PRMBench: leaderboard
Process Reward Model benchmark evaluating fine-grained mathematical reasoning step verification across multiple error types.
Metric: Overall Score. Source: prmbench.github.io. Saturation forecast: Estimated already saturated. 26 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O1 Mini | 68.8 | #323 |
| 2 | GPT-4o | 66.8 | #343 |
| 3 | Gemini 2.0 Flash (Preview) | 66 | #370 |
| 4 | QwQ 32B-Preview | 63.6 | #651 |
| 5 | DeepSeek R1 Distill Qwen 32B | 60.2 | #655 |
| 6 | DeepSeek R1 Distill Llama 70B | 57.5 | #594 |
| 7 | Qwen2.5-Math-72B | 57.4 | #727 |
| 8 | DeepSeek R1 Distill Llama 8B | 52.7 | #1291 |
| 9 | DeepSeek-R1-Distill-Qwen-7B | 52.6 | #1415 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=prmbench · How It Works · Data refreshed daily, snapshot 2026-10-20.