IF-RewardBench - Overall Assessment (System-Prompt): leaderboard
Metric: Kendall tau-b (-1 to 1) between the judge ranking of the responses to each instruction (about seven) and the human-verified preference graph, on the 247 system-prompt steerability instructions; general LLMs compare every response pair (pairwise prompt, Elo aggregation), reward models score each response; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 17 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 0.49 | #93 |
| 2 | GPT-5 Mini | 0.41 | #176 |
| 3 | DeepSeek V3.2 | 0.21 | #198 |
| 4 | GLM-4.6 (Thinking) | 0.19 | #246 (GLM-4.6) |
| 5 | QwQ-32B | 0.11 | #410 |
| 6 | GLM 4.5 Air (Thinking) | 0.08 | #377 (GLM 4.5 Air) |
| 7 | Qwen 3 32B (Thinking) | 0.04 | #424 (Qwen 3 32B) |
| 8 | Qwen 2.5 72B Instruct | 0 | #436 |
| 9 | Llama 3.3 70B Instruct | -0.06 | #520 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=if-rewardbench-overall-assessment-system-prompt · How It Works · Data refreshed daily, snapshot 2026-10-11.