List your company
ResearchFormerly WizWand

Mathematical Reasoning on GSM8K (avg@K)

94.26Accuracy (avg@K)CL-EGRSD
86.907288.816190.72592.6339May 13, 2026
Updated 5mo ago

Evaluation results

MethodLinks
CL-EGRSDBackbone=Qwen3-4B, Training Budget=100-step, Thinking Mode=True, K (Samples per prompt)=4, Temperature=1.0, Top-p=0.95, Max Tokens=32,7682026.0594.26
CRISPBackbone=Qwen3-4B, Training Budget=100-step, Thinking Mode=True, K (Samples per prompt)=4, Temperature=1.0, Top-p=0.95, Max Tokens=32,7682026.0594.01
Baseline (no-train)Backbone=Qwen3-4B, Training Budget=N/A, Thinking Mode=True, K (Samples per prompt)=4, Temperature=1.0, Top-p=0.95, Max Tokens=32,7682026.0593.96
GRPOBackbone=Qwen3-4B, Training Budget=100-step, Thinking Mode=True, K (Samples per prompt)=4, Temperature=1.0, Top-p=0.95, Max Tokens=32,7682026.0593.95
OPSDBackbone=Qwen3-4B, Training Budget=100-step, Thinking Mode=True, K (Samples per prompt)=4, Temperature=1.0, Top-p=0.95, Max Tokens=32,7682026.0593.89
EGRSDBackbone=Qwen3-4B, Training Budget=100-step, Thinking Mode=True, K (Samples per prompt)=4, Temperature=1.0, Top-p=0.95, Max Tokens=32,7682026.0593.88
SFTBackbone=Qwen3-4B, Training Budget=100-step, Thinking Mode=True, K (Samples per prompt)=4, Temperature=1.0, Top-p=0.95, Max Tokens=32,7682026.0587.19