Uncertainty Quantification for LLMs

Talk-Titel: Uncertainty Quantification for LLMs

Speaker: Dr. Lynton Ardizzone

Abstract: LLM outputs look the same whether they're correct or not. For applications where being wrong has consequences, this is a core problem. Classical uncertainty quantification doesn't straightforwardly apply when outputs are free-form text.
This talk works through a progression of approaches, from token-level scores to semantic consistency across samples, each expanding the scope of what's being estimated. Each comes with trade-offs, and none fully solve the problem.
In agentic settings, uncertainty estimates become control signals for decision-making under incomplete information, connecting LLMs to classical decision theory. Getting UQ right is what separates principled agents from vibes-based ones.

Bio: Lynton Ardizzone holds a PhD in machine learning from Heidelberg University, where he developed deep generative models with built-in uncertainty quantification. His thesis on conditional invertible neural networks was awarded the Ruprecht-Karls-Prize. His methods have found applications in astrophysics, medical imaging, particle physics, and mechanical engineering. From 2022 to 2025, he was Head of Machine Learning at Copresence, an AI startup for 3D avatar reconstruction. He now works as an independent researcher, helping startups and smaller teams assess feasibility, prototype, and make technical decisions around AI.

We are with IMBIT this time and meet in the Nexus Lab. We have the Meetup account thinks to Prior Labs. Averbis provides beers and bezels after the talk, and we invite you to stay for a chat.
Thanks for the support from Prior Labs, Averbis, and IMBIT!