Foundation Models for Automatic CAD Generation

Short summary

Researchers introduce LLMForge, a multi-model text-to-CAD framework that uses foundation models to generate parametric 3D mechanical designs from natural language. The study evaluates seven LLMs on 97 engineering problems using two critique regimes: IterTracer (analytic visual metrics) and IterVision (VLM-based semantic critique via Qwen2.5-VL-72B). Compact instruction-tuned models matched larger systems under analytic scoring, while VLM critique achieved 100% watertight mesh generation on the leading model but struggled with rotationally symmetric geometries.

• •LLMForge framework converts natural language to parametric CAD using 7 foundation models on 97 benchmark problems
• •IterTracer analytic scoring shows compact models matching larger ones at 98.97% mesh success
• •IterVision VLM critique achieves 100% watertight meshes but diverges from analytic scoring on cylindrical geometries

Generated with AI, which can make mistakes.