List your company

Process-level verification on MATH ProcessBench (test)

9.6Error RateRLHFlow-PRM-Mistral-8B
7.90819.32930.7542.171May 20, 2025

Evaluation results

MethodLinks
RLHFlow-PRM-Mistral-8Btraining_data_source=automatically constructed datasets2025.059.645.215.8
RLHFlow-PRM-Deepseek-8Btraining_data_source=automatically constructed datasets2025.0510.951.916.9
Skywork-PRM-1.5Btraining_data_source=automatically constructed datasets2025.0513.632.819.2
Skywork-PRM-7Btraining_data_source=automatically constructed datasets2025.051441.921
Math-Shepherd-PRM-7Btraining_data_source=automatically constructed datasets2025.0514.27323.8
SCOPE (Step Skipping)Step Compression Strategy=Step Skipping2025.0515.97126
SCOPE (Step Replacement)Step Compression Strategy=Step Replacement2025.0516.568.526.6
Math-Shepherd-PRM-7Btraining_data_source=automatically constructed datasets2025.05188229.5
RLHFlow-PRM-Deepseek-8Btraining_data_source=automatically constructed datasets2025.0521.48033.8
RLHFlow-PRM-Mistral-8Btraining_data_source=automatically constructed datasets2025.0521.772.233.4
EurusPRM-Stage1training_data_source=automatically constructed datasets2025.0521.924.523.1
EurusPRM-Stage2training_data_source=automatically constructed datasets2025.0523.119.120.9
SCOPE (w/o Code Translation)AST Normalization=true, Code Translation=false2025.0526.556.836.1
Qwen2.5-Math-7B-PRM800Ktraining_data_source=high-quality manually annotated process-level data (PRM800K)2025.0529.886.144.3
SCOPE (w/o AST Normalization)AST Normalization=false, Code Translation=true2025.0529.857.339.2
SCOPEAST Normalization=true, Code Translation=true, training_data_source=automatically constructed datasets2025.0531.462.241.7
EurusPRM-Stage1training_data_source=automatically constructed datasets2025.0533.338.235.6
SCOPE (Step Replacement)Step Compression Strategy=Step Replacement2025.0535.58349.8
EurusPRM-Stage2training_data_source=automatically constructed datasets2025.0536.43535.7
Skywork-PRM-1.5Btraining_data_source=automatically constructed datasets2025.0537.965.248
SCOPE (Step Skipping)Step Compression Strategy=Step Skipping2025.053884.552.5
Skywork-PRM-7Btraining_data_source=automatically constructed datasets2025.0543.862.253.6
SCOPE (w/o Code Translation)AST Normalization=true, Code Translation=false2025.054775.657.9
Qwen2.5-Math-7B-PRM800Ktraining_data_source=high-quality manually annotated process-level data (PRM800K)2025.054890.162.6
SCOPEAST Normalization=true, Code Translation=true, training_data_source=automatically constructed datasets2025.0550.874.960.6
SCOPE (w/o AST Normalization)AST Normalization=false, Code Translation=true2025.0551.974.661.2