We are glad to announce that our paper, NURBGen: High-Fidelity Text-to-CAD Generation through LLM-Driven NURBS Modeling, has been accepted at the Fortieth AAAI Conference on Artificial Intelligence (AAAI) 2026 Singapore. Co-led by Muhammad Sadil Khan and Muhammad Usama under the supervision of Didier Stricker and Muhammad Zeshan Afzal, our research introduces the first framework to generate high-fidelity, editable 3D CAD models directly from text by fine-tuning a large language model to produce structured NURBS surface parameters. To overcome the limitations of existing mesh-based or design-history systems, we propose a hybrid symbolic representation that combines untrimmed NURBS with analytic primitives to robustly handle trimmed surfaces and degenerate regions while maintaining token efficiency. Evaluated on our new partABC dataset of 300k annotated CAD components, NURBGen demonstrates strong performance in geometric fidelity and dimensional accuracy, as confirmed by expert evaluations.

Project Page: https://muhammadusama100.github.io/NURBGen/

Code: https://github.com/SadilKhan/NURBGen

Poster: https://mdsadilkhan.onrender.com/publications/data/nurbgen_poster.png

Contact: mohammad.khan@dfki.de