Jerry Liu
PhD student at Stanford ICME · Hazy Research · DOE CSGF Fellow · jerrywliu@stanford.edu
About
I’m a rising fifth-year PhD student in the Institute of Computational & Mathematical Engineering at Stanford, advised by Chris Ré. I am supported by the DOE Computational Science Graduate Fellowship. Previously, I completed my undergraduate degrees in Math and Computer Science at Duke, where I was advised by Cynthia Rudin. My path has been shaped by many kind and brilliant researchers, including Atri Rudra, Michael Mahoney, Jin Yao, and Kenny Weiss.
Research Interests
My recent work studies memory in language models: how knowledge is encoded in model weights and how architectural choices determine what models can remember and at what cost. I’m especially interested in the tradeoffs between learning in weights and learning in context, and in how architecture, data, and optimization interact to determine what models learn. My goal is to use these insights to build more parameter-, data-, and compute-efficient models.
I’m currently developing recurrent architectures for efficient in-context memory, with a focus on recall capacity and the interaction between architecture and training.
I’m also interested in machine learning for numerical tasks, especially in scientific settings such as differential equations. My previous work examined why standard architectures struggle with high-precision numerical computation, and developed methods to improve precision in PDEs and continuous-valued regression.
I’m currently exploring research roles beginning in late 2026 or early 2027. Please feel free to reach out by email!
News
| Apr 07, 2026 | Featured in a DEIXIS profile on my work at the intersection of machine learning, numerics, and scientific reasoning. |
|---|---|
| March 2026 | Constructing Machine-Precision Neural Networks with Quasi-Interpolants was selected for an oral presentation (8/113) at the AI&PDE Workshop at ICLR 2026. |
| Jun 30, 2025 | Presented BWLer at the Theory of AI for Scientific Computing (TASC) Workshop at COLT, where it received the Best Paper Award. |
Selected Publications
- MLPs are Hebbians: Constructing Efficient Fact-Storing MLPs for TransformersCOLM, 2026TL;DR: We show how MLPs can store facts as key-value mappings within Transformers, giving an explicit, information-theoretically optimal construction that improves on prior constructions by up to two orders of magnitude. Our analysis precisely characterizes how embedding geometry and query noise limit fact-storage capacity.
- BWLer: Barycentric Weight Layer Elucidates a Precision-Conditioning Tradeoff for PINNsIn Workshop on the Theory of AI for Scientific Computing @ COLT, 2025Best Paper AwardTL;DR: BWLer uses barycentric interpolation within physics-informed neural networks to decouple solution parameterization from derivative computation, improving accuracy by up to three orders of magnitude on benchmarks and revealing a precision-conditioning tradeoff.
- Towards Learning High-Precision Least Squares Algorithms with Sequence ModelsICLR, 2025TL;DR: We identify architectural and optimization bottlenecks that prevent standard Transformers from learning high-precision least-squares algorithms, then develop sequence models and a training recipe that learn gradient-descent iterates to near machine precision.