Reputation-based large language model multi-agent cooperation maintenance excitation method and device

Abstract

The invention discloses a reputation-based large language model multi-agent cooperation maintenance incentive method and device, and aims to effectively maintain cooperation in a large model multi-agent hybrid motivation scene, relieve public tragedies and improve the overall efficiency of a system. Comprising the following steps: determining a first network decision for maintaining a relationship or terminating a relationship between agents based on a large model language model according to updated first companion reputation, a t-hour first-order direct interaction event and a first network decision function; and according to the updated second partner reputation and the second network decision function, determining a second network decision for increasing or reducing business exchange between the agents based on the large model language model.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and more specifically to a method and device for maintaining incentives for multi-agent cooperation in a large language model based on reputation. Background Art

[0002] As large language model-based multi-agent systems (LLM-MAS) are implemented in scenarios such as resource scheduling, automated trading, and complex game-playing, academics have increasingly discovered that when the number of agents increases and they share common resources, the level of cooperation often declines rapidly over time, leading to a classic "tragedy of the commons" scenario: resources are over-extracted, system performance declines, and even the entire task fails. Therefore, maintaining long-term cooperation among generative large-scale model agents and curbing opportunistic behavior has become a core technical bottleneck hindering further commercialization in this field.

[0003] In early traditional multi-agent systems, research generally used "reputation" as an incentive: dynamically scoring agents based on their historical contributions to guide rational partner selection, thereby improving the level of group cooperation. For example, the document "Multi-agent Deep Reinforcement Learning Method Based on Counterfactual Returns" improves the problem of sparse rewards and unfair distribution in fully cooperative tasks by splitting individual rewards into distributable "reputation." However, because it does not involve generative large models or public resource games, it mainly stays in situations with small parameter scales and limited strategy spaces. The document "General Multi-Agent Artificial Intelligence Architecture" allows for hierarchical management of agent behavior in game and simulation environments. The article also includes interface design for trust / reputation modules, but overall it is more oriented towards small-scale entities and fixed tasks, and lacks adaptability to new features of large-model agents such as open text interaction and non-stationary language strategies.

[0004] To prevent centralized servers from failing or acting maliciously, the industry has gradually adopted blockchain for decentralized reputation storage and incentive settlement: the document proposes a "multi-dimensional distributed reputation scoring system," which can record the reputation values of entities in multiple behavioral dimensions in a peer-to-peer network, improving query and tamper-proof capabilities. However, this method does not address the on-chain write overhead caused by high-frequency updates, resulting in high on-chain costs. The document combines smart contracts with Internet of Things (IoT) devices to achieve blockchain-based asset control and reputation verification, emphasizing device-level identity security. However, it also faces problems such as "limited computing resources" and "high latency," making it difficult to directly reuse in large-model multi-agent system scenarios with a huge number of parameters.

[0005] At present, most reputation incentive schemes focus on IoT, Internet of Vehicles, or small-scale simulated agents. They lack special adaptation for LLM-Agents with natural language reasoning / generation capabilities and a large strategy space, resulting in object limitations. Furthermore, on-chain storage ensures immutability, but leads to long reputation update cycles and high costs, which cannot meet the decision-making needs of large-model multi-agents in second-level interactions, resulting in a conflict between real-time performance and cost. Existing methods mostly focus on a single reputation scoring or allocation algorithm, and fail to organically combine reputation evaluation with the dynamic evolution of the agent interaction topology (such as prioritizing connection to high-reputation nodes and eliminating bad nodes). Therefore, they lack the ability to suppress the "tragedy of the commons" at the root level. At the same time, existing technologies are mostly verified by offline simulation or static games, and have not yet been tested on a large scale in new large-model multi-agent platforms with open text, multi-round dialogue, and non-stationary resource games.

[0006] In summary, although the industry has explored various reputation incentive mechanisms in traditional multi-agent systems and blockchain fields, there are still problems such as real-time performance, adaptability, and insufficient network coordination in generative large-scale multi-agent scenarios. Summary of the Invention

[0007] The embodiments of the present invention provide a reputation-based large language model multi-agent cooperation maintenance incentive method and device, which can simultaneously complete the accurate update of reputation, instant incentive redemption and adaptive reconstruction of the interactive network, effectively maintain cooperation, alleviate the tragedy of the commons, and improve the overall efficiency of the system in large model multi-agent mixed motivation scenarios.

[0008] The embodiment of the present invention provides a reputation-based large language model multi-agent cooperation maintenance incentive method, including:

[0009] When two agents have a first-order direct interaction at time t, the initial first peer reputation is updated based on the large model language model according to the first-order direct interaction event at time t, the initial first peer reputation, and the first reputation function, to obtain an updated first peer reputation; the initial self-reputation is updated based on the large model language model according to the first-order direct interaction event at time t, the initial self-reputation, the self's basic information, and the second reputation function, to obtain an updated self-reputation; the first network decision of maintaining or terminating the relationship between the agents is determined based on the large model language model according to the updated first peer reputation, the first-order direct interaction event at time t, and the first network decision function;

[0010] When rumors appear again among intelligent agents, a second-order rumor occurs at time t. According to the second-order rumor event at time t, the initial second peer reputation and the third reputation function, the initial second peer reputation is updated based on the large model language model to obtain an updated second peer reputation; according to the updated second peer reputation and the second network decision function, the second network decision to increase or decrease business transactions between intelligent agents is determined based on the large model language model.

