Layered feedback collaboration method, device and equipment of multi-agent system and medium

Abstract

The invention relates to the field of artificial intelligence, can be applied to business system platforms of financial science and technology, medical health and the like, and discloses a hierarchical feedback collaboration method, device and equipment of a multi-agent system and a medium, and the method comprises the steps: carrying out the role division of a plurality of agents according to a preset hierarchical feedback architecture; outputting an initial generation content by the generation agent; the evaluation supervision agent performs multi-dimensional evaluation on the initially generated content and feeds back and outputs structured feedback information; the review agent optimizes and modifies the initial generation content according to the feedback information, and outputs the optimized generation content; and the main supervision agent performs quality evaluation on the optimization generation content, determines whether a preset output condition is met or not, performs iterative feedback and optimization if the preset output condition is not met, and outputs the latest optimization generation content to the outside if the preset output condition is met. Through an evaluation supervision and main supervision double-layer scheduling structure, multi-dimensional evaluation feedback results are integrated and output to guide content optimization, the feedback efficiency and reliability are improved, and the system stability is ensured.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a hierarchical feedback collaboration method, device, equipment and medium for a multi-agent system. Background Art

[0002] Currently, multi-agent systems (MAS) driven by large language models (LLMs) are gradually demonstrating powerful capabilities in applications such as complex task processing, content generation, and information evaluation.

[0003] For example, in financial transactions and business processes in the financial field, multi-agent systems can be used to detect abnormal transaction behaviors and fraudulent activities, identify potential fraud risks by analyzing transaction patterns, user behavior and historical data; or act as intelligent customer service to answer customers' questions in real time, provide financial product information and business advice, and so on.

[0004] For example, in the field of medical health, multi-agent systems can generate personalized health recommendations based on the user's health status, living habits, and query needs. For example, when a user queries "how to improve sleep quality", the system can generate targeted suggestion text based on factors such as age, gender, and lifestyle; or generate health Q&A content in real time to help users quickly obtain the information they need.

[0005] However, most existing systems use a flat evaluation method, that is, all evaluation agents evaluate the generated results in parallel or sequentially, and then the system simply integrates these feedbacks. Although this approach realizes a basic feedback loop, due to the duplication and even conflict of fragmented information, the feedback process is lengthy and even falls into a loop or stagnation, thereby reducing the performance and stability of the multi-agent system and restricting the performance of the multi-agent system in high-demand tasks. Summary of the Invention

[0006] In view of the above-mentioned deficiencies in the prior art, the purpose of the present invention is to provide a hierarchical feedback collaboration method, device, equipment and medium for a multi-agent system that can be applied to the medical field, financial technology or other related fields. Its main purpose is to improve the feedback efficiency and reliability of the multi-agent system, achieve high-quality collaborative feedback to improve the performance and stability of the multi-agent system.

[0007] The technical solutions of the present invention are as follows:

[0008] A first aspect of the present invention provides a hierarchical feedback collaboration method for a multi-agent system, comprising:

[0009] The multiple agents in the multi-agent system are divided into roles according to a preset hierarchical feedback architecture to obtain a generating agent, a reviewing agent, and a supervising agent, wherein the supervising agent includes a main supervising agent and an evaluating supervising agent;

[0010] The master supervisory agent controls the generating agent to receive external input tasks and respond to them, outputting corresponding initial generated content;

[0011] The evaluation supervision agent performs multi-dimensional evaluation feedback on the initially generated content and outputs structured feedback information;

[0012] The review agent optimizes and modifies the initially generated content according to the structured feedback information, and outputs corresponding optimized generated content;

[0013] The master supervisory agent performs a quality assessment on the optimized generated content to confirm whether the preset output conditions are currently met. If not, the optimized generated content is iteratively fed back and optimized. If so, the latest optimized generated content is output to the outside.

[0014] A second aspect of the present invention provides a hierarchical feedback coordination device for a multi-agent system, comprising:

[0015] A hierarchical partitioning module is used to divide the roles of multiple agents in the multi-agent system according to a preset hierarchical feedback architecture to obtain generation agents, review agents and supervisory agents, respectively. The supervisory agents include a main supervisory agent and an evaluation supervisory agent;

[0016] An initial generation module is configured to control the generation agent to receive external input tasks and respond to them, and output corresponding initial generation content;

[0017] An evaluation and feedback module, configured to allow the evaluation and supervision agent to perform multi-dimensional evaluation and feedback on the initially generated content and output structured feedback information;

[0018] An optimization module, configured for the review agent to optimize and modify the initially generated content according to the structured feedback information, and output corresponding optimized generated content;

[0019] The output confirmation module is used to perform a quality assessment on the optimized generated content through the main supervisory intelligent agent to confirm whether the preset output conditions are currently met. If not, the optimized generated content is iteratively fed back and optimized. If it is met, the latest optimized generated content is output to the outside.

[0020] A third aspect of the present invention provides a computer device comprising at least one processor; and

[0021] a memory communicatively connected to the at least one processor; wherein,

[0022] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the above-mentioned hierarchical feedback collaboration method for the multi-agent system.

[0023] A fourth aspect of the present invention provides a non-volatile computer-readable storage medium, which stores computer-executable instructions. When the computer-executable instructions are executed by one or more processors, the one or more processors can execute the above-mentioned hierarchical feedback collaboration method of the multi-agent system.

