My research focuses on LLM self-explanations: the natural language explanations LLMs give to justify their own decision-making. I measure whether these explanations are faithful to a model's true internal reasoning, and I develop new training objectives to incentivise faithfulness. I'm motivated by AI safety and think monitoring verbalised reasoning is one of the most promising short-to-medium-term safety methods. I also work on safety and capability evals, including the LingOly and LingOly-TOO reasoning benchmarks, and the Measuring what Matters review paper. My work has been published at NeurIPS, ICML, and ICLR. I'm supervised by Prof. Adam Mahdi (OxRML) and Prof. Jakob Foerster (BOLD).

Previously, I was an Astra Fellow with Owain Evans (Truthful AI) and completed the SPAR programme with Noah Siegel (Google DeepMind).

I mainly write about AI safety, explainability, and evals.

I'm a final-year PhD researcher. I've had an unusual path, having originally studied economics.

I've held several teaching positions including TA-ing the Social Data Science MSc at Oxford and tutoring Stanford computer science students. Previous students have gone on to the CS Master's at Stanford and various PhD positions at Oxford.