List your company
ResearchFormerly WizWand

Math Reasoning on AIME24

76.7AccuracyDiffMAS
24.738.251.765.2Sep 25, 2025Oct 30, 2025Dec 4, 2025Jan 8, 2026Feb 12, 2026Mar 19, 2026Apr 23, 2026
Updated 6mo ago

Evaluation results

MethodLinks
DiffMASBackbone=Qwen3-14B2026.0476.7
DiffMASBackbone=Mistral3-8B2026.0473.3
DiffMASBackbone=DeepSeek-R1-Distill-Qwen-32B2026.0470
LatentMASBackbone=Qwen3-14B2026.0466.7
SingleBackbone=DeepSeek-R1-Distill-Qwen-32B2026.0466.7
CE-GPPOBase Model=DS-R1-Distill-Qwen-7B, Training Algorithm=CE-GPPO, Beta1 Hyperparameter=0.75, Beta2 Hyperparameter=12025.0966
SingleBackbone=Qwen3-14B2026.0463.3
TextMASBackbone=Qwen3-14B2026.0463.3
CE-GPPOBase Model=DS-R1-Distill-Qwen-7B, Training Algorithm=CE-GPPO, Beta1 Hyperparameter=0.5, Beta2 Hyperparameter=12025.0962.3
DAPOBase Model=DS-R1-Distill-Qwen-7B, Training Algorithm=DAPO2025.0959.7
GRPOBase Model=DS-R1-Distill-Qwen-7B, Training Algorithm=GRPO2025.0955.3
DS-R1-Distill-Qwen-7BBase Model=DS-R1-Distill-Qwen-7B2025.0954.5
SingleBackbone=Mistral3-8B2026.0450
TextMASBackbone=DeepSeek-R1-Distill-Qwen-32B2026.0450
LatentMASBackbone=DeepSeek-R1-Distill-Qwen-32B2026.0450
CE-GPPOBase Model=DS-R1-Distill-Qwen-1.5B, Training Algorithm=CE-GPPO, Beta1 Hyperparameter=0.75, Beta2 Hyperparameter=12025.0943.6
CE-GPPOBase Model=DS-R1-Distill-Qwen-1.5B, Training Algorithm=CE-GPPO, Beta1 Hyperparameter=0.5, Beta2 Hyperparameter=12025.0942
DAPOBase Model=DS-R1-Distill-Qwen-1.5B, Training Algorithm=DAPO2025.0940
GRPOBase Model=DS-R1-Distill-Qwen-1.5B, Training Algorithm=GRPO2025.0933.4
LatentMASBackbone=Mistral3-8B2026.0433.3
DS-R1-Distill-Qwen-1.5BBase Model=DS-R1-Distill-Qwen-1.5B2025.0929.2
TextMASBackbone=Mistral3-8B2026.0426.7