Many math questions are easy to verify, and it's a classic benchmark for reasoning -> so it's a good hill to climb.

I agree with your meta-point that better benchmarks testing more types of task would be good!