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.
This paper investigates whether repetition loops in long factual enumeration tasks by Gemma 4 models can be fixed by editing a single neuron. It finds that targeted weight edits on a small set of MLP neurons can significantly reduce loop failures, though not completely eliminate doom looping in larger models.
Contrastive neuron attribution (CNA) identifies a sparse set of MLP neurons that distinguish harmful from benign prompts, enabling effective behavioral steering in instruction-tuned LLMs without degrading output quality. The method reduces refusal rates by over 50% on jailbreak benchmarks while preserving fluency.