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 Erhong Yang · 2026

Read PR-CAD: Progressive Refinement for Unified Controllable and Faithful Text-to-CAD Generation with Large Language Models by Jiyuan An, Jiachen Zhao, Fan Chen, Liner Yang, Zhenghao Liu, Hongyan Wang Weihua An Meishan Zhang Erhong Yang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-19773.txt.

Introduction

An and eight colleagues present PR-CAD, a progressive refinement framework for controllable, faithful text-to-CAD modeling with large language models. It joins generation with editing. Building CAD models by hand is labor-intensive and demands expert skill. Current text-to-CAD tools split making and editing into separate tasks. PR-CAD is backed by a high-fidelity interaction dataset spanning the full CAD life cycle. It covers multiple CAD representations and both qualitative and quantitative descriptions.

Catalog record

CREATORJiyuan An, Jiachen Zhao, Fan Chen, Liner Yang, Zhenghao Liu, Hongyan Wang Weihua An Meishan Zhang Erhong Yang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE39762
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.19773

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

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 Erhong Yang 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. 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. 1 INTRODUCTION Computer-Aided Design (CAD) is a cornerstone of modern manufacturing, engineering, and industrial design, enabling the creation of precise and complex 3D models. However, traditional CAD modeling remains a labor-intensive process that demands significant expertise, with designers needing to master intricate software interfaces and possess a deep understanding of geometric modeling principles. This steep learning curve limits accessibility and efficiency, highlighting a long-standing research challenge: enabling the intuitive creation and manipulation of CAD models through natural language. The advent of large language models (LLMs) has spurred progress in text-to-CAD generation, with recent work demonstrating the feasibility of converting textual descriptions into CAD operations. Despite these advancements, a critical limitation remains: current approaches largely treat generation and editing as separate, disjoint tasks. In practice, design is an inherently iterative process. Designers rarely produce a perfect model on their first attempt; rather, they refine their designs through a series of edits (modifying, adding, or removing) features in response to evolving requirements or identified improvements. Existing systems lack a unified framework for this iterative refinement, forcing users to either settle for suboptimal initial models or revert to manual editing outside of the generative process. This disconnection severely limits the practicality and user-friendliness of text-to-CAD systems. Moreover, many current methods rely on highly detailed, technical text prompts, which are often unnatural for non-expert users to formulate. While some efforts have begun to explore editing, they are often constrained by training data that predominantly features simplistic, randomly generated edits, failing to capture the nuanced, intent-driven nature of real-world design refinements. There is a pressing need for a unified solution that seamlessly integrates generation with controllable, faithful editing based on both qualitative (e.g., ”make the base thicker”) and quantitative (e.g., ”reduce the radius by 6mm”) instructions. To address these challenges, we propose PR-CAD, a progressive refinement framework that unifies controllable and faithful text-to-CAD generation. Our approach is built on three key innovations: First, we curate a high-quality, human-like interaction dataset that spans the entire CAD lifecycle, systematically defining and generating diverse edit operations alongside both qualitative and quantitative descriptions. Second, we introduce a reinforcement learning-enhanced reasoning framework that integrates intent understanding, parameter estimation, and precise edit localization into a single agent, enabling an ”all-in-one” solution for both generation and iterative refinement. Third, we employ a structured chain-of-thought (SCoT) methodology to guide the LLM’s reasoning, breaking down complex tasks into manageable steps for robust and interpretable generation. Extensive experiments demonstrate that PR-CAD achieves state-of-the-art performance on public benchmarks, significantly outperforming existing methods in both generation and editing tasks across multiple metrics, including geometric accuracy (Chamfer Distance) and faithfulness to user intent (VLM-Eval). Crucially, human evaluations confirm that our progressive refinement paradigm dramatically improves usability and success rates for both expert and novice users, making professional-grade CAD modeling more accessible. In summary, our contributions are: • We propose PR-CAD, a novel progressive refinement framework that unifies text-driven CAD generation and editing within a single, controllable agent, providing a seamless workflow for both creation and modification tasks. • We introduce a high-fidelity interaction dataset for the full CAD lifecycle, supporting diverse edit types, descriptive modalities, and ensuring comprehensive coverage of CAD interactions from start to finish. • We develop a reinforcement learning-enhanced reasoning method combining supervised fine-tuning (SFT), structured chain-of-thought (SCoT), and reinforcement learning (RL) to optimize for geometric fidelity, executability, and alignment with user intent, validated through empirical experiments using ChatCAD. 2 RELATED WORK CAD Model Generation . Traditional CAD modeling is a highly manual process, requiring deep domain expertise and precision in design. In recent years, research has emerged that aims to automate or simplify this process, particularly by integrating natural language input for CAD generation. Parametric CAD 22 generation has been an important focus, where CAD models are represented as a sequence of operations (e.g., sketching and extrusion). This approach not only describes the geometric structure of the model but also captures the historical process and evolution of the design. With the introduction of large-scale datasets like DeepCAD ( 33 ) , significant progress has been made in text-to-CAD model generation based on Transformer architectures ( 29 ) . 10 proposed Text2CAD, a Transformer-based framework that effectively generates CAD models from textual descriptions. This was followed by additional works leveraging the DeepCAD dataset, such as Text-to-CADQuery ( 34 ) and GeoCAD ( 38 ) . GeoCAD introduces local geometric controls into the text-to-CAD generation process, enabling more flexible models with fewer input details. Furthermore, Seek-CAD ( 14 ) introduces a self-improvement mechanism, where the model iteratively refines its designs based on feedback, improving the accuracy of CAD models and enhancing their applicability in real-world scenarios. 20 , 30 , 16 , 13 and 15 are also paying attention to this research area. However, these methods heavily rely on detailed and highly technical text inputs, making it challenging for users to write instructions that ensure accurate CAD model generation. For example, in the Text2CAD dataset, an average of 100 words is required to describe a single s-E operation, which significantly impacts the practical usability of these models. CAD Model Editing . To address issues such as localized errors or changing requirements in CAD models, the task of CAD model editing has been proposed. Methods like FLEXCAD ( 39 ) explore the random editing of CAD models, allowing users to modify various model parameters without needing to regenerate the entire design. These techniques increase design flexibility but often fall short in high-precision CAD modeling tasks due to the inherent randomness of the editing process. On the

Read the complete text