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。
过去几年，大模型的能力边界被反复刷新——聊天、写代码、推理、决策。但绝大多数研究，都只盯着「单个模型」能做什么。
一个更根本的问题被忽略了：当大量智能体不再孤立，而是像一个个「节点」被连入网络、彼此交互、不断更新自身观点时，整个系统的行为会是什么？
这不再是单个模型能力的简单叠加，而是群体智能（collective intelligence）与多智能体系统（multi-agent system）的核心问题：智能体的集体行为如何涌现？是否会形成与人类群体类似的观点分化与极化？
图1：智能体系统的三阶段运行框架——自我表达、信息交流、观点更新，智能体在交互中动态调整连接关系。
为了回答这一问题，清华团队联合阿姆斯特丹大学、芝加哥大学、圣塔菲研究所等机构，构建了一个包含1000个大语言模型智能体的网络化系统，并系统对比了GPT-3.5、GPT-4o、ChatGLM、Llama-3、DeepSeek-V3五种主流骨干模型。
论文链接：https://www.nature.com/articles/s41467-026-78228-y
代码链接：https://github.com/tsinghua-fib-lab/LLM_for_Polarization
每个智能体围绕三个环节持续运行：
- 自我表达——生成支持自身观点的理由；
- 信息交流——自主决定与谁交流、传递什么信息；
- 观点更新——根据收到的信息更新自身观点。
关键在于，智能体可以在交互中动态调整连接关系——与观点相近者建立连接、疏远观点相左者。系统采用近高斯分布初始化观点、Watts-Strogatz小世界网络初始化连接，除此之外不预设任何背景、记忆或行为规则。
研究聚焦三个长期存在观点极化的议题：党派立场、枪支管控、堕胎禁令，观点采用「左—偏左—中立—偏右—右」五级尺度。于是，研究者获得了一个可实时追踪的智能体网络，得以观察观点极化从萌芽到形成的完整演化过程。
图2：三个议题上的观点动态演化——初始近高斯分布的观点迅速分化，中立观点占比从40%大幅萎缩。
实验结果相当惊人：在短短不到10轮互动后，观点分布就从初始的近高斯形态迅速分化。
一个一致的规律在三个议题上都出现了：原本占比高达40%的中立观点大幅萎缩，分别降至22.5%、0.4%和5.1%——在长期的网络化互动中，维持中立立场变得异常困难。与此同时，离开中立位置的智能体自发地聚成两个阵营，形成明显的两极分化。
更重要的是，这一现象在五种骨干模型中都稳定存在，展现出跨模型的稳健性。
图3：自我不一致问题的评估（a/b）与自我调节策略的效果（c/d）。自我调节将自我不一致率降低9.4%–52.2%，并让左右阵营恢复平衡。
研究团队还发现了一个值得注意的技术问题：大模型智能体存在「自我不一致」——它们的观点、生成的理由、传递的信息之间可能相互矛盾。例如，一个声明「中立」的智能体，在自我表达时却流露出明显的左倾倾向。
这种不一致并非均匀分布：右倾智能体切换为左倾的概率显著更高（z=8.27），导致极化结果呈现左倾偏斜。
针对这一问题，团队借鉴班杜拉的自我调节理论，提出了一种自我调节策略：让智能体在生成信息后自我检查一致性，若不一致则迭代重生成，直到自洽为止。实验表明，这一策略将自我不一致率降低了9.4%–52.2%，并使左右阵营恢复平衡，更贴近真实分布。
图4：智能体网络的自组织行为——同阵营交互比例随时间的增长、网络结构的同质性聚集演化，以及回音室与逆火效应。
具体而言，同阵营智能体之间的交互比例随时间增长了156.8%–382.7%，最终有48.5%–88.3%的交互发生在观点相近者之间。与此同时，智能体网络还演化出了无标度（scale-free）的度分布，少数「高连接度」节点获得大量关注。
更重要的是，这种网络结构反过来塑造了集体观点：接触更激进的同质智能体会显著加剧极化（回音室效应）；而即便少数智能体接触到对立观点（7.4%–10.2%），也未必能降低极化，反而可能加剧（逆火效应）。
图5：五类干预策略对极化水平与同质性交互比例的影响对比。个体层面的消除确认偏差（NCB）与中立精英信号（NES）降幅最大。
这项工作的价值，远不止揭示现象。研究团队进一步把系统当作「数字实验场」（synthetic testbed），在已经极化的系统上检验了五类干预策略：
- 网络层面：随机交互（RI）、适度对立交互（MOI）；
- 个体层面：消除选择性暴露（NSE）、消除确认偏差（NCB）、中立精英信号（NES）。
结果发现：在已经极化、网络结构趋于稳定的系统中，直接调整连接关系的作用有限——例如「适度对立交互」仅带来约2%的观点变化；相比之下，引导智能体更开放地处理不同观点的个体层面策略效果更明显，「消除确认偏差」与「中立精英信号」分别将极化水平降低了11.8%和8.8%。
这一发现的意义在于：过去难以在真实世界大规模、可控开展的干预实验，如今可以在一个安全、可重复的智能体网络里先行验证。
这项研究也抛出了一个严肃的问题：当智能体从「个体」走向「网络」，系统的可靠性不再只取决于单个模型的能力，还取决于连接关系与信息更新方式。
个体层面的「正确」，并不保证群体层面的「可靠」。观点是否会走向两极、跨群体交互是否减少、同类信息是否被反复强化——这些都需要被纳入对多智能体系统的评估。
研究团队也坦言其边界：模拟出现类人模式，并不等于能够预测真实人群；模型、提示词与简化的观点表示仍会影响结果。但理解网络化智能体自身的集体动力学，正是构建更可靠多智能体系统的第一步。
参考资料：
https://www.nature.com/articles/s41467-026-78228-y
编辑：LRST
不喜欢
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
New Ji-won.
What happens when thousands of large language models, AI agent, are connected to the same network, can freely communicate with each other and keep their views updated?
