Regime-Dependence and the LLM Individuation Problem.
Under prompting, fine-tuning and inference-time intervention, what makes a representation “the same” content? Drawing on the persona topology experiments, the paper argues that individuating what a model represents depends on the regime it is acting under.