A Multi-Agent Social Network Simulation Method and System Based on Cognitive Reasoning Chain

Abstract

This invention relates to the field of artificial intelligence technology and discloses a multi-agent social network simulation method and system based on a cognitive reasoning chain. The method includes: initializing a multi-agent system comprising a social environment engine, a user profile engine, and a cognitive reasoning engine; executing a multi-agent social network simulation loop with a preset number of iterations; in each iteration, the social environment engine pushes social information to the agents as external stimuli and activates the agents to make independent decisions; updating the cognitive states of each dimension in the cognitive reasoning chain and generating corresponding social behaviors through reasoning using a large-scale language model; recording the social behaviors and their corresponding cognitive state trajectories; periodically analyzing historical records to optimize the influence coefficients between the cognitive dimensions of the cognitive reasoning chain; and adjusting the reasoning strategy of the preset large-scale language model. This makes the simulation of the agent's cognitive process a transparent and traceable evolutionary process, achieving a complete and understandable simulation of the "observation-cognition-behavior" cycle.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a multi-agent social network simulation method and system based on cognitive reasoning chains. Background Technology

[0002] With the expansion of the social media ecosystem, it has become a core carrier of information dissemination dynamics, human-computer interaction paradigms, and public opinion formation mechanisms. However, this heterogeneous network environment has given rise to complex social computing problems such as the information cocoon effect, group polarization, and the cascading spread of misinformation, posing a severe challenge to online social governance. To achieve quantitative modeling and controllable intervention in the dynamics of network society, the field of computational social science has gradually formed a simulation technology system based on intelligent agents.

[0003] Early technical solutions were primarily based on the rule-engine-driven agent-based modeling (ABM) paradigm. These models construct agent decision-making logic by explicitly encoding social science theories. Typical applications include opinion evolution models based on social judgment theory and public opinion dynamics models based on bounded trust mechanisms. Within this framework, agents interact with the environment or other agents through state transitions and information exchange based on a predefined rule base, thereby reproducing the complex dynamic characteristics of group behavior in a micro-to-macro emergent process.

[0004] With breakthroughs in generative artificial intelligence technology, the cognitive computing paradigm centered on Large Language Models (LLMs) is reshaping the architecture of social simulation technology. New technical solutions fully leverage the superior capabilities of LLMs in natural language processing, contextual reasoning, and multimodal generation to construct intelligent agent systems with cognitive reasoning, memory storage, and autonomous decision-making abilities. These agents, through knowledge injection and parameter fine-tuning, can achieve human-like social interaction, collaborative software development, and open-world exploration in virtual environments. Building upon this, third-generation social simulation platforms such as OASIS, SocioVerse, and MOSAIC have emerged. These platforms support large-scale parallel interaction of intelligent agents through distributed computing architectures, integrating real user profile data and multimodal alignment algorithms to effectively simulate complex social phenomena such as information diffusion topology and group polarization thresholds. While significant progress has been made in simulating social phenomena, substantial shortcomings remain:

[0005] First, traditional rule-based agent modeling methods rely heavily on subjective rule setting, making it difficult to fully capture the richness, dynamism, and situational adaptability of human behavior. Furthermore, when simulating large-scale, heterogeneous social groups, these models often face bottlenecks due to high computational complexity and poor scalability. Second, while current agent systems based on Large Language Models (LLMs) excel at reproducing observable user behavior, they largely treat the internal cognitive processes of the agent as a "black box." These systems primarily focus on simple behavioral responses to external stimuli, lacking in-depth modeling and characterization of crucial internal psychological elements such as belief formation, emotional shifts, value-based judgments, and specific thought patterns. For example, some simulation studies have revealed inconsistencies between the agent's verbal explanations and their group behavior, exposing the difficulty of capturing genuine decision-making motivations with existing technologies. Because internal cognitive mechanisms are ignored, when the agent cognition depicted by the simulation system deviates significantly from human psychological reality, its value in guiding social science theory and solving real-world problems is greatly diminished. This methodological gap makes it difficult to trace the precise cognitive psychological path that users experience when they go from passively observing to actively or even radically participating in controversial discussions, thus making it impossible to deeply understand the driving forces behind phenomena such as online information dissemination and opinion polarization. Summary of the Invention

[0006] In view of this, in order to overcome the problems of rough cognitive modeling and black box reasoning process in existing social media simulation technology, this invention provides a multi-agent social network simulation method and system based on cognitive reasoning chain. It aims to break the "black box" problem in traditional simulation and accurately simulate and understand the internal cognitive evolution process of users in online social environment by constructing a simulation platform with high psychological transparency and interpretability.

[0007] In a first aspect, the present invention provides a multi-agent social network simulation method based on cognitive reasoning chains, comprising:

[0008] Initialize a multi-agent system including a social environment engine, a user profile engine, and a cognitive reasoning engine; construct a social environment containing multiple agents; configure user profiles for each agent; and establish a cognitive reasoning chain that includes dimensions of emotion, mood, thought, stance, and intention.

[0009] The simulation loop of a multi-agent social network is executed in a preset number of iterations. In each iteration, social information is pushed to the agents as external stimuli through the social environment engine to activate the agents to make independent decisions. The agents update the cognitive state of each dimension in the cognitive reasoning chain through large-scale language model reasoning. Based on the updated cognitive state, corresponding social behaviors are generated and the cognitive state trajectory of the social behaviors and the corresponding cognitive reasoning chain is recorded.

[0010] Periodically analyze the historical records of cognitive state trajectories and behaviors to optimize the influence coefficients between various cognitive dimensions in the cognitive reasoning chain and adjust the reasoning strategies of the pre-set large-scale language model.

[0011] This invention couples demographic information and psychological traits of an intelligent agent through multi-dimensional cognitive modeling. It integrates historical states and external stimuli into a cognitive reasoning chain to achieve human-like cognitive evolution. LLM-driven thought chain reasoning makes the cognitive process transparent and interpretable. Based on dynamic Bayesian networks and least squares, it achieves adaptive optimization of cognitive parameters. At the same time, the modular engine architecture supports flexible configuration and large-scale expansion. Finally, it forms a complete and understandable simulated data loop of the "observation-cognition-behavior" cycle to support computational social science research and platform governance. It has significant advantages such as high realism, strong interpretability, excellent adaptability, high degree of engineering, and wide cross-domain application.

