Opening the Black Box: Interpretability from Natural Language Processing to Large Language Models
Date:
Interpretability is central to opening the “black box” of AI: explaining the structure and behavior of models and improving their trustworthiness and safety. It is key to the compliant use and continued improvement of AI technology. This talk reviews the course of interpretability research for large language models. It first revisits interpretability methods from traditional natural language processing and their applications in dialogue systems and question answering. It then turns to the capability boundaries of large language models and the new interpretability challenges they raise, and presents a framework for mechanistic interpretability, focusing on sparse autoencoders for disentangling superposed features, circuit analysis for uncovering causal relations among features, and sparse replacement models for constructing essential circuits. Finally, it looks ahead to scalable interpretability research as a theoretical and technical foundation for trustworthy large models.