ResearchBenchmarksMathematical Reasoning on MATH500 (Accuracy, Drop)Follow50AccuracyBF1645.8446.924849.08May 25, 2026Evaluation ResultsMethodMethodLinksAccuracyDropBF16shot-count=4-shot, Lea...shot-count=4-shot, Learning Rate=10^-52026.0550—HiF8 QATshot-count=4-shot, Lea...shot-count=4-shot, Learning Rate=10^-5, amax_algo=max, history=64, BF16 warmup=500 steps2026.0546—