Whenever you think of xsl:function (or xsl:sequence) as source of
intermediate result,
that you won't output in the tree immediately, you notice that xsl:sequence
is more efficient:

<xsl:function name="t:add" as="xs:integer">
 <xsl:param name="first" as="xs:integer"/>
 <xsl:param name="second" as="xs:integer"/>
 <xsl:sequence select="$first + $second"/>
</xsl:function/>

Why is  <xsl:sequence select="$first + $second"/> more efficient than
<xsl:value-of select="$first + $second"/>?

I can understand that math on atomic values ought to be more efficient
than math on text nodes,

but is the benefits something we can measure
in any meaningful sense?