Tag
Presents output vector editing, a constrained-optimization weight edit to mitigate memorization in LLMs by modifying MLP neuron output vectors instead of zeroing activations, achieving up to 87.9% suppression with minimal locality failures.