The ML part of CS is in a sense funding replication: there are a decent number of papers whose only premise is "we compare 5 recent papers on this benchmark (and fill a couple pages with discussion)" or "we made a new benchmark, here's how commonly cited papers compare (plus a couple pages of discussion)"

Outside the high-profile cases it seems accepted norm that papers perform far worse when scored against somebody else's benchmark. The real measure of quality is how big the gap is.