💬 NLP

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

PR-CAD introduces a progressive refinement framework that unifies text-to-CAD generation and editing into a single, controllable agent by leveraging a high-fidelity interaction dataset and reinforcement learning to achieve state-of-the-art faithfulness and efficiency in CAD modeling.

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

Published 2026-04-23

📖 5 min read🧠 Deep dive

CC BY 4.0

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

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are an architect who wants to build a complex 3D house, but instead of using blueprints and tools, you have to talk to a very smart, but slightly literal, robot assistant.

The Problem with Old Robots
In the past, asking a computer to build a 3D model (like a chair, a car part, or a toy) was like trying to program a video game character by typing lines of code. You needed to be an expert. If you wanted to change the chair's leg from round to square, you couldn't just say, "Make the leg square." You had to know the exact mathematical coordinates and delete the old code, then write new code.

Even newer "AI" assistants that could understand text had a major flaw: they were great at building the first draft, but terrible at fixing it. If you said, "The leg is too long," the robot would either ignore you or crash. It treated "building" and "fixing" as two completely different jobs, forcing you to start over or do the editing manually.

The Solution: PR-CAD (The "Progressive Refinement" Architect)
The paper introduces PR-CAD, a new system that acts like a master architect who doesn't just build once and leave. Instead, it works with you in a continuous conversation to get the design just right.

Here is how it works, using some simple analogies:

1. The "Human-Like" Training (The Internship)

Before PR-CAD could talk to you, the researchers had to teach it how real designers think.

The Old Way: They trained robots on random, robotic instructions like "Add a cylinder at X,Y,Z."
The PR-CAD Way: They created a massive library of human-like interactions. Imagine a master designer working with a junior apprentice. The master says, "Make the base thicker," or "Shrink this wheel by 6 millimeters." The system learned to understand both vague ideas ("make it look sturdier") and precise math ("reduce radius by 6mm").
The Analogy: It's like teaching a student not just by giving them a textbook of formulas, but by having them shadow a master chef for years, watching how the chef tastes the soup and says, "Needs more salt," or "Too spicy, add water."

2. The "All-in-One" Brain (The Reasoning Agent)

PR-CAD uses a special "brain" (a Large Language Model) that has been supercharged with three tools:

Supervised Fine-Tuning (SFT): Like memorizing the rules of the game.
Structured Chain of Thought (SCoT): This is the most important part. Instead of guessing the answer, the AI is forced to think out loud in steps.
- Step 1: "What does the user want?" (Intent)
- Step 2: "What does the current model look like?" (Analysis)
- Step 3: "How much do I need to change the numbers?" (Math)
- Step 4: "Where exactly do I make the cut?" (Location)
Reinforcement Learning (RL): This is the "trial and error" coach. If the AI builds a chair that looks like a blob, it gets a "bad grade" (a penalty). If it builds a perfect chair, it gets a "gold star." Over time, it learns to avoid mistakes and aim for perfection.

3. The "Progressive Refinement" (The Conversation)

This is the magic sauce. With PR-CAD, you don't have to get it perfect on the first try.

You: "Build me a coffee mug."
AI: Builds a basic mug.
You: "The handle is too thin."
AI: Thickens the handle.
You: "Actually, make the handle a bit higher up."
AI: Moves the handle.
You: "Now, make the bottom wider so it doesn't tip over."
AI: Widens the base.

The system remembers the whole history. It doesn't forget the mug it built in step one; it just tweaks the specific part you asked for. It's like sculpting clay: you don't throw the clay away and start a new statue every time you want to fix the nose; you just pinch the clay and shape it.

Why This Matters

For Experts: It speeds up work. Instead of clicking through 50 menus to change a dimension, they just type a sentence.
For Beginners: It removes the fear. You don't need to know engineering jargon. You can say, "Make it look like a toy," and the AI figures out the technical details.
The Result: The paper shows that PR-CAD is much more accurate and less likely to make "glitchy" models than previous methods. It understands that design is a conversation, not a one-time command.

In a Nutshell:
PR-CAD turns the difficult, technical process of 3D modeling into a natural conversation. It's the difference between trying to write a computer program to build a house versus simply talking to a contractor who understands your vision and knows exactly how to adjust the plans until you say, "Yes, that's it!"

1. Problem Statement

Traditional Computer-Aided Design (CAD) modeling is a labor-intensive process requiring specialized expertise and complex software interfaces. While recent Large Language Model (LLM) research has enabled text-to-CAD generation, existing approaches suffer from three critical limitations:

Disjoint Tasks: Current systems treat generation (creating from scratch) and editing (modifying existing designs) as separate, disjoint tasks. In reality, design is an iterative process requiring continuous refinement (adding, modifying, deleting features), which existing unified frameworks fail to support.
Lack of Controllability and Faithfulness: Existing methods often require highly technical, verbose prompts (e.g., 100+ words for a single operation) and struggle to interpret natural language intent, particularly when mixing qualitative descriptions (e.g., "make it thicker") with quantitative constraints (e.g., "reduce radius by 6mm").
Data Limitations: Training data for editing is often synthetic and random, failing to capture the nuanced, intent-driven nature of real-world human design iterations.

