HoWToBench - Completion: leaderboard
HoWToBench (2026) tests long-form Chinese writing against curated human references; GPT-4o rates content traits 1 to 10 and format traits 0, 5 or 10, and the Tree-of-Writing method combines them with explicit per-instruction weights. Mean score on completion tasks, where the model restores paragraphs removed from a human text; higher is better.
Metric: Tree-of-Writing score: GPT-4o rates each content trait 1-10 against the human reference and format traits 0, 5 or 10, combined with per-instruction LLM-assigned weights that sum to 1, mean over the 379 completion tasks (fill in removed paragraphs of a human text) of HoWToBench; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 10 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o (2024-11-20) | 6.6 | #859 |
| 2 | O3 Mini (2025-01-31) | 6.16 | #496 |
| 3 | HoWToBench DeepSeek-R1 (checkpoint unspecified) | 6.1 | |
| 4 | Doubao-pro-241225 | 5.58 | #810 |
| 5 | Claude 3.5 Sonnet (20241022) | 5.55 | #648 |
| 6 | HoWToBench DeepSeek-V3 (checkpoint unspecified) | 5.44 | #666 |
| 7 | Gemini 2.0 Flash | 5.43 | #723 |
| 8 | GLM-4-Plus-250111 | 5.19 | #831 |
| 9 | Claude 3 Haiku (20240307) | 5.12 | #2396 |
| 10 | Llama 3.3 70B Instruct | 4.36 | #1385 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Example tasks
有人说,唐诗是中国人精神生活的最大公约数。无论牙牙学语的儿童,还是两鬓斑斑的老人,谁都有几行烂熟于胸的诗句。
比如,君不见黄河之水天上来,奔流到海不复回;再如,无边落木萧萧下,不尽长江滚滚来;还有,莫愁前路无知己,天下谁人不识君。
但人们不知道,这些佳句能够流传至今,得益于《全唐诗》以及精选出的《唐诗三百首》。
那么,《全唐诗》又从何而来?这就与一位名叫胡震亨的明代海盐人密不可分。今天,我们一起来认识胡震亨和他的著作《唐音统签》。
一
[fill in the blank]
《唐音统签》全书以十干为纪,从《甲签》至《癸签》共计1033卷,收录了当时传世的盛唐、中唐、晚唐,乃至五代的诗,还包括断篇零句、词曲、歌谣、谚语、酒令。
书中还给每位诗人留下了个人小传和文学评论,因此,也有了李白、杜甫并称。
除了唐诗,作者胡震亨对唐诗体裁变迁、比兴体格、字句声调等的“研究见解”也收…
Interactive version: theaggregate.ai/benchmark?slug=howtobench-completion · How It Works · Data refreshed daily, snapshot 2026-11-21.