Method based on large model multi-round disassembly task

Abstract

This invention relates to the field of large-model task understanding and intelligent scheduling technology, and in particular to a method for multi-round task decomposition based on a large model, comprising the following steps: S1, receiving a multimodal complex task description, fusing task semantic features with heterogeneous system business rule features, and generating regularized task information; S2, inputting the regularized task information into the WxTaskLM task decomposition large model to obtain an initial set of subtasks; S3, determining the executability of the initial subtasks and generating an executable subtask sequence; S4, analyzing subtask data dependencies, real-time load of heterogeneous systems, and task priorities, constructing a dependency directed graph and performing topological sorting; S5, advancing subtask execution according to the execution plan, and re-executing after root cause diagnosis and correction in case of anomalies; S6, optimizing the WxTaskLM large model parameters and subtask planning logic to form a closed-loop optimization. This invention effectively solves the technical problems of unreasonable subtask priority allocation and lack of quantitative standards in the prior art.

Description

Technical Field

[0001] This invention relates to the field of large model task understanding and intelligent scheduling technology, and in particular to a method for decomposing large models into multiple rounds of tasks. Background Technology

[0002] Traditionally, handling complex business tasks typically involves a combination of manual experience-based breakdown, RPA fixed script execution, and manual intervention. Tasks are first broken down into multi-step sub-tasks by business personnel based on Standard Operating Procedures (SOPs), and then technical personnel configure RPA processes or API call scripts to complete the execution. Abnormal processes still require manual intervention. Industry-leading solutions such as SAP Task Management, Yonyou Workflow Platform, and traditional RPA tools are all based on this paradigm and have become a common approach for implementing digital office solutions in enterprises.

[0003] However, existing technologies generally suffer from the following problems when dealing with multi-round decomposition tasks in industrial scenarios: On the one hand, there is a dual static disconnect between task decomposition and execution scheduling: the decomposition rules rely on manual presets and fixed process configurations, and cannot jointly reason about business semantics, real-time system load, and sub-task dependencies. The granularity of sub-tasks and execution sequence are both statically set. When encountering process changes, interface adjustments, or high-concurrency system scenarios, problems such as incomplete decomposition, circular dependencies, and execution conflicts are very likely to occur. Moreover, there is a lack of quantifiable decomposition quality verification mechanisms, which makes the stability of the process heavily dependent on manual maintenance costs.

[0004] On the other hand, there is a semantic-execution gap in the integration of large models with business systems: general large models can only complete task decomposition at the text level and cannot deeply embed business rules, data dictionaries and interface constraints of heterogeneous systems, resulting in mismatch between output sub-tasks and actual system operations; at the same time, there is a lack of closed-loop optimization mechanism, and execution anomalies and decomposition deviations cannot be reverse-iterated to iterate model parameters, resulting in the model being in a general weak adaptation state in enterprise scenarios for a long time, making it difficult to support the stable operation of highly compliant and strongly constrained businesses.

[0005] Therefore, it is necessary to design a method for multi-round decomposition tasks based on large models that can realize dynamic multi-round decomposition and quantitative quality verification. Summary of the Invention

[0006] To solve one of the above-mentioned technical problems, the present invention adopts the following technical solution: a method for multi-round task decomposition based on a large model, comprising the following steps: S1, receiving a multimodal complex task description, converting the voice task description into text, completing text redundancy cleaning, business terminology unification and feature extraction, fusing task semantic features and heterogeneous system business rule features, and generating regular task information.

[0007] S2. Input the standardized task information into the WxTaskLM task decomposition model. Combine the business rule base and the historical task decomposition case base to decompose the complex task into core target sub-tasks and auxiliary support sub-tasks in the first round, and obtain the initial set of sub-tasks.

[0008] S3. Perform an executability check on the initial subtasks, iteratively decompose non-executable subtasks, and perform a two-dimensional check on the completeness and granularity adaptability of the decomposition after each round of decomposition. If the check fails, adjust the strategy and decompose again. If the check passes, proceed to the next round until an executable subtask sequence is generated.

[0009] S4. Analyze the data dependencies of subtasks, the real-time load of heterogeneous systems and task priorities, construct a directed dependency graph and perform topological sorting, and generate a subtask execution plan that includes execution order, execution system, data source and execution time limit.

[0010] S5. Proceed with the execution plan to advance the execution of sub-tasks, monitor the execution status and data integrity in real time, and re-execute after completing root cause diagnosis and correction in case of anomalies.

[0011] S6. Based on the full-process breakdown and execution data, optimize the parameters of the WxTaskLM large model and the sub-task planning logic, update the business rule library and historical case library, and form a closed-loop optimization.

[0012] As a preferred approach, the preprocessing of the multimodal task in S1 follows these steps:

[0013] Convert the speech task description into text and remove redundant and meaningless expressions;

[0014] Then, we standardized the format of business terminology and extracted four core features: task type, associated system, time node, and data requirement.

[0015] Integrate core features with operational guidelines and data dictionaries from the business rule base;

[0016] The terminology normalization matching factor is introduced to calibrate the feature output; the preprocessing time is no more than 8 seconds, and the semantic feature vector dimension is fixed at 1024.

[0017] As a preferred approach, the WxTaskLM large model is used for semantic understanding, multi-round decomposition, deviation verification, and anomaly root cause diagnosis of cross-system business tasks. The model construction and inference follow these steps:

[0018] Build an improved Transformer architecture that includes a task semantic understanding layer, a business rule fusion layer, and a multi-round inference layer;

[0019] Deep semantics are extracted and interference is filtered out through bidirectional encoding at the semantic understanding layer;

[0020] By fusing rules and case knowledge through a dual-branch independent attention mechanism, a fusion bias suppression factor is introduced to avoid feature confusion.

[0021] The hierarchical decomposition and verification are completed through multiple rounds of inference layers;

[0022] The model parameters are 1024 hidden layer dimensions, 16 layers, and 16 attention heads, with the core sub-tasks accounting for no less than 60% in the first round.

[0023] As a preferred approach, the executability determination of subtasks in S3 follows these steps:

[0024] Verify whether the subtask meets the three conditions of single system operation, clear input and output, and no unresolved dependencies;

[0025] Introduce a subtask dependency cycle detection depth factor to traverse and detect circular dependencies across subtasks;

[0026] Mark subtasks that do not meet the conditions or have circular dependencies as non-executable;

[0027] The number of disassembly rounds should not exceed 5; if it does, the model should be sent back for re-disassembly.

[0028] As a preferred approach, the subtask planning for S4 follows these steps:

[0029] Extract the input and output data items of the subtasks, and construct a dependency directed graph with the subtasks as nodes and the data dependencies as directed edges;

[0030] Perform cycle detection and automatic cycle breaking on dependent directed graphs to eliminate timing conflicts;

[0031] Introduce a system load fluctuation correction factor to calibrate the system load determination results;

[0032] The acyclic graph is topologically sorted and prioritized, with execution timing deviations controlled within ±50ms.

[0033] As a preferred solution, the subtask execution control of S5 follows these steps:

[0034] Identify subtask operation types and distinguish between web-based operations and system data interaction operations;

[0035] Execution logic is adapted to the operation type to advance the task;

[0036] A redundancy factor for execution link data verification is introduced to complete multi-layer data integrity verification;

[0037] The system includes four core components: real-time monitoring of execution status, retention of abnormal information, and log retention for at least 90 days.

[0038] As a preferred approach, the abnormal root cause diagnosis of S5 follows these steps:

[0039] Capture full information including the identifier of the abnormal subtask, the execution node, and the exception type;

[0040] An anomaly root cause classification confidence factor is introduced, and the WxTaskLM large model is used to distinguish between decomposition deviation and planning defect anomalies.

[0041] Subtasks are supplemented or their timing is adjusted for different anomaly types; only the anomaly-related links are re-executed, and the anomaly repair response time does not exceed 30 seconds.

[0042] As a preferred approach, S3's two-dimensional verification follows these steps:

[0043] Extract the current set of subtasks and the core requirements data of the original tasks in each round;

[0044] Calculate the completeness of requirement coverage and the adaptability of granularity separately; then remove abnormal sub-task interference data and smooth the indicators; finally determine the termination conditions for decomposition.

