CV

Contact Information

Name Jiarui (Jerry) Yan
Professional Title MS in Computational Data Science, Carnegie Mellon University
Email jerryy2@cs.cmu.edu

Professional Summary

Master’s student at Carnegie Mellon working on LLM pre-training data, RL post-training, and long-horizon agents. Software engineer intern at Google in summer 2026.

Education

  • 2025 - 2026

    Pittsburgh, PA

    MS
    Carnegie Mellon University, School of Computer Science
    Computational Data Science
    • Large Language Models, Advanced NLP, Multimodal Machine Learning, Deep Learning, Cloud Computing, Distributed Systems
  • 2021 - 2025

    Seattle, WA

    BS
    University of Washington, Information School
    Informatics: Data Science (major), Applied Mathematics (minor)
    • Magna Cum Laude (top 3.5%); Dean’s List every term.
    • Coursework: Machine Learning, NLP, Numerical Analysis, Probability I & II, Linear Algebra, Methods in Data Science.

Research (Project Lead)

  • ReScraper: Pre-training Data Curation (01/2026 – present) — Graduate Researcher with Prof. Chenyan Xiong (CMU LTI); co-first author, under review. One 0.6B model (Qwen3-0.6B) replaces the heuristic HTML scraper and rule-based filter stack: it extracts a page’s main content, then keeps, edits, deletes or rewrites it, trained by two-stage SFT on labels from three teachers. Pretraining 400M, 1.4B and 2.8B models on its output gives the best DCLM Core at every scale, +3.8–4.7% relative over the strongest baseline. Paper, code, model and 7.44B-token corpus released.
  • TraceML: Long-Horizon Auto-Research Agents (01/2026 – present) — Graduate Researcher with Prof. Yiming Yang (CMU LTI); co-first author, NeurIPS 2026 Evaluations & Datasets Track. Captures how ML solutions are developed, not just where they end up: every scored code version of 4,465 human and 207 agent Kaggle trajectories, with each edit labeled by what it did and why. Agents loop where human experts explore: Codex keeps re-weighting one ensemble, MLEvolve keeps mutating one model, and only human pivots pay off. A ~1K-token planning prompt written from these gaps lifts scores on 5 of 7 competitions with no regressions; the data and an open-source toolkit are released.

Research (Collaborations)

  • Predicting RL Utility Across Scale (05/2026 – 09/2026) — Graduate Researcher with Prof. Sanmi Koyejo (Stanford SAIL); under review. Which environments should an LLM learn from with RL, and does that choice hold as models scale: predicting how much RL post-training will improve a model from measurable properties of the environment, with the RL algorithm held fixed across math, coding and tool-use tasks.
  • Agentic Systems (01/2026 – 05/2026) — Graduate Researcher with Prof. William W. Cohen (CMU MLD, sponsored by Google DeepMind); co-first author on one paper, both under review. Built a modular agentic-LLM framework around pseudo-tools and ran a study across 19 evaluation suites showing static workflows beat dynamic planners (0.64 vs 0.56 average correctness at half the cost); learned reusable agent subroutines from experience.

Earlier Research

  • University of Washington (2023 – 2024) — transfer-learning matrix completion for multi-source recommendation with Prof. Wanning Chen (RMSE 0.6 → 0.35 on NetEase data); large-scale GitHub / Hugging Face mining of fork behavior in open source with Prof. Jane Tan and Prof. Yong Tan (first author, POMS 2025).

Teaching

  • Teaching Assistant, University of Washington (2023 – 2024) — INFO 498 Text Mining & Analytics and INFO 330 Databases & Data Modeling with Prof. Lucy Lu Wang; designed NLP assignments and led weekly SQL recitations for 35 students.

Skills

Languages: Python, C++, Java, TypeScript, JavaScript, SQL, R, Shell
ML / LLM: PyTorch, vLLM, Verl, Hugging Face, TensorFlow, Pydantic-AI, LiteLLM, LangChain, CUDA
Data & Infra: Spark, Hadoop, HDFS, Kafka, Airflow, AWS, GCP, Docker, Kubernetes