GB/T-Bench - Direct Review DMTR@8: 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 8 exactly diagnosed errors (each has 10 to 18 seeded), percent; human experts reach 94.4 (Table 2). Higher is better.
Metric: Documents with at least 8 correct diagnoses (%). Source: arxiv.org. Saturation forecast: Around February 2027. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol | 54.35 | #22 |
| 2 | GPT-5.5 | 54.1 | #33 |
| 3 | Claude Fable 5 | 47.83 | #20 |
| 4 | GPT-5.4 | 38.73 | #83 |
| 5 | Qwen 3.7 Plus | 34.43 | #98 |
| 6 | Gemini 3.5 Flash | 30.12 | #66 |
| 7 | Claude Opus 4.8 | 23.57 | #39 |
| 8 | DeepSeek V4 Pro | 21.93 | #112 |
| 9 | GPT-5.3 | 18.44 | #104 |
| 10 | Qwen 3 235B A22B 2507 Instruct | 14.96 | #315 |
| 11 | DeepSeek V4 Flash | 14.55 | #155 |
| 12 | GPT-5.4 Mini | 10.25 | #228 |
| 13 | Qwen 2.5 VL 72B Instruct | 8.2 | #383 |
| 14 | MiniMax-M2.7 | 6.56 | #270 |
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-8 · How It Works · Data refreshed daily, snapshot 2026-11-05.