Search papers, labs, and topics across Lattice.
This paper introduces TraceCAD, a recovery layer designed to enhance the performance of LLM-based CAD agents by maintaining a persistent state that links features, modeling steps, and repair outcomes. By diagnosing faulty operations and executing localized searches for bounded edits, TraceCAD significantly improves the geometric quality of generated CAD programs while reducing retries and latency. The empirical results show that the removal of persistent state and localized search drastically decreases recovery scores and increases geometric regression, underscoring the importance of these mechanisms in achieving reliable CAD generation.
Persistent state and localized search in TraceCAD can dramatically enhance CAD generation quality and reliability, revealing a critical dependency in agentic design.
LLM-based CAD agents produce executable parametric programs, but their correction loops may lose evidence about satisfied requirements, faulty operations, and prior repairs. We introduce TraceCAD, a recovery layer that links requested features, modeling steps, failure evidence, and candidate outcomes as persistent state. TraceCAD diagnoses likely faulty operations, searches bounded edits in their dependency regions, validates candidates through execution and preservation checks, and retains successful and failed repair outcomes in reusable skill memory. On DeepCAD-derived benchmarks with 200-model ablations and a 1K-model comparison, TraceCAD achieves competitive geometric quality in terms of IoU, Chamfer distance, and Hausdorff distance. Removing persistent state nearly halves recovery score; removing localized search more than doubles geometric regression and doubles code-agent invocations. Initializing the skill store on disjoint training models further reduces retries, token cost, and latency. These results demonstrate that persistent, localized, and reusable recovery improves final CAD quality and repair reliability.
