ResearchBenchmarksMathematical Reasoning on Math500, OlympiadBench, Minerva, AIME (24/25), and AMCFollow85.6Math500 AccuracyDMPO62.7268.6674.680.54May 19, 2026Evaluation ResultsMethodMethodLinksMath500 AccuracyOlympiadBench AccuracyMinerva AccuracyAIME 2024 AccuracyAIME 2025 AccuracyAMC AccuracyAverage AccuracyDMPO2026.0585.648.441.220.21556.344.5GRPO2026.0583.844.939.717.814.154.742.5GSPO2026.0583.244.739.318.812.653.942.1FlowRL2026.0582.44437.61611.150.240.2ClipCov2026.0581.446.738.619.113.854.542.3GPG2026.0579.642.740.416.310.648.839.7Backbone2026.0563.628.321.7146.945.630