The big picture

The results suggest leading vision models recognize patterns rather than reason about scenes. Decoys reveal a gap that standard benchmarks miss.