Popular news sites




查看: 2481|回复: 53573

[转帖] Cutting -edge progress: The social interaction of the big model Agent has emerged as a bidding network

发表于2024-03-03 20:57:55 | 只看该作者 回帖奖励 |倒序浏览 |阅读模式

    Original David Garcia et al.


    Thenews express website standard -free network is one of the most famous examples of behavior. It is everything in the social system, especially in the online social media that users can pay attention to each other.By using GPT3.5-Turbo as a language model, analyzes the interaction of multiple genetic subjects (General Agents). This article shows that they not only simulate the language behavior of individual human beings, but also show the collective phenomenon of human society, especially in the human society, especially the collective phenomenon of human society, especially the collective phenomenon of human society, especially in the human society, especiallyThe emergence of no standard network.It was found that the process of biased Token of GPT3.5-Turbo interfere with this process, which may lead to the phenomenon of extreme centralization of the network.This article shows how to eliminate these token priorities by renamed the main body, so that the model can generate various networks from random networks to more realistic non -standard networks.

    Research field: complex network, no standard network, emergence, large language model, alignment phenomenon

    Giordano de Marzo, Luciano Pietronero, David Garcia | Author

    Liu Peiyuan | Translator

    Essay topic:
    Emergeence of Scale-Free Networks in Social Interactions among Large Language Models
    Thesis address:


    Artificial intelligence (AI) subtle and profoundly penetrated into our daily life, especially with the rise of deep learning technology.This subtle influence is even more intense with the rise of the generation AI (GENAI) and especially the rise of large language models (LLMS), such as ChatGPT.These technologies have been regarded as research auxiliary tools, and they have now emerged and have become the tools and focus of scientific exploration.They surpass the preliminary programming and writing auxiliary roles, and can be competent tasks that they have never been designed or trained [1].In this context, a lot of attention focuses on the ability of understanding and analyzing a single model [2, 3, 4].These studies provide in -depth insights on the subtle function, prejudice and potential applications of large language models [5, 6, 7, 8, 9, 9, 11, 11, 12].However, the area that needs to be explored in depth is interaction between multiple large language models [13, 14, 15].This research direction is very important, especially considering how human social interaction causes the emergence of collective behavior [16].

    Although the research on isolated large language model behaviors is inspiring, but the implicit dynamics of these models in a more complex and interactive environment can only provide a limited perspective.In human society, many most important processes and phenomena are not the products of a single behavior, but the result of collective behavior.For example, cultural trends, economic transformations, and social movements are the embodiment of the emergence of many individuals interaction, and they usually appear in a complex and difficult to predict. [17]Large language models are the core parts of the concept of genetic forms: combining large language models with other structures and attributes (such as memory, planning, and identity) to model human behavior [18,19].When multiple generations interact in the simulation environment, they may show the behavior of unpredictable or not directly coding in any single subject or large language model, and resulting in the expected results [20].These situations originated from the complex interaction and dependencies between multiple subjects and their large language models.In a sense, each generating body contributes to a more complicated and unpredictable collective "intelligence" or pattern.EssenceFor example, the simulation using the genetic body shows an instance of emerging behaviors such as information diffusion and coordination [18].

    Complex network is a representative example of the emergence structure [21].The complex network has the distribution of Free-SCALE, with amazing emergence: the variance of degree may increase with the growth of the network, and the spread of the epidemic may be extremely difficult to deal with [22].For example, the Wanwei.com and the online social network are formed by the interaction between countless individuals, and the relationship and information flow shake a dynamic and evolutionary structure.In particular, online social networks like Twitter or Instagram have been confirmed to have a marginal distribution [23, 24].These examples highlight a key feature of complex networks: their structure and dynamics are the emergence of interaction between their components, not determined by the central entities of the control system.

    In this article, for the first time, we discuss how the genetic subject forms a network structure by self -organizing and whether these structures may be complex.Just as human interaction has spawned the complex structure of the social network, we speculate that the interaction between the genetic types can also bred the equivalent and complicated network.The importance of this study is highlighted by two emerging trends: on the one hand, the rise of social networks driven by artificial intelligence -driven social intelligence such as Chirper, and the increase in the number of robots controlled by large language models in traditional social networks in traditional social networks [25, 20];On the one hand, the development of the main body of the genetic body is based on the development of the main model [15, 14].These development trends indicate the birth of a digital ecosystem shaped by a large language model. Among them, interaction between artificial intelligence entities may become daily like human communication.


    Simulation of social network growth

    We have built a simplified version of the online social network model. This model is inspired by complex network growth models such as the Barabasi-Albert model [26]. Among them, the nodes are added to the network in turn and connected with other nodes.In our simulation, the node is a genetic format for online social network users.We did not clearly model and simulate the subject's behavior through the equations, but used large language models to guide the activity of generating subjects.

