like LeWM are strong predictors but still rely on online trajectory optimization (e.g., CEM) that solves every state-goal pair from scratch, paying iterative cost each replanning step. LeFlow instead learns a reusable latent trajectory prior offline. A rectified-flow model generates a latent path between start and goal embeddings, an inverse dynamics decoder converts latent transitions to action chunks, and the frozen world model verifies candidates via autoregressive rollout. Across four goal-conditioned pixel-control benchmarks, LeFlow replaces iterative action-space search with fixed-budget generation and selection, achieving consistent success-rate gains and an order-of-magnitude reduction in planning time. This shows that latent world models can support not only prediction but also reusable planning priors, amortizing planning effort across queries.
Reusable latent trajectories improve world-model control while cutting planning time
LeFlow generates plans in a world model’s latent space, decodes them into actions, and verifies candidates through simulated rollouts.
Top University
Hsiang-Wei Huang · Jianxu Shangguan · Junbin Lu · Jenq-Neng Hwang
University of Washington
Research Digest··2 min read
In one line
LeFlow replaces iterative action-space planning with amortized latent trajectory generation, cutting planning time by an order of magnitude while improving success rates.
Before you read this
Three ideas this paper builds onlatent world modelA learned internal, compressed representation of environment dynamics, enabling future-aware planning, generation, and reasoning directly in latent space.44 papers →amortized planningA learned model generates plans directly for new instances, replacing expensive per-instance search with a fast, generalizable forward pass.9 papers →rollout verificationVerification of generated sequences to detect and correct errors that accumulate during long-horizon autoregressive generation.22 papers →
Read it as
Huang et al. replace per-query iterative action optimization with a learned trajectory prior conditioned on current and goal images.
The authors trained a rectified-flow model to generate latent trajectories between current-state and goal embeddings produced by a learned world model.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.
What it released
Code