According to a new arXiv paper, LLM-based agents are becoming increasingly capable of generating complex 3D structures, raising hopes that they could reshape how objects are designed and realized in the physical world. However, the authors note that producing elegant geometry is fundamentally different from producing something that can actually be built and used.
To address this gap, the paper introduces LMBuild, a benchmark designed to evaluate LLM agents on generating buildable and functional structures. The work suggests that while current models may excel at geometric creativity, their outputs often fall short when judged against practical constraints.
The paper does not present results in the abstract, but the framing makes clear that the authors see buildability and functionality as essential criteria for any real-world application. The benchmark appears intended to push the field beyond purely aesthetic generation toward more physically grounded design.