CADReview 🏗️

🌐 Homepage | 📖 arXiv | 🤗 HuggingFace Dataset

This repo contains the codebase for our paper CADReview: Automatically Reviewing CAD Programs with Error Detection and Correction

🎉 ACL 2025 main

Introduction

We introduce the CAD review task, which aims to automatically detect and correct errors in CAD programs by comparing them with reference images. To support this task, we propose ReCAD, a multimodal large language model (MLLM)-based framework that generates feedback and edits code for accurate 3D reconstruction. We also present CADReview, a large-scale dataset with over 20,000 program–image pairs featuring diverse geometric structures and real-world error types. Our results show that ReCAD significantly outperforms existing models, offering a practical solution for AI-assisted CAD debugging and refinement.

Training and Inference

1. Our training and inference are conducted using the ms-swift framework. Environment configuration: ms-swift >= 3.3, vllm >= 0.7.3.
2. The alignment training for GCR and SGO can be found in: ./training_and_inference/alignment_gcr and ./training_and_inference/alignment_sgo.
3. Training for $\phi_1$ and $\phi_2$ can be found in: ./training_and_inference/feedback_gen and ./training_and_inference/code_editor.
4. The inference script can be found at: ./training_and_inference/inference.py.

Evaluation

Run ./evaluate/eval.sh to perform evaluation.