Latent world models compress an environment into a learned latent space and then plan by searching for an action sequence whose predicted terminal latent is closest to the goal latent. The cost function for that search typically uses an ambient norm, such as Euclidean distance. That choice implicitly assumes the latent space is flat and that straight-line distances correspond to feasible transitions.
The paper argues this assumption is flawed. The latent space is a manifold, and the Euclidean distance between two points can cut through regions that are not reachable or meaningful. Planning with such a cost may therefore prefer action sequences that look good in the ambient space but do not actually lead to the goal.
Instead, the authors suggest that latent world models should plan with distances measured along the latent manifold itself. This is a conceptual correction rather than a new architecture, but it has direct implications for goal-conditioned planning and model-based reinforcement learning: the right notion of distance is the one that respects the geometry of the learned representation, not the one that is easiest to compute.