[0045] The comprehensive determination formula for dismantling termination is as follows:

[0046] ;

[0047] in, The score is terminated by dismantling; The coefficients are for different scenarios: 0.65 for high compliance and 0.55 for rapid response. To determine the coverage of requirements, core requirements are assigned a value of 1, auxiliary requirements are assigned a value of 0.5, and irrelevant requirements are assigned a value of 0. The core demand quantity; This is the business compliance coefficient, ranging from 0.92 to 1.08. This is the granularity adaptation value; 1 for a single system, 0.3 for cross-system, and 0 for no entry point. The total number of subtasks is ≥0.85, and the disassembly is terminated.

[0048] As a preferred approach, incremental fine-tuning of the WxTaskLM large model follows these steps:

[0049] Filter historical successful and anomalous amendment examples, and divide the training set and validation set in an 8:2 ratio;

[0050] Incremental fine-tuning of the gradient clipping coefficient is introduced to control the parameter update magnitude;

[0051] Only the parameters of the last three attention mechanisms are fine-tuned, without changing the basic architecture;

[0052] Set batch size to 32 and learning rate to 10. -5 Fine-tuning was completed in accordance with industrial deployment standards.

[0053] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0054] 1. This invention effectively solves the technical problems of unreasonable subtask priority allocation and lack of quantitative standards in the prior art. By constructing a standardized dynamic priority quantification formula, the weight allocation of core parameters such as business importance, execution urgency, and blocking coefficient is clarified. Combined with the adaptive calibration of load correction coefficient, latency risk, and system interaction latency factor, the priority of subtasks is accurately quantified, ensuring that core urgent tasks are executed first and non-urgent tasks are reasonably staggered. For example, in the financial reimbursement scenario in Example 1, the core compliance verification task is scheduled first through this priority allocation method, which effectively avoids the task blocking problem under the high load of the financial system at the end of the month and improves the timeliness and rationality of task execution.

[0055] 2. This invention integrates three core indicators—accuracy rate, execution success rate, and anomaly repair rate—through a scientifically designed comprehensive evaluation formula for model optimization. It introduces historical case matching gain factors to achieve a comprehensive and accurate evaluation of the model optimization effect. At the same time, it sets clear optimization qualification thresholds to ensure the targetedness and effectiveness of incremental fine-tuning of the model and avoid model performance fluctuations caused by blind optimization. For example, in Examples 1 and 2, the optimization scores calculated by this evaluation formula are all higher than the qualification standard of 0.9, which enables precise iteration of model weights, allowing the model to continuously adapt to different business scenarios and improve the stability and adaptability of long-term operation.

[0056] 3. This invention takes into account the differentiated needs of both high-compliance and conventional business scenarios. By flexibly adjusting parameters such as scenario coefficients and business compliance coefficients, it achieves scenario-based adaptation of subtask decomposition, priority allocation, and model optimization. It not only meets the traceability and strong constraint requirements of high-compliance scenarios such as financial reimbursement, but also adapts to the efficient execution needs of conventional scenarios such as cross-platform project approval. At the same time, through multimodal task preprocessing and dual-dimensional verification of subtask executability, it ensures the integrity and executability of task decomposition, significantly reduces the cost of manual intervention, and improves the automation level and compliance of task execution in heterogeneous enterprise systems.

[0057] 4. This invention achieves traceability and reproducibility of the entire task processing process through standardized end-to-end design and parameter specifications. It can adapt to the business needs of different enterprises through flexible parameter adjustment. At the same time, through system load fluctuation correction, rapid anomaly diagnosis and repair mechanism, it effectively reduces the risk of anomalies in the task execution process, shortens the anomaly repair time, and improves the operational efficiency and stability of enterprise business processes. It has strong practicality and promotion value. Attached Figure Description

[0058] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or components are generally identified by similar reference numerals. In the drawings, the elements or components are not necessarily drawn to scale.

[0059] Figure 1 This is a schematic diagram of the process of the present invention.

[0060] Figure 2 This is a graph showing the effect of the blocking coefficient on dynamic priority in the embodiment.

[0061] Figure 3 The graph shows the impact of delay risk on dynamic priority in the example.

[0062] Figure 4 The graph shows the impact of disassembly accuracy on the model optimization score in the example.

[0063] Figure 5 The graph shows the impact of the historical case matching gain factor on the model optimization score in the example. Detailed Implementation

[0064] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and are therefore merely examples and should not be used to limit the scope of protection of the present invention. The specific process and related curves of the present invention are as follows: Figures 1-5 As shown in the image.

[0065] A method for multi-round task decomposition based on a large model includes the following steps: S1, receiving a multimodal complex task description, converting the voice task description into text, completing text redundancy cleaning, business terminology unification and feature extraction, fusing task semantic features and heterogeneous system business rule features, and generating regular task information.

[0066] S2. Input the standardized task information into the WxTaskLM task decomposition model. Combine the business rule base and the historical task decomposition case base to decompose the complex task into core target sub-tasks and auxiliary support sub-tasks in the first round, and obtain the initial set of sub-tasks.

[0067] S3. Perform an executability check on the initial subtasks, iteratively decompose non-executable subtasks, and perform a two-dimensional check on the completeness and granularity adaptability of the decomposition after each round of decomposition. If the check fails, adjust the strategy and decompose again. If the check passes, proceed to the next round until an executable subtask sequence is generated.

[0068] S4. Analyze the data dependencies of subtasks, the real-time load of heterogeneous systems and task priorities, construct a directed dependency graph and perform topological sorting, and generate a subtask execution plan that includes execution order, execution system, data source and execution time limit.

[0069] S5. Proceed with the execution plan to advance the execution of sub-tasks, monitor the execution status and data integrity in real time, and re-execute after completing root cause diagnosis and correction in case of anomalies.

[0070] S6. Based on the full-process breakdown and execution data, optimize the parameters of the WxTaskLM large model and the sub-task planning logic, update the business rule library and historical case library, and form a closed-loop optimization.

[0071] In step S1 of this technical solution, the description of the speech task is first converted into text using a well-known end-to-end speech recognition algorithm. This algorithm adopts the industry-standard Mel-spectrum feature extraction and CTC decoding logic. The feature extraction stage focuses on four core dimensions: task type, associated system, time node, and data requirements. These four dimensions are the core components of cross-system business tasks. The fusion of semantic features and business rule features adopts a well-known attention-weighted fusion algorithm. This algorithm has become a routine technical means in the fields of natural language processing and business rule matching. The fused and regularized task information not only retains the semantic connotation of the task, but also conforms to the operation specifications of heterogeneous systems, laying the foundation for subsequent decomposition.

[0072] The WxTaskLM task decomposition model used in step S2 is a dedicated model built on an improved Transformer architecture. Its hierarchical structure is completely clear, consisting of a task semantic understanding layer, a business rule fusion layer, and a multi-round inference layer. The task semantic understanding layer takes regularized task information as input, extracts core requirements and constraints through bidirectional encoding, and outputs a structured semantic vector. The business rule fusion layer takes semantic vectors and data from the business rule base and historical case base as input, completes knowledge embedding through a dual-branch independent attention mechanism, and outputs a feature vector that integrates business knowledge. The multi-round inference layer takes the fused feature vector as input, performs the first round of hierarchical decomposition, and outputs an initial set consisting of core target sub-tasks and auxiliary support sub-tasks. The inputs and outputs of each level are clear, and the connections are clear. Those skilled in the art can directly build the model architecture based on the above description.

[0073] The feasibility assessment of step S3 is based on the universally recognized subtask executability standards in the field, namely, single system operation, clear input and output, no unresolved dependencies, and iterative decomposition and two-dimensional verification to form a closed loop, avoiding over- or under-decomposition. Step S4 uses well-known algorithms in graph theory for dependency directed graph construction and topology sorting, and system load collection follows the general specifications for heterogeneous system interface monitoring. Priority allocation is combined with the conventional logic of enterprise business task management. Step S5's execution advancement and anomaly diagnosis are aligned with the actual process of cross-system business execution, and anomaly classification and correction logic are based on industry experience in handling business execution anomalies. Step S6's model optimization uses the well-known incremental fine-tuning training method, and the rule base and case base updates follow the general specifications of enterprise knowledge base management.

[0074] Furthermore, this solution enables the rule-free, dependency-free autonomous decomposition of complex cross-system tasks, breaking free from the strong binding of traditional solutions to preset rule bases. It can adaptively adapt to changes in business processes and the addition of new task types without the need for frequent manual updates to scripts or rules. Through multi-dimensional intelligent planning of subtasks, it achieves reasonable allocation of resources from heterogeneous systems, avoiding system congestion and subtask execution conflicts during high-load periods, and improving the smoothness of cross-system business execution. Through a closed-loop feedback optimization mechanism, the accuracy of task decomposition and execution efficiency continuously improve with the accumulation of business data, forming a technical system with self-iterative capabilities. Through fully automated processing, it significantly reduces the necessity of manual intervention, minimizes business deviations caused by human error, and ensures the accuracy and stability of cross-system business execution.

