Step-level correctness assessment on Calendar (test)
0.789PR-AUCReProbe
Evaluation Results
| Method | Links | |
|---|---|---|
| ReProbe# Sample=32K, Architecture=Hidden States, Annotation=DeepSeek-anno2025.11 | 0.789 | |
| ReProbe# Sample=32K, Architecture=Attn+Logit, Annotation=DeepSeek-anno2025.11 | 0.786 | |
| ReProbe# Sample=32K, Architecture=Hidden States, Annotation=Self-anno2025.11 | 0.785 | |
| ReProbe# Sample=32K, Architecture=Attn+Logit, Annotation=Self-anno2025.11 | 0.779 | |
| Qwen2.5-Math-PRM-7B# Sample=860K2025.11 | 0.745 | |
| Qwen2.5-Math-7B-PRM800k# Sample=265K2025.11 | 0.727 | |
| Universal-PRM-Qwen2.5-Math-7B# Sample=690K2025.11 | 0.691 | |
| Math-Shepherd-PRM-7B# Sample=440K2025.11 | 0.657 | |
| RLHFlow-PRM-Deepseek-8B# Sample=253K2025.11 | 0.504 | |
| Skywork-PRM-1.5B# Sample=Unk2025.11 | 0.502 | |
| RLHFlow-PRM-Mistral-8B# Sample=273K2025.11 | 0.499 | |
| H4-Qwen2.5-PRM-1.5B-0.2# Sample=369K2025.11 | 0.495 | |
| Semantic Entropy# Sample=-2025.11 | 0.492 | |
| P(True)# Sample=-2025.11 | 0.49 | |
| Lexical Similarity# Sample=-2025.11 | 0.49 | |
| Random# Sample=-2025.11 | 0.486 | |
| Degree Matrix# Sample=-2025.11 | 0.484 | |
| MaxProb# Sample=-2025.11 | 0.483 | |
| Self-Certainty# Sample=-2025.11 | 0.482 | |
| CCP# Sample=-2025.11 | 0.452 | |
| MaxEntropy# Sample=-2025.11 | 0.443 | |
| Perplexity# Sample=-2025.11 | 0.418 |