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

TL;DR - MIRROR is a self-supervised reinforcement-learning method that transfers reasoning behavior between textual, visual, and combined views of geometry problems. It improves both accuracy and cross-modal consistency compared with standard RL.

• Introduces ODA-Data, a paired geometry dataset containing equivalent text-dominant, image-dominant, and combined views.
• Identifies complementary, modality-dependent failures in vision-language models.
• Selects the best-performing view as a teacher and aligns other views toward it using a reverse-KL objective.
• Reports gains across geometry reasoning benchmarks without specifying numerical results in the provided abstract.