[0011] An embodiment of the present invention provides a reputation-based large language model multi-agent cooperation maintenance incentive device, comprising:

[0012] A first updating unit is configured to, when two agents have a first-order direct interaction at time t, update the initial first peer reputation based on the large model language model according to the first-order direct interaction event at time t, the initial first peer reputation, and the first reputation function, to obtain an updated first peer reputation; and to update the initial self reputation based on the large model language model according to the first-order direct interaction event at time t, the initial self reputation, the self's basic information, and the second reputation function, to obtain an updated self reputation;

[0013] a first decision-making unit, configured to determine, based on the large model language model and according to the updated first peer reputation, the first-order direct interaction event at time t, and the first network decision function, a first network decision to maintain or terminate a relationship between the agents;

[0014] The second updating unit is configured to update the initial second peer reputation based on the large model language model according to the second-order rumor event at time t, the initial second peer reputation, and the third reputation function when a rumor reappears between the agents, thereby obtaining an updated second peer reputation;

[0015] The second decision unit is used to determine a second network decision of increasing or decreasing business transactions between intelligent agents based on the large model language model according to the updated second peer reputation and the second network decision function.

[0016] An embodiment of the present invention provides a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes any one of the above-mentioned multi-agent cooperation maintenance incentive methods.

[0017] An embodiment of the present invention provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes any one of the above-mentioned multi-agent cooperation maintenance incentive methods.

[0018] The embodiment of the present invention provides a method and device for maintaining incentives for multi-agent cooperation based on a large language model based on reputation. The method abstracts all reputation items into five-tuples, uses a large language model to perform deep semantic understanding of natural language evaluation to obtain continuous quantitative scores, and solves the technical problem that existing methods can only use discrete labels or manual thresholds and are difficult to maintain consistency in open text interactions, bringing the dual advantages of cross-scenario compatibility and explainable evaluation; the method uses direct interaction and rumor events to update reputation information, which solves the problem of delayed reputation updates in sparse interaction situations, and can still quickly update reputations under low interaction density, greatly improving the credibility of reputation information; furthermore, the method uses direct interaction and rumor events to update reputation information, which solves the problem of delayed reputation updates in sparse interaction situations, and can still quickly update reputations under low interaction density, greatly improving the credibility of reputation information; furthermore, the method uses Through output decision-making, reputation changes are mapped to edge addition and deletion decisions of the directed graph in real time, forming a closed loop of "interaction → reputation → topology → interaction", which solves the problem of reducing the risk of betrayal such as "free riding" at the level of the relationship, strengthens the relationship between high-reputation groups, and significantly alleviates the tragedy of the commons phenomenon; the method provided by the embodiment of the present invention realizes a highly scalable, low-cost, and strong real-time cooperation maintenance mechanism in a multi-agent system driven by a large language model, and provides a systematic solution to avoid the "tragedy of the commons"; it can simultaneously complete accurate reputation updates, instant incentive redemption, and adaptive reconstruction of the interaction network, effectively maintain cooperation, alleviate the tragedy of the commons, and improve the overall efficiency of the system in large-model multi-agent mixed-motivation scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0020] Figure 1 A flowchart of a reputation-based large language model multi-agent cooperation maintenance incentive method provided by an embodiment of the present invention;

[0021] Figure 2 A schematic diagram of the functional module structure of the reputation-based large language model multi-agent cooperation maintenance incentive method provided in an embodiment of the present invention;

[0022] Figure 3 A flow chart of a method for maintaining incentives for multi-agent cooperation in a large language model based on reputation and function modules provided by an embodiment of the present invention;

[0023] Figure 4 A schematic diagram of the structure of a reputation-based large language model multi-agent cooperation maintenance incentive device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0024] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0025] Figure 1 A flow chart of a reputation-based large language model multi-agent cooperation maintenance incentive method provided by an embodiment of the present invention; Figure 1 As shown, the method includes the following steps:

[0026] Step 101: When two agents have a first-order direct interaction at time t, the initial first peer reputation is updated based on the large model language model according to the first-order direct interaction at time t, the initial first peer reputation, and the first reputation function, to obtain an updated first peer reputation; the initial self-reputation is updated based on the large model language model according to the first-order direct interaction at time t, the initial self-reputation, the self's basic information, and the second reputation function, to obtain an updated self-reputation; and the first network decision of maintaining or terminating the relationship between the agents is determined based on the large model language model according to the updated first peer reputation, the first-order direct interaction at time t, and the first network decision function.

[0027] In step 102, when rumors appear again among the agents, a second-order rumor event at time t occurs. According to the second-order rumor event at time t, the initial second peer reputation and the third reputation function, the initial second peer reputation is updated based on the large model language model to obtain an updated second peer reputation; according to the updated second peer reputation and the second network decision function, a second network decision to increase or decrease business transactions between the agents is determined based on the large model language model.

[0028] The execution subject of the embodiment of the present invention is a processor, which can build an interactive network and generate a local reputation database for each intelligent agent.

[0029] Specifically, before step 101, the processor first constructs the interactive network G (0) = <V,E (0) >, where V represents the set of agents, E (0) Represents edges, i.e., potential interactions between agents. Furthermore, a local reputation database for each agent is generated. In this embodiment of the present invention, reputation information includes self-reputation and peer reputation. Self-reputation represents an agent's cognitive evaluation of itself based on its interaction experience; peer reputation represents an agent's evaluation of other agents' behaviors during interaction. In this embodiment of the present invention, peer reputation includes first peer reputation and second peer reputation.

[0030] Specifically, reputation is represented by a five-tuple consisting of basic identity information, scenario, role, natural language description, and quantitative score, and an index is created for all entries.<a,s,o,c,μ> In the model, a represents the basic identity information of the evaluated agent (such as ID and name); since reputation is scenario-dependent, s and o represent the scenario and role corresponding to the reputation, respectively; c represents the reputation content described in natural language; μ is a quantitative score of c, ranging from [-1, 1], where a higher score indicates a better reputation.

[0031] This method uses a unified five-tuple reputation representation to make numerical scores correspond one-to-one with natural language reasons, which facilitates traceability, evaluation and inspection, and improves the transparency and trust level of the system.