[0024] Beneficial effects: The present invention discloses a hierarchical feedback collaboration method, device, equipment and medium for a multi-agent system. Compared with the existing technology, the embodiment of the present invention divides the roles of multiple agents in the multi-agent system according to a preset hierarchical feedback architecture to obtain a generating agent, a reviewing agent and a supervisory agent respectively. The supervisory agent includes a main supervisory agent and an evaluation supervisory agent; the main supervisory agent controls the generating agent to receive external input tasks and respond, and outputs the corresponding initial generated content; the evaluation supervisory agent performs multi-dimensional evaluation feedback on the initial generated content and outputs structured feedback information; the reviewing agent optimizes and modifies the initial generated content according to the structured feedback information and outputs the corresponding optimized generated content; the main supervisory agent performs quality evaluation on the optimized generated content to confirm whether the preset output conditions are currently met. If not, the optimized generated content is iteratively fed back and optimized. If it is met, the latest optimized generated content is output to the outside. By introducing a two-layer scheduling structure of evaluation supervision and main supervision, the multi-dimensional evaluation feedback results are integrated and output to guide content optimization, effectively alleviating the problems of evaluation conflicts and redundancy, improving feedback efficiency and reliability, and ensuring the performance and stability of the multi-agent system. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the solutions in the present invention, a brief introduction is given below to the drawings required for use in describing the embodiments of the present invention. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0026] Figure 1 A schematic diagram of an application environment for the hierarchical feedback collaboration method for a multi-agent system provided by an embodiment of the present invention;

[0027] Figure 2A flowchart of a hierarchical feedback collaboration method for a multi-agent system provided by an embodiment of the present invention;

[0028] Figure 3 A flowchart of step S202 in the hierarchical feedback collaboration method for a multi-agent system provided by an embodiment of the present invention;

[0029] Figure 4 A flowchart of step S203 in the hierarchical feedback collaboration method for a multi-agent system provided by an embodiment of the present invention;

[0030] Figure 5 A flowchart of step S204 in the hierarchical feedback collaboration method for a multi-agent system provided by an embodiment of the present invention;

[0031] Figure 6 A schematic diagram of the functional modules of a hierarchical feedback coordination device for a multi-agent system according to an embodiment of the present invention;

[0032] Figure 7 A schematic diagram of the hardware structure of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0033] To make the objectives, technical solutions, and effects of the present invention more clear and distinct, the present invention is further described in detail below. It should be understood that the specific embodiments described herein are merely for the purpose of explaining the present invention and are not intended to limit the present invention. The embodiments of the present invention are described below with reference to the accompanying drawings.

[0034] The hierarchical feedback coordination method of the multi-agent system provided by the embodiment of the present invention can be applied in the following situations: Figure 1 In an application environment, the system includes a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 is a medium for providing a communication link between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired and / or wireless communication links, etc.

[0035] The user may use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 via the network 104 to receive or send messages, etc. Various communication client applications may be installed on the first terminal device 101, the second terminal device 102, and the third terminal device 103, such as knowledge reading applications, web browser applications, search applications, instant messaging tools, email clients, and / or social platform software (for example only).

[0036] The first terminal device 101 , the second terminal device 102 , and the third terminal device 103 may be various electronic devices having display screens and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, desktop computers, and the like.

[0037] The server 105 may be a server that provides various services, such as a backend server that provides support for the content browsed by the user using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (for example only). The backend server may analyze and process the received user requests and other data, and feed back the processing results (such as web pages, information, or data obtained or generated according to the user request) to the terminal device. The server 105 may be a cloud server, also known as a cloud computing server or cloud host, which is a host product in the cloud computing service system to solve the defects of difficult management and weak business scalability in traditional physical hosts and VPS services ("Virtual Private Server", or "VPS" for short). The server 105 may also be a server for a distributed system, or a server combined with a blockchain.

[0038] It should be noted that the hierarchical feedback coordination method for a multi-agent system provided in the embodiments of the present application can generally be executed by the first terminal device 101, the second terminal device 102, or the third terminal device 103. Accordingly, the hierarchical feedback coordination apparatus for a multi-agent system provided in the embodiments of the present invention can also be set in the first terminal device 101, the second terminal device 102, or the third terminal device 103. Alternatively, the hierarchical feedback coordination method for a multi-agent system provided in the embodiments of the present invention can also generally be executed by the server 105. Accordingly, the hierarchical feedback coordination apparatus for a multi-agent system provided in the embodiments of the present invention can generally be set in the server 105.

[0039] It should be understood that the numbers of the above terminal devices, networks and servers are merely illustrative and any number of terminal devices, networks and servers may be provided as required.

[0040] like Figure 2 As shown, the hierarchical feedback collaboration method for a multi-agent system provided by an embodiment of the present invention specifically includes the following steps:

[0041] S201. Divide the roles of multiple agents in the multi-agent system according to a preset hierarchical feedback architecture to obtain a generation agent, a review agent and a supervision agent respectively. The supervision agent includes a main supervision agent and an evaluation supervision agent.

[0042] In this embodiment, a multi-agent system driven by a large language model (LLM) is a system composed of multiple agents that collaborate to complete complex tasks and simulate complex real-world environments. Each agent is an independent computing unit with its own model, goal, environmental perception, and output behavior, but they communicate and plan through collaborative channels and jointly pursue shared goals. According to the preset hierarchical feedback architecture, the roles of multiple agents in the multi-agent system are divided into three categories, namely, generation agents, review agents, and supervision agents. The supervision agents are further divided into main supervision agents and evaluation supervision agents. The main supervision agent is used for global control, the generation agent is used for content generation, the review agent is responsible for content optimization, and the evaluation supervision agent is used to evaluate and feedback the generated content.

[0043] At system startup, each agent is initialized according to pre-defined roles. Each agent is assigned specific tasks and permissions. For example, the generation agent can load pre-trained language models (such as GPT or BERT), the review agent can load text editing and optimization tools, and the supervisory agent can load evaluation criteria and decision-making algorithms. This creates a layered architecture for collaborative feedback, improving the system's organization and efficiency while reducing task conflicts and redundancy.

[0044] S202: The master supervisory agent controls the generating agent to receive external input tasks and respond to them, outputting corresponding initial generated content.

