Machine Learning

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

Wen Ye, Yuxiao Qu, Aviral Kumar, Xuezhe Ma

Featured July 25, 2026

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

When a smart computer can solve a math problem from words but not from a picture, this method helps it learn from its own success in words to get better at understanding pictures, making it smarter and more consistent.

In depth

The paper introduces MIRROR (Modality-Informed Reciprocal Reasoning Optimization), a reinforcement learning framework that addresses the inconsistency of vision-language models (VLMs) across different input modalities (text, image, combined). It exploits the observation that a VLM might solve a problem in one view but fail in another, using the successful view as a self-supervision signal to improve weaker views through on-policy reverse-KL distillation and adaptive teacher selection.

Key Takeaways

1
VLMs exhibit modality asymmetry, meaning they can solve a problem from one input view (e.g., text) but fail on an equivalent problem presented in another view (e.g., image).
2
MIRROR leverages this asymmetry by adaptively selecting the best-performing view for each problem as an internal teacher, then uses on-policy reverse-KL distillation to guide weaker student views.
3
The approach significantly improves multimodal reasoning accuracy and cross-view consistency on geometry problems, outperforming standard reinforcement learning baselines with less training data.

Conceptual Flow

HIGH LEVEL

Methodology: Learning from Self-Disagreement

The computer tries to solve the same puzzle using words, pictures, and both. If it solves it one way, that good answer helps teach it how to solve it the other ways.

Puzzle in Words

Puzzle in Pictures

Puzzle in Both

Find Best Answer

Best Answer (Teacher)