RL for LLM reasoning often reaches an early performance plateau as the policy distribution becomes increasingly sharp. Recent work has studied policy entropy…
机构:阿里
来源:arXiv 2611.05942 | AI4Papers 论文推荐平台