ResearchBenchmarksMathematical Reasoning on GSM8K (Accuracy Percentiles & Token Costs)Follow93.07Mean Accuracy (±STD)ACTIVE92.23892.45492.6792.886May 7, 2026Evaluation ResultsMethodMethodLinksMean Accuracy (±STD)Accuracy (Q1)Accuracy (Worst-25%)Token Cost (Select)Token Cost (Train)Token Cost (Test)ACTIVESetup=ACTIVESetup=ACTIVE2026.0593.0792.8292.772.10.183.54RANDOMSetup=RANDOMSetup=RANDOM2026.0592.7892.4992.06—0.173.56NO TRAINSetup=NO TRAINSetup=NO TRAIN2026.0592.2791.7391.61——4.5