Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

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

TL;DR

STEP-LLM addresses the challenge of translating natural language into manufacturable CAD STEP models by introducing a DFS-based reserialization of STEP graphs, CoT-style structural cues, retrieval-augmented fine-tuning, and a geometry-aware reinforcement learning objective. The approach pairs a ~40K caption–STEP dataset with a three-stage pipeline (dataset preprocessing, SFT with RAG, RL alignment) to achieve higher geometric fidelity than a strong Text2CAD baseline, while maintaining competitive renderability. Key contributions include the DFS serialization strategy, RAG grounding, and a Scaled Chamfer Distance–based reward guiding geometry, validated through quantitative metrics and qualitative comparisons. The work demonstrates the feasibility of direct NL-to-STEP generation, offering a path toward democratizing CAD design for manufacturing.

Abstract

Computer-aided design (CAD) is vital to modern manufacturing, yet model creation remains labor-intensive and expertise-heavy. To enable non-experts to translate intuitive design intent into manufacturable artifacts, recent large language models-based text-to-CAD efforts focus on command sequences or script-based formats like CadQuery. However, these formats are kernel-dependent and lack universality for manufacturing. In contrast, the Standard for the Exchange of Product Data (STEP, ISO 10303) file is a widely adopted, neutral boundary representation (B-rep) format directly compatible with manufacturing, but its graph-structured, cross-referenced nature poses unique challenges for auto-regressive LLMs. To address this, we curate a dataset of ~40K STEP-caption pairs and introduce novel preprocessing tailored for the graph-structured format of STEP, including a depth-first search-based reserialization that linearizes cross-references while preserving locality and chain-of-thought(CoT)-style structural annotations that guide global coherence. We integrate retrieval-augmented generation to ground predictions in relevant examples for supervised fine-tuning, and refine generation quality through reinforcement learning with a specific Chamfer Distance-based geometric reward. Experiments demonstrate consistent gains of our STEP-LLM in geometric fidelity over the Text2CAD baseline, with improvements arising from multiple stages of our framework: the RAG module substantially enhances completeness and renderability, the DFS-based reserialization strengthens overall accuracy, and the RL further reduces geometric discrepancy. Both metrics and visual comparisons confirm that STEP-LLM generates shapes with higher fidelity than Text2CAD. These results show the feasibility of LLM-driven STEP model generation from natural language, showing its potential to democratize CAD design for manufacturing.

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

TL;DR

STEP-LLM addresses the challenge of translating natural language into manufacturable CAD STEP models by introducing a DFS-based reserialization of STEP graphs, CoT-style structural cues, retrieval-augmented fine-tuning, and a geometry-aware reinforcement learning objective. The approach pairs a ~40K caption–STEP dataset with a three-stage pipeline (dataset preprocessing, SFT with RAG, RL alignment) to achieve higher geometric fidelity than a strong Text2CAD baseline, while maintaining competitive renderability. Key contributions include the DFS serialization strategy, RAG grounding, and a Scaled Chamfer Distance–based reward guiding geometry, validated through quantitative metrics and qualitative comparisons. The work demonstrates the feasibility of direct NL-to-STEP generation, offering a path toward democratizing CAD design for manufacturing.

Abstract

Computer-aided design (CAD) is vital to modern manufacturing, yet model creation remains labor-intensive and expertise-heavy. To enable non-experts to translate intuitive design intent into manufacturable artifacts, recent large language models-based text-to-CAD efforts focus on command sequences or script-based formats like CadQuery. However, these formats are kernel-dependent and lack universality for manufacturing. In contrast, the Standard for the Exchange of Product Data (STEP, ISO 10303) file is a widely adopted, neutral boundary representation (B-rep) format directly compatible with manufacturing, but its graph-structured, cross-referenced nature poses unique challenges for auto-regressive LLMs. To address this, we curate a dataset of ~40K STEP-caption pairs and introduce novel preprocessing tailored for the graph-structured format of STEP, including a depth-first search-based reserialization that linearizes cross-references while preserving locality and chain-of-thought(CoT)-style structural annotations that guide global coherence. We integrate retrieval-augmented generation to ground predictions in relevant examples for supervised fine-tuning, and refine generation quality through reinforcement learning with a specific Chamfer Distance-based geometric reward. Experiments demonstrate consistent gains of our STEP-LLM in geometric fidelity over the Text2CAD baseline, with improvements arising from multiple stages of our framework: the RAG module substantially enhances completeness and renderability, the DFS-based reserialization strengthens overall accuracy, and the RL further reduces geometric discrepancy. Both metrics and visual comparisons confirm that STEP-LLM generates shapes with higher fidelity than Text2CAD. These results show the feasibility of LLM-driven STEP model generation from natural language, showing its potential to democratize CAD design for manufacturing.
Paper Structure (13 sections, 3 equations, 4 figures, 4 tables)

This paper contains 13 sections, 3 equations, 4 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Background and Related Work
  3. Large Language Models for CAD Generation
  4. Learning-based B-rep Generation and STEP File Analysis
  5. The STEP-LLM Framework
  6. Dataset Construction and Preprocessing
  7. Supervised Fine-Tuning with RAG
  8. Reinforcement Learning Alignment
  9. Experiments and Evaluation
  10. Experimental Setup
  11. Advantages of STEP-LLM over Baseline
  12. Ablation Studies on Module Importance
  13. Effectiveness of RL

Figures (4)

  • Figure 1: The typical workflow of CAD from design to manufacturing.
  • Figure 2: Framework of STEP-LLM. The top panel illustrates data preprocessing: raw STEP files are rendered into nine views and captioned with GPT-4o (left), and their internal DAG structures (the numbers represent the identifier of the entity in a STEP file) are reserialized into locality-preserving trees (right). These captions and DFS-style STEP files together form the paired STEP-caption dataset for SFT (bottom-left). A specific geometric reward based on scaled Chamfer Distance is designed for further RL training (bottom-right).
  • Figure 3: Entity count distribution of generated STEP files vs. ground truth.
  • Figure 4: The visual comparison between Text2CAD, STEP-LLM-noRAG and STEP-LLM.