Foundation Models for Automatic CAD Generation

arXiv 2026 57.9 method

TLDR

Empirical study of foundation models for text-to-CAD generation, introducing LLMForge with iterative refinement and visual feedback, evaluated on 97 mechanical part designs.

Reasoning

The paper provides a thorough empirical evaluation with multiple models and a curated benchmark, demonstrating strong results. However, the scope is limited to four geometry families and mechanical parts, and the abstract does not detail limitations or comparisons to prior work.

Read-first score

Read-first score 57.9, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 75.

Recency 8%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=benchmark,evaluation,metric,validation

Topical relevance 42%

46.9

Uses existing LLM keyword relevance scores normalized to 0-100. AI for CAD,computer-aided design,neural CAD,generative CAD,parametric CAD,B-Rep,boundary representation,constructive solid geometry,CSG,sketch extrusion,CAD generation,CAD reconstruction,text-to-CAD,image-to-CAD,point cloud to CAD,CAD program

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 9.

Keyword Scores

CAD generation

text-to-CAD

AI for CAD

computer-aided design

generative CAD

parametric CAD

CAD program

neural CAD

B-Rep

boundary representation

constructive solid geometry

CSG

sketch extrusion

CAD reconstruction

image-to-CAD

point cloud to CAD

Deep Analysis

Innovations

LLMForge: a multi-model text-to-CAD framework with JSON-schema validation, analytic feature scoring, mesh synthesis, and multi-round iterative refinement
Two critique regimes: IterTracer (analytic visual metrics via ray-tracing) and IterVision (VLM semantic critic with chain-of-thought visual reasoning)
Unified evaluation pipeline and curated benchmark of 97 engineering design problems across four geometry families
Empirical comparison of seven foundation models revealing that compact instruction-tuned models can match larger systems

Methodology

LLMForge integrates LLMs for text-to-CAD generation with JSON-schema validation, analytic feature scoring, and mesh synthesis, using iterative refinement under two critique types: IterTracer with Phong-shaded ray-tracing and analytic metrics, and IterVision with a VLM (Qwen2.5-VL-72B) performing chain-of-thought visual reasoning. The framework is evaluated on a benchmark of 97 mechanical design problems spanning plates, boxes, flanged cylinders, and L-brackets, using seven foundation models.

Key Results

Under IterTracer, the top four models achieve a tight cluster of mean scores (0.885–0.890) and 98.97% mesh success, showing compact models can rival larger ones. IterVision attains 100% watertight mesh generation for the leading model but exposes difficulties with rotationally symmetric parts like cylinders, where visual and semantic scoring diverge.

Limitations

Systematic difficulty with rotationally symmetric geometries (e.g., cylinders) where visual and semantic scoring diverge
Benchmark limited to four canonical geometry families of mechanical parts
VLM-based critic may not fully align with analytic metrics on certain shapes

Resources

Paper