HoWToBench: 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 over all instructions in completion, outline-guided and open-ended writing; 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 all 1,302 Chinese writing instructions (completion, outline-guided and open-ended writing in 12 genres) of HoWToBench; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 9 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1HoWToBench DeepSeek-R1 (checkpoint unspecified)6.1
2O3 Mini (2025-01-31)5.86#496
3GPT-4o (2024-11-20)5.81#859
4Claude 3.5 Sonnet (20241022)5.58#648
5Gemini 2.0 Flash5.43#723
6HoWToBench DeepSeek-V3 (checkpoint unspecified)5.42#666
7GLM-4-Plus-2501115.34#831
8Claude 3 Haiku (20240307)5.01#2396
9Llama 3.3 70B Instruct4.59#1385

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=howtobench · How It Works · Data refreshed daily, snapshot 2026-11-21.