Arik Reuter

Answering causal questions lies at the core of nearly all scientific disciplines, yet the estimation of causal quantities has traditionally relied on bespoke methods tailored to narrow sets of assumptions and problem settings. In contrast, recent advances in machine learning have demonstrated the power of foundation models: single, general-purpose models capable of solving a wide range of tasks, often matching or surpassing specialized approaches while yielding new conceptual insights. This PhD project aims to investigate how the foundation model paradigm can be extended to causal inference. The research will focus on developing training strategies for causal foundation models, incorporating causal assumptions in a principled manner, and providing theoretical analyses of their behavior. A central objective is to ensure that such models are reliable and scientifically meaningful, which requires rigorous treatment of uncertainty, robustness to misspecification, and interpretability.