Detecting Misbehaviors of Large Vision-Language Models by Evidential Uncertainty Quantification

Name: Detecting Misbehaviors of Large Vision-Language Models by Evidential Uncertainty Quantification
Uploaded: 2026-10-23T08:01:59.325822+00:00
Duration: 897 s
Description: Evidential Uncertainty Quantification (EUQ) is a training-free method that detects harmful or unreliable outputs from large vision-language models by measuring internal contradictions and knowledge gaps. It uses Dempster-Shafer Theory to decompose uncertainty into conflict and ignorance, improving detection across hallucinations, jailbreaks, adversarial attacks, and OOD failures.

Evidential Uncertainty Quantification (EUQ) is a training-free method that detects harmful or unreliable outputs from large vision-language models by measuring internal contradictions and knowledge gaps. It uses Dempster-Shafer Theory to decompose uncertainty into conflict and ignorance, improving detection across hallucinations, jailbreaks, adversarial attacks, and OOD failures.

Open the full episode