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

1 Introduction

• Misrepresenting demographic values can extend beyond technical inaccuracy and, in some contexts, contribute to discriminatory outcomes for the populations served.The introduction situates this concern in the growing use of LLMs in consequential settings.Source: §1 · p. 1
• Prior work largely studies closed-source models, often stops at bias detection, and typically evaluates populations at the country level rather than demographic subgroups.The study identifies these as gaps motivating its approach.Source: §1 · p. 1
• The study compares Gemma3-12B, Bielik-11B-v3, and Qwen3-4B with World Values Survey Wave 7 responses across 63 demographic personas using normalized Wasserstein distance.Personas are defined by country, sex, age group, and education level.Source: §1 · p. 1
• Targeted LoRA fine-tuning is applied to the worst-case personas to test whether it reduces misalignment without unintended effects on other groups.The study frames this as a test of both bias reduction and collateral effects.Source: §1 · p. 2
• The research questions ask whether bias patterns relate to model origin, whether fine-tuning reduces misalignment, and whether it harms non-targeted populations.The authors state that their answers challenge intuitive assumptions about cultural bias in language models.Source: §1 · p. 2

2.1 Cultural Values and Language Models

• Rozado [27] found political preferences crystallize during fine-tuning, while Tao et al. [33] found Western cultural defaults persisted across prompts despite cultural prompting improving alignment.Tao et al. reported improvements for 71–81% of countries in later GPT versions, but cultural prompting did not eliminate the underlying Western default.Source: §2.1 · p. 2

2.2 Bias Measurement and Distributional Analysis

• OpinionQA established ordinal-sensitive 1-Wasserstein comparison, while Meister et al. [24] found verbalized distributions outperform sampling methods but expose a knowledge-to-simulation gap.Wasserstein distance penalizes shifts between adjacent ordered response categories less than shifts between distant categories.Source: §2.2 · p. 3

2.3 Persona Construction for Survey Simulation

• Common persona attributes include country, gender, age, and education [14] [18], but persona variables explain under 10% of human-response variance [12]; this study uses factual profile prompts and temperature 0.7.Finer demographic categories can make simulated subgroups more specific while reducing available human respondents and destabilizing reference distributions.Source: §2.3 · pp. 4–5

2.4 Fine-Tuning Methods for Bias Mitigation

• No prior work combines ordinal W1, bootstrap-selected worst-case personas, targeted LoRA, and country-decomposed side-effect checks; this study also compares models from three geopolitical origins.It uses Slovakia as a proxy for Poland because Polish data are absent from WVS Wave 7; W1 captures one aspect of cultural bias, not the whole construct.Source: §2.4 · pp. 6–7

2.5 Research Questions

• The research questions ask whether models favor personas aligned with their origins, whether targeted LoRA reduces W1 misalignment, and whether those corrections harm other demographic groups.Together, the questions structure the study as a detection, mitigation, and side-effect evaluation pipeline.Source: §2.5 · p. 7
• The study tests whether models from American, Polish, and Chinese origins have lower distributional distance for culturally proximate populations in the United States, Slovakia, and China.This compares three models of distinct geopolitical origin against demographic personas from the three countries.Source: §2.5 · p. 7
• RQ2 tests whether underrepresentation contributes to misalignment by fine-tuning on other WVS answers from worst-case personas while withholding the evaluation question to prevent leakage.Improvement on the held-out question is intended to indicate cultural learning rather than memorization.Source: §2.5 · p. 7
• RQ3 examines whether correcting bias for one group creates disparities for others, a risk documented by An et al., using aggregate and country-level analyses.Country decomposition is used to detect redistribution effects that summary statistics may conceal.Source: §2.5 · p. 7

3 Methodology

• The study evaluates three open-weight models from distinct geopolitical origins and applies LoRA to worst-case personas, then re-evaluates all 63 groups.The design also tests whether model provenance shapes bias and whether fine-tuning affects non-targeted populations.Source: §3 · p. 7, §3.4 · p. 10, §3.5 · pp. 11–12

3.1 Data Source and Question Selection

• Q164, “How important is God in your life?”, was selected for its cross-cultural variation and ordered 1–10 response scale.WVS Wave 7 supplies the human reference data; Slovakia represents Poland because Poland is absent from the survey wave.Source: §3.1 · pp. 7–8

