List your company
ResearchFormerly WizWand

Mathematical Reasoning on AIME25 (Avg@8 Accuracy)

74Accuracy (Avg@8)CERL full
14.376829.855945.33560.8141May 22, 2026May 29, 2026Jun 5, 2026Jun 12, 2026Jun 19, 2026Jun 26, 2026Jul 3, 2026
Updated 4mo ago

Evaluation results

MethodLinks
CERL fullTraining=SFT+RL2026.0574
Correctness-only RLTraining=RL2026.0570.8
HaloTraining=SFT+RL2026.0570.3
Long-answer SFTTraining=SFT2026.0568.2
Fixed-ratio erasure + free answerTraining=None2026.0568
Length-penalty RLTraining=RL2026.0567.3
Base Qwen3-8BTraining=N/A2026.0567.1
TokenSkipTraining=SFT2026.0562.9
ACPOModel=DeepSeek-R1-Distill-Qwen-7B2026.0737.04
SAPOModel=DeepSeek-R1-Distill-Qwen-7B2026.0734.03
ACPOModel=Qwen3-8B-Base2026.0730
GTPOModel=DeepSeek-R1-Distill-Qwen-7B2026.0728.48
80/20Model=DeepSeek-R1-Distill-Qwen-7B2026.0727.48
DAPOModel=DeepSeek-R1-Distill-Qwen-7B2026.0726.67
SAPOModel=Qwen3-8B-Base2026.0726.43
ACPOModel=Qwen2.5-Math-7B2026.0725.69
GTPOModel=Qwen3-8B-Base2026.0725.45
GRPOModel=DeepSeek-R1-Distill-Qwen-7B2026.0724.83
80/20Model=Qwen3-8B-Base2026.0724.5
DAPOModel=Qwen3-8B-Base2026.0722.92
GTPOModel=Qwen2.5-Math-7B2026.0722.77
SAPOModel=Qwen2.5-Math-7B2026.0721.25
80/20Model=Qwen2.5-Math-7B2026.0720.5
DAPOModel=Qwen2.5-Math-7B2026.0719.56
GRPOModel=Qwen2.5-Math-7B2026.0716.71
GRPOModel=Qwen3-8B-Base2026.0716.67