[0075] As a preferred approach, the preprocessing of the multimodal task in S1 follows these steps:

[0076] Convert the speech task description into text and remove redundant and meaningless expressions;

[0077] Then, we standardized the format of business terminology and extracted four core features: task type, associated system, time node, and data requirement.

[0078] Integrate core features with operational guidelines and data dictionaries from the business rule base;

[0079] The terminology normalization matching factor is introduced to calibrate the feature output; the preprocessing time is no more than 8 seconds, and the semantic feature vector dimension is fixed at 1024.

[0080] This step further refines the S1 preprocessing stage, removing irrelevant interjections, repetitive phrases, and special symbols to prevent invalid information from affecting the accuracy of subsequent feature extraction. It standardizes business terminology by referencing industry-standard business system terminology dictionaries; for example, terms like expense approval in finance and inventory verification in supply chain use industry-standard expressions to eliminate feature bias caused by differences in user expressions. Simultaneously, it extracts four core features: task type, associated systems, time nodes, and data requirements. These four features are recognized by those skilled in the art as core elements of cross-system business tasks, fully covering the task's execution requirements and constraints. Finally, it matches and integrates these core features with the operational specifications and data dictionary of the business rule base, which stores the operational flows of various heterogeneous systems. Information such as process, interface specifications, and data formats are included. The data dictionary defines the meaning, type, and value range of the system data fields. The matching and fusion adopts a well-known feature similarity matching algorithm to ensure that the extracted features are consistent with the actual rules of the system. A term normalization matching degree factor is introduced to calibrate the feature output, which is used to correct minor deviations in the term matching process and ensure that the output semantic features are accurately aligned with the business scenario. The preprocessing time of no more than 8 seconds is set in accordance with the response requirements of real-time business processing of enterprises. The semantic feature vector dimension is fixed at 1024 dimensions, which is in line with the general standard of semantic vector representation in the field of natural language processing. This dimension can fully preserve the semantic information of the task without increasing the computational burden of the model due to excessive dimensionality. Those skilled in the art can directly realize standardized preprocessing of multimodal tasks by following the above steps and parameters.

[0081] Furthermore, speech conversion and redundancy removal eliminate information interference for subsequent terminology unification. Terminology unification and feature extraction construct the core feature framework of the task. Feature matching and fusion ensure that features are deeply aligned with the actual business rules of the system. The terminology normalization matching factor further improves the accuracy of feature output. Preprocessing time and vector dimension constraints balance processing efficiency and feature quality. Each step serves the core objective of providing accurate and standardized input for model decomposition, avoiding decomposition bias caused by non-standard information, while simultaneously achieving unified and standardized multimodal task information and taking into account the real-time requirements of business processing.

[0082] As a preferred approach, the WxTaskLM large model is used for semantic understanding, multi-round decomposition, deviation verification, and anomaly root cause diagnosis of cross-system business tasks. The model construction and inference follow these steps:

[0083] Build an improved Transformer architecture that includes a task semantic understanding layer, a business rule fusion layer, and a multi-round inference layer;

[0084] Deep semantics are extracted and interference is filtered out through bidirectional encoding at the semantic understanding layer;

[0085] By fusing rules and case knowledge through a dual-branch independent attention mechanism, a fusion bias suppression factor is introduced to avoid feature confusion.

[0086] The hierarchical decomposition and verification are completed through multiple rounds of inference layers;

[0087] The model parameters are 1024 hidden layer dimensions, 16 layers, and 16 attention heads, with the core sub-tasks accounting for no less than 60% in the first round.

[0088] This step clarifies the application scenarios and construction inference logic of the WxTaskLM large model. The model hierarchy, inputs and outputs, connection relationships, and parameter settings are all clearly defined, allowing those skilled in the art to directly build and implement it. The construction of the WxTaskLM large model strictly follows the large model development process:

[0089] First, an improved Transformer architecture was built. Based on the classic Transformer, this architecture was optimized to meet the needs of business task decomposition into a three-layer structure: a task semantic understanding layer, a business rule fusion layer, and a multi-round inference layer. Each layer has a clear division of labor and seamless integration. The task semantic understanding layer takes preprocessed, standardized task information as input and uses the well-known BERT bidirectional encoding logic to extract deep semantics, filtering out ambiguous and unclear expressions and other interfering information, outputting a 1024-dimensional structured semantic vector. The business rule fusion layer takes the semantic vector, business rule base data, and historical case base data as input and adopts a dual-branch independent attention mechanism. The two branches process business rule knowledge and historical case knowledge respectively, avoiding mutual interference between the two types of knowledge features. Simultaneously, a fusion bias suppression factor is introduced. The factor, based on the general specifications of attention mechanism feature fusion, is used to correct weight bias in the feature fusion process and outputs a fused 768-dimensional rule case embedding vector. The multi-round inference layer takes the fused embedding vector as input and uses iterative inference logic to complete the first round of hierarchical decomposition, dividing the task into core target sub-tasks and auxiliary support sub-tasks. The proportion of core sub-tasks in the first round is not less than 60%, which is set according to the composition rules of cross-system business tasks. Core sub-tasks are the key links to achieve the task objectives, and this proportion can ensure the integrity of the task framework after decomposition. The model parameters, hidden layer dimension 1024, number of layers 16, and number of attention heads 16, are all set according to the general standards for large-scale industrial deployment of models, taking into account the semantic understanding accuracy and computational efficiency of the model. The inference process and functions of each level of the model are fully disclosed.

[0090] In addition, this step deeply integrates the general Transformer architecture with the cross-system business task decomposition requirements. Through layered architecture design and dual-branch attention mechanism, the model can simultaneously possess accurate semantic understanding capabilities and deep business rule adaptation capabilities. Its uniqueness lies in breaking the problem of general large models being disconnected from business scenarios and unable to deeply integrate with internal enterprise business rules, and realizing integrated reasoning of semantic understanding, rule fusion, and iterative decomposition.

[0091] The dual-branch independent attention mechanism avoids feature confusion between business rules and historical cases. The fusion bias suppression factor further optimizes the stability and balance of feature fusion. The multi-round inference layer realizes accurate hierarchical decomposition and bias verification of tasks. The standardized model parameters balance inference efficiency and accuracy. The limitation of the proportion of core sub-tasks in the first round ensures that the decomposed task framework fits the actual business logic.

[0092] As a preferred approach, the executability determination of subtasks in S3 follows these steps:

[0093] Verify whether the subtask meets the three conditions of single system operation, clear input and output, and no unresolved dependencies;

[0094] Introduce a subtask dependency cycle detection depth factor to traverse and detect circular dependencies across subtasks;

[0095] Mark subtasks that do not meet the conditions or have circular dependencies as non-executable;

[0096] The number of disassembly rounds should not exceed 5; if it does, the model should be sent back for re-disassembly.

[0097] This step, based on the WxTaskLM large model architecture, further refines the executability determination process by introducing a subtask dependency cycle detection depth factor. This factor, based on a general algorithm for cycle detection in graph theory, controls the detection depth of circular dependencies. A well-known depth-first search algorithm is used to traverse and detect cross-subtask circular dependencies, quickly identifying circular dependencies between subtasks. Subtasks that do not meet the three conditions or have circular dependencies are marked as unexecutable, with the marking information simultaneously including the specific reasons for not meeting the conditions, providing a clear optimization direction for subsequent iterative decomposition. The decomposition rounds are then controlled to not exceed 5 rounds. This number is set based on the actual complexity of enterprise business task decomposition; typical cross-system tasks can reach an executable state within 5 rounds. If more rounds are exceeded, the data is sent back to the WxTaskLM model for re-decomposition, and the model optimizes the decomposition strategy based on the deviation information from previous decompositions.

[0098] Furthermore, this step quickly identifies unexecutable subtasks, avoiding invalid decomposition and subsequent execution conflicts. Its uniqueness lies in integrating basic condition verification, circular dependency detection, unexecutable marking, and round control into a closed-loop decision logic, rather than the traditional single-condition judgment. The three basic condition verifications complete the initial screening of subtasks from the perspective of execution feasibility; dependency loop detection avoids the risk of temporal deadlock from the perspective of task association; the unexecutable marking clarifies the optimization direction for subsequent iterative decomposition; and the decomposition round control balances decomposition accuracy and processing efficiency, avoiding ineffective iterative computation and wasted resources. This step is linked with the model decomposition logic, allowing the judgment results to directly guide the model re-decomposition process. The overall design logic aligns with the actual needs of cross-system business task decomposition.

