Task processing method, multi-agent system and electronic equipment

Abstract

The invention discloses a task processing method, a multi-agent system and electronic equipment, and the task processing method comprises the steps: determining a target task and first information through an information agent based on user input information; the first information represents at least one first task slot position corresponding to the target task and at least one piece of filling information corresponding to the at least one first task slot position; updating the first information by using an information agent based on the task processing information; the task processing information represents processing information related to the target task; and based on the updated first information, at least generating an execution instruction for executing a current sub-task in the target task by using a decision agent.

Description

Technical Field

[0001] This disclosure relates to, but is not limited to, the field of computer technology, and in particular to a task processing method, a multi-agent system, and an electronic device. Background Technology

[0002] An intelligent agent is an entity capable of perceiving its environment and making decisions and taking actions based on the perceived environmental information to achieve specific goals or tasks. Currently, intelligent agents are widely used in fields such as automated tasks, decision support, interactive systems, games, and intelligent transportation, providing great convenience to people's lives and work. Summary of the Invention

[0003] In view of this, this disclosure provides at least one task processing method, a multi-agent system, and an electronic device.

[0004] The technical solution disclosed herein is implemented as follows:

[0005] On the one hand, this disclosure provides a task processing method, the method comprising:

[0006] Based on user input information, an information agent is used to determine the target task and the first information; the first information represents at least one first task slot corresponding to the target task, and at least one filling information corresponding to the at least one first task slot.

[0007] Based on task processing information, the first information is updated using an information agent; the task processing information represents the processing information related to the target task.

[0008] Based on the updated first information, the decision-making agent generates at least one execution instruction for executing the current subtask in the target task.

[0009] In some implementations, task processing information indicates that the filling information of a second task slot in at least one first task slot is empty;

[0010] Based on task processing information, the information agent updates the first information, including:

[0011] In response to the fact that the filling information of the second task slot in at least one of the first task slots is empty, the information agent interacts with the user to determine the filling information of the second task slot.

[0012] Based on the filling information of the second task slot, the first information is updated using the information agent.

[0013] In some implementations, task processing information represents the execution result of the first historical subtask adjacent to the current subtask;

[0014] Based on task processing information, the first information is updated using an information intelligent agent, including: based on task processing information, the target task and the first information are updated using an information intelligent agent;

[0015] Based on task processing information, the information agent updates the target task and primary information, including:

[0016] Based on the execution result of the first historical subtask, the third task slot is determined using an information agent; the third task slot represents a task slot that is different from at least one first task slot.

[0017] By using an information agent to interact with the user, the filling information for the third task slot can be determined;

[0018] Based on the filling information of the third task slot, the information agent updates the subtask information after the first historical subtask in the target task and the first information.

[0019] In some implementations, task processing information represents the context information of the current subtask and conflict information between it and the context information of at least one previously executed historical subtask; the context information includes at least one of task execution perception information and execution action information.

[0020] Based on task processing information, the information agent updates the first information, including:

[0021] Based on conflict information, the information agent interacts with the user to determine the first feedback information;

[0022] Based on the first feedback information, the information agent updates the current subtask and the first information.

[0023] In some implementations, before updating the first information using the information agent based on task processing information, the method further includes:

[0024] The decision-making agent performs a consistency judgment on the context information of the current subtask, the context information of at least one historical subtask, and the first information to determine conflicting information.

[0025] In some implementations, a decision-making agent performs a consistency assessment on the context information of the current subtask, the context information of at least one historical subtask, and the first information to determine conflicting information, including:

[0026] Based on the contextual information of at least one second historical subtask, a reflective agent is used to generate summary information.

[0027] The decision-making agent uses the context information, summary information, context information of at least one third historical subtask, and first information to make a consistency judgment to determine conflicting information.

[0028] Among them, at least one second historical subtask represents multiple subtasks executed before at least one third historical subtask.

[0029] In some implementations, updating the first information using an information agent based on task processing information includes:

[0030] Based on the execution result of the first historical subtask adjacent to the current subtask, a reflection result is generated using a reflective agent; the reflection result includes a first reflection result and a second reflection result; the first reflection result is used to indicate whether the first historical subtask was executed; the second reflection result is used to indicate whether the execution result of the first historical subtask conforms to the first information;

[0031] In response to the reflection result representation that the first historical subtask was successfully executed, the first information is updated using the information agent based on the task processing information.

[0032] In some implementations, the current subtask represents an operation task targeting the application interface of the target application;

[0033] Based on the updated first information, the decision-making agent generates at least the following execution instructions for executing the current subtask in the target task:

[0034] Based on the updated first information and the current subtask, a decision agent is used to generate a call command for the target tool, so that the target tool can operate the application interface based on the call command.

[0035] On the other hand, this disclosure also provides a multi-agent system, including an information agent and a decision agent; wherein,

[0036] An information intelligent agent is used to determine a target task and first information based on user input information; and to update the first information based on task processing information; the first information represents at least one first task slot corresponding to the target task, and at least one filling information corresponding to the at least one first task slot; the task processing information represents task processing information related to the target task.

[0037] A decision-making agent is used to generate at least the execution instructions for the current subtask in the target task, based on the updated first information.

[0038] Furthermore, this disclosure also provides an electronic device, including at least one processor and a memory; wherein,

[0039] A memory for storing multiple computer instructions; at least one processor for loading and executing the computer instructions to run information agents and decision-making agents; wherein,

[0040] An information intelligent agent determines a target task and first information based on user input information; updates the first information based on task processing information; the first information represents at least one first task slot corresponding to the target task, and at least one filling information corresponding to the at least one first task slot; the task processing information represents task processing information related to the target task.

[0041] The decision-making agent generates execution instructions for the current subtask in the target task, based on the updated first information.

[0042] It should be understood that the above general description and the following detailed description are merely exemplary and explanatory, and are not intended to limit the technical solutions of this disclosure. Attached Figure Description

[0043] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the specification, serve to illustrate the technical solutions of this disclosure.

[0044] Figure 1 This is a schematic diagram of an embodiment of using an intelligent agent to perform user tasks in related technologies;

[0045] Figure 2 A schematic diagram illustrating the implementation flow of a task processing method provided in this disclosure;

[0046] Figure 3 A schematic diagram illustrating an embodiment of performing a food ordering task according to the task processing method provided in this disclosure;

[0047] Figure 4 This is a system block diagram of a multi-agent system provided in this disclosure;

[0048] Figure 5 This is a schematic diagram of the hardware entity of an electronic device provided in this disclosure. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this disclosure clearer, the technical solutions of this disclosure are further described in detail below with reference to the accompanying drawings and embodiments. The described embodiments should not be regarded as limitations on this disclosure. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0050] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0051] The terms “first / second / third” are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that “first / second / third” may be interchanged in a specific order or sequence where permitted, so that the embodiments of this disclosure described herein can be implemented in an order other than that illustrated or described herein.

[0052] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure belongs. The terminology used herein is for descriptive purposes only and is not intended to limit this disclosure.

[0053] In intelligent agent applications, the following three information dilemmas are often encountered. For example, consider the task of ordering food using a mobile agent with a multimodal large model:

[0054] The initial instructions are vague, meaning that the initial instructions issued by the user lack key information, making it difficult for the intelligent agent to accurately understand the user's intentions during task decomposition and execution.

[0055] Incremental task information refers to incremental tasks that may arise during task execution. For example, if a product needs to specify its specifications, the user may need to supplement incremental task information.

[0056] Information conflict occurs when there is a discrepancy between the task execution status and the user's task instructions; for example, the user's budget does not match the price of the goods.

[0057] For example, such as Figure 1 As shown, in the scenario where an intelligent agent is used to perform the task of ordering food, the task execution process is as follows:

[0058] First, user 110 inputs information 101 into agent 120; the content of input information 101 is "Order me a Starbucks coffee, budget 30 yuan";

[0059] Then, the intelligent agent 120 extracts information from the input information 101, understands the user's intent, makes decisions, reflects on the task execution process, and manages the context information during the task processing.