[0045] In this embodiment, the main supervisory agent is used for global control, including task allocation, process monitoring, quality assessment, and decision-making on whether iteration is required, etc. When the user inputs the corresponding task description information in the form of text or voice through the terminal, the external input task description is received under the global control of the main supervisory agent and assigned to the generating agent. After receiving the task, the generating agent starts the content generation process and calls the pre-trained language model (such as GPT) according to the received task requirements to generate preliminary content to quickly respond to user needs, providing a basis for content evaluation and optimization.

[0046] For example, in a financial intelligent customer service scenario, a user submits a question through an online customer service system. The main supervisory agent receives the question and assigns it to a generating agent. The generating agent generates a preliminary response based on the question. For example, if a customer asks about credit card repayment methods, the generating agent responds with "You can repay your credit card through mobile banking, online banking, or at the bank counter."

[0047] In a healthcare Q&A scenario, users submit questions through the health Q&A system. The main supervisory agent receives the question and assigns it to the generating agent. The generating agent generates a preliminary response based on the question. For example, if a user asks, "How can I relieve the symptoms of a cold?" the generating agent responds, "Drink more water, get enough rest, and take cold medicine to relieve symptoms."

[0048] S203: The evaluation supervision agent performs multi-dimensional evaluation feedback on the initially generated content and outputs structured feedback information.

[0049] In this embodiment, the generating agent submits the initial generated content to the main supervisory agent. After receiving the content, the main supervisory agent passes it to the evaluation supervisory agent for further evaluation. The evaluation agent provides feedback on the initial generated content in different dimensions, such as logic, accuracy, and innovation. After evaluating the initial generated content from multiple factors across different dimensions, the feedback information from different dimensions is integrated and organized into an organized and easy-to-understand format. This allows the evaluation supervisory agent to pre-process different feedback situations and then generate structured feedback information. Feedback integration is embedded as an intermediate step in the evaluation process to alleviate problems such as evaluation conflicts and redundancy in traditional feedback.

[0050] For example, in the financial intelligent customer service scenario in the financial field, the evaluation and supervision agent evaluates and provides feedback on the initially generated content from the dimensions of language accuracy and compliance, etc., and evaluates and confirms that the answers are grammatically correct and meet financial regulatory requirements. The evaluation and supervision agent integrates these feedbacks into structured feedback information, such as "Language accuracy: 90 points; Compliance: 95 points; Recommendations: None".

[0051] In the health question-and-answer scenario in the medical field, the evaluation supervision agent evaluates and provides feedback on the initially generated content from the dimensions of medical accuracy and readability. The evaluation confirms that the answers are consistent with medical common sense and easy to understand. The evaluation supervision agent integrates these feedbacks into structured feedback information, such as "Medical accuracy: 92 points; User readability: 88 points; Recommendations: None".

[0052] S204: The review agent optimizes and modifies the initially generated content according to the structured feedback information, and outputs corresponding optimized generated content.

[0053] In this embodiment, after the evaluation supervisory agent submits structured feedback information to the primary supervisory agent, the primary supervisory agent passes the feedback information to the review agent, allowing the review agent to further optimize the content based on the structured feedback information. Specifically, the review agent receives the structured feedback information, analyzes the key issues and improvement suggestions, determines the corresponding optimization direction, and uses a pre-trained language model (such as GPT) to optimize and modify the initial generated content, obtaining the modified optimized generated content. The review agent then improves the content based on the feedback information, improving its quality and making it more consistent with the task requirements and standards.

[0054] For example, in a financial intelligent customer service scenario, the review agent receives structured feedback from the evaluation and supervision agent, such as "Language accuracy: 90 points; Compliance: 95 points; Recommendations: None." The review agent then optimizes and modifies its initial response based on this feedback. For example, it might change the response from "You can repay your credit card through mobile banking, online banking, or at a bank counter" to "You can repay your credit card through mobile banking, online banking, or at a bank counter. For detailed procedures, please refer to the bank's official website or call customer service."

[0055] In a healthcare Q&A scenario, the review agent receives structured feedback from the evaluation and supervision agent, such as "Medical Accuracy: 92; User Readability: 88; Recommendations: None." The review agent then optimizes and modifies the initial response based on this feedback. For example, it might change the initial response from "Drink more water, get plenty of rest, and take cold medicine to relieve symptoms" to "Drink more water, get plenty of rest, and take cold medicine to relieve symptoms. If symptoms persist, seek medical attention."

[0056] S205. The master supervisory agent performs a quality assessment on the optimized generated content to confirm whether the preset output conditions are currently met. If not, the optimized generated content is iteratively fed back and optimized. If so, the latest optimized generated content is output to the outside.

[0057] In this embodiment, the master supervisory agent receives the optimized content and conducts a comprehensive evaluation based on preset quality standards. These quality standards may include indicators such as logic, accuracy, innovation, and language fluency. For example, the master supervisory agent can check whether the optimized content addresses key issues, meets improvement suggestions, and complies with industry standards. Based on the quality evaluation results, the master supervisory agent determines whether the current optimized content meets the preset output conditions. If the optimized content does not meet the output conditions, the master supervisory agent returns the content to the generation agent or review agent for regeneration or optimization before further evaluation. This forms a closed-loop mechanism of "generation-evaluation-correction-re-evaluation." Each round of evaluation is completed by a hierarchical structure of agents until the optimized content meets the output conditions. The master supervisory agent then outputs the final optimized content to an external system or user interface. Through comprehensive evaluation and iterative optimization of the optimized content, the quality of the content is gradually improved, ensuring that the quality of the output content meets the preset standards.

[0058] For example, in a financial intelligent customer service scenario in the financial sector, the main supervisory agent evaluates the quality of the response optimized by the review agent, for example, assessing the accuracy, completeness, and compliance of the response. If the response meets the preset output conditions, such as "the response is accurate, complete, and compliant," the response is output to the customer. If the output conditions are not met, such as "some information in the response is inaccurate," the main supervisory agent returns the response to the generating agent or review agent for regeneration or optimization.

