The major difference is that the tools compose with the language better. Instead of having to keep "I'm working in a multicore environment" in your head all the time when you're writing your client code (as you would in Java, worrying about concurrent mutation and so on), Haskell enforces immutability throughout the language. Often you get the extremely convenient situation where you observe that something could be parallelized, change map to pmap and you're now using all your cores.
Compare Guy Steele's talk "Get rid of Cons" at the International Conference on Functional Programming 2009. For the slides go to http://news.ycombinator.com/item?id=814632, the video should also be available online somewhere.
Haskell does keep track of mutations in your program, because you have to wrap those in a monad. Since monads are part of the type system, you are right--but I felt that needed more explanation.
How is Haskell better at auto-parallelizing an algorithm compared to say, java's fork/join or .NET's equivalent?