See What You Are Told: Visual Attention Sink in Large Multimodal Models

• URL: http://arxiv.org/abs/2503.03321v1
• Date: Wed, 05 Mar 2025 09:55:07 GMT
• Title: See What You Are Told: Visual Attention Sink in Large Multimodal Models
• Authors: Seil Kang, Jinyeong Kim, Junhyeok Kim, Seong Jae Hwang,
• Abstract summary: Large multimodal models (LMMs) "see" images by leveraging the attention mechanism between text and visual tokens in the transformer decoder.<n>Recent findings indicate that LMMs have an extraordinary tendency to consistently allocate high attention weights to specific visual tokens.<n>We introduce Visual Attention Redistribution ( VAR), a method that redistributes attention in image-centric heads.
• Score: 4.024850952459758
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Large multimodal models (LMMs) "see" images by leveraging the attention mechanism between text and visual tokens in the transformer decoder. Ideally, these models should focus on key visual information relevant to the text token. However, recent findings indicate that LMMs have an extraordinary tendency to consistently allocate high attention weights to specific visual tokens, even when these tokens are irrelevant to the corresponding text. In this study, we investigate the property behind the appearance of these irrelevant visual tokens and examine their characteristics. Our findings show that this behavior arises due to the massive activation of certain hidden state dimensions, which resembles the attention sink found in language models. Hence, we refer to this phenomenon as the visual attention sink. In particular, our analysis reveals that removing the irrelevant visual sink tokens does not impact model performance, despite receiving high attention weights. Consequently, we recycle the attention to these tokens as surplus resources, redistributing the attention budget to enhance focus on the image. To achieve this, we introduce Visual Attention Redistribution (VAR), a method that redistributes attention in image-centric heads, which we identify as innately focusing on visual information. VAR can be seamlessly applied across different LMMs to improve performance on a wide range of tasks, including general vision-language tasks, visual hallucination tasks, and vision-centric tasks, all without the need for additional training, models, or inference steps. Experimental results demonstrate that VAR enables LMMs to process visual information more effectively by adjusting their internal attention mechanisms, offering a new direction to enhancing the multimodal capabilities of LMMs.

Related papers

• Introducing Visual Perception Token into Multimodal Large Language Model [53.82301522384719]
  Multimodal Large Language Model (MLLM) relies on the perception process of its vision encoder.<n>MLLM still lacks the autonomous capability to control its own visual perception processes.<n>We propose the concept of Visual Perception Token, aiming to empower MLLM with a mechanism to control its visual perception processes.
  arXiv  Detail & Related papers  (2025-02-24T18:56:12Z)
• AdaFV: Rethinking of Visual-Language alignment for VLM acceleration [7.9213473377478865]
  Some approaches to reduce the visual tokens according to the self-attention of VLMs, which are biased, result in inaccurate responses.<n>We propose a self-adaptive cross-modality attention mixture mechanism that dynamically leverages the effectiveness of visual saliency and text-to-image similarity.<n>The proposed approach achieves state-of-the-art training-free VLM acceleration performance, especially when the reduction rate is sufficiently large.
  arXiv  Detail & Related papers  (2025-01-16T13:34:33Z)
• [CLS] Token Tells Everything Needed for Training-free Efficient MLLMs [66.5266435598799]
  Multi-language Large Language Models (MLLMs) have recently demonstrated strong performance across a wide range of vision tasks.<n>However, their efficient deployment remains a substantial challenge due to high computational costs and memory requirements.<n>We introduce a simple yet effective method for train-free visual compression, called VTC- compression.
  arXiv  Detail & Related papers  (2024-12-08T05:29:39Z)
• What's in the Image? A Deep-Dive into the Vision of Vision Language Models [20.669971132114195]
  Vision-Language Models (VLMs) have recently demonstrated remarkable capabilities in comprehending complex visual content. In this paper, we conduct a thorough empirical analysis, focusing on attention modules across layers. We reveal several key insights about how these models process visual data.
  arXiv  Detail & Related papers  (2024-11-26T14:59:06Z)
• Shifting Focus with HCEye: Exploring the Dynamics of Visual Highlighting and Cognitive Load on User Attention and Saliency Prediction [3.2873782624127834]
  This paper examines the joint impact of visual highlighting (permanent and dynamic) and dual-task-induced cognitive load on gaze behaviour. We show that state-of-the-art saliency models increase their performance when accounting for different cognitive loads.
  arXiv  Detail & Related papers  (2024-04-22T14:45:30Z)
• LLaVA-PruMerge: Adaptive Token Reduction for Efficient Large Multimodal Models [35.88374542519597]
  Large Multimodal Models (LMMs) have shown significant visual reasoning capabilities by connecting a visual encoder and a large language model. Recent LMMs incorporate more complex visual inputs, such as high-resolution images and videos, which further increases the number of visual tokens significantly. We propose PruMerge, a novel adaptive visual token reduction strategy that significantly reduces the number of visual tokens without compromising the performance of LMMs.
  arXiv  Detail & Related papers  (2024-03-22T17:59:52Z)
• Chain-of-Spot: Interactive Reasoning Improves Large Vision-Language Models [81.71651422951074]
  Chain-of-Spot (CoS) method is a novel approach that enhances feature extraction by focusing on key regions of interest. This technique allows LVLMs to access more detailed visual information without altering the original image resolution. Our empirical findings demonstrate a significant improvement in LVLMs' ability to understand and reason about visual content.
  arXiv  Detail & Related papers  (2024-03-19T17:59:52Z)
• Multi-modal Auto-regressive Modeling via Visual Words [96.25078866446053]
  We propose the concept of visual tokens, which maps the visual features to probability distributions over Large Multi-modal Models' vocabulary. We further explore the distribution of visual features in the semantic space within LMM and the possibility of using text embeddings to represent visual information.
  arXiv  Detail & Related papers  (2024-03-12T14:58:52Z)
• Unveiling Typographic Deceptions: Insights of the Typographic Vulnerability in Large Vision-Language Model [23.764618459753326]
  The Typographic Attack has also been expected to be a security threat to LVLMs. We verify typographic attacks on current well-known commercial and open-source LVLMs. To better assess this vulnerability, we propose the most comprehensive and largest-scale Typographic dataset to date.
  arXiv  Detail & Related papers  (2024-02-29T13:31:56Z)
• Incorporating Visual Experts to Resolve the Information Loss in Multimodal Large Language Models [121.83413400686139]
  This paper proposes to improve the visual perception ability of MLLMs through a mixture-of-experts knowledge enhancement mechanism. We introduce a novel method that incorporates multi-task encoders and visual tools into the existing MLLMs training and inference pipeline.
  arXiv  Detail & Related papers  (2024-01-06T02:02:34Z)
• Collaborative Attention Mechanism for Multi-View Action Recognition [75.33062629093054]
  We propose a collaborative attention mechanism (CAM) for solving the multi-view action recognition problem. The proposed CAM detects the attention differences among multi-view, and adaptively integrates frame-level information to benefit each other. Experiments on four action datasets illustrate the proposed CAM achieves better results for each view and also boosts multi-view performance.
  arXiv  Detail & Related papers  (2020-09-14T17:33:10Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.