All terms
Safety & Alignment
Interpretability
The study of a model's internal workings to explain why it behaves the way it does.
Definition
Interpretability is the study of a model's internal workings, aiming to explain why it behaves as it does. It ranges from post-hoc methods like attention visualization to mechanistic interpretability, which dissects individual neurons and circuits to find human-readable features and algorithms inside the weights. Better understanding supports debugging, trust, and safety audits.