Debating to Verify: A Robust and Explainable Multi-Agent LLM System for Fact-Checking

Large language models have advanced automated fact verification, yet single-agent prompting remains prone to hallucination, while fine-tuned models suffer from limited scalability and cross-domain generalization. This paper proposes FC-MAD, a training-free multi-agent debate framework that coordinates multiple LLMs through structured critique, context summarization, and judge-guided consensus reasoning. Extensive experiments on Vietnamese (ViFactCheck), multilingual (X-Fact), and English (FEVER) benchmarks show that FC-MAD consistently outperforms strong fine-tuned and prompting-based baselines, achieving state-of-the-art performance on ViFactCheck and FEVER, while delivering robust gains across X-Fact languages. These results highlight the effectiveness of structured multi-agent reasoning for reliable AI-based fact-checking systems.