RMGAP - Best-of-4: leaderboard
RMGAP (Zhou and Li, 2026) tests whether reward models follow a preference stated in the prompt: each of 1,097 instances pairs four responses in deliberately different styles with prompts that each ask for one of those styles (three paraphrases each), across chat, writing, reasoning and safety. Scalar reward models score each response directly, and DPO-trained models by their implicit reward against the reference model. Best-of-4 accuracy is the percentage of cases in which the requested response scores above all three alternatives, averaged over the four domains; higher is better.
Metric: Best-of-4 accuracy (%; chosen response scored above all three alternatives, mean of chat, writing, reasoning and safety; scalar reward models score each response, DPO models by their implicit reward against the reference model). Source: arxiv.org. Saturation forecast: Estimated already saturated. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama-3.1-Tulu-3-8B-DPO | 31.07 | #3885 |
| 2 | zephyr-7B-beta | 29.91 | #7159 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=rmgap-best-of-4 · How It Works · Data refreshed daily, snapshot 2026-11-12.