Hi, I’m Shaoyi Zheng, currently a Ph.D. student in Computer Science at the Courant Institute of Mathematical Sciences, New York University, where I am advised by Prof. Shengjie Wang. I completed my undergraduate studies at NYU Shanghai, majoring in Computer Science with a minor in Mathematics.
World Model & Robotic Learning — Using video world models as a substitute source of experience, so robot policies can scale on generated interaction rather than teleoperation hours. The catch is latency: a world model worth rolling out is far too slow to close a control loop, so the other half of this is making it fast enough to act on.
Robotic Agentic System — Hierarchical rather than monolithic: a deliberative layer that plans over long horizons, and a reactive action model that runs at control rate. I am interested in how cleanly the two decouple — what the planner hands down, and how they stay consistent when plan and reality disagree.
Efficiency Foundation Model — Making large generative models fast without giving up quality: sparse and hardware-aligned attention, token merging, and selective KV-cache recomputation for long contexts. FLOP savings mean little if they fight the hardware, so I care about speedups that hold up on the clock.
