Process-level Reward Modeling on PROCESSBENCH Omni-MATH
2.8Error RateSPARE-Llama3-8B
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| SPARE-Llama3-8B# Train=40.5K2025.06 | 2.8 | 82.2 | 5.4 | |
| RLHFlow-Deepseek-8B# Train=250K2025.06 | 10.9 | 51.9 | 16.9 | |
| Skywork-7B2025.06 | 14 | 41.9 | 21 | |
| SPARE-Qwen2.5-3B# Train=40.5K2025.06 | 14 | 83.8 | 23.9 | |
| Math-Shepherd-7B# Train=440K2025.06 | 14.2 | 73 | 23.8 | |
| Qwen-2.5-Math-7B-PRM800K (Human)# Train=250K2025.06 | 29.8 | 86.1 | 44.3 |