AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Trading Complexity for Expressivity Through Structured Generalized Linear Token Mixing

Erwan Fagnou, Paul Caillon, Blaise Delattre, Alexandre Allauzen

ICML 2026 · Architectures and efficiency › Attention and transformers · approach: State space and recurrent sequence models · 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
use cache-efficient variant (ablation)f(k)=2^k · Copyaccuracy (%): 92.63 -> 75.47Hurt
use cache-efficient variant (ablation)f(k)=2^k · Associative recallaccuracy (%): 49.03 -> 52.59Improved
use cache-efficient variant (ablation)f(k)=k^2+1 · Copyaccuracy (%): 99.66 -> 91.59Hurt
use cache-efficient variant (ablation)f(k)=k^2+1 · Associative recallaccuracy (%): 53.61 -> 54.56Improved
use dense recurrence instead of attentiontoken mixer · Multi-hop recallaccuracy (%): 39.21 -> 99.80Improved
use k-th order recurrence instead of diagonal SSMrecurrence · Copyaccuracy (%): 42.98 -> 74.66Improved
use f(k)=k^2+1 instead of f(k)=2^ktranslation-invariant pattern · Copyaccuracy (%): 92.63 -> 99.66Improved
use f(k)=k^2+1 instead of f(k)=2^ktranslation-invariant pattern · Multi-hop recallaccuracy (%): 34.85 -> 35.68No clear effect
use Mamba parameterization of A and B instead of the authors' one (ablation)first-order recurrence · OpenWebText 125Mperplexity: 31.28 -> 30.64Improved
use local attention instead of dense recurrencetoken mixer · Copyaccuracy (%): 100.00 -> 23.75Hurt

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