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

CADBench has been released as a comprehensive benchmark for evaluating AI-assisted CAD program generation, encompassing 18,000 samples across six benchmark families and five input modalities, including various mesh and render types. It assesses eleven CAD-specialized and general-purpose vision-language models on metrics such as geometric fidelity and executability, revealing that specialized models significantly outperform general code-generating VLMs under ideal conditions, though they struggle with geometric complexity and modality shifts. This benchmark serves as a critical tool for practitioners to measure advancements in editable 3D reconstruction and multimodal CAD understanding, with implications for improving model robustness and performance in practical applications.