Latent world models, particularly those based on the JEPA (Joint Embedding Predictive Architecture) framework, are increasingly used for model-predictive control (MPC). In these systems, the Euclidean distance between the current latent state and a goal latent is often used as the cost function to rank candidate action sequences. However, a strong decoding of task variables does not guarantee that this distance metric accurately reflects real task progress.
To address this gap, researchers introduce the concept of decision-metric alignment—the property that latent distances preserve the ranking of action sequences by actual task success. They propose two new diagnostics: Plan-Real Spearman, which measures rank agreement on random plans, and CEM-stage Spearman, which measures agreement as the cross-entropy method (CEM) search concentrates its proposals.
The analysis identifies three key factors that control alignment: encoder distortion, terminal rollout error, and candidate margins. Guided by these insights, the team develops DA-LeWM, which augments the existing LeWM model with inverse-dynamics and demonstration-conditioned goal-action heads. These action-conditioned objectives improve the latent geometry specifically for Euclidean-cost, CEM-based MPC.
In experiments, DA-LeWM consistently accelerates convergence and achieves higher online success rates than LeWM, while probe scores (a measure of representation quality) remain similar. This demonstrates that optimizing for decision-metric alignment can substantially improve planning performance without degrading the learned representations.