MMLU-Redux: leaderboard

5,700 MMLU questions, 100 per subject across all 57, re-annotated by an Edinburgh-led team (2024) with error labels such as wrong ground truth or unclear options; 6.49% of MMLU estimated faulty.

Source: arxiv.org.

Interactive version: theaggregate.ai/benchmark?slug=mmlu-redux · How It Works · Data refreshed daily, snapshot 2026-09-05.