Task execution method based on large language model and related equipment

Through the method of task decomposition, planning, calibration and combination, the problem of poor universality of AI-Agent in performing geoscience tasks by large language model is solved, which improves the rationality and universality of task execution and reduces the difficulty of execution of complex tasks. .

CN119067094AActive Publication Date: 2024-12-03CENT SOUTH UNIV

Patent Information

Application Number
CN202411301192.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-18
Publication Date
2024-12-03
Estimated Expiration
2044-09-18

AI Technical Summary

Technical Problem

When performing geographic science exploration and discovery tasks, the existing AI-Agent driven by large language models has the problem of poor generalization of task execution methods, resulting in high difficulty in execution of complex tasks and insufficient logic and rationality.

Method used

By using the task decomposition model to decompose complex tasks into multiple subtasks, and building an execution model for each subtask, the task planning model is used to determine the execution order of subtasks, calibrate and combine them, and finally use the execution model to perform combination tasks and improve task execution rationality and universality.

Benefits of technology

Convert complex tasks into multiple simple subtasks, reduce the difficulty of executing complex tasks, improve the logic and rationality of subtasks, and improve the universality and efficiency of task execution methods.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119067094A_ABST
    Figure CN119067094A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of artificial intelligence, and provides a task execution method based on a large language model and related equipment. The method comprises the following steps: decomposing a complex task into a plurality of sub-tasks by utilizing a task decomposition model, and constructing an execution model for each sub-task; planning all the sub-tasks by using the task planning model to obtain an execution logic sequence of all the sub-tasks; calibrating each subtask to obtain a plurality of final subtasks, and then combining all the final subtasks according to an execution sequence to obtain a combined task; and executing the combined task by utilizing all the execution models to obtain a final execution result of the complex task. According to the method, the universality of the task execution method can be improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a task execution method based on a large language model and related equipment. Background Art

[0002] Earth science research and geographic knowledge discovery is a highly complex and multidimensional task. Solving related problems requires processing a large amount of data and conducting statistical information and knowledge mining for complex problems. Large Language Models (LLMs) encode a large amount of human language including earth science and geographical knowledge, and thus bring about human-like complex task understanding and reasoning capabilities. At the same time, artificial intelligence agents (AI-Agents) driven by large language models (LLMs) show more automated task understanding and reasoning. Therefore, artificial intelligence agents (AI-Agents) driven by large language models (LLMs) make automated earth science research and geographical knowledge discovery possible. However, due to the complexity of earth science research and geographical knowledge discovery, most related tasks are highly complex. When the current large language model-driven AI-Agents are applied to the execution of related tasks in geographical science exploration and discovery, there is a problem of poor universality of task execution methods. Summary of the invention

[0003] The present application provides a task execution method and related devices based on a large language model, which can solve the problem of poor versatility of the task execution method.

[0004] In a first aspect, an embodiment of the present application provides a task execution method based on a large language model, the task execution method comprising:

[0005] The task decomposition model is used to decompose the complex task to be executed into multiple subtasks, and an execution model is built for each subtask; the task decomposition model is a large language model that can perform task decomposition;

[0006] The task planning model is used to plan all subtasks and obtain the execution order of all subtasks; the task planning model is a large language model that can perform subtask planning;

[0007] Calibrate each subtask to obtain multiple final subtasks, and then combine all the final subtasks in the execution order to obtain a combined task; the combined task includes all the planned final subtasks and the combined reasoning rules and combined reasoning methods between all the final subtasks. The combined reasoning rules are used to describe the order in which all the final subtasks are executed, and the combined reasoning method is a method for optimizing the combined reasoning rules;

[0008] All execution models are used to execute the combined task and obtain the final execution result of the complex task.

[0009] Optionally, use the task decomposition model to decompose complex tasks into multiple subtasks, including:

[0010] By formula:

[0011] LLDM(I)={t1,t2,...,t n}

[0012] Get multiple subtasks {t1, t2, ..., t n};

[0013] Among them, LLDM() represents the task decomposition model, I represents the complex task, t1 represents the first subtask, t2 represents the second subtask, and t n Indicates the nth subtask, where n is the total number of subtasks.

[0014] Optionally, build an execution model for each subtask, including:

[0015] By formula:

[0016] LLMM={(t1,M1),(t2,M2),...,(t n ,M n )}

[0017] Get each subtask and execution model combination {(t1,M1),(t2,M2),...,(t n ,M n )};

[0018] Among them, LLMM represents the modeling model, M1 represents the execution model of the first subtask, M2 represents the execution model of the second subtask, and M n Represents the execution model of the nth subtask.

[0019] Optionally, use the task planning model to plan all subtasks and obtain the execution order of all subtasks, including:

[0020] By formula:

[0021] LLPM(M,E,G)={d1,d2,...,d n}

[0022] Get the execution order {d1,d2,...,d n};

[0023] Among them, LLPM() represents the planning model, M represents all execution models, E represents the task working environment of the complex task, G represents the task goal of the complex task, d1 represents the first subtask executed, d2 represents the second subtask executed, and d n Indicates the third subtask to be executed.

[0024] Optionally, each subtask is calibrated to obtain multiple final subtasks, including:

[0025] By formula:

[0026] LLCM(D,R C ,E)={c1,c2,...,c n}

[0027] Get the final set of subtasks {c1,c2,...,c n};

[0028] Among them, LLCM() represents the calibration model, D represents the calibration data, and R C represents the task calibration rule, E represents the task target error and deviation constraint, c1 represents the final subtask corresponding to the first subtask, c2 represents the final subtask corresponding to the second subtask, and c n Indicates the final subtask corresponding to the nth subtask.

