CASHEW: Stabilizing Multimodal Reasoning via Iterative Trajectory Aggregation
CASHEW introduces CASHEW and CASHEW-RL to stabilize unstable multi-step reasoning in vision-language models by aggregating multiple candidate reasoning trajectories at inference time. CASHEW uses visual verification during aggregation to ground claims in evidence, while CASHEW-RL learns this aggregation policy through Group Sequence Policy Optimization (GSPO) using a composite reward that balances correctness, evidence quality, and task difficulty.