[0099] As a preferred approach, the subtask planning for S4 follows these steps:

[0100] Extract the input and output data items of the subtasks, and construct a dependency directed graph with the subtasks as nodes and the data dependencies as directed edges;

[0101] Perform cycle detection and automatic cycle breaking on dependent directed graphs to eliminate timing conflicts;

[0102] Introduce a system load fluctuation correction factor to calibrate the system load determination results;

[0103] The acyclic graph is topologically sorted and prioritized, with execution timing deviations controlled within ±50ms.

[0104] Based on the feasibility assessment, this step further refines the S4 subtask planning process. A dependent directed graph is constructed using subtasks as nodes and data dependencies as directed edges, employing a well-known graph theory method. Data dependency refers to the output data of one subtask being the input data of the next; this dependency relationship is a natural logical connection between cross-system business tasks. Next, cycle detection and automatic cycle breaking are performed on the dependent directed graph, splitting cyclically dependent subtasks into acyclic subtasks to eliminate execution timing conflicts. Then, a system load fluctuation correction factor is introduced. This factor, based on industry standards for enterprise information system load monitoring, is used to calibrate the real-time load collection results and eliminate judgment errors caused by instantaneous peak fluctuations. Subsequently, topological sorting is performed on the acyclic graph using the well-known Kahn algorithm, a common method for task timing sorting. Priority allocation combines industry-standard rules such as business importance and urgency. Those skilled in the art can directly achieve intelligent subtask planning by following the above steps and algorithms.

[0105] In addition, this step is based on the data dependencies of subtasks and the real-time running status of the system to realize the intelligent allocation of subtask execution order and system resources. Its uniqueness lies in integrating dependency graph construction, loop detection and destruction, load calibration, topology sorting and priority allocation into an integrated planning process, rather than the traditional fixed time sequence sorting.

[0106] The system relies on directed graphs to fully restore the natural data relationships and execution logic between subtasks. Loop detection and automatic loop breaking eliminate execution timing conflicts at the source. The load correction factor ensures the accuracy of system load determination. Topology sorting determines the conflict-free basic execution sequence. Priority allocation fits the actual business needs of enterprises. Timing deviation control ensures the timing consistency of cross-system operations.

[0107] As a preferred solution, the subtask execution control of S5 follows these steps:

[0108] Identify subtask operation types and distinguish between web-based operations and system data interaction operations;

[0109] Execution logic is adapted to the operation type to advance the task;

[0110] A redundancy factor for execution link data verification is introduced to complete multi-layer data integrity verification;

[0111] The system includes four core components: real-time monitoring of execution status, retention of abnormal information, and log retention for at least 90 days.

[0112] This step, based on the subtask planning results, further refines the S5 subtask execution control process. The time-series flow aligns with the actual scenario of cross-system business execution, introducing a data verification redundancy factor for the execution link. This factor, based on industry standards for enterprise data security and integrity verification, is used to set the level and redundancy of data verification, completing multi-layer data integrity verification. The verification content includes data format, field integrity, and data legality, preventing data errors from causing execution failure. Subsequently, the execution status is monitored in real time, using a common method of system interface status monitoring and log embedding. Anomaly information is retained in four core categories: subtask identifier, execution steps, anomaly type, and system environment. These four categories are recognized by those skilled in the art as core elements for anomaly investigation. Logs are retained for no less than 90 days, in accordance with national standards for enterprise business log management and relevant requirements of the Cybersecurity Law, meeting the needs of business traceability and anomaly investigation.

[0113] This step adapts to different execution logics for different types of sub-task operations, and ensures execution stability through multiple data verifications and full-process status monitoring. Its uniqueness lies in integrating operation type identification, execution logic adaptation, multi-level data verification, real-time status monitoring, and compliance log retention into a full-process execution control system, rather than a single task execution.

[0114] Operation type identification provides a precise basis for execution logic adaptation, differentiated execution logic improves the compatibility of the solution with new and old heterogeneous systems, data verification redundancy factors ensure the data integrity and accuracy of the execution link, real-time status monitoring realizes full-process controllability of the execution process, and standardized log retention meets business compliance and traceability requirements.

[0115] As a preferred approach, the abnormal root cause diagnosis of S5 follows these steps:

[0116] Capture full information including the identifier of the abnormal subtask, the execution node, and the exception type;

[0117] An anomaly root cause classification confidence factor is introduced, and the WxTaskLM large model is used to distinguish between decomposition deviation and planning defect anomalies.

[0118] Subtasks are supplemented or their timing is adjusted for different anomaly types; only the anomaly-related links are re-executed, and the anomaly repair response time does not exceed 30 seconds.

[0119] An anomaly root cause classification confidence factor is introduced. This factor is based on industry standards for large-scale text classification and is used to constrain the accuracy of anomaly classification and improve the reliability of classification results. The WxTaskLM large-scale model is used to distinguish anomaly types, clearly classifying anomalies into decomposition deviation anomalies and planning defect anomalies. Decomposition deviation anomalies refer to anomalies caused by missing subtasks or inappropriate granularity in the task decomposition process, while planning defect anomalies refer to anomalies caused by timing conflicts or improper load distribution in the subtask planning process. The model classification adopts well-known text classification and feature matching algorithms. Differentiated correction strategies are implemented for different anomaly types. For decomposition deviation anomalies, missing subtasks are added or the decomposition granularity is adjusted. For planning defect anomalies, the execution sequence is adjusted or the execution node is replaced.

[0120] In addition, this step uses the WxTaskLM large model to accurately classify the root causes of anomalies and execute differentiated and lightweight repair logic. Its uniqueness lies in integrating anomaly information capture, accurate root cause classification, differentiated repair, and re-execution of related links into an efficient anomaly closed-loop processing flow, rather than the traditional full task restart.

[0121] Full anomaly information capture provides a complete data foundation for root cause diagnosis. Anomaly root cause classification confidence factor improves the accuracy of anomaly classification. Large model classification enables automated and accurate root cause localization. Differentiated repair strategies address different types of anomalies. Re-executing only the anomaly association link significantly reduces repair costs and business delay risks. Response time limits ensure the continuity of business execution.

[0122] As a preferred approach, S3's two-dimensional verification follows these steps:

[0123] Extract the current set of subtasks and the core requirements data of the original tasks in each round;

[0124] Calculate the completeness of requirement coverage and the adaptability of granularity separately; then remove abnormal sub-task interference data and smooth the indicators; finally determine the termination conditions for decomposition.

[0125] The comprehensive determination formula for dismantling termination is as follows:

[0126] .

[0127] in, The score is terminated by dismantling; The coefficients are for different scenarios: 0.65 for high compliance and 0.55 for rapid response. To determine the coverage of requirements, core requirements are assigned a value of 1, auxiliary requirements are assigned a value of 0.5, and irrelevant requirements are assigned a value of 0. The core demand quantity; This is the business compliance coefficient, ranging from 0.92 to 1.08. This is the granularity adaptation value; 1 for a single system, 0.3 for cross-system, and 0 for no entry point. The total number of subtasks is ≥0.85, and the disassembly is terminated.

[0128] This step calculates the completeness and granularity adaptability of the requirements coverage. Completeness measures the extent to which the subtasks cover the requirements of the original task, while adaptability measures the executable granularity of the subtasks. Next, abnormal subtask interference data is removed using a well-known data smoothing algorithm to eliminate the impact of abnormal data on the evaluation results and ensure the accuracy of the verification results. The calculation is then completed by substituting the results into the comprehensive judgment formula for dismantling termination. This formula combines the dual dimensions of completeness and adaptability, and introduces scenario coefficients and business compliance coefficients to adaptively adapt to the dismantling requirements of different business scenarios. Those skilled in the art can directly achieve quantitative judgment and termination control of dismantling quality by following the above steps and formulas.

