ClaimVerAgents: A Multi-Agent Retrieval-Augmented Claim Verification Framework

The spread of fake news has had major impact on public discourse and trust. Detection methods rely heavily on evidence quality and verdict accuracy. Traditional approaches, often based on static sources, struggle with outdated or incomplete information, especially for new or obscure claims. Large Language Models (LLMs) offer promising reasoning and generation capabilities but face similar challenges, including outdated knowledge and limited coverage. To address these challenges, we present ClaimVerAgents<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup>, a novel, retrieval-augmented, modular, and interpretable multi-agent system that leverages LLMs for real-time fake news verification. Each autonomous agent fulfills a specialized sub-task: claim extraction, query generation, evidence evaluation, verdict decision, and explanation generation, all within a transparent, confidenceaware pipeline. Extensive experiments on the PolitiFact dataset show that ClaimVerAgents outperforms both classical and LLM-based baselines in accuracy and robustness. Importantly, the system generates structured, humanreadable explanations alongside its verdicts, enhancing trust and interpretability.<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup>Code and data are available at https://anonymous.4open.science/r/ ClaimVerAgents-832E