Research map / Architectures and efficiency / Attention and transformers
Towards Understanding Massive Activations in Attention Sink Mechanism
Haiyu Wang, Yuanyuan Lin
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| zero out massive activations of the sink token (ablation) | pretrained LLM (LLaMA3.1-8B) · sink rate eps=0.3 | Sink_1 (%): 98.91 -> 9.19 | Hurt |
| additionally zero the sink token value vector (h = v = 0) after zeroing massive activations (ablation) | pretrained LLM (LLaMA3.1-8B) · sink rate eps=0.3 | Sink_1 (%): 9.19 -> 96.60 | Improved |
| additionally zero the sink value vector after zeroing massive activations (ablation) | LLaMA2-7B · WikiText | perplexity: 9774.18 -> 29.04 (original 5.37) | Improved |
| zero out top-k massive activations before RMSNorm (ablation) | LLaMA2-7B, LLaMA3-8B · sink rate | Sink_1 (%): decreases sharply to near zero | Hurt |
| zero out top-k largest entries after RMSNorm (ablation) | LLaMA2-7B, LLaMA3-8B · sink rate | Sink_1 (%): moderate drop | Hurt |
| rescale the sink hidden state in all subsequent RMSNorm layers (All layers) (ablation) | LLaMA2-7B, LLaMA3-8B · sink rate | Sink_1 (%) | No clear effect |
| rescale the sink hidden state at only one layer (One layer) (ablation) | LLaMA2-7B, LLaMA3-8B · sink rate | Sink_1 (%) | Hurt |
| remove massive activations from sink tokens (ablation) | Llama2-7B · token perturbation propagation | representational distance | Hurt |
| increase sink token hidden-state norm (ablation) | LLaMA2-7B, LLaMA3-8B · Jacobian norm of sink token | Jacobian norm | Improved |
Closest papers by meaning
- When Attention Sink Emerges in Language Models: An Empirical View (ICLR 2025 · similarity 0.89)
- Interpreting the Repeated Token Phenomenon in Large Language Models (ICML 2025 · similarity 0.89)
- Unveiling and Harnessing Hidden Attention Sinks: Enhancing Large Language Models without Training through Attention Calibration (ICML 2024 · similarity 0.88)
- Attention Sinks Induce Gradient Sinks: Massive Activations as Gradient Regulators in Transformers (NeurIPS 2026 · similarity 0.87)
- Sink vs. diagonal patterns as mechanisms for attention switch and oversmoothing prevention (NeurIPS 2026 · similarity 0.87)
- From Attention to Activation: Unraveling the Enigmas of Large Language Models (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
