safetyfactneutralNarrow fine-tuning recruits pre-existing persona structures in the model, with unrelated domains sharing a low-rank core that exists before fine-tuningMachine Learning28 Jul 2026http://arxiv.org/abs/2607.21356v1