Dynamic convergence control method and device for iterative task in intelligent application

Abstract

The invention discloses a dynamic convergence control method and device for an iterative task in intelligent application, and belongs to the technical field of computers. Task data are acquired and initialized to obtain a plurality of data units, and a corresponding high-dimensional state vector is constructed for each data unit; the high-dimensional state vector is used for representing the change of content embedding, and performing iterative loop processing on the task data; after execution of one iteration loop processing period is completed, the high-dimensional state vectors of the related data units are updated, and updated high-dimensional state vectors are obtained; and based on a preset convergence condition, judging the updated high-dimensional state vector by using a semantic convergence controller to obtain a judgment result. According to the method, the high-dimensional state vector is judged and processed, so that resources are saved, and a high-quality result is generated; by introducing the semantic convergence controller, whether semantic oscillation occurs or not is judged, meanwhile, invalid optimization can be recognized, and the iterative task processing efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of computer, and particularly relates to a dynamic convergence control method and device for iterative tasks in intelligent application. BACKGROUND

[0002] In recent years, with the continuous development of AI technology, modern AI applications have been widely applied in various life scenarios. The core feature of modern AI applications is the ability to perform complex tasks that require iteration and feedback, such as multiple rounds of modifying a report or a design scheme. The underlying execution framework must support loop structures. How to intelligently control and terminate loops is the focus of existing technologies.

[0003] The existing technology of AI application loop implementation includes the following ways: (1) Microsoft's AutoGen and LangChain project's LangGraph library in the existing technology have provided the ability to build loop workflows. AutoGen forms iteration by allowing multiple agents to repeatedly dialogue in a group chat. LangGraph library allows processes to return to previous nodes based on specific conditions by defining state graphs and setting conditional edges to form loops; (2) In the AI Agent design pattern (such as Google ADK or OpenAI Cookbook's Actor-Critic pattern), an LLM (Large Language Model) is used to evaluate another LLM's output quality. One LLM is used as a Critic or Judge, and another LLM is used as an Actor or Writer. The loop logic generates content for the Actor, and the Critic evaluates the content. If the Critic considers the quality to be substandard, the loop continues until the Critic gives a satisfactory signal.

[0004] But the above-mentioned methods all have corresponding disadvantages, such as: (1) resource waste and low result quality: the termination of the cycle depends on fixed rounds, keyword matching or manual intervention, which all sacrifice the core value of automation, and the framework cannot independently judge the semantic change of the content output by this iteration compared with the last time, cannot intelligently answer the core question of "whether the task has been truly completed", and leads to the problems of resource waste or low result quality when the result is good enough or the cycle is exited too early when the satisfactory effect is not reached; (2) unable to detect semantic shock and identify invalid optimization: when the views of the Actor and the Critic are inconsistent, it is easy to fall into a tug-of-war, such as the Actor repeatedly generating similar content and the Critic repeatedly giving similar negative opinions, leading to idle circulation, wasting of computing resources, and when the content quality is already high, the Actor may still have some insignificant or even negative minor modifications, which are difficult to effectively identify by the quality score of the Critic.

[0005] Therefore, based on the above-mentioned disadvantages, how to provide an effective technical solution to solve the problems of resource waste, low result quality, inability to detect semantic shock and identify invalid optimization in the prior art has become a difficult problem to be solved in the prior art. SUMMARY

[0006] The purpose of the present application is to provide a dynamic convergence control method and device for iterative tasks in intelligent applications to solve the above-mentioned problems in the prior art.

[0007] In order to achieve the above-mentioned purpose, the present application adopts the following technical solutions: In a first aspect, the present application provides a dynamic convergence control method for iterative tasks in intelligent applications, comprising: obtaining task data, initializing the task data to obtain a plurality of data units, for each data unit, constructing a corresponding high-dimensional state vector, the high-dimensional state vector being used to represent the change of content embedding, and performing iterative loop processing on the task data; after an iteration loop processing period is executed, updating the high-dimensional state vector of the related data unit to obtain an updated high-dimensional state vector; based on a preset convergence condition, using a semantic convergence controller to judge the updated high-dimensional state vector to obtain a judgment result, the judgment result being used to indicate the next round of iteration loop processing flow or terminate the cycle for the task data.

