OrderProbe: leaderboard
OrderProbe (Kang et al., ACL 2026) gives a model a four-character Simplified Chinese, Traditional Chinese, Japanese or Korean expression with its characters scrambled and asks for the canonical order plus a one-line meaning; every expression appears in all 23 non-identity permutations. The score is the exact-match recovery rate in percent under the zero-shot prompt, higher is better.
Metric: Exact Recovery Rate (%; zero-shot prompt). Source: arxiv.org. Saturation forecast: Around 2031. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 14B | 31.06 | #1235 |
| 2 | Gemini 2.5 Flash | 30.49 | #452 |
| 3 | GPT-4o | 23.25 | #741 |
| 4 | DeepSeek V3.2 | 20.1 | #403 |
| 5 | DeepSeek R1 | 18.6 | #467 |
| 6 | Claude Haiku 4.5 | 18.02 | #568 |
| 7 | Gemma 3 4B (IT) | 17.09 | #3848 |
| 8 | Qwen 3 8B | 15.79 | #1838 |
| 9 | Llama 2 7B Chat | 12.58 | #8862 |
| 10 | Llama 3 8B Instruct | 12.17 | #4652 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=orderprobe · How It Works · Data refreshed daily, snapshot 2026-11-12.