Researchers unveiled PR-CAD, a progressive refinement system that merges text‑to‑CAD generation and editing into a single workflow. By leveraging large language models, the framework accepts natural language prompts and produces editable CAD geometry without separate tools. The approach targets the long‑standing bottleneck of manual modeling, promising faster, more accessible design creation, and supports iterative refinement through user feedback loops.

To train the agent, the team compiled a high‑fidelity interaction dataset covering the full CAD lifecycle, including multiple representations and both qualitative and quantitative descriptions. The dataset defines edit operation types and generates human‑like dialogue. Coupled with a CAD representation tuned for LLMs, a reinforcement‑learning‑enhanced reasoning module learns intent, parameter estimation, and precise edit localization in one pass, and enables cross‑modal reasoning between sketches and parameters.

Extensive experiments on public benchmarks show PR-CAD achieving state-of-the-art controllability and faithfulness for both initial generation and subsequent refinement. Users reported markedly shorter modeling cycles, confirming the system’s efficiency gains. The results demonstrate that unifying generation and editing not only improves accuracy but also streamlines the designer’s workflow, establishing a new baseline for text‑driven CAD tools, and reduces error propagation.