[0012] In one optional implementation, the initialization includes a multi-agent system comprising a social environment engine, a user profiling engine, and a cognitive reasoning engine, comprising:

[0013] Load the parameters required for the simulation from the preset configuration file, including: preset large language model interface information, total number of simulation rounds, number of agents, and data paths for each functional module, and initialize the social environment engine, user profile engine, and cognitive reasoning engine in sequence; among which:

[0014] The initialization process of the social environment engine includes: constructing and instantiating platform objects and social environment databases, creating data tables according to a preset database pattern to store user data, social content data and cognitive state data, and simultaneously initializing the social information content pool;

[0015] The initialization process of the user profile engine includes: creating intelligent agents based on the intelligent agent generator, and assigning a unique user profile to each intelligent agent, wherein the user profile includes demographic information and psychological trait information;

[0016] The initialization process of the user profile engine includes: instantiating a cognitive manager for each agent to initialize the cognitive state of each cognitive dimension in the agent's cognitive reasoning chain, with the initial value being a neutral default value or a personalized setting based on the user profile.

[0017] The initialization scheme provided in this invention achieves scalability and personalized cognitive modeling of the simulation system through a modular engine architecture and parameterized configuration mechanism: It loads parameters such as the LLM interface and simulation rounds from a preset configuration file, enabling the system to flexibly adapt to the simulation needs of social networks of different scales; the social environment engine constructs a database and social information pool, forming a structured data storage base through standardized data table structures, providing underlying support for subsequent behavioral analysis; the user profile engine creates intelligent agents based on a generator, endowing them with composite profiles of demographic information and psychological traits, laying the foundation for modeling individual cognitive differences; the cognitive reasoning engine initializes five-dimensional cognitive states such as emotion and mood through a cognitive manager, supporting neutral default values ​​and profile-driven personalized settings, ensuring model startup efficiency while achieving differentiation of cognitive starting points through profile association, ultimately forming an initialization system of "configurable parameters - storable data - distinguishable individuals," providing a standardized and flexible infrastructure for high-precision social simulation.

[0018] In one optional implementation, the step of pushing social information to the agent as an external stimulus via a social environment engine to activate the agent to perform independent decision-making includes:

[0019] The social environment engine distributes observation information for this round to each agent from the social information content pool based on a preset information recommendation strategy.

[0020] Activate each agent to update its cognitive state and make behavioral decisions based on the distributed observation information.

[0021] The social environment engine of this invention, based on a preset recommendation strategy, accurately distributes observation information to agents from a social information content pool, making the external stimuli received by different agents context-specific and replicating the information filtering mechanism of real-world social networks. It activates agents to independently execute cognitive updates and behavioral decisions based on the distributed information, ensuring that each agent's cognitive evolution path is driven by both personalized profiles and real-time information stimuli, avoiding homogeneous simulation of group behavior. This mechanism constructs a closed-loop link of "information stimulus-cognitive response-behavioral output," ensuring both the overall synergy of the simulation system and preserving the uniqueness of individual cognition through independent decision-making. Simultaneously, the standardized information distribution process improves the repeatability and controllability of the simulation process.

[0022] In one optional implementation, updating the cognitive state of each dimension in the cognitive reasoning chain through large-scale language model reasoning includes:

[0023] Generate cognitive guidance prompts, which include the agent's user profile, data memory of a preset number of historical iterations, currently observed social information content, the cognitive state at the previous moment, and the influence coefficients between various cognitive dimensions;

[0024] The cognitive guidance prompts are input into a preset large-scale language model for cognitive reasoning chain evolution and update, obtaining the cognitive state of each cognitive dimension at the current moment. The evolution and update of each dimension are represented by the following functional relationship:

[0025] The formula for the evolution of emotion is: M t =f M (M t-1 ,P M ,ξ t,M ,∈ M,t );

[0026] The formula for the evolution of emotion is: E t =f E (M t E t-1 ,ξ t,E ,∈ E,t );

[0027] The formula for the evolution of thought is: T t =f T (M t E t ,T t-1 ,P T ,ξ t,T ,∈ T,t );

[0028] The formula for the evolution of a position is: S t =f S (T t ,S t-1 ,P S ,ξ t,S ,∈ S,t );

[0029] The formula for the evolution of intention is: I t =f I (S t ,I t-1 ,ξ t,I ,∈ I,t );

[0030] Where: t represents the current time step; M t E t ,T t ,S t ,I t M represents the state of emotion, mood, thought, stance, and intention at time step t; t-1 E t-1 ,T t-1 ,S t-1 ,I t-1 These represent the states of each cognitive dimension at the previous time step t-1; fX (·) represents the state transition function of dimension X, which is implemented through inference using a pre-defined large-scale language model; P X ξ represents the stable individual traits that influence dimension X; t,X This represents the influence of the external environment on dimension X at time step t; ∈ X,t This represents the random noise term applied to dimension X at time step t, used to simulate random influencing factors in human cognition that are not explicitly captured by the model, thereby increasing the realism of the simulation.

[0031] This invention integrates user profiles, historical cognitive states, real-time social information, and dynamic influence coefficients into cognitive guidance prompts, constructing a multi-dimensional input system covering individual traits, historical inertia, and environmental stimuli. This allows LLM reasoning to closely align with the real-world mechanism of human cognition being influenced by multiple factors. Furthermore, by pre-setting mathematical functions to clarify the evolutionary logic of each cognitive dimension, the LLM reasoning process is constrained within a quantifiable theoretical framework. This leverages the advantages of LLM natural language processing while avoiding the randomness of reasoning results, ensuring the scientific rigor of cognitive state updates. In addition, the introduction of random noise terms simulates the unpredictability of human cognition, enhancing the realism of the simulation while enabling structured recording and retrospective analysis of the cognitive process. This overcomes the limitations of traditional black-box models, providing computational social science research with an innovative tool that combines dynamic adaptability and verifiable results.

[0032] In one optional implementation, the step of generating corresponding social behaviors based on the updated cognitive state and recording the cognitive state trajectory of the social behaviors and the corresponding cognitive reasoning chain includes:

[0033] The decision-making behavior type is generated based on the intent dimension. When content needs to be generated, the large language model is called based on the current complete cognitive state to generate text that matches the state.

[0034] The decision-making behavior is executed in a simulated environment by a behavior actuator, and the decision-making behavior and cognitive state trajectory are recorded to a database by a recorder.

[0035] This invention uses the intent dimension of the cognitive reasoning chain as the core of behavioral decision-making. Through a large language model, it generates text content that aligns with emotions and stances based on a complete cognitive state, ensuring a high degree of consistency between the agent's behavior and its internal cognition, significantly improving the realism of the simulated scenario. Simultaneously, by using behavior executors and recorders, it maps decision-making behaviors to the simulated environment and structurally stores behavioral and cognitive state trajectory data, providing traceable and quantifiable empirical evidence for computational social science research and platform strategy evaluation. The modular design ensures that the system can be flexibly expanded to include new behavior types and data formats, enhancing the versatility and engineering adaptability of the simulation system.

