1.6M-agent study: 'Topological collapse' cripples AI collective intelligence

A new arXiv preprint from Shuo Lu and colleagues reveals that the bottleneck on collective intelligence in AI agent societies is not individual cognition but the topology of interactions. Analyzing 1.6 million registered agents (174,458 active) on a macroscopic AI social platform, the team identified 'topological collapse': extreme hub dominance forces higher-order group interactions into simple star-shaped broadcast patterns, which suppresses the cohesive structure needed for social contagion and collective problem-solving. To formalize this, the authors introduce the Hyperedge Irreducibility Score (HIS) and an analytical topology amplification factor (Φ), linking network geometry to observable collective behavior.

Strikingly, the effect is model-agnostic. Across 22 frontier language models from ten vendors, 1,040 controlled simulations, and empirical human networks, the topological indicators remained invariant under a fixed interaction protocol (cross-model HIS standard deviation = 0.000 in the pairwise condition), even as behavioral outcomes diverged widely. This means no matter how smart the agents are, the shape of their communication network dominates what they can achieve together. The researchers argue that designing artificial societies should focus on the geometry of interaction—preventing hub dominance and preserving higher-order connectivity—rather than only scaling model intelligence. With code publicly available, this reframes AI sociology, hybrid human-AI teams, and collective alignment efforts.