An AI CAD system plans mechanical operations instead of writing long programs

Why a convincing shape is not enough

Computer-aided design, or CAD, describes geometry and how surfaces, edges and solid bodies connect. These connections are called topology. For mechanical design, a recognizable silhouette is insufficient: holes, separate bodies and shared axes must also match the requirements.

In an arXiv preprint posted on November 8, 2026, Yitian Pang, Yifan Li, Xiaowei Yue and Li Zheng introduce GOPR-CAD. Their central distinction is between successfully executing a CAD program and actually delivering the requested mechanical structure. Software can export a valid solid even when an important feature is absent.

The reported findings are preliminary preprint results, not independent confirmation of engineering reliability.

Letting AI choose operations, not implement geometry

GOPR-CAD uses a language model as a decision-making agent over a fixed library of executable CAD operations. Each operation has a contract specifying its parameters, supported relationships and deterministic geometric implementation.

The process has four stages. First, the model records construction intent: required features, counts, dimensions and relationships. Second, it selects suitable operations. Third, it connects those operations through explicit relationships, such as a hole belonging to a particular body or two features sharing an axis. Finally, deterministic rules complete parameters using stated values, relationships and operation-specific constraints.

The result is an operation graph: a network whose nodes represent construction actions and whose connections record their relationships. A deterministic compiler converts the completed graph into a plan executed by the CAD software. No further language-model decisions occur during this realization step.

This separates deciding what to build from implementing its geometry. It requires no additional model training, although it depends on an existing operation library. Contracts restrict incompatible choices; they do not guarantee that the model chooses the right design. Missing parameters or unsupported construction roles can still prevent completion.

Checking the exported model

The researchers built MHCAD, containing 300 standard mechanical-part requirements derived from Machinery’s Handbook. They evaluated exported STEP files, a CAD exchange format, after importing them back into the geometric software. Structural checks inspect the resulting boundary representation, or B-rep: the surfaces, edges and connections defining the solid.

The six check categories cover bores, material bodies, coaxial alignment, mounting holes, circular patterns and rolling-element counts. Evaluation is reference-free: checks come from requirements rather than comparison with one supposedly unique correct model.

Two scores distinguish quality from delivery. Structural Fidelity measures the fraction of applicable checks passed among valid imported models. Structural Yield includes all applicable checks, giving no credit when output is missing or invalid. Neither score is the percentage of completely correct parts.

An important limitation is coverage. MHCAD has 342 structural checks across 196 prompts. The other 104 prompts contribute to file-production and validity statistics, but not to structural scores.

What the experiments found

With Qwen3.5-9B as its default agent, GOPR-CAD produced valid models for 270 of 300 prompts. It passed 283 of 315 checks attached to valid models, giving 89.8% Structural Fidelity. Against all 342 applicable checks, its Structural Yield was 82.7%.

Text2CAD produced STEP files for 94.0% of prompts but achieved 37.4% Structural Yield. NURBGen produced files for 97.3% but achieved 20.8% yield. In these tests, producing more files did not mean preserving more required structure.

The authors estimated a 95% bootstrap interval of 78.0–87.1% for GOPR-CAD’s yield, resampling prompts together with their associated checks. This captures sampling uncertainty within the benchmark, not every possible source of error.

On an external CADPrompt subset, GOPR-CAD passed 41 of 47 checks, or 87.2%, compared with CADFusion’s 32 of 47, or 68.1%. That subset contained only 30 of the source collection’s 200 cases, selected for scope and checker applicability independently of outputs.

What remains uncertain

Replacing the typed executable pipeline with prose stages followed by direct code generation reduced yield to 0.9%. However, that comparison removed several components together; it does not isolate the compiler or graph schema as the sole cause.

The checkers detected all 86 deliberately introduced topology mutations tested by the authors. That supports sensitivity to those disruptions, not comprehensive detection of design errors. Exact gear-tooth counts and some small details lie outside the validated checks. Some categories also have very small samples.

The work therefore supports a narrower conclusion than automated engineering: structured operation planning improved delivery of selected mechanical properties in these experiments. Complete dimensional accuracy, manufacturability and complex assembly behavior are not established by the reported structural scores.