Keyu Li Keyu Li 李克钰

elsewhere

Long-horizon agents, from pre-training to post-training to evaluation.

Keyu Li (李克钰), second-year Ph.D. student at Shanghai Jiao Tong University and Shanghai Innovation Institute, advised by Dequan Wang and Pengfei Liu. Working on the full LLM pipeline (continual pre-training & mid-training, post-training, agentic RL) and the data flywheels that drive it, plus multi-agent systems and frontier AI evaluation.

Now a research intern at Xiaohongshu (ACE Top Talent Program), owning data pipelines and post-training for visual coding: web, image, 3D and game generation. Previously on the Tencent WorkBuddy team, building the data loop, CPT / SFT and agentic RL for coding agents, and at GAIR Lab, where I led AgencyBench and DatasetResearch. Heavy daily user of Claude Code and Codex.

CV · English (PDF) → 简历 · 中文 (PDF) → chlorophyll@sjtu.edu.cn WeChat · 13370856505

News

  • Oct 2026Joined Xiaohongshu as a research intern (ACE Top Talent Program), working on data pipelines and post-training for visual coding, after wrapping up my internship on the Tencent WorkBuddy team.
  • May 2026Joined the Tencent WorkBuddy team as a research intern, working on LLM CPT / SFT and agentic RL for coding agents.
  • Apr 2026AgencyBench accepted to the ACL 2026 Main Conference.
  • Mar 2026Two papers accepted to ICLR 2026: Aligned Agents, Biased Swarm and InnovatorBench.
  • Sep 2025Released LIMI: Less is More for Agency.
  • Aug 2025Released DatasetResearch, a benchmark for demand-driven dataset discovery.

Publications Google Scholar →

Experience

  • Xiaohongshu · Research Intern, ACE Top Talent Program 2026.10 – now

    Data pipelines and model post-training for visual coding tasks spanning web, image, 3D and game generation.

  • Tencent WorkBuddy Team · Research Intern 2026.05 – 2026.10

    Built a coding-agent data loop from 6.75M production trajectories (3.0M training candidates, 1.5M traceable SFT examples); continued pre-training and SFT of a 30B-A3B model on 60B tokens with 128 / 256 H200s, +5–10% on SWE-bench Verified and WorkBuddy Bench; online RL and on-policy distillation with veRL / GRPO.

  • Shanghai Innovation Institute · GAIR Lab · Research Intern 2025.05 – 2026.05

    Post-training and evaluation of LLM agents with Prof. Pengfei Liu: AgencyBench, DatasetResearch, LIMI.

  • Soochow Securities · Fintech · Quant Research Intern 2024.06 – 2024.09

    Deep-learning factor mining on Level-2 market data (CNN / RNN / Transformer / KAN).

Education

  • Shanghai Jiao Tong University · Ph.D., Computer Science and Technology 2024 – 2029

    Joint program with Shanghai Innovation Institute. Advisors: Dequan Wang, Pengfei Liu.

  • Shanghai Jiao Tong University · B.S., Mathematics and Applied Mathematics 2020 – 2024

    GPA 4.03 / 4.3, rank 4 / 45. Shanghai Outstanding Graduate, Guo Chenchen Scholarship. Thesis: Towards Visualizing and Understanding Bayesian Flow Networks.

Visitors details →

Thanks for stopping by. Readers so far, by country:

Visitors by country