DERI Seminar with David Mguni, Lecturer in Machine Learning at the EECS
When: Thursday, May 21, 2026, 11:00 AM - 12:00 PM
Where: Teams
Speaker: Dr David Mguni, Lecturer in Machine Learning at QMUL
Title: The Boundary Problem in AI: Can Machines Discover Their Own Limits?
Abstract: Large language models can generate increasingly sophisticated reasoning and decision-making behaviour, yet they remain surprisingly poor judges of when and why they will fail. This limitation becomes especially important as general-purpose AI systems are deployed in scientific, legal, and other high-stakes settings, where hidden failure modes may be amplified through interaction, automation, and scale.
In recent theoretical work, we argued that some limitations of large AI systems are not merely engineering artefacts, but reflect deeper capability constraints associated with uncertainty, interaction, and bounded knowledge. If failure is unavoidable in sufficiently complex environments, then static benchmarking alone may be insufficient for understanding the true boundaries of AI capability.
In this talk, I will present a framework we have recently developed that reframes AI self-evaluation as an active scientific process. The framework adopts a split-entity philosophy in which an AI system divides into complementary reasoning and critical roles: one focused on solving tasks, the other on probing for weaknesses, uncertainty boundaries, and hidden failure modes through adaptive experimentation. We use game-theoretic ideas to formalise this process of self-discovery and adversarial introspection, framing capability discovery as a dynamic process of hypothesis formation, targeted probing, evidence gathering, and refinement.
More broadly, the work explores whether future AI systems may require explicit mechanisms for modelling the boundaries of their own competence: constructing internal representations of where their reasoning is reliable, where it becomes fragile, and how those boundaries evolve through interaction and experience. The talk will focus on the conceptual foundations and empirical observations underlying this framework, with the aim of opening discussion around whether future AI systems may need not only the ability to solve problems, but also the ability to discover and represent the limits of their own reasoning.
Bio: David Mguni is a Lecturer in Machine Learning at the School of Electronic Engineering and Computer Science at Queen Mary University of London, and a member of the Centre for Multimodal AI. His research lies at the intersection of ML, game theory, and statistical learning theory, with a particular focus on multi-agent systems, reinforcement learning, and the reliability of general-purpose AI systems. Prior to academia, he led research teams in industry working on multi-agent learning systems.
________________________________________________________________________________
Microsoft Teams meeting
Join: https://teams.microsoft.com/meet/340185578341617?p=Hak9hlw6E5cgLy90yU
Meeting ID: 340 185 578 341 617
Passcode: ez3La9Df
For organizers: Meeting options