RA-CAD Improves Text-to-CAD with Learnable Post-Execution Critique.

Generating Computer-Aided Design (CAD) models from natural language descriptions can significantly lower the barrier to entry for design, but current text-to-CAD methods often struggle with effectively utilizing feedback for correction. Existing systems typically rely on fixed or externally supplied critique mechanisms, which don't always optimize the translation of feedback into actionable revisions during the generation process. Researchers have introduced RA-CAD (ReAct Agent for CAD), a state-aware agent designed to bridge this feedback-utilization gap. RA-CAD operates within a Generate-Execute-Critique-Rewrite loop, where it executes the generated CAD code, observes the outcome, and then generates an explicit post-execution critique. This critique, conditioned on the design instruction, current code, and execution feedback, either validates the result for termination or provides specific guidance for the next rewrite. The system employs two key training phases: CAD Code Bootstrapping (CCB) for foundational parametric CAD coding, and Feedback-Driven Agent Optimization (FAO). FAO uses trajectory-level Group Relative Policy Optimization to train both code and critique sequences, making the critique an outcome-aligned, learnable policy decision rather Pre-execution critique is an outcome-aligned, learnable policy decision rather than a separate output. This approach has led to state-of-the-art execution validity and geometric quality on benchmarks like CADFusion and Text2CAD, outperforming existing methods and even strong proprietary language models.