[0059] In healthcare Q&A scenarios, the main supervisory agent evaluates the quality of the answers optimized by the reviewer, assessing, for example, their medical accuracy and user readability. If the answer meets pre-defined output criteria, such as "medically accurate and understandable," it is presented to the user. If it fails to meet these criteria, such as "some medical terms in the answer are difficult to understand," the main supervisory agent returns the answer to the generator or reviewer for regeneration or optimization.

[0060] In the above embodiment, the present invention discloses a hierarchical feedback collaboration method for a multi-agent system. The multi-agent system is divided into roles according to a preset hierarchical feedback architecture, and a generation agent, a review agent, and a supervisory agent are obtained respectively. The supervisory agent includes a main supervisory agent and an evaluation supervisory agent. The main supervisory agent controls the generation agent to receive external input tasks and respond to them, outputting corresponding initial generated content. The evaluation supervisory agent performs multi-dimensional evaluation feedback on the initial generated content and outputs structured feedback information. The review agent optimizes and modifies the initial generated content based on the structured feedback information and outputs corresponding optimized generated content. The main supervisory agent performs quality evaluation on the optimized generated content to confirm whether the preset output conditions are currently met. If not, the optimized generated content is iteratively fed back and optimized. If it is met, the latest optimized generated content is output to the outside. By introducing a dual-layer scheduling structure of evaluation supervision and main supervision, the multi-dimensional evaluation feedback results are integrated and output to guide content optimization, effectively alleviating the problems of evaluation conflict and redundancy, improving feedback efficiency and reliability, and ensuring the performance and stability of the multi-agent system.

[0061] In one embodiment, Figure 3 As shown, step S202 includes:

[0062] S401, in response to a task start instruction, the master supervisory agent controls the generating agent to receive an externally input task description text;

[0063] S402: The generating agent parses the task description text to obtain task objectives and generation requirements;

[0064] S403: Perform text generation processing according to the task objectives and generation requirements, and output corresponding initial generated content.

[0065] In this embodiment, the master supervisory agent receives externally input task initiation instructions through a system interface. The task initiation instructions contain a task description, such as a user's question or requirement. The master supervisory agent parses the task initiation instructions, extracts the task description, and assigns it to the generating agent. After receiving the task description, the generating agent sends a confirmation signal to the master supervisory agent to ensure that the task has been correctly assigned. Preferably, when there are multiple tasks, the master supervisory agent can place the tasks in a task queue and assign them to the generating agents in sequence according to task priority or a first-in-first-out (FIFO) principle. Alternatively, the master supervisory agent can dynamically assign tasks based on the load of the generating agents to improve overall efficiency.

[0066] After receiving the task description text, the generative agent uses natural language processing (NLP) technology to parse the text and extract key information. Specifically, through semantic analysis and keyword extraction, the generative agent identifies the task objective, such as "answering user questions" or "generating a report." It also extracts specific requirements for the generated content, such as "formal language style" or "inclusion of specific data." The generative agent then uses a pre-trained language model to generate text, generating initial content that meets the task objective and generation requirements. This rapid generation of initial content that meets these requirements provides a foundation for subsequent optimization.

[0067] In one embodiment, Figure 4 As shown, step S203 includes:

[0068] S401, the evaluation supervision agent distributes the initially generated content to multiple evaluation agents, each evaluation agent having a different evaluation dimension;

[0069] S402, using multiple evaluation agents to independently evaluate the initially generated content according to their respective evaluation dimensions, and generating evaluation opinions in multiple dimensions;

[0070] S403: The evaluation supervision agent processes the evaluation opinions of the multiple dimensions and outputs structured feedback information.

[0071] In this embodiment, after the generation agent completes the preliminary output, the evaluation supervision agent is started under the control of the main supervisor, and the evaluation supervision agent receives the initial generated content. According to preset rules or task requirements, the evaluation supervision agent determines the dimensions that need to be evaluated, such as logic, accuracy, innovation, language style, etc., and distributes the initial generated content to multiple evaluation agents within the evaluation supervision agent according to the dimensions that need to be evaluated. Each evaluation agent has different evaluation dimensions and performs independent evaluation processing on the content from a specific dimension.

[0072] Each evaluation agent independently assesses the initially generated content based on its own evaluation dimension, avoiding interference. Based on the evaluation results, it generates an opinion, which can include a score and comments, reflecting the quality of the content in that specific dimension. For example, a logic evaluation agent might give a rating of "Logic score: 85 points, Comment: The logic is clear, but some arguments are insufficient."

[0073] Each evaluation agent sends its opinion back to the evaluation supervisor. This means that the opinions generated by multiple evaluation agents do not directly contribute to the generated content. Instead, they are returned to the intermediate evaluation supervisor for overall data processing, such as denoising, aggregation, and structuring, before outputting structured feedback information. This allows feedback integration to be embedded as an intermediate step in the evaluation process, preventing fragmented feedback from leading to a lengthy feedback process, or even looping or stagnating, and improving the stability and reliability of feedback.

[0074] In one embodiment, step S403 includes:

[0075] Performing data preprocessing on the evaluation opinions of the multiple dimensions to obtain preprocessed evaluation data in a unified format;

[0076] Extract key information from the pre-processed evaluation data under each dimension and add corresponding feedback type labels;

[0077] Aggregating the pre-processed evaluation data of all dimensions according to the feedback type labels to generate comprehensive feedback data for each feedback type;

[0078] The comprehensive feedback data of each feedback type is converted and processed according to a preset structured format to output structured feedback information.