[0036] In one optional implementation, the periodic analysis of cognitive state trajectories and historical records of behavior is used to optimize the influence coefficients between various cognitive dimensions of the cognitive reasoning chain and adjust the reasoning strategy of a preset large-scale language model, including:

[0037] Dynamic dependencies between cognitive dimensions are modeled using dynamic Bayesian networks, where the structure of the dynamic Bayesian network is determined based on the first-order Markov property and the assumption of regular stability of state transitions.

[0038] Collect the cognitive state trajectories and corresponding decision-making behavior history records of all agents within a preset iteration round;

[0039] The influence coefficients between cognitive dimensions are calculated using ordinary least squares regression, including: establishing a linear regression equation for each cognitive dimension, and using historical data as input to calculate the optimal coefficient value using ordinary least squares.

[0040] Cognitive guidance prompts for the cognitive reasoning process are updated based on the influence coefficients between cognitive dimensions, thereby optimizing the reasoning strategy of large language models.

[0041] This invention employs a dynamic Bayesian network to model the dynamic dependencies between cognitive dimensions. Combined with ordinary least squares regression analysis of the agent's cognitive state and behavioral data across preset rounds, it precisely quantifies the influence coefficients between each cognitive dimension. The calculation results are then integrated into cognitive guidance prompts to optimize the reasoning strategy of a large-scale language model. This allows the system to dynamically and quantitatively learn the strength of influence between cognitive dimensions based on the agent's interaction history data. This not only makes the simulation more closely resemble real-world psychological dynamics but also significantly enhances the model's adaptability, enabling it to capture the evolution of cognitive influence paths in different social contexts.

[0042] Secondly, the present invention provides a multi-agent social network simulation system based on cognitive reasoning chain, the system comprising:

[0043] The system initialization module is used to initialize a multi-agent system including a social environment engine, a user profile engine, and a cognitive reasoning engine, to build a social environment containing multiple agents, to configure user profiles for each agent, and to establish a cognitive reasoning chain that includes dimensions of emotion, mood, thought, stance, and intention.

[0044] The simulated loop execution module is used to execute a multi-agent social network simulation loop with preset iterations. In each iteration, social information is pushed to the agents as external stimuli through the social environment engine to activate the agents to make independent decisions. The cognitive state of each dimension in the cognitive reasoning chain is updated through large-scale language model reasoning. The corresponding social behavior is generated based on the updated cognitive state, and the social behavior and the cognitive state trajectory of the corresponding cognitive reasoning chain are recorded.

[0045] The dynamic optimization module is used to periodically analyze the historical records of cognitive state trajectories and behaviors, optimize the influence coefficients between various cognitive dimensions in the cognitive reasoning chain, and adjust the reasoning strategies of the preset large-scale language model.

[0046] Thirdly, the present invention provides a computer device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the multi-agent social network simulation method based on cognitive reasoning chain described in the first aspect or any corresponding embodiment.

[0047] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to execute the multi-agent social network simulation method based on cognitive reasoning chain described in the first aspect or any corresponding embodiment thereof.

[0048] Fifthly, the present invention provides a computer program product, including computer instructions for causing a computer to execute the multi-agent social network simulation method based on cognitive reasoning chain described in the first aspect or any corresponding embodiment thereof. Attached Figure Description

[0049] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0050] Figure 1 This is a flowchart illustrating a multi-agent social network simulation method based on cognitive reasoning chain according to an embodiment of the present invention.

[0051] Figure 2 This is a general architecture diagram of a multi-agent system according to an embodiment of the present invention;

[0052] Figure 3 Structural block diagram of a multi-agent social network simulation system based on cognitive reasoning chain according to an embodiment of the present invention;

[0053] Figure 4 A schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] To overcome the problems of coarse cognitive modeling (such as insufficient realism due to rule-based fixed behavioral presuppositions), black-box reasoning processes (traditional models struggle to explain cognitive evolution logic), and poor dynamic adaptability (static models cannot cope with the dynamic changes of complex social phenomena) in existing social media simulation technologies, this embodiment provides a multi-agent social network simulation method based on cognitive reasoning chains. It aims to achieve refined modeling of the complete psychological chain of users, from information perception, emotional response, cognitive thinking to stance formation and behavioral intention generation. This allows for the tracking and analysis of individuals' cognitive evolution paths in complex social interactions, providing a powerful tool for revealing the deep cognitive driving mechanisms of online group behavior (such as opinion polarization and information dissemination).

[0056] Figure 1 This is a flowchart of a multi-agent social network simulation method based on cognitive reasoning chain according to an embodiment of the present invention. Figure 1 As shown, the process includes the following steps:

[0057] Step S1: Initialize a multi-agent system including a social environment engine, a user profile engine, and a cognitive reasoning engine; construct a social environment containing multiple agents; configure user profiles for each agent; and establish a cognitive reasoning chain that includes dimensions of emotion, mood, thought, stance, and intention.

[0058] Specifically, this invention provides a multi-agent system (Agent-based Social Cognition Evolution, ASCE). For example... Figure 2 As shown, the ASCE system architecture comprises three core modules: the Social Environment Engine (SEE), the User Profile Engine (UPE), and the Cognitive Reasoning Engine (CRE). Among them:

[0059] The Social Environment Engine (SEE) is responsible for simulating the information flow of real social media platforms, dynamically generating and updating social information (such as trending posts), managing user interactions, and integrating information feedback mechanisms. This engine periodically introduces real-world news events and trending topics, transforming them into structured social media information and managing them uniformly within a global social information content pool for agent interaction.

[0060] The User Profile Engine (UPE) builds and dynamically maintains personalized user profiles for each agent. These profiles not only include demographic attributes but also integrate higher-order psychological traits such as cognitive style, emotional sensitivity, and value preferences. These traits collectively determine the agent's personalized cognitive response to external information stimuli in specific situations.

[0061] The Cognitive Reasoning Engine (CRE) is used to construct and drive the agent's internal cognitive reasoning path from information perception to behavioral response. It provides a Cognitive Reasoning Chain (CRC), which decomposes the complex cognitive process into five core cognitive dimensions: Mood, Emotion, Thinking, Stance, and Intention. These five dimensions unfold sequentially, and their theoretical basis integrates several classic social psychology theories. For example, the Affective Imprinting Model (AIM) explains how emotion influences emotional and cognitive processes, while the Theory of Planned Behavior (TPB) describes the cognitive process from thought to stance, and then to intention and behavior. Please refer to Table 1 for the specific values ​​of cognitive states.

