The following schedule is a work in progress. Once we have the full seminar details including title, abstract and speaker bio, we will publish these talks under upcoming seminars.
| Date | Speaker | Title | Abstract |
|---|---|---|---|
| 2026-11-10 10:30 | Dr Russell Tsuchida | Uncertainty quantification in the age of foundation models | Uncertainty quantification — an important consideration in almost all fields of science, social science, and engineering — allows practitioners to make risk-aware decisions, know when to trust a prediction, improve explainability, and keep humans in the loop. Recently, foundation models and AI have promised to fundamentally change the landscape of uncertainty quantification research. In this talk, I will describe an evolving research program in uncertainty quantification, spanning work before and after the rise of foundation models. I will first present published work on squared families (the topic of my DECRA), sequential decision making, and Bregman divergences, with applications in point cloud alignment and label distribution learning. Bregman divergences form a natural mathematical language for evaluating and training uncertainty-aware models, including foundation models. Building on this, I will then look ahead to two directions: using tabular foundation models for uncertainty quantification, and using Bregman divergences to advance diffusion and related generative models.This seminar will include a mix of published results and a longer-term research vision for uncertainty quantification in the age of foundation models. |
| 2026-11-12 11:00 | |||
| 2026-11-19 11:00 | Paul Thomas | Evaluating Conversational Agents at Scale | Microsoft 365 Copilot helps millions of people around the world navigate their digital world and get things done. Measuring and improving Copilot’s quality relies on constant and in-depth evaluations, but these evaluations are made difficult by precisely the same features that make the product interesting: it is multimodal, multilingual, multiturn, serves a huge number and diversity of people, and has strong guarantees on data protection and privacy. I’ll outline some of the challenges of evaluating such a system at scale, and some of the techniques we've developed to help understand and improve the product. |
| 2026-11-26 11:00 | Chen Soon Ong | To score or not to score | Scoring the performance of a machine learning model is a key choice we all make, whether it is the loss for training a model, a gap for understanding alignment, or a metric such as accuracy for predictive performance. This talk considers the score we use in adaptive experimental design for scientific discovery, where we use bandit type approaches to choose experiments. I show some recent theoretical and practical work on adaptive design and discuss the needs of cross disciplinary collaboration for AI in science, for example to empirically validate bandit designs in a genomics lab. On the other side, machine learning researchers may be tempted to generalise the idea of optimising a score to other aspects of work and life. I argue that it is problematic to convert the diversity of human experience into a single scalar score, and then optimising that score. Finally I discuss some preliminary results on what could be in between scoring and not scoring. |
| 2026-11-30 11:00 | Bryon Aragam | ||
| 2026-12-03 11:00 | tbd | ||
| 2026-12-10 11:00 | Pushkar Gopalakrishna | Autonomous LLM agents can become trapped in repetitive execution loops, excessive tool calls, or recursive delegation while continuing to consume computational resources. Detecting these failures is challenging because legitimate long-running tasks can exhibit similar behavior. This talk presents an experimental investigation of resource governance and runaway-behavior detection in autonomous agents, drawing on approximately 16,500 simulated agent runs and offline analysis of over 2,200 real-agent trajectories from publicly available benchmarks. The presentation examines hierarchical budget enforcement, loop detection, circuit breakers, and the trade-offs between preventing runaway execution and preserving legitimate agent workflows. It also discusses failure cases where improvements to legitimate-task completion significantly increased the cost of stopping runaway agents, highlighting broader challenges in reliable agent evaluation. | |
| 2026-12-15 10:30 | MoonJeong and Saemi | ||
| 2026-12-17 11:00 | Alexander Soen |