MMLU is a standardized benchmark for testing knowledge and reasoning across multiple academic and professional subjects and languages. It helps compare how well language models handle different kinds of questions. section 3.3
MMLU also exposes language-related inequality. GPT-4 exceeds 85% accuracy on English tasks but falls to about 62% for Telugu, showing how training-data imbalance can affect real users. section 3.3
English and Telugu performance
Comparing English and Telugu results illustrates that strong overall benchmark performance does not guarantee equally reliable reasoning across languages, scripts, grammatical systems, or cultural contexts. section 3.3
