NAVER LABS Europe virtual seminars are open to the public. Please register here for your participation (Zoom event).
Date: 10th June 2025, 11:00 AM (CEST)
About the speaker: Florian Le Bronnec is currently a PhD student jointly affiliated with Université Paris-Dauphine and Sorbonne University. He specializes in the study of generative models for text, with a particular interest in model faithfulness and evaluation methods. Alongside his research, he teaches in various programs within PSL, where he gives lectures on LLMs to both students and professionals.
Abstract: This talk introduces the concept of faithfulness in the context of large language models, discussing why it is critical for reliable text generation. I will present SCOPE, a self-supervised training method designed to improve model faithfulness in a variety of conditional generation settings. The method leverages preference learning from automatically constructed data and shows consistent improvements over standard fine-tuning approaches. The talk will conclude with a discussion on evaluation strategies for faithfulness, including both automatic and human-based assessments.