When Sinks Help or Hurt: Unified Framework for Attention Sink in Large Vision-Language Models

• URL: http://arxiv.org/abs/2604.03316v1
• Date: Wed, 01 Apr 2026 09:59:09 GMT
• Title: When Sinks Help or Hurt: Unified Framework for Attention Sink in Large Vision-Language Models
• Abstract summary: This paper first categorizes visual sinks into two distinct categories: ViT-emerged sinks (V-sinks) and LLM-emerged sinks (L-sinks)<n>We propose Layer-wise Sink Gating (LSG), a lightweight, plug-and-play module that dynamically scales the attention contributions of V-sink and the rest visual tokens.
• Score: 36.88302523716978
• License: http://creativecommons.org/licenses/by-nc-nd/4.0/
• Abstract: Attention sinks are defined as tokens that attract disproportionate attention. While these have been studied in single modality transformers, their cross-modal impact in Large Vision-Language Models (LVLM) remains largely unexplored: are they redundant artifacts or essential global priors? This paper first categorizes visual sinks into two distinct categories: ViT-emerged sinks (V-sinks), which propagate from the vision encoder, and LLM-emerged sinks (L-sinks), which arise within deep LLM layers. Based on the new definition, our analysis reveals a fundamental performance trade-off: while sinks effectively encode global scene-level priors, their dominance can suppress the fine-grained visual evidence required for local perception. Furthermore, we identify specific functional layers where modulating these sinks most significantly impacts downstream performance. To leverage these insights, we propose Layer-wise Sink Gating (LSG), a lightweight, plug-and-play module that dynamically scales the attention contributions of V-sink and the rest visual tokens. LSG is trained via standard next-token prediction, requiring no task-specific supervision while keeping the LVLM backbone frozen. In most layers, LSG yields improvements on representative multimodal benchmarks, effectively balancing global reasoning and precise local evidence.

Related papers

• Look Less, Reason More: Block-wise Attention Skipping for Efficient Multimodal LLMs [89.7106332677868]
  We present Visual-Skip, a training-free inference paradigm that decouples spatial interaction from semantic evolution.<n>V-Skip effectively bypasses redundant vision attention to achieve block-wise sparsity, maintaining a 94.16% to 100.31% performance retention across diverse MLLMs.
  arXiv  Detail & Related papers  (2026-06-07T08:32:13Z)
• SLASH the Sink: Sharpening Structural Attention Inside LLMs [55.75072688988806]
  Large Language Models (LLMs) show remarkable semantic understanding but struggle with structural understanding when processing graph topologies in a serialized format.<n>We propose StructuraL Attention SHarpening (SLASH), which amplifies this internal structural understanding via a plug-and-play attention redistribution.
  arXiv  Detail & Related papers  (2026-05-11T12:59:07Z)
• On the Nature of Attention Sink that Shapes Decoding Strategy in MLLMs [38.05844382560401]
  OutRo is a lightweight inference-time strategy that leverages the sink token to enhance contextual representations.<n>Based on experiments, OutRo consistently improves performance across representative MLLMs on seven video QA benchmarks.
  arXiv  Detail & Related papers  (2026-03-15T12:05:35Z)
• TagaVLM: Topology-Aware Global Action Reasoning for Vision-Language Navigation [70.23578202012048]
  Vision-Language Navigation (VLN) presents a unique challenge for Large Vision-Language Models (VLMs) due to their inherent architectural mismatch.<n>We propose TagaVLM (Topology-Aware Global Action reasoning), an end-to-end framework that explicitly injects topological structures into the VLM backbone.<n>To enhance topological node information, an Interleaved Navigation Prompt strengthens node-level visual-text alignment.<n>With the embedded topological graph, the model is capable of global action reasoning, allowing for robust path correction.
  arXiv  Detail & Related papers  (2026-03-03T13:28:07Z)
• Look Carefully: Adaptive Visual Reinforcements in Multimodal Large Language Models for Hallucination Mitigation [51.743225614196774]
  Multimodal large language models (MLLMs) have achieved remarkable progress in vision-language reasoning.<n>They remain vulnerable to hallucination, where generated content deviates from visual evidence.<n>Recent vision enhancement methods attempt to address this issue by reinforcing visual tokens during decoding.<n>We propose Adaptive Visual Reinforcement (AIR), a training-free framework for MLLMs.
  arXiv  Detail & Related papers  (2026-02-27T14:18:51Z)
• Understanding and Enhancing Encoder-based Adversarial Transferability against Large Vision-Language Models [19.899086203883254]
  Large vision-language models (LVLMs) have achieved impressive success across multimodal tasks.<n>Their reliance on visual inputs exposes them to significant adversarial threats.<n>Existing encoder-based attacks perturb the input image by optimizing solely on the vision encoder, rather than the entire LVLM.<n>We present the first systematic study towards encoder-based adversarial transferability in LVLMs.
  arXiv  Detail & Related papers  (2026-02-10T05:51:02Z)
• Can LLMs See Without Pixels? Benchmarking Spatial Intelligence from Textual Descriptions [18.455501447828343]
  Spatial Intelligence (SI) has predominantly relied on Vision-Language Models (VLMs)<n>We introduce SiT-Bench, a novel benchmark designed to evaluate the SI performance of Large Language Models (LLMs) without pixel-level input.<n>We find that explicit spatial reasoning significantly boosts performance, suggesting that LLMs possess latent world-modeling potential.
  arXiv  Detail & Related papers  (2026-01-07T05:13:52Z)
• To Sink or Not to Sink: Visual Information Pathways in Large Vision-Language Models [34.902254997726835]
  Vision Transformer (ViT) encodes visual content into a sequence of image tokens.<n>Large Language Model (LLM) interprets these tokens to perform high-level reasoning.<n>We identify a class of high-norm visual tokens from ViT, referred to as ViT attention sinks.
  arXiv  Detail & Related papers  (2025-10-09T17:44:42Z)
• Spectral Insights into Data-Oblivious Critical Layers in Large Language Models [7.486925126518052]
  We introduce a data-oblivious approach to identify intrinsic critical layers in pre-fine-tuned language models.<n>We show that layers with significant shifts in representation space are also those most affected during fine-tuning.
  arXiv  Detail & Related papers  (2025-05-31T04:21:39Z)
• Boosting Multimodal Large Language Models with Visual Tokens Withdrawal for Rapid Inference [59.91176945361035]
  We introduce Visual Tokens Withdrawal (VTW), a plug-and-play module to boost MLLMs for rapid inference.<n>VTW strategically withdraws vision tokens at a certain layer, enabling only text tokens to engage in subsequent layers.<n>Our approach can cut computational overhead by over 40% across diverse multimodal tasks while maintaining performance.
  arXiv  Detail & Related papers  (2024-05-09T14:38:53Z)

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.