what's meaning of "prediction confidence"？

[BohanSi]

I see the issue#85, and I have the same question. In the paper "Attention mask for generative pretraining" part, you said “scGPT selects the top 1/K genes from the unknown set with the highest prediction confidence to be included as known genes in the next iteration i + 1.” I'd like to know how to calculate the prediction confidence of generated genes.
And I see the Supplementary Fig. 1b, does it mean the highest attention score among the unknown genes?