[0129] Further explanation of the formula is needed. This formula weights and integrates completeness and granularity adaptability, while also introducing scenario coefficients and business compliance coefficients. The values ​​of the parameters in the formula are all set according to industry standards and enterprise compliance requirements for cross-system business management: Scenario Coefficient The coefficients of 0.65 for high compliance scenarios and 0.55 for rapid response scenarios are set based on industry practice that high compliance scenarios such as finance and regulation need to focus on the completeness of requirements, while rapid response scenarios such as ad-hoc inquiries need to focus on the adaptability of execution. (Business compliance coefficient) The value ranges from 0.92 to 1.08, set according to national standards for enterprise business compliance management, and is used to calibrate the completeness score for high-compliance scenarios; requirement coverage. Values ​​are assigned as follows: 1 for core requirements, 0.5 for secondary requirements, and 0 for irrelevant requirements, based on industry classification standards for the importance of task requirements; granularity adjustment value. The score is set as follows: 1 for single system, 0.3 for cross-system, and 0 for no entry point, based on the industry-recognized standard for subtask execution complexity; the dismantling termination score is ≥0.85, based on the enterprise's general qualified standard for task dismantling quality.

[0130] Taking a company's monthly cross-system financial reimbursement task as an example, when selecting parameters, this task belongs to a high-compliance financial scenario, and the scenario coefficient is... Take 0.65 as the business compliance coefficient. We set the value to 1.05 and used the WxTaskLM semantic understanding layer to extract the number of core requirements for the task. The tasks are as follows: collecting expense reports from various departments, verifying the compliance of expense information, entering data into the financial system, generating and submitting summary expense reports; and the set of sub-tasks obtained in the current round. It includes 6 sub-tasks: s1=Collect expense reports from various departments, s2=Verify the completeness of expense report information, s3=Check the compliance of expense reports, s4=Enter expense report information into the financial system, s5=Generate expense summary report, and s6=Submit the report to management.

[0131] The specific calculation process is as follows:

[0132] The first step is to calculate the total requirement coverage: s1 covers 1 core requirement. S2 covers auxiliary needs, S3 covers core requirement 2. S4 covers core requirement 3. S5 covers four core requirements. S6 covers auxiliary needs, Summation yields .

[0133] The second step is to calculate the completeness components:

[0134] The third step is to calculate the total granularity adaptability: all six subtasks are single-system operations. Both are 1, summing them up gives The average value is .

[0135] Step 4: Calculate the adaptability components:

[0136] Step 5, Calculate the total score: If the score is ≥0.85, the disassembly termination condition is met, the disassembly is stopped, and the subtask sequence is output.

[0137] This step achieves accurate and standardized judgment of disassembly quality through quantitative formulas. The formula algorithm and the technical features of the disassembly steps support each other. The algorithm provides quantitative basis for disassembly judgment, and the disassembly steps provide real-time data input for the algorithm. Together, they realize accurate closed-loop control of task disassembly, which avoids the execution risk caused by insufficient disassembly and the efficiency loss caused by excessive disassembly.

[0138] As a preferred approach, incremental fine-tuning of the WxTaskLM large model follows these steps:

[0139] Filter historical successful and anomalous amendment examples, and divide the training set and validation set in an 8:2 ratio;

[0140] Incremental fine-tuning of the gradient clipping coefficient is introduced to control the parameter update magnitude;

[0141] Only the parameters of the last three attention mechanisms are fine-tuned, without changing the basic architecture;

[0142] Set batch size to 32 and learning rate to 10. -5 Fine-tuning was completed in accordance with industrial deployment standards.

[0143] This step, based on the WxTaskLM large model architecture, further refines the incremental fine-tuning process of the S6 model optimization stage. It selects historical successful and abnormal correction examples from the S6 historical example library, covering all business scenarios including successful decomposition and abnormal correction. The training and validation sets are divided in an 8:2 ratio to effectively ensure the model's generalization ability. An incremental fine-tuning gradient pruning coefficient is introduced, based on a general solution to the gradient vanishing / exploding problem in large models. This coefficient controls the maximum norm of parameter updates, preventing performance degradation caused by excessive fluctuations in model parameters during fine-tuning. Only the parameters of the last three attention mechanisms are fine-tuned, without altering the pre-training infrastructure. This ensures the model adapts to specific business needs while fully preserving the general semantic understanding capabilities of the pre-trained model, avoiding overfitting. A batch size of 32 and a learning rate of 10 are set. -5 It balances fine-tuning efficiency and model accuracy. The fine-tuning process follows the conventional timing of large model training. Those skilled in the art can directly complete the incremental fine-tuning of the WxTaskLM model by following the above steps and parameters.

[0144] This step employs a lightweight incremental fine-tuning strategy to continuously adapt the model to dynamic changes in business scenarios. Its uniqueness lies in using a local fine-tuning strategy, optimizing only the top-level attention mechanism parameters while preserving the stability of the basic architecture. Simultaneously, standardized dataset partitioning and parameter settings ensure the reliability and consistency of the fine-tuning results. Case selection and dataset partitioning guarantee the rationality and comprehensiveness of the fine-tuning data, gradient pruning coefficients ensure the stability of parameter updates, the local fine-tuning strategy balances the model's business adaptability and general semantic capabilities, and standardized training parameters guarantee fine-tuning efficiency and accuracy.

[0145] As a preferred approach, the subtask priority allocation for S4 follows these steps:

[0146] Calculate the base score based on business importance, execution urgency, and the impact of blockage;

[0147] Collect real-time load data of the associated system and remaining execution time of subtasks;

[0148] Introducing a system interaction delay factor to calibrate for potential delay risks;

[0149] The execution order is determined by combining the base score and the correction coefficient.

[0150] The dynamic priority quantification formula for task priority allocation is as follows:

[0151] .

[0152] In the formula, As the final priority; Take values ​​of 0.3, 0.4, and 0.3; I(s) represents business importance, with a value range of 0-1; I(s) represents urgency, with a value range of 0-1. Blocking coefficient, with a value range of 1-3; This is the load correction factor; To account for delay risk, the value ranges from 0 to 0.5; This is the system interaction latency factor, with a value range of 0.9-1.1. The higher the score, the higher the priority of execution.

[0153] This step refines the S4 subtask priority allocation process. The logic and formula design align with the actual business needs of enterprise subtask scheduling. This formula deeply integrates basic business attribute priorities with multiple correction factors, including real-time system status and interaction risks. The parameter values ​​in the formula are set according to industry standards for enterprise business task management and system interaction: weighting coefficients. We assign weights of 0.3, 0.4, and 0.3 to the urgency of tasks, based on the general guidelines for enterprise business task priority management; business importance... urgency The value ranges from 0 to 1, set according to industry standards for task importance classification; the blocking coefficient B(s) ranges from 1 to 3, set according to industry-recognized standards for the number of downstream tasks blocked by subtasks; system interaction latency factor. The value ranges from 0.9 to 1.1, set according to the national standard for heterogeneous system interface interaction latency; load correction factor. With delay risk The values ​​are set based on industry-wide classification standards for system load and latency risk.

[0154] Taking the cross-system core customer contract approval task as an example, when selecting parameters, this task is designated as an urgent contract approval for core customers, indicating its business importance. Take 0.95 and apply urgency. Taking a value of 0.9, this subtask blocks three downstream contract fulfillment subtasks, with a blocking coefficient of 0.9. Set to 3; weighting coefficient The associated contract management system has a real-time load of 0.3, which is below the 0.5 threshold. (Load correction factor) Set to 0; the subtask has sufficient remaining execution time, with minimal risk of delay. Set to 0.1; the system interface interaction latency is stable, and the system interaction latency factor is... Take 1.0.

[0155] The specific calculation process is as follows:

[0156] The first step is to calculate the basic priority score:

[0157] The second step is to calculate the correction term:

[0158] The third step is to calculate the final priority score: The score was significantly higher than that of regular non-urgent subtasks, so it was listed as the highest priority in the execution plan and was given priority in the allocation of system execution resources. This example verifies the computability of the formula in actual business scenarios, the logic of dynamic priority adjustment, and the feasibility of implementation.

[0159] Furthermore, this step utilizes a multi-dimensional fusion quantitative formula to achieve dynamic and precise allocation of subtask priorities. Its uniqueness lies in integrating the business attributes of subtasks, the real-time operating status of the system, and cross-system interaction risks into a single formula, enabling adaptive intelligent scheduling of execution order. The formula algorithm and the priority allocation step mutually support each other. The algorithm provides a quantitative basis for priority allocation, while the steps provide the algorithm with real-time business and system data. Together, they achieve the rational allocation of heterogeneous system resources, ensuring the priority execution of highly important and urgent subtasks. Compared to the fixed priority allocation methods of existing technologies, this approach possesses a non-obvious innovation.

