Tag
This paper investigates how much structure a task needs from a world model, showing that the objective's dimensionality determines how many predictive directions the model installs, with the common scalar reward objective being only the rank-one corner of value equivalence.