Same Events, Different Opinions: Cross-Platform Opinion Divergence in LLM-Agent Societies

Social media platforms expose users to different information environments, which may lead communities discussing the same events to develop increasingly different collective opinions. We investigate whether such cross-platform divergence can emerge in LLM-agent societies, what platform mechanisms drive it, and whether cross-platform participation can mitigate it. To study these questions, we build a multi-platform simulation in which heterogeneous LLM agents generate posts, consume platform-ranked feeds, and update topic-specific beliefs over repeated interactions. We first place initially neutral agent populations under the same external events and find that distinct platform environments alone can produce substantial opinion divergence. Through controlled ablations, we further show that differences in feed recommendation are a much stronger driver of this divergence than the tested platform-specific surface writing styles. We then vary the proportion of agents who consume content from a second platform and observe a clear, consistent reduction in cross-platform divergence. Most of this reduction remains when each agent is counted only once according to its primary platform, indicating that the pattern cannot be explained by overlapping group membership alone. The main findings also persist when stylized platform networks are replaced with network structures derived from Weibo, Zhihu, and Reddit data. These findings highlight the potential of LLM-agent societies as controlled testbeds for studying social dynamics.