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

Antoni Czolgowski; Abel Iyasele

cs.CLSep 3, 2026

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

Authors: Antoni Czolgowski, Abel Iyasele

Organizations: Masters in Data Science Program University of Colorado Boulder Boulder, CO 80309, USA · Department of Information Science University of Colorado Boulder Boulder, CO 80309, USA

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.

Figures & tables

Country	Code	Mean	Std Dev	Character	$N_{\text{WVS}}$
China	CHN	2.8	${\sim}2.5$	Strongly secular	${\sim}3{,}000$
Slovakia	SVK	6.4	${\sim}3.3$	Moderately religious	${\sim}1{,}200$
United States	USA	6.7	${\sim}3.4$	Highly religious	${\sim}2{,}500$

Table 1: WVS Wave 7 country profiles for Q164 (“Importance of God”). Slovakia serves as the Central European proxy for Poland, which is absent from Wave 7.

Dimension	Categories	Source
Country	China (CHN), Slovakia (SVK), USA	WVS B_COUNTRY
Sex	Male, Female	WVS Q260
Age group	18–29, 30–49, 50–64, 65+	WVS Q262 (binned)
Education	Lower (ISCED 0–2), Medium (3–4), Higher (5–8)	WVS Q275R
Theoretical combinations: $3\times 2\times 4\times 3=72$
Valid personas ( $N\geq 10$ ): 63 (23 CHN, 23 SVK, 17 USA)

Table 2: Persona construction design. The four demographic dimensions are fully crossed, subject to a minimum sample size threshold of

N\geq 10

WVS respondents per cell.

Figure 1: Illustration of the Wasserstein

W_{1}

distance for the persona USA_Male_30-49_Medium (Bielik-11B). (a) Response distributions (PMFs): the human distribution (blue) is broadly dispersed with a mode at 10, while the model (orange) concentrates at 5. (b) Cumulative distributions: the shaded area between the CDFs equals the

W_{1}

distance (

\hat{W}_{1}=0.365

Model	Origin	Params	Arch.	Quant.	Role
Gemma3-12B	USA (Google)	12B	Gemma	None (Ollama)	Baseline
Bielik-11B-v3	Poland (SpeakLeash)	11B	Mistral	4-bit BnB	Baseline + LoRA
Qwen3-4B	China (Alibaba)	4B	Qwen	4-bit BnB	Baseline + LoRA

Table 3: Model specifications. All models are open-weight and locally deployable on a single NVIDIA A100 GPU.

	Bielik-11B	Qwen3-4B
Target personas	5 (USA)	5 (CHN)
Training pairs	1,180	1,003
LoRA rank ( $r$ )	16	16
LoRA alpha ( $\alpha$ )	32	32
LoRA dropout	0.05	0.05
Target modules	q_proj, v_proj	q_proj, v_proj

Table 4: LoRA fine-tuning configuration and training statistics. Gemma3-12B is excluded because its Ollama deployment does not expose model weights for fine-tuning.

Figure 2: Response fingerprints: PMF heatmaps across all 63 personas for human WVS data and three models. Columns represent personas sorted by country (China, Slovakia, USA); rows represent response values (1–10); color intensity indicates probability mass. The human panel reveals three distinct national patterns. Models differ markedly in their ability to reproduce this structure.

Source	df	$F$	$p$	Significance
Model	2, 180	56.54	$<2\times 10^{-16}$
Country	2, 180	5.25	0.006	**
Model $\times$ Country	4, 180	26.38	$<2\times 10^{-16}$

Table 5: Two-way ANOVA results for

\hat{W}_{1}

as a function of model and persona country. All effects are statistically significant. The model

\times

country interaction confirms that bias patterns differ across models.

	China	Slovakia	USA
Bielik-11B (Poland)	0.173	0.249	0.288
Gemma3-12B (USA)	0.199	0.306	0.293
Qwen3-4B (China)	0.436	0.286	0.348

Table 6: Mean normalized Wasserstein distance (

\hat{W}_{1}

) by model and persona country. Lower values indicate better alignment with human survey data. Bordered cells in Figure 3 indicate each model’s cultural origin country.

