AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

You Need Better Attention Priors

Elon Litman, Gabe Guo

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 uniform prior with learnable GOAT prior (vs ALiBi)125M decoder on C4 · C4 in-distribution perplexity at L=2048perplexity: 32.49 -> 30.95Improved
replace RoPE with GOAT prior for length extrapolation1.23B Llama-style decoder on FineWeb · FineWeb perplexity at 16k (trained at 2048)perplexity: 417.43 -> 24.77Improved
replace RoPE with GOAT at the training length1.23B Llama-style decoder on FineWeb · FineWeb perplexity at 2kperplexity: 20.82 -> 20.75No clear effect
use GOAT prior instead of RoPE, position-interpolated RoPE, ALiBi or sinusoidal PEGPT-style LM · Passkey retrieval and Needle-in-a-Haystackspan token accuracyImproved
use GOAT prior instead of learnable absolute positional embedding in a ViTViT on ImageNet-1k · ImageNet-1k top-1, trained at 224px, tested at higher resolutiontop-1 accuracyImproved
use GOAT instead of RoPE on human genome sequencesDNA language model · human reference genome next-token modelingvalidation NLLImproved
add explicit key-only sink bias u(j) separate from content keysGOAT · 125M C4 / synthetic copy-mixturelearned prior structureImproved
use fixed ALiBi slopes versus learned priorALiBi · Needle-in-a-Haystackspan token accuracyHurt
use ALiBi for length extrapolation instead of RoPE125M decoder on C4 · C4 length extrapolationperplexity: in-distribution 32.49 (ALiBi) vs 33.24 (RoPE)Mixed

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