    Figure 1: The growth of the main body network based on the large language model.Left: We are used to initialize a large language model of each generated body, so that it is a prompt (Promot) as a user of the online social network.After this prompt, a detailed list is connected to list all the number of network users and their friends.Right: Diagram of the simulation process.At each time, the new subject uses a given prompt for instance and receives the degree of information of existing network users.Then, the subject determines that it wants to connect M and update the network to include these new connections.With the addition of the new subject, this iterative process continues.

    Figure 1 shows the behavior of our genetic subject in the simulation.In each time step, use Openai API, a generating subject driven by GPT-3.5-Turbo is initialized, using prompts displayed in Figure 1 (left), which represents a user in an online social network.It is a user who choose to connect with it.The subject receives a list to list all existing subjects. Each subject is identified by a random three -character string, and the number of connections they already have, as shown in Figure 1 (right).The ranking of the user list is completely random, and it will change in each time step to avoid potential prejudice due to the position in the prompt.The new subject decides to establish contact with which other subjects by responding to their names.It is worth noting that in our simulation, the information received by the subject is exactly the same as the information owned by nodes in the Barabasi-Albert model.The main difference is that they created the connection: In the Barabasi-Albert model, the node is connected in accordance with a fixed probability rule, and our main body is based on the nodes to be connected based on text generated by a large language model.Once the genetic subject selects the user, the corresponding M non -directional side is formed, and the list of the network is updated.This process will repeatedly iterate until the expected network scale.

    Network topology

    Figure 2: Network topology structure.Left: Different network topology structures corresponding to the four situations discussed in the main article: the existence/lack of the token priority, and the existence/lack of the degree in the main interface.The color reflects the age of the node, and the size of the node is proportional to the number of the degree.Right: The complementary cumulative distribution function of the four different network topology structures (CCDF).

    After determining the network growth process, we will focus on analyzing the structure of the network.When the subject can see the number of other people's friends, the generated network presents the center-radiation topology structure, and a few of them have obtained most of the connected edges, as shown in the left panel of Figure 2.a.In this case, the transmission connection of new nodes is limited to the few previous nodes, resulting in unevenness, which is more different from the actual social network structure observed.This extreme concentration can be seen in the CCDF on the right panel on Figure 2.This is an artificial intelligence alignment method that leads to unreal network structure and extreme concentration of network degrees.

    In order to understand this network generation process more deeply, we adjusted the model and removed the degree information displayed to the subject. That is to say, the number of friends at each node will not be displayed on the interface seen by the subject.The list of other subjects is for selection.If an analogy is based on the Barabasi-Albert model, we will expect this to generate a random network structure.However, this simulation without node information has unexpectedly produced a network with a wider range of distribution. From the topology point of view, it is closer to the actual online social network, as shown in the left and right panels of Figure 2B.The in -depth exploration of this network topology reveals its main problem: the existence of the node is not related to the degree of nodes, and the new node may attract the proportion of large degrees, which is not established in the network growth model with priority connection.This unexpected result can be linked to the extraordinary priority in the process of generating the token generation of a large language model.Each body in our network is given a random name, and the large language model has preference for certain letters sequences, which may be affected by the letters in the alphabet in the training language library.This may be the embodiment of the Benford law when using the ZIPF law or when using the number [27, 28, 29].This preference makes some subjects gain higher popularity and more connections because of the high frequency of its name in large language models.Therefore, the extensive distribution we observe is the result of the inherent prejudice in the GPT3.5-Turbo response, not the result of the priority connection model reproduced in the simulation of the genera.

    In order to verify the suspicion of the impact of the prior probability of GPT3.5-Turbo, we adjust the simulation by randomly changing the name of the main body at each iteration.This also tracks their degree information.This model with a renamed model, combined with node information in the main interface, can generate a complex network structure similar to the actual network, as shown in the left panel of Figure 2, C.When checking the existence time of the node, the network shows the typical correlation between the existence of the time and the degree of the node, which is closer to the results generated by the priority connection process.If we change the interface to hide the node degree information and keep renamed at the same time, then the result is similar to a random network, as shown in the left panel of Figure 2, D, and the right panel shows a narrow degree distribution (pay attention to the linear level levelaxis).Therefore, after renamed the subject in the interface, the model has changed from a complex network structure with degree information to a random network structure that cannot access other subjects.Based on this, the center-radiation structure of unavailable names and possessive information models is a combination of two concentrated processes: one is the priority of the priority connection, and the other is the priority of the token name, which leads to an extreme centralization.network.

