Séminaire Agorantic 2026/27 #1 - Maxime Dassen

Titre : FACTUM: Mechanistic Detection of Citation Hallucination in Long-Form RAG

Résumé : Retrieval-Augmented Generation (RAG) systems are only as trustworthy as the citations they produce. A citation hallucination occurs when a model cites a source that does not support its claim, even when the underlying statement may be correct. In this talk, I will present FACTUM, a mechanistic framework that studies citation hallucination through internal model signals related to contextual grounding and parametric knowledge. Using four interpretable scores, FACTUM outperforms existing detection baselines by up to 37.5% in AUC. I will also discuss how these metrics were developed, including their connection to ReDeEP, what they reveal about the mechanisms behind citation hallucination, and how this line of work could extend to other model families and new internal-signal-based metrics. Finally, I will discuss practical and broader implications, including applications in high-stakes settings such as law and how such techniques could be incorporated into real-world RAG systems.

Oratrice : Maxime Dassen is a PhD student in Artificial Intelligence / Computer Science at the IRLab of the University of Amsterdam, advised by Andrew Yates and Evangelos Kanoulas. Their research aims to understand how foundation models acquire, represent, and manipulate information across modalities and how that understanding can be used to build AI systems that are more reliable, interpretable, controllable, and efficient. In 2025, they were a visiting PhD student at the Johns Hopkins University HLTCOE as part of the SCALE program, where they worked on FACTUM.

Page personnelle : https://maximeswd.github.io/

Site de la FR Agorantic : https://agorantic.univ-avignon.fr/