AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Short Window Attention Enables Long-Term Memorization

Loïc Cabannes, Maximilian Beck, Maria Lomeli, Gergely Szilvasy, Matthijs Douze, Jade Copet, Pierre-Emmanuel Mazare, Gabriel Synnaeve, Herve Jegou

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: Linear attention approximations · 10 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
shorten sliding window from 2048 to 128 in trainingSWAX 1.4B · RULER NIAH (avg over tasks and lengths)accuracy: outperforms 2048 window by 16 accuracy points; 88.9% higher recallImproved
increase sliding window from 128 to 2048SWAX 1.4B · Code validation perplexityval_PPL: 2.551 -> 2.523Improved
increase sliding window from 128 to 2048SWAX 1.4B · Short-context reasoning averageaverage accuracy: 39.81 -> 40.88Improved
add xLSTM layers to SWA (hybrid SWAX:128 vs pure SWA:128)SWA 1.4B · Code validation perplexityval_PPL: 3.036 -> 2.551Improved
add stochastic window size (128/2048) instead of fixed 2048SWAX 1.4B · Code validation perplexityval_PPL: 2.523 -> 2.502Improved
add stochastic window size instead of fixed 128SWAX 1.4B · Short-context averageaverage accuracy: 39.81 -> 40.81Improved
add stochastic window size instead of fixed 2048SWAX 7B · Short-context averageaverage accuracy: 49.32 -> 49.52Improved
add stochastic window size instead of fixed 2048SWAX 1.4B · Short-context averageaverage accuracy: 40.88 -> 40.81No clear effect
add stochastic window size instead of fixed 2048SWAX 7B · RULER NIAH averageniah_average: 29.52 -> 37.73Improved
add stochastic window size instead of fixed 2048SWAX 1.4B · Babilongaverage score: 7.29 -> 9.35Improved
add stochastic window size instead of fixed 2048 (Gated DeltaNet)SWAX/GDN 1.4B · Ruler multiquery NIAHaccuracy: 28.80 -> 40.90Improved
add stochastic window size instead of fixed 2048 (Gated DeltaNet)SWAX/GDN 1.4B · LongBenchaverage score: 11.42 -> 9.94Hurt

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