KOR-Bench - Counterfactual: leaderboard
Answering within 25 counterfactual worlds whose stated rules contradict real-world knowledge; the score is accuracy under the fictional rules. Chat models, zero-shot. Higher is better. Source: arxiv.org/abs/2410.06526.
Metric: Accuracy (%; zero-shot, 250 questions). Source: arxiv.org. Saturation forecast: Estimated already saturated. 38 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O1 Preview (2024-09-12) | 92.8 | #220 |
| 2 | GPT-4o (2024-05-13) | 92 | #338 |
| 3 | O1 Mini (2024-09-12) | 91.6 | #414 |
| 4 | Claude 3.5 Sonnet (20240620) | 91.2 | #309 |
| 5 | Command-R+ (08-2024) | 90.4 | #839 |
| 6 | Llama 3 70B Instruct | 89.6 | #643 |
| 7 | Qwen 2 72B Instruct | 89.6 | #584 |
| 8 | Mistral Large 2 (Jul) | 89.2 | #498 |
| 9 | Qwen 2.5 32B Instruct | 88.8 | #511 |
| 10 | Qwen 2.5 14B Instruct | 88.8 | #657 |
| 11 | Qwen 2 7B Instruct | 88.8 | #983 |
| 12 | c4ai-command-r-08-2024 | 88.4 | #945 |
| 13 | Llama 3.1 70B Instruct | 88 | #570 |
| 14 | Phi-3.5-MoE-instruct | 88 | #706 |
| 15 | Llama 3.1 405B Instruct | 87.6 | #472 |
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-counterfactual · How It Works · Data refreshed daily, snapshot 2026-11-07.