交叉接触激进化引擎：观点动力学随机复合模型中的平台设计、注意力与地理因素
Radicalization Kinetics under Algorithmic Exposure in a Stochastic Multiplex Model of Opinion Dynamics
AI总结：
该研究构建随机复合观点动力学模型，分析社交媒体平台设计、注意力与地理因素对激进化的影响，发现算法同质性可悖论性抑制极端主义，未策划接触是激进化饱和水平的决定因素。
AI中文摘要：
社交媒体平台的极化是因为它们隐藏了分歧，还是因为它们让人们暴露于分歧中？我们研究了一个连续时间模型，其中智能体通过布朗运动在物理空间中扩散，同时在有限注意力预算下通过自适应的、算法策划的数字网络进行交互。当影响完全是同化性的（有界置信度）时，无观点偏见的远程接触会修复由局域性导致的碎片化，而算法同质性则会构建回音室。一旦影响包含对强烈对立观点的排斥反应，平台设计的排名就会反转：中立的、未策划的平台会导致最大极化，观点被固定在极端；寻求争议的参与算法几乎同样具有激进化作用；而强算法同质性悖论性地保护人群免受极端主义影响。双 bloc 简化显示，由参与核设定的稳态跨 bloc 注意力分数控制激进化速率，以及数字注意力份额中的有限时间跨度交叉点λ_c(T)——与模拟结果一致，且无针对 onset 数据拟合的参数。因此，激进化是速率受限的，而非阈值受限的。在排斥参数、核和规模达N=1600的情况下，未策划的接触设定了激进化的饱和水平。 mobility-注意力相图显示，一旦平台拥有智能体的大部分注意力，物理移动性就变得无关紧要。在观点独立的移动性下，我们未检测到地理观点结构；仅当佩克莱特阈值χℓ/D~1以上时，谢林型同质性漂移强度χ才会恢复该结构。该模型为现场观察提供了一种机制，即策划的交叉接触会增加极化，并暗示接触多样性干预的效果可正可负，取决于负面影响的普遍性。
英文摘要：
We study how physical mobility, algorithmic exposure, and repulsive social influence interact in a stochastic multiplex model of opinion dynamics. Agents diffuse in physical space while a directed digital network rewires under a conserved attention budget, so digital exposure displaces rather than supplements local interaction. With purely assimilative bounded-confidence influence, opinion-blind long-range exposure reduces locality-induced fragmentation whereas homophilic recommendation preserves echo chambers. When a contested repulsive response to sufficiently distant opinions is activated, this ordering reverses at the reference parameters: a neutral platform reaches the maximal polarization permitted by the bounded opinion space, controversy-seeking curation drives faster initial separation but slows sharply near the boundary, and homophilic curation delays radicalization by suppressing cross-bloc exposure. In a late-stage symmetric two-bloc reduction, any curation kernel maps to a state-dependent cross-bloc exposure profile $p(y)$ and an exact quadrature for the radicalization time. Pointwise-ordered profiles inherit a global kinetic ordering; crossing profiles yield target- and horizon-dependent rankings. For similarity-driven curation the quadrature has a closed form involving the exponential integral. Simulations, finite-size scans to $N=1600$, structural controls, and a well-mixed particle comparison support the mechanism. Heavy-tailed influence strengths are not required for the inversion; in the well-mixed heavy-tail regime they additionally produce a non-self-averaging stable-weighted asymptotic description. Finally, opinion-independent Brownian mobility produces no detectable geographic opinion structure in the explored regime, whereas opinion-dependent drift produces spatial domains through a Péclet-controlled crossover near $χ\ell/D \sim 1$.
