Step-level reasoning verification on Aggregate ID/OOD
37.3ID ScoreQwen2.5-Math-7B-PRM800k
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Qwen2.5-Math-7B-PRM800k# Sample=265K2025.11 | 37.3 | 53.5 | 47.4 | |
| Qwen2.5-Math-7B# Sample=860K2025.11 | 34.8 | 51.7 | 45.3 | |
| Universal-PRM-Qwen2.5-Math-7B# Sample=690K2025.11 | 32 | 47.4 | 41.6 | |
| ReProbe, Attn+Logit, Qwen3-8B-anno# Sample=32K2025.11 | 30 | 53.8 | 44.8 | |
| Math-Shepherd-PRM-7B# Sample=440K2025.11 | 20.8 | 47.9 | 37.8 | |
| Skywork-PRM-1.5B# Sample=Unk2025.11 | 19.5 | 37.1 | 30.5 | |
| RLHFlow-PRM-Deepseek-8B# Sample=253K2025.11 | 19.1 | 42.2 | 33.5 | |
| RLHFlow-PRM-Mistral-8B# Sample=273K2025.11 | 14.3 | 35.3 | 27.4 | |
| H4-Qwen2.5-PRM-1.5B-0.2# Sample=369K2025.11 | 11.3 | 35.4 | 26.4 | |
| MaxProb# Sample=-2025.11 | 11.1 | 39.1 | 28.6 | |
| MaxEntropy# Sample=-2025.11 | 9.9 | 37.3 | 27 | |
| Perplexity# Sample=-2025.11 | 9.4 | 35.2 | 25.5 | |
| Random# Sample=-2025.11 | 7.5 | 31.9 | 22.8 |