{"id":"arxiv:2508.07629","title":"Klear-Reasoner: Advancing Reasoning Capability via Gradient-Preserving Clipping Policy Optimization","date":"2025-08-11","section":"Post-training","quality":-0.228184,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/","source_url":"https://arxiv.org/abs/2508.07629","text_available":true,"authors":["Zhenpeng Su","Leiyu Pan","Xue Bai","Dening Liu"],"outgoing":[{"source":"arxiv:2508.07629","target":"arxiv:2402.03300","evidence":"title","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:2508.07629","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:2508.07629","target":"arxiv:2506.13585","evidence":"title","id":"arxiv:2506.13585","title":"MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/5beb2ec6c7e6362f9d957219d7a205dd1adcc53f739acc0c7fd5f1f2bded4d83/"}],"incoming":[{"source":"arxiv:2512.13961","target":"arxiv:2508.07629","evidence":"id","id":"arxiv:2512.13961","title":"Olmo 3","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/8daa11a9729da9879a7e95c15d4eef3bab8579d2885a14102ed09db111221802/"}],"text_parts":[{"start":0,"end":15351,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-1.html"},{"start":15351,"end":30325,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-2.html"},{"start":30325,"end":46188,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-3.html"},{"start":46188,"end":58208,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-4.html"}],"contents":[{"title":"Klear-Reasoner: Advancing Reasoning Capability via Gradient-Preserving Clipping Policy Optimization","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-1.html#text"},{"title":"Abstract","offset":103,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-1.html#text"},{"title":"1 Introduction","offset":1720,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-1.html#text"},{"title":"2 Preliminary","offset":5809,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-1.html#text"},{"title":"2.1 Proximal Policy Optimization (PPO)","offset":5932,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-1.html#text"},{"title":"2.2 Group Relative Policy Optimization (GRPO)","offset":7980,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-1.html#text"},{"title":"2.3 Raise the Ceiling: Clip-Higher","offset":9894,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-1.html#text"},{"title":"2.4 Token-Level Policy Gradient Loss","offset":11281,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-1.html#text"},{"title":"3 Method","offset":12785,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-1.html#text"},{"title":"3.1 long Chain-of-Thought Supervised Fine-tuning","offset":12936,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-1.html#text"},{"title":"3.2 Reinforcement Learning","offset":15293,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-1.html#text"},{"title":"3.2.1 Data curation","offset":15325,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-1.html#text"},{"title":"3.2.2 Gradient-Preserving Clipping Policy Optimization","offset":16985,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-2.html#text"},{"title":"3.2.3 RL Training with SFT Loss","offset":27451,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-2.html#text"},{"title":"3.2.4 Reward Design for Math and Code RL","offset":28805,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-2.html#text"},{"title":"4 Evaluation","offset":30035,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-2.html#text"},{"title":"4.1 Training Details","offset":30299,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-2.html#text"},{"title":"4.2 Benchmarks and Baselines","offset":32102,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-3.html#text"},{"title":"4.3 Evaluation Results","offset":32860,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-3.html#text"},{"title":"5 Analysis","offset":34918,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-3.html#text"},{"title":"5.1 SFT Analyses","offset":35363,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-3.html#text"},{"title":"5.1.1 Impact of data correctness","offset":35385,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-3.html#text"},{"title":"5.1.2 Ablation of teacher models","offset":38419,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-3.html#text"},{"title":"5.1.3 Data quality and diversity balance","offset":39717,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-3.html#text"},{"title":"5.2 RL Analyses","offset":41910,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-3.html#text"},{"title":"5.2.1 Effectiveness of Gradient-Preserving Clipping Policy Optimization","offset":41931,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-3.html#text"},{"title":"5.2.2 Impact of SFT loss","offset":44579,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-3.html#text"},{"title":"5.2.3 Soft Reward vs. Hard Reward","offset":45721,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-3.html#text"},{"title":"5.2.4 Improving Code RL via Test Case Filtering","offset":47156,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-4.html#text"},{"title":"5.2.5 Impact of Zero-Advantage Sample Filtering","offset":48164,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-4.html#text"},{"title":"6 Conclusions","offset":49978,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-4.html#text"},{"title":"Appendix A Appendix","offset":50951,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-4.html#text"},{"title":"A.1 CISPO Gradient Analysis","offset":50975,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-4.html#text"},{"title":"A.2 The General Form of GPPO","offset":54226,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-4.html#text"},{"title":"A.3 License for Scientific Artifacts","offset":57014,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-4.html#text"},{"title":"Instructions for reporting errors","offset":57244,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/part-4.html#text"}],"text_code_points":58208,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a232533e4e78ad28d2ffe9fafe799b66cf626b18ad523d0088d9cc12f1ba1b76/source.md"}