Definition
Plain language
Changing one specific fact stored inside an AI without retraining the whole thing.
As stated in the literature
Techniques that directly modify a model's weights to update or insert a targeted fact; known to often fail to propagate the change to multi-hop or downstream queries that depend on the edited fact.
Why it matters: It matters because retraining a large model for every small fact change is slow and expensive, but an edit that doesn't carry through to related questions can leave the model quietly inconsistent.
For example, if a company changes its CEO, knowledge editing lets you swap in the new name without retraining the entire model from scratch.
Heard on the show
“The full annotated version is on paperdive dot AI — every term tap-to-define, with links to the related work on knowledge editing and interpretability, grouped by theme.”Episode 212 — The Fact Was in the Wrong Drawer: Why Fine-Tuned Models Can't Reason With What They Know