Lisps pass some objects (lists, hash tables) by reference, but others (integers, symbols) by value. In languages that offer it like C++, explicit pass-by-reference is useful because A) these languages don't deal with objects as references by default and B) to pass extra information to the caller of the function (e.g. having a boolean reference that's set to true if some condition is met).
However, A is a non-issue in Lisps and B is trivially solvable by returning a pair, so I don't see how pass-by-reference could be useful.
Consider deleting a node from a binary tree. You want a function that looks at a node and, if it's a match, unlinks it from its parent. Passing by reference allows you to do this cleanly. The alternative is to peek ahead at each subnode (messy) or to pass information up the call stack (inefficient, as the function can no longer be tail-recursive).
I'm not advocating C++ references, which I think are too implicit. If a function call foo(x) can change the value of x, there needs to be some visual indication of this. I'd prefer something like plain C pointers, with which you can pass the address of an object: foo(&x).
Of course an alternative is store objects wrapped in containers, and this isn't too bad a solution.
A Lisp variable is implicitly a pointer to an object. Passing it to a function generates a copy of the pointer, which is pass-by-value. With pass-by-reference, the function would get a reference to the pointer. So e.g. reverse could be called for side effect: you could say (rev xs) instead of (= xs (rev xs)).
