Agentic generation of web-backend security benchmark tasks with functional tests, contrastive exploits, and an executable benchmark release.
Hey! I'm a Data Science Master's student at ETH Zurich; currently visiting the University of Cambridge.
I'm interested in machine learning, security, and language models. I've worked on LLMs + security at the ETH SRI Lab and at UBS, pretraining and memorization dynamics with the ETH AI Center, and applied machine learning across oncology and the geospatial domain.
I enjoy hybrid athletics, playing the guitar, discussing ethics; and have plenty of mini-obsessions such as chess, geoguessr, fractals, and many other things.
I'm always happy to discuss interesting ideas---please reach out!
tobiasvonarx [at] proton [dot] me
CV,
LinkedIn,
GitHub, and
X.
Research
Controlled 100B+ token pretraining experiments, studying FIM vs. left-to-right memorization through probabilistic extraction and attention patterns.
Devloped novel autoencoder and variational autoencoder approaches for cell-type deconvolution from bulk RNA-seq data.