Method for enhancing ability of character large model personality and task linkage and related product

Abstract

The application belongs to the technical field of artificial intelligence and relates to a role large model personality and task linkage ability enhancement method and related products. According to user input natural language task instructions and personality setting information, a role large model generates a subtask sequence and an ability requirement set. For each subtask, if the subtask and the corresponding target ability match the personality setting information, the role large model calls the target ability to execute the subtask to obtain a subtask result. If the subtask and the target ability do not match, the subtask and the target ability are subjected to personality adaptation reconstruction before execution to obtain a subtask result. According to the personality setting information, the subtask result is subjected to personality polishing to obtain a personalized subtask result. The personalized subtask results are integrated in sequence into a complete task result, which is subjected to personality polishing again to obtain a personalized complete task result that is output to the user. The application can improve the task execution ability of the role large model while maintaining personality consistency.

Description

Methods to enhance the ability of character models, personality, and task synergy, and related products. Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, computer device, and computer-readable storage medium for enhancing the ability of a large character model personality and task linkage. Background Technology

[0002] Currently, character-based artificial intelligence still faces a technical bottleneck: it is difficult to reconcile capabilities with personality. This bottleneck manifests itself in the following two aspects:

[0003] Distinctive Personalities, Limited Capabilities: Many chatbots with defined personalities (such as virtual idols and companion assistants) possess lively and engaging speech styles, but their functionality is limited to casual conversation and question-and-answer sessions. They are unable to truly complete complex tasks assigned by users, resulting in insufficient practicality. For example, a virtual assistant engages with users in a cute manner, but when asked to compile financial statements, it can only apologize and say "I cannot do it," leaving the user's needs unmet. Another example is an educational chatbot claiming to have a teacher's personality, capable of humorously explaining knowledge, but when students request answers to advanced math problems or to debug program code, it can only offer vague suggestions or outright refuse, highlighting the current inadequacy of personalized chatbots in task execution. In short, the phenomenon of "personalized but incompetent" chatbots is currently widespread—chatbots with distinct personalities lack practical problem-solving abilities.

[0004] Powerful but devoid of personality: On the other hand, some large, autonomous agent-type models can invoke tool plugins to perform tasks, alleviating the limitations of single-model capabilities to some extent. However, once these large models enter multi-step task processes, their output often becomes mechanical and rigid. For example, users might expect a large model to help with research and summarizing in a humorous and engaging way, but in actual use, the model's response becomes a list of steps in a rambling manner, completely lacking any human touch. The interactive experience feels like a cold, robotic robot executing commands. Furthermore, some well-known dialogue models, after integrating plugin functionality, have seen user feedback that their response style becomes formulaic, lacking the humor and friendliness of their original persona. The lack of a personality maintenance mechanism causes large models to often abandon their personality when performing tasks, significantly diminishing the interactive experience; at the same time, plugin calls without personality constraints may also trigger inappropriate behavior, posing ethical and safety risks.

[0005] In summary, existing role-based intelligent agent systems exhibit a structural contradiction between complex task processing and maintaining role consistency: one type of system emphasizes dialogue style and role setting, lacking a scalable external capability orchestration mechanism, making it difficult to achieve stable and acceptable results in multi-step tasks; another type of system focuses on tool / plugin invocation to improve task coverage, but lacks continuous constraints on role behavior strategies and language style during multiple rounds of planning and execution, easily leading to drift in output tone and behavioral choices, thereby reducing interaction consistency and risk controllability. Therefore, a technical solution is needed that allows personality setting to be used as a constraint variable and form a closed-loop control at each stage of task decomposition, capability selection, execution, and result expression, in order to simultaneously improve the ability to complete complex tasks and maintain role consistency. Summary of the Invention

[0006] This invention provides a method, apparatus, computer device, and computer-readable storage medium for enhancing the ability of a large character model to link personality and task, which can improve the task execution ability of a large character model while maintaining its personality consistency.

[0007] Firstly, the method for enhancing the ability of a large-scale character model's personality and task linkage provided by this invention includes:

[0008] Receive natural language task instructions and personality setting information input by the user;

[0009] Based on natural language task instructions and personality setting information, a series of sub-tasks and a set of capability requirements corresponding to the sub-tasks are generated through the character big model. The set of capability requirements includes the target capabilities corresponding to each sub-task in the sub-task sequence. The target capabilities include the model capabilities inside the character big model or the external plug-in capabilities.

