StepCAD: Mesh-to-CAD Code Generation via LLM Policy and Geometry-Guided Search

StepCAD. Mesh-to-CAD Code Generation via LLM Policy and Geometry-Guided Search. By Ghadi Nehme and one other, from the Massachusetts Institute of Technology. Published on the first of October. Its code repository is a placeholder: a readme, and nothing to run yet.

Turning a 3D mesh back into an executable CAD program is hard for two reasons at once. CAD construction is compositional, built up operation by operation, and every step mixes discrete modelling choices with continuous parameters. Many learning-based methods predict the whole program in a single pass, and lean mostly on sketch-extrude representations. That limits how varied the operations can be, and it leaves little opportunity to correct geometric errors while the reconstruction is under way.

StepCAD treats reconstruction as generative optimization. It pairs a CAD policy that is conditioned on state with a search that is guided by geometry. The policy proposes; the search checks the shape and refines.

Given an input mesh, the policy predicts construction actions conditioned on two things: the target geometry, and the intermediate geometry built so far. An IoU-guided tree search then refines the resulting program through local edits. Alongside the method, the authors introduce a large dataset, ARCADE-1.5M. It holds one point five million executable CAD programs spanning diverse operations, sequences reaching more than a hundred and fifty counted operations, and twelve point five million intermediate state-action transitions.

Across several CAD reconstruction benchmarks, StepCAD reaches state-of-the-art geometric reconstruction accuracy, with consistently high validity. Its largest margin is a relative IoU improvement of up to eighty-seven point two percent over the strongest evaluated baseline, and the gains are particularly large on complex shapes.

Building a CAD model step by step, while looking at what has been built, gives the method something a single-pass predictor lacks: a chance to notice where the geometry has gone wrong and edit it. The dataset, with its millions of recorded intermediate states, is a contribution in its own right, and the reported gains are at their largest on the complex shapes.