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

2026-07-23

AI Summary

This work addresses the inconsistency in cross-modal behavior of current vision-language models on visual reasoning tasks such as geometry, which limits their reliability compared to large language models. To tackle this, the authors introduce ODA-Data, a paired geometric reasoning dataset encompassing three perspectives: text-dominant, image-dominant, and multimodal fusion. They propose MIRROR, a novel method that systematically reveals the complementary nature and divergent failure modes across these views, leveraging the best-performing view as a teacher to guide the training of others. Through a reinforcement learning framework based on reverse KL divergence, combined with multi-view self-evaluation and cross-modal knowledge transfer, MIRROR enables self-supervised optimization. Experiments demonstrate that this approach significantly outperforms standard reinforcement learning baselines on multiple geometric reasoning benchmarks, substantially improving both accuracy and consistency in cross-modal reasoning.

Abstract

Unlike large language models (LLMs) that exhibit strong reasoning capabilities, vision-language models (VLMs) struggle with visual reasoning, even on geometry problems that admit equivalent text, diagram, and combined diagram+text views. We show that these views often elicit different behaviors: a model may solve a problem from text but fail on the corresponding diagram, or succeed visually while failing textually. This inconsistency suggests that different views expose complementary reasoning paths and failure modes that standard multimodal post-training does not fully exploit. To study and exploit this phenomenon, we construct ODA-Data, a high-quality paired multimodal geometry dataset with text-dominant, image-dominant, and combined image+text views of the same problems, together with splits for training and evaluating modality-dependent reasoning behaviors. We then develop Modality-Informed Reciprocal Reasoning Optimization (MIRROR), a reinforcement learning approach for improving multimodal reasoning via self supervision. For each problem, MIRROR evaluates the model under all views, selects the best-performing view as a teacher, and trains other views with a reverse-KL objective towards the teacher. Across reasoning benchmarks that evaluate on geometry problems, MIRROR improves over standard RL and yields more accurate and consistent behavior across modalities

Problem

• multimodal reasoning
• vision-language models
• modality inconsistency
• geometric reasoning
• complementary views

Innovation

• multimodal reasoning
• reinforcement learning
• self-supervision
• modality consistency
• vision-language models