[0032] When two agents in the interaction network have a first-order direct interaction at time t-1, the first-order direct interaction at time t-1 can be encapsulated as a first-order direct interaction event at time t-1 according to the type. Then, the first-order direct interaction event at time t-1 generates the agent's initial first peer reputation for the interaction object agent according to the prompt word, and also generates its own initial self-reputation. Then, the first-order direct interaction event at time t-1 is written into the local memory of the two agents in a JSON structure for caching. The prompt word here can be P direct .

[0033] In practical applications, when agent i and related agent j finish interacting at time t-1, the agents in the interactive network may directly spread rumors, that is, a second-order rumor event at time t-1 occurs at time t-1. The second-order rumor event at time t-1 is encapsulated as a second-order rumor event at time t-1 according to the type. The second-order rumor event at time t-1 generates the initial second peer reputation of the agent to the rumor target agent according to the prompt word, and then writes the second-order rumor event at time t-1 into the local memory of the agent in a JSON structure for caching. The prompt word here can be P direct .

[0034] For example, in a large-scale multi-agent system simulating a business cooperation scenario, there are three agents: Agent A, Agent B, and Agent C. They represent different business companies and carry out interactive activities such as cooperative projects and information exchange in this system.

[0035] The processor builds the initial interaction network G (0) = <V,E (0) >. Here, vertices V = (A, B, C), edges E (0) Represents potential interaction relationships, that is, it is initially believed that there may be cooperation opportunities between any two agents. At the same time, a local reputation database is generated for each agent. in, Represents the local reputation database of Agent A, Represents the local reputation database of Agent B, Represents the local reputation database of Agent C.

[0036] Specifically, assume that at a certain moment, Agent A and Agent B had a direct business cooperation interaction. The cooperation content was that Agent B provided a batch of raw materials for Agent A. After this interaction, the first-order direct interaction at t-1 was encapsulated as the first-order direct interaction event at t-1 Written in JSON structure into the local memory banks of Agent A and Agent B [[ID=I0]] and cache. Further, the first-order direct interaction event at t-1 Can also be based on the prompt word Ρ direct , the first peer reputation of Agent A towards Agent B can be obtained respectively At the same time, generate the self-reputation of Agent A Among them, Represents the first-order direct interaction event between Agent A and Agent B at t-1, Represents the local memory bank of Agent A, Represents the local memory bank of Agent B, Represents the initial first peer reputation of Agent A towards Agent B at t-1, [[ID=3I]] Represents the initial self-reputation of Agent A at t-1.

[0037] Taking Agent A as an example, its initial reputation database includes the following entries stored: the initial first peer reputation and the initial self-reputation. Since Agent A has interacted with Agent B, the initial first peer reputation here is about Agent B. The initial first peer reputation is: <B, project cooperation, supplier delivers high-quality products on time, 0.8>, which means that Agent A believes that Agent B has a good reputation of delivering high-quality products on time when acting as a supplier in the project cooperation scenario, and the quantitative score is .8; the initial self-reputation of Agent A is: <A, market promotion, leader, has innovative marketing ideas, 0.9>, which is Agent A's self-reputation evaluation in the market promotion scenario; among them, B in the initial first peer reputation represents the name of the evaluated agent, project cooperation and supplier are the corresponding scenarios and roles under this reputation respectively, delivering high-quality products on time is the reputation content described in natural language, and 0.8 is the quantitative score of delivering high-quality products on time; A in the initial self-reputation represents the name of the evaluated agent, market promotion and leader are the corresponding scenarios and roles under this reputation respectively, having innovative marketing ideas is the reputation content described in natural language, and 0.9 is the quantitative score of delivering high-quality products on time.

[0038] In practical applications, the reputation of the initial first peer and the initial self-reputation are both represented by five-tuples. Each initial reputation database establishes an index for these entries (reputations) for fast query.

[0039] When agent C hears a rumor about agent A from other channels, the content is that agent A has a serious quality problem in a recent project. The second-order rumor event at time t-1 is encapsulated as the second-order rumor event at time t-1. Then write to the local memory of agent C Cache, further, the second-order rumor event at time t-1 According to the prompt word P gossip Generate the initial second peer reputation of agent C towards agent A in, represents the second-order rumor event about agent A at time t-1, represents the local memory of agent C, represents the initial second peer reputation of agent C towards agent A at time t-1.

[0040] It should be noted that in actual applications, when the interaction network includes three intelligent agents, namely the first intelligent agent, the second intelligent agent and the third intelligent agent, when the first intelligent agent and the second intelligent agent have a first-order direct interaction at time t-1, it is possible that a second-order rumor event at time t-1 occurs between the first intelligent agent and the third intelligent agent at time t-1, or a second-order rumor event at time t-1 occurs between the second intelligent agent and the third intelligent agent at time t-1.

[0041] Similarly, when the first agent and the third agent have a first-order direct interaction at time t-1, it is possible that a second-order rumor event at time t-1 occurs between the first agent and the second agent.

[0042] Or a second-order rumor event occurs between the second agent and the third agent at time t-1.

[0043] Similarly, when the second agent and the third agent have a first-order direct interaction at time t-1, it is possible that a second-order rumor event at time t-1 occurs between the first agent and the second agent.

[0044] Or a second-order rumor event occurs between the first agent and the third agent at time t-1.

[0045] In step 101, when two intelligent agents have a first-order direct interaction at time t, the first-order direct interaction at time t is encapsulated as a first-order direct interaction event at time t according to its type, and then the first-order direct interaction event at time t, the initial first peer reputation and the first reputation function are updated based on the large model language model to obtain an updated first peer reputation; accordingly, according to the first-order direct interaction event at time t, the initial self-reputation, its own basic information and the second reputation function, the initial self-reputation is updated based on the large model language model to obtain an updated self-reputation.

[0046] Furthermore, the first network decision of maintaining or terminating the network relationship between the agents is determined based on the large model language model according to the updated first peer reputation, the first-order direct interaction event at time t, the basic description of the agent itself and the first network decision function.

