ReMMD Detects Multimodal Misinformation with Agentic Verification.

This paper introduces ReMMD, a comprehensive framework designed to tackle the growing challenge of multimodal misinformation detection, particularly in complex, real-world scenarios. Existing benchmarks often fall short by focusing on isolated elements like short captions or single images. ReMMD addresses this by providing ReMMDBench, a new benchmark dataset featuring 500 real-world samples with multiple images, multilingual narratives, varying text lengths, and detailed veracity and distortion labels, along with evidence provenance. Central to the framework is ReMMD-Agent, a persistent-memory verifier that intelligently decomposes misinformation posts into atomic points, constructs a reusable evidence set, and predicts structured veracity outputs. This agentic approach significantly outperforms proprietary systems and open LVLMs in five-way veracity performance, achieving 41.80% accuracy and 39.12% macro-F1 with GPT-5.2. Crucially, ReMMD-Agent also demonstrates substantial cost reductions compared to other agentic verification methods, making it a more practical solution for realistic misinformation detection.