Junwei Zhang

CS Undergraduate at UCAS | Exploring Recursive Self-Improvement for AI

my_personal_photo.jpg

Hi! I am a fourth-year undergraduate majoring in Computer Science at UCAS, where I am fortunate to be advised by Prof. Fei Sun.

My research interests lie in AI agents, particularly learning from experience and recursive self-improvement (RSI). My earlier work explored hallucinations and knowledge mechanisms in LLMs.

During my second year, I explored chip design through the ysyx project. That experience sparked a lasting interest in computing systems and infrastructure, including machine learning systems (MLSys) and high-performance computing (HPC).

Beyond academics, I enjoy classical music, tinkering with Linux, and discovering new tools and projects on GitHub.

Internships & Education

Internships

Meituan Inc.

Feb. 2026 – Present

Research Intern / RSI Team

Education

University of Chinese Academy of Sciences

2023 – 2027 (Expected)

B.S. in Computer Science
GPA 3.97/4.00 | Rank 1/86

Honors and Awards

  • National Scholarship (Ministry of Education of P.R. China)

    2024, 2025

  • Finalist, Interdisciplinary Contest in Modeling (Consortium for Mathematics and its Applications)

    2025

  • First Class Academic Scholarship (University of Chinese Academy of Sciences)

    2024, 2025

News

Our new preprint Strong yet Controllable LLM Abstention without Learning to Abstain is now available on arXiv! Oct 14, 2026
Our paper BaseCal: Unsupervised Confidence Calibration via Base Model Signals has been accepted to the ACL 2026 main conference! May 23, 2026
Our new preprint Beyond Reasoning: Reinforcement Learning Unlocks Parametric Knowledge in LLMs is now available on arXiv! May 10, 2026
This website is launched! 🎉 Dec 19, 2025

Selected Publications

  1. arXiv Preprint
    Strong yet Controllable LLM Abstention without Learning to Abstain
    Junwei Zhang*, Yifei Wang*, Wanli Yang, and 6 more authors (* Equal contribution)
    arXiv 2026
  2. arXiv Preprint
    Beyond Reasoning: Reinforcement Learning Unlocks Parametric Knowledge in LLMs
    Wanli Yang, Hongyu Zang, Junwei Zhang, and 5 more authors
    arXiv 2026
  3. ACL 2026
    BaseCal: Unsupervised Confidence Calibration via Base Model Signals
    Hexiang Tan, Wanli Yang, Junwei Zhang, and 7 more authors
    In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics, 2026