Phun-Bench - Homophone Recall: leaderboard
Phun-Bench (2026) tests whether text-only language models understand the sound of Chinese words, using tasks that cannot be solved by memorized pinyin alone. Homophone Recall gives a generated homophone of a common four-character idiom and asks for the original idiom; accuracy in percent, higher is better.
Metric: Accuracy (%) of recovering a four-character chengyu from an automatically generated homophone of it (1,098 items); thinking models run on 300 sampled items per setting, DeepSeek-V3 and GPT-4o on 500, the others on the full set; recommended sampling temperatures; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek R1 0528 | 97.7 | #428 |
| 2 | DeepSeek V3 (0324) | 66.4 | #709 |
| 3 | GPT-4o (2024-11-20) | 61.8 | #866 |
| 4 | Qwen 3 8B (Thinking) | 56.3 | #1830 (Qwen 3 8B) |
| 5 | Qwen 3 8B (Non-reasoning) | 15.2 | #1830 (Qwen 3 8B) |
| 6 | GLM-4 9B Chat | 12.4 | #2759 |
| 7 | Llama 3.2 3B Instruct | 0.7 | #7147 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=phun-bench-homophone-recall · How It Works · Data refreshed daily, snapshot 2026-11-13.