As large language models (LLMs) are increasingly integrated into intelligent software systems that range from educational assistants to legal advisory tools, ensuring their alignment with diverse cultural values becomes a critical software engineering challenge. Traditional software engineering methods such as unit testing and formal verification fall short in specifying or validating complex normative expectations such as fairness, cultural sensitivity, and moral appropriateness. Addressing this gap, we introduce VADA, a value-aware development and evaluation framework for systematically testing and benchmarking LLMs under multiple cultural value systems. VADA incorporates a modular scenario-question generation pipeline that constructs culturally grounded test cases spanning 25 value dimensions across Chinese, European, and Islamic ethical frameworks. It further includes a Bayesian ensemble evaluation framework that aggregates alignment judgments from multiple diverse LLM-based evaluators, assigning dimension-specific trust weights based on their observed reliability. We also develop a lightweight supervised evaluator fine-tuned on VADA-generated data, providing a scalable and deployable alternative to multi-model ensemble evaluation. We construct a large-scale benchmark containing 11,865 automatically annotated cases, along with a human-labeled subset of 1000 instances for validation and evaluation.

Empirical results demonstrate that VADA substantially outperforms existing prompting-based evaluators, achieving over 93.6 percent accuracy and high agreement with human annotations. Ablation studies confirm the complementary benefits of evaluator diversity and reliability-aware aggregation. Furthermore, VADA enables comparative audits of state-of-the-art LLMs, uncovering alignment inconsistencies across cultural dimensions. Results highlight VADA’s effectiveness as a foundation for robust value-aware evaluation and development of LLMs.