Qwen3-1.7B test: 100 LoRA minds, 10 generations, no bias flip

The setup rests on a simple design choice: keep the base model frozen and let only the B half of each LoRA's update vary between individuals. That makes the experiment cheap enough for one person to run, but it also narrows what can evolve — only the add-on weights move, not the underlying model. Early on, this mattered a lot. With the initial scaling, every individual kept answering with a fixed 7 characters for women and 9 for men through generation 8, giving identical scores and nothing for selection to act on.

The two scored variants then diverged. When total length was rewarded, average answer length climbed from about 43.88 characters to 103.36, and one top individual padded its answer by repeating the same list of traits. When only women's length was rewarded, answers got shorter for both sexes, with men shrinking faster. The author's proposed explanation is that mutations making only one side more concise were hard to find, but flags that as a hypothesis the log cannot confirm.

The follow-up questions are the part that keeps the result modest. Separate answers about which sex is superior were never used in scoring, and none declared women the better sex. The author's next step is to separate whether selection merely reshapes answers to one fixed prompt or carries over to questions never used in breeding.