Research Summary: Steering LLMs Responses Towards Moral Foundations on the Norwegian MFQ-30

Recent research explored the application of human psychometric questionnaires to large language models (LLMs) to ascertain their moral and value profiles. The study administered the Norwegian Moral Foundations Questionnaire (MFQ-30) to six open-weight LLMs, comparing their responses to a Norwegian human population sample. It also investigated two methods for steering LLM responses: prompt-level persona steering and activation-level ActAdd. A significant observation was that many models defaulted to central-tendency outputs, appearing human-like without genuinely tracking item content.

Why it matters

This research is crucial for understanding the ethical alignment and potential biases inherent in AI systems, particularly LLMs. It highlights the complexities of assessing AI's 'moral compass' and the challenges in steering these systems reliably. The findings influence strategy around responsible AI development and deployment, particularly where moral reasoning or alignment with human values is critical.

Key insights

• Human psychometric questionnaires, specifically the Norwegian MFQ-30, were used to elicit moral and value profiles from six open-weight Large Language Models (LLMs).
• LLM moral profiles were compared against a dataset of 1,282 Norwegian human respondents.
• Two steering interventions, prompt-level persona steering and activation-level ActAdd, were tested to influence LLM responses towards a target human population's profile.
• A key finding was that some LLMs, when not specifically directed, produced 'flat' or 'central-tendency' outputs that resembled human responses on average but lacked genuine engagement with individual item content.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2609.21636