Visual Reasoning with Chain of Thought: An Empirical Study

摘要

We investigate Chain of Thought prompting for visual language models.

Introduction

Visual question answering requires complex reasoning over image content.

Method

We propose a visual CoT prompting strategy that breaks down reasoning into steps.

Experiments

Our experiments span 5 VLM models and 3 visual QA datasets.

Analysis

We analyze failure cases and identify key factors for CoT success.

Conclusion

CoT is particularly effective for multi-step visual reasoning tasks.