The article discusses a new approach to multimodal reasoning that addresses the limitations of current Process Reward Models, which may overlook individual dimension failures. By focusing on worst dimension optimization, the method aims to enhance the integrity and validity of reasoning across various constraints.