Foundation Models for Automatic CAD Generation

Abstract: Recent advances in LLMs and Vision-LLMs (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.

• The paper introduces LLMForge, a multi-stage pipeline that converts natural language into structured CAD models using iterative analytic and semantic feedback.
• It employs both deterministic metrics (e.g., silhouette IoU, mesh soundness) and VLM-based semantic critique to refine design outputs.
• Empirical results show that top-tier foundation models achieve near-saturated performance on benchmark canonical geometries with high mesh integrity.

Foundation Models for Automatic CAD Generation: A Technical Analysis

Introduction

"Foundation Models for Automatic CAD Generation" (2607.05573) systematically investigates the application of state-of-the-art LLMs and VLMs for the generation of parametric 3D mechanical designs from natural language specifications. The work introduces LLMForge—a modular, multi-model text-to-CAD pipeline—and conducts a detailed empirical study across two critique regimes, analytic and semantic. The evaluation leverages a curated benchmark of 97 engineering design problems covering four canonical geometry families. This essay presents a technical summary and analysis of the framework, evaluation methodology, experimental findings, failure modes, and implications for the automated design of engineering components.

Framework and Evaluation Pipeline

The LLMForge framework operationalizes automatic CAD generation as a multi-stage process:

1. Parsing natural language part descriptions into structured JSON schemas encoding geometry and feature information,
2. Converting JSON specifications into triangulated watertight meshes using deterministic CAD libraries (Trimesh, Shapely),
3. Iterative refinement of candidate designs through integration of feedback from analytic and/or VLM-based critics over multiple rounds.

Two critique regimes are defined:

• IterTracer: Employs deterministic, differentiable visual metrics (silhouette IoU, hole visibility, edge clearance, aspect-ratio conformance, cross-sectional checks) computed from Phong-shaded rendered views. This approach delivers rapid, reproducible, geometry-aware feedback.
• IterVision: Augments analytic feedback with a VLM-based semantic critic (Qwen2.5-VL-72B), which processes rendered multi-view images and JSON summaries through chain-of-thought reasoning, returning a "semantic match" score and actionable critique.

Problem specifications and model outputs are strictly structured to ensure schema compliance and deterministic downstream meshing. The framework scores solutions over four axes (validation, mesh soundness, feature adherence, and visual fidelity) in IterTracer, and adds a fifth semantic dimension in IterVision.

Benchmark and Experimental Design

The benchmark comprises 97 engineering problems stratified across rectangular plates with holes, multi-feature boxes, flanged cylinders, and L-brackets. Each problem is annotated with ground-truth feature specifications to enable analytic scoring. Seven competitive foundation models are evaluated: DeepSeek-V3.2, Qwen3-235B-A22B, Llama-3.3-70B, Gemma-3-27B, GLM-4.5, MiniMax-M2.1, and INTELLECT, under constant backend/temperature conditions to ensure evaluative parity.

Each model iterates up to four rounds per problem, with structured feedback injected via system prompts. Early exit occurs if a sufficiently high composite score is achieved; otherwise, the best round output is retained.

Results and Comparative Analysis

Under the IterTracer regime, a convergence in performance is observed among four top-tier instruction-tuned models (DeepSeek-V3.2, Qwen3-235B-A22B, Llama-3.3-70B, Gemma-3-27B), all achieving mean overall scores in $[0.885,\,0.890]$ and mesh success rates at 98.97%. The minor variance among these models signifies saturation on this class of canonical geometries and metrics. The lower-tier models (GLM-4.5, MiniMax-M2.1, INTELLECT) exhibit greater failure rates—especially false schema outputs and malformed meshes—correlating with reduced scores and increased variance.

IterVision's introduction of VLM-based semantic critique increases assessment stringency, reflected in a roughly 0.04 reduction in overall score for top models, but yields improved mesh integrity. Notably, Gemma-3-27B demonstrates a perfect 100% watertight mesh success rate. The VLM axis captures orthogonal failure modes overlooked by analytic metrics, surfacing discrepancies in semantic alignment versus geometric validity, particularly for complex and rotationally symmetric geometries (e.g., cylinders).

The per-round analysis shows that analytic feedback yields a sharp initial correction (most top-model improvements are in round 1), while semantic feedback via VLM sustains beneficial refinement through later rounds, supporting a higher "refinement depth."

Bold claims include the finding that compact instruction-tuned models can approach or match the reliability of much larger LLMs for structured CAD generation, without model-specific fine-tuning or retrieval augmentation.

Failure Modes and Limitations

The work presents an incisive taxonomy of failure cases:

• Saturated analytic metrics make further distinction among top models difficult; the addition of VLM axes restores evaluative headroom.
• Schema non-conformance and geometric failures are the primary failure points for weaker models, often leading to non-meshable outputs even if they appear plausible upon rendering.
• Visual–geometric decoupling is evident where VLM scores are favorable despite analytic failures, spotlighting the necessity for composite, multi-modal evaluation.

Cylindrical geometries are systematically more difficult across models and pipelines—attributed both to the paucity of visual features and the challenge for VLMs in assessing rotational symmetry and feature correspondence.

The current benchmark is intentionally limited to single-part problems with standard features, which excludes freeform surfaces, assemblies, and functional constraints.

Broader Implications and Future Directions

The results indicate the viability of using top-tier foundation models for automatic generation of dimensionally and topologically valid parametric CAD under practical conditions (natural language input, multi-round feedback, moderate temperature). The iterative, feedback-driven protocol is well-suited for scalable deployment in industrial PDM workflows, with semantic VLM critique enabling selective escalation for complex, ambiguous, or critical parts.

Potential extensions encompass:

• Expansion to complex assemblies, freeform surfacing, and incorporation of manufacturing constraints (e.g., GD&T, material selection),
• Fine-tuning LLMs/VLMs on structured CAD-centric feedback to bridge residual model gaps,
• Replacement or augmentation of rendering-based critique with simulation-in-the-loop protocols (e.g., FEA-based validation),
• Integration into end-to-end applied AI design pipelines with human-in-the-loop verification.

Conclusion

"Foundation Models for Automatic CAD Generation" provides a rigorous empirical foundation for the application of instruction-tuned LLMs and VLMs in automatic CAD workflow automation (2607.05573). Analytic and semantic critique regimes reveal both the strengths and residual limitations of current models for structured parametric design. The work substantiates that, on benchmark canonical geometries, top-tier foundation models achieve near-saturation in analytic correctness and approach full geometric validity under multi-round feedback. Semantic (VLM-based) axes surface additional quality headroom and reveal orthogonal semantic-alignment challenges missed by geometry-only scoring. Future research should address open challenges in assembly complexity, semantic geometric understanding, and integration with simulation-oriented feedback essential for robust, industrial-grade CAD automation.