Rongyi Yu

Rongyi Yu

Rongyi Yu is a Master student at Peking University.

I am Rongyi Yu (余荣毅), a Master student at Peking University, Center for Machine Learning Research. I am currently involved in research and engineering on large model training and alignment, with a focus on data attribution, dynamic data selection, and training efficiency optimization for Pretrain / SFT / RLHF (PPO) settings.

My recent work centers on data-centric AI and AI infrastructure, aiming to improve the alignment quality, robustness, and cost efficiency of large language models. I am comfortable with the full research pipeline from algorithm design to engineering implementation, including training, evaluation, reproduction, and performance diagnosis.

My current and previous research interests include:

  • Data Attribution
  • Data-Centric AI
  • LLM Training and Alignment
  • Dynamic Data Selection
  • AI Infrastructure
  • Diffusion Models

🔥 News

  • 2026:  🎉 I will join Peking University as a master’s student at the Center for Machine Learning Research.
  • 2026:  📝 Our paper “LongVidSearch: An Agentic Multi-hop Search Benchmark for Long Videos” is under submission to SIGIR 2026.
  • 2025:  🚀 Participated in research and engineering for DCAI-DataFlex, an open-source data-centric toolkit and dynamic training framework for LLMs.
  • 2025:  🏆 Received multiple honors including National First Prize in the Chinese Mathematics Competition and Top 10 Students at Harbin Institute of Technology.

📝 Publications

  • Rongyi Yu, Chenyuan Duan, et al.
    LongVidSearch: An Agentic Multi-hop Search Benchmark for Long Videos
    SIGIR 2026 Submission

🎖 Honors and Awards

  • 2025 National First Prize, Chinese Mathematics Competition (Mathematics Category)
    Highest-level national honor for mathematics undergraduates.
  • 2025 Top 10 Students, Harbin Institute of Technology
    Highest individual student honor at HIT; only 10 students selected across all campuses from freshman to senior year.
  • 2022, 2024 National Scholarship ×2
    Awarded to only about 2% of undergraduate students each year.
  • 2025 Bronze Medal, China International College Students’ Innovation Competition

📖 Education

  • 2026.09 - 2028.06 Peking University, Center for Machine Learning Research, M.S.
    I will join Peking University in Fall 2026 for my master’s study, co-supervised by Prof. Weinan E and Prof. Wentao Zhang.
  • 2022.09 - 2026.06 Harbin Institute of Technology, School of Mathematics, B.S.
    GPA: 99.53/100, Rank: 1/45
    CET-4: 585, CET-6: 571

🔬 Research Experience

  • 2025.09 - 2026.02 Research Assistant, Peking University, Center for Machine Learning Research
    Advisors: Prof. Wentao Zhang, Dr. Hao Liang
    • Worked on data-centric LLM training and parameter-efficient fine-tuning (PEFT).
    • Studied data quality, data selection, and training stability for supervised fine-tuning.
    • Participated in the development of the DataFlex dynamic training framework.
    • Supported the full pipeline of data scoring / sampling / training / evaluation to improve generation quality and training efficiency.
  • 2024.03 - 2025.08 Research Assistant, Harbin Institute of Technology & HKUST School of Mathematics
    • Collaborated closely with Prof. Yao Li and Prof. Yang Xiang.
    • Conducted research on diffusion-model-based theory and implementation for satellite image alignment and fusion.

💻 Selected Projects

DCAI-DataFlex

Open-source data-centric toolkit / dynamic training framework for large models

  • Problem: In SFT/RLHF training, noisy, redundant, and distribution-shifted data can lead to unstable generation quality and high training cost.
  • Method: As a major contributor, I helped build the dynamic training framework, designed multiple offline domain data selection operators, and developed domain-mixing strategies compatible with Qwen, Llama, and related training pipelines.
  • Result: Verified the framework’s effectiveness in SFT training: using only about 5% of the training data can achieve near full-data MMLU accuracy.
  • Impact: The method can be extended to pretraining, reinforcement learning, and hard-example mining, improving alignment quality and system stability.

🛠 Skills

  • Programming & ML: Python, PyTorch, Transformers, vLLM, DeepSpeed, LLaMA-Factory
  • Mathematics: Probability and Statistics, Stochastic Processes, Optimization, Numerical Analysis
  • Tools: Git, LaTeX, Markdown, Linux