Retrospective Progress-Aware Self-Refinement for LLM Agent Training

In three linesRePro, a training framework for LLM agents, teaches models to retrospectively self-generate progress signals via a forward-then-reflect rollout paradigm. Tested on WebShop, ALFWorld, and Sokoban with Qwen family, RePro achieves up to 12% absolute success rate gains without continuous external supervision.

Summary generated by Claude — human-verified