[0008] In a possible design, the high-dimensional state vector includes a content embedding vector, a history embedding list, a stability score and a quality score; and the construction of a corresponding high-dimensional state vector for each data unit includes: The pre-trained deep learning model is used to calculate the data unit to obtain a content embedding vector; Based on the content embedding vector, a history embedding list is obtained; The content embedding vector and the history embedding list are calculated to obtain a stability score; Based on a preset business standard, the content of the current data unit is evaluated to obtain a quality score.

[0009] In a possible design, the calculation expression of the stability score is: ; In the formula, is the stability score, is the number of iteration loop processing, is the content embedding vector, is the history embedding list, is the number of current iteration loop processing.

[0010] In a possible design, the judgment result includes performing a next round of iteration loop processing or terminating the loop; after obtaining the judgment result, the following steps are further included: If the judgment result is to perform the next round of iteration loop processing, at least one data unit with the lowest stability score or quality score is selected, the data unit is input into a processing operator to obtain an updated data unit, and the next round of iteration loop processing is performed based on the updated data unit; If the judgment result is to terminate the loop, a convergence audit log is generated.

[0011] In a possible design, the preset convergence condition at least includes any one of the following: Whether the stability average value or the minimum value of the stability score is higher than a preset stability threshold value in a preset period, the stability average value being calculated based on the stability scores of all data units; Whether the quality scores of all data units are greater than or equal to a preset quality threshold value; Whether the quality growth rate is lower than a preset gain threshold value in a preset period, the quality growth rate being calculated based on the quality scores of all data units; Whether at least one data unit has a high-dimensional state vector that repeatedly switches among multiple historical states using SSC detection, the historical state being a high-dimensional state vector corresponding to all data units in all iteration loop processing periods in the history embedding list.

[0012] In a possible design, the calculation expression of the quality growth rate is: ; In the formula, is a quality growth rate, t is a number of iteration loop processing times, is a quality average value of the tth iteration loop processing, is a quality average value of the t-1th iteration loop processing.

[0013] In a possible design, the data unit is a minimum data object in the task data.

[0014] In a second aspect, the present application provides a dynamic convergence control device for an iterative task in an intelligent application, comprising: an acquisition construction unit configured to acquire task data, initialize the task data, obtain a plurality of data units, construct a corresponding high-dimensional state vector for each data unit, and perform iterative loop processing on the task data, wherein the high-dimensional state vector is used to represent changes in content embedding; an update processing unit configured to update the high-dimensional state vector of the related data unit after an iteration loop processing period is executed, and obtain an updated high-dimensional state vector; a convergence determination unit configured to determine the updated high-dimensional state vector based on a preset convergence condition using a semantic convergence controller, and obtain a determination result, wherein the determination result is used to indicate whether to perform a next round of iteration loop processing on the task data or terminate the loop.

[0015] In a possible design, the device further comprises a data unit repository, a process execution engine, and a state vector manager; the data unit repository is configured to store the data unit and the corresponding high-dimensional state vector; the process execution engine is configured to modify the content of the data unit by calling a processing operator; the state vector manager is configured to maintain and update the high-dimensional state vector.

[0016] In a third aspect, the present application provides a computer device comprising a memory, a processor, and a transceiver connected in sequence and in communication, wherein the memory is configured to store a computer program, the transceiver is configured to receive and send messages, and the processor is configured to read the computer program and execute the dynamic convergence control method for an iterative task in an intelligent application according to any one of the above aspects.

[0017] In a fourth aspect, the present application provides a computer program product comprising instructions, which, when executed on a computer, cause the computer to execute the dynamic convergence control method for an iterative task in an intelligent application according to any one of the above aspects.

