CAVE: A Structured Credit Assignment Approach for Fragmented Visual Evidence Reasoning

In three linesCAVE is a credit assignment method based on GRPO to improve fragmented visual reasoning in VLMs. It evaluates intermediate steps via three signals: belief update, evidence acquisition, and adaptive focus control. TRACER-Bench, a new benchmark, assesses reasoning across four nonlocal and semantically confusable dimensions.

Summary generated by Claude — human-verified