Christophe Roux
I am the research area lead for iol.OML (Optimization and Learning) at the Zuse Institute Berlin (ZIB). I defended my PhD at the Institute of Mathematics, TU Berlin in June 2026, under the supervision of Sebastian Pokutta and David Martínez-Rubio. I am also a member of the Berlin Mathematical School (BMS).
My research focuses on algorithms and theory for large-scale continuous optimization, with an emphasis on first-order methods for problems in modern machine learning. Specifically, I have worked on first order methods on Riemannian manifolds, online optimization, min-max formulations, Frank-Wolfe and federated learning. I have also recently started working on efficiency in deep learning, applying optimization techniques in compression/pruning methods for LLMs.
- New preprint on arXiv: The KV-Surgeon: Compensation-Aware Offline KV-Cache Compaction
- Two papers accepted at NeurIPS 2026:
- New preprint on arXiv: LinearPFN: Amortized Variable Selection for Linear Models with Interactions
- I successfully defended my PhD thesis First-Order Methods for Geodesically Convex Optimization on Riemannian Manifolds with distinction! Many thanks to my committee: Sebastian Pokutta, David Martínez-Rubio, Martin Skutella, Suvrit Sra, and Tobias Breiten.
- New preprint on arXiv: Restart and Adaptive Acceleration in Stochastic Gradient Methods