ResearchBenchmarksSTEM Reasoning on SATFollow0.893AccuracyAGPO0.755720.791360.8270.86264May 20, 2026Evaluation ResultsMethodMethodLinksAccuracyAGPOclipping=adaptive, ATS...clipping=adaptive, ATS=true2026.050.893AGPOclipping=adaptive, ATS...clipping=adaptive, ATS=false2026.050.886GRPOATS=trueATS=true2026.050.879Adaptive-KL PPO2026.050.843GRPOclipping=fixed εclipping=fixed ε2026.050.806DPO2026.050.777PPO2026.050.761