Enginuity - Parts Description F1 (CoT): leaderboard
Enginuity (2026) asks vision-language models to read engineering diagrams from U.S. military repair-parts manuals: in Task 1 the model lists every numbered part callout in a diagram, with part number and description, as JSON, and is scored against the parts table of the manual over 494 figures. This board, with a zero-shot chain-of-thought prompt, is the token F1 of the predicted part descriptions against the manual wording, multiplied by recall so missed parts count against it, in percent; higher is better.
Metric: Penalised token F1 (%) of matched part descriptions against the manual text, multiplied by Recall@all; zero-shot chain-of-thought prompt (think step by step before the JSON); Task 1 of Enginuity: 494 engineering diagrams from U.S. military repair-parts manuals; the model lists every called-out part (item number, part number, description) as JSON from the diagram image alone, scored against the manuals' parts tables; temperature 0, 4,096 output tokens; higher is better. Source: arxiv.org. Saturation forecast: Around July 2028. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 17.9 | #79 |
| 2 | GPT-5.2 Instant | 12.7 | #293 |
| 3 | Qwen 3 VL 32B Instruct | 9 | #585 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=enginuity-parts-description-f1-cot · How It Works · Data refreshed daily, snapshot 2026-11-13.