“Due to rapid model improvements, evaluation benchmarks often become saturated and need to be replaced after approximately one to three model generations.”