Here's a writeup about a recent study that showed interviewers couldn't predict GPA any better when the interviewees answers were accurate versus random: http://www.danielwillingham.com/1/post/2013/10/why-job-inter...

Anyone who supports math puzzles (or whatever else) in an interview would have to argue that their perception of the candidates performance offers a clear enough data point that it doesn't dilute other information available to them. Given Google's study finding data otherwise, they certainly have the burden of proof.