[0060] Afterwards, the intelligent agent 120 invokes a tool based on the decision result to operate the ordering application running on the mobile phone 130.

[0061] Finally, based on the perception results of the ordering application interface, when the agent 120 determines that the current task is to fill in the delivery address, it encounters an information dilemma because the input information 101 does not include delivery address information. In this situation, the agent 120 in the related technology has two possible execution strategies: one is to use the previously saved delivery address as the current delivery address, which may be an incorrect delivery address; the other is to output a prompt message indicating that the task execution failed because there is no delivery address.

[0062] Therefore, in order for intelligent agents to complete user tasks more accurately when faced with information dilemmas, related technologies clarify instructions and refine tasks by having multiple rounds of dialogue with the user before task execution.

[0063] However, the solutions proposed in the related technologies have the following drawbacks:

[0064] First, there is a lack of efficient dynamic information management mechanisms. Second, excessively long task context prompts may lead to the loss of critical information and hallucination problems.

[0065] Secondly, it failed to address the potential information conflict issues that may arise during the execution of automated tasks by intelligent agents.

[0066] If an agent cannot effectively handle information dilemmas, it will produce erroneous or deviate from the user's preferences or intentions in task execution, thereby reducing the agent's task execution efficiency and success rate.

[0067] Based on this, the task processing method provided in this disclosure firstly determines the target task and first information using an information agent based on user input information, wherein the first information represents at least one first task slot corresponding to the target task and at least one filling information corresponding to the at least one first task slot; then, based on task processing information, the first information is updated using an information agent, wherein the task processing information represents processing information related to the target task; finally, based on the updated first information, an execution instruction for executing the current subtask in the target task is generated using a decision agent. In this way, on the one hand, this disclosure utilizes a multi-agent system including at least an information agent and a decision agent to process user input information, with different agents performing different functions, realizing the specialized and division of labor among multiple agents to execute user tasks. Therefore, it can improve the output quality and task execution effect of the agents. Simultaneously, in the case of division of labor in task execution, the context information of different agents is relatively reduced, thereby avoiding the loss of key information and illusion problems caused by excessively long task context prompts; on the other hand, a slot mechanism is established, utilizing task slots (e.g., at least one first task slot) and slot information (e.g., at least one filling information). Using a structured storage method to store information related to the target task can help each agent in a multi-agent system quickly and accurately locate key information, avoiding ambiguity in the understanding of user input information between different agents. At the same time, task execution and / or interaction with the user can be performed based on the slot and slot information, which can improve the accuracy and efficiency of task execution. Furthermore, based on the task processing information related to the target task, the information agent can dynamically update the first information (i.e., at least one first task slot and its corresponding filling information), which can realize dynamic information management throughout the entire task cycle, thereby improving the accuracy of task execution and user satisfaction.

[0068] The method provided in this disclosure can be executed by an electronic device, which can be a laptop, tablet, desktop computer, set-top box, mobile device (e.g., mobile phone, portable music player, personal digital assistant, dedicated messaging device, portable gaming device), or a server. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0069] The technical solutions in the embodiments of this disclosure will now be clearly and completely described with reference to the accompanying drawings.

[0070] Figure 2 This is a schematic diagram illustrating the implementation flow of a task processing method provided in this disclosure, such as... Figure 2 As shown, the method includes the following steps S21 to S23:

[0071] Step S21: Based on user input information, use an information agent to determine the target task and the first information; the first information represents at least one first task slot corresponding to the target task, and at least one filling information corresponding to the at least one first task slot.

[0072] Here, input information refers to the initial instructions that the user inputs into the intelligent agent, that is, the user query.

[0073] In some implementations, the input information can be information entered by the user into the intelligent agent in any manner. In some implementations, the input information can be user data collected using at least one sensor in the electronic device, such as user voice collected using a microphone, user gestures or facial expressions collected using a camera, etc. In some implementations, the input information can be information entered by the user through an information input window provided by the intelligent agent.

[0074] In some implementations, the input information can be of any type. For example, the input information can be voice information, text information, or image information, etc.

[0075] An information agent is an intelligent agent in a multi-agent system that has information management functions. Information management functions refer to information acquisition, analysis, and processing, such as analyzing user input to determine corresponding user tasks and generating queries for user interaction.

[0076] In some implementations, the first model, serving as the brain of the information intelligent agent, can be implemented as any type of model, such as a language model, a visual-language model, a speech model, etc. In some implementations, the first model may include multiple models; for example, it may include a language model and a speech model, or it may include a visual-language model and a speech model, etc.

[0077] The target task refers to the user task generated by an information agent based on user input information. Here, the information agent can input the user input information into a first model to perform semantic analysis on the user input information, thereby determining the target task. In some implementations, the target task can be any type of task, such as multimodal perception tasks, language understanding tasks, path planning tasks, resource allocation tasks, strategy optimization tasks, cross-modal generation tasks, and various interactive tasks. In some implementations, the target task can be an operation task targeting the application interface of a target application, such as an operation task targeting the interface of a food ordering application or a control task targeting the interface of a video application.

[0078] The first information refers to information related to the target task determined by an information agent based on user input and the target task. The first information includes at least one first task slot and at least one filling information for that first task slot.

[0079] In this context, at least one first task slot refers to at least one information category determined by an information agent based on the type of the target task. For example, in the case of a food ordering task, at least one first task slot may include "dish name," "budget," "delivery address," and "recipient." Similarly, in a flight booking task, at least one first task slot may include "departure point," "destination," "date," and "passenger." In some implementations, the information agent can generate at least one first task slot corresponding to the target task using a first model; that is, it can determine at least one first task slot using the world knowledge of the first model. In some implementations, the information agent can query a specified database to determine at least one first task slot based on the type of the target task; wherein, the specified database is used to store the correspondence between different task types and task slots.

[0080] At least one piece of fill information refers to the fill information corresponding to at least one first task slot obtained by the information agent parsing the user input information according to at least one first task slot. For example, in a food ordering task, if the user inputs "Order me a Starbucks coffee, budget 30 yuan, delivered to the office", then after analyzing the user input information, the information agent can determine that the fill information corresponding to "dish name" in at least one first task slot is "Starbucks coffee", the fill information corresponding to "budget" is "30 yuan", the fill information corresponding to "delivery address" is "office", and the fill information corresponding to "recipient" is "me" (here, based on device ownership, the information agent can determine that "me" is the current device owner).

[0081] Step S22: Based on the task processing information, update the first information using an information agent; the task processing information represents the processing information related to the target task.

[0082] Here, task processing information can be any information related to the processing of the target task.

[0083] In some implementations, task processing information may be descriptive information about information dilemmas encountered during the processing of a target task, such as descriptive information about an unclear target task, descriptive information about incremental task information, or descriptive information about information conflicts.

[0084] In some implementations, task processing information can be the perception information of any agent in the multi-agent system regarding the task execution environment during the target task processing process, such as the perception information of the state of the target application being operated, the perception information of the physical environment, the perception information of the user's state, etc.

[0085] In some implementations, task processing information may be intermediate processing results of the target task, such as information on the success or failure of the execution of subtasks within the target task.

[0086] In this way, the information agent parses the task processing information and updates the first information based on the parsing results. In some embodiments, the information agent can update at least one first task slot in the first information based on the parsing results, for example, by adding or removing a first task slot. In some embodiments, the information agent can update at least one filler information in the first information based on the parsing results, for example, by modifying at least one filler information, and / or, in response to the addition or removal of a first task slot, by supplementing or deleting filler information.

