CADBench: A Multimodal Benchmark for AI-Assisted CAD Program Generation

arXiv 2026 81.9 method

TLDR

CADBench is a unified multimodal benchmark for evaluating AI-assisted CAD program generation from images and 3D data, covering 18,000 samples, five modalities, and six metrics.

Reasoning

The paper's strength lies in providing a comprehensive, standardized benchmark that addresses fragmentation in existing evaluations, with diverse modalities and metrics. Weaknesses include reliance on existing models and no novel method; the abstract does not discuss limitations like potential dataset biases or computational costs.

Read-first score

Read-first score 81.9, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 105.

Methodology quality 25%

100

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

Recency 8%

100

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

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=True; dataset=False; markers=code,dataset,github

Topical relevance 42%

65.6

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

Field roles

FrontierMethodology anchorReproducibility anchor

Rank sensitivity

Stability: stable; rank range: 0.

Keyword Scores

CAD program

AI for CAD

computer-aided design

CAD generation

CAD reconstruction

generative CAD

parametric CAD

image-to-CAD

neural CAD

B-Rep

boundary representation

point cloud to CAD

constructive solid geometry

CSG

sketch extrusion

text-to-CAD

Deep Analysis

Innovations

Unified multimodal benchmark CADBench with 18,000 samples from six dataset families, five input modalities, and six metrics for CAD program generation.
Stratification by B-rep face count and diversity sampling to enable controlled analysis across geometric complexity and object variation.
Large-scale evaluation of 11 CAD-specialized and general-purpose VLMs, generating 1.4M CAD programs and identifying three recurring failure modes.

Methodology

CADBench aggregates 18,000 evaluation samples from six benchmark families (DeepCAD, Fusion 360, ABC, MCB, Objaverse) across five input modalities (clean/noisy meshes, single-view, photorealistic, multi-view renders) and six metrics (geometric fidelity, executability, program compactness). STEP-based families are stratified by B-rep face count and all families are diversity-sampled. Eleven models are benchmarked by generating over 1.4 million CAD programs.

Key Results

Specialized mesh-to-CAD models substantially outperform code-generating VLMs under idealized inputs; three failure modes emerge: reconstruction quality degrades with geometric complexity, CAD-specialized models are brittle under modality shift, and model rankings change across metrics.

Resources

Paper Code