[0079] In this embodiment, the evaluation supervision agent receives evaluation opinions from various evaluation agents. The specific evaluation opinions may include different forms such as scores, comments, and marks. Since evaluation opinions of different dimensions are independently output by different evaluation agents, they may exist in different formats. Through data preprocessing, the evaluation opinions of multiple dimensions are formatted in a standardized manner, and the evaluation opinions in different formats are converted into a unified format. For example, all scores are unified into a percentage system, and the comments are unified into a text format, etc., so as to improve the efficiency of subsequent data processing.

[0080] Key information is extracted from the pre-processed evaluation data under each dimension. Key information such as scores, main issues, and advantages are extracted from the pre-processed evaluation data, and corresponding feedback type labels are added to classify them into different feedback types, such as "scores", "logical issues", "data accuracy issues", "language style issues", "improvements", etc. The amount of data processing is reduced through key information extraction to avoid information overload. According to the extracted feedback type labels, the pre-processed evaluation data of all dimensions are classified and aggregated, so that the data of each feedback type can be comprehensively analyzed and integrated to obtain comprehensive feedback for each feedback type, such as data obtained, such as calculating comprehensive scores, comprehensive key issues, comprehensive improvement suggestions, etc., which provides a clear basis for subsequent optimization and facilitates rapid problem location and improvement.

[0081] The output format for each feedback type is then defined according to a pre-set structured format, such as a table or JSON format. The comprehensive feedback data is converted into the pre-set structured format to ensure the consistency and readability of the output information. The converted structured feedback information is then output to the review agent for content optimization. Preferably, the structured feedback information can also be presented in a visual format, such as through charts or dashboards, to improve readability and allow users to intuitively understand the generation and evaluation optimization process of the agent system.

[0082] In one embodiment, after aggregating the pre-processed evaluation data of all dimensions according to the feedback type label to generate comprehensive feedback data for each feedback type, the method further includes:

[0083] De-duplicate the comprehensive feedback data for each feedback type;

[0084] Conflict detection is performed on the deduplicated integrated feedback data, and logical optimization processing is performed on the deduplicated integrated feedback data according to the conflict detection result.

[0085] In this embodiment, after the comprehensive feedback data of each feedback type is aggregated and generated, the comprehensive feedback data is processed by a deduplication algorithm, such as hash value-based deduplication (determining duplication by calculating the hash value of the text) and similarity-based deduplication (determining duplication by calculating the similarity between texts), etc., to identify repeated key information, such as the same rating or similar problem description, and delete the repeated data to ensure the uniqueness of the data and avoid data redundancy.

[0086] After deduplication, the comprehensive feedback data is analyzed to identify contradictory or inconsistent information, such as situations where different evaluation dimensions may have inconsistent evaluations of the same issue. Specifically, it can be detected through conflict identification algorithms, such as rule-based conflict detection or logical reasoning-based conflict detection, etc., to mark the detected conflict information and record the type and location of the conflict. Based on the recorded conflict detection results, the conflict information is logically optimized, such as merging conflict information, that is, if the conflict information is due to different expressions but the same problem, they can be merged into a unified description; or priority adjustment, that is, if the conflict information cannot be merged, according to the preset priority rules, more important information is retained, and secondary information is deleted or adjusted, etc. The comprehensive feedback data is updated based on the results of the logical optimization processing. Through conflict detection and optimization processing, the logical consistency of the comprehensive feedback data is ensured, and the interference of contradictory information on subsequent optimization is avoided, providing a clearer and more reliable decision-making basis for the review intelligent body, and improving the optimization efficiency.

[0087] In one embodiment, Figure 5As shown, step S204 includes:

[0088] S501: The review agent receives and parses the structured feedback information to extract comprehensive feedback data from each field.

[0089] S502: Confirm the key issues to be optimized and the corresponding optimization strategies based on the comprehensive feedback data in each field;

[0090] S503: Optimize and modify the content corresponding to the key question in the initially generated content according to the optimization strategy;

[0091] S504. After all key issues to be optimized have been optimized and modified, the corresponding optimized generated content is integrated and output.

[0092] In this embodiment, the proofreading agent receives the structured feedback information output by the evaluation and supervision agent, and parses the structured feedback information into multiple fields through data parsing, such as scores, problem descriptions, improvement suggestions, etc., and extracts specific comprehensive feedback data from each field, such as extracting specific problems from the "problem description" field and extracting optimization suggestions from the "improvement suggestions" field. Based on the comprehensive feedback data, the key issues to be optimized and the optimization strategies of teammates are identified. Specifically, the key issues to be optimized are identified from the "problem description" field, such as "lack of logic" and "inaccurate data". And according to the content in the "improvement suggestions" field, the corresponding optimization strategy is matched for each key problem. For example, for the problem of "lack of logic", the optimization strategy may be "reorganizing the paragraph structure", and the problem is associated with the corresponding optimization strategy to ensure the targeted optimization.

[0093] When optimizing content, first identify the sections of the initially generated content that correspond to key issues. For example, if the issue is "lack of logic," locate the paragraphs with disorganized logical structures. Modify the identified content based on the optimization strategy, such as reorganizing the paragraph structure to improve logic. This will address key issues in the initially generated content and improve its quality. Integrate all modified sections into the initially generated content, and perform a consistency check on the integrated optimized content to ensure its logical and stylistic coherence, resulting in complete, high-quality optimized content and optimized processing of the initially generated content.

[0094] In one embodiment, step S503 includes:

[0095] Prioritize all key issues to be optimized;

[0096] In descending order of priority, the corresponding contents in the initially generated contents are optimized and modified based on the optimization strategies of the key issues.

[0097] In this embodiment, when there are multiple key issues to be optimized, all key issues to be optimized are prioritized. Specifically, they can be prioritized according to preset priority rules, such as accuracy issues, logic issues, and fluency issues, with the priorities decreasing in sequence, etc., so that when performing optimization and modification, the corresponding content in the initially generated content can be optimized and modified in order from high to low priority based on the optimization strategy of each key issue, ensuring that more important issues are optimized first and improving resource utilization.

