AI Paper Atlas

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

Exposing Mixture and Annotating Confusion for Active Universal Test-Time Adaptation

Jiayao Tan, Fan Lyu, Chenggong Ni, Fuyuan Hu, Wei Feng, Rui Yao

ICLR 2026 · Label-efficient and robust learning › Test-time adaptation · approach: Realistic and stable 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 EMAC active annotation on top of UTTAOPTTT-style UTTA baselines · DomainNet avgAH: 49.8 -> 52.2Improved
use EMAC over strongest UTTA baseline on VisDA-COPTTT · VisDA-C SVHNAH: 69.2 -> 72.3Improved
use EMAC instead of generic active learning selectors (same B=1000)TENT-wrapped AL methods · DomainNet avgAH: 49.5 -> 52.2Improved
raise annotation budget B from 800 to 1500 (ablation)EMAC · DomainNet avgAH: 50.8 -> 53.1Improved
add exposing mixed candidates (EMSC) to SC+BTPO (ablation)EMAC · DomainNetAH: 50.8 -> 52.2Improved
add cluster-contrastive balanced loss (BTPO) to EMSC+SC (ablation)EMAC · VisDA-CAH: 71.7 -> 76.5Improved
add reward-guided selection (SC) to EMSC+BTPO (ablation)EMAC · DomainNetAHImproved
add sliding window buffer and EMA for GMM at batch size 1 (ablation)EMAC · DomainNet S2RAH: 16.3 -> 47.3Improved
select annotation samples from GMM mixed region rather than entropy-defined region (ablation)pseudo-label error analysis · Painting/Real/Sketcherror rate in mixed region: Xmix error 82.6/86.4/84.0 vs entropy 77.8/71.2/68.8Improved
inject 5-10% annotation noise (ablation)EMAC (B=800) · DomainNet avgAH-type avg: 50.8 -> 48.7 (10% error)Hurt

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