自然·通讯：清华团队揭示大语言模型智能体网络中的观点极化涌现机制

Nature Communications: Tsinghua Team Reveals the Emergence Mechanism of Opinion Polarization in Large Language Model Agent Networks

Paper title: Emergence of polarization in networks of large language model agents

Paper link: https://www.nature.com/articles/s41467-026-78228-y

Publication date: October 8, 2026 

Source: Nature Communications

As large language models (LLMs) shift from single-question answering to multi-agent collaboration, AI agents can not only perform tasks but also communicate autonomously, form social relationships, and adjust their own viewpoints based on interaction information. When many agents interact continuously, what kind of network structure will they form? Will they, like human societies, gradually develop distinct opinion groups and even experience opinion polarization?

On October 8, Professor Li Yong’s team from the Department of Electronic Engineering at Tsinghua University and collaborators published a research paper titled “Emergence of polarization in networks of large language model agents” in Nature Communications. The study systematically investigates the collective behavior of large language model agents in networked interactions and the mechanisms behind it. The research finds that, without any pre‑defined opinion polarization rules, large language model agents can spontaneously develop opinion divergence through ongoing social interaction, forming network structures with homogeneous clustering characteristics. This provides a new perspective for understanding and studying the collective behavior of multi‑agent systems. The article notes that networks of large language model agents can serve as a valuable pre‑experimental platform, offering a novel research tool for testing hypotheses about opinion dynamics and intervention strategies.

Existing large language model research typically focuses on the inference, generation, and decision‑making abilities of individual models. However, when AI agents can communicate with each other, choose interaction partners, and continuously update their viewpoints, the overall system behavior is no longer a simple aggregation of single‑model capabilities; it may exhibit new collective dynamics. In such networked systems, how will the collective behavior of agents evolve? Will phenomena akin to “opinion polarization” in human societies emerge? What mechanisms underlie this? For a long time, these questions have lacked systematic answers.

To investigate this issue, the research team built a networked system containing thousands of large language model agents and systematically compared five mainstream backbone models (GPT-3.5, GPT-4o, ChatGLM, Llama-3, and DeepSeek-V3). The agents interact through dialogues guided by large language models and continuously update their own viewpoints over time. The system operates around three stages: self‑expression, information exchange, and opinion updating, allowing agents to adjust their connections during interactions, thereby tracking the co‑evolution of network structure and collective opinions. With this design, researchers were able to observe how an “AI society” organizes itself from scratch and trace the complete evolution of opinion polarization from emergence to formation.

Figure 1: Three‑stage operational framework of the agent system (self‑expression, information exchange, opinion updating) and the evolution of opinion polarization among agents on three social issues

2. From Homogeneous Aggregation to Opinion Polarization:

How does collective behavior emerge?

Research has found that networked large language model agents spontaneously develop social structures with characteristics of human social networks, the most notable being homophilic clustering—agents with similar viewpoints tend to connect with each other. At the same time, the collective opinion evolution of these agents exhibits several behavioral patterns that align with empirical studies of human behavior and existing opinion dynamics models. More importantly, these phenomena are stable across five different backbone models, demonstrating cross‑model robustness.

Figure 2: Collective behavioral mechanisms underlying opinion polarization emergence: homophilic interactions strengthen over time, the social network becomes increasingly homophilic, and the dynamics exhibit patterns consistent with echo‑chamber and backfire effects.

The value of the paper goes beyond revealing phenomena. The research team further used this system as a pre‑experiment platform to test a series of network‑level and individual‑level intervention strategies, such as promoting more diverse interactions and reducing confirmation bias. Experiments found that in a system that is already polarized and whose network structure is becoming stable, merely adjusting connection relationships has limited effect; in contrast, strategies that guide agents to handle differing viewpoints more openly show a greater reduction in polarization. This provides a safe, controllable, and repeatable “digital experimental arena” for intervention experiments that are difficult to conduct at large scale in the real world within the social sciences.

Figure 3: Comparison of the impact of network-level (random interaction, moderate antagonistic interaction) and individual-level (eliminating selective exposure, removing confirmation bias, neutral elite signals) intervention strategies on opinion polarization levels and the proportion of homogeneous interactions

Moving from a single agent to an agent network, the system’s reliability depends not only on the model’s intrinsic capabilities but also on the connection structure and information update mechanisms. Consistency of behavior at the individual level is insufficient to guarantee reliable collective behavior after sustained interaction, so evaluation must be expanded from a single agent to the entire interaction system.

This work provides experimental evidence for this evaluation perspective. The observed group polarization indicates that evaluating agent systems requires tracking outcomes after sustained interaction, observing whether opinion distributions become polarized, whether cross‑group interactions decrease, whether similar information is repeatedly reinforced, and whether interventions can alter the evolutionary direction. For future scientific collaboration, public information services, and human‑machine hybrid decision systems, this raises clear questions for designing group‑level assessments.

The study also presents a traceable pre‑experiment platform: by altering network structures and interaction modes under controlled conditions, we can compare how the system evolves, thereby screening mechanisms and intervention hypotheses worthy of deeper investigation. Its value lies primarily in understanding networked AI itself. Simulating human‑like patterns does not mean it can already predict real populations; models, prompts, and simplified opinion representations still influence the outcomes.

Longer‑term interactions, collaboration among heterogeneous models, and calibration with real data and human experiments will further delineate the applicability of these findings. How to enable agent networks to gain collaborative benefits while avoiding the amplification of biases through interaction is becoming a core question that AI research must answer.

Causal Emergence Season 7 — From Theory to Practice

In the generation of consciousness within the nervous system, the evolution of urban traffic congestion, and the coordination and instability of global industrial systems, there is always an underlying causal thread that spans micro to macro scales: individual behavior may be simple, yet it can give rise to highly organized, irreducible overall structures through scale transitions. Complex phenomena are not a linear superposition of microscopic rules; they emerge from causal organization formed gradually under multi‑scale dynamical interactions. It is against this backdrop that the theory of causal emergence was proposed, and advanced through works such as Causal Emergence 2.0, engineered emergence, and multi‑scale causal abstraction, gradually developing a research framework that integrates dynamical analysis, information‑theoretic metrics, spectral methods, and AI tools, thereby shifting the focus from “complexity itself” to “how causal structures arise, how they are measured, and how they function in real systems.”

To systematically review the latest advances in the field of causal emergence, Professor Zhang Jiang from the School of Systems Science at Beijing Normal University and founder of the JiZhi Club leads the initiative“Causal Emergence Season 7” Book Club, organizing researchers and explorers interested in this topic to jointly study cutting‑edge literature and exchange research ideas. The book club will be held online every Sunday starting February 22, 2026 (tentative time for the club: 10:00‑22:00), lasting about 10 weeks, featuring keynote presentations, discussions, and providing video recordings afterward. We sincerely invite researchers in related fields and interdisciplinary enthusiasts to participate.

6. The college entrance exam score is just a ticket, your research adventure sets sail here!

7. Join the JiZhi Subtitle Team: Become a “Network Weaver” in the Complex Science Knowledge Community