Text-to-CAD stops writing long programs: GOPR-CAD builds an operation graph and delivers 82.7% structural yield
Synopsis
Most text-to-CAD systems have a large language model write the CAD program directly, so geometric, topological and dimensional constraints can fail early in a long code sequence and propagate through later construction, while the kernel may still accept the program and export a valid solid. GOPR-CAD has the language model decide progressively over encapsulated CAD operations, build an executable operation graph, and hand it to a deterministic compiler, with no additional training. On 300 standard mechanical-part requirements derived from Machinery's Handbook it reaches 82.7% Structural Yield and 89.8% Structural Fidelity.
Interpretation
The construction process moves from having the language model write a complete CAD program in one pass to having it make progressive construction decisions over encapsulated CAD operations, with geometric realization left to a deterministic compiler. Earlier approaches generate CAD programs directly, so the model must hold geometric, topological and dimensional constraints together across a long sequence of dependent instructions, and early errors propagate through later construction. On the 300 standard mechanical-part requirements of MHCAD, GOPR-CAD reaches 82.7% Structural Yield and 89.8% Structural Fidelity; by comparison, Text2CAD produces STEP files for 94.0% of prompts but reaches only 37.4% Structural Yield, and NURBGen reaches 97.3% STEP coverage with 20.8% Structural Yield.
Each operation carries a contract stating its mechanical role, the relations it may join, its parameter interface, and a binding to a deterministic CAD implementation, so the agent only selects operations and fixes relations. The operation library, compiler and CAD kernel stay fixed when the agent backbone changes, so substituting a different language model needs no retraining and no change to the CAD system. Across five backbones from four model families, Structural Yield ranges from 77.2% to 87.4%, with Phi-4-14B highest at 87.4% and Granite 4.1 8B at 84.5%, and performance is not monotonic in parameter count.
Evaluation moves to checking mechanical structure directly on reimported STEP/B-rep models rather than only whether the program executes or the render looks right. Execution metrics verify only that an artifact is valid, rendered views can miss internal structural errors, and comparison with a single reference restricts the several valid constructions a requirement may admit. MHCAD's 342 structural checks span 196 prompts, and the checkers detect all 86 targeted topology mutations that keep the B-rep valid; on the independently constructed 30-case, 47-check CADPrompt subset, GOPR-CAD reaches 87.2% Structural Yield against 68.1% for the strongest baseline there, CADFusion.
Perspective
The demonstrated scope is single standard mechanical parts: MHCAD's 300 requirements come from Machinery's Handbook, and external validation uses 30 single-part cases and 47 checks selected from CADPrompt by checker applicability. Structural scoring covers only requirements supported by validated deterministic checks, so details such as exact gear-tooth counts fall outside it. The authors propose extending the operation graph to complex assemblies with mating constraints, motion relations and cross-part dependencies.
Structural Fidelity is conditional on valid B-reps, while Structural Yield counts missing or invalid outputs as zero, so the two metrics must be read together. GOPR-CAD's remaining errors concentrate on body realization and relation-sensitive placement, and the mounting-hole and circular-pattern groups in the CADPrompt subset contain only three checks each, so they should be read with their small denominators.
