Attention Head
Attention heads, the fundamental building blocks of transformer-based models, are crucial for processing information in sequence data. Current research focuses on understanding their functional specialization during training, optimizing their efficiency for large language models (LLMs) through techniques like sparse attention and head clustering, and leveraging their internal representations for improved model interpretability and performance in various tasks. This work is significant because it addresses both the computational challenges of deploying LLMs and the need for better understanding and control over their internal mechanisms, ultimately leading to more efficient and effective AI systems.
Papers
Active-Dormant Attention Heads: Mechanistically Demystifying Extreme-Token Phenomena in LLMs
Tianyu Guo, Druv Pai, Yu Bai, Jiantao Jiao, Michael I. Jordan, Song Mei
On the Role of Attention Heads in Large Language Model Safety
Zhenhong Zhou, Haiyang Yu, Xinghua Zhang, Rongwu Xu, Fei Huang, Kun Wang, Yang Liu, Junfeng Fang, Yongbin Li
Differentiation and Specialization of Attention Heads via the Refined Local Learning Coefficient
George Wang, Jesse Hoogland, Stan van Wingerden, Zach Furman, Daniel Murfet
Listening to the Wise Few: Select-and-Copy Attention Heads for Multiple-Choice QA
Eduard Tulchinskii, Laida Kushnareva, Kristian Kuznetsov, Anastasia Voznyuk, Andrei Andriiainen, Irina Piontkovskaya, Evgeny Burnaev, Serguei Barannikov