KOR-Bench - Base Models (3-shot): leaderboard
Overall KOR-Bench accuracy for pretrained base checkpoints, given three generic worked examples per rule. Higher is better. Source: arxiv.org/abs/2410.06526.
Metric: Accuracy (%; three-shot, 1,250 questions). Source: arxiv.org. Saturation forecast: Estimated already saturated. 24 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Llama 3.1 405B | 39.68 | #477 |
| 2 | Qwen 2.5 72B | 37.28 | #514 |
| 3 | Qwen 2.5 32B | 37.28 | #535 |
| 4 | Llama 3 70B | 35.2 | #635 |
| 5 | Qwen 2 72B | 34.32 | #517 |
| 6 | Llama 3.1 70B | 33.84 | #592 |
| 7 | Gemma 2 27B | 33.36 | #805 |
| 8 | Qwen 2.5 14B | 33.28 | #718 |
| 9 | Yi 1.5 34B | 30.08 | #808 |
| 10 | Yi-1.5-9B | 29.2 | #1219 |
| 11 | Qwen 2.5 7B | 28.8 | #898 |
| 12 | Qwen 2 7B | 27.44 | #1072 |
| 13 | Llama 3.1 8B | 26 | #1134 |
| 14 | Gemma 2 9B | 25.52 | #927 |
| 15 | Llama 3 8B | 24.96 | #1231 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #6 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=kor-bench-base-models-3-shot · How It Works · Data refreshed daily, snapshot 2026-11-07.