I work toward Agentic Recursive Self-Improvement to advance AutoResearch and Long-Horizon Problem Solving.
I am a third-year Ph.D. student in Machine Learning at Georgia Tech, co-advised by Bo Dai and Chao Zhang.
Currently, I am a Student Researcher at ByteDance Seed, improving model behavior for long-horizon coding tasks and building AutoResearch agents that identify worthwhile questions and develop promising ideas.
Previously, I received my B.Eng. in Computer Science from Tsinghua University, where I was a member of the TSAIL group advised by Jianfei Chen and Jun Zhu.
Always happy to talk about research and new ideas. Feel free to reach me at changhao.li@gatech.edu.
Research
My research spans the core technical stack for Agentic RSI, including:
Harness Evolution
Learned model control and hierarchical orchestration to improve how agents solve long-horizon tasks.
Efficient On-Policy Learning
Efficient policy learning from on-policy experience with dense expert feedback.
Data, Environments & Rubrics
Executable tasks and environments for experimentation, paired with adaptive rubrics for research evaluation.
Training Infrastructure
A modular system connecting task execution, rollout collection, and model updates for agent training.
News
-
Started as a Student Researcher at ByteDance Seed in San Jose, working on long-horizon coding agents and AutoResearch.
-
Released STACX, an open-source modular infrastructure for end-to-end agentic RL.
-
Released Agent DAgger, combining on-policy experience with dense expert feedback to train long-horizon coding agents.
-
EDO is accepted by TMLR 2026.
-
DREAM, an agentic evaluation framework for Deep Research agents, is accepted to ACL 2026 as an Oral presentation.
-
Matryoshka Pilot and MLE-Dojo are accepted by NeurIPS 2025.
-
TS-ZSQ is accepted by ICCV 2025.