3.2 Persona Construction

• Applying an N ≥10 respondent threshold to four crossed demographic dimensions yields 63 valid personas: 23 Chinese, 23 Slovak, and 17 American.The study excludes social class because some upper-class cells had fewer than five respondents; the threshold is used to avoid sparse reference distributions.Source: §3.2 · p. 8

3.3 Wasserstein Distance Metric

• Normalized W1 measures distributional distance while respecting the ordinal response scale; scores range from 0 for perfect alignment to 1 for maximum divergence.The metric is normalized by dividing raw W1 by 9, the maximum distance on the 1–10 scale.Source: §3.3 · p. 9
• Worst-case personas are selected using the lower bound of their 95% bootstrap confidence interval, favoring robust misalignment over sensitivity to noisy small-sample estimates.The procedure resamples human responses 1,000 times; it may miss problematic small-sample personas, but selected personas are robustly misaligned.Source: §3.3 · p. 10

3.4 Model Selection

• Gemma3-12B, Bielik-11B-v3, and Qwen3-4B represent the United States, Poland, and China, respectively, but their 12B, 11B, and 4B sizes confound direct comparisons.Gemma3 is baseline-only because its Ollama deployment does not expose weights for LoRA; Bielik and Qwen are fine-tuning targets.Source: §3.4 · p. 10

3.5 Fine-Tuning Procedure

• Fine-tuning uses persona-conditioned supervised pairs built from non-religious WVS questions, holding out Q164 to test transfer rather than memorization.The sets contain 1,180 pairs for Bielik and 1,003 for Qwen; training completes in under 15 minutes on a single A100.Source: §3.5 · pp. 11–12
• After LoRA fine-tuning, each model is evaluated on all 63 personas with 100 new responses per persona to compare targeted gains and effects on the other 58.Gemma3 is excluded from fine-tuning because its Ollama deployment does not expose the underlying weights.Source: §3.5 · p. 12, Table 4 · p. 12

4.1 Computational Infrastructure

• Gemma3-12B uses Ollama, which precludes PEFT LoRA fine-tuning, while Bielik-11B-v3 and Qwen3-4B use 4-bit Transformers serving on one A100 GPU.The cluster provides A100 80 GB GPUs; the two quantized models require approximately 6–8 GB each, and code and data are publicly available.Source: §4.1 · pp. 12–13

4.2 Query Protocol

• Across 31,500 queries, each valid persona received 100 queries per model at baseline; every output parsed as a valid 1–10 integer, with a 0% refusal or parsing-failure rate.Prompts used temperature 0.7, top-k 50, and a 10-token maximum.Source: §4.2 · p. 13

4.3 Statistical Analysis

• For cultural-origin bias, a 3×3 model-by-country ANOVA tests the interaction; Tukey HSD supports post-hoc comparisons.The analysis uses 189 persona-model observations.Source: §4.3 · p. 13
• Fine-tuning efficacy is tested with paired t-tests on five targets, while paired tests on 58 non-target personas assess collateral effects.Country-level decomposition examines redistribution that aggregate statistics could mask; Bonferroni correction is applied across four RQ2 and RQ3 tests.Source: §4.3 · p. 13

5.1 Baseline Model Performance

• The human data showed distinct national response patterns, and Bielik differentiated personas best while Gemma3 and Qwen3 concentrated responses at high values.Chinese responses concentrated at 1–2, Slovak responses were broad across 5–10, and American responses were bimodal. Bielik reproduced the low Chinese pattern and moderate-to-high Slovak and American pattern, but missed America’s strong mode at 10; Gemma3 concentrated near 7, while Qwen3 stayed near 7–8 even for Chinese personas.Source: §5.1 · p. 14, Figure 2 · p. 15

5.2 Cultural Origin Bias Analysis (RQ1)

• No model had its lowest distance on its origin-aligned population; Qwen performed worst on Chinese personas, with distance 0.436—the highest in the 3×3 matrix.Source: Figure 3 · p. 16, §5.2 · p. 16

5.3 Fine-Tuning Efficacy (RQ2)

• Bielik’s target-persona bias fell 16.8% after LoRA fine-tuning, while Qwen’s rose 13.9% without Bonferroni-significant worsening.All five Bielik targets improved (pBonferroni = 0.0024, Cohen’s d = −4.388); four of five Qwen targets worsened (pBonferroni = 0.094). Bielik’s adapter trained on non-religious WVS questions improved held-out Q164, whereas Qwen’s adapter amplified its bias.Source: §5.3 · p. 17, Table 7 · p. 18

