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

Researchers introduced an industry-style CAD generation task that produces assembled multi-part STEP files from engineering briefs.

They added a text blueprint schema and a 21-view renderer as supervision signals and feedback tools.

Finite element analysis was used to validate designs against structural requirements, not just visual similarity.

The added signals improved reconstruction performance on S2O and Fusion360.

Results also showed that current frontier agents like GPT-5.5 and Claude Code still struggle with first-pass FEA success.