Before working with generative AI, my research used to focus on formal semantics that capture the behavoir of sets of programs (as opposed to single programs) in the context of program synthesis.

My current research aims to understand and influence generative AI models (such as LLMs) under this paradigm. Generative models can be thought of as defining a probability distribution over a set of programs; my research focuses on understanding and manipulating these distributions through the lens of semantics, with the goal of getting generative models to more reliably output text that meets a desired property.

For an example, check out Diffinity, where we utilize information about distributions defined by diffusion models, to guide them generate text that matches a regular expression.