Borrowing a Digital Twin: Grounding Individual-Level LLM Personas in a Matched Donor's Profile

Persona-conditioned language models reproduce human responses well at the population level but remain unreliable for any single person. They usually rely on demographic attributes, and those attributes do not pin down what one individual thinks. Prior work grounds personas in a person's own past behavior or measured psychological traits, but such signals are rarely available for the population one wants to simulate. We explore a different route. When a target's own psychology cannot be measured, we borrow it from a well-matched donor drawn from a deeply profiled panel such as Twin-2K-500, and use the donor's measured profile to ground the target's persona. We describe a three-phase pipeline (demographic filtering, behavioral-fit donor selection, and profile-grounded rationale generation) and report a preliminary study on OpinionQA, where donor-grounded rationales perform on par with a psychological-scaffold baseline and above demographics-only prompting.