SCOPE: Spatially-Constrained Parametric Editing for Text-Guided CAD Models

Text-based CAD editing (e.g., CAD-Editor) has emerged as a promising approach for automating CAD modifications from natural language instructions. However, existing methods lack explicit spatial understanding, limiting their ability to accurately interpret instructions that involve relative positions or geometric constraints. To address this gap, we introduce $\texttt{SCOPE}$, an extension of the locate-then-infill framework that integrates language-guided spatial reasoning into text-guided CAD editing. $\texttt{SCOPE}$ enhances training by synthesizing spatially grounded editing samples and enables the model to learn spatial relations between CAD features (e.g., “drill a hole on the left panel above the rectangle”). Furthermore, we integrate spatial context into both the locate and infill stages, improving target region identification and spatially consistent modifications. Experiments on a public CAD dataset demonstrate that incorporating spatial reasoning significantly improves the accuracy of text-based CAD editing with more precise region localization and control while retaining the efficiency of the original framework.