2026 Li FutureCAD BRep Primitive Grounding

The field of Computer-Aided Design (CAD) generation has made significant progress in recent years, yet a critical paradigm gap limits AI-driven modeling for complex industrial design. This paper presents FutureCAD, a novel framework that leverages large language models (LLMs) and a B-Rep grounding transformer (BRepGround) for high-fidelity CAD generation. Our approach generates executable CadQuery scripts and introduces a text-based query mechanism that allows LLMs to specify geometric selections via natural language, which are then grounded to target primitives by BRepGround. To train our framework, we constructed a new dataset comprising real-world CAD models. We applied supervised fine-tuning (SFT) to establish fundamental CAD generation capabilities, followed by reinforcement learning (RL) to improve generalization. Experimental results demonstrate that FutureCAD achieves state-of-the-art CAD generation performance, showcasing the potential of LLM-driven methods in bridging the gap between parametric modeling and B-Rep synthesis and enhancing the capabilities of CAD systems. Code and dataset are available at https://github.com/JohanStackk/FutureCAD.