Researchers propose DA-WAM, a framework that combines predictive representation learning and decision-making for safe autonomous driving. DA-WAM aims to improve world models by directly linking predicted future states to the selection of trajectories. The framework uses an online encoder and momentum target to maintain predictive supervision during planner optimization. An action-conditioned predictor generates distinct future latent states per trajectory candidate, which are then evaluated by a factorized scorer. Experiments on NAVSIM-v1 and NAVSIM-v2 demonstrate state-of-the-art performance. --- Why it matters: This matters because it addresses the challenge of ensuring that predicted futures in world models inform decision-making for autonomous driving. DA-WAM's ability to link future states to trajectory selection could improve safety and efficiency in self-driving vehicles. Source: https://arxiv.org/abs/2608.19085