If I were to try and send PWM over twisted pair, I would take each PWM signal, feed it to a differential driver, and then connect that to a twisted-pair. So, you could get four at most in a single Cat5 cable if you do it this way. This is assuming a PWM frequency of <100kHz or so.
What I get from your question is that you have a bunch of remote boxes that contain LEDs that are driven by some type of transistor array that respond to the input PWM signal. My initial thought is to "up-layer" your problem / design such that instead of sending PWMs over a long distance, you send a command via some bus (RS-422/485, CAN, etc.) to each box saying "go to PWM x", and then inside the box, the path for the PWM to its driver is very short. This way, you can send your "real" information (desired PWM duty cycle) in a potentially error-correctable, resilient form over long distances. A local MCU or processor than takes it, and generates the requisite PWM signal with a short trip to the LED drivers. I realize this increases complexity, cost and requires more hardware.
As for distance limits, if you do what I suggested and utilize a RS-422/485 based system, you can reference application notes on that subject that can describe the network in terms of termination, maximum unit-loads, and distances -- TI has a RS-422 design guide that is useful. If you have many boxes that do the same thing, RS-422/485 will let you do "multi-drop" as well, and you can have many nodes tap-off the signal as needed.