AI Paper Atlas

Research map / Label-efficient and robust learning / Test-time adaptation

SaTeen: Learning Structural Alignment for Continual Test-Time Adaptation

Chang Liu, Ruotong Zhao, Li Gao, Yupei Zhang

ICML 2026 · Label-efficient and robust learning › Test-time adaptation · approach: Pseudo-label test-time adaptation

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
add inter-sample structure alignment to intra-sample alignment (ablation)SaTeen (ResNet50-GN) · ImageNet-C single TTAaccuracy: 47.1 -> 49.9Improved
add inter-sample structure alignment to intra-sample alignment (ablation)SaTeen (VitBase-LN) · ImageNet-C single TTAaccuracy: 65.1 -> 65.4Improved
add inter-sample structure alignment to intra-sample alignment (ablation)SaTeen (ResNet50-GN) · ImageNet-C continual TTAaccuracy: 43.2 -> 47.1Improved
add inter-sample structure alignment to intra-sample alignment (ablation)SaTeen (VitBase-LN) · ImageNet-C continual TTAaccuracy: 62.7 -> 64.6Improved
use SaTeen instead of DeYOResNet18-BN · ColoredMNISTworst-group accuracy: 52.7 -> 67.2Improved
use SaTeen instead of DeYOResNet50-BN · Waterbirdsworst-group accuracy: 74.6 -> 76.3Improved
use SaTeen instead of DeYOResNet50-GN · ImageNet-C continual TTAaccuracy: 40.1 -> 47.1Improved
use SaTeen instead of DeYOVitBase-LN · ImageNet-C continual TTAaccuracy: 61.6 -> 64.6Improved
use SaTeen instead of DeYOResNet50-GN · ImageNet-C single TTAaccuracy: 42.4 -> 49.9Improved
use intra-sample alignment alone without sample selection (ablation)SaTeen · ImageNet-C continual TTAaccuracyImproved
define negative sample by randomly drawing another batch instance (ablation)SaTeen · ImageNet-C mild scenariosaccuracyMixed
use SaTeen instead of TentVitBase-LN · ImageNet-C continual TTA timeaccuracy: 49.0 -> 64.6Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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