Task adaptive analysis processing method, system and equipment based on large-model multi-agent and medium

Abstract

The invention relates to the technical field of multi-agent task analysis, and discloses a task adaptive analysis processing method, system and device based on large-model multi-agent, and a medium, and the method comprises the steps: inputting a task; in the outer circulation processing stage, the task is disassembled into sub-tasks, intelligent matching execution agents and parallel processing are carried out, results are integrated, the quality is evaluated through a double-track scoring mechanism, and a scoring result is obtained; for the subtasks with the complexity exceeding a first threshold value, triggering an internal loop processing stage, performing secondary disassembly to obtain atomic tasks, processing the atomic tasks by a fine-grained agent, integrating results, and returning to an external loop; and determining whether to iteratively optimize according to a scoring result, and outputting an optimization scheme and a quality report. According to the method, dynamic adjustment can be automatically carried out according to task complexity through an internal and external circulation double-layer processing mechanism, so that the method not only can process macroscopic strategic problems, but also can cope with fine operation requirements, and the adaptability and the processing depth of multi-level and high-complexity tasks are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of multi-agent task analysis technology, and in particular to a task adaptive analysis and processing method, system, device and medium based on a large model multi-agent. Background Technology

[0002] With the rapid development of artificial intelligence technology, large-scale models such as large language models have demonstrated powerful capabilities in natural language understanding, content generation, and complex reasoning. Building on this, multi-agent systems, through the division of labor and collaboration among multiple specialized agents, offer a new paradigm for solving complex tasks. These systems can decompose macro-level tasks and have different agents process sub-tasks in parallel, theoretically significantly improving the efficiency and quality of handling complex problems.

[0003] The task decomposition mechanism lacks dynamism, often relying on preset fixed rules for task splitting. It fails to adjust the decomposition granularity based on complexity indicators such as the number of processing steps and dependencies in subtasks, leading to redundant and wasteful decomposition of simple tasks and fragmented results due to overly coarse decomposition of complex tasks. The agent matching dimension is singular, considering only a single factor in capability matching or load balancing, without achieving optimal allocation through a multi-factor decision model that integrates capability matching and real-time load rate. This easily results in problems such as overload and delay of high-capability agents or mismatch between agent capabilities and tasks. The quality assessment and iteration mechanism is imperfect, lacking a dual-track scoring mechanism that simultaneously covers language scoring for semantic consistency and logical coherence, and rule scoring for format standardization and content completeness. Furthermore, there is no design to select the iteration path based on the problem type when the threshold is not reached, resulting in uncontrollable result quality and low iteration optimization efficiency. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a task adaptive analysis and processing method and system based on a large model and multiple agents to solve the problems of lack of dynamism in current task decomposition mechanisms, mismatch between agent capabilities and tasks, and low optimization efficiency.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a task adaptive analysis and processing method based on a large model multi-agent, comprising: inputting a task, receiving and verifying the task through a standardized interface, encapsulating the task, and starting an outer loop processing flow. In the outer loop processing stage, the task is broken down into sub-tasks, intelligent agents are matched to execute the tasks, the results are processed in parallel and integrated, and the quality is evaluated through a dual-track scoring mechanism to obtain a score result. For subtasks whose complexity exceeds the first threshold, the inner loop processing stage is triggered to perform a second decomposition into atomic tasks, which are then processed by a fine-grained agent and the results are integrated before being returned to the outer loop. The decision to iterate and optimize is based on the scoring results, with a maximum number of iterations set until the quality meets the standard or the process is forcibly terminated. Output optimization solutions and quality reports, support multiple delivery methods, release resources and record complete processing logs.

[0007] As a preferred embodiment of the task adaptive analysis and processing method based on a large model multi-agent described in this invention, in the outer loop processing stage, the task is decomposed into sub-tasks, including: The task decomposition intelligent agent is built based on a large language model to parse task objects and use semantic analysis to perform intent recognition, key information extraction and task element analysis. During the parsing process, the agent performs multiple rounds of understanding of the task text, performs grammatical parsing and semantic role labeling, and identifies the main actions, objects and modifying conditions in the task; By using knowledge graph association technology, the concepts involved in the task are matched with existing knowledge bases; Perform task boundary delineation, obtain the task scope and optional extended content, in order to complete the parsing.

