AI-generated CAD just got a revision loop.
Researchers introduced IterCAD, a multimodal agent framework that treats computer-aided design as a multi-turn conversation rather than a single prompt-and-pray generation step. The system works inside an executable CAD sandbox, handling three task types: converting drawings to code, generating code from text descriptions, and editing designs interactively. To train it, the team built a data pipeline that produces multi-view engineering drawings, code-editing examples, and recorded interaction sequences that mirror how engineers actually work. The agent was then optimized using supervised fine-tuning followed by geometry-aware reinforcement learning — a technique that penalizes outputs failing to produce valid, executable geometry.
The gap this targets is real. Most AI CAD tools today work in open-loop fashion: generate once, hope for the best, start over if it's wrong. That bears no resemblance to how engineers use actual CAD software, where designs go through dozens of revision cycles. A closed-loop agent that can receive feedback and adjust — without losing geometric coherence — is a more honest fit for industrial workflows. The researchers also introduce a new benchmark metric, the Chamfer Distance Tolerance-Recall curve, designed to avoid the survivor bias baked into metrics that only grade outputs that successfully compile.
IterCAD outperforms prior approaches on both code executability and geometric precision by its own benchmarks, which is worth noting — self-reported comparisons on self-designed metrics are a reason to read the eventual peer review carefully before updating your priors.