💻 computer science

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

HierCAD is a hierarchical text-to-CAD framework that enhances structural consistency and parameter accuracy in complex designs by decomposing CAD generation into object-level and part-level reasoning trajectories, guided by a unified Structure Alignment and Parameter Grounding (SAPG) learning strategy.

Original authors: Jimin Xu, Tianbao Wang, Tao Jin, Zhou Zhao

Published 2026-07-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jimin Xu, Tianbao Wang, Tao Jin, Zhou Zhao

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're trying to teach a super-smart robot how to build a complex Lego castle, but you can only talk to it in plain English. You say, "Build a tower with a round window and a square door," and the robot tries to write out the exact instructions for every single brick.

For a while, the best robots (like the ones in previous studies called Text2CAD and CADmium) were trying to do this by writing one giant, long list of instructions from start to finish. But the authors of this new paper, HierCAD, found a major problem with that approach. They noticed that when the robot tries to write the whole list at once, it gets confused. It mixes up the big picture (like "build a tower") with the tiny details (like "this specific brick is 0.5 inches wide"). It's like trying to write a novel, a grocery list, and a math equation all in the same sentence without any commas. The robot often gets the shape right but messes up the numbers, or it builds a tower that looks like a blob because it forgot how the pieces fit together.

The Big Idea: Breaking the Task into "Chapters"

To fix this, the researchers at Zhejiang University came up with a new way to teach the robot. Instead of asking for one giant list, they broke the building process into two distinct "chapters" or layers of thinking, kind of like how a director plans a movie before the actors start filming.

The "Big Picture" Chapter (Object-Level): First, the robot has to figure out the main parts. "Okay, I need a base, then I'll add a wedge, and finally, I'll cut a hole." It plans the order of operations without worrying about the exact measurements yet.
The "Blueprint" Chapter (Part-Level): Once the robot knows what parts to build, it zooms in on each part to figure out the shape. "For this base, I need a loop made of four lines and one arc."

By separating the "what" from the "how," the robot stops getting tangled up. It's like telling a builder, "First, build the frame. Once the frame is up, then worry about the exact size of the windows." This method, called hierarchical reasoning, helps the robot keep the structure of the design consistent, even for very complex objects with many loops and parts.

The "Shortcut" Problem and the "Grounding" Fix

Even with the new two-step plan, the robot had a sneaky habit. It loved to take shortcuts. If it had just written down a number like "5.0" for a circle's radius, it might lazily copy that "5.0" for a completely different measurement, like the distance of a cut, just because it was a number it had seen recently. It wasn't actually reading your text to find the right number; it was just guessing based on what it had already typed.

The paper calls this "shortcut learning," and it leads to weird results, like a door that is the same size as a tiny screw.

To stop this, the authors introduced a special training trick called Structure Alignment and Parameter Grounding (SAPG).

Structure Alignment: They made sure the robot's "blueprint" thoughts matched the final "construction" instructions. If the robot thought the shape was a square loop, the final numbers had to describe a square loop.
Parameter Grounding: They taught the robot to be punished if it just copied numbers. They would give it a text description and then a "trick" version where the numbers were slightly changed but the shape stayed the same. The robot had to learn that the text was the only thing that mattered, not the numbers it had just typed. It had to "ground" its numbers in the story you told it.

Did it Work? (The Proof)

The team tested their new HierCAD robot against the old champions (Text2CAD and CADmium) using a massive dataset of real CAD designs. The results were clear:

Better Shapes: When they measured how well the robot could draw lines, arcs, and circles, HierCAD scored higher. For example, on "Arc F1" (a score for how well it drew curved lines), HierCAD hit 79.43, while the previous best, CADmium, was at 75.68.
Fewer Mistakes: The old robots made invalid designs (like shapes that couldn't actually be built) about 4.12% of the time. HierCAD dropped that error rate to just 1.45%.
Closer to Reality: When they measured how close the robot's 3D model was to the real thing (using something called Chamfer Distance), HierCAD was the closest, with a score of 35.22, beating CADmium's 44.52.

The authors suggest that this approach doesn't just make the robot smarter; it makes it more reliable. By forcing the robot to think in steps and stop taking lazy shortcuts, it can build complex, multi-part machines that actually look like what you asked for.

What the Paper Doesn't Claim

It's important to know what this paper doesn't say. The authors don't claim that text-to-CAD is a "solved" problem. They explicitly state that generating complex objects from text is "far from solved." They also don't say their method works for every type of design imaginable, but rather that it significantly improves the current state of the art for the specific types of CAD sequences they tested. They didn't simulate this in a vacuum; they measured it against real data and found that while the old methods struggled with long, complex designs, HierCAD held its ground much better.

In short, HierCAD is like giving a robot architect a better set of instructions: "Plan the rooms first, then measure the walls, and don't just copy-paste numbers." The result is a robot that builds things that actually make sense.

