“Pathfinding algorithms like Dijkstra's and Bellman-Ford are non-invertible, as different input graphs can yield the same output, making them impossible to model with the symmetry groups used in geometric deep learning.”