Definition
Plain language
Tracing which specific connections inside a model carry a particular behavior by intervening on them.
As stated in the literature
An interpretability method that intervenes on specific edges between components in a network to identify which causal pathways carry a behavior.
Why it matters: Pinpointing which internal pathways carry a behavior is what turns interpretability from suggestive into mechanistic.
For example, you swap the output of one attention head's edge to a downstream MLP and watch whether the model's answer changes.