Output Vector Editing for Memorization Mitigation in Large Language Models

In three linesMemorization suppression method in LLMs via output vector editing of MLP neurons. Tested on 4 models (360M-7B parameters), achieves 87.9% suppression on OLMo-7B with 6831 memorized sequences. Complementary approach to existing neuron ablation methods.

Summary generated by Claude — human-verified