State-of-the-art AI agents cannot figure out how to stack blocks — and a new benchmark exists specifically to rub that in.
Researchers introduced BuilderBench, a benchmark pairing a physics simulator with a suite of over 50 target structures. Agents must learn to build those structures by interacting with the environment across multiple episodes — not by recalling training data, but by genuinely experimenting. The tasks test physics intuition, mathematical reasoning, and long-horizon planning. When the team ran both frontier language-model-based agents and tabula rasa reinforcement learning algorithms through the suite, none of them solved any non-trivial task.
The failure reveals something the field mostly talks around: current AI learns by mimicry, not by doing. When a problem requires "embodied reasoning" — the kind expressed through action and iteration rather than words — models trained on text hit a wall. BuilderBench is designed to make that wall visible and measurable, giving researchers a concrete target for training agents that can actually explore and adapt.
For all the breathless coverage of AI agents booking flights and writing code, this result is a useful corrective: give a frontier model a pile of blocks and a goal, and it is no more capable than a tabula rasa algorithm starting from zero.