社交媒体平台导致两极分化，是因为它们掩盖了分歧，还是因为它们使人们暴露于分歧之中？我们研究了一个连续时间模型：智能体在物理空间中通过布朗运动扩散，同时在有限注意力预算下，经由一个自适应的、算法驱动的数字网络进行交互。当影响纯为同化型（即有界置信度）时，无视观点的长程暴露可修复因局部性导致的碎片化，而算法同质性则构建回音室；一旦影响中包含对强烈对立观点的排斥反应，平台设计的排序即发生逆转：一个中立、未经筛选的平台导致最大程度的两极分化，意见被锁定在极端位置；追求争议性内容的参与度算法几乎同样具有激进化效应；而强算法同质性却悖论式地使群体免于极端主义。一个双阵营简化模型表明，由参与度核函数设定的稳态跨阵营注意力比例控制着激进化速率，且存在一个数字注意力份额的有限时间尺度临界值 $λ_c(T)$——该结果与模拟一致，且未对激进化起始数据拟合任何参数。因此，激进化是速率受限的，而非阈值受限的。在不同排斥参数、核函数及规模（最大至 $N=1600$）下，未经筛选的暴露均设定了激进化的饱和水平。一个流动性–注意力相图显示，一旦平台占据个体绝大部分注意力，物理流动性便变得无关紧要。在观点无关的流动性假设下，我们未检测到地理层面的意见结构；仅当施加强度为 $χ$ 的类似谢林（Schelling）模型的同质性漂移，且其佩克莱特数 $χ\ell/D\sim 1$ 超过某一阈值时，该结构才得以恢复。本模型为实证研究中观察到的‘经算法筛选的交叉暴露反而加剧两极分化’现象提供了一种机制解释，并暗示‘暴露多样性’本身并非充分条件……

Do social-media platforms polarize because they hide disagreement, or because they expose people to it? We study a continuous-time model in which agents diffuse in physical space by Brownian motion while interacting through an adaptive, algorithmically curated digital network under a finite attention budget. When influence is purely assimilative (bounded confidence), opinion-blind long-range exposure heals locality-induced fragmentation and algorithmic homophily builds echo chambers. Once influence includes a repulsive response to strongly opposed views, the ranking of platform designs inverts: a neutral, uncurated platform drives maximal polarization with opinions pinned at the extremes; a controversy-seeking engagement algorithm is nearly as radicalizing; and strong algorithmic homophily paradoxically shields the population from extremism. A two-bloc reduction shows that the stationary cross-bloc attention fraction set by the engagement kernel controls the radicalization rate and a finite-horizon crossover $λ_c(T)$ in the digital attention share---consistent with simulations, with no parameters fitted to the onset data. Radicalization is thus rate-limited, not threshold-limited. Across repulsion parameters, kernels, and sizes up to $N=1600$, uncurated exposure sets the saturation level of radicalization. A mobility--attention phase diagram shows that once the platform owns most of an agent's attention, physical mobility becomes irrelevant. Under opinion-independent mobility we detect no geographic opinion structure; a Schelling-type homophilic drift of strength $χ$ restores it only above a Péclet threshold $χ\ell/D\sim 1$. The model offers a mechanism for field observations in which curated cross-cutting exposure increased polarization, and implies that exposure-diversity interventions can have either sign depending on the prevalence of negative influence.