From Failure to Feedback: Group Revision Unlocks Hard Cases in Object-Level Grounding

Researchers propose a group-revision reinforcement learning method for object-level grounding in large vision-language models.

The approach samples initial answers, generates revised candidates, and uses improvement-based shaping signals to refine rewards and advantages.

It addresses weak response-level rewards on hard grounding cases and improves performance across referring segmentation, reasoning segmentation, REC, and counting benchmarks.