Yikai Wu

My name is Yikai Wu (pronounced ee-keye uu; Chinese: 吴溢恺). I am a fifth-year Ph.D. candidate in Computer Science at Princeton University, advised by Prof. Sanjeev Arora.

Research interests: machine learning and algorithms, particularly large language model (LLM) reasoning, interpretability, continual learning, LLM agents for program optimization, and AI for combinatorial optimization.

Research brief: My current work studies how language models acquire, represent, and combine problem-solving skills, and how this understanding can guide better training and evaluation. My research includes:

• Mathematical reasoning and training (ongoing). I am developing synthetic math data with controlled combinations of skills and investigating how targeted supervised fine-tuning and reinforcement learning affect reasoning capabilities.
• Skill representations and continual learning. I extract interpretable skill representations from model gradients and use them to guide fine-tuning, with the goal of learning new skills while preserving existing capabilities.
• LLM agents for program optimization. I contributed to AlgoTune, a benchmark for evaluating whether language-model agents can speed up numerical programs while preserving correctness.
• AI for combinatorial optimization. I benchmarked AI methods for Maximum Independent Set against strong classical solvers and analyzed why some methods resemble greedy heuristics.

Before Princeton, I was a Ph.D. student at Yale University (2021-2022), where I worked with Prof. Nisheeth Vishnoi on differentially private matrix optimization. I received my B.S. summa cum laude from Duke University in 2021, double majoring in Computer Science and Mathematics with a minor in Physics. At Duke, I worked with Prof. Rong Ge on deep learning theory and Prof. Ashwin Machanavajjhala on differential privacy.