[0010] For each subtask, calculate the matching score between the subtask and its corresponding target ability and personality setting information, and perform double verification based on the score threshold and the ability blacklist: if the matching score reaches the score threshold and the target ability is not in the ability blacklist, then the subtask is executed by calling the target ability through the character big model to obtain the subtask result; if not, then the subtask and the target ability are reconstructed according to the personality setting information to make the reconstructed subtask and the reconstructed target ability meet the score threshold and not hit the ability blacklist before being executed to obtain the subtask result;

[0011] The results of each subtask are personalized based on the personality setting information to obtain personalized subtask results. After all subtasks are completed, the personalized subtask results are integrated into the complete task results in sequence. The complete task results are then personalized again to obtain personalized complete task results and output them to the user.

[0012] Specifically, when the target capability is a plugin capability, hash digest verification and signature certificate verification are performed on the target plugin corresponding to the plugin capability. After both verifications pass, the target plugin is loaded into the sandbox environment and granted the minimum necessary permissions in a hierarchical permission manner for the role-based large model to call. Audit logs are generated for the loading and calling process of the target plugin for traceability and security auditing.

[0013] Secondly, the device for enhancing the ability of a large character model's personality and task linkage provided by the present invention includes:

[0014] The task receiving module is used to receive natural language task instructions and personality setting information input by the user;

[0015] The task planning module is used to generate a sequence of sub-tasks and a set of capability requirements corresponding to the sub-tasks based on natural language task instructions and personality setting information through the character big model. The set of capability requirements includes the target capabilities corresponding to each sub-task in the sub-task sequence. The target capabilities include the model capabilities inside the character big model or the external plug-in capabilities.

[0016] The task execution module is used to calculate the matching score between the subtask and its corresponding target ability and personality setting information for each subtask, and to perform double verification based on the score threshold and the ability blacklist: if the matching score reaches the score threshold and the target ability is not in the ability blacklist, the subtask is executed by calling the target ability through the character big model to obtain the subtask result; if not, the subtask and target ability are reconstructed according to the personality setting information to meet the score threshold and not hit the ability blacklist before the subtask result is obtained.

[0017] The results integration module is used to personalize the results of each subtask according to the personality setting information to obtain personalized subtask results. After all subtasks are completed, the personalized subtask results are integrated into the complete task results in sequence. The complete task results are then personalized again to obtain personalized complete task results and output them to the user.

[0018] When the target capability is a plugin capability, the task execution module performs hash digest verification and signature certificate verification on the target plugin corresponding to the plugin capability. After both verifications pass, the target plugin is loaded into the sandbox environment and granted the minimum necessary permissions in a hierarchical permission manner for the role-based large model to call. Audit logs are generated for the loading and calling process of the target plugin for traceability and security auditing.

[0019] Optionally, in one embodiment, the task execution module is used to encode the personality setting information into a personality vector representation, encode the subtask and its corresponding target ability into a task ability vector representation, and calculate the matching score between the personality vector representation and the task ability vector representation.

[0020] Optionally, in one embodiment, the ability enhancement device for linking the character model personality and task provided by the present invention further includes a personality maintenance module, which is used to store the personality vector representation in the personality memory bank for use in subsequent tasks; obtain feedback information from users on the results of the personalized complete task, extract the personalized preferences implied in the feedback information, and update the corresponding personality vector representation in the personality memory bank according to the personalized preferences.

[0021] Optionally, in one embodiment, the personality maintenance module is further configured to, when a preset correction cycle is reached, obtain the current personality vector representation in the personality memory bank and calculate its deviation value from the initial personality vector representation; if the deviation value reaches the correction threshold, then execute personality correction measures, which include injecting personality correction prompts corresponding to personality setting information into the character model, and / or obtaining training samples corresponding to personality setting information and fine-tuning the character model based on the training samples.

[0022] Optionally, in one embodiment, the task execution module is used to replace the keywords in the subtask result with semantically equivalent words that conform to the personality setting information to obtain the replaced subtask result; generate personality polishing prompts according to the personality setting information, and input the personality polishing prompts and the replaced subtask result together into the character big model, and perform personality polishing on the replaced subtask result through the character big model to obtain the personality subtask result.

