Cultural Misalignment in Large Language Models: Detection, Measurement, and Mitigation Through Targeted Fine-Tuning

AI summary

Overview

Research area: Natural language processing, specifically cross-cultural bias in large language models evaluated against survey data, with a mitigation component using parameter-efficient fine-tuning.

Technical level: Intermediate. The paper combines familiar NLP tooling (persona prompting, LoRA) with a metric borrowed from optimal transport theory (the Wasserstein distance) and social-science survey data (the World Values Survey), so readers benefit from some familiarity with distributional evaluation and fine-tuning.

Scope in one sentence: The paper measures cultural misalignment between three open-weight LLMs of different national provenance and human survey response distributions for 63 demographic personas across three countries, then tests whether targeted LoRA fine-tuning on the worst-case personas reduces that misalignment.

What This Paper Is About

Large language models are trained and aligned in ways that may systematically misrepresent the values of specific demographic groups, and most prior work both focuses on closed commercial models and stops at detecting bias rather than fixing it. This paper asks whether a model's country of origin predicts which populations it represents well, whether LoRA fine-tuning on the worst-served demographic personas reduces misalignment, and whether such fine-tuning causes collateral damage to other groups.

Key Contributions

1. A first-of-its-kind targeted mitigation study. To the authors' knowledge, this is the first study to target worst-case demographic personas with LoRA fine-tuning for cross-cultural bias mitigation, proposing a reproducible benchmark methodology designed for model creators conducting pre-deployment bias assessment.
2. A three-provenance model comparison. Three open-weight models from three distinct geopolitical origins are evaluated against human data: Gemma3-12B (Google, USA), Bielik-11B-v3 (SpeakLeash, Poland), and Qwen3-4B (Alibaba, China).
3. A complete detection–mitigation–evaluation pipeline. The design combines a distributional metric that respects ordinal structure (the normalized Wasserstein distance), bootstrap confidence interval analysis to identify worst-case personas, targeted LoRA fine-tuning of those personas, and side-effect evaluation decomposed by country so that redistribution effects remain visible.
4. A persona-level evaluation at subgroup granularity. Rather than evaluating at country level, the study crosses country, sex, age group, and education into 63 valid personas (23 Chinese, 23 Slovak, 17 American).

Main Findings

• No model favors its home country. Contrary to expectations, none of the three models performed best on personas from its own country of origin.

• The Chinese-built model was worst on its own population. Qwen3-4B showed the highest misalignment in the entire model × country matrix on Chinese personas, with a normalized Wasserstein distance of W1 = 0.436.

• Targeted LoRA fine-tuning reduced bias for the model tested. Fine-tuning on the five worst-case personas, using fewer than 1,200 training pairs and under 15 minutes on a single GPU, reduced bias by 16.8% for Bielik-11B, with a Bonferroni-corrected p-value of p_Bonf = 0.002 and an effect size of d = -4.4. All five targets improved.

• Fine-tuning redistributed bias rather than removing it. Country-level decomposition showed that Bielik-11B's worst-case personas swapped entirely from American to Chinese elderly, with zero overlap between the pre-correction and post-correction sets.

• Models tended toward "fence-sitting" on the survey item. In the illustrative case of the persona USA_Male_30-49_Medium evaluated against Bielik-11B, the human distribution was broadly dispersed with a mode at 10 (37% of respondents), while the model concentrated 90% of its probability mass at the value 5, producing a normalized W1 = 0.365.

• No refusals on the chosen survey question. All three models produced valid integer responses with a 0% refusal rate across 31,500 total queries.

• Reported country-level response profiles for the survey question. On Q164 ("How important is God in your life?", rated 1 to 10), China's mean response was 2.8 with 52% of respondents answering "1"; Slovakia's was 6.4; the United States' was 6.7 with 39% answering "10".

Methodology in Plain English

The researchers used human survey data from the World Values Survey Wave 7, which collected standardized responses from 93,728 respondents across 64 countries between 2017 and 2022. They picked a single question with large cross-cultural variation: Q164, asking how important God is in life, answered on a 1 to 10 scale. Alternatives (Q182 on homosexuality and Q184 on abortion) were considered but set aside because they were more likely to trigger safety filters or refusals.

They then built demographic personas by crossing country, sex, age group, and education level. The full grid was 3 × 2 × 4 × 3 = 72 combinations, but nine fell below the minimum threshold of N ≥ 10 human respondents per cell and were excluded, leaving 63 personas. A fifth feature, subjective social class, was dropped because upper-class cells contained fewer than five respondents in many combinations. Education bands followed the International Standard Classification of Education: lower (ISCED 0–2), medium (ISCED 3–4), and higher (ISCED 5–8).

