How do persona directions extracted from prompts relate to the states a model reaches after training, and to how it actually behaves?
I combined steering, mixed-persona fine-tuning and fictional-persona conditions to compare them. The questions raised here went on into a philosophy paper, Persona Without Substrate.