2026 Qi Pointer CAD (1)

Constructing computer-aided design (CAD) models is labor-intensive but essential for engineering and manufacturing. Recent advances in Large Language Models (LLMs) have inspired LLM-based CAD generation by representing CAD as command sequences. However, existing methods struggle in practical scenarios as command sequence representation lacks support for entity selection, which limits their ability to facilitate complex editing operations like chamfers or fillets, and can introduce topological errors during sketch and extrude operations. To address these limitations, we introduce Pointer-CAD, a novel LLM-based CAD generation framework that incorporates a pointer-based command sequence representation to integrate the geometric information of B-rep models into sequential modeling. Pointer-CAD decomposes CAD model generation into sequential steps, conditioning the generation of each step on both the textual description and the B-rep from prior steps. When a specific geometric entity needs to be selected, the LLM predicts a Pointer that selects the most feature-consistent candidate from the available set, thereby reducing quantization error. We also develop a data annotation pipeline to create expert-level natural language descriptions and compile a dataset of approximately 575K CAD models. Experimental results demonstrate that Pointer-CAD supports the generation of complex geometric structures and significantly mitigates topological inaccuracies caused by quantization error.