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

TL;DR

This paper introduces EAH, a training-free method that enhances attention heads in multimodal large language models to reduce hallucinations by focusing more on image tokens, improving model reliability.

Contribution

The paper proposes a novel, training-free approach called EAH that strengthens attention to image tokens in shallow layers to mitigate hallucinations in MLLMs.

Findings

EAH significantly reduces hallucinations across various models.

Enhancing shallow layer attention sinks improves model focus on image content.

The method demonstrates broad effectiveness and generality in experiments.

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.