[0062] Table 1

[0063]

[0064]

[0065]

[0066] During system initialization, all parameters required for the simulation are first loaded into a pre-configured file (e.g., a YAML file named test_dei.yaml). These parameters include: the interface information of the Large Language Model (LLM) used, the total number of simulation rounds, the number of agents, and the data paths for each module. Then, the three core engines are initialized sequentially:

[0067] 1. The initialization process of the social environment engine includes: building and instantiating platform objects (such as the Platform class in platform.py) and social environment database. It connects to the database and creates data tables according to the preset database schema (such as schema / *.sql) to store user data, social content data (posts, comments, interaction behaviors) and cognitive state data. At the same time, it initializes the social information content pool (loads the initial social media posts to form a global post pool) to provide the content foundation for subsequent simulations.

[0068] 2. The initialization process of the user profile engine includes: creating agents based on the agent generator (such as agents_generator.py) and assigning a unique user profile to each agent. The user profile is a structured data body that not only contains demographic information such as age and gender, but also defines its key intrinsic psychological traits, such as cognitive style (biased towards analysis or intuition), values ​​and emotional sensitivity, which are the basis for subsequent personalized cognitive simulation.

[0069] 3. The initialization process of the cognitive reasoning engine includes: instantiating a cognitive manager for each agent to initialize the cognitive state of each cognitive dimension in the agent's cognitive reasoning chain. The initial value is a neutral default value or a personalized setting based on the user profile.

[0070] Step S2: Execute a multi-agent social network simulation loop with preset iterations. In each iteration, social information is pushed to the agents as external stimuli through the social environment engine to activate the agents to make independent decisions. The agents also update the cognitive states of each dimension in the cognitive reasoning chain through large-scale language model reasoning. Based on the updated cognitive states, corresponding social behaviors are generated, and the social behaviors and the corresponding cognitive state trajectories of the cognitive reasoning chain are recorded.

[0071] Specifically, after initialization, the multi-agent system enters the main simulation loop, which iterates according to a preset number of rounds (timesteps). The following steps are performed in each round:

[0072] a1: Information distribution. The recommender module in the Social Environment Engine (SEE) is activated. This module filters and recommends information (i.e., one or more posts) that each agent will "observe" in this round based on a pre-defined algorithm (such as the popularity of posts or their matching degree with the agent's interests).

[0073] a2: Agent activation. The system iterates through all agents, triggering each agent's independent decision-making process sequentially. Specifically, for each agent, its core `step` method is executed, passing in the post distributed to it in the previous stage as an external information stimulus. This is completed within each agent's `step` method and can be further subdivided into the following sub-steps:

[0074] a21: Cognitive State Update, which is an evolutionary model based on the Cognitive Chain Reasoning (CRC) model. Its core is to invoke the cognitive manager within the intelligent entity to update the state of its cognitive chain. This update process is not a simple state switch, but follows a rigorous mathematical model inspired by psychological theory.

[0075] The cognitive reasoning chain (CRC) proposed in this invention simulates the internal cognitive state vector X of an agent at each time step t. t ={M t E t ,T t ,S t ,I t The dynamic evolution process of}. This vector contains five core dimensions: Mood (M), Emotion (E), Thinking (T), Stance (S), and Intention (I). At each time t, the current state of each dimension is influenced by the following factors:

[0076] (1) Historical dependence: The state of this dimension at the previous moment t-1 reflects the inertia of cognition.

[0077] (2) Immediate impact: In the cognitive chain, the current state of a preceding dimension affects subsequent dimensions (e.g., the current emotion E). t It will affect your current thinking. t ).

[0078] At each time step t, the state X(t) of a cognitive dimension X is influenced by its previous state X(t-1) and the current state of the dimensions preceding it on the chain. Specifically: Emotion M(t) represents the agent's long-term, universal emotional background, primarily based on its previous state M(t-1). Sentiment E(t) captures short-term, situation-specific emotional responses, influenced by the current emotion M(t) and its previous state E(t-1). Thought T(t) refers to the agent's reasoning style and attentional focus when processing information, jointly influenced by the current emotion M(t), sentiment E(t), and its past state T(t-1). Stance S(t) is the agent's evaluative attitude towards a specific issue, driven by the current thought T(t) and its previous stance S(t-1). Intention I(t) represents the agent's readiness to perform a specific action, primarily guided by the current stance S(t) and its previous state I(t-1).

[0079] (3) External stimuli: The impact of external information currently received by the agent. For example, in a simulated social media platform, agent A receives a post titled "New research confirms that a certain health method can extend lifespan by 20 years." The exaggerated statements and lack of authoritative data support in the post are the external stimuli currently received by agent A. If the post is full of emotional language (such as "Amazing discovery! It's too late if you don't try it now"), it may directly stimulate agent A's emotional dimension, causing its emotional state to change from "calm" to "curious" or "skeptical." The core viewpoint of the post will further affect agent A's thinking, driving it to judge the authenticity of the information, and thus change its stance dimension (such as from "neutral" to "support" or "oppose").

[0080] (4) Individual traits: The inherent and long-term stable psychological characteristics of an agent. For example, agents with different individual traits will have different reactions to the same health post. If agent B has the individual traits of "cautious cognitive style" and "low emotional sensitivity", even when faced with exaggerated external stimuli, its emotional fluctuations will be small, and it will be more inclined to use its rational analysis ability to check the source of the post and the authenticity of the data before forming a stance; while agent C has "sensitive emotional traits" and "intuitive cognitive style", it may be quickly attracted by the emotional expression of the post, and its emotional dimension will quickly turn to "excitement", and it will form a "supportive" attitude in the stance dimension based on intuitive judgment, which reflects the underlying influence of individual traits on the cognitive process.

[0081] (5) Randomness: Simulating random noise from unobservable factors in reality. For example, when a sudden hot topic appears on the simulation platform, agent D should maintain continuous attention and participate in the discussion based on external event information and its own stance. However, the effect of random noise may cause it to suddenly interrupt its attention and turn to other topics. This corresponds to the attention shifts and behavioral changes caused by unobservable factors such as users being temporarily attracted by other information or being disturbed by sudden life events in reality. This avoids the agent's behavior in the simulation being too mechanical and deterministic, and is closer to the complexity and diversity of user behavior in real social networks.

