aiqbench analysis
This diagnostic study reveals that multi-horizon latent consistency acts as a specific control knob capable of constraining the expansion geometry in passive video world models, effectively pushing them toward a near-contractive band. However, researchers should note this effect is highly domain-limited and does not automatically transfer to action-conditioned environments like Pendulum or CartPole despite improvements in prediction error. The findings provide a stochastic-forcing law that unifies control domains but explicitly caution against assuming universal applicability across different world-model architectures without further validation.
Multi-Horizon Consistency as Geometry: When Latent Dynamics Contract, and When They Do Not
arXiv:2607.21645v1 Announce Type: cross Abstract: Multi-horizon latent consistency is a common training knob in video predictors and world models, but practitioners rarely know what it does to transition geometry. We treat lambda, the weight on multi-step latent agreement, as a diagnostic control and measure an empirical expansion proxy L20,q95 together with horizon-20 prediction error E20. On Moving-MNIST (n=6 seeds at the critical pair), raising lambda from 0 to 0.8 cuts L20 from 4.96 +/- 2.01 to 1.01 +/- 0.06 (paired t p=0.005, Wilcoxon p=0.031) and halves E20 (0.365 to 0.177, paired t p=1.1e-13). Four of six seeds cross L<1 at lambda=0.8. The same loss does not produce population L<1 on action-conditioned Pendulum-v1 or CartPole-v1, nor on KTH Actions video, even when E20 improves. An associational mediation analysis on MMNIST gives r-hat=0.94 (95% CI [0.88, 1.00], n=27, B=2000); lambda was not randomized. Defensive checks (architectural baselines, exogenous stress, WorldTest, MPC, scaling) mostly support a narrow claim: soft consistency can push passive video toward a near-contractive band, and that band is domain-limited. A stochastic-forcing law L20 ~ 1.23 + 1.82 eta at lambda=0.8 (bootstrap slope CI [1.73, 1.92], R^2=0.96) unifies control domains on the same curve via calibrated eta_eff. Complete joint slices at lambda in {0.4, 1.2} (30/30 cells, 5 eta x 3 seeds) show comparable linear L20(eta) slopes (~1.69 and ~2.00); we do not fit a continuous (lambda, eta) surface. We do not report DreamerV3 or TD-MPC2 returns.
Abstract:Multi-horizon latent consistency is a common training knob in video predictors and world models, but practitioners rarely know what it does to transition geometry. We treat lambda, the weight on multi-step latent agreement, as a diagnostic control and measure an empirical expansion proxy L20,q95 together with horizon-20 prediction error E20. On Moving-MNIST (n=6 seeds at the critical pair), raising lambda from 0 to 0.8 cuts L20 from 4.96 +/- 2.01 to 1.01 +/- 0.06 (paired t p=0.005, Wilcoxon p=0.031) and halves E20 (0.365 to 0.177, paired t p=1.1e-13). Four of six seeds cross L<1 at lambda=0.8. The same loss does not produce population L<1 on action-conditioned Pendulum-v1 or CartPole-v1, nor on KTH Actions video, even when E20 improves. An associational mediation analysis on MMNIST gives r-hat=0.94 (95% CI [0.88, 1.00], n=27, B=2000); lambda was not randomized. Defensive checks (architectural baselines, exogenous stress, WorldTest, MPC, scaling) mostly support a narrow claim: soft consistency can push passive video toward a near-contractive band, and that band is domain-limited. A stochastic-forcing law L20 ~ 1.23 + 1.82 eta at lambda=0.8 (bootstrap slope CI [1.73, 1.92], R^2=0.96) unifies control domains on the same curve via calibrated eta_eff. Complete joint slices at lambda in {0.4, 1.2} (30/30 cells, 5 eta x 3 seeds) show comparable linear L20(eta) slopes (~1.69 and ~2.00); we do not fit a continuous (lambda, eta) surface. We do not report DreamerV3 or TD-MPC2 returns.
22 pages, 9 figures. Diagnostic study of multi-horizon latent consistency, expansion proxies (L20), and a stochastic-forcing law for world-model geometry. Code and experiment logs to be released publicly
Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
I.2.6; I.2.10; I.4.8
arXiv:2607.21645 [cs.LG]
(or arXiv:2607.21645v1 [cs.LG] for this version)
Full text under the publisher's open licence · 277 words
Related coverage
4DGS-JEPA: Temporally Compositional Joint-Embedding Prediction for Dynamic Gaussian Splatting
arXiv · Sep 24, 2026
Automating coherent long-form video generation
Google Research · Sep 24, 2026
Depth-Regularized JEPA World Models Learn More Transferable Representations from Real Outdoor Robot Data
arXiv · Jul 21, 2026
Marionette: Predicting World States, Rendering Geometry, Painting Appearance
arXiv · Aug 17, 2026