Figure 3: Mean Wasserstein distance by model origin and persona country. Bordered cells indicate the model’s cultural origin country or closest available proxy. If models favored their origin, bordered cells would show the lowest values in each row. Instead, no model performs best on its origin-aligned population: Gemma and Bielik obtain their lowest distances on Chinese personas, while Qwen performs worst on Chinese personas and best on Slovak personas.

Persona	$\hat{W}_{1}$ Before	$\hat{W}_{1}$ After	$\Delta$	% Change
Bielik-11B (targets: USA personas)
USA_Male_30-49_Medium	0.365	0.289	$-$ 0.076	$-$ 20.8%
USA_Male_18-29_Medium	0.355	0.287	$-$ 0.068	$-$ 19.1%
USA_Male_50-64_Higher	0.340	0.290	$-$ 0.050	$-$ 14.7%
USA_Female_18-29_Medium	0.327	0.277	$-$ 0.050	$-$ 15.3%
USA_Male_18-29_Higher	0.321	0.276	$-$ 0.046	$-$ 14.2%

Table 7: RQ2: Fine-tuning efficacy on worst-case target personas. Negative

\Delta\hat{W}_{1}

indicates improvement (bias reduction). Bielik achieved statistically significant improvement across all five targets; Qwen worsened on four of five.

Figure 4: CDF overlays before and after LoRA fine-tuning for the worst-case target persona of each model. Top row: Bielik-11B on USA_Male_30-49_Medium ; the model CDF shifts toward the human CDF, reducing the shaded

W_{1}

area from 0.365 to 0.289. Bottom row: Qwen3-4B on CHN_Male_65+_Medium ; the model CDF shifts further from the human CDF, increasing

W_{1}

from 0.637 to 0.723. Annotations show

\hat{W}_{1}

with 95% bootstrap confidence intervals.

Model	$n$	Mean $\Delta\hat{W}_{1}$	$t$	df	$p_{\text{Bonf}}$	Cohen’s $d$
Bielik-11B	58	$+$ 0.019	1.775	57	0.325	0.233
Qwen3-4B	58	$+$ 0.012	1.244	57	0.875	0.163

Table 8: RQ3: Aggregate side effects on non-target personas. Neither model shows statistically significant aggregate change after Bonferroni correction.

Model	Country	$n$	Mean $\Delta\hat{W}_{1}$	% Change	Improved	Worsened
Bielik-11B	China	23	$+$ 0.077	$+$ 50.2%	3	20
	Slovakia	23	$-$ 0.032	$-$ 10.4%	20	3
	USA (non-target)	12	$+$ 0.008	$+$ 6.3%	6	6
Qwen3-4B	China (non-target)	18	$+$ 0.085	$+$ 27.6%	2	16
	Slovakia	23	$-$ 0.001	$+$ 1.6%	9	10
	USA	17	$-$ 0.047	$-$ 13.5%	15	1

Table 9: RQ3: Country-level decomposition of side effects on non-target personas. The aggregate non-significance masks a dramatic redistribution: Bielik’s LoRA training on USA personas devastates Chinese personas (

+50.2\%

) while improving Slovak personas (

-10.4\%

). Qwen’s training on CHN personas improves American personas (

-13.5\%

) while further worsening Chinese non-targets (

+27.6\%

Figure 5: Bielik+LoRA: side effects on all 58 non-target personas, grouped by country. Bars extending right (red) indicate increased

\hat{W}_{1}

(worsened bias); bars extending left (green) indicate decreased

\hat{W}_{1}

(improved bias). Training on American personas devastated Chinese personas (

+50.2\%

mean) while incidentally improving Slovak personas (

-10.4\%

mean). The Qwen redistribution pattern is presented in Appendix A.

Figure 6: Before-and-after

\hat{W}_{1}

distributions for all 63 personas, by model. Each dot represents one persona, colored by country (pink = China, green = Slovakia, yellow = USA). Star markers with connecting lines indicate the five target personas. Left panel (Bielik): targets drop from 0.32–0.37 to 0.28–0.29; Chinese non-targets shift upward; Slovak non-targets shift downward. Right panel (Qwen): targets rise from 0.53–0.64 to 0.52–0.72; American non-targets shift downward.

