i always upsample to at least 96k if not 192k
and 32-bit float
because the EQ plugins i use sound much more accurate and tighter in the whole spectrum, e.g. PEQ Orange, PLPAR EQ. but especially bass, if i want bass control with an EQ i have to upsample to 192k for the control i need in this region of the audio, no question about it.
also if i'm using any tube emulation or tape emulation, essentially anything nonlinear, it comes back sounding much better and with the intended effects, even after downsample, i don't know for sure, but i would guess it's because since there are more samples for the plugins to shape the way they do, it sounds smoother in the end.
even the compression plugs sound better and smoother in my own experience with it.
i'm using secret rabbit code, btw, for my conversions.
and sometimes Audacity, which uses the same sinc interpolation that secret rabbit code does.
so the SRC and the quality of it may also play a large factor with all of this.
but even before when i was using r8brain pro, still much better sounds with an upsample at the beginning of the process.
and to go further, it still sounded better to me way back when i was using wavelabs built in src plug, which tested poorly against the others.