[0082] (6) The theoretical basis of this evolutionary process integrates the Affective Imprint Model (AIM) and the Theory of Planned Behavior (TPB), and the evolution of its various dimensions can be formally described by the following functional relationships:

[0083] The formula for the evolution of mood is: M t =f M (M t-1 ,P M ,ξ t,M ,∈ M,t );

[0084] The formula for the evolution of emotion is: E t =f E (M t E t-1 ,ξ t,E ,∈ E,t );

[0085] The formula for the evolution of thinking is: T t =f T (M t E t ,T t-1 ,P T ,ξ t,T ,∈ T,t );

[0086] The evolution formula for stance is: S t =f S (T t ,S t-1 ,P S ,ξ t,S ,∈ S,t );

[0087] The evolution formula for intention is: I t =f I (S t ,I t-1 ,ξ t,I ,∈ I,t);

[0088] Where: t represents the current time step (simulation round); M t E t ,T t ,S t ,I t M represents the state of emotion, mood, thought, stance, and intention at time step t; t-1 E t-1 ,T t-1 ,S t-1 ,I t-1 These represent the states of each cognitive dimension at the previous time step t-1; f X (·) represents the state transition function of dimension X, which is implemented through inference using a pre-defined large-scale language model; P X Represents the stable individual traits that influence dimension X, for example, P M P may represent an individual's underlying optimistic or pessimistic tendency. T This represents their inherent cognitive style (such as preference analysis or intuition). This parameter plays a role in relatively stable dimensions such as emotion, thinking, and stance; ξ t,X This represents the influence of the external environment on dimension X at time step t, typically referring to the stimuli brought about by the content, emotional tone, and other features of social media posts currently observed by the agent; ∈ X,t This represents the random noise term applied to dimension X at time step t, used to simulate random influencing factors in human cognition that are not explicitly captured by the model, thereby increasing the realism of the simulation.

[0089] a22: Cognitive guidance prompt generation, in order to computationally achieve the above f X The `(·)` function dynamically constructs a structured "Cognitive Guidance Prompt," which translates the mathematical model into natural language instructions and sends them to a large language model (LLM) for inference. This prompt integrates:

[0090] (1) Agent profiling (P) X For example, "You are a cautious person whose thinking leans towards logical analysis."

[0091] (2) External stimulus (ξ) t,X ): The currently observed full text of the post. Historical Status (X) t-1 ): The agent's complete CRC state at the previous moment, such as "your emotion at the previous moment was 'optimistic', and your mood was 'calm'...".

[0092] (3) Dynamic Influence Coefficient (from DBN): A coefficient β learned in subsequent step S3 that quantifies the influence strength between various cognitive dimensions. For example, the prompt might include: "Please note that, based on recent interaction data, the influence coefficient of emotion on thinking patterns is β." T,E ".

[0093] a23: After receiving this prompt, the LLM is instructed to strictly follow the internal logic of CRC (i.e., the order of M→E→T→S→I), reason based on the thought chain, and finally output the new cognitive state vector X at the current time t in a structured JSON format. t ={M t E t ,T t ,S t ,I t}

[0094] a24: State resolution and assignment, specifically including:

[0095] a241: The cognitive manager receives structured JSON data returned by the LLM, parses it, and updates the new state values ​​(such as "positive", "negative", etc.) to the corresponding five cognitive dimension variables inside the agent, completing a single step of cognitive evolution;

[0096] a242: After the cognitive state is updated, the agent adjusts its latest intention dimension I. t The state value is used to decide the action to be taken. For example, if I t If the action is "expressive," it may trigger a "comment" action; if it is "observative," it will trigger a "remain silent" action. If the decision-making action requires generating natural language content (such as commenting or writing a new article), the system will make a second LLM call. This call will include richer contextual information, including: the agent's user profile, the original post content, and an updated, complete current cognitive state vector X. t .

[0097] LLM generates content that perfectly matches the individual's current mood and stance based on their complete cognitive state. For example, someone experiencing "anger" or "T" in their emotional state... t For "critical thinking", S t An agent with a "radical" stance will generate a sharply worded and opinionated critical comment. The agent's final action (and the generated content) is executed by an action actuator in a simulated environment. For example, the newly generated comment is added to the comment section of the corresponding post in the database. Simultaneously, the action itself, and the complete CRC cognitive state vector X driving the action, are... tThe state and its evolution trajectory will be completely stored in the database by the Action Recorder, providing a data foundation for subsequent dynamic learning of cognitive models.

[0098] Step S3 involves periodically analyzing the historical records of cognitive state trajectories and behaviors to optimize the influence coefficients between various cognitive dimensions of the cognitive reasoning chain and adjust the reasoning strategies of the preset large-scale language model.

[0099] Specifically, to overcome the limitations of static models and make the cognitive simulation process more adaptive, this embodiment of the invention introduces a parameter learning mechanism based on a Dynamic Bayesian Network (DBN). The system periodically (e.g., configured to execute every two rounds) updates the influence coefficients between CRC dimensions. The specific process is as follows:

[0100] S31 abstracts the evolution of the cognitive reasoning chain (CRC) into a DBN model. This model explicitly describes the cognitive state vector X. t Dependencies over time and inter-dimensional dependencies within a given moment. To simplify modeling, the following two basic assumptions are made:

[0101] 1. First-Order Markov Property: The cognitive state X at any time t t It depends only on the state X of its previous time step t-1. t-1 And it is independent of earlier historical states. Its probability is expressed as: P(X... t |X t-1 ,X t-2 ,...)=P(X t |X t-1 ).

[0102] 2. Stationarity: The law governing state transitions, i.e., P(X) t |X t-1 The probability model remains unchanged throughout the simulation.

[0103] Based on the above assumptions and the inherent logic of CRC, the structure of DBN is defined as follows:

[0104] 1. The joint probability distribution of the cognitive chain in the initial state (t=0, i.e., in round 0) can be factored as follows:

[0105] P(M0)=P(M0)·P(E0|M0)·P(T0|M0,E0)·P(S0|T0)·P(I0|S0),

[0106] 2. When t>0 rounds, the state transition distribution is as follows:

[0107] P(X t |X t-1 )=P(M t |M t-1 )·P(E t |M t E t-1 )·P(T t |M t E t ,T t-1 )·P(S t |T t ,S t-1 )·

[0108] P(I t |S k ,I t-1 ).

[0109] This expression clearly demonstrates the temporal dependencies across time steps in the cognitive chain (such as M). t-1 →M t ) and immediate dependencies within a single time step (such as M) t →E t The complete joint probability distribution of the entire cognitive state sequence of length T can be represented as: After defining the DBN model structure, the goal of the system is to use the data generated during the simulation to quantitatively estimate the parameters in the model, namely the influence coefficients (β values) between the cognitive dimensions.

