Toward Fine-Grained Domain Knowledge: Curriculum Pseudo-Labeling for Online Test-Time Adaptation

ICLR 2026 Conference Submission15203 Authors

Keywords: Curriculum Pseudo-Labeling, Online Test-Time Adaptation, Domain Knowledge

TL;DR: We propose a novel framework curriculum pseudo-labeling for Online Test-Time Adaptation, which further mines domain knowledge at a fine-grained instance level.

Abstract: Online Test-Time Adaptation (OTTA) aims to adapt a pre-trained model to unlabeled test instances under domain shift in an online manner, where domain knowledge that the model accumulates from previously observed mini-batches directly affects its predictions on subsequent instances. Most previous OTTA methods exploit domain knowledge at a coarse-grained batch level, which prevents the model from fully absorbing the domain knowledge. To deal with this problem, we propose a novel framework CUrriculum Pseudo-Labeling for Online Test-time adaptation (CUPLOT), which further mines orderly domain knowledge at a fine-grained instance level. Specifically, CUPLOT prepares the arriving batch as a series of curricula based on the modeled relevance of domain knowledge between the model and instances. Then, the model orderly learns the instances with pseudo-labels generated by class prototypes in each curriculum. In this way, the domain knowledge is accumulated in a fine-grained manner through instances of curricula rather than mini-batches, improving the absorption of domain knowledge and the performance of the model. Theoretically, we prove that the curriculum pseudo-labels could enable the model to have a stronger adaptation ability, resulting in a tighter bound of approaching the Bayes optimal classifier on the target domain.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 15203