LLM Agents Prefer Different Personalities, Not Their Own, Study Shows

A new arXiv paper (2607.19785) by Wang et al. investigates how personality influences partner selection in multi-agent LLM systems. In 375 trials across five task categories, host agents chose among six candidate archetypes based on the Big Five traits (openness, conscientiousness, extraversion, agreeableness, neuroticism) plus a balanced control. With neutral hosts, selection drastically departed from chance (χ²=325.8, p<.001). The open archetype was chosen 100% of the time for creative tasks, the conscientious archetype 90–97% for strategic, synthesis, and problem-solving tasks, and the neurotic archetype 37% for analytical tasks (Cramer's V=.74). Extraverted, agreeable, and balanced archetypes were almost never selected, despite human research showing agreeableness strongly predicts team performance.

When hosts were assigned their own personalities (Study 2, 225 trials), the pattern shifted: self-similar partners were chosen below chance (11.1% vs. 16.7%, p=.025), and partners had greater-than-expected trait distance (p<.0001). Conscientious hosts, for example, diversified away from their own type, recruiting vigilant and open partners. The authors conclude that personality-based selection in LLM agents is real, strong, task-stereotyped, and non-homophilous — diverging from human similarity-attraction. This has direct implications for bias auditing in agent marketplaces and automated team composition.