[0008] As a preferred embodiment of the task adaptive analysis and processing method based on a large model and multiple agents described in this invention, the method includes: intelligent matching of executing agents, parallel processing and integration of results, including: The task distribution agent assigns a suitable execution agent to the sub-task based on a multi-factor decision model, and the assignment of the execution agent is performed by a matching degree calculation method. During the matching degree calculation process, based on the agent's historical performance data and dynamic weight adjustment, the values ​​of WP and WI are automatically adjusted according to the task type and urgency to increase the proportion of load weight. Each agent receives a task and works in parallel. The execution process includes parameter parsing, model invocation, and result generation.

[0009] As a preferred embodiment of the task adaptive analysis and processing method based on a large model and multiple agents described in this invention, the quality is evaluated through a dual-track scoring mechanism, including: the task correction agent initiates the dual-track scoring mechanism to perform quality evaluation, including language and rule scoring, wherein: Language scoring assesses deeper quality such as semantic consistency and logical coherence through a large language model; The rule scoring is based on a predefined set of rules to check for compliance requirements regarding format conformity and content completeness.

[0010] As a preferred embodiment of the task adaptive analysis and processing method based on a large model and multiple agents described in this invention, the complexity is calculated by comprehensively analyzing the number of processing steps of sub-tasks and the quantitative indicators of dependency relationships. After the inner loop is triggered, the task allocation agent takes over the corresponding complex subtask and performs a secondary decomposition to generate several atomic tasks. Atomic tasks are assigned to corresponding fine-grained agents for execution, with each agent optimized for a specific task type.

[0011] The beneficial effects of this preferred technical solution are: through the dual-layer processing mechanism of macroscopic decomposition in the outer loop and microscopic refinement in the inner loop, it can automatically and dynamically adjust according to the complexity of the task.

[0012] As a preferred embodiment of the task adaptive analysis and processing method based on a large model and multiple agents described in this invention, the method includes: determining whether to iterate and optimize based on the scoring results, setting a maximum number of iterations, and continuing until the quality meets the standard or is forcibly terminated, including: Set a second threshold; When either value in the dual-track scoring fails to reach the second threshold, the iteration path is intelligently selected based on the nature of the problem. If the problem involves the global structure or core logic, the correction task is sent back to the task decomposition step for reorganization; if it is only a local content or format problem, it is directly distributed to the corresponding agent for quick revision. After each outer loop execution, the dual-track scoring is performed again. If any score fails to meet the standard, iterative optimization continues, with a maximum number of iterations set, until the quality meets the standard or is forcibly terminated.

[0013] The beneficial effects of this preferred technical solution are as follows: the introduction of a dual-track scoring mechanism and intelligent iterative optimization can quantitatively evaluate the processing results of each round, ensuring the high quality and compliance of the final output solution, while realizing closed-loop optimization of the processing process.

[0014] As a preferred embodiment of the task adaptive analysis and processing method based on a large model and multiple agents described in this invention, the task input content is natural language text, and the text length is set within a valid range.

[0015] Secondly, the present invention provides a task adaptive analysis and processing system based on a large model multi-agent system, comprising: The input module is used to input tasks, receive and verify the tasks through a standardized interface, encapsulate them, and start the outer loop processing flow. The outer loop processing module is used to break down tasks into subtasks, intelligently match and execute intelligent agents, process and integrate results in parallel, evaluate quality through a dual-track scoring mechanism, and obtain a score result. The inner loop processing module is used to trigger the inner loop processing stage for subtasks whose complexity exceeds the first threshold. The subtasks are then further decomposed into atomic tasks, which are processed by a fine-grained intelligent agent. The results are then integrated and returned to the outer loop. The iterative optimization module is used to determine whether to perform iterative optimization based on the scoring results, set the maximum number of iterations, and continue until the quality meets the standard or is forcibly terminated. The output module is used to output optimization solutions and quality reports, supports multiple delivery methods, releases resources, and records complete processing logs.

[0016] Thirdly, the present invention provides a computer device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of a task adaptive analysis and processing method based on a large model multi-agent system.

[0017] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the task adaptive analysis and processing method based on a large model multi-agent system.

