Natural‑language Computer‑Aided Design (CAD) code generation aims to translate design intent into executable, editable parametric programs. With the rise of large language models (LLMs), this goal becomes increasingly practical, yet a usable system must preserve the construction process behind the rendered geometry. Existing benchmarks and methods mainly evaluate how closely the generated CAD model matches a reference geometry, often using Intersection over Union (IoU) or similar metrics. Such metrics can overlook errors in part decomposition, construction hierarchy, Boolean operations, sketch structure, and geometric relations.
To address this gap we introduce the CIT‑CAD framework. It first infers a Constraint Intent Tree (CIT) from the natural‑language description, encoding the intended entities, hierarchy, operations, and constraints. The tree serves two purposes: it guides CAD code generation and defines the set of constraints that should hold for verification. After generating the program, the framework extracts the actual constraints, compares them with the expected constraints from the CIT, and uses any mismatches to localize and repair design violations.
Experiments demonstrate that CIT‑CAD improves CAD generation performance, with especially large gains on complex multi‑entity designs. By turning design intent into an explicit, checkable object, this work represents the first step toward moving text‑to‑CAD generation beyond rendered‑geometry matching toward construction‑aware synthesis, verification, and repair.
Review