Definition
Plain language
Reaching inside a model and swapping a piece of its internal state to see whether that piece is actually causing a behavior.
As stated in the literature
An umbrella term for interventions like activation patching and path patching that substitute internal model state from one run into another to establish causal rather than correlational claims about components.
Why it matters: It lets interpretability researchers move beyond 'this neuron lights up' correlations to actually testing which internal pieces cause a model's behavior.
For example, a researcher copies the hidden state from a run where the model correctly named Paris and pastes it into a run about Berlin to see if the answer flips.