[0087] In this way, given the task processing information, the first information is updated using the information agent based on the task processing information, so that the first task slot or filling information can be dynamically updated to adapt to various processing information that appears during the target task processing.

[0088] Step S23: Based on the updated first information, use the decision agent to generate at least one execution instruction for executing the current subtask in the target task.

[0089] A decision-making agent is an agent in a multi-agent system that has decision-making capabilities. Here, decision-making capability refers to the ability to determine action strategies based on user tasks and task context information to complete the user tasks.

[0090] In some implementations, the second model, which serves as the brain of the decision-making agent, can be implemented as any type of model, such as a rule-based decision model, a mathematical model-based decision model, a heuristic-based decision model, a machine learning-based decision model, a hybrid / cross-modal decision model, etc.

[0091] In some implementations, after determining the target task, the information agent can break down the target task into multiple subtasks, and then the decision agent can iteratively execute each subtask; wherein, the multiple subtasks can be executed sequentially or in parallel.

[0092] In this way, after updating the first information, the decision-making agent reads the current subtask and the updated first information, and inputs the read information into the second model to generate execution instructions for executing the current subtask.

[0093] As can be seen from the above, in the task processing method provided by this disclosure, firstly, based on user input information, an information agent is used to determine the target task and first information, wherein the first information represents at least one first task slot corresponding to the target task and at least one filling information corresponding to the at least one first task slot; then, based on task processing information, the information agent is used to update the first information, wherein the task processing information represents processing information related to the target task; finally, based on the updated first information, a decision agent is used to generate at least an execution instruction for executing the current subtask in the target task. Thus, on the one hand, this disclosure utilizes a multi-agent system including at least an information agent and a decision agent to process user input information, with different agents performing different functions, realizing the specialized and division of labor among multiple agents to execute user tasks. Therefore, it can improve the output quality and task execution effect of the agents. Simultaneously, in the case of division of labor in task execution, the context information of different agents is relatively reduced, thereby avoiding the loss of key information and illusion problems caused by excessively long task context prompts; on the other hand, a slot mechanism is established, utilizing task slots (e.g., at least one first task slot) and slot information (e.g., at least one filling information). Using a structured storage method to store information related to the target task can help each agent in a multi-agent system quickly and accurately locate key information, avoiding ambiguity in the understanding of user input information between different agents. At the same time, task execution and / or interaction with the user can be performed based on the slot and slot information, which can improve the accuracy and efficiency of task execution. Furthermore, based on the task processing information related to the target task, the information agent can dynamically update the first information (i.e., at least one first task slot and its corresponding filling information), which can realize dynamic information management throughout the entire task cycle, thereby improving the accuracy of task execution and user satisfaction.

[0094] In some implementations, the task processing information indicates that the filling information of the second task slot in the at least one first task slot is empty.

[0095] Here, the second task slot is a task slot that does not have corresponding filling information in at least one of the first task slots.

[0096] For example, if a user inputs "Order me a Starbucks coffee, budget 30 yuan," the information agent can determine multiple first task slots: "Dish Name," "Budget," "Delivery Address," and "Recipient." Simultaneously, based on the analysis of the user input, the agent can determine that the "Dish Name" should be "Starbucks coffee," the "Budget" should be "30 yuan," the "Recipient" should be "me," and the "Delivery Address" should be empty. Therefore, the information management agent will use the "Delivery Address" as the second task slot.

[0097] Since the filling information corresponding to each first task slot is the necessary information for executing the target task, the filling information of the second task slot is empty, indicating that the target task has an information dilemma of ambiguous initial instructions (i.e., user input information), which leads to the target task being unable to be executed accurately.

[0098] Thus, the step of updating the first information based on task processing information using an information agent, i.e., the above step S22, can be implemented as the following steps S221 to S222:

[0099] Step S221: In response to the fact that the filling information of the second task slot in the at least one first task slot is empty, the information agent interacts with the user to determine the filling information of the second task slot.

[0100] Here, when the filling information of the second task slot is empty, the information agent can proactively generate interaction information with the user based on the second task slot and output the interaction information to the user; then, based on the user feedback, the filling information of the second task slot is determined.

[0101] For example, in the above-mentioned ordering task, the information agent can generate various query messages such as "Please provide your delivery address." or "What is your delivery address?"

[0102] In some implementations, the information agent can generate a first prompt message based on a second task slot and a preset prompt message template; then, the first prompt message is input into a first model to generate interaction information with the user using the first model.

[0103] In some implementations, the information agent can input the second task slot into the first model; then, the first model combines a preset prompt information template to generate a first prompt information, and then generates interactive information with the user based on the first prompt information.

[0104] In some implementations, the information agent can determine the filling information of the second task slot through one or more rounds of interaction with the user.

[0105] Step S222: Update the first information using the information agent based on the filling information of the second task slot.

[0106] Here, after determining the filling information for the second task slot, the information agent updates the filling information to the second task slot in the first information.

[0107] In some implementations, while or after updating the first information, the information agent also updates the target task based on the filling information of the second task slot. For example, in the above-mentioned ordering task, based on the user's initial input information, the information agent can determine the target task as "order a Starbucks coffee, budget 30 yuan, recipient is the phone owner"; after obtaining the user's supplementary delivery address information, the information agent can update the target task to "order a Starbucks coffee, budget 30 yuan, recipient is the phone owner, delivery address is the phone owner's office".

[0108] In some implementations, when the information agent breaks down the target task into multiple subtasks, the information agent updates at least one subtask related to the second task slot among the multiple subtasks simultaneously or after updating the first information, based on the filling information of the second task slot.

[0109] In the embodiments provided in this disclosure, during the preprocessing stage for the target task, if there is a second task slot with empty information (i.e., the agent encounters an information dilemma with ambiguous initial instructions), the information agent can actively interact with the user and dynamically update the first information to clarify the user's intention, update the target task, and thus improve the task execution accuracy and user satisfaction.

[0110] In some implementations, the task processing information represents the execution result of a first historical subtask adjacent to the current subtask.

[0111] Here, the task processing information represents the execution result of the first historical subtask. The first historical subtask is a historical subtask adjacent to the current subtask; that is, a subtask that is adjacent to the current subtask and executed before the current subtask. The first historical subtask and the current subtask are executed sequentially.

[0112] In some implementations, the first historical subtask can be a single subtask or multiple subtasks executed in parallel.

[0113] In some implementations, the execution result of the first historical subtask can represent any type of result presented after the completion of the first historical subtask. In some implementations, when the first historical subtask is an operation task targeting the application interface of a target application, the execution result of the first historical subtask can represent the current application interface state of the target application. In some implementations, when the first historical subtask is a control task targeting an intelligent robot, the execution result of the first historical subtask can represent the current state information of the intelligent robot and its perception information of the surrounding environment.

[0114] Thus, the step of updating the first information using an information agent based on task processing information, i.e., step S22 above, can be implemented as the following step S223:

[0115] Step S223: Based on the task processing information, update the target task and the first information using the information agent.

[0116] Here, while updating the first information based on task processing information, the target task is also updated using an information agent so that the updated target task conforms to the user's true intention or matches the dynamically changing task processing information.

[0117] Specifically, step S223 can be implemented as follows: steps S2231 to S2233:

[0118] Step S2231: Based on the execution result of the first historical subtask, the information agent is used to determine the third task slot; the third task slot represents a task slot that is different from the at least one first task slot.

[0119] Here, after the first historical subtask is completed, the information agent performs incremental task analysis on its execution result to obtain the analysis result; if the analysis result indicates that there is no incremental task, then there is no need to determine the third task slot; if the analysis result indicates that there is an incremental task, then the third task slot is determined based on the incremental task.

[0120] In some implementations, when the target task is an application interface operation task for the target application, the information agent can determine whether there is an incremental task based on the application interface state corresponding to the first historical subtask. For example, it can determine whether there is a new control on the current application interface of the target application and whether the new control is related to the target task, thereby determining whether there is an incremental task.

