Is that how you read the OPs intent? I thought about it, but if the requirement is to retain the final level of entities, then his hardcoded, 3 decodes is going belly up whenever he processes any that has been encoded less than 3 4 times.

Even so, the logic of testing for a change in length works. You just have to retain 2 levels of 'undo' at each iteration. If the data being processed isn't too many megabytes each time, then something as simple as this would work regardless of how many times the content has been entity encoded:

#! perl -slw
use strict;
use HTML::Entities;

my $data = '<p><b><i>AT&amp;T &lt;grin></i></b></p>';

$data = HTML::Entities::encode( $data ) for 1 .. rand( 10 );

my @saved = $data;
my $l1 = length $data;
{
    my $l2 = length( $data = HTML::Entities::decode( $data ) );
    if( $l2 < $l1 ) {
        push @saved, $data;
        $l1 = $l2;
        redo;
    }
}

$data = $saved[-2];
print $data;
__END__
P:\test>junk2
<p><b><i>AT&amp;T &lt;grin></i></b></p>

P:\test>junk2
<p><b><i>AT&amp;T &lt;grin></i></b></p>

P:\test>junk2
<p><b><i>AT&amp;T &lt;grin></i></b></p>
[download]

I still think that the logic shown in the OPs code $title =~ s/strip_stuff_like_html_and_cdata_tags//g;, plus his description

Before working on the text we find inside title tags

suggests that he is interested in manipulating the content, not the markup.

And that if this is ever destined to be redisplayed in a browser, (of which I see no mention?), it will probably be in a completely different context to that from which it was fetched.

Which suggests to me that it would be better to extract the text content, remove all entities to allow for DB storage, pattern matching etc. and if it is ever going to be redisplayed in a browser, re-encode the content before combining it with the new markup.

But you could be right.