Researchers have proposed BrickBench, a new benchmark for what they call agentic text-conditioned LEGO-set design. The task asks an agent, given a text prompt, to produce an assembly that satisfies semantic and design criteria while remaining physically buildable.
Meeting that bar is not trivial. The agent must choose parts from a discrete library and reason jointly about local and global constraints, balancing what the prompt asks for against what the available pieces can actually support.
The benchmark scores designs on validity, alignment, and design across three settings that vary in scale and part availability. To support this, the authors also release BrickAgent, an environment in which coding agents can construct, inspect, and validate their own designs.
According to the authors, leading agents largely satisfy the verifiable physical and semantic requirements of the task, but their outputs still fall short of human designs. The benchmark and environment are released at http://www.brickben.ch.