Andreas Madsen

Founding Research Scientist at Guide Labs, developing interpretable foundation models. My research focuses on enabling LLMs to introspect and avoid hallucinations in their reasoning and explanations via post-training, continuing my Ph.D. research at Mila from 2020 to 2024.

I am also technical lead for our internal training framework, our post-training pipeline, and our service platform, Clarity, which combines custom inference, agentic tool use, and interpretability.

Before my Ph.D., I was an independent researcher, publishing in Distill and at ICLR. My work has received spotlights at ICML and ICLR, and I have given invited talks and participated in interviews at major venues. I also have extensive open-source experience: my packages were downloaded 243 million times in 2025, and I was a steering committee member at Node.js.