I don't think regex is needed here, and you really should make the code somewhat more forgiving to data variations. What if you use all letters of the alphabet instead of just a and b?

use strict;
use warnings;

my @temp=("a_1","b_1","a_2","a_3","a_4","b_2","a_5","b_3","a_6","b_4")
+;

my ($key, $val, %hash);
for (@temp) {
    ($key, $val) = split '_';
    if (exists($hash{$key})) { push @{$hash{$key}}, $val; }
    else { $hash{$key} = [$val]; }
}

for (sort keys %hash) {
    print "$_: " . join(' ', @{$hash{$_}}) . "\n";
}
[download]

You could probably optimize this further by allocating the entire expected space for each array when initially created, but I'll leave that for the next guy.