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

TL;DR

This work investigates why multimodal LLMs struggle with small visual details and whether they can be steered to better perceive them without training. It shows that perception deteriorates with smaller visual concepts and demonstrates, via intervention, a causal link to concept size while revealing that MLLMs often know where to look. The authors introduce ViCrop, training-free visual cropping methods that exploit internal attention and gradient signals to crop regions of interest, significantly boosting accuracy on detail-sensitive VQA benchmarks while preserving performance on larger concepts. Through extensive experiments on multiple MLLMs and datasets, ViCrop proves to be a practical, inference-time solution to mitigate small-detail perception risks in MLLMs. The work highlights both the limitations and potential of internal-state–driven interventions for robust multimodal reasoning.

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.

Figures (8)

• Figure 1: The effect of visual cropping on the probability of answers predicted by BLIP-2 FlanT5$_\mathrm{XL}$ zero-shot VQA model. The x-axis labels are indices to the respective cropped images displayed under each plot that the model sees at each step. The model gradually finds the correct answer.
• Figure 2: Examples of MLLMs knowing where to look despite answering incorrectly. The right panel in each example displays relative attention to the image (defined in \ref{['sec:where_to_look']}) of one layer in the MLLM.
• Figure 3: MLLMs' attention ratio across all layers (average with $95\%$ CI over TextVQA). The attention ratio measures how significantly the MLLM is attending to the ground-truth bounding box (defined in \ref{['sec:where_to_look']}). We observe that it is greater than 1 in most layers, showing that the MLLMs know where to look in the image even when they fail to answer correctly.
• Figure 4: Illustration of the proposed visual cropping approach applied to two MLLMs.
• Figure 5: Examples of rel-att helping MLLMs correct their mistakes (cyan-colored bounding box shows cropped region by rel-att; zoom-in insets are displayed for better readability).