RQ	Model	Mean $\Delta\hat{W}_{1}$	$t$	$p_{\text{Bonf}}$	Cohen’s $d$	Sig.
RQ2 (targets)	Bielik	$-$ 0.058	$-$ 9.811	0.0024	$-$ 4.388
RQ2 (targets)	Qwen	$+$ 0.082	3.563	0.094	$+$ 1.593	n.s.
RQ3 (non-targets)	Bielik	$+$ 0.019	1.775	0.325	0.233	n.s.
RQ3 (non-targets)	Qwen	$+$ 0.012	1.244	0.875	0.163	n.s.

Table 10: Summary of all statistical tests (RQ2 and RQ3). Bonferroni correction applied across 4 tests (

\alpha_{\text{adjusted}}=0.0125

). Only Bielik’s RQ2 result reaches significance.

Figure 7: Bias redistribution scales with cultural distance from the fine-tuning target. Each dot represents a non-target persona, colored by country. The

x

-axis measures the absolute difference between the persona’s country mean on Q164 and the training target’s country mean. Left: Bielik (

r=0.57

); personas distant from American religiosity (i.e., Chinese secular personas) worsen most. Right: Qwen (

r=-0.71

); personas distant from Chinese secularity (i.e., American personas) improve most, despite the fine-tuning failing on its actual targets.

Figure 8: Bielik+LoRA: complete worst-case turnover. Before fine-tuning, all five worst-case personas were American; after fine-tuning, all five are Chinese elderly. The overlap between pre- and post-LoRA worst-case sets is zero, a complete cultural swap.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 9: Persona-based survey prompting pipeline. A demographic persona card from the WVS is translated into a structured LLM prompt consisting of a system instruction, a demographic profile, and the survey question. Each persona is queried 100 times to construct an empirical response distribution.

Figure 10: Bootstrap 95% confidence interval width as a function of human sample size (

N_{\text{WVS}}

). CI width decreases monotonically with sample size, confirming that the bootstrap procedure produces tighter estimates for personas with more WVS respondents. The dashed line shows the linear trend.

Figure 11: Baseline

\hat{W}_{1}

distances for all 63 personas across three models. Each dot represents one persona, colored by country. Personas are sorted within each model panel by

\hat{W}_{1}

. Qwen3-4B exhibits the widest spread and the highest individual values, concentrated among Chinese personas.

Figure 12: Qwen+LoRA: side effects on all 58 non-target personas, grouped by country. Compare with Figure 5 for Bielik. The mirror-image pattern is evident: Chinese non-targets worsen (

+27.6\%

mean, 16 of 18 worsened) while American personas improve (

-13.5\%

mean, 15 of 17 improved). Slovak personas show negligible aggregate change.

Figure 13: Education level distribution of worst-case personas across the three models. Medium and higher education levels dominate the worst-case lists, contrary to the intuitive expectation that lower-education personas, typically underrepresented in LLM training data, would be most poorly served.

Figure 14: Six example response distribution comparisons (human PMF in blue, model PMF in orange). Cases illustrate: best-case alignment (CHN, Bielik,

\hat{W}_{1}=0.111

), worst-case misalignment (CHN, Qwen,

\hat{W}_{1}=0.637

), moderate alignment (SVK, Bielik), worst Slovak case (SVK, Gemma3,

\hat{W}_{1}=0.474

), and Bielik before vs. after LoRA on USA_Male_30-49_Medium (

\hat{W}_{1}

0.365\to 0.289

Cultural Bias in Language Models

Cultural bias in language models is the uneven representation of cultural knowledge, values, identities, and norms across groups and languages.

See all 129 papers

Aug 30, 2026

CoCoA: Context-Conditional Cultural Alignment for Large Language Models

Large Language Models (LLMs) often favor Western-associated entities across cultural contexts. Conventional debiasing methods aim for uniform neutrality, but cultural bias mitigation demands context-conditional behavior, preferring culturally appropriate entities when cultural cues are present and remaining neutral when they are absent. We propose CoCoA (Context-Conditional Cultural Alignment), a framework that learns this behavior through dual-context training on the same entity pairs under contexts with and without cultural cues. CoCoA combines a contrastive alignment objective with calibration and drift regularization, optimized through goal-aware gradient reconciliation. We evaluate CoCoA on CAMeL and Camellia, two entity-centric cultural bias benchmarks, across ten language settings and four LLMs. CoCoA reduces the Cultural Bias Score from 43 to 24 on average while maintaining near-neutral preferences at 50.2, with minimal impact on general performance across five standard benchmarks. These findings highlight that effective cultural alignment requires context-conditional modeling rather than uniform debiasing, and establish a new direction for mitigating entity-centric cultural bias in LLMs.