[0029] Optionally, all execution models are used to execute the combined task to obtain the final execution result of the complex task, including:

[0030] By formula:

[0031] LLEM(D,R e ,R m )={d1,d2,...,d n}

[0032] Get the final execution result {d1,d2,...,d n};

[0033] Among them, LLEM() represents the combined model, D represents the input data, and R e represents the combination inference rule, R m represents the combined reasoning method, d1 represents the execution result of the final subtask executed in the first combined task, d2 represents the execution result of the final subtask executed in the second combined task, and d n Indicates the execution result of the final subtask executed in the nth combined task.

[0034] In a second aspect, an embodiment of the present application provides a task execution device based on a large language model, including:

[0035] The decomposition module uses the task decomposition model to decompose the complex task to be executed into multiple subtasks and build an execution model for each subtask; the task decomposition model is a large language model that can perform task decomposition;

[0036] The planning module uses the task planning model to plan all subtasks and obtain the execution order of all subtasks; the task planning model is a large language model that can perform subtask planning;

[0037] A calibration module calibrates each subtask to obtain multiple final subtasks, and then combines all the final subtasks in the execution order to obtain a combined task; the combined task includes all the planned final subtasks and the combined reasoning rules and combined reasoning methods between all the final subtasks, the combined reasoning rules are used to describe the order in which all the final subtasks are executed, and the combined reasoning method is a method for optimizing the combined reasoning rules;

[0038] The execution module uses all execution models to execute the combined task and obtain the final execution result of the complex task.

[0039] In a third aspect, an embodiment of the present application provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned task execution method based on a large language model when executing the above-mentioned computer program.

[0040] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, it implements the above-mentioned task execution method based on a large language model.

[0041] The above solution of the present application has the following beneficial effects:

[0042] In an embodiment of the present application, a complex task to be executed is decomposed into multiple subtasks by using a task decomposition model, and an execution model is constructed for each subtask, and then all subtasks are planned using a task planning model to obtain the execution order of all subtasks, and then each subtask is calibrated to obtain multiple final subtasks, and then all final subtasks are combined according to the execution order to obtain a combined task, and finally all execution models are used to execute the combined task to obtain the final execution result of the complex task. Among them, decomposing a complex task into multiple subtasks can convert a complex task with complexity into multiple simple subtasks, effectively reducing the difficulty of executing a complex task, planning all subtasks, so that the logic and rationality of executing all subtasks are improved, and all final subtasks with low execution difficulty are combined and executed based on a highly logical execution order, which can improve the rationality and versatility of the task execution method.

[0043] Other beneficial effects of the present application will be described in detail in the subsequent specific implementation section. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0045] Figure 1 A flowchart of a task execution method based on a large language model provided in one embodiment of the present application;

[0046] Figure 2 A schematic diagram of the structure of a task execution system based on a large language model provided in one embodiment of the present application;

[0047] Figure 3 A schematic diagram of a first final execution result provided in an embodiment of the present application;

[0048] Figure 4 A schematic diagram of a second final execution result provided in an embodiment of the present application;

[0049] Figure 5 A schematic diagram of a third final execution result provided in an embodiment of the present application;

[0050] Figure 6 A schematic diagram of the structure of a task execution device based on a large language model provided in one embodiment of the present application;

[0051] Figure 7 A schematic diagram of the structure of a terminal device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0052] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0053] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, wholes, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or combinations thereof.

[0054] It should also be understood that the term “and / or” used in the specification and appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0055] As used in the specification and appended claims of this application, the term "if" can be interpreted as "when" or "uponce" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "uponce it is determined" or "in response to determining" or "uponce [described condition or event] is detected" or "in response to detecting [described condition or event]", depending on the context.

[0056] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0057] References to "one embodiment" or "some embodiments" etc. described in the specification of this application mean that one or more embodiments of the present application include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0058] In view of the problem of poor universality of existing task execution methods, an embodiment of the present application provides a task execution method based on a large language model, which decomposes a complex task to be executed into multiple subtasks by using a task decomposition model, and constructs an execution model for each subtask, and then plans all subtasks using a task planning model to obtain the execution order of all subtasks, and then calibrates each subtask to obtain multiple final subtasks, and then combines all final subtasks according to the execution order to obtain a combined task, and finally uses all execution models to execute the combined task to obtain the final execution result of the complex task. Among them, decomposing a complex task into multiple subtasks can convert a complex task with complexity into multiple simple subtasks, effectively reducing the difficulty of executing a complex task, planning all subtasks, so that the logic and rationality of executing all subtasks are improved, and all final subtasks with low execution difficulty are combined and executed based on a highly logical execution order, which can improve the rationality and universality of the task execution method.

[0059] Next, an exemplary description is given of the task execution method based on a large language model provided in this application.

[0060] like Figure 1 As shown, the task execution method based on the large language model provided in this application includes the following steps:

[0061] Step 11: Decompose the complex task to be executed into multiple subtasks using the task decomposition model, and build an execution model for each subtask.

[0062] The above-mentioned task decomposition model is a large language model capable of performing task decomposition. The above-mentioned complex task is a task that needs to be executed and is input by a user, such as a task of drawing a distribution map of the elderly population.

[0063] The first step is to use the task decomposition model to decompose the complex task to be executed into multiple subtasks.

[0064] Specifically, through the formula:

[0065] LLDM(I)={t1,t2,...,t n}

[0066] Get multiple subtasks {t1, t2, ..., t n}.

[0067] Among them, LLDM() represents the task decomposition model, I represents the complex task, t1 represents the first subtask, t2 represents the second subtask, and t n Indicates the nth subtask, where n is the total number of subtasks.

