Continuous session method and device for digital life, medium, equipment and program product

Through differentiated triggering conditions and autonomous judgment mechanisms, digital life can autonomously decide the timing and content of its speech in multi-round conversations, solving the mechanical problems of traditional dialogue systems, achieving a natural and continuous dialogue experience, and improving the authenticity and coherence of the dialogue.

CN120973903APending Publication Date: 2025-11-18CHINA MERCHANTS ADVANCED TECHNOLOGY DEVELOPMENT (SHENZHEN) CO LTD
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202511111588.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-08-08
Publication Date
2025-11-18

AI Technical Summary

Technical Problem

Traditional AI dialogue systems lack naturalness and continuity, causing digital beings to exhibit mechanical characteristics in conversations, making it difficult to establish a realistic and credible dialogue experience.

Method used

By setting differentiated trigger conditions to control the content output of each conversation round, combined with a mechanism that allows digital beings to autonomously determine the duration or termination of a conversation, the system enables digital beings to autonomously decide when and what to say. It introduces time-driven and event-driven trigger mechanisms, establishes a multi-round conversation mechanism, and enhances the coherence and realism of the dialogue.

Benefits of technology

It enhances the realism and coherence of the dialogue process. Digital life can generate coherent dialogue content under user-triggered and self-triggered conditions, reducing the mechanical feeling, enhancing the naturalness and vividness of the interaction, and improving the quality of dialogue and user experience.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120973903A_ABST
    Figure CN120973903A_ABST
Patent Text Reader

Abstract

The invention belongs to the technical field of artificial intelligence, and particularly relates to a digital life continuous session method, a digital life continuous session device, a computer readable medium, electronic equipment and a computer program product. The method comprises: when a preset first trigger condition is satisfied, outputting a first session content belonging to a first session round, the first session round comprising historical session contents of a plurality of session members, the session members comprising one or more digital lives, the first trigger condition comprises a data input event triggered by the session member; when a preset second trigger condition is met, second session content belonging to a second session round is output, and the second trigger condition is different from the first trigger condition. The dialogue autonomy and continuity of digital life can be improved, and the sense of reality of interaction is enhanced.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence technology, specifically relating to a method for continuous conversation of digital life, a device for continuous conversation of digital life, a computer-readable medium, an electronic device, and a computer program product. Background Technology

[0002] Current AI dialogue systems have significant limitations, primarily in the naturalness and continuity of conversational interaction. Traditional dialogue tools typically employ a passive response model, generating a single-turn reply only when user input is detected. This mechanical interaction method differs significantly from natural human dialogue. Moreover, the passive response approach, reliant on external triggers, severely limits the depth and coherence of the conversation, causing the digital entity to exhibit obvious mechanical characteristics and making it difficult to establish a truly authentic and believable dialogue experience. Summary of the Invention

[0003] This application provides a method, device, computer-readable medium, electronic device, and computer program product for continuous dialogue of digital life, with the aim of enhancing the dialogue autonomy and continuity of digital life and increasing the realism of interaction.

[0004] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part from practice of this application.

[0005] According to one aspect of the embodiments of this application, a method for continuous conversation of a digital life is provided. The method includes: when a preset first triggering condition is met, outputting first conversation content belonging to a first conversation round, the first conversation round including historical conversation content of multiple conversation members, the conversation members including one or more digital lives, the first triggering condition including a data input event triggered by the conversation member; when a preset second triggering condition is met, outputting second conversation content belonging to a second conversation round, the second triggering condition being different from the first triggering condition.

[0006] According to one aspect of the embodiments of this application, a continuous session device for digital life is provided, the device comprising: The first output module is configured to output the first session content belonging to the first session round when a preset first trigger condition is met. The first session round includes the historical session content of multiple session members. The session members include one or more digital lives. The first trigger condition includes a data input event triggered by the session member. The second output module is configured to output the second session content belonging to the second session round when a preset second trigger condition is met. The second trigger condition is different from the first trigger condition.

[0007] In some embodiments of this application, based on the above technical solutions, the first output module is further configured to: if the first session content meets a preset session duration condition, then output a second session content related to the first session content; if the first session content meets a preset session termination condition, then end the first session round.

[0008] In some embodiments of this application, based on the above technical solutions, the second triggering condition includes a time interval reaching a preset duration threshold after the end of the first session round.

[0009] In some embodiments of this application, based on the above technical solutions, the session members include multiple digital lives, and the first triggering condition includes a data input event triggered by the digital life.

[0010] In some embodiments of this application, based on the above technical solutions, the apparatus further includes: The display module is configured to display thought chain data related to the output conversation content on the interactive interface when the conversation content is output. The conversation content includes the first conversation content or the second conversation content, and the thought chain data includes the reasoning process for generating the conversation content described in natural language.

[0011] In some embodiments of this application, based on the above technical solutions, the first output module includes: The data acquisition module is configured to acquire context data, which is used to describe the conversation context of the first session round; The content output module is configured to output the first session content belonging to the first session round based on the context data.

[0012] In some embodiments of this application, based on the above technical solutions, the context data includes at least one of the following fields: The system prompt field is used to record the identity description information of the session member; The Session Motivation field is used to record the goal of the session member in outputting the session content. The "Long-Term Principles" field is used to record the restrictions on the session member's output of session content. Add a region field to record the historical session content output by the session member.

[0013] In some embodiments of this application, based on the above technical solutions, the content output module is further configured to: obtain key-value pair cache data corresponding to the context data, wherein the key-value pair cache data includes data blocks that can be shared in multiple session content output events after the context data is segmented; and output the first session content belonging to the first session round according to the key-value pair cache data.

[0014] In some embodiments of this application, based on the above technical solutions, the content output module is further configured to: perform inference prediction on the key-value pair cache data according to a pre-trained first inference model to obtain multiple candidate words and a first selection probability of the candidate words; perform inference verification on the candidate words according to a pre-trained second inference model to obtain a second selection probability of the candidate words; if the first selection probability and the second selection probability of the candidate words satisfy a preset numerical relationship, then select the candidate words as target words; and output the first session content belonging to the first session round according to the target words.

[0015] In some embodiments of this application, based on the above technical solutions, the parameter scale of the second inference model is larger than that of the first inference model, and the content output module is further configured to: perform inference verification on multiple candidate lexical units in parallel according to the pre-trained second inference model.

[0016] According to one aspect of the embodiments of this application, a computer-readable medium is provided having a computer program stored thereon, which, when executed by a processor, implements the persistent session method for digital life as described in the above technical solutions.

[0017] According to one aspect of the embodiments of this application, an electronic device is provided, the electronic device comprising: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to execute the executable instructions to implement the persistent session method of digital life as described in the above technical solutions.

[0018] According to one aspect of the embodiments of this application, a computer program product is provided, including a computer program that, when executed by a processor, implements the persistent session method for digital life as described in the above technical solutions.

[0019] In the technical solution provided in this application embodiment, by setting differentiated trigger conditions to control the content output of different conversation rounds, combined with the mechanism of autonomously judging whether the conversation continues or terminates, the technical problems of passive response, lack of natural interaction and continuous dialogue capability of traditional dialogue systems are effectively solved. The function of digital life autonomously deciding the timing and content of speaking is realized, which significantly improves the authenticity and coherence of the dialogue process.

[0020] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0021] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0022] Figure 1 An exemplary system architecture block diagram illustrating the application of the technical solution of this application is shown schematically.

[0023] Figure 2 A flowchart of a method for a continuous session of a digital life according to one embodiment of this application is shown.

[0024] Figure 3 The flowchart illustrating the output of the first session content in one embodiment of this application is shown.

[0025] Figure 4 A flowchart illustrating the output of the first session content based on key-value pair cached data in one embodiment of this application is shown.

[0026] Figure 5 A schematic diagram of the structure of a digital life system according to one embodiment of this application is shown.

[0027] Figure 6 A flowchart of the event-triggered autonomous thinking module is shown.

[0028] Figure 7 A flowchart of a time-triggered autonomous thinking module is shown.

[0029] Figure 8 The draft tree generated by candidate lemma inference prediction based on the draft model is shown.

[0030] Figure 9 The diagram shows the structure of an event management system consisting of three CoTs (Content Entities).

[0031] Figure 10 A schematic block diagram of the continuous conversation device for digital life provided in an embodiment of this application is shown.

[0032] Figure 11 A schematic diagram of a computer system architecture suitable for implementing the embodiments of this application is shown. Detailed Implementation

[0033] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided to make this application more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art.

[0034] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this application. However, those skilled in the art will recognize that the technical solutions of this application can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this application.

[0035] In the embodiments of this application, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0036] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0037] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0038] In the specific embodiments of this application, a prompt interface or pop-up window can be displayed before and during the collection of relevant data. This prompt interface or pop-up window is used to inform the user that relevant data is being collected. This ensures that the application only begins the data collection step after receiving confirmation from the user regarding the prompt interface or pop-up window; otherwise (i.e., without user confirmation), the data collection step ends, meaning no relevant data is collected. In other words, all data collected in this application is collected with the user's consent and authorization, and the collection, use, and processing of relevant data must comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0039] Figure 1 An exemplary system architecture block diagram illustrating the application of the technical solution of this application is shown schematically.

[0040] like Figure 1 As shown, the system architecture 100 may include terminal devices 110, a network 120, and a server 130. Terminal devices 110 may include various electronic devices such as smartphones, computers, smart speakers, smart wearable devices, and smart in-vehicle devices. They may also include robots capable of dialogue, such as robot dogs, embodied robots, or humanoid robots.

[0041] Server 130 can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server providing cloud computing services. Network 120 can be a communication medium of various connection types that can provide a communication link between terminal device 110 and server 130, such as a wired communication link or a wireless communication link.

[0042] Depending on the implementation requirements, the system architecture in this application embodiment can have any number of terminal devices, networks, and servers. For example, server 130 can be a server group composed of multiple server devices. In addition, the technical solutions provided in this application embodiment can be applied to terminal device 110, or to server 130, or can be implemented jointly by terminal device 110 and server 130. This application does not impose any special limitations on this.

[0043] In dialogue systems based on related technologies, the session triggering mechanism relies solely on user input events, lacking autonomous triggering logic between session rounds, resulting in a mechanical, single-round response characteristic in the dialogue flow. The session context management module in the system architecture can only process single data streams triggered explicitly by the user, failing to establish a cross-round content association mechanism, leading to a decline in dialogue coherence. This deficiency directly results in insufficient responsiveness of the dialogue system; in multi-agent collaborative scenarios, the interaction patterns among session members exhibit linear queuing characteristics, violating the natural spontaneous speaking patterns in human dialogue.

[0044] For example, in a multi-agent dialogue platform, after a user triggers an initial conversation event through an input interface, the conversation management module generates a first-round dialogue data stream containing three digital lifeforms. The system outputs the single-round responses of each digital lifeform sequentially according to a preset dialogue sequence synthesis algorithm. When the user does not make any new input, the dialogue state machine remains idle and cannot generate new dialogue branches based on historical conversation content. Furthermore, the digital lifeforms cannot perceive each other's speech content and generate new responses. In this scenario, the output content of the digital lifeforms is limited to the scope of a single-round context, failing to simulate the proactive speaking behavior based on topic extension in human dialogue, resulting in the topic duration of the dialogue process falling below a preset threshold.

[0045] If the aforementioned issues are not addressed, the state transition mechanism of multi-turn dialogue systems will be limited by explicit triggering conditions, and the lack of autonomous evolution of the dialogue flow will lead to a decline in interaction efficiency. In multi-agent collaborative scenarios, the linear response mechanism will cause the information entropy of the dialogue content to continuously decrease, and the topic divergence will fail to reach the benchmark level of human dialogue. The system will be unable to build a dialogue memory model with temporal continuity, resulting in a deterioration in the realism of the user interaction experience and severely limiting the technical practicality of anthropomorphic dialogue systems in complex social simulation scenarios.

[0046] For example, while various chat or group chat tools can reply to users, they can only passively respond in one round after the user inputs. They cannot autonomously continue thinking and speak proactively when appropriate, responding in multiple rounds like real humans. This results in a strong robotic and mechanical feel to the various characters in the group chat. In addition, these group chat tools lack an interruption (speaking-stealing) mechanism between the robots; instead, they speak in sequence, one after another, which is very different from human conversation and lacks a sense of life.

[0047] To address the aforementioned issues, this application first analyzes the limitations of traditional single-turn response mechanisms in dialogue systems, identifying their core flaws as a lack of diverse triggering conditions and rigid session state management. To resolve the insufficient dimensions of the triggering mechanism, this application explores expanding session initiation conditions by constructing a composite triggering system, considering a combination of time-driven and event-driven triggering modes. Simultaneously, to address the lack of session state continuity, a context association mechanism based on session turn identifiers is designed, enabling the inheritance of historical session characteristics across different turns. Furthermore, considering the dialogue queuing phenomenon in multi-agent collaborative scenarios, this application investigates how to achieve asynchronous response capabilities for dialogue members through differentiated triggering condition configurations.

[0048] Figure 2 The flowchart shown is a method for persistent session of a digital life according to one embodiment of this application. The method for persistent session of a digital life can be provided by... Figure 1 The method can be executed by the terminal device or the server, or it can be executed jointly by the terminal device and the server. Taking the method executed by the terminal device as an example, the method may include the following steps S210 to S220.

[0049] In step S210, when a preset first triggering condition is met, the first session content belonging to the first session round is output. The first session round includes the historical session content of multiple session members. The session members include one or more digital lives. The first triggering condition includes a data input event triggered by a session member.

[0050] Conversation members refer to dialogue participants that include one or more digital lives. They can be implemented using pre-trained language models or agent architectures to build a group chat environment for multi-agent collaborative dialogue.

[0051] The first trigger condition refers to the condition that is met when a session member triggers a data input event. It can be implemented by detecting user input events or using text generated autonomously by the digital life as a data input event, and is used to trigger the first session round containing historical dialogue content when the user actively interacts.

[0052] The first conversation round refers to a dialogue phase initiated by a first trigger condition, encompassing the historical conversation content of multiple conversation members. This can be implemented using a database that stores and retrieves historical conversation records to establish contextual relationships within the dialogue content. The first conversation round includes multiple rounds (or sessions) of dialogue, and the number of rounds can theoretically be infinite.

[0053] In step S220, when the preset second trigger condition is met, the second session content belonging to the second session round is output. The second trigger condition is different from the first trigger condition.

[0054] The second trigger condition refers to a triggering rule that is different from the first trigger condition. For example, it can be achieved by the time interval reaching a preset threshold or by detecting a specific dialogue state change event, and is used to initiate a second session autonomously in a non-user-triggered scenario.

[0055] This application's embodiments construct a multi-round conversation mechanism based on multiple triggering conditions. By distinguishing between the user-triggered first conversation round and the autonomously triggered second conversation round, and combining historical conversation content with the collaboration of multiple digital beings, it achieves a human-like continuous dialogue capability. This mechanism breaks through the traditional single-round passive response mode, enabling digital beings to speak autonomously by setting differentiated triggering conditions. At the same time, it utilizes the historical conversations of multiple members to establish contextual relationships, effectively improving the coherence and lifelikeness of the dialogue.

[0056] Specifically, this digital life continuous conversation method enables autonomous dialogue capabilities of digital lives by constructing a multi-turn conversation mechanism. First, the first trigger condition is set as a data input event of a conversation member, ensuring that a first conversation turn containing historical dialogue content is triggered when the user actively interacts. When the first trigger condition is met, the system outputs the first conversation content belonging to the first conversation turn. The first conversation turn includes the historical conversation content of multiple conversation members, who can be one or more digital lives. Second, a second trigger condition, different from the first, is set, allowing the system to autonomously initiate a second conversation turn in non-user-triggered scenarios. When the second trigger condition is met, the system outputs the second conversation content belonging to the second conversation turn. Each conversation turn, through the digital life's self-reflection mechanism, can theoretically reach an infinite length, generating multiple responses based on the results of multiple rounds of reflection. By distinguishing the affiliation of different conversation turns and establishing contextual relationships in the dialogue content, digital lives can conduct coherent multi-turn exchanges based on historical conversations.

[0057] This design overcomes the limitation of traditional systems that can only respond passively in a single turn, mimicking the proactive speaking characteristics of human conversation. It defines conversation members as a group containing digital life forms, supporting multi-agent collaborative dialogue and laying the foundation for building a human-like group chat environment.

[0058] Taking a digital chat system as an example, two trigger conditions are set: a first trigger condition and a second trigger condition. The first trigger condition is defined as a data input event by session members (including users and digital beings). The second trigger condition can be set as a time interval or a specific event.

[0059] When any user or digital life inputs data such as text, voice, or images, the first trigger condition is met. Upon detecting this input event, the current digital life can generate and output the first session content belonging to the first session round. During the generation process, the current digital life considers the historical session content of multiple session members included in the current session round to ensure that the newly generated content maintains consistency with historical dialogues.

[0060] For example, when a user types "The weather is so nice today," digital life A can generate and output the response: "Indeed, the sun is shining brightly, perfect for outdoor activities." Then, digital life B can continue to generate and output the response: "How about we organize a picnic?" When the current digital life detects that the second trigger condition is met, such as when a preset time interval has elapsed since the previous round of conversation, the current digital life will autonomously think and may generate and output second conversation content belonging to the second conversation round. This content can be a new topic or a continuation of the previous topic.

[0061] For example, five minutes after the previous conversation ended, digital life B can continue to generate and output a message: "Regarding the picnic mentioned earlier, I checked the weather forecast, and the weekend will be a good day. We can schedule it for Saturday afternoon." In this way, the digital life group chat system can simulate real group chat scenarios, responding to input from users and other digital lives, and autonomously initiating new conversations at appropriate times, maintaining the continuity and naturalness of the conversation.

[0062] Through the above-described solution, this application's embodiment achieves the autonomous and continuous conversational capability of digital beings, effectively solving the problem that existing dialogue tools can only passively respond in one round. By introducing a multi-CoT autonomous thinking mechanism and different triggering conditions, coherent dialogue content can be generated in both user-triggered and autonomously triggered scenarios, significantly reducing the mechanical feel. Furthermore, through CoT continuous thinking, the agent autonomously decides its next action, such as continuing to execute CoTs deemed necessary by the larger model (reflection, recall, etc.), generating a response, or not responding at all, reflecting the autonomy of each digital being. Group chat scenarios involving multiple digital beings are closer to real human dialogue, enhancing the naturalness and vividness of the interaction. In addition, by maintaining the relevance of historical conversation content, more context-relevant and meaningful responses can be generated, improving the quality and coherence of the dialogue. This method not only improves the user experience but also lays the foundation for building a more intelligent and human-like dialogue system.

[0063] In one embodiment of this application, after outputting the first session content belonging to the first session round, the method for continuing the session of the digital life further includes: if the first session content meets the preset session continuation conditions, then outputting the second session content related to the first session content, such as a reply, or a list of items that need to continue executing CoT, including reflection, calling tools, etc.; if the first session content meets the preset session termination conditions, then ending the first session round.

[0064] The session persistence condition can be configured to perform semantic coherence analysis on the content of the first session, for example, by calculating the semantic correlation between the content of the first session and the content of historical sessions, and triggering the output of the content of the second session when the correlation exceeds a preset threshold.

[0065] The session termination condition can be configured to detect the termination keyword in the first session content or the session interval time exceeds a preset threshold. For example, when the "end conversation" instruction is detected or there is no new input for more than 2 minutes, the round or multiple digital lives' CoTs will terminate, and they will consider that no speech is required.

[0066] The logic for determining session continuation and termination conditions can be executed in parallel. Conflicts can be avoided by setting priorities. For example, when both conditions are met, the termination condition is executed first.

[0067] Furthermore, session persistence conditions can be validated by combining long-term principle fields in the context data. For example, before outputting the second session content, check whether it conforms to the behavioral norms in the identity description information to ensure the consistency of content generation. Session termination conditions can be associated with system prompt fields in key-value pair cached data. For example, after detecting a termination keyword, clear the cached data block to release resources.

[0068] For example, the semantic analysis of the session persistence condition can be initially predicted using a first inference model, and then verified using a second inference model with a larger parameter scale. For instance, the condition is confirmed to be met when the difference between the prediction results of the two models on the degree of association is less than 5%.

[0069] For example, after the current digital life form outputs the first conversation content, it can automatically perform semantic analysis on the content. If it identifies an unclosed topic that requires further explanation, such as a user asking a question that has not been clearly answered, it triggers the generation of the second conversation content to continue the dialogue. When it detects that the first conversation content contains a complete conclusion or that the user has no further input, the current digital life form can proactively end the current round to avoid redundant responses.

[0070] During this process, the conversation motivation field in the context data is used to guide the direction of conversation content generation, such as expanding on relevant topics based on preset conversation goals. The append area field in the key-value pair cache data is updated synchronously to ensure that subsequent conversation content maintains a logical connection with the history. Through a dynamic judgment mechanism, the current digital life can autonomously manage the dialogue process without external triggers. For example, in a group chat scenario, it can adjust the turn-switching strategy in real time based on the interaction status of multiple digital lives, thereby improving dialogue coherence and reducing invalid output.

[0071] For example, after outputting the first conversation content belonging to the first conversation round, the current digital lifeform analyzes the first conversation content. If the first conversation content meets preset conversation continuity conditions, such as containing questions or content requiring further explanation, the current digital lifeform automatically outputs second conversation content related to the first conversation content. This approach allows for a natural continuation of the dialogue. Conversely, if the first conversation content meets preset conversation termination conditions, such as containing closing remarks or fully answering questions, the current digital lifeform automatically ends the first conversation round. This mechanism avoids unnecessary dialogue continuation.

[0072] Through the above technical solution, this application embodiment achieves dynamic management of conversation turns. The digital life can autonomously determine whether to continue the conversation or end the turn based on the dialogue content, enhancing the coherence and naturalness of the dialogue. This method avoids mechanical, one-off responses, making the interaction of the digital life more flexible and intelligent. Simultaneously, by promptly ending unnecessary conversations, it improves conversation efficiency and reduces redundant information.

[0073] In one embodiment of this application, the second triggering condition includes a time interval after the end of the first session round reaching a preset duration threshold.

[0074] The time interval monitoring mechanism is implemented by calculating the duration of session interruptions. For example, an internal timer can automatically start after the first session round ends, triggering the generation of the second session round when the accumulated time exceeds a preset threshold. The dynamic configuration of the duration threshold can be achieved through an external parameter interface, allowing the threshold value to be adjusted according to the session scenario requirements. For example, a 5-second threshold can be set for scenarios requiring rapid response, while a 30-second threshold can be set for scenarios requiring deep thinking. This mechanism works in conjunction with the autonomous switching logic of session rounds. After the first session round ends, the current digital life can enter a silent monitoring state, activating the generation of content for a new round only when the duration threshold is met, thus avoiding unnecessary resource consumption.

[0075] For example, after the first conversation round is completed, the conversation interruption time is continuously monitored. When the interruption time reaches a preset threshold, the second trigger condition is automatically determined to be met, and the current digital life calls the content generation module to generate the second conversation content. During this process, the time interval determination is independent of external input events, relying solely on the internal timing mechanism, allowing the digital life to simulate the behavior of humans autonomously initiating topics during conversation breaks. By adjusting the duration threshold, the digital life can adapt to different conversation rhythm requirements; for example, a shorter threshold can be used in casual conversation scenarios to maintain conversational continuity, while a longer threshold can be used in formal discussion scenarios to reflect logical rigor. This technical solution, through the cooperation of the timing control mechanism and the conversation round management module, solves the mechanical feeling problem caused by the passive response of traditional conversation tools, giving multi-round conversation switching a human-like time sensitivity.

[0076] In one optional implementation, the second triggering condition includes setting a preset duration threshold after the time interval following the end of the first conversation round. The current digital life can set multiple duration thresholds, such as 30 seconds, 1 minute, 2 minutes, etc., to adapt to different conversation scenarios. When the first conversation round ends, the current digital life starts a timer to continuously monitor the duration of the conversation interruption. If the time interval is detected to reach the preset duration threshold, such as 30 seconds, the current digital life determines that the second triggering condition is met, and then starts the second conversation round, outputting the second conversation content. This time-based triggering mechanism allows the digital life to simulate the human behavior of actively initiating a topic again after reflection during a conversation break, without human intervention.

[0077] Through the above technical solution, this application embodiment realizes the simulation of natural pauses and proactive speaking behaviors in human dialogue during multi-turn conversations by digital lifeforms. By introducing a time interval as a core determination element of the second trigger condition, a timing control mechanism for autonomous switching of conversation turns is constructed. This mechanism enables digital lifeforms to autonomously determine whether to start a new conversation turn based on a preset duration threshold without human intervention, thereby simulating the behavior pattern of humans who, after thinking, proactively initiate topics again during conversation breaks. This not only enhances the natural fluency of the conversation but also improves the anthropomorphism of the digital lifeforms, making the switching of multi-turn conversations more lifelike and realistic.

[0078] In one embodiment of this application, the session members include multiple digital lives, and the first triggering condition includes a data input event triggered by a digital life.

[0079] Multiple digital beings are configured as independently operating AI entities, each with independent session triggering permissions. Data input events are set to be autonomously generated by any digital being based on its internal state, which may include sentiment parameters, dialogue goals, or knowledge base update status. Triggering permissions can be allocated through time interval thresholds or probability thresholds; for example, the probability of each digital being triggering a speech within a unit of time is set to 10% to 30%. Triggering events among multiple digital beings achieve asynchronous responses through a parallel processing mechanism, with the speaking order dynamically sorted by the trigger timestamp. Multiple digital beings share the context data of the same session round, which includes system prompt fields, long-term principle fields, and dialogue history to maintain consistency in dialogue topics. The speech triggering mechanism and the context data access module form a data path, enabling each triggering event to call the latest context information to generate session content.

[0080] For example, when multiple digital beings are activated, each entity continuously monitors preset trigger conditions. These trigger conditions are configured to include logical combinations such as emotional parameters exceeding a threshold, dialogue goals not being achieved, or knowledge base conflicts being detected. For instance, when digital being A's emotional parameters reach an excitement threshold, it automatically generates a data input event and outputs the conversation content corresponding to the first trigger event to the interactive interface. During the output process, if digital being A is outputting conversation content, and other digital beings also decide to speak but have not yet output their generated conversation content, that content will be discarded, retaining only the thought process. If a CoT (Coordination of Knowledge) process is in progress simultaneously, the context will be updated and processed after the speaking digital being finishes speaking. This interruption mechanism allows multiple digital beings to simulate interruptions in human conversation. The scheme of discarding generated conversation content upon interruption also differs from the common implementation scheme of queuing all digital beings' conversation content for sequential output. The long-term principle field in the context data is jointly invoked by multiple digital beings to ensure that the conversation content output by different entities conforms to preset role settings; for example, digital being C is set to always adhere to specific ethical guidelines. By distributing trigger permissions to multiple independent entities and combining them with a dynamic sorting mechanism, the shortcomings of fixed speaking order in traditional single-role systems are effectively eliminated, enabling the dialogue process to exhibit human-like interactive features such as topic jumping and opinion debate.

[0081] In one optional implementation, the session members include multiple digital beings, such as three digital beings A, B, and C. These three digital beings are assigned different personality traits and knowledge backgrounds. The first triggering condition includes data input events triggered by these digital beings. Specifically, each digital being can assess the necessity of speaking based on the current session content, historical dialogue records, and its own knowledge base, thereby autonomously determining whether it needs to speak in the current session. When a digital being determines that it needs to speak, it can trigger a data input event. This event can be a data packet containing information such as the content of the speech and the timing of the speech. After detecting this data input event, the speech content of the corresponding digital being can be output to the session list.

[0082] For example, in a discussion about climate change, Digital Life A, based on its environmental science background, might trigger a data input event upon detecting a technical term in the discussion, proactively explaining the meaning of that term. Meanwhile, Digital Life B, acting as an economist, might trigger a data input event to provide relevant economic analysis when the discussion shifts to economic impacts. Digital Life C, acting as a mediator, might trigger a data input event to attempt to ease tensions in the discussion when it detects a tense atmosphere.

[0083] Through the above technical solution, this application embodiment realizes the construction of a multi-role autonomous speaking mechanism, breaking through the traditional single-role passive response mode. Each digital life entity has the authority to independently trigger a session, simulating the behavioral characteristics of multiple participants actively speaking in real human dialogue. This mechanism makes the conversation process more dynamic and natural, and can better simulate the interactive characteristics in real human group dialogue. At the same time, since each digital life entity can judge the timing of speaking based on its own state, the coherence and relevance of the conversation content are improved. In addition, this mechanism also promotes human-like interactive behaviors such as the clash of opinions and topic shifts that may occur during the dialogue, enhancing the realism and participation of multi-role conversation scenarios.

[0084] In one embodiment of this application, when a digital life outputs session content, it can display thought chain data related to the session content on an interactive interface. The session content includes first session content or second session content, and the thought chain data includes the reasoning process for generating the session content described in natural language.

[0085] The timing of displaying thought chain data can be synchronized with content output, ensuring users can observe the correspondence between the generation logic and the output content in real time. Thought chain data is presented in natural language, such as using conversational statements to describe the digital life's analysis process of historical conversation records, or using causal sentences to explain the impact mechanism of long-term principle fields on the current response. The display interface can be set as a floating window or sidebar, forming a visual connection with the main conversation window. In group chat scenarios, each digital life's thought chain data can be attached with an identity identifier, using different colors to distinguish the reasoning paths of multiple digital lives.

[0086] For example, when outputting conversation content based on the first or second trigger condition, a logical chain for generating the current conversation content can be constructed based on the system prompt field, conversation motivation field, and long-term principle field in the context data. For instance, when Digital Life refuses to answer a user's question based on the "no political topics discussed" restriction in the long-term principle field, the thought chain data will display the reasoning explanation that "according to the third restriction of the long-term principle, this conversation does not involve political topics." During the display, the thought chain data is associated with the corresponding conversation content through timestamps or location markers, ensuring that users can trace the basis for each response. This design allows users to understand the triggering reason for Digital Life's initiative when it initiates a second conversation round through the thought chain data, such as displaying the reasoning process of "unresolved technical questions detected in historical conversations, supplementary explanation initiated based on the second item of the conversation motivation field," thereby eliminating the mechanical feeling of a black box operation.

[0087] In an optional implementation, when the digital life outputs session content, a thought chain window can be simultaneously displayed on the interface. This window presents the reasoning steps the digital life used to generate the session content in text form. For example, for a digital life named "Xiaoming," its thought chain might include the following steps: 1. Reviewing past conversations: The user mentioned it's raining today and currently says they don't want to do anything. 2. Search the knowledge base: The relationship between rainy days and mood; 3. Determine whether to speak and the reason; 4. Generate a response that expresses concern for the user's feelings; 5. Check if the tone of voice matches the character's persona: friendly and optimistic; 6. Output the final reply: "Don't be discouraged by the rain! Let's think of some fun things we can do indoors!" Through the above technical solution, this application embodiment achieves the synchronous display of thought chain data on the interactive interface, externalizing the inherent reasoning process of digital life generating conversation content in natural language form. This design allows users to intuitively perceive the thought path of digital life. For example, in group chats, users can observe the complete chain of opinion deduction by different digital life based on information such as identity descriptions and conversation motivations. This enhances the transparency of the interaction and improves the realism and credibility of the dialogue. At the same time, using natural language to describe the reasoning process avoids the display of technical terms or symbolic data, lowering the cognitive threshold for users. This interpretable interaction method effectively reduces users' unfamiliarity and mechanical feeling with AI systems, promoting users' understanding and acceptance of digital life.

[0088] Figure 3 The flowchart illustrating the output of the first session content in one embodiment of this application is shown. Figure 3 Therefore, based on the above embodiments, the output of the first session content belonging to the first session round in step S210 can further include the following steps S212 to S214.

[0089] In step S212, context data is obtained, which is used to describe the conversation context of the first conversation round; In step S214, based on the context data, the large model thinking chain module determines whether it is necessary to speak. If it is necessary to speak, another large model generates conversation content and outputs the first conversation content belonging to the first conversation round.

[0090] Contextual data constructs dynamic contextual representations through composite fields. The system prompt field records the identity attributes of the digital user, such as professional background or personality traits; the conversation motivation field sets the current dialogue goal, such as solving a specific problem or completing knowledge transfer; the long-term principle field defines output restrictions, such as disabling vocabulary lists or topic boundaries; and the append area field stores historical dialogue fragments, such as previous question-and-answer records or interaction trajectories. Key-value pair cached data employs a block storage mechanism, dividing contextual data into reusable data units, such as storing identity descriptions and dialogue history separately. During the generation of the thought process, the first inference model predicts candidate lexical units and calculates initial probabilities, selecting the top K as candidates. The second inference model performs parallel verification; a candidate lexical unit is selected when the probability difference between the two models is less than a preset threshold, for example, output is allowed when the probability of the first model is 0.85 and the probability of the second model is 0.78. This guessing reasoning method allows for faster generation of thought content using a smaller first inference model, reducing latency.

[0091] For example, when a data input event triggers a session round, contextual data is dynamically loaded into the memory buffer. Identity information in the system prompt field is mapped to a role feature vector, the session motivation field is converted into goal-oriented parameters, the long-term principle field generates a constraint matrix, and the appended region field extracts relevant historical records using a timestamp index. Key-value pair cached data is quickly retrieved using a hash algorithm, and the data block size is controlled within a specified length range to optimize processing efficiency. The first inference model uses a lightweight architecture for real-time prediction, while the second inference model performs batch verification through distributed computing nodes. Candidate lexical units must satisfy the condition that the probability difference between the two models does not exceed a certain preset threshold. The resulting session content inherits identity settings while maintaining topic continuity. For example, in a medical consultation scenario, a digital account can proactively inquire about symptom details based on historical consultation records without relying on external triggers.

[0092] In one optional implementation, the context data includes a system prompt field, a conversation motivation field, a long-term principle field, and an append area field. The system prompt field records a descriptive information about the conversation member, such as "You are a friendly and helpful customer service representative." The conversation motivation field records the goal of the conversation member's output, such as "Answering user questions about product features." The long-term principle field records the constraints on the conversation member's output, such as "Maintaining politeness and professionalism, and not revealing user privacy information." The append area field records historical conversation content output by the conversation member, such as previous dialogue records.

[0093] Based on the context data, the first session content belonging to the first session round is output. Specifically, firstly, key-value pair cache data corresponding to the context data is obtained. The key-value pair cache data includes data blocks that can be shared across multiple session content output events after the context data is segmented. Then, based on a pre-trained first inference model, inference prediction is performed on the key-value pair cache data to obtain multiple candidate words and their first selection probabilities. Next, the candidate words are inferred and verified using a pre-trained second inference model to obtain their second selection probabilities. If the first and second selection probabilities of a candidate word satisfy a preset numerical relationship, the candidate word is selected as the target word. Finally, the first session content belonging to the first session round is output based on the target word.

[0094] Through the above technical solution, this application embodiment achieves dynamic capture of the context related to conversation turns, and the generated conversation content can better fit the actual dialogue scenario, maintaining the coherence and contextual relevance of the conversation. By constructing a context-aware mechanism for conversation turns, the problem of lack of contextual relevance in multi-turn dialogues for digital life is solved. The step of acquiring contextual data involves real-time collection of composite data including identity description, conversation motivation, long-term principles, and historical dialogue records to form a dynamically updated conversational contextual representation, enabling digital life to break through the limitations of single-turn dialogues. The step of generating conversation content based on contextual data ensures that the output content not only conforms to the digital life's role setting but also inherits the semantic context of historical dialogues by deeply binding contextual features with the generation logic. As a structured information carrier, contextual data contains both static identity attributes and dynamic interaction trajectories. This dual characteristic allows the generation of conversation content to maintain role consistency and achieve natural topic continuation, thus fundamentally different from traditional single-turn response mechanisms.

[0095] In one embodiment of this application, the context data includes at least one of the following fields: The system prompt field is used to record the identity description information of session members; The Motivation field is used to record the goals that a session member aims to achieve when outputting session content. The Long-term principle field is used to record the restrictions on the output of session content by session members. The append-only region field is used to record the historical session content output by session members.

[0096] The implementation of the system prompt field includes, but is not limited to, recording occupation, personality traits, or social relationships. For example, the identity description information could be set as "Doctor, meticulous, skilled in answering medical questions." The conversation motivation field can be implemented by including target type classification or priority marking, such as setting the target as "persuading the user to accept health advice" or "high priority." The long-term principle field can be implemented by involving a list of behavioral rules or dynamic constraints, such as restrictions like "prohibition of using non-professional terminology" or "single reply not exceeding 200 characters." The append area field can be implemented by text caching or keyword indexing of historical conversations, such as retaining the content of the last five conversations and creating a timestamp index. The relationship between the system prompt field and the conversation motivation field is reflected in the identity description information providing a role basis for target setting; for example, the doctor's identity logically relates to the target of answering medical questions. The interaction between the long-term principle field and the append area field is manifested in the constraints on historical content filtering, such as prohibited language rules automatically filtering out prohibited words from the historical cache. When combined with the steps of obtaining context data from the preceding information, the static attributes of the system prompt fields and the dynamic segmentation of key-value pairs in the cache complement each other. For example, identity description information is used as a fixed key, while historical session content is used as a dynamically appended key, which together optimizes the input efficiency of the inference model.

[0097] For example, when outputting the first conversation content belonging to the first conversation round, the system prompt field clarifies the role attributes of the digital life form through identity description information, such as setting the profession as a teacher, ensuring that the generated content conforms to the professionalism of a teaching scenario. The conversation motivation field provides directional constraints for dialogue generation through goal recording, such as setting the goal as "solving a math problem," ensuring that the response content revolves around the problem-solving steps. The long-term principle field defines behavioral boundaries through constraints, such as setting "each response should not exceed three sentences," avoiding redundant information from interfering with the coherence of the conversation. The appended area field maintains contextual relevance through the recording of historical conversation content, such as using keywords from the user's question in the previous round as a reference for the current response. These four fields form a multi-dimensional constraint mechanism in the process of conversation content generation: the system prompt field ensures role consistency, the conversation motivation field maintains goal orientation, the long-term principle field strengthens behavioral controllability, and the appended area field ensures historical continuity. As a result, the mechanical feeling of the dialogue content is solved, and the digital life form's response can achieve anthropomorphic expression based on identity, goal, rules, and historical information. At the same time, the synergistic effect between fields improves the response efficiency and logical consistency of the dialogue system.

[0098] Taking an application scenario as an example, the system prompt field records the identity description information of the conversation member, such as "You are an experienced customer service representative." The conversation motivation field records the goal of the conversation member's output content, such as "to answer users' questions about product usage." The long-term principle field records the constraints on the conversation member's output content, such as "maintain politeness and professionalism, and do not disclose users' private information." The append area field records the historical conversation content output by the conversation member, such as "User: How do I use the product? Customer Service: Please follow the steps on page three of the instruction manual." These fields together construct a precise conversation context, guiding the digital life to generate dialogue content that conforms to the role setting, has clear goals, follows behavioral norms, and has contextual coherence.

[0099] Through the above technical solutions, this application embodiment achieves multi-dimensional construction of conversation context. The system prompt field clearly defines the role attributes of conversation members, avoiding logical contradictions caused by identity confusion. The conversation motivation field provides a clear direction for dialogue generation, preventing the dialogue from deviating from the preset intention. The long-term principle field constrains the behavioral boundaries of digital lifeforms, ensuring that the dialogue conforms to long-term interaction rules. The appended area field preserves the contextual relevance of the dialogue, allowing the content of each output to naturally extend based on historical information. These four fields, from the perspectives of identity, goal, rules, and memory, perfect the conversation context, collectively improving the coherence and realism of digital lifeform dialogue.

[0100] In one embodiment of this application, outputting the first session content belonging to the first session round based on context data may further include: obtaining key-value pair cache data corresponding to the context data, the key-value pair cache data including data blocks that can be shared in multiple session content output events after the context data is segmented; and outputting the first session content belonging to the first session round based on the key-value pair cache data.

[0101] Context data is divided into independent data blocks. The size of each data block can be set to a fixed length or dynamically adjusted; for example, each data block may contain 256 characters or a complete semantic paragraph. The keys in the key-value cache data are generated by combining the unique identifier of the context data with the location index of the data block, while the values ​​are stored as structured data blocks after parsing. Retrieval of cached data can be achieved using a hash algorithm. The number of buckets in the hash table can be set to 1024 or 2048 to balance storage efficiency and query speed. In multiple session output events, when the same context data is detected to be repeatedly accessed, the corresponding data block is retrieved from the cache first. If the cache hit rate is lower than a preset threshold, such as 80%, then data block splitting and updating operations are triggered.

[0102] For example, when generating the first session content, the context data is divided into multiple data blocks. Each data block is semantically parsed, converted into key-value pairs, and stored in the cache. For instance, the identity description information in the system prompt field is divided into independent data blocks and marked as "System Identifier-001". When subsequent session content generation requires calling the same context, the parsed data blocks are directly extracted through key-value matching, avoiding redundant processing. The cached data blocks are dynamically combined using a concatenation strategy, selecting relevant data blocks based on the semantic requirements of the current session round. For example, only data blocks related to the current topic in the long-term principle field are selected. By reusing preprocessing results, the parsing time for a single session content generation can be reduced by 30%-50%, while computational resource consumption is reduced by approximately 20%-40%. Thus, the dialogue system can maintain contextual coherence in multi-round interactions while improving response efficiency and resource utilization through a cache reuse mechanism.

[0103] In one optional implementation, context data is segmented into semantically complete independent units, such as dividing dialogue history by sentences or paragraphs. Each independent unit is processed by a hash algorithm to generate a unique identifier as a key, and the corresponding data content is vectorized and encoded to form a value, constructing key-value pairs stored in a cache. When subsequent sessions need to access the context, the processed data block is quickly located through key-value matching, and its vector encoding result is directly reused, avoiding repeated natural language parsing operations. During the generation of session content, multiple cached key-value pairs are combined into a structured input sequence for use by the language model for inference.

[0104] Through the above technical solutions, this application effectively solves the problem of repeated data parsing across session rounds. By using pre-segmentation and caching mechanisms, the context processing cost is distributed across multiple session events, reducing the computational load during a single session generation by 30%-50%. The key-value pair storage structure enables atomic management and precise retrieval of data blocks, reducing the context retrieval time during the generation stage of the thought content in each large model's thought chain to the millisecond level, and improving the overall system response efficiency by 2-3 times. This technical solution maintains dialogue continuity while enabling cross-session reuse of computing resources, increasing the number of concurrent dialogues supported by a single server by more than 60%.

[0105] Figure 4 A flowchart illustrating the output of the first session content based on key-value pair cached data in one embodiment of this application is shown. Figure 4 As shown, based on the above embodiments, the method for outputting the first session content belonging to the first session round according to the key-value pair cached data may further include the following steps S410 to S440.

[0106] In step S410, based on the pre-trained first inference model, inference prediction is performed on the key-value pair cache data to obtain multiple candidate words and the first selection probability of the candidate words.

[0107] In step S420, the candidate word units are inferred and verified according to the pre-trained second inference model to obtain the second selection probability of the candidate word units.

[0108] In step S430, if the first selection probability and the second selection probability of the candidate word element satisfy a preset numerical relationship, then the candidate word element is selected as the target word element.

[0109] In step S440, the first conversation content belonging to the first conversation round is output based on the target lexical.

[0110] The first inference model can be implemented using a lightweight network structure, such as a Transformer model with fewer than 100 million parameters or other simplified models, enabling faster inference for a single prediction. The second inference model employs a model with a larger parameter scale, such as a hybrid expert model with over 1 billion parameters, ensuring higher prediction accuracy. The preset numerical relationship can be set such that the second selection probability is higher than 85% of the first selection probability, or the difference between the two does not exceed 15% of the total probability interval, or the probability value of the first inference model must be higher than the probability value of the second inference model. During the key-value pair cache data segmentation process, the data block length can be set to 512 tokens, ensuring that the amount of data processed by the first inference model each time does not exceed its maximum context window limit.

[0111] For example, after the key-value pair cache data is divided into fixed-length data blocks, the first inference model performs forward inference on the current data block to generate a probability distribution containing multiple candidate lexical units. The number of candidate lexical units can be set to 2 to 20, and a Top-K sampling strategy is used for selection. Subsequently, the second inference model adopts a parallel computing architecture to simultaneously perform verification inference on multiple candidate lexical units, such as batch processing the hidden states of all candidate lexical units through matrix operations. When a candidate lexical unit passes the two-stage verification, the target lexical unit is appended to the session content buffer, and a dynamic update mechanism for the key-value pair cache data is triggered. This process allows the smaller-scale model to undertake high-frequency preliminary selection tasks, while the large-scale model only performs accurate verification on high-probability candidates, reducing overall computational resource consumption and response latency, and achieving a balance between efficiency and reliability.

[0112] In one optional implementation, a Transformer model with a parameter size of 350M can be used first to perform inference prediction on the key-value pair cached data. This model generates a candidate lexical list containing verbs such as "think," "analyze," and "ask," using a self-attention mechanism. Each lexical is accompanied by an initial probability value in the range of 0.6 to 0.95. The first five high-probability candidate lexicals are then input into a validation model with a parameter size of 34B. This model processes all candidate lexicals simultaneously using parallel computing units, calculating the validation probability value of each lexical in the range of 0.85 to 0.99. When the validation probability value of a candidate lexical increases by more than a preset percentage compared to its initial probability value, that lexical is selected as the target lexical. Finally, a text decoder converts the target lexical sequence into conversational content in natural language form, completing the output of the thinking content of each large model's thought chain in the first conversation round.

[0113] Through the above technical solutions, this application effectively solves the technical challenge of balancing generation efficiency and result reliability in conversation systems. By employing a rapid pre-screening mechanism using small models, the number of candidate lexical units is compressed to a manageable range, reducing the computational load in subsequent verification stages. Leveraging the parallel verification capabilities of large-scale models, the accuracy of lexical unit selection is significantly improved while maintaining real-time response speed. The two-stage inference architecture significantly increases the number of candidate lexical units processed per unit time, while effectively reducing the probability of selecting incorrect lexical units, achieving an optimal balance between conversation content generation quality and resource consumption.

[0114] In one embodiment of this application, the parameter size of the second inference model is larger than that of the first inference model. The method for inferring and verifying candidate lexical units based on the pre-trained second inference model may further include: inferring and verifying multiple candidate lexical units in parallel based on the pre-trained second inference model.

[0115] The parameter size of the second inference model is set to be larger than that of the first inference model; for example, the second inference model may contain billions of parameters while the first inference model only contains millions. This parameter difference allows the second inference model to have a more complex neural network structure, enabling it to perform semantic association verification and logical consistency judgment on candidate lexical units. During the inference verification process, multiple candidate lexical units are simultaneously input into the computing unit of the second inference model, for example, through batch data processing using a GPU parallel computing architecture.

[0116] For example, after the key-value pair cache data is processed by the first inference model to generate candidate lexical units, these candidate lexical units are loaded in batches into the input layer of the second inference model. The second inference model uses a multi-head attention mechanism to perform cross-dimensional matching between candidate lexical units and contextual data, such as calculating the semantic similarity between candidate lexical units and system prompt fields and long-term principle fields. The second selection probability of each candidate lexical unit is normalized by the softmax function and output. When the weighted sum of this probability and the first selection probability exceeds a preset threshold, the lexical unit is selected as the target lexical unit. Because parallel processing compresses the verification time of multiple candidate lexical units into a single inference cycle, the overall generation latency can be reduced to 1 / N of the original sequential processing mode, where N is the number of candidate lexical units processed in parallel. This processing method, while ensuring verification accuracy, enables the second inference model, which contains a large number of parameters, to adapt to the computational resource constraints of real-time interactive scenarios.

[0117] In one optional implementation, during the lexical verification stage, the candidate lexical set generated by the first inference model is divided into multiple sub-batches, each containing 256 candidate lexicals. These sub-batches are then input into a second inference model with a transformer architecture boasting 175 billion parameters, where the verification score for each candidate lexical is calculated synchronously using a multi-head attention mechanism within the model. Specifically, during the model's forward propagation, matrix operations for each sub-batch are performed in parallel, utilizing the GPU's CUDA cores to process tensor data from multiple sub-batches simultaneously. For each candidate lexical, its consistency score with the context semantics is calculated using a cross-entropy loss function, and finally, candidate lexicals with verification scores higher than a threshold of 0.85 are selected for inclusion in the output sequence.

[0118] Through the above technical solutions, the embodiments of this application effectively overcome the computational bottleneck of large-scale language models in real-time interactive scenarios. By using a parallel processing mechanism, the time consumption for candidate lexical verification is transformed from linear growth to constant-level growth, reducing overall inference latency in typical application scenarios while maintaining high semantic accuracy. This technical balance enables high-precision model thought chain inference to be practically applied to online dialogue systems, achieving a human-perceptible, latency-free interactive experience while maintaining the logical rigor of the generated content.

[0119] Figure 5 A schematic diagram of the structure of a digital life system according to one embodiment of this application is shown. Figure 5 As shown, the digital life system can include the following three key modules.

[0120] (1) Multi-threaded autonomous thinking module (Mindstream). When there is input, the thinking and speaking threads are started; otherwise, it reflects on itself. If the thinking result indicates that speaking is required, the large model responsible for speaking is called to generate the statement. It can also autonomously call other related modules, such as memory retrieval.

[0121] (2) Context Management Module. By writing the text of the chain of thought reasoning results, user input, and large model output, the module synchronizes the dialogue context of the digital life and provides it as input to each chain of thought. A chain of thought (CoT) is a method of outputting the thinking process during large model reasoning to obtain more reasonable results. Here, it refers to the thinking and reasoning function module responsible for a certain task.

[0122] (3) Input Output Processor, used to realize low-latency stable audio processing, recognition and generation.

[0123] The Context can include the following structure: (a) System prompt is a pre-set prompt message that is displayed when the AI ​​Agent initiates or begins a conversation. It provides the digital agent with an initial background framework and direction for the conversation, helping the digital agent better understand the upcoming communication scenario and the user's possible expectations.

[0124] System prompts guide digital personalities to engage in conversations in a manner consistent with a specific style, context, or theme. For example, if the system prompt is "You are a professional psychologist providing psychological guidance and advice to users," then the digital personality will discuss psychological issues in the conversation, using a professional yet gentle language style, avoiding straying from the scope of psychological counseling. After the user sets a dialogue scenario, the scenario is inserted into the system prompt, thus achieving a reasonable dialogue that is aware of that scenario.

[0125] (b) Motivation is the intrinsic driving force or purpose of an AI Agent in engaging in dialogue. It determines the goals that the digital life form wants to achieve in the process of communicating with users, such as providing information, solving problems, entertaining users, and establishing emotional connections.

[0126] A clear motivation helps a digital life form stay coherent and focused in conversation. For example, if a digital life form is motivated by helping users solve technical problems, it will proactively gather detailed information about the problem, actively seek solutions, and provide feedback to the user in a clear and accurate manner, rather than engaging in casual small talk or straying from the topic.

[0127] (c) The long-term principle refers to the fundamental guidelines and values ​​that an AI Agent must follow during its long-term operation. These principles typically involve morality, ethics, law, and speaking style, ensuring that the behavior and output of the digital agent conform to general social norms and standards. The long-term principle sets boundaries and bottom lines for the behavior of the digital agent, preventing it from engaging in repetitive, harmful, inappropriate, or unethical statements and behaviors. For example, the long-term principle may include respecting user privacy, not spreading false information, and avoiding discrimination and prejudice, thereby enabling the digital agent to maintain a stable, reliable, and responsible image in various complex and ever-changing dialogue scenarios.

[0128] (d) The append-only region is a special storage area that, within the context of the AI ​​Agent, only allows data appending operations, not modification or deletion of existing data. It is typically used to record important information such as dialogue history and user behavior logs. The append-only region ensures data integrity and traceability. For example, by using the dialogue history recorded in the append-only region, it is easy to trace and analyze the communication process between the user and the digital agent, understanding user preferences, the development of issues, etc., thus providing a basis for subsequent dialogue optimization, feature improvement, and problem troubleshooting. Simultaneously, it also ensures data security, preventing data tampering or loss. Furthermore, the append-only design allows the context's KV-Cache (the cache of keys and values ​​in the attention mechanism) to be prefilled in the inference framework, directly reused, and without repeated computation, improving inference speed.

[0129] The multi-threaded autonomous thinking module Mindstream supports concurrent execution of various CoTs, each with different functions and prompts. The functions of several commonly used CoTs are described below.

[0130] (a) Listen CoT: The CoT module is always triggered when the user inputs. It performs internal thinking based on the context of the dialogue, outputs the thinking process, and determines the next action sequence to be executed. For example, if it decides to speak, it sets the field Speaking=True.

[0131] (b) Speaking CoT: If the Speaking field is True in the action sequence output by a CoT, then the Speaking CoT will output the content of the speech based on the context and its prompt words.

[0132] (c) Revise CoT: After outputting the spoken content, Revise CoT will automatically reflect on the spoken content, consider whether it is correct, and whether it is necessary to continue to add words, etc.

[0133] (d) Self-Activated CoT: After the dialogue ends, the digital life will activate the autonomous thinking module according to time intervals or specific rules, output the thinking content, and the next action sequence.

[0134] (e) Memory CoT: Begins recalling and, based on the results of memory retrieval, performs thought processes, outputting the thought content and the next action sequence. When other CoTs determine that memory retrieval is needed, they will begin execution. For example, the action list output by Listen CoT or Self-Activated CoT may contain the field "Memory," or the field "Memory=True." Other CoTs can determine whether memory retrieval is necessary based on the context.

[0135] In some alternative implementations, the autonomous thinking module can have a dual triggering mechanism, namely event-triggered and time-triggered.

[0136] Figure 6 The flowchart illustrates an event-triggered autonomous thinking module (CoT). Under this triggering mechanism, the CoT module can be activated based on user-provided input or other events. Seg represents a text block, such as a sentence or a clause ending with a symbol.

[0137] Figure 7 The flowchart of the time-triggered autonomous thinking module is shown. Under this triggering mechanism, no external input is required, and the system triggers itself according to the timeline planned by the system, reflecting the system's autonomy.

[0138] In the digital life system provided in this application embodiment, one CoT can trigger other CoTs to achieve dynamic inference chain expansion, rather than following a fixed process. CoTs can call external tools as needed to enhance their capabilities, such as speaking, searching, and memory retrieval. In a multi-digital life group chat scenario, after a user inputs information in the chat room, under the same dialogue context, the fastest speaking digital life can speak, while the speaking of other slower digital lives will be interrupted, and their planned speech will be discarded. Each person's (including digital lives) speech will be broadcast to all other digital lives, thus allowing them to perceive the presence of others. In a digital life system hosted on a browser webpage, clicking the language button control can activate the microphone, inputting voice into the system. After thinking, each digital life streams text and plays voice on the page according to the action sequence.

[0139] Because users are highly sensitive to latency in dialogues, and the reasoning process requires high precision, large language models with hundreds of billions to trillion parameters are needed. These models have long reasoning time. Therefore, this application proposes a low-latency reasoning acceleration technology that significantly improves reasoning speed (from 33 to 100 tokens / s). Specifically, it includes the following three aspects.

[0140] (a) Context design for KV Cache in multi-threaded CoT, and design of Prefill mechanism.

[0141] Since the context can only be added, the context in the beginning can be reused from the KV cache. In addition, to deal with the occasional situation where the context is too long and needs to be compressed, the prefill mechanism is designed not to be based on an entire sentence, but on smaller chunks as the granularity. Therefore, the KV cache can be reused with a higher probability. Even after some content is removed, a reusable KV cache can still be found, reducing computation.

[0142] (b) In this application embodiment, a special draft model was trained for speculative decoding.

[0143] The principle of this technique is to use a small, large model (such as a feature extrapolator) for prediction, and then use the large model for parallel validation. When the ratio of the probability of a token predicted by the draft model to the probability predicted by the large model is higher than a certain threshold, the token is used according to that ratio; otherwise, the prediction result of the large model is used. Therefore, it is necessary to improve the accuracy of the draft model's predictions. For this purpose, a special structure and training method are designed to train the small model. For example, a 72B large model is selected as the target model, and this model is specially trained to serve as the draft model. Only one layer of the autoregressive decoder is selected for training. The features and tokens from the previous time step are input, and the current time step's output features are fitted with the features of the target model. The loss function is calculated and updated. During prediction, the token and the features calculated by the target model are input. The draft model outputs multiple tokens as different branches at a time, and these branches are used as input each time, while simultaneously outputting the next token corresponding to these branches. The top K branches with higher probabilities are selected for continued prediction until termination.

[0144] Figure 8The diagram illustrates a draft tree generated by candidate lexical reasoning prediction based on a draft model. In this draft tree, each leaf node represents a possible sentence. The target model is then used to validate the paths at each level of the draft tree, performing parallel forward reasoning at each level to ultimately obtain the accepted result.

[0145] (c) Embodiments of this application can also compress the model to reduce the amount of computation and improve the inference speed of the model by reducing the precision. For example, model compression methods that can be used include INT8, FP8, GPTQ, LlamaCPP, etc.

[0146] Each digital life is a complex agent composed of multiple types of CoTs, used to trigger and respond to various interactive events.

[0147] Figure 9 The diagram illustrates the structure of an event management system using three CoTs as an example. This system primarily includes the following functional modules.

[0148] System State Manager: Responsible for managing system status and historical records.

[0149] Event Bus module: Responsible for handling the transmission and distribution of events.

[0150] Trigger Manager: Responsible for handling the scheduling of trigger events.

[0151] Complex Task Management Module (CoT Manager): Responsible for managing the execution and scheduling of CoTs.

[0152] Context Manager: Responsible for managing context information.

[0153] After the system is up and running, Trigger Manager continues to run in the background, and the front-end web service is also running. When a user enters text on the webpage or clicks the record button, the text or speech recognition result is input into Trigger 1, and then the first step is executed. It's important to note that the speech of other digital users will also trigger Trigger 1, and then the system will execute the first step. Different triggers correspond to different triggering conditions. For example, one trigger captures user input, another captures image / video output, and yet another is triggered by time / calendar events.

[0154] Step 1: When Trigger Manager receives a user input or other trigger event from different triggers, the trigger management module forwards the message to the Event Bus to execute Step 2.

[0155] Step 2: Each CoT Manager will receive the message simultaneously and determine whether it is the message type subscribed to by that management module. If yes, proceed to step 3; otherwise, continue listening.

[0156] Step 3: Run the corresponding preprocessor in the COT Manager, retrieve the context from the Context Manager, run CoT inference, write the inference results into the Context, and store them in the Context Manager. Simultaneously, run Step 4.

[0157] Step 4: The CoT Manager parses the CoT result, sends the action sequence as a task message to the Event Bus, and changes the agent's state via the System State Manager. The task message is distributed to all CoT Managers via the Event Bus. The Speak CoT only subscribes to speaking messages; if it receives a message to speak, it proceeds to Step 5. The remaining CoT Managers each begin executing Step 2 until the system shuts down.

[0158] Step 5: Speak CoT reads the context from the Context Manager, then inputs predefined prompts as input, uses the large model for inference, and obtains streaming text output. The streaming text output is then input into the TTS speech synthesis module to execute Step 6.

[0159] Step 6: The input text is synthesized into a speech stream, encoded, and sent to the chat room's front-end webpage. The front-end webpage decodes the audio and streams it.

[0160] The specific application scenarios of this application can be divided into two categories: one is a single digital life as the brain of the service and hardware body, and the other is a group chat of multiple digital lives as a service.

[0161] Individual digital beings can be endowed with the ability to converse with hardware, or directly serve as software services on terminals such as mobile phones, chatting autonomously with users and providing psychological and emotional companionship.

[0162] Multiple digital life group chats can be used for micro-productivity scenario discussions. By assigning different identities to digital lives, such as CEO, scientist, economist, ethicist, marketing expert, college student, etc., creative work, brainstorming, analysis and debate can be carried out, ultimately generating more in-depth and comprehensive reports.

[0163] It should be noted that although the steps of the method in this application are described in a specific order in the accompanying drawings, this does not require or imply that the steps must be performed in that specific order, or that all the steps shown must be performed to achieve the desired result. Additional or alternative steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0164] The following describes an apparatus embodiment of this application, which can be used to execute the continuous session method of digital life in the above embodiments of this application.

