Appealing results... do we think there is scope to close the accuracy gap against AR models? or even leverage the "Bidirectional Reasoning and Self-Correction" into an overall advantage?