DVMap: Fine-Grained Pluralistic Value Alignment via Demographic Mapping
The newly introduced DVMap (High-Consensus Demographic-Value Mapping) framework seeks to overcome the challenges posed by broad national labels in aligning large language models (LLMs) with diverse human values. Existing methods frequently neglect the value diversity within countries, resulting in weak alignment. DVMap replaces national labels with multi-dimensional demographic constraints, pinpointing groups that exhibit predictable, high-consensus value preferences. This framework compiles a dataset of 56,152 samples from the World Values Survey (WVS), focusing on respondents who maintain consistent value preferences within the same demographic categories. Additionally, it incorporates a Structured Chain-of-Thought (CoT) mechanism to facilitate the alignment process. The findings are elaborated in arXiv:2605.14420.
Key facts
- DVMap proposes fine-grained pluralistic value alignment for LLMs.
- Current methods rely on coarse national labels, obscuring intra-country heterogeneity.
- The framework uses multi-dimensional demographic constraints.
- A corpus of 56,152 samples is drawn from the World Values Survey (WVS).
- Only respondents with consistent value preferences under identical demographics are retained.
- A Structured Chain-of-Thought (CoT) mechanism guides alignment.
- The paper is available on arXiv with ID 2605.14420.
- The approach aims to improve value alignment by capturing demographic diversity.
Entities
Institutions
- World Values Survey (WVS)
- arXiv