List your company
ResearchFormerly WizWand

Step-level correctness assessment on SciQA (test)

55.3PR-AUCReProbe
10.78822.34433.945.456Nov 9, 2025
Updated 6mo ago

Evaluation results

MethodLinks
ReProbe# Sample=32K, Architecture=Hidden States, Annotation=Self-anno2025.1155.3
RLHFlow-PRM-Deepseek-8B# Sample=253K2025.1151.8
ReProbe# Sample=32K, Architecture=Hidden States, Annotation=DeepSeek-anno2025.1145
Qwen2.5-Math-PRM-7B# Sample=860K2025.1142.9
Math-Shepherd-PRM-7B# Sample=440K2025.1141.5
RLHFlow-PRM-Mistral-8B# Sample=273K2025.1141.5
Skywork-PRM-1.5B# Sample=Unk2025.1140.8
Qwen2.5-Math-7B-PRM800k# Sample=265K2025.1140.4
ReProbe# Sample=32K, Architecture=Attn+Logit, Annotation=Self-anno2025.1140.4
ReProbe# Sample=32K, Architecture=Attn+Logit, Annotation=DeepSeek-anno2025.1136.1
Universal-PRM-Qwen2.5-Math-7B# Sample=690K2025.1133
Semantic Entropy# Sample=-2025.1126.5
P(True)# Sample=-2025.1126.3
MaxProb# Sample=-2025.1125.9
Degree Matrix# Sample=-2025.1125.8
Lexical Similarity# Sample=-2025.1125.4
Self-Certainty# Sample=-2025.1124.3
CCP# Sample=-2025.1123.5
Perplexity# Sample=-2025.1122.8
H4-Qwen2.5-PRM-1.5B-0.2# Sample=369K2025.1122.8
MaxEntropy# Sample=-2025.1122.7
Random# Sample=-2025.1112.5