Agent evolution
Algorithm discovery, recursive self-improvement, harness evolution, and feedback-system design.
Incoming Ph.D. student in Machine Learning (CSE)
Shuhan Zhang
Georgia Institute of Technology
I am an incoming Ph.D. student in Machine Learning (CSE) at Georgia Tech, advised by Prof. Kai Wang. My research lies at the intersection of agent evolution, decision-making, and optimization.
I study how intelligent systems can improve not only a solution, but also the process that produces it: the algorithms they search, the feedback they use, and the evaluation mechanisms that guide them. I am also interested in interpretable choice models and reinforcement learning through the lens of optimization.
Previously, I completed my B.S. at CUHK-Shenzhen, visited the University of Pennsylvania, conducted research at UT Austin, and worked as a research intern at Microsoft Research Asia.
News
I will join Georgia Tech as an incoming Ph.D. student in Machine Learning (CSE).
I joined Microsoft Research Asia as a research intern, studying feedback in agentic algorithm discovery.
DeepHalo was presented as a Spotlight at NeurIPS 2025.
Logic-Logit was accepted to ICLR 2025.
Research interests
AI for decision-making
Choice modeling, decision-focused learning, preference learning, and human–AI interaction.
Optimization & learning
Reinforcement learning, Wasserstein geometry, differentiable optimization, and online learning.
Selected publications & projects
Full list on Google Scholar ↗2026
Is Auxiliary Feedback a Free Lunch for Agentic Algorithm Discovery?
Working paper · Agentic AI · Algorithm discovery
A mechanism-driven study of when richer feedback helps—or misleads—coding agents searching for better algorithms.
2025
2025
Logic-Logit: A Logic-Based Approach to Choice Modeling
ICLR 2025 · Poster · Interpretable ML · Discrete choice
An interpretable choice model built from sparse logical rules and trained through column generation and Frank–Wolfe optimization.
2026
Wasserstein Proximal Policy Gradient
Under review · Reinforcement learning · Wasserstein geometry
Policy optimization through Wasserstein geometry, with expressive implicit policy classes and distributional updates.
Education & experience
Ph.D. in Machine Learning (CSE)
Georgia Institute of Technology
Incoming student; advised by Prof. Kai Wang
Research Intern
Microsoft Research Asia
Agentic algorithm discovery and feedback systems
Summer Researcher
The University of Texas at Austin
Reinforcement learning and policy optimization
Visiting Student
University of Pennsylvania
Optimization, Bayesian analysis, and game theory
B.S. in Data Science
The Chinese University of Hong Kong, Shenzhen
GPA 3.9/4.0 · Dean’s List