Megadose AI progress, ranked and analyzed.

CADWorld: Computer-Use Benchmark for Long-Horizon Computer-Aided Design

· HF Daily Papers ·
The best tested agent completed only 17.5% of CADWorld tasks, versus 87.0% for the expert reference.

CADWorld is a 200-task FreeCAD benchmark meant to test long-horizon CAD work through screenshots and GUI actions. The tasks span 11 workflow categories, including sketching, part modeling, assembly, CAM, FEM, mesh processing, and technical drawing. Success is checked against saved FreeCAD artifacts and related outputs, not just visible UI progress. The paper says weaker agents often fail before making a valid artifact, while stronger ones fail on geometry, structure, and construction requirements. HF Daily Papers' note

score 5

Categories: Research