AI Paper Atlas

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

Adapt in the Wild: Test-Time Entropy Minimization with Sharpness and Feature Regularization

Shuaicheng Niu, Guohao Chen, Deyu Chen, Yifan Zhang, Jiaxiang Wu, Zhiquan Wen, Yaofo Chen, Peilin Zhao, Chunyan Miao, Mingkui Tan

TPAMI 2026 · Label-efficient and robust learning › Test-time adaptation · approach: Pseudo-label test-time adaptation · 4 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
add sharpness-aware reliable entropy minimization instead of plain entropy minimizationSAR · ImageNet-C L5 imbalanced label shift (ResNet50-GN)accuracy (%): 22.0 -> 37.2Improved
add feature redundancy and inequity regularizers with feature bank (ablation)SAR2 · ImageNet-C L5 imbalanced label shift (ResNet50-GN)accuracy (%): 37.2 -> 51.7Improved
add feature regularization at batch size 1 (ablation)SAR2 · ImageNet-C L5 batch size 1 (ViT-Base-LN)accuracy (%): 56.3 -> 67.4Improved
add feature regularization on mixed shifts (ablation)SAR2 · ImageNet-C L5 mixture (ResNet50-GN)accuracy (%): 38.3 -> 43.3Improved
add feature regularization on mixed shifts (ablation)SAR2 · ImageNet-C L5 mixture (ViT-Base-LN)accuracy (%): 57.1 -> 61.2Improved
add feature regularization in continuous adaptation (ablation)SAR2 · ImageNet-C L5 continuous adaptation (ResNet50-GN)accuracy (%): 27.4 -> 49.3Improved
use group norm instead of batch norm for entropy minimization (ablation)Tent · ImageNet-C L5 imbalanced label shiftaccuracy (%): 2.1 -> 22.0Improved
use SAR2 over DeYO under imbalanced label shiftSAR2 · ImageNet-C L5 imbalanced label shift (ResNet50-GN)accuracy (%): 43.9 -> 51.7Improved
DeYO selective entropy minimization in continuous adaptationDeYO · ImageNet-C L5 continuous adaptation (ResNet50-GN)accuracy (%): 43.9 -> 9.7Hurt
add feature regularization costSAR2 · ImageNet-C Gaussian L5 run timeGPU time (s): 115 -> 226Hurt
select samples for feature regularizers (SAR2 dagger variant) (ablation)SAR2 · ImageNet-C mixtureaccuracy (%): 163 s vs 226 s, 0.4% accuracy dropHurt
add inequity regularizer or redundancy regularizer individually (ablation)SAR2 · wild TTA settingsaccuracy (%)Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 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