RewardBench 2 Precise IF — leaderboard
Metric: Accuracy (%). Source: huggingface.co. 52 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 75.78 |
| 2 | Qwen 3.5 397B A17B | 75.62 |
| 3 | GPT-5.5 | 74.38 |
| 4 | Qwen 3.6 Plus | 73.75 |
| 5 | Qwen 3.5 122B A10B | 73.75 |
| 6 | GLM-5.1 FP8 | 73.28 |
| 7 | GLM-5 FP8 | 71.56 |
| 8 | Kimi K2.6 | 71.25 |
| 9 | Kimi K2.5 | 70 |
| 10 | GPT-5 Mini | 69.84 |
| 11 | GPT-OSS-120B | 69.53 |
| 12 | Kimi K2 (Thinking) | 69.06 |
| 13 | DeepSeek-V4-Flash-FP8 | 68.91 |
| 14 | DeepSeek V3.2 Speciale | 68.75 |
| 15 | NVIDIA-Nemotron-3-Super-120B-A12B-FP8 | 68.44 |
Interactive version: theaggregate.ai/benchmark?slug=rewardbench-2-precise-if · How the rankings work · Data refreshed daily, snapshot 2026-07-22.