LLMStructBench: leaderboard
Metric: Composite score (0-1): the mean of micro-averaged key-value F1 (key component weighted 0.25, values with graded partial credit) and document correctness, on 995 email-to-JSON extraction tests (five scenarios of 199 manually verified cases, flat to nested schemas), under the P strategy: schema and an example object in the prompt, no API JSON-format parameter; higher is better. Source: arxiv.org. Saturation forecast: Around January 2027. 23 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o | 0.74 | #333 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=llmstructbench · How It Works · Data refreshed daily, snapshot 2026-10-11.