Active Visual Thinking Improves Reasoning

This paper introduces FIGR, a novel approach that integrates active visual thinking into multi-turn reasoning. It addresses the limitations of text-based reasoning in handling complex spatial, geometric, and structural relationships. The use of reinforcement learning to control visual reasoning and the construction of visual representations are key innovations. The paper's significance lies in its potential to improve the stability and reliability of reasoning models, especially in domains requiring understanding of global structural properties. The experimental results on challenging mathematical reasoning benchmarks demonstrate the effectiveness of the proposed method.

• •FIGR integrates active visual thinking into multi-turn reasoning.
• •Uses reinforcement learning to control visual reasoning.
• •Improves performance on challenging mathematical reasoning benchmarks.
• •Addresses limitations of text-based reasoning in handling spatial and structural relationships.