A Multi-Agent Debate Approach Based on Large Language Models for Scientific Misinformation Detection

Abstract

The rapid advancement of generative artificial intelligence (GenAI) has accelerated the dissemination of misinformation, making it swifter and more covert, and posing significant risks such as public misperception, erosion of scientific authority, and trust crises. To address this challenge, this study investigates the representational characteristics of scientific misinformation and proposes MAD-MID (Multi-Agent Debate for Misinformation Detection in Science), a novel LLM-based detection framework. After a literature review and content analysis of 200 samples, we extract features across semantic structure, emotional tone, linguistic style, and technical specificity to build a domain-specific representation. MAD-MID employs three debating agents—proponent, opponent, and moderator—that interact dynamically to enhance contextual understanding, interpretability, and accuracy while reducing reliance on extensive pre-training. Experimental results show that MAD-MID achieves superior performance and robustness, contributing theoretical and methodological advances for a healthier information ecosystem in science and technology.