Role Steering of Language Models for Social Simulations

Short summary

This paper introduces an activation-steering screening workflow for role-conditioned language model agents in social simulations. Applied to OLMo-3-7B-Instruct across 275 roles, role-specific steering directions achieve higher alignment scores (63.2 vs 41.1) than prior persona-vector approaches while preserving lexical diversity. The key practical finding: 38 of 275 roles decline with increased steering, so coefficients should be tuned per role rather than set uniformly.

• •Activation-steering workflow for screening role-conditioned agents before simulation deployment
• •Role-specific directions outperform persona-vector controls on alignment (63.2 vs 41.1) while preserving diversity
• •38 of 275 roles decline with stronger steering, requiring per-role coefficient tuning
• •Code and evaluation artifacts publicly available

Generated with AI, which can make mistakes.