On July 6, 2026, researchers J de Curtò, Victoria Guillén, and I. de Zarzà presented their findings on foundation models for automatic Computer-Aided Design (CAD) generation. The study illustrates how recent advancements in Large Language Models (LLMs) and Vision-Language Models (VLMs) facilitate the automatic creation of parametric 3D designs from natural language descriptions.
Empirical Study of CAD Generation Models
The chapter details an empirical study evaluating a unified pipeline that integrates a curated benchmark of 97 engineering design problems. The framework, known as LLMForge, combines JSON-schema validation, analytic feature scoring, mesh synthesis, and iterative refinement in its approach to CAD generation.
Two critique regimes were employed in the study: IterTracer and IterVision. IterTracer utilizes a Phong-shaded ray-trace renderer for lightweight geometry-aware feedback, assessing metrics such as silhouette IoU and edge clearance. Conversely, IterVision employs a VLM semantic critic, Qwen2.5-VL-72B, to evaluate rendered views through chain-of-thought visual reasoning.
Benchmarking Foundation Models
The study evaluated seven foundation models, including DeepSeek-V3.2, Qwen3-235B-A22B, and Llama-3.3-70B. Under the IterTracer regime, the top four models achieved a mean score between 0.885 and 0.890, with an impressive 98.97% mesh success rate. This suggests that compact instruction-tuned models can compete with much larger systems.
