Elias Kempf

Elias Kempf is an ELLIS PhD student in interpretability and AI safety at the University of Freiburg and Oxford, jointly supervised by Thomas Brox and Fazl Barez. His research studies how learned model behaviors emerge, generalize, and transfer, with a focus on safety-relevant failure modes. Recent work spans domain and compositional generalization in vision-language models, hidden-bias transfer via subliminal learning, and black-box model diffing for comparing language model behaviors. His master’s thesis on CLIP’s domain and compositional generalization received the Förderpreis Künstliche Intelligenz and led to a Spotlight paper at ICML 2025. He has also been a MATS Research Fellow, mentored by Arthur Conmy (prev. Google DeepMind), and previously worked on deep learning for radiotherapy dosimetry at PTW Freiburg.