Biography

My research focuses on self-improving agents and autonomous research, reinforcement learning for agentic LLM post-training, and decision-making. Download my CV.

Education

• Ph.D. candidate, Computer Science, University of Illinois Urbana-Champaign (2024.08 - present).

  • Advisor: Prof. Tong Zhang.

• MPhil, Artificial Intelligence, The Hong Kong University of Science and Technology (2021.09 - 2024.08).

  • Advisor: Prof. Tong Zhang.

• B.S., Statistics, University of Science and Technology of China (2017.09 - 2021.06).

Research Experience

• Research Intern, Meta MSL / FAIR (2026.05 - present).

  • Managers: Sainbayar Sukhbaatar and Jason Weston.

  • Learning agent skills and file-system memory from experience for continual learning and synthetic data generation.

  • Distilling agent reflections into reusable task-authoring and rubric-design skills.

  • Investigating compute-efficient training for autonomous research agents on long-horizon, multi-turn tasks.

• Applied Scientist Intern, Amazon (2025.05 - 2026.05).

  • Hosts: Dr. Zhou Yu and Dr. Ziji Zhang.

  • Developed Adaptive Layerwise Perturbation (ALP) for robust off-policy RL under system noise and training-inference mismatch, with applications to mathematical and tool-integrated reasoning.

  • Proposed the Process Consistency Filter (PROF) and developed PROF-GRPO to improve answer accuracy and intermediate reasoning quality.

• Ph.D. Research, University of Illinois Urbana-Champaign (2024.09 - present).

  • Advisor: Prof. Tong Zhang.

  • Developed Reinforce-Ada for adaptive sampling in online RL post-training and a self-rewarding correction framework for mathematical reasoning.

  • Developed an online iterative RLHF framework, implemented it with online DPO, and extended it to general preference models.

• MPhil Research, The Hong Kong University of Science and Technology (2021.09 - 2024.08).

  • Advisor: Prof. Tong Zhang.

  • Formulated RLHF as a reverse-KL-regularized contextual bandit problem and developed statistically efficient algorithms with finite-sample guarantees.

  • Developed uncertainty-weighted algorithms for corruption-robust reinforcement learning across online and offline settings.

• Visiting Research Scholar, University of California, Los Angeles (2023.08 - 2023.12).

  • Host: Prof. Quanquan Gu.

  • Studied reinforcement learning algorithms robust to adversarial corruption in online and offline decision-making.

Professional Service

• Conference reviewer: ICML, NeurIPS, ICLR, AISTATS.

• Journal reviewer: JMLR, Machine Learning, Artificial Intelligence.

Honors and Awards

• Gold Prize for Outstanding Student Scholarship (1/40), 2020.09.

• Bronze Prize for Outstanding Student Scholarship, 2019.09 and 2018.09.

Skills

• Programming and ML: Python, PyTorch, C++.

• Developer tools: Git, Docker.