Research

My recent research focuses on understanding and improving multi-step inference processes in neural networks. I am interested in how internal representations evolve across iterative steps, and how these dynamics can be exploited to improve efficiency, controllability, generalization, and even the model itself.

Diffusion Models

I study how diffusion models allocate computation across layers and time steps, with an emphasis on dynamic routing, token/state selection, and accelerated inference.

Video Generation

I work on planning-generation decomposition for video models, scalable spatio-temporal architectures, and methods that improve controllability while maintaining generation quality.

Agentic Systems

I am interested in agentic systems that leverage multi-step inference and tool use, exploring how generative models can act, plan, and self-improve within interactive environments.