The Structural Origin of Attention Sink: Variance Discrepancy, Super Neurons, and Dimension Disparity

• URL: http://arxiv.org/abs/2605.06611v1
• Date: Thu, 07 May 2026 17:28:55 GMT
• Title: The Structural Origin of Attention Sink: Variance Discrepancy, Super Neurons, and Dimension Disparity
• Abstract summary: 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.
• Score: 22.040532283910522
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Despite the prevalence of the attention sink phenomenon in Large Language Models (LLMs), where initial tokens disproportionately monopolize attention scores, its structural origins remain elusive. This work provides a \textit{mechanistic explanation} for this phenomenon. First, we trace its root to the value aggregation process inherent in self-attention, which induces a systematic variance discrepancy. We further demonstrate that this discrepancy is drastically amplified by the activation of super neurons within Feed-Forward Network (FFN) layers. Specifically, the channel-sparse down-projections trigger a dimension disparity of the first-token representation, necessitating the formation of attention sinks as a structural anchor. Then, we validate this causal chain through two controlled interventions: (i) isolating the aggregation effect via attention mask modifications and (ii) amplifying the variance of targeted token representations. Both interventions can replicate attention sinks at arbitrary positions. Our mechanistic understanding offers a foundation for the systematic control of sink formation. Finally, as a proof of concept, we propose \textit{head-wise RMSNorm}, an architectural modification that stabilizes value aggregation outputs during pre-training. Our experiments demonstrate that restoring statistical parity across positions significantly accelerates convergence.

Related papers

• Mechanistic Interpretability of Chain-of-Thought Reasoning via Sequential Activation Patching [0.9940728137241215]
  We investigate where Chain-of-Thought (CoT)-related causal effects emerge across the generated reasoning trajectory.<n>We introduce a sequential activation patching framework that traces CoT-conditioned attention-head activations across token positions.<n>Our results provide evidence for distributed reasoning-support sub-circuits associated with CoT-conditioned exemplars.
  arXiv  Detail & Related papers  (2026-08-23T10:05:16Z)
• Disentangling Semantic Attention from Structural Bias in the Attention Manifold [61.68044165974505]
  Multimodal Large Language Models (MLLMs) exhibit disproportionate attention toward certain semantically uninformative visual tokens.<n>We introduce Saliency-guided Purification and Adaptive Redistribution (SPAR), a training-free, plug-and-play intervention.
  arXiv  Detail & Related papers  (2026-07-27T05:30:19Z)
• Cognitive Pivot Points and Visual Anchoring: Unveiling and Rectifying Hallucinations in Multimodal Reasoning Models [65.57985131861399]
  hallucinations are strongly correlated with cognitive bifurcation points that often exhibit high entropy states.<n>We propose V-STAR, a lightweight, holistic training paradigm designed to internalize visually aware reasoning capabilities.
  arXiv  Detail & Related papers  (2026-04-11T13:59:05Z)
• OrthoFormer: Instrumental Variable Estimation in Transformer Hidden States via Neural Control Functions [0.0]
  Transformer architectures excel at sequential modeling yet remain fundamentally limited by correlational learning.<n>We propose OrthoFormer, a causally grounded architecture that embeds instrumental variable estimation directly into Transformer blocks via neural control functions.
  arXiv  Detail & Related papers  (2026-03-08T03:05:16Z)
• 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)
• A Unified View of Attention and Residual Sinks: Outlier-Driven Rescaling is Essential for Transformer Training [86.64715217940274]
  Outliers function jointly with normalization.<n>Outliers serve more as rescale factors rather than contributors.<n>Outliers can be absorbed into learnable parameters or mitigated via explicit gated rescaling.
  arXiv  Detail & Related papers  (2026-01-30T13:29:45Z)
• On Evolution-Based Models for Experimentation Under Interference [7.262048441360133]
  We study an evolution-based approach that investigates how outcomes change across observation rounds in response to interventions.<n>We highlight causal message passing as an instantiation of this method in dense networks.<n>We discuss the limits of this approach, showing that strong temporal trends or endogenous interference can undermine identification.
  arXiv  Detail & Related papers  (2025-11-26T18:53:46Z)
• Noise & pattern: identity-anchored Tikhonov regularization for robust structural anomaly detection [58.535473924035365]
  Anomaly detection plays a pivotal role in automated industrial inspection, aiming to identify subtle or rare defects in otherwise uniform visual patterns.<n>We tackle structural anomaly detection using a self-supervised autoencoder that learns to repair corrupted inputs.<n>We introduce a corruption model that injects artificial disruptions into training images to mimic structural defects.
  arXiv  Detail & Related papers  (2025-11-10T15:48:50Z)
• Feature Dynamics as Implicit Data Augmentation: A Depth-Decomposed View on Deep Neural Network Generalization [18.72807692009739]
  We show that temporal consistency extends to unseen and corrupted data, but collapses when semantic structure is destroyed.<n>Together, these findings suggest a conceptual perspective that links feature dynamics to generalization.
  arXiv  Detail & Related papers  (2025-09-24T17:23:56Z)
• "Principal Components" Enable A New Language of Images [79.45806370905775]
  We introduce a novel visual tokenization framework that embeds a provable PCA-like structure into the latent token space.<n>Our approach achieves state-of-the-art reconstruction performance and enables better interpretability to align with the human vision system.
  arXiv  Detail & Related papers  (2025-03-11T17:59:41Z)
• 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)
• Active-Dormant Attention Heads: Mechanistically Demystifying Extreme-Token Phenomena in LLMs [77.66717051042032]
  Practitioners have consistently observed three puzzling phenomena in transformer-based large language models. These phenomena are characterized by certain so-called "sink tokens" receiving disproportionately high attention weights. We elucidate the mechanisms behind extreme-token phenomena.
  arXiv  Detail & Related papers  (2024-10-17T17:54:06Z)

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.