Weighting our benchmark functions using profile data (as discussed in the conclusion of the blog post) should address this problem, at least partially.