Researchers unveil role-steering AI for social simulations

A team led by Isaac Song from Georgia Tech has developed a novel activation-steering workflow to improve role-conditioned behavior in AI agents for social simulations. The method defines role profiles, extracts role-specific activation directions, and systematically evaluates steering coefficients to maximize role-profile alignment.

When tested on the OLMo-3-7B-Instruct model with a 275-role inventory and 228 role-agnostic questions, their approach achieved mean role-profile alignment scores of 63.2 versus 41.1 for baseline persona-vector controls—a 54% improvement. Crucially, the method preserved high lexical diversity even at higher steering coefficients, unlike baselines where diversity dropped sharply. The research highlights the need for role-specific coefficient tuning, as 38 roles showed declining performance across all measured dimensions when using uniform high-strength settings.