Alex L. Zhang is a PhD student at MIT CSAIL researching how language models use tools, context and computation. His work includes Recursive Language Models and KernelBench, a benchmark that checks whether model-generated GPU kernels are both correct and fast.[1][3][5]
Start with Recursive Language Models →
Education
Links
Website
GitHub 
X LinkedIn Google Scholar PhD student, CSAIL · Massachusetts Institute of Technology
Open Research Resident · Laude Institute
Research Fellow · Prime Intellect
Career
At MIT, Zhang is advised by Omar Khattab and Tim Kraska. He is also an Open Research Resident with Laude and a research fellow at Prime Intellect. He completed undergraduate computer-science study at Princeton in 2024, where the department awarded him its Phillip Goldman Senior Prize for academic excellence. His current research asks how the computing environment around a model changes what it can accomplish.[1][2]
Contributions
Working with context through code
Recursive Language Models place a large input in a programming environment where a model can inspect parts of it and make further model calls. Instead of repeatedly asking the model to absorb an entire document in its prompt, the system lets it choose which pieces to examine and combine. This changes the inference procedure rather than requiring a new base model.[3]
Beginning tool work before generation ends
Zhang’s speculative programmatic tool-calling work then asks whether some expensive calls can begin while the surrounding code is still being generated, overlapping computation that would otherwise run in sequence.[4]
Testing speed alongside correctness
KernelBench evaluates a model’s ability to turn PyTorch workloads into GPU kernels, the low-level programs that perform computation on an accelerator. A generated kernel must first produce correct results, then exceed a chosen speedup threshold over the baseline. That metric prevents an incorrect or merely runnable answer from counting as an optimization. Zhang is a coauthor of the framework and paper; its published experiments use execution and profiling feedback to study where models improve and where difficult workloads remain.[5]
Tools, writing & talks
Writing, research & talks
Start here
Recursive Language Modelsalexzhang13.github.ioAn inference strategy that lets models inspect large inputs through a programming environment and recursively call models on selected material.
Start with the illustrated overview to understand why the prompt becomes data in a programming environment. Basic Python and model-API familiarity help with the code, but the opening explains the change in workflow.
Research coauthor[3]- Academia is for Ambition — Alex Zhang, MITlatent.space
Latent Space interviews MIT's Alex Zhang on Recursive Language Models, context offloading, subagent swarms, KernelBench and GPU Mode, and treating harnesses as compositional generalizers.
Catalog workOriginal source Alex Zhang, the MIT researcher behind Recursive Language Models, discusses harnesses as compositional generalizers, programmatic subagent calling, agent swarms, AI-written GPU kernels and capability overhang in current models.
Catalog workOriginal source- Speculative Programmatic Tool Callingalexzhang13.github.io
A harness technique for launching tool calls before code generation finishes, overlapping work to reduce waiting.
Author[4]Original source - KernelBencharxiv.org
A benchmark measuring whether generated GPU kernels are correct and faster than PyTorch baselines.
Paper and framework coauthor[5]