D2ART
Dual-modal prompting for vision–language models at test time
D2ART adapts CLIP at test time by updating prompts in both modalities rather than text alone. Two pieces do the work: entropy-guided prompt updates, which use the model’s own confidence as the adaptation signal in the absence of labels, and EMA feature stabilization, which keeps that unsupervised signal from drifting as adaptation accumulates over a test stream.
On ImageNet-A — the natural-adversarial split where CLIP struggles most — this improved top-1 accuracy by 2.3% over prior test-time adaptation methods.
Done at the OV³ Lab, Wangxuan Institute of Computer Technology, Peking University, with Prof. Jiahuan Zhou.
Placeholder — to be replaced with the dual-modal prompting diagram.