Balancing Stability and Plasticity: Dynamic Supervision Regulation for Multi-Label Continual Test-Time Adaptation

Continual test-time adaptation (CTTA) aims to adapt a source-trained model to continuously changing unlabeled test distributions without accessing source data during deployment. Existing CTTA methods are predominantly developed for single-label classification, where predictions compete through softmax. In multi-label recognition, however, each category is independently modeled with a sigmoid output, and online adaptation simultaneously involves multiple positive, negative, and uncertain label predictions. As a result, persistent pseudo-label supervision can repeatedly reinforce erroneous label decisions, whereas soft consistency alone, although stable, may remain overly conservative under challenging distribution shifts. In this work, we study continual test-time adaptation for source-trained multi-label classifiers and propose Dynamic Supervision Regulation (DSR) to balance adaptation stability and plasticity. DSR maintains teacher–student soft consistency throughout the target stream as a persistent stability anchor, while selectively injecting tri-state hard supervision as an additional plasticity signal. Specifically, confident teacher predictions are assigned positive or negative pseudo-labels, whereas ambiguous labels are excluded from hard supervision. A temporal Bernoulli student–teacher discrepancy is further used as an online control signal to determine when tri-state hard supervision should be activated, thereby avoiding continuous exposure to potentially noisy discrete pseudo-labels. Experiments on VOC-C, COCO-C, and NUS-WIDE-C demonstrate that DSR provides stable and effective continual adaptation across diverse multi-label distribution shifts while mitigating prediction drift and excessive positive predictions caused by overly aggressive supervision.