©BAIOSPHERE
Lecture
From Concepts to Certificates: Towards Controllable Generative Models
Invited Talk by Francesco Leofante on Making Generative Models Safer and More Controllable
20.11.2026
2:15 pm - 6:00 pm
Oettingenstraße 67, Room 151, Munich
Generative models such as text-to-image diffusion models are used at scale, but their behavior remains hard to control. They can amplify stereotypes, produce explicit or protected content, and resist attempts to remove such behaviors. Current safeguards mostly filter inputs and outputs, which leaves the model’s internal causes untouched and can be bypassed.
In this talk, Francesco Leofante proposes a pipeline that extracts what a model has learned, uses it to guide targeted edits, and verifies that those edits hold, with concept unlearning as the running example. He will introduce concept graphs, which show how a model links concepts in its text encoder and denoiser, and use them to reduce bias and make unlearning more robust. He will also show why low attack success rates can be misleading, present a framework that certifies how much of an unlearned concept can still leak, and introduce eval-unlearn, an open-source library for comparing unlearning methods.
Francesco Leofante is an Assistant Professor in the Department of Computing at Imperial College London, where he leads the CRUMBS lab and co-directs the Centre for Explainable AI. His research combines learning and reasoning to build rigorous, robust, and useful explanations of AI systems.
Everyone is welcome, and there will be time for questions and discussion after the talk.