揭示Transformer中注意力集中在无意义词元的现象及其影响
Attention Sink in Transformers: A Survey on Utilization, Interpretation, and Mitigation
- 分析注意力集中在少数无信息词元的机制
- 指出该现象导致模型不可解释且加剧幻觉问题
- 为研究者提供系统性综述与未来方向指引
作为现代机器学习的基础架构，Transformer 在多个 AI 领域推动了显著进展。然而，各类 Transformer 中普遍存在注意力陷阱（Attention Sink, AS）问题，即大量注意力集中在少数特定但缺乏信息的词元上。这不仅削弱了模型的可解释性，还显著影响训练与推理动态，加剧幻觉等缺陷。近年来，学界已开展大量研究以理解并利用 AS，但尚无系统性综述整合相关成果并指导未来发展。为此，本文首次全面综述 AS 研究，围绕三大核心维度构建框架：基础利用、机制解释与策略缓解。我们梳理了该领域的关键概念与主要趋势，助力研究者在当前 Transformer 框架下有效管理 AS，同时激发下一代 Transformer 的创新突破。论文列表详见 https://github.com/ZunhaiSu/Awesome-Attention-Sink。
原文摘要 · Abstract (English)
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 makes a pivotal contribution by highlighting the key concepts and main trends in the field, guiding researchers through the evolution of AS-related studies. We envision this survey as a valuable resource, empowering researchers 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.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。