Director of the Computational Linguistics Postgraduate Programme · UCL PaLS
Member of the European Laboratory for Learning and Intelligent Systems · UCL ELLIS Unit
My research explores the information processing principles that underly the ability to use and learn language — both in human and artificial language processing systems. I'm also increasingly interested in biological and artificial cognitive systems more broadly, and by extension to extra-linguistic aspects of perception, action, and interaction. I am committed to advancing the science of AI evaluation, with a focus on language as well as broader aspects of agency and safety. In 2018, I co-authored a paper that, according to the very kind Aaron Mueller, introduced the first causal (mechanistic?) interpretability method for language models.
- 2025Senior Research Scientist, UK AI Security Institute, Science of Evaluation & Testing Teams
- 2023–25Postdoctoral Fellow, ETH Zurich, Department of Computer Science, working with Ryan Cotterell
- 2019–23PhD, University of Amsterdam, Institute for Logic, Language and Computation, advised by Raquel Fernández
Two papers were accepted at ICML 2026: A Behavioural and Representational Evaluation of Goal-Directedness in Language Model Agents, which combines behavioural evaluation with probing analyses of internal representations to study how LLM agents represent and pursue goals, and Skewed Score: A statistical framework to assess autograders, which introduces a Bayesian statistical framework for measuring autograder bias, disagreement, and uncertainty in LLM evaluation.
Two papers were accepted at ACL 2026. One paper, Probing for Reading Times, studies whether language model representations contain signatures of human reading times. The other, Surprisal Minimisation over Goal-directed Alternatives Predicts Production Choice in Dialogue, uses a new approach to generating open-ended goal-agnostic and goal-oriented alternatives to study speaker choice sensitivity to various information-theoretic notions of comprehension and production cost. It is coming out very soon and was selected as an oral!
We have announced the call for papers for the LM Playschool Workshop and Challenge. We invite submissions exploring the frontier of language agents that learn, adapt, and improve through situated interaction, with a focus on conversational, collaborative, goal-oriented, and multi-turn environments.
New work with colleagues from ETH AI Centre and Institute of Neuroinformatics. We introduce an active probabilistic reasoning task, inspired by cognitive neuroscience, that isolates two core elements of decision-making under uncertainty -- sampling and inference -- and use it to evaluate LLMs at scale. We find that reasoning aligns the cognitive profiles of LLMs with those of humans.
Work with collaborators from Edinburgh on extending information-theoretic models of language production to visually grounded settings was accepted at EACL 2026.