[0018] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention employs a dual-layer processing mechanism of macroscopic decomposition in the outer loop and microscopic refinement in the inner loop, enabling automatic dynamic adjustment based on task complexity. It can handle both macro-strategic issues and fine-grained operational needs, enhancing adaptability and processing depth for multi-level, highly complex tasks. Combined with a task distribution agent that makes decisions based on a matching degree model integrating capability matching and real-time load rate, optimal matching between tasks and agents can be achieved. Furthermore, by introducing a dual-track scoring mechanism and intelligent iterative optimization process, the system can quantitatively evaluate the processing results of each round, ensuring the high quality and compliance of the final output solution. Simultaneously, it achieves closed-loop optimization of the processing process, resulting in an overall improvement in system throughput, response speed, and resource utilization efficiency. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the overall process of a task adaptive analysis and processing method based on a large model and multiple agents according to an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the overall process architecture for task adaptive analysis and processing based on a large model and multiple agents, according to an embodiment of the present invention. Detailed Implementation

[0022] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0023] Example 1, referring to Figure 1 As an embodiment of the present invention, a task adaptive analysis and processing method based on a large model multi-agent is provided, comprising: S100: Input the task, receive and verify the task through a standardized interface, encapsulate it, and start the outer loop processing flow; S200: In the outer loop processing stage, the task is broken down into sub-tasks, intelligent agents are matched to execute the task, the results are processed in parallel and integrated, and the quality is evaluated through a dual-track scoring mechanism to obtain the scoring result. S300: For subtasks whose complexity exceeds the first threshold, the inner loop processing stage is triggered to perform a second decomposition into atomic tasks, which are then processed by a fine-grained agent and the results are integrated before being returned to the outer loop. S400: Determine whether to iterate and optimize based on the scoring results, set the maximum number of iterations, and continue until the quality meets the standard or the process is forcibly terminated. S500: Outputs optimization solutions and quality reports, supports multiple delivery methods, releases resources, and records complete processing logs.

[0024] It should be noted that current multi-agent designs generally rely on pre-defined, fixed rules to break down macro-level tasks, such as decomposing tasks based on keywords or simple hierarchical structures. Therefore, regardless of whether the task is a simple query or a multi-step strategic planning process, the same decomposition template is used, leading to uneven distribution of resources such as computing power. Furthermore, forcibly breaking down simple tasks into multiple sub-tasks introduces unnecessary intermediate steps, increasing computational overhead and response latency; complex tasks, on the other hand, result in logical gaps and incompleteness. The evaluation of results is also not comprehensive enough, currently focusing more on execution than adaptive governance.

[0025] Therefore, to address the aforementioned main issues, a two-layer processing mechanism is designed for S100-S500, employing an outer loop for macroscopic decomposition and an inner loop for microscopic refinement. This mechanism automatically adjusts dynamically based on task complexity. The outer loop decomposes macroscopic tasks into subtasks, while the inner loop further decomposes highly complex subtasks into atomic tasks, which are then executed by dedicated fine-grained agents. This structure enables the method to handle both macroscopic strategic problems and fine-grained operational needs, significantly improving its adaptability and processing depth for multi-level, highly complex tasks.

[0026] Example 2, refer to Figures 1-2 This is one embodiment of the present invention. Based on the above embodiment, a task adaptive analysis and processing method based on a large model multi-agent is provided.

[0027] In this embodiment of the application, in step S100, the task is input, the task is received and verified through a standardized interface, and the outer loop processing flow is started after encapsulation. Specifically, the input channel supports submitting plain string task descriptions via HTTP POST. The standardized interface can be POST / api, etc., and the text can be encapsulated in JSON body.

[0028] In this embodiment of the application, the task input content in step S100 is natural language text, and the text length is set within a valid range.

[0029] In one optional implementation, during the task input phase, the user-submitted original task requirements are received through a preset standardized data interface. The user input content is natural language text, and its length must be strictly controlled within a set valid range, specifically, the number of characters L satisfies the inequality Lmin ≤ L ≤ Lmax. The minimum length threshold Lmin can be set to 5 characters, and the maximum length threshold Lmax can be set to 10,000 characters. Upon receiving the data, the interface module immediately initiates a triple verification mechanism: first, it checks the text length by calculating the number of characters and verifying whether it is within the valid range; then, it activates the outer loop processing flow, passing the task object to the outer loop controller to formally start the processing flow.

