List your company
ResearchFormerly WizWand

Step-level correctness assessment on ProofNet (test)

32.9PR-AUCUniversal-PRM-Qwen2.5-Math-7B
10.95616.65322.3528.047Nov 9, 2025
Updated 6mo ago

Evaluation results

MethodLinks
Universal-PRM-Qwen2.5-Math-7B# Sample=690K2025.1132.9
Qwen2.5-Math-PRM-7B# Sample=860K2025.1131
Qwen2.5-Math-7B-PRM800k# Sample=265K2025.1130.1
ReProbe# Sample=32K, Architecture=Hidden States, Annotation=DeepSeek-anno2025.1128.1
ReProbe# Sample=32K, Architecture=Hidden States, Annotation=Self-anno2025.1126.4
ReProbe# Sample=32K, Architecture=Attn+Logit, Annotation=Self-anno2025.1126
ReProbe# Sample=32K, Architecture=Attn+Logit, Annotation=DeepSeek-anno2025.1124.3
MaxProb# Sample=-2025.1119.4
MaxEntropy# Sample=-2025.1118.5
Perplexity# Sample=-2025.1117.6
Semantic Entropy# Sample=-2025.1117.3
P(True)# Sample=-2025.1117.2
Lexical Similarity# Sample=-2025.1117
CCP# Sample=-2025.1116.8
H4-Qwen2.5-PRM-1.5B-0.2# Sample=369K2025.1115.9
Self-Certainty# Sample=-2025.1115.5
Random# Sample=-2025.1115.3
Degree Matrix# Sample=-2025.1114.7
Skywork-PRM-1.5B# Sample=Unk2025.1114.7
Math-Shepherd-PRM-7B# Sample=440K2025.1114.7
RLHFlow-PRM-Deepseek-8B# Sample=253K2025.1113.6
RLHFlow-PRM-Mistral-8B# Sample=273K2025.1111.8