Using eval for this kind of thing might be a little over the top. It works, yes, but being a dynamically compiled thing can cause problems. Personally, I would have implemented it without the eval. The code below is a little more "verbose", but most of it is standard issue variable declarations:

sub FindRecord
{
    my ($self) = shift;
    my ($key)  = @_;

    my $ref = \{$self->{$key}};

    if (defined $ref->{path})
    {
         my $path = $ref->{path};
         foreach my $bit (@$path)
         {
              $ref = $ref->{$bit};
         }
    }

    return $ref;
}

# ...
sub SomeFunc
{
    my ($self) = shift;

    # ...

    ${$self->FindRecord($key)} = $value;
}
[download]

This, of course, could also be declared as an 'lvalue' function which would eliminate the ${} prefix. lvalue functions are fun, and I hope that they don't take them away any time soon.

NB: Code untested, used for demonstration purposes.