Basically a flow is a set of packets out of a stream that have something in common.
It's easy to determine flows with protocols like TCP that have sequence numbers, specifically tag packets with them, and are connection oriented.
UDP based protocols require the WAN accelerator to work harder since nothing "built-in" is keeping track of the flow.
Let's say you are caling someone using a UDP-based VoIP protocol, to a remote office bridged by two WAN accelerators, and the line is silent for 20 seconds. So, during this time, your VoIP protocol is basically transmitting a flow of empty VoIP packets - with a header and other overhead. Crossflow compression sounds like it would be able to recognize this flow, and compress accordingly (perhaps instead of a whole UDP packet containing data representing no sound, it just sends a single byte representing that packet - the "flow" comes into play where the WAN accelerator knows this byte "stands for" data in your VoIP call.)
The VoIP example is likely not the best but hopefully you get the general idea.