Yilun Du

I am an Assistant Professor at Harvard in the Kempner Institute and Computer Science, where I run the Embodied Minds Lab. I am also a part-time research scientist at NVIDIA Research. I received my PhD from MIT EECS, advised by Leslie Kaelbling, Tomas Lozano-Perez, and Joshua B. Tenenbaum; previously, I was a research fellow at OpenAI and a senior research scientist at Google DeepMind.

My research asks how intelligent systems can solve unfamiliar problems by iteratively generating, evaluating, and refining possible solutions. Instead of relying on a single feedforward prediction, I build generative models of possible futures, actions, and paths of reasoning, and use inference-time computation—including sampling, energy minimization, search, planning, and interaction—to adapt them to new goals. My long-term vision is to build compositional world models that support reliable reasoning and action in open-ended physical environments.

My work has helped shape several parts of this agenda. Early work on deep energy-based models demonstrated scalable high-dimensional generation, refinement, and composition using Langevin dynamics, contributing to renewed interest in energy-based generative modeling; subsequent work showed how energy functions and diffusion models can be composed. Diffuser and Diffusion Policy helped establish diffusion models as tools for planning, decision-making, and robot control. UniPi was an early demonstration that text-conditioned video generation could serve as a policy for control. Our work on multi-agent debate helped establish iterative interaction among language models as an inference-time strategy for improving reasoning and factuality.

Across these projects, the common principle is that intelligence can emerge through structured, iterative inference with reusable, composable models—over pixels, trajectories, possible futures, and the reasoning of other agents. My lab is extending this principle to compositional world models, continuous reasoning, long-horizon embodied planning, and economies of interacting agents.