2. Methodology: PR-CAD Framework

The authors propose PR-CAD, a unified framework that integrates generation and editing into a single, progressive refinement workflow. The methodology consists of three core components:

A. High-Fidelity Human-Like Interaction Dataset

To address data scarcity and quality, the authors curated a comprehensive dataset spanning the full CAD lifecycle:

Generation Data: Derived from the DeepCAD dataset, augmented with Qualitative (high-level design intent) and Quantitative (specific numerical parameters) instructions. These are generated using Large Vision Models (LVMs) rendering 9 views of CAD models and Large Language Models (LLMs) processing JSON descriptions.
Editing Data: A multi-stage simulation of human editing:
1. Deletion/Addition Simulation: Systematically removing or adding Sketch-Extrude (S-E) operations from existing models to create model pairs.
2. Instruction Generation: Generating natural language instructions for these changes based on the removed/added parts.
3. Verification: Using the trained model to regenerate the modified design, ensuring the new model and original form a valid "edit pair."
Representations: The dataset supports multiple CAD serialization formats, including Domain-Specific Languages (DSL), Structured Text (ST), and General-Purpose Languages (GPL).

B. Progressive Refinement Model Architecture

The model is built upon the Qwen2.5-7B-Instruct base and employs a three-stage post-training strategy:

Structured Chain-of-Thought (SCoT): The model is guided to break down tasks into four logical steps:
- Intent Understanding: Aligning text instructions with the current CAD state.
- Modeling Analysis: Identifying geometric elements using special markers (e.g., <sketch>).
- Parameter Computation: Calculating necessary transformations (rotations, displacements).
- Position Identification: Locating specific parts of the model to edit.
Supervised Fine-Tuning (SFT): The model is fine-tuned using cross-entropy loss on the curated dataset to learn the mapping between natural language and CAD operations (specifically using LogoUp 3D for generation and Structured Text for editing).
Reinforcement Learning (RL): A Generalized Reward Optimization (GRPO) approach is used to further refine the model. The reward function ( $R$ $R$ ) combines four components:
- $R_{chamfer}$ : Geometric similarity (Chamfer Distance) between generated and ground truth.
- $R_{format}$ : Penalty for output format violations.
- $R_{exec}$ : Penalty for non-executable code/sequences.
- $R_{length}$ : Penalty for unnecessarily long outputs to encourage efficiency.

C. Unified Agent

The system acts as a single "all-in-one" agent capable of handling both initial design creation and iterative refinement (add, modify, delete) based on user feedback, eliminating the need to switch between different tools or models.

3. Key Contributions

PR-CAD Framework: A novel unified framework that seamlessly integrates text-driven CAD generation and editing, enabling progressive refinement through a single agent.
High-Fidelity Dataset: The creation of a comprehensive interaction dataset covering the full CAD lifecycle, featuring diverse edit types and both qualitative/quantitative modalities, which simulates human-like design iterations.
Advanced Training Strategy: A combination of SFT, SCoT, and RL that optimizes for geometric fidelity, executability, and intent alignment, validated through the ChatCAD system.

4. Experimental Results

The authors evaluated PR-CAD on public benchmarks against state-of-the-art methods (including GPT-4o, Text2CAD, FLEXCAD, and CAD-Editor).

Quantitative Performance:
- Generation: PR-CAD achieved the lowest Invalidity Ratio (0.62%) and Mean Chamfer Distance (5.87), significantly outperforming GPT-4o (74.22% IR) and Text2CAD (27.68 CD).
- Editing: In editing tasks, PR-CAD achieved an IR of 0.91% and Mean CD of 6.30, surpassing specialized editors like CAD-Editor.
- Intent Preservation: PR-CAD scored highest on VLM-Eval (77.83), indicating superior alignment with user intent compared to other models.
Human Evaluation:
- Success Rate: Experts achieved a 100% success rate, and novices achieved 81.25% using multi-turn progressive refinement, compared to ~56% and ~31% for single-turn generation.
- Usability: System Usability Scale (SUS) scores were exceptionally high (93.1 for experts, 80.6 for novices), demonstrating that the iterative approach makes professional CAD accessible to non-experts.
Ablation Studies: Removing SFT, RL, or SCoT resulted in significant performance drops, confirming that all three components are essential for the system's success.

5. Significance

PR-CAD represents a paradigm shift in automated CAD design by moving from static, single-shot generation to dynamic, iterative refinement.

Democratization: By supporting natural language (both qualitative and quantitative) and iterative dialogue, it lowers the barrier to entry for CAD modeling, making it accessible to novices without sacrificing professional-grade precision.
Practicality: It solves the "disjoint task" problem, allowing designers to refine models in real-time without reverting to manual editing or regenerating entire designs from scratch.
Efficiency: The system significantly improves modeling efficiency, reducing the time required for design iterations and increasing the success rate of generating faithful CAD models.

In conclusion, PR-CAD establishes a new state-of-the-art for controllable and faithful text-to-CAD generation, bridging the gap between high-level design intent and precise geometric execution through a unified, reinforcement-learned agent.