Preprint · cad-bench/v0.5 · sweep 2026-04-12open · MIT
CAD·Benchv0.5
TASKS · pilot subset

28 prompts across 20 categories

Each task carries a verbatim natural-language prompt, a canonical reference STEP (sha-256 in the listing), numerical ground-truth quantities (volume, surface area, Euler χ, genus, named features), and a difficulty class 1-5. Click a task to see the held-out reference, candidate output viewers, and metric scores per agent.

Geometric Primitivesw = 0.10 · 2/24 shown

Boolean Robustnessw = 0.20 · 2/18 shown

BREP Fidelityw = 0.40 · 1/14 shown

Free-form Surfacesw = 0.30 · 1/12 shown

Parametric Mechanical Partsw = 0.25 · 3/30 shown

Assembly & Matingw = 0.22 · 2/16 shown

Standards Compliancew = 0.15 · 1/18 shown

Sheet-Metal Bodiesw = 0.13 · 1/14 shown

Sealing-Groove Designw = 0.10 · 1/10 shown

Kinematic Mechanismsw = 0.15 · 1/12 shown

DFM · 3-Axis CNCw = 0.30 · 1/18 shown

DFM · Injection Mouldw = 0.30 · 1/20 shown

DFM · FDM 3D Printw = 0.20 · 1/14 shown

CAM Toolpath Validityw = 0.20 · 1/12 shown

Constraint Solving & Editabilityw = 0.18 · 2/18 shown

Reverse Engineeringw = 0.20 · 2/22 shown

2-D Sketch Constraintsw = 0.10 · 1/20 shown

Functional Intent · FEA-Gatedw = 0.25 · 2/16 shown

Paraphrase Robustnessw = 0.15 · 1/20 shown

Confidence Calibrationw = 0.12 · 1/15 shown