Foundation Models Automate Parametric 3D CAD Design Generation

A recent empirical study investigates the capability of foundation models to automatically generate parametric 3D Computer-Aided Design (CAD) models directly from natural language specifications. The research introduces LLMForge, a comprehensive text-to-CAD framework designed for evaluating and refining mechanical part designs. LLMForge incorporates several advanced features, including JSON-schema validation for structural integrity, analytic feature scoring, mesh synthesis, and a multi-round iterative refinement process. This refinement can operate under two critique regimes: IterTracer, which uses a ray-trace renderer with analytic visual metrics, and IterVision, which employs a Vision-Language Model (VLM) for semantic visual reasoning. The study benchmarked seven foundation models across four geometry families, revealing that compact instruction-tuned models can achieve performance comparable to much larger systems, with high mesh success rates. VLM-based critique further improved watertight mesh generation, though it highlighted systematic difficulties with rotationally symmetric geometries. This work has significant implications for industrial design workflows and scalable automated mechanical design.