Improving Multimodal Reasoning via Worst Dimension Optimization

In three linesPaper introduces worst dimension optimization to improve multimodal reasoning. Current Process Reward Models equally weight factors like visual grounding and logic consistency, potentially concealing individual dimension failures. The approach aims to ensure overall validity of the reasoning process.

Summary generated by Claude — human-verified