Chao Zhang

Chao Zhang

James Edenfield Associate Professor

School of Computational Science and Engineering
College of Computing, Georgia Institute of Technology

I am an Associate Professor (and Amazon Scholar) at Georgia Tech. I received my Ph.D. from the University of Illinois at Urbana-Champaign, advised by Jiawei Han. My research builds LLM agents for open-ended discovery — from automating ML research to scientific discovery.

Google Scholar LinkedIn X / Twitter

Research

I build LLM agents for open-ended discovery — agents that navigate vast design spaces, form and revise hypotheses, and learn from sparse, delayed, and noisy feedback. My group develops diversity-driven search and long-horizon learning methods, validated on ML research automation and scientific discovery.

Diversity-Driven Agent Search

MCTS vs A* search in ToolChain*: two tree search strategies compared

Discovery spaces are combinatorially vast: molecular space contains more candidates than atoms in the observable universe; the space of possible ML algorithms, architectures, and training procedures is effectively unbounded. Random exploration is hopeless in these spaces, and naive language model sampling wastes the structured knowledge that could guide search toward promising regions.

The core challenge is not just longer reasoning traces; it is choosing which hypotheses, candidates, or experiments to try next when the space is vast and each evaluation is expensive. We combine tree search over compositional action and design spaces, quality-diversity methods that preserve breadth, and uncertainty-guided selection that prioritizes the most informative next evaluations.

Long-Horizon Agent Learning

Long-horizon agent learning: discovery trajectory with sparse outcome and intermediate feedback signals flowing into multi-turn RL, self-rewarding, and domain adaptation methods

Discovery agents must learn from feedback that is sparse, delayed, expensive, and noisy. A failed experiment is not wasted — it carries information. Even strong execution-agent recipes break down when supervision comes as sparse, delayed, domain-specific outcomes rather than correct traces or preference labels.

We study how discovery agents should update from weak real-world feedback: multi-turn RL from partial outcomes, continual pre-training to strengthen core agent capabilities, and lightweight adaptation methods that specialize models to new domains without overwriting prior capabilities.

Application Domains

Automating ML Research

We deploy the full stack — search, learning, and environments — to build ML agents that automate the research cycle from hypothesis to experiment to iteration. MLE-Dojo provides interactive training environments that capture the structure of real ML research workflows. MLE-Smith scales this with automated multi-agent pipelines. These systems enable agents to tackle open-ended research tasks end-to-end.

Science: Chemistry, Materials, and Molecular Design

In chemistry and materials science, we navigate molecular and materials design spaces where each candidate requires synthesis and physical characterization. Our work spans LLM-augmented synthesis planning, evolutionary search over chemical space, autonomous materials discovery, and uncertainty-aware molecular property prediction.

Molecular discovery evolutionary loop with LLM-guided mutation and crossover Molecular dynamics simulation

Awards & Recognition

Awards

Funding

Supported by NSF (CAREER-2144338, IIS-2403240, ACED-2435754, IIS-2106961, IIS-2008334), ONR MURI, and industry partners including Amazon, Google, Meta, Adobe, HomeDepot, and Kolon.

Group

Prospective students: I am always looking for strong and motivated students to join our group. If you are interested in working with me, you can either email me or fill out this form.

Current PhD Students

Alumni

Publications

(* denotes equal contribution)

2026
2025
2024
2023
2022
2021
2020
2019
2018
Earlier

Teaching

Contact

Office: CODA E1358B
Address: 756 W Peachtree St NW, Atlanta, GA 30308
Email: chaozhang@gatech.edu