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

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

Summary

The MIRROR (Modality-Informed Reciprocal Reasoning Optimization) framework addresses the challenge of vision-language models (VLMs) in visual reasoning, particularly for geometry problems, where they often underperform large language models (LLMs). The research identifies that VLMs exhibit inconsistent reasoning across different modalities—text, diagram, and combined diagram+text views—even for equivalent problems. To study and exploit this, the authors constructed ODA-Data, a high-quality paired multimodal geometry dataset featuring text-dominant, image-dominant, and combined views. MIRROR is a reinforcement learning approach that employs self-supervision: it evaluates a model across all available views for a given problem, designates the best-performing view as a teacher, and then trains the other views using a reverse-KL objective to align with the teacher's output. This method significantly improves multimodal reasoning on geometry benchmarks, yielding more accurate and consistent behavior compared to standard RL techniques.

Key takeaway

For AI Scientists developing vision-language models, if you are struggling with inconsistent multimodal reasoning, consider implementing self-supervision techniques like MIRROR. This approach, which uses a "teacher" view from the best-performing modality to train others via a reverse-KL objective, can significantly improve accuracy and consistency on complex tasks like geometry problems. You should explore creating paired multimodal datasets to facilitate such training.

Key insights

Multimodal models show inconsistent reasoning across views, suggesting complementary paths that self-supervision can exploit.

Principles

• Different modalities offer complementary reasoning.
• Inconsistent VLM behavior reveals failure modes.
• Self-supervision can align multimodal reasoning.

Method

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

In practice

• Construct paired multimodal datasets.
• Use best-performing view as a teacher.
• Apply reverse-KL for cross-modal alignment.

Topics

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

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

Related on AIssential

• MIRROR: Learning from the Other View for Multi-Modal Reasoning — MIRROR improves VLM reasoning consistency by using self-supervision from a problem's best-performing modality view.
• PaLMR: Towards Faithful Visual Reasoning via Multimodal Process Alignment — Aligning multimodal reasoning processes with visual evidence, not just final answers, is crucial to prevent hallucinations.
• MIRROR: A Hierarchical Benchmark for Metacognitive Calibration in Large Language Models — LLMs possess partial self-knowledge but universally fail to translate it into appropriate agentic action without external architectural control.
• BUS: Brain-Inspired Unsupervised Self-Reflection for Advanced Multimodal Reasoning — BUS enhances VLM reflective reasoning through unsupervised, brain-inspired backward prediction, eliminating annotated data reliance.
• Iterative Visual Thinking: Teaching Vision-Language Models Spatial Self-Correction through Visual Feedback — Iterative Visual Thinking enables VLMs to self-correct spatial predictions through visual feedback, closing a critical performance gap.
• 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.

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 Takara TLDR - Daily AI Papers.