Self-Improving CAD Generation Agents with Finite Element Analysis as Feedback

A new benchmark for computer-aided design generation shows that leading AI agents fail to produce a single structurally valid artifact on first attempt, according to research posted on arXiv. The study introduces finite element analysis as a feedback mechanism to align model output with engineering requirements. The work, authored by Guijin Son and submitted on 17 May 2026, proposes a task formulation that requires a model to generate a fully assembled multi-part STEP file from a free-form engineering brief [1]. The output is then validated through finite element analysis, a computational method used to predict how a product reacts to real-world forces, vibration, heat, and other physical effects [1]. When tested, agents built on Codex (GPT-5.5) and Claude Code (Opus-4.7) did not produce a single strict-passing artifact in the main first-attempt sweep [1]. The best configuration met only about 20% of typed requirements on average [1]. These results underscore a gap between current learned CAD generators and the iterative, evaluation-driven workflows common in industrial engineering pipelines [2]. To narrow that gap, the researchers introduced two additional supervision signals: a text-only blueprint schema and a 21-view image renderer that aids the agent's visual inspection [1]. These tools are designed to better align the generation loop with how engineers iterate in practice [2]. When applied to geometric reconstruction tasks on the S2O and Fusion360 datasets, the feedback tools yielded measurable improvements. The GPT-5.5/xhigh configuration rose from 0.444 to 0.592 Box-IoU on S2O and from 0.397 to 0.505 on Fusion360 [1]. Computer-aided design remains central to modern industrial design, yet prior work has typically treated CAD generation as two separate steps—part synthesis and assembly—with the former graded by proximity to a reference and the latter reduced to a constraint-solving problem [2]. The new formulation treats the entire process as a single, end-to-end task evaluated against physical and structural criteria, moving CAD programs toward artifacts that are not only visually plausible but also checked against engineering requirements [2].