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

One-line Summary

The paper introduces Evidential Uncertainty Quantification (EUQ), a method to detect misbehaviors in large vision-language models by assessing internal conflicts and knowledge gaps, outperforming existing methods in identifying issues like hallucinations and adversarial vulnerabilities.

Plain-language Overview

Large vision-language models (LVLMs), which understand and generate content from both images and text, sometimes produce unreliable or harmful outputs, especially when given tricky inputs. This can be dangerous in critical applications. The paper proposes a new method called Evidential Uncertainty Quantification (EUQ) to better detect these issues by identifying internal conflicts and gaps in the model's knowledge. The researchers tested EUQ on various problems like hallucinations and adversarial attacks, finding it more effective than current techniques.

Technical Details