RMGAP - EndoRM: 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. The reward is the model's discounted sum of token log-probabilities of the response (EndoRM). Pairwise accuracy is the percentage of comparisons in which the requested response scores above an alternative, averaged over the four domains; higher is better.
Metric: Pairwise accuracy (%; chosen response scored above a rejected one, prompts asking for one of four styles, mean of chat, writing, reasoning and safety; discounted sum of the model's token log-probabilities of the response (gamma 0.93)). Source: arxiv.org. Saturation forecast: Estimated already saturated. 2 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2.5 7B Instruct | 61.3 | #2583 |
| 2 | Qwen 3 8B | 59.89 | #1838 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=rmgap-endorm · How It Works · Data refreshed daily, snapshot 2026-11-12.