Back to feed
arXiv cs.LG
arXiv cs.LG
7/27/2026
Multi-Horizon Consistency as Geometry: When Latent Dynamics Contract, and When They Do Not

Multi-Horizon Consistency as Geometry: When Latent Dynamics Contract, and When They Do Not

Short summary

This paper treats the multi-horizon latent consistency weight (lambda) as a diagnostic control and measures its effect on transition geometry in video predictors and world models. On Moving-MNIST, raising lambda from 0 to 0.8 cuts expansion proxy L20 from 4.96 to 1.01 and halves prediction error, pushing passive video toward near-contractive dynamics. However, the same loss does not produce contraction on action-conditioned Pendulum, CartPole, or KTH Actions, and a stochastic-forcing law unifies domains via calibrated noise levels.

  • •Multi-horizon consistency loss pushes passive video toward near-contractive latent dynamics on Moving-MNIST
  • •Contraction effect is domain-limited — not observed on control tasks or KTH Actions
  • •Stochastic-forcing law L20 ~ 1.23 + 1.82*eta unifies control and video domains on one curve

Generated with AI, which can make mistakes.

#research-breakthrough
Read full article at arXiv cs.LG

Is this a good recommendation for you?

Comments

Failed to load comments. Please try again.

Explore more

When Do Conservation Laws Survive Learned Representations? Certified Horizons for Latent World Models
arXiv cs.LG
Interpreting Latent CoT Reasoning as Dynamical Systems
arXiv CS.AI
Grounded Iterative Language Planning: How Parameterized World Models Reduce Hallucination Propagation in LLM Agents
arXiv CS.AI
AI/ML Research Digest — Jun 27, 2026
Dev.to