List your company
ResearchFormerly WizWand

Math Reasoning on AIME 25 (Accuracy)

76.7AIME 25 AccuracyBaseline
-3.06817.64138.3559.059Sep 25, 2025Nov 8, 2025Dec 22, 2025Feb 4, 2026Mar 20, 2026May 3, 2026Jun 16, 2026
Updated 3mo ago

Evaluation results

MethodLinks
BaselineBackbone=Qwen3-MoE-30B-A3B, Type=–2026.0676.7
OursQBackbone=Qwen3-MoE-30B-A3B, Sparsity=25%, Quantization=4-bit2026.0675
BaselineSparsity=0%2025.1172.9
OursBackbone=Qwen3-MoE-30B-A3B, Sparsity=50%2026.0664
PuzzleMoESparsity=25%2025.1161.5
CE-GPPOBase Model=DS-R1-Distill-Qwen-7B, Training Algorithm=CE-GPPO, Beta1 Hyperparameter=0.75, Beta2 Hyperparameter=12025.0951.4
CE-GPPOBase Model=DS-R1-Distill-Qwen-7B, Training Algorithm=CE-GPPO, Beta1 Hyperparameter=0.5, Beta2 Hyperparameter=12025.0949.1
DAPOBase Model=DS-R1-Distill-Qwen-7B, Training Algorithm=DAPO2025.0948.7
GRPOBase Model=DS-R1-Distill-Qwen-7B, Training Algorithm=GRPO2025.0940.3
DS-R1-Distill-Qwen-7BBase Model=DS-R1-Distill-Qwen-7B2025.0939.1
CE-GPPOBase Model=DS-R1-Distill-Qwen-1.5B, Training Algorithm=CE-GPPO, Beta1 Hyperparameter=0.5, Beta2 Hyperparameter=12025.0933.9
CE-GPPOBase Model=DS-R1-Distill-Qwen-1.5B, Training Algorithm=CE-GPPO, Beta1 Hyperparameter=0.75, Beta2 Hyperparameter=12025.0931
DAPOBase Model=DS-R1-Distill-Qwen-1.5B, Training Algorithm=DAPO2025.0928.4
GRPOBase Model=DS-R1-Distill-Qwen-1.5B, Training Algorithm=GRPO2025.0928.1
DS-R1-Distill-Qwen-1.5BBase Model=DS-R1-Distill-Qwen-1.5B2025.0924.1
HC-SMoESparsity=25%2025.110