Work

A Mechanistic Account of Attention Sinks in GPT-2: One Circuit, Broader Implications for Mitigation

Year: 2026

Type: conference-paper

Abstract: Transformers commonly exhibit an attention sink: disproportionately high attention to the first position.We study this behavior in GPT-2-style models with learned query biases and absolute positional embeddings.Combining structural analysis with causal interventions, validated across natural-language, mathematical, and code inputs, we find that the sink arises from the interaction among (i) a learned query bias, (ii) the first-layer MLP transformation of the positional encoding, and (iii) structure in the key projection.Crucially, each component we identify is individually dispensable: architectures omitting each of them robustly exhibit sinks.This indicates that attention sinks may arise through distinct circuits across architectures.These findings inform mitigation of sinks, and motivate broader investigation into why sinks emerge.* Equal contribution. 1 However, some architectures have been reported to have little to no sink (Qiu et al., 2025;Endy et al., 2025).

Source: Proceedings of the conference - Association for Computational Linguistics. Meeting

DOI: 10.18653/v1/2026.acl-short.9

Authors: Yuval Ran-Milo, Hila Ofek, Shahar Mendel

Institution: Tel Aviv University

Language: English

Topic: Topic Modeling

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: hybrid

Funders: Israel Science Foundation, Tel Aviv University

Award: G4730508171