Automated machine learning is evolving from the optimization of individual model parameters towards the coordination of entire experimentation processes. How AI agents can select tools, plan individual steps and support the development of new models was at the heart of the sixth ELLIS Summer School on AutoML at TU Dortmund University. The five-day event, held from 27 to 31 July 2026, was organized by the ELLIS Unit NRW with the involvement of the Lamarr Institute. Lamarr Principal Investigators Matthias Feurer and Katharina Eggensperger served as two of the main organizers. The program combined a focus on Agentic AutoML with other current research topics, including neural architecture search for edge AI.
AutoML: From Optimization to AI Agents
Lectures and hands-on sessions covered both fundamental concepts and current research approaches. Topics included Bayesian optimization and hyperparameter optimization, preference optimization, novel foundation models for tabular data, scaling laws and the scientific evaluation of agentic AutoML systems. AutoML refers to methods that automate recurring decisions in the development of AI models. Agentic approaches extend this principle. They can not only optimize individual settings but also connect multiple tools and tasks into more comprehensive experimentation processes, thereby automating additional stages of the workflow.
Architecture Search for Resource-Constrained Devices
Another part of the program focused on neural architecture search. In this field, automated methods examine different neural network configurations to identify a suitable model architecture for a particular task. A hands-on session demonstrated how these methods can be applied. A lecture on neural architecture search for edge AI also showed how models can be developed for smartphones, sensors and microcontrollers with limited memory and computing capacity. In doing so, the Summer School also addressed questions of resource-aware machine learning, one of the Lamarr Institute’s core research areas.
ELLIS Unit NRW
Approved and established in May 2026, the ELLIS Unit NRW connects eight founding institutions, including TU Dortmund University, and builds on existing regional AI structures, including the Lamarr Institute. With Lamarr PI Katharina Eggensperger serving as Co-Director, the Dortmund location is also represented in the Unit’s leadership. Joint research activities, ELLIS summer schools and the training of early-career researchers are among the new Unit’s planned activities. At the same time, the Summer School demonstrated how scientific exchange extends beyond the structures of the Unit: Bernd Bischl, Director of the Munich Center for Machine Learning (MCML), is a member of the Summer School series’ organizing team and spent several days in Dortmund as a speaker. As the first summer school organized by the ELLIS Unit NRW, the event brought together participants and speakers in Dortmund, including representatives of partner institutions in Bonn, Cologne and Bochum.