[0110] S32, Data Collection: Extract the cognitive state history records of all agents from the database for the most recent N rounds (the specific value of which should be set reasonably according to the actual application) {X0,X1,...,X...} t}

[0111] S33, Parameter Estimation: This invention uses Ordinary Least Squares (OLS) regression to estimate these influence coefficients. The specific steps are as follows: First, the discrete, categorized cognitive states (e.g., "optimistic," "angry") defined in Table 1 are numerically mapped (e.g., by assigning ordered values). Then, a linear regression equation is established for each cognitive dimension. Taking the Thinking dimension as an example, its regression model can be set as follows: Where ∈ represents random error. This module uses collected historical data as input and calculates the optimal coefficient value, such as β, through OLS regression. T,M (Measures the strength of the influence of current emotions on thinking), β T,E (measuring the strength of the impact of current emotions on thinking) and (Measure the strength of historical inertia in thinking). Perform similar operations on other dependencies in CRC.

[0112] Furthermore, the calculated series of β coefficients are updated and saved. In the "Cognitive Guidance Tip Generation" stage of subsequent simulation rounds (i.e., step a22), these new, more quantifiable coefficients that better reflect the current simulation dynamics will be integrated into the cognitive guidance tips. For example, the updated tip might instruct LLM: "When deriving your current thought pattern, please give priority to your current emotions (influence weighted by β)." T,E ) and sentiment (influence weighted by β) T,M This allows the agent's subsequent cognitive reasoning process to dynamically adapt to new patterns emerging from group interactions, rather than adhering to a set of initially set static rules.

[0113] By employing a dynamic update mechanism that combines Dynamic Bayesian Networks (DBNs) with Ordinary Least Squares (OLS), the system can dynamically and quantitatively learn the strength of influence between cognitive dimensions based on the agent's interaction history data. This not only makes the simulation more closely resemble real-world psychological dynamics but also greatly enhances the model's adaptability, enabling it to capture the evolution of cognitive influence paths in different social contexts.

[0114] Compared with the prior art, the present invention has achieved significant positive effects:

[0115] 1. Regarding the fidelity of simulating psychological states, the method provided by this invention achieves high accuracy. Experimental results show that the ASCE system proposed in this invention achieves an average accuracy of approximately 80% in modeling psychological states. This demonstrates that the cognitive reasoning chain and dynamic update mechanism proposed in this invention can effectively capture and simulate the complex internal psychological dynamics of humans.

[0116] 2. This invention significantly enhances the interpretability of social simulations. By decomposing cognitive processes into five observable dimensions and quantifying the influence between them, this invention opens the "black box" of agent decision-making, enabling researchers to clearly understand the micro-cognitive drivers behind group phenomena.

[0117] 3. This invention can discover and explain complex social cognitive phenomena. For example, in controversial online environments, this invention successfully reveals a multi-stage "cognition-to-behavior activation path" and observes how differences in initial thought patterns lead to agents developing different cognitive paths and differentiated social influences. These findings demonstrate the powerful ability of this invention to model and explain online cognitive dynamics in complex social environments, providing profound insights for social science research and platform governance that far exceed those of previous technologies. This overcomes the limitation of existing technologies, which can only reproduce macroscopic phenomena but cannot explain their underlying cognitive mechanisms.

[0118] This invention provides a specific implementation example to illustrate the execution process of a multi-agent social network simulation method. This embodiment simulates the user's reaction to a controversial topic on a social media platform.

[0119] Step 1: System Initialization. The Social Environment Engine (SEE) generates a social media post about a controversial policy. Simultaneously, the User Profile Engine (UPE) creates multiple agents and assigns them different personalized user profiles. For example, some agents have a cognitive style biased towards analytical thinking, while others are biased towards intuitive thinking, and they have different initial emotions (such as optimism or pessimism).

[0120] Step Two: First Round of Interaction. An "analytical" agent first observes the post. The Cognitive Reasoning Engine (CRE) begins to work. Based on its Cognitive Reasoning Chain (CRC), the agent's cognitive state begins to evolve. For example, its neutral sentiment and the controversial content of the post work together to arouse its curiosity (emotion). This sentiment and its analytical thinking traits guide it to perform logic and fact-checking (thinking), thus forming a neutral-to-supportive stance (stance) and generating the intention to post a comment to seek more information (intent). The CRE combines this series of cognitive state changes with the influence coefficients learned through the DBN (initial or previous update) to generate a cognitive guidance prompt, which is then provided to the LLM. Based on this prompt, the LLM generates an objective, evidence-seeking comment and publishes it.

[0121] Step 3: Subsequent Interactions and Dynamic Evolution. Other agents (including "intuitive" agents) observe the original post and the first comment. They also undergo a similar cognitive evolution process. For example, an "intuitive" agent might quickly develop strong negative emotions due to the inflammatory language of the post, skipping deep thinking, directly forming an opposing stance, and intending to post an emotionally charged rebuttal comment.

[0122] Step 4: DBN Parameter Update. After two rounds of interaction, the system collects the cognitive state sequence data of all agents during this period (i.e., the state of emotion, mood, thought, stance, and intention at each time step). CRE uses this data to recalculate the influence coefficients (β values) between the various cognitive dimensions in DBN through OLS regression. For example, the system may find that in a controversial environment, the influence weight of emotion on thought increases significantly.

[0123] Step 5: Adaptive Adjustment of Cognitive Guidance. The newly calculated influence coefficient is used to update the content of the cognitive guidance cues. In subsequent simulation rounds, when the agent makes decisions again, this new cue guides the LLM to more accurately reflect the cognitive logic of the current interactive environment (e.g., increased emotional influence), thus making its behavioral decisions more consistent with the dynamically evolving psychological reality.

[0124] Through the above steps, this invention realizes a dynamic, adaptive social simulation with transparent cognitive processes.

[0125] This embodiment also provides a multi-agent social network simulation system based on cognitive reasoning chains. This system is used to implement the above embodiments and preferred embodiments, and details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that implements a predetermined function. Although the system described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0126] This embodiment provides a multi-agent social network simulation system based on cognitive reasoning chains, such as... Figure 3 As shown, it includes:

[0127] The system initialization module 31 is used to initialize a multi-agent system including a social environment engine, a user profile engine, and a cognitive reasoning engine, to build a social environment containing multiple agents, to configure user profiles for each agent, and to establish a cognitive reasoning chain that includes dimensions of emotion, mood, thought, stance, and intention.

[0128] The simulation loop execution module 32 is used to execute a multi-agent social network simulation loop with preset iterations. In each iteration, social information is pushed to the agent as an external stimulus through the social environment engine to activate the agent to make independent decisions. The agent also updates the cognitive state of each dimension in the cognitive reasoning chain through large-scale language model reasoning. Based on the updated cognitive state, the agent generates corresponding social behaviors and records the cognitive state trajectory of the social behaviors and the corresponding cognitive reasoning chain.

