Zheng Zifan, Wang Yezhaohui, Huang Yuxin, Song Shichao, Yang Mingchuan, Tang Bo, Xiong Feiyu, Li Zhiyu
Institute for Advanced Algorithms Research (IAAR), Shanghai, China.
Institute for AI Industry Research (AIR), Tsinghua University, Beijing, China.
Patterns (N Y). 2025 Feb 6;6(2):101176. doi: 10.1016/j.patter.2025.101176. eCollection 2025 Feb 14.
Large language models (LLMs) have demonstrated performance approaching human levels in tasks such as long-text comprehension and mathematical reasoning, but they remain black-box systems. Understanding the reasoning bottlenecks of LLMs remains a critical challenge, as these limitations are deeply tied to their internal architecture. Attention heads play a pivotal role in reasoning and are thought to share similarities with human brain functions. In this review, we explore the roles and mechanisms of attention heads to help demystify the internal reasoning processes of LLMs. We first introduce a four-stage framework inspired by the human thought process. Using this framework, we review existing research to identify and categorize the functions of specific attention heads. Additionally, we analyze the experimental methodologies used to discover these special heads and further summarize relevant evaluation methods and benchmarks. Finally, we discuss the limitations of current research and propose several potential future directions.
大语言模型（LLMs）在长文本理解和数学推理等任务中展现出接近人类水平的性能，但它们仍然是黑箱系统。理解大语言模型的推理瓶颈仍然是一项关键挑战，因为这些限制与它们的内部架构紧密相关。注意力头在推理中起着关键作用，并且被认为与人类大脑功能有相似之处。在这篇综述中，我们探讨注意力头的作用和机制，以帮助揭开大语言模型内部推理过程的神秘面纱。我们首先引入一个受人类思维过程启发的四阶段框架。利用这个框架，我们回顾现有研究，以识别和分类特定注意力头的功能。此外，我们分析用于发现这些特殊头的实验方法，并进一步总结相关评估方法和基准。最后，我们讨论当前研究的局限性，并提出几个潜在的未来研究方向。
