AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Anatomy of Massive Activations and Attention Sinks

Shangwen Sun, Alfredo Canziani, Yann LeCun, Jiachen Zhu

ICML 2026 · Architectures and efficiency › Attention and transformers · approach: Sparse and structured attention

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
add per-channel conditional gating (ablation)Llama-style 7B · C4 sentences, 7B from scratchsink ratio: 46.0% -> 4.5%Improved
add static unconditional channel gating (ablation)Llama-style 7B · C4 sentences, 7B from scratchsink ratio: 46.0% -> 42.2%No clear effect
add sandwich normalization (ablation)Llama-style 7B · C4 sentences, 7B from scratchspike magnitude: 3,818 -> 520Improved
add QK-Norm to sandwich normalization (ablation)Llama-style 7B · C4 sentences, 7B from scratchspike magnitude: 3,818 -> 92Improved
replace pre-norm with DynamicTanh (ablation)Llama-style 7B · C4 sentences, 7B from scratchsink ratio: 46.0% -> 61.0%Improved
replace SwiGLU FFN with linear or attention-only blocks (ablation)Llama-style 7B · C4 sentences, 7B from scratchspike magnitude: 3,818 -> 688Improved
increase head dimension from 8 to 128 (ablation)Llama-style 7B · C4 sentences, 7B from scratchsink ratio: 4.1% -> 46.0%Improved
train only on long-context positions 2048-4096 (ablation)Llama-style 7B · C4 sentences, 7B from scratchsink ratio: 46.0% -> 1.2%Hurt
train with maximum context 256 instead of 4096 (ablation)Llama-style 7B · C4 sentences, 7B from scratchsink ratio: 46.0% -> 42.1%No clear effect
disable weight decay (ablation)Llama-style 7B · C4 sentences, 7B from scratchspike magnitude: 3,818 -> 12,275Hurt
extend training from 100B to 200B tokens (ablation)Llama-style 7B · C4 sentences, 7B from scratchsink ratio: 46.0% -> 63.3%Improved
raise AdamW beta2 from 0.95 to 0.999 (ablation)Llama-style 7B · C4 sentences, 7B from scratchsink ratio: 46.0% -> 20.9%Hurt

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