[0047] It should be noted that the first reputation function here is ShapeRepuPeer, the second reputation function is ShapeRepuSelf, and the first network decision function is InteractEdgeShape. The agent's basic description includes basic information similar to that of a human, such as name, age, occupation, and values. The agent's basic description needs to be determined based on the specific scenario and is not specifically limited here.

[0048] Furthermore, the processor may store the first-order direct interaction event at time t of the agent, the updated self-reputation, the updated first peer reputation and the first network decision into the local reputation databases of the two agents.

[0049] For example, suppose that at a certain time t, Agent A and Agent B once again have a direct business cooperation interaction. After this interaction ends, the first-order direct interaction content at time t is encapsulated as the first-order direct interaction event at time t If agent A is satisfied with agent B's service in this cooperation, the processor generates agent A's first peer reputation for agent B at time t based on the first-order direct interaction event at time t, the initial first peer reputation, and the first reputation function ShapeRepuPeer and the large model language model. This can also be called the updated first peer reputation, which is the updated first peer reputation obtained after updating the initial first peer reputation. Assume that the updated first peer reputation is (Compared to the previous initial first companion reputation improved).

[0050] At the same time, the initial self-reputation can be updated based on the first-order direct interaction event at time t, the initial self-reputation, the basic information of the self, and the second reputation function ShapeRepuSelf based on the large model language model to obtain the updated self-reputation. Assume that the updated self-reputation of the generated agent A (Because of the successful completion of this cooperation, my recognition of my own abilities has also increased).

[0051] Furthermore, the processor updates the first peer reputation of agent B based on agent A First-order direct interaction event at time t And the basic description of the agent itself, call the large language model to execute the first network decision function DirectEdgeShape() to output the first network decision of whether agent A and agent B maintain the network relationship or terminate the network relationship. In this embodiment, due to High, and the interaction process is smooth, the large language model outputs the first network decision Indicates that Agent A will maintain its network relationship with Agent B, that is, it will continue to consider cooperating with Agent B in the future. It represents the network relationship between agent A and agent B output at time t, and "Y" represents the establishment or maintenance of the network relationship.

[0052] The processor will be the first-order direct interaction event at time t Update the first companion's reputation to Update your reputation and First Network Decision Stored in the local reputation database of Agent A and Agent B.

[0053] In step 102, in actual applications, when the agent receives any rumor, that is, receives a second-order rumor event occurring at time t, the processor updates the initial second peer reputation based on the second-order rumor event at time t, the initial second peer reputation, and the third reputation function based on the large model language model to obtain an updated second peer reputation;

[0054] Furthermore, the processor updates the second peer reputation, the second-order rumor event at time t, the basic description of the agent itself and the second network decision function and outputs a second network decision to increase or decrease business transactions between agents based on the large model language model.

[0055] It should be noted that the third reputation function here is ShapeRepuCossip, and the second network decision function is GossipEdgeShape.

[0056] Furthermore, the processor may store the second-order rumor event about the agent at time t, the updated second peer reputation, and the second network decision into the local reputation database of the two agents.

[0057] For example, when agent C receives a second-order rumor event about agent A at time t After that, the processor generates the second-order rumor event at time t Initial second companion reputation And the third reputation function ShapeRepuGossip and based on the voice model, the second peer reputation is updated In this process, it is assumed that the large language model combines the rumor content and agent C’s previous reputation cognition of agent A to derive agent C’s updated second peer reputation for agent A. (Reputation score dropped due to rumors of quality issues).

[0058] Furthermore, the processor updates the second peer reputation of agent A based on agent C Call the second network decision function of the large language model GossipEdgeShape() to output the second network decision of increasing or decreasing business transactions between agent C and agent A. In this embodiment, because the second peer reputation is updated has decreased, so the large language model outputs the second network decision Indicates that Agent C decides to terminate its network relationship with Agent A, that is, it is unlikely to cooperate with Agent A in the future. It represents the network relationship between agent C and agent A output at time t, and "N" indicates the termination of the network relationship.

[0059] The processor will send the second-order rumor event at time t Update the first companion's reputation to Updated second companion reputation and second network decision Stored in the local reputation database of Agent A and Agent C.

[0060] The embodiment of the present invention provides a multi-agent cooperation maintenance incentive method, the execution subject of which is a processor, which includes multiple Figure 2 The functional modules shown below are combined Figure 2 The functional modules shown are used as the execution body to introduce the multi-agent cooperation maintenance incentive method provided by the embodiment of the present invention.

[0061] like Figure 2 As shown, the reputation network initialization module 201 is responsible for building the initial interaction network G (0) = <V,E (0) >, where V represents the set of agents, E (0)Represents potential interaction relationships; at the same time, reads the system configuration file and generates a local reputation database for each agent

[0062] The interaction event collection module 201 intercepts two types of interactions through the environment engine: (i) first-order direct interaction events (ii) Second-order rumor events in, Indicates that at time t, agents i and j have a first-order direct interaction, Indicates a second-order rumor event about agent y at time t. First-order direct interaction events and second-order rumor events are written to the agent's local memory in a JSON structure. cached and pushed to the reputation update module 203, where represents the first-order direct interaction event between agent i and agent j at time t, represents the second-order rumor event about agent y at time t-1, Represents the local memory bank of agent i.

[0063] Reputation update module 203, which includes direct interaction reputation submodule 2031 and rumor inference reputation submodule 2032;

[0064] Direct Interaction Reputation Submodule 2031: In Capture Then, call the large language model inference module 206 and press the prompt P direct and the initial first companion reputation is updated with the first companion reputation Similarly, update your reputation in, represents the updated first peer reputation of agent i to agent j at time t, represents the initial first-peer reputation of agent i toward agent j at time t-1.

