ResearchBenchmarksSTEM Reasoning on OCWFollow17.3AccuracyAGPO7.83610.29312.7515.207May 20, 2026Evaluation ResultsMethodMethodLinksAccuracyAGPOclipping=adaptive, ATS...clipping=adaptive, ATS=true2026.0517.3GRPOATS=trueATS=true2026.0516.3AGPOclipping=adaptive, ATS...clipping=adaptive, ATS=false2026.0516.2GRPOclipping=fixed εclipping=fixed ε2026.0515.4Adaptive-KL PPO2026.0514.6DPO2026.059.1PPO2026.058.2