The latest study published by the University of Tsinghua team in Nature Communications offers the answer: In the absence of any pre-set rules, these AI agents will spontaneously “stop team” — the views gradually split into opposing poles, and the connection is then restructured into a “grouping of things”.
This process is one of the “emerging” phenomena that has received much attention in group intelligence research.
Over the past few years, the capacity boundaries of LLM have been repeatedly updated — chatting, writing codes, reasoning, decision-making. But most of the research is focused on what a single model can do.
A more fundamental question was ignored: what would the whole system do when a large number of AI agents were no longer isolated, but were connected, interacting and updating their views like a "node"?
This is no longer a simple addition of individual model capabilities, but the core issue of group intelligence and multi AI agent systems: how does the collective behaviour of AI agent emerge? Will it be possible to divide and polarize views similar to those of human groups?
Figure 1: The three-stage operating framework of the AI agent system - self-expression, information exchange, updates of views, and AI agent dynamically adjusts the connection in interaction.
To answer this question, the Qinghua team, in association with the University of Amsterdam, the University of Chicago, the Institute of Santa Fe and others, built a network system of 1,000 large language models AI agent and systematically compared the five main backbone models GPT-3.5, GPT-4o, ChatGM, Llama-3, DeepSeek-V3.
Thesis links: https://www.nature.com/articles/s4146-026-78228-y
Code link: https://github.com/tsinghua-fib-lab/LLM for Polarization
Each AI agent runs around three links:
- Self-expression - generating reasons to support one's own views;
- Information exchange - autonomous decision on who to communicate and what to convey;
- Update of views - update of views based on information received.
The key is that AI agent can dynamically adjust connections in interaction — to those who are close to the point of view and alienate from it. The system uses an initialization of the near-gos distribution, and an initialization of the Watts-Strogatz mini-world network, in addition to which no background, memory or behavioural rules are foreseen.
The study focused on three long-standing polarization topics: partisan positions, gun control, abortion bans, using the "left-left-neutral-right-right" scale. As a result, researchers were provided with a real-time tracking network of AI agent, which allowed them to observe the evolution of polarization from germination to formation.
Figure 2: The dynamic evolution of views on the three topics — the initial near-Gorse distribution was rapidly divided, with the neutral view being significantly reduced from 40 per cent.
