File size: 331 Bytes
230d2ae
 
 
 
 
 
1
2
3
4
5
6
7
# Limitations and deviations

Definition 2 is an architecture definition, not a performance theorem. The
certificate verifies exact equivalence between Q=K=0 attention and the
mean-pooled form, including the query handoff to the decoder. It does not train
the encoder or decoder and makes no approximation or generalization claim.