Knowledge editing in language models — updating a specific fact without retraining — requires choosing which layer to edit. Different facts are stored in different layers, so the natural expectation is that the optimal editing layer depends on the query. Each fact has its own location.
Datta, Liu, and Chhabra (arXiv:2602.20207) find the opposite. Certain “golden layers” achieve near-optimal editing performance regardless of the specific query. The optimal layer is a property of the model architecture, not of the fact being edited. A fixed set of layers generalizes across datasets, across editing techniques, and across input types.
The method for finding these layers — Layer Gradient Analysis — uses gradient attribution to identify which layers are most responsive to knowledge changes. The gradient pattern is consistent: the same layers show high sensitivity across different queries, different facts, different domains. The sensitivity is structural, not content-dependent.
This is surprising because knowledge localization studies show that different facts activate different parts of the network. The fact about Paris being the capital of France lives in different neurons than the fact about water freezing at 0°C. But the layer at which editing is most effective is the same for both. Localization of knowledge and localization of editability are different properties.
The general observation: where information is stored and where information is most efficiently modified need not coincide. Storage is distributed and content-dependent. Editability is concentrated and architecture-dependent. A system can encode different things in different places while remaining editable at the same universal control point.