Victor Morand Understanding LLM Entity Representations for Better knowledge manipulation
- Date: 7 septembre 2026 à 13h00
- Salle: 65-66 304
Large Language Models (LLMs) have revolutionized how we interact with information, yet how they internally organize and manipulate real-world knowledge remains a black box. In this talk, we’ll dive into LLMs magic through the lens of Entities — fundamental “atoms of knowledge” within text.
Our recent mechanistic interpretability research confirms a widespread intuition: LLMs naturally learn to encode entity structures during pretraining, processing multi-token entities almost as if they were single words in their native vocabulary. We will explore how these hidden representations are built, manipulated and how we can exploit them to extract entity mentions efficiently without any schema specification. Finally, we will discuss how uncovering these internal mechanisms allows us to build lightweight, generalist mention detectors, paving the way for more grounded, interpretable, and powerful information retrieval systems.