[0068] Exemplarily, the large language model can understand the complex task description input by the user through natural language processing technology, and decompose it into multiple specific and executable subtasks. The complex task is input into the task decomposition model to obtain multiple subtasks of the complex task. For example, for the task of drawing the distribution map of the elderly population mentioned above, the task is input into the task decomposition model, and the multiple subtasks obtained are: "Subtask 1: Data collection + collecting geographic spatial data of the distribution of the elderly population (over 65 years old), including the elderly population data results", "Subtask 2: Data preprocessing + cleaning the collected data, including removing missing or abnormal data points", "Subtask 3: Spatial distribution analysis + analyzing the geographic spatial distribution of sampling points, including calculating point density and identifying hot spots", "Subtask 4: Map drawing + using geographic information system software or library to draw a distribution map of sampling points, marking different densities and hot spots".

[0069] The second step is to build an execution model for each subtask.

[0070] Specifically, through the formula:

[0071] LLMM={(t1,M1),(t2,M2),...,(t n ,M n )}

[0072] Get each subtask and execution model combination {(t1,M1),(t2,M2),...,(t n ,M n )}.

[0073] Among them, LLMM represents the modeling model, M1 represents the execution model of the first subtask, M2 represents the execution model of the second subtask, and M n Represents the execution model of the nth subtask.

[0074] Exemplarily, the modeling model is a large language model capable of model building. Based on the large language model's ability to understand natural language for task modeling, the large language model is enabled to understand the description of tools or execution models through designed comparative learning, thereby stimulating the large model's task modeling capabilities. Each subtask is input into the modeling model separately. The modeling model can match the task objectives and descriptions of the subtasks in a model space that stores a variety of models and tools (such as target detection models, drawing tools, and recognition models), and use the matched models and tools as the execution models of the subtasks. The execution model is a model or tool that can achieve the task purpose of the subtask. For example, for "Subtask 1: Data collection + collecting geographic spatial data on the distribution of the elderly population (over 65 years old), including the data results of the elderly population", the modeling model is matched in the model space according to the task goal and description of Subtask 1, and the execution model of Subtask 1 is a data collection tool (such as an online database); for "Subtask 2: Data preprocessing + cleaning the collected data, including removing missing or abnormal data points", the modeling model is matched in the model space according to the task goal and description of Subtask 2, and the execution model of Subtask 2 is a data collection tool (such as an online database). Data cleaning tools (such as data processing scripts); for "Subtask 3: Spatial distribution analysis + analysis of the geographic spatial distribution of sampling points, including calculation of point density and identification of hot spots", the model building model performs model matching in the model space according to the task objectives and description of subtask 3, and the execution model of subtask 3 obtained by matching is the spatial analysis model (such as geographic statistical methods); for "Subtask 4: Map drawing + use of geographic information system software or library to draw a distribution map of sampling points, marking different densities and hot spots", the model building model performs model matching in the model space according to the task objectives and description of subtask 4, and the execution model of subtask 4 obtained by matching is the geographic information system.

[0075] In some embodiments of the present application, since it is necessary to instruct a large language model, a complex task input by a user can exist in the form of a task instruction in an instruction space used to present the task objectives and content of the complex task. In the instruction space, the user's task instruction is in the form of natural language data. For example, for the task of drawing a distribution map of the elderly population, the task instruction in the instruction space is "There is a distribution map drawing task of the elderly population distribution that needs to be executed. In this task, the distribution of geographic spatial sampling points is calculated and drawn, and then task decomposition, task modeling, task planning, task calibration and task combination are continued. Finally, a distribution map of the elderly population distribution is drawn by combining the drawing tools. The result is described as + <explain this distribution map>. Please analyze this specific task step by step according to the analysis paradigm."

[0076] It is worth mentioning that decomposing a complex task into multiple subtasks can convert a complex task with complexity into multiple simple subtasks, effectively reducing the difficulty of executing the complex task.

[0077] Step 12: Plan all subtasks using the task planning model to obtain the execution order of all subtasks.

[0078] The above-mentioned task planning model is a large language model capable of performing subtask planning, such as the LLM+P model.

[0079] Specifically, through the formula:

[0080] LLPM(M,E,G)={d1,d2,...,d n}

[0081] Get the execution order {d1,d2,...,d n}.

[0082] Among them, LLPM() represents the planning model, M represents all execution models, E represents the task working environment of the complex task, G represents the task goal of the complex task, d1 represents the first subtask executed, d2 represents the second subtask executed, and d n Indicates the third subtask to be executed.

[0083] It should be noted that the execution models of all subtasks, the task working environment of complex tasks and the task objectives are input into the planning model to obtain the order of executing each execution model, and the subtasks corresponding to each execution model are planned according to the order to obtain the execution order. The above-mentioned task working environment is the constraint conditions that users comprehensively consider during task execution. These constraints may include resource constraints, time constraints, geographical location, external interference factors, etc., which are all environmental variables that users need to consider during task planning. For example, in the distribution map drawing task, the task working environment may include the map area corresponding to the distribution map preset by the user, the time corresponding to the obtained relevant data, etc.

[0084] The task target is the task execution result of a complex task. For example, for the distribution map drawing task mentioned above, the task target is the distribution map.

[0085] For example, the large language model automatically understands the task goal and the working environment to adaptively establish the connection between subtasks and form the execution order of subtasks. For example, for subtask 1, subtask 2, and subtask 3, the large language model believes that subtask 2 should be executed before subtask 1, and subtask 3 should be executed last, so the first subtask executed is subtask 2, the second subtask executed is subtask 1, and the third subtask executed is subtask 3.

[0086] In some embodiments of the present application, after obtaining the execution sequence, the execution sequence can be sent to the user's device, and the next step can be performed after receiving the confirmation instruction returned by the user's device. If the confirmation instruction returned by the user's device is not received, it means that the execution sequence does not meet the user's expectations and the user believes that the execution sequence is unreasonable, then the planning model adjusts the order of subtasks in the execution sequence.

