Researchers have introduced LMBuild, a benchmark designed to evaluate LLM-based agents on generating 3D structures that can actually be built and that perform their intended functions — not merely objects that look good. The work argues that producing elegant geometry is fundamentally different from producing physically realizable objects, and that existing evaluations largely overlook physical realizability.
LMBuild represents generated objects as assembled structures made up of part decompositions, joints, materials, and assembly sequences. To support reproducible evaluation, the authors provide a unified framework with three components: an interactive environment where agents use tools to retrieve, create, and place components; a curated benchmark that repurposes established CAD datasets and augments them with knowledge from Wikipedia; and an evaluation framework covering structural soundness, functional affordance, design quality, and physical realization.
The benchmark includes LMBuild-Core, comprising 200 objects, and 12 hierarchical metrics organized into four groups: Soundness (connectivity, collision, stability), Affordance (functional geometry, parts, kinematics), Design (decomposition, aesthetics, alignment — with human-calibrated VLM judges), and Realization (assembly sequence, material, and simulation-based operability). The underlying data draws on CAD sources such as LEGO/BrickComposer, Fusion 360, Artiverse, PartNeXt, and open-source product CAD.
Evaluations across 30 systems — including frontier closed APIs, open-source (M)LLMs, and domain-specific generators such as BrickGPT, PartCrafter, PartPacker, Cube3D, and PhysX-Anything — surfaced three main findings. First, soundness and visual alignment are no longer the primary bottleneck for frontier closed-source models, while functional affordance and physical operability remain substantially harder; even the best systems score only about 20 on simulation-based operability. Second, stronger models tend to create new components, whereas weaker models rely on retrieval — and allowing part creation helps frontier models but can hurt weaker ones. Third, explicitly stating functional requirements substantially improves part completeness, kinematics, and physical operability.
The authors conclude that generating real-world structures requires deeper reasoning about functional affordances, mechanics, and the design of novel components, and they expect LMBuild to serve as a foundation for measuring progress toward agents that generate buildable and functional structures. Code, data, and evaluation are released as open source, with the dataset hosted on Hugging Face and the project available on GitHub.