FormStruct-Bench - Schema Similarity: 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. Schema-nTED is one minus the normalized tree edit distance between the predicted and gold schema trees (field names, containers and hierarchy, values ignored), macro-averaged over the 1,100 human-reviewed test pages in percent; invalid outputs score 0 and higher is better.
Metric: Schema-nTED (%; 1 minus the normalized tree edit distance between predicted and ground-truth schema trees that keep field names, container types and hierarchy but drop values; 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 2029. 14 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 54.15 | #48 |
| 2 | Seed 2.1 Pro | 52.81 | #133 |
| 3 | Claude Sonnet 5 | 52.22 | #108 |
| 4 | Gemini 3.5 Flash | 48.59 | #98 |
| 5 | Qwen 3.6 35B A3B | 48.46 | #461 |
| 6 | Kimi K2.6 | 48.2 | #173 |
| 7 | Qwen 3.5 9B | 45.58 | #803 |
| 8 | Qwen 3.7 Plus | 34.01 | #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-schema-similarity · How It Works · Data refreshed daily, snapshot 2026-11-12.