I study how intelligent systems can learn reusable knowledge and reason with it at inference time to solve problems not anticipated during training.
I pursue this question across generative models, language and multi-agent reasoning, perception, and robotics. My long-term goal is to build physical agents that use compositional world models to imagine possible futures, plan, 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.
Research
This division between learning and inference shapes four areas of my work.
During trainingLearn reusable knowledgeConcepts, dynamics, and solution spaces
sample · optimize · search · plan
At inference timeSolve new problemsNew goals, constraints, and environments
Instead of producing one answer in a single pass, generative models represent many possible solutions. I develop iterative methods that search and refine these possibilities, so additional test-time computation can improve results or satisfy new constraints.
Rather than train a separate predictor for every combination of goals and constraints, I compose reusable models at inference time and search for solutions that satisfy them together.
World models learn how environments can evolve. Planning over their predicted futures lets robots choose actions for new goals and adapt when the environment changes.
When the reusable components are entire language models, their interactions can serve as an inference procedure. I study how debate, verification, and decentralized coordination allocate test-time computation to propose, evaluate, and refine candidate solutions.