TraceCAD: Trace-Guided Repair for Agentic CAD Generation

TraceCAD is a new recovery layer for agentic CAD generation that tries to stop LLM-based modeling systems from forgetting what they already got right. Instead of treating each correction pass as a clean slate, it keeps persistent state linking requested features, modeling steps, failure evidence, and candidate outcomes.

The system then diagnoses likely bad operations, searches for bounded edits in the dependency region around those operations, and validates candidates with execution plus preservation checks. Successful and failed repairs are stored in a reusable skill memory, so later models can benefit from earlier recovery attempts rather than repeating them.

The numbers are the part developers should care about. On DeepCAD-derived benchmarks, TraceCAD was tested with 200-model ablations and a 1K-model comparison, and it delivered competitive IoU, Chamfer distance, and Hausdorff distance. Removing persistent state nearly halved recovery score, while removing localized search more than doubled geometric regression and doubled code-agent invocations.

For teams building CAD agents or other code-generating tools, the practical takeaway is that repair quality depends on memory, locality, and reuse, not just raw model capability. Initializing the skill store on disjoint training models also reduced retries, token cost, and latency, which makes the approach interesting for production pipelines where every failed regeneration has a real compute bill.