[0030] For example, S100 receives the original task requirements submitted by the user through a preset standardized data interface. This interface uses the HTTP communication protocol and JSON as the data exchange format to ensure compatibility with diverse clients. The user input content is natural language text, and the length is controlled within the set valid range. Upon receiving data, the interface module immediately initiates a triple verification mechanism: First, it checks the text length by calculating the number of characters and verifying that it is within a valid range; next, it performs encoding format verification to ensure that the text uses UTF-8 encoding to avoid parsing errors; finally, it performs content security detection by scanning for potential security risks or non-compliant content using a predefined regular expression rule base. Only inputs that pass all verifications are encapsulated into a structured task object, which contains four key attributes: a task identifier generated in a timestamp-server identifier-sequence number format (ensuring global uniqueness in a distributed environment), a text field storing the original content, a timestamp recording the accurate reception time, and a priority identifier dynamically divided according to real-time load (high / medium / low levels); after encapsulation, the outer loop processing flow is activated, and the task object is passed to the outer loop controller to formally start the processing flow. This rigorous design at the source ensures the standardization, security, and processability of the input data, laying a solid foundation for subsequent complex processing.

[0031] S200: In the outer loop processing stage, the task is broken down into sub-tasks, intelligent agents are matched to execute the task, the results are processed in parallel and integrated, and the quality is evaluated through a dual-track scoring mechanism to obtain the scoring result. In this embodiment of the application, step S200, during the outer loop processing stage, decomposes the task into sub-tasks, including the following steps A1-A4: A1: Task decomposition intelligent agent, built on a large language model, parses task objects and uses semantic analysis to perform intent recognition, key information extraction and task element analysis; For example, intent recognition identifies the user's core goal or purpose. It can guide the LLM to classify intent types based on the prompt. For example, if the goal is to analyze the AI ​​development trend in the past year and generate a market report based on this, including risk assessment, then the main intent in the recognition result is: to analyze trends and generate a report.

[0032] For example, key information extraction is the extraction of core entities and facts from text, which can be achieved by using named entity recognition or information extraction models, such as extracting time "the most recent year" and topic "AI development trends".

[0033] For example, task element analysis involves breaking down a task into its components, including input requirements, output expectations, constraints, and dependencies. For instance, input elements could be current AI market data and expert knowledge bases; output elements could be trend summaries, reports, and risk matrices; constraints could be data timeliness; and dependencies could be trend analysis and risk assessment.

[0034] A2: During the parsing process, the agent performs multiple rounds of understanding of the task text, performs grammatical parsing and semantic role labeling, and identifies the main actions, objects and modifying conditions in the task; For example, the main actions can be understood as the core verbs or operation sequences in the task, such as words like "analyze" and "generate". For example, the object can be understood as the target or object of the action, usually a noun phrase, such as: AI development trend, report, etc. For example, modifying conditions can be understood as adverbial clauses that limit the action or object, such as time, scope, quality requirements, etc.

[0035] A3: Using knowledge graph association technology, the concepts involved in the task are matched with existing knowledge bases; A4: Perform task boundary delineation, obtain the task scope and optional extended content, in order to complete the parsing.

[0036] In one optional implementation, A3-A4 use knowledge graph association technology to match the concepts mentioned in the task with existing knowledge bases, enriching the semantic information of the task; finally, task boundaries are defined to clarify the scope of the core task and optional extended content; after parsing, the agent logically decomposes the macro-task into multiple discrete subtasks. Each subtask is accompanied by a clear description of the deliverables. (Specified output content) and acceptance criteria (Define quality indicators), and clearly define the dependencies between subtasks to ensure a reasonable execution order; the decomposition process can adopt a top-down approach, first determining the main objectives of the task, then breaking it down into several key stages, and each stage is further refined into specific executable subtasks; for each subtask, its input requirements, processing logic, and output format are defined in detail to ensure that the interfaces between subtasks are clear and unambiguous; the dependencies between subtasks are managed through a directed acyclic graph to ensure the scientific and reasonable execution order.

