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

Universal Test-Time Adaptation (UTTA) tackles the challenge of handling both class and domain shifts in unsupervised settings with streaming test data. However, existing UTTA methods are often limited to minor shifts and heavily rely on heuristic rules. To advance UTTA under dual shifts, we propose a novel framework, Active Universal Test-Time Adaptation (AUTTA), and instantiate it with Exposing Mixture and Annotating Confusion (EMAC), which incorporates active human annotation into the UTTA setting. To select appropriate samples for annotation in AUTTA, we first identify the mixed regions of target-domain samples under dual shifts, thereby exposing reliable candidate samples. We then design a reward-guided active selection strategy to prioritize annotating the most representative samples within this set, maximizing the effectiveness of limited annotations. In addition, to balance pseudo-labels with scarce annotations, we introduce an adaptation objective that mitigates the imbalance and alleviates decision-boundary ambiguity. Extensive experiments demonstrate that AUTTA significantly improves performance and achieves state-of-the-art results under dual-shift scenarios. * Equal contribution † Corresponding authors: Wei Feng (first), Fuyuan Hu 2 RELATED WORK 2.1 UNIVERSAL TEST-TIME ADAPTATION Universal Test-Time Adaptation (UTTA) (Schlachter et al., 2025) is designed to address the challenge of domain and class shifts that are prevalent in open-world environments. Unlike traditional TTA (Sun et al., 2020) , which assumes the test data to be somewhat aligned with the source domain,