Assistant Professor of Computer ScienceHarvard University · Kempner Institute
I study how intelligent systems can learn reusable models, then compose and reason with them at inference time.
I explore this question through four interfaces between representation and inference: search over generative solution spaces, composition of reusable models, planning with world models, and reasoning through multi-agent interaction.
Across these directions, a central focus of my work is building robots that can plan, adapt, and act reliably in unfamiliar environments.
I lead the Embodied Minds Lab at Harvard and am a part-time research scientist at NVIDIA Research. I received my PhD from MIT EECS, advised by Leslie Kaelbling, Tomas Lozano-Pérez, and Joshua B. Tenenbaum. Before Harvard, I held research roles at OpenAI and Google DeepMind.
Different representations support different forms of inference. Our group studies four interfaces between what models learn and how they reason at inference time.
KnowledgeSolution spacesReasoningSample and refine
Generative models represent a space of possible solutions. Iterative inference lets us search that space and refine candidate solutions to improve results or meet new constraints.
Examples: deep EBMs, Reduce, Reuse, Recycle, and reasoning with sampling.
KnowledgeReusable modelsReasoningCompose and search
Reusable models capture different sources of knowledge. Composing them at inference time lets us adapt to new combinations of goals and constraints.
Examples: composable diffusion models, Diffuser, and compositional energy minimization.
KnowledgeWorld dynamicsReasoningPredict and plan
World models capture how environments evolve and how actions change them. Planning over possible futures enables robots to solve new tasks robustly and adapt as conditions change.
Examples: UniPi, Video Language Planning, and Large Video Planner.
KnowledgeDistributed across agentsReasoningInteract and coordinate
Different language models capture complementary knowledge and capabilities. Composing their interactions at inference time lets them function as a unified system that evaluates and refines candidate solutions.
Examples: multi-agent debate, multi-agent verification, and Economy of Minds.
* Equal contribution; † or + equal advising.