Reward-Adaptation: A Novel Test-Time Adaptation Method With Reward Model

Test-Time Adaptation (TTA) effectively addresses the domain shift problem with minimal computational overhead. Many TTA methods adapt to the target domain using features and pseudo-labels from the source model, but unreliable pseudolabels can degrade performance or even cause reverse optimization. Selecting reliable pseudo-labels is thus crucial. Existing filtering methods, such as entropy- or logit-based approaches, often yield inconsistent results across datasets. To address this, we propose Reward-Adaptation, which trains a lightweight model, called the Rewarder, to filter pseudo-labels. The Rewarder is plug-and-play, introduces minimal computational overhead, and consistently improves performance on both natural and medical datasets, enhancing existing TTA methods efficiently.