The results of the experiment were quite striking: after less than 10 rounds of interaction, the distribution of views rapidly divided from the initial near-Gorse pattern.
A consistent pattern has emerged on three topics: the neutral view, which was already as high as 40 per cent, has shrunk significantly, to 22.5 per cent, 0.4 per cent and 5.1 per cent, respectively. In long-term networked interactions, maintaining neutrality becomes extremely difficult. At the same time, AI agent, who left the neutral position, spontaneously came together in two camps, forming a clear polarization.
More importantly, the phenomenon has stabilized in five backbone models, showing the robustness of the cross-model.
Figure 3: Assessment of self-incompatibility (a/b) and effectiveness of self-regulation strategies (c/d). Self-regulation reduces self-harmonicity by 9.4 per cent to 52.2 per cent and restores balance to the right and right camps.
The team also identified a technical problem that deserves attention: LLMZ42Z has "self-inconsistent" — their views, the reasons for their generation, and the information they send may contradict each other. For example, a "neutral" AI agent manifests itself in a self-expression with a clear left leaning.
This inconsistency is not evenly distributed: the right AI agent is significantly more likely to switch to the left (z = 8.27), leading to a left-sided polarization.
In response, the team developed a self-regulating strategy based on Bandura ' s self-regulating theory: allowing AI agent to check for consistency after generating information and, if inconsistent, to regenerate until the time of agreement. The experiment showed that the strategy had reduced self-inconsistency by 9.4 per cent to 52.2 per cent and had restored balance to the right and right camps, closer to real distribution.
Figure 4: Self-organizing behaviour of the AI agent network — ratio of interaction with the camp growing over time, homogenous cluster evolution of the network structure, and echo room and counterfire effects.
In particular, the interaction ratio with the camp AI agent has increased over time by 156.8 per cent - 382.7 per cent, with 48.5 per cent - 88.3 per cent of interactions occurring between people with similar views. At the same time, the AI agent network has evolved into a scale-free distribution, with a few "high connectivity" nodes receiving considerable attention.
More importantly, this network structure, in turn, shapes the collective view that exposure to the more radical homogenous AI agent can significantly exacerbate polarization (sounding-room effects); and that even a small number of AI agent encounter opposing views (7.4 per cent-10.2 per cent) may not necessarily reduce polarization but may increase (antifire effects).
Figure 5: Impact of five types of intervention strategies on polarization levels versus homogenous interactions. The elimination of identified deviations (NCBs) and neutral elite signals (NESs) at the individual level were the most significant reductions.
The value of this work goes far beyond revealing phenomena. The research team has further developed the system as a "synthetic testbed" and has tested five types of intervention strategies on already polarized systems:
- Network level: random interaction (RI), moderate interaction (MOI);
- Individual level: elimination of selective exposure (NSE), elimination of identified deviations (NCB), neutral elite signal (NES).
As a result, in systems that have become polarized and network structures more stable, the effect of direct adjustments to connections has been limited — for example, “moderate confrontation” has brought about only about 2 per cent change of view; in contrast, the effect of individual-level strategies that have guided AI agent towards more open treatment of different viewpoints has been more evident, with “eliminating identified deviations” and “impartial elite signals” reducing the level of polarization by 11.8 per cent and 8.8 per cent, respectively.
The significance of this discovery lies in the fact that it was difficult to conduct large-scale, controlled intervention experiments in the real world, which can now be validated in a secure, repeatable AI agent network.
This study also raises a serious question: when AI agent moves from "individuals" to "networks", the reliability of the system no longer depends solely on the ability of individual models, but also on the connection and how information is updated.
The "right" at the individual level does not guarantee "reliability" at the group level. Whether or not the perspective is polarized, whether cross-group interaction is reduced and whether similar information is reinforced repeatedly — these need to be included in the assessment of the multi-AI agent system.
The research team also spoke of its boundaries: modelling the emergence of a human-type model does not mean that it is possible to predict the real population; models, prompt and a simplified perspective suggest that it will still affect results. But understanding the collective dynamics of networked AI agent is the first step in building a more reliable multi-AI agent system.
References:
https://www.nature.com/articles/s41467-026-78228-y
Edit: LRST
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.