MLLMs Know Where to Look: Training-free Perception of Small Visual Details with Multimodal LLMs

2025-02-24

AI Summary

This work identifies a “localization-accuracy dissociation” phenomenon in multimodal large language models (MLLMs): while their self-attention mechanisms accurately localize small objects in images, downstream reasoning frequently overlooks fine-grained visual details, causing substantial performance degradation on small-object visual question answering (VQA) as object size decreases. To address this, we propose a training-free, plug-and-play visual intervention method that dynamically enhances fine-grained visual representations by fusing self-attention maps with gradient-based saliency maps. Our approach requires no fine-tuning, additional data, or architectural modifications, and is compatible with arbitrary MLLMs. Extensive experiments across two mainstream MLLMs and seven standard VQA benchmarks demonstrate consistent and significant improvements in fine-detail VQA accuracy. The results validate the method’s generality, effectiveness, and zero-training-overhead advantage.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Large Vision Models
• Natural Language Processing: Language Grounding & Multi-modal NLP

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Search and Retrieval-Augmented AI: Large language models for search

Abstract

Multimodal Large Language Models (MLLMs) have experienced rapid progress in visual recognition tasks in recent years. Given their potential integration into many critical applications, it is important to understand the limitations of their visual perception. In this work, we study whether MLLMs can perceive small visual details as effectively as large ones when answering questions about images. We observe that their performance is very sensitive to the size of the visual subject of the question, and further show that this effect is in fact causal by conducting an intervention study. Next, we study the attention patterns of MLLMs when answering visual questions, and intriguingly find that they consistently know where to look, even when they provide the wrong answer. Based on these findings, we then propose training-free visual intervention methods that leverage the internal knowledge of any MLLM itself, in the form of attention and gradient maps, to enhance its perception of small visual details. We evaluate our proposed methods on two widely-used MLLMs and seven visual question answering benchmarks and show that they can significantly improve MLLMs' accuracy without requiring any training. Our results elucidate the risk of applying MLLMs to visual recognition tasks concerning small details and indicate that visual intervention using the model's internal state is a promising direction to mitigate this risk.

Problem

• MLLMs' perception of small visual details
• Impact of visual subject size on MLLMs
• Training-free methods to enhance MLLMs' accuracy

Innovation

• Training-free visual intervention methods
• Leverage attention and gradient maps
• Enhance perception of small details