Prototype-Driven Contrastive Alignment for Survey-Grounded Cultural-Value Adaptation in Large Language Models

Culturally inclusive large language models (LLMs) should adapt to heterogeneous value systems without conflating national survey aggregates with comprehensive intercultural competence. We introduce Prototype-driven Contrastive Cultural Alignment (ProCA), a training framework that initializes learnable country-level prototypes from World Values Survey (WVS) answer distributions, generates prototype-guided synthetic social interactions, combines WVS-distribution Kullback–Leibler (KL) supervision with goal-conditioned realization–prototype contrastive learning, and applies score-based iterative refinement. Across eight instruction-tuned backbones and five WVS Wave 7 country strata, the complete ProCA configuration yields lower in-domain KL-D and higher respondent-level accuracy than the reported Standard and Cultural Prompting baselines. On GPT-OSS 120B, KL-D changes from 0.3909 to 0.3395 and accuracy from 0.4425 to 0.4832; data-matched Naive Dialogue SFT reaches 0.3652 and 0.4667, respectively. A paired 10,000-resample bootstrap over the five culture-level KL-D differences gives a mean reduction of 0.05136 versus Standard (95% percentile interval [0.04728, 0.05646]), but this interval captures across-culture variation rather than training-seed uncertainty. Crucially, the archived experiments use the same WVS item–country marginals for KL supervision and primary evaluation. Completed numerical runs are unavailable for the prespecified 30/7/7 held-out-item split, the independent CultureBank evaluation, the supervision-matched KL-only LoRA control, the no-refinement and randomized-prototype controls, and the multi-seed reruns. We therefore interpret the current evidence as objective-matched WVS cultural-value adaptation, not as proof of held-out cultural generalization, broad intercultural competence, or the causal contribution of each ProCA component. The revised protocol specify the exact held-out split, matched baselines, component ablations, seed-level uncertainty analysis, prediction-output schema, prototype construction/refinement records, and runtime logging required for a confirmatory evaluation.