What are you sinking? A geometric approach on attention sink

Valeria Ruscio, Umberto Nanni, Fabrizio Silvestri

Keywords: Attention sink, Transformer geometry, Position encoding, Self-attention

TL;DR: Attention sink in LLMs serves as geometric reference frames that anchor token representations in high-dimensional space, emerging during training as optimal solutions to the coordinate system problem, shaped by architecture and position encodings.

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.

Primary Area: Deep learning (e.g., architectures, generative models, optimization for deep networks, foundation models, LLMs)

Submission Number: 9301