TraceCAD: Trace-Guided Repair for Agentic CAD Generation

2026-08-03

AI Summary

This work addresses the degradation in generation quality observed in large language model (LLM)-driven CAD agents during iterative repair, primarily caused by the loss of user requirements, operation history, and failure evidence. To mitigate this, the authors propose a persistent tracking mechanism that jointly models user intent, modeling steps, failure evidence, and candidate outputs. The approach enables reliable repair through diagnosing faulty operations, performing bounded edit searches within their local dependency regions, and integrating execution validation with retention checks. For the first time, the method introduces persistent state tracking, localized dependency-aware editing, and reusable skill memory, significantly improving both repair success rates and geometric fidelity. Evaluated on DeepCAD, the system achieves state-of-the-art performance: ablating persistent state reduces repair success by nearly 50%, while removing local search doubles geometric error and API call count; pre-populating the skill library effectively lowers retry frequency, token consumption, and latency.

Abstract

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.

Problem

• CAD generation
• repair reliability
• persistent state
• fault diagnosis
• geometric quality

Innovation

• persistent state
• localized repair
• skill memory
• trace-guided diagnosis
• agentic CAD generation