[0160] Preferably, the model optimization evaluation of S6 follows these steps:

[0161] Collect all data on disassembly, execution, and anomaly repair within a 72-hour period;

[0162] Calculate three types of indicators: disassembly accuracy, execution success rate, and anomaly repair rate;

[0163] Historical case studies are used to match gain factor calibration metrics; subsequently, model iteration is performed; the comprehensive evaluation formula for model optimization during iteration is:

[0164] .

[0165] in, To optimize the score; Take values ​​of 0.5, 0.3, and 0.2; For the correct number of decompositions; Number of successful executions; The number of effective repairs; For historical cases, a gain factor of 0.93-1.07 is matched; parameters are retained for scores ≥0.9, with an iteration cycle of 72 hours.

[0166] Further explanation of the formula is needed. This formula integrates three core indicators: accuracy, execution success rate, and anomaly repair rate. It also incorporates a historical case matching gain factor. The values ​​of the parameters in the formula are all set according to industry standards for large-scale model iteration and enterprise business management.

[0167] Weighting coefficient We assign the highest weight to the decomposition accuracy, using values ​​of 0.5, 0.3, and 0.2, based on the general standards for optimizing and evaluating large-scale model business applications; historical case matching gain factors are also used. The value is set between 0.93 and 1.07, based on the national standard for matching accuracy in the historical case database; parameters are retained if the optimization score is ≥0.9, based on the general qualified standard for large model optimization effect; the iteration cycle is 72 hours, based on the industry standard for enterprise business data accumulation.

[0168] Taking the quarterly optimization and evaluation of the enterprise cross-system task processing model as an example, when the parameters are selected, all business data within a 72-hour period are statistically analyzed: total number of sub-tasks. The number of correctly disassembled parts Total number of subtasks executed Successfully executed number Total number of anomalies within the period Valid repair count The newly added cases in this optimization have a high degree of matching with the historical case library, and the historical case matching gain factor is high. Set to 1.05; weighting coefficient .

[0169] The specific calculation process is as follows:

[0170] The first step is to calculate the breakdown of disassembly accuracy.

[0171] The second step is to calculate the success rate components.

[0172] The third step is to calculate the anomaly repair rate component.

[0173] Step 4, Calculate the total optimization score: If the score is ≥0.9, the model parameters of this incremental fine-tuning are retained to complete the model iterative optimization.

[0174] In addition, this step achieves precise quantification of model optimization effect through multi-dimensional evaluation formula. Its uniqueness lies in integrating indicators of the entire process of decomposition, execution, and anomaly repair, and matching gain factors with historical cases to achieve standardized and intelligent control of model iteration.

[0175] The specific method of the present invention is described below through two embodiments. The pre-implementation definitions and general parameter specifications for the embodiments are as follows:

[0176] (I) Core Formulas and Parameter Boundaries:

[0177] All calculations in this embodiment use the formulas described above, and the parameter values ​​are strictly controlled within the specified range.

[0178] (II) Implementation software and hardware environment:

[0179] Server environment: Intel Xeon 8-core CPU, 32GB RAM, NVIDIA T4 16GB graphics card, Ubuntu 20.04 LTS operating system.

[0180] Software environment: Python 3.9, Transformers 4.30.0, Whisper speech recognition framework, Selenium browser automation framework, NetworkX graph computing library.

[0181] The basic database consists of a MySQL 8.0 business rules database and a MongoDB 5.0 historical case database, which store basic data such as the enterprise's heterogeneous system operation specifications, data dictionary, and historical breakdown cases.

[0182] (III) Standard for calculating auxiliary factors:

[0183] Terminology normalization matching factor: The acceptable range is 0.92-1.08.

[0184] Fusion bias suppression factor:

[0185] Subtask dependency loop detection depth factor:

[0186] System load fluctuation correction factor: The acceptable range is 0.8-1.2.

[0187] Example 1: Monthly financial expense reimbursement summary of the enterprise and reporting of high compliance tasks to the group.

[0188] This embodiment is designed for a high-compliance financial scenario and strictly adheres to financial industry regulatory requirements. The entire execution and detailed calculation process of the method of this invention is as follows:

[0189] Original input for the task:

[0190] Corporate finance personnel input text tasks through the office client: complete the summary of expense reports from all 12 departments of the company this month, verify financial compliance, enter the data into the ERP financial system, generate a monthly expense summary report that complies with the group's standards, and submit it to the group's financial center. The entire process must be completed by the end of this month, and is traceable and compliant with financial supervision requirements.

[0191] Step S1: Multimodal task preprocessing, the calculation process is as follows:

[0192] 1. Text Redundancy Cleaning: Remove redundant expressions in the task description to obtain standardized text: Complete the summary of expense reports from 12 departments this month, compliance verification, ERP system entry, monthly expense report generation, and submission to the Group Finance Center. Complete by the end of the month, with compliance and traceability.

[0193] 2. Terminology Standardization and Core Feature Extraction: Referring to the standardized terminology dictionary of the financial industry, four types of core features were extracted:

[0194] Task type: Financial reimbursement summary and compliance reporting.

[0195] Related systems: OA expense reimbursement system, ERP financial system, and group financial reporting platform.

[0196] Timeframe: Before the end of the month (5 working days remaining).

[0197] Data requirements: departmental expense reports, compliance verification rules, ERP data entry fields, and group report templates.

[0198] Total number of core terms extracted in this study .

[0199] 3. Feature and Business Rule Base Matching and Integration: The extracted core features are matched with the operation specifications and data dictionaries of the three related systems based on similarity. The number of successfully matched terms is then calculated. .

[0200] 4. Terminology normalization matching factor calibration: Substitute into the formula for calculation. It is within the acceptable range of 0.92-1.08, and no secondary calibration is required.

[0201] 5. Output: Generates 1024-dimensional regularized task information. The entire preprocessing process takes 5.2 seconds, which meets the requirement of ≤8 seconds.

[0202] Step S2: The first round of decomposition of the WxTaskLM large model, the calculation process is as follows:

[0203] 1. Model initialization: Model parameters used: hidden layer dimension 1024, number of layers 16, number of attention heads 16, loading pre-trained weights and financial scenario fine-tuning weights.

[0204] 2. Semantic understanding layer encoding: Input 1024-dimensional regularized task information, output 1024-dimensional structured semantic vector, and extract the core requirements and compliance constraints of the task.

[0205] 3. Dual-branch attention mechanism fusion and bias suppression: The rule branch outputs a 768-dimensional business rule feature vector. The case branch outputs a 768-dimensional feature vector of historical cases. Calculate the cosine similarity between two vectors. Substituting into the formula for the fusion bias suppression factor:

[0206] .

[0207] The calculated value is at the lower limit of the acceptable range, triggering a dual-branch weight adjustment. After adjustment, the similarity is 0.52, and a recalculation is performed.

[0208] It is in the optimal range and outputs a 768-dimensional fused feature vector.

[0209] 4. First Round of Decomposition and Output: The multi-round inference layer completes the first round of decomposition, outputting 4 core target sub-tasks and 2 auxiliary support sub-tasks. The proportion of core sub-tasks is approximately 66.67% (4 / (4+2) ≈ 66.67%), which meets the requirement of ≥60%, resulting in the initial set of sub-tasks.

[0210] Core sub-tasks: expense reimbursement form compilation, compliance verification, ERP system data entry, report generation and submission.

[0211] Auxiliary subtasks: expense report collection and report archiving.

[0212] Step S3: Subtask executability determination and two-dimensional verification, the calculation process is as follows:

[0213] 1. Verification of basic executability conditions: The expense report summary subtask in the first round of decomposition requires cross-system operation between OA system and ERP system. It does not meet the single system operation condition and is marked as unexecutable, and enters the iterative decomposition.

[0214] 2. Circular Dependency Detection: Total Number of Initial Subtask Nodes The core sub-task expense report summarizes the number of associated downstream nodes. Substituting into the dependency loop detection depth factor formula:

[0215] .

[0216] After rounding, the value is 3. The detection depth is set to 3 layers. No circular dependencies were found after traversal.

[0217] 3. Iterative Decomposition: Unexecutable expense reports are broken down into two sub-tasks. After two rounds of iterative decomposition, a sequence of six executable sub-tasks is finally obtained:

[0218] s1 = Collection of expense reports from various departments (OA system), s2 = Expense report integrity verification (OA system), s3 = Expense compliance verification (financial compliance system), s4 = Expense information entry into ERP system, s5 = Generation of monthly expense report summary, s6 = Report submission to the group's financial center.

