Probability is the default language for uncertainty in machine learning, but it is a demanding one: it asks for a full distribution even when the evidence is less precise. Outer probability measures relax that demand, they keep a measure-theoretic framework while letting ignorance be expressed directly, rather than being forced into a prior probability. Where probability integrates, possibility theory takes suprema, so the operations of summarising, updating and propagating a belief become optimisation problems instead of often-intractable integrals. I will only assume knowledge of the familiar aleatoric/epistemic distinction, and then show how possibility theory can be implemented in general via a maxitive analogue of the Donsker-Varadhan formula that recasts inference as optimisation, leading to a drop-in replacement for the cross-entropy loss that gives deep classifiers a sense of when to abstain.

About Jeremie

Jeremie Houssineau is an Assistant Professor in the Division of Mathematical Sciences at the Nanyang Technological University as well as an honorary Associate Professor at the University of Warwick. His research interests include possibility theory and Bayesian statistics. He received a Ph.D. in statistical signal processing from Heriot-Watt University, Edinburgh, in 2015, was with the Department of Statistics at the National University of Singapore from 2016 to 2018, and was an Assistant Professor in the Department of Statistics at the University of Warwick from 2019 to 2023. 

AI systems must act on queries that are often underspecified, ambiguous, or missing information necessary to respond well. Yet today’s language models are typically trained on static data and lack principled mechanisms for recognizing what they don’t know and acquiring it through interaction. In this talk, I will present a line of work on building language models that resolve this incompleteness by asking the right questions. I’ll begin with work that trains models to ask targeted follow-up questions that efficiently infer users’ latent preferences, outperforming static prompting and survey-style approaches. I’ll then describe follow-up work that formulates preference elicitation in a Bayesian decision-theoretic framework, equipping language models with an explicit Bayesian user model. Finally, I’ll introduce QuestBench, a benchmark for evaluating interactive question-asking behavior in controlled environments. Together, these results outline a path toward language models that recognize the limits of their own knowledge and resolve it through natural interaction, enabling more reliable, personalized, and collaborative AI systems.

About Belinda

Belinda Z. Li is a researcher at Anthropic and an incoming Assistant Professor at the University of Chicago. Belinda recently received a PhD from MIT, where Belinda focused on building AI systems that have coherent and interpretable models of the world, the user, and themselves. Belinda has developed methods for understanding LMs' implicit world models, eliciting user models through proactive question-asking, and training LMs to faithfully explain their internal computations. Belinda's work aims to make AI systems more transparent, reliable, and amenable to human collaboration. Belinda is a recipient of a Rising Stars in EECS Award, a Clare Boothe Luce Fellowship, and an NDSEG Fellowship.

TBD :)

About Fazl

Fazl Barez is a Senior Researcher and Principle Investigator at the University of Oxford, where he leads the Technical Safety & Governance Lab and serve as Technical Director of the AI Governance initiative. His research spans mechanistic interpretability, AI safety, and technical governance --- focused on understanding the internal workings of neural networks and using that understanding to make AI systems more reliable and auditable. He teaches Oxford AI Safety and Alignment course and is Principal Scientist at Martian. His work is supported by OpenAI, Anthropic, Schmidt Sciences, and NVIDIA.

TBD!

About Dylan Hadfield-Menel

Dylan Hadfield-Menel is an Associate Professor of EECS at MIT. Dylan runs the Algorithmic Alignment Group in the Computer Science and Artificial Intelligence Laboratory (CSAIL). Dylan's research develops methods to ensure that AI systems behavior aligns with the goals and values of their human users and society as a whole, a concept known as 'AI alignment'. Dylan's group work to address alignment challenges in multi-agent systems, human-AI teams, and societal oversight of machine learning. The goal is to enable the safe, beneficial, and trustworthy deployment of AI in real-world settings.