Do LLMs have a Gender (Entropy) Bias?

Researchers discovered that large language models exhibit gender bias at the individual question level, creating different amounts of information for men versus women despite appearing unbiased at category levels. A new benchmark dataset called RealWorldQuestioning was developed, and a simple prompt-based debiasing approach was shown to improve response quality in 78% of cases.