Rollout-Decoded Reconstruction Improves Long-Horizon Latent World Model Predictions.

Latent world models are trained to predict future states by encoding observations into a latent space and then decoding these latents back into observations. A key challenge arises because the decoder is typically trained on latents directly linked to real observations, but during evaluation, it must operate on latents generated by the model's own free-running predictions, often hundreds of steps beyond the last actual observation. This discrepancy can lead to performance degradation over long prediction horizons. To bridge this gap, researchers propose Rollout-Decoded Reconstruction (RDR). RDR introduces a straightforward yet effective loss term that forces the model to train under conditions identical to its evaluation. During training, the model is allowed to free-run, generating its own sequence of latent states. Each of these rollout latents is then decoded, and the reconstruction error against the true ground truth is penalized. This method adds no new parameters to the model and only incurs computational cost during training. Experiments on the chaotic Kuramoto-Sivashinsky equation demonstrated a substantial improvement, nearly doubling the valid prediction time. This enhancement was consistent across multiple configurations, suggesting RDR is a robust and efficient way to boost the long-horizon predictive capabilities of latent world models.