Erle Zhu's Homepage
Erle Zhu

Erle Zhu | Allen

Chinese Name: 祝尔乐 (meaning: wish you happy)

PhD Student, Department of Computer Science and Technology, Tsinghua University

I am a PhD student at Tsinghua University and a member of the CoAI research group. My research focuses on building generalizable foundation agents that can perceive and interact with the world. I am currently working on agents in digital world, advancing CLI-Agent with Reinforcement Learning. I also worked on data-efficient RLVR (CROPI) and scientific agent (MAPS) before.

Email GitHub Google Scholar Twitter(X) Xiaohongshu

News

  • September 2026: I wrote a new blog post, Some Thoughts on the Development of AI.
  • August 2026: GLM 5.3, Z.ai's latest coding and cyber-defense model, has been released.
  • June 2026: Our blog on GLM 5.2 has been released. Check out our recipe for long-horizon RL.
  • May 2026: Our paper Skill-Evolver on online learning for CLI-Agent (with meta-skill) has been released on arXiv, checkout the paper.
  • April 2026: My original GitHub account, Lez-3f, was suspended due to an unknown issue😭. I'm using erlezhu02 as a temporary account.
  • April 2026: Our paper CROPI has been accepted to ACL 2026. See you in San Diego.
  • March 2026: We released PhD-Zero, a skill-based research agent that excels at AI research.

Experience

  • 2025.10–2026.8: Z.AI Coding Agent Team
    Led post-training for terminal-agent scenarios and contributed to AgentRL infrastructure.
  • 2025.2–2025.10: Z.AI RLHF Team
    Contributed to improving the instruction-following capabilities of GLM-4.5.
  • 2023.12–2025.10: Tsinghua AI4Edu Project, now AIcosmos
    Developed agents for programming and circuit design.
  • 2023.10–2024.10: Z.AI RLHF Team
    Researched LLM alignment algorithms, including reward design and DPO variants.
  • 2023.6–2023.7: Research Intern, Energy Internet Research Institute, Tsinghua University
    Conducted data science research on indoor energy consumption.
  • 2022.7–2023.4: Research Intern, LeapLab
    Advised by Gao Huang; conducted research on unsupervised reinforcement learning.
  • 2024–Present: Ph.D. Student in Computer Science, Tsinghua University
    Member of the CoAI research group.
  • 2020–2024: Undergraduate Student, Tsinghua University
    Majored in Mathematics and Physics Science & Electrical Engineering; Major Rank 2/30.

Publications

  1. SkillEvolver: Skill Learning as a Meta-Skill
    Genrui Zhang*, Erle Zhu*, Jinfeng Zhou, Caiyan Jia, Hongning Wang
    arXiv 2026 paper
  2. CROPI: Data-Efficient RLVR via Off-Policy Influence Guidance
    Erle Zhu*, Dazhi Jiang*, Yuan Wang, Xujun Li, Jiale Cheng, Yuxian Gu, Yilin Niu, Aohan Zeng, Jie Tang, Minlie Huang, Hongning Wang
    ACL 2026 (Main) paper code poster
  3. MAPS: Advancing Multi-Modal Reasoning in Expert-level Physical Science
    Erle Zhu, Yadi Liu, Zhe Zhang, Xujun Li, JinZhou, Xinjie Yu, Minlie Huang, Hongning Wang
    ICLR 2025 paper code
  4. SR: Augmenting Unsupervised Reinforcement Learning with Self-Reference
    Andrew Zhao*, Erle Zhu*, Rui Lu, Matthieu Gaetan Lin, Ning Jia, Yong-jin Liu, Gao Huang
    Neural Networks 2025 paper

Projects

  • GLM 5.3: contributed to optimizations for Terminal-Bench 3 and Agent's Last Exam.
  • GLM 5.2: led SFT and RL data construction for terminal coding scenarios; improved Terminal-Bench performance; and contributed to CLI-agent RL infrastructure and training verification. These efforts helped make GLM 5.2 the state-of-the-art open-source model at the time, with strong capabilities in general terminal use and code-repository reconstruction.
  • GLM5: contributed to agentic RL and data synthesis in terminal-agent scenarios.
  • GLM4.5 Series: worked on RLHF data pipeline and RL infrastructure for multi-task agentic RL.
  • Slime: contributed to asynchronous RL training infrastructure for agentic post-training.
  • THU-AI4EDU: led educational agent work for circuit theory and programming at Tsinghua University.

Talks

  • RL Data Selection and Scaling for Complex LLM Task Scenarios
    April 23, 2026, Generative AI Conference, Beijing - LLM RL Workshop
    Slides Photo

Rewards

  • Comprehensive-Merit Scholarship of Tsinghua University, 2021 and 2022
  • Technology-Innovation Scholarship of Tsinghua University, 2022
  • Academic-Merit Scholarship of Tsinghua University, 2021
  • Third Prize, 23rd THU-EDC Electronic Design Competition, 2021
  • First Prize, 13th Chinese Mathematics Competition for College Students, 2021

Services

  • Teaching Assistant: LLM Summer School 2024; Linear Algebra, Fall 2024; Introduction to AI, Spring 2026
  • Reviewer: NeurIPS 2024, NeurIPS 2025
  • Minister, Academic Department, Student Association of Science and Technology, Weiyang College, Tsinghua University