Junteng Liu
Third-year Ph.D. student · HKUST NLP Group
I am advised by Professor Junxian He at the HKUST NLP Group. I received my B.Eng. in Automation from Shanghai Jiao Tong University (IEEE Honor Class) in 2024.
My research focuses on recursive self-improvement (RSI), coding agents, and long-horizon AI development. I study how verifiable task design and reinforcement learning can improve agents for inference-system development, logical reasoning, and tool use.
At Apple MLR, I developed and publicly released RSI-Inference (first author), a benchmark of 37 real-world tasks for coding agents that develop and maintain vLLM and SGLang on H100 and B200 GPUs.
Previously, I worked on logical reasoning RL and search-agent RL at MiniMax, as a core contributor to MiniMax-M1 and MiniMax-M2.
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?
News
| Oct 2026 | RSI-Inference is now public: 37 real-world tasks for long-horizon coding agents across model support, inference optimization, and bug fixing. Code. |
|---|---|
| Apr 2026 | Started a Research Scientist internship at Apple MLR in Cupertino, mentored by Yizhe Zhang and Ruixiang Zhang. |
| Oct 2025 | Our paper The Tool Decathlon was released, benchmarking language agents for diverse and long-horizon task execution. |
| Sep 2025 | Our paper WebExplorer was released on arXiv for long-horizon web-agent training. |
| May 2025 | SynLogic was accepted to NeurIPS 2025. |