OrderProbe - Chain-of-Thought: leaderboard
OrderProbe's exact-match recovery rate in percent (higher is better) when the prompt asks for one line of reasoning before the recovered expression, over the same scrambled four-character Chinese, Japanese and Korean expressions in all 23 non-identity permutations (Kang et al., ACL 2026).
Metric: Exact Recovery Rate (%; chain-of-thought prompt). Source: arxiv.org. Saturation forecast: Around August 2027. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V3.2 | 62.24 | #403 |
| 2 | Gemini 2.5 Flash | 41.56 | #452 |
| 3 | GPT-4o | 37.69 | #741 |
| 4 | Qwen 3 14B | 31.53 | #1235 |
| 5 | DeepSeek R1 | 27.73 | #467 |
| 6 | Llama 3 8B Instruct | 24.43 | #4652 |
| 7 | Claude Haiku 4.5 | 21.65 | #568 |
| 8 | Llama 2 7B Chat | 18.75 | #8862 |
| 9 | Gemma 3 4B (IT) | 18.03 | #3848 |
| 10 | Qwen 3 8B | 12.59 | #1838 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=orderprobe-chain-of-thought · How It Works · Data refreshed daily, snapshot 2026-11-12.