STEP-LLM: Generating CAD STEP Models from Natural Language with Large Language Models

STEP-LLM bridges large language models with the STEP (ISO 10303) boundary-representation CAD standard, enabling direct generation of manufacturable CAD models from natural language. The framework curates ~40K STEP-caption pairs, introduces a depth-first-search-based reserialization that linearizes cross-references while preserving locality, grounds generation with retrieval-augmented fine-tuning, and refines quality via reinforcement learning with a Chamfer-distance geometric reward.

Authors: Xiangyu Shi, Junyang Ding, Xu Zhao, Sinong Zhan, Payal Mohapatra, Daniel Quispe, Kojo Welbeck, Jian Cao, Wei Chen, Ping Guo, Qi Zhu

@inproceedings{shi2026stepllm, title={STEP-LLM: Generating CAD STEP Models from Natural Language with Large Language Models}, author={Shi, Xiangyu and Ding, Junyang and Zhao, Xu and Zhan, Sinong and Mohapatra, Payal and Quispe, Daniel and Welbeck, Kojo and Cao, Jian and Chen, Wei and Guo, Ping and Zhu, Qi}, booktitle={Design, Automation and Test in Europe Conference (DATE)}, year={2026}, url={https://arxiv.org/abs/2601.12641} }