Kaiwei Lei

Output vector editing can suppress up to 87.9% of memorized sequences in large language models, significantly outperforming traditional neuron-level methods.