If you were to do type inference across mutable reference assignments, you would have to do some semi unification. I don't think haskell does, even ML treats refs very strangely.