Khondo (PolySeq): leaderboard
Khondo (2026) document packet splitting on the PolySeq variant of Bangladeshi government form packets (forms of two or more domains concatenated in order), from page images. Packet score from -0.5 to 1: the mean of a clustering score (V-measure and Rand index) and a page-ordering score (Kendall tau-a); higher is better.
Metric: Packet score (-0.5 to 1; mean of the clustering score, the average of V-measure and Rand index, and the ordering score, mean Kendall tau-a over predicted multi-page documents; forms of two or more domains concatenated in order; 210 test packets of 5 to 20 page images of Bangladeshi government forms (Bangla and English, 14 domains); zero-shot order-aware prompt, all page images in one request, JSON output, temperature 1.0, reasoning disabled, invalid outputs retried up to three times). Source: arxiv.org. Saturation forecast: Estimated already saturated. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.6 Plus (Non-reasoning) | 0.92 | #222 (Qwen 3.6 Plus) |
| 2 | Gemini 3.5 Flash (Minimal) | 0.91 | #101 (Gemini 3.5 Flash) |
| 3 | GPT-5.4 (Non-reasoning) | 0.83 | #123 (GPT-5.4) |
| 4 | GLM-4.6V (Non-reasoning) | 0.83 | #701 (GLM-4.6V) |
| 5 | Kimi K2.5 (Non-reasoning) | 0.77 | #256 (Kimi K2.5) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=khondo-polyseq · How It Works · Data refreshed daily, snapshot 2026-11-13.