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 at Georgia Tech and an Amazon Scholar. I received my Ph.D. from the University of Illinois at Urbana-Champaign, advised by Jiawei Han. My research builds scalable training and evaluation loops for LLM agents on long-horizon tasks. We develop inference-time search, post-training algorithms, and agent harnesses that turn agent trajectories into better data, faster feedback, and stronger models — applied to scientific discovery.

Google Scholar LinkedIn X / Twitter

Research

Long-horizon tasks remain hard for LLM agents: action spaces are large, feedback is sparse or delayed, evaluation is expensive, and errors compound over many steps. Our group works across three axes — inference-time search and planning, post-training algorithms and data from agent trajectories, and scalable environments and harnesses for agent evaluation. We apply these methods to scientific discovery in chemistry and materials science.

Search and Planning

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

When the space of possible hypotheses, tool sequences, or molecular candidates is combinatorially vast and each evaluation is expensive, naive LLM sampling is hopeless. The core problem is selecting which candidate to evaluate next given a limited budget.

Our work in this area includes A* tree search over compositional tool chains, evolutionary quality-diversity methods that maintain breadth across chemical space, and uncertainty-guided selection that routes evaluation budget toward the most informative candidates.

Post-Training Algorithms and Data

Agent training from research trajectories: multi-turn RL, self-rewarding, and continual pre-training methods

Long-horizon agent tasks demand sample-efficient training from multi-step trajectories, where errors compound and feedback is sparse. Recent work includes multi-turn RL from end-to-end task outcomes, on-policy distillation into smaller models to avoid error compounding, continual pre-training that builds core agent capabilities, and self-rewarding methods that eliminate external reward models.

Agent Environments and Harnesses

Long-horizon agents need more than algorithms — they need realistic environments to act in, execution harnesses that manage tool calls and trace collection, and evaluation pipelines that measure progress reliably.

We build environments and harnesses for studying LLM agents in realistic workflows. MLE-Dojo provides interactive environments that capture real ML research tasks. MLE-Smith scales evaluation with automated multi-agent pipelines. Our ongoing work targets scalable orchestration and cost-efficient evaluation.

AI for Scientific Discovery

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

We deploy our agent methods to scientific domains where each proposed candidate requires real synthesis and physical characterization — making every evaluation expensive and every failed attempt informative. Current focus: chemistry, materials science, and molecular design.

Our work spans LLM-augmented synthesis planning that proposes viable reaction routes, evolutionary search that navigates molecular space under synthesizability constraints, autonomous materials discovery systems, and uncertainty-aware molecular property prediction.

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