Reproduce “IntroConformal: Conformal Factuality Guarantees for Large Vision-Language Models via Introspective Signals” 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
Md. Atabuzzaman, Christian Alexander, Chris Thomas
IntroConformal filters claims in vision-language descriptions by decomposing responses, scoring claims with image-conditioned verification probability or layer stability, and calibrating thresholds on labeled examples. On held-out LLaVA-1.5 MSCOCO scene understanding, verification probability yields 0.054 response-level risk, 25% abstention, and 0.581 claim-level F1, versus CONFLVLM’s 0.102, 57%, and 0.504. The concurrent bound is α′ ≤ 0.170, conditional on calibration assumptions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.