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

Recent research evaluated the cultural alignment of three open-weight Large Language Models (LLMs) – Gemma3-12B (USA), Bielik-11B-v3 (Poland), and Qwen3-4B (China) – against World Values Survey data for 63 demographic personas across three countries. Contrary to expectations, models did not favor their country of origin, with the Chinese-built Qwen3-4B showing the highest misalignment with its own Chinese population. The study also demonstrated that targeted low-rank adaptation (LoRA) fine-tuning can significantly reduce cultural bias with minimal resources.

Why it matters

The findings highlight the inherent cultural biases in Large Language Models and the inadequacy of geographic origin as a proxy for cultural alignment. This research provides a scalable and resource-efficient method for mitigating such biases, which is crucial for the global deployment and trustworthiness of AI systems across diverse populations and markets.

Key insights

• Three open-weight LLMs (Gemma3-12B, Bielik-11B-v3, Qwen3-4B) were evaluated for cultural alignment using World Values Survey Wave 7 data for 63 demographic personas across three countries.
• Cultural misalignment was quantified using normalized Wasserstein distance.
• No observed favoritism for an LLM's home country; specifically, Qwen3-4B performed worst on its own Chinese population (W1 = 0.436).
• Targeted LoRA fine-tuning on five worst-case personas, using fewer than 1,200 training pairs and under 15 minutes on a single GPU, reduced bias by 16.8% for Bielik-11B.
• All five targeted personas showed improvement post-fine-tuning, indicating the efficacy of the mitigation strategy.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2609.04485