[0018] The present application has the following beneficial effects: The application discloses a dynamic convergence control method and device for an iterative task in an intelligent application, and the method comprises the following steps: obtaining task data, initializing the task data, obtaining a plurality of data units, constructing a corresponding high-dimensional state vector for each data unit, the high-dimensional state vector is used for representing the change of content embedding, and the task data is subjected to iterative loop processing; when an iterative loop processing period is executed, the high-dimensional state vector of the related data unit is updated to obtain an updated high-dimensional state vector; based on a preset convergence condition, a semantic convergence controller is used to judge the updated high-dimensional state vector to obtain a judgment result, and the judgment result is used for indicating whether the next round of iterative loop processing flow or the termination loop is performed on the task data. The application constructs a corresponding high-dimensional state vector for the data unit of the task data, creatively introduces stability as a new dimension orthogonal to quality, realizes real closed-loop feedback, solves the problem that semantic shock and invalid optimization cannot be identified in the prior art, intelligently judges whether the iterative loop process should be terminated, intelligently answers the core problem of "whether the task has been completed", saves resources, and generates a result with high quality; the semantic convergence controller is used for global convergence judgment, the semantic convergence controller detects whether the iterative loop processing process is trapped in semantic shock in combination with the preset convergence condition, can identify invalid optimization, and improves the iterative task processing efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 A flow chart of the dynamic convergence control method provided for the first aspect of the application; Figure 2 A module diagram of the dynamic convergence control device provided for the second aspect of the application; Figure 3 An architectural schematic diagram of the dynamic convergence control device provided for the second aspect of the application. DETAILED DESCRIPTION

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the application will be briefly introduced below in combination with the drawings and the description of the embodiments or the prior art. Obviously, the following description of the drawings is only some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor. It should be noted that the description of these embodiments is used to help understand the application, but does not constitute a limitation on the application.

[0021] It should be understood that although the terms first, second, etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are only used to distinguish one unit from another. For example, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit, without departing from the scope of the exemplary embodiments of the invention.

[0022] It should be understood that the term "and / or" that may appear in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" that may appear in this document describes another relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " that may appear in this document generally indicates that the related objects before and after it are in an "or" relationship.

[0023] Example: like Figure 1 As shown, the first aspect of this embodiment provides a dynamic convergence control method for iterative tasks in intelligent applications, which can be executed, but is not limited to, by a computer device or virtual machine with certain computing resources, such as a personal computer or smartphone, or by a virtual machine; the dynamic convergence control method includes, but is not limited to, the following steps: S1. Obtain task data, initialize the task data to obtain multiple data units, and construct a corresponding high-dimensional state vector for each data unit. The high-dimensional state vector is used to characterize the changes in content embedding. Perform iterative loop processing on the task data. The high-dimensional state vector represents the change in the content embedding vector, and the high-dimensional state vector includes, but is not limited to, stability score, quality score, content embedding vector, and historical embedding list.

[0024] In this embodiment, a data unit is the smallest data object in the task data that can be independently processed and evaluated. Examples include a paragraph of text, a summary, a region in an image, or a function in code. The high-dimensional state vector is a data structure used to comprehensively describe the dynamic state of a data unit at any given time. It contains quantitative indicators that characterize the intrinsic semantics and extrinsic quality of the data.

[0025] For example, the task data is "AI automatically generates a market analysis report, which includes three parts: market size, major players, and future trends". The task data is initialized to obtain three data units: "market size", "major players", and "future trends".

[0026] Specifically, in step S1, for each data unit, a corresponding high-dimensional state vector is constructed, including: S11. Calculate the data unit using a pre-trained deep learning model to obtain a content embedding vector; Preferably, the training process of the deep learning model includes: constructing a deep learning model; obtaining historical data units, labeling the historical data units based on a preset label value to obtain historical labeled data units, inputting the historical labeled data units into the deep learning model as training data, obtaining a predicted value, calculating the predicted value and the preset label value using a loss function to obtain a loss value, updating the parameters of the deep learning model based on the loss value to obtain an updated deep learning model, inputting the training data into the updated deep learning model, repeating the above updating steps until a preset update termination condition is reached, obtaining a trained deep learning model, the preset update termination condition includes but is not limited to a loss value less than a preset threshold or reaching a preset training number.