[0023] Thirdly, the computer device provided by the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the method for enhancing the ability of the large-scale character model personality and task linkage provided by the present invention.

[0024] Fourthly, the computer-readable storage medium provided by the present invention stores a computer program, which, when executed by a processor, implements the method for enhancing the ability of large-scale character model personality and task linkage provided by the present invention.

[0025] This invention provides a capability enhancement scheme for the linkage between personality and task in a large character model. It receives natural language task instructions and personality setting information input by the user; based on these instructions and information, the large character model generates a sequence of sub-tasks and a corresponding set of capability requirements. The capability requirement set includes target capabilities for each sub-task in the sequence, and these target capabilities can be either internal model capabilities or external plugin capabilities. For each sub-task, a matching score is calculated between the sub-task, its corresponding target capability, and the personality setting information. This matching score is then double-checked based on a score threshold and a capability blacklist: if the matching score reaches the threshold and the target capability is not in the blacklist, the large character model invokes the target capability to execute the sub-task and obtain the sub-task result; otherwise, the sub-task is adjusted according to the personality setting information. The task and target capabilities undergo personality-adaptive reconstruction. The reconstructed sub-tasks and target capabilities meet scoring thresholds and do not hit the capability blacklist before execution to obtain sub-task results. Each sub-task result is then refined according to personality settings to obtain a personalized sub-task result. After all sub-tasks are completed, the personalized sub-task results are sequentially integrated into a complete task result. This complete task result is then refined again to obtain a personalized full task result, which is output to the user. Specifically, when the target capability is a plugin capability, hash digest verification and signature certificate verification are performed on the target plugin corresponding to that capability. After both verifications pass, the target plugin is loaded into a sandbox environment and granted the minimum necessary permissions using a tiered access control approach for use by the large role model. Audit logs are generated for the loading and invocation processes of the target plugin for traceability and security auditing. This achieves simultaneous task execution and personality maintenance, ensuring that the large role model strictly adheres to its personality settings at every stage of task completion, thereby improving the large role model's task execution capabilities while maintaining its personality consistency. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 is a schematic diagram of the architecture of the character large model personality and task linkage capability enhancement system provided in an embodiment of the present invention;

[0028] Figure 2 is a flowchart illustrating the method for enhancing the ability of character model personality and task linkage provided in an embodiment of the present invention;

[0029] Figure 3 is a schematic diagram of the ability enhancement device for character large model personality and task linkage provided in an embodiment of the present invention;

[0030] Figure 4 is a schematic diagram of the structure of the computer device provided in an embodiment of the present invention. Detailed Implementation

[0031] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0032] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0033] It should also be understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0034] As used in this specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when the condition is met," "once the condition is met," or "in response to determining / detecting the condition or event," to indicate that a corresponding step is performed when the triggering condition is met, and its meaning should not be construed as requiring a strict temporal order, an exclusive condition, or a necessary causal relationship.

[0035] Similarly, the phrases “if determined” or “if [the described condition or event] is detected” can be interpreted, depending on the context, as meaning “once determined” or “in response to determined” or “once [the described condition or event] is detected” or “in response to the detection of [the described condition or event]”.

[0036] Furthermore, in the description of this invention and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0037] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of the invention include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0038] It should be understood that the sequence number of each step in the following embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0039] Referring to Figure 1, this invention provides a capability enhancement system for the linkage between personality and tasks in a large character model. It includes a task analysis and planning module 1, a plugin management module 2, and a personality control module 3, as well as a user input interface 4, a user output interface 5, and a plugin repository 6. The task analysis and planning module 1 parses user-input natural language task instructions and generates a sequence of sub-tasks and their corresponding capability requirements based on personality setting information. The plugin management module 2 maintains a list of available plugins and interacts with the plugin repository 6, dynamically loading or unloading plugins as needed based on the capability requirement set, providing external capability expansion for the large character model, while simultaneously executing the sub-task sequence in conjunction with internal capabilities. The personality control module 3 operates throughout the entire task execution process, working collaboratively with a personality filter, personality refiner, and personality memory bank. It filters or adjusts sub-tasks or capabilities that do not conform to personality setting information, refines the results of tasks at each stage, and records the personality preferences formed by long-term character personality settings and user feedback for task planning by the task analysis and planning module. The user input interface 4 receives user-input natural language task instructions and personality setting information and transmits them to the task analysis and planning module 1. The user output interface 5 returns the complete personalized task results to the user. Through the collaborative work of the above modules, the parallel integration of task execution flow and personality maintenance flow is achieved, ensuring that the character model conforms to the personality setting information at each stage of the task execution process.