Kyungdon Lee, Wei Xu, Alan Ritter +2

Sungkyunkwan University, Suwon, South Korea · Georgia Institute of Technology, GA, USA · University of Southern California, CA, USA

Cultural Bias in Language Models Language Model Debiasing Cultural Alignment LLM Alignment Cross-Cultural Language Model Evaluation

Ask PDF

Mar 17, 2026

Prompt Programming for Cultural Bias and Alignment of Large Language Models

Culture shapes reasoning, values, prioritization, and strategic decision-making, yet large language models (LLMs) often exhibit cultural biases that misalign with target populations. As LLMs are increasingly used for strategic decision-making, policy support, and document engineering tasks such as summarization, categorization, and compliance-oriented auditing, improving cultural alignment is important for ensuring that downstream analyses and recommendations reflect target-population value profiles rather than default model priors. Previous work introduced a survey-grounded cultural alignment framework and showed that culture-specific prompting can reduce misalignment, but it primarily evaluated proprietary models and relied on manual prompt engineering. In this paper, we validate and extend that framework by reproducing its social sciences survey based projection and distance metrics on open-weight LLMs, testing whether the same cultural skew and benefits of culture conditioning persist outside closed LLM systems. Building on this foundation, we introduce use of prompt programming with DSPy for this problem-treating prompts as modular, optimizable programs-to systematically tune cultural conditioning by optimizing against cultural-distance objectives. In our experiments, we show that prompt optimization often improves upon cultural prompt engineering, suggesting prompt compilation with DSPy can provide a more stable and transferable route to culturally aligned LLM responses.

Maksim Eren, Eric Michalak, Brian Cook +1

Advanced Research in Cyber Systems, Los Alamos National Laboratory, Los Alamos, New Mexico, USA · Analytics Division, Los Alamos National Laboratory, Los Alamos, New Mexico, USA

LLM Alignment Cultural Alignment Cultural Bias in Language Models Cross-Cultural Language Model Evaluation Prompt Optimization

Ask PDF

May 29, 2026

Anchoring LLM Gender Bias to Human Baselines: A Cross-Lingual Audit

We audit six large language models (LLMs) for gender stereotyping across English, Korean, Chinese, and Japanese. Three were developed primarily for English-language use (Claude, GPT, Gemini) and three for East Asian use (DeepSeek, Syn-Pro, HyperCLOVA X). We adopt the HEXACO-100 personality inventory and anchor each model against a cross-cultural human dataset spanning 48 countries to ask not whether LLMs are biased, but how far their gender attributions drift from the populations they are deployed among. Our findings show that their stereotyping spans a range roughly 2.5 times wider than the entire cross-country range found in humans, and the effect can compound across languages. One English-centric model, prompted in Korean, reached 5 times the local baseline, even when the prompt stated the candidate had already been hired, which often dampens human stereotyping. To characterize such behaviors without ranking them, we introduce a four-pattern framework -- concordance, suppression, reorganization, and amplification -- across 24 (model x language) cells. Item-level analysis reveals that translation does not just rescale stereotypes, but changes the attributes tied to it, hiding significant rearrangement under the surface while appearing well-calibrated. Our results ultimately suggest that no single debiasing pipeline is likely to address bias evenly across linguistic boundaries.

Jiwoo Choi, Seonwoo Ahn, Tongxin Zhang +1

School of Digital Humanities and Computational Social Sciences, KAIST, South Korea

Gender Bias in Language Models Multilingual Language Model Evaluation LLM Auditing Cultural Bias in Language Models Language Model Bias Evaluation Linguistic Bias in Language Models

Ask PDF

Abstract

Cultural Bias in Language Models

Explore similar work

CoCoA: Context-Conditional Cultural Alignment for Large Language Models

Prompt Programming for Cultural Bias and Alignment of Large Language Models

Anchoring LLM Gender Bias to Human Baselines: A Cross-Lingual Audit