AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Head-in-Head in Linear Attention

Shijie Mei, Man Yao, Jiabo Tong, Bo XU, Guoqi Li

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 Head-in-Head intra-head mask grouping (r=4, data-dependent) to Gated-DeltaNetGated-DeltaNet 1.3B · Wikitext / LAMBADAperplexity: 16.30 -> 16.10 (Wiki); 14.52 -> 13.05 (Lamb.)Improved
add Head-in-Head to Gated-DeltaNetGated-DeltaNet 340M · Commonsense reasoning avgaccuracy (%): 41.79 -> 42.43Improved
add Head-in-Head to Gated-DeltaNetGated-DeltaNet 340M · Recall-intensive tasksaverage score: 24.88 -> 27.28Improved
add Head-in-Head to DeltaNet / Gated-DeltaNetDeltaNet, Gated-DeltaNet · MAD benchmarkaverage score: 74.4 -> 76.5 (DN); 74.9 -> 77.2 (GDN)Improved
add Head-in-Head to Gated-DeltaNetGated-DeltaNet 1.3B · LongBenchaverage score: 17.06 -> 17.57Improved
increase intra-head groups from r=1 to r=2 and r=4 (static mask) (ablation)Gated-DeltaNet 1.3B · CSR-avgaccuracy (%): 51.76 -> 52.17 -> 52.24Improved
increase intra-head groups from r=2 to r=4 (ablation)Gated-DeltaNet 1.3B · Recall-avgrecall average: 36.48 -> 37.62Improved
switch from static learnable mask to data-dependent (time-varying) mask at r=4 (ablation)Head-in-Head GDN 1.3B · Wiki/Lamb perplexity vs Recall-avgperplexity and recall: Lamb 13.61 -> 13.05; Recall-avg 37.62 -> 37.07Mixed
add Head-in-Head grouping (more intra-head groups) to GDN (ablation)Gated-DeltaNet · PG-19 memory-state similarityinformation entropy: 0.756 -> 0.813 (rank2) -> 0.823 (rank4)Improved
add Head-in-Head to Gated-DeltaNet: effect on length extrapolationGated-DeltaNet 1.3B · PG-19 up to 16kperplexityImproved
add time-varying mask vs fixed mask: inference latency (ablation)HH-GDN 1.3B · decoding 1000 tokenslatency relative to baseline: 1.2x (time-varying), 1.05x (fixed)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