Unicoon: Hypergraph-based Multi-Agent Simulation of Information Cocoons

摘要

Information cocoons pose significant challenges to democratic discourse and social cohesion. While extensive research has examined how social networks and recommender systems independently contribute to this phenomenon, their coevolving dynamics remain unexplored. We present Unicoon, the first unified computational framework that simultaneously models both social network propagation and algorithmic content delivery using LLM-based multi-agent simulation. Our key innovation lies in a hypergraph formulation where agents constitute nodes, social relationships form edges, and recommender-delivered content creates dynamic hyperedges, elegantly capturing heterogeneous information diffusion patterns within a single mathematical structure. To address the adaptive nature of recommendation algorithms, we introduce a dynamic hyperedge construction technique that computationally matches trending content with interested user cohorts in real-time. Extensive experiments on synthetic and real-world networks reveal that the synergistic interplay of social and algorithmic mechanisms creates qualitatively distinct polarization patterns, with a critical finding that larger macro-structures paradoxically accelerate micro-level cohort polarization.