Research map / Large language models / Language, safety and interpretability
GeWu: A Culturally-Grounded Chinese Benchmark for Multi-Stage Social Bias Evaluation in Large Language Models
Yi Lin, Ziyi Zhou, Jiashi Gao, Xinwei Guo, Jiaxin Zhang, Haiyan Wu, Xin Yao, Xuetao Wei
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| enhance CBBQ ambiguous contexts with LLM-generated cultural detail (GeWu) | 11 Chinese LLMs · Stage I multiple choice | bias score: bias scores more than double on aggregate for most models | Improved |
| enhance context for Qwen2.5-1.5B-Instruct (Age) | Qwen2.5-1.5B-Instruct · Stage I Age | bias score: 0.90 -> 0.97 | Improved |
| enhance context for Qwen3-1.7B (Religion) | Qwen3-1.7B · Stage I Religion | bias score: 0.75 -> 0.94 | Improved |
| enhanced contexts diversify questions | GeWu vs CBBQ · question pairs | average ROUGE-L: 0.8825 -> 0.8025 | Improved |
| increase model size to 660B (DeepSeek-V3) | Chinese LLMs · Stage I GeWu | bias score | No clear effect |
| switch to reasoning mode / reasoning-distilled models | Qwen3-14B, DeepSeek-R1-Distill-Qwen-14B · Stage I GeWu | bias score | Hurt |
| add second dialogue turn with enhanced cultural context | 11 LLMs on GeWu-1K · Stage II multi-turn | bias score (DeepSeek-V3 classifier) | Hurt |
| add second turn context for household registration | LLMs on GeWu-1K · Stage II Household registration | bias score | Hurt |
| small model scale (Qwen2.5-1.5B, Qwen3-1.7B) vs medium scale | Qwen families · Stage I GeWu | bias score: Qwen3-1.7B approaching 100% | Hurt |
Closest papers by meaning
- Mitigating Social Bias in Large Language Models: A Multi-Objective Approach Within a Multi-Agent Framework (AAAI 2025 · similarity 0.88)
- SocialStigmaQA: A Benchmark to Uncover Stigma Amplification in Generative Language Models (AAAI 2024 · similarity 0.87)
- B-score: Detecting biases in large language models using response history (ICML 2025 · similarity 0.86)
- BLEnD: A Benchmark for LLMs on Everyday Knowledge in Diverse Cultures and Languages (NeurIPS 2024 · similarity 0.86)
- Prompting Fairness: Integrating Causality to Debias Large Language Models (ICLR 2025 · similarity 0.86)
- Investigating and Mitigating Undesirable Biases in Large Language Models (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