[0098] It should be noted that there is not necessarily a certain order between the above steps. A person skilled in the art can understand, based on the description of the embodiments of the present invention, that in different embodiments, the above steps may have different execution orders, that is, they may be executed in parallel, or may be executed interchangeably, etc.

[0099] Further references Figure 6 , as a response to the above Figure 2 The present invention provides an embodiment of a hierarchical feedback coordination device for a multi-agent system. Figure 2 Corresponding to the method embodiment shown, the device can be specifically applied to various electronic devices.

[0100] like Figure 6 As shown, the hierarchical feedback coordination device 60 of the multi-agent system described in this embodiment includes:

[0101] A hierarchical division module 601 is used to divide the multiple agents in the multi-agent system into roles according to a preset hierarchical feedback architecture, thereby obtaining a generating agent, a reviewing agent, and a supervisory agent, wherein the supervisory agent includes a main supervisory agent and an evaluation supervisory agent;

[0102] Initial generation module 602, configured to control the generation agent to receive external input tasks and respond to them, and output corresponding initial generation content;

[0103] Evaluation and feedback module 603, configured to perform multi-dimensional evaluation and feedback on the initially generated content by the evaluation supervision agent, and output structured feedback information;

[0104] Optimization module 604, configured for the review agent to optimize and modify the initially generated content according to the structured feedback information, and output corresponding optimized generated content;

[0105] The confirmation output module 605 is used to perform a quality assessment on the optimized generated content through the main supervisory intelligent agent to confirm whether the preset output conditions are currently met. If not, iterative feedback and optimization are performed on the optimized generated content. If satisfied, the latest optimized generated content is output to the outside.

[0106] The module referred to in the present invention refers to a series of computer program instruction segments that can perform specific functions. It is more suitable for describing the hierarchical feedback collaborative execution process of a multi-agent system than a program. For the specific implementation of each module, please refer to the corresponding method embodiment above, which will not be repeated here.

[0107] In one embodiment, the initial generation module 602 includes:

[0108] A task receiving unit, configured to control the generating agent to receive an externally input task description text in response to a task start instruction;

[0109] A parsing unit, configured to control the generating agent to parse the task description text to obtain task objectives and generation requirements;

[0110] The text generation unit is used to perform text generation processing according to the task objectives and generation requirements, and output corresponding initial generated content.

[0111] In one embodiment, the evaluation feedback module 603 includes:

[0112] a task distribution unit, configured to control the evaluation supervision agent to distribute the initially generated content to a plurality of evaluation agents, each evaluation agent having a different evaluation dimension;

[0113] An independent evaluation unit, configured to independently evaluate the initially generated content according to respective evaluation dimensions through a plurality of evaluation agents, and generate evaluation opinions in a plurality of dimensions;

[0114] The data integration processing unit is used to control the evaluation supervision agent to process the evaluation opinions of the multiple dimensions and output structured feedback information.

[0115] In one embodiment, the data integration processing unit includes:

[0116] A preprocessing unit, configured to perform data preprocessing on the evaluation opinions of the multiple dimensions to obtain preprocessed evaluation data in a unified format;

[0117] An extraction unit, configured to extract key information from the pre-processed evaluation data under each dimension and add a corresponding feedback type label;

[0118] an aggregation unit, configured to aggregate the pre-processed evaluation data of all dimensions according to the feedback type label to generate comprehensive feedback data for each feedback type;

[0119] The structured conversion unit is used to convert the comprehensive feedback data of each feedback type according to a preset structured format and output structured feedback information.

[0120] In one embodiment, the data integration processing unit further includes:

[0121] A deduplication unit, used for performing deduplication processing on the comprehensive feedback data of each feedback type;

[0122] The conflict detection optimization unit is used to perform conflict detection on the integrated feedback data after deduplication processing, and perform logic optimization processing on the integrated feedback data after deduplication processing according to the conflict detection result.

[0123] In one embodiment, the optimization module 604 includes:

[0124] A feedback parsing unit, configured to control the review agent to receive and parse the structured feedback information, and extract comprehensive feedback data from each field;

[0125] A confirmation unit is used to confirm the key issues to be optimized and the corresponding optimization strategies based on the comprehensive feedback data in each field;

[0126] An optimization unit, configured to optimize and modify the content corresponding to the key issue in the initially generated content according to the optimization strategy;

[0127] The output unit is used to integrate and output the corresponding optimization generated content after all key issues to be optimized have completed corresponding optimization modifications.

[0128] In one embodiment, the optimization unit includes:

[0129] Priority processing unit, used to prioritize all key issues to be optimized;

[0130] The text modification unit is used to optimize and modify the corresponding content in the initially generated content in order from high to low priority based on the optimization strategy of the key issues.

[0131] In the above embodiment, the present invention discloses a hierarchical feedback collaboration device for a multi-agent system. The multi-agent system is divided into roles according to a preset hierarchical feedback architecture, and a generation agent, a review agent, and a supervisory agent are obtained respectively. The supervisory agent includes a main supervisory agent and an evaluation supervisory agent. The main supervisory agent controls the generation agent to receive external input tasks and respond to them, outputting corresponding initial generated content. The evaluation supervisory agent performs multi-dimensional evaluation feedback on the initial generated content and outputs structured feedback information. The review agent optimizes and modifies the initial generated content based on the structured feedback information and outputs corresponding optimized generated content. The main supervisory agent performs quality evaluation on the optimized generated content to confirm whether the preset output conditions are currently met. If not, the optimized generated content is iteratively fed back and optimized. If it is met, the latest optimized generated content is output to the outside. By introducing a dual-layer scheduling structure of evaluation supervision and main supervision, the multi-dimensional evaluation feedback results are integrated and output to guide content optimization, effectively alleviating the problems of evaluation conflict and redundancy, improving feedback efficiency and reliability, and ensuring the performance and stability of the multi-agent system.

