Parametric CAD Bench measures text and image CAD design accuracy
Parametric CAD Bench v3: Evaluating Agentic Parametric CAD Authoring from Text and Engineering Drawings
Artificial IntelligenceComputer Vision and Pattern RecognitionGraphicsMachine Learning
Summary
Creating a CAD script alone isn't enough to prove a design is correct; you must also check the resulting shapes, sizes, and modeling steps. The paper introduces Parametric CAD Bench v3, a set of 100 tasks that test software agents on turning text descriptions and engineering drawings into accurate parametric CAD models. The authors measure success by how closely the final models match intended geometry and dimensions, not just by simple pass/fail. Their study includes 13 systems, finding the top system can reconstruct designs from images with over 80% accuracy. They share the tasks and results for others to try.
What this means in practice
- •For mechanical design teams: Test CAD automation tools by checking if generated models meet exact drawing specifications and dimensions from text or images.
- •For cad software developers: Evaluate and improve parametric CAD model generation from engineering drawings and written requirements using a standardized benchmark.
Authors
Mei Chen, Yu Xia, Haipeng Mao
Abstract
Producing an executable CAD script is not sufficient to establish that an agent has satisfied an engineering design request. Evaluation must also consider the resulting geometry, specified dimensions, and native CAD modeling history. We present Parametric CAD Bench v3, a 100-task benchmark comprising 30 text-to-CAD creation tasks, 30 text-to-CAD create-and-edit tasks, and 40 tasks that reconstruct mechanical parts from engineering drawings. Agents operate in a pinned FreeCAD environment and submit native CAD artifacts to a separate programmatic verifier. The primary metric is mean continuous task reward, combining geometry similarity and specification consistency rather than binary task success. A September 24, 2026 snapshot covers 13 model-agent systems; Claude Code with Claude Opus 5.5 achieves the highest overall point estimate, 61.03%, and an image-to-CAD mean reward of 84.66%. These measurements describe complete systems, not isolated model capabilities. We release versioned tasks, evaluation configurations, and inspectable result artifacts.