Zijian Zhang

PhD student in Electrical and Computer Engineering, University of Minnesota

prof_pic.jpg

University of Minnesota

Minneapolis, MN, USA

I am a PhD student in the Department of Electrical and Computer Engineering at the University of Minnesota, advised by Prof. Mingyi Hong. I joined in Fall 2025 after receiving my B.S. in Computer Science from East China University of Science and Technology. Before that, I was a research assistant in Prof. Huaxiu Yao’s lab at UNC Chapel Hill, the Machine Learning Group at Microsoft Research Asia, and the InternLM2 team at Shanghai AI Lab.

My research focuses on self-improving LLM agents via reinforcement learning:

  • Agentic RL and multi-agent systems for real-world coding tasks such as GPU kernel optimization.
  • Recursive self-improvement of the RL post-training loop itself, where the system refines its own scaffolds, rewards, data, and evaluation to keep pushing the capability frontier.

news

Nov 19, 2026 New blog posts on AutoScaffold and Is One Layer Enough?
Nov 10, 2026 New preprint: AutoScaffold, an AI-for-AI framework that turns a student model’s RL training experience into adaptive scaffolds.
Jul 01, 2026 New preprint: Is One Layer Enough? Training a single transformer layer can match full-parameter RL training.
May 08, 2026 New preprint: CUDAHercules, a benchmark for hardware-aware, expert-level CUDA optimization with LLMs.
Apr 30, 2026 StitchCUDA is accepted to ICML 2026.
Mar 03, 2026 New preprint: StitchCUDA, a multi-agent framework for end-to-end GPU programming with rubric-based agentic RL.
Feb 2026 Gave an invited talk at Amazon: LLM for Kernel: Automated Agentic GPU Programming Framework — From Single Kernel to E2E Programs.
Jan 31, 2026 GRAPE is accepted to ICRA 2026.

selected publications

  1. Preprint
    AutoScaffold: AI-Guided Scaffolding for Reinforcement Learning of Language Models
    Zijian Zhang, Rizhen Hu, Jiangweizhi Peng, Shiyang Li, Sujay Bhatt, Sihan Zeng, and Mingyi Hong
    arXiv preprint arXiv:2611.14515, 2026
    Under review
  2. Preprint
    Is One Layer Enough? Training A Single Transformer Layer Can Match Full-Parameter RL Training
    Zijian Zhang, Rizhen Hu, Athanasios Glentis, Dawei Li, Chung-Yiu Yau, Hongzhou Lin, and Mingyi Hong
    arXiv preprint arXiv:2607.01232, 2026
    Under review
  3. ICML
    StitchCUDA: An Automated Multi-Agent End-to-End GPU Programming Framework with Rubric-based Agentic Reinforcement Learning
    Shiyang Li*, Zijian Zhang*, Winson Chen, Yuebo Luo, Mingyi Hong, and Caiwen Ding
    In International Conference on Machine Learning (ICML), 2026
  4. Preprint
    CudaForge: An Agent Framework with Hardware Feedback for CUDA Kernel Optimization
    Zijian Zhang, Rong Wang, Shiyang Li, Yuebo Luo, Mingyi Hong, and Caiwen Ding
    arXiv preprint arXiv:2511.01884, 2025
    Under review
  5. NeurIPS
    InfantAgent-Next: A Multimodal Generalist Agent for Automated Computer Interaction
    Bin Lei, Weitai Kang, Zijian Zhang, Winson Chen, Xi Xie, Shan Zuo, Mimi Xie, Ali Payani, Mingyi Hong, Yan Yan, and Caiwen Ding
    In Advances in Neural Information Processing Systems (NeurIPS), 2025
  6. ICRA
    GRAPE: Generalizing Robot Policy via Preference Alignment
    Zijian Zhang, Kaiyuan Zheng, Zhaorun Chen, Joel Jang, Yi Li, Chaoqi Wang, Mingyu Ding, Dieter Fox, and Huaxiu Yao
    In IEEE International Conference on Robotics and Automation (ICRA), 2026
    Also presented at an ICLR 2025 workshop
  7. Preprint
    CUDAHercules: Benchmarking Hardware-Aware Expert-level CUDA Optimization for LLMs
    Shiyang Li, Zijian Zhang, Guangyan Sun, Yuebo Luo, Winson Chen, Yanzhi Wang, Mingyi Hong, and Caiwen Ding
    arXiv preprint arXiv:2605.08467, 2026
    Under review
  8. ICLR
    Anyprefer: An Agentic Framework for Preference Data Synthesis
    Yiyang Zhou, Zhaoyang Wang, Tianle Wang, Shangyu Xing, Peng Xia, Bo Li, Kaiyuan Zheng, Zijian Zhang, Zhaorun Chen, Wenhao Zheng, Xuchao Zhang, Chetan Bansal, Weitong Zhang, Ying Wei, Mohit Bansal, and Huaxiu Yao
    In International Conference on Learning Representations (ICLR), 2025