Persona Vectors Form Early in LLM Pretraining

A new study from arXiv traces the formation of persona vectors—linear directions in internal activations corresponding to high-level behaviors like sycophancy—across the pretraining of OLMo-3-7B. These vectors form within 0.22% of pretraining and remain effective for steering fully post-trained instruct models. Although core representations emerge early, they continue to refine geometrically and semantically throughout training. The research addresses a gap in AI safety interpretability, as persona vectors are routinely used to inspect and steer model behavior.