Researchers achieve 6x improvement in visual hierarchy detection by modeling objects as spacetime trajectories using Lorentzian geometry instead of Euclidean space

Worldline Slot Attention treats objects as persistent trajectories through spacetime rather than independent points, enabling hierarchical part-to-whole relationships

Lorentzian worldlines achieved 0.479-0.661 level accuracy across three datasets, dramatically outperforming Euclidean worldlines which scored 0.078 (below random chance of 0.33)

Results replicated consistently over 20+ independent runs, demonstrating that visual hierarchies require causal/temporal structure rather than tree-like radial branching

Lorentzian geometry outperformed hyperbolic embeddings, suggesting causal structure is more fundamental than traditional hierarchical embeddings for object discovery