[0132] Another embodiment of the present invention provides a computer device, such as Figure 7 As shown, the computer device 70 includes:

[0133] One or more processors 701 and memory 702, Figure 7 In the description, a processor 701 is used as an example. The processor 701 and the memory 702 can be connected via a bus or other means. Figure 7 The bus connection is taken as an example.

[0134] The processor 701 is used to complete various control logics of the computer device 70. It can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), a single-chip microcomputer, an ARM (Acorn RISC Machine) or other programmable logic device, discrete gate or transistor logic, discrete hardware components or any combination of these components. In addition, the processor 701 can also be any traditional processor, microprocessor or state machine. The processor 701 can also be implemented as a combination of computing devices, for example, a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors combined with a DSP and / or any other such configuration.

[0135] Memory 702, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer executable programs, and modules, such as program instructions corresponding to the hierarchical feedback collaboration method for a multi-agent system in the embodiments of the present invention. Processor 701 executes the non-volatile software programs, instructions, and modules stored in memory 702 to execute various functional applications and data processing of computer device 70, thereby implementing the hierarchical feedback collaboration method for a multi-agent system in the above-mentioned method embodiments.

[0136] The memory 702 may include a program storage area and a data storage area, wherein the program storage area may store an operating system and application programs required for at least one function; the data storage area may store data created according to the use of the computer device 70, etc. In addition, the memory 702 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 702 may optionally include a memory remotely located relative to the processor 701, and these remote memories may be connected to the computer device 70 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof. One or more units are stored in the memory 702, and when executed by one or more processors 701, the steps of the hierarchical feedback collaboration method of the multi-agent system in any of the above-mentioned method embodiments are executed.

[0137] In the above embodiment, the present invention discloses a computer device, which divides the roles of multiple agents in a multi-agent system according to a preset hierarchical feedback architecture, thereby obtaining a generating agent, a reviewing agent, and a supervisory agent, respectively. The supervisory agent includes a main supervisory agent and an evaluation supervisory agent; the main supervisory agent controls the generating agent to receive external input tasks and respond to them, outputting corresponding initial generated content; the evaluation supervisory agent performs multi-dimensional evaluation feedback on the initial generated content and outputs structured feedback information; the reviewing agent optimizes and modifies the initial generated content based on the structured feedback information and outputs corresponding optimized generated content; the main supervisory agent performs quality evaluation on the optimized generated content to confirm whether the preset output conditions are currently met. If not, it iterative feedback and optimization are performed on the optimized generated content. If it is met, the latest optimized generated content is output to the outside. By introducing a two-layer scheduling structure of evaluation supervision and main supervision, the multi-dimensional evaluation feedback results are integrated and output to guide content optimization, effectively alleviating the problems of evaluation conflict and redundancy, improving feedback efficiency and reliability, and ensuring the performance and stability of the multi-agent system.

[0138] An embodiment of the present invention provides a non-volatile computer-readable storage medium, which stores computer-executable instructions. When the computer-executable instructions are executed by one or more processors, the steps of the hierarchical feedback collaboration method of the multi-agent system in any of the above method embodiments are executed.

[0139] In the above embodiment, the present invention discloses a non-volatile computer-readable storage medium, which divides the roles of multiple agents in a multi-agent system according to a preset hierarchical feedback architecture to obtain a generating agent, a reviewing agent, and a supervisory agent, respectively. The supervisory agent includes a main supervisory agent and an evaluation supervisory agent; the main supervisory agent controls the generating agent to receive external input tasks and respond to them, outputting corresponding initial generated content; the evaluation supervisory agent performs multi-dimensional evaluation feedback on the initial generated content and outputs structured feedback information; the reviewing agent optimizes and modifies the initial generated content based on the structured feedback information and outputs corresponding optimized generated content; the main supervisory agent performs quality assessment on the optimized generated content to confirm whether the preset output conditions are currently met. If not, it iterative feedback and optimization are performed on the optimized generated content. If it is met, the latest optimized generated content is output to the outside. By introducing a two-layer scheduling structure of evaluation supervision and main supervision, the multi-dimensional evaluation feedback results are integrated and output to guide content optimization, effectively alleviating the problems of evaluation conflict and redundancy, improving feedback efficiency and reliability, and ensuring the performance and stability of the multi-agent system.

[0140] Through the description of the above embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus the necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better embodiment. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for enabling a terminal device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in each embodiment of the present invention.

[0141] The present invention can be used in a wide variety of general or special computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, and the like. The present invention can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present invention can also be practiced in distributed computing environments in which tasks are performed by remote processing devices connected via a communications network. In a distributed computing environment, program modules can be located in local and remote computer storage media, including storage devices.

[0142] In summary, the hierarchical feedback collaboration method, device, equipment and medium of the multi-agent system disclosed in the present invention include: dividing the roles of multiple agents in the multi-agent system according to a preset hierarchical feedback architecture, and obtaining a generating agent, a reviewing agent and a supervisory agent respectively, and the supervisory agent includes a main supervisory agent and an evaluation supervisory agent; the main supervisory agent controls the generating agent to receive external input tasks and respond, and outputs the corresponding initial generated content; the evaluation supervisory agent performs multi-dimensional evaluation feedback on the initial generated content and outputs structured feedback information; the reviewing agent optimizes and modifies the initial generated content according to the structured feedback information, and outputs the corresponding optimized generated content; the main supervisory agent performs quality evaluation on the optimized generated content to confirm whether the preset output conditions are currently met. If not, the optimized generated content is iteratively fed back and optimized. If it is met, the latest optimized generated content is output to the outside. By introducing a two-layer scheduling structure of evaluation supervision and main supervision, the multi-dimensional evaluation feedback results are integrated and output to guide content optimization, effectively alleviating the problems of evaluation conflicts and redundancy, improving feedback efficiency and reliability, and ensuring the performance and stability of the multi-agent system.

