The bleakest finding was that the more advanced that people’s math skills were, the more likely it was that their political views, whether liberal or conservative, made them less able to solve the math problem.

So, my first thought was that this might be a trivial statement. If any person is 30% likely to give the wrong answer for partisan reasons, but the chance of getting it wrong depends on innumeracy, then partisanship will be the dominant reason for error among highly numerate people, because it's basically the only source of error, even though total error is lower.

But turns out this is a rather meaningful statement (as shown by Figure 6 of the paper). Among liberals, numeracy has basically no impact on whether or not they correctly answered the political question when it disagreed with their priors (which is, frankly, horrifying). Conservatives who scored 8 or 9 (the max score) on numeracy were slightly better at giving the answer that went against their political position than less numerate conservatives, but only slightly. When the answer supported by the data supports the expected preconceived notion, both conservatives and liberals saw a positive relationship between numeracy and giving the correct answer. (Oddly, for conservatives the relationship looks linear, but for liberals it's flat-ish then sharply positive.)