Probabilistic test-time adaptation framework handles data shifts without labels

A new paper from researchers Daniel Corrales and David Ríos Insua introduces a probabilistic framework for online test-time adaptation, a critical problem in machine learning where a model trained on labeled data must adapt to unlabeled test data under distributional shift. The framework is built on a state-space modeling architecture that jointly characterizes parameter learning, temporal evolution of parameters, prior tuning, and predictive inference. This allows the model to continuously update its beliefs about model parameters as new unlabeled test samples arrive, without requiring ground-truth labels. The approach is particularly relevant for real-world scenarios where data distributions drift over time—such as changes in user behavior, sensor degradation, or environmental conditions.

The key technical contribution is a principled probabilistic treatment that extends classical Bayesian online learning to the test-time setting. By treating the model parameters as latent states evolving via a Markov process, the framework can infer posterior distributions over parameters given the observed test data stream. This enables uncertainty-aware predictions and robust adaptation even with limited or noisy observations. The authors also discuss how to tune priors and handle non-stationary environments. While the paper is theoretical, it provides a foundation for practical algorithms that could benefit applications like autonomous driving, medical monitoring, or recommendation systems, where retraining is expensive or impossible. The work is published on arXiv (2606.26457) and aims to bridge the gap between offline training and online deployment.