I am a machine learning researcher with experience working on the entire LLM stack, ranging from pretraining with novel optimizers, over model merging, and post-training to test-time scaling and applications like LLMs in education. A lot of my work has focused on showing that LLMs can be improved by using uncertainty without increasing cost by learning weight uncertainty during training and using it in downstream applications and decision-making.
I have just defended my ELLIS PhD supervised by Prof. Iryna Gurevych at UKP Lab, TU Darmstadt and Prof. Mrinmaya Sachan at LRE, ETH Zürich and am now working with Emtiyaz Khan at TU Darmstadt while preparing my final submission.
Previously, I have received a M.Sc. in Data Science from RWTH Aachen University and worked as a student reasearcher at the Chair for Human Language Technology and Pattern Recognition of Prof. Hermann Ney.
Currently, my research interests include the following topics:
- Improving LLMs using uncertainty without increasing cost
- Pretraining with novel optimizers
- Model merging
- Test-time scaling
- LLM post-training & Exploration in post-training
- Novel architectures for efficient & adaptive computation
- LLM applications in education