Try this. It does it in 1 pass and creates only 1 temporary list as opposed to 2 half passes, 3 temporary lists (5 if you count the ranges inside the slices), and 2 temporary arrays, so I think it should work out to be faster.

I realise that I am omitting the lists used to create the hashes, but we both use one, so that balances out.

I've been trying to benchmark the two variations, but for reasons I haven't yet fathomed, I have yet to succeed. I'll update if I do.

sub remove_item {
    my ($i, %del) = (0); 
    @del{@_} = undef;
    @items = grep { $i++ and !exists $del{$_} and $index -= ($i < $ind
+ex) } @items;
}
[download]

Update:I got the benchmarking to go.

#! perl -sw
use strict;
use Benchmark;

my @i = qw( a b c x d e f a g h i j k l z m n o p q r t s t b u v w x 
+y z );
my @dups = qw( x a z t b );
my $index = 13;
my @items = @i;

sub remove_item {
    my %del = map { $_ => 1 } @_;

    my @part_1 = grep { !exists $del{$_} } @items[0..$index];
    my @part_2 = grep { !exists $del{$_} } @items[$index+1..$#items];

    $index = $#part_1;
    @items = (@part_1, @part_2);
}

sub remove_items {
    my ($i, %del) = (0);
    @del{@_} = undef;
    @items = grep { $i++ and !exists $del{$_} and $index -= ($i < $ind
+ex) } @items;
}

print "@dups\n";
@items = @i; $index = 13;
print "$index: $items[$index] : @items\n";
remove_item @dups;
print "$index: $items[$index] : @items\n";
print $/ x 2;

print "@dups\n";
@items = @i; $index=13;
print "$index: $items[$index] : @items\n";
remove_item @dups;
print "$index: $items[$index] : @items\n";
print $/ x 2;

Benchmark::cmpthese( 1000, {
    mine => sub { @items = @i; $index=13; remove_items @dups; },
    yours => sub { @items = @i; $index=13; remove_item @dups; },
});

__DATA__
# Output
C:\test>195796
x a z t b
13: l : a b c x d e f a g h i j k l z m n o p q r t s t b u v w x y z
9: l : c d e f g h i j k l m n o p q r s u v w y


x a z t b
13: l : a b c x d e f a g h i j k l z m n o p q r t s t b u v w x y z
9: l : c d e f g h i j k l m n o p q r s u v w y


Benchmark: timing 1000 iterations of mine, yours...
      mine:  0 wallclock secs ( 0.51 usr +  0.00 sys =  0.51 CPU) @ 19
+60.78/s (n=1000)
     yours:  1 wallclock secs ( 0.64 usr +  0.00 sys =  0.64 CPU) @ 15
+62.50/s (n=1000)
        Rate yours  mine
yours 1562/s    --  -20%
mine  1961/s   25%    --

C:\test>
[download]

I agree that this isn't the way I would tackle the overall problem, but I enjoyed playing with it!