Yanjin He

Yanjin He

Peking University

I’m Yanjin He (何彦瑾), a 4th-year undergraduate student at School of Mathematical Sciences, Peking University. I visited the University of Notre Dame in the summer of 2024 (fortunate to be supervised by Meng Jiang), and I visited UIUC in the summer of 2025 (fortunate to be supervised by Tong Zhang).

My research interests focus on generative models (especially diffusion models) and large language models (LLMs). I have a soild foundation in mathematics, particularly in probability and statistics, which has provided strong theoretical support for my research.

I am applying for a USA PhD (2026 Fall)

  • GPA: 93.1/100, Major core GPA: 94.1/100
  • Top 1, Data Science and Big Data Technology major, School of Mathematical Sciences, Peking University
  • Selected Courses:
    • Mathematical Analysis (I) (96), Geometry (100), Advanced Algebra (I) (99), Abstract Algebra (96), Mathematical Statistics (98)
    • Foundations of Machine Learning (97), Data Structure and Algorithm (96), Introduction to Computation (97), Introduction to Finance (96)
  • If you are interested, please feel free to contact me. Thank you for the opportunity.

🔥 News

  • 2025.11:  ✈️ Presented our poster at EMNLP 2025 in Suzhou.
  • 2025.09:  🎉 1 Paper (as first author) accepted by EMNLP 2025 （Main Conference).
  • 2025.09:  🏅 Named as Merit Student.
  • 2025.09:  🏅 Awarded the Leo KoGuan Scholarship.
  • 2025.06:  ✈️ Going to UIUC for summer internship.

📝 Publications

  • Pre-trained Models Perform the Best When Token Distributions Follow Zipf’s Law
    Yanjin He, Qingkai Zeng, Meng Jiang
    EMNLP 2025（Main Conference)
    [pdf] [arxiv]

🏅 Talent Programs

  • Elite Undergraduate Program in Applied Mathematics and Statistics (应用数学及统计拔尖人才计划), Peking University, 2024
  • Mathematical Talent Program, Peking University, 2022

🏅 Honors and Awards

  • Merit Student, Peking University, 2025
  • Leo KoGuan Scholarship, Peking University, 2025
  • Merit Student, Peking University, 2024
  • Ding Shisun-Gui Linlin Academic Scholarship (丁石孙-桂琳琳优秀学生奖学金), Peking University, 2024
  • WizardQuant Scholarship, Peking University (1 awardee in the School), 2024
  • First Prize, 16th National College Student Mathematics Competition, Beijing, 2024
  • Panasonic Scholarship, Peking University (5 awardees in the University), 2023
  • Outstanding Study Award, Peking University, 2023
  • First Prize, 15th National College Student Mathematics Competition, Beijing, 2023

    Selected High School Awards

  • Gold Medalist, Chinese Mathematical Olympiad (CMO Finals), 2021
  • Outstanding Graduate, Affiliated High School of South China Normal University, 2021
  • Top 10 School Stars, Affiliated High School of South China Normal University, 2020

📖 Education

  • 2022.09 - 2026.07, Undergraduate Student, School of Mathematical Sciences, Peking University
  • 2021.09 - 2022.07, Senior High School, Affiliated High School of South China Normal University
  • 2018.09 - 2021.07, Junior High School, Affiliated High School of South China Normal University

🔬 Research Experiences

Analyze ODE based diffuison under Fokker-Plank Equation

  • Supervisor: Tong Zhang (UIUC)
  • Doing theoretical analysis of the gap of ODE-based diffusion and SDE-based diffusion under the Fokker-Plank Equation assumption, set up a theoretical bound of the Wasserstein-2 distance of the distribution lead by ODE/SDE diffusion.
  • Analysing the convergence rate of ODE based diffusion, getting a good bound for ODE-based diffuison by controlling the divergence error and the Fokker-Plank error of score function, which could motivate us a better training objective for diffusion model.

Investigate the Impact of Tokenizers and decide the optimal tokenizer for pre-trained models

  • Supervisor: Meng Jiang (University of Notre Dame)
  • Doing extensive experiments across NLP, genomics, and chemistry demonstrate that models consistently achieve peak performance when the token distribution closely adheres to Zipf’s law, which inspire us a criterion for deciding the tokenizer.
  • Proposed a method for selecting the most suitable tokenizer for different datasets prior to pre-training and fine-tuning process. Experiments over several domains has proven the robustness nad effectiveness of our method.
  • Pre-trained Models Perform the Best When Token Distributions Follow Zipf’s Law, EMNLP 2025（Main Conference)

Inference-Time Guidance of ODE-Based Diffusion Models via Langevin Dynamics

  • Supervisors: Liwei Wang, Di He (Peking University)
  • Combining the Langevin process with the energy-based distribution lead by KL-regularized Reinforcement Learning Objective, we derive a inference time refinement for diffusion model to increase the expected reward of generated images under a specfic differentiable reward function. This method could not only be used to increase the image quality, but also guide the diffusion model toward generating images of specific categories.

🔗 Links

(Alphabetical Order)