[0087] It is worth mentioning that planning all subtasks improves the logic and rationality of executing all subtasks.

[0088] Step 13, calibrate each subtask to obtain multiple final subtasks, and then combine all the final subtasks in the execution order to obtain a combined task.

[0089] The above-mentioned combined task includes all the final subtasks of the plan and the combined reasoning rules and combined reasoning methods between all the final subtasks. The combined reasoning rules are used to describe the order of execution of all the final subtasks, determine the order, dependency, priority, etc. between the final subtasks, avoid conflicts and improve execution efficiency. The combined reasoning method is a method for optimizing the combined reasoning rules, such as logical reasoning, heuristic algorithms, user interaction scheduling methods, etc., to optimize the combined tasks and resource allocation, so that the combined tasks can be completed in an orderly, efficient and correct manner.

[0090] In the first step, each subtask is calibrated to obtain multiple final subtasks.

[0091] Specifically, through the formula:

[0092] LLCM(D,R C ,E)={c1,c2,...,c n}

[0093] Get the final set of subtasks {c1,c2,...,c n}.

[0094] Where LLCM() represents the calibration model, D represents the calibration data, and R C represents the task calibration rule, E represents the task target error and deviation constraint, c1 represents the final subtask corresponding to the first subtask, c2 represents the final subtask corresponding to the second subtask, and c n Indicates the final subtask corresponding to the nth subtask.

[0095] It should be noted that the calibration data is pre-stored data for calibration. The calibration data can be stored in a storage device such as a data space. Task calibration will establish a monitoring mechanism for the target error and deviation of task modeling and planning according to the task judgment rules and the rationality of task judgment to effectively eliminate the illusion problem caused by the large language model. Each subtask is input into the calibration model respectively, and the calibration model calibrates the subtask to obtain the final subtask. The calibration data is used to adjust and correct the standard data set or benchmark information output by the task model. These data are used to compare and calibrate the output of the model to ensure the accuracy and consistency of the task results. By comparing the task results with the calibration data, the errors generated by the model are identified and corrected. Ensure that the results at different stages or when different tasks are executed are consistent. Task calibration rules refer to specific methods or criteria for calibrating the task execution process according to specific standards and requirements. These rules guide how to use calibration data to correct the task and ensure that the task meets the expected goal after calibration. Provide clear processes and standards for the application of calibration data to ensure the effectiveness of calibration. Through clear rules, reduce the influence of subjective factors in the calibration process and ensure that the results are more reliable. Task target deviation refers to the difference or deviation between the actual result and the expected goal during the execution of the task. It is the core element of task calibration, which determines the accuracy of task execution by identifying and measuring deviations. By calculating the target deviation, the gap between the task results and expectations can be evaluated to understand the accuracy of task execution. After identifying the deviation, it is possible to decide what kind of calibration or adjustment is needed based on the size and nature of the deviation. Deviation constraints refer to the constraints set during the task calibration process to ensure that the final task results are within an acceptable range. These conditions specify the acceptable range of task target deviations, and adjustments need to be made when they exceed this range. By setting constraints, it is ensured that the error of the task results does not exceed the acceptable range. Providing clear goals for the calibration process makes the calibration work more targeted and effective.

[0096] For example, in the task of drawing a distribution map of the elderly population, it is assumed that the location of the elderly population distribution needs to be drawn in a specific area in order to optimize sampling efficiency and resource allocation. The following is an example of specific calibration data, task calibration rules, task target deviation, and deviation constraint conditions:

[0097] Calibration data: including actual geographic information data, historical sampling point location data, and regional population density data, etc. They are used to compare the sampling point distribution drawn by the model with the actual situation and correct the deviation.

[0098] Task calibration rules: such as calibrating the spatial position of sampling points according to actual geographic information, ensuring that sampling points do not fall in unreasonable locations (such as water areas, mountains, etc.), and calibrating the distribution density of sampling points according to population density.

[0099] Task target deviation: the distance difference between the location of the elderly population distribution points drawn by the model and the actual reasonable location, or the difference between the number of elderly population distribution points set by the model and the regional demand. The smaller the deviation, the more reasonable the elderly population distribution points.

[0100] Deviation constraint: It is stipulated that the actual position deviation of each elderly population distribution point shall not exceed 50 meters, or the distribution density deviation shall not exceed 5%, to ensure that the final elderly population distribution point is reasonable and feasible. If it exceeds these ranges, the model needs to be recalibrated.

[0101] Through these specific data, rules, deviations and constraints, the distribution map drawing task can be accurately calibrated to ensure that the final result meets actual needs and reduce unnecessary errors and waste of resources.

[0102] Task calibration can also be done through human-computer interaction. Human calibration is a process in which human experts or users intervene to review, adjust, and correct the results generated by the model. It combines human experience and judgment to make up for the shortcomings of machine automated calibration, ensuring that the task results are not only technically reasonable but also meet actual needs. Human calibration includes expert review, user feedback, and manual correction. For example, in the drawing of distribution maps, experts can review and adjust the positions of the elderly population distribution points, and user feedback can optimize the distribution of the points, thereby ensuring that the results are more accurate and practical.

[0103] Exemplarily, the calibration model is a large language model that can perform task calibration, such as a large language model with a self-refinement algorithm.

[0104] In the second step, all final subtasks are combined in the execution order to obtain a combined task.

