A team from Tsinghua University, in collaboration with multiple institutions, has found that in networked free-interaction scenarios without preset rules, a thousand large language model agents spontaneously exhibit opinion polarization: opinions that initially approximate a Gaussian distribution rapidly diverge within just over ten rounds of interaction, the proportion of neutral agents shrinks significantly, agents spontaneously form opposing camps, and through dynamic reorganization of connections they develop a homophilous, clustered network structure in which like attracts like. This phenomenon holds stable across five mainstream backbone models. The study also reveals a self-inconsistency problem among large model agents—namely, inconsistencies among opinions, expressions, and transmitted information—and the self-regulation strategy proposed in response can effectively reduce the inconsistency rate and balance camp distribution. This networked agent system can serve as a digital experimental field for testing the regulatory effects of different intervention strategies on polarization. Experiments show that intervention strategies at the individual level are more effective at reducing polarization than interventions that directly adjust network connections. The study points out that the reliability of multi-agent systems depends not only on the capability of individual models but is also shaped by connection relationships and information update methods, offering an important reference for building more reliable multi-agent systems.