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

2026-07-23 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

MIRROR (Modality-Informed Reciprocal Reasoning Optimization) is a new reinforcement learning approach designed to enhance vision-language models' (VLMs) reasoning capabilities, particularly for geometry problems. Unlike large language models (LLMs) which excel at reasoning, VLMs often exhibit inconsistent performance across different input modalities—text, diagram, or combined views—for the same problem. This inconsistency suggests complementary reasoning paths. To address this, researchers developed ODA-Data, a high-quality paired multimodal geometry dataset. MIRROR leverages this by evaluating a model across all views for a given problem, identifying the best-performing view as a "teacher," and then training the other views using a reverse-KL objective to align with the teacher's superior performance. This self-supervised method significantly improves over standard reinforcement learning, leading to more accurate and consistent multimodal reasoning behavior on geometry benchmarks.

Key takeaway

For Machine Learning Engineers developing vision-language models for complex reasoning tasks, particularly geometry, standard multimodal post-training may not suffice. You should recognize that different input modalities expose complementary reasoning paths and inconsistencies. Implementing a self-supervised approach, such as MIRROR's method of training weaker modalities towards a stronger "teacher" view using a reverse-KL objective, can significantly improve reasoning accuracy and consistency across modalities in your models.

Key insights

MIRROR improves VLM reasoning consistency by using self-supervision from a problem's best-performing modality view.

Principles

• Inconsistent VLM performance across modalities reveals complementary reasoning paths.
• Learning from a "best view" teacher can improve other views' reasoning.

Method

MIRROR evaluates a model across all problem views, selects the best-performing view as a teacher, and trains other views with a reverse-KL objective.

In practice

• Construct paired multimodal datasets for modality-dependent reasoning.
• Apply self-supervision with a "best view" teacher for VLM consistency.

Topics

• MIRROR
• Vision-Language Models
• Multimodal Reasoning
• Reinforcement Learning
• Geometry Problems
• Self-Supervision

Best for: Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• MIRROR: Learning from the Other View for Multi-Modal Reasoning — Multimodal models show inconsistent reasoning across views, suggesting complementary paths that self-supervision can exploit.
• Reason, Then Re-reason: Cross-view Revisiting Improves Spatial Reasoning — Revisiting spatial hypotheses with synthesized complementary views significantly enhances MLLM reasoning from egocentric videos.
• Reinforcing 3D Understanding in Point-VLMs via Geometric Reward Credit Assignment — Targeted reward assignment and geometric consistency verification significantly improve 3D spatial reasoning in Point-VLMs.
• Seeing Together:Multi-Robot Cooperative Egocentric Spatial Reasoning with Multimodal Large Language Models — Multi-robot egocentric spatial reasoning benefits from physics-informed MLLMs and specialized cooperative datasets.
• Self-Questioning Vision-Language Models: Reinforcement Learning for Compositional Visual Reasoning — Rewarding VLMs for generating intermediate sub-questions enables them to autonomously discover compositional reasoning strategies for complex visual tasks.
• Visual-OPSD: Cross-Modal On-Policy Self-Distillation for Efficient Unified Multimodal Reasoning — Visual-OPSD distills reasoning from expensive visual thought generation pathways into efficient text-only models, achieving substantial speedup and performance gains.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.