Interpretability and explainability
We study how trained models work inside, and what it would take to explain them. Much of the work is on language models: where in the model a behaviour comes from, and what changes there when the model is trained further.
The rest is about evidence. An explanation, a measurement of a model’s insides, or a model’s report on its own reasoning can be convincing without being true. We ask what each of them really shows, drawing on philosophy of science and on neuroscience, which faced the same questions about brains.
Themes
- Mechanistic interpretabilityWhere inside a model a behaviour comes from, and what training changes.
- Philosophy of machine learningWhen an explanation, a measurement or a model’s account of itself can be believed.
- EvaluationWhether a test, an audit or a safety claim checks what it says it checks.
Papers
2026
2026