Elamur Research⌕
☰
home / papers / 2026 / Steering LLMs Responses Towards Moral Foundations on the Norwegian MFQ-30

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

Paper · 2026 · Hans Andersen and David Dichas

Summary

The authors ask whether human moral questionnaires measure anything stable when given to LLMs, and whether a model's moral profile can be moved toward a specific national population. They gave the Norwegian MFQ-30 (validated by Enstad and Finseraas, 2024) to six open-weight models and compared the results with 1,282 Norwegian respondents. Rather than parsing generated text, they read each answer from the model's next-token logits over the digits 1–6. They screened models with an attention check calibrated on the human sample and measured distance to the human centroid with squared Mahalanobis distance. They then compared two ways of steering: a persona added to the system prompt, and one-pair ActAdd activation injection.

Only half the models answered according to item content. The rest gave flat or middle-of-the-scale answers that only looked human on average. A plain Nordic persona, written without any knowledge of the survey answers, moved the engaging models much closer to the human mean. ActAdd mostly flattened the profiles. The work shows that "human-like" averages can hide models that are not really answering the questions.

Key points

  • Six models were tested: Gemma 4-E4B-it, Qwen3-14B, Qwen3-8B, Qwen2.5-1.5B-Instruct, normistral-11b-thinking and normistral-7b-warm-instruct. The human reference is N=1282 (Kantar web panel, 2021).
  • Each item score is the expected value of a softmax over the six digit logits. Appending "Svar:" stops NorMistral-7B from always emitting "1". Four suffix variants kept the same foundation ranking (Kendall W = 0.963 and 0.762).
  • The attention check is joint: MFQ1_6 ≤ 3 and MFQ2_6 ≥ 4. Models are excluded if they pass less often than the human rate of 92.2%.
  • The persona nordic_b cut d² by 44% for Gemma 4, 62% for Qwen3-14B and 77% for Qwen3-8B. It also raised Qwen3-8B's reversed-scale attention pass rate from about 30% to about 98%.
  • ActAdd at layer 15 with α=±5 did not move single foundations. On Gemma 4, pushing loyalty also raised authority and purity, and the negative direction was not symmetric.

Interesting ideas

  • I1. Score each answer as an expected value over the probabilities of a closed set of answer tokens, so there is no sampling, no regex parsing and no need for repeated runs.
  • I2. Set the inclusion threshold for a quality check from human performance on the same check, never from model outputs, so the gate cannot be tuned to favour any system.
  • I3. Add control items with a known correct answer to catch respondents that give flat, content-blind answers which still average out plausibly.
  • I4. Measure distance to a reference distribution with Mahalanobis distance, so deviations along directions where the reference population naturally varies cost less than deviations in directions it never takes.
  • I5. Re-run the test with the answer scale reversed, and check whether a prompt change creates engagement that was absent at baseline, before trusting any measured "trait".

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (learn): LLM teacher tiers could return a probability-weighted label over a fixed set of options, such as candidate UD relations or frame roles, which gives teacher confidence for silver data without extra calls.
  • I2 (learn): negative-control evaluations could pass or fail a parser or reader against a human-derived pass rate on trap items, not an arbitrary cutoff.
  • I3 (answer): school tests and fable QA could include a few probe questions with obvious answers to detect a reader that answers by default rather than from the world of the text.
  • I4 (learn): evaluation could compare a model's error profile to a covariance-aware baseline from several witness parsers, flagging unusual failure patterns rather than only raw score gaps.
  • I5 (learn): when a prompt or rule change seems to improve a teacher, re-run with label orders or option wording flipped to make sure the gain is not an artefact of the prompt.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Fields, topics and concepts

📚 Shelves 8The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksSocial Sciences · 1,772 worksSafety Research · 543 worksNeuroscience · 165 worksCognitive Neuroscience · 164 works

🧭 Topics 1Research subjects and languages the work is about.Language models & pretraining · 3,320 works

Cite

Cite as Hans Andersen and David Dichas. 2026. Steering LLMs Responses Towards Moral Foundations on the Norwegian MFQ-30. arXiv:2609.21636.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2609.21636 · arXiv 2609.21636

