Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note2 min read

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

Source

arXiv

arxiv.org/abs/2607.21552v1

Read the source

Source

Plain-Language Brief

The paper presents MIRROR, a method to improve multi-modal reasoning in AI models by learning from different views (text, image, combined text and image) of the same problems. The study reveals that vision-language models struggle with visual reasoning, even on simple geometry problems, and suggests that different views expose complementary reasoning paths and failure modes. To address this, they create a high-quality dataset called ODA-Data with multiple views of geometry problems for training and evaluation. MIRROR is a reinforcement learning approach that uses self-supervision to improve multimodal reasoning by evaluating the model under all views, selecting the best-performing view as a teacher, and training other views towards the teacher. The method improves over standard reinforcement learning and results in more accurate and consistent behavior across modalities.

What the Document Says

  • The paper discusses the inconsistencies in visual reasoning of vision-language models (VLMs) compared to large language models (LLMs).
  • A new dataset, ODA-Data, is constructed with text-dominant, image-dominant, and combined image+text views of geometry problems.
  • MIRROR, a reinforcement learning approach, is developed for improving multimodal reasoning via self-supervision.

Key Concepts Mentioned

  • multi-modal reasoning
  • vision-language models
  • large language models
  • ODA-Data
  • MIRROR

Claims Made by the Document

  • VLMs struggle with visual reasoning on simple geometry problems.
  • Different views expose complementary reasoning paths and failure modes.

Evidence Used

  • Unlike large language models that exhibit strong reasoning capabilities, vision-language models (VLMs) struggle...
  • This inconsistency suggests...

Methods / Approach

  • Construction of ODA-Data dataset
  • Development of MIRROR reinforcement learning approach

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

The study does not provide information about the limitations of the MIRROR method.

Open Questions

  • What specific failure modes are exposed by different views?
  • How does MIRROR compare to other methods in terms of improving multi-modal reasoning?

Notes for Future Q&A

May include interpretation. Labeled where it does.

The paper focuses on geometry problems, so it's unclear if the inconsistencies and benefits of MIRROR apply to other problem types as well.

Raw Abstract / Excerpt

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