Kilian Merkelbach

I'm moving from applied LLM research into frontier AI safety. Through Constellation's Astra Fellowship, I'm working with Marius Hobbhahn (Apollo Research) on automated red-teaming for a scheming monitor, and I've trained deliberative monitors for black-box scheming detection. I care about whether we can reliably tell when a model is scheming. My transition into AI safety is supported by Coefficient Giving.

This builds on almost a decade in ML: a PhD at RWTH Aachen on making sense of complex temporal data using autoencoders, and fine-tuning and evaluating LLMs on eBay's search team. If you're working on scheming, evaluations, or monitoring, let's talk.