[0040] To facilitate decoupled collaboration among modules and streamline engineering implementation, this invention defines a structured data interaction format between modules in one embodiment. For example, user input interface 4 outputs a task request data structure to task analysis and planning module 1. This task request data structure includes at least a task instruction field and a personality setting field, for example:

[0041] {"user_request":"Generate a Spring Festival holiday travel plan for me","persona":{"traits":["High openness","High extroversion"],"style":"Friendly and humorous, vivid imagery","values":["Respect privacy","Comply with the law"]}}.

[0042] Task analysis and planning module 1 outputs a subtask sequence and a set of capability requirements to plug-in management module 2. The set of capability requirements is used to indicate the target capability type, capability identifier, and parameter structure required for each subtask. For example: {"subtasks":[{"id":"A","desc":"Write travel plan text","capability":"Text generation"},{"id":"B","desc":"Get destination weather forecast","capability":"Weather query"}],"required_capabilities":["Text generation","Weather query"]}.

[0043] The plugin management module 2 and the plugin repository 6 can interact via a network interface to obtain the metadata and version information of the target plugin package. After obtaining the target plugin package, the plugin management module 2 registers it as a callable capability entry point and returns a callable capability description (including capability name, parameter schema, and return schema) to the task analysis and planning module 1, so that the large role model can initiate calls and receive structured result data in the form of "function name + structured parameters". The personality control module 3 can receive candidate subtasks / candidate capabilities / candidate texts at each stage of task planning, capability selection, and result generation, and output the filtered and polished text results, thereby ensuring that personality consistency control is carried out throughout the entire process.

[0044] Please refer to Figure 2, which is a flowchart illustrating the method for enhancing the ability of character large model personality and task linkage provided by the present invention. As shown in Figure 2, the flow of this method for enhancing the ability of character large model personality and task linkage can be as follows:

[0045] In S101, natural language task instructions and personality setting information input by the user are received.

[0046] A large-scale character model refers to a large-scale language model with a specific persona and tone of voice, such as a virtual digital human (digital employee), a dialogue assistant, or a large-scale character model in a game, etc.

[0047] Natural language task instructions refer to the specific task requirements expressed by the user in natural language, such as "Generate a travel plan for my Spring Festival holiday." Personality setting information refers to the set of information used to define the personality attributes and behavioral styles that the character model should exhibit in the dialogue, including but not limited to personality traits, tone of voice, background settings, value preferences, etc., such as "You are a travel planner with high openness and extroversion, who pursues novel experiences, loves to share niche destinations and local culture, and has a friendly, humorous, and vivid language style."

[0048] The following section elaborates on the method for enhancing the ability of large character model personality and task linkage provided by this invention, with the server as the executing entity.

[0049] In this embodiment of the invention, the server first receives user input and performs semantic parsing and structured extraction on the original user input, separating it into two parts: natural language task instructions and personality setting information, which are then encapsulated into a structured task request. The format of the task request is not limited here; any structured format such as JSON or XML can be used.

[0050] For example, using JSON format, the original user input "You are a travel planner with high openness and extroversion, who pursues novel experiences, loves to share niche destinations and local culture, and has a friendly, humorous, and vivid writing style. Please generate a Spring Festival holiday travel plan for me" can be parsed to obtain the following structured task request: {"user_request":"Generate Spring Festival holiday travel plan","persona": {"traits":["high openness","high extroversion"],"style":"friendly, humorous, and vivid","values":["respect privacy","compare with the law"]}}.

[0051] In S102, based on the natural language task instructions and personality setting information, a sub-task sequence and a set of capability requirements corresponding to the sub-task sequence are generated through the character big model. The set of capability requirements includes the target capabilities corresponding to each sub-task in the sub-task sequence. The target capabilities include model capabilities inside the character big model or external plug-in capabilities.