[0027] S12. Obtain a historical embedding list based on the content embedding vector; Wherein, the historical embedding list is The content embedding vector of the data unit in the past N iterations is stored in the historical embedding list; S13. Calculate the content embedding vector and the historical embedding list to obtain a stability score; Further, the calculation expression of the stability score is: ; In the formula, is the stability score, is the number of iteration loop processing, is the content embedding vector, is the historical embedding list, is the number of current iteration loop processing.

[0028] S14. Evaluate the content of the current data unit based on a preset business standard to obtain a quality score.

[0029] Specifically, at least one evaluation operator is used to evaluate the content of the current data unit, wherein the evaluation operator is a functional unit that encapsulates specific evaluation logic, has a clear input and output interface, and is used to perform quantitative evaluation of data, models or systems; The preset business standard can be set according to different businesses for different modules and implementations. Here, for example, the business standard is the specification degree of text normalization or the completion degree of logical processing.

[0030] S2. When one iteration cycle processing period is executed, the high-dimensional state vector of the relevant data unit is updated to obtain an updated high-dimensional state vector; S3. Based on a preset convergence condition, the updated high-dimensional state vector is judged by using a semantic convergence controller (SCC) to obtain a judgment result.

[0031] Specifically, the preset convergence condition at least includes any one of the following: whether the minimum value of the stability average value or the stability score is higher than a preset stability threshold value within a preset period, the stability average value being calculated based on the stability scores of all data units; whether the quality scores of all data units are greater than or equal to a preset quality threshold value; whether the quality growth rate is lower than a preset gain threshold value within a preset period, the quality growth rate being calculated based on the quality scores of all data units; whether the high-dimensional state vector of at least one data unit repeatedly switches among multiple historical states is detected by using the SCC, the historical state being the high-dimensional state vector corresponding to all data units within all iteration cycle processing periods in the historical embedding list.

[0032] It should be noted that in the embodiment, the stability scores of all data units are calculated to obtain the stability average value; the quality scores of all data units are calculated to obtain the quality average value, and the quality growth rate is obtained according to the quality average value.

[0033] wherein, the calculation expression of the quality growth rate is: ; In the formula, is the quality growth rate, t is the number of iteration cycle processing, is the quality average value of the tthiteration cycle processing, is the quality average value of the (t-1)thiteration cycle processing.

[0034] For example, the stability threshold value is set to 0.98, the gain threshold value is set to 0.9, after the first iteration cycle processing, the AI generates a preliminary draft, the quality score of "market size" is 0.8, the quality score of "main players" is 0.85, and the quality score of "future trend" is 0.7, at this time, the SCC judges that it does not meet the standard, and enters the next round of iteration process.

[0035] In a possible design, the judgment result includes performing the next round of iteration cycle processing process or terminating the cycle; after obtaining the judgment result, the method further includes: S4. If the result of the judgment is to proceed with the next round of iteration cycle processing, select at least one data unit with the lowest stability score or quality score, input the data unit into the processing operator, obtain an updated data unit, and perform the next round of iteration cycle processing based on the updated data unit; S5. If the result of the judgment is to terminate the cycle, generate a convergence audit log.

[0036] For example, according to the first evaluation result, the AI optimizes the data unit of “future trend”, based on the optimized data unit, after the second iteration cycle processing, the quality score of “market size” is 0.82, the quality score of “main players” is 0.85, and the quality score of “future trend” is 0.88, and the SCC still does not meet all the standards, and the next round of iteration process is entered again; the AI optimizes all data units, and after the third iteration cycle processing, the quality score of “market size” is 0.91, the quality score of “main players” is 0.92, and the quality score of “future trend” is 0.90, at this time the SCC makes a global judgment, and meets the termination cycle condition, and the iteration cycle processing is terminated.

[0037] It should be noted that in the embodiment, since the high-dimensional state vector represents the change of content embedding, it is only necessary to track the change of content embedding, judge the iteration cycle processing according to the change of content embedding, and realize the convergence control of the iteration task; preferably, for each data unit, a corresponding content embedding vector is constructed, that is, only the change data of the content embedding vector of the data unit with the iteration cycle processing period needs to be tracked, and the change data is judged by the semantic convergence controller to realize the convergence control of the iteration task.

