AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

PLASH: Provably Linear-Time Attention with Selective Higher-Order Feature Sketching

Yuwen Huang, Xiang Pan

ICML 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
use PLASH M=16 instead of Standard FlashAttention-2Qwen3-4B · PG-19 N=16,384EMA training loss: 25.2 -> 23.5Improved
use PLASH M=32 instead of Standard at 32KQwen3-4B · PG-19 N=32,768EMA training loss: 40.1 -> 40.1No clear effect
increase prototypes M from 8 to 16 (ablation)PLASH · PG-19 N=16,384EMA training loss: 30.1 -> 23.5Improved
increase prototypes M from 16 to 32 (ablation)PLASH · PG-19 N=16,384EMA training loss: 23.5 -> 24.5Hurt
increase prototypes M from 8 to 32 (ablation)PLASH · PG-19 N=32,768EMA training loss: 41.2 -> 40.1Improved
increase prototypes M from 8 to 32 (ablation)PLASH · Forward time, N=32Ktime (ms): 1809 -> 2536Hurt
use PLASH M=8 instead of Standard FlashAttention-2Qwen3-4B · Forward time, N=32Ktime (ms): 4358 -> 1809Improved
use PLASH M=64 instead of Vanilla attentionInformer · Run time, length 11264latency (ms): 58.95 -> 2.69Improved
increase M from 64 to 512 (ablation)PLASH · Run time, length 11264latency (ms): 2.69 -> 4.89Hurt
increase sketch size Dk from 128 to 512 at M=512 (ablation)PLASH · ETTh2 (L=96)MSE: 0.4700 -> 0.8952Hurt
increase M from 64 to 512 at Dk=128 (ablation)PLASH · ETTh2 (L=96)MSE: 1.0193 -> 0.4700Improved
use PLASH M=512 Dk=128 instead of VanillaInformer · WTH (L=96)MSE: 0.3121 -> 0.3009Improved

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