Specialization after generalization: towards understanding test-time training in foundation models