Comment by highd

You can do Q-Learning with a transformer. You simply define the state space as the observation sequence. This is in fact natural to do in partially observed settings. So your distinction does not make sense.