Closing the Reflection Gap: A Free Calibration Bonus for Agentic RL

In three linesRefGRPO closes the reflection gap in LLM agents: they mis-assess outputs despite correct answers after environment feedback. Method adds a free calibration bonus (contrasting agent reflection vs actual outcome) to standard RL. On text-to-SQL: underconfidence rate 44.4%→7.7%, accuracy 75.1%→76.5%.

Summary generated by Claude — human-verified