Collective Intelligence or Collective Error? Social-Epistemic Dynamics in Diverse Multi-Agent Societies

Multi-agent systems increasingly rely on societies of language-model agents that differ in capability, reasoning behavior, functional role, and social position. While such diversity can provide complementary information, interaction can also cause agents to overweight socially influential information and amplify collective error. We study this problem through a social-epistemic framework that treats collective reasoning as an evolving process shaped jointly by agent diversity, interaction topology, social status, and the alignment between social influence and epistemic competence. Across controlled experiments spanning heterogeneous language models, multiple reasoning benchmarks, and the interactive ALFWorld environment, we systematically vary these factors while separating evidence-driven belief updates from source-driven social influence. Our results show that diversity alone does not reliably produce collective intelligence: unstructured heterogeneity yields only modest and inconsistent gains, while competence-aligned social influence produces substantially larger improvements in collective accuracy. In contrast, when social influence is concentrated on less competent agents, repeated interaction systematically degrades performance and can erase useful epistemic disagreement. The effect persists in sequential decision making, where competence-aligned interaction improves task success while adversarial influence amplifies errors. Mechanistically, we find that source status changes the weighting of otherwise identical evidence, with high-status sources receiving disproportionately greater weight when their evidence is weak or incorrect. This source-dependent distortion compounds over repeated interaction and provides a concrete pathway from local social bias to collective error. Finally, an evidence-gated updating mechanism selectively reduces socially induced updating when available evidence conflicts with source influence, substantially recovering performance from the adversarial regime without eliminating interaction altogether. Together, these results show that the reliability of a multi-agent society depends not simply on how capable or diverse its agents are, but on whether its social structure amplifies epistemically reliable information.