Tag
This paper rethinks the role of grouping in critic-free reinforcement learning for LLMs and proposes negative token filtering to enable stable training with a single rollout per prompt, achieving comparable or better performance on reasoning and agentic tasks.