AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Causal Attention with Lookahead Keys

Zhuoqing Song, peng sun, Huizhuo Yuan, Quanquan Gu

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
replace standard causal attention with lookahead keys (CASTLE)Transformer LM · FineWeb-Edu 50B tokens, 1.3Bvalidation loss: 2.426 -> 2.391Improved
replace standard causal attention with lookahead keys (CASTLE)Transformer LM · FineWeb-Edu 50B tokens, 160Mvalidation loss: 2.798 -> 2.792Improved
add lookahead keys (CASTLE)Transformer LM · 5-shot downstream average, 1.3Baverage accuracy (%): 58.11 -> 60.50Improved
add sliding window to lookahead keys (CASTLE-SWL)CASTLE · FineWeb-Edu 50B tokensvalidation loss: 2.391 -> 2.389 (1.3B)No clear effect
add causal keys alongside lookahead keys (ablation)CASTLE · FineWeb-Edu 25B tokens, 120Mvalidation loss: 3.021 -> 2.920Improved
add lookahead keys while matching key count to baseline (ablation)CASTLE-M-16 · FineWeb-Edu 25B tokens, 340Mvalidation loss: 2.742 -> 2.716Improved
apply SiLU to lookahead-key scores (ablation)CASTLE-SWL-XL · FineWeb-Edu 25B tokenstraining stabilityImproved
keep SiLU instead of lowering learning rate to stabilize (ablation)CASTLE-SWL-XL · FineWeb-Edu 25B tokensvalidation loss: 2.520 -> 2.503Improved
use sliding window 512 instead of 64 for lookahead keys (ablation)CASTLE-SWL · FineWeb-Edu 25B tokens, XLvalidation loss: 2.525 -> 2.503Improved
vary sliding window between 128 and 1024 (ablation)CASTLE-SWL · FineWeb-Edu 25B tokens, medium/XLvalidation loss: medium: best-worst gap 0.0038; XL: 0.0099No clear effect
use SiLU instead of tanh clipping of lookahead scores (ablation)CASTLE-M · FineWeb-Edu 50B tokens, 350Mvalidation loss: 2.634 -> 2.615Improved
use SiLU instead of constant scaling of lookahead scores (ablation)CASTLE-SWL-XL · FineWeb-Edu 50B tokensvalidation loss: 2.417 -> 2.389Improved

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