DDR-Bench - GLOBEM - Item-Averaged Message-Wise: leaderboard
Evaluates autonomous database exploration with SQL/Python tools on MIMIC, GLOBEM and 10-K. GPT-5 mini checks coverage of 2,058 checklist items across 291 task entities. Message-wise and trajectory-wise insights use sample- or item-averaging. Paper v1 scores are preserved; current repository defaults include a 100-turn safety cap. Released financial-trajectory labels resolve eight Qwen instruct/checkpoint identities; ambiguous Qwen3-30B-A3B and Kimi K2 snapshots remain explicit source-local rows. This series reports DDR-Bench - GLOBEM - Item-Averaged Message-Wise using Checklist Coverage Accuracy (%).
Metric: Checklist Coverage Accuracy (%). Source: arxiv.org. Saturation forecast: Around 2030. 22 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-4.6 | 41.61 | #265 |
| 2 | Claude Sonnet 4.5 | 40.23 | #148 |
| 3 | GPT-5.2 | 38.39 | #115 |
| 4 | DeepSeek V3.2 | 38.16 | #210 |
| 5 | GPT-5.1 | 37.88 | #145 |
| 6 | MiniMax-M2 | 37.24 | #327 |
| 7 | GPT-5 Mini | 36.09 | #193 |
| 8 | Qwen 3 Next 80B A3B Instruct | 35.4 | #381 |
| 9 | Gemini 3 Flash | 35.29 | #103 |
| 10 | Gemini 2.5 Pro | 33.79 | #159 |
| 11 | Qwen2.5-7B-Instruct-1M | 29 | #886 |
| 12 | Gemini 2.5 Flash | 28.95 | #255 |
| 13 | Qwen2.5-14B-Instruct-1M | 28.8 | #581 |
| 14 | Qwen 2.5 72B Instruct | 28.74 | #469 |
| 15 | Qwen 2.5 7B Instruct | 27.36 | #863 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=ddr-bench-globem-item-averaged-message-wise · How It Works · Data refreshed daily, snapshot 2026-11-05.