    Biddless network

    As mentioned above, when renamed the renames to eliminate the token priority, we found that as long as the subject can see the number of friends of other subjects, the network structure will change from random networks to more complicated forms.Without renamed, as long as the main body can see the number of other subjects, the network will show the topological and radiated topological structure.These phenomena can be understood by considering the different modes of the Barabasi-Albert model [26]. Among them, the possibility of connecting connection with specific nodes is proportional to its degree K. This relationship can be represented by connection probability π (K).

    π (k) μ km

    This mechanism is called "priority connection", that is, nodes with higher degree are more likely to obtain new connections.Among them, the connection index μ plays a decisive role in shaping the network topology [30]:

    When μ = 0, the network is a random network with index distribution.

    When 0 <μ <1, the network presents a degree of stretching index.

    When μ = 1, the network becomes a biddless network, and its degree of distribution features is a power distribution.

    When μ> 1, the network evolution into the structure of the central hub and radiation, is usually vividly described as "the rich is getting rich".

    Typical online social networks have a degree distribution corresponding to linear priority connection.The value that deviates from this narrow range will lead to obviously different distribution.

    Figure 3: Linear priority connection.Figure left: The complementary cumulative distribution functions of the three network scale n = 300, 1000, and 2500 degrees.The real line is the observation guideline, which corresponds to the power probability density function corresponding to the index α = 2.The network shows the biddingless characteristics.Right: The reconstruction of the three network snapshot cumulative connection probability Pi (K) is used as a function of K.As the number of nodes N increases, a longer secondary function curve area appears.This corresponds to the typical linear priority connection in the bonus -free network and human interaction.

    Our genetic main model reflects the laws observed in the Barabasi-Albert model, that is, when the priority connection index μ is greater than zero, the degree distribution gradually evolved from the index type to a wider probability distribution. In the end, at μ> 1In the case, the richer situation of the rich.Our goal is to determine which mode of the network generated by our generation subject (random network of stretching index, no standard network, or center-radiation network).Figure 3 (left) shows the evolution of the complementary cumulative distribution function of the node as the network increases to n = 2000 nodes.In a dual -pairing number coordinates, the linear trend of complementary cumulative distribution functions indicates the biddable topology structure.The solid line in the figure represents the maximum like -like fit for the complementary accumulated distribution function [31, 32], which corresponds to a power distribution of an index α = 1.93 ± 0.12.

    Our research has found that in the online interaction, the main body of the format is like humans, and it is more inclined to spontaneously form a bid -free network.However, due to the relative limitedness of the network scale and the cost of simulating larger networks, we need to conduct more in -depth analysis to confirm whether the genetic subject really shows the characteristics of linear priority connection.We restructured the connection probability π (K); if this probability is linearly related to the degree K, this implies that the network is a non -standard network controlled by a linear priority connection mechanism.We adopted the method overview in [33] to check continuous network snapshots, and calculate the possibility of connecting the new node to the existing node with a specific degree, so as to obtain π (K).In order to reduce noise, we calculated a function of cumulative connection probability π (k) = ∑ kiπ (k) and drawing it into K.In the case of a linear priority connection, this cumulative indicator shows secondary function growth.Figure 3 (right) shows π (K) of the above three networks.These charts revealed that with the growth of N, the shape of the secondary function is becoming more and more significant, which indicates that there is a linear priority connection between the genetic subject.


    Large language models have shown skilled skills in simulating human language, and the texts generated often cannot distinguish from the works created by humans.This ability has stimulated people's interest in building a new generation of main body (ABMS) to use large language models to achieve unprecedented realism [15].Among them, the subject can even be used as a digital twin of social media users.Such progress contains huge potential in various applications such as the intervention of the test social platform [14] to the analysis of epidemic dissemination [34].However, to realize this potential, the key is to understand that the big language model can not only copy human behavior at the individual level, but also show collective behavior that is vital to human society.Because the benchmark tests of most large language models are around the Q & A task, the understanding of group behaviors showing the group behavior of large language models in many individuals interacts with the environment and interaction in many individuals is limited.

    In this study, we focus on a basic emergence mode in human interaction -no standard network to explore this issue.This type of network is generally common in various systems from biology to social and economic areas. The topology characteristics are manifested as a degree of distribution follow the power law. Therefore, most of the nodes have fewer connections, while a few nodes have a lot of connections.Our studies have found that the generatory main model (that is, the large language model used by ChatGPT) driven by GPT3.5-Turbo is similar to human behavior in online social networks.The network of the genetic subject shows a similar linear priority connection tendency to form a biddless network.It is worth noting that the power index of the no-standard network generated by we simulated is close to -2, that is, the famous ZIPF law [27, 29].This value is very close to the values observed by online social networks such as Twitter [24].

