Less spiky workload, especially noticeable if the server weight is high, e.g. we had weights like 100 and 137 on servers since two-digit weights weren't giving sufficient precision to achieve equal load between different hardware generations.
Since the change reduces the peak load you get lower 99% latencies and thus need fewer machines to achieve the same 99% latency on servers
I generally use leastconn with haproxy (obvious why this isn't ideal, but it was better than pure RR) as some our posts either take 10ms or 5500ms