A latent rollout is an imagined multi-step future inside a World Model. The model feeds its predicted latent state back into itself and uses that prediction to predict the following state.
One-step prediction
Teacher forcing gives the model the real state at each step:
real A + action 1 → predicted B
real B + action 2 → predicted C
real C + action 3 → predicted DThis trains accurate local predictions, but the model does not experience its own earlier mistakes.
Rollout prediction
A rollout uses earlier predictions as later inputs:
real A + action 1 → predicted B
predicted B + action 2 → predicted C
predicted C + action 3 → predicted DThis is the meaning of feeding the predictor output back into the input.
Why errors accumulate
If predicted B is slightly wrong, predicted C starts from the wrong state. The next prediction can add another error. This causes drift over long horizons.
V-JEPA 2 and GeoWorld both use teacher-forcing loss and a two-step rollout loss. GeoWorld also uses Geometric Reinforcement Learning to shape how errors propagate across a Hyperbolic Latent Space.
Horizon distinction
The rollout horizon is the number of future transitions that the world model imagines. An action-chunk horizon in SmolVLA is the number of actions produced by the policy in one inference pass. Both use future action sequences, but they serve different roles.
My way of understanding rollout
The rollout loss feeds the predictor's output back into the input. This helps the model learn multi-step future predictions instead of only learning one-step predictions.
This feels a bit like sequential models such as an RNN, GRU, or LSTM, where information from the previous step is fed into the next step. The important difference is that GeoWorld feeds a predicted latent state into an action-conditioned predictor. It does not use the same recurrent architecture as an RNN.
My simplest picture is:
real A + action 1 → predicted B
predicted B + action 2 → predicted C
predicted C + action 3 → predicted DIf predicted B is already slightly wrong, predicted C starts from that wrong representation. This is how the error starts to accumulate over a long horizon.