Research map / Architectures and efficiency / Attention and transformers
The Structural Origin of Attention Sink: Variance Discrepancy, Super Neurons, and Dimension Disparity
Siquan Li, Kaiqi Jiang, Jiacheng Sun, Tianyang Hu
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 |
|---|---|---|---|
| add head-wise RMSNorm after value aggregation | Llama-2 152M softmax attention · OpenWebText pretraining | Validation loss: 2.7812 ± 0.0109 -> 2.7421 ± 0.0066 | Improved |
| add head-wise RMSNorm after value aggregation | Llama-2 152M softmax attention · OpenWebText pretraining | Train loss: 2.7483 ± 0.0118 -> 2.7073 ± 0.0095 | Improved |
| add head-wise RMSNorm after value aggregation | Llama-2 152M softmax attention · OpenWebText pretraining | Effective rank: 343.71 ± 15.63 -> 445.96 ± 5.37 | Improved |
| add head-wise RMSNorm after value aggregation | Llama-2 152M softmax attention · OpenWebText pretraining | Dimension disparity: 82.67 ± 8.09 -> 33.74 ± 2.73 | Improved |
| replace softmax with unnormalized sigmoid attention (ablation) | Llama-2 152M · OpenWebText pretraining | Validation loss / convergence | Hurt |
| replace softmax with sigmoid attention (ablation) | Llama-2 152M · OpenWebText pretraining | Attention sink | Improved |
| block token 10 from attending to earlier tokens (mask) (ablation) | Llama-2-7B · WikiText-2 | Attention to token | Improved |
| amplify variance of token 10 by lambda (ablation) | Llama-2-7B · random tokens | Attention received | Improved |
| scale norm by lambda instead of variance (control) (ablation) | Llama-2-7B · random tokens | Attention received | No clear effect |
Closest papers by meaning
- When Attention Sink Emerges in Language Models: An Empirical View (ICLR 2025 · similarity 0.91)
- 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 0.89)
- The emergence of sparse attention: impact of data distribution and benefits of repetition (NeurIPS 2025 · similarity 0.88)
- Sink vs. diagonal patterns as mechanisms for attention switch and oversmoothing prevention (NeurIPS 2026 · similarity 0.88)
- On the Nature of Attention Sink that Shapes Decoding Strategy in Omni-LLMs (NeurIPS 2026 · similarity 0.88)
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.
