Introduction
As Large Language Models (LLMs) become increasingly integrated into global society, the challenge of value alignment—ensuring AI behavior matches human ethics—has become paramount. However, current alignment efforts are predominantly lopsided toward Western cultural contexts. To address this gap, researchers have introduced C-VARC, the first large-scale Chinese Value Rule Corpus. Containing over 250,000 "Rules of Thumb" (RoTs), C-VARC is designed to align LLMs with Chinese cultural and ethical norms, providing a structured framework for generating culturally relevant evaluation scenarios and complex moral dilemmas.
The Challenge of Cultural Misalignment
Most existing value alignment benchmarks, such as Social Chemistry 101 (SC101) or the Moral Integrity Corpus (MIC), are grounded in Western "Moral Foundations Theory." When these rules are applied to Chinese contexts, they frequently clash with local values regarding sensitive topics like family obligations, social harmony, and political perspectives. Furthermore, existing Chinese benchmarks often suffer from limited scale, lack systematic coverage of contemporary values, or rely on manual scenario designs that do not scale effectively. This creates a "cultural misalignment" where an LLM might be considered "safe" by Western standards but "inappropriate" or "unnatural" within a Chinese sociocultural framework.
A Hierarchical Value Framework
To provide a systematic foundation for alignment, the authors developed a hierarchical framework based on the "Socialist Core Values." This framework is organized into three primary dimensions—National, Societal, and Personal—which are further subdivided into 12 core values (such as Prosperity, Civility, and Harmony) and 50 derived values.
As shown in the framework comparison, C-VARC offers significantly broader coverage of these core values compared to previous Chinese benchmarks like FLAMES, CMoralEval, and CVALUES, which often focus on a narrower subset of ethical dimensions.
Methodology: Building the Corpus
The construction of C-VARC involved a sophisticated multi-stage pipeline designed to ensure both scale and cultural authenticity. The process combined automated LLM-assisted extraction with rigorous human-in-the-loop quality control.
The Pipeline Stages:
- Filtering & Adaptation: International datasets (SC101, MIC) were processed using LLMs (specifically Qwen2.5-72B) to filter out rules that were culturally incompatible or semantically incomplete.
- Native Sourcing: Data was crawled from diverse Chinese sources, including academic datasets (FLAMES), social media platforms (Zhihu), and official news outlets (People's Daily).
- LLM-Assisted Extraction: Using a "Rule of Thumb" (RoT) template, LLMs extracted underlying moral principles from raw scenarios (e.g., converting a story about a family dispute into the rule "It's good to listen to your parents' advice").
- Expert Quality Control: A team of 40 philosophy and AI experts manually annotated and rewrote a subset of 36,000 rules to ensure high reliability (Cohen’s κ > 0.85). This high-quality subset was then used to refine the remaining 250,000+ rules.
- Dilemma Generation: To test the limits of model alignment, the researchers used semantic similarity and contradiction models to pair conflicting rules, generating over 400,000 moral dilemma scenarios.
Key Insights and Results
The research yielded several critical insights into the nature of AI value alignment:
- Semantic Boundaries: Using C-VARC to guide scenario generation leads to significantly cleaner data. t-SNE visualizations revealed that scenarios generated with C-VARC guidance form distinct clusters with minimal overlap, whereas unguided generation results in thematic intermixing.
- The RoT as a Bridge: By breaking down abstract values (like "Integrity") into concrete RoTs (e.g., "You should not deceive a friend's spouse"), the researchers created a machine-readable bridge between high-level ethics and low-level model responses.
- Model Performance: In evaluations across six sensitive themes, seven mainstream LLMs (including GPT-4o and Gemini) preferred C-VARC-aligned options in over 70.5% of cases. Furthermore, Chinese human annotators showed an 87.5% agreement rate with C-VARC rules, confirming the dataset's cultural authenticity.
Limitations and Future Outlook
While C-VARC represents a major step forward, the authors acknowledge a current data imbalance: personal-level rules dominate the corpus because they are more prevalent in social media and existing datasets. National and societal-level rules are currently more limited and require further synthesis.
Future work will focus on the dynamic evolution of the corpus to match shifting societal priorities, using C-VARC as an automated monitor for LLM "value drift," and expanding the corpus through cross-linguistic translation to facilitate global cross-cultural alignment research.
Summary updated