Hi! I’m Zhi.

I am a PhD student at UC San Diego advised by Tianhao Wang.

My research centers on the predictability and reliability of AI systems. Three questions I’m working on as a PhD student:

1. How do we know if an optimizer works? I approach this through theory and predictive science: proving how optimization behaves, or forecasting what training will do to help set up and monitor it. Examples include convergence bounds for adaptive optimizers and predicting or monitoring loss of plasticity.

2. How do we monitor and steer agents to make agentic systems more reliable? This connects to safety and credit assignment in long-horizon RL. I’m building RCP, a control hub to make agents more reliable for scientific research, with humans in the loop and an auditable record of their work.

3. How do we get VLM agents to perform robotic tasks reliably and with low latency? I’m interested in tasks with different horizons and different demands on the speed of perception and action.

Publications

• Lee Eon, Andrés R. Vindas-Meléndez, and Zhi Wang. Generalized Snake Posets, Order Polytopes, and Lattice-Point Enumeration. Discrete Mathematics (2026): 115072. We derived the Ehrhart and h-star polynomials for the generalized snake posets, with closed-form formulae for the two extreme cases, and obtained a monotonicity result using a novel technique that bridges combinatorics and lattice polytopes, extending several previous results.

People I’ve worked with

Tianhao Wang, Difan Zou, Andrés R. Vindas Meléndez, Eon Lee, Quan Wen, Misha Belkin.

Huge thanks to all my mentors and collaborators! Here is my CV. (last updated: Apr 2026)