ResearchBenchmarksMathematical Reasoning on Hendrycks MATH (GRPO Convergence Metrics)Follow36.7Peak AccuracyGXPO33.99634.69835.436.102May 7, 2026Evaluation ResultsMethodMethodLinksPeak AccuracySteps to Match GRPOStep SpeedupTime to Match GRPO (h)Time SpeedupBackpropagations to Match GRPOBP SpeedupGXPOModel=Llama3.2-3B, k=10Model=Llama3.2-3B, k=102026.0536.76046.062.331801.33GXPOModel=Llama3.2-3B, k=3Model=Llama3.2-3B, k=32026.0536.19653.696.722.11951.23SFPOModel=Llama3.2-3B, k=10Model=Llama3.2-3B, k=102026.0535.2580318.680.768700.28SFPOModel=Llama3.2-3B, k=3Model=Llama3.2-3B, k=32026.0534.658039.071.563200.75GRPOModel=Llama3.2-3B, k=–Model=Llama3.2-3B, k=–2026.0534.1240114.1412401