Attention Sinks Induce Gradient Sinks

• URL: http://arxiv.org/abs/2603.17771v1
• Date: Wed, 18 Mar 2026 14:31:21 GMT
• Title: Attention Sinks Induce Gradient Sinks
• Abstract summary: We show that under causal mask, attention sinks can induce pronounced gradient concentration.<n>Massive activations can be understood as an adaptive response to this localized gradient pressure.<n>These results support the interpretation that gradient sink is a key training-time mediator linking attention sinks and massive activations.
• Score: 41.85762183893736
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Attention sinks and massive activations are recurring and closely related phenomena in Transformer models. Existing studies have largely focused on the forward pass, making it unclear whether their connection is direct or mediated by a training-time mechanism. We study this question from the perspective of backpropagation. Empirically and theoretically, we show that under causal mask, attention sinks can induce pronounced gradient concentration, which we term gradient sinks. Furthermore, in pre-norm architectures with RMSNorm, massive activations can be understood as an adaptive response to this localized gradient pressure during training. To test this hypothesis, we introduce V-scale, a modification that adjusts value-path backpropagated gradients. In pretrained V-scale models, attention sinks are preserved whereas massive activations are suppressed. These results support the interpretation that gradient sink is a key training-time mediator linking attention sinks and massive activations.

Related papers

• Gradient Perturbation: Learning to Perturb Gradients for Adaptive Training [3.2608238154676843]
  gradient perturbation has received little systematic investigation.<n>We propose Learning to Perturb Gradients (LPG), which adaptively perturbs logit-level gradients at the class level to achieve category-aware training.<n> Experiments on balanced classification, long-tail classification, and noisy label learning demonstrate LPG consistently outperforms existing methods and can be combined with them as a plug-in module.
  arXiv  Detail & Related papers  (2026-05-28T07:19:36Z)
• 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)
• 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)
• Spectral Gradient Descent Mitigates Anisotropy-Driven Misalignment: A Case Study in Phase Retrieval [13.218607858857295]
  Spectral gradient methods modify gradient updates by preserving directional information while discarding scale.<n>We investigate the mechanisms underlying these gains through a dynamical analysis of a nonlinear phase retrieval model.
  arXiv  Detail & Related papers  (2026-01-30T07:12:58Z)
• Attention Sinks and Compression Valleys in LLMs are Two Sides of the Same Coin [41.97400057649291]
  We present a surprising connection between attention sinks and compression valleys, tracing both to the formation of massive activations in the residual stream.<n>We propose the Mix-Compress-Refine theory of information flow, as an attempt to explain how LLMs organize their computation in depth by controlling attention and representational compression via massive activations.
  arXiv  Detail & Related papers  (2025-10-07T21:27:24Z)
• Generative Model Inversion Through the Lens of the Manifold Hypothesis [98.37040155914595]
  Model inversion attacks (MIAs) aim to reconstruct class-representative samples from trained models.<n>Recent generative MIAs utilize generative adversarial networks to learn image priors that guide the inversion process.
  arXiv  Detail & Related papers  (2025-09-24T14:39:25Z)
• Towards Training Without Depth Limits: Batch Normalization Without Gradient Explosion [83.90492831583997]
  We show that a batch-normalized network can keep the optimal signal propagation properties, but avoid exploding gradients in depth. We use a Multi-Layer Perceptron (MLP) with linear activations and batch-normalization that provably has bounded depth. We also design an activation shaping scheme that empirically achieves the same properties for certain non-linear activations.
  arXiv  Detail & Related papers  (2023-10-03T12:35:02Z)
• Extreme Memorization via Scale of Initialization [72.78162454173803]
  We construct an experimental setup in which changing the scale of initialization strongly impacts the implicit regularization induced by SGD. We find that the extent and manner in which generalization ability is affected depends on the activation and loss function used. In the case of the homogeneous ReLU activation, we show that this behavior can be attributed to the loss function.
  arXiv  Detail & Related papers  (2020-08-31T04:53:11Z)
• The Break-Even Point on Optimization Trajectories of Deep Neural Networks [64.7563588124004]
  We argue for the existence of the "break-even" point on this trajectory. We show that using a large learning rate in the initial phase of training reduces the variance of the gradient. We also show that using a low learning rate results in bad conditioning of the loss surface even for a neural network with batch normalization layers.
  arXiv  Detail & Related papers  (2020-02-21T22:55:51Z)

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.