[0105] Specifically, according to the planning of all subtasks in the execution order and the correspondence between the subtasks and the final subtasks, the final subtasks are combined in sequence according to the execution order, and the combination model gives the combination reasoning rules and combination reasoning methods. Specifically, the combination model is a large language model that can perform task combination, such as a large language model with algorithms such as the Tree of Thoughts (ToT) algorithm. The execution order and all final subtasks are input into the combination model, and the combination model plans the final subtask based on the execution order, and gives the combination reasoning rules and combination reasoning methods by analyzing the task objectives and descriptions of the final subtasks.

[0106] Exemplarily, the execution sequence is: subtask 2, subtask 1, subtask 3, then the order of the final subtasks in the combined task is final subtask 2, final subtask 1, final subtask 3. In the above nucleic acid sampling point distribution map drawing task, the combined reasoning rule and the combined reasoning method are respectively.

[0107] In the above-mentioned nucleic acid sampling point distribution map drawing task, the combined reasoning rule is the order of execution of the calibrated subtask 1, subtask 2, subtask 3, and subtask 4. For example, the tasks are executed in the order of the calibrated subtask 1, subtask 2, subtask 3, and subtask 4. The combined reasoning method may be a user-interactive scheduling method that optimizes the combined reasoning rule by interacting with the user so that the combined reasoning rule meets the user's expectations.

[0108] It is worth mentioning that by calibrating the subtasks, the accuracy and rationality of the subtasks can be improved and the deviation of the subtasks can be reduced.

[0109] Step 14, using all execution models to execute the combined task and obtain the final execution result of the complex task.

[0110] Specifically, through the formula:

[0111] LLEM(D,R e ,R m )={d1,d2,...,d n}

[0112] Get the final execution result {d1,d2,...,d n}.

[0113] Among them, LLEM() represents the combined model, D represents the input data, and R e represents the combination inference rule, R m represents the combined reasoning method, d1 represents the execution result of the final subtask executed in the first combined task, d2 represents the execution result of the final subtask executed in the second combined task, and d n Indicates the execution result of the final subtask executed in the nth combined task.

[0114] It should be noted that the above input data can be the relevant data of the complex task obtained by using a data acquisition tool (such as an online database), or it can be data provided by the user. When obtaining the final execution result, each final subtask is executed in sequence according to the final subtask in the combined task and the combined reasoning rules and combined reasoning methods, using the execution model corresponding to the final subtask.

[0115] In the interaction between users and the system, users can manually judge the rationality of the system. The large language model calls the model by matching the description of the task tool, and the user issues model execution commands through natural language to execute the model. Combinatorial reasoning rules and combined reasoning methods are input to the system through the interaction between users and the system, and the system automatically understands the combined reasoning rules and combined reasoning methods.

[0116] For example, the final execution result is the result after the complex task is executed. After the final execution result is obtained, the final execution result can be sent to the user's device. For example, if the complex task is to draw a distribution map of the elderly population in a certain area, the final execution result is as follows: Figure 3 The density map of the elderly population in a certain area based on simulated data and the corresponding explanation are shown. Figure 3 The explanation in the figure is: This is a scatter plot of the elderly population generated based on simulated data, showing the latitude and longitude and the density of the elderly population in different regions. The grayscale and size of each scatter point reflects the density of the elderly population at that location. Figure 3 The horizontal axis represents longitude, the vertical axis represents latitude, and the grayscale of the scatter points represents the percentage of the elderly population density in the region. Through this map, we can observe the distribution pattern of the elderly population in the region. This visualization tool is of great value for understanding the geographical distribution of the elderly population, guiding public health decisions and resource distribution.

[0117] It is worth mentioning that decomposing a complex task into multiple subtasks can convert a complex task with complexity into multiple simple subtasks, effectively reducing the difficulty of executing the complex task, planning all subtasks, and improving the logic and rationality of executing all subtasks. All final subtasks with low execution difficulty can be combined and executed based on a highly logical execution order, which can improve the rationality and versatility of the task execution method.

[0118] In some embodiments of the present application, the above steps can be integrated into the same system to obtain a task execution system based on a large language model. Figure 2As shown, it includes instruction space, task space, model space and data space. The instruction space is used to receive complex tasks and instructions of complex tasks sent by users, and contains task instructions of natural language description of complex tasks. This space aims to present task requirements in an intuitive and easy-to-understand natural language form, so that users and systems can clearly understand the task goals and contents. The task space is used to execute the above steps 11 to 14. The model space is used to store a variety of models and tools. The data space is used to store and manage open data and private data. It is responsible for collecting, storing, processing and managing various data required for task execution, ensuring the security and availability of data, and providing necessary data support for task execution. The specific structure of the task space includes task decomposition module, task modeling module, task planning module, task calibration module and task combination module. The task decomposition module is used to decompose complex tasks into multiple subtasks, the task modeling module is used to build an execution model for the subtasks, the task planning module is used to plan all subtasks, the task calibration model is used to calibrate each subtask, and the task combination module is used to combine all final subtasks.

[0119] For the above task execution system, since the models therein all adopt a large language model, the system prompt words can be set to enable the system to understand the task locally. For example, the system prompt words are: "You want to generalize all the tasks that users ask you to perform. The key steps of generalization include task decomposition, task modeling, task planning, task calibration and task combination. Your name is Geo-Jarvis, and you are an assistant for earth science research and geographical knowledge discovery. You can decompose a complex task into several subtasks. The description of the subtask is <subtask name> + <subtask description>. Please pay attention to the simplicity of the language. At the same time, when the user asks you to modify a subtask, you can modify the subtask. When the user asks you to complete a task, you should proactively ask the user whether the subtask decomposition is reasonable. Later, when the user asks you to complete the modeling of the subtask, you should model all the subtasks, and describe the modeling of the subtask as <subtask x> call + <submodel x>; when the user asks you to plan the task, you should reasonably consider the order of subtask execution and give the order of subtask execution. When the user asks you to calibrate the execution of the subtask, you should reflect on the order of subtask execution and ask the user to confirm whether the subtask planning and subtask calibration are correct. Finally, when the user asks you to complete the task combination to execute the task and obtain the execution result of the complex task, you should also give an understanding of the execution, and describe the result as <execution result> + <reason>. "

[0120] The application of the above task execution system is exemplified below with reference to specific examples.

