Why is our top-k or top-p worse than SUEPRT?

[forrestbao]

SUPERT shows that that top-k is a good way to extract salient sentences to generate pseudo-reference.
However, our results show otherwise.

SUPERT shows that extractive heuristics are not good at extracting salient sentences.
But how come we are worse than SUPERT?