A new framework called Stereotypes-to-Decisions (S2D) has been developed to systematically evaluate regional bias in large language models (LLMs). This framework assesses six different LLMs across all 34 provincial-level administrative regions of China, examining stereotypes related to warmth and competence, as well as decision-making in education, occupation, and social interaction. The study found significant and consistent regional biases across models, which correlate with regional development indicators and show mixed human-like stereotypes. These biases largely persist regardless of whether prompts are in Chinese or English, highlighting the prevalence and systematic nature of regional bias in LLMs. AI
IMPACT Highlights the need for more regionally aware LLM evaluation and mitigation strategies to address prevalent biases.
RANK_REASON Academic paper introducing a new evaluation framework for LLM bias. [lever_c_demoted from research: ic=1 ai=1.0]
- arXiv
- China
- Education
- English
- large language models (LLMs)
- Occupation
- Social Interaction
- Standard Chinese
- Stereotypes-to-Decisions (S2D)
- Warmth
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
