Tag
This paper investigates how transformer language models implement multiple mental spaces (counterfactual, belief, fictional, etc.) using a shared router mechanism over a value slot, showing that a single low-rank subspace controls all space types and drives inference.