CV

Download CV (PDF)

Curriculum vitae of Junteng Liu.

Contact Information

Name Junteng Liu
Professional Title Third-year Ph.D. Student
Email jliugi@connect.ust.hk

Professional Summary

Third-year Ph.D. student at HKUST NLP Group studying recursive self-improvement (RSI) and coding agents for long-horizon AI development, with verifiable feedback and reinforcement learning for logical reasoning, search, and inference systems.

Experience

  • 2026.04 - 2026.09

    Cupertino, USA

    Research Scientist Intern
    Apple MLR
    • Mentored by Yizhe Zhang and Ruixiang Zhang.
    • Developed and publicly released RSI-Inference (first author), a benchmark of 37 real-world tasks evaluating whether coding agents can autonomously develop and maintain vLLM and SGLang on H100 and B200 GPUs.
    • Day-0 model support (10 tasks): Built tasks for integrating previously unsupported models, evaluated through correctness and serving-feature checks.
    • Inference optimization (20 tasks): Built tasks requiring source-code changes to improve serving throughput under fixed deployment settings.
    • Bug fix (7 tasks): Built tasks requiring agents to identify and repair performance regressions.
    • Sonnet-5.5 and Opus-5.5 ranked first and second overall, achieving 1.68× and 1.48× geometric-mean speedups across the 27 optimization and bug-fix tasks, but only 26.7% and 13.3% pass@1 on Day-0 Model Support, respectively. DeepSeek-V4.1-Flash achieved 0.0% support pass@1 and 1.24× speedup.
  • 2025.02 - 2026.02

    Shanghai, China

    Research Intern
    MiniMax
    • Core contributor to MiniMax-M1 and MiniMax-M2.
    • Logical reasoning RL: Developed SynLogic (first author, NeurIPS 2025), synthesizing rule-verifiable data across 35 logical task types and training reasoning models. SynLogic-32B outperformed DeepSeek-R1-Distill-Qwen-32B on BBEH (25.5% vs. 19.2%). This data and RL work contributed to MiniMax-M1.
    • Search-agent RL: Developed WebExplorer (first author) and conducted RL training experiments for long-horizon web agents, using challenging search tasks synthesized through web exploration and query evolution. WebExplorer-8B surpassed WebSailor-72B on BrowseComp-en/zh (Avg@4: 15.7% vs. 12.0% / 32.0% vs. 30.1%). The data was used in MiniMax-M2 and formed the basis for subsequent MiniMax-M3 search-data improvements.
  • 2024.06 - 2024.09

    Shanghai, China

    Research Intern
    Tencent WXG
  • 2023.06 - 2023.12

    Shanghai, China

    Research Intern
    Shanghai AI Lab
    • Mentored by Prof. Yu Cheng.
    • Investigated a universal truthfulness hyperplane using probes trained on 40+ datasets; found that data diversity improves cross-dataset generalization. First-author paper: On the Universal Truthfulness Hyperplane Inside LLMs (EMNLP 2024).

Research Questions

  • How can we train agents to do ultra-long-horizon RSI or AI4AI development?
  • Beyond coding agents, what other domains can we explore?

Education

  • 2024 - Present

    Hong Kong

    Ph.D.
    Hong Kong University of Science and Technology
    Computer Science
    • HKUST NLP Group; advisor: Prof. Junxian He.
  • 2020 - 2024

    Shanghai, China

    B.Eng.
    Shanghai Jiao Tong University
    Automation (IEEE Honor Class)

Publications

Awards

  • 2020–2024
    Zhiyuan Honor Scholarship
    Shanghai Jiao Tong University

Skills

Research Topics (Advanced): Recursive Self-Improvement, Coding Agents, Long-Horizon AI Development, Verifiable Tasks, Reinforcement Learning, Inference Systems, Logical Reasoning, Tool Use

Languages

Chinese : Native
English : Fluent