phenomenon · first seen 4 Oct 2026
policy entropy collapse
Also: entropy drift, policy collapse, entropy collapse, premature convergence
Premature drop in a policy's output entropy during reinforcement learning, which limits exploration and stalls improvement.
Mentions per week
PapersCodeDiscussion
Earliest sources
- Long-MDR: Long-Context Reinforcement Learning for Multimodal Deep-Research AgentsPapers · 4 Oct 2026 · On Tai Tang, Huanyao Zhang, Jiepeng Zhou, Wentao Zhang
- GRPODropout: Less is More for Online Reinforcement Learning RolloutsPapers · 8 Oct 2026 · Hexuan Deng, Zihao Yan, Xuebo Liu, Shuo Nie, Yue Wang, Chen Wang, Zhaohua Zhang,…
- A Sign Law for Entropy Drift under Sampler Mismatch in RLVRPapers · 25 Oct 2026 · Zhiyuan Wang, Jacob Steeves, Chang Liu
- Collective Dynamics of Self-Reinforcement: How Cross-Task Coupling Drives Policy Collapse in Unsupervised RLVRPapers · 27 Oct 2026 · Peize Ding
- There Is No Free Lunch in RLVR: Every pass@1 Gain Has a pass@k PricePapers · 29 Oct 2026 · Yinuo Liu
- When Does Exploration Prevent Collapse? Three Regimes and the Path of the Reward in Self-Consistency TrainingPapers · 31 Oct 2026 · Jaeho Hwang, Uiwon Hwang
Latest sources
- Reward Budgeting Reduces Premature Convergence in Reinforcement Learning for LLM ReasoningPapers · 4 Nov 2026 · Mengni Jia, Mengyu Zhou, Xiaoxi Jiang, Guanjun Jiang
- When Does Exploration Prevent Collapse? Three Regimes and the Path of the Reward in Self-Consistency TrainingPapers · 31 Oct 2026 · Jaeho Hwang, Uiwon Hwang
- There Is No Free Lunch in RLVR: Every pass@1 Gain Has a pass@k PricePapers · 29 Oct 2026 · Yinuo Liu
- Collective Dynamics of Self-Reinforcement: How Cross-Task Coupling Drives Policy Collapse in Unsupervised RLVRPapers · 27 Oct 2026 · Peize Ding
- A Sign Law for Entropy Drift under Sampler Mismatch in RLVRPapers · 25 Oct 2026 · Zhiyuan Wang, Jacob Steeves, Chang Liu
- GRPODropout: Less is More for Online Reinforcement Learning RolloutsPapers · 8 Oct 2026 · Hexuan Deng, Zihao Yan, Xuebo Liu, Shuo Nie, Yue Wang, Chen Wang, Zhaohua Zhang,…
- Long-MDR: Long-Context Reinforcement Learning for Multimodal Deep-Research AgentsPapers · 4 Oct 2026 · On Tai Tang, Huanyao Zhang, Jiepeng Zhou, Wentao Zhang