Dylan Iddings and Daniel Seredensky

Abstract:

Our project is a solution for the TREC 2025 DRAGUN track. This track’s goal is to find the most effective way to verify the trustworthiness of online articles by providing users with the necessary background and context to make their own decisions. Our group focused on Task 2, which asks us to write short, fact-checked, well-sourced reports that provide the answers to the questions readers would likely have regarding an article. Our team made an AI-powered pipeline that updates reports through different rounds of revision. Every time the system evaluates the previous version, it figures out which aspects are missing and searches for additional evidence to create a better version. To generate context, we use traditional keyword searches as a baseline and then apply AI-based reranking to get the most relevant passages to be included. The last versions of the reports are split into concise sections where each idea is backed up by citations.