When they say "improved in XYZ", what does that mean? "Improved" on synthetic benchmarks is guaranteed to translate to the rest of the problem space? If not tha

When they say "improved in XYZ", what does that mean? "Improved" on synthetic benchmarks is guaranteed to translate to the rest of the problem space? If not that, is there any guarantees of no regressions?