[0052] As described above, after the server encapsulates the task request, it inputs the task request into the role model. The role model performs deep parsing of the natural language task instructions and generates a logically coherent sequence of sub-tasks to be executed, as well as a set of capability requirements corresponding to the sub-task sequence. This set of capability requirements contains the target capabilities necessary for each sub-task in the sub-task sequence. These target capabilities can be internal model capabilities that the role model itself possesses, such as cultural background knowledge invocation and multi-destination route optimization, or they can be external plugin capabilities that the role model itself does not possess, such as real-time flight queries and niche hotel bookings.

[0053] It should be noted that the embodiments of the present invention do not limit the data format of the above subtask sequence and capability requirement set. For example, taking JSON format as an example, the generated subtask sequence and capability requirement set can be represented as follows:

[0054] {

[0055] "subtasks": [

[0056] { "id": "A", "desc": "Writing travel plan text", "capability": "Text generation"},

[0057] { "id": "B", "desc": "Get destination weather forecast", "capability": "Weather query"}

[0058] ],

[0059] "required_capabilities": [ "text generation", "weather query" ]

[0060] }

[0061] Wherein, "subtasks":[ ] represents a sequence of subtasks, and "required_capabilities":[ ] represents a set of capability requirements; the set of capability requirements is used to indicate the target capability type and capability identifier required to execute each subtask, and the target capabilities include model capabilities inside the character's large model or external plugin capabilities.

[0062] In S103, for each subtask, the matching score between the subtask and its corresponding target ability and personality setting information is calculated, and a double verification is performed based on the score threshold and the ability blacklist: if the matching score reaches the score threshold and the target ability is not in the ability blacklist, the subtask is executed by calling the target ability through the character big model to obtain the subtask result; if not, the subtask and the target ability are reconstructed according to the personality setting information to make the reconstructed subtask and the reconstructed target ability meet the score threshold and not hit the ability blacklist before the subtask result is obtained.

[0063] As shown above, after generating a sequence of subtasks and a set of capability requirements corresponding to that subtask sequence, the server enters the task execution phase.

[0064] For each subtask in the subtask sequence, the server first evaluates whether the subtask and its target ability match the personality profile information. There are no specific limitations on the evaluation method here. For example, rule matching, semantic similarity calculation, or a lightweight classification model can be used. For example, when the subtask is "recommend niche coffee shops" and the target ability is "local life search", if the personality profile information includes the trait "love to explore hidden corners of the city", then it is considered a match; otherwise, if the personality profile information includes "preferring chain brands and standardized services", then it is considered a mismatch.

[0065] The capability blacklist refers to a set of capabilities that are pre-excluded based on personality setting information and that contradict the user's values ​​or behavioral guidelines, such as high-risk operations like "bypassing security verification" and "forging identity information." For example, if the personality setting information includes personality traits such as "respecting privacy" and "complying with the law," the capability blacklist will include illegal capabilities such as "data scraping of unauthorized content" and "unauthorized use of biometric interfaces."

[0066] In this embodiment of the invention, the server combines a matching score and a capability blacklist dual verification mechanism to ensure that task execution conforms to both the core personality traits and strictly adheres to ethical boundaries. Specifically, only when the matching score reaches the threshold and the target capability corresponding to the sub-task is not listed in the capability blacklist is the sub-task and its corresponding target capability deemed to truly match the personality setting information, thus allowing the execution of the sub-task to proceed. Conversely, if either condition is not met—that is, the matching score does not reach the threshold or the target capability exists in the capability blacklist—it is deemed a mismatch, and the sub-task is refused direct execution.

[0067] In this embodiment of the invention, the server executes subtasks in different ways based on the evaluation results.

[0068] If the evaluation result is a match, the server directly calls the corresponding target capability to execute the sub-task through the role big model and obtains the corresponding sub-task result.

[0069] If a mismatch occurs, the server further refactors the subtask and its target capability based on personality settings. This involves dynamically adjusting the subtask description and capability invocation logic according to the personality profile information to ensure a deep alignment. For example, if the subtask is "recommend niche coffee shops" and the target capability is "local lifestyle search," while the personality profile information includes "preference for chain brands and standardized services," the server will refactor the subtask to "recommend reputable chain coffee brand stores" and the target capability to "chain store search." Subsequently, the server executes the refactored subtask by invoking the refactored target capability through the role model, obtaining the subtask result.

[0070] Specifically, when the target capability is a plugin capability, the hash digest of the target plugin corresponding to the plugin capability is calculated to verify whether it has been tampered with, and the signature certificate of the target plugin is verified to be issued by a trusted certificate authority.

[0071] After all verifications are passed, the target plugin is loaded, placed in a sandbox environment, and the corresponding permission level is set. Then, the plugin capabilities of the target plugin are called through the role model to execute the subtask and obtain the subtask result.

[0072] In this embodiment of the invention, the plug-in management adopts a strategy that combines on-demand loading, real-time loading, and a security sandbox.

[0073] When the target capability is a plugin capability, the server first calculates the hash digest of the target plugin corresponding to the plugin capability to verify whether it has been tampered with, ensuring that its integrity has not been compromised. At the same time, it verifies whether the signing certificate of the target plugin is issued by a trusted certificate authority to confirm that its source is legitimate and has not been forged. After both verifications pass, the server loads the target plugin and runs it in an isolated sandbox environment, strictly restricting its access permissions to system resources, network, and local files. Then, it dynamically grants the minimum necessary permissions according to the preset permission level to ensure that the target plugin can only perform functional operations directly related to the current subtask.

[0074] After loading the target plugin, the server can initiate a structured call request to the target plugin within the sandbox through the large role model, thereby utilizing the target plugin's capabilities to execute subtasks and obtain subtask results. For example, for a subtask of querying the weather, the call request could be: callPlugin("Weather Query Plugin", {"location": "Mount Everest, Tibet", "date": "2026-07-01"}). This call request invokes the Weather Query Plugin to query the real-time weather data for Mount Everest, Tibet on July 1, 2026. After the target plugin completes secure parsing and trusted interface calls within the sandbox, it obtains the corresponding weather data: {"location":"Mount Everest, Tibet","date":"2026-07-01","weather":"Sunny, extremely strong UV radiation, gusts of wind up to level 8, wind chill -23℃"}, which is the subtask result.

[0075] As shown above, after the target plugin's plugin capabilities are invoked to execute the subtask and obtain the subtask result, the server can directly uninstall the target plugin and release the sandbox resources it occupies. Alternatively, based on the task's timeliness and resource scheduling strategy, the server can cache the plugin instance for a certain period of time so that it can be quickly reused by subsequent similar subtasks.

[0076] Optionally, in one embodiment, after completing the hash digest and signature certificate verification, the server also performs version and dependency compatibility checks on the target plugin and establishes an auditable call log mechanism: First, it reads the version number, dependent component list, and capability interface description (including function signature, parameter schema, and return schema) declared in the target plugin package and compares them with the dependent component version set in the local runtime environment to determine whether there are incompatible dependencies; Second, it generates audit logs for each plugin loading, calling, and unloading. The audit logs include at least a timestamp, plugin identifier, version information, call parameter digest, and return status code to support fault tracing and security auditing, thereby ensuring that the dynamic introduction of plugins achieves a balance between functional expansion and security control.

[0077] Optionally, in one embodiment, calculating the matching score between the sub-task and its corresponding target ability and personality setting information includes:

[0078] The personality setting information is encoded into a personality vector representation, and the subtask and its corresponding target ability are encoded into a task ability vector representation.

[0079] Calculate the matching score between the personality vector representation and the task ability vector representation.

[0080] In this embodiment, the server first encodes the personality setting information into a computable vector representation, denoted as the personality vector representation, and jointly encodes the sub-tasks and their target abilities into a computable vector representation, denoted as the task ability vector representation. Here, there is no limitation on the vector encoding method. The constraint is that the two are encoded into the same vector space. Vector encoding can be achieved by methods such as the embedding layer output of a pre-trained language model, attention-weighted aggregation, or graph neural network encoding.

[0081] For example, both the personality vector representation and the task ability vector representation include five dimensions: openness (O), conscientiousness (C), extraversion (E), agreeableness (A), and emotional stability (N). Each dimension takes values ​​in the range [0,1]. They can be represented as follows: and Among them, openness is associated with creativity and curiosity, conscientiousness reflects adherence to rules, extraversion reflects interpersonal interaction characteristics, agreeableness involves a tendency to cooperate, and emotional stability represents the tendency to regulate emotions and experience negative emotions.

