Sebastian Bordt

Hi there! I'm a postdoctoral researcher interested in large language models and interpretability. I work at the University of Tübingen with Ulrike von Luxburg.

Most of my current work is empirical language model research. I'm intersted in the learning dynamics of pre- and post-training, and in how academica can stay relevant in an age of ever increasing compute demands. In some of my recent work, we studied how pretraining hyperparameter optimization impacts performance after post-training (ICML'26), and proposed an approch that can make experimentation during pretraining more efficient (ICLR'26). I've also worked on data contamination (ICML'25), and contributed to theory research on learning dynamics (NeurIPS'25).

During my PhD, I worked on a variety of different topics in explainable machine learning. For example, I have worked on the connections between post-hoc methods and interpretable models, and on the suitability of explanation algorithms for regulation. I wrote up my perspective on explainable machine learning in this ICML'25 position paper.