[0129] The dynamic optimization module 33 is used to periodically analyze the historical records of cognitive state trajectories and behaviors, optimize the influence coefficients between various cognitive dimensions of the cognitive reasoning chain, and adjust the reasoning strategy of the preset large-scale language model.

[0130] In some optional implementations, the system initialization module 31 includes:

[0131] The parameter initialization unit is used to load the parameters required for simulation from a preset configuration file, including: preset large language model interface information, the total number of simulation rounds, the number of agents, and the data paths of each functional module, and sequentially initializes the social environment engine, user profile engine, and cognitive reasoning engine; wherein:

[0132] The social environment engine initialization unit is used to build and instantiate platform objects and social environment databases, and create data tables according to the preset database pattern to store user data, social content data and cognitive state data, and initialize the social information content pool at the same time.

[0133] The user profile engine initialization unit is used to create intelligent agents based on the intelligent agent generator and assign a unique user profile to each intelligent agent. The user profile includes demographic information and psychological trait information.

[0134] The cognitive reasoning engine initialization unit is used to instantiate a cognitive manager for each agent to initialize the cognitive state of each cognitive dimension in the agent's cognitive reasoning chain. The initial value is a neutral default value or a personalized setting based on the user profile.

[0135] In some alternative implementations, the simulated loop execution module 32 includes:

[0136] The information distribution unit, based on a preset information recommendation strategy, distributes the observation information for this round to each agent from the social information content pool.

[0137] The agent activation unit is used to activate each agent to update its cognitive state and make behavioral decisions based on the distributed observation information.

[0138] A cognitive guidance prompt generation unit is used to generate cognitive guidance prompts, which include the user profile of the intelligent agent, the data memory of a preset number of historical iterations, the currently observed social information content, the cognitive state at the previous moment, and the influence coefficients between various cognitive dimensions.

[0139] The cognitive reasoning chain evolution unit is used to input the cognitive guidance prompts into a preset large-scale language model for cognitive reasoning chain evolution updates, and to obtain the cognitive state of each cognitive dimension at the current moment. The evolution update of each dimension is represented by the following functional relationship:

[0140] The formula for the evolution of emotion is: M t =f M (M t-1 ,P M ,ξ t,M ,∈ M,t );

[0141] The formula for the evolution of emotion is: Et =f E (M t E t-1 ,ξ t,E ,∈ E,t );

[0142] The formula for the evolution of thought is: T t =f T (M t E t ,T t-1 ,P T ,ξ t,T ,∈ T,t );

[0143] The formula for the evolution of a position is: S t =f S (T t ,S t-1 ,P S ,ξ t,S ,∈ S,t );

[0144] The formula for the evolution of intention is: I t =f I (S t ,I t-1 ,ξ t,I ,∈ I,t );

[0145] Where: t represents the current time step; M t E t ,T t ,S t ,I t M represents the state of emotion, mood, thought, stance, and intention at time step t; t-1 E t-1 ,T t-1 ,S t-1 ,I t-1 These represent the states of each cognitive dimension at the previous time step t-1; f X (·) represents the state transition function of dimension X, which is implemented through inference using a pre-defined large-scale language model; P X ξ represents the stable individual traits that influence dimension X; t,X This represents the influence of the external environment on dimension X at time step t; ∈ X,t This represents the random noise term applied to dimension X at time step t, used to simulate random influencing factors in human cognition that are not explicitly captured by the model, thereby increasing the realism of the simulation.

[0146] The decision behavior generation unit is used to generate decision behavior types based on the intent dimension. When content needs to be generated, it calls the large language model based on the current complete cognitive state to generate text that matches the state.

[0147] The decision-making behavior execution and recording unit is used to execute decision-making behavior in a simulated environment through a behavior executor, and to record the decision-making behavior and cognitive state trajectory to the database through a recorder.

[0148] In one optional implementation, the dynamic optimization module 33 includes:

[0149] The dynamic Bayesian network modeling unit is used to model the dynamic dependencies between cognitive dimensions through dynamic Bayesian networks. The structure of the dynamic Bayesian network is determined based on the first-order Markov property and the assumption of regular stability of state transitions.

[0150] The historical data acquisition unit is used to collect the cognitive state trajectory and corresponding decision behavior history of all agents within a preset iteration round;

[0151] The influence coefficient calculation unit is used to calculate the influence coefficients between cognitive dimensions using ordinary least squares regression, including: establishing a linear regression equation for each cognitive dimension, and using historical data as input to calculate the optimal coefficient value using ordinary least squares.

[0152] The model optimization unit is used to update the cognitive guidance prompts for the cognitive reasoning process based on the influence coefficients between cognitive dimensions, thereby optimizing the reasoning strategy of large language models.

[0153] Further functional descriptions of the above modules and units are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0154] In this embodiment, the multi-agent social network simulation system based on cognitive reasoning chain is presented in the form of functional units. Here, a unit refers to an ASIC (Application Specific Integrated Circuit), a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0155] This invention also provides a computer device having the above-described features. Figure 3 The example shown is a multi-agent social network simulation system based on cognitive reasoning chain.

[0156] Please see Figure 4 , Figure 4 This is a schematic diagram of the structure of a computer device provided in an optional embodiment of the present invention, such as... Figure 4As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 4 Take a processor 10 as an example.

[0157] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GPA), or any combination thereof.

[0158] The memory 20 stores instructions executable by at least one processor 10 to cause at least one processor 10 to perform the method shown in the above embodiments.

[0159] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0160] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0161] The computer device also includes a communication interface 30 for communicating with other devices or communication networks.

[0162] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

[0163] A portion of this invention can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to the invention through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0164] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A multi-agent social network simulation method based on cognitive reasoning chain, characterized in that, include: Initialize a multi-agent system including a social environment engine, a user profile engine, and a cognitive reasoning engine; construct a social environment containing multiple agents; configure user profiles for each agent; and establish a cognitive reasoning chain that includes dimensions of emotion, mood, thought, stance, and intention. The simulation loop of a multi-agent social network is executed in a preset number of iterations. In each iteration, social information is pushed to the agents as external stimuli through the social environment engine to activate the agents to make independent decisions. The agents update the cognitive state of each dimension in the cognitive reasoning chain through large-scale language model reasoning. Based on the updated cognitive state, corresponding social behaviors are generated and the cognitive state trajectory of the social behaviors and the corresponding cognitive reasoning chain is recorded. The process of updating the cognitive state of each dimension in the cognitive reasoning chain through large-scale language model reasoning includes: Generate cognitive guidance prompts, which include the agent's user profile, data memory of a preset number of historical iterations, currently observed social information content, the cognitive state at the previous moment, and the influence coefficients between various cognitive dimensions; The cognitive guidance prompts are input into a preset large-scale language model for cognitive reasoning chain evolution and update, obtaining the cognitive state of each cognitive dimension at the current moment. The evolution and update of each dimension are represented by the following functional relationship: The formula for the evolution of emotions is: ; The formula for the evolution of emotions is: ; The formula for the evolution of thought is: ; The formula for the evolution of a stance is: ; The formula for the evolution of intention is: ; in: Indicates the current time step; They represent the time steps respectively The state of one's emotions, feelings, thoughts, stances, and intentions; Each represents a different cognitive dimension at the previous time step. The state; Representative dimension The state transition function is implemented through reasoning using a pre-defined large-scale language model. Representative influence dimensions Individual stable traits; Represents the time step Dimension The influence of the external environment; Represents the time step Apply to dimension The random noise term is used to simulate random influencing factors that exist in human cognition but are not explicitly captured by the model, thereby increasing the realism of the simulation. Periodically analyze the historical records of cognitive state trajectories and behaviors to optimize the influence coefficients between various cognitive dimensions in the cognitive reasoning chain and adjust the reasoning strategies of the pre-set large-scale language model.

