Seeing Clearly by Layer Two: Enhancing Attention Heads to Alleviate Hallucination in LVLMs

TL;DR

<3-5 sentence high-level summary> The paper tackles hallucinations in multimodal large language models by analyzing how attention sinks concentrate image-token information in early layers. It introduces Enhancing Attention Heads (EAH), a training-free plug-and-play method that identifies the densest vision-sink head in shallow layers and broadcasts its attention to other heads to enforce a dense, global attention to image content. The authors demonstrate EAH's effectiveness across multiple LVLMs and even some LLMs on hallucination benchmarks (e.g., POPE, CHAIR) and general vision-language tasks, without adding inference cost. This work highlights the critical role of early-layer information flow and offers a practical, generalizable tool for reducing image-token–driven hallucinations in multimodal models.

Abstract

The hallucination problem in multimodal large language models (MLLMs) remains a common issue. Although image tokens occupy a majority of the input sequence of MLLMs, there is limited research to explore the relationship between image tokens and hallucinations. In this paper, we analyze the distribution of attention scores for image tokens across each layer and head of the model, revealing an intriguing and common phenomenon: most hallucinations are closely linked to the pattern of attention sinks in the self-attention matrix of image tokens, where shallow layers exhibit dense attention sinks and deeper layers show sparse attention sinks. We further analyze the attention heads of different layers and find that heads with high-density attention sink in the image part play a positive role in alleviating hallucinations. In this paper, we propose a training-free method named \textcolor{red}{\textbf{E}}nhancing \textcolor{red}{\textbf{A}}ttention \textcolor{red}{\textbf{H}}eads (EAH), an approach designed to enhance the convergence of image tokens attention sinks in the shallow layers. EAH identifies the attention head that shows the vision sink in a shallow layer and extracts its attention matrix. This attention map is then broadcast to other heads in the layer, thereby strengthening the layer to pay more attention to the image itself. With extensive experiments, EAH shows significant hallucination-mitigating performance on different MLLMs and metrics, proving its effectiveness and generality.

Figures (19)

• Figure 1: We found a common phenomenon through the attention map: In the range of image token, the attention head of shallow Sparse attention sink is prone to hallucination, while the attention head of Dense attention sink is much less likely to hallucinate.
• Figure 2: Definition of dense vision sink head and its layer-wise distribution. In this case, $\beta$ = 0.0015, $\gamma$ = 15%
• Figure 3: Relationship between text tokens and the average proportion of dense vision sink heads within a single layer by layer2, analyzed across 5,000 randomly selected MSCOCO images using LLaVA1.5-7B.
• Figure 4: (a) A example of distribution of dense vision head and the corresponding proportions/densities of vision sinks within these heads when model output hallucination token; (b) Relationship between the average skewness and CHAIR$_{I}$ on 150 randomly selected MSCOCO images, using LLaVA1.5-7B for captioning; (c) Comparative skewness scatter plot for Hallucination and Non-Hallucination classification on 150 randomly selected MSCOCO images, using LLaVA1.5-7B for VQA.
• Figure 5: Differences in analysis perspectives between Label words label-words, DOPRA DOPRA, OPERA opera and our method.