Foundation Models for Automatic CAD Generation
06:00 · July 8, 2026 · arXiv cs.AI RSS
Recent advances in Large Language Models (LLMs) and Vision-Language Models (VLMs) enable the automatic generation of parametric 3D designs from natural-language specifications. This chapter presents an empirical study of foundation models for automatic Computer-Aided Design (CAD) generation of mechanical parts, using a unified evaluation pipeline and a curated benchmark of 97 engineering design problems. We introduce LLMForge, a multi-model text-to-CAD framework integrating JSON-schema validation, analytic feature scoring, mesh synthesis, and multi-round iterative refinement, studied under two critique regimes. IterTracer uses a Phong-shaded ray-trace renderer with analytic visual metrics (silhouette IoU, hole visibility, edge clearance, aspect-ratio conformance) for lightweight geometry-aware feedback across rounds. IterVision replaces the analytic scorer with a VLM semantic critic (Qwen2.5-VL-72B) that evaluates rendered views via chain-of-thought visual reasoning, assessing spatial coherence and design intent. On a benchmark spanning four canonical geometry families (plates with holes and bolt circles, multi-feature boxes, flanged cylinders, and L-brackets), we evaluate seven foundation models: DeepSeek-V3.2, Qwen3-235B-A22B, Llama-3.3-70B, Gemma-3-27B, GLM-4.5, MiniMax-M2.1, and INTELLECT. Under IterTracer, the four highest-ranked models form a tight cluster (overall mean in [0.885, 0.890]) with 98.97% mesh success, showing that compact instruction-tuned models can match substantially larger systems. VLM-based critique in IterVision yields 100% watertight mesh generation on the leading model while surfacing systematic difficulty on rotationally symmetric geometries such as cylinders, where visual and semantic scoring diverge most. We discuss benchmark design, failure modes, CAD-oriented prompting, and implications for industrial workflows and scalable automated mechanical design.
Summary
Recent advances in large language models and vision-language models have made it feasible to translate natural-language specifications into parametric three-dimensional CAD models of mechanical parts. This empirical study examines that capability through a unified evaluation pipeline and a benchmark of 97 design problems drawn from four canonical geometry families: plates with holes and bolt circles, multi-feature boxes, flanged cylinders, and L-brackets.
The authors present LLMForge, a multi-model text-to-CAD framework that combines JSON-schema validation, analytic feature scoring, mesh synthesis, and up to four rounds of iterative refinement. Two critique regimes are compared. IterTracer supplies lightweight geometry-aware feedback by rendering each candidate with a Phong-shaded ray tracer and scoring it on silhouette IoU, hole visibility, edge clearance, and aspect-ratio conformance. IterVision replaces the analytic scorer with a chain-of-thought VLM critic (Qwen2.5-VL-72B) that assesses spatial coherence and design intent from rendered views.
Seven foundation models were evaluated: DeepSeek-V3.2, Qwen3-235B-A22B, Llama-3.3-70B, Gemma-3-27B, GLM-4.5, MiniMax-M2.1, and INTELLECT. Under the IterTracer regime the four strongest models produced overall scores tightly clustered between 0.885 and 0.890, each achieving a 98.97 % mesh success rate. The IterVision regime raised mesh success on the leading model to 100 % while revealing systematic difficulties with rotationally symmetric parts such as cylinders, where visual and semantic scores diverged most noticeably. The study also characterises failure modes, prompting strategies, and implications for automated industrial design workflows.
Why it matters
This research is highly relevant for the Dutch AI market, particularly for its strong high-tech manufacturing and engineering sectors. The introduction of automated, iterative text-to-CAD generation offers actionable insights for researchers and enterprises looking to optimize industrial workflows using state-of-the-art foundation models.
