The Massive Multitask Language Understanding (MMLU) dataset is a dataset of high school, college, and professional multiple choice exams that test expert subject knowledge. It was constructed by Hendrycks et al. (2021). Hypermind forecasters were commissioned to predict state-of-the-art performance on June 30, 2022, '23, '24, and '25. The 2022 result of 67.5% was significantly outside forecasters' prediction intervals, so we're seeing what the updated forecasts are for 2023, '24, and '25.
What will be state-of-the-art accuracy on the Massive Multitask dataset in the following years?
These questions should resolve identically to the Hypermind forecasts:
"These questions resolve as the highest performance achieved on MMLU by June 30 in the following years by an eligible model. Eligible models must not have been specifically trained on data from the MMLU dataset. A model need not be publicly released, as long as the resulting performance itself is reported in a published paper (on arxiv or a major ML conference) or through an official communication channel of an industry lab (e.g. claimed in a research blog post on the OpenAI blog, or a press release). If there's uncertainty about whether something counts, we will defer to this leaderboard."