List your company
ResearchFormerly WizWand

Math Reasoning on MATH500 (Accuracy)

95.6AccuracyCE-GPPO
52.75263.8767586.124May 16, 2025Jul 22, 2025Sep 27, 2025Dec 4, 2025Feb 9, 2026Apr 17, 2026Jun 24, 2026
Updated 4mo ago

Evaluation results

MethodLinks
CE-GPPOBase Model=DS-R1-Distill-Qwen-7B, Training Algorithm=CE-GPPO, Beta1 Hyperparameter=0.75, Beta2 Hyperparameter=12025.0995.6
DAPOBase Model=DS-R1-Distill-Qwen-7B, Training Algorithm=DAPO2025.0995.1
CE-GPPOBase Model=DS-R1-Distill-Qwen-7B, Training Algorithm=CE-GPPO, Beta1 Hyperparameter=0.5, Beta2 Hyperparameter=12025.0994.9
GRPOBase Model=DS-R1-Distill-Qwen-7B, Training Algorithm=GRPO2025.0993.7
DS-R1-Distill-Qwen-7BBase Model=DS-R1-Distill-Qwen-7B2025.0993.6
ReST-MCTSBackbone=Gemma-3-27B2025.1193.2
ReST-MCTSBackbone=GPT-OSS-120B2025.1191.6
APRMBackbone=GPT-OSS-120B2025.1191.4
APRMBackbone=Gemma-3-27B2025.1191.4
ReST-MCTSBackbone=GPT-OSS-20B2025.1191
APRMBackbone=GPT-OSS-20B2025.1191
CE-GPPOBase Model=DS-R1-Distill-Qwen-1.5B, Training Algorithm=CE-GPPO, Beta1 Hyperparameter=0.5, Beta2 Hyperparameter=12025.0991
CE-GPPOBase Model=DS-R1-Distill-Qwen-1.5B, Training Algorithm=CE-GPPO, Beta1 Hyperparameter=0.75, Beta2 Hyperparameter=12025.0990.9
VRPOModel Backbone=Qwen3-8B(Strong Model), Optimization Method=VRPO2025.0890.2
rStar MathBackbone=GPT-OSS-120B2025.1190
DAPOBase Model=DS-R1-Distill-Qwen-1.5B, Training Algorithm=DAPO2025.0990
rStar MathBackbone=Gemma-3-27B2025.1189.5
NL GuidanceBackbone=Qwen3-8B-Base2026.0389.4
Reward ShapingBackbone=Qwen3-8B-Base2026.0389.2
Ctrl-RBackbone=Qwen3-8B-Base, beta=0.22026.0389.2
GRPOModel Backbone=Qwen3-8B(Strong Model), Optimization Method=GRPO2025.0889.2
rStar MathBackbone=GPT-OSS-20B2025.1189
Reinforce++Model Backbone=Qwen3-8B(Strong Model), Optimization Method=Reinforce++2025.0889
DAPOBackbone=Qwen3-8B-Base2026.0388.4
GRPOBase Model=DS-R1-Distill-Qwen-1.5B, Training Algorithm=GRPO2025.0988.3
ToTBackbone=Gemma-3-27B2025.1188
Dr.GRPOModel Backbone=Qwen3-8B(Strong Model), Optimization Method=Dr.GRPO2025.0887.8
ToTBackbone=GPT-OSS-120B2025.1187.6
ToTBackbone=GPT-OSS-20B2025.1187.6
BaseModel Backbone=Qwen3-8B(Strong Model), Optimization Method=Base2025.0887.4
LLM-JBackbone=Gemma-3-27B2025.1186.4
PPOModel Backbone=Qwen3-8B(Strong Model), Optimization Method=PPO2025.0886.4
AutoPSVBackbone=GPT-OSS-120B2025.1186
DS-R1-Distill-Qwen-1.5BBase Model=DS-R1-Distill-Qwen-1.5B2025.0986
Math S.Backbone=GPT-OSS-120B2025.1185.8
TaHModel Size=4B*2025.1185.8
TaH+Model Size=4B*2025.1185.6
LLM-JBackbone=GPT-OSS-120B2025.1185.2
LLM-JBackbone=GPT-OSS-20B2025.1185.2
Verify SbS.Backbone=GPT-OSS-120B2025.1185
SoftThinkModel Size=4B*2025.1185
CoT-SCBackbone=GPT-OSS-120B2025.1184.4
StandardModel Size=4B*2025.1184.2
AutoPSVBackbone=GPT-OSS-20B2025.1184
OuroModel Size=4B*2025.1183.8
AutoPSVBackbone=Gemma-3-27B2025.1183
Math S.Backbone=GPT-OSS-20B2025.1182
Math S.Backbone=Gemma-3-27B2025.1181.5
ReST-MCTSBackbone=Gemma-3-12B2025.1181.4
Verify SbS.Backbone=GPT-OSS-20B2025.1180
Verify SbS.Backbone=Gemma-3-27B2025.1180
APRMBackbone=Gemma-3-12B2025.1180
CGDTeacher Model=S1.1-32B2025.0579.6
CGDTeacher Model=Claude Sonnet 3.72025.0579.4
CFTTeacher Model=GPT-4o2025.0579.2
rStar MathBackbone=Gemma-3-12B2025.1179
