EvalYaks CEFR B2 (Without Descriptors): leaderboard
Metric: Average acceptable accuracy (%) across speaking-test parts 1-4, predicting 1-5 rubric scores from expert-reviewed synthetic transcripts; without descriptors; the paper counts accurate, partly accurate and acceptable multi-criterion responses, not just exact matches; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 1.0 Pro | 69 | #916 |
| 2 | vicuna-33B-v1.3 | 66 | #1332 |
| 3 | Claude 3 Haiku | 62 | #789 |
| 4 | Llama 2 70B Chat | 56 | #1223 |
| 5 | Llama 2 7B Chat | 53 | #1481 |
| 6 | Mixtral 8x7B Instruct (v0.1) | 47 | #1058 |
| 7 | Qwen-72B-Chat | 46 | #833 |
| 8 | Mistral Medium | 46 | #767 |
| 9 | GPT-3.5 Turbo (0125) | 43 | #969 |
| 10 | Mistral 7B Instruct (v0.2) | 40 | #1316 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=evalyaks-cefr-b2-without-descriptors · How It Works · Data refreshed daily, snapshot 2026-10-12.