My research interest is in how discrete and recombinable concepts and skills can emerge in learning systems; my current line of work focuses on library learning, i.e., tool creation. Given a system (typically an LLM) that solves a stream of tasks (e.g., by generating Python code), the question is whether we can learn reusable and composable tools (e.g., well designed, non-trivial, reusable Python functions). More broadly, I am interested in the larger problem of library learning for program synthesis, and how hierarchical skills can be built up over time.

Other research interests include multi-modal representations of text and vision, and model evaluation. I have done research measuring compositional generalization to new combinations of textual and visual primitives, as well as research into the evaluation of generative dialogue agents.