Vision models can recognize a cat but may not actually see one.
Researchers tested ResNets and Vision Transformers on a deceptively simple binary task: classify images containing a closed square versus an identical square with one boundary pixel flipped. The classes are visually near-identical at the local level — which is the point. Using a framework called syntactic distance, the researchers measured whether two classes can be told apart by local statistical shortcuts. When syntactic distance is zero, those shortcuts fail, and a model must reason globally to succeed. Every architecture tested collapsed to random-guessing accuracy once image scale crossed a critical threshold. Bigger training sets and larger models only delayed the inevitable. Vision Transformers, despite their attention mechanisms, actually hit the wall sooner.
The finding matters because it exposes a structural ceiling, not a tuning problem. Scaling — the industry's default answer to benchmark gaps — cannot fix this. The researchers frame it bluntly: general visual intelligence may require models to build new representational systems rather than recycle the statistical patterns baked into existing ones. That is a longer road than another training run.
The benchmark leaderboards will keep improving, but this work is a reminder that high accuracy on standard datasets is not the same as understanding — a gap the field has debated for years without fully closing.