2026 Liu CADIR Agentic CAD Representation (1)

Largelanguagemodelshavemadeitpossibletogenerateexe- cutablecomputer-aideddesign(CAD)programsfromnatural- language descriptions or images. However, existing methods represent modeling processes as backend-specific sequential scriptswithimplicitdependenciesorasstaticgeometry,mak- ing it difficult to simultaneously preserve construction his- tory,stabletopologicalreferences,andfeature-leveleditability across different CAD systems. We present CADIR, an agent- friendlyexecutableintermediaterepresentationforCADgen- eration and cross-backend editing. Built on the OCCT geom- etrykernelviaOCP,CADIRprovidesexplicit,compositional modeling operations and fine-grained execution diagnostics. During program execution, CADIR records modeling opera- tions, parameter dependencies, constraints, and topology se- lections in a construction graph. To enable reliable cross- backend reconstruction, we introduce Geometric Signature Matching, which identifies corresponding edges and faces despite parameter changes and backend differences, allow- ing adapters to reconstruct native editable feature histories in FreeCAD,SolidWorks,andFusion360.Buildingonthisrep- resentation,wefurtherproposeaconstruction-graphretrieval method for text and image queries that supports both full-graphandsubgraphretrieval,enablingagentstoleveragecom- plete models and modeling substructures. Extensive experi- ments demonstrate that CADIR achieves higher geometric fi- delityandexecutionreliabilitythanexistingCADrepresenta- tions,thatconstruction-graphretrievalfurtherimprovesmodel generation quality, and that cross-backend editing enables re- liable model reconstruction and post-reconstruction editing acrossmultipleCADenvironments.