[0037] In this embodiment of the application, step S200 involves intelligent matching and execution of intelligent agents, parallel processing and integration of results, including the following steps A5-A7: A5: The task distribution agent assigns a suitable execution agent to the sub-task based on a multi-factor decision model, and the assignment of the execution agent is performed by a matching degree calculation method; A6: During the matching degree calculation process, based on the agent's historical performance data and dynamic weight adjustment, the system is automatically adjusted according to the task type and urgency. and The value of this parameter is used to increase the proportion of load weight. In another alternative implementation, based on the above implementation, the task distribution agent assigns a suitable execution agent to the sub-task based on a multi-factor decision model. This decision relies on accurate matching degree calculation, wherein the matching degree formula can be expressed as: ； In the formula, The ability matching degree is represented by the semantic similarity between the subtask requirements and the agent's capabilities, calculated using cosine similarity. The real-time load rate is represented by the following formula: In the formula, This is the current queue length. To maximize processing capacity, and Configurable weighting coefficients and satisfying By selecting the agent with the highest matching degree for each subtask, both capability matching and load balancing can be optimized.

[0038] In another optional implementation, based on the above implementation, the historical performance data of the agent will be considered during the matching degree calculation process, including multiple dimensions such as task completion quality, average processing time, and professional skill matching degree, to ensure the comprehensiveness and accuracy of the allocation decision; and dynamic weight adjustment will be supported, automatically adjusting according to task type and urgency. and The value of will appropriately increase the proportion of the load weight. For example, for tasks with high timeliness requirements, the proportion of the load weight will be appropriately increased to ensure that the task can be processed quickly.

[0039] A7: Each agent receives a task and works in parallel. The execution process includes: parameter parsing, model invocation, and result generation. Specifically, in the parameter parsing stage, the intelligent system carefully analyzes the specific requirements of the subtasks, extracts key parameters, and processes constraints. During the model invocation phase, an appropriate processing model is selected based on the task type. Models include text generation models, data analysis models, image processing models, etc. In the results generation phase, standardized outputs are generated according to the requirements of the deliverable description.

[0040] In another optional implementation, after each agent in A7 completes its task, it sends the output results to the result collection center. In addition to the main output content, the results also include metadata information, such as processing time, model version used, and confidence score. After all sub-tasks are completed, the task-submitting agent synthesizes the scattered results according to the preset integration rules to generate a preliminary solution document. The integration process adopts a progressive approach. First, the integration order is determined according to the dependencies between sub-tasks. Then, the results are assembled according to the content logic. Next, the format is standardized. Finally, an overall logical consistency check is performed to ensure that the results of each sub-task can be organically combined.

[0041] In this embodiment of the application, the quality is evaluated through a dual-track scoring mechanism in step S200, including: Step A8: A8: The task correction agent initiates a dual-track scoring mechanism for quality assessment, including language and rule-based scoring, where: Language score The deep quality of semantic consistency and logical coherence is evaluated using a large language model. Rule scoring The system checks compliance requirements for format conformity and content integrity based on a predefined set of rules.

[0042] Specifically, the language score SI assessment process includes multiple dimensions, such as content relevance, information accuracy, logical rigor, and language fluency, with detailed scoring criteria for each dimension. Specifically, rule scoring This focuses on formal requirements, including document structural integrity, format standardization, terminology consistency, and accuracy of citations.

[0043] In one alternative implementation, language scoring Large Language Models (LLMs) can be used to evaluate deep quality. For semantic consistency, the similarity between the output and the embedding of the original task intent can be compared. For logical coherence, the argument chain can be checked, such as ensuring there are no jumps in causal relationships, and prompts can be used to guide LLM scoring; multi-turn prompt chains can also be supported to improve accuracy.

[0044] In another alternative implementation, rule scoring It can check surface quality based on rule sets. For formatting compliance, it verifies the structure, such as whether the format of headings, paragraphs, and lists matches the template. For content completeness, it verifies whether all subtask outputs are covered and whether there are any missing fields, and the rule engine can be customized to calculate scores.

[0045] S300: For subtasks whose complexity exceeds the first threshold, the inner loop processing stage is triggered to perform a second decomposition into atomic tasks. These tasks are then processed by a fine-grained agent, and the results are integrated before returning to the outer loop. This includes the following steps B1-B3: In this embodiment of the application, in step S300, B1: the complexity By comprehensively analyzing the number of processing steps of sub-tasks Dependency number Quantitative indicators were calculated. Specifically, the complexity calculation can be expressed as: ; in, and These are weighting coefficients, which are adjusted according to the different task types. As for the first threshold The settings need to be based on the actual task situation.

