LO
CapabilityCognitive automation2025-02-28
GPT-4 wrote CAD scripts that compiled 96.5% of the time on CADPrompt, 200 design prompts, while the researchers note a compiled script may not meet the specified requirements
Mechanical engineeroccupation page →
Event date / reported
2025-02-28
Evidence stage
CapabilityA demo, benchmark or paper shows the task can be done. Updates what the technology can do — not what employers will do.
Tasks this bears on
CAD modelling and generative design
Building 3D models and drawings of parts and assemblies, and using generative tools to propose shapes within set constraints.
Being augmented✓ Evidence-backed
Where this applies
An academic benchmark of 200 natural-language prompts paired with expert-annotated CAD scripting code. It reports that GPT-4 produced code that compiled 96.5% of the time, compared with 85% for Gemini and 73.5% for CodeLlama, and the authors note that the 3D object a generated script renders may not meet the specified requirements, which is why they propose a verification step. It measures whether scripts run and resemble the target, not whether parts are manufacturable or correct.
What this means
Models can write CAD code that runs almost every time. Whether the part it draws is the part that was asked for is a separate question, and the authors built a checking step because the answer is often no.
What it does not yet show
A benchmark of short prompts; it does not measure real product design or manufacturability.
What you can check
Open arXiv 2410.05340 and find "may not meet the specified requirements" in the abstract.
Does it change the assessment?
No — and this stage does not move it either. A "Capability" record is real evidence, but it does not upgrade a task judgement on its own. The 1 linked judgement above stand where they were.
Source
Alrashedy et al. (Georgia Institute of Technology) — Generating CAD Code with Vision-Language Models for 3D Designs (CADCodeVerify / CADPrompt), arXiv 2410.05340 (v2, 28 Feb 2025; ICLR 2025) · verified 2026-09-29 · Claude (VOLO agent) · interpreted 2026-09-29 · Claude (VOLO agent)
Primary source — published by the party that did this, or the authority of record. No co-signature needed.