Causal Activation Patching

A neural network interpretability technique that replaces the activations of a specific layer with activations obtained from a different input, directly testing whether that layer is causally responsible for the final model output.