Reproduce “CAD-Judge: Toward Efficient Morphological Grading and Verification for Text-to-CAD Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zheyuan Zhou, Jiayi Han, Liang Du, Naiyu Fang, Lemiao Qiu, Shuyou Zhang
CAD-Judge enhances text-to-CAD generation by developing a compiler-driven feedback system for efficiently aligning Large Language Models. This framework achieves superior geometric fidelity and significantly reduces the invalidity ratio of generated CAD sequences, demonstrating strong generalization across diverse datasets.
Key insights
- Direct compiler feedback, leveraging compilation success and geometric comparison (Chamfer Distance), provides a more efficient and verifiable reward signal for LLM alignment than traditional VLM-based methods.
- An agentic generation-review loop during inference, mediated by a CAD compiler, significantly enhances the robustness and syntactic correctness of generated CAD sequences by iteratively correcting errors.
- Optimizing generative utility using a binary preference framework (KTO), grounded in prospect theory, effectively guides LLMs to produce both geometrically accurate and syntactically valid CAD models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.