List your company
ResearchFormerly WizWand

Mathematical Reasoning on AIME 2024 (Accuracy@32, Average)

67.8Average Score (avg@32)DS-R1-Distill-Qwen-7B + IC-DAPO
1.510418.720235.9353.1398Mar 10, 2026Mar 26, 2026Apr 11, 2026Apr 28, 2026May 14, 2026May 30, 2026Jun 16, 2026
Updated 4mo ago

Evaluation results

MethodLinks
DS-R1-Distill-Qwen-7B + IC-DAPOBackbone=Qwen-7B, Zero-shot mode=true, Time/Step (s)=315.62026.0367.866.5—
GRPOParams=4B2026.0567.6——
DS-R1-Distill-Qwen-7B + CE-GPPOBackbone=Qwen-7B, Zero-shot mode=true, Time/Step (s)=292.52026.0367.364.2—
DS-R1-Distill-Qwen-7B + DAPOBackbone=Qwen-7B, Zero-shot mode=true, Time/Step (s)=303.12026.0365.362—
SRaRModel Size=32B2026.0564.17——
RaRModel Size=32B2026.0563.65——
DS-R1-Distill-Qwen-7BBackbone=Qwen-7B, Zero-shot mode=true2026.0361.854.5—
DS-R1-Distill-Qwen-7B + GRPOBackbone=Qwen-7B, Zero-shot mode=true, Time/Step (s)=305.62026.0361.455.3—
DAPOModel Size=32B2026.0561.25——
SRaRModel Size=8B2026.0561.15——
RaRModel Size=8B2026.0557.5——
E3RLBackbone=Qwen3-8b, RL Strategy=E3RL2026.0657.5—76.7
DS-R1-Distill-Qwen-1.5B + IC-DAPOBackbone=Qwen-1.5B, Zero-shot mode=true, Time/Step (s)=477.22026.0356.445.6—
DS-R1-Distill-Qwen-1.5B + GSPOBackbone=Qwen-1.5B, Zero-shot mode=true, Time/Step (s)=437.32026.0355.742.5—
DS-R1-Distill-Qwen-1.5B + CE-GPPOBackbone=Qwen-1.5B, Zero-shot mode=true, Time/Step (s)=464.02026.0355.742.8—
DS-R1-Distill-Qwen-1.5B + DAPOBackbone=Qwen-1.5B, Zero-shot mode=true, Time/Step (s)=459.62026.0353.940—
DAPOBackbone=Qwen3-8b, RL Strategy=DAPO2026.0653.8—80
DAPOModel Size=8B2026.0552.6——
GSPOBackbone=Qwen3-8b, RL Strategy=GSPO2026.0652.1—83.3
GRPOBackbone=Qwen3-8b, RL Strategy=GRPO2026.0651.9—76.7
SODParams=1.7B2026.0550.83——
SAPOBackbone=Qwen3-8b, RL Strategy=SAPO2026.0650.7—76.7
E3RLBackbone=Qwen3-4b, RL Strategy=E3RL2026.0650.6—83.3
DS-R1-Distill-Qwen-1.5B + GRPOBackbone=Qwen-1.5B, Zero-shot mode=true, Time/Step (s)=457.42026.0350.333.4—
DS-R1-Distill-Qwen-1.5B + CISPOBackbone=Qwen-1.5B, Zero-shot mode=true, Time/Step (s)=466.32026.0348.832.9—
RGR-GRPOModel Size=32B2026.0547.71——
SAPOBackbone=Qwen3-4b, RL Strategy=SAPO2026.0647.5—83.3
DS-R1-Distill-Qwen-1.5BBackbone=Qwen-1.5B, Zero-shot mode=true2026.0346.329.2—
RGR-GRPOModel Size=8B2026.0545.31——
GRPO-VPSModel Size=8B2026.0544.58——
GSPOBackbone=Qwen3-4b, RL Strategy=GSPO2026.0644.5—83.3
GRPOBackbone=Qwen3-4b, RL Strategy=GRPO2026.0644.2—83.3
OPDParams=1.7B2026.0543.86——
DAPOBackbone=Qwen3-4b, RL Strategy=DAPO2026.0643.6—80
GRPO-VPSModel Size=32B2026.0543.02——
GRPOModel Size=32B2026.0537.92——
GRPOModel Size=8B2026.0536.56——
VanillaBackbone=Qwen3-4b, RL Strategy=Vanilla2026.0635.8—63.3
VanillaBackbone=Qwen3-8b, RL Strategy=Vanilla2026.0634.9—60
OPSDhintParams=1.7B2026.0534.42——
OPSDgtParams=1.7B2026.0533.85——
BaselineModel Size=32B2026.0527.19——
BaselineModel Size=8B2026.0526.88——
SFTParams=1.7B2026.0526.77——
GRPOParams=1.7B2026.0525.63——
SODParams=0.6B2026.0520.84——
OPDParams=0.6B2026.0516.82——
OPSDgtParams=0.6B2026.0512.63——
VanillaParams=1.7B2026.059.9——
OPSDhintParams=0.6B2026.059.77——
VanillaParams=0.6B2026.057.71——
SFTParams=0.6B2026.055.67——
GRPOParams=0.6B2026.054.06——
GRPOReward condition=Clean, Training steps=Full training2026.06—19.3—
GRPOReward condition=25% reward flipping, Training steps=200 steps2026.06—13—
Naive GRPOReward condition=Clean, Training steps=Full training2026.06—20—
Qwen3-4B-BaseReward condition=Clean, Training steps=Full training2026.06—8.6—
VIMPOReward condition=Clean, Training steps=Full training2026.06—21.7—
VIMPOReward condition=25% reward flipping, Training steps=200 steps2026.06—18.3—