Rethinking Human Intent-to-CAD: Parametric CAD Model Generation via Cooperative Multi-Task Alignment and Spatial-Aware Reinforcement Learning

Qingwang Zhang; Jiahao Li; Xiangdong Zhou

cs.CVOct 11, 2026

Rethinking Human Intent-to-CAD: Parametric CAD Model Generation via Cooperative Multi-Task Alignment and Spatial-Aware Reinforcement Learning

Authors: Qingwang Zhang, Jiahao Li, Xiangdong Zhou

Organizations: College of Computer Science and Artificial Intelligence, Fudan University, Shanghai, China

Abstract

Parametric Computer-Aided-Design (CAD) modeling from human intent remains challenging, particularly during the conceptual design stage, where design goals are expressed through incomplete and unstructured modalities (e.g., hand-drawn sketches and textual descriptions). In this work, we rethink the human intent-to-CAD pipeline and propose a unified method that directly maps multi-level human intents to executable codes, without assuming the prior existence of target CAD models. To support our study, we construct HiCAD, the first large-scale dataset aligning hand-drawn sketches, textual descriptions, and parametric CAD codes. Based on this, we introduce HiCAD, a two-stage framework comprising Cooperative Multi-Task Alignment to bridge the representational gap between heterogeneous inputs, and Spatial-Aware Reinforcement Learning to enforce geometric and topological consistency. Extensive experiments demonstrate that our method significantly outperforms existing baselines across multiple tasks, validating its effectiveness and robustness in transforming heterogeneous human intents into high-fidelity parametric CAD models. Our project page: https://zqwlearning.github.io/HiCAD.

Figures & tables

Figure 1: We rethink the human intent-to-CAD pipeline, aiming to explore multi-level human intent-to-CAD tasks expressed using four types of hand-drawn sketches and textual descriptions.

Figure 2: Illustration of the creation process for synthetic and real hand-drawn sketches.

Figure 3: Textual description generation pipeline. First, we execute CadQuery code to obtain a B-rep of the CAD model for rendering the image, and a Mesh representation of the CAD model for calculating geometric attributes ( e.g. , length, width, and height). Then, we inject the original CadQuery code, the rendered image, and the geometric attributes together into a carefully designed prompt template, which is subsequently provided to a VLM to generate a complete textual description.

Figure 4: An overview of the proposed HiCAD framework, consisting of Stage I: Cooperative Multi-Task Alignment (CMTA) and Stage II: Spatial-Aware Reinforcement Learning (SARL) . In the CMTA stage, a pre-trained VLM is collaboratively aligned across multiple tasks. In the SARL stage, the HiCAD -CMTA model obtained from CMTA is used to initialize the policy model $\pi_{\theta}$ ; multiple rollouts are performed for a query $q$ , and a spatial-aware reward is computed to update $\pi_{\theta}$ via GSPO.

Method	HDS-to-CAD Task				CTD-to-CAD Task				HDS & PTD-to-CAD Task				HDS & CTD-to-CAD Task
Method	IoU (%) $\uparrow$	IR (%) $\downarrow$	Mean CD $\downarrow$	Med. CD $\downarrow$	IoU (%) $\uparrow$	IR (%) $\downarrow$	Mean CD $\downarrow$	Med. CD $\downarrow$	IoU (%) $\uparrow$	IR (%) $\downarrow$	Mean CD $\downarrow$	Med. CD $\downarrow$	IoU (%) $\uparrow$	IR (%) $\downarrow$	Mean CD $\downarrow$	Med. CD $\downarrow$
GPT-4o	20.49	7.00	53.52	35.31	47.64	20.80	25.86	5.93	37.41	16.60	38.41	16.93	51.52	17.20	22.57	4.50
GPT-5-mini	9.51	16.00	67.53	53.23	53.61	23.60	21.87	3.98	38.20	9.80	33.39	13.08	52.96	14.40	19.58	3.42
Gemini-2.5	23.72	19.60	44.70	23.51	52.94	35.80	28.22	4.95	40.87	28.20	34.85	11.48	54.01	23.20	26.09	3.27
Claude-3.7	19.48	15.80	49.51	31.70	44.21	28.60	30.11	6.74	34.11	22.20	38.86	18.68	46.81	26.20	28.99	6.91
Qwen-VL-Max	22.21	14.20	44.58	23.17	46.46	54.20	28.83	6.21	36.86	47.20	35.06	16.59	52.73	53.20	23.77	4.48

Table 1: Results for four multi-level human intent-to-CAD tasks on the HiCAD dataset.

{\dagger}

denotes training on the HiCAD dataset. Median CD is denoted as Med. CD . CD is multiplied by

10^{3}

. Bold and underline indicate the best and the second best result.

Figure 5: Qualitative comparison of different methods across four human intent-to-CAD tasks. refers to an invalid and invisible model.

