AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Pretrain–Test Task Alignment Governs Generalization in In-Context Learning

Mary Letey, Jacob A Zavatone-Veth, Yue M. Lu, Cengiz Pehlevan

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: Training dynamics of in-context learning

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 the derived e_misalign alignment measure instead of simpler measures to predict ICL error2-layer softmax Transformer on linear regression ICL · Power-law and low-rank test covariancesSpearman correlation with ICL error: 0.99 (ours) vs 0.98, 0.96, 0.39Improved
increase pretraining task diversity kappa when train and test are poorly aligned (ablation)Linear attention ICL model · Solvable model, fixed C_train, varied C_testICL test errorMixed
pretrain on a more low-dimensional (higher spectral power) task covariance than the test one, at low task diversity (ablation)Linear attention ICL model · Power-law task covariancesICL test errorImproved
raise pretraining spectral power too far above the test power (ablation)Linear attention ICL model · Power-law task covariancesICL test errorHurt
use CKA or population C_test C_train^-1 as alignment predictor instead of e_misalignICL error prediction · Linear attention, power-law C_trainmonotonic relation to ICL errorHurt

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