[0038] Based on the above disclosure, the present application can directly compare the semantic state of the final output result to determine whether to converge without setting an external reward signal. The present application is a self-contained and pre-training-free deterministic engineering control method. Instead of determining the specific actions within the loop, the present application determines whether the loop process has reached the target and whether to terminate. The present application upgrades the loop control from mechanical control to intelligent control, greatly expands the application scenarios of automatic processes, automatically makes the most economic decision between "continuing iteration to pursue higher quality" and "stopping iteration to save computing cost" without human intervention, provides cost control capability, has significant commercial value, and finally generates a convergence audit log, providing a complete and traceable evidence chain for dynamic convergence control process, facilitating developers to troubleshoot and optimize algorithms. The present application creatively introduces stability as a new dimension orthogonal to quality, constructs a real closed-loop feedback process, solves the problem of semantic shock and invalid optimization that cannot be solved by single-point quality evaluation in the prior art, and can produce a synergistic effect of "1+1>2". At the same time, the present application breaks through the technical bias of the industry to give control to "external code" or "single judge", proposes that the control logic should "come from the state evolution of data itself over time", and realizes a cognitive leap from "evaluation of static results" to "control of dynamic processes".

[0039] As shown in Figures 2-3 The second aspect of the present embodiment provides a dynamic convergence control device for an iterative task in an intelligent application, comprising: An acquisition construction unit is configured to acquire task data, initialize the task data, obtain a plurality of data units, construct a corresponding high-dimensional state vector for each data unit, wherein the high-dimensional state vector is used to represent the change of content embedding, and perform iterative loop processing on the task data. An update processing unit is configured to update the high-dimensional state vector of the related data unit after an iteration loop processing period is completed, and obtain an updated high-dimensional state vector. A convergence determination unit is configured to use a semantic convergence controller to determine the updated high-dimensional state vector based on a preset convergence condition, and obtain a determination result, wherein the determination result is used to indicate whether to perform a next round of iterative loop processing on the task data or terminate the loop.

[0040] In a possible design, the device further comprises a data unit repository, a process execution engine, and a state vector manager. The data unit repository is configured to store the data units and the corresponding high-dimensional state vectors. The process execution engine is configured to call a processing operator to modify the content of the data units. The state vector manager is configured to maintain and update the high-dimensional state vectors.

[0041] It should be noted that the data unit storage in the embodiment and the data unit storage in the data unit storage repository in the embodiment are actually hardware facilities of the data unit storage. Figure 3

[0042] The working process, working details and technical effects of the dynamic convergence control device provided in the second aspect of the embodiment can be referred to the dynamic convergence control method described in the first aspect, and will not be repeated here.

[0043] The third aspect of the embodiment provides a computer device, which comprises a memory, a processor and a transceiver connected in sequence, wherein the memory is used to store a computer program, the transceiver is used to transceive messages, and the processor is used to read the computer program and execute the dynamic convergence control method described in the first aspect. Specifically, the memory can include, but is not limited to, a random access memory (RAM), a read-only memory (ROM), a flash memory, a first input first output (FIFO) memory and / or a first input last output (FILO) memory, etc.; and the processor can be, but is not limited to, a microprocessor with a model number of STM32F105 series. In addition, the computer device can further include, but is not limited to, a power module, a display screen and other necessary components.

[0044] The working process, working details and technical effects of the computer device described above provided in the third aspect of the embodiment can be referred to the dynamic convergence control method described in the first aspect, and will not be repeated here.

[0045] The fourth aspect of the embodiment provides a computer program product, which comprises a computer program or instructions, and the computer program or the instructions are used to implement the dynamic convergence control method described in the first aspect when executed by a computer.

[0046] The working process, working details and technical effects of the computer program product described above provided in the fourth aspect of the embodiment can be referred to the dynamic convergence control method described in the first aspect, and will not be repeated here.