CoT-SCBackbone=Gemma-3-27B2025.1178.8
Math S.Backbone=Gemma-3-12B2025.1178.5
CoT-SCBackbone=GPT-OSS-20B2025.1178
AutoPSVBackbone=Gemma-3-12B2025.1178
ToTBackbone=Gemma-3-12B2025.1177.6
Verify SbS.Backbone=Gemma-3-12B2025.1177.5
CoT-SCBackbone=Gemma-3-12B2025.1176.8
Dr.GRPOModel Backbone=Qwen2.5-7B-Cold Start(Weak Model), Optimization Method=Dr.GRPO2025.0873
LLM-JBackbone=Gemma-3-12B2025.1172.8
TaH+Model Size=1.7B2025.1172.6
VRPOModel Backbone=Qwen2.5-7B-Cold Start(Weak Model), Optimization Method=VRPO2025.0872.2
CFTTeacher Model=S1.1-32B2025.0571.6
BaseModel Backbone=Qwen2.5-7B-Cold Start(Weak Model), Optimization Method=Base2025.0871.6
TaHModel Size=1.7B2025.1171.4
Ctrl-RBackbone=Qwen3-1.7B-Base, beta=0.22026.0371
Ctrl-RBackbone=Qwen3-1.7B-Base, beta=0, Importance Sampling (IS)=false2026.0369.6
Ctrl-RBackbone=Qwen3-1.7B-Base, beta=1, Importance Sampling (IS)=true2026.0369.6
NL GuidanceBackbone=Qwen3-1.7B-Base2026.0369
Reward ShapingBackbone=Qwen3-1.7B-Base2026.0369
DAPOBackbone=Qwen3-1.7B-Base2026.0368.4
OuroModel Size=1.7B2025.1168.2
StandardModel Size=1.7B2025.1167.8
PPOModel Backbone=Qwen2.5-7B-Cold Start(Weak Model), Optimization Method=PPO2025.0867.4
SoftThinkModel Size=1.7B2025.1164.8
Reinforce++Model Backbone=Qwen2.5-7B-Cold Start(Weak Model), Optimization Method=Reinforce++2025.0863.8
AlwaysThinkModel Size=1.7B2025.1163.2
ICLBackbone=Qwen2.5-14B-Instruct, Alignment Strategy=In-Context Learning (System Prompt)2026.0658.4
BaseBackbone=Qwen2.5-14B-Instruct, Alignment Strategy=Base2026.0658.2
PolicyAlignBackbone=Qwen2.5-14B-Instruct, Alignment Strategy=Policy-Based Safety Alignment2026.0658.2
IndividualModel Architecture=Qwen3-30B-A3B, HARC Calibration=false2026.0657.65
NSPOBackbone=Qwen2.5-14B-Instruct, Alignment Strategy=Null-space constrained policy optimization2026.0657.4
DARE w/ HARCMerging Strategy=DARE, Model Architecture=Qwen3-30B-A3B, HARC Calibration=true2026.0657.35
GRPO+PolicyBackbone=Qwen2.5-14B-Instruct, Alignment Strategy=Group Relative Policy Optimization with policy-conditioned reward2026.0657.2
TIES-Merging w/ HARCMerging Strategy=TIES-Merging, Model Architecture=Qwen3-30B-A3B, HARC Calibration=true2026.0656.98
Fisher Merging w/ HARCMerging Strategy=Fisher Merging, Model Architecture=Qwen3-30B-A3B, HARC Calibration=true2026.0656.94
Weight AveragingMerging Strategy=Weight Averaging, Model Architecture=Qwen3-30B-A3B, HARC Calibration=false2026.0656.78
DAREMerging Strategy=DARE, Model Architecture=Qwen3-30B-A3B, HARC Calibration=false2026.0656.45
Weight Averaging w/ HARCMerging Strategy=Weight Averaging, Model Architecture=Qwen3-30B-A3B, HARC Calibration=true2026.0656.23
WUDI-Merging w/ HARCMerging Strategy=WUDI-Merging, Model Architecture=Qwen3-30B-A3B, HARC Calibration=true2026.0656.19
TIES-MergingMerging Strategy=TIES-Merging, Model Architecture=Qwen3-30B-A3B, HARC Calibration=false2026.0656.14
Qwen2.5-Math-7B (Base)2025.0555.4
Fisher MergingMerging Strategy=Fisher Merging, Model Architecture=Qwen3-30B-A3B, HARC Calibration=false2026.0655.38
WUDI-MergingMerging Strategy=WUDI-Merging, Model Architecture=Qwen3-30B-A3B, HARC Calibration=false2026.0655.18
AlphaAlignBackbone=Qwen2.5-14B-Instruct, Alignment Strategy=RL-based alignment2026.0654.4