CED: CLIP-guided entropy dynamics for robust test-time adaptation in harsh visual conditions

Test-Time Adaptation (TTA) aims to enhance model robustness against unseen data distribution shifts, particularly under harsh visual conditions such as corruptions, motion blur, and low-quality imaging. Prevailing methods rely on self-supervised signals like entropy minimization, yet they are prone to generating overconfident predictions that ultimately lead to model collapse. Although recent approaches attempt to mitigate this issue by modifying gradients, the fundamental reasons for this failure remain poorly understood. To deeply understand the intrinsic mechanism of model collapse induced by entropy minimization, this paper pioneers an analysis of the process from a learning dynamics perspective. We reveal the existence of a squeezing effect during entropy minimization, identifying it as the root cause of the model's overconfident predictions. Based on this finding, we propose a novel adaptation framework, named CED (CLIP-Guided Entropy Dynamics). First, we introduce a frozen CLIP model as an external semantic signal to counteract the squeezing effect, thereby ensuring an unbiased estimation of entropy. Second, inspired by system control theory, we propose Trajectory-Aware Regularization (TAR) to ensure the model does not deviate excessively from its original robust state during adaptation. Extensive experiments on corrupted image and video benchmarks (e.g., CIFAR-10-C, ImageNet-C, Mini Kinetics-C) demonstrate that CED significantly outperforms state-of-the-art methods under harsh visual conditions, validating its effectiveness as a unified solution for robust visual adaptation. Our work highlights the potential of leveraging foundation models like CLIP for stabilizing and enhancing test-time adaptation in challenging environments. © 2026 Elsevier Ltd