Actionable Intelligence: Transforming the predictor into a recurrent mental simulator for zero-shot planning.
Until 2026, JEPAs were mostly used for single-step prediction. **LeWorldModel (LeWM)** shifts the architecture toward Recurrent Rollouts. The predictor is applied repeatedly to its own output, allowing the agent to "imagine" the future 20, 50, or 100 steps ahead in latent space.
In this mode, the predictor becomes a function of both state and action: $P(s_t, a_t) \to s_{t+1}$. By evaluating thousands of action sequences in parallel, the agent can pick the path that minimizes cost before taking a single physical step.
Use the Joystick to set a constant action (thrust). The green particles represent the model's Mental Rollout. Notice how the uncertainty expands over time—this is the model acknowledging multiple possible outcomes based on environmental noise.
LeWorldModel doesn't just guess; it optimizes. It uses the Cross-Entropy Method (CEM) to sample thousands of "Imagined Rollouts," selects the top 10% with the lowest cost, and refines its action sequence. This process happens in milliseconds, allowing for fluid, brain-like motor control.
Interested in how other AI architectures handle sequence generation? Explore the Learning LLM section to compare World Models with Autoregressive Transformers.