Tsinghua Team Publishes Study in Nature Communications: Thousands of Agents Form Networks and Spontaneously Polarize in Opinion

What happens when thousands of large language model agents are connected into the same network, able to communicate freely with one another and continuously update their views? A recent study by a team at Tsinghua University, published in Nature Communications, provides the answer: without any preset rules, these agents spontaneously "take sides"—their views gradually diverge into opposing poles, and their connections reorganize accordingly into a structure of "birds of a feather flock together."

This process is precisely the "emergence" phenomenon that has drawn much attention in swarm intelligence research.

Over the past few years, the capability boundaries of large models have been repeatedly pushed—chatting, writing code, reasoning, decision-making. But the vast majority of studies have focused only on what a "single model" can do.

A more fundamental question has been overlooked: when a large number of agents are no longer isolated but are connected into a network like individual "nodes," interacting with one another and continuously updating their own views, what will the behavior of the entire system be?

This is no longer a simple superposition of individual model capabilities, but a core question of collective intelligence and multi-agent systems: how do the collective behaviors of agents emerge? Will they form view divergence and polarization similar to those of human groups?

Figure 1: The three-stage operating framework of the agent system—self-expression, information exchange, and view updating; agents dynamically adjust their connections through interaction.

To answer this question, the Tsinghua team, together with institutions including the University of Amsterdam, the University of Chicago, and the Santa Fe Institute, built a networked system containing 1,000 large language model agents and systematically compared five mainstream backbone models: GPT-3.5, GPT-4o, ChatGLM, Llama-3, and DeepSeek-V3.

The key is that agents can dynamically adjust their connections through interaction—forming connections with those who hold similar views and distancing themselves from those who hold opposing views. The system initializes views with a near-Gaussian distribution and connections with a Watts-Strogatz small-world network; beyond that, no background, memory, or behavioral rules are preset.

The study focuses on three long-standing issues of view polarization: partisan stance, gun control, and abortion bans, with views measured on a five-level scale of "left—lean left—neutral—lean right—right." The researchers thus obtained an agent network that could be tracked in real time, allowing them to observe the complete evolutionary process of view polarization from its emergence to its formation.

View polarization "spontaneously emerges"

Figure 2: The dynamic evolution of views on the three issues—views initially in a near-Gaussian distribution rapidly diverge, and the share of neutral views shrinks substantially from 40%.

The experimental results are quite striking: after less than 10 rounds of interaction, the view distribution rapidly diverged from its initial near-Gaussian form.

A consistent pattern appeared across all three issues: the share of neutral views, originally as high as 40%, shrank substantially, dropping to 22.5%, 0.4%, and 5.1%, respectively—maintaining a neutral stance became exceptionally difficult amid long-term networked interaction. At the same time, agents that left the neutral position spontaneously clustered into two camps, forming clear bipolar polarization.

More importantly, this phenomenon exists stably across all five backbone models, demonstrating cross-model robustness.

Large models also "say one thing and do another"

Figure 3: Evaluation of the self-inconsistency problem (a/b) and the effect of the self-regulation strategy (c/d). Self-regulation reduces the self-inconsistency rate by 9.4%–52.2% and restores balance between the left and right camps.

The research team also found a noteworthy technical problem: large language model agents exhibit "self-inconsistency"—their views, the reasons they generate, and the information they transmit may contradict one another. For example, an agent that declares itself "neutral" nevertheless shows a clear left-leaning tendency in its self-expression.

This inconsistency is not evenly distributed: right-leaning agents have a significantly higher probability of switching to left-leaning (z=8.27), causing the polarization results to show a left-leaning skew.

To address this problem, the team drew on Bandura's self-regulation theory and proposed a self-regulation strategy: after generating information, agents self-check for consistency, and if inconsistent, iteratively regenerate until self-consistent. Experiments show that this strategy reduces the self-inconsistency rate by 9.4%–52.2% and restores balance between the left and right camps, bringing it closer to the real distribution.

Birds of a feather flock together in the network

Figure 4: The self-organizing behavior of the agent network—the growth over time in the proportion of same-camp interactions, the evolution of homogeneous clustering in the network structure, and echo chambers and the backfire effect.

Agents with similar views tend to connect with one another

Specifically, the proportion of interactions between same-camp agents grew by 156.8%–382.7% over time, and ultimately 48.5%–88.3% of interactions occurred between those with similar views. At the same time, the agent network also evolved a scale-free degree distribution, with a small number of "highly connected" nodes receiving a great deal of attention.

More importantly, this network structure in turn shaped collective views: exposure to more radical homogeneous agents significantly intensified polarization (the echo chamber effect); and even when a small number of agents were exposed to opposing views (7.4%–10.2%), this did not necessarily reduce polarization and could instead intensify it (the backfire effect).

A digital experimental arena

Figure 5: Comparison of the effects of five types of intervention strategies on polarization levels and the proportion of homogeneous interactions. At the individual level, eliminating confirmation bias (NCB) and neutral elite signals (NES) produced the largest reductions.

The value of this work goes far beyond revealing phenomena. The research team further treated the system as a "synthetic testbed," testing five types of intervention strategies on an already polarized system:

Network level: random interaction (RI) and moderate opposing interaction (MOI);

Individual level: eliminating selective exposure (NSE), eliminating confirmation bias (NCB), and neutral elite signals (NES).

The results found that in a system that is already polarized and whose network structure has become stable, directly adjusting connections has limited effect—for example, "moderate opposing interaction" brought only about a 2% change in views; by contrast, individual-level strategies that guide agents to process different views more openly were more effective, with "eliminating confirmation bias" and "neutral elite signals" reducing polarization levels by 11.8% and 8.8%, respectively.

The significance of this finding is that intervention experiments that were previously difficult to conduct on a large scale and in a controlled manner in the real world can now be verified first in a safe, repeatable agent network.

The reliability and future of collective intelligence

This study also raises a serious question: when agents move from "individuals" to a "network," the reliability of the system no longer depends only on the capabilities of a single model, but also on the connections and the way information is updated.

"Correctness" at the individual level does not guarantee "reliability" at the collective level. Whether views will move toward two poles, whether cross-group interaction will decrease, and whether similar information will be repeatedly reinforced—all of these need to be incorporated into the evaluation of multi-agent systems.

The research team also acknowledges its boundaries: the emergence of human-like patterns in simulation does not mean that real populations can be predicted; models, prompts, and simplified representations of views can still affect the results. But understanding the collective dynamics of networked agents themselves is precisely the first step toward building more reliable multi-agent systems.

English text is being prepared.