Comp-PointLLM: An Explainable Multimodal Large Language Model for 3D Object Comparison

Abstract Object comparison, a core cognitive ability for human understanding and decision-making, is essential in complex 3D environments, such as enabling intelligent robotic systems to autonomously select appropriate tools and plan sequences of object manipulations, supporting Computer-Aided Design (CAD) in assessing modifications against previous models, and ensuring quality control by verifying whether manufactured products match their original designs. However, there is a lack of research on directly comparing and explaining complex 3D objects in various formats, such as mesh, voxel, and point cloud. To address this gap, we propose Comp-PointLLM, a novel Multimodal Large Language Model (MLLM) framework for explainable comparison of 3D point cloud objects. Comp-PointLLM enhances 3D object understanding through a hybrid architecture that integrates 3D geometric features and 2D visual features. Furthermore, we propose an automated data generation pipeline to construct comparison-captioning and question-answering datasets based on LLMs Experimental results demonstrate that Comp-PointLLM significantly outperforms baseline models across diverse object categories and comparison criteria, and exhibits strong zero-shot generalization to unseen categories. Ablation studies confirm that the hybrid architecture, two-stage training, and integrated data strategy all contribute to performance gains. Comp-PointLLM lays a solid foundation for comparing real-world 3D objects in product design, engineering, and intelligent robotic systems, paving the way for more advanced AI applications.