[0121] In some implementations, when the target task is to operate the intelligent robot, the information agent can determine whether there is an incremental task based on the current environmental state of the intelligent robot after the first historical subtask has been completed. For example, it can determine whether there is a new control element in the current environment and whether the new control element is related to the target task, thereby determining whether there is an incremental task.

[0122] If an incremental task is determined to exist, the information agent can generate a corresponding task slot based on the incremental task, and take the task slot that is different from the first task slot in the first information as the third task slot.

[0123] For example, in the aforementioned ordering task, if the first historical subtask is "clicking the 'Select Specification' button" and the execution result of the first historical subtask is the product specification selection interface of the ordering application, the information agent, based on analysis of this interface, determines that it includes incremental tasks such as "cup type selection" and "temperature selection," even though the user did not provide information related to these tasks when specifying to purchase Starbucks coffee. In this case, the information agent can generate corresponding third task slots for the identified incremental tasks, such as "cup type" and "temperature."

[0124] Step S2232: The information agent interacts with the user to determine the filling information of the third task slot.

[0125] Here, since the third task slot is a different task slot from the already determined first task slot, the filling information for the third task slot is empty. Thus, in response to the empty filling information of the third task slot, the information agent can proactively generate interaction information with the user based on this third task slot and output this interaction information to the user; then, based on user feedback, it determines the filling information for the third task slot.

[0126] For example, in the above-mentioned ordering task, for the incremental task of "cup type selection", the information agent can generate diverse query information such as "Please provide cup type information." or "Which cup type of coffee would you like to order?"

[0127] In some implementations, the specific implementation method of the information agent interacting with the user based on the third task slot is the same as the specific implementation method of interacting with the user based on the second task slot. Therefore, you can refer to the detailed description of step S221 above.

[0128] Step S2233: Based on the filling information of the third task slot, use the information agent to update the subtask information after the first historical subtask in the target task and the first information.

[0129] Here, after determining the filling information of the third task slot, the information agent updates the third task slot and its filling information to the first information; at the same time, the information agent updates the subtask information in the target task after the first historical subtask based on the third task slot and its filling information.

[0130] For example, in the incremental task of the aforementioned ordering task, if the information agent determines, based on the results of user interaction, that the filling information for the third task slot "cup type" is "medium" and the filling information for "temperature" is "add ice," the information agent will update the first information with the cup type, temperature, and other information provided by the user as context information for the target task, while simultaneously updating the task information for the current subtask and subsequent subtasks. For example, if the current subtask is "select coffee sweetness," the updated current subtask will be "select coffee cup type, temperature, and sweetness."

[0131] In the embodiments provided in this disclosure, when incremental tasks are encountered during the execution of the target task, the information agent actively interacts with the user and updates the first information and sub-task information based on the user feedback. This effectively addresses the incremental task information that occurs during task execution and enables the dynamic evolution of the task, thereby improving the accuracy of task execution and user satisfaction. At the same time, updating the first information with user feedback information in the form of slots and filling information enables other agents in the multi-agent system to quickly and accurately determine user needs, further improving task execution efficiency and accuracy.

[0132] In some implementations, the task processing information represents the context information of the current subtask and conflict information between it and the context information of at least one previously executed historical subtask. The context information includes at least one of task execution perception information, execution action information, and execution result information.

[0133] Task execution perception information within the context information refers to any type of information perceived by the agent during or after task execution. In some implementations, task execution perception information may include sensor sampling information, multimodal perception information, etc. Sensor sampling information refers to any type of environmental information collected using at least one sensor in an electronic device, such as temperature information, time information, geographical location information, audio information, image information, etc.; multimodal perception information refers to information obtained by analyzing task-related target applications or pages using a multimodal model.

[0134] The execution action information in the context information refers to the actions taken during task execution and the purpose of those actions. In some implementations, the execution action information includes the tools invoked by the decision-making agent during task execution and the instructions for invoking those tools.

[0135] The execution result information in the context information is of the same type as the execution result information of the first historical subtask described above, and will not be repeated here.

[0136] In some implementations, the context information may further include reflective information regarding task execution perception information, execution action information, and / or execution result information. In some implementations, a reflective agent with reflective capabilities in a multi-agent system can be used to generate reflective information based on user input information, initial information, task execution perception information, execution action information, and / or execution result information.

[0137] In this way, when the decision-making agent executes the current subtask, it analyzes the context information of the current subtask and the context information of at least one historical subtask. If there is a conflict between the two, the information related to the conflict is taken as the conflict information.

[0138] For example, in a food ordering task, if the current subtask is "selecting a 'medium' cup", but based on its perception of the product specification selection interface, the decision-making agent determines that the control corresponding to "medium" is unselectable, then the decision-making agent considers that there is conflict information.

[0139] In the presence of conflicting information, the decision-making agent cannot continue to execute the current subtask entirely based on the user input or previously fed back information.

[0140] Thus, the step of updating the first information based on task processing information using an information agent, i.e., step S22 above, can be implemented as steps S224 to S225:

[0141] Step S224: Based on the conflict information, the information agent interacts with the user to determine the first feedback information.

[0142] Here, when conflicting information occurs, the decision-making agent sends the conflicting information to the information agent, so that the information agent can actively interact with the user and obtain the first feedback information from the user.

[0143] For example, in the ordering task where conflicting information appears, the information agent can generate diverse queries such as "Medium coffee is sold out, what other cup size can I choose?" and "Medium coffee is sold out, would you like to choose a large?"

[0144] In some implementations, the information agent can generate a second prompt based on conflict information and a preset prompt template; then, the second prompt is input into a first model to generate interaction information with the user.

[0145] In some implementations, the information agent can input conflict information into a first model; then, the first model, combined with a preset prompt information template, generates a second prompt information, and then generates interactive information with the user based on the second prompt information.

[0146] In some implementations, the information agent can determine the first feedback information through one or more rounds of interaction with the user.

[0147] Step S225: Based on the first feedback information, update the current subtask and the first information using the information agent.

[0148] Here, after receiving the user's first feedback, the information agent performs semantic understanding of the first feedback, determines the user's intent, and then updates the current subtask and the first information.

[0149] In some implementations, based on the first feedback information, the information agent can update the filling information of the first task slot corresponding to the conflicting information in the first information. For example, in the ordering task where conflicting information occurs, if the user's first feedback information is "select large", the information agent can modify the filling information of the task slot "cup type" to "large".

[0150] In some implementations, based on the first feedback information, the information agent can update the filling information of task slots other than the task slot corresponding to the conflicting information in the first information. For example, in the ordering task where conflicting information occurs, if the user's first feedback information is "change to a medium-sized vanilla coffee", the information agent will not update the filling information of the task slot "cup type", but will modify the task slot "coffee flavor" to "vanilla flavor" (for example, the coffee flavor before the update was "pandan flavor").

[0151] In some implementations, the information agent updates the content of the current subtask based on the first feedback information. For example, the current subtask may be updated from "select medium cup size" to "select large cup size," or from "select medium cup size" to "select vanilla coffee." Simultaneously, if the first feedback information is related to at least one subtask following the current subtask, at least one subtask following the current subtask is updated based on the first feedback information. For example, when updating the current subtask from "select medium cup size" to "select vanilla coffee," subsequent ordering subtasks need to be updated based on the updated first information, user input information, and prior multi-round user interaction information.

[0152] In some implementations, the information agent can determine to terminate the current subtask based on the first feedback information. For example, in the aforementioned ordering task with conflicting information, if the user's first feedback information is "cancel order," the information agent cancels the current subtask and subsequent subtasks based on this first feedback information.

