Inference-time agentic framework that decomposes multimodal misinformation detection into four modules: visual veracity, cross-modal consistency, retrieval-augmented checking, calibrated judgment. 81.65% F1 on MMFakeBench. arXiv:2510.17590 (formerly MIRAGE).
fact-checking llm vision-language-model llm-agent agentic-framework retrival-augmented-generation misinformation-detection misinformation-research mmfakebench
-
Updated
May 25, 2026 - Python