Šimon Váry

portrait Title: MATS Research Fellow
Location: Berkeley, California
Research: Optimization, learning theory, LLM inference
Email: simon.vary@stats.ox.ac.uk
Education:

 

News: New preprint on Masked Language Flow Models!
News: I will present our paper On-Average Stability of Multipass Preconditioned SGD and Effective Dimension at COLT 2026; slides here.
News: I taught a hands-on Introduction to LLMs at Oxford and Imperial College London.

I am a MATS Research Fellow working with Roger Grosse on data attribution methods for large language models. My broader research interests include statistical learning theory, optimization, interpretability, and efficient inference.

Previously, I was a postdoctoral researcher in Statistics at the University of Oxford, working with Prof. Patrick Rebeschini on optimization and statistical learning theory. Before that, I was a visiting lecturer and postdoctoral researcher at UCLouvain, hosted by Prof. Pierre-Antoine Absil and Prof. Julien Hendrickx. I completed my DPhil in Mathematics at Oxford under the supervision of Prof. Jared Tanner, where I worked on robust PCA and low-rank plus sparse matrix sensing (thesis).