Skip to content
AI360Xpert
Glossary
Definition

Activation Patching

An interpretability technique that proves a network component handles a specific concept by pasting its clean activations into a corrupted forward pass and w...

Think of It Like This

Like bypassing a broken house fuse with a battery to prove that specific fuse controls the kitchen lights.

This causal intervention is critical for mechanistic interpretability. Probing only shows correlation, but overwriting a live activation proves that the overwritten component was the causal source of the behavior. By sweeping this patch across layers, researchers can locate the exact circuits responsible for a specific concept.