Reproduce “Training-Free Active Test-Time Adaptation for Vision-Language Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jihwan Bang, Sumyeong Ahn, Hwanjun Song, Jae-Gil Lee

ACTOR uses delayed queried labels as retrieval anchors to improve future predictions from a frozen vision-language model under distribution shift. It first predicts, then stores returned labels and cautiously reuses nearby examples when two anchors agree and are sufficiently similar. With TDA, it raises CLIP ViT-B/16 mean accuracy from 68.14% to 72.66% across ten datasets, though not every dataset improves.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.