EVL
MMLU-Pro
Harder multitask QA
What this measures
A harder, cleaned-up MMLU: more reasoning-heavy questions with 10 answer choices instead of 4, so lucky guessing matters less. A better separator of strong models.
Sample
A reasoning-heavy version of an MMLU question with 10 options (A–J) where several look plausible and you must actually work it out.
Leaderboard · 100 models
Accuracy (% correct, 10 options) — higher is betterBenchmark health
saturatingTop models cluster near the ceiling — differences here are getting too small to mean much. Weight newer, harder benchmarks more.
88.0
top-10 mean
89.8
best
100
models
What improves this score
Datasets & environments shown to raise it.
No data yet.
Research using this benchmark
all →NITP: Next Implicit Token Prediction for LLM Pre-trainingThe Chain Holds, the Answer Folds: Trace-Answer Dissociation in Reasoning Models Under Adversarial PressureResolution Diagnostics for Paired LLM EvaluationReasoning that Travels: Dissecting How Chain-of-Thought Transfers Across ModelsLaw of Neural Interaction: Depth-Width Shape, Interaction Efficiency, and Generalization
Matched by name in title/abstract.