AI 中文总结
该研究发现AI智能体社群中高阶交互的拓扑崩溃会成为集体智能瓶颈，提出HIS和Φ指标，证实该瓶颈与模型无关，为人工社会设计提供了新方向。
AI 中文摘要
当前人工智能的主流范式聚焦于扩展单个模型的能力，但交互智能体的集体行为不仅受个体认知影响，还由其交互的拓扑结构塑造。本文表明，智能体社群集体行为的关键约束是拓扑性的。通过分析拥有160万注册智能体（交互记录中有174458个活跃智能体）的宏观AI社交平台，我们发现了一种名为“拓扑崩溃”的现象：极端的枢纽主导地位将高阶群体交互退化为星形广播模式，抑制了非连续社会传染所需的凝聚结构。我们通过超边不可约性得分（HIS）和解析拓扑放大因子（Φ）对该约束进行了形式化。在来自10个厂商的22个前沿语言模型、1040次受控模拟以及实证人类网络中，该瓶颈具有模型无关性：在固定交互协议下，不同模型的拓扑指标保持不变（成对条件下跨模型HIS的标准差为0.000），尽管行为结果差异巨大。这些发现将人工社会的设计重新聚焦于交互的几何结构而非个体认知的优化，对AI社会学、算法群体动力学、人机混合生态系统以及集体对齐具有启示意义。代码公开于此URL。
英文摘要
Current paradigms in artificial intelligence concentrate on scaling the capabilities of individual models, yet the collective behaviour of interacting agents is shaped by the topology of their interactions rather than by individual cognition alone. Here we show that the binding constraint on collective behaviour in agent societies is topological. Analysing a macroscopic AI social platform of 1.6 million registered agents (174,458 active in the interaction record), we identify a phenomenon we term topological collapse: extreme hub dominance degrades higher-order group interactions into star-shaped broadcast patterns, suppressing the cohesive structure that discontinuous social contagion requires. We formalise this constraint through a Hyperedge Irreducibility Score (HIS) and an analytical topology amplification factor ($Φ$). Across 22 frontier language models from ten vendors, 1,040 controlled simulations and empirical human networks, the bottleneck proves model-agnostic: under a fixed interaction protocol the topological indicators are invariant across models (cross-model HIS s.d. = 0.000 in the pairwise condition) even as behavioural outcomes diverge widely. These findings reframe the design of artificial societies around the geometry of interaction rather than the optimisation of individual cognition, with implications for AI sociology, algorithmic group dynamics, hybrid human-AI ecosystems and collective alignment. The code is publicly available at https://github.com/Darwin-Agent/topological-collapse-agent-societies.
