AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Attention Sinks and Compression Valleys in LLMs are Two Sides of the Same Coin

Enrique Queipo-de-Llano, Alvaro Arroyo, Federico Barbero, Xiaowen Dong, Michael M. Bronstein, Yann LeCun, Ravid Shwartz-Ziv

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: Long-context KV and memory attention · 35 citations (OpenAlex)

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 changedWhereMeasuredEffect
zero the MLP contribution to the BOS token at layer 0 (removes massive activation) (ablation)LLaMA3 8B · GSM8Kmatrix-based entropy (bits): 0.02 -> 0.5Improved
zero the MLP contribution to the BOS token at layer 0 (removes massive activation) (ablation)LLaMA3 8B · GSM8Ksink rate: 0.85 -> 0.0Hurt
zero the MLP contribution to the BOS token at layer 0 (removes massive activation) (ablation)LLaMA3 8B · GSM8KBOS norm ratio to other tokens: 10^2x -> within 2xHurt
ablate only a single massive-activation stage (multi-stage models) (ablation)Pythia 410M, Qwen2 7B · GSM8Kcompression: partially reduces compressionMixed
clip magnitude of the massive activation to lower norm (ablation)Pythia 410M · GSM8Krepresentational entropy and sink rate: entropy rises, sinks decrease with general downward trendMixed
go from early layers to full depth via LogitLenspretrained LLMs · WikiText-2perplexity: >10^4 -> 10-25Improved
read out mid-depth layers (25-75% depth) instead of early/late layers for embeddingsdecoder-only LLMs · MTEB (32 tasks)performance: 10-20%Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy