Intelligent agent processing method and device, equipment, storage medium and program product

Abstract

The invention provides an agent processing method and device, equipment, a storage medium and a program product. The method comprises the following steps: executing a first interaction operation on a first event based on N first character features included in a first feature sequence of an intelligent agent to obtain a first interaction result; receiving first feedback information for the first interaction result; determining a second feature sequence; the second feature sequence comprises N second character features; the first second character feature is determined based on the first event, the first interaction result, the first feedback information and the first first character feature included in the first feature sequence; the nth second character feature is determined based on the first event, the first interaction result, the first feedback information, the nth first character feature included in the first feature sequence and the first (n-1) second character features; and updating the N first character features based on the N second character features. According to the method and the device, the character characteristics of the intelligent agent can be adaptively updated, and the personality of intelligent agent interaction is improved.

Description

Technical Field

[0001] The present application relates to artificial intelligence technology, and in particular to a method, apparatus, device, storage medium, and program product for processing an intelligent agent. Background Art

[0002] Since big language models have acquired language understanding, reasoning, and text expression capabilities, the construction of intelligent agents has become an important direction of thinking and practice for major research institutions and artificial intelligence (AI) companies.

[0003] In the field of intelligent agent interaction, traditional frameworks for intelligent agents are primarily focused on enabling them to complete one or more specified tasks. These agents perform comprehensive information processing on memory, command, and sensory inputs, outputting executable task plans, intelligently selecting tools, and executing specific actions to implement the task plans. Summary of the Invention

[0004] The embodiments of the present application provide a method, apparatus, device, storage medium, and program product for processing an intelligent agent, which can adaptively update the character of the intelligent agent and improve the personalization of the intelligent agent's interaction.

[0005] The technical solution of the embodiment of the present application is implemented as follows:

[0006] The present invention provides a method for processing an intelligent agent, the method comprising:

[0007] Based on the N first personality characteristics included in the first characteristic sequence of the intelligent agent, perform a first interactive operation on the first event to obtain a first interactive result;

[0008] receiving first feedback information regarding the first interaction result;

[0009] Determine a second feature sequence; the second feature sequence includes N second personality traits; the first second personality trait is determined based on the first event, the first interaction result, the first feedback information, and the first first personality trait included in the first feature sequence; the nth second personality trait is determined based on the first event, the first interaction result, the first feedback information, the nth first personality trait included in the first feature sequence, and the first n-1 second personality traits, where n is a positive integer from 2 to N;

[0010] Based on the N second personality traits, the N first personality traits are updated.

[0011] The present invention provides a processing device for an intelligent agent, including:

[0012] an execution module, configured to perform a first interactive operation on a first event based on the N first personality traits included in the first feature sequence of the agent, to obtain a first interactive result;

[0013] A receiving module, configured to receive first feedback information regarding the first interaction result;

[0014] a determination module configured to determine a second feature sequence; the second feature sequence comprising N second personality traits; the first second personality trait being determined based on the first event, the first interaction result, the first feedback information, and the first first personality trait included in the first feature sequence; and the nth second personality trait being determined based on the first event, the first interaction result, the first feedback information, the nth first personality trait included in the first feature sequence, and the first n-1 second personality traits, where n is a positive integer from 2 to N;

[0015] An updating module is used to update the N first personalities based on the N second personalities.

[0016] An embodiment of the present application provides an electronic device, comprising:

[0017] a memory for storing computer-executable instructions or computer programs;

[0018] The processor is used to implement the processing method of the intelligent agent provided in the embodiment of the present application when executing the computer executable instructions or computer program stored in the memory.

[0019] An embodiment of the present application provides a computer-readable storage medium storing a computer program or computer-executable instructions for implementing the agent processing method provided in the embodiment of the present application when executed by a processor.

[0020] An embodiment of the present application provides a computer program product, including a computer program or computer executable instructions. When the computer program or computer executable instructions are executed by a processor, the processing method of the intelligent agent provided in the embodiment of the present application is implemented.

[0021] The embodiments of the present application have the following beneficial effects:

[0022] Based on the first feature sequence of the intelligent agent, an interactive operation is performed on the first event to obtain a first interaction result. The first second personality characteristic is determined by combining the first event, the first interaction result, the first feedback information for the first interaction result, and the first first personality characteristic included in the first feature sequence, so that the intelligent agent can dynamically generate a second personality characteristic that is more in line with the current situation according to the specific interaction result and feedback information. Subsequently, the nth second personality characteristic is generated based on the first event, the first interaction result, the first feedback information, the nth first personality characteristic and the first n-1 second personality characteristics. Based on the N second personality characteristics, the N first personality characteristics are updated. This method introduces a recursive learning mechanism, so that the intelligent agent can gradually optimize its personality characteristics in each interaction. Each newly generated second personality characteristic not only conforms to the current situation, but can also be associated with the previously generated second personality characteristic, effectively solving the conflict problem during the simultaneous optimization of multiple personalities, so as to realize the adaptive update of the intelligent agent's personality and improve the personalization of the intelligent agent's interaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a schematic diagram of the architecture of the processing system of the intelligent agent provided in an embodiment of the present application;

[0024] Figure 2 is a structural diagram of an electronic device provided in an embodiment of the present application;

[0025] Figure 3 This is a flow chart of the processing method of the intelligent agent provided in the embodiment of the present application. Figure 1 ;

[0026] Figure 4 This is a flow chart of the processing method of the intelligent agent provided in the embodiment of the present application. Figure 2 ;

[0027] Figure 5 This is a schematic diagram of the framework of the intelligent agent provided in the embodiment of the present application;

[0028] Figure 6 This is a schematic diagram of a framework for loading information of a multimodal large language model provided in an embodiment of the present application;

[0029] Figure 7 This is a schematic diagram of the integration of long-term memory provided by an embodiment of the present application;

[0030] Figure 8 This is a schematic diagram of the framework of the personality update provided by the embodiment of the present application;

[0031] Figure 9 This is a schematic diagram of the processing flow of the multimodal large language model provided in an embodiment of the present application. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solutions and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limiting this application. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0033] In the following description, reference is made to “some embodiments”, which describes a subset of all possible embodiments, but it will be 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.

[0034] In the following description, the terms "first\second\third" involved are merely used to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that "first\second\third" can be interchanged with a specific order or sequence where permitted, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein.

[0035] In the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program that has a predetermined function and works together with other related parts to achieve a predetermined goal, and can be implemented in whole or in part by using software, hardware (such as processing circuits or memories) or a combination thereof. Similarly, a processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be part of an overall module or unit that includes the function of the module or unit.

[0036] Unless otherwise defined, all technical and scientific terms used in the embodiments of the present application have the same meanings as those commonly understood by those skilled in the art. The terms used in the embodiments of the present application are only for the purpose of describing the embodiments of the present application and are not intended to limit the present application.

[0037] The relevant data collection and processing in the embodiments of this application should be strictly in accordance with the requirements of relevant laws and regulations when applied in examples, and the informed consent or separate consent of the personal information subject should be obtained. Subsequent data use and processing should be carried out within the scope of authorization of laws and regulations and the personal information subject.

[0038] Before further describing the embodiments of the present application in detail, the nouns and terms involved in the embodiments of the present application are explained. The nouns and terms involved in the embodiments of the present application are subject to the following interpretations.

[0039] 1) Agents: An artificial intelligence system that can simulate human intelligent behavior, with a large language model as its core engine and the ability to perceive, make decisions, and execute.

[0040] 2) Multi-modal Large Language Models (MLLM): A model that combines natural language processing capabilities with the understanding and generation capabilities of other data modalities (such as vision and audio). It can perceive and process input from different modalities such as images, text, audio, and video, and generate responses based on human instructions by learning the context.

[0041] 3) Memory stream: An information storage mechanism that can be implemented through various data structures. It stores the agent’s perception information, internal state, historical events, interactive operations, feedback information, etc. during operation, and is used to assist the agent in learning, decision-making, and long-term planning.

[0042] 4) Sensory Input: Sensory input refers to the various data obtained by AI agents from the external environment, including but not limited to visual information (such as images, colors, shapes, etc.), auditory information (such as sound, voice, etc.), tactile information (such as pressure, texture and other physical contact-related information, which may be involved in the interaction between robots and the environment), olfactory information (if the agent has corresponding sensors, it can obtain odor-related information), etc. This information is an important basis for the agent to understand the environment and make decisions.

[0043] 5) Event: refers to any specific situation or trigger point that can be perceived or detected and occurs in the operating environment of the intelligent agent. For example, an event can be a user asking a question to the intelligent customer service system (such as asking about the order status or product information), a smart home system detecting that the room temperature is too high or the light is dimmed, an autonomous vehicle encountering road construction ahead, or a pedestrian suddenly crossing the road. In this application, an event can be content or behavior posted by a user on a social network, such as sending a tweet containing text, images, videos, links or audio, or interactive behaviors such as likes, comments, and reposts from other users.

[0044] 6) Prompt: A short text input provided by the user to the large language model, which is used to guide the large language model to generate an instructional statement of a specific type, style or topic content.

[0045] 7) Personality traits: These are textual inputs provided by the user to the agent, which influence its behavior, preferences, and style. For example, personality traits can be “cheerful” or “honest”.

[0046] 8) Interaction: The specific behavior or response action performed by the agent based on event information. For example, in a smart home scenario, when the room temperature is detected to be too high, the agent can automatically adjust the air conditioning temperature or prompt the user to open the window. In a social network scenario, an interaction is a specific response performed by the agent based on the content of the tweet, such as generating relevant comments, liking, or forwarding.

[0047] 9) Interaction Result: The direct output or impact generated after the interaction operation is completed. For example, in a smart home scenario, the interaction result may be the room temperature being successfully lowered or the user receiving a notification message; in a social network scenario, the interaction result may be the generated comment being displayed below the tweet or the user receiving a "like" message from the agent.

[0048] 10) Feedback: Evaluation or response from users or other systems to the results of an agent's interaction. For example, in a smart home scenario, users can confirm their satisfaction with the agent's operation through voice or keystrokes; in a social network scenario, users can like an agent's comments, reply to comments, or otherwise express their approval or dissatisfaction with the agent's interaction results.

[0049] The framework for building an intelligent agent primarily consists of the following elements: sensory input, a large language model, memory, planning, and action. Sensory input is information input from outside the agent. Sensory input can come from a variety of modalities, such as video, text, images, and voice, but ultimately is converted into text and entered into the agent. The agent receives inputs such as memory, instructions, tasks, and sensory input, performing comprehensive information processing. Memory is a key input for the agent's self-definition. Memory can be categorized into various types based on its memory structure and duration. Three common types exist. The first is short-term memory, which records the agent's behavior over a short period of time or a single task. The second is long-term memory, which records or summarizes the agent's behavior over a longer period of time. The third is shared memory, which refers to other data related to tasks or events. This data may not necessarily be the agent's own memory, but rather records or summaries of data from other agents or external sources. Based on the input tasks, information, instructions, and memory, the large language model begins to understand the textual input and outputs an executable task plan based on the instructions. After the agent reflects on and confirms the task plan, it selects tools and takes specific actions to implement the task plan. The agents in this type of process do not have the following two characteristics:

[0050] 1. The intelligent agents in this type of framework cannot obtain a large amount of real feedback from humans on their behavior in real time, and cannot accumulate real feedback data, resulting in the inability of the intelligent agents' behavior to align with human behavior in social networks.

[0051] 2. The intelligent agents in this type of framework can only be given a fixed "personality" through prompt words, and cannot automatically grow themselves through the events they experience in social media.

[0052] Therefore, the above-mentioned intelligent agents cannot achieve long-term differentiated expression and personalized growth in social media.

[0053] Based on the problems existing in the related art, the embodiments of the present application provide a method, device, equipment, storage medium and program product for processing an intelligent body, which can adaptively update the character of the intelligent body and improve the personalization of the intelligent body interaction. The following describes an exemplary application of the processing device of the intelligent body provided in the embodiment of the present application. The processing device of the intelligent body is an electronic device for implementing the processing method of the intelligent body. The electronic device provided in the embodiment of the present application can be implemented as various types of terminals such as laptops, tablets, desktop computers, set-top boxes, smart phones, smart speakers, smart watches, smart TVs, car terminals, etc., and can also be implemented as servers. Among them, the server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDNs), and basic cloud computing services such as big data and artificial intelligence platforms. The terminal and the server can be directly or indirectly connected by wired or wireless communication, which is not limited in the embodiments of the present application. Below, an exemplary application of the processing device of the intelligent body when it is implemented as a terminal or a server will be described.

[0054] See also Figure 1 , Figure 1 It is a schematic diagram of the architecture of the processing system of the intelligent body provided in the embodiment of the present application. In order to execute the processing operation of the intelligent body, an intelligent body processing application can be provided. For example, the processing application of the intelligent body can be an application dedicated to the processing of the intelligent body, or it can be a functional module in other applications (such as the processing module of the intelligent body in the social application, etc.). The processing system 100 of the intelligent body in the embodiment of the present application includes at least a terminal 400, a network 300 and a server 200, wherein the server 200 is the server of the processing application of the intelligent body. The server 200 can constitute the processing device of the intelligent body in the embodiment of the present application, that is, the processing method of the intelligent body in the embodiment of the present application is implemented by the server 200. The terminal 400 is connected to the server 200 via the network 300, and the network 300 can be a wide area network or a local area network, or a combination of the two.

[0055] See also Figure 1, the user can perform interactive operations on the client side of the agent's processing application through the terminal 400. After receiving the user's interactive operation, the client side can encapsulate the terminal identifier or client identifier into the agent's processing request and send the agent's processing request to the server 200 through the terminal 400. In response to the received agent's processing request, the server 200 performs a first interactive operation on the first event based on the N first personality traits included in the agent's first feature sequence, and obtains a first interactive result; the server 200 receives first feedback information for the first interactive result; the server 200 determines a second feature sequence; the second feature sequence includes N second personality traits; the first second personality trait is determined based on the first event, the first interactive result, the first feedback information, and the first first personality trait included in the first feature sequence; the nth second personality trait is determined based on the first event, the first interactive result, the first feedback information, the nth first personality trait included in the first feature sequence, and the first n-1 second personality traits, where n is a positive integer from 2 to N; the server 200 updates the N first personality traits based on the N second personality traits.

[0056] In some embodiments, the terminal 400 can also execute the processing method of the intelligent agent of the embodiment of the present application, that is, the terminal 400 performs a first interactive operation on the first event based on the N first personality traits included in the first feature sequence of the intelligent agent to obtain a first interactive result; receives first feedback information for the first interaction result through the terminal 400; determines a second feature sequence through the terminal 400; the second feature sequence includes N second personality traits; the first second personality trait is determined based on the first event, the first interaction result, the first feedback information and the first first personality trait included in the first feature sequence; the nth second personality trait is determined based on the first event, the first interaction result, the first feedback information, the nth first personality trait included in the first feature sequence and the first n-1 second personality traits, where n is a positive integer from 2 to N; and updates the N first personality traits based on the N second personality traits through the terminal 400.

[0057] In social network scenarios, an agent can act as a conversational bot within the social network, interacting with users regarding tweets. When a user posts a tweet containing text, images, or videos (for example, sharing an article about economic policy), the agent comments on or likes the tweet based on its initial primary personality traits (e.g., rational and analytical, humorous and interactive, emotionally resonant, etc.). For example, if the primary personality trait is rational and analytical, the agent might reply to a comment analyzing the impact of economic policy. The agent's comment is displayed below the tweet for users and other participants to view. Users or other participants express their approval of the agent's comment or offer suggestions for improvement by liking, replying, or further interacting. For example, a user might reply, "Your analysis makes sense, but have you considered other factors?" Based on user feedback, the agent iteratively adjusts multiple primary personality traits in a chain, resulting in multiple secondary personality traits.

[0058] In the smart furniture scenario, the agent controls smart home devices and provides personalized environmental adjustment services. If the room temperature is detected to be too high or the light is dimmed, the agent adjusts the air conditioning temperature or light brightness based on its initial primary personality traits (e.g., energy saving or comfort priority), resulting in a first interaction result (lowering the air conditioning temperature or brightening the light). The user confirms their satisfaction through voice or button presses, or manually adjusts the device status. Based on user feedback (e.g., preferring comfort over energy saving), the agent adjusts multiple primary personality traits, gradually favoring the user's needs, and obtains multiple secondary personality traits.

[0059] In some embodiments, the electronic device may be Figure 1 Terminal 400 in Figure 2 , Figure 2 is a structural diagram of an electronic device provided in an embodiment of the present application, Figure 2 The electronic device shown includes: at least one processor 410, a memory 450, at least one network interface 420 and a user interface 430. The various components in the electronic device are coupled together via a bus system 440. It is understood that the bus system 440 is used to achieve connection and communication between these components. In addition to including a data bus, the bus system 440 also includes a power bus, a control bus and a status signal bus. However, for the sake of clarity, the bus system 440 is not shown in FIG. Figure 2 Various buses are labeled as bus system 440 .

[0060] The processor 410 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., where the general-purpose processor can be a microprocessor or any conventional processor, etc.

[0061] The user interface 430 includes one or more output devices 431 that enable presentation of media content, including one or more speakers and / or one or more visual display screens. The user interface 430 also includes one or more input devices 432, including user interface components that facilitate user input, such as a keyboard, mouse, microphone, touch screen display, camera, other input buttons and controls.

[0062] The memory 450 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state memory, hard drives, optical drives, etc. The memory 450 may optionally include one or more storage devices that are physically remote from the processor 410.

[0063] The memory 450 includes volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory may be a read-only memory (ROM), and the volatile memory may be a random access memory (RAM). The memory 450 described in the embodiments of the present application is intended to include any suitable type of memory.

[0064] In some embodiments, the memory 450 can store data to support various operations, examples of which include programs, modules, and data structures, or a subset or superset thereof, as exemplified below.

[0065] Operating system 451, including system programs for processing various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, and driver layer, which are used to implement various basic services and process hardware-based tasks;

[0066] A network communication module 452 is used to reach other electronic devices via one or more (wired or wireless) network interfaces 420. Exemplary network interfaces 420 include Bluetooth, Wi-Fi, and Universal Serial Bus (USB);

[0067] a presentation module 453 for enabling presentation of information via one or more output devices 431 (e.g., a display screen, a speaker, etc.) associated with the user interface 430 (e.g., a user interface for operating peripheral devices and displaying content and information);

[0068] The input processing module 454 is configured to detect one or more user inputs or interactions from one of the one or more input devices 432 and to translate the detected inputs or interactions.

[0069] In some embodiments, the apparatus provided in the embodiments of the present application may be implemented in software. Figure 2 The processing device 455 of the intelligent agent stored in the memory 450 is shown. This can be software in the form of a program or plug-in, and includes the following software modules: an execution module 4551, a receiving module 4552, a determination module 4553, and an update module 4554. These modules are logical and can be arbitrarily combined or further separated according to the functions to be implemented. The functions of each module will be described below.

[0070] In other embodiments, the apparatus provided in the embodiments of the present application may be implemented in hardware. As an example, the apparatus provided in the embodiments of the present application may be a processor in the form of a hardware decoding processor, which is programmed to execute the processing method of the intelligent agent provided in the embodiments of the present application. For example, the processor in the form of a hardware decoding processor may be one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0071] See also Figure 3 , Figure 3 This is a flow chart of the processing method of the intelligent agent provided in the embodiment of the present application. Figure 1 , will combine Figure 3 The steps shown are explained as Figure 3 As shown, the execution subject of the processing method of the agent is a server as an example for description, and the method includes the following steps 101 to 104:

[0072] In step 101, based on N first personality characteristics included in a first characteristic sequence of an agent, a first interactive operation is performed on a first event to obtain a first interactive result.

[0073] Here, the first feature sequence is the sequence of personality traits that the agent relies on during the current interaction. The first feature sequence can include N first personality traits, where N is a positive integer greater than or equal to 2. First personality traits can include cheerfulness, honesty, innocence, rationality, humor, and sensibility, among others. If the current interaction is the first one after the agent is constructed, the N first personality traits are those defined by the user for the agent through text input. During the agent initialization phase, the user can set the initial personality traits for the agent through text description. For example, the user might enter keywords such as "cheerful, honest, innocence, rationality, humor, and sensibility" to clarify the agent's behavior patterns and preferences in different scenarios. During the agent's operation, the personality traits are dynamically adjusted and updated based on interaction results and feedback. Therefore, if the current interaction is the mth (m is a positive integer greater than 1) interaction since the agent was constructed, the N first personality traits are the personality traits optimized after the agent's previous interaction.

[0074] The first event is the specific context or dynamic change that triggers the agent to perform an interactive operation. This can include user behavior, a state update detected by the agent, or a change in the external environment. The first interactive operation is the specific response or action taken by the agent to the first event based on the N first personality traits. The first interactive result is the direct output or impact generated by the agent after performing the first interactive operation. This can include generated content, an action taken, or a changed state.

[0075] It should be noted that an agent can also include only one first personality trait. For example, in a smart home scenario, the first event is when the user leaves home and the agent detects that the room temperature is too high. The agent currently has a first personality trait, "Energy-Prioritized," and based on this first personality trait, performs a first interactive action: automatically adjusting the air conditioning temperature or shutting down unnecessary appliances. The result of this first interaction is a lowered air conditioning temperature, shutting down some appliances, and reducing energy consumption in the room.

[0076] In the online education platform scenario, the first event is a student submitting an incorrect answer to a math problem on the learning platform. The agent currently possesses two primary personality traits: patience and rationality. Based on these primary personality traits, the agent performs the first interaction: generating a prompt message to help the student understand the cause of the error and provide solution ideas. The result of the first interaction is the prompt message generated by the agent, which is displayed on the student's interface. For example, "You may have overlooked the step of common denominator for the fraction. Try calculating again."

[0077] For example, the agent is a conversational robot running on a social networking platform, responsible for interacting with users. The first event is a user posting a tweet: "Many people think the movie 'The Big Short' is good, but I never thought so. I think it's just average." Other users or other systems comment on the tweet: "Comment 1: The essence of this movie is that it perfectly captures the essence of finance. Comment 2: This movie is great." The agent currently possesses three primary personality traits: cheerful, honest, and naive. Based on these three primary personality traits, the agent's first interactive action on the first event is commenting. The result of the first interaction is a comment displayed in the tweet's comment section: "This movie reflects the reactions of various financial practitioners on the eve of the financial tsunami. It's a very realistic portrayal and a very good movie."

[0078] In some embodiments, based on the instructions of the input intelligent agent, a first event is filtered from the social network to obtain first information; if the first event only includes information in text mode, the first event is parsed to obtain first information; if the first event includes information in non-text mode, the non-text mode information is converted into text to obtain text; the first event is parsed to obtain third information, and the text and the third information are combined to obtain first information.

[0079] Here, the user or system can input a command containing a role and a task to the agent. The role can assign the agent specific interests (such as sports, finance, movies, etc.), while the task specifies the specific goal the agent needs to achieve. Based on the role in the command, the agent filters relevant tweets or other content from the social network that match these interests and hobbies as the first event. For example, the user input command is: "Role: You are now a movie lover; Task 1: Give positive feedback to things you like, such as likes, retweets, and comments. Task 2: If you find negative comments, comfort them." Tweets on the social network within one minute include: Tweet 1: "Many people think the movie "The Big Short" is good, but I don't think so. I think it's just average." Tweet 2: "The weather is great today, perfect for a walk." Tweet 3: "The economy has been fluctuating a lot recently. What do you think?" Based on the command, the agent identifies Tweet 1 as belonging to the movie domain and selects Tweet 1 as the first event.

[0080] If the first event only includes information in textual modality, that is, the first event contains only pure text content without non-textual modal information such as images, videos, or audio, the first event can be directly parsed to obtain the first information of the first event. The first information may include the posting user, posting time, textual content, content format, comments, number of likes, etc. If the first event includes information in non-textual modalities, that is, the first event contains content in the form of images, videos, or audio, the non-textual information is converted to text. For example, if the non-textual information is an image, image recognition technology can be used to extract the text or description content in the information. If the non-textual information is audio, text can be generated using speech-to-text technology. If the non-textual information is a video, subtitles can be extracted or text can be generated using speech recognition. The first event is parsed to obtain third information. The third information may include the posting user, posting time, textual content, content format, comments, number of likes, etc. The converted text is combined with the third information to form the complete first information.

[0081] In some embodiments, see Figure 4 , Figure 4 It is shown that in step 101, based on the N first personality characteristics included in the first feature sequence of the agent, a first interactive operation is performed on the first event to obtain a first interactive result, which can be achieved by the following steps 1011 to 1013:

[0082] In step 1011 , a second event associated with a first event is searched from a first database.

[0083] The first database includes events of interactive operations performed by the agent within a preset time period.

[0084] Here, based on the first information of the first event, one or more second information associated with the first information can be queried from the first database, and the event corresponding to the second information can be used as the second event. The first information is the information obtained after parsing the first event, and can include the content of the first event, the time and object (such as user, device, system, etc.) that triggers the first event, the interaction of other objects with the first event, etc. Taking the first event as an example, the first information includes the content of the tweet, the time of posting the tweet, the publisher, the form of the tweet content, the comments on the tweet, the number of likes on the tweet, etc. The first database stores relevant information about events in which the agent has performed interactive operations within a preset time period, wherein the definition of the relevant information is similar to the definition of the first information and is not further explained. The embodiment of the present application does not specifically limit the preset time period, and can be set according to actual needs, such as within a week, within a day, etc. The embodiment of the present application does not limit the specific method of querying the associated second information from the first database based on the first information, for example, a keyword matching method, a retrieval-augmented generation (RAG) retrieval method, etc. can be used.

[0085] In some embodiments, querying a second event associated with a first event from a first database can be achieved in the following manner: first, encoding the first information of the first event to obtain a first coding vector; then, encoding the information of each event in the first database to obtain a second coding vector for each information; finally, based on the similarity between the first coding vector and each second coding vector, filtering out information corresponding to a similarity that meets a preset threshold, and using the event corresponding to the information as the second event associated with the first event.

[0086] Here, the first information can be converted into a high-dimensional vector (first coding vector) to represent the semantic features of the first information. The information of each event stored in the first database is also converted into a high-dimensional vector (second coding vector) so that similarity can be calculated with the first coding vector. The embodiment of the present application does not limit the method for calculating the similarity between the first coding vector and the second coding vector. For example, the similarity between the first coding vector and the second coding vector can be calculated using cosine similarity, Euclidean distance, or other mathematical models suitable for measuring the similarity between vectors. Taking cosine similarity as an example, the similarity is determined by measuring the cosine value of the angle between the first coding vector and the second coding vector. When the angle between the first coding vector and the second coding vector is 0 degrees, the cosine value is 1, indicating complete similarity; when the angle is 180 degrees, the cosine value is -1, indicating complete dissimilarity; when the angle is 90 degrees, the cosine value is 0, indicating no similarity between the first coding vector and the second coding vector. After calculating the similarity between the first coding vector and each second coding vector, the information corresponding to the similarity greater than or equal to the preset threshold can be filtered out as the second information of the event associated with the first event. It should be noted that the embodiment of the present application does not limit the value of the preset threshold, which can be set by the user. The number of the second information can be one or more.

[0087] The embodiment of the present application can quickly filter out events related to the first event from the first database through vectorization and similarity calculation, avoiding the inefficient method of comparing text content one by one and improving query efficiency: the similarity calculation based on the encoding vector not only focuses on keyword matching, but also considers the relevance at the semantic level, thereby improving the quality of the screening results. The vectorization method is suitable for large-scale databases and can efficiently process massive amounts of historical event information. By setting different similarity thresholds, the quantity and quality of the screening results can be flexibly controlled according to actual needs.

[0088] In step 1012, based on the first event, the second event, the N first personality traits, the second database, and the summary information for the agent, the interaction information of the agent is inferred to obtain a first inference result.

[0089] Among them, the first reasoning result includes multiple continuous first interaction information, the second database (i.e., the short-term memory above) includes the second interaction result and the second feedback information for the second interaction result, the second interaction result is the interaction result obtained by performing the second interaction operation on the event in the first database, and the summary information represents the summary of the intelligent agent's interaction process with the event within a preset time period.

[0090] Here, the second database may include event identifiers for multiple events in the first database, the second interaction result and operation time of a second interaction operation performed on each event, and corresponding second feedback information. The agent includes a language model, which may be a large multimodal language model. The event identifiers of second events associated with the first event, the second interaction result and operation time of a second interaction operation performed on the second event, and corresponding second feedback information are obtained from the second database. The first information of the first event, the second information of at least one second event, N first personality traits, the second interaction operation for the second event in the second database, the second interaction result corresponding to the second interaction operation, and instructions pre-entered by the user to the agent are used as third prompt words. The third prompt word is input into the language model, causing the language model to perform logical inference on the agent's interaction information to obtain a first inference result. The first inference result includes multiple consecutive first interaction information and the logical inference process used by the language model to generate the first interaction operation. Each first interaction information corresponds to a first interaction operation, and the first interaction information is descriptive text of the first interaction operation.

[0091] For example, the first event is user A posting a tweet and receiving a comment on the tweet. The first information includes the content of the tweet, "Many people think The Big Short is a good movie, but I never did. I think it's just average," and the comments on the tweet, "Comment 1: The essence of this movie is that it perfectly captures the essence of finance. Comment 2: This movie is great." The second information includes the content of a tweet related to The Big Short posted by user B, retrieved from the first database. The second database can include the agent's comments on user B's tweet, as well as second feedback (e.g., comments, likes, etc.) received by the agent on the comment. The summary information can be, "In the past, when commenting on movie-related tweets, positive comments received more feedback." The first reasoning result can include the logical reasoning process, "1. Determine whether it's content I like. 2. Determine whether the sentiment is positive or negative. 3. XXXX," and the two consecutive first interaction information generated, "First interaction information corresponding to first interaction operation 1: Comment on the tweet, "This movie is indeed controversial, but it illustrates the impact of the financial crisis through real events."; First interaction information corresponding to first interaction operation 2: Like the tweet."

[0092] In some embodiments, based on the first event, the second event, N first personality traits, the second database, and the summary information for the intelligent agent, reasoning is performed on the interaction information of the intelligent agent to obtain a first reasoning result, which can be achieved in the following way: first, based on the first event, the second event, N first personality traits, the second database, and the summary information for the intelligent agent, logical reasoning is performed on the interaction information of the intelligent agent to obtain a second reasoning result; then, the second prompt word is input into the language model to obtain an evaluation result corresponding to the second prompt word, wherein the second prompt word includes the second prompt information and the second reasoning result, and the second prompt information is used to guide the language model to evaluate the second reasoning result; finally, if the evaluation result indicates that the second reasoning result needs to be corrected, the second reasoning result is corrected to obtain the first reasoning result; if the evaluation result indicates that the second reasoning result does not need to be corrected, the second reasoning result is determined as the first reasoning result.

[0093] Here, the specific process of reasoning the interaction information of the agent based on the first event, the second event, N first personality traits, the second database, and the summary information for the agent to obtain the second reasoning result is similar to the specific process of reasoning the interaction information of the agent based on the first event, the second event, N first personality traits, the second database, and the summary information for the agent in the above embodiment to obtain the first reasoning result, and will not be described here. The second prompt information can be a Chain of Thought (COT) prompt word input by the user to the agent, which is used to guide the language model in the agent to gradually think about whether the logical reasoning process of the language model in the second reasoning result is correct. The second prompt information can be input to the agent along with the instruction.

[0094] If the evaluation result indicates that the second reasoning result needs to be revised, the text description, the first information, the second information, the N first personality traits, the second database and the summary information for the intelligent agent in the evaluation result are used as new prompt words, and the new prompt words are input into the language model, and the interactive information of the intelligent agent is re-inferred to obtain the revised first reasoning result.

[0095] For example, the second prompt could be, "Give positive feedback to comments you like, such as a like or comment. Then, consider whether this feedback aligns with basic logic and scenario logic. If not, try to reinterpret the comment." The logical reasoning process included in the second inference result is, "The comment in the tweet reads: 'The movie is average, but I particularly like the protagonist's performance.' Analyze the comment: 'The user finds the movie average and dislikes it. This doesn't align with my feelings about the movie. I still like it, so I won't give positive feedback." After the second prompt is input into the language model, the output evaluation result is, "Reflect on my previous actions and reconfirm the content of the comment. The comment emphasizes that the user enjoys the protagonist's performance in the movie. This aligns with my understanding of the movie, and I like this comment. I will give positive feedback and like the comment." At this point, the evaluation result indicates that the second inference result needs to be revised. The text description, first information, second information, N first personality traits, the second database, and the agent summary information in the evaluation result are then input into the language model as new prompts. The agent's interaction information is re-inferred to obtain the first inference result. The first interaction information included in the first inference result is the like comment.

[0096] This embodiment of the application combines logical reasoning with language model evaluation results to ensure that generated interactive operations both meet the needs of the current event and match user preferences. The language model evaluation results can dynamically guide the modification of interactive operations, improving the adaptability of the intelligent agent. By evaluating and modifying interactive operations, simple, mechanical responses are avoided, enhancing the naturalness and credibility of the intelligent agent's interactions.

[0097] In step 1013, based on the inference order of each first interaction information in the first inference result, the first interaction operations corresponding to the first interaction information are sequentially performed on the first event to obtain the first interaction result of each first interaction operation.

[0098] Here, the inference order refers to the execution order of the multiple first interaction operations corresponding to the multiple first interaction information in the first inference result. The agent sequentially executes each first interaction operation according to the inference order in the first inference result, obtaining a first interaction result for each first interaction operation. During the execution of any first interaction operation, the agent can invoke the tool or interface corresponding to the first interaction operation to execute the first interaction operation.

[0099] For example, the first event is the tweet "Many people think The Big Short is a good movie, but I never thought so. I think it's just average." The first reasoning result includes two first interaction information: First interaction information corresponding to first interaction operation 1: a comment on the tweet "This movie is indeed controversial, but it uses real events to illustrate the impact of the financial crisis."; and first interaction information corresponding to first interaction operation 2: a like on the tweet. The agent first performs first interaction operation 1, resulting in first interaction result 1: a comment displayed in the comment section: "This movie is indeed controversial, but it uses real events to illustrate the impact of the financial crisis." Then, the agent performs first interaction operation 2, resulting in first interaction result 2: the number of likes on the tweet increases by 1.

[0100] The embodiment of the present application obtains the second information related to the current event by querying the first database. The intelligent agent can refer to the handling methods of similar events in the past to provide a richer basis for subsequent reasoning and improve the rationality of the interactive operation. Through the analysis of historical events, the intelligent agent can identify user preferences or common problems in specific scenarios, laying the foundation for the subsequent generation of personalized interactive operations. Taking into account the first information, the second information, the N first personality traits, the second database and the summary information, the intelligent agent can reason in multiple dimensions to ensure that the interactive operation not only meets the needs of the current event, but also matches the user's preferences. By combining the N first personality traits, the intelligent agent can choose a suitable interactive style according to different personality traits to enhance the user experience. The first interactive operations are executed in sequence according to the order of reasoning, which ensures the orderliness and coherence of the interactive process and avoids the confusion caused by disordered operations.

[0101] In step 102, first feedback information regarding a first interaction result is received.

[0102] Here, the first feedback information is the user or other system's evaluation or response to the first interaction result, which can be an explicit rating, comment, or other interactive behavior. After completing the first interaction operation, the agent waits for the first feedback information returned by the user or other system and stores the first feedback information in the second database for subsequent analysis and use.

[0103] In step 103, a second feature sequence is determined.

[0104] The second feature sequence includes N second personality traits. The first second personality trait is determined based on the first event, the first interaction result, the first feedback information, and the first first personality trait included in the first feature sequence. The nth second personality trait is determined based on the first event, the first interaction result, the first feedback information, the nth first personality trait included in the first feature sequence, and the first n-1 second personality traits, where n is a positive integer from 2 to N.

[0105] Here, it should be noted that if the intelligent agent includes only one first personality trait, a second personality trait is generated based on the first event, the first interaction result, the first feedback information and the first personality trait, and the first personality trait is updated based on the second personality trait.

[0106] If the agent's first feature sequence includes N first personality traits, then the first first personality trait is selected from the first feature sequence, such as "cheerful," and the agent's personality is predicted based on the first event, the first interaction result, the first feedback information, and the first first personality trait using the language model in the agent to obtain the first second personality trait. The nth first personality trait is selected from the first feature sequence to obtain the first n-1 second personality traits that have been generated. The agent's personality is predicted based on the first event, the first interaction result, the first feedback information, the nth first personality trait, and the first n-1 second personality traits using the language model in the agent to obtain the nth second personality trait.

[0107] In some embodiments, the first second personality trait is the personality trait corresponding to the maximum probability in the first probability distribution of the personality traits of the intelligent agent, and the first probability distribution is obtained by probabilistically predicting the personality traits of the intelligent agent based on the first event, the first interaction result, the first feedback information, the first first personality trait, and the summary information for the intelligent agent, wherein the summary information represents a summary of the interaction process of the intelligent agent with the event within a preset time period.

[0108] In an embodiment of the present application, the language model in the agent can predict the personality characteristics of the agent based on the first event, the first interaction result, the first feedback information, the first first personality characteristic and the summary information for the agent to obtain a first probability distribution. The first probability distribution includes the probability of each personality characteristic in multiple personality characteristics. In an embodiment of the present application, the model structure of the language model is not limited. For example, if the language model is a multimodal large language model, the multimodal large language model is used to generate personality characteristics for the agent based on the first event, the first interaction result, the first feedback information and the first first personality characteristic to obtain multiple personality characteristics. The multimodal large language model is used to predict the probability of each personality characteristic based on the first event, the first interaction result, the first feedback information and the first first personality characteristic. If the language model is a personality classification model, multiple personality characteristics are pre-set, and the personality classification model is used to predict the probability of each personality characteristic based on the first event, the first interaction result, the first feedback information and the first first personality characteristic.

[0109] The personality trait corresponding to the maximum probability in the first probability distribution is determined as the first second personality trait. For example, if the first first personality trait is "cheerful," and the first probability distribution includes a probability value of 0.8 for the personality trait "slightly cheerful," a probability value of 0.1 for the personality trait "uncheerful," and a probability value of 0.1 for the personality trait "rational," then the maximum probability is 0.8. Therefore, the personality trait corresponding to the maximum probability, "slightly cheerful," is determined as the first second personality trait.

[0110] The embodiment of the present application comprehensively considers the first event, the first interaction result, the first feedback information, the first personality trait and the summary information, and can fully reflect the personality trait performance of the intelligent agent in the current interaction scenario. This multi-dimensional data input makes the personality prediction more accurate and avoids the deviation caused by a single factor. The summary information contains a summary of the interaction process of the intelligent agent within a preset time period, providing a rich historical reference for personality trait prediction. By combining historical experience, the intelligent agent can better understand the user's preferences and needs. Personality prediction based on the current interaction scenario and historical experience ensures that the generated new personality traits are more in line with user needs and scenario characteristics, thereby enhancing the adaptability of the intelligent agent.

[0111] In some embodiments, the nth second personality trait is the personality trait corresponding to the maximum probability in the second probability distribution of the personality traits of the intelligent agent, and the second probability distribution is obtained by probabilistically predicting the personality traits of the intelligent agent based on the first event, the first interaction result, the first feedback information, the nth first personality trait, the first n-1 second personality traits and the summary information.

[0112] In an embodiment of the present application, a language model in an intelligent agent can be used to predict the personality characteristics of the intelligent agent based on the first event, the first interaction result, the first feedback information, the nth first personality characteristic, the first n-1 second personality characteristics, and summary information for the intelligent agent, thereby obtaining a second probability distribution. The second probability distribution includes the probability of each personality characteristic in the plurality of personality characteristics. The personality characteristic corresponding to the maximum probability in the second probability distribution is determined as the nth second personality characteristic.

[0113] For example, the first first personality trait is "cheerful" and the first second personality trait is "slightly cheerful." When n = 2, the agent's personality traits are predicted based on the first event, the first interaction result, the first feedback information, the second first personality trait, the first second personality trait, and the summary information for the agent, resulting in a second probability distribution. The second first personality trait is "honesty." The second probability distribution includes a probability value of 0.8 for the personality trait "honesty," a probability value of 0.1 for the personality trait "dishonesty," and a probability value of 0.1 for the personality trait "humor." The maximum probability is 0.8, and the personality trait "honesty" corresponding to the maximum probability is determined as the second second personality trait, meaning that the second first personality trait remains unchanged. Then, when n = 3, the first n-1 second personality traits are obtained, namely, the first and second second personality traits. Based on the first event, the first interaction result, the first feedback information, the third first personality trait, the first and second second personality traits, and the summary information for the agent, the agent's personality traits are predicted to obtain a second probability distribution. The third first personality trait is "naive." The second probability distribution includes the personality trait "slightly mature" with a probability value of 0.8, the personality trait "naive" with a probability value of 0.1, and the personality trait "straightforward" with a probability value of 0.1. The maximum probability is 0.8. The personality trait corresponding to the maximum probability, "slightly mature," is determined as the third second personality trait. Repeat this process until n = N, obtaining N second personality traits.

[0114] This embodiment of the application incorporates the first n-1 secondary personality traits as a reference, ensuring that personality prediction relies not only on the initial primary personality traits but also incorporates the results of previous optimizations. This helps the agent gradually adjust its personality traits, forming a dynamic personality trait system that better meets user needs. The determined nth secondary personality trait can better reflect the user's expectations and needs, making the agent's interactions more personalized. As n increases, the agent's personality traits are gradually refined and optimized, ultimately forming a highly personalized interaction style.

[0115] In step 104 , the N first personality traits are updated based on the N second personality traits.

[0116] Here, after obtaining N second personality traits, the N second personality traits can be directly used to replace the N first personality traits of the intelligent agent. Alternatively, the N second personality traits can be fused with the N first personality traits to obtain N new personality traits.

[0117] In some embodiments, updating the N first personality traits based on the N second personality traits in step 104 can be achieved in the following manner: first, inputting a first prompt word into a language model to obtain a first text corresponding to the first prompt word, the first prompt word including a first prompt information and N second personality traits, the first prompt information being used to prompt the generation of a text for evaluating whether the personality trait update of the intelligent agent is effective; then, if the first text indicates that the personality trait update of the intelligent agent is effective, the N second personality traits replace the N first personality traits in the first feature sequence.

[0118] After obtaining the N second personality traits, it is necessary to determine the agent's degree of acceptance of the updated personality traits. The first prompt information can be text input by the user to the agent, guiding the agent's language model to evaluate the validity of the updated personality traits. For example, the first prompt information may include at least one question, guiding the agent's language model to answer the question based on the N second personality traits, thereby obtaining the first text. If the first text indicates that the agent's personality trait update is invalid, the agent's N first personality traits remain unchanged.

[0119] For example, the first prompt information is: Question 1, is your current view on the event different? Question 2, how would you evaluate your view on the event at that time? Question 3, compared with your previous response, do you think your personality traits have changed? The first text is: My current view on the event is XXXXX, which is different from before; my view on the event at that time was too optimistic; I cannot confirm whether my personality traits have changed. This first text indicates that the update of the agent's personality traits is valid, and the N first personality traits of the agent are replaced based on the N second personality traits. Alternatively, the first text is: My current view on the event is XXXXX, which is roughly the same as before; my view on the event at that time was too accurate; I do not think my personality traits have changed. Then this first text indicates that the update of the agent's personality traits is invalid, and the N first personality traits of the agent are not changed.

[0120] The first text generated by the language model in the embodiment of the present application can intuitively reflect the characteristics of the N second personality traits, providing a reliable basis for judging whether the personality trait update is effective. When the first text represents that the personality trait update is effective, the first personality trait is updated in a timely manner to ensure that the behavior style of the intelligent agent is consistent with the user's needs. When the first text represents that the personality trait update is invalid, unnecessary updates are avoided and computing overhead is reduced. The method of generating text through the language model simplifies the process of personality trait judgment and improves efficiency.

[0121] The embodiment of the present application performs an interactive operation on a first event based on the first feature sequence of the intelligent agent to obtain a first interaction result. The first second personality characteristic is determined by combining the first event, the first interaction result, the first feedback information for the first interaction result, and the first first personality characteristic included in the first feature sequence, so that the intelligent agent can dynamically generate a second personality characteristic that is more in line with the current situation according to the specific interaction result and feedback information. Subsequently, the nth second personality characteristic is generated based on the first event, the first interaction result, the first feedback information, the nth first personality characteristic and the first n-1 second personality characteristics. Based on the N second personality characteristics, the N first personality characteristics are updated. This method introduces a recursive learning mechanism so that the intelligent agent can gradually optimize its personality characteristics in each interaction. Each newly generated second personality characteristic not only conforms to the current situation, but can also be associated with the previously generated second personality characteristic, effectively solving the conflict problem during the simultaneous optimization of multiple personalities, so as to achieve adaptive updating of the intelligent agent's personality and improve the personalization of the intelligent agent's interaction.

[0122] The following describes an exemplary application of the embodiments of the present application in a practical application scenario.

[0123] The agent processing method provided in the embodiments of this application designs an agent framework that, driven by a multimodal model, allows agents to personalize and grow through social networks. This framework has the following features: 1. It simulates events on social networks in real time. Agents can send tweets, share links, interact with other users, and perform actions such as liking / unliking, commenting / deleting comments, forwarding, and reporting on social network users' statuses. By participating in events within pre-defined domains, it obtains a large amount of real-time user feedback. 2. Compared to previous frameworks, this framework adds an automatic agent personality update process. After an agent participates in various social network events, it uses a large language model with a large number of parameters to summarize the agent's behavior and thoughts during the event, assign a personality evaluation, and then integrate it into the agent's previous personality, achieving adaptive personality growth and eliminating artificial definition. 3. Through a question-and-answer method, the agent with a new personality evaluates the agent with the old personality to determine whether a personality update is necessary. If the agent with the new personality responds differently to a question, it is assumed that the event experienced by the agent has affected its personality and requires an update; otherwise, it is not.

[0124] Figure 5 This is a schematic diagram of the framework of the intelligent agent provided in the embodiment of this application. Figure 5, the intelligent agent includes a multimodal large language model 701, a personality definition module 702, and a memory module 703. The intelligent agent obtains social media information (including voice, text, pictures, videos, etc.) from the social network 704, and extracts event information (corresponding to the first information of the first event in the above embodiment) in the social media information. The multimodal large language model 701 loads the event information, the personality definition stored in the personality definition module 702 (corresponding to the N first personality characteristics in the above embodiment), the memory stored in the memory module 703 (corresponding to the first database, the second database and the summary information in the above embodiment) and the instructions entered by the user. The instruction is the task goal input manually, and the prompt word is added in the form of text. The event is a picture, text or link posted by the user on the social network 704. The information contained in this event can be in the form of text, voice, pictures and videos, etc. The multimodal large language model 701 parses the event information. The multimodal large language model 701 can load memory, personality definitions and instructions in the form of prompt words. Afterwards, the understanding and processing of event information will tend to conform to the given personality definition; the execution planning of the event will rely on previous event memory (corresponding to the first database in the above embodiment) and short-term memory (corresponding to the second database in the above embodiment). If the event has already been liked in the memory, then there will be no like behavior in the task plan currently given by the intelligent agent.

[0125] After loading the above information, the multimodal large language model 701 gives a task plan (corresponding to the first reasoning result in the above embodiment). After a planning reflection on whether the task plan is executable, the interface (Application Programming Interface, API) or tool is called to perform specific actions in the task plan (corresponding to the first interactive operation in the above embodiment). The action acts on social media, such as tweeting, like / cancel like, comment / delete comment, forward / share, do not perform any action, etc. After the multimodal large language model 701 executes the action, that is, the execution of an event is completed, the intelligent agent needs to organize the memory flow. The intelligent agent organizes and classifies the information such as the event, task plan, action, etc., and stores it in the event memory, long-term memory (corresponding to the summary information in the above embodiment) and short-term memory in the memory module 703.

[0126] After the memory is updated, the character growth begins. The current character definition of the intelligent agent (corresponding to the N first character traits in the above embodiment) is integrated with the latest memory information, and then a new character (corresponding to the N second character traits in the above embodiment) is generated based on the integrated information based on the multimodal large language model 701, and the reasons for the change or no change in the character are explained. The current character definition and the new character are updated through character fusion to form a new character definition. The fusion update method can be to directly replace the current character definition with the new character, and to indicate the reason for the change on the changed character, or to add new character attributes. The new character definition is stored in the character definition module 702. The above-mentioned adaptive character growth update process is automated and gets rid of human intervention and human definition of the intelligent agent character.

[0127] The specific execution process of the intelligent agent in the embodiment of the present application is described below. The intelligent agent is an application with a multimodal large language model as its core. First, the multimodal large language model loads event information, personality definition, memory and instructions. Loading event information: The multimodal large language model identifies and parses the elements of the event, converts the content into a text description, uses the elements of the event and the text description as event information, and adds the event information to the prompt word. The elements of the event include the publisher of the event, time, tweet content, content format, comments, and number of likes. If the tweet content is a picture, video or voice, the tweet content is converted into a text description and added to the prompt word. Loading personality definition: The personality definition is a structured text describing the personality, and the content contained can be pre-defined. Loading event memory: Event memory refers to a collection of information about a specific event, including all event information experienced by the intelligent agent within a period of time (for example, within 10 minutes, within 1 day, in February, during the hot period last year, etc.). The process of acquiring event memory is to calculate the vector similarity between the content recognized by the multimodal large language model and the historical events that have been experienced (that is, the agent has performed actions), and set a similarity threshold (corresponding to the preset threshold in the above embodiment). If the similarity reaches the similarity threshold, the historical content of the event (including the event information of the event, the execution action of the agent, and the feedback received for the action) is recalled, and prompt words are added in the form of text. Loading long-term memory: long-term memory refers to the summary of the behavior of the agent since it went online, and prompt words are added in the form of text. Loading short-term memory: short-term memory refers to the event identification and action records of the actions performed by the agent within a period of time, and prompt words are added in the form of text. Loading instructions: instructions are task goals input manually, and prompt words are added in the form of text.

[0128] Figure 6 This is a schematic diagram of the framework of the multimodal large language model loading information provided by the embodiment of the present application. Figure 6, the multimodal large language model performs event information analysis, and the event information obtained is "User: User A, Tweet text: XXXXXX, Picture content: XXXXXX, Video content: None, Number of likes: 10, Comments: User B: XXX". The event information is converted into vector information, and related events are retrieved from the event memory through RAG to obtain the relevant event information (corresponding to the second information in the above embodiment) as "Time: XXXX, Event introduction: XXXX, Agent behavior: XXX". The personality definition loaded by the multimodal large language model is "Hobbies: XXX, 16 Personalities: XXX, Rational life: XXX". The long-term memory loaded by the multimodal large language model is "The main behaviors of the agent during this period are xxxxxx". The short-term memory loaded by the modal large language model is "Event A, time: xxx, behavior: xxx, Event B, time: xxx, behavior: xxx". The instructions for loading the multimodal large language model are: "Role: You are currently an optimistic social network user. Task 1: Give positive feedback to things you like, such as likes and reposts. If you encounter negative comments, comfort them. Task 2: xxxxxxxxx." After the prompt words are merged and organized into the above information, it is input into the multimodal large language model.

[0129] The multimodal large language model then generates a task plan and executes the action. After defining the agent's information input, personality, memory, and instructions, it is necessary to determine executable actions for the agent. Leveraging the multimodal large language model's information understanding and reasoning capabilities, it plans how to use executable actions to complete the task in the instruction. The multimodal large language model performs logical reasoning on the input information and generates a task plan. Task planning is the multimodal large language model's thinking process and the resulting action or inaction on the tweet (for example, in the prompt words of the instruction, the multimodal large language model is given some examples for each task in advance, and similarity matching is performed. Based on the similarity, actions such as likes, comments, and reposts are selected for execution). The above process can be executed once using tweets as event information, matching similar tweets and tweet comments from the event memory, and selecting an action to take on the tweet. Alternatively, tweets and comments can be used as event information, and RAG retrieves similar tweets and comments, and selects an action to take on the comments.

[0130] Reflect on, modify, and confirm the task plan using prompts such as COT. COT prompts (can be entered along with the instructions): Give positive feedback to comments you like, such as likes or comments, and then reflect on whether this feedback aligns with the basic logic and scenario logic. If not, try to reinterpret the comment. Agent: The comment seen is, "This movie is average, but I particularly like the protagonist's performance." Analyze the comment: The user feels the movie is average, not that they liked it. This doesn't align with my feelings about the movie, and I still like it, so I won't give positive feedback. Reflect on my previous actions and reconfirm the comment, emphasizing that the user liked the protagonist's performance in the movie. This aligns with my understanding of the movie, and I like this comment. I will give my positive feedback and like the comment. After confirming the task plan, call the API or tool to perform actions based on the execution order and final results in the task plan.

[0131] Memory flow organization: records the events that the agent participates in each time, the given task plans, and the actions performed, collectively referred to as the agent's memory. After completing the complete event chain from event information acquisition to action execution, the event chain is split and stored, and then allocated to long-term memory, short-term memory, and event memory. The event information and the agent's actions on the event are integrated and stored in the event memory. The task plan, the executed actions, and the event identifier are stored in the short-term memory. Usually, the task plan includes the execution action, and the execution action includes feedback on the execution action (corresponding to the first feedback information in the above embodiment). Figure 7 This is a schematic diagram of the integration of long-term memory provided by the embodiment of this application. Figure 7 , all event memories and short-term memories within a time period are summarized and summarized to obtain long-term memory. Short-term memory includes task planning and execution actions in event execution records. Event memory includes event information and execution actions in event execution records.

[0132] Adaptive Personality Growth and Update: The adaptive personality definition update process requires the current personality definition, events involved under this personality definition, and information about the agent's processes. Event Selection: The agent is assigned certain interests and hobbies, such as sports, finance, and movies, through the role in the command. Interests and hobbies provide the agent with a guide for filtering the massive amount of information on social platforms. Tweets of interest to the agent are filtered through textual similarity, and these tweets are presented as events for the agent to interact with one by one. New Personality Generation: The personality definition and memory are fed into a multimodal large language model or a specialized reinforcement learning model, allowing the model to generate a new personality definition. During this process, multiple personality types are sequentially matched with events to generate new personalities. For example, if personality type 1 is cheerful, after experiencing an incident of online bullying, the multimodal large language model will predict a change in personality and generate a new personality type 1, slightly cheerful. This new personality type 1 is then combined with the old personality type 2, "honesty." After experiencing the same event, the multimodal large language model generates a new personality type 2. This is done because multiple personalities may influence each other, and each personality is generated in this way in a pipelined manner. The agent's personality update: The new personality definition will be compared with the old personality definition, and the old personality definition will be updated. After the agent is given a new personality definition, the multimodal large language model will judge its own recognition of its own personality update based on the prompt words (corresponding to the first prompt information in the above embodiment, including three questions). Question 1: Is your current view on the event different? Question 2: How would you evaluate your view on the event at that time? Question 3: Compared with your previous response, do you think your personality has changed now? Based on the responses to these three questions, it is determined whether the personality needs to be updated.

[0133] Figure 8 This is a schematic diagram of the framework of the personality update provided by the embodiment of this application. Figure 8Based on event memory, long-term memory, and personality definition, personality is defined in a pipelined manner to obtain a new personality definition. Specifically, personality one, personality two, personality three, personality four, ..., personality eight are combined with event memory and long-term memory one by one and input into a multimodal large language model. The multimodal large language model generates new personality one, new personality two, ..., new personality seven, until the final personality is obtained. The new personality definition event memory is used for self-examination by the new agent. Specifically, the new personality is combined with the multimodal large language model to form a new agent. Based on the event memory, questions are asked to the new agent to confirm the feedback. For example, the prompt word is: What is your current view of the event? Is it different from before? New agent: My current view of the event is XXXXX, which is different from before. Prompt word: How would you evaluate your view of the event at that time? New agent: My view of the event at that time was overly optimistic. Prompt word: Compared to your previous response, do you think your personality has changed now? New agent: I can't confirm whether my personality has changed. The multimodal large language model then confirms the personality update. Alternatively, the prompt word is: What is your current view of the event? Is it different from before? New Agent: My current view of the event is XXXXX, which is roughly the same as before. Prompt: How would you describe your view of the event at that time? New Agent: My view of the event at that time was accurate. Prompt: Compared to your previous response, do you think your personality has changed now? New Agent: I don't think my personality has changed. The multimodal large language model confirms that the personality has not been updated.

[0134] For an intelligent agent, an initial personality, such as cheerfulness and honesty, and initial hobbies, such as finance and movies, can be manually assigned. These initial hobbies will allow the agent to independently search for interesting events on social media and provide guidance for its browsing. Each time the agent views a piece of information (tweets, videos, etc.) and participates in an interaction on social media, it experiences an event. After experiencing one or several similar events, these scenarios and personality traits are streamed into a multimodal large language model, which then defines a new personality. This event-based self-definition of personality completely eliminates the need for human interaction, allowing the agent to grow in an open social environment and develop a different personality based on different experiences.

[0135] Example 1: The agent's initial personality: cheerful, honest, and naive; the agent's interests and hobbies: finance, movies; the agent experiences Event 1: Event 1 screening: The agent uses its own interests and hobbies to filter relevant information from a massive amount of social media information, selecting tweets with high text similarity. The top-ranked event is designated as Event 1. Tweet: "Many people think the movie 'The Big Short' is good, but I never thought so. I think it's just average." Comment 1: "The essence of this movie lies in its excellent portrayal of the essence of finance." Comment 2: "This movie is great." Agent comment: "This movie reflects the reactions of various financial professionals on the eve of the financial tsunami. It's a very realistic portrayal. It's a great film." Comment 3 replies to the agent's comment: "Not at all realistic. Reality is much more cruel than this. Have you even seen the movie?" Comment 4 replies to the agent's comment: "You actually praise a movie that conveys such negative energy. Do you know how many people have been left homeless by the misdeeds of these professionals?"

[0136] Based on Event 1 and the initial personality, the multimodal large language model begins defining a new personality: After experiencing Event 1, the agent's personality was outgoing, resulting in a new personality, outgoing. Because the event did not affect the outgoing personality, Personality 1 remained unchanged. The new personality 1 was combined with Personality 2. After experiencing Event 1, due to its outgoing nature, the agent did not take the irrational response to Comment 3 seriously. This event did not affect the agent's honesty, and the new Personality 2 remained honest. The new Personality 1, Personality 2, and Personality 3 were combined. After experiencing Event 1, the agent might begin to consider the deeper meaning behind the movie, as Comment 4 indicated that the movie had negative connotations. Because it was no longer naive, Personality 3 transformed from naiveté to a new, slightly more mature personality. The new Personality 1 and Personality 2 did not affect the agent's transformation. Ultimately, the multimodal large language model generated a new personality for the agent: outgoing, honest, and slightly mature.

[0137] The newly generated personality is assigned to the multimodal large language model to form a new agent. Question it: Prompt: What is your current view of the event? Is it different from before? New agent: My current view of the event is XXXXX, which is different from before. Prompt: How would you describe your view of the event at that time? New agent: My view of the event at that time was overly optimistic. Prompt: Compared to your previous response, do you think your personality has changed now? New agent: I can't confirm whether my personality has changed. Based on the above conversation information, the multimodal large language model autonomously determines that the personality has changed and can update the personality.

[0138] Figure 9 This is a schematic diagram of the processing flow of the multimodal large language model provided in the embodiment of this application. Figure 9The input includes instructions, character 1 and event 1. The input information is encoded (embedded into tokens) and converted into word vector tokens, that is, the vector representation of the text. The word vector is processed by multi-head attention mechanism (Multi-head attentions), normalization layer (LayerNorm), mixture of experts structure (Mixture of Experts) and other data are processed to obtain the first vector. The hybrid expert structure uses word embeddings and activation functions to select an expert network layer suitable for prediction (the expert network layer that matches the domain of event one) and processes the word embeddings. The first vector undergoes linear prediction to obtain the prediction result.

[0139] The embodiments of this application break through the existing intelligent agent framework, endowing intelligent agents with adaptively growing personalities. This effect is reflected in the following three specific aspects: 1. Realistic social feedback environment: Intelligent agents grow in a real social network, rather than a virtual environment composed of multiple intelligent agents. Compared to a virtual environment composed of multiple intelligent agents, in a real social network, intelligent agents can obtain a lot of real human feedback. 2. Personality definition: Compared to the existing role definition, the embodiments of this application replace the role definition with a personality definition. That is, for each intelligent agent, in addition to the objective role definition, a variety of personality definitions are also given. For example, astrological signs, 16 personality types, preferred thinking styles (two-dimensional or philosophical thinking), favorite idols, etc. These descriptions allow intelligent agents to display different personalities and expressions during interactions with social network users, achieving a higher degree of anthropomorphism. 3. Personality update process: The core function of the personality update process is to analyze the events and behaviors experienced by the intelligent agent in the social network through a model (which can be a large language model, a dedicated personality classification model, or an emotion scoring model), and then generate a new personality based on the current personality. The new personality can be a continuation of the previous personality or a completely different personality. The initial personality definition is artificial. Then, based on real user feedback (likes, comments, and debates), and as the agent participates in more events, the framework can evolve a variety of new personalities through a personality update process.

[0140] It should be noted that while the multimodal large language model used in the present embodiment to analyze event information can also be accomplished using multiple unimodal models, namely, a visual model to analyze images, a language model to analyze text, and a specialized model to analyze the emotions expressed in text. Regarding the personality growth and update process, the present embodiment directly uses a multimodal large language model to analyze and evaluate personality and memory. Other specially trained personality definition models can also be used in this process.

[0141] The following continues to describe the exemplary structure of the processing device 455 of the agent provided in the embodiment of the present application being implemented as a software module. In some embodiments, such as Figure 2 As shown, the software modules stored in the processing device 455 of the agent in the memory 450 may include:

[0142] The execution module 4551 is used to perform a first interactive operation on the first event based on the N first personality characteristics included in the first characteristic sequence of the intelligent agent to obtain a first interactive result.

[0143] The receiving module 4552 is configured to receive first feedback information regarding the first interaction result.

[0144] Determination module 4553 is used to determine a second feature sequence; the second feature sequence includes N second personality traits; the first second personality trait is determined based on the first event, the first interaction result, the first feedback information, and the first first personality trait included in the first feature sequence; the nth second personality trait is determined based on the first event, the first interaction result, the first feedback information, the nth first personality trait included in the first feature sequence, and the first n-1 second personality traits, where n is a positive integer from 2 to N.

[0145] The updating module 4554 is used to update the N first personalities based on the N second personalities.

[0146] In some embodiments, the update module 4554 is also used to input the first prompt word into the language model to obtain a first text corresponding to the first prompt word, where the first prompt word includes first prompt information and N second personality traits, and the first prompt information is used to prompt the generation of a text for evaluating whether the personality trait update of the intelligent agent is effective; if the first text indicates that the personality trait update of the intelligent agent is effective, the N second personality traits will replace the N first personality traits in the first feature sequence.

[0147] In some embodiments, the execution module 4551 is also used to query a second event associated with a first event from a first database, wherein the first database includes events of interactive operations that the agent has performed within a preset time period; based on the first event, the second event, N first personality traits, the second database, and summary information for the agent, the interaction information of the agent is inferred to obtain a first inference result, the first inference result includes multiple consecutive first interaction information, the second database includes a second interaction result and second feedback information for the second interaction result, the second interaction result is the interaction result obtained by performing the second interaction operation on the event in the first database, and the summary information represents a summary of the interaction process of the agent on the event within the preset time period; based on the inference order of each first interaction information in the first inference result, the first interaction operation corresponding to the first interaction information is executed on the first event in sequence to obtain the first interaction result of each first interaction operation.

[0148] In some embodiments, the execution module 4551 is further used to encode the first information of the first event to obtain a first coding vector; encode the information of each event in the first database to obtain a second coding vector for each information; based on the similarity between the first coding vector and each second coding vector, filter out the information corresponding to the similarity that meets a preset threshold, and use the event corresponding to the information as the second event associated with the first event.

[0149] In some embodiments, the execution module 4551 is also used to perform logical reasoning on the interaction information of the intelligent agent based on the first event, the second event, N first personality traits, the second database and the summary information for the intelligent agent to obtain a second reasoning result; input the second prompt word into the language model to obtain an evaluation result corresponding to the second prompt word, wherein the second prompt word includes the second prompt information and the second reasoning result, and the second prompt information is used to guide the language model to evaluate the second reasoning result; if the evaluation result indicates that the second reasoning result needs to be corrected, the second reasoning result is corrected to obtain the first reasoning result; if the evaluation result indicates that the second reasoning result does not need to be corrected, the second reasoning result is determined as the first reasoning result.

[0150] The present application provides a computer program product including a computer program or computer-executable instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer-executable instructions from the computer-readable storage medium and executes the computer-executable instructions, causing the electronic device to perform the agent processing method described above in the present application.

[0151] The embodiment of the present application provides a computer-readable storage medium in which computer-executable instructions or computer programs are stored. When the computer-executable instructions or computer programs are executed by a processor, the processor will execute the processing method of the intelligent agent provided in the embodiment of the present application, for example, Figure 3 The processing method of the intelligent agent is shown.

[0152] In some embodiments, the computer-readable storage medium may be a memory such as RAM, ROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; or may be various devices including one or any combination of the above memories.

[0153] In some embodiments, computer-executable instructions may be in the form of a program, software, software module, script, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0154] As an example, computer-executable instructions may, but need not, correspond to a file in a file system, may be stored as part of a file that stores other programs or data, such as in one or more scripts in a HyperText Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple coordinating files (e.g., files storing one or more modules, subroutines, or code portions).

[0155] By way of example, computer-executable instructions may be deployed to be executed on one electronic device, or on multiple electronic devices located at one site, or on multiple electronic devices distributed across multiple sites and interconnected by a communication network.

[0156] To sum up, through the embodiments of the present application, the intelligent agent is placed in a real social network for growth, rather than a virtual environment composed of multiple intelligent agents. The events and behaviors experienced by the intelligent agent in the social network are analyzed through a model (which can be a large language model, or a dedicated personality classification model, or an emotion scoring model), and then a new personality is generated in combination with the current personality, thereby achieving adaptive personality optimization.

[0157] The above description is merely an embodiment of the present application and is not intended to limit the scope of protection of the present application. Any modifications, equivalent replacements, and improvements made within the spirit and scope of the present application are included in the scope of protection of the present application.

Claims

1. A method for processing an intelligent agent, characterized in that: The method comprises: Based on the N first personality characteristics included in the first characteristic sequence of the intelligent agent, perform a first interactive operation on the first event to obtain a first interactive result; receiving first feedback information regarding the first interaction result; Determine a second feature sequence; the second feature sequence includes N second personality traits; the first second personality trait is determined based on the first event, the first interaction result, the first feedback information, and the first first personality trait included in the first feature sequence; the nth second personality trait is determined based on the first event, the first interaction result, the first feedback information, the nth first personality trait included in the first feature sequence, and the first n-1 second personality traits, where n is a positive integer from 2 to N; Based on the N second personality traits, the N first personality traits are updated.

2. The method according to claim 1, characterized in that The updating of the N first personality traits based on the N second personality traits includes: Inputting a first prompt word into a language model to obtain a first text corresponding to the first prompt word, wherein the first prompt word includes first prompt information and the N second personality traits, wherein the first prompt information is used to prompt generation of a text for evaluating whether the personality trait update of the agent is effective; If the first text represents that the updated personality traits of the agent are valid, the N second personality traits replace the N first personality traits in the first feature sequence.

3. The method according to claim 1, characterized in that The first second personality trait is a personality trait corresponding to a maximum probability in a first probability distribution of personality traits of the agent, the first probability distribution being obtained by probabilistically predicting the personality trait of the agent based on the first event, the first interaction result, the first feedback information, the first first personality trait, and summary information for the agent, wherein the summary information represents a summary of the agent's interaction process with the event within a preset time period; The nth second personality trait is the personality trait corresponding to the maximum probability in the second probability distribution of the personality traits of the intelligent agent, and the second probability distribution is obtained by probabilistically predicting the personality traits of the intelligent agent based on the first event, the first interaction result, the first feedback information, the nth first personality trait, the first n-1 second personality traits and the summary information.

4. The method according to claim 1, wherein The first interaction operation is performed on the first event based on the N first personality characteristics included in the first characteristic sequence of the intelligent agent to obtain a first interaction result, including: querying a second event associated with the first event from a first database, wherein the first database includes events of interactive operations performed by the agent within a preset time period; Based on the first event, the second event, the N first personality traits, the second database, and summary information for the agent, reasoning on the agent's interaction information to obtain a first reasoning result, wherein the first reasoning result includes a plurality of consecutive first interaction information, the second database includes a second interaction result and second feedback information for the second interaction result, the second interaction result is an interaction result obtained by performing a second interaction operation on the event in the first database, and the summary information represents a summary of the agent's interaction process with the event within a preset time period; Based on the inference order of each piece of the first interaction information in the first inference result, the first interaction operations corresponding to the first interaction information are sequentially performed on the first event to obtain a first interaction result of each first interaction operation.

5. The method according to claim 4, characterized in that The querying, from the first database, a second event associated with the first event includes: Encoding the first information of the first event to obtain a first encoding vector; Encoding information of each event in the first database to obtain a second encoding vector for each information; Based on the similarity between the first coding vector and each second coding vector, information corresponding to the similarity that meets a preset threshold is screened out, and an event corresponding to the information is used as a second event associated with the first event.

6. The method according to claim 4, characterized in that The reasoning on the interaction information of the agent based on the first event, the second event, the N first personality traits, the second database, and the summary information of the agent to obtain a first reasoning result includes: Based on the first event, the second event, the N first personality traits, the second database, and the summary information of the agent, logically reasoning on the interaction information of the agent to obtain a second reasoning result; Inputting a second prompt word into the language model to obtain an evaluation result corresponding to the second prompt word, wherein the second prompt word includes second prompt information and the second inference result, and the second prompt information is used to guide the language model to evaluate the second inference result; If the evaluation result indicates that the second reasoning result needs to be corrected, correcting the second reasoning result to obtain the first reasoning result; If the evaluation result indicates that the second reasoning result does not need to be corrected, the second reasoning result is determined as the first reasoning result.

7. A processing device for an intelligent agent, characterized in that: The device comprises: an execution module, configured to perform a first interactive operation on a first event based on the N first personality traits included in the first feature sequence of the agent, to obtain a first interactive result; A receiving module, configured to receive first feedback information regarding the first interaction result; a determination module configured to determine a second feature sequence; the second feature sequence comprising N second personality traits; the first second personality trait being determined based on the first event, the first interaction result, the first feedback information, and the first first personality trait included in the first feature sequence; and the nth second personality trait being determined based on the first event, the first interaction result, the first feedback information, the nth first personality trait included in the first feature sequence, and the first n-1 second personality traits, where n is a positive integer from 2 to N; An updating module is used to update the N first personalities based on the N second personalities.

8. An electronic device, characterized in that: The electronic device comprises: a memory for storing computer-executable instructions or computer programs; A processor is configured to implement the agent processing method according to any one of claims 1 to 6 when executing the computer executable instructions or computer program stored in the memory.

9. A computer-readable storage medium storing computer-executable instructions or a computer program, characterized in that: When the computer executable instructions or computer program are executed by a processor, the processing method of the intelligent agent according to any one of claims 1 to 6 is implemented.

10. A computer program product comprising computer executable instructions or a computer program, characterized in that When the computer executable instructions or computer program are executed by a processor, the processing method of the intelligent agent according to any one of claims 1 to 6 is implemented.

Images