Search papers, labs, and topics across Lattice.
Tsinghua AIHKUHuaweiLongCat TeamOhio StateShanghai AI LabUMichXiamen UniversityApr 11, 2026arXiv:2604.10098This paper surveys the phenomenon of "Attention Sink" (AS) in Transformers, where a small set of uninformative tokens attract disproportionate attention. It synthesizes research on AS across three dimensions: utilization, mechanistic interpretation, and mitigation strategies. By consolidating existing knowledge and identifying research trends, the survey aims to guide researchers in managing AS and inspiring innovations for future Transformer architectures.
Attention Sink, where Transformers fixate on seemingly irrelevant tokens, is more than just a quirk – it's a fundamental challenge impacting training, inference, and even causing hallucinations, demanding a systematic approach to understanding and mitigating its effects.
As the foundational architecture of modern machine learning, Transformers have driven remarkable progress across diverse AI domains. Despite their transformative impact, a persistent challenge across various Transformers is Attention Sink (AS), in which a disproportionate amount of attention is focused on a small subset of specific yet uninformative tokens. AS complicates interpretability, significantly affecting the training and inference dynamics, and exacerbates issues such as hallucinations. In recent years, substantial research has been dedicated to understanding and harnessing AS. However, a comprehensive survey that systematically consolidates AS-related research and offers guidance for future advancements remains lacking. To address this gap, we present the first survey on AS, structured around three key dimensions that define the current research landscape: Fundamental Utilization, Mechanistic Interpretation, and Strategic Mitigation. Our work provides a pivotal contribution by clarifying key concepts and guiding researchers through the evolution and trends of the field. We envision this survey as a definitive resource, empowering researchers and practitioners to effectively manage AS within the current Transformer paradigm, while simultaneously inspiring innovative advancements for the next generation of Transformers. The paper list of this work is available at https://github.com/ZunhaiSu/Awesome-Attention-Sink.
