In reply to giulio95:
This question is the Holy Grail of Functional Coverage. If you were using assertions, a formal tool can trace through the design and exhaustively prove that the assertion is valid or not. But given a testbench, it can be next to impossible to figure out what stimulus needs to change to hit a functional coverage point.
Some are experimenting with machine learning algorithms to see if there is anything that could be done to the constraints to improve coverage, but I would say that is mostly in the academic and design specific cases.
Some things that I would suggest is looking at your coverage models and making sure you are not over-specifying your bins(like trying to cover an entire address space) and looking at formal analysis in certain areas to achieve functional coverage rather than dynamic simulation.