TW-LegalBench - Judgment Prediction: leaderboard
Criminal judgment prediction in TW-LegalBench (2026), a Traditional Chinese benchmark built from Taiwan official legal examinations and court judgments: ROUGE-L times 100 between the predicted holding and the main result of the court for first-instance criminal judgments, zero-shot with chain of thought. Higher is better.
Metric: ROUGE-L x100 (0-100), Jieba-segmented, of the predicted judgment holding against the court's main result, 535 first-instance criminal judgments over 107 offense categories, zero-shot chain-of-thought. Source: arxiv.org. Saturation forecast: Around June 2027. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.5 | 53.5 | #250 |
| 2 | GPT-5.2 | 42.7 | #176 |
| 3 | GPT-5 | 39.9 | #167 |
| 4 | Qwen 3 235B A22B | 32.5 | #623 |
| 5 | Llama 3.1 405B | 29.8 | #1076 |
| 6 | GPT-4o (2024-08-06) | 28.4 | #740 |
| 7 | GPT-OSS-120B | 26.3 | #700 |
| 8 | GPT-OSS-20B | 24.6 | #1147 |
| 9 | nemotron-3-nano-30B-a3B | 22.5 | #1875 |
| 10 | Qwen 2.5 7B | 21.7 | #2703 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=tw-legalbench-judgment-prediction · How It Works · Data refreshed daily, snapshot 2026-11-13.