List your company

Mathematical Reasoning on DeepScaleR (test)

39.2Greedy SuccessAvg Probability
8.707216.623624.5432.4564Feb 3, 2026

Evaluation results

MethodLinks
Avg ProbabilityBackbone=Qwen 3B, Group size (G)=322026.0239.228.79-0.59-0.852.33460—
ProbabilityBackbone=Qwen 3B, Group size (G)=322026.0238.6836.83-1.68-2.7315.28481.15—
Base RLBackbone=Qwen 3B, Group size (G)=322026.0238.338.56-2.38-3.5534.85453.69—
Avg LogprobBackbone=Qwen 3B, Group size (G)=322026.0238.1225.82-0.57-0.72.02463.46—
Log-probBackbone=Qwen 3B, Group size (G)=322026.0237.9225.79-0.56-0.591.8532.79—
Log-probBackbone=Llama 3B, Group size (G)=322026.0232.417.97-0.79-0.832.3437.11—
Avg LogprobBackbone=Llama 3B, Group size (G)=322026.0232.218.32-0.8-0.842.32371.88—
ProbabilityBackbone=Llama 3B, Group size (G)=322026.0231.7528.47-1.63-2.3210.28340.01—
Avg ProbabilityBackbone=Llama 3B, Group size (G)=322026.0230.7521.71-0.85-1.313.71335.89—
Base RLBackbone=Llama 3B, Group size (G)=322026.0228.5528.54-2.01-2.7716.89294.52—
Base modelBackbone=Qwen 3B, Group size (G)=322026.0214.0211—-2.5713.12218.12—
SFT (no CoT)Backbone=Qwen 3B, Group size (G)=322026.0211.526.48-0.76-0.762.145—
Base modelBackbone=Llama 3B, Group size (G)=322026.0210.055.2—-4.92137.25354.91—
SFT (no CoT)Backbone=Llama 3B, Group size (G)=322026.029.885.14-1.01-1.012.745—
AdaBackModel Size=1B2025.06——————9
AdaBackModel Size=3B2025.06——————10.6
Base+RLModel Size=1B2025.06——————6.8
Base+RLModel Size=3B2025.06——————6.6
SFT+AdaBackModel Size=1B2025.06——————9.5
SFT+AdaBackModel Size=3B2025.06——————12.5
SFT+RLModel Size=1B2025.06——————7.1
SFT+RLModel Size=3B2025.06——————9.1