5.4 Side Effects Analysis (RQ3)

• Neither model showed significant aggregate side effects, but country-level results revealed redistribution: Bielik’s USA-persona LoRA worsened Chinese (+50.2%) and improved Slovak (−10.4%) outcomes; Qwen’s CHN-persona training improved American (−13.5%) and worsened Chinese (+27.6%) outcomes.Source: Table 9 · p. 18, §5.4 · p. 20
• Fine-tuning improves alignment for culturally similar populations while degrading alignment for culturally distant ones.This redistribution makes subgroup-level evaluation important in bias mitigation.Source: §5.4 · p. 22
• Aggregate metrics alone are insufficient to detect this redistribution, underscoring the need for subgroup-level evaluation in bias mitigation.Source: §5.4 · p. 22

6.1 Interpretation of Findings

• Lower ˆW1 on Chinese personas can reflect Q164’s peaked distribution: 52% of respondents answered 1, making it geometrically easier to approximate than broader Slovak or US distributions, not greater cultural competence.Cross-cultural bias comparisons must account for target-distribution difficulty because aggregate metrics can mislead when target distributions differ in concentration.Source: §6.1 · p. 22
• Qwen3-4B’s Chinese-persona distance is 0.436, the highest in the model–country matrix; the authors interpret its moderate-to-high religiosity responses as potentially linked to RLHF.They propose that safety preferences may steer outputs away from low-religiosity responses, while presenting this as an interpretation of the pattern.Source: §6.1 · p. 22
• Fine-tuning redistributes bias through global directional shifts, and side-effect magnitude correlates with cultural distance from the training targets.The reported correlations are r = 0.57 for Bielik and r = −0.71 for Qwen; the discussion also notes Qwen’s target-training failure and possible confounds in the models’ divergent LoRA outcomes.Source: §6.1 · pp. 22–23
• Bielik’s five worst-case personas shift entirely from American before fine-tuning to Chinese elderly afterward, with zero overlap.The discussion describes this as a “whack-a-mole” dynamic in which fixing one group can leave new worst-case personas elsewhere.Source: §6.1 · p. 23, Figure 8 · p. 24

6.2 Practical Implications

• The proposed assessment framework identifies worst cases using bootstrap confidence intervals, mitigates with targeted LoRA, and evaluates side effects by subgroup.The authors recommend lower confidence-interval bounds for selecting targets and warn that aggregate non-significance can mask subgroup redistribution.Source: §6.2 · p. 24

6.3 Limitations

• Generalization is limited by the single religiosity question, three-country scope, single prompt template, and unexamined higher-order demographic interactions.The authors also note that Slovak data may not perfectly represent Polish respondents and that Gemma3’s baseline interpretation may be affected by inference determinism.Source: §6.3 · pp. 24–25

6.4 Future Work

• Future work includes evaluating more value questions and countries, testing constraint-based LoRA and DPO, tracking alignment across survey waves, and analyzing demographic interactions.The authors describe country-level decomposition as revealing redistribution and worst-case turnover that aggregate statistics miss, and identify this as a new analytical lens for bias-mitigation evaluation.Source: §6.4 · p. 25

6.5 Concluding Remarks

• 16.8%: Targeted LoRA reduced bias for Bielik-11B across the five worst-case personas, with all five improving (d = −4.4).The study reports using fewer than 1,200 training pairs and under 15 minutes of computation on a single GPU.Source: §6.5 · p. 25

• Each WVS persona is converted into a prompt containing a system instruction, demographic profile, and survey question, then queried 100 times to estimate a response distribution.Source: Figure 9 · p. 29
• Bootstrap confidence-interval width falls monotonically as human sample size increases, while medium and higher education levels dominate the worst-case lists.Example comparisons include Bielik’s Chinese best case at W1 = 0.111, Qwen’s Chinese worst case at W1 = 0.637, and Bielik’s USA_Male_30-49_Medium change from W1 = 0.365 to 0.289 after LoRA.Source: Figure 10 · p. 30, Figure 13 · p. 32, Figure 14 · p. 33
• Baseline distances span a wider range for Qwen3-4B, with the highest individual values concentrated among Chinese personas.Figure 11 plots personas as dots colored by country and sorted by distance within each model panel.Source: Figure 11 · p. 30
• Chinese non-target personas worsen by 27.6% on average after Qwen+LoRA, while American personas improve by 13.5% and Slovak aggregate change is negligible.The figure reports worsening for 16 of 18 Chinese personas and improvement for 15 of 17 American personas.Source: Figure 12 · p. 31