Process-level Reward Modeling on PROCESSBENCH MATH
6.1Error RateSPARE-Llama3-8B
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| SPARE-Llama3-8B# Train=40.5K2025.06 | 6.1 | 91.6 | 11.4 | |
| SPARE-Qwen2.5-3B# Train=40.5K2025.06 | 16 | 89.2 | 27.1 | |
| Math-Shepherd-7B# Train=440K2025.06 | 18 | 82 | 29.5 | |
| RLHFlow-Deepseek-8B# Train=250K2025.06 | 21.4 | 80 | 33.8 | |
| Skywork-7B2025.06 | 43.8 | 62.2 | 53.6 | |
| Qwen-2.5-Math-7B-PRM800K (Human)# Train=250K2025.06 | 48 | 90.1 | 62.6 |