Why do LLMs attend to the first token?
This provides a new practical perspective on attention sinks in LLMs, potentially improving understanding of training patterns, though it appears incremental as it builds on existing studies of the phenomenon.
The paper investigates why Large Language Models (LLMs) heavily attend to the first token (attention sink), arguing theoretically and empirically that this mechanism helps LLMs avoid over-mixing of information, with experiments showing how factors like context length and depth influence this behavior.
Large Language Models (LLMs) tend to attend heavily to the first token in the sequence -- creating a so-called attention sink. Many works have studied this phenomenon in detail, proposing various ways to either leverage or alleviate it. Attention sinks have been connected to quantisation difficulties, security issues, and streaming attention. Yet, while many works have provided conditions in which they occur or not, a critical question remains shallowly answered: Why do LLMs learn such patterns and how are they being used? In this work, we argue theoretically and empirically that this mechanism provides a method for LLMs to avoid over-mixing, connecting this to existing lines of work that study mathematically how information propagates in Transformers. We conduct experiments to validate our theoretical intuitions and show how choices such as context length, depth, and data packing influence the sink behaviour. We hope that this study provides a new practical perspective on why attention sinks are useful in LLMs, leading to a better understanding of the attention patterns that form during training.