CodeMixQA - Grammar Forcing (Tgt): leaderboard
CodeMixQA (Winata et al., 2026) questions produced by grammar forcing on the target language: GPT-5.2 writes natural code-switched text constrained to keep the target language's grammatical structure, across 16 code-switched variants (English mixed into Bengali, Chinese, French, Hindi, Indonesian, Italian, Japanese, Korean, Marathi, Spanish, Toba Batak and Urdu, with romanised forms of the four Indic languages). SimpleQA F1 from a Qwen3-30B-A3B judge, 0-100; higher is better.
Metric: SimpleQA F1 (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4.1 | 36.86 | #492 |
| 2 | GPT-OSS-120B | 21.09 | #699 |
| 3 | Qwen 3 30B A3B (Thinking) | 20.44 | #1124 (Qwen 3 30B A3B) |
| 4 | Llama 3.3 70B Instruct | 17.67 | #1353 |
| 5 | GPT-OSS-20B | 15.84 | #1150 |
| 6 | Gemma 3 27B | 11.18 | #1638 |
| 7 | Qwen 3 4B | 8.16 | #2297 |
| 8 | Gemma 3 12B | 7.91 | #2013 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=codemixqa-grammar-forcing-tgt · How It Works · Data refreshed daily, snapshot 2026-11-12.