Driving-model preprint reports approximately 44 percent reduction in prediction error

Researchers report that a training method reduced a driving world model's 10-second final displacement error by approximately 44 percent against an unconstrained baseline when trained with 3-step rollout supervision. The preprint describes Directional Pullback Regularization, which penalizes internal model transitions that predict sideways vehicle motion while retaining the existing model architecture. The method lowered final displacement error against both comparison models across four training regimes. Models trained with 10-step rollouts delivered the strongest combined improvements in displacement error, lateral motion consistency and preservation of forward motion. The evaluation tested open-loop motion predictions over a 10-second horizon.