LLM-driven text-to-CAD now claims true controllability.
The paper introduces a pipeline that feeds natural‑language prompts to a large language model, which then produces a structured script for a CAD engine. The system includes a feedback loop that checks the generated geometry against explicit constraints supplied by the user, iterating until the design satisfies them. Experiments on a benchmark of 500 prompt‑shape pairs show a 23% improvement in fidelity over prior text‑to‑CAD approaches.
If the claim holds, designers could skip manual sketching for routine parts, speeding up early‑stage prototyping. The controllable aspect also reduces the trial‑and‑error typical of generative CAD, making the output more reliable for downstream simulation.
The work remains academic, with no commercial product announced, and it still relies on heavyweight LLMs that may be costly to run at scale.