The score has gone down a very little bit over each of the middle releases and escaped notice.
It does not seem to be related to GC performance. There are big differences in the compiled code of the hot method
spec.benchmarks.compress.Compress::performAction

There are much more instructions in 18 than 9
9:
Compress.compress:instructions:u thrpt 6122454411.511 #/op
18-b24:
Compress.compress:instructions:u thrpt 6749065336.960 #/op