Definition
Plain language
Taking a snapshot of what an AI is thinking at one internal stage and dropping it into a different stage of the same run to see if the answer improves.
As stated in the literature
An interpretability intervention that relocates a representation from one layer to another within the same forward pass, used when no correct 'clean run' exists for standard activation patching; demonstrates causally that needed knowledge already resides in the network but is misrouted.
Also called: self-patch
Why it matters: It matters because it can prove a model already contains the knowledge it needs but is simply routing it to the wrong place, pointing to fixes that don't require teaching it anything new.
For example, researchers can copy the model's internal state from an early processing stage into a later one during the same answer and check whether the final response becomes correct.
Heard on the show
“Their move is called self-patching.”Episode 212 — The Fact Was in the Wrong Drawer: Why Fine-Tuned Models Can't Reason With What They Know