I'm a PhD student at the Technical University of Munich, supervised by Prof. Daniel Rückert and Prof. Georgios Kaissis. My research spans the full LLM stack, from pretraining and reinforcement learning to efficient inference and practical applications.
My current research is on looped, depth-recurrent transformers and on benchmarking LLM agents in clinical decision workflows. I've previously published on differentially private learning and hyperbolic neural networks. Best Graduate Award (M.Sc. Applied Computer Science, 2024), NeurIPS 2026 Spotlight, NeurIPS 2025 Top Reviewer.
We introduce the first efficient end-to-end inference method for depth-adaptive looped language models, which are taking less compute for "easy" tokens and more for "hard" tokens, translating per-token compute savings into practical inference speedups through continuous depth batching.
We measure the capacity of parameters of looped LMs with a scaling law exponent 𝜑 and use Δ𝜑 as a diagnostic: truncated BPTT lowers it (worse loop), hyperconnections raise it (real gain), even though both improve validation loss. Model designers should make each loop worth its compute.
We release MIMIC-Note-Extract (MNE), a structured clinical dataset extracted from 300,000+ MIMIC discharge summaries using schema-guided LLMs. We use MNE to build a clinical decision-making benchmark for LLM agents.
Diffusion models, used as image priors, can reconstruct training data from differentially private models, exposing a gap between DP's theoretical bounds and what actually leaks in practice.
First framework for combining active learning with DP-SGD in standard supervised settings. We introduce step amplification to fix privacy budget waste and show which acquisition functions still work under DP, and which fall apart.
HCNN: a CNN that lives entirely in hyperbolic space. We generalize the convolutional layer, batchnorm, and the classifier head to the Lorentz model, so the network leverages hyperbolic geometry at every layer instead of just the head.