Native Boundary Representation Understanding with Large Language Models
1Northwestern Polytechnical University · 2Shanghai Artificial Intelligence Laboratory · 3National University of Singapore · 4Hong Kong University of Science and Technology
BrepLLM is a novel framework that enables large language models (LLMs) to directly understand and reason over native boundary representation (B-rep) data. Unlike previous approaches that rely on intermediate formats such as point clouds, meshes, or CAD command sequences, BrepLLM operates directly on the original geometric and topological structure of CAD models.
It introduces a hierarchical B-rep encoder that captures both geometric and topological information, aligned with language representations through cross-modal learning. The framework follows a two-stage design: cross-modal alignment + multi-stage fine-tuning.
This page showcases BrepLLM's ability to understand and reason about CAD models. Each sample below is a real STEP CAD file processed by BrepLLM.
For each model, BrepLLM answers questions about three aspects:
Click any 3D model below to view the conversation.
If our work is helpful for your research or applications, please cite:
@misc{deng2026brepllmenablinglargelanguage,
title={BrepLLM: Enabling Large Language Models to Understand Boundary Representations},
author={Liyuan Deng and Hao Guo and Yongkang Dai and Yunpeng Bai and Yifan Zhu and Yuanyuan Gao and Huaxi Huang and Yilei Shi},
year={2026},
eprint={2512.16413},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2512.16413},
}