MAD-Sherlock: Multi-Agent Debate for Visual Misinformation Detection

TL;DR

MAD-Sherlock introduces a multi-agent debate framework for visual misinformation detection that leverages external information retrieval to assess cross-context consistency between images and captions without domain-specific fine-tuning. By employing asynchronous debates with human-like perception and detailed explanations, the system achieves state-of-the-art accuracy on NewsCLIPpings, VERITE, and MMFakeBench, and provides interpretable reasoning to improve trust among experts and non-experts. The approach combines debate modelling, prompt engineering, and a two-stage retrieval pipeline (API-based retrieval and LLM summarization) to incorporate up-to-date external context into multimodal reasoning. User studies show that MAD-Sherlock enhances human accuracy and confidence in judging misinformation, signaling practical utility for citizen intelligence and public safety. The work lays a foundation for scalable, general-purpose misinformation detection with transparent explanations, while outlining limitations and future work in multilingual, video-text, and high-throughput deployments.

Abstract

One of the most challenging forms of misinformation involves pairing images with misleading text to create false narratives. Existing AI-driven detection systems often require domain-specific finetuning, limiting generalizability, and offer little insight into their decisions, hindering trust and adoption. We introduce MAD-Sherlock, a multi-agent debate system for out-of-context misinformation detection. MAD-Sherlock frames detection as a multi-agent debate, reflecting the diverse and conflicting discourse found online. Multimodal agents collaborate to assess contextual consistency and retrieve external information to support cross-context reasoning. Our framework is domain- and time-agnostic, requiring no finetuning, yet achieves state-of-the-art accuracy with in-depth explanations. Evaluated on NewsCLIPpings, VERITE, and MMFakeBench, it outperforms prior methods by 2%, 3%, and 5%, respectively. Ablation and user studies show that the debate and resultant explanations significantly improve detection performance and improve trust for both experts and non-experts, positioning MAD-Sherlock as a robust tool for autonomous citizen intelligence.

Figures (10)

• Figure 1: Overview of MAD-Sherlock: Two or more independent agents see the same image-text input and are tasked with detecting whether the input is misinformation or not. After the agents form their independent opinions, they participate in a debate until they converge on the same response or when $n$ debate rounds are completed (whichever is earlier).
• Figure 2: Debating Strategies: We experiment with multiple debating strategies to evaluate which performs best on our task.
• Figure 3: Structure of the external information retrieval module: We use the Bing Visual Search API to obtain web pages related to a given image, which are then summarised using Llama-13B llama. This summary is then passed to the debating agents as a part of the initial prompt.
• Figure 4: Russian President Vladimir Putin has called Ukraine's move into Kursk a "major provocation". Image and caption taken from the BBC article here (Accessed at 17:43 on Aug 11, 2024): https://www.bbc.co.uk/news/articles/cze5pkg5jwlo
• Figure 5: Model initialization prompt.