I develop methods for proactive agents, sample-efficient evaluation, failure discovery, and automated red teaming. My research builds on expertise in Bayesian optimization and decision-making — modeling uncertainty to guide sequential information gathering and making optimal decisions under uncertainty.

Helping humans understand AI

• Gaussian process probes for uncertainty-aware understanding
• Belief graphs of agents for interpretability
• Proactive failure discovery and efficient evaluation in generative AI
• Learning the underlying constitutions of models

Aligning AI to humans

• Pre-training Gaussian processes to align with expert beliefs
• Proactively gathering information from humans in image and multimodal content generation, and reasoning tasks
• Steering behavior via constitution-guided concept edits

To demonstrate proactive agents that ask clarification questions to seek information, I created the Proactive Co-Creator in Google AI Studio. You can also explore these concepts on my research showcase.

Try Proactive Co-Creator

I was a Visiting Lecturer at Harvard University in 2025 and have served as an Area Chair for top-tier AI/ML conferences including ICML, ICLR, AISTATS, and NeurIPS since 2023. My teaching experience also includes guest lecturing at HEC Montréal, Boston University, Brown University, and Harvard. I completed my Ph.D. in Computer Science at MIT in 2020, advised by Leslie Kaelbling and Tomás Lozano-Pérez. View my CV →

News

• Our paper ProEval: Proactive Failure Discovery and Efficient Performance Estimation for Generative AI Evaluation was accepted at ICML 2026!
• Our paper Interpreting and Controlling Model Behavior via Constitutions for Atomic Concept Edits was accepted at AISTATS 2026!