Deeper Thought, Weaker Aim: Understanding and Mitigating Perceptual Impairment during Reasoning in Multimodal Large Language Models

TL;DR

This paper investigates perceptual impairments in multimodal large language models during reasoning, identifies attention dispersion as a key cause, and proposes a training-free attention reweighting method to improve visual focus and reasoning accuracy.

Contribution

The paper introduces a novel analysis of attention dispersion in MLLMs and proposes VRGA, a training-free framework that enhances visual focus during reasoning tasks.

Findings

Attention dispersion correlates with reasoning prompts and reduced focus on relevant regions.

VRGA improves visual grounding and reasoning accuracy on benchmarks.

The method offers interpretable insights into visual information processing in MLLMs.

Abstract

Multimodal large language models (MLLMs) often suffer from perceptual impairments under extended reasoning modes, particularly in visual question answering (VQA) tasks. We identify attention dispersion as the underlying cause: during multi-step reasoning, the model's visual attention becomes scattered and drifts away from question-relevant regions, effectively "losing focus" on the visual input. To better understand this phenomenon, we analyze the attention maps of MLLMs and observe that reasoning prompts significantly reduce attention to regions critical for answering the question. We further find a strong correlation between the model's overall attention on image tokens and the spatial dispersiveness of its attention within the image. Leveraging this insight, we propose a training-free Visual Region-Guided Attention (VRGA) framework that selects visual heads based on an entropy-focus criterion and reweights their attention, effectively guiding the model to focus on question-relevant regions during reasoning. Extensive experiments on vision-language benchmarks demonstrate that our method effectively alleviates perceptual degradation, leading to improvements in visual grounding and reasoning accuracy while providing interpretable insights into how MLLMs process visual information.