Research
My research focuses on understanding and improving multi-step inference processes in generative models and agentic systems. I am interested in how internal states evolve across iterative computation, and how these dynamics can be exploited for efficiency, controllability, long-horizon generation, collective intelligence, and self-improvement.
Diffusion and Video Generation
Within this broader view of multi-step inference, I model video generation as an iterative spatiotemporal refinement process using diffusion models. I study how to allocate computation efficiently across layers, tokens, frames, and denoising steps to build scalable architectures, accelerate inference, enable controllable generation, and support long-context video synthesis.
Recursive Self-Improvement
I am increasingly interested in AI systems that improve by reasoning over their own multi-step traces, failures, generated data, and interaction histories. The long-term goal is to build models that can inspect and refine their own computation: not only generating better outputs, but also improving the procedures, memories, tools, and training signals that produce those outputs.
Agentic Systems
I study agentic systems that combine multi-step reasoning, tool use, and social interaction. I am especially interested in how scaling the number of agents, the structure of their communication, and the quality of their feedback loops can produce stronger planning and collective intelligence.
Approach
Beyond research topics, I care deeply about how problems are chosen and built. See Research Approach for a more personal view of my research style.