[0065] Rumor inference reputation submodule 2032: Receive rumors in agent l When, according to the initial second companion reputation, prompt word P gossip Calculate and update the reputation of the second peer in, represents the updated second peer reputation of agent l to agent y at time t, represents the initial second peer reputation of agent l towards agent y at time t-1.

[0066] A network adaptive reconstruction module 204, which includes a direct interaction reconstruction submodule 2041 and a rumor reasoning reconstruction submodule 2042;

[0067] Direct interaction reconstruction submodule 2041: Set Represents agent i’s decision on the network connection of j at time t: Determines whether to keep or disconnect edge (i, j). i Represents the basic description of agent i itself.

[0068] Rumor inference reconstruction submodule 2042: After receiving the rumor Post-call Dynamically adjust the network structure. Among them, D l Represents the basic description of agent l itself.

[0069] The data storage module 205 is responsible for writing the updated reputation information, interaction information and interaction logs into the local reputation database of each agent. and agent local memory for the next iteration to read.

[0070] The large language model inference module 206 provides a unified interface, Infer(prompt, context). The prompt is populated by modules 201-204 according to the prompt template, and the context is historical information, such as reputation or interaction information. The model returns structured JSON, which is parsed by the data storage module and written to the corresponding database.

[0071] The following is a description of the multi-agent cooperation maintenance incentive method provided by the embodiment of the present invention in combination with the functional modules shown in 2. Figure 3 As shown, the method includes the following process:

[0072] In a large-scale multi-agent system, the interactions between agents can be modeled using a network that determines the likelihood of encounters and interactions between agents. In this embodiment, reputation information can be divided into two types: (i) self-reputation, which is an agent's self-evaluation based on its own interaction experiences; and (ii) first-peer reputation, which is an agent's evaluation of other agents' behavior during interactions.

[0073] Step 301: Call the reputation network initialization module 201 to implement the initialization operation of the system.

[0074] Step 302: Interaction event perception. The interaction event collection module 202 encapsulates the interaction content into direct interaction events according to the type at the end of each interaction between agent i and related agent j. Or second-order rumors And push it to the reputation update module 103.

[0075] Step 303, reputation update driven by interaction events. Based on the interaction event perceived in step 302, if it is a direct interaction, execute step 3031; if it is a rumor, execute step 3032.

[0076] Step 3031: First-order direct interaction driven reputation update. The direct interaction reputation submodule 2031 calls the large language model inference module 206 to execute prompt P direct , respectively generate self-reputation information Reputation information with the first companion The generated results are written into the ontology database via the data storage module 105

[0077] Step 3032: Reputation update driven by second-order rumor events. The rumor reasoning reputation submodule 2032 calls the large language model reasoning module 206 and executes prompt P. gossip , calculate and update the reputation of agent y

[0078] Step 304, network adaptive reconstruction driven by reputation update. According to the reputation update type in step 303, if the reputation update is driven by first-order direct interaction, step 3041 is executed; if the reputation update is driven by second-order rumor events, step 3042 is executed.

[0079] Step 3041: First-order direct interaction driven network adaptive reconstruction, direct interaction reconstruction submodule 2041 based on the latest first peer reputation and interactive information Call the large language model inference module 206, execute the prompt irectEdgeShape(), the first network decision in, It represents the network relationship between agent i and agent j output at time t, "Y" indicates the establishment or maintenance of the network relationship, and "N" indicates the termination of the network relationship.

[0080] Step 3042: Second-order rumor event-driven network adaptive reconstruction, rumor reasoning reconstruction submodule 2042 based on the latest first peer reputation Call the large language model inference module 206, execute the prompt GossipEdgeShape(), and make the second decision in, Represents the network relationship between agent l and agent y output at time t.

[0081] Step 305: Data write-back and next iteration, all updated interaction information, reputation information, network structure are written back and stored in each agent's local reputation database R by calling data storage module 205. i and the agent's local memory Mi , providing historical information for the next round of steps 301 - 305.

[0082] Suppose there are three agents in a large - model multi - agent system in the e - commerce field: supplier agent S, retailer agent R, and logistics agent L.

[0083] Based on the above scenario, the following provides a multi - agent cooperation maintenance incentive method. Specifically, an embodiment taking retailer agent R as an example includes the following steps:

[0084] Step 401, the processor constructs an initial interaction network G (0) =(V, E (0) >) where V=(S, R, L), and the edge E (0) represents potential interaction relationships, that is, initially, it is considered that there may be cooperation opportunities between any two agents.

[0085] Generate a local reputation database for each agent Taking retailer agent R as an example, there are such entries in its initial reputation database: <S, product supply, supplier,

[0086] The product quality is good and the supply is relatively stable, 0.7>, <R, market sales, retailer,

[0087] Having a certain market promotion ability and customer base, 0.8>; where, the first entry represents the performance of supplier S as a supplier in the product supply scenario as perceived by retailer agent R, with a quantitative score of 0.7; the second entry represents retailer R's self - reputation evaluation of itself in market sales. Here, in the five - tuple, a is the supplier S and retailer R respectively, s is "product supply" and "market sales" respectively, o is "supplier" and "retailer" respectively, c is the corresponding natural - language description, and μ is the quantitative score.

[0088] Step 402, at a certain moment, retailer R has a direct transaction (first - order direct interaction) with supplier S, and supplier S supplies a batch of electronic products to retailer R. After the transaction, the first - order direct interaction is encapsulated as a first - order direct interaction event Record the transaction details in JSON structure, such as the number of products is 100 pieces, the price is 500 yuan per piece, the delivery time is 2 days earlier than the agreement, and the product quality feedback is excellent, and then write it into the local memory banks and cache of retailer R and supplier S.

[0089] Later, logistics company L heard rumors about retailer R from other logistics companies, saying that retailer R frequently changed the delivery address during a recent delivery, resulting in a significant increase in logistics costs and a decrease in delivery efficiency. This second-order rumor is encapsulated as a second-order rumor event. Write to the local memory of logistics provider L cache.

