Multi-Horizon Consistency Impacts Latent Dynamics Geometry in Video Predictors.

Multi-horizon latent consistency is a widely used training technique in video predictors and world models, yet its precise impact on the geometry of latent transitions is often unclear to practitioners. This research treats the lambda weight, which governs multi-step latent agreement, as a diagnostic control to measure its effects. The study quantifies an empirical expansion proxy, L20, and the horizon-20 prediction error, E20. On the Moving-MNIST dataset, increasing lambda from 0 to 0.8 significantly reduced L20, indicating a contraction of latent dynamics, and also halved E20, improving prediction accuracy. This suggests that soft consistency can effectively guide passive video models towards a near-contractive state. However, the same loss function did not produce similar population-level contraction (L<1) on action-conditioned environments like Pendulum-v1 or CartPole-v1, nor on KTH Actions video, even when prediction error improved. This domain-limited effect implies that while multi-horizon consistency can be powerful, its ability to induce contractive latent dynamics is not universal across all types of environments. The findings are supported by various checks, including architectural baselines and exogenous stress tests, reinforcing the claim that soft consistency can push passive video models towards a near-contractive band, but this band's existence and accessibility are dependent on the specific domain.