Herbert Woisetschläger

I recently joined IBM Research Europe in Zurich, Switzerland, where I am part of the AI Foundations department. My work is centered on agentic memory, routing, and tooling for efficient LLM training and ops. The common thread for all my work is making agentic systems both more capable with fewer resources and more economical to operate.

Before moving to IBM Research, I completed a PhD in computer science at the Technical University of Munich, Germany, specializing in distributed and cost-efficient machine learning applications. Throughout my PhD, I was fortunate to collaborate with fantastic people on federated learning in resource-constrained environments, LLM pre-training, and cost optimal inference using LLM zoos. We managed to place our work in top-tier conferences like NeurIPS, ICML, ICLR, IJCAI, and ACM Middleware. I was advised by Prof. Hans-Arno Jacobsen (U. of Toronto, Canada & TUM) and mentored by Prof. Shiqiang Wang (U. of Exeter, UK). For a broader perspective on what matters in industry, I also spent multiple summers at the IBM T.J. Watson Research Center in the U.S., working on various aspects of pre- and post-training of IBM Granite models.

Before I started my PhD, I spent two years working as a management consultant for the German and Swiss branch of Capgemini Invent specializing in performance and cost optimization programs for consumer good retailers, financial institutions, and manufacturers. I also worked for Detecon, Inc. in Silicon Valley focusing on corporate startup incubation as well as accelerator programs for the public sector. This is where I built interest in optimizing and automating business processes through quantitative methods.

For my recent research contributions, please check out my Google Scholar profile or the list of working papers below. The working papers I post are typically under submission at the time they appear here. If you have comments or critique for them, please let me know. I am always happy to discuss.