Action Inference by Maximising Evidence: Zero-Shot Imitation from Observation with World Models | Read Paper on Bytez