Summary
Large language models (LLMs) are computer programs that understand and generate language, and it's important to make sure they reflect social values appropriately. This study found that LLMs do not always express the same social values when using different languages, especially for values tied closely to culture. To test this, the researchers created a large set of questions based on Chinese social values and compared how LLMs responded in six languages. They also developed a way to guide the models to align better with these values without changing their main training. Their method worked across languages and suggests that social values in LLMs depend on both the model and the language it uses.
Large language modelsSocial valuesMultilingual evaluationCultural valuesChinese social valuesContrastive probingValue alignmentFine-tuning-free steeringCross-lingual transfer
Authors
Yuemei Xu, Kexin Xu, Jian Zhou, Haoyu Lu, Yequan Wang, Aishan Liu
Abstract
As Large Language Models (LLMs) are increasingly integrated into human society, aligning them with pluralistic social values has become a critical priority. However, whether LLMs exhibit consistent value preferences across languages remains underexplored, particularly for culturally grounded values, which are more abstract and difficult to evaluate and align than safety-centric principles. We investigate this issue through Chinese Social Values (CSV), a value system rooted in Chinese culture and comprising $12$ dimensions across national, societal, and personal levels. We construct C-Voices, the first comprehensive multilingual contrastive probe dataset for CSV, with 86,400 dilemma-based instances in six languages, each pairing a CSV-aligned action with a value-conflicting alternative. Building on the contrastive probes of C-Voices, we then propose a fine-tuning-free value vector steering method that derives value directions from hidden-state discrepancies and selectively intervenes on value-sensitive layers during inference. Experiments on six languages show that CSV-oriented preferences are model-dependent and language-sensitive, with the same dilemma eliciting divergent responses across languages. Our method achieves effective CSV steering, supports cross-lingual transfer of value vectors, and generalizes to existing FLAMES and ValuePrism.