Large Language Models for Computer-Aided Design (LLM4CAD) Fine-Tuned: Dataset and Experiments

arXiv 2025 51.7 method, benchmark, application

TLDR

Fine-tunes GPT-3.5 models on a new multimodal CAD dataset with text descriptions to improve text-to-CAD generation, evaluated via parsing rate and IoU.

Reasoning

The paper addresses a clear gap (lack of datasets for text-to-CAD) and provides a practical fine-tuning pipeline with comparative experiments. However, the dataset is limited to simple mechanical components and only GPT-3.5 is fine-tuned, leaving generalizability and broader model comparisons unexplored.

Read-first score

Read-first score 51.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 67.

Methodology quality 25%

100

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=baseline,dataset,evaluation,experiment,metric

Recency 8%

86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Topical relevance 42%

41.9

Uses existing LLM keyword relevance scores normalized to 0-100. AI for CAD,computer-aided design,neural CAD,generative CAD,parametric CAD,B-Rep,boundary representation,constructive solid geometry,CSG,sketch extrusion,CAD generation,CAD reconstruction,text-to-CAD,image-to-CAD,point cloud to CAD,CAD program

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=False; code=False; dataset=False; markers=dataset

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 21.

Keyword Scores

text-to-CAD

computer-aided design

CAD generation

CAD program

generative CAD

AI for CAD

neural CAD

parametric CAD

image-to-CAD

sketch extrusion

CAD reconstruction

B-Rep

boundary representation

constructive solid geometry

CSG

point cloud to CAD

Deep Analysis

Innovations

Development of a multimodal dataset of CAD models, sketches, images, and natural-language descriptions for mechanical components (gears, shafts, springs) via Amazon Mechanical Turk.
Use of CAD programs as an intermediate bridge to convert LLM textual output into precise 3D CAD designs.
A pipeline to generate fine-tuning training data for GPT-3.5 for text-to-CAD generation.
Fine-tuning of four GPT-3.5 models with different data sampling strategies based on CAD program length.
Comparative evaluation using parsing rate and Intersection over Union (IoU) metrics against a GPT-4 baseline without fine-tuning.

Methodology

The authors created a multimodal dataset of CAD models, sketches, images, and crowdsourced text descriptions for mechanical parts. They used CAD programs as a bridge to translate LLM text outputs into 3D CAD designs. Four GPT-3.5 models were fine-tuned with sampling strategies varying by CAD program length, and evaluated via parsing rate and IoU against GPT-4.

Key Results

The comparative study of four fine-tuned models yielded guidance on selecting sampling strategies for text-to-CAD fine-tuning, balancing part complexity, model performance, and cost.

Limitations

The dataset is limited to a few representative mechanical component types (gears, shafts, springs).
Only GPT-3.5 models were fine-tuned; findings may not generalize to other LLMs.
Evaluation relies solely on parsing rate and IoU, which may not fully capture CAD generation quality.
The sampling strategies involve a trade-off between part complexity, model performance, and cost, indicating inherent limitations in optimizing all aspects simultaneously.