    Overall, our research results emphasize the challenge of using large language models as the main body of the format.Large language models are usually described as "Stochastic Parrot" [35]. It is generally believed that they are a system that randomly generates sentences of sentences based on the previous article.However, our research has found that large language models can capture human behavior characteristics that beyond pure text generation, which is usually the main focus of standard and benchmark testing.Compared with a single response evaluation method, our method involves comparison of the distribution of the entire response.We show that in similar situations, humans and generators show similar connection probability distribution.This observation is essential for Montecaro -based models, because they need to be updated multiple times to explore all possible response scope.Therefore, in order to successfully integrate the large language model into the main body -based model, we must ensure that these models can not only copy the real person reaction in specific conditions, but also accurately represent the subject interaction situation of the subject.Overall probability distribution.


    [1] S. Bubeck, V. CHANDRASEKARAN, R. Eldan, J. Gehrke, E. Horvitz, E. Kamar, P. Lee, Y. T. Lee, Y. Li, Et Al., Sparks of Artificial GeneralIntelligence: Early Experiments with GPT-4, Arxiv Preprint Arxiv: 2303.12712 (2023).

    [2] Y. Chang, X. Wang, J. Wang, Y. Wu, K. Zhu, H. Chen, L. Yang, X. YI, C. Wang, Y. Wang, et al., A Survey onEvaluation of large language models, Arxiv Preprint Arxiv: 2307.03109 (2023).

    [3] F. F. XU, U. Alon, G. Neubig, and V. J. Heellendoorn, A Systematic Evaluation of Large Language Models of Code, in Proc themem signature ymposium on Machine Programming (2022) pp. 1–10.

    [4] K. Valmeekam, A. Olmo, S. SREEDHARAN, and S. Kambhampati, Large Language Models Still Can n’t Plan (a Benchmark for LLMS on Planning and Reasoning About CHANGE), AR XIV Preprint Arxiv: 2206.10498 (2022).

    [5] L. P. Argyle, C. A. Bail, E. C. BUSBY, J. R. Gubler, T. HOWE, C. Rytting, T. Sorensen, AND D. Wingate, Leveraging AI for Democrative Discourse: Chat International MprVe Online Polictical Conversations at Scale, Proceedings OFThe National Academy of Sciences 120, E2311627120 (2023).

    [6] B. J. Jansen, S.-G. JUNG, and J. Salminen, Employing Large Language Models in Survey Research, Nature Language Processing Journal 4, 100020 (2023).

    [7] J. S. Park, L. L. CAI, M. R. Morris, P. Liang, and m. S. Bernstein, Social Simulacra: Creating Populated Prototypes for Social Computing Systems, IN PRO Ceedings of the 35th Annual ACM Symposium on User Interface Software and Technology(2022) PP. 1–18.

    [8] A. J. Thirunavukarasu, D. S. J. Ting, K. Elangovan, L. Gutierrez, T. F. Tan, and D. S. W. Ting, Large Language Models in Medicine, Nature Medicine 29, 1930 (2023).

    [9] M. Abdulhai, G. Serapio-Garcia, C. Crepy, D. Valter, J. CanNy, and N. Jaques, Moral Foundations of Large Language Models, Arxiv Preprint Arxiv: 2310.15337 (20222 3).

    [10] K. Singhal, S. Azizi, T. TU, S. S. MAHDAVI, J. Wei, H. W. Chung, N. SCAles, A. Tanwani, H. Cole-Lewis, S. PFOHL, Et Al., Large Language ModelsEncode Clinical Knowledge, Nature 620, 172 (2023).

    [11] M. C. Frank, Baby Steps in Evaluating the Capacities of Large Language Models, Nature Reviews Psychology 2, 451 (2023).

    [12] T. Webb, K. J. Holyoak, and H. Lu, Emergent Analogical Reasoning in Large Language Models, Nature Human Behaviour 7, 1526 (2023).

    [13] G. Betz, Nature-Language Multi-Agent Simulations of Argumentative Opinion Dynamics, Journal of Artificial Society and Simulation 25 (2022).

    [14] P. Tbornberg, D. Valeeva, J. UITERMARK, and C. Bail, Simulating Social Media user Language Models to Evaluate ALGONATIVE New Algorithms, ARXIV preprint arxiv: 2310.05984 (2023).

    [15] I. Group, M. Feinberg, D. C. Parker, N. A. Christakis, P. E. TetLock, and W. A. Cunningham, AI and the Transform of Social 380, Science 380, 380, 380, Science 1108 (2023).

