École Nationale Supérieure de
Techniques Avancées

Abstract

During my internship, I worked on the challenge of hallucination detection in Large Language Models, a key obstacle to their reliable deployment in real-world applications. My work focused on designing an evaluation framework that distinguishes between extrinsic and intrinsic hallucinations and on extending an attention-based uncertainty quantification algorithm with new attention aggregation strategies. I evaluated these methods across several open-source models and benchmarks, comparing them against state-of-the-art baselines. The results showed that sampling-based methods are more effective for detecting extrinsic hallucinations, while the proposed attention-based approaches perform better for intrinsic hallucinations. This internship allowed me to gain experience in large-scale experimentation, uncertainty quantification, and evaluation of LLMs in collaboration with both academic and industrial teams. A major outcome of this work was the acceptance of a research paper based on these contributions.