Tyler has part of it. Each receiver input also has its own resistance to that side’s GND2 (the input impedance, around 96k for a 1/8 unit load device, more or less), and the bias network ties A and B toward one side’s isolated supply and ground. Each isolated side is floating, so it settles wherever those resistances and the leakage across the isolation barrier put it. The two GND2s drift toward each other through the bus wires themselves. On a 2 m bench with nothing injecting current, that’s close enough, so it works.
Over 300 m between buildings you get induced common-mode voltages and lightning transients. With nothing defining the common mode, one side can sit far outside the receiver’s range and the link gets flaky, or the isolation barrier takes surges it doesn’t need to.
Standard practice: run a third conductor as a signal reference, connect it to GND2 at each node through about 100 ohm to limit current if the grounds disagree, and earth the cable shield at one end only. Michael, that’s why you were taught three wires, it’s the right habit. Keep the bias on one end and termination at both ends as you have it.