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

• URL: http://arxiv.org/abs/2604.14722v1
• Date: Thu, 16 Apr 2026 07:32:37 GMT
• Title: A Mechanistic Account of Attention Sinks in GPT-2: One Circuit, Broader Implications for Mitigation
• Abstract summary: Transformers commonly exhibit an attention sink: disproportionately high attention to the first position.<n>We study this behavior in GPT-2-style models with learned query biases and absolute positional embeddings.
• Score: 1.2744523252873352
• License: http://creativecommons.org/licenses/by/4.0/
• 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.

Related papers

• The Structural Origin of Attention Sink: Variance Discrepancy, Super Neurons, and Dimension Disparity [22.040532283910522]
  This work provides a textitmechanistic explanation for the phenomenon.<n>We trace its root to the value aggregation process inherent in self-attention.<n>As a proof of concept, we propose textithead-wise RMSNorm, an architectural modification that stabilizes value aggregation outputs during pre-training.
  arXiv  Detail & Related papers  (2026-05-07T17:28:55Z)
• The Spike, the Sparse and the Sink: Anatomy of Massive Activations and Attention Sinks [32.60957674853853]
  We study two recurring phenomena in Transformer language models.<n>Massive activations, in which a small number of tokens exhibit extreme outliers in a few channels, and attention sinks, in which certain tokens attract disproportionate attention mass regardless of semantic relevance.
  arXiv  Detail & Related papers  (2026-03-05T18:59:04Z)
• How Attention Sinks Emerge in Large Language Models: An Interpretability Perspective [67.08306259095778]
  We identify a simple mechanism that enables the model to recognize token at position zero and induce an attention sink within two transformer blocks.<n>By analyzing training traces from a 30B A3B MoE model trained from scratch, we find that this mechanism emerges early in training and becomes increasingly concentrated in the first two layers.
  arXiv  Detail & Related papers  (2026-02-04T11:10:34Z)
• What One Cannot, Two Can: Two-Layer Transformers Provably Represent Induction Heads on Any-Order Markov Chains [64.31313691823088]
  In-context learning (ICL) is a capability of transformers, through which trained models learn to adapt to new tasks by leveraging information from the input context.<n>We show that a two-layer transformer with one head per layer can indeed represent any conditional k-gram.
  arXiv  Detail & Related papers  (2025-08-10T07:03:01Z)
• What are you sinking? A geometric approach on attention sink [6.552700667389349]
  Attention sink (AS) is a consistent pattern in transformer attention maps where certain tokens disproportionately attract attention from other tokens.<n>We show that in transformers, AS is not an architectural artifact, but it is the manifestation of a fundamental geometric principle.
  arXiv  Detail & Related papers  (2025-08-04T15:59:15Z)
• On the Emergence of Position Bias in Transformers [59.87743433861665]
  This paper presents a graph-theoretic framework for analyzing position biases in multilayer positions.<n>Our framework offers a principled foundation for understanding positional interplay in transformers.
  arXiv  Detail & Related papers  (2025-02-04T02:53:07Z)
• Enhancing Transformers for Generalizable First-Order Logical Entailment [51.04944136538266]
  This paper studies the generalizable first-order logical reasoning ability of transformers with their parameterized knowledge.<n>We propose TEGA, a logic-aware architecture that significantly improves the performance in first-order logical entailment.
  arXiv  Detail & Related papers  (2025-01-01T07:05:32Z)
• Unveiling Induction Heads: Provable Training Dynamics and Feature Learning in Transformers [54.20763128054692]
  We study how a two-attention-layer transformer is trained to perform ICL on $n$-gram Markov chain data. We prove that the gradient flow with respect to a cross-entropy ICL loss converges to a limiting model.
  arXiv  Detail & Related papers  (2024-09-09T18:10:26Z)
• Transformers with Competitive Ensembles of Independent Mechanisms [97.93090139318294]
  We propose a new Transformer layer which divides the hidden representation and parameters into multiple mechanisms, which only exchange information through attention. We study TIM on a large-scale BERT model, on the Image Transformer, and on speech enhancement and find evidence for semantically meaningful specialization as well as improved performance.
  arXiv  Detail & Related papers  (2021-02-27T21:48:46Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.