[0046] B2: After the inner loop is triggered, the task allocation agent takes over the corresponding complex subtasks and performs secondary decomposition to generate several atomic tasks. ; It should be noted that these atomic tasks are characterized by fine granularity, simple operation, and clearly defined input and output specifications. The decomposition process of atomic tasks follows the single responsibility principle, with each atomic task completing only one specific function. Tasks remain relatively independent to minimize coupling. B3: Atomic tasks are assigned to corresponding fine-grained agents for execution. Each fine-grained agent is optimized for a specific task type to ensure processing accuracy.

[0047] In one optional implementation, during the execution of the B3 fine-grained agent, operation logs and intermediate results are fully recorded to facilitate problem tracking and performance analysis. After all atomic tasks are completed, the inner loop task output module initiates a result integration process, including data verification (to ensure the accuracy of results), format unification (to maintain data consistency), and content synthesis (to organize scattered results into a complete output). Finally, the integrated sub-task results are returned to the outer loop result collection center to participate in the initial scheme integration steps of the outer loop, and then the inner loop instance is closed.

[0048] It should be noted that this inner loop mechanism can effectively solve the problem of excessive complexity within a single subtask, and improves the method's adaptability and processing depth for multi-level complex tasks through layered processing.

[0049] In this embodiment of the application, step S400 determines whether to iterate and optimize based on the scoring results, sets the maximum number of iterations, and continues until the quality meets the standard or is forcibly terminated, including the following steps C1-C4: C1: Set the second threshold T; C2: When either value in the dual-track score fails to reach the second threshold, the iteration path is intelligently selected based on the nature of the problem. C3: If the problem involves the global structure or core logic, the correction task will be sent back to the task decomposition step for reorganization; if it is only a local content or format problem, it will be directly distributed to the corresponding agent for quick revision. C4: After each outer loop execution, the dual-track scoring is re-performed. If any score fails to meet the standard, iterative optimization continues, with a maximum number of iterations set, until the quality meets the standard or the process is forcibly terminated.

[0050] For example, a second threshold T can be set to 80 points if and only if ≥80 and If the score is ≥80, the process proceeds to the next stage; if the score is not met, specific modification suggestions are generated, clearly pointing out the existing problems and directions for improvement. The correction task is then reinjected into the outer loop starting point or directly distributed to a specific agent, initiating a new round of optimization iterations. The generation process of modification suggestions is targeted, analyzing in detail the reasons for each point deduction and providing specific improvement suggestions, such as: which parts need to be supplemented, which expressions need to be adjusted, and which formats need to be modified, providing clear guidance for subsequent iterative optimization.

[0051] To prevent infinite loops, an iteration count counter is provided. When the number of iterations reaches a preset upper limit... The process is forcibly terminated at any time, and the current optimal version is output; the only condition for process termination is that both track scores are met simultaneously. ≥T and ≥T), mark the current version as the final solution document and enter the output preparation state; the entire iteration termination process reflects the design principle of prioritizing quality and balancing efficiency; during the iteration process, the scoring change trend of each iteration will be recorded, and the processing parameters will be dynamically adjusted by analyzing these data, such as adjusting the granularity of task decomposition, optimizing the weight coefficient of agent matching, etc., to achieve continuous optimization of processing effect.

[0052] S500: Outputs optimization solutions and quality reports, supports multiple delivery methods, releases resources, and records complete processing logs.

[0053] Specifically, in the final output stage, the final solution document, which has undergone multiple rounds of iteration and optimization, will be delivered to users in various ways, including direct display on the interface for instant viewing, file download supporting multiple formats such as PDF and DOCX, and email delivery. A detailed quality assessment report will be generated, which will not only include the specific values ​​of the final language score and rule score of the final solution and the score details of each dimension, but also summarize the key indicators of the processing process, providing users with a comprehensive and transparent basis for quality verification and processing traceability. The quality assessment report can be presented in a structured manner, including an execution summary, detailed scoring analysis, processing statistics, and improvement suggestions, to help users quickly understand the quality and efficiency of the solution. After the entire output process is completed, the task status is updated to completed, the occupied computing and storage resources are released, and a complete processing log is recorded, including the inputs and outputs of each stage, agent decision records, scoring history, etc., for subsequent auditing, analysis, and continuous optimization of the method.

