Changhu Wang

This paper studies the problem of test-time adaptation for vision-language models (VLMs). Recent approaches typically measure the prediction entropy to store a confident cache for logit refinement. However, these confident samples tend to approach prototypes with limited coverage of data distribu...