New Method Improves AI Adaptation for Streaming Tabular Data.

Deep learning models often struggle with tabular data when the test distribution shifts from the training distribution, a common problem in real-world applications. Existing test-time adaptation (TTA) methods, largely inherited from computer vision, typically rely on batch-level statistics. However, new research reveals that these methods perform poorly in strict streaming scenarios where individual data points arrive one at a time, as batch statistics become unavailable or unreliable. To address this, the paper proposes Prequential Logit-Origin Centering (PLOC), a novel and lightweight approach specifically designed for singleton tabular TTA. PLOC keeps the source model frozen and instead shifts the logit space at each step by simply maintaining a running mean of past logits. This method requires no labels, estimates no priors, and avoids weight updates, making it highly efficient. Experiments across various tabular benchmarks, architectures, and checkpoints demonstrate that PLOC significantly outperforms strong existing TTA baselines, particularly in preserving AUROC by applying a static shift that maintains source ranking.