CSMR: A Framework for Visual Evidence Acquisition in Multimodal Reasoning

A recent publication on arXiv (2605.28160) presents CSMR, a framework for multimodal reasoning designed to overcome structural shortcomings found in current methodologies. Existing techniques typically either translate visual data into text prior to reasoning—resulting in the loss of intricate details—or engage in end-to-end reasoning within a single domain, leading to linguistic bias and diminished adherence to visual information. CSMR introduces a cognitive scheduling method, enabling a language model to determine the optimal moments to activate a separate visual perception module for gathering visual evidence during the reasoning process. This framework seeks to enhance the timing and integration of visual information in reasoning tasks.