List your company

STEM Reasoning on GaokaoQA

41.1AccuracyAGPO
37.35638.32839.340.272May 20, 2026

Evaluation results

MethodLinks
AGPOclipping=adaptive, ATS=true2026.0541.1
AGPOclipping=adaptive, ATS=false2026.0540.6
GRPOclipping=fixed ε2026.0539.8
GRPOATS=true2026.0539.2
Adaptive-KL PPO2026.0538.5
PPO2026.0537.7
DPO2026.0537.5