Role Steering of Language Models for Social Simulations
Source
Isaac Song, Mohammed Rehan Parwani, Glenn Matlin, Emile Anand, Akhil Theerthala, Arjun Chatterjee, Maria Kostylew, Yonadav G. Shavit, Sebastien Krier, Mark Riedl
Author
Isaac Song, Mohammed Rehan Parwani, Glenn Matlin, Emile Anand, Akhil Theerthala, Arjun Chatterjee, Maria Kostylew, Yonadav G. Shavit, Sebastien Krier, Mark Riedl
Date
Key takeaways · AI-distilled
Role-specific activation directions beat the prior persona-vector assistant-axis control decisively — 63.2 vs 41.1 mean judged role-alignmentThe work of making AI systems actually pursue what their builders and users intend, rather than something subtly or dangerously different.Full definition → across the tested grid on OLMo-3-7B-Instruct with 275 roles.
Role directions also preserve lexical diversity, while the assistant-axis control drops sharply at larger steering coefficients.
38 of 275 roles (~14%) degrade across all six measured dimensions at every steering strength — so coefficients must be chosen per role via screening, never deployed as one uniform high-strength setting.
The reusable recipe: define a role profile, extract its direction, sweep four coefficients, judge alignment against the profile, then pass or flag each configuration before it enters a simulation.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
alignment — The work of making AI systems actually pursue what their builders and users intend, rather than something subtly or dangerously different.
activation steering — Changing a model's behavior by nudging its internal activations directly at runtime — no retraining, no prompt changes.
Why it matters
The paper shows role-specific activation steeringChanging a model's behavior by nudging its internal activations directly at runtime — no retraining, no prompt changes.Full definition → beats persona-vector control (63.2 vs 41.1 alignment) for controlling AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → personas in social simulations, but flags that ~14% of roles degrade regardless of steering strength, arguing practitioners need per-role validation before deployment rather than uniform tuning.
Key quotes
“Role-specific directions receive higher judged role-profile alignment than an assistant-axis directional control from prior persona-vector work, with mean overall scores of 63.2 versus 41.1 across the tested grid.”
“The role-level screen is the main practical output: most roles improve as steering increases, but 38 roles decline across all six measured dimensions, showing why simulation builders should choose coefficients per role rather than deploy a uniform high-strength setting.”
articleDoes Fine-Tuning Undo Activation Steering? Behavioural Recovery Without Weight-Edit ReversalPhilipp E. Glass, Allan Tucker, Yongmin Li, Alina Miron
articleEven More Deception: Objective Misalignment in Mixed-Motive LLM Multi-Agent SystemsMarylou Fauchard, Florian Carichon, Margarida Carvalho, Golnoosh Farnadi
articleWhat a crowdsourced game revealed about steering Olmo 3allenai.org