[0153] In the embodiments provided in this disclosure, when conflicting information occurs during the execution of the target task (i.e., the agent encounters an information conflict dilemma), the information agent actively interacts with the user and updates the first information and the current sub-task based on the user's first feedback information. This effectively addresses the conflicting information that occurs during task execution and enables the dynamic evolution of the task, thereby improving the accuracy of task execution and user satisfaction. At the same time, updating the first feedback information to the first information in the form of slots and filling information enables other agents in the multi-agent system to quickly and accurately determine user needs, further improving task execution efficiency and accuracy.

[0154] In some implementations, before updating the first information using the information agent based on task processing information, that is, before step S22 above, the following step S24 is further included:

[0155] Step S24: Use the decision-making agent to perform a consistency judgment on the context information of the current subtask, the context information of at least one historical subtask, and the first information to determine the conflict information.

[0156] Here, before the decision-making agent executes the current subtask, the decision-making agent performs a consistency judgment on the context information of the current subtask, the context information of at least one historical subtask, and the first information to determine whether there is conflicting information.

[0157] In some implementations, the decision agent can generate a third prompt based on the context information of the current subtask, the context information of at least one historical subtask, the first information, and a preset prompt template; then, the decision agent inputs the third prompt into a second model to perform a consistency judgment and thereby determine the conflict information.

[0158] In some implementations, the decision-making agent can input the context information of the current subtask, the context information of at least one historical subtask, and the first information into the second model; then, the second model combines the first information with a preset prompt information template to generate the third prompt information, thereby performing a consistency judgment and determining conflict information based on the third prompt information.

[0159] In the embodiments provided in this disclosure, performing a consistency judgment before the decision-making agent executes the current sub-task can reduce invalid operations or operations that do not meet user expectations due to information conflicts, thereby improving the accuracy of task execution. In addition, using the decision-making agent to perform a consistency judgment, and then having the information agent actively interact with the user regarding the conflicting information, reflects the division of labor and cooperation among different agents in completing the user's task. This allows each agent to focus on the task it needs to perform and reduces the amount of contextual information that each agent needs to process, thereby improving the efficiency and accuracy of the agent's task execution.

[0160] In some implementations, the step of using the decision-making agent to perform a consistency judgment on the context information of the current subtask, the context information of the at least one historical subtask, and the first information to determine the conflict information, i.e., the above step S24 can be implemented as the following steps S241 to S242:

[0161] Step S241: Based on the contextual information of at least one second historical subtask, generate summary information using a reflective agent.

[0162] Here, the second historical subtask refers to the subtask executed in the target task before the current subtask.

[0163] A reflective agent is an agent in a multi-agent system that performs reflective functions. In some implementations, the third model, which serves as the brain of the reflective agent, can be any type of model, such as a large language model, a reinforcement learning model, or a memory enhancement model.

[0164] Generating summary information corresponding to the context information of at least one second historical subtask using a reflective agent refers to using a reflective agent to analyze various types of information in the context information of at least one second historical subtask, as well as the decision logic when executing each second historical subtask, thereby generating structured summary information. In some embodiments, the summary information includes at least the task objective of each second historical subtask and whether it was successfully executed.

[0165] Thus, during implementation, the reflective agent inputs the contextual information of at least one second historical subtask into the third model to generate summary information using the third model.

[0166] In some implementations, this summary information is also used to guide the decision-making agent in generating execution instructions for the current subtask.

[0167] Step S242: The decision-making agent performs a consistency judgment on the context information of the current subtask, the summary information, the context information of at least one third historical subtask, and the first information to determine the conflict information.

[0168] The at least one second historical subtask represents a plurality of subtasks executed prior to the at least one third historical subtask.

[0169] Here, the third historical subtask refers to the subtask executed in the target task before the current subtask.

[0170] At least one second historical subtask represents multiple subtasks executed before at least one third historical subtask; that is, each second historical subtask is a subtask executed earlier than each third historical subtask. For example, at least one third historical subtask may refer to a target number of historical subtasks that are adjacent to the current subtask and executed earlier; at the same time, all historical subtasks executed before the target number of historical subtasks in the target task are considered as second historical subtasks.

[0171] Thus, in some implementations, a decision-making agent is used to concatenate the context information of the current subtask, the summary information of at least one second historical subtask, the context information of at least one third historical subtask, and the first information. Then, a consistency judgment is made on the concatenation result to determine the conflict information between the context information of the current subtask and the context information of multiple historical subtasks and the first information.

[0172] In the embodiments provided in this disclosure, when determining conflict information, the reflective agent performs summary processing on historical subtasks whose execution order is far from the current subtask. This can reduce the amount of information input to the decision agent during consistency judgment, thereby reducing the impact of redundant information on the decision-making process of the decision agent and improving the data processing speed and the accuracy of the decision agent in judging conflict information.

[0173] In some implementations, updating the first information based on task processing information using an information agent, i.e., step S22 above, can be implemented as steps S225 to S226:

[0174] Step S225: Based on the execution result of the first historical subtask adjacent to the current subtask, a reflection result is generated using a reflection agent; the reflection result includes a first reflection result and a second reflection result; the first reflection result is used to indicate whether the first historical subtask was executed; the second reflection result is used to indicate whether the execution result of the first historical subtask conforms to the first information.

[0175] The specific implementation method of the execution result of the first historical subtask is the same as the specific description above, and will not be repeated here.

[0176] Here, a reflective agent is used to reflect on the execution results of the first historical subtask, including reflection on the execution results at the task level and information level, and generates the first reflection result and the second reflection result respectively.

[0177] In some implementations, the reflective agent can determine the first reflection result in any way.

[0178] In some implementations, the reflective agent can determine the first reflection result based on its perception of the state of the operation object corresponding to the first historical subtask. For example, when the operation object of the first historical subtask is the application interface of the target application, the reflective agent can determine the first reflection result by perceiving whether the application interface of the target application has changed. That is, if it has changed, the first reflection result indicates that the first historical subtask has been executed; otherwise, the first reflection result indicates that the first historical subtask has not been executed.

[0179] In some implementations, the reflective agent can determine a first reflection result based on its perception of the environment in which the manipulated object is located. For example, if the first historical subtask is a path control task for an intelligent robot, the reflective agent can determine the first reflection result based on environmental information collected by at least one sensor to see if the current position of the intelligent robot has changed. That is, if the position has changed, the first reflection result indicates that the first historical subtask has been executed; otherwise, the first reflection result indicates that the first historical subtask has not been executed.

[0180] In some implementations, the reflective agent can determine the second reflection result in any way.

[0181] In some implementations, the reflective agent can determine the second reflection result based on the state of the operation object corresponding to the first historical subtask. For example, when the operation object corresponding to the first historical subtask is the application interface of the target application, the reflective agent can determine the first reflection result by judging whether the page elements or user interface (UI) components in the current application interface of the target application match the first information. That is, if they match, the second reflection result indicates that the execution result of the first historical subtask is consistent with the first information; otherwise, the second reflection result indicates that the execution result of the first historical subtask is inconsistent with the first information. For example, if the first historical subtask is "fill in the recipient information", by sensing the current application interface of the ordering application, if the reflective agent determines that the recipient information filled in the current application interface matches the filling information of the task slot "recipient" in the first information, then the reflective agent uses the second reflection result to indicate that the execution result of the first historical subtask is consistent with the first information.

[0182] In some implementations, the reflective agent can determine the second reflection result by sensing environmental information from the electronic device. For example, if the first historical subtask is a path control task for an intelligent robot, the reflective agent can determine the second reflection result by judging whether the environmental information collected by at least one sensor installed in the intelligent robot matches the environmental information stored in the task slot in the first information. That is, if they match, the second reflection result indicates that the execution result of the first historical subtask is consistent with the first information; otherwise, the second reflection result indicates that the execution result of the first historical subtask is inconsistent with the first information.

[0183] Step S226: In response to the reflection result indicating that the first historical subtask was successfully executed, the first information is updated using the information agent based on the task processing information.