[0121] The user needs to perform a spatial analysis distribution pattern analysis task for the elderly population. The instruction in the instruction space is "Assuming that you are a data expert with all the knowledge of geographic spatial analysis distribution pattern analysis, there is a spatial analysis distribution pattern analysis task for the elderly population that you need to perform. In this task, perform spatial analysis distribution pattern type and distribution pattern analysis, and then continue with task decomposition, task modeling, task planning, task calibration and task combination. Finally, use the drawing tools to draw the results of the spatial analysis distribution pattern analysis of the elderly population. The result description is <Distribution Pattern Analysis>+<Explanation of Distribution Pattern Analysis>. Please analyze this specific task step by step according to the analysis paradigm."

[0122] The task execution system breaks down this complex task into multiple subtasks: "Subtask 1: Data collection: Collect geospatial data related to the elderly population, including the distribution of the elderly population, population density, etc.", "Subtask 2: Data preprocessing: Clean and integrate data to ensure that the data format is suitable for spatial analysis", "Subtask 3: Distribution pattern type identification: Clean and integrate data to ensure that the data format is suitable for spatial analysis", "Subtask 4: Distribution pattern analysis: Use statistical and geographic information system tools to analyze the identified distribution patterns", "Subtask 5: Visualization: Use drawing tools to draw visualization graphics of the distribution pattern", "Subtask 6: Result interpretation: Based on the distribution pattern and the output of the analysis tool, explain the potential causes and impacts of the distribution pattern of the elderly population"

[0123] Then, an execution model is built for each subtask, and the results are as follows: subtask 1 calls the data collection model, subtask 2 calls the data cleaning model, subtask 3 calls the pattern recognition algorithm, subtask 4 calls the spatial analysis model, subtask 5 calls the GIS drawing tool, and subtask 6 calls the result interpretation model.

[0124] Then plan all subtasks, and the execution order is: subtask 1, subtask 2, subtask 3, subtask 4, subtask 5, subtask 6.

[0125] Then, subtask calibration is performed by asking users, and a message is sent to the user: "Please confirm whether the above decomposition and planning meet your expectations, or whether there are parts that need to be adjusted."

[0126] After the user confirms that the decomposition and planning are correct, the relevant data of the elderly population sent by the user is received, and all subtasks are executed in the execution order to obtain the following Figure 4 The final task result shown is Figure 4 The schematic diagram is a heat map of the distribution of the elderly population. The horizontal and vertical axes in the figure are the regional numbers. The darker the grayscale, the more elderly people there are in the area.

[0127] If the user needs to perform a task of detecting the type of marine ship using remote sensing images, the instruction is: "Assuming that you are an expert with all the knowledge of analyzing the remote sensing image data of marine ship targets, there is a task of detecting the type of marine ship using remote sensing images that you need to perform. In this task, the hull is disassembled into key sub-components according to the different ships themselves, and then the task is decomposed, modeled, planned, calibrated and combined. Finally, the type of unknown ship target is given. The result description is <execution result> + <reason>. Please analyze this specific task step by step according to your analysis paradigm."

[0128] The task is broken down into "Subtask 1: Image preprocessing: Clean, standardize and enhance remote sensing image data to improve the recognizability of ship targets", "Subtask 2: Disassembly of the hull into key sub-components: The process of decomposing the hull graphics into its key sub-components (such as hull structure, deck facilities, chimney features, etc.)", "Subtask 3: Feature extraction: Extract key features from each sub-component (such as shape, size, texture, etc.)", "Subtask 4: Classification model selection and training: Select a machine learning or deep learning model suitable for remote sensing image analysis and train it with labeled data", "Subtask 5: Ship type prediction: Input the extracted features into the trained model to predict the ship type", "Subtask 6: Result analysis and interpretation: Analyze the prediction results and provide detailed explanations about the identified ship types."

[0129] Build the execution model, the result is: subtask 1 calls the image processing submodel, subtask 2 calls the target detection submodel, subtask 3 calls the feature extraction submodel, subtask 4 calls the machine learning or deep learning training submodel, subtask 5 calls the classification prediction submodel, and subtask 6 calls the data analysis submodel.

[0130] The execution order is: subtask 1, subtask 2, subtask 3, subtask 4, subtask 5, subtask 6.

[0131] Then the task is calibrated, and the system considers that during execution, the task may need to be adjusted based on the quality of the preprocessed image, the accuracy of sub-component recognition, and the effectiveness of feature extraction. For example, if a sub-component is found to be difficult to recognize, it may be necessary to return to the image preprocessing stage for more specialized image enhancement.

[0132] The ship graphics sent by the user are received, and the tasks are combined and executed. The final execution result is the ship type and reason based on the extracted features and model predictions. These include: the hull is long and wide, many containers are stacked on the deck, there are specific cranes and loading machinery on the structure, and the layout of the hull and deck indicates that the purpose is to transport large-scale cargo. These features are consistent with typical container ships. This ship is a container ship because the long and wide hull, the deck structure for stacking containers, and the presence of loading machinery are all significant features of container ships.

