Xuechen Li

PhD Candidate, Stanford Computer Science
Stanford Artificial Intelligence Laboratory (SAIL)
Stanford Machine Learning Group
Stanford NLP Group

lxuechen [at] cs [dot] stanford [dot] edu

CV, Google Scholar, Github, Twitter, goodreads

Research

I'm a PhD candidate at Stanford CS. I go by Chen, the second segment of my first name.

My long-term research interest is to build safe and general problem-solving AI. I currently focus on data and evals for large models.

My research is supported by a Stanford Graduate Fellowship and a Meta PhD Fellowship in Security and Privacy.

Fun Past Projects

The Alpaca Series: I started the work that eventually evolved into the Alpaca Series (Stanford Alpaca, AlpacaFarm) in a joint effort with wonderful labmates. We showed that strong instruction-following language models can be created cheaply by fine-tuning pretrained models. In addition, we built a simulator to help researchers study RLHF. This line of work made people in the open community realize that they can build their own chatbots easily and cheaply.

Differentially Private Deep Learning: I did a body of work on DP deep learning which substantially improved the privacy-utility tradeoff and computational efficiency of methods [1, 2, 3, 4], in collaboration with friends in both academia and industry. We showed that with careful design and engineering, DP training results in deployable performance and its computational efficiency approaches standard training. This body of work led to the first deployment of DP machine learning at Microsoft. DP machine learning is now used in products such as Outlook and by internal teams at Microsoft, protecting users' privacy and providing substantial monetary savings.

Generative Modeling and Numerical Methods: I worked a lot on generative modeling before my PhD. In my last undergrad year, I co-developed a numerical scheme for computing gradients through a special type of differential equations, involving a combination of ideas from stochastic analysis, numerical analysis, and deep learning. As part of the project, I wrote the first version of the numerical solver codebase torchsde. Since then, this codebase has been used in projects like Stable Diffusion web UI. In addition, the PyPI package reached a million monthly downloads in 2023.

Selected Research Articles (full list see google scholar)