MEND: low-rank gradient-transform networks for local model editing

What changed

Learns small shared editor networks on factorized per-example gradients, producing targeted weight changes while training against edit-neighborhood and locality losses without modifying the base model before an edit.

Scientific scope

Learns small shared editor networks on factorized per-example gradients, producing targeted weight changes while training against edit-neighborhood and locality losses without modifying the base model before an edit. Requires editor training and chosen editable layers, edit/locality distributions and objectives. Locality is measured on sampled unrelated inputs, not guaranteed everywhere. O(d) editor parameterization is not an all-resource constant-cost claim.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Requires editor training and chosen editable layers, edit/locality distributions and objectives. Locality is measured on sampled unrelated inputs, not guaranteed everywhere. O(d) editor parameterization is not an all-resource constant-cost claim.

Contributors

• Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.

• Fast Model Editing at Scale

Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction