ReMMD: Realistic Multilingual Multi-Image Agentic Verification for Multimodal Misinformation Detection

A new framework called ReMMD aims to improve the detection of multimodal misinformation by verifying posts that combine multiple images and languages, according to research published on arXiv [1]. The system introduces a benchmark of 500 real-world samples and an agent that decomposes content into verifiable claims [1]. The ReMMD framework addresses a gap in current detection methods, which often focus on single images, short captions, or binary labels, failing to capture the complexity of modern viral posts that weave together long narratives, several images, and subtle framing errors [1]. The project includes ReMMDBench, a benchmark containing 500 samples and 2,756 images across five monolingual languages and two cross-lingual settings, with posts spanning three text-length tiers and labeled with five-way veracity and eight distortion categories [1]. At the core of the system is ReMMD-Agent, a persistent-memory verifier that breaks a post into atomic points, assembles a reusable evidence set, and then predicts structured outputs at three levels of detail [1]. When tested against proprietary systems, open large vision-language models, and other agent-based approaches, ReMMD-Agent achieved the highest five-way veracity performance, recording 41.80% accuracy and 39.12% macro-F1 using GPT-5.2 [1]. The researchers also reported cost reductions of 17.5% compared to MMD-Agent and 79.9% compared to T2-Agent [1]. The work arrives as the volume of multimodal misinformation continues to grow, with posts increasingly mixing languages and image provenance in ways that evade simpler detection tools [1]. The benchmark's inclusion of evidence provenance and rationales is designed to support more transparent verification pipelines, moving beyond black-box classification toward auditable, evidence-backed judgments [1]. The project page is publicly available at the authors' GitHub site [1].