FormStruct-Bench - Path Accuracy: leaderboard
FormStruct-Bench (2026) evaluates structure recognition of table-form documents, generated from 70 annotated real-world templates in seven languages, by asking a system to output the page as a hierarchical JSON record. TSR-path is the share of gold leaf field paths whose full path and value the output reproduces exactly, macro-averaged over the 1,100 human-reviewed test pages in percent; invalid outputs score 0 and higher is better.
Metric: TSR-path (%; share of ground-truth leaf field-path and value pairs reproduced exactly, a strict leaf recall; macro-averaged over the 1,100 human-reviewed pages of the template-disjoint test split of table-form documents in seven languages; the model outputs the page as FormStruct-Bench JSON (deterministic decoding where available; pipelines as released), and missing, unparsable or schema-invalid outputs score 0). Source: arxiv.org. Saturation forecast: Around 2032. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.6 35B A3B | 17.91 | #461 |
| 2 | GPT-5.5 | 17.15 | #48 |
| 3 | Seed 2.1 Pro | 15.9 | #133 |
| 4 | Gemini 3.5 Flash | 14.55 | #98 |
| 5 | Claude Sonnet 5 | 13.77 | #108 |
| 6 | Kimi K2.6 | 12.96 | #173 |
| 7 | Qwen 3.5 9B | 12.58 | #803 |
| 8 | Qwen 3.7 Plus | 9.75 | #203 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=formstruct-bench-path-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-12.