LLM4CAD-Editor lets you edit designs via natural language with 96% accuracy
No more fragile coordinate-based editing—just describe the change in plain English
Real-world computer-aided design (CAD) workflows are inherently iterative—engineers constantly tweak dimensions, swap operations, or change functions. While large language models can generate CAD programs from scratch, they struggle with editing because they rely on fragile coordinate-based reasoning. To solve this, Yuewan Sun and Zhenghui Sha propose LLM4CAD-Editor, an intent-aware framework that uses a structured domain-specific language (LLM4CAD-DSL) to represent CAD geometry via feature names instead of raw coordinates. This shift transforms editing into natural-language-based operations, allowing the model to understand high-level intents like "increase the radius of the three holes" without needing explicit coordinate references.
The team built a multimodal CAD editing dataset containing 35,139 instruction-program pairs, covering functional-, operation-, and parameter-level edits through DSL-based augmentation and vision-language instruction synthesis. They fine-tuned a 32B-parameter language model to generate DSL edits directly. Results show high parsing accuracy (96.3% for parameter-level edits) and strong intent satisfaction (82% for functional instructions). The model achieved an average Intersection-over-Union (IoU) of 0.935 for parameter edits, 0.871 for operation edits, and 0.708 for functional edits, with corresponding editing distances improving as edits become more granular. Comparative studies show a 1.4x improvement in editing robustness over traditional Python-based CAD scripting.
- LLM4CAD-Editor uses a custom DSL (LLM4CAD-DSL) to replace coordinate-based geometry with feature names, enabling natural-language editing
- A dataset of 35,139 instruction-program pairs was created for fine-tuning a 32B-parameter LLM, covering three editing levels
- Achieves 96.3% parsing accuracy for parameter edits and 82% intent satisfaction on functional instructions, outperforming Python scripting by 1.4x
Why It Matters
LLMs can now handle iterative CAD editing reliably, bridging natural language and professional design workflows without fragile coordinate dependencies.