[0054] In summary, this invention employs a two-tiered processing mechanism—an outer loop for macroscopic decomposition and an inner loop for microscopic refinement—that automatically adjusts dynamically based on task complexity. Specifically, the outer loop decomposes macroscopic tasks into subtasks, while the inner loop further decomposes highly complex subtasks into atomic tasks, which are then executed by dedicated fine-grained intelligent agents. This structure enables the entire method to handle both macroscopic strategic problems and fine-grained operational needs, enhancing its adaptability and processing depth for multi-level, highly complex tasks. Furthermore, by making decisions based on a matching degree model that integrates capability matching degree and real-time load rate, the task distribution agent achieves optimal matching between tasks and agents, which not only fully leverages the professional capabilities of each agent, but also achieves system load balancing and avoids single-point overload. Finally, by introducing a dual-track scoring mechanism and an intelligent iterative optimization process, the system can quantitatively evaluate the processing results of each round. Tasks that fail to meet the standards will be intelligently corrected by selecting the optimal iteration path based on the nature of the problem, until the quality meets the standards or the iteration limit is reached. This mechanism ensures the high quality and compliance of the final output solution, while simultaneously achieving closed-loop optimization of the processing process. Overall, the combination of parallel task processing and resource release mechanisms significantly improves the system's throughput, response speed, and resource utilization efficiency.

[0055] Example 3 illustrates a schematic scheme for a task adaptive analysis and processing method based on a large-model multi-agent system. It should be noted that the technical solution of this task adaptive analysis and processing system based on a large-model multi-agent system is based on the same concept as the aforementioned task adaptive analysis and processing method based on a large-model multi-agent system. Details not described in detail in this example can be found in the description of the aforementioned task adaptive analysis and processing method based on a large-model multi-agent system.

[0056] This embodiment also provides another task adaptive analysis and processing system based on a large model multi-agent system, including: The input module is used to input tasks, receive and verify the tasks through a standardized interface, encapsulate them, and start the outer loop processing flow. The outer loop processing module is used to break down tasks into subtasks, intelligently match and execute intelligent agents, process and integrate results in parallel, evaluate quality through a dual-track scoring mechanism, and obtain a score result. The inner loop processing module is used to trigger the inner loop processing stage for subtasks whose complexity exceeds the first threshold. The subtasks are then further decomposed into atomic tasks, which are processed by a fine-grained intelligent agent. The results are then integrated and returned to the outer loop. The iterative optimization module is used to determine whether to perform iterative optimization based on the scoring results, set the maximum number of iterations, and continue until the quality meets the standard or is forcibly terminated. The output module is used to output optimization solutions and quality reports, supports multiple delivery methods, releases resources, and records complete processing logs.

[0057] This embodiment also provides a computer device applicable to a task adaptive analysis and processing method based on a large model multi-agent, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the task adaptive analysis and processing method based on a large model multi-agent proposed in the above embodiment.

[0058] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, it implements a task adaptive analysis and processing method based on a large model multi-agent as proposed in the above embodiments.

[0059] The storage medium proposed in this embodiment belongs to the same inventive concept as the method for implementing a task adaptive analysis and processing method based on a large model and multiple agents proposed in the above embodiments. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0060] From the above description of the implementation methods, those skilled in the art will clearly understand that the present invention can be implemented using software and necessary general-purpose hardware. 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. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0061] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A task adaptive analysis and processing method based on a large model and multiple agents, characterized in that, include: The task is input, received and verified through a standardized interface, and then encapsulated to start the outer loop processing flow. In the outer loop processing stage, the task is broken down into sub-tasks, intelligent agents are matched to execute the tasks, the results are processed in parallel and integrated, and the quality is evaluated through a dual-track scoring mechanism to obtain a score result. For subtasks whose complexity exceeds the first threshold, the inner loop processing stage is triggered to perform a second decomposition into atomic tasks, which are then processed by a fine-grained agent and the results are integrated before being returned to the outer loop. The decision to iterate and optimize is based on the scoring results, with a maximum number of iterations set until the quality meets the standard or the process is forcibly terminated. Output optimization solutions and quality reports, support multiple delivery methods, release resources and record complete processing logs.

