Archive

Role-Agent framework uses dual-role LLM co-evolution to boost agent performance

Role-Agent demonstrates that a single LLM can bootstrap its own agent training by self-generating both process rewards and targeted practice tasks, achieving consistent gains over strong baselines without requiring separate environment models.