How to Trust AI Benchmarks in Software Development?

Benchmarks can often be misleading, as demonstrated by the varying scores of Muse Spark in different tests. How do you determine the reliability of benchmarks in your projects? What criteria do you use to assess their validity?