RewardBench Chat: leaderboard
RewardBench chat slice evaluates preference-model accuracy on conversational response comparisons.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 178 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 3 70B Instruct | 97.63 | #634 |
| 2 | tulu-2-dpo-70B | 97.49 | #1052 |
| 3 | tulu-2-dpo-7B | 97.49 | #1400 |
| 4 | Llama 3.1 70B Instruct | 97.21 | #561 |
| 5 | Llama 3.1 405B Instruct | 97.21 | #457 |
| 6 | ArmoRM-Llama3-8B-v0.1 | 96.93 | #1394 |
| 7 | GPT-4o (2024-05-13) | 96.65 | #325 |
| 8 | stablelm-2-12B-chat | 96.65 | #1138 |
| 9 | stablelm-2-zephyr-1.6B | 96.65 | #1485 |
| 10 | Claude 3.5 Sonnet (20240620) | 96.37 | #299 |
| 11 | Hermes 3 - Llama-3.1 70B | 96.23 | #927 |
| 12 | GPT-4o (2024-08-06) | 96.09 | #336 |
| 13 | tulu-2-dpo-13B | 95.81 | #1340 |
| 14 | zephyr-7B-gemma-v0.1 | 95.81 | #1345 |
| 15 | GPT-4 Turbo | 95.25 | #404 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=rewardbench-chat · How It Works · Data refreshed daily, snapshot 2026-10-20.