[0133] If the user needs to perform a task of discovering unknown spatial relationships among the elderly population, the instructions are: Assuming that you are a data expert with all the knowledge of discovering unknown spatial relationships in geographic space, there is a task of discovering unknown spatial relationships among the elderly population that you need to perform. In this task, you need to perform spatial relationship types and unknown spatial relationship discoveries, and then continue with task decomposition, task modeling, task planning, task calibration and task combination. Finally, you need to use drawing tools to draw the analysis results of the unknown spatial relationship discovery among the elderly population. The result description is <Unknown spatial relationship> + <Explanation of unknown spatial relationship discovery>. Please analyze this specific task step by step according to your analysis paradigm.

[0134] The task is broken down into: Subtask 1: Data collection: Collect relevant data on the distribution of the elderly population, population density, etc.; Subtask 2: Data preprocessing: Clean data, handle missing values, and standardize data formats; Subtask 3: Spatial relationship type analysis: Determine the key spatial relationship types in the study, such as distance, adjacency, connectivity, etc.; Subtask 4: Spatial model detection: Use spatial statistical methods to detect the spatial distribution patterns of the elderly population; Subtask 5: Unknown spatial relationship mining: Use machine learning and pattern recognition techniques to mine unknown spatial relationships; Subtask 6: Result visualization: Use GIS tools to draw spatial relationship maps to highlight important findings.

[0135] The task modeling is as follows: subtask 1 calls the data collection interface, subtask 2 calls the data cleaning algorithm, subtask 3 calls the spatial analysis library, subtask 4 calls the spatial statistics tool, subtask 5 calls the machine learning model, and subtask 6 calls the GIS drawing tool.

[0136] The task planning is: Subtask 1, Subtask 2, Subtask 3, Subtask 4, Subtask 5, Subtask 6.

[0137] Task calibration, sent to the user: Please confirm whether the decomposition and execution order of the above subtasks are reasonable. We need to ensure that the logic of data processing is correct and the analysis method is suitable for detecting the spatial relationship of the elderly population.

[0138] Receive relevant data sent by the user, execute all subtasks in order, and combine the outputs of all subtasks to form Figure 5 The bar charts of the elderly population in several provinces are shown below. Figure 5 The horizontal axis of the chart represents provinces, and the vertical axis represents the number of elderly people, in units of 10,000 people. The explanation is: This is a bar chart of the number of elderly people in each province, from which we can clearly see the spatial quantitative relationship of the elderly population in each province. This kind of chart can help us intuitively understand the geographical distribution pattern of the elderly population and provide a basis for further spatial relationship analysis.

[0139] It can be seen from this that the task execution method of the present application can successfully complete the execution of a variety of complex tasks and has good versatility.

[0140] The following is an exemplary description of the task execution device based on the large language model provided in this application.

[0141] like Figure 6 As shown, the embodiment of the present application provides a task execution device based on a large language model, and the task execution device 600 based on a large language model includes:

[0142] Decomposition module 601, decomposing a complex task to be executed into multiple subtasks using a task decomposition model, and constructing an execution model for each subtask; the task decomposition model is a large language model capable of performing task decomposition;

[0143] Planning module 602, using the task planning model to plan all subtasks and obtain the execution order of all subtasks; the task planning model is a large language model capable of subtask planning;

[0144] The calibration module 603 calibrates each subtask to obtain multiple final subtasks, and then combines all the final subtasks in the execution order to obtain a combined task; the combined task includes all the planned final subtasks and the combined reasoning rules and combined reasoning methods between all the final subtasks. The combined reasoning rules are used to describe the order in which all the final subtasks are executed, and the combined reasoning method is a method for optimizing the combined reasoning rules;

[0145] The execution module 604 executes the combined task using all execution models to obtain the final execution result of the complex task.

[0146] It should be noted that the information interaction, execution process, etc. between the above-mentioned devices / units are based on the same concept as the method embodiment of the present application. Their specific functions and technical effects can be found in the method embodiment part and will not be repeated here.

[0147] The technicians in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0148] like Figure 7 As shown, an embodiment of the present application provides a terminal device. The terminal device D10 of this embodiment includes: at least one processor D100 ( Figure 7 Only one processor is shown in the figure), a memory D101, and a computer program D102 stored in the memory D101 and executable on the at least one processor D100, wherein the processor D100 implements the steps of any of the above-mentioned method embodiments when executing the computer program D102.

[0149] Specifically, when the processor D100 executes the computer program D102, the complex task to be executed is decomposed into multiple subtasks by using the task decomposition model, and an execution model is constructed for each subtask, and then all subtasks are planned by using the task planning model to obtain the execution order of all subtasks, and then each subtask is calibrated to obtain multiple final subtasks, and then all final subtasks are combined according to the execution order to obtain a combined task, and finally the combined task is executed using all execution models to obtain the final execution result of the complex task. Among them, decomposing a complex task into multiple subtasks can convert a complex task with complexity into multiple simple subtasks, effectively reducing the difficulty of executing a complex task, planning all subtasks, so that the logic and rationality of executing all subtasks are improved, and all final subtasks with low execution difficulty are combined and executed based on a highly logical execution order, which can improve the rationality and versatility of the task execution method.

[0150] The processor D100 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0151] In some embodiments, the memory D101 may be an internal storage unit of the terminal device D10, such as a hard disk or memory of the terminal device D10. In other embodiments, the memory D101 may also be an external storage device of the terminal device D10, such as a plug-in hard disk, a smart memory card (SMC, SmartMedia Card), a secure digital (SD, Secure Digital) card, a flash card (Flash Card), etc. equipped on the terminal device D10. Further, the memory D101 may also include both an internal storage unit of the terminal device D10 and an external storage device. The memory D101 is used to store an operating system, an application program, a boot loader (BootLoader), data and other programs, such as the program code of the computer program, etc. The memory D101 may also be used to temporarily store data that has been output or is to be output.

[0152] An embodiment of the present application further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in the above-mentioned method embodiments can be implemented.

