ReGRPO — Grounded Reflection and Recovery
Learning structured reflection and corrective tool use; ECCV 2026.
ReGRPO learns grounded reflection and recovery for tool-using multimodal agents. Accepted at ECCV 2026; first author.
- A reflective data engine executes near-miss actions to collect grounded failure observations.
- Structured ErrorType / Evidence / FixPlan reflections pair those failures with corrective tool actions for supervised warm-start training.
- Policy optimization jointly learns reflections and corrective actions within local trajectories, with a cost term discouraging unnecessary reflection.
- The underlying GRPO estimator is retained; the contribution lies in grounded recovery data, trajectory structure, and the training protocol.