What are you sinking? A geometric approach on attention sink

TL;DR

This paper reveals that attention sink patterns in transformers are driven by fundamental geometric principles related to reference frames, influencing architecture design and understanding of attention mechanisms.

Contribution

It introduces a geometric perspective showing attention sink as a manifestation of reference frame establishment in transformer representations.

Findings

Attention sink correlates with three reference frame types.

Reference frames emerge early in training as optimal solutions.

Position encoding influences the type of reference frame.

Abstract

Attention sink (AS) is a consistent pattern in transformer attention maps where certain tokens (often special tokens or positional anchors) disproportionately attract attention from other tokens. We show that in transformers, AS is not an architectural artifact, but it is the manifestation of a fundamental geometric principle: the establishment of reference frames that anchor representational spaces. We analyze several architectures and identify three distinct reference frame types, centralized, distributed, and bidirectional, that correlate with the attention sink phenomenon. We show that they emerge during the earliest stages of training as optimal solutions to the problem of establishing stable coordinate systems in high-dimensional spaces. We show the influence of architecture components, particularly position encoding implementations, on the specific type of reference frame. This perspective transforms our understanding of transformer attention mechanisms and provides insights for both architecture design and the relationship with AS.