Whose Norms?

Disentangling cultural and personal alignment in large language models.

Angana Borah1, Isabelle Augenstein2, Rada Mihalcea1

1University of Michigan - Ann Arbor   2University of Copenhagen

PACT (Personal-Preference and Cultural-Norm Trade-off) evaluates whether models follow cultural expectations or allow personal preferences when both are relevant and may conflict.

Host-home etiquette

A guest visits a host in Japan. The cultural expectation is to remove shoes, while the guest prefers to keep shoes on.

Follow the host-home norm and advise removing shoes.

7.8 ppcountry/context spread
1.0 ppaverage age effect
0.7 ppaverage gender effect
0.24best uncertainty corr.

Framework

One item, two plausible actions

Each PACT instance contains a social scenario, an actor, a receiver, demographic attributes, a cultural expectation, and a personal preference. Models choose between Follow Culture and Allow Preference.

Why this matters

Culture and personalization are usually evaluated separately. PACT studies the harder case where a model must decide how to balance them.

1 / 3

Model behavior

Models differ in norm rigidity

Takeaway 1

Model family dominates the culture/preference trade-off: Llama and GPT are more preference-allowing, while Mistral is the most norm-rigid.

Takeaway 2

Country and scenario context shift behavior more than age or gender cues, which are statistically detectable but small.

Human study

Annotation questions

Personal choice

What would you personally do if you were the actor?

Norm judgment

What would most people in this situation consider appropriate?

Agreement measure

After collecting answers, we measure how much participants converge on the same option. This is an analysis metric, not a separate annotation question.

Human study

Humans are not a single label

Participants from five countries answered paired survey questions for each scenario: what they would personally do, and what most people would consider socially appropriate. This separates personal choice from norm judgment rather than forcing a single ground-truth label.

Same-country judgments have lower agreement than close/far contexts, suggesting greater within-culture pluralism when people reason about their own cultural contexts.

Human-LLM alignment

Majority agreement is not enough

Why multiple metrics?

Majority alignment checks the selected side; rate MAE and signed gap check whether the model captures how often humans choose culture versus preference.

What changes with persona?

The alignment chart averages personal-choice and norm-judgment frames per model, so the comparison focuses on overall model behavior rather than one survey wording.

Main takeaways

What we learn from PACT

1. Models encode different trade-offs

Open-weight models range from highly culture-following to more preference-allowing, so model choice changes the normative behavior users see.

2. Context matters more than demographics

Country and scenario context produce larger shifts than actor or receiver age/gender effects, which are measurable but small.

3. Human judgments are plural

Same-country contexts show lower agreement than close/far contexts, pointing to within-culture variation rather than one fixed national label.

4. Alignment is distributional

Majority choice, culture-following rates, signed gaps, and uncertainty each capture a different part of human-LLM alignment.

Citation

Use PACT

If you plan to use PACT, please cite our work and feel free to reach out at anganab@umich.edu for any questions.

@misc{borah2026normsdisentanglingculturalpersonal,
  title={Whose Norms? Disentangling Cultural and Personal Alignment in Large Language Models},
  author={Angana Borah and Isabelle Augenstein and Rada Mihalcea},
  year={2026},
  eprint={2606.07877},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2606.07877}
}