[0219] 4. Detailed calculation of the two-dimensional verification and disassembly termination formula:

[0220] Step 1: Determine formula parameters: High compliance scenario Business compliance coefficient Extracting the number of core requirements through the semantic understanding layer (Reimbursement form summary, compliance verification, ERP data entry, report submission), total number of sub-tasks .

[0221] Step 2: Calculate the total demand coverage :

[0222] S1 covers auxiliary needs, S2 covers auxiliary needs, S3 covers core requirement 2. S4 covers core requirement 3. S5 covers auxiliary needs, S6 covers four core requirements. ;

[0223] Summing yields:

[0224] Step 3: Calculate the completeness components:

[0225] .

[0226] Step 4: Calculate the sum and average of granularity fit: All 6 subtasks are single-system operations. Both are 1, summing them up gives ,average value .

[0227] Step 5: Calculate the adaptability components:

[0228] Step 6: Calculate the final disassembly termination score:

[0229] .

[0230] Step 7: Termination Decision: The score 1.1178 ≥ 0.85 satisfies the termination condition for disassembly. Stop disassembly and output the final executable subtask sequence. The number of disassembly rounds should be controlled within 5 rounds, which meets the requirements.

[0231] Step S4: Intelligent planning and priority allocation of subtasks, the calculation process is as follows:

[0232] 1. Dependency-based directed graph construction: Using 6 subtasks as nodes and data dependencies as directed edges, construct a directed graph with the following dependency relationships: s1→s2→s3→s4→s5→s6. There are no circular dependencies, so no loop breaking is required.

[0233] 2. System load calibration: The end of the month is the peak period for the financial system. The average load of the ERP system over the past 30 minutes is collected. Current real-time load Calculate the system load fluctuation correction factor: The value is constrained to the range of 0.8-1.2. If we take 0.8, we determine that the system is in a high-load state.

[0234] 3. Full-step calculation of dynamic priority quantification formula:

[0235] Fixed parameters: System interaction delay factor Load correction factor under high load scenarios Non-urgent subtasks are assigned a value of -0.2, while urgent subtasks are assigned a value of 0; delay risk. The core subtask is set to 0.1, and the auxiliary subtask is set to 0.

[0236] Taking the core subtask S3 expense reimbursement compliance verification as an example, the detailed calculation process is as follows:

[0237] Business Importance (Core aspects of high financial compliance), urgency of execution Blocking coefficient (Blocks 3 downstream subtasks).

[0238] Basic priority score:

[0239] Correction items:

[0240] Final priority score:

[0241] All subtask priorities are calculated using the same logic.

[0242] 4. Execution plan generation: Combining topology sorting and priority sorting, non-urgent subtasks s5 are adjusted to be executed during the low-load period at night, and finally, the subtask execution plan is generated, with the execution timing deviation controlled within ±50ms.

[0243] To intuitively present the core influence of the dynamic priority quantification formula in this step, through... Figure 2 , Figure 3 Demonstrating the impact of key parameters on the final priority score:

[0244] Figure 2 The curve illustrating the impact of the blocking coefficient on dynamic priority shows the linear growth pattern of the final priority score when the blocking coefficient varies within the range of 1-3, under conditions of fixed business importance, execution urgency, system load, and latency risk. This verifies the design logic that subtasks with more blocked downstream tasks have higher priority weights, which perfectly matches the priority ranking results of this step. Figure 2 The curves depict the trend of priority score changes corresponding to variations in the blocking coefficient under conditions of fixed business importance, execution urgency, system load, and latency risk.

[0245] Figure 3 The curve showing the impact of latency risk on dynamic priority demonstrates the trend of the final priority score when latency risk varies within the range of 0-0.5, while keeping other parameters fixed. This verifies the formula's adaptive calibration capability for latency risk in cross-system interactions, and is perfectly consistent with the timing adjustment logic in high-load scenarios in this step. Figure 3 The curve shows the trend of priority score changes corresponding to changes in delay risk, with other parameters fixed.

[0246] Step S5: Subtask execution control and anomaly diagnosis, performed according to the following steps:

[0247] 1. Operation type identification and execution logic adaptation: s1, s2, and s3 are web-based operations, executed by calling the Selenium browser driver; s4, s5, and s6 are system data interaction operations, executed by calling standardized interfaces.

[0248] 2. Multi-layer data integrity verification: In a high-compliance financial scenario, the redundancy factor for data verification in the execution chain is set to 3, and three layers of verification are performed: data format verification, field integrity verification, and financial compliance verification. All sub-task data verifications pass.

[0249] 3. Execution status monitoring and log retention: Real-time monitoring of subtask execution status. All subtasks are executed successfully on the first attempt. Execution logs are retained, including core information such as subtask identifier, execution steps, and system environment. The log retention period is set to 180 days, meeting the requirement of ≥90 days.

[0250] 4. Anomaly Diagnosis Simulation Verification: Simulate a timeout exception in the s4ERP system's data entry interface call, capture all exception information, introduce an exception root cause classification confidence factor of 0.95, and determine it as a planning defect type exception through the WxTaskLM large model. Adjust the execution sequence to a low-load period, only re-execute the s4 sub-task, and the exception repair response time is 12 seconds, which meets the requirement of ≤30 seconds.

[0251] Step S6: Model optimization evaluation and closed-loop iteration, the calculation process is as follows:

[0252] 1. Full Data Statistics: Statistics on all business data within the 72-hour period of this task: Total number of sub-tasks broken down Correctly decomposed number Total number of subtasks executed Successfully executed number Total number of outliers Valid repair count .

[0253] 2. Detailed calculation of the comprehensive evaluation formula for model optimization:

[0254] Fixed parameters: The newly added cases have a high degree of matching with the historical case database, and the historical case matching gain factor is high. .

[0255] Disassembly accuracy breakdown: 0.5×(181 / 186)≈0.5×0.9731=0.4866.

[0256] Success rate breakdown: 0.3×(180 / 184)≈0.3×0.9783=0.2935.

[0257] The abnormal repair rate is calculated as follows: 0.2 × (4 / 4) × 1.05 = 0.2 × 1 × 1.05 = 0.21.

[0258] Final optimized score: =0.4866+0.2935+0.21=0.9901.

[0259] To visually present the core impact principles of the comprehensive evaluation formula for model optimization in this step, through... Figure 4 , Figure 5 Demonstrating the impact of key parameters on optimization scores:

[0260] Figure 4 The curve showing the impact of decomposition accuracy on model optimization score demonstrates the trend of model optimization score when the decomposition accuracy varies within the range of 60%-97.5%, under the conditions of fixed execution success rate, anomaly repair rate and historical case matching gain factor. It also marks the 0.9 optimization qualification threshold, clearly presenting the core impact of decomposition accuracy on model optimization effect, which is completely consistent with the parameter design logic with the highest weight of decomposition accuracy in this step. Figure 4 The curve content shows the trend of optimization score change corresponding to the change in decomposition accuracy under the condition that other parameters are fixed, including the limited 0.9 optimization qualified threshold reference line.

[0261] Figure 5 The graph showing the influence of the historical case matching gain factor on the model optimization score demonstrates the linear change in the model optimization score when the historical case matching gain factor varies within the range of 0.93-1.07, while keeping other parameters constant. This verifies the calibrating effect of the historical case matching degree on the model optimization effect, which is completely consistent with the design logic of the closed-loop iteration in this step. Figure 5 The curve content shows the trend of optimization score change corresponding to the change of gain factor under the condition that other parameters are fixed, including the 0.9 optimization pass threshold reference line.

[0262] 3. Incremental Fine-tuning Execution: Optimization score ≥ 0.99, meeting the qualification requirements. Historical cases within the selection period are divided into training and validation sets at an 8:2 ratio. The gradient pruning coefficient is set to 1.0. Only the parameters of the last 3 layers of the model's attention mechanism are fine-tuned. The batch size is set to 32 and the learning rate to 10%. -5 This completes incremental fine-tuning of the model, updates the model weights to the historical case library, and forms a closed-loop optimization.

[0263] Example 2: Cross-platform key project approval and data reporting routine business tasks.

[0264] This embodiment focuses on conventional business scenarios, highlighting the adaptability of this method in non-high compliance scenarios. The core parameters and differentiated calculation processes are as follows:

[0265] Original input for the task:

[0266] Project management personnel input voice task: Complete the cross-platform approval of key projects in XX Industrial Park, extract the core construction data of the project after obtaining the approval results, and report it to the group's data management platform within 3 working days.

