MBBQ: A Dataset for Cross-Lingual Comparison of Stereotypes in Generative LLMs

Name: MBBQ: A Dataset for Cross-Lingual Comparison of Stereotypes in Generative LLMs
Uploaded: 2026-10-17T08:20:13.736566+00:00
Duration: 587 s
Description: Researchers created MBBQ, a multilingual benchmark extending an English dataset to Dutch, Spanish, and Turkish to study LLM stereotypes across languages. They found that non-English models often show more bias than English ones when controlling for culture. The work suggests evaluating bias by category rather than overall model performance.

Researchers created MBBQ, a multilingual benchmark extending an English dataset to Dutch, Spanish, and Turkish to study LLM stereotypes across languages. They found that non-English models often show more bias than English ones when controlling for culture. The work suggests evaluating bias by category rather than overall model performance.

Open the full episode