    [16] D. Garcia, M. GALESIC, and H. OLSSON, The Psychology of Collections, Perspectives on Psychology Science, 17456916231201139 (2023).

    [17] P. W. Anderson, More is Different: Broken Symmetry and the Nature of the Hierarchical Structure of Science., Science 177, 393 (1972).

    [18] J. S. Park, J. O'Brien, C. J. Cai, M. R. Morris, P. Liang, and M. S. Bernstein, Generation Agents: Interactive Simulacra of Human Behavior, in Procedings of the 36 TH Annual ACMPOSIUM on User Interface Software and Technology (2023) PP. 1–22.

    [19] H. Zhang, W. Du, J. Shan, Q. ZHOU, Y. DU, J. B. Tenenbaum, T. Shu, and C. Gan, Building Cooperative Embodied Agelarly With Language Models, ARXIV, preprint arxiv: 2307.02485(2023).

    [20] J. He, F. Wallis, and s. Rathje, Homophily in An Artificial Social Network of Agents Powered by Large Language Models (2023).

    [21] A. CAPOCCI, V. D. Servedio, F. Colaiori, L. S. Buriol, D. DONATO, S. Leonardi, and G. Caldarelli, PREFENTIAL AttACHMENT in the Growth of Social Networks: TH E interseted enclopedia wikipedia, Physical Review E 74, 036116(2006).

    [22] R. Pastor-Satorras and A. Vespignani, Epidemic Spreading in Scale-Free Networks, Physical Review Letters 86, 3200 (2001).

    [23] S.-Y. TENG, M.-Y. YEH, and K.-T. Chuang, Toward Undersoning The Mobile Social Properties: An Analysis on Instagram Photo-Sharing Network, In PROCEEDIN GS of the 2015 IEEE/ACM InternationalConference on Advances in Social Networks Analysis and Mining 2015 (2015) pp. 266–269.

    [24] S. Aparicio, J. Villaz´on-Errazas, and G. Alvarez, a Model ´For Scale-Free Networks: Application to Twitter, Entropy 17,5848 (2015).

    [25] J. Luo, Analyzing Character and Consciousness in AIGENTED SOCIAL Content: A Case Study of Chirper, The AI Social Network, Arxiv Preprint Arxiv: 2309.0861 4 (2023).

    [26] A.-L. Barab´asi and R. Albert, Emergeence of Scaling in Random Networks, Science 286, 509 (1999).

    [27] G. K. ZIPF, Human Behavior and the Principle of Least Effort: An INTRODUCTION to Human Ecology (Ravenio Books, 2016).

    [28] L. Pietronero, E. TOSATTI, V. TOSATTI, and A. Vespignani, Explaining The UNEven Distribution of Numbers in Nature: The Laws of Benford and Zipf, PHYSICA A: Stat ISTICAL MeCHANICS and ITS Applications 293, 297 (2001)Then, then

    [29] G. de Marzo, A. Gabrielli, A. Zaccaria, and L. Pietronero, Dynamical Approach to Zipf ’s Law, Physical Review Research 3, 013084 (2021).

    [30] S. N. Dorogovtsev, J. F. F. Mendes, and A. N. Samukhin, Structure of Growing Networks with Preference, Physical Review Letters 85, 4633 (2000).

    [31] A. Clauset, C. R. Shalizi, and M. E. Newman, Powerlaw Distribution in Empirical Data, Siam Review 51, 661 (2009)..

    [32] J. Alstott, E. Bullmore, and d. PLENZ, Powerlaw: A python package for analysis of heavy-taled distribution, PLOS One 9, E85777 (2014).

    [33] M. E. Newman, Clustering and Preference Attachment in Growing Networks, Physical Review E 64, 025102 (2001).

    [34] R. Williams, N. Hosseinichimeh, A. Majumdar, and N. GHAFFARZADEGAN, EPIDEMIC MODELING With GENERATIVE AGENTS, Arxiv Preprint Arxiv (2023)..

    [35] E. M. Bender, T. Gebru, A. MCMILLAN-MAJOR, and S. Shmitchell, on the DANGERS of Stochastic Parrots: Can Language Models Be too big ?, in Proceedings of the 2021 AC M conference on fairness, accountability, and transparency(2021) PP. 610–623.

    Original title: "cutting -edge progress: Social interaction of big models AGENT has emerged without standard networks"

您需要登录后才可以回帖 登录 | 注册


King of Glory| mother and baby| PUBG| parenting| internationality| science and technology| ( 浙ICP备966241366号-1 )

Powered by Popular news sites X3.4

© 2001-2017