Activation Steering Enhances Role-Conditioned Agents in Social Simulations

A new arXiv preprint (2608.00023) introduces an activation-steering screening workflow for role-conditioned language model agents in social simulations. The method involves defining a role profile, extracting a role-specific direction, sweeping four steering coefficients, evaluating role-profile alignment, and passing or flagging candidate configurations. Applied to OLMo-3-7B-Instruct with a mixed 275-role inventory and 228 role-agnostic questions, the approach uses GPT-4.1-mini for prompted role references and judging. Role-specific directions achieved higher judged role-profile alignment (mean overall score 63.2) compared to an assistant-axis directional control (41.1) across the tested grid, while preserving high lexical diversity. The role-level screen is the main practical output, enabling pre-screening of agents before population deployment.