For averaging multiple different benchmarks, the geometric mean makes sense. For removing measurement noise from a single benchmark, min or median is usually the right statistic.