cv

Research experience, education, publications, honors, and technical skills.

Contact Information

Name Yikai Wu
Professional Title Ph.D. Candidate in Computer Science
Email yikai.wu@cs.princeton.edu
Website https://yikai-wu.github.io/

Professional Summary

Ph.D. candidate at Princeton University, advised by Prof. Sanjeev Arora. Research focuses on mathematical reasoning in large language models, synthetic data, gradient-based skill representations, and continual learning, with earlier work in combinatorial optimization and differential privacy.

Experience

  • 2026 - present
    Synthetic Math Data Generation
    Princeton University
    Advisor: Prof. Sanjeev Arora
    • Develop pipelines to generate, solve, check, and refine difficult math questions with controlled skill combinations.
    • Annotate mathematical skills and filter questions using answer consistency across teacher models.
  • 2025 - present
    Gradient Skill Decomposition for Continual Learning
    Princeton University
    Advisor: Prof. Sanjeev Arora
    • Combine layer-wise factor estimation and supervised skill recovery to extract interpretable skill vectors from per-example LoRA-B gradients.
    • Use recovered skill vectors to project fine-tuning gradients away from protected skill subspaces and reduce forgetting.
  • 2025 - 2026
    Skill-Based Training for Mathematical Reasoning
    Princeton University
    Advisor: Prof. Sanjeev Arora
    • Use stronger language models to diagnose missing mathematical skills and construct targeted training data.
    • Study skill-targeted supervised fine-tuning and reinforcement learning; targeted RL did not consistently outperform standard RL.
  • 2025 - 2025
    Language-Model Agents for Numerical Program Optimization
    Princeton University
    Mentor: Dr. Ofir Press
    • Contributed to AlgoTune, a benchmark for numerical programming with correctness checks and runtime comparisons against reference implementations.
    • Evaluated AlgoTuner, an agent that edits, compiles, profiles, verifies, and selects optimized code.
    • Published at NeurIPS 2025, Datasets and Benchmarks Track.
  • 2025 - 2025
    Step-by-Step Chain-of-Thought Reasoning Analysis
    Princeton University
    Advisor: Prof. Sanjeev Arora
    • Studied reasoning decomposed into iterative steps across multiple forward passes to investigate reliance on memorized patterns.
    • Analyzed reasoning robustness and failure modes including context loss, step fragmentation, and premature termination.
  • 2023 - 2025
    Benchmarking AI for Combinatorial Optimization
    Princeton University
    Advisor: Prof. Sanjeev Arora
    • Benchmarked generative and reinforcement learning methods for Maximum Independent Set against strong classical solvers.
    • Introduced serialization analysis to explain similarities between non-backtracking AI methods and greedy heuristics.
    • Published in Transactions on Machine Learning Research, April 2026.
  • 2021 - 2022
    Differentially Private Optimization on Unitary Orbits
    Yale University
    Mentor: Prof. Nisheeth Vishnoi
    • Designed efficient differentially private optimization algorithms on unitary orbits, with applications to PCA and rank-k matrix approximation.
    • Established upper and lower utility bounds; published at COLT 2022.
  • 2020 - 2021
    Hessian Structure in Neural Networks
    Duke University
    Mentor: Prof. Rong Ge
    • Explained common neural-network Hessian structures using Kronecker factorization, with applications to PAC-Bayes generalization bounds.
  • 2020 - 2021
    Differential Privacy for Multiple Analysts
    Duke University
    Mentor: Prof. Ashwin Machanavajjhala
    • Formulated criteria for differential privacy with multiple analysts and designed algorithms with provable guarantees.
    • Published at VLDB 2021, following a preliminary version at TPDP 2020.
  • 2019 - 2020
    Quantum Information and Algorithms
    Duke University
    Mentor: Prof. Iman Marvian
    • Investigated operators for ion-trap quantum computers and designed algorithms for quantum-state purification.
  • 2018 - 2019
    Differential Privacy for Summation Queries
    Duke University
    Mentor: Prof. Ashwin Machanavajjhala
    • Developed a data-dependent truncation algorithm for answering numerical summation queries under differential privacy.
    • Presented at TPDP 2019.

Education

  • 2022 - present

    Princeton, NJ, USA

    Ph.D. candidate
    Princeton University
    Computer Science
    • Advisor: Prof. Sanjeev Arora
  • 2021 - 2022

    New Haven, CT, USA

    Ph.D. studies
    Yale University
    Computer Science
  • 2017 - 2021

    Durham, NC, USA

    B.S., summa cum laude
    Duke University
    Computer Science and Mathematics; minor in Physics
    • GPA: 3.97/4.00
    • Phi Beta Kappa

Publications

Honors and Awards

  • 2020
    Outstanding Undergraduate Researcher Honorable Mention
    Computing Research Association (CRA)
  • 2019
    Meritorious Winner
    Mathematical Contest in Modeling (MCM)
  • 2018
    Karl Menger Award in Mathematics
    Duke University
  • 2017, 2018
    4th Place, Mid-Atlantic Regional
    International Collegiate Programming Contest (ICPC)
  • 2017
    National Rank 142.5; Top 4 at Duke
    William Lowell Putnam Mathematical Competition
  • 2016
    Gold Award; Individual Prize
    Singapore Mathematical Olympiad (Open)
  • 2015
    Gold Medal
    Singapore National Olympiad in Informatics (NOI)
  • 2014
    Bronze Medal
    Asia-Pacific Informatics Olympiad (APIO)

Skills

Programming: Python, C++, C, R, Mathematica
Frameworks: PyTorch, verl
Methods: Factor analysis, supervised fine-tuning (SFT), reinforcement learning with verifiable rewards (RLVR), continual learning, LLM evaluation, synthetic data generation