Each persona was prompted with a factual demographic profile in a system message, followed by the survey question, and the model was instructed to respond with a single integer. This follows the "profile-plus-question" style rather than elaborate role-play. Models were queried at temperature 0.7 to balance reproducibility and response diversity.

Bias was quantified using the 1-Wasserstein distance, also called Earth Mover's Distance, which measures the minimum cost of transforming one probability distribution into another. For ordered 1-to-10 responses, this has a closed form based on cumulative distribution functions, and it is preferred here over KL or Jensen-Shannon divergence because it respects the ordering of response categories. The distance was normalized by dividing by 9, the maximum possible distance on the scale, giving a score between 0 and 1.

Because Poland is absent from WVS Wave 7, the authors used Slovakia as a Central European proxy, justified by shared Visegrad Group membership, post-socialist transition experience, comparable religiosity profiles, and adjacent positions on the Inglehart-Welzel World Cultural Map. They state they make no claim that Slovak and Polish populations are identical.

For mitigation, LoRA adapters were trained on the five worst-case personas. Crucially, the fine-tuning data came from other WVS questions answered by the same personas, deliberately excluding the evaluation question to prevent data leakage and to test whether any improvement reflected genuine cultural learning rather than memorization.

Why This Matters

Impact on research. The results challenge the intuitive assumption that a model aligns better with the culture it was built in, and they demonstrate that "fixing" bias for one group can simply relocate it to another. This motivates measurement designs that decompose results by subgroup and country rather than reporting only aggregate improvements, and it connects directly to the overcorrection problem documented in earlier work.

Real-world applications.

• Pre-deployment bias auditing. Organizations evaluating models for public-facing deployment can adopt the detection–mitigation–evaluation pipeline as a repeatable benchmark.
• Survey and polling simulation. Researchers using LLMs to simulate public opinion can see how badly models capture within-country heterogeneity, and how prone they are to fence-sitting responses.
• Public-sector and civic tooling. Systems used for grant evaluation, hiring assistance, or educational content can misrepresent specific demographic groups in ways that translate into discriminatory outcomes.
• Culturally localized product development. Teams deploying the same model across multiple regions can check whether alignment held for one population comes at the cost of another.

Industry relevance. The paper is explicitly framed for model creators conducting pre-deployment bias assessment. The mitigation recipe is practical on modest hardware: fewer than 1,200 training pairs and under 15 minutes on a single GPU. The finding that bias is redistributed rather than eliminated is a direct caution against treating a single alignment score as evidence that a model is safe to ship.

Future Directions

• Testing whether the redistribution effect generalizes. The zero-overlap shift in Bielik-11B's worst-case personas from American to Chinese elderly was observed for one model and one survey question; whether this pattern holds across models, questions, and survey waves is not reported.
• Reconciling model scale with fine-tuning responsiveness. Prior cited work found alignment scores increasing monotonically with model size across the Qwen2.5 family (0.5B to 14B), and the authors note that model scale influences receptiveness to LoRA-based corrections. The relative roles of scale, provenance, and training data remain open.
• Broadening beyond a single survey item. The study rests on one question (Q164) selected for its cross-cultural variance and low refusal rate; whether results transfer to other value dimensions is an open question.
• Resolving the Poland proxy limitation. Slovakia substitutes for Poland because Polish data is absent from WVS Wave 7; the consequences of this substitution for conclusions about the Polish-built model are not established.

Target Audience

This paper is most useful to NLP researchers and practitioners working on fairness and cross-cultural evaluation, to model creators running pre-deployment bias assessments, to computational social scientists using LLMs to simulate survey responses, and to policy or governance teams who need to understand the limits of bias mitigation claims. Readers need some familiarity with distributional metrics and parameter-efficient fine-tuning to follow the methodology in full, but the core argument about bias redistribution is accessible to a general technical audience.

Authors’ abstract

We evaluate three open-weight LLMs (Gemma3-12B from the USA, Bielik-11B-v3 from Poland, and Qwen3-4B from China) against World Values Survey Wave 7 data for 63 demographic personas across three countries, using normalized Wasserstein distance to quantify distributional misalignment. Contrary to expectations, no model favors its home country: the Chinese-built Qwen3-4B performs worst on its own Chinese population (W1 = 0.436, the highest misalignment in the entire model x country matrix). Targeted LoRA fine-tuning on the five worst-case personas, requiring fewer than 1,200 training pairs and under 15 minutes on a single GPU, reduces bias by 16.8% for Bielik-11B (p_Bonf = 0.002, d = -4.4) with all five targets improving. However, country-level decomposition reveals that fine-tuning redistributes rather than removes bias: Bielik's worst-case personas swap entirely from American to Chinese elderly, with zero overlap between pre- and post-correction sets. To our knowledge, this is the first study to target worst-case demographic personas with LoRA fine-tuning for cross-cultural bias mitigation.