Optimizing Latent Visual Representations for Multimodal Reasoning

Research identifies limitations in two-stage multimodal training involving visual chain-of-thought. The authors propose improvements to address suboptimal latent representations from off-the-shelf vision encoders and rigid RL regularization during the refinement stage.