That's because many of the tests were redesigned in the 1990s to eliminate their correlation with g, the general intelligence factor. This was explicitly done to eliminate the gaps between sexes and races.
Girls have been closing the gap in the study of mathematically precocious youth, and the international science Olympiads, too. The SAT and the IMO et. al. have been re-engineered? News to me.
How the heck do you make a math test with the g correlation eliminated, anyway? That seems pretty hard.
I'm not entirely sure if the g-loading changes were made in the math portion, or just the verbal. As to how to decorrelate math, I suspect it canbe done by simply reducing the number of chained dependencies. For example, have one question involving sin(x) and another involving x^2, instead of one question involving sin(x^2).
If you eliminated g loading of a mathematical test, you'd need to be able to take a developmentally challenged person, and expect, statistically, for them to get the same score as a fields medalist. I just don't see how you're going to swing that.
RE: The Mensa site. They claim that the GRE and the SAT and the ACT no longer correlate with IQ. Or g. That's just wrong. The old SAT correlates positively with the new SAT. The old SAT correlates positively with g. You'd have to have an extremely unusual statistical correlation between the old SAT and some other variables to yield 0 correlation with g.
They could be concerned that it's less well correlated. Or that preparatory courses from Kaplan, et. al. were killing the correlation.
But in all seriousness, who's keeping count. There are pretty severe conceptual problems with g anyway. I mean, all it is is the principal factor in a kind of multivariate linear regression analysis of a bunch of scores on tests, which are arbitrarily selected and claimed to represent 'intelligence.' You can't expect that linear statistical relationships mapping a few scores from a few tests chosen ad hoc really capture the breadth and diversity and range of human intelligence. It's crazy! If we look for some linear relationship, we'll find it. But it can't possibly tell the whole story.
For a math test, reducing g loading means having simple problems that rely on previously-learned information, rather than multi-step problems and problems that require invention. This is trivial to arrange.
Obviously this will not make a dull person able to pass the test, but we're talking about college prep tests, where the competition is among the people on the top part of the g bell curve.
The MENSA issue is one of grossly insufficient correlation, since the organization is based on intelligence, not educational history.
g is the psychometric measure most highly correlated with practical outcomes like annual income and mortality risk. I suspect the second most correlated measure is attentional control (focus versus ADD), which is the fuel for the brain's g engine. (Time preference likely has more practical importance than either, but we don't yet know how to measure it.)