[0090] Step 403: The processor captures the first-order direct interaction event between the retailer R and the supplier S. Then, call the large language model inference module. Assume that the prompt P direct The large language model is required to focus on factors such as product quality and delivery time to adjust the reputation score. It is known that the first peer reputation of retailer R to supplier S According to the interaction details such as the quality of the goods in this transaction is excellent (a bonus item), the delivery is 2 days ahead of schedule (a bonus item), etc., through the formula The large language model outputs the updated first-peer reputation of retailer R for supplier S

[0091] At the same time, because retailer R successfully completed this transaction and enhanced its own market competitiveness, the updated self-reputation of retailer R is generated through a similar formula Assume before The large language model generates the updated self-reputation of retailer R The generated results are written into the local reputation database of retailer R through the data storage module

[0092] Accordingly, logistics provider L receives the rumor Then, according to the prompt word P gossip Calculate the second peer reputation of retailer R. It is known that the initial second peer reputation of logistics provider L to retailer R The large language model integrates the rumor content (negative information such as frequent changes in delivery addresses) and the logistics company L’s previous reputation perception of R, and uses the formula Get the updated second peer reputation of logistics provider L to retailer R

[0093] Step 404: The processor updates the first peer reputation of the supplier S according to the retailer R. and first-order direct interaction events Call the large language model inference module and execute the prompt DirectEdgeShape(). Assume that the prompt DirectEdgeShape() requires the large language model to maintain the network relationship when the reputation score is higher than 0.8 and the interaction is smooth. The basic description D of the retailer R is known. R , according to the formula The large language model outputs the first network decision It indicates that retailer R decides to maintain and strengthen its cooperation relationship with supplier S, and will consider increasing the order quantity or expanding the cooperation categories in the future.

[0094] Based on the updated second peer reputation of retailer R by logistics provider L The large language model inference module is called to execute the prompt GossipEdgeShape(). Assume that the GossipEdgeShape() prompt requires the large language model to terminate the network relationship when the reputation score is lower than 0.6. Given the basic description D of logistics provider L itself L , according to the formula The large language model outputs the second network decision It means that logistics provider L decides to reduce its business dealings with retailer R and will prefer to cooperate with other retailers in future distribution task assignments.

[0095] Step 405, the processor updates the interaction information Reputation information And the network structure decision Are written back and stored in the local reputation database R of each agent i And the local memory bank M of the agent i , providing historical information for the next iteration starting from the initialization of the reputation network.

[0096] The following is an embodiment taking logistics provider agent L as an example:

[0097] Step 501, the processor constructs the initial interaction network G (0) =<V,E (0) >>, where V=(S,R,L), and the edge E (0) Represents potential interaction relationships, that is, there may be cooperation opportunities between any two agents initially.

[0098] Generate a local reputation database for logistics provider agent L There are such entries in its initial reputation database: <R, Logistics distribution, Retailer, Delivery address accurate and high cooperation degree, 0.75>, <S, Goods packaging, Supplier, Packaging firm and labels clear, 0.8>; among them, the first entry indicates that logistics provider L believes that retailer R performs well as a retailer in the logistics distribution scenario, and the quantitative score is 0.75; the second entry represents the evaluation of L on S as a supplier in the goods packaging scenario. In this five-tuple, taking <R, Logistics distribution, Retailer, Delivery address accurate and high cooperation degree, 0.75> as an example, a is R, s is "Logistics distribution", o is "Retailer", c is "Delivery address accurate and high cooperation degree", and μ is 0.75.

[0099] Step 502: At a certain moment, retailer R and logistics provider L conduct a direct logistics delivery interaction (first-order direct interaction), and logistics provider L delivers a batch of electronic products to retailer R. After the interaction is completed, the first-order direct interaction is encapsulated as a first-order direct interaction event. Record the following in a JSON structure: 200 items, 500 km delivery distance, 5 hours less than expected delivery time, 100% integrity rate, and write it to the local memory of retailer R and logistics provider L. and cache.

[0100] Later, logistics company L heard rumors about supplier S from its peers, saying that some of the packaging of the goods recently supplied by supplier S was damaged, which might affect the safety of logistics and transportation. This second-order rumor was packaged as a second-order rumor event. Write to the local memory of logistics provider L cache.

[0101] Step 503: The processor captures the first-order direct interaction event between the retailer R and the logistics provider L. Then, call the large language model inference module. direct The large language model is required to adjust the reputation based on factors such as delivery time and cargo integrity rate. It is known that the reputation of the first peer of the logistics company L to the retailer R is According to the interactive details such as the shortened actual delivery time (bonus item) and the 100% intact rate of goods (bonus item), the formula The large language model outputs the updated first-peer reputation of logistics provider L for retailer R

[0102] At the same time, logistics company L has improved its recognition of its own capabilities due to the successful completion of this delivery task. Assuming that the initial self-reputation of logistics company L is After calculation by the large language model, the updated self-reputation of logistics provider L is generated through a similar formula The generated results are written into the local reputation database of logistics provider L

[0103] Accordingly, when the logistics provider L receives the rumor Then, according to the prompt word ρ gossip Calculate the second peer reputation of supplier S. It is known that the initial second peer reputation of logistics company L for supplier S The large language model combines the rumor content (the packaging has a risk of damage) and the logistics company L's previous reputation perception of the supplier S, and uses the formula Obtain the updated second peer reputation of logistics provider L for supplier S

[0104] Step 504: The processor updates the first peer reputation of the retailer R based on the logistics provider L. and first-order direct interaction events Call the large language model inference module and execute the prompt DirectEdgeShape(). Assume that DirectEdgeShape() prompts that the reputation is higher than 0.8 and the interactive experience is good, then maintain cooperation. Known logistics company L's basic description D L , according to the formula Large language model output first network decision This indicates that logistics company L has decided to maintain and strengthen its cooperative relationship with retailer R, and will give retailer R a certain preference in allocating distribution resources in the future.

