AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Norm$\times$Direction: Restoring the Missing Query Norm in Vision Linear Attention

Weikang Meng, Yadan Luo, Liangyu Huo, Yingjian Li, Yaowei Wang, Xin Li, Zheng Zhang

ICML 2026 · Architectures and efficiency › Attention and transformers · approach: Sparse and structured attention · 0 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
add cosine direction similarity to ReLU + constant power linear attention (ablation)FLatten-Swin-T · ImageNet-1KTop-1 accuracy (%): 82.1 -> 82.5Improved
upgrade constant power to query-norm-aware power function (ablation)FLatten-Swin-T with cosine direction · ImageNet-1KTop-1 accuracy (%): 82.5 -> 82.9Improved
replace attention with NaLaFormer in Swin-TSwin-T · ImageNet-1KTop-1 accuracy (%): 81.2 -> 82.9Improved
QK-Norm instead of FLatten baseline (ablation)FLatten Swin / PVT · ImageNet-1KTop-1 accuracy (%): -0.04% on Swin, +0.15% on PVTMixed
NaLaFormer instead of FLatten baseline (ablation)FLatten Swin / PVT · ImageNet-1KTop-1 accuracy (%): +0.21% on Swin, +0.50% on PVTImproved
apply NaLa feature map to ESRT for super-resolutionESRT · Set5/Set14/BSD100/Urban100 x4PSNR: Urban100 x4: 25.85 -> 25.83No clear effect
apply NaLa to DCTLSA for super-resolutionDCTLSA · Urban100 x4PSNR: 26.41 -> 26.74Improved
replace SiLU with query-norm-aware feature map in DeltaNetDeltaNet 340M · common-sense reasoning averageaverage accuracy: 43.6 -> 44.3Improved
replace SiLU with query-norm-aware feature map in Gated DeltaNetGated DeltaNet 340M · common-sense reasoning averageaverage accuracy: 44.3 -> 44.8Improved
replace attention with NaLaFormer in diffusion transformerSiT S/2 · ImageNet-1KFID: 58.61 -> 53.08Improved

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