You say it performs well when two characters have a single trait which is different between them. I wonder how much better it performs better when you give character A many masculine traits, and B many feminine traits, without directly stating A is male and B is female, compared to if you randomize those traits for A or B.
In general, assigning traits which correlate highly with each other should give better results. Perhaps a problem is that the more characters and traits you assign, the less correlated those traits are with all the other traits, and so far lower performance is seen.
You say it performs well when two characters have a single trait which is different between them. I wonder how much better it performs better when you give character A many masculine traits, and B many feminine traits, without directly stating A is male and B is female, compared to if you randomize those traits for A or B.
In general, assigning traits which correlate highly with each other should give better results. Perhaps a problem is that the more characters and traits you assign, the less correlated those traits are with all the other traits, and so far lower performance is seen.