ArtisanCAD CAD agent: CAD-IR cuts Chamfer from 14.83 to 9.88

ArtisanCAD, an industrial-level CAD agent submitted to arXiv on 7 Jul 2026 and revised 8 Jul 2026, introduces CAD intermediate representation, or CAD-IR, to turn vague design intent into executable CAD procedures. The paper by Yunhan Xu and 12 coauthors shows CAD-IR reduces mean Chamfer Distance on the Text2CAD benchmark from 14.83 to 9.88.

What is CAD-IR and how does ArtisanCAD use it?

CAD-IR is an executable procedural representation that encodes parameters, ordered operations, MCP tool bindings, dependencies, generated entities, and verification rules. ArtisanCAD uses CAD-IR for two purposes: to carry distilled expert CAD procedures as reusable parameterized skills, and to provide a procedural scaffold that expands intermediate or ambiguous prompts into complete executable CAD operations. The agent retrieves expert-derived skills, instantiates and revises CAD-IR, and executes the resulting procedure through a dedicated CATIA-MCP backend.

Ailerons of the system include skill distillation from expert artifacts such as CATIA operation recordings, macro logs, drawing notes, and engineering descriptions. After execution the pipeline uses multi-view visual feedback for iterative refinement and produces production-ready boundary representation models, abbreviated in the paper as B-Rep.

How did CAD-IR perform on benchmarks and real components?

CAD-IR improved Text2CAD generation from intermediate prompts by reducing mean Chamfer Distance from 14.83 to 9.88, demonstrating measurable gains in converting ambiguous textual intent into geometric output. The paper also reports results on four complex automotive components, where CAD-IR enabled expert CATIA recordings to be distilled into reusable skills and allowed ArtisanCAD to generate editable CATIA-native B-Rep models for new variant requests.

The system executes distilled procedures through a CATIA-MCP backend and relies on multi-view visual feedback for iterative refinement, which the authors present as part of the end-to-end pipeline that reaches production-grade B-Rep execution. The evaluation therefore combines a quantitative benchmark improvement on Text2CAD with practical demonstration on automotive parts.

Why it matters

CAD workflows for industrial components require long-horizon procedural modeling, robust feature dependencies, editable parametric geometry, and production-grade B-Rep execution. By encoding expert operations into a reusable procedural representation, ArtisanCAD aims to bridge gaps when user prompts are ambiguous or underspecified, turning high-level design intent into executable CAD programs. The measurable Chamfer Distance reduction on Text2CAD shows a concrete improvement in geometry generation from intermediate prompts, while the four-component use case illustrates applicability to realistic automotive tasks.

This approach points to a path where recorded expert activity—operation logs, macros, and notes—become first-class training material for automation, rather than relying solely on large-language-model style end-to-end synthesis of geometry. For engineering teams that need editable, native CAD outputs, the paper emphasizes producing CATIA-native B-Rep models rather than only mesh approximations.

What to watch

Whether the CAD-IR approach scales beyond the four automotive components reported and whether independent evaluations replicate the Text2CAD Chamfer Distance improvements are the immediate signals to follow. Also watch for broader release or tooling around the CATIA-MCP backend, and for more examples of expert procedure distillation from varied industrial CAD artifacts.