Chinese LLMs Show Social Identity Biases, Research Finds

A new research paper explores social identity biases within Chinese large language models (LLMs) by using Mandarin-specific prompts. The study evaluated ten different LLMs, focusing on how ingroup versus outgroup framings affect sentiment and toxicity scores across 240 social groups relevant to the Chinese context. Findings indicate that while instruction tuning can reduce sentiment bias, toxicity disparities persist, and the use of feminine pronouns is linked to increased toxicity in some models. AI

IMPACT Highlights potential biases in Chinese LLMs, informing developers and users about language-specific risks.

RANK_REASON Academic paper detailing a new evaluation framework and findings on LLM bias. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• Chinese LLMs
• Geng Liu
• Mandarin

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →