MFC-Bench introduces a comprehensive framework for measuring factual accuracy in large vision-language models (LVLMs), addressing critical concerns about AI trustworthiness.

• Evaluates LVLMs on their ability to detect multimodal misinformation across diverse domains
• Tests models on manipulation detection and veracity classification of text-image pairs
• Reveals significant performance gaps even in state-of-the-art systems like GPT-4V
• Provides a standardized method to improve factual reliability in next-gen AI systems

Security Impact: As LVLMs become more prevalent in critical applications, MFC-Bench offers a crucial tool for identifying and mitigating potential security vulnerabilities related to misinformation propagation and manipulated content detection.

MFC-Bench: Benchmarking Multimodal Fact-Checking with Large Vision-Language Models