What they did

The authors trained a rectified-flow model to generate latent trajectories between current-state and goal embeddings produced by a learned world model. An inverse-dynamics decoder converts transitions along those latent paths into chunks of executable actions.

Rather than executing generated plans directly, LeFlow evaluates a fixed number of candidates through autoregressive rollouts in the frozen world model and selects among them. The method was tested on four goal-conditioned control benchmarks using image observations.

Key findings

• LeFlow consistently improved success rates across all four evaluated benchmarks relative to iterative action-space planning baselines.
• Amortizing planning through a reusable latent trajectory prior reduced planning time by approximately an order of magnitude.
• The approach avoided restarting an iterative optimizer for every state-goal pair, while retaining model-based verification through simulated rollouts.
• Planning operated directly in the world model’s learned representation rather than treating the model solely as a black-box action-sequence simulator.

Why it matters

The results suggest that world models can learn reusable structure for planning, not just predict the consequences of proposed actions. This could make visual model-based control more practical when agents must replan frequently under tight latency constraints.

Caveats

The reported evidence covers four goal-conditioned pixel-control benchmarks; generalization to substantially different environments, longer horizons, real robots, or goals outside the training distribution remains open. LeFlow also depends on the accuracy of the frozen world model and inverse-dynamics decoder, whose errors may compound during rollout.