[0105] The processor updates the second partner reputation of the supplier S based on the logistics provider L Call the large language model inference module and execute the prompt GossipEdgeShape(). Assume that GossipEdgeShape() prompts to reduce cooperation when the reputation is lower than 0.7. Known logistics company L's basic description D L , according to the formula Large language model outputs second network decision This means that logistics company L has decided to reduce its cooperation with supplier S and will give priority to other suppliers with more reliable packaging quality in subsequent logistics and transportation arrangements.

[0106] Step 505: The processor updates the interaction information Reputation Information and network structure decisions Write back and store it in the local reputation database R of the logistics agent L (and other related agents) L and the agent's local memory M L , providing historical information for the next round of iteration starting from the initialization of the reputation network.

[0107] Embodiments of the present invention provide a reputation-based large language model multi-agent cooperation maintenance incentive method and device. This method abstracts all reputation items into five-tuples and uses a large language model to perform deep semantic understanding of natural language evaluations to obtain continuous quantitative scores. This solves the technical problem that existing methods can only use discrete labels or manual thresholds and are difficult to maintain consistency in open text interactions, bringing the dual advantages of cross-scenario compatibility and explainable evaluation. The method uses direct interaction and rumor events to update reputation information, solving the problem of delayed reputation updates in sparse interaction situations. Reputation can still be quickly updated under low interaction density, significantly improving the credibility of reputation information. Furthermore, the method maps reputation changes in real time to edge addition and deletion decisions of a directed graph by outputting decisions, forming a closed loop of "interaction → reputation → topology → interaction", solving the problem of reducing the risk of betrayal such as "free riding" at the level, strengthening the relationship between high-reputation groups, and significantly alleviating the tragedy of the commons phenomenon. The method provided by the embodiment of the present invention realizes a highly scalable, low-cost, and strong real-time cooperation maintenance mechanism in a multi-agent system driven by a large language model, providing a systematic solution to avoid the "tragedy of the commons".

[0108] Based on the same inventive concept, an embodiment of the present invention provides a reputation-based large language model multi-agent cooperation maintenance incentive device. Since the principle of this device in solving technical problems is similar to the reputation-based large language model multi-agent cooperation maintenance incentive method, the implementation of this device can refer to the implementation of the method, and the repeated parts will not be repeated.

[0109] like Figure 4 As shown, the apparatus includes a first updating unit 601 , a first decision unit 602 , a second updating unit 603 and a second decision unit 604 .

[0110] The first updating unit 601 is configured to, when two agents have a first-order direct interaction at time t, update the initial first peer reputation based on the large model language model according to the first-order direct interaction event at time t, the initial first peer reputation, and the first reputation function, to obtain an updated first peer reputation; and update the initial self reputation based on the large model language model according to the first-order direct interaction event at time t, the initial self reputation, the self's basic information, and the second reputation function, to obtain an updated self reputation;

[0111] A first decision unit 602 is configured to determine a first network decision of maintaining or terminating a relationship between the agents based on a large model language model according to the updated first peer reputation, the first-order direct interaction event at time t, and a first network decision function;

[0112] The second updating unit 603 is configured to update the initial second peer reputation based on the large model language model according to the second-order rumor event at time t, the initial second peer reputation, and the third reputation function when a rumor re-emerges between agents and a second-order rumor event at time t occurs at time t, thereby obtaining an updated second peer reputation.

[0113] The second decision unit 604 is configured to determine a second network decision of increasing or decreasing business transactions between agents based on the large model language model according to the updated second peer reputation and the second network decision function.

[0114] Preferably, the first updating unit 601 is further configured to construct an initial interaction network, determine multiple agents included in the interaction network and potential cooperative relationships between the agents, and determine a local reputation database for each agent; wherein the reputation stored in the reputation database is represented by a five-tuple consisting of basic identity information, scenario, role, natural language description, and quantitative score;

[0115] When two agents have a first-order direct interaction at time t-1, the first-order direct interaction at time t-1 is encapsulated as a first-order direct interaction event at time t-1, and the first-order direct interaction event at time t-1 generates an initial first peer reputation and an initial self-reputation according to the first prompt word;

[0116] The second updating unit 603 is used to generate a second-order rumor at time t-1 when a rumor first appears between intelligent agents, encapsulate the second-order rumor at time t-1 into a second-order rumor event at time t-1, and generate an initial second peer reputation according to the second prompt word for the second-order rumor event at time t-1.

[0117] Preferably, the first reputation function is ShapeRepuPeer, the second reputation function is ShapeRepuSelf, the third reputation function is ShapeRepuGossip, the first network decision function is InteractEdgeShape, and the second network decision function is GossipEdgeShape.

[0118] It should be understood that the units included in the above-mentioned reputation-based large language model multi-agent cooperation maintenance incentive device are merely logical divisions based on the functions implemented by the device. In actual applications, the above-mentioned units can be superimposed or split. In addition, the functions implemented by the reputation-based large language model multi-agent cooperation maintenance incentive device provided in this embodiment correspond one-to-one to the reputation-based large language model multi-agent cooperation maintenance incentive method provided in the above-mentioned embodiment. The more detailed processing flow implemented by the device has been described in detail in the above-mentioned method embodiment 1 and will not be described in detail here.

[0119] Another embodiment of the present invention also provides a computer device, which includes: a processor and a scenario database; the scenario database is used to store computer program code, and the computer program code includes computer instructions; when the processor executes the computer instructions, the electronic device executes each step of the reputation-based large language model multi-agent cooperation maintenance incentive method in the method flow shown in the above method embodiment.

[0120] Another embodiment of the present invention also provides a computer-readable storage medium, which stores computer instructions. When the computer instructions are executed on a computer device, the computer device executes each step of the reputation-based large language model multi-agent cooperation maintenance incentive method in the method flow shown in the above method embodiment.

