Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Learning CAD Modeling Sequences via Projection and Part Awareness
Authors: Yang Liu, Daxuan Ren, Yijie Ding, Jianmin Zheng, Fang Deng
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show that Part CAD significantly outperforms existing methods for CAD instruction generation in both accuracy and robustness. The work sheds light on part-driven reconstruction of interpretable CAD models, opening new avenues in reverse engineering and CAD automation. 4 Experiments |
| Researcher Affiliation | Academia | Yang Liu Beijing Institute of Technology China, Beijing EMAIL Daxuan Ren Nanyang Technological University Singapore EMAIL Yijie Ding Nanyang Technological University Singapore EMAIL Jianmin Zheng Nanyang Technological University Singapore EMAIL Fang Deng Beijing Institute of Technology China, Beijing EMAIL |
| Pseudocode | No | The paper describes methods and architecture in Section 3 and Appendix A.2, but does not contain explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper provides comprehensive implementation details, including architecture configurations, training hyperparameters, loss formulations, and dataset processing steps. Key components such as autoregressive part decomposition, triplane feature guidance, and CAD instruction generation are fully described in Section 3, Section 4, and Appendix A.2. All experimental protocols follow standard benchmarks with clearly stated evaluation metrics. These details are sufficient to reproduce the main results even without access to the source code. |
| Open Datasets | Yes | Dataset: We train and evaluate our model on the Deep CAD dataset [42], and further perform cross-dataset validation on the Fusion 360 Gallery [12]. |
| Dataset Splits | Yes | Following prior work [43], we remove duplicate data based on geometry similarities, yielding 140k training and 7k test/validation samples. |
| Hardware Specification | Yes | All experiments are conducted on 8 NVIDIA A100-40GB GPUs, with each training run taking around 18 hours. |
| Software Dependencies | No | The paper mentions using the Adam W optimizer [54] and Python OCC [10] but does not provide specific version numbers for these or other key software dependencies. |
| Experiment Setup | Yes | The latent dimension in Part CAD is set to 512. We use k = 40 neighbors for 3D point cloud encoding and k = 60 for projection feature extraction. For adaptive point cloud projection, we set the normal threshold δnormal = 0.5 and grid spacing δgrid = 1 10 6. The model is trained for 200 epochs with a batch size of 32 using the Adam W optimizer [54] (initial learning rate 1 10 4) and an Exponential LR scheduler. Loss weights are set as λskt = 2, λext = 1, λval = 5, and λrot = 1. |