Test-time adaptation (TTA) lets pretrained models adjust to distribution shifts they encounter after deployment. A new arXiv paper takes aim at 'corruption drift' — the way model performance degrades as input corruption accumulates — and argues that hidden trajectories offer a more reliable corrective signal than commonly used final-layer representations.
The paper notes that most existing TTA methods depend on the weight representations of the final layer or on signals derived from the model's outputs. These approaches, the authors suggest, may miss information contained in intermediate layers. The proposed method instead tracks hidden trajectories — the evolving activations through the network — to correct for drift during adaptation.
Because the abstract is truncated, full experimental details are not available in the source. The title, 'Hidden Trajectories Don't Lie,' signals the core claim: intermediate states carry trustworthy evidence for correcting corruption drift. The work positions itself as an alternative to final-layer-centric TTA.