GB/T-Bench - Direct Review Diagnosis Recall: 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 seeded errors whose section, review dimension and error type all match, percent; human experts reach 83.1 (Table 2). Higher is better.
Metric: Diagnosis recall (%; exact section, dimension and error-type match). Source: arxiv.org. Saturation forecast: Around March 2028. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol | 52.03 | #22 |
| 2 | GPT-5.5 | 50.52 | #33 |
| 3 | Claude Fable 5 | 49.93 | #20 |
| 4 | Qwen 3.7 Plus | 43.51 | #98 |
| 5 | GPT-5.4 | 42.87 | #83 |
| 6 | Gemini 3.5 Flash | 40.87 | #66 |
| 7 | DeepSeek V4 Pro | 39.61 | #112 |
| 8 | GPT-5.3 | 36.2 | #104 |
| 9 | DeepSeek V4 Flash | 34.56 | #155 |
| 10 | GPT-5.4 Mini | 30.52 | #228 |
| 11 | Claude Opus 4.8 | 29.04 | #39 |
| 12 | Qwen 2.5 VL 72B Instruct | 27.31 | #383 |
| 13 | Qwen 3 235B A22B 2507 Instruct | 25.51 | #315 |
| 14 | MiniMax-M2.7 | 25.16 | #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-diagnosis-recall · How It Works · Data refreshed daily, snapshot 2026-11-05.