To expand a bit on why this is relevant, VSX tries to hide a lot of how it works. Really, it's VRFs (R80.30 and earlier) or network namespaces (R80.40). Each physical member gets a VRF or NetNS for each VS you define. These VRFs/NetNSs are then clustered like any other Check Point cluster. To provide the illusion of the VS being a single thing which can move between the physical cluster members, you only have to specify one IP address for each interface. That IP is the cluster VIP. In the background, the management then sets up the VRFs/NetNSs on the physical members with automatically-allocated off-net IP addresses.
Check Point includes wrappers for many Linux commands which makes them work in the individual VRF/NetNS and show results from the cluster. 'ifconfig', for example, is wrapped and shows you the cluster IPs rather than the member IPs on the interfaces. If you use /sbin/ifconfig, you will see the member IPs.
As a result of this, all IPs on the same network as the VS's IP are inherently not on the same network as the members' IPs. Therefore you have to use this ip_nonlocal_bind trick to use arping to resolve addresses in the VS's networks.