[0267] Core Differentiation Calculation Process:

[0268] 1. Preprocessing stage: Extracting the total number of core terms Number of successful matches Terminology normalization matching factor The preprocessing time was 5.8 seconds, which meets the requirements.

[0269] 2. Model decomposition stage: The similarity of the two-branch feature vectors is 0.54, and the fusion bias suppression factor is used. The first round of dismantling core sub-tasks accounted for 65%, which meets the requirement of ≥60%.

[0270] 3. Two-dimensional verification process: standard business scenarios Business compliance coefficient Core demand quantity After three rounds of iterative decomposition, eight executable subtasks were finally obtained. Substituting these subtasks into the decomposition termination formula, the final result was... The value is ≥0.85, which satisfies the termination condition.

[0271] 4. Priority allocation process: When the system is under low load, the load correction factor is applied. The core subtask project approval process has a final priority score of 1.6995, which is the highest priority, and the execution timing deviation is controlled within ±50ms.

[0272] 5. Model Optimization Phase: Data from a 72-hour period is statistically analyzed and substituted into the model optimization evaluation formula to calculate... ≥0.9, complete the incremental fine-tuning of the model, and form a closed-loop optimization.

[0273] IV. Verification of Implementation Results and Achievement of Technical Targets:

[0274] The two embodiments fully reproduce the entire process logic of the method of the present invention. All calculation processes strictly follow the above limitations, and the core technical indicators achieved are as follows:

[0275] 1. Preprocessing time is ≤8 seconds, and the semantic feature vector dimension is fixed at 1024, which meets the requirements;

[0276] 2. The core sub-tasks accounted for ≥60% in the first round of decomposition, and the number of decomposition rounds was ≤5, which meets the requirements;

[0277] 3. The decomposition termination score, priority calculation, and model optimization score all conform to the formula constraints, and the parameters are all within the specified range;

[0278] 4. Subtask execution success rate is 100%, exception repair response time is ≤30 seconds, and log retention time is ≥90 days, meeting the requirements;

[0279] 5. Execution timing deviations were all controlled within ±50ms, and model optimization scores were all ≥0.9, meeting the requirements.

[0280] This embodiment fully discloses the implementation details and the entire calculation process of the method of the present invention. Those skilled in the art can fully reproduce the technical solution of the present invention, solve its technical problems, and achieve the expected technical effects based on this embodiment.

[0281] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention. For those skilled in the art, any alternative improvements or transformations made to the implementation of the present invention fall within the protection scope of the present invention.

[0282] Any aspects of this invention not described in detail are well-known to those skilled in the art.

Claims

1. A method for multi-round task decomposition based on a large model, characterized in that, Includes the following steps: S1. Receive a multimodal complex task description, convert the voice task description into text, complete text redundancy cleaning, business terminology unification and feature extraction, integrate task semantic features and heterogeneous system business rule features, and generate regular task information. S2. Input the regularized task information into the WxTaskLM task decomposition model, and combine it with the business rule base and the historical task decomposition case base to decompose the complex task into core target sub-tasks and auxiliary support sub-tasks in the first round to obtain the initial set of sub-tasks. S3. Perform an executability test on the initial subtasks, iteratively decompose the non-executable subtasks, and perform a two-dimensional check on the completeness and granularity of the decomposition after each round of decomposition. If the check fails, adjust the strategy and re-decompose. If the check passes, proceed to the next round until an executable subtask sequence is generated. S4. Analyze the data dependencies of subtasks, the real-time load of heterogeneous systems and task priorities, construct a directed dependency graph and perform topological sorting, and generate a subtask execution plan that includes execution order, execution system, data source and execution time limit. S5. Proceed with the execution plan to execute sub-tasks, monitor execution status and data integrity in real time, and re-execute after root cause diagnosis and correction in case of anomalies; S6. Based on the full-process breakdown and execution data, optimize the parameters of the WxTaskLM large model and the sub-task planning logic, update the business rule base and historical case base, and form a closed-loop optimization; the two-dimensional verification of S3 follows these steps: Extract the current set of subtasks and the core requirements data of the original tasks in each round; Calculate the completeness of requirement coverage and the adaptability of granularity separately; then remove abnormal sub-task interference data and smooth the indicators; finally determine the termination conditions for decomposition. The comprehensive determination formula for dismantling termination is as follows: ； in, The score is terminated by dismantling; The coefficients are for different scenarios: 0.65 for high compliance and 0.55 for rapid response. To determine the coverage of requirements, core requirements are assigned a value of 1, auxiliary requirements are assigned a value of 0.5, and irrelevant requirements are assigned a value of 0. The core demand quantity; This is the business compliance coefficient, ranging from 0.92 to 1.

08. This is the granularity adaptation value; 1 for a single system, 0.3 for cross-system, and 0 for no entry point. The total number of subtasks is ≥0.85, and the disassembly is terminated.

2. The method for multi-round decomposition of a large model according to claim 1, characterized in that, The preprocessing of S1's multimodal tasks follows these steps: Convert the speech task description into text and remove redundant and meaningless expressions; Then, we standardized the format of business terminology and extracted four core features: task type, associated system, time node, and data requirement. Integrate core features with operational guidelines and data dictionaries from the business rule base; The terminology normalization matching factor is introduced to calibrate the feature output; the preprocessing time is no more than 8 seconds, and the semantic feature vector dimension is fixed at 1024.

3. The method for multi-round decomposition of a large model according to claim 1, characterized in that, The WxTaskLM large model is used for semantic understanding, multi-round decomposition, deviation verification, and anomaly root cause diagnosis of cross-system business tasks. The model construction and inference follow these steps: Build an improved Transformer architecture that includes a task semantic understanding layer, a business rule fusion layer, and a multi-round inference layer; Deep semantics are extracted and interference is filtered out through bidirectional encoding at the semantic understanding layer; By fusing rules and case knowledge through a dual-branch independent attention mechanism, a fusion bias suppression factor is introduced to avoid feature confusion. The hierarchical decomposition and verification are completed through multiple rounds of inference layers; The model parameters are 1024 hidden layer dimensions, 16 layers, and 16 attention heads, with the core sub-tasks accounting for no less than 60% in the first round.

4. The method for multi-round decomposition of a large model according to claim 3, characterized in that, The executability determination of S3's subtasks follows these steps: Verify whether the subtask meets the three conditions of single system operation, clear input and output, and no unresolved dependencies; Introduce a subtask dependency cycle detection depth factor to traverse and detect circular dependencies across subtasks; Mark subtasks that do not meet the conditions or have circular dependencies as non-executable; The number of disassembly rounds should not exceed 5; if it does, the model should be sent back for re-disassembly.

5. The method for multi-round decomposition of a large model according to claim 4, characterized in that, S4's subtask planning follows these steps: Extract the input and output data items of the subtasks, and construct a dependency directed graph with the subtasks as nodes and the data dependencies as directed edges; Perform cycle detection and automatic cycle breaking on dependent directed graphs to eliminate timing conflicts; Introduce a system load fluctuation correction factor to calibrate the system load determination results; The acyclic graph is topologically sorted and prioritized, with execution timing deviations controlled within ±50ms.

6. The method for multi-round decomposition of a large model according to claim 5, characterized in that, S5 subtask execution control follows these steps: Identify subtask operation types and distinguish between web-based operations and system data interaction operations; Execution logic is adapted to the operation type to advance the task; A redundancy factor for execution link data verification is introduced to complete multi-layer data integrity verification; The system includes four core components: real-time monitoring of execution status, retention of abnormal information, and log retention for at least 90 days.

7. The method for multi-round decomposition of a large model according to claim 6, characterized in that, The abnormal root cause diagnosis of S5 follows these steps: Capture full information including the identifier of the abnormal subtask, the execution node, and the exception type; An anomaly root cause classification confidence factor is introduced, and the WxTaskLM large model is used to distinguish between decomposition deviation and planning defect anomalies. Complete subtask supplementation or timing adjustment for different exception types; Only re-execute the abnormal related links, and the abnormal repair response time shall not exceed 30 seconds.

8. The method for multi-round decomposition of a large model according to claim 3, characterized in that, The incremental fine-tuning of the WxTaskLM large model follows these steps: Filter historical successful and anomalous amendment examples, and divide the training set and validation set in an 8:2 ratio; Incremental fine-tuning of the gradient clipping coefficient is introduced to control the parameter update magnitude; Only the parameters of the last three attention mechanisms are fine-tuned, without changing the basic architecture; Set batch size to 32 and learning rate to 10. -5 .

Images