Yikai Wang

A Ph.D. Student at UNC

ykwang@unc.edu

Welcome to My Homepage!

Yikai selfie

I'm Yikai Wang, currently pursuing a Ph.D. degree in the Statistics and Operations Research Department (STOR) at the University of North Carolina at Chapel Hill (UNC).

I received my B.S. degree in Computer Science from Zhejiang University, China.

My research interests include large language models, reinforcement learning and optimization.

Outside research, I enjoy playing Weiqi, astronomy, history, movies, music, all kinds of sports, and the many curious facts and stories that make the world endlessly interesting. I also play the bamboo flute and guitar.

Get in Touch

Research

Why Agentic Design is Necessary for Data Analytics

Working paper

Calibrating Conditional Risk

Working paper

Risk Profiling and Modulation for LLMs

arXiv preprint (arXiv), 2025; INFORMS Annual Meeting talk, 2025

REMAST Real-time Emotion-based Music Arrangement with Soft Transition

IEEE Transactions on Affective Computing (IEEE TAFFC), 2025

SongDriver Real-time Music Accompaniment Generation without Logical Latency nor Exposure Bias

Proceedings of the 30th ACM International Conference on Multimedia (ACM MM), 2022
View All Research

Blogs

tmux Is Not a Shortcut Cheat Sheet. It Is a Terminal Workflow.

March 19, 2026
tmux becomes a tool you actually want to use every day only when sessions, windows, panes, reconnects, configuration, and common mistakes all fit into one mental model.

March 18 Group Meeting Notes: Agentic Vending, Pre-Training Data, and Constitutional Classifiers

March 19, 2026
Cleaned-up listener notes from a March 18 group meeting hosted by my professor, covering Project Vend 2, data poisoning, token-level filtering, duplication, replayed pre-training data, and Anthropic's next-generation constitutional classifiers.

verl, vLLM, and FlashAttention: How the Stack Actually Fits Together

March 17, 2026
A practical guide to what verl, vLLM, and FlashAttention each do, why they appear in the same post-training setup, and where their responsibilities actually differ.

GRPO and Its Variants: What Actually Changes, and Why It Matters

March 17, 2026
A long-form engineering guide to GRPO, Dr.GRPO, DAPO, BNPO, REINFORCE++, RLOO, and newer trainer variants, with an emphasis on normalization, length bias, and practical training choices for LLM RL.

PPO Is Not Just a Clip Trick

March 17, 2026
Why the practical success of PPO comes from the whole implementation stack rather than the clipping term alone.
View All Blogs

Teaching

  • STOR 113: Decision Models for Business and Economics

    Terms: 2023 Fall | Role: Teaching Assistant

    An introduction to multivariable quantitative models in economics. Mathematical techniques for formulating and solving optimization and equilibrium problems are developed, including elementary models under uncertainty.

  • STOR 455: Methods of Data Analysis

    Terms: 2024 Spring, 2024 Fall | Role: Teaching Assistant

    Review of basic inference; two-sample comparisons; correlation; introduction to matrices; simple and multiple regression (including significance tests, diagnostics, and variable selection); analysis of variance; use of statistical software. Honors version available.

  • STOR 155: Introduction to Data Models and Inference

    Terms: 2024 Summer, 2025 Summer | Role: Instructor

    Data analysis; correlation and regression; sampling and experimental design; basic probability (random variables, expected values, normal and binomial distributions); hypothesis testing and confidence intervals for means, proportions, and regression parameters; use of spreadsheet software. Honors version available.

  • STOR 120: Foundations of Statistics and Data Science

    Terms: 2025 Spring, 2025 Fall | Role: Lab Instructor

    The course teaches critical concepts and skills in computer programming and statistical inference, in conjunction with hands-on analysis of real-world datasets, including economic data, document collections, geographical data, and social networks. It delves into social issues surrounding data analysis such as privacy and design. Honors version available.