Fathom: A fast and modular evidence based Automated Fact-Checking system

Abstract

The growing spread of misinformation, driven by social media and generative AI, has underscored the urgent need for scalable Automated Fact-Checking (AFC) systems. These tools are essential for verifying claims efficiently, as manual efforts cannot keep pace with the volume and speed of online content. We present an evidence-based verification pipeline designed for both efficiency and real-world applicability. Evaluated on the AVeriTeC 2025 shared task, our system achieved an Ev2R score of 0.3423 on the development set and 0.2043 on the test set, with an average claim verification time of under 22 seconds. To test real-world effectiveness, we constructed a new dataset by scraping claims from PolitiFact and Snopes, pairing them with up to-date supporting evidence and metadata such as source credibility and content type. On this dataset, our system achieved 87% accuracy on Snopes and 68% on PolitiFact, demonstrating strong performance in dynamic, real-world fact-checking beyond benchmark constraints.