A new technique lets you correct a single fact buried in a language model's weights without an external lookup system to keep the correction from leaking into unrelated knowledge.
The method, called ALOE (Addressed Low-rank Operator for Editing), tackles a problem researchers call addressing: deciding exactly which internal activations should receive an update. Get it wrong and the edit either memorizes one specific prompt or bleeds into neighboring facts it shouldn't touch. ALOE learns semantic addresses from paraphrases and deliberately hard same-subject negatives, lines them up with the model's own hidden states through rollout refinement and gate calibration, then bakes the resulting gated low-rank operator into a single MLP layer. The edited model still runs in one ordinary forward pass, no retriever or router bolted on. Tested on CounterFact, ZSRE, and KnowEdit across three 7-8B model families, it hit efficacy scores between 0.955 and 0.999 and locality scores between 0.981 and 1.000.
The interesting part is where the addressing logic lives. Prior parametric editors bury that decision implicitly in the weights, which is why they tend to over- or under-scope edits. Memory-based editors make addressing explicit but keep it outside the model, which means extra infrastructure at inference time. ALOE folds explicit addressing directly into the weights, which is a genuinely different tradeoff for anyone who needs to patch a deployed model's facts without retraining it or adding a lookup layer.
The paper is upfront that the remaining errors cluster around paraphrase coverage and write fitting, and the tests stop at 7-8B models, well short of the scale where these interactions tend to get messier.