KOR-Bench - Counterfactual: leaderboard

Metric: Accuracy (%; zero-shot, 250 questions). Source: arxiv.org. Saturation forecast: Estimated already saturated. 38 models tracked.

Top models

#ModelScore
1O1 Preview (2024-09-12)92.8
2GPT-4o (2024-05-13)92
3O1 Mini (2024-09-12)91.6
4Claude 3.5 Sonnet (20240620)91.2
5Command-R+ (08-2024)90.4
6Llama 3 70B Instruct89.6
7Qwen 2 72B Instruct89.6
8Mistral Large 2 (Jul)89.2
9Qwen 2.5 32B Instruct88.8
10Qwen 2.5 14B Instruct88.8
11Qwen 2 7B Instruct88.8
12c4ai-command-r-08-202488.4
13Llama 3.1 70B Instruct88
14Phi-3.5-MoE-instruct88
15Llama 3.1 405B Instruct87.6

Interactive version: theaggregate.ai/benchmark?slug=kor-bench-counterfactual · How It Works · Data refreshed daily, snapshot 2026-09-24.