DocSplit-Poly-Rand: leaderboard
Metric: Packet score: half the clustering score (mean of V-measure and Rand index of the predicted subdocument grouping) plus half the mean Kendall's tau of the page order within each multi-page subdocument, from -0.5 to 1; multi-category packets with every page shuffled; document packets assembled from RVL-CDIP-MP test documents (13 document types); the model reads each page as Textract markdown and returns every subdocument's type, its pages and their order as JSON (temperature 0, 4,096-token output); higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V3.1 | 0.92 | #260 |
| 2 | Claude Haiku 4.5 | 0.91 | #271 |
| 3 | Claude Sonnet 4.5 | 0.91 | #138 |
| 4 | Gemma 3 27B | 0.88 | #596 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=docsplit-poly-rand · How It Works · Data refreshed daily, snapshot 2026-10-11.