Summarisation models that do not follow instructions already exist! Fixing exfiltration is good, low hanging fruit.
But for a summarisation task, whole classes of typical instruction following behaviour are totally off target!
Has the field moved on much since the pre-LLM-era models like T5?