Teaching

Modern NLP: Mechanistic Interpretability of Large Language Models

Course held within the national PhD program in AI and Society (20 hours). The course focuses on recent topics related to mechanistic interpretability of Large Language Models (LLMs) including:

• Probing

• Outliers in LLMs

• Activation Steering

• Sparse Autoencoders

With guest lectures from Fabio Brau, Gabriele Sarti and William Rudman