[0184] The reflection result indicates that the first historical subtask was successfully executed, including a first reflection result indicating that the first historical subtask has been executed, and a second reflection result indicating that the execution result of the first historical subtask is consistent with the first information.

[0185] Here, if the first historical subtask is successfully executed, the decision agent reads and executes the current subtask. During the execution of the current subtask, the first information is updated based on the task processing information and the information agent, thereby realizing the iterative execution of multiple subtasks in the target task.

[0186] In the embodiments provided in this disclosure, the reflective agent performs task-level and information-level reflections on the execution results of the first historical subtask, which can improve the accuracy of verifying the task execution results. On the other hand, by starting the execution of the current subtask after determining that the first historical subtask has been successfully executed, task execution deviations or errors can be detected and corrected in a timely manner, thereby increasing the probability that the target task will be successfully executed.

[0187] In some implementations, the current subtask represents an operation task targeting the application interface of the target application.

[0188] The target application can be any type of application running on an electronic device, such as shopping apps, food ordering apps, ticketing apps, navigation apps, etc.

[0189] The target application has multiple application interfaces. By operating these multiple application interfaces according to the specified operation path, the corresponding application functions can be realized.

[0190] In some implementations, the current subtask can be any type of operation task targeting the application interface, such as a task to click on a specified location or control in the application interface, a task to slide on the application interface, an information input task, or other operation tasks that simulate user interaction.

[0191] Thus, based on the updated first information, the decision-making agent generates at least one execution instruction for executing the current subtask in the target task. That is, step S23 above can be implemented as the following step S231:

[0192] Step S231: Based on the updated first information and the current subtask, the decision agent generates a call instruction for the target tool, so that the target tool operates the application interface based on the call instruction.

[0193] Here, "tools" refers to various components or services that help decision-making agents complete specific tasks and extend their functionality. In some implementations, tools can be implemented as Application Programming Interfaces (APIs), function libraries, plugins, etc. In some implementations, the tools that the decision-making agent can invoke can be data query tools, computation and analysis tools, automated task tools, language processing tools, speech and vision tools, etc.

[0194] The target tool refers to the tool selected by the decision-making agent from a pool of available tools. In some implementations, the decision-making agent can dynamically determine the target tool based on factors such as the type of the current subtask, the capabilities of each available tool, and the current state of the target application's interface.

[0195] Different tools correspond to different invocation mechanisms. Therefore, after determining the target tool, the decision-making agent can generate invocation instructions through the invocation mechanism corresponding to the target tool. Then, the target tool can operate the application interface of the target application based on the invocation instructions to complete the execution of the current subtask.

[0196] In some implementations, the task processing method provided in this disclosure can be used to construct a multi-agent collaborative dynamic information management framework (Dynamic Information Architecture, DynamicIA). Below, in conjunction with... Figure 3 An embodiment of executing a food ordering task using this dynamic information management framework is described below; wherein, the dynamic information management framework includes an information agent 320, a decision agent 330, a reflective agent 340, and a memory module 350. This embodiment includes the following steps S301 to S3012: wherein steps S301 to S305 are the preprocessing stage of the food ordering task, and steps S306 to S3012 are the iterative execution stage for multiple sub-tasks in the food ordering task.

[0197] Step 301: User 310 inputs information into information agent 320; the content of the input information is "Order me a Starbucks coffee, budget 30 yuan"; then, step S302 is executed.

[0198] Step 302: The intelligent agent 320 understands and analyzes the input information and generates an ordering task, namely, "Order a Starbucks coffee, budget 30 yuan, recipient is the phone owner"; it generates multiple task slots, namely, "Dish Name", "Budget", "Recipient", and "Delivery Address"; based on the analysis of the input information, it determines that the information to be filled in for "Dish Name" is "Starbucks coffee", the information to be filled in for "Budget" is "30 yuan", the information to be filled in for "Recipient" is "Phone Owner", and the information to be filled in for "Delivery Address" is empty; then, step S303 is executed;

[0199] In step S303, based on the task slot "delivery address", the information agent 320 generates interactive information for the user such as "Please fill in the delivery address." or "What is the delivery address?", and sends it to the user 310; then, step S304 is executed.

[0200] In step S304, the information agent 320 receives feedback information from the user 310 and determines that the information to be filled in for the "delivery address" is "owner's office"; accordingly, the information agent 320 updates the ordering task to "order a Starbucks coffee, budget 30 yuan, recipient is the owner, delivery address is the owner's office"; then, step S305 is executed.

[0201] In step S305, the information agent 320 splits the updated ordering task into multiple sub-tasks and stores the user's input information, the updated ordering task, the split sub-tasks, the generated task slots and slot information into the memory module 350.

[0202] Thus, at step S305, the preprocessing stage of the ordering task ends.

[0203] After determining the accurate ordering task and its multiple subtasks, multiple task slots and their filling information, the multi-agent system iteratively executes the multiple subtasks according to their execution order. The following describes the process of the multi-agent system collaboratively executing one of the target subtasks using steps S306 to S3012:

[0204] Step S306: Information agent 320 reads the current application interface of the ordering application from memory module 350 and analyzes the interface to determine whether there is an incremental task; if there is, proceed to step S307; if not, proceed to step S308.

[0205] In some implementations, after the previous subtask of the target subtask has been completed, the decision agent 330 uses a tool to take a screenshot of the current application interface of the ordering application and stores the screenshot in the memory module 350. In this way, the information agent 320, the decision agent 330, and the reflective agent 340 can read the screenshot from the memory module 350 to perceive the current application interface state of the ordering application.

[0206] Here, the information agent 320 determines whether there is an incremental task by analyzing whether there is a newly added control in the current application interface of the ordering application, and whether the newly added control is related to the ordering task. For example, if the information agent 320 determines that a "cup shape" control appears in the current application interface of the ordering application, then the information agent 320 determines "select cup shape" as an incremental task and generates a new task slot "cup shape" accordingly.

[0207] Step S307: Based on the incremental task, the information agent 320 generates interaction information with the user and sends it to the user; in response to the user's feedback information, the information agent 320 updates the target subtask and its subsequent subtask information.

[0208] Here, the information agent 320 generates interactive information with the user based on the new task slot "cup type", such as "Which cup type would you like to choose?". In response to the user's feedback "Choose medium cup", the information agent 320 fills the task slot "cup type" with "medium cup" and updates the target subtask and its subsequent subtask information.

[0209] After updating the task slots, target subtasks, and their subsequent subtask information, the information agent 320 stores the updated information in the memory module 350.

[0210] In step S308, the decision-making agent 330 reads the target subtask, the context information of the target subtask, the context information of the n historical subtasks closest to the target subtask, and the summary information of the m historical subtasks farther away from the target subtask from the memory module 350 to determine whether there is conflict information; if there is, proceed to step S309; ​​if there is no conflict, proceed to step S3010.

[0211] Here, m historical subtasks are historical tasks of n historical subtasks; m and n are both integers greater than or equal to 1.

[0212] In some implementations, the reflective agent 340 performs summary processing on the context information of m historical subtasks to obtain corresponding summary information, and stores the summary information in the memory module 350.

[0213] In some implementations, the decision agent reads long-term memory information and preference information related to the user 310 from the memory module 350, and then determines conflict information based on the long-term memory information, preference information, context information of the target subtask, context information of n historical subtasks, and summary information of m historical subtasks.

[0214] In step S309, the decision-making agent 330 stores the conflict information in the memory module 350; the information agent 320 reads the conflict information from the memory module 350, and then actively interacts with the user based on the conflict information to obtain the corresponding feedback information; based on the feedback information, the information agent 320 updates the task slots and their filling information, updates the target subtasks and their subsequent subtask information, and stores the updated information in the memory module 350.

