Reproduce “PROVE: Probabilistic Visual Grounding via Patch-Level Evidence 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
Mayank Nautiyal, Tianru Zhang, Li Ju, Ekta Vats, Prashant Singh
PROVE adds evidential grounding to frozen CLIP by training an adapter on gRefCOCO supervision. It predicts each patch’s relevance and evidential support with a Beta distribution, plus vacuity and an image-level absence score. On gRefCOCO testA, it achieves absence AUROC 0.849 ± 0.010, versus 0.823 for a same-architecture sigmoid control; this indicates ranking ability, not calibrated absence probabilities.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.