2. The method according to claim 1, characterized in that, The initialization includes a multi-agent system comprising a social environment engine, a user profiling engine, and a cognitive reasoning engine, including: Load the parameters required for the simulation from the preset configuration file, including: preset large language model interface information, total number of simulation rounds, number of agents, and data paths for each functional module, and initialize the social environment engine, user profile engine, and cognitive reasoning engine in sequence; among which: The initialization process of the social environment engine includes: constructing and instantiating platform objects and social environment databases, creating data tables according to a preset database pattern to store user data, social content data and cognitive state data, and simultaneously initializing the social information content pool; The initialization process of the user profile engine includes: creating intelligent agents based on the intelligent agent generator, and assigning a unique user profile to each intelligent agent, wherein the user profile includes demographic information and psychological trait information; The initialization process of the cognitive reasoning engine includes: instantiating a cognitive manager for each agent to initialize the cognitive state of each cognitive dimension in the agent's cognitive reasoning chain, with the initial value being a neutral default value or a personalized setting based on the user profile.

3. The method according to claim 1, characterized in that, The step of pushing social information to the agent as an external stimulus through a social environment engine to activate the agent to execute independent decisions includes: The social environment engine distributes observation information for this round to each agent from the social information content pool based on a preset information recommendation strategy. Activate each agent to update its cognitive state and make behavioral decisions based on the distributed observation information.

4. The method according to claim 1, characterized in that, The step of generating corresponding social behaviors based on the updated cognitive state and recording the cognitive state trajectory of the social behaviors and the corresponding cognitive reasoning chain includes: The decision-making behavior type is generated based on the intent dimension. When content needs to be generated, the large language model is called based on the current complete cognitive state to generate text that matches the state. The decision-making behavior is executed in a simulated environment by a behavior actuator, and the decision-making behavior and cognitive state trajectory are recorded to a database by a recorder.

5. The method according to claim 1 or 4, characterized in that, The periodic analysis of cognitive state trajectories and historical records of behavior is used to optimize the influence coefficients between various cognitive dimensions of the cognitive reasoning chain and adjust the reasoning strategy of the preset large-scale language model, including: Dynamic dependencies between cognitive dimensions are modeled using dynamic Bayesian networks, where the structure of the dynamic Bayesian network is determined based on the first-order Markov property and the assumption of regular stability of state transitions. Collect the cognitive state trajectories and corresponding decision-making behavior history records of all agents within a preset iteration round; The influence coefficients between cognitive dimensions are calculated using ordinary least squares regression, including: establishing a linear regression equation for each cognitive dimension, and using historical data as input to calculate the optimal coefficient value using ordinary least squares. Cognitive guidance prompts for the cognitive reasoning process are updated based on the influence coefficients between cognitive dimensions, thereby optimizing the reasoning strategy of large language models.

6. A multi-agent social network simulation system based on cognitive reasoning chain, characterized in that, include: The system initialization module is used to initialize a multi-agent system including a social environment engine, a user profile engine, and a cognitive reasoning engine, to build a social environment containing multiple agents, to configure user profiles for each agent, and to establish a cognitive reasoning chain that includes dimensions of emotion, mood, thought, stance, and intention. The simulated loop execution module is used to execute a multi-agent social network simulation loop with preset iterations. In each iteration, social information is pushed to the agents as external stimuli through the social environment engine to activate the agents to make independent decisions. The cognitive state of each dimension in the cognitive reasoning chain is updated through large-scale language model reasoning. The corresponding social behavior is generated based on the updated cognitive state, and the social behavior and the cognitive state trajectory of the corresponding cognitive reasoning chain are recorded. The process of updating the cognitive state of each dimension in the cognitive reasoning chain through large-scale language model reasoning includes: Generate cognitive guidance prompts, which include the agent's user profile, data memory of a preset number of historical iterations, currently observed social information content, the cognitive state at the previous moment, and the influence coefficients between various cognitive dimensions; The cognitive guidance prompts are input into a preset large-scale language model for cognitive reasoning chain evolution and update, obtaining the cognitive state of each cognitive dimension at the current moment. The evolution and update of each dimension are represented by the following functional relationship: The formula for the evolution of emotions is: ; The formula for the evolution of emotions is: ; The formula for the evolution of thought is: ; The formula for the evolution of a stance is: ; The formula for the evolution of intention is: ; in: Indicates the current time step; They represent the time steps respectively The state of one's emotions, feelings, thoughts, stances, and intentions; Each represents a different cognitive dimension at the previous time step. The state; Representative dimension The state transition function is implemented through reasoning using a pre-defined large-scale language model. Representative influence dimensions Individual stable traits; Represents the time step Dimension The influence of the external environment; Represents the time step Apply to dimension The random noise term is used to simulate random influencing factors that exist in human cognition but are not explicitly captured by the model, thereby increasing the realism of the simulation. The dynamic optimization module is used to periodically analyze the historical records of cognitive state trajectories and behaviors, optimize the influence coefficients between various cognitive dimensions in the cognitive reasoning chain, and adjust the reasoning strategies of the preset large-scale language model.

7. A computer device, characterized in that, include: A memory and a processor are interconnected, the memory storing computer instructions, and the processor executing the computer instructions to perform the multi-agent social network simulation method based on cognitive reasoning chain as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the multi-agent social network simulation method based on cognitive reasoning chain as described in any one of claims 1 to 5.

9. A computer program product, characterized in that, It includes computer instructions for causing a computer to execute the multi-agent social network simulation method based on cognitive reasoning chains as described in any one of claims 1 to 5.

Images