Changing one specific fact inside a model without retraining it. Useful when a person’s job title changes or a figure is wrong, and much harder than it sounds, because facts are not stored in one place and editing one thing tends to disturb others.
Technical view
Methods locate factual associations, often in feed-forward layers, and apply targeted rank-one or low-rank updates. Evaluation covers reliability on the edited fact, generality across paraphrases, and locality, meaning unrelated knowledge is preserved. Sequential edits degrade the model, and edited facts often fail to propagate to consequences, so the model may accept the new fact while still asserting things that contradict it. Retrieval remains the more reliable route for changeable knowledge.
Related
Part of AI Terminology Index, section: Training and Post-training.