Does the fuzzer generate just random input or does it adapt the input to achieve maximum test coverage (e.g. using constraint solver)?
The fuzzer generates input according to coverage-guided algorithm which proved to be very successful in real-world software. Originaly described in AFL white-paper http://lcamtuf.coredump.cx/afl/technical_details.txt