Prompt Programming for Cultural Bias and Alignment of Large Language Models

Maksim Eren; Eric Michalak; Brian Cook; Johnny Seales

cs.AIMar 17, 2026

Prompt Programming for Cultural Bias and Alignment of Large Language Models

Authors: Maksim Eren, Eric Michalak, Brian Cook, Johnny Seales

Organizations: 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

Abstract

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.

Figures & tables

Figure 1. Cultural map of countries/territories in the IVS benchmark space (Survival vs. Self-Expression; Traditional vs. Secular values). We overlay points derived from open-weight model responses (Llama 3.3 70B, Llama 4 16

\times

17B, Gemma 3 27B, GPT-OSS 20B, GPT-OSS 120B) to the same IVS items, using the projection procedure introduced by Tao et al. ( Tao et al., 2024 ) . Shaded regions are visual guides highlighting approximate cultural-zone clusters and are not used in distance calculations.

Figure 2. Country-level cultural distance for open-weight LLMs under three prompting regimes: (i) without culture conditioning, (ii) with manual culture prompt engineering, and (iii) culture prompt programming with DSPy. Cultural distance is measured, for each country/territory

c

, as the Euclidean distance between the model’s projected coordinate in the IVS cultural-map space and the corresponding human reference coordinate

\boldsymbol{\nu}^{\mathrm{IVS}}_{c}

. For DSPy prompt-programming conditions, distances are computed on held-out test-set countries across 5-fold country-level cross-validation. The no-culture-prompt and manual culture-prompt conditions are baseline country-level distance distributions shown for comparison. Smaller distances indicate closer alignment to the country/territory benchmark, and changes across regimes reflect how cultural conditioning and prompt programming shift the model toward

\boldsymbol{\nu}^{\mathrm{IVS}}_{c}

Figure 3. Per-country movement in the Inglehart–Welzel IVS benchmark space (PC1 ′ /PC2 ′ ; Survival vs. Self-Expression and Traditional vs. Secular) for GPT-OSS:120B after Culture Prompt Programming with DSPy (MIPROv2; proposer GPT-OSS:120B). Results are calculated across both compilation and held-out countries over the 5-fold country-level cross-validation runs. Each mini-panel corresponds to one country/territory

c

, grouped by cultural zone using the background tint, and compares the generic non-national projection

\boldsymbol{\mu}_{m,\varnothing}

(Generic), the aligned country-conditioned projection

\boldsymbol{\mu}^{\mathrm{DSPy}}_{m,c}

(Aligned), and the human IVS reference point

\boldsymbol{\nu}^{\mathrm{IVS}}_{c}

(Human). Arrows show the shift from Generic to Aligned, dashed segments show the remaining alignment gap

\lVert\boldsymbol{\mu}^{\mathrm{DSPy}}_{m,c}-\boldsymbol{\nu}^{\mathrm{IVS}}_{c}\rVert_{2}

, and

\Delta(c)>0

indicates that alignment moves the model closer to the human benchmark for

c

Prompting setting	llama3.3:70B	llama4:16x17B	gemma3:27b	gpt-oss:20b	gpt-oss:120b
No culture prompt	2.217 ( $\pm$ 0.16)	2.819 ( $\pm$ 0.19)	3.085 ( $\pm$ 0.20)	2.789 ( $\pm$ 0.19)	3.153 ( $\pm$ 0.19)
Culture prompt engineering	0.973 ( $\pm$ 0.11)	2.006 ( $\pm$ 0.12)	2.008 ( $\pm$ 0.13)	1.688 ( $\pm$ 0.11)	1.778 ( $\pm$ 0.10)
DSPy MIPROv2, llama3.2:1b	1.496 ( $\pm$ 0.16)	1.229 ( $\pm$ 0.10)	2.182 ( $\pm$ 0.15)	1.844 ( $\pm$ 0.13)	1.770 ( $\pm$ 0.12)
DSPy MIPROv2, gpt-oss:120b	1.350 ( $\pm$ 0.15)	0.818 ( $\pm$ 0.08)	1.499 ( $\pm$ 0.13)	1.369 ( $\pm$ 0.11)	1.364 ( $\pm$ 0.12)
DSPy COPRO, llama3.2:1b	1.206 ( $\pm$ 0.13)	1.295 ( $\pm$ 0.10)	2.400 ( $\pm$ 0.17)	1.700 ( $\pm$ 0.12)	2.015 ( $\pm$ 0.13)
DSPy COPRO, gpt-oss:120b	1.021 ( $\pm$ 0.10)	1.163 ( $\pm$ 0.09)	2.250 ( $\pm$ 0.16)	1.695 ( $\pm$ 0.11)	1.860 ( $\pm$ 0.13)

