MIRROR: Learning from the Other View for Multi-Modal Reasoning

The paper addresses the challenge of improving visual reasoning in vision-language models (VLMs), which often struggle with problems that can be represented in text, diagrams, or a combination of both. The authors identify that different modalities can lead to varying reasoning paths and failure modes, suggesting that these views can provide complementary insights that are not fully utilized by standard multimodal training approaches.

To explore this, the authors introduce ODA-Data, a dataset containing geometry problems with text-dominant, image-dominant, and combined views. They propose a novel reinforcement learning framework called Modality-Informed Reciprocal Reasoning Optimization (MIRROR), which selects the best-performing view for each problem as a teacher and trains the other views to align with it using a reverse-KL objective.

The results demonstrate that MIRROR significantly improves reasoning accuracy and consistency across modalities compared to standard reinforcement learning approaches. The method achieves notable performance gains on both the ODA-Data and external geometry benchmarks, indicating its effectiveness in leveraging cross-view supervision for enhanced multimodal reasoning.

The study highlights the potential of using modality asymmetry as a self-supervision signal, turning cross-view inconsistencies into a strength rather than a limitation. This approach not only improves accuracy but also increases the set of problems that can be solved consistently across different modalities.

The authors suggest that future work could extend this approach to other domains beyond geometry, such as scientific figures or robotics, and explore more sophisticated teacher selection mechanisms. The findings point to a broader opportunity to use cross-modal inconsistencies as a training signal for developing more robust and versatile multimodal models.