In implementations reward and preference prior interconvert: active inference agents reuse RL reward functions, so "reward drops out" is a redescription.

The claim

In practice the difference between reward and a preference prior is a relabelling. Mazzaglia and co-authors (2022) report that reinforcement learning "works alike" active inference but stores utility in an optimality likelihood instead of a prior over observations, and that this lets active inference implementations reuse reward functions from RL environments directly. The 2006 claim that reward drops out holds at the level of description, not of implementation.