ResearchBenchmarksReasoning on MMLU (Detailed Accuracy Statistics and Token Costs)Reasoning on MMLU (Detailed Accuracy Statistics and Token Costs)Follow78.32Mean AccuracyACTIVE75.574476.28727777.7128May 7, 2026Evaluation resultsMethodMethodLinksMean AccuracyAccuracy STDAccuracy Q1Accuracy 25th PercentileToken Cost (Select Phase)Token Cost (Train Phase)Token Cost (Test Phase)ACTIVESetup=ACTIVESetup=ACTIVE2026.0578.321.876.4776.250.80.120.67RANDOMSetup=RANDOMSetup=RANDOM2026.0576.873.7273.272—0.110.67NO TRAINSetup=NO TRAINSetup=NO TRAIN2026.0575.682.175.1673.41——1.04