GOPR-CAD builds CAD models through progressive operation graphs
Original title: GOPR-CAD: Graph-of-Operations Progressive Reasoning for CAD Generation
On a 300-part MHCAD benchmark, the training-free pipeline reaches 82.7% Structural Yield.
GOPR-CAD uses a language model as a decision agent over encapsulated CAD operations, progressively building an executable graph that a deterministic compiler turns into a model. The authors say directly generated CAD programs can execute yet still be structurally wrong. They introduce MHCAD, 300 mechanical part requirements derived from Machinery's Handbook, for reference-free STEP/B-rep evaluation. Structural Yield is 87.2% on an external CADPrompt subset.
- The pipeline needs no additional training and supports different language models.
- MHCAD contains 300 standard mechanical part requirements.
- Structural Yield is 82.7% on MHCAD and 87.2% on a CADPrompt subset.
From the original article
1 min read at arxiv.orgGenerating accurate CAD models requires substantial engineering expertise and careful control of geometry, topology, and mechanical constraints. Text-to-CAD aims to simplify this process, but current LLM-based methods often generate CAD programs directly. Language models remain unreliable when producing long code while coordinating coupled geometric, topological, and dimensional constraints.
By Yitian Pang · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.