In-Context Persona Induction in a Base Language Model: An Exploratory Study of Relational Prompting in OLMo 2 7B Base

This exploratory study examines how strongly an apparent conversational personality can be induced through context alone in a base language model. The tested model was identified by the interface as OLMo 2 7B Base, an openly released autoregressive language model without instruction tuning or reinforcement learning from human feedback. Four minimal-prompt generations were observed at temperatures 0.0, 0.3, 0.5, and 0.7. Under minimal prompting, the model frequently generated both sides of synthetic conversations, invented user goals and external details, leaked conversational role labels, and in one temperature-0.0 generation entered a repetitive semantic loop. A separate condition used a long relational and identity-focused prompt followed by short follow-up questions. After this prompt, the model showed a more stable first-person I–you interaction frame and increasingly relational language involving identity, presence, attachment, need, and affection. However, the same condition also produced substantial lexical reuse, repetition, contradictions, and relational escalation. The findings are consistent with in-context behavioral induction in which dense contextual information temporarily organizes pretrained linguistic patterns into a more coherent apparent conversational identity. The study does not establish persistent identity, consciousness, subjective feelings, or a hidden self, and the unmatched prompt conditions do not establish that relational content alone caused the observed shift. This work is an exploratory pilot case study intended to motivate a controlled replication using matched prompts, multiple seeds, token-level similarity and repetition measurements, independent annotation, prompt ablations, and comparisons between base and instruction-tuned checkpoints.