Latent world models let agents plan by searching for action sequences that minimise a cost, typically the Euclidean distance between the predicted final latent and a goal latent. But the latent representation is not entirely signal: the regulariser used to prevent collapse during training also encourages variance in directions that carry no useful information, so the planning cost can be swayed by noise.
The paper shows that the same regulariser provides the structure needed to prune that noise. If the latent sequence is treated as a Gauss-Markov process, the transition is linear and its covariance is determined by the fitted dynamics matrix A. Constraining A to be a normal matrix keeps the eigenbasis fixed across timesteps, giving a stable coordinate system in which each latent direction has its own uncertainty.
The correction is post hoc: the encoder, predictor, rollout, and planner are left untouched, and only A is fit. The l2 cost is then restricted to the k-dimensional subspace that the dynamics carry forward with the least uncertainty.
Across six environments, the largest gains appear in the 2D navigation tasks TwoRoom and PointMaze, where the method raises success rate over the plain l2 cost by up to 54.7 points under a harder planning setting.