RewardBench 2 Precise IF: leaderboard
RewardBench 2 precise-instruction-following slice evaluates preference accuracy on tightly constrained instructions.
Metric: Score (%). Source: huggingface.co. Saturation forecast: Around January 2027. 249 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 75.78 | #58 |
| 2 | Qwen 3.5 397B A17B | 75.62 | #146 |
| 3 | GPT-5.5 | 74.38 | #31 |
| 4 | Qwen 3.6 Plus | 73.75 | #120 |
| 5 | Qwen 3.5 122B A10B | 73.75 | #173 |
| 6 | GLM-5.1 FP8 | 73.28 | #109 |
| 7 | GLM-5 FP8 | 71.56 | #143 |
| 8 | Kimi K2.6 | 71.25 | #104 |
| 9 | Kimi K2.5 | 70 | #144 |
| 10 | GPT-5 Mini | 69.84 | #186 |
| 11 | GPT-OSS-120B | 69.53 | #341 |
| 12 | Kimi K2 (Thinking) | 69.06 | #243 (Kimi K2) |
| 13 | DeepSeek V3.2 Speciale | 68.75 | #128 |
| 14 | NVIDIA-Nemotron-3-Super-120B-A12B-FP8 | 68.44 | #223 |
| 15 | Qwen 3.6 35B A3B | 68.12 | #226 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=rewardbench-2-precise-if · How It Works · Data refreshed daily, snapshot 2026-10-20.