PR-CAD: Progressive Refinement for Unified Controllable and Faithful Text-to-CAD Generation with Large Language Models

Jiyuan An; Jiachen Zhao; Fan Chen; Liner Yang; Zhenghao Liu; Hongyan Wang; Weihua An; Meishan Zhang

cs.CLMar 27, 2026

PR-CAD: Progressive Refinement for Unified Controllable and Faithful Text-to-CAD Generation with Large Language Models

Authors: Jiyuan An, Jiachen Zhao, Fan Chen, Liner Yang, Zhenghao Liu, Hongyan Wang, Weihua An, Meishan Zhang,

Organizations: National Language Resources Monitoring and Research Center for Print Media, Beijing Language and Culture University, China · School of Information Science, Beijing Language and Culture University, China · School of Computer Science and Technology, Beijing Jiaotong University, China · School of Computer Science and Engineering, Northeastern University, China ·

Abstract

The construction of CAD models has traditionally relied on labor-intensive manual operations and specialized expertise. Recent advances in large language models (LLMs) have inspired research into text-to-CAD generation. However, existing approaches typically treat generation and editing as disjoint tasks, limiting their practicality. We propose PR-CAD, a progressive refinement framework that unifies generation and editing for controllable and faithful text-to-CAD modeling. To support this, we curate a high-fidelity interaction dataset spanning the full CAD lifecycle, encompassing multiple CAD representations as well as both qualitative and quantitative descriptions. The dataset systematically defines the types of edit operations and generates highly human-like interaction data. Building on a CAD representation tailored for LLMs, we propose a reinforcement learning-enhanced reasoning framework that integrates intent understanding, parameter estimation, and precise edit localization into a single agent. This enables an "all-in-one" solution for both design creation and refinement. Extensive experiments demonstrate strong mutual reinforcement between generation and editing tasks, and across qualitative and quantitative modalities. On public benchmarks, PR-CAD achieves state-of-the-art controllability and faithfulness in both generation and refinement scenarios, while also proving user-friendly and significantly improving CAD modeling efficiency.

Figures & tables

Figure 1: PR-CAD enables user-friendly and controllable CAD generation through progressive refinement.Below each subfigure is the user’s intended input described via text instructions, with the corresponding CAD model output above. PR-CAD allows users to generate and progressively refine (add, modify, and delete) their designs from scratch. Users can flexibly choose either qualitative or quantitative descriptions of their intent until the model matches their envisioned design.

Figure 2: High-quality data annotation pipeline for generation task. Based on the DeepCAD dataset, it generates both generation instructions and CAD sequence representations. Multimodal large models are utilized to produce quantitative and qualitative CAD generation instructions from nine views and JSON descriptions. Code translation and large language models are applied to convert JSON descriptions into various CAD sequence representations.

Figure 3: Human-like instruction annotation pipeline for CAD model editing task. In the first stage, CAD model pairs are generated by removing S-E operations or loops, forming deletion and addition pairs as training data (as shown in I). Simultaneously, qualitative add instructions are generated based on the deleted portions. Next, using the trained add/delete model, new CAD models are generated based on the above instructions (as shown in II). The resulting new model and the original model form the edited CAD model pair. Finally, the JSON descriptions are converted into editing instructions and sequence representations in the same manner as for generation tasks (Figure 2 ).

Figure 4: Overview of the PR-CAD post-training process. The post-training process consists of two stages: supervised fine-tuning and reinforcement learning. Using the highly human-like interaction data curated earlier, we employ Qwen2.5-7B-Instruction as the base model. First, DeepSeek ( Liu et al., 2024 ) is used to construct a structured chain of thought by analyzing the input and output data. Then, the model is fine-tuned through supervised learning with cross-entropy loss to learn the task patterns. Afterward, reinforcement learning is applied to the fine-tuned model using rewards based on chamfer distance, among others. For generation tasks, the Current CAD Model is initially empty; for editing tasks, chamfer distance is calculated only for the edited portion of the model.

Method	Generation Task				Editing Task
Method	Quantitative		Qualitative		Quantitative		Qualitative
	IR ↓	Mean CD ↓	IR ↓	VLM-Eval ↑	IR ↓	Mean CD ↓	IR ↓	VLM-Eval ↑
GPT-4o (zero-shot)	74.22	133.52	66.48	55.85	25.81	23.30	27.76	61.01
GPT-4o (few-shot)	55.95	77.49	49.24	58.91	15.47	15.52	13.47	66.18
Text2CAD	0.97	27.68	3.41	66.35	✗	✗	✗	✗
Text-to-CadQuery	6.62	11.32	✗	✗	✗	✗	✗	✗

Table 1: Comparison of PR-CAD with existing methods for both generation and editing tasks. The evaluation metrics include quantitative measures (IR, Mean CD) and qualitative evaluations (IR, VLM-Eval). PR-CAD significantly outperforms other methods across both tasks, demonstrating superior controllability and model faithfulness. CD values are scaled by 10 3 . ”✗” indicates that the method does not support the corresponding task. ↑: the higher, the better; ↓: the lower, the better. Best performance is highlighted in bold .

Figure 5: Visualization comparison among different methods or models, including the closed-source model GPT-4o relying on context learning capabilities, Text-to-CADQuery for CAD model generation, CAD-Editor designed for CAD editing, and our proposed method.

