RewardBench Reasoning: leaderboard
RewardBench reasoning slice evaluates preference-model choices on reasoning-heavy responses.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 171 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | ArmoRM-Llama3-8B-v0.1 | 97.35 | #1394 |
| 2 | Gemini 1.5 Pro (0514) | 91.99 | #512 |
| 3 | Qwen 1.5 7B Chat | 90.41 | #1324 |
| 4 | Gemini 1.5 Pro (Sept) | 90.22 | #448 |
| 5 | Qwen 1.5 14B Chat | 89.61 | #977 |
| 6 | stablelm-2-12B-chat | 89.45 | #1138 |
| 7 | bagel-dpo-34B-v0.5 | 88.89 | #962 |
| 8 | Llama 3.1 405B Instruct | 87.15 | #457 |
| 9 | GPT-4 Preview (0125) | 86.92 | #486 |
| 10 | GPT-4o (2024-08-06) | 86.61 | #336 |
| 11 | Llama 3.1 70B Instruct | 85.99 | #561 |
| 12 | Qwen 1.5 72B Chat | 85.54 | #776 |
| 13 | Gemini 1.5 Flash (001) | 85.12 | #577 |
| 14 | GPT-4o (2024-05-13) | 84.87 | #325 |
| 15 | Claude 3.5 Sonnet (20240620) | 84.69 | #299 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=rewardbench-reasoning · How It Works · Data refreshed daily, snapshot 2026-10-20.