Origami is the simplest possible physical construction task. The action space contains a single primitive: the fold. The material is a flat sheet. The constraints are geometric and fully specified. If any physical reasoning benchmark should yield to scale, it is this one.
Agarwal, Wu, Jian, Hu, Mansoor, Li, Peng, Dai, Ding, and Sansone built OrigamiBench, an interactive environment where models propose folds and receive feedback on physical validity. The result: scaling model size does not reliably produce causal reasoning about physical transformations. Models fail to generate coherent multi-step folding strategies. Visual and language representations remain weakly integrated — the model can describe a fold and can recognize a folded shape, but cannot reliably connect the action to the outcome across a sequence of steps.
The failure is informative precisely because the task is so constrained. This is not open-ended manipulation, dexterous control, or contact-rich physics. It is planar geometry with a single operation. The model doesn't need to simulate friction, estimate mass, or plan around obstacles. It needs to track how a flat region maps to another flat region through a crease line, then compose these maps across steps.
The through-claim is about what scaling buys. More parameters improve pattern recognition — the ability to classify, describe, and generate plausible text about physical scenes. But pattern recognition is not causal reasoning. Recognizing that a folded shape looks like a crane is a different computation from determining which sequence of folds produces that shape from a flat sheet. The first is classification. The second is simulation. Scaling reliably improves the first. The evidence from OrigamiBench is that it does not reliably improve the second, even in the simplest geometric domain where simulation should be easiest.