SMSP: A Plug-and-Play Strategy of Multi-Scale Perception for MLLMs to Perceive Visual Illusions

TL;DR

This paper identifies the high-frequency attention bias in Multimodal Large Language Models (MLLMs) as a cause of vulnerability to visual illusions and proposes a multi-scale perception strategy (SMSP) to improve their perception, significantly boosting accuracy.

Contribution

The paper introduces SMSP, a plug-and-play framework that aligns MLLMs' perception with human visual strategies by suppressing distracting high-frequency backgrounds.

Findings

SMSP improves MLLMs' accuracy on illusion images from 13.0% to 84.0%.

High-frequency attention bias causes MLLMs to overlook hidden patterns.

SMSP effectively mitigates the impact of visual illusions on MLLMs.

Abstract

Recent works have shown that Multimodal Large Language Models (MLLMs) are highly vulnerable to hidden-pattern visual illusions, where the hidden content is imperceptible to models but obvious to humans. This deficiency highlights a perceptual misalignment between current MLLMs and humans, and also introduces potential safety concerns. To systematically investigate this failure, we introduce IlluChar, a comprehensive and challenging illusion dataset, and uncover a key underlying mechanism for the models' failure: high-frequency attention bias, where the models are easily distracted by high-frequency background textures in illusion images, causing them to overlook hidden patterns. To address the issue, we propose the Strategy of Multi-Scale Perception (SMSP), a plug-and-play framework that aligns with human visual perceptual strategies. By suppressing distracting high-frequency backgrounds, SMSP generates images closer to human perception. Our experiments demonstrate that SMSP significantly improves the performance of all evaluated MLLMs on illusion images, for instance, increasing the accuracy of Qwen3-VL-8B-Instruct from 13.0% to 84.0%. Our work provides novel insights into MLLMs' visual perception, and offers a practical and robust solution to enhance it. Our code is publicly available at https://github.com/Tujz2023/SMSP.