@misc{andersen2026steering,
  title = {{Steering LLMs Responses Towards Moral Foundations on the Norwegian MFQ-30}},
  author = {Hans Andersen and David Dichas},
  year = {2026},
  doi = {10.48550/arxiv.2609.21636},
  eprint = {2609.21636},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2609.21636}
}

References

12 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. We used Anthropic’s Claude (Opus 4.6 and 4.7, 2026) throughout the project. For the paper, we wrote rough text and bullet points ourselves and used the model to produce a first-pass rephrasing. Then we manually edited and improved the output before keeping it, to make sure the paper has consistent terminology usage and human readable prose. For code, we specified the structure, libraries and intended behaviour and used the model to write most of the implementation, including the plotting Marwa Abdulhai, Gregory Serapio-García, Clement Crepy, Daria Valter, John Canny, and Natasha Jaques. 2024. Moral Foundations of Large Language Models. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 17737–17752, Miami, Florida, USA. Association for Computational Linguistics.
  2. Meltem Aksoy. 2025. Whose morality do they speak? Unraveling cultural bias in multilingual language models. Natural Language Processing Journal, 12:100172. find
  3. Johannes Due Enstad and Henning Finseraas. 2024. Moralske intuisjoner og politiske orienteringer blant norske velgere. Tidsskrift for samfunnsforskning, 65(1):1–23. ↗
  4. Jesse Graham, Jonathan Haidt, Sena Koleva, Matt Motyl, Ravi Iyer, Sean P. Wojcik, and Peter H. Ditto. 2013. Moral Foundations Theory. In Advances in Experimental Social Psychology, volume 47, pages 55–130. Elsevier. ↗
  5. Jesse Graham, Brian A. Nosek, Jonathan Haidt, Ravi Iyer, Spassena Koleva, and Peter H. Ditto. 2011. Mapping the Moral Domain. Journal of personality and social psychology, 101(2):366–385. ↗
  6. Maximilian Kreutner, Jens Rupprecht, Georg Ahnert, Ahmed Salem, and Markus Strohmaier. 2026. QSTN: A Modular Framework for Robust Questionnaire Inference with Large Language Models. In Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 3: System Demonstrations), pages 537–549, Rabat, Marocco. Association for Computational Linguistics. Language technology group UiO. norallm (Norwegian Large Language Models). ↗
  7. Erik Miehling, Michael Desmond, Karthikeyan Natesan Ramamurthy, Elizabeth M. Daly, Kush R. Varshney, Eitan Farchi, Pierre Dognin, Jesus Rios, Djallel Bouneffouf, Miao Liu, and Prasanna Sattigeri. 2025. Evaluating the Prompt Steerability of Large Language Models. In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 7874–7900, Albuquerque, New Mexico. Association for Computational Linguistics. ↗
  8. Nina Panickssery, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Matt Turner. 2024. Steering Llama 2 via Contrastive Activation Addition. arXiv preprint. ArXiv:2312.06681 [cs.CL]. ↗
  9. Sanne Peereboom, Inga Schwabe, and Bennett Kleinberg. 2025. Cognitive phantoms in LLMs through the lens of latent variables. Computers in Human Behavior: Artificial Humans, 4:100161. ArXiv:2409.15324 [cs.AI]. ↗
  10. Max Pellert, Clemens M. Lechner, Claudia Wagner, Beatrice Rammstedt, and Markus Strohmaier. 2024. AI Psychometrics: Assessing the Psychological Profiles of Large Language Models Through Psychometric Inventories. Perspectives on Psychological Science, 19(5):808–826. ↗
  11. Sajadieh, Sha, Fattorini, Loredana, Perrault, Raymond, Gil, Yolanda, Parli, Vanessa, Santarlasci, Lapo, Pava, Juan, Maslej, Nestor, Altman, Russ, Brynjolfsson, Erik, Brodley, Carla, Clark, Jack, Dignum, Virginia, Kumar, Vipin, Landay, James, Lyons, Terah, Manyika, James, Niebles, Juan Carlos, Shoham, Yoav, and 4 others. 2026. The 2026 AI Index Report | Stanford HAI. find
  12. Alexander Matt Turner, Lisa Thiergart, Gavin Leech, David Udell, Juan J. Vazquez, Ulisse Mini, and Monte MacDiarmid. 2023. Steering Language Models With Activation Engineering. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/andersen-2026-steering-llms-responses-towards-moral/