Large language models show different social values across languages
Same Values, Different Languages? From Multilingual Probing to Steering LLMs Toward Chinese Social Values
Computation and LanguageArtificial Intelligence
Summary
Large language models (LLMs) need to align with social values to work well in society. This paper shows that LLMs do not always prefer the same social values when responding in different languages, especially for Chinese cultural values. The authors created a big dataset to test these values in six languages and developed a method to adjust models towards Chinese social values without extra training. Their approach can transfer value adjustments across languages and to other value systems.
What this means in practice
- •For multilingual ai developers: Adjust language models to better reflect culturally specific social values across different languages during inference.
- •For content moderation teams: Improve detection and steering of culturally sensitive value conflicts in multilingual content generated by AI systems.
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.