I work on generative models, world models, and reinforcement learning. I'm a PhD student at Mila and University of Montreal, advised by Aishwarya Agrawal. This summer I'll join Qualcomm Research in Amsterdam as an intern.

Research

The central question I work on: can a model learn to simulate the world well enough to act in it? We are entering the era of experience, where the data that matters most is not human-labeled but agent-generated.

My current research focuses on diffusion and flow models as world models. These models are expressive, but sampling from them is slow, which makes them nearly unusable for planning. I work on making them fast, temporally consistent, and structured enough to support a control loop. The adjacent question is representation: what does a model need to internalize about the world for its latent space to be useful for downstream decision-making?

A few deep-learning "recipes" I keep coming back to: the target network trick from RL, which later echoed through SSL and flow-based models; and asymmetric views in self-supervised learning.

See Google Scholar for a full list. *denotes equal contribution.

Academic Service