Unified Controllable and Faithful Text-to-CAD Generation with LLMs

Researchers propose PR-CAD, a framework that uses large language models to generate and edit CAD models from text descriptions, unifying what were previously separate tasks. The system leverages a new dataset covering the full CAD lifecycle and employs reinforcement learning to understand user intent, estimate parameters, and localize edits in a single agent. PR-CAD achieves state-of-the-art performance in both design creation and refinement while significantly improving CAD modeling efficiency.