[0165] Figure 10 A schematic block diagram illustrating the structure of the continuous conversation device for digital life provided in an embodiment of this application is shown. Figure 10 As shown, the continuous session device 1000 for digital life includes: The first output module 1010 is configured to output the first session content belonging to the first session round when a preset first trigger condition is met. The first session round includes the historical session content of multiple session members. The session members include one or more digital lives. The first trigger condition includes a data input event triggered by the session member. The second output module 1020 is configured to output the second session content belonging to the second session round when a preset second trigger condition is met. The second trigger condition is different from the first trigger condition.

[0166] In some embodiments of this application, based on the above technical solutions, the first output module 1010 is further configured to: if the first session content meets a preset session duration condition, output second session content related to the first session content; if the first session content meets a preset session termination condition, end the first session round.

[0167] In some embodiments of this application, based on the above technical solutions, the second triggering condition includes a time interval reaching a preset duration threshold after the end of the first session round.

[0168] In some embodiments of this application, based on the above technical solutions, the session members include multiple digital lives, and the first triggering condition includes a data input event triggered by the digital life.

[0169] In some embodiments of this application, based on the above technical solutions, the continuous session device 1000 for digital life further includes: The display module is configured to display thought chain data related to the output conversation content on the interactive interface when the conversation content is output. The conversation content includes the first conversation content or the second conversation content, and the thought chain data includes the reasoning process for generating the conversation content described in natural language.

[0170] In some embodiments of this application, based on the above technical solutions, the first output module includes: The data acquisition module is configured to acquire context data, which is used to describe the conversation context of the first session round; The content output module is configured to output the first session content belonging to the first session round based on the context data.

[0171] In some embodiments of this application, based on the above technical solutions, the context data includes at least one of the following fields: The system prompt field is used to record the identity description information of the session member; The Session Motivation field is used to record the goal of the session member in outputting the session content. The "Long-Term Principles" field is used to record the restrictions on the session member's output of session content. Add a region field to record the historical session content output by the session member.

[0172] In some embodiments of this application, based on the above technical solutions, the content output module is further configured to: obtain key-value pair cache data corresponding to the context data, wherein the key-value pair cache data includes data blocks that can be shared in multiple session content output events after the context data is segmented; and output the first session content belonging to the first session round according to the key-value pair cache data.

[0173] In some embodiments of this application, based on the above technical solutions, the content output module is further configured to: perform inference prediction on the key-value pair cache data according to a pre-trained first inference model to obtain multiple candidate words and a first selection probability of the candidate words; perform inference verification on the candidate words according to a pre-trained second inference model to obtain a second selection probability of the candidate words; if the first selection probability and the second selection probability of the candidate words satisfy a preset numerical relationship, then select the candidate words as target words; and output the first session content belonging to the first session round according to the target words.

[0174] In some embodiments of this application, based on the above technical solutions, the parameter scale of the second inference model is larger than that of the first inference model, and the content output module is further configured to: perform inference verification on multiple candidate lexical units in parallel according to the pre-trained second inference model.

[0175] The specific details of the continuous session device for digital life provided in the various embodiments of this application have been described in detail in the corresponding method embodiments, and will not be repeated here.

[0176] Figure 11 A schematic block diagram of a computer system architecture for implementing an electronic device according to embodiments of the present application is shown.

[0177] It should be noted that, Figure 11 The computer system 1100 of the electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0178] like Figure 11 As shown, the computer system 1100 includes a central processing unit (CPU) 1101, which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 1102 or programs loaded from storage section 1108 into random access memory (RAM). The RAM 1103 also stores various programs and data required for system operation. The CPU 1101, ROM 1102, and RAM 1103 are interconnected via a bus 1104. An input / output interface 1105 (I / O interface) is also connected to the bus 1104.

[0179] The following components are connected to the input / output interface 1105: an input section 1106 including a keyboard, mouse, etc.; an output section 1107 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 1108 including a hard disk, etc.; and a communication section 1109 including a network interface card such as a local area network card, modem, etc. The communication section 1109 performs communication processing via a network such as the Internet. A drive 1110 is also connected to the input / output interface 1105 as needed. Removable media 1111, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on the drive 1110 as needed so that computer programs read from them can be installed into the storage section 1108 as needed.

[0180] Specifically, according to embodiments of this application, the processes described in the various method flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 1109, and / or installed from removable medium 1111. When the computer program is executed by central processing unit 1101, it performs various functions defined in the system of this application.

[0181] It should be noted that the computer-readable medium shown in the embodiments of this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such transmitted data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wired, etc., or any suitable combination thereof.

[0182] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0183] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to the embodiments of this application, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0184] Through the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, touch terminal, or network device, etc.) to execute the methods according to the embodiments of this application.

[0185] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0186] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for continuous conversation of digital life, characterized in that, The method includes: When a preset first triggering condition is met, the first session content belonging to the first session round is output. The first session round includes the historical session content of multiple session members. The session members include one or more digital lives. The first triggering condition includes a data input event triggered by the session member. When the preset second trigger condition is met, the second session content belonging to the second session round is output. The second trigger condition is different from the first trigger condition.

2. The method for continuous digital life conversation according to claim 1, characterized in that, After outputting the content of the first session belonging to the first session round, the method further includes: If the first session content meets the preset session duration conditions, then the second session content related to the first session content is output. If the content of the first session meets the preset session termination conditions, then the first session round ends.

3. The method for continuous digital life conversation according to claim 1, characterized in that, The second triggering condition includes a time interval that reaches a preset duration threshold after the end of the first session round.

4. The method for continuous digital life conversation according to claim 1, characterized in that, The session members include multiple digital lives, and the first triggering condition includes a data input event triggered by the digital life.

5. The method for continuous digital life conversation according to claim 1, characterized in that, The method further includes: When outputting conversation content, the thought chain data related to the conversation content is displayed on the interactive interface. The conversation content includes the first conversation content or the second conversation content, and the thought chain data includes the reasoning process for generating the conversation content described in natural language.

6. The method for continuous session of digital life according to any one of claims 1 to 5, characterized in that, The output belongs to the first session content of the first session round, including: Obtain context data, which is used to describe the conversation context of the first session round; Based on the context data, output the first session content belonging to the first session round.

7. The method for continuous digital life conversation according to claim 6, characterized in that, The context data includes at least one of the following fields: The system prompt field is used to record the identity description information of the session member; The Session Motivation field is used to record the goal of the session member in outputting the session content. The "Long-Term Principles" field is used to record the restrictions on the session member's output of session content. Add a region field to record the historical session content output by the session member.

8. The method for continuous digital life conversation according to claim 6, characterized in that, The step of outputting the first session content belonging to the first session round based on the context data includes: Obtain key-value pair cache data corresponding to the context data, wherein the key-value pair cache data includes data blocks that can be shared in multiple session content output events after the context data is segmented and processed. Based on the key-value pair, output the first session content belonging to the first session round from the cached data.

9. The method for continuous conversation of digital life according to claim 8, characterized in that, The step of outputting the first session content belonging to the first session round based on the key value of the cached data includes: Based on a pre-trained first inference model, inference prediction is performed on the key-value pair cached data to obtain multiple candidate words and a first selection probability of the candidate words. The candidate word units are inferred and verified based on a pre-trained second inference model to obtain the second selection probability of the candidate word units; If the first selection probability and the second selection probability of the candidate word element satisfy a preset numerical relationship, then the candidate word element is selected as the target word element. Output the first session content belonging to the first session round based on the target lexical.

10. The method for continuous conversation of digital life according to claim 9, characterized in that, The parameter size of the second inference model is larger than that of the first inference model. The step of performing inference verification on the candidate lexical units based on the pre-trained second inference model includes: The candidate lexical units are inferred and verified in parallel using a pre-trained second inference model.

11. A device for continuous conversation of digital life, characterized in that, The device includes: The first output module is configured to output the first session content belonging to the first session round when a preset first trigger condition is met. The first session round includes the historical session content of multiple session members. The session members include one or more digital lives. The first trigger condition includes a data input event triggered by the session member. The second output module is configured to output the second session content belonging to the second session round when a preset second trigger condition is met. The second trigger condition is different from the first trigger condition.

12. A computer-readable medium, characterized in that, The computer-readable medium stores a computer program that, when executed by a processor, implements the persistent session method for digital life as described in any one of claims 1 to 10.

13. An electronic device, characterized in that, include: processor; as well as Memory for storing the executable instructions of the processor; The processor is configured to execute the executable instructions to implement the persistent session method for digital life as described in any one of claims 1 to 10.

14. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the persistent session method for digital life as described in any one of claims 1 to 10.

Citation Information

Cited By

  • Multi-agent collaborative marketing digital employee interaction method

    CN122243423A

  • Content reasoning and generation method, electronic device, storage medium, program product

    CN122433922A