PRMBench: leaderboard
Process Reward Model benchmark evaluating fine-grained mathematical reasoning step verification across multiple error types.
Metric: Overall Score. Source: prmbench.github.io. Status: saturated. 26 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | O1 Mini | 68.8 |
| 2 | GPT-4o | 66.8 |
| 3 | DeepSeek R1 Distill Qwen 32B | 60.2 |
| 4 | DeepSeek R1 Distill Llama 70B | 57.5 |
| 5 | Qwen2.5-Math-72B | 57.4 |
| 6 | DeepSeek R1 Distill Llama 8B | 52.7 |
| 7 | DeepSeek-R1-Distill-Qwen-7B | 52.6 |
Interactive version: theaggregate.ai/benchmark?slug=prmbench · How It Works · Data refreshed daily, snapshot 2026-09-05.