Modern neural networks are increasingly powerful, but most of them still use a fixed amount of computation for every input, even when some inputs may need far less work than others. I am interested in building models that can decide how to use computation adaptively, such as which paths to take, which modules to activate, which agents to communicate with, how much to recurse, and how much sampling is actually needed.

This leads to two closely related areas. The first is efficient neural networks and conditional computation, where I study routing and adaptive architectures trained with or without reinforcement learning. The second is generative modeling, especially diffusion models, where better sampling can make generation more efficient and more useful for world modeling, video generation, and discrete domains.

Email · Google Scholar · GitHub · Twitter / X