STEP-LLM: Generating CAD STEP Models from Natural Language with Large Language Models
TLDR
STEP-LLM generates CAD STEP models from natural language using LLMs with RAG and RL, addressing graph-structured B-rep challenges.
Reasoning
The paper introduces a novel method for generating STEP files (B-rep) from text, with a curated dataset and preprocessing techniques. Strengths include addressing a practical manufacturing format and using RAG/RL; weaknesses include limited baseline comparison and no explicit real-world validation.
Read-first score
Read-first score 64.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 85.
Field roles
Rank sensitivity
Stability: volatile; rank range: 4.
Keyword Scores
Deep Analysis
Innovations
- Curated dataset of ~40K STEP-caption pairs
- DFS-based reserialization to linearize cross-references while preserving locality
- Chain-of-thought-style structural annotations for global coherence
- Retrieval-augmented generation (RAG) to ground predictions in relevant examples
- Reinforcement learning with Chamfer Distance-based geometric reward
Methodology
STEP-LLM fine-tunes an LLM on ~40K STEP-caption pairs with preprocessing including DFS-based reserialization and chain-of-thought structural annotations. It integrates retrieval-augmented generation and reinforcement learning with a Chamfer Distance-based geometric reward, and is evaluated against the Text2CAD baseline using geometric fidelity, completeness, renderability, and Chamfer Distance metrics.
Key Results
STEP-LLM achieves consistent gains in geometric fidelity over Text2CAD; RAG improves completeness and renderability, DFS reserialization strengthens overall accuracy, and RL reduces geometric discrepancy, confirmed by metrics and visual comparisons.