Language-Model Interpretability and Alignment

We study how large language models represent knowledge and use it during computation. Through mechanistic interpretability, we examine the internal features and pathways that contribute to model behavior.

Our work includes organizing learned features into maps of related concepts and testing how changes to internal representations affect subsequent reasoning.

These studies help distinguish a change in the model’s answer from a change in the computation producing it. Our broader goal is to develop a stronger basis for evaluating and guiding model behavior toward reliable, intended outcomes.