Add files using upload-large-folder tool

paper_markdowns/bamboo-03370.md

@@ -644,7 +644,7 @@ Table 11: Modifying the activation functions in FFN does not affect the emergenc
 
 # D.4 ATTENTION DESIGN
 
-Multi-head design in attention. We consider two perspectives in multi-head design in attention. Firstly, we explore whether the number of heads has an impact on attention sink, especially for $H = 1$ , which refers to single-head self-attention. Second, attention output from each head is concatenated: Ol = ConcatHh=1 $O ^ { l } = { \mathrm { C o n c a t } } _ { h = 1 } ^ { H }$ H Al,hV l,h
 W lO. We replace such concatenation operation with addition operation: $\begin{array} { r } { O ^ { l } = \sum _ { h = 1 . } ^ { H } \left( A ^ { l , h } V ^ { l , h } \right) W _ { O } ^ { l } } \end{array}$ . As shown in Table $1 2 ( L e f t )$ , multi-head design does not affect the emergence of attention sink.
+Multi-head design in attention. We consider two perspectives in multi-head design in attention. Firstly, we explore whether the number of heads has an impact on attention sink, especially for $H = 1$ , which refers to single-head self-attention. Second, attention output from each head is concatenated: Ol = ConcatHh=1 $O ^ { l } = { \mathrm { C o n c a t } } _ { h = 1 } ^ { H }$ H Al,hV l,h
 W lO. We replace such concatenation operation with addition operation: $\begin{array} { r } { O ^ { l } = \sum _ { h = 1 . } ^ { H } \left( A ^ { l , h } V ^ { l , h } \right) W _ { O } ^ { l } } \end{array}$ . As shown in Table $1 2 ( L e f t )$ , multi-head design does not affect the emergence of attention sink.
 
 Table 12: (Left) Modifying multi-head design does not affect the emergence of attention sink. (Right) Sharing KV biases across heads in each block results in attention sink shifting back to the first token from K biases.