[0121] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A reputation-based large language model multi-agent cooperation maintenance incentive method, characterized by: include: When two agents have a first-order direct interaction at time t, the initial first peer reputation is updated based on the large model language model according to the first-order direct interaction event at time t, the initial first peer reputation, and the first reputation function, to obtain an updated first peer reputation; and the initial self reputation is updated based on the large model language model according to the first-order direct interaction event at time t, the initial self reputation, the self's basic information, and the second reputation function, to obtain an updated self reputation; Determining a first network decision of maintaining or terminating a relationship between the agents based on a large model language model according to the updated first peer reputation, the first-order direct interaction event at time t, and the first network decision function; When rumors appear again among intelligent agents, a second-order rumor occurs at time t. According to the second-order rumor event at time t, the initial second peer reputation and the third reputation function, the initial second peer reputation is updated based on the large model language model to obtain an updated second peer reputation; according to the updated second peer reputation and the second network decision function, the second network decision to increase or decrease business transactions between intelligent agents is determined based on the large model language model.

2. The method according to claim 1, wherein Before the two agents have a first-order direct interaction at time t, the process further includes: Construct an initial interaction network, identify multiple agents included in the interaction network and potential cooperative relationships between agents, and determine a local reputation database for each agent. The reputation stored in the reputation database is represented by a five-tuple consisting of basic identity information, scenario, role, natural language description, and quantitative score. When two agents have a first-order direct interaction at time t-1, the first-order direct interaction at time t-1 is encapsulated as a first-order direct interaction event at time t-1, and the first-order direct interaction event at time t-1 generates an initial first peer reputation and an initial self-reputation according to the first prompt word; Before the rumor appears again between the agents, it also includes: When a rumor first appears between agents, a second-order rumor at time t-1 occurs at time t-1, and the second-order rumor at time t-1 is encapsulated as a second-order rumor event at time t-1. The second-order rumor event at time t-1 generates an initial second peer reputation according to the second prompt word.

3. The method according to claim 2, wherein The initial interaction network includes at least three intelligent agents; When the first agent and the second agent have a first-order direct interaction at time t-1; the first agent and the third agent have a second-order rumor event at time t-1; or the second agent and the third agent have a second-order rumor event at time t-1; When the first agent and the third agent have a first-order direct interaction at time t-1; a second-order rumor event at time t-1 occurs between the first agent and the second agent at time t-1; or a second-order rumor event at time t-1 occurs between the second agent and the third agent at time t-1; When the second agent and the third agent have a first-order direct interaction at time t-1; a second-order rumor event at time t-1 occurs between the first agent and the second agent at time t-1; or a second-order rumor event at time t-1 occurs between the first agent and the third agent at time t-1.

4. The method according to claim 1, wherein The first reputation function is ShapeRepuPeer, the second reputation function is ShapeRepuSelf, the third reputation function is ShapeRepuGossip, the first network decision function is InteractEdgeShape, and the second network decision function is GossipEdgeShape.

5. The method according to claim 1, wherein Also includes: The first-order direct interaction event at time t, the second-order rumor event at time t, the updated second peer reputation, the updated self-reputation, the updated first peer reputation, the first network decision and the second network decision are stored in the local reputation database of each agent.

6. A reputation-based large language model multi-agent cooperative maintenance incentive device, characterized by: include: A first updating unit is configured to, when two agents have a first-order direct interaction at time t, update the initial first peer reputation based on the large model language model according to the first-order direct interaction event at time t, the initial first peer reputation, and the first reputation function, to obtain an updated first peer reputation; and to update the initial self reputation based on the large model language model according to the first-order direct interaction event at time t, the initial self reputation, the self's basic information, and the second reputation function, to obtain an updated self reputation; a first decision-making unit, configured to determine, based on the large model language model and according to the updated first peer reputation, the first-order direct interaction event at time t, and the first network decision function, a first network decision to maintain or terminate a relationship between the agents; The second updating unit is configured to update the initial second peer reputation based on the large model language model according to the second-order rumor event at time t, the initial second peer reputation, and the third reputation function when a rumor reappears between the agents, thereby obtaining an updated second peer reputation; The second decision unit is used to determine a second network decision of increasing or decreasing business transactions between intelligent agents based on the large model language model according to the updated second peer reputation and the second network decision function.

7. The device according to claim 6, characterized in that The first updating unit is further configured to: Construct an initial interaction network, identify multiple agents included in the interaction network and potential cooperative relationships between agents, and determine a local reputation database for each agent. The reputation stored in the reputation database is represented by a five-tuple consisting of basic identity information, scenario, role, natural language description, and quantitative score. When two agents have a first-order direct interaction at time t-1, the first-order direct interaction at time t-1 is encapsulated as a first-order direct interaction event at time t-1, and the first-order direct interaction event at time t-1 generates an initial first peer reputation and an initial self-reputation according to the first prompt word; The second updating unit is configured to: When a rumor first appears between agents, a second-order rumor at time t-1 occurs at time t-1, and the second-order rumor at time t-1 is encapsulated as a second-order rumor event at time t-1. The second-order rumor event at time t-1 generates an initial second peer reputation according to the second prompt word.

8. The device according to claim 6, wherein The first reputation function is ShapeRepuPeer, the second reputation function is ShapeRepuSelf, the third reputation function is ShapeRepuGossip, the first network decision function is InteractEdgeShape, and the second network decision function is GossipEdgeShape.

9. A computer device, characterized in that: The computer device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the reputation-based large language model multi-agent cooperation maintenance incentive method as described in any one of claims 1 to 5.

10. A computer-readable storage medium, characterized in that A computer program is stored, and when the computer program is executed by a processor, the processor executes the reputation-based large language model multi-agent cooperation maintenance incentive method as described in any one of claims 1 to 5.

Images