Saved in:
Bibliographic Details
Main Authors: Ran-Milo, Yuval, Ofek, Hila, Mendel, Shahar
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2604.14722
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908969873702912
author Ran-Milo, Yuval
Ofek, Hila
Mendel, Shahar
author_facet Ran-Milo, Yuval
Ofek, Hila
Mendel, Shahar
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2604_14722
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle A Mechanistic Account of Attention Sinks in GPT-2: One Circuit, Broader Implications for Mitigation
Ran-Milo, Yuval
Ofek, Hila
Mendel, Shahar
Machine Learning
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.
title A Mechanistic Account of Attention Sinks in GPT-2: One Circuit, Broader Implications for Mitigation
topic Machine Learning
url https://arxiv.org/abs/2604.14722