What's this about?
AnisoWM learns different variance along latent dimensions so Euclidean distance weights planning errors in a way that better matches task outcomes.
What exactly is anisotropic geometry?
Different directions count differently—like one metre uphill costing more effort than one metre across flat ground—even when ordinary distance treats them equally.
Why does this matter?
5 HF upvotes. The change leaves the predictor and planner intact yet improves success in four visual-control environments, isolating representation geometry as a planning variable.