[0047] Finally, it should be noted that: the above only describes the preferred embodiments of the present application, and is not used to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.​

Claims

1. A method for dynamic convergence control of iterative tasks in intelligent applications, characterized in that, The method comprises the following steps: obtaining task data, initializing the task data, obtaining a plurality of data units, for each data unit, constructing a corresponding high-dimensional state vector, the high-dimensional state vector is used to represent the change of content embedding, and the task data is iteratively processed; after an iteration cycle processing period is executed, the high-dimensional state vector of the related data unit is updated to obtain an updated high-dimensional state vector; based on a preset convergence condition, using a semantic convergence controller to judge the updated high-dimensional state vector to obtain a judgment result, the judgment result is used to indicate the next round of iteration cycle processing flow or termination cycle of the task data.

2. The method of claim 1, wherein, The high-dimensional state vector comprises a content embedding vector, a history embedding list, a stability score and a quality score. For each data unit, a corresponding high-dimensional state vector is constructed, which comprises: using a pre-trained deep learning model to calculate the data unit to obtain a content embedding vector; based on the content embedding vector, a history embedding list is obtained; the content embedding vector and the history embedding list are calculated to obtain a stability score; based on a preset business standard, the content of the current data unit is evaluated to obtain a quality score.

3. The method of claim 2, wherein, The calculation expression of the stability score is: ； wherein is a stability score, is a number of iteration loop processes, is a content embedding vector, is a history embedding list, is a number of current iteration loop processes.

4. The method of claim 2, wherein, The judgment result comprises the next round of iteration cycle processing flow or termination cycle; after obtaining the judgment result, it further comprises: if the judgment result is the next round of iteration cycle processing flow, at least one data unit with the lowest stability score or quality score is selected, the data unit is input into the processing operator to obtain an updated data unit, and the next round of iteration cycle processing is performed based on the updated data unit; if the judgment result is termination cycle, a convergence audit log is generated.

5. The method of claim 2, wherein the method further comprises: The preset convergence condition at least comprises any one of the following: whether the stability average value or the minimum value of the stability score is higher than the preset stability threshold value within a preset period, the stability average value is calculated based on the stability scores of all data units; whether the quality scores of all data units are greater than or equal to the preset quality threshold value; whether the quality growth rate is lower than the preset gain threshold value within a preset period, the quality growth rate is calculated based on the quality scores of all data units; whether at least one high-dimensional state vector of the data unit repeatedly switches among a plurality of history states is detected by using SSC, the history state is the high-dimensional state vector corresponding to all data units within all iteration cycle processing periods in the history embedding list.

6. The method of claim 5, wherein, The calculation expression of the quality growth rate is: ； In the formula, is the mass growth rate, t is the number of iteration loop processing, is the mass average value of the tth iteration loop processing, is the mass average value of the t-1th iteration loop processing.

7. An apparatus for dynamic convergence control of iterative tasks in intelligent applications, for implementing the method of any one of claims 1 to 6, characterized in that, The method comprises the following steps: an obtaining unit is used to obtain task data, initialize the task data, obtain a plurality of data units, for each data unit, construct a corresponding high-dimensional state vector, the high-dimensional state vector is used to represent the change of content embedding, and the task data is iteratively processed; an updating processing unit is used to update the high-dimensional state vector of the related data unit to obtain an updated high-dimensional state vector after an iteration cycle processing period is executed; A convergence determination unit is configured to determine the updated high-dimensional state vector using the semantic convergence controller based on a preset convergence condition, and obtain a determination result, which is used to indicate whether to perform a next round of iteration loop processing flow or terminate the loop on the task data.

8. The device of claim 7, wherein, The data unit repository, the process execution engine, and the state vector manager are further included. The data unit repository is configured to store data units and corresponding high-dimensional state vectors. The process execution engine is configured to call a processing operator to modify the content of the data unit. The state vector manager is configured to maintain and update the high-dimensional state vector.

9. A computer device, comprising: The memory is configured to store a computer program, the transceiver is configured to receive and send messages, and the processor is configured to read the computer program and execute the dynamic convergence control method for an iterative task in an intelligent application according to any one of claims 1-6.

10. A computer program product comprising computer programs or instructions, characterized in that, The computer program or the instructions realize the dynamic convergence control method for an iterative task in an intelligent application according to any one of claims 1-6 when executed by a computer.

Images