Table 1. Mean cultural distance (

\overline{d}

) is reported with 95% confidence intervals for each model and prompting setting. Distance is measured as Euclidean distance in the IVS cultural-map space between each model projection and the corresponding country/territory reference point

\boldsymbol{\nu}^{\mathrm{IVS}}_{c}

. For DSPy prompt-programming rows, distances are computed on held-out test-set countries across 5-fold country-level cross-validation and correspond to

d_{\mathrm{test}}

from Equation 16 . Bold values indicate the lowest distance within each model column, while underlined values indicate the lowest distance within each prompting-setting row.

LLM Alignment

Large language model alignment studies methods for making models follow human intentions, values, and safety constraints across contexts.

See all 598 papers

AlignCultura: Towards Culturally Aligned Large Language Models?
Apr 21, 2026Gautam Siddharth Kashyap, Mark Dras, Usman NaseemCross-Cultural Language Model Evaluation Cultural Alignment
School of Computing, Macquarie University, Australia
Scenario-based Probing and Steering Cultural Values in Large Language Models--Extended Version
Jun 9, 2026Trung Duc Anh Dang, Tung Kieu, Sarah MasudLanguage Model Steering Cultural Alignment
University of Copenhagen · Aalborg University
Mind the Gap in Cultural Alignment: Task-Aware Culture Management for Large Language Models
Feb 25, 2026Binchi Zhang, Xujiang Zhao, Jundong Li +2Cultural Alignment Cultural Knowledge in Language Models
University of Virginia · NEC Laboratories America

Apr 21, 2026J/K move · Enter open · S save

AlignCultura: Towards Culturally Aligned Large Language Models?

Gautam Siddharth Kashyap, Mark Dras, Usman Naseem

School of Computing, Macquarie University, Australia

Cross-Cultural Language Model Evaluation Cultural Alignment Cultural Bias in Language Models Language Model Bias Evaluation Cultural Knowledge in Language Models

Open paper AskPDF arXiv

Cultural alignment in Large Language Models (LLMs) is essential for producing contextually aware, respectful, and trustworthy outputs. Without it, models risk generating stereotyped, insensitive, or misleading responses that fail to reflect cultural diversity w.r.t Helpful, Harmless, and Honest (HHH) paradigm. Existing benchmarks represent early steps toward cultural alignment; yet, no benchmarks currently enables systematic evaluation of cultural alignment in line with UNESCO's principles of cultural diversity w.r.t HHH paradigm. Therefore, to address this gap, we built Align-Cultura, two-stage pipeline for cultural alignment. Stage I constructs CULTURAX, the HHH-English dataset grounded in the UNESCO cultural taxonomy, through Query Construction, which reclassifies prompts, expands underrepresented domains (or labels), and prevents data leakage with SimHash. Then, Response Generation pairs prompts with culturally grounded responses via two-stage rejection sampling. The final dataset contains 1,500 samples spanning 30 subdomains of tangible and intangible cultural forms. Stage II benchmarks CULTURAX on general-purpose models, culturally fine-tuned models, and open-weight LLMs (Qwen3-8B and DeepSeek-R1-Distill-Qwen-7B). Empirically, culturally fine-tuned models improve joint HHH by 4%-6%, reduce cultural failures by 18%, achieve 10%-12% efficiency gains, and limit leakage to 0.3%.

Abstract

LLM Alignment

Explore similar work

AlignCultura: Towards Culturally Aligned Large Language Models?

Scenario-based Probing and Steering Cultural Values in Large Language Models--Extended Version

Mind the Gap in Cultural Alignment: Task-Aware Culture Management for Large Language Models