A new benchmark called BrickBench grades AI agents on designing LEGO sets that have to actually hold together, not just look good.
Researchers behind BrickBench give an agent a text prompt and task it with assembling a LEGO set from a discrete library of parts. The agent has to reason about local constraints, like whether two pieces actually click together, and global ones, like whether the whole structure stands up on its own. Scoring covers three criteria, physical validity, how well the build matches the prompt, and overall design quality, tested across three settings that vary set size and the number of available parts. The team also built BrickAgent, a companion environment where coding agents can construct, inspect, and test their designs before submitting them.
This is a clean stress test for whether language models can reason about real-world physical constraints, not just generate plausible-looking text or images. That distinction matters because physical common sense, whether something actually fits or actually stands, is a different skill than predicting the next word, and most AI benchmarks never touch it.
The researchers found that leading agents mostly satisfy the verifiable physical and semantic requirements, but still fall short of human designers, a reminder that even a kids' toy can expose the gap between pattern matching and real-world reasoning.