental Results","offset":41009,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-3.html#text"},{"title":"Appendix F Implementation Details","offset":41191,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-3.html#text"},{"title":"F.1 Hyperparameters","offset":41229,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-3.html#text"},{"title":"F.2 Low-Variance pass@ k k Estimation","offset":41449,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-3.html#text"},{"title":"F.3 Baseline Implementation Details","offset":42194,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-3.html#text"},{"title":"SFT.","offset":42235,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-3.html#text"},{"title":"GRPO.","offset":42597,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-3.html#text"},{"title":"DAPO.","offset":42776,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-3.html#text"},{"title":"QuestA.","offset":42946,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-3.html#text"},{"title":"NuRL.","offset":43311,"level":5,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-3.html#text"},{"title":"Appendix G Limitations and Future Work","offset":43656,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-3.html#text"},{"title":"Appendix H Hardware Setup","offset":44464,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-3.html#text"},{"title":"Appendix I Prompt for Subproblem Generation","offset":44643,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-3.html#text"},{"title":"Appendix J Chat Template","offset":44691,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-3.html#text"},{"title":"J.1 Chat Template of Curriculum Learning","offset":44720,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-3.html#text"},{"title":"J.2 Chat Template of Original Problem","offset":44766,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-3.html#text"},{"title":"Appendix K Case Study","offset":44809,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-3.html#text"},{"title":"Instructions for reporting errors","offset":44910,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/part-3.html#text"}],"text_code_points":45874,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/e51c03db071e17a4d5403433f1037e79c3cda795e0b20f6a490df35167411dd0/source.md"}