Technical Summary: HierCAD

Problem Statement

Recent text-to-CAD approaches leveraging Large Language Models (LLMs) have demonstrated promise but face two critical bottlenecks when generating complex designs:

Structural Inconsistency: Current methods typically formulate CAD generation as a flat autoregressive sequence prediction. This approach entangles heterogeneous information (object composition, loop topology, and continuous geometric parameters) within a single token stream, making it difficult for models to maintain long-horizon procedural dependencies, especially for multi-part objects and sketches with numerous loops.
Parameter Grounding Failures: Even when the coarse topology is correct, parameter prediction remains fragile. Autoregressive models often rely on "shortcut learning," spuriously reusing previously generated legal values rather than grounding parameters in the textual description. This leads to inter-field parameter confusion (e.g., confusing extrusion parameters with coordinates) and intra-field parameter collapse (e.g., assigning multiple distinct points to identical coordinates), degrading geometric fidelity and mesh quality.

Methodology

The authors propose HierCAD, a hierarchical text-to-CAD framework designed to improve structural reasoning and parameter grounding through two main innovations:

1. Hierarchical CAD Reasoning

Instead of learning a flat sequence, HierCAD disentangles the CAD construction tree into multiple reasoning layers, reformulating generation as progressive reasoning:

Disentanglement: The CAD sequence is abstracted into a construction tree $T = \{P, F, L\}$ , separating object-level composition ( $P$ ), part-level sketch faces ( $F$ ), and closed loops ( $L$ ) from continuous geometric parameters.
Object-Level Global Procedural Reasoning: The model first predicts a global supervision sequence $G$ describing the step-by-step composition of parts and Boolean operations (New, Join, Cut, Intersect). This acts as a coarse-grained procedural plan.
Part-Level Local Topology Reasoning: For each part, the model predicts a local supervision sequence $P_i$ that determines the existence and primitive sequence (e.g., line|arc|line) of loops before generating specific coordinates.
Progressive Generation: The generation probability is factorized as $P(Y|X) = P(G|X) \prod P(P_i|G, X)$ , allowing the model to resolve compositional structure before instantiating geometric parameters.

2. Unified Structure Alignment and Parameter Grounding (SAPG)

To further improve generation fidelity, HierCAD introduces a unified learning strategy:

Structure Alignment: This component aligns the topology reasoning trajectories with their corresponding parametric CAD spans in the representation space. By minimizing the distance between the pooled hidden states of the topology reasoning and the parametric sequence, the model is encouraged to preserve the predicted loop structure in the final output.
Parameter Grounding: To mitigate shortcut learning, the authors employ structure-preserving parameter perturbations. They generate perturbed targets ( $y^-$ ) where numerical parameters are transformed (via anisotropic scaling and translation) to be inconsistent with the text but consistent with the structure. A ranking-based supervision loss ( $L_{param}$ ) is then applied to ensure the model assigns a higher likelihood to the original text-grounded target than to the perturbed counterpart, forcing the model to rely on long-range textual descriptions rather than spurious value reuse.

Key Contributions

Hierarchical Reformulation: A new text-to-CAD generation paradigm that decomposes flat supervision into global procedural reasoning and local topology reasoning, providing stronger structural priors for complex CAD generation.
SAPG Learning Strategy: A novel learning approach combining structure alignment and parameter grounding to explicitly address topology drift and shortcut-driven parameter errors in autoregressive generation.
Empirical Validation: Comprehensive experiments demonstrating significant performance improvements over state-of-the-art methods in both generated CAD sequence accuracy and reconstructed CAD model quality.

Experimental Results

Experiments were conducted on the DeepCAD dataset using expert-level annotations from Text2CAD and CADmium. The model was built upon Qwen 2.5-Coder-14B.

Quantitative Performance: HierCAD outperformed baselines (Text2CAD and CADmium) across all metrics.
- Sequence Generation: Achieved the highest F1 scores for Line, Arc, Circle, and Extrusion primitives. Notably, it showed substantial gains in Arc F1, indicating better preservation of complex multi-primitive loop structures.
- Model Reconstruction: Achieved the lowest invalidity ratio (1.45%) and best Chamfer Distance (35.22 mean). It also excelled in topological metrics such as Watertightness, EECM, and DMCD, demonstrating superior manifold validity and local surface geometry preservation.
Qualitative Analysis: Visual comparisons showed that while baselines often collapsed loops into simplified shapes, hallucinated parts, or drifted in Boolean operations, HierCAD successfully recovered exact part-level topological structures and Boolean sequences.
Ablation Studies: Removing the SAPG module (HierCAD w/o SAPG) resulted in "numerically plausible but semantically incorrect" values, where the model reused coordinates from unrelated parts of the sequence (shortcut learning). The full HierCAD with SAPG successfully grounded these parameters to the specific textual descriptions.

Significance and Claims

The paper claims that HierCAD addresses the core limitations of existing autoregressive CAD generators: weak long-horizon structural reasoning and shortcut-driven parameter errors. The authors assert that progress in text-to-CAD generation depends not only on stronger foundation models but also on better supervision reformulation and training objectives tailored to CAD construction processes.

By decomposing the generation task into hierarchical reasoning and enforcing structure-parameter alignment, HierCAD provides a path toward more accurate, valid, and editable text-driven CAD generation. The authors position this framework as a step toward reliable CAD copilots for engineering and design applications, capable of handling complex, multi-part, and multi-loop designs that previous methods struggled to resolve.