Method	HDS-to-CAD Task				CTD-to-CAD Task				HDS & PTD-to-CAD Task				HDS & CTD-to-CAD Task
Method	IoU (%) $\uparrow$	IR (%) $\downarrow$	Mean CD $\downarrow$	Med. CD $\downarrow$	IoU (%) $\uparrow$	IR (%) $\downarrow$	Mean CD $\downarrow$	Med. CD $\downarrow$	IoU (%) $\uparrow$	IR (%) $\downarrow$	Mean CD $\downarrow$	Med. CD $\downarrow$	IoU (%) $\uparrow$	IR (%) $\downarrow$	Mean CD $\downarrow$	Med. CD $\downarrow$
ST-SFT (Vanilla)	63.84	1.27	7.68	0.66	71.01	1.49	8.91	0.31	66.61	1.44	8.07	0.54	74.35	1.97	6.53	0.27
CMTA	66.87	1.20	5.69	0.48	71.90	0.95	8.08	0.29	72.39	1.07	5.43	0.33	79.28	1.10	4.05	0.22
CMTA + SARL ( Ours )	69.09	0.22	4.57	0.46	73.79	0.30	7.26	0.30	76.03	0.42	3.91	0.30	80.90	0.42	3.04	0.22

Table 2: Ablation study on the effectiveness of CMTA and SARL. CMTA and SARL represent our Cooperative Multi-Task Alignment and Spatial-Aware Reinforcement Learning, respectively. Notations are the same as Table 1 .

Method	HDS-to-CAD Task		HDS & CTD-to-CAD Task
Method	IoU (%) $\uparrow$	IR (%) $\downarrow$	IoU (%) $\uparrow$	IR (%) $\downarrow$
GPT-4o	7.93	20.00	18.78	10.00
Gemini-2.5	9.52	36.00	20.28	42.00
Claude-3.7	8.10	66.00	23.71	56.00
CAD-Coder ${\dagger}$	35.72	8.00	54.57	3.00
Cadrille ${\dagger}$	17.74	87.00	59.20	9.00

Table 3: Results of different methods on HDS-to-CAD and HDS & CTD-to-CAD tasks based on real sketches drawn by humans.

Figure 6: Results of training on different tasks during the SARL stage. “ALL” indicates that SARL is performed on all tasks.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: The complete prompt template used in the text description generation pipeline.

Figure 8: Qualitative comparison of different methods across four human intent-to-CAD tasks. refers to an invalid and invisible model.

Figure 9: Some examples of our HiCAD framework handling real hand-drawn sketches. The red dashed boxes indicate areas that are clearly incorrect.

Parametric CAD Modeling

Parametric CAD modeling represents 3D designs as editable, constraint-driven construction histories and develops methods to generate, reconstruct, and revise them.

See all 81 papers

HierCAD: Hierarchical Text-to-CAD Design via Structure Alignment and Parameter Grounding
Jul 13, 2026Jimin Xu, Tianbao Wang, Tao Jin +1Parametric CAD Modeling Text-to-CAD Generation
College of Computer Science and Technology, Zhejiang University, Hangzhou, China
Zero-to-CAD: Agentic Synthesis of Interpretable CAD Programs at Million-Scale Without Real Data
Apr 27, 2026Mohammadmehdi Ataei, Farzaneh Askari, Kamal Rahimi Malekshan +1Synthetic Data Generation Parametric CAD Modeling
Autodesk Research
ReliCAD: From Uncertain LLM Generation to Reliable Parametric CAD Modeling
Sep 16, 2026Peng Zheng, Xintong Dong, Chuanyang Li +5Parametric CAD Modeling Text-to-CAD Generation
University of Science and Technology of China, Hefei, Anhui 230026, China

Jul 13, 2026J/K move · Enter open · S save

HierCAD: Hierarchical Text-to-CAD Design via Structure Alignment and Parameter Grounding

Jimin Xu, Tianbao Wang, Tao Jin, Zhou Zhao

College of Computer Science and Technology, Zhejiang University, Hangzhou, China

Parametric CAD Modeling Text-to-CAD Generation

Open paper AskPDF arXiv

Recent text-to-CAD approaches have shown promising results by leveraging large language models, but they often struggle with maintaining structural consistency in complex designs and accurately grounding geometric parameters. To address these issues, we propose HierCAD, a hierarchical text-to-CAD framework that improves both structural reasoning and parameter prediction. HierCAD reformulates CAD generation as progressive reasoning by decomposing CAD construction trees into object-level procedural reasoning and part-level topology reasoning trajectories. To further improve generation fidelity, we introduce a unified Structure Alignment and Parameter Grounding (SAPG) learning strategy. Structure alignment aligns topology reasoning trajectories with their corresponding parametric CAD spans, while parameter grounding mitigates shortcut learning through structure-preserving parameter perturbations and ranking-based supervision. Experiments demonstrate that HierCAD outperforms prior state-of-the-art methods on both CAD sequence generation and reconstructed CAD model evaluation. Our code is available at https://github.com/Collab-Gen/HierCAD.

Abstract

Parametric CAD Modeling

Explore similar work

HierCAD: Hierarchical Text-to-CAD Design via Structure Alignment and Parameter Grounding

Zero-to-CAD: Agentic Synthesis of Interpretable CAD Programs at Million-Scale Without Real Data

ReliCAD: From Uncertain LLM Generation to Reliable Parametric CAD Modeling