A Resource-Efficient Framework for Cultural Alignment in Large Language Models (LLMs): The Chinese Context

• We propose a cost-effective Human-AI cultural alignment design framework to enhance LLMs’ cultural understanding and conversational capabilities in Chinese contexts, focusing on cultural values, culturally specific content, and conversational logic design. • We introduced a targeted fine-tuning framework that combines LoRA fine-tuning, Retrieval-Augmented Generation (RAG), and Chain-of- Thought (CoT) prompting to address cultural challenges without extensive resource requirements, emphasizing resource-efficient design. • We conduct extensive evaluations across six datasets using several stateof-the-art LLMs, demonstrating consistent and significant improvements in cultural sensitivity, comprehension accuracy, and user satisfaction,validating the effectiveness of our design approach. The rapid development of Artificial Intelligence (AI) has underscored the urgent need to align technological innovation with ethical standards and Environmental, Social, and Governance (ESG) principles. A critical challenge lies in the design and deployment of Large Language Models (LLMs), predominantly trained on English-centric data, that can understand and respond to diverse cultural contexts without perpetuating bias, discrimination, or societal inequalities, posing a risk to responsible AI system design. In this study, we propose a low-cost and scalable Human-AI cultural alignment design framework to enhance LLMs cultural sensitivity and ethical alignment, with a focus on the Chinese context as a representative case. Our approach uses LoRA fine-tuning, Retrieval-Augmented Generation (RAG), and Chain-of-Thought (CoT) prompting to address three design dimensions: (1) embedding cultural values into model reasoning, (2) improving comprehension of culturally specific content, and (3) fostering inclusive and ethically consistent conversational logic design. Experimental results across six downstream tasks demonstrate significant improvements in cultural awareness, fairness, and user trust, achieving performance comparable to GPT-4 and Gemini Pro while requiring fewer computational resources. Beyond cultural alignment, our method advances ESG goals: environmentally, by reducing the carbon footprint through lightweight fine-tuning; socially, by promoting inclusivity and reducing cultural bias through better UX design; and governance-wise, by offering a transparent, adaptable, and resource-efficient alignment strategy. This work highlights how culturally adaptive AI systems can contribute to a more equitable, sustainable, and ethically responsible technological future.