Step Verification Efficiency Evaluation on MATH (test)
4Runtime (s)Universal-PRM-Qwen2.5-Math-7B
Evaluation results
| Method | Links | |
|---|---|---|
| Universal-PRM-Qwen2.5-Math-7BBatch size=1, Model scale category=PRMs 750× to 810× Larger than ReProbes2025.11 | 4 | |
| Math-Shepherd-PRM-7BBatch size=1, Model scale category=PRMs 750× to 810× Larger than ReProbes2025.11 | 5 | |
| RLHFlow-PRM-Mistral-DataBatch size=1, Model scale category=PRMs 750× to 810× Larger than ReProbes, Annotation source=Mistral-Data2025.11 | 5 | |
| RLHFlow-PRM-Deepseek-DataBatch size=1, Model scale category=PRMs 750× to 810× Larger than ReProbes, Annotation source=Deepseek-Data2025.11 | 5 | |
| H4-Qwen2.5-PRM-1.5B-0.2Batch size=1, Model scale category=PRMs 150× Larger than ReProbes2025.11 | 8 | |
| ReProbe, Hidden States, DeepSeek-annoBatch size=1, Model scale category=Reasoning Probes (ReProbes), Annotation source=DeepSeek-anno2025.11 | 13 | |
| ReProbe, Hidden States, Self-annoBatch size=1, Model scale category=Reasoning Probes (ReProbes), Annotation source=Self-anno2025.11 | 14 | |
| Skywork-PRM-1.5BBatch size=1, Model scale category=PRMs 150× Larger than ReProbes2025.11 | 17 | |
| Qwen2.5-Math-PRM-7BBatch size=1, Model scale category=PRMs 750× to 810× Larger than ReProbes2025.11 | 34 | |
| Qwen2.5-Math-7B-PRM800kBatch size=1, Model scale category=PRMs 750× to 810× Larger than ReProbes2025.11 | 37 |