Title
The evolving science of AI evaluation
Abstract
As AI foundation models produce increasingly open-ended outputs, participate in multi-turn, environment-grounded interactions, and act over extended trajectories, the object of AI evaluation is changing.
This talk examines the methodological shift required to evaluate AI systems under these conditions. I argue that evaluation should characterise the latent capabilities and decision processes that generate observed behaviour, allowing us to diagnose failures, distinguish between competing explanations, and predict performance under changes in task and interaction conditions. Doing so requires careful evaluation design, clearly established construct validity, calibrated measurement instruments, and causal tests linking higher-level variables, such as beliefs, goals, and plans, to observed behaviour. I also consider how these methods fit within a broader evaluation ecosystem, including the roles of expert judgement, continual evaluation, and independent third-party auditing throughout the AI lifecycle. Together, these developments can support a more explanatory and predictive science of AI evaluation.
Bio
Mario Giulianelli is an Associate Professor of Computational Linguistics at University College London, a Member of the European Laboratory for Learning and Intelligent Systems, and Co-founder and Scientific Director of Parallax. Previously, he was a Senior Research Scientist at the UK AI Security Institute and a Postdoctoral Fellow in the Department of Computer Science at ETH Zürich. His research lies at the interface of artificial intelligence, linguistics, and cognitive science and examines the information processing principles underlying human and artificial language processing systems. Mario is also committed to developing a rigorous science of AI evaluation that combines behavioural evidence with analysis of the internal mechanisms underlying the behaviour of AI systems. His work includes the first causal interpretability method for neural network language models and research on LLM judges, agent benchmarks, goal-directedness, and the safety evaluation of increasingly autonomous AI systems.
Speaker is presenting in person in Oslo.
Practical
The Tuesday Seminar Series covers a range of topics relevant to Integreat’s research focus. Presenters from within and beyond the Integreat community give 40-minute presentations, followed by group discussions.
Seminars are open.
For those unable to attend in person: https://uio.zoom.us/j/65632027408
Point of contacts:
- Organiser and moderator: Andrey Kutuzov, University of Oslo
- Integreat Administrative Leader Maria Dikova, University of Oslo