WEval: leaderboard
WEval (2026) tests writing reward models and LLM judges on ranking candidate responses by how many of the specific requirements of a writing instruction each one satisfies. This board is the pairwise rank correlation with the gold ranking, scaled from -100 to 100; higher is better.
Metric: Ranking correlation (Kendall-style pairwise sign agreement with the gold ranking, x100, -100 to 100, averaged over instructions); WEval writing reward-model test: 2,750 writing instructions with five requirements each; for every instruction DeepSeek-R1 answers the full instruction and versions with 1 to 4 requirements dropped, and the gold ranking orders the five responses by how many requirements were kept; LLMs are prompted to rank the responses, reward models score them; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 5 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2.5 72B Instruct | 93.6 | #1117 |
| 2 | Qwen 2.5 7B Instruct | 85.4 | #2723 |
| 3 | Skywork-Reward-V2-Qwen3-8B | 82.5 | |
| 4 | Writing-Critic-7B | 80.6 | |
| 5 | Skywork-Reward-V2-Llama-3.1-8B | 75.4 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=weval · How It Works · Data refreshed daily, snapshot 2026-11-21.