Walk the Talk: Bridging the Reasoning-Action Gap for Thinking with Images via Multimodal Agentic Policy Optimization

TL;DR

This paper introduces MAPO, a new training method for multimodal models that aligns visual actions with textual reasoning, improving accuracy in visual reasoning tasks by reducing reasoning-action discrepancies.

Contribution

MAPO is a novel approach that explicitly links visual actions with textual descriptions, enhancing multimodal reasoning and addressing the reasoning-action gap in large language models.

Findings

MAPO reduces gradient variance and improves training stability.

Models trained with MAPO outperform existing methods on visual reasoning benchmarks.

Explicit textual descriptions of visual content enhance reasoning accuracy.

Abstract

Recent advancements in Multimodal Large Language Models (MLLMs) have incentivized models to ``think with images'' by actively invoking visual tools during multi-turn reasoning. The common Reinforcement Learning (RL) practice of relying on outcome-based rewards ignores the fact that textual plausibility often masks executive failure, meaning that models may exhibit intuitive textual reasoning while executing imprecise or irrelevant visual actions within their agentic reasoning trajectories. This reasoning-action discrepancy introduces noise that accumulates throughout the multi-turn reasoning process, severely degrading the model's multimodal reasoning capabilities and potentially leading to training collapse. In this paper, we introduce Multimodal Agentic Policy Optimization (MAPO), bridging the gap between textual reasoning and visual actions generated by models within their Multimodal Chain-of-Thought (MCoT). Specifically, MAPO mandates the model to generate explicit textual descriptions for the visual content obtained via tool usage. We then employ a novel advantage estimation that couples the semantic alignment between these descriptions and the actual observations with the task reward. Theoretical findings are provided to justify the rationale behind MAPO, which inherently reduces the variance of gradients, and extensive experiments demonstrate that our method achieves superior performance across multiple visual reasoning benchmarks.