A 72-problem revision of MLE-Bench that retains high-signal original tasks, adds 2025-2026 ML competitions, allows up to three leaderboard submissions, and scores against a reference solution distribution.
Refreshes saturated or low-signal competitions with newer practical ML problems. Measures iterative engineering against a calibrated hidden-test reference distribution.
The task set, interaction protocol, and normalized reward are not comparable with the original public 75-competition medal-rate leaderboard. Public reproducibility and independent coverage remain limited.
Higher is better. Mean normalized reward relative to a reference distribution.
Max score: 100
The task set, interaction protocol, and normalized reward are not comparable with the original public 75-competition medal-rate leaderboard.
No scores available for this benchmark yet.