AI

LMBuild: Evaluating LLM Agents for Generating Buildable and Functional Structures

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 06/10/2026, 16:18:24 UTC.

Briefings

This article introduces LMBuild, a benchmark designed to evaluate Large Language Model (LLM) agents on generating physically buildable and functional 3D structures, moving beyond geometric quality to assess real-world operability. The study utilizes a unified framework with an interactive environment, a CAD dataset augmented with Wikipedia knowledge, and a four-dimensional evaluation covering structural soundness, functional affordance, design quality, and physical realization. Across 30 evaluated systems, the research identifies that while soundness and alignment are no longer primary bottlenecks for frontier closed-source models, functional affordance and physical operability remain significantly more challenging. Stronger models demonstrate superior component creation, whereas weaker models rely more on retrieval. Providing functional specifications substantially improves part completeness, kinematics, and physical operability. The findings suggest that generating real-world structures requires deeper reasoning about mechanics and novel component design.

Key points

• The LMBuild benchmark evaluates LLM agents on generating buildable and functional structures using a unified framework with interactive environments, CAD datasets augmented with Wikipedia knowledge, and a four-dimensional evaluation covering structural soundness, functional affordance, design quality, and physical realization.
• Across 30 evaluated systems, the study found that while soundness and alignment are no longer primary bottlenecks for frontier closed-source models, functional affordance and physical operability remain substantially more challenging for agents.

Context

Existing evaluations of LLM agents for 3D generation focus primarily on geometric quality, often overlooking the critical requirement that generated objects must be physically buildable and functional. This research addresses that gap by introducing LMBuild, a benchmark specifically designed to measure an agent's ability to create structures that can be assembled and perform intended functions in the physical world.

Limitations

The evidence is limited to findings from 30 evaluated systems across a single study period. Results are based on observed performance patterns rather than long-term forecasts or causal claims about future capabilities. The benchmark relies on existing CAD datasets and Wikipedia knowledge, which may not fully represent all real-world construction constraints or novel component designs.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 22/10/2026, 05:05:55 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeidnqgy6phhllc2tz4fyhnkw3gjapv52klobxidqzjxh52nfy7ejp4