Better loss improvement

I have changed the API of loss_improvement to accept only a point instead of a list of points because we won't ever use it in any other way in the current implementation.