GB/T-Bench - Direct Review DMTR@9: leaderboard
GB/T-Bench (arXiv 2608.06312v1): review of 488 Chinese national standard documents seeded with 7,306 errors across 25 error types, by direct LLM review without the authors' multi-agent reviewer. Share of documents with at least 9 exactly diagnosed errors (each has 10 to 18 seeded), percent; human experts reach 90.2 (Table 2). Higher is better.
Metric: Documents with at least 9 correct diagnoses (%). Source: arxiv.org. Saturation forecast: Around April 2027. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 39.34 | #33 |
| 2 | GPT-5.6 Sol | 36.96 | #22 |
| 3 | Claude Fable 5 | 32.61 | #20 |
| 4 | GPT-5.4 | 25 | #83 |
| 5 | Qwen 3.7 Plus | 18.85 | #98 |
| 6 | Gemini 3.5 Flash | 17.21 | #66 |
| 7 | Claude Opus 4.8 | 14.34 | #39 |
| 8 | DeepSeek V4 Pro | 12.3 | #112 |
| 9 | GPT-5.3 | 8.81 | #104 |
| 10 | Qwen 3 235B A22B 2507 Instruct | 8.2 | #315 |
| 11 | DeepSeek V4 Flash | 7.38 | #155 |
| 12 | GPT-5.4 Mini | 3.89 | #228 |
| 13 | MiniMax-M2.7 | 2.87 | #270 |
| 14 | Qwen 2.5 VL 72B Instruct | 2.46 | #383 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=gb-t-bench-direct-review-dmtr-9 · How It Works · Data refreshed daily, snapshot 2026-11-05.