The paper claims in its conclusion that it has provided evidence (what is more, "further evidence") yet I can't find any in there.

It argues that you can achieve a certain useful separation between programs together when one produces data for the other.

This can be achieved in a very satisfactory way with explicit streams (i.e. lazy lists). It can be satisfied with delimited closures, coroutines, threads and often with lexical closures. Not to mention Icon-style generators.

Lazy lists can be incorporated into the language so that their cell are first-class objects and substitute for regular eager cells smoothly. (Thank you, OOP).

The paper is actually wrong there, because laziness alone will not provide the kind of separation that g can begin executing, such that f then only executes when an item is required. Not for an arbitrary f! Suppose f traverses a graph structure recursively and yields some interesting items. Lazy eval alone isn't going to allow the f traversal to behave as a coroutine controlled by g, proceeding only as far as g continues to be interested in further items. The author is attributing to lazy evaluation magical powers that it doesn't have.