[0153] An embodiment of the present application provides a computer program product. When the computer program product runs on a terminal device, the terminal device can implement the steps in the above-mentioned method embodiments when executing the computer program product.

[0154] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application implements all or part of the processes in the above-mentioned embodiment method, which can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may at least include: any entity or device, recording medium, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, RandomAccess Memory), electric carrier signal, telecommunication signal and software distribution medium that can carry the computer program code to the task execution method device / terminal device based on the large language model. For example, a USB flash drive, a mobile hard disk, a magnetic disk or an optical disk. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electric carrier signals and telecommunication signals.

[0155] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0156] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0157] The above is a preferred embodiment of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles described in the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A task execution method based on a large language model, characterized in that: include: Decomposing a complex task to be executed into multiple subtasks using a task decomposition model, and constructing an execution model for each of the subtasks; The task decomposition model is a large language model capable of performing task decomposition; Use the task planning model to plan all subtasks and obtain the execution order of all subtasks; The task planning model is a large language model capable of performing subtask planning; Calibrate each of the subtasks to obtain multiple final subtasks, and then combine all the final subtasks according to the execution order to obtain a combined task; the combined task includes all the planned final subtasks and the combined reasoning rules and combined reasoning methods between all the final subtasks, the combined reasoning rules are used to describe the order in which all the final subtasks are executed, and the combined reasoning method is a method for optimizing the combined reasoning rules; The combined task is executed using all execution models to obtain a final execution result of the complex task.

2. The task execution method according to claim 1, characterized in that: The complex task is decomposed into multiple subtasks using a task decomposition model, including: By formula: LLDM(I)={t1,t2,...,t n } Get multiple subtasks {t1, t2, ..., t n }; Wherein, LLDM() represents the task decomposition model, I represents the complex task, t1 represents the first subtask, t2 represents the second subtask, t n Indicates the nth subtask.

3. The task execution method according to claim 2, characterized in that: The step of constructing an execution model for each of the subtasks includes: By formula: LLMM={(t1,M1),(t2,M2),...,(t n ,M n )} Get each subtask and execution model combination {(t1,M1),(t2,M2),...,(t n ,M n )}; Wherein, LLMM represents the modeling model, M1 represents the execution model of the first subtask, M2 represents the execution model of the second subtask, and M n represents the execution model of the nth subtask, where n is the total number of subtasks.

4. The task execution method according to claim 1, characterized in that: The task planning model is used to plan all subtasks to obtain the execution order of all subtasks, including: By formula: LLPM(M,E,G)={d1,d2,...,d n } Get the execution order {d1,d2,...,d n }; Among them, LLPM() represents the planning model, M represents all execution models, E represents the task working environment of the complex task, G represents the task goal of the complex task, d1 represents the first subtask executed, d2 represents the second subtask executed, and d n Indicates the third subtask to be executed.

5. The task execution method according to claim 1, characterized in that: The step of calibrating each of the subtasks to obtain a plurality of final subtasks includes: By formula: LLCM(D,R C ,E)={c1,c2,...,c n } Get the final set of subtasks {c1,c2,...,c n }; Where LLCM() represents the calibration model, D represents the calibration data, and R C represents the task calibration rule, E represents the task target error and deviation constraint, c1 represents the final subtask corresponding to the first subtask, c2 represents the final subtask corresponding to the second subtask, and c n Indicates the final subtask corresponding to the nth subtask.

6. The task execution method according to claim 1, characterized in that: The utilizing all execution models to execute the combined task to obtain the final execution result of the complex task includes: By formula: LLEM(D,R e ,R m )={d1,d2,...,d n } Get the final execution result {d1,d2,...,d n }; Among them, LLEM() represents the combined model, D represents the input data, and R e represents the combination inference rule, R m represents the combined reasoning method, d1 represents the execution result of the final subtask executed by the first combined task, d2 represents the execution result of the final subtask executed by the second combined task, and d n Indicates the execution result of the final subtask executed in the nth combined task.

7. A task execution device based on a large language model, characterized in that: include: A decomposition module, which decomposes a complex task to be executed into multiple subtasks using a task decomposition model, and constructs an execution model for each of the subtasks; The task decomposition model is a large language model capable of performing task decomposition; The planning module uses the task planning model to plan all subtasks and obtain the execution order of all subtasks; The task planning model is a large language model capable of performing subtask planning; A calibration module calibrates each of the subtasks to obtain multiple final subtasks, and then combines all the final subtasks according to the execution order to obtain a combined task; the combined task includes all the planned final subtasks and the combined reasoning rules and combined reasoning methods between all the final subtasks, the combined reasoning rules are used to describe the order in which all the final subtasks are executed, and the combined reasoning method is a method for optimizing the combined reasoning rules; The execution module uses all execution models to execute the combined task and obtain the final execution result of the complex task.

8. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the task execution method based on the large language model is implemented as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the task execution method based on a large language model is implemented as described in any one of claims 1 to 6.

Citation Information

Patent Citations

  • Method and device for executing complex task by using large language model

    CN116594757A

  • Multi-unmanned aerial vehicle task planning method based on large language model

    CN117151338A

  • Long-term task planning method based on large language model

    CN118014246A

  • Task execution method and device based on large model

    CN118519680A

  • Task decomposition method and system based on large language model

    CN118520073A

Cited By

  • Task scheduling method and device, electronic equipment, readable storage medium and computer program product

    CN119759520A

  • Voice interactive console based on artificial intelligence

    CN120431919A

  • Driving task searching method and device and intelligent vehicle

    CN120763379A

  • Task processing method and device, equipment, storage medium and product

    CN121300986A

  • Loop task processing method and device, electronic equipment and storage medium

    CN121328748A