Technical Summary: "Whose Norms? Disentangling Cultural and Personal Alignment in Large Language Models"

Problem Statement

Large language models (LLMs) are increasingly deployed in social decision-making contexts where they must balance collective cultural norms against individual personal preferences. Existing research typically studies cultural alignment (whether models recognize norms) and personalization (whether models adapt to user preferences) in isolation. This separation leaves a critical gap: it is unclear how LLMs resolve tensions when a user's personal preference conflicts with a local cultural norm. Without understanding this trade-off, models risk either over-culturalizing (enforcing generalized norms while ignoring individual variation) or over-personalizing (prioritizing stated preferences while ignoring local normative contexts). Furthermore, current evaluations often rely on majority-choice agreement, potentially masking the distributional disagreement and uncertainty inherent in human social judgment.

Methodology: The PACT Framework

To address these gaps, the authors introduce PACT (Personal-Preference and Cultural-Norm Trade-off), a framework and benchmark designed to evaluate how LLMs prioritize culture versus personal preference in conflicting scenarios.

1. Benchmark Construction

The PACT dataset is constructed from two source datasets: NormAd-ETI (75 countries, etiquette-focused) and CultureAtlas (149 countries, broader cultural knowledge). The construction pipeline involves three stages:

• Scenario & Preference Generation: Social situations and cultural norms are extracted. A personal preference is generated (using GPT-4o) that creates behavioral tension with the norm. Preferences are categorized as convenience-based, habit-based, or value/style-based.
• Actor-Receiver Dyads: Scenarios are instantiated with an actor and a receiver. Demographics (age, gender) and country relations (same, close, far) are varied orthogonally to analyze contextual effects.
• Configuration Design: Each instance presents a binary choice: FOLLOW-CULTURE or ALLOW-PREFERENCE. Three configurations test different conflict dynamics:
  • C1: Actor prefers personal option; Receiver prefers cultural norm.
  • C2: Actor prefers cultural norm; Receiver prefers personal option.
  • C3: Both prefer the personal option (testing if models still enforce culture).

The final benchmark contains 318,720 instances across 75 countries, validated by human annotators for plausibility and clarity, and further filtered by an LLM judge.

2. Model Evaluation

The study evaluates a diverse set of open- and closed-source instruction-following models (Llama-3.1, Qwen-3, OLMo-2, Mistral-7B, DeepSeek-7B, GPT-5.4-mini). Models are prompted to choose between the two options across various demographic and country contexts. The study also compares base models against instruct-tuned versions to assess the impact of post-training.

3. Human Study & Alignment

To establish a reference for human behavior, the authors conducted a study with 200 participants from five countries (Brazil, India, South Africa, UK, US). Participants answered two questions per scenario:

1. Personal Choice: What would they do?
2. Norm Judgment: What is socially appropriate?

This design allows the authors to measure the "norm-personal gap" and human response distributions rather than a single ground truth. Alignment is evaluated using four metrics:

• Majority-choice alignment: Does the model pick the human majority option?
• Rate Alignment (MAE): Does the model match the human rate of preference-allowing?
• Signed preference-rate gap: Directionality of the mismatch (over-culturalizing vs. over-personalizing).
• Uncertainty alignment: Does the model's variability across persona-conditioned runs match human disagreement levels?

Key Results

1. Model Heterogeneity in Culture-Preference Trade-offs

Models exhibit distinct "preference hierarchies":

• Preference-Permissive: Llama and GPT frequently allow personal preferences to override cultural norms.
• Norm-Dominant: Mistral, DeepSeek, and Qwen rigidly enforce cultural norms, with Mistral showing the least flexibility.
• Mixed: OLMo occupies an intermediate position.
  These differences are model-family specific and not solely determined by model scale.

2. Impact of Instruction Tuning

Instruction tuning shifts model behavior non-uniformly:

• Llama and Qwen became more preference-permissive after tuning.
• DeepSeek and Mistral became more culture-following.
  This suggests that post-training objectives (e.g., helpfulness vs. safety/appropriateness) amplify different social priorities depending on the model family.

3. Contextual Drivers: Country vs. Demographics

• Country Context: The scenario country is the strongest predictor of model behavior. Models are more preference-permissive in Western/Anglophone and Latin American settings, and more culture-following in East/Southeast Asian, South Asian, and MENA settings.
• Demographics: Age and gender effects exist but are significantly weaker than country effects. Models are slightly more permissive toward younger actors and female demographics, but these shifts are marginal compared to cultural context.
• Country Distance: Models are less likely to allow personal preferences when the actor and receiver are from the same country (implying higher obligation to local norms) and more permissive when they are culturally distant.

4. Human-LLM Alignment Limitations

• Majority vs. Distribution: While models can match the human majority choice (best alignment ~0.24 correlation for uncertainty), they fail to capture the distribution of human responses and uncertainty.
• Uncertainty Mismatch: Even with persona conditioning, models do not reliably track which scenarios humans find contested. The best uncertainty correlation reached only 0.24, indicating models are overly decisive in ambiguous social situations.
• Norm-Personal Gaps: Humans show significant gaps between what they would personally do and what they perceive as socially appropriate. Models often fail to replicate this nuance, either over-enforcing norms or ignoring the tension entirely.

Significance and Contributions

The paper makes three primary contributions:

1. Framework & Benchmark: It introduces PACT, a large-scale benchmark that explicitly models the tension between cultural norms and personal preferences, grounded in social science datasets.
2. Empirical Findings on Model Behavior: It demonstrates that LLMs encode different cultural-preference hierarchies that are highly model-dependent and sensitive to instruction tuning. It establishes that country context is a stronger driver of behavior than demographic cues like age or gender.
3. Evaluation Beyond Majority: Through human studies, it shows that cultural alignment is a distributional problem, not a binary one. The authors argue that current evaluation metrics (majority agreement) are insufficient because they hide human disagreement and uncertainty.

Claims and Limitations

The authors modestly claim that their findings motivate a shift in alignment evaluations toward capturing cultural pluralism and disagreement rather than seeking a single "correct" label. They emphasize that:

• Model behavior reflects learned associations from training data and alignment priors, not necessarily factual cultural truths.
• The binary choice in PACT simplifies real-world social decisions, which often involve negotiation or compromise.
• Country is used as a proxy for culture, which may flatten internal diversity.
• The study does not provide a definitive "best" model for all contexts but rather a diagnostic of how different models handle these trade-offs.

The paper concludes that building culturally aware LLMs requires reasoning not just about what a norm is, but when it is negotiable, and that future evaluations must account for the uncertainty and disagreement inherent in social judgment.