Interaction Method	Experts				Novices
Interaction Method	Success Rate	Avg. Turns	Avg. Time	SUS	Success Rate	Avg. Turns	Avg. Time	SUS
Single-turn Generation	56.25%	1	6’48”	38.125	31.25%	1	9’27”	25.125
Multi-turn Interaction	100%	3.375	3’24”	93.125	81.25%	4.53	5’16”	80.625

Table 2: Human Interaction Results. Comparison of experts and novices using two interaction methods: end-to-end generation and Progressive Refinement. Metrics evaluated include success rate, average number of turns, average time per turn, and System Usability Scale (SUS) scores.

Figure 6: Examples of CAD modeling through multi-turn dialogues. (a) In this step, the specified cube is enlarged based on the earlier description. (b) The next step accurately identifies the cube from the previous operation and completes the enlargement. (c) In this step, the edited cube is identified and moved towards the base edge. (d) Finally, this step precisely undoes all the operations performed in the previous three steps.

Training Strategy	IR	Mean CD	Median CD	VLM-Eval
Qwen2.5-7B	✗	✗	✗	✗
Qwen2.5-7B w/o SFT	✗	✗	✗	✗
Qwen2.5-7B w/o RL	12.67	84.63	6.54	56.44
Qwen2.5-7B w/o SCoT	2.48	11.34	2.08	68.81
Qwen2.5-7B t/o/o Generation	9.45	42.07	5.49	60.56
Qwen2.5-7B t/o/o Editing	6.94	20.05	1.90	65.54

Table 3: Ablation Studies for Post-Tuning LLMs with Various Training Configurations. We assess the impact of different training configurations, including baseline models, reinforcement learning-only fine-tuning (w/o SFT), supervised fine-tuning-only (w/o RL), training without structured reasoning (w/o SCoT), and single-task data training. The notation t/o/o (trained only on) indicates that the model was trained exclusively on one type of data. Best performance is highlighted in bold .

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: The difference in model performance caused by using different CAD sequence representations, while keeping all other supervised fine-tuning strategies identical. Best performance is highlighted in bold.

Figure 8: Numerical issues due to scaling. The above image shows a typical example where accuracy errors caused by scaling result in internal through-holes, which should have the same radius, displaying ladder.

Figure 9: Robustness Example. When the user requests the outer ring radius to be smaller than the inner ring radius, the model automatically reduces the inner ring radius to ensure the model can generate correctly.

Text-to-CAD Generation

Text-to-CAD generation turns natural-language design intent into executable, editable parametric 3D models and assemblies.

See all 48 papers

TOOLCAD: Exploring Tool-Using Large Language Models in Text-to-CAD Generation with Reinforcement Learning
Apr 9, 2026Yifei Gong, Xing Wu, Wenda Liu +1Text-to-CAD Generation LLM Tool Use
School of Computer Engineering & Science, Shanghai University
CADFS: A Big CAD Program Dataset and Framework for Computer-Aided Design with Large Language Models
May 3, 2026Vladislav Pyatov, Gleb Bobrovskikh, Saveliy Galochkin +6Parametric CAD Modeling Program Synthesis
1Applied AI Institute · 2AI Foundation and Algorithm Lab · 3AXXX +1
Towards High-Fidelity CAD Generation via LLM-Driven Program Generation and Text-Based B-Rep Primitive Grounding
Mar 12, 2026Jiahao Li, Qingwang Zhang, Qiuyu Chen +3Text-to-CAD Generation Parametric CAD Modeling
Fudan University · Shanghai Jiao Tong University

Apr 9, 2026J/K move · Enter open · S save

TOOLCAD: Exploring Tool-Using Large Language Models in Text-to-CAD Generation with Reinforcement Learning

Yifei Gong, Xing Wu, Wenda Liu, Kang Tu

School of Computer Engineering & Science, Shanghai University

Text-to-CAD Generation LLM Tool Use LLM Agent Training RL for Tool Use

Open paper AskPDF arXiv

Computer-Aided Design (CAD) is an expert-level task that relies on long-horizon reasoning and coherent modeling actions. Large Language Models (LLMs) have shown remarkable advancements in enabling language agents to tackle real-world tasks. Notably, there has been no investigation into how tool-using LLMs optimally interact with CAD engines, hindering the emergence of LLM-based agentic text-to-CAD modeling systems. We propose ToolCAD, a novel agentic CAD framework deploying LLMs as tool-using agents for text-to-CAD generation. Furthermore, we introduce an interactive CAD modeling gym to rollout reasoning and tool-augmented interaction trajectories with the CAD engine, incorporating hybrid feedback and human supervision. Meanwhile, an end-to-end post-training strategy is presented to enable the LLM agent to elicit refined CAD Modeling Chain of Thought (CAD-CoT) and evolve into proficient CAD tool-using agents via online curriculum reinforcement learning. Our findings demonstrate ToolCAD fills the gap in adopting and training open-source LLMs for CAD tool-using agents, enabling them to perform comparably to proprietary models, paving the way for more accessible and robust autonomous text-to-CAD modeling systems.

Abstract

Text-to-CAD Generation

Explore similar work

TOOLCAD: Exploring Tool-Using Large Language Models in Text-to-CAD Generation with Reinforcement Learning

CADFS: A Big CAD Program Dataset and Framework for Computer-Aided Design with Large Language Models

Towards High-Fidelity CAD Generation via LLM-Driven Program Generation and Text-Based B-Rep Primitive Grounding