An AI/ML group and ART-AI Seminar
We are pleased to have Alex Modell, who is a mathematician and computer scientist at Imperial College London, join us for this joint AI/ML Group and ART-AI Seminar entitled ‘An introduction to mechanistic interpretability of large language models’.
This seminar will take place in person in 1W 2.01, on Wednesday 17th June 2026, 15.15pm-16.05pm (BST). Edward Milsom will chair the seminar. There is also an option to join online. For more information, please e-mail [email protected].
Title
An introduction to mechanistic interpretability of large language models
Abstract
In this talk, Alex will introduce the new science of understanding the inner working of large language models. While humans designed the scaffolding of these systems, the mechanisms that allow them to intelligently reason and communicate are effectively “grown” through training, rather than built, and to date, remain a relative mystery. Alex will focus on the problem of understanding how neural networks geometrically represent information, and explaining some key hypotheses and discoveries through the lens of mathematics and statistics. Finally, there will be a discussion about some recent work on representation manifolds in large language models.
Bio
Alex is a mathematician and computer scientist at Imperial College London, where he currently holds a Chapman Research Fellowship. He is interested in studying the internals of AI systems, such as large language models, to understand how they represent information and reason about the world. The fundamental goal of his research is to ensure that the AI systems that we build and rely upon are safe and aligned with human values. Alex was previously a post-doc with Nick Heard on the NeST programme, and he completed his PhD at the University of Bristol, entitled “Spectral embedding of large graphs and dynamic networks”, supervised by Patrick Rubin-Delanchy.
