AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Long-Context Generalization with Sparse Attention

Pavlo Vasylenko, Hugo Pitorro, Andre Martins, Marcos Vinicius Treviso

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: Linear attention approximations · 21 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 adaptive scaling (ASEntmax over fixed-alpha Entmax) (ablation)Entmax · MQMTAR 1024x lengthexact match accuracy (%): 9.3 -> 95.3Improved
add adaptive scaling (ASEntmax over Entmax) (ablation)Entmax · Reverse 8x lengthexact match accuracy (%): 0.2 -> 56.7Improved
add adaptive scaling (ASEntmax over Entmax) (ablation)Entmax · RULER S-NIAH-1 16Kaccuracy: 1.2 -> 97.4Improved
add adaptive scaling (ASEntmax over Entmax) (ablation)Entmax · ArXiv perplexity 16Kperplexity: 10.07 -> 10.01No clear effect
add adaptive scaling (ASEntmax over Entmax) (ablation)Entmax · Lambada perplexityperplexity: 47.9 -> 41.6Improved
replace softmax with alpha-entmaxSoftmax · ArXiv perplexity 16Kperplexity: 12.71 -> 10.07Improved
replace softmax with alpha-entmaxSoftmax · Reverse 4x lengthexact match accuracy (%): 0.0 -> 28.5Improved
replace softmax with alpha-entmaxSoftmax · Copy 64x lengthexact match accuracy (%): 85.5 -> 0.0Hurt
replace softmax with fixed Top-K (K=32)Softmax · MQMTAR 4x lengthexact match accuracy (%): 100.0 -> 6.2Hurt
replace SSMax with ASEntmaxSSMax · Copy 64x lengthexact match accuracy (%): 95.8 -> 86.6Hurt
replace SSMax with ASEntmaxSSMax · MQMTAR 1024x lengthexact match accuracy (%): 26.7 -> 95.3Improved
replace softmax with ASEntmaxSoftmax · PIQAaccuracy: 65.1 -> 63.8No clear effect

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