YOMI-Bench - Kanji Reading QA (Multiple Readings): leaderboard
Yes/no questions on whether a given reading is correct for a kanji inside a word, for kanji with several possible readings, from YOMI-Bench (2026). Accuracy in percent averaged over five paraphrased prompts, 4-shot; higher is better.
Metric: Accuracy (%; yes/no judgement of whether a given reading is correct for a kanji in a word, 4-shot with two positive and two negative examples; 450 items on kanji with several readings; mean over five paraphrased prompts; accuracy as a fraction x100). Source: arxiv.org. Saturation forecast: Estimated already saturated. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 98.48 | #167 |
| 2 | Gemini 2.5 Flash | 98.17 | #453 |
| 3 | GPT-4o | 94.04 | #741 |
| 4 | Mistral Medium 3.1 | 87.37 | #852 |
| 5 | Llama-3.1-Swallow-8B-Instruct-v0.5 | 77.82 | #2704 |
| 6 | Claude Sonnet 4.5 | 77.33 | #250 |
| 7 | Ministral-8B-Instruct-2410 | 63.55 | #4684 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=yomi-bench-kanji-reading-qa-multiple-readings · How It Works · Data refreshed daily, snapshot 2026-11-13.