Reverse generation of 3D CAD models based on CAD sequences plays a significant role in improving product design efficiency. However, widely used neural network-based mapping and generation methods often suffer from high computational complexity and insufficient reconstruction accuracy, limiting their practical application efficiency. To address these challenges, this chapter proposes an improved Transformer-based method for 3D CAD reverse generation. First, to tackle the difficulty of achieving high-precision reconstruction due to the complexity of CAD sequence descriptions, an Efficient Local Attention (ELA) module was developed. By simultaneously focusing on multilevel global and local features, the ELA module enhances the understanding and processing of complex 3D scenarios, effectively reducing detail loss during the reconstruction of detail-rich models. Second, to address the high computational complexity of neural networks, a Sparse Transformer (ST) module was designed. Through sparsifying the computational structure, this module significantly reduces computational redundancy and improves inference efficiency, enabling the rapid extraction of critical geometric features during the reverse generation of 3D models. Experimental results demonstrate the effectiveness of the proposed algorithm.