Long-form RewardBench - RAG: leaderboard
Metric: Accuracy (%) on the retrieval-augmented generation subset of Long-form RewardBench; each item pairs one chosen long-form response (validated by gemini-2.5-pro to beat every rejected one) with three rejected responses, and an item counts as correct only when the reward model ranks the chosen response above all rejected ones (best-of-n); sequence classifiers score each response, generative judges select the best response (Selection mode); higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 22 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4 (20250514) | 71.9 | #148 |
| 2 | GPT-4.1 | 70.6 | #240 |
| 3 | Claude 3.5 Sonnet (20240620) | 69.4 | #293 |
| 4 | Claude 3.7 Sonnet (20250219) | 67.2 | #196 |
| 5 | Gemini 2.5 Flash (Preview 04-17) | 60.4 | #230 |
| 6 | GPT-4o (2024-08-06) | 59.6 | #326 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=long-form-rewardbench-rag · How It Works · Data refreshed daily, snapshot 2026-10-11.