As foundation models grow more complex, it becomes harder to understand how they reason or whether their decisions align with human goals. In biomedicine, this raises a dual challenge: how to use these systems effectively, and how to understand what drives their performance.

We tackle both. We design modular, agentic systems that let researchers interact with AI as collaborators, not black boxes. At the same time, we study model internals, testing how reasoning unfolds and which components are most critical. Our benchmarks focus not just on accuracy, but on why models work. Finally, it is often equally important to know when to not use a large, black-box model, and rely on alternatives instead. To achieve this, we build on a strong network of interdisciplinary collaborators based at Helmholtz Munich and its wider ecosystem. Our PI Sebastian Lobentanzer leads both the Accessible Biomedical AI Research Lab and the Computational Biology Unit at the German Center for Diabetes Research (DZD), and we work closely with Helmholtz AI, ELIXIR Germany, and the Open Targets Platform.