Stich Group
Optimization and collaborative machine intelligence
We are the research group of Sebastian U. Stich at the CISPA Helmholtz Center for Information Security in Saarbrücken and St. Ingbert, Germany.
We develop foundations and algorithms for modern machine learning and optimization, with a focus on training and updating models efficiently and reliably at scale. Our work spans modern optimization methods (including decentralized and non-convex optimization), communication-efficient and privacy-aware learning (federated and decentralized training), and learning under heterogeneity and distribution shift, including continual learning, personalization, and unlearning. We also study how models can share knowledge across sites and even across architectures to enable long-running, collaborative learning systems.
Methodologically, we connect mathematical understanding with algorithm design and careful experiments. We especially like problems where real constraints like limited communication, privacy, and changing data lead to clean research questions and practical impact.
Upcoming talks
No talks scheduled right now — check back soon.
news
| Sep 30, 2026 | Our group gathered at Parkhotel Weiskirchen from September 28–30 for our group retreat. It was a chance to exchange research ideas, discuss ongoing work, and spend time together outside the office.  |
|---|---|
| Sep 27, 2026 | Three papers from our group and collaborators have been accepted at NeurIPS 2026: Forgetting Has Neighbors: Localized Collateral Forgetting in Machine Unlearning, Efficient Gradient Methods for Distributed Saddle Problems, and What's in a Smoothness Constant? Tighter Rates for Local SGD with Bounded Second-order Heterogeneity. |
| Jul 05, 2026 | This week, we are presenting two papers at the International Conference on Machine Learning (ICML) in Seoul, Korea. In Enhancing LLM Training via Spectral Clipping [CODE], we propose SPECTRA, a simple wrapper that consistently improves coordinate-wise optimizers (AdamW, Signum, etc.), leading to better convergence and generalization. In Improved Convergence Analysis of Topology Dependence in Decentralized SGD, we develop a tighter convergence analysis of decentralized SGD, offering a more precise understanding of how network topology affects its convergence rate. |
| May 08, 2026 | We were honored to take part in the ELLIIT Symposium in Lund. ELLIIT’s focus period is a fantastic initiative, launched in 2017 with its first program on Large-Scale and Distributed Optimization, and this year centers on Optimization for Learning. We are especially grateful that Yuan Gao was invited to stay on to participate in the focus period and collaborate with other visiting scholars. |
| Apr 23, 2026 | Xiaowen Jiang presents oral paper at ICLR 2026 |
