A mechanistic interpretability technique that identifies which parts of a neural network are responsible for specific outputs.