There isn't a single paper that says to use random starting points when we have multiple kangaroos per herd; on the contrary, they start from well-defined positions. To use random start points is an arbitrary decision: first of all, on average, it degrades the expected runtime; second, it's used for spin-off strategies, like in Bernstein's pre-computation paper "Computing discrete logs faster".
Well, papers are about theory mostly.
Again, in practical implementation you have to use random start points because you don't know exact number of GPUs you will use during solving (because it's a long process and some GPUs can be turned on and off, or added/removed), also sometimes some GPUs will start new kangs.