[arXiv]score: 0.29

Marginal Coverage Failure in Zero-Shot VLM Conformal Prediction

August 21, 2026

Conformal prediction methods used as abstention layers for CLIP and SigLIP fail to guarantee class-conditional safety under distribution shift. While marginal coverage may remain high, tail coverage for specific classes can drop to near zero, meaning the model's uncertainty estimates are unreliable for rare categories.

HOW THIS AFFECTS YOU

●

builderDo not rely on aggregate marginal coverage scores to ensure safety or reliability for all object classes.

●

researcherThis highlights a critical gap in applying conformal prediction to zero-shot vision-language models under shift.

●

policyYou should be aware that uncertainty guarantees for VLMs can be deceptive when deployed in new environments.

read original ↗arxiv.org

DAILY DIGEST

you don't check 9 sources — we do. one email every morning, read in 2 min. free. unsubscribe anytime. privacy

← back to feed