[0143] Of course, those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware (such as a processor, controller, etc.) through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes in the above-described method embodiments. The storage medium can be a memory, a magnetic disk, a floppy disk, a flash memory, an optical storage device, etc.

[0144] It should be noted that if any software tools or components not developed by our company appear in the examples of this application, they are for illustration purposes only and do not represent actual use. It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications shall fall within the scope of protection of the appended claims.

Claims

1. A hierarchical feedback collaboration method for a multi-agent system, characterized in that: include: The multiple agents in the multi-agent system are divided into roles according to a preset hierarchical feedback architecture to obtain a generating agent, a reviewing agent, and a supervising agent, wherein the supervising agent includes a main supervising agent and an evaluating supervising agent; The master supervisory agent controls the generating agent to receive external input tasks and respond to them, outputting corresponding initial generated content; The evaluation supervision agent performs multi-dimensional evaluation feedback on the initially generated content and outputs structured feedback information; The review agent optimizes and modifies the initially generated content according to the structured feedback information, and outputs corresponding optimized generated content; The master supervisory agent performs a quality assessment on the optimized generated content to confirm whether the preset output conditions are currently met. If not, the optimized generated content is iteratively fed back and optimized. If so, the latest optimized generated content is output to the outside.

2. The hierarchical feedback collaboration method for a multi-agent system according to claim 1, characterized in that: The master supervisory agent controls the generating agent to receive external input tasks and respond to them, and output corresponding initial generated content, including: In response to a task start instruction, the master supervisory agent controls the generating agent to receive an externally input task description text; The generating agent parses the task description text to obtain the task objectives and generation requirements; Perform text generation processing according to the task objectives and generation requirements, and output corresponding initial generated content.

3. The hierarchical feedback collaboration method for a multi-agent system according to claim 1, characterized in that: The evaluation supervision agent performs multi-dimensional evaluation feedback on the initially generated content and outputs structured feedback information, including: The evaluation supervision agent distributes the initial generated content to multiple evaluation agents, each evaluation agent having a different evaluation dimension; Using multiple evaluation agents to independently evaluate the initially generated content according to their respective evaluation dimensions, and generate evaluation opinions in multiple dimensions; The evaluation supervision agent processes the evaluation opinions of the multiple dimensions and outputs structured feedback information.

4. The hierarchical feedback collaboration method for a multi-agent system according to claim 3, characterized in that: The evaluation supervision agent processes the evaluation opinions of the multiple dimensions and outputs structured feedback information, including: Performing data preprocessing on the evaluation opinions of the multiple dimensions to obtain preprocessed evaluation data in a unified format; Extract key information from the pre-processed evaluation data under each dimension and add corresponding feedback type labels; Aggregating the pre-processed evaluation data of all dimensions according to the feedback type labels to generate comprehensive feedback data for each feedback type; The comprehensive feedback data of each feedback type is converted and processed according to a preset structured format to output structured feedback information.

5. The hierarchical feedback collaboration method for a multi-agent system according to claim 4, characterized in that: After aggregating the pre-processed evaluation data of all dimensions according to the feedback type labels to generate comprehensive feedback data of each feedback type, the method further includes: De-duplicate the comprehensive feedback data for each feedback type; Conflict detection is performed on the deduplicated integrated feedback data, and logical optimization processing is performed on the deduplicated integrated feedback data according to the conflict detection result.

6. The hierarchical feedback collaboration method for a multi-agent system according to claim 1, characterized in that: The review agent optimizes and modifies the initially generated content according to the structured feedback information and outputs corresponding optimized generated content, including: The review agent receives and parses the structured feedback information, extracting comprehensive feedback data from each field; Identify key issues to be optimized and corresponding optimization strategies based on the comprehensive feedback data in each field; Optimizing and modifying the content corresponding to the key issues in the initially generated content according to the optimization strategy; After all key issues to be optimized have been optimized and modified accordingly, the corresponding optimized generated content is integrated and output.

7. The hierarchical feedback collaboration method for a multi-agent system according to claim 6, characterized in that: Optimizing and modifying the content corresponding to the key issue in the initially generated content according to the optimization strategy includes: Prioritize all key issues to be optimized; In descending order of priority, the corresponding contents in the initially generated contents are optimized and modified based on the optimization strategies of the key issues.

8. A hierarchical feedback coordination device for a multi-agent system, characterized in that: include: A hierarchical partitioning module is used to divide the roles of multiple agents in the multi-agent system according to a preset hierarchical feedback architecture to obtain generation agents, review agents and supervisory agents, respectively. The supervisory agents include a main supervisory agent and an evaluation supervisory agent; An initial generation module is configured to control the generation agent to receive external input tasks and respond to them, and output corresponding initial generation content; An evaluation and feedback module, configured to allow the evaluation and supervision agent to perform multi-dimensional evaluation and feedback on the initially generated content and output structured feedback information; An optimization module, configured for the review agent to optimize and modify the initially generated content according to the structured feedback information, and output corresponding optimized generated content; The output confirmation module is used to perform a quality assessment on the optimized generated content through the main supervisory intelligent agent to confirm whether the preset output conditions are currently met. If not, the optimized generated content is iteratively fed back and optimized. If it is met, the latest optimized generated content is output to the outside.

9. A computer device, characterized in that: comprising at least one processor; and, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the hierarchical feedback collaboration method for a multi-agent system as described in any one of claims 1-7.

10. A non-volatile computer-readable storage medium, characterized in that: The non-volatile computer-readable storage medium stores computer-executable instructions, which, when executed by one or more processors, enable the one or more processors to execute the hierarchical feedback collaboration method for a multi-agent system as described in any one of claims 1-7.

Images