Trained Experience versus Prompted Personas: Population-Level Diversity in LLM Moral Judgment

Alignment is usually treated as convergence on a single, static value system. We show that a population of LLM-based agents can instead sustain and grow genuinely different values, with "experience" as the source of that diversity. Each agent is a LoRA adapter over one frozen model, trained with no task loss or reward on its own subset of Japanese narratives, mimicking how humans form values through what they encounter and live through. Experience content alone pushes agents well past a same-subset noise floor (2.86 vs. 5.06), and six generations of novelty search grow that divergence further, to 7.65, rather than collapsing it toward consensus. This transfers largely intact to an independent moral-judgment benchmark (JETHICS), with small gains on two categories and only a modest shift on the rest, and it does not change much when agents exchange experiences directly. Persona-prompting a static model with no training produces real but far smaller spread (4.63) than trained experience does (7.65), and richer real personas outspread simpler synthetic ones at equal group size: "experience" does something prompting alone cannot fully replicate. Checked against human sentence-level safety judgments, agents trained on human-sourced experience track human moral judgment better than those trained on synthetic content. We discuss the quantity vs. quality of experience behind human simulations, an approach gaining visibility across alignment research.