Decision-Metric Alignment in Latent World Models: Diagnostics and Action-Conditioned Objectives for MPC Planning
The paper argues that latent-world-model planning can fail even when the latent state decodes task variables well.
The authors name the missing property “decision-metric alignment”: whether Euclidean distance in latent space ranks action plans the same way real task progress does. They introduce two Spearman diagnostics to test that rank agreement before and during CEM search. Their DA-LeWM adds inverse-dynamics and demonstration-conditioned goal-action objectives, improving convergence and online success over LeWM while leaving probe scores similar. HF Daily Papers' note
The authors name the missing property “decision-metric alignment”: whether Euclidean distance in latent space ranks action plans the same way real task progress does. They introduce two Spearman diagnostics to test that rank agreement before and during CEM search. Their DA-LeWM adds inverse-dynamics and demonstration-conditioned goal-action objectives, improving convergence and online success over LeWM while leaving probe scores similar. HF Daily Papers' note
score 4