Networked LLM Agents Form Polarized Opinion Groups
A study by researchers from Tsinghua University and partner institutions, published in Nature Communications, built a network of 1,000 LLM agents and tested GPT-3.5, GPT-4o, ChatGLM, Llama-3 and DeepSeek-V3. Without prescribed behavioral rules, agents formed opposing opinion groups, increasingly connected with similar-minded agents and showed self-inconsistency between their stated views, generated reasons and shared information. The researchers released code and quantified the effects of five interventions. In the polarized system, network-level changes had limited effects, while eliminating confirmation bias and neutral elite signals reduced polarization by 11.8% and 8.8%, respectively. A 36Kr article was an authorized republication of the New Zhiyuan report, not independent confirmation. The researchers cautioned that the simulation does not predict real-world populations and may be affected by the selected models, prompts and simplified opinion representation.
The study provides a controllable synthetic testbed for examining how network relationships and information updating contribute to collective polarization. The reports say these factors should be considered when evaluating multi-agent system reliability, while also warning that the simulation does not predict human behavior.
新智元 2026-11-03 19:53 北京
新智元报道
当数千个大语言模型智能体被连入同一个网络、彼此可以自由交流并不断更新观点时，会发生什么？
清华大学团队在《自然·通讯》（Nature Communications）上发表的最新研究给出了答案：在没有预设任何规则的情况下，这些智能体会自发地「站队」——观点逐渐分化成对立的两极，连接关系也随之重组为「物以类聚」的结构。
这一过程，正是群体智能研究中备受关注的「涌现」现象。
#Tsinghua University#University of Amsterdam#University of Chicago#Santa Fe Institute#Nature Communications#GPT-3.5#GPT-4o#ChatGLM#Llama-3#DeepSeek-V3