Massive Activations in LLMs Function as Gradient Regulators, Study Reveals
Researchers from Tsinghua and Peking University discover that massive activations in transformer models act as adaptive gradient regulators, with attention sinks creating gradient sinks that RMSNorm suppresses, revealing a backward-propagation mechanism linking attention concentration to activation scaling.
Background: Attention Sinks and Massive Activations
In large language models, a recurring phenomenon is that a few tokens at the sequence start attract disproportionate attention, known as attention sinks . Simultaneously, certain internal activations grow extremely large, termed massive activations . These two phenomena often co-occur, but their connection remains unclear.
Shift to Backward Propagation
Prior work mainly examines forward computation. This NeurIPS 2026 paper from Tsinghua University and Peking University investigates the other half: backward propagation . The authors find that attention sinks also concentrate training signals onto a few tokens, creating gradient sinks . Crucially, massive activations can attenuate these locally concentrated gradients via RMSNorm .
01 Normalized Activations: Why Are They Trained So Large?
Modern transformers use pre-norm architectures where each token's hidden state passes through RMSNorm before entering attention or MLP sublayers. RMSNorm normalizes the vector direction, so scaling the vector magnitude leaves the normalized input nearly unchanged. This raises a question: if forward computation normalizes away scale, why are massive activation norms trained in the first place?
The answer may lie in backward propagation. The scale eliminated in the forward pass still influences gradient flow during backpropagation.
02 Attention as Both Forward Value Router and Backward Gradient Router
Forward attention computes a weighted sum of Value vectors. During backpropagation, the same attention weights determine how gradients are distributed back to those positions. If many subsequent tokens attend to the same position, their gradients also flow back along those connections.
In causal attention, early tokens can be accessed by many later positions. Once a token becomes an attention sink, it structurally receives large amounts of backward training signals. The exact concentration strength depends on gradient directions and correlations.
The team measured models trained from scratch and several open-source models. At the attention input after RMSNorm, gradient concentration is pervasive. The first token's gradient norm exceeds other early positions by one to two orders of magnitude. This local gradient concentration is termed gradient sinks : attention sinks are also gradient sinks .
Further decomposing Query, Key, and Value paths reveals sharper differences. Key and Value paths show prominent peaks at the first token, with Value path being most significant; Query path is relatively flat. Local gradient concentration exhibits a clear path structure.
03 Massive Activations as Gradient "Pressure Relief Valves"
How do gradient sinks relate to massive activations? Observing gradient scale changes across the attention sublayer shows that while the first token's gradient is sharply amplified after the attention sublayer, its residual stream gradient scale remains relatively stable. This is because the gradients concentrated by the attention layer are compressed by RMSNorm.
The authors theoretically analyze RMSNorm's gradient propagation properties. The core result: under otherwise equal conditions, larger input activation scale gives RMSNorm greater contraction capability on backward gradients . Specifically, the operator norm upper bound of RMSNorm's Jacobian matrix is inversely proportional to the input's root-mean-square scale. This makes activation scale a local gradient regulation mechanism.
Thus the paper proposes a mechanism chain: attention concentrates on few positions, causing gradient sinks on the Value path, while larger residual stream activations attenuate the branch gradients flowing back to the residual stream via RMSNorm . From this perspective, massive activations can be understood as an adaptive regulatory structure formed during model training .
04 Experimental Verification: Providing an Alternative Gradient Regulation Path
If the explanation holds, a testable prediction emerges: if the model gains another path to attenuate local gradients, its reliance on massive activations should weaken even if attention sinks persist.
To test this, the team introduces V-scale : after Value projection but before attention aggregation, scale the Value vectors by their norm, applying stronger decay to backpropagated gradients corresponding to small-norm Values. The Value path is chosen because it carries the most prominent gradient peaks and attention sink positions tend to have small Value norms.
Results confirm the prediction. In a 0.3B model, the median gradient sink ratio drops from ~40.3 to 5.6. Meanwhile, the first token's residual stream activation norm decreases significantly—massive activations are substantially mitigated while the model retains clear attention sink behavior. The mitigation mainly stems from reduced MLP output norms.
Since full V-scale also alters forward Value propagation, a control modifying only the backward rule was designed. The backward-only intervention yields the same decoupling: first token residual stream and MLP output norms shrink, attention sinks remain prominent. This provides experimental support for massive activations participating in regulating local gradient pressure .
05 Understanding Model Anomalies Requires Examining Training Signal Flow
This work connects two frequently discussed phenomena—attention sinks and massive activations—and discovers their link in backward propagation: gradient sinks . Attention weights govern both forward information aggregation and backward gradient routing. Normalization layers affect both forward representation scale and backward signal transmission. Considering both together enables a more complete understanding of why extreme activations form during training.
The research offers a new perspective: when we observe "anomalous" model behaviors, we can further ask what role they play during the training process.
Paper Information
Title: Attention Sinks Induce Gradient Sinks: Massive Activations as Gradient Regulators in Transformers
Authors: Yihong Chen, Zhouchen Lin, Quanming Yao
ArXiv: https://arxiv.org/abs/2603.17771
Code: https://github.com/LARS-research/GradientSink
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
