Prompt and Activation-Level Steering Moves LLMs Toward Target Moral Profiles
September 21, 2026
Researchers tested steering interventions on six open-weight LLMs using the Norwegian MFQ-30 questionnaire. Persona-level prompting improved alignment with the Norwegian mean by 44-77% in Mahalanobis distance, while testing activation-level ActAdd interventions.
HOW THIS AFFECTS YOU
●
researcherYou can use these steering methods to evaluate how effectively model values can be modulated.
●
policyThis highlights the technical feasibility of adjusting model alignment to meet specific regional or cultural norms.