A new paper on arXiv proposes BrickBench, a benchmark for evaluating AI agents that design LEGO sets from text prompts. The task requires an agent to produce an assembly that satisfies the prompt's semantic and design requirements while also being physically viable.
BrickBench appears aimed at pushing agentic systems beyond simple text-to-image or text-to-3D generation, into a domain where the output must be a constructible, physical object. By focusing on LEGO, the benchmark offers a constrained but rich space for testing planning, spatial reasoning, and constraint satisfaction.
The abstract is brief, so details on evaluation metrics and baselines are not yet available. The authors position BrickBench as a step toward more robust evaluation of text-conditioned assembly agents.