Research map / Large language models / Language, safety and interpretability
CAD-Llama: Leveraging Large Language Models for Computer-Aided Design Parametric 3D Model Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Adapts LLaMA3-8B to generate parametric CAD command sequences by converting DeepCAD data into hierarchically annotated Python-like code (SPCC), followed by adaptive pretraining and LoRA instruction tuning.
Problem
Generating parametric CAD construction sequences with LLMs is hard because LLMs see little parametric CAD data in pretraining and CAD sequences lack semantic descriptions, limiting prior methods to simple models and weak generalization.
Approach
Convert CAD command sequences into a Python-like code format (SPCC) enriched with hierarchical VLM-generated descriptions (component and global), then pretrain an LLM on similarity-grouped SPCC contexts and instruction-tune it with LoRA on CAD-centric tasks.
Key results
- Text-to-CAD: CAD-Llama-INS reaches ACCT 84.72 vs 70.36 for CAD-Translator and 69.91 for Text2CAD.
- Unconditional generation: SR 99.90 vs 79.11 for HNC-CAD, with MMD 0.96 and JSD 0.66 (best among compared methods).
- CAD-related tasks: average score 63.58% vs 47.88% for GPT-4.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CAD-Llama instead of HNC-CAD | Unconditional CAD generation · Unconditional generation | SR: 79.11 -> 99.90 | Helped |
| Use CAD-Llama instead of HNC-CAD | Unconditional CAD generation · Unconditional generation | COV: 80.46 -> 79.93 | Hurt |
| Use CAD-Llama-INS instead of CAD-Translator | Text-to-CAD generation · Text-to-CAD | ACCT: 70.36 -> 84.72 | Helped |
| Use CAD-Llama-INS instead of GPT-4 | CAD captioning · CAD captioning | BLEU@1: 27.34 -> 43.44 | Helped |
| Use CAD-Llama-INS instead of GPT-4 | CAD model deletion · Deletion | EM: 66.20 -> 81.93 | Helped |
| Use SPCC rather than single-description CAD code (ablation) | Text-to-CAD generation · Text-to-CAD | ACCcmd: 42.62 -> 80.41 | Helped |
| Use SPCC rather than single-description CAD code (ablation) | Text-to-CAD generation · Text-to-CAD | SR: 21.46 -> 99.30 | Helped |
| Use hierarchical descriptions with CAD command sequences rather than single descriptions (ablation) | Text-to-CAD generation · Text-to-CAD | ACCparam: 25.56 -> 47.32 | Helped |
| Use CAD code rather than CAD command sequences with hierarchical descriptions (ablation) | Text-to-CAD generation · Text-to-CAD | ACCparam: 47.32 -> 59.09 | Helped |
| Use CAD code rather than CAD command sequences with hierarchical descriptions (ablation) | Text-to-CAD generation · Text-to-CAD | SR: 98.71 -> 99.30 | Helped |
| Use SPCC input and output for deletion instead of CAD code (ablation) | GPT-4 CAD editing · Deletion* | EM: 66.20 -> 90.41 | Helped |
| Use SPCC input and output for deletion instead of CAD code (ablation) | CAD-Llama-INS CAD editing · Deletion* | EM: 81.93 -> 89.55 | Helped |
Surprising findings
- GPT-4 outperforms CAD-Llama-INS on the plain deletion task (66.20 vs 81.93 EM is actually lower for GPT-4, so this holds only for the deletion* variant where GPT-4 scores 90.41 vs 89.55).
Limitations
- Relies on GPT-4o-generated annotations, which may hallucinate; two-shot prompting used to reduce this.
- Evaluated only on DeepCAD-derived data after removing simple models; generalization to other CAD datasets not shown.
- Only an 8B backbone is studied; the authors note larger models and corpora may help.
Experimental setup
Datasets: DeepCAD
Benchmarks: DeepCAD test set
Metrics: COV MMD JSD SR Novel ACCT MCD BLEU Rouge ACCcmd
Baselines: DeepCAD SkexGen HNC-CAD Text2CAD CAD-Translator GPT-4 GPT-3.5 LLaMA3-8B Mistral-7B
Builds on: LLaMA3 DeepCAD LoRA CLIP
Closest papers by meaning
- Text-to-CAD Generation Through Infusing Visual Feedback in Large Language Models (ICML 2025 · similarity 0.90)
- From 2D CAD Drawings to 3D Parametric Models: A Vision-Language Approach (AAAI 2025 · similarity 0.90)
- Generating CAD Code with Vision-Language Models for 3D Designs (ICLR 2025 · similarity 0.89)
- FlexCAD: Unified and Versatile Controllable CAD Generation with Fine-tuned Large Language Models (ICLR 2025 · similarity 0.89)
- CAD-GPT: Synthesising CAD Construction Sequence with Spatial Reasoning-Enhanced Multimodal LLMs (AAAI 2025 · similarity 0.89)
- CAD-Coder: Text-to-CAD Generation with Chain-of-Thought and Geometric Reward (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.