LLM Agents Create Reliable CAD Models with Solver Feedback.

While large language models (LLMs) can generate plausible CAD scripts, industrial-grade computer-aided design (CAD) requires more than just syntactically correct code; every geometric feature and assembly relation must be precisely accepted by an exact geometric kernel and remain parametrically editable. This research presents "Embodied CAD," a system of LLM agents designed for parametric B-Rep assembly modeling that addresses these challenges. Instead of generating a complete script in one go, these agents operate iteratively. The agents select actions from a structured CAD skill library, translate them into geometric operations, and execute them within a CAD backend. Crucially, they use the solver's feedback to refine their planning, correct errors, and learn from experience. This framework integrates action grammar constraints, deterministic parameter resolution, and solver-derived rewards for both supervised initial training and reinforcement learning refinement. Evaluated on complex mechanical assembly tasks, Embodied CAD demonstrates high executable rates and significant progress towards reliable, long-horizon policy prediction for CAD modeling.