List your company
ResearchFormerly WizWand

Science Reasoning Accuracy on GPQA

54.67Accuracy (GPQA)EAPO
-2.186812.574127.33542.0959Aug 5, 2025Sep 28, 2025Nov 22, 2025Jan 16, 2026Mar 11, 2026May 5, 2026Jun 29, 2026
Updated 4mo ago

Evaluation results

MethodLinks
EAPOModel=Qwen-3-8B-Base2026.0654.67
DAPOModel=Qwen-3-8B-Base2026.0652.27
OPDModel=Qwen-3-8B-Base2026.0652.27
MOPDModel=Qwen-3-8B-Base2026.0650.51
GRPOModel=Qwen-3-8B-Base2026.0650.44
EAPOModel=Qwen-2.5-Math-7B2026.0643.18
EAPO w/o sISModel=Qwen-2.5-Math-7B2026.0641.98
Trajectory ReplayModel=Qwen-2.5-Math-7B2026.0639.67
OPDModel=Qwen-2.5-Math-7B2026.0639.39
DAPOModel=Qwen-2.5-Math-7B2026.0639.27
MOPDModel=Qwen-2.5-Math-7B2026.0638.76
GRPOModel=Qwen-2.5-Math-7B2026.0634.34
MENTORBackbone=Qwen2.5-7B-Base2025.1030.8
MENTORBackbone=LLaMa3.1-8B-Base2025.1030.3
QuestABackbone=Qwen2.5-7B-Base2025.1030.3
On-policy RLBackbone=Qwen2.5-7B-Base2025.1029.3
LUFFYBackbone=LLaMa3.1-8B-Base2025.1027.8
LUFFYBackbone=Qwen2.5-7B-Base2025.1026.8
On-policy RLBackbone=LLaMa3.1-8B-Base2025.1025.8
QuestABackbone=LLaMa3.1-8B-Base2025.1025.3
MENTORBackbone=Qwen2.5-3B-Base2025.1022.7
On-policy RLBackbone=Qwen2.5-3B-Base2025.1017.7
QuestABackbone=Qwen2.5-3B-Base2025.1016.2
LUFFYBackbone=Qwen2.5-3B-Base2025.1015.2
BaseBackbone=Qwen2.5-7B-Base2025.1011.1
Reinforce++Model Backbone=Qwen3-8B(Strong Model), Optimization Method=Reinforce++2025.084.35
VRPOModel Backbone=Qwen3-8B(Strong Model), Optimization Method=VRPO2025.084.35
VRPOOptimization Type=test-time optimization, Reward Type=rule-based rewards2025.084.17
Dr.GRPOModel Backbone=Qwen3-8B(Strong Model), Optimization Method=Dr.GRPO2025.083.99
VRPOModel Backbone=Qwen2.5-7B-Cold Start(Weak Model), Optimization Method=VRPO2025.083.44
BaseModel Backbone=Qwen3-8B(Strong Model), Optimization Method=Base2025.083.1
GRPOOptimization Type=test-time optimization, Reward Type=rule-based rewards2025.083.08
BaseBackbone=Qwen2.5-3B-Base2025.103
GRPOModel Backbone=Qwen3-8B(Strong Model), Optimization Method=GRPO2025.082.98
PPOOptimization Type=test-time optimization, Reward Type=rule-based rewards2025.082.54
PPOModel Backbone=Qwen2.5-7B-Cold Start(Weak Model), Optimization Method=PPO2025.082.54
Cold StartOptimization Type=test-time optimization, Reward Type=rule-based rewards2025.082.53
GRPOModel Backbone=Qwen2.5-7B-Cold Start(Weak Model), Optimization Method=GRPO2025.082.36
Reinforce++Model Backbone=Qwen2.5-7B-Cold Start(Weak Model), Optimization Method=Reinforce++2025.082.17
PPOModel Backbone=Qwen3-8B(Strong Model), Optimization Method=PPO2025.082.17
BaseModel Backbone=Qwen2.5-7B-Cold Start(Weak Model), Optimization Method=Base2025.081.45
Dr.GRPOModel Backbone=Qwen2.5-7B-Cold Start(Weak Model), Optimization Method=Dr.GRPO2025.081.27
BaseBackbone=LLaMa3.1-8B-Base2025.100