2. The task adaptive analysis and processing method based on a large model and multiple agents as described in claim 1, characterized in that, During the outer loop processing phase, the task is broken down into subtasks, including: The task decomposition intelligent agent is built based on a large language model to parse task objects and use semantic analysis to perform intent recognition, key information extraction and task element analysis. During the parsing process, the agent performs multiple rounds of understanding of the task text, performs grammatical parsing and semantic role labeling, and identifies the main actions, objects and modifying conditions in the task; By using knowledge graph association technology, the concepts involved in the task are matched with existing knowledge bases; Perform task boundary delineation, obtain the task scope and optional extended content, in order to complete the parsing.

3. The task adaptive analysis and processing method based on a large model and multiple agents as described in claim 2, characterized in that, Intelligent matching executes intelligent agents, processes and integrates results in parallel, including: The task distribution agent assigns a suitable execution agent to the sub-task based on a multi-factor decision model, and the assignment of the execution agent is performed by a matching degree calculation method. During the matching degree calculation process, based on the agent's historical performance data and dynamic weight adjustment, the values ​​of WP and WI are automatically adjusted according to the task type and urgency to increase the proportion of load weight. Each agent receives a task and works in parallel. The execution process includes parameter parsing, model invocation, and result generation.

4. The task adaptive analysis and processing method based on a large model and multiple agents as described in claim 3, characterized in that, Quality is assessed through a dual-track scoring mechanism, including: the task-correcting agent initiates the dual-track scoring mechanism for quality assessment, comprising language and rule-based scoring, wherein: Language scoring assesses deeper quality such as semantic consistency and logical coherence through a large language model; The rule scoring is based on a predefined set of rules to check for compliance requirements regarding format conformity and content completeness.

5. The task adaptive analysis and processing method based on a large model and multiple agents as described in claim 4, characterized in that, The complexity is calculated by comprehensively analyzing the number of processing steps of the subtasks and the quantitative indicators of dependencies; After the inner loop is triggered, the task allocation agent takes over the corresponding complex subtask and performs a secondary decomposition to generate several atomic tasks. Atomic tasks are assigned to corresponding fine-grained agents for execution, with each agent optimized for a specific task type.

6. The task adaptive analysis and processing method based on a large model and multiple agents as described in claim 5, characterized in that, The decision to iterate and optimize is based on the scoring results, setting a maximum number of iterations until the quality target is met or forced termination occurs, including: Set a second threshold; When either value in the dual-track scoring fails to reach the second threshold, the iteration path is intelligently selected based on the nature of the problem. If the problem involves the global structure or core logic, the correction task is sent back to the task decomposition step for reorganization; if it is only a local content or format problem, it is directly distributed to the corresponding agent for quick revision. After each outer loop execution, the dual-track scoring is performed again. If any score fails to meet the standard, iterative optimization continues, with a maximum number of iterations set, until the quality meets the standard or is forcibly terminated.

7. The task adaptive analysis and processing method based on a large model and multiple agents as described in claim 6, characterized in that, The task input is natural language text, and the text length is set within a valid range.

8. A task adaptive analysis and processing system based on a large model multi-agent system, using the method described in any one of claims 1-7, characterized in that, include: The input module is used to input tasks, receive and verify the tasks through a standardized interface, encapsulate them, and start the outer loop processing flow. The outer loop processing module is used to break down tasks into subtasks, intelligently match and execute intelligent agents, process and integrate results in parallel, evaluate quality through a dual-track scoring mechanism, and obtain a score result. The inner loop processing module is used to trigger the inner loop processing stage for subtasks whose complexity exceeds the first threshold. The subtasks are then further decomposed into atomic tasks, which are processed by a fine-grained intelligent agent. The results are then integrated and returned to the outer loop. The iterative optimization module is used to determine whether to perform iterative optimization based on the scoring results, set the maximum number of iterations, and continue until the quality meets the standard or is forcibly terminated. The output module is used to output optimization solutions and quality reports, supports multiple delivery methods, releases resources, and records complete processing logs.

9. A computer device, characterized in that, include: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the task adaptive analysis and processing method based on a large model multi-agent as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores computer-executable instructions, which, when executed by a processor, implement the steps of the task adaptive analysis and processing method based on a large model multi-agent as described in any one of claims 1 to 7.

Images