This article investigates an ML-like language with by-name semantics for polymorphism: polymorphic objects are not evaluated once for all at generalization time, but re-evaluated at each specialization. Unlike the standard ML semantics, the by-name semantics works well with polymorphic references and polymorphic continuations: the naive typing rules for references and for continuations are sound with respect to this semantics. Poly-morphism by name leads to a better integration of these imperative features into the ML type discipline. Practical experience shows that it retains most of the efficiency and predictability of polymorphism by value.