I work on the models underneath intelligent agents: how they are pretrained, and how they act in the physical world.

I am a B.S. candidate in Computer Science at the Institute for Interdisciplinary Information Sciences, Tsinghua University. Since July 2026 I have been doing pretraining research at 九坤至知研究院 (Ubiquant), where I study hyper-connections: replacing the single residual stream of a Transformer with several learnable, interacting streams.

Before that I worked on embodied intelligence: Vision-Language-Action foundation models and flow-matching control policies at Galaxea, and LLM-based embodied agents and multi-agent simulation with Prof. Chuang Gan at UMass Amherst.

  • LLM Pretraining
  • Hyper-Connections
  • Embodied AI
  • Vision-Language-Action Models
  • LLM Agents
  • Flow-Matching Policies

Experience

  1. LLM Pretraining ResearchJul 2026 – Present
    九坤至知研究院 · Ubiquant

    Pretraining research on hyper-connections: widening the residual stream into multiple streams with learned connections between them.

  2. VLA Algorithm InternNov 2025 – May 2026
    Galaxea Inc.

    Algorithmic improvements for Vision-Language-Action foundation models, focusing on robust continuous action generation. Developed GeCO, a time-unconditional flow-matching framework for adaptive and robust robotic control.

  3. Embodied AI InternFeb 2025 – Aug 2025
    UMass Amherst · with Prof. Chuang Gan

    LLM-based embodied agents and interactive multi-agent simulation. Contributed to Ella and Virtual Community.

  4. B.S. in Computer Science2022 – 2027
    Institute for Interdisciplinary Information Sciences, Tsinghua University

    GPA 3.73 / 4.0. Coursework: Deep Learning, Deep Reinforcement Learning, Algorithm Design, Probability and Statistics, AI: Principles and Technology.

Selected Papers

All papers →
  • NeurIPS 2026

    Generative Control as Optimization: Time Unconditional Flow Matching for Adaptive and Robust Robotic Control

    Zunzhe Zhang*, Runhan Huang*, Yicheng Liu, Shaoting Zhu, Linzhan Mou, Hang Zhao
    * equal contribution
    arXivProject
  • ICLR 2026

    Virtual Community: An Open World for Humans, Robots, and Society

    Qinhong Zhou, Hongxin Zhang, Xiangye Lin, Zheyuan Zhang, Yutian Chen, Wenjun Liu, Zunzhe Zhang, Sunli Chen, Lixing Fang, Qiushi Lyu, Xinyu Sun, Jincheng Yang, Zeyuan Wang, Bao Chi Dang, Zhehuan Chen, Daksha Ladia, Quang Vinh Dang, Jiageng Liu, Chuang Gan
    arXivProjectCode
  • ICML 2025 MAS Workshop

    Ella: Embodied Social Agents with Lifelong Memory

    Hongxin Zhang, Zheyuan Zhang, Zeyuan Wang, Zunzhe Zhang, Lixing Fang, Qinhong Zhou, Chuang Gan
    arXivProjectCode

Writing

All posts →
13 篇

大模型规模化训练

从显存账本、Roofline 到 5D 并行、FlashAttention、FP8 和推理，逐层拆开。

 6 篇

从零训练一个 LLM

消融方法、架构、优化器与数据、训练马拉松、后训练和基础设施。

 5 篇

模型技术报告精读

DeepSeek-V4、Qwen3-Next、Kimi K2 / Linear / K3，每个新组件从原理讲到细节。

Awards

  • Gold Medal, National Olympiad in Informatics (NOI), China2021
  • International Gold Medal, Asia-Pacific Informatics Olympiad (APIO)2021
  • Gold Medal, CCPC Guilin Regional Contest2022
  • School Scholarship, Tsinghua University2023
  • Freshman Scholarship, Tsinghua University2022