Unlearning

Deleting Paris from a Language Model

A single rank-1 weight edit suppresses one learned fact while leaving the rest of the model intact. No fine-tuning. No retraining. Just a feature subtracted from one layer's gate matrix — with a receipt.