My research focuses on advancing generative AI systems that can reason formally with respect to a specification. A key theme in my work is the translation from foundational capabilities and algorithms (i.e., what can be computed about a generative system’s probability distribution) to practical inference schemes improving the reliability and efficiency of generative AI. I am particularly interested in using these methods to embed domain knowledge into generative AI in a reliable manner, with applications in learning biological networks, software testing, and robotics.
Language Model Inference
How do we improve the
accuracy, control, and alignment of large language models through
practical inference-time strategies?
Probabilistic Foundations
What can we characterize about the behavior of a generative AI system using
queries on its underlying probability distribution? What probabilistic queries can be computed
efficiently? How do we
design and
learn generative models with these query-answering capabilities?
Bayesian and Causal Methods
How do we model
uncertainty and the effect of
interventions on a system in a principled manner? How can we computationally
interpret and
make decisions based on these models?
Verification and Guarantees
How do we know that a given AI system will
behave as intended? Can we use this knowledge to
improve the robustness and accuracy of the AI system?
See Publications for a full list of my work, or my Google Scholar page.