[0215] For example, if the target subtask is "select medium", the conflict information is that the user selected "medium", but the current ordering application interface shows that medium sizes are sold out. Based on this, the information agent 320 generates the interaction information "medium sizes are sold out, would you like to change to large sizes?" with the user 310. Thus, if the user 310 responds by changing to large sizes, the information agent 320 updates the task slot "cup type" to "large" and updates the target subtask to "select large".

[0216] In step S3010, the decision-making agent 330 determines the execution instruction corresponding to the target sub-task and generates a call instruction for the target tool 360 to operate the application interface of the ordering application running in the electronic device 370 using the target tool 360; after performing the operation on the application interface of the ordering application, the target tool 360 takes a screenshot of the application interface of the ordering application and stores the screenshot in the memory module 350; then, step S3011 is executed.

[0217] Step S3011: The reflecting agent 340 reads the application interface screenshot, target subtask, multiple task slots and their filling information after executing the target subtask from the memory module 350, so as to reflect on the execution result of the target subtask and store the reflection result in the memory module 350; then step S3012 is executed.

[0218] In step S3012, if the reflection result indicates that the target subtask has been successfully executed, the next subtask of the current target subtask is updated to the new target subtask, and steps S306 to S3012 are repeated; if the reflection result indicates that the target subtask has not been successfully executed, steps S308 to S3010 are repeated based on the current target subtask.

[0219] As can be seen from the above, in the embodiments provided in this disclosure, on the one hand, multiple agents in the multi-agent system can use the memory module as a communication medium and actively interact with the user by calling tools to realize an iterative task execution process of "information management - action decision-making - reflection," thereby improving the efficiency and accuracy of device operation and task execution. On the other hand, the task execution process is divided into a preprocessing stage and an iterative task execution stage. In the task preprocessing stage, the information agent performs fuzzy analysis on the initial instructions and interacts with the user based on the dynamic task slot generation and updating method to clarify the instructions. In each round of task iteration, the information agent analyzes the incremental task, dynamically generates and updates the task slots corresponding to the incremental task, thereby realizing the dynamic evolution of the task. In addition, the decision agent judges whether there is conflicting information when making decisions and collaborates with the information agent to resolve the conflicting information problem. At the same time, the reflection agent reflects on the execution results to verify whether the task has been successfully executed. It can be seen that in this dynamic information management framework, three information dilemmas can be addressed through the dynamic evolution of the entire task cycle, and task execution results that better meet user expectations are obtained through active interaction with the user.

[0220] The following section evaluates the experimental results to demonstrate the performance of this dynamic information management framework in real-world task scenarios.

[0221] In this evaluation experiment, four complex task scenarios were selected, including hotel booking, flight booking, food delivery ordering, and online shopping. Chinese and English applications were selected for testing, and task instructions with ambiguity and information dilemmas during execution were constructed.

[0222] Table 1 shows some of the fuzzy task instructions:

[0223]

[0224]

[0225] Table 1, Examples of Fuzzy Commands

[0226] Furthermore, in the evaluation experiments, GPT-4 (Generative Pre-trained Transformer 4Omni) was used as the backbone network for each agent in the dynamic information management framework. Simultaneously, during the evaluation experiments, the same application operational knowledge was injected into the first agent, the second agent, and the dynamic information management framework—serving as comparison objects—to guide the navigation decisions within the application.

[0227] The evaluation results are shown in Table 2. As can be seen from Table 2, in the evaluation of indicators at both the task level and the information level, the dynamic information management framework significantly outperforms the first and second intelligent agents in all indicators.

[0228]

[0229] Table 2, Evaluation of Experimental Results

[0230] Based on the above embodiments, this disclosure also provides a multi-agent system. For example... Figure 4 As shown, the multi-agent system 400 includes an information agent 410 and a decision agent 420; wherein,

[0231] The information agent 410 is used to determine a target task and first information based on user input information; and to update the first information based on task processing information; the first information represents at least one first task slot corresponding to the target task, and at least one filling information corresponding to the at least one first task slot; the task processing information represents task processing information related to the target task.

[0232] The decision-making agent 420 is used to generate at least one execution instruction for executing the current subtask in the target task based on the updated first information.

[0233] In some embodiments, the task processing information indicates that the filling information of the second task slot in the at least one first task slot is empty; the information agent 410 is used for:

[0234] In response to the fact that the fill information of the second task slot in the at least one first task slot is empty, the system interacts with the user to determine the fill information of the second task slot.

[0235] Update the first information based on the filling information of the second task slot.

[0236] In some implementations, the task processing information represents the execution result of a first historical subtask adjacent to the current subtask;

[0237] The information agent 410 is used to update the target task and the first information based on task processing information;

[0238] Specifically, the information agent 410 is used for:

[0239] Based on the execution result of the first historical subtask, a third task slot is determined; the third task slot represents a task slot that is different from the at least one first task slot.

[0240] Interact with the user to determine the filling information for the third task slot;

[0241] Based on the filling information of the third task slot, update the subtask information after the first historical subtask in the target task and the first information.

[0242] In some implementations, the task processing information represents the context information of the current subtask and conflict information between it and the context information of at least one previously executed historical subtask; the context information includes at least one of task execution perception information and execution action information.

[0243] The information agent 410 is used for:

[0244] Based on the conflict information, interact with the user to determine the first feedback information;

[0245] Based on the first feedback information, update the current subtask and the first information.

[0246] In some implementations, the decision agent 420 is used to perform a consistency judgment on the context information of the current subtask, the context information of the at least one historical subtask, and the first information to determine the conflict information.

[0247] In some implementations, the multi-agent system further includes a reflective agent;

[0248] The reflexive agent is used to generate summary information based on contextual information from at least one second historical subtask.

[0249] The decision-making agent 420 is used to perform a consistency judgment on the context information of the current subtask, the summary information, the context information of at least one third historical subtask, and the first information pair to determine the conflict information;

[0250] The at least one second historical subtask represents a plurality of subtasks executed prior to the at least one third historical subtask.

[0251] In some implementations, the multi-agent system further includes a reflective agent;

[0252] The reflective agent is used to generate a reflection result based on the execution result of a first historical subtask adjacent to the current subtask; the reflection result includes a first reflection result and a second reflection result; the first reflection result is used to indicate whether the first historical subtask was executed; the second reflection result is used to indicate whether the execution result of the first historical subtask conforms to the first information;

[0253] The information agent 410 is used to update the first information based on the task processing information in response to the reflection result indicating that the first historical subtask was successfully executed.

[0254] In some implementations, the current subtask represents an operation task targeting the application interface of the target application;

[0255] The decision-making agent 420 is used to generate a call instruction for the target tool based on the updated first information and the current subtask, so that the target tool can operate the application interface based on the call instruction.

[0256] The descriptions of the system embodiments above are similar to those of the method embodiments above, and have similar beneficial effects. In some embodiments, the functions or units included in the system provided by this disclosure can be used to perform the methods described in the method embodiments above. For technical details not disclosed in the system embodiments of this disclosure, please refer to the descriptions of the method embodiments of this disclosure for understanding.

[0257] Based on the foregoing embodiments, this disclosure also provides an electronic device. For example... Figure 5 As shown, the electronic device 500 includes at least one processor 510 and a memory 520; wherein,

[0258] The memory 520 is used to store multiple computer instructions; the at least one processor 510 is used to load and execute the computer instructions to run an information agent and a decision agent; wherein,

[0259] The information agent determines a target task and first information based on user input information; updates the first information based on task processing information; the first information represents at least one first task slot corresponding to the target task, and at least one filling information corresponding to the at least one first task slot; the task processing information represents task processing information related to the target task.

[0260] The decision-making agent generates an execution instruction for executing the current subtask in the target task based on the updated first information.

[0261] In some embodiments, the electronic device 500 further includes a reflective agent; wherein the reflective agent is used to reflect on the task execution result corresponding to the current subtask.

