Mechanistic interpretability

Mechanistic interpretability investigates neural networks by identifying internal components, representations, and computations that causally produce particular observed behaviors.