AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Householder-Diagonalized Linear Attention (HDLA): Utilizing Enhanced Decay Mechanism for Efficient Sequence Modeling

Jiefu Zhang, Zhen Qin, Jiabo Tong, Shijie Mei, JiaKui Hu, Yuqi Pan, Anjie Hu, Man Yao, Bo XU, Guoqi Li

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: Linear attention approximations

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
HDLA decay vs Gated DeltaNetlinear attention · MAD (2048)average score: 68.17 -> 72.97Improved
Expand hidden state from 2048 to 4096 (ablation)HDLA · MADaverage score: 72.97 -> 74.17Improved
Use HDLA instead of Gated DeltaNet at 2.8Blinear attention LM · FineWeb-Edu avg perplexity (2.8B)perplexity: 19.60 -> 18.58Improved
Use HDLA instead of Llama at 2.8Bsoftmax baseline · FineWeb-Edu avg perplexity (2.8B)perplexity: 20.71 -> 18.58Improved
Use HDLA instead of Gated DeltaNet on S-NIAH-31.45B model · S-NIAH-3 length 2048accuracy: 7.0% -> 65.2%Improved
Use HDLA instead of Gated DeltaNet on S-NIAH-31.45B model · S-NIAH-3 length 1024accuracy: 50.6% -> 82.0%Improved
Expand hidden state (nh 12 to 4) (ablation)HDLA 0.4B · AVG-RETaccuracy: 10.54 -> 13.91Improved
Expand hidden state (nh 12 to 3) (ablation)HDLA 0.17B · AVG-RETaccuracy: 7.20 -> 9.62Improved
Restrict beta range from [0,2] to [0,1] (ablation)HDLA 0.4B · AVG-RETaccuracy: 11.71 -> 12.51No clear effect
Raise learning rate from 3e-4 to 6e-4 (ablation)HDLA 0.4B · AVG-RETaccuracy: 11.7 -> 13.3Improved
Use SiLU vs 1+elu on keys/values (ablation)HDLA 0.4B · AVG-RETaccuracy: 11.71 -> 12.93No clear effect
Use HDLA instead of Gated DeltaNetlinear attention · Training throughputtokens per second: 19277.18 -> 7340.01Hurt

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