[0262] In some embodiments, when the at least one processor loads the computer instructions, it can run the information agent, the decision agent, and the reflective agent, and the actions and functions that the information agent, the decision agent, and the reflective agent can perform correspond to the information agent, the decision agent, and the reflective agent in the above-mentioned multi-agent system, respectively, which will not be elaborated here.

[0263] It should be noted that the descriptions of the above device embodiments are similar to those of the above method and system embodiments, and have similar beneficial effects. In some embodiments, the functions or included units of the electronic device provided in this disclosure can be used to perform the methods described in the above method and system embodiments. For technical details not disclosed in the device embodiments of this disclosure, please refer to the descriptions of the method and system embodiments of this disclosure for understanding.

[0264] If the technical solution disclosed herein involves personal information, the product using this technical solution has clearly informed the user of the personal information processing rules and obtained the user's voluntary consent before processing the personal information. If the technical solution disclosed herein involves sensitive personal information, the product using this technical solution has obtained the user's separate consent before processing the sensitive personal information, and also meets the requirement of "express consent". For example, at personal information collection devices such as cameras, clear and prominent signs are set up to inform users that they have entered the scope of personal information collection and that personal information will be collected. If an individual voluntarily enters the collection scope, it is deemed that they have agreed to the collection of their personal information; or on the personal information processing device, with clear signs / information informing users of the personal information processing rules, authorization is obtained from the individual through pop-up information or by asking the individual to upload their personal information; wherein, the personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the types of personal information processed.

[0265] It should be noted that, in the embodiments of this disclosure, if the above-described multi-agent task processing method is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of this disclosure, or the part that contributes to related technologies, can be embodied in the form of a software product. This software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, external hard drive, read-only memory (ROM), magnetic disk, or optical disk. Thus, the embodiments of this disclosure are not limited to any specific hardware, software, or firmware, or any combination of hardware, software, and firmware.

[0266] This disclosure provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements some or all of the steps in the above-described method. The computer-readable storage medium may be transient or non-transient.

[0267] This disclosure provides a computer program including computer-readable code, wherein when the computer-readable code is executed in a computer device, a processor in the computer device performs some or all of the steps in the above-described method.

[0268] This disclosure provides a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program. When the computer program is read and executed by a computer, it implements some or all of the steps in the above-described method. This computer program product can be implemented specifically through hardware, software, or a combination thereof. In some embodiments, the computer program product is specifically embodied as a computer storage medium; in other embodiments, the computer program product is specifically embodied as a software product, such as a software development kit (SDK), etc.

[0269] It should be noted that the descriptions of the various embodiments above tend to emphasize the differences between them, while their similarities or commonalities can be referenced interchangeably. The descriptions of the above embodiments of the device, storage medium, computer program, and computer program product are similar to the descriptions of the above method embodiments and have similar beneficial effects. For technical details not disclosed in the embodiments of the device, storage medium, computer program, and computer program product of this disclosure, please refer to the descriptions of the method embodiments of this disclosure for understanding.

[0270] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0271] In the several embodiments provided in this disclosure, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components may be combined, or integrated into another system, or some features may be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed may be through some interfaces, and the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0272] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0273] In addition, each functional unit in the various embodiments of this disclosure can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0274] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as mobile storage devices, read-only memory (ROM), magnetic disks, or optical disks.

[0275] Alternatively, if the integrated units described above are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this disclosure, or the part that contributes to related technologies, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, magnetic disks, or optical disks.

[0276] The above description is merely an embodiment of this disclosure, but the scope of protection of this disclosure is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A task processing method, comprising: Based on user input, an information intelligence agent is used to determine the target task and primary information. The first information represents at least one first task slot corresponding to the target task, and at least one filling information corresponding to the at least one first task slot; Based on task processing information, the first information is updated using an information agent; The task processing information represents processing information related to the target task; Based on the updated first information, the decision-making agent generates at least one execution instruction for executing the current subtask in the target task.

2. The method according to claim 1, wherein the task processing information indicates that the filling information of the second task slot in the at least one first task slot is empty; The step of updating the first information using an information agent based on task processing information includes: In response to the fact that the filling information of the second task slot in the at least one first task slot is empty, the information agent interacts with the user to determine the filling information of the second task slot. The first information is updated using the information agent based on the filling information of the second task slot.

3. The method according to claim 1 or 2, wherein the task processing information represents the execution result of a first historical subtask adjacent to the current subtask; The step of updating the first information using an information agent based on task processing information includes: updating the target task and the first information using an information agent based on task processing information; Based on task processing information, the target task and the first information are updated using an information agent. Based on the execution results of the first historical sub-task, the third task slot is determined using the information agent. The third task slot represents a task slot that is different from the at least one first task slot. The information agent interacts with the user to determine the filling information of the third task slot; Based on the filling information of the third task slot, the information agent updates the subtask information after the first historical subtask in the target task and the first information.

4. The method according to claim 1 or 2, wherein the task processing information represents the context information of the current subtask and the conflict information between the context information of at least one previously executed historical subtask; The context information includes at least one of task execution perception information, execution action information, and execution result information; The step of updating the first information using an information agent based on task processing information includes: Based on the conflict information, the information agent interacts with the user to determine the first feedback information; Based on the first feedback information, the information agent updates the current subtask and the first information.

5. The method according to claim 4, further comprising, before updating the first information using the information agent based on task processing information: The decision-making agent performs a consistency judgment on the context information of the current subtask, the context information of at least one historical subtask, and the first information to determine the conflict information.

6. The method according to claim 5, wherein the step of using the decision-making agent to perform a consistency judgment on the context information of the current subtask, the context information of the at least one historical subtask, and the first information to determine the conflict information includes: Based on the contextual information of at least one second historical subtask, a reflective agent is used to generate summary information. The decision-making agent performs a consistency judgment on the context information of the current subtask, the summary information, the context information of at least one third historical subtask, and the first information to determine the conflict information; The at least one second historical subtask represents a plurality of subtasks executed prior to the at least one third historical subtask.

7. The method according to claim 3 or 4, wherein updating the first information based on task processing information using an information agent comprises: Based on the execution result of the first historical subtask adjacent to the current subtask, a reflection result is generated using a reflective agent; The reflection results include the first reflection results and the second reflection results; The first reflection result is used to indicate whether the first historical subtask was executed; The second reflection result is used to indicate whether the execution result of the first historical subtask is consistent with the first information; In response to the reflection result indicating that the first historical subtask was successfully executed, the first information is updated using the information agent based on the task processing information.

8. The method according to claim 1, wherein the current subtask represents an operation task targeting the application interface of the target application; Based on the updated first information, the step of using a decision-making agent to generate at least one execution instruction for executing the current subtask in the target task includes: Based on the updated first information and the current subtask, the decision-making agent generates a call instruction for the target tool, so that the target tool operates the application interface based on the call instruction.

9. A multi-agent system, comprising an information agent and a decision agent; wherein, The information agent is used to determine a target task and first information based on user input information; and to update the first information based on task processing information; the first information represents at least one first task slot corresponding to the target task, and at least one filling information corresponding to the at least one first task slot; the task processing information represents task processing information related to the target task. The decision-making agent is used to generate, based on the updated first information, at least an execution instruction for executing the current subtask in the target task.

10. An electronic device comprising at least one processor and the memory; wherein, The memory is used to store multiple computer instructions; the at least one processor is used to load and execute the computer instructions to run an information agent and a decision agent; wherein, The information agent determines a target task and first information based on user input information; updates the first information based on task processing information; the first information represents at least one first task slot corresponding to the target task, and at least one filling information corresponding to the at least one first task slot; the task processing information represents task processing information related to the target task. The decision-making agent generates an execution instruction for executing the current subtask in the target task based on the updated first information.

Images