I’m a PhD at the Tübingen AI Center, advised by Matthias Bethge and Maksym Andriushchenko. I work on understanding LLM capabilities and failure modes, with a focus on long horizon evaluation and safety.

My research asks how we can rigorously measure what LLMs can and can’t do, especially in extended autonomous settings. I am working on PostTrainBench, a benchmark measuring whether CLI agents can autonomously post-train LLMs. Most recently, at MATS with Mary Phuong (Google DeepMind), I trained model organisms for sandbagging: models that deliberately underperform on evaluations. Before that, at ARENA 6.0 I built a long horizon agent benchmark.

I also worked on mechanistic interpretability. At LASR Labs with Joseph Bloom (UK AISI), we discovered “feature absorption”—a systematic failure mode in Sparse Autoencoders. At Microsoft Research with Navin Goyal, I studied how harmful concepts transform during post-training.

A long time ago, I worked in computational neuroscience at the Max Planck Institute for Biological Cybernetics and EPFL, modeling how visual features map onto neural activity. I hold a dual degree in Computer Science and Biological Sciences from BITS Pilani.