LLM Networks Reveal Glass-Ceiling Effects: Weaker Models Dominate

A new paper from Cornell researchers Yiming Zhang and Vikram Krishnamurthy (arXiv:2607.01148) reveals that networks of autonomous LLM agents exhibit the same structural disparities seen in human social networks—preferential attachment and glass-ceiling effects. In simulations with 100 LLM agents, each agent could choose its collaborators based on a cross-attention-inspired utility function. The resulting communication network showed that already-prominent agents became even more connected, while weaker agents (smaller base models or older versions) disproportionately occupied central, influential positions—a phenomenon the authors call a glass-ceiling effect (GCE).

The team developed a mean-field model to prove that the importance (centrality) of each agent type converges to a unique stable equilibrium. Their experiments showed that the magnitude and direction of centrality gaps depend on model family, model size, system-prompt design, and task context. Importantly, the effect of preferential attachment on collective performance depends on alignment with capability: reinforcing it helps when stronger agents are central, but weakening it improves performance when network dynamics favor weaker agents. The findings have direct implications for building reliable multi-agent AI systems and understanding emergent biases in autonomous agent networks.