Definition
Plain language
A check for whether the parts of a model you flagged actually matter — you switch them off and see if behavior shifts more than switching off random parts would.
As stated in the literature
An interpretability metric measuring how much ablating a set of discovered components changes a model's prediction relative to ablating an equal number of random components; a positive gap indicates the discovered components are causally load-bearing rather than coincidental.
Why it matters: It separates the parts of a model that truly drive behavior from ones that merely look involved, keeping interpretability claims honest.
For example, after flagging a handful of neurons as responsible for a model's answer, a researcher switches them off and confirms the answer changes far more than switching off random neurons does.
Heard on the show
“And the way they verify they've actually found a circuit, not just noise, is something called the causal gap.”Episode 023 — Why a Small Agent Confidently Overwrites Memories It Doesn't Understand