SDTTA: Semantic Dependent Test-Time Adaptation with Ordinary Differential Equation meta-networks

Test-Time Adaptation (TTA) typically adapts pre-trained models to the target domain via entropy minimization, which often leads to error accumulation and catastrophic forgetting due to the lack of label guidance. Approaches that freeze the backbone to optimize attached meta-networks effectively avoid catastrophic forgetting but typically necessitate a source-domain warm-up, violating the source-free premise. To address these challenges, we propose Semantic Dependent Test-Time Adaptation (SDTTA), a robust, warm-up-free meta-network paradigm. First, we introduce a Semantic-Guided (SG) adaptation mechanism. By utilizing on-the-fly linear probes to quantify the semantic intensity of intermediate features, SG prioritizes adapting universal low-level features while shielding high-level discriminative boundaries. Second, to mitigate the noise from random initialization that is difficult to rectify in source-free scenarios, we model the meta-networks using Neural Memory Ordinary Differential Equations (nmODE). Leveraging its stable dynamical structure and theoretically bounded mapping, nmODE establishes structural stability by suppressing initialization noise and ensuring theoretically bounded feature transformations. Extensive experiments on CIFAR10-C, CIFAR100-C, and ImageNet-C demonstrate that SDTTA achieves competitive performance against state-of-the-art methods. The code is available at https://github.com/bgx666/sdtta .