ZHENG FANG, Chenbo ZENG, Xiaoyu DONG, Fufang Wen, Yujie Chen, Guanlin Li, LU Zhengping, Yin Yang

CAD agents can edit 3D models using natural-language instructions and test feedback, yet repairing errors while preserving other design requirements remains challenging. Code execution and local test results alone give an incomplete picture of repair quality. We introduce MUSE0.1, a benchmark for evaluating how CAD agents detect and repair design errors. It covers basic parametric parts, complex structures with multiple holes or internal channels, and multi-component models, alongside existing text-to-CAD outputs. Each CAD model is paired with explicit design requirements. The evaluation measures error detection, target localization, repair success, and preservation of the remaining requirements. We check dimensions, positions, and connectivity through direct geometric measurements, and examine structural changes before and after repair in representative cases. Experiments show that passing local checks does not ensure a correct repair: agents may edit the wrong feature or fix the target error while changing structures that should remain intact. MUSE0.1 provides a testbed for developing CAD agents that correct errors while preserving the intended design.

Paper In "CVPR"｜Click Here To Get The Full Paper