AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Softplus Attention with Re-weighting Boosts Length Extrapolation in Large Language Models

Bo Gao, Michael W. Spratling, Letizia Gionfrida

ICML 2026 · Architectures and efficiency › Attention and transformers · approach: Theory of self-attention dynamics · 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 Softmax with LSSA (Softplus, l1-norm, entropy scale)GPT-2 small · FineWeb-10B val, 8Kvalidation loss: 6.2823 -> 5.9403Improved
add re-weighting p=15 to LSSA (ablation)LSSA · FineWeb-10B val, 8Kvalidation loss: 5.9403 -> 3.3171Improved
add re-weighting p=15 to Softmax (ablation)Softmax · FineWeb-10B val, 1Kvalidation loss: 3.1911 -> 5.3878Hurt
add re-weighting p=3 to Softmax (ablation)Softmax · FineWeb-10B val, 8Kvalidation loss: 6.2823 -> 6.2339Improved
add re-weighting p=15 to Sigmoid (b=-log N, l1-norm) (ablation)Sigmoid l1 · FineWeb-10B val, 8Kvalidation loss: 6.1846 -> 3.8587Improved
add l1-normalisation to Sigmoid (b=-log N) (ablation)Sigmoid · FineWeb-10B val, 8Kvalidation loss: 9.5939 -> 6.1846Improved
use b=-log N instead of -log L in Sigmoid attention (ablation)Sigmoid · FineWeb-10B val, 8Kvalidation loss: 14.4995 -> 9.5939Improved
use re-weighting p=3 instead of p=15 in LSSAR (ablation)LSSAR · FineWeb-10B val, 8Kvalidation loss: 3.3171 -> 6.3007Hurt
replace Softmax with LSSAR (p=15) in passkey retrievalGPT-2 small · passkey retrieval, 1Kaccuracy: 64% -> 86%Improved
replace Softmax with LSSAR (p=15) in passkey retrieval beyond 1.5KGPT-2 small · passkey retrieval, 1.5Kaccuracy: 0% -> 45%Improved
replace Softmax with LSSAR (p=15) on ARC-EGPT-2 small · ARC-Enormalised accuracy: 39.77 -> 40.57Improved
replace Softmax with LSSAR (p=15) on ARC-CGPT-2 small · ARC-Cnormalised accuracy: 23.72 -> 22.61Hurt

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