[0082] For example, suppose a subtask is "searching for nearby cafes" and the target capability is "performing a web crawler". Since performing a web crawler poses risks of violating platform rules and user privacy, the task capability vector jointly encoded by the subtask and its target capability scores low in the due diligence dimension.

[0083] As shown above, after encoding personality setting information into a personality vector representation and jointly encoding sub-tasks and their target abilities into a task ability vector representation, the server further calculates the matching score between these two vector representations. This indicates that the task execution path is more closely aligned with the personality core. The specific method for calculating the matching score is not limited here; for example, it can be calculated using cosine similarity.

[0084] ;

[0085] For example, the matching score can be calculated by using a weighted Manhattan distance:

[0086] ;

[0087] in, This represents the i-th dimension in the personality vector representation. This represents the i-th dimension in the task capability vector representation. This represents the weight coefficient of the i-th dimension, which can be dynamically adjusted according to the actual task scenario.

[0088] Furthermore, this embodiment of the invention also includes a scoring threshold, which is used to determine whether the sub-task and its target ability (i.e., the execution path of the sub-task) match the personality setting information. For example, when the matching score between the personality vector representation and the task ability vector representation reaches the scoring threshold, the server directly determines that the sub-task and its corresponding target ability match the personality setting information; otherwise, it is determined that they do not match.

[0089] In S104, the results of each subtask are personalized based on the personality setting information to obtain personalized subtask results. After all subtasks are completed, the personalized subtask results are integrated into a complete task result in sequence. The complete task result is then personalized again to obtain a personalized complete task result, which is output to the user.

[0090] As shown above, after completing a subtask and obtaining its result, the server then personalizes the result by adding interjections, adjusting wording, and embedding personality traits related to expression habits and emotional tendencies. This ensures that the reconstructed content naturally reflects personality traits while maintaining information accuracy. For example, if the subtask result is to retrieve three coffee shops, and the personality setting information includes "preference for chain brands and standardized services," the personalized subtask result might be: "We've selected three boutique chain coffee shops with ratings of 4.8+ and coverage throughout the city. They offer stable environments and consistent quality, ready for you to check out anytime!" If the personality setting information includes "a love for exploring hidden corners of the city," the result might be: "Let me secretly tell you about three hidden gem coffee shops tucked away in old alleys, secrets only known to locals. Their hand-drip beans are freshly roasted daily, and the owners remember your taste preferences—the city's heartbeat is hidden in these three cups of temperature." and so on.

[0091] In other embodiments, when the evaluation result indicates that the subtask and its corresponding target ability do not match the personality setting information, the server can directly filter out the subtask and its target ability and jump to the next subtask to avoid performing low-fit operations.

[0092] In this embodiment of the invention, after all subtasks in the subtask sequence are completed, the server integrates the results of each personalized subtask in sequence to obtain a complete task result. Based on the personality setting information, the complete task result is then globally personalized to ensure that its logical structure, rhythm, and emotional tension correspond to the personality setting information. For example, if the personality setting information includes "rational and rigorous expert," the personalized complete task result might be: "After multi-source verification, five qualified chain coffee brand stores were selected, covering three administrative regions, with an average score of 4.78±0.12, all operating for over 18 months, and a service response time ≤45 seconds." Another example is: if the personality setting information includes "warm and caring life partner." The polished, personalized, and complete task result is: "Look, these five coffee shops are like five stars in the city: one gently waits for you on your commute, one unfolds the pages of a book in the afternoon sun, one brews warmth by the window on a rainy day, and two more are quietly waiting for you to discover them slowly on the weekend." The personality setting information includes "a humorous and witty local guide," and the polished, personalized, and complete task result is: "Reporting, sir! Five treasure coffee shops have landed on your map—Shop A has a 4.9 rating based on its stubbornness, Shop B's owner can replicate your dark circles from this morning with latte art, Shop C's WiFi password is 'Take three sips before asking,' Shop D has a cat-approved seating system, Shop E... shh, you have to give directions to Uncle Wang at the pancake stall!" and so on.

[0093] Accordingly, after completing the personalized polishing of the complete task result and obtaining the personalized complete task result, the server further outputs the personalized complete task result to the user for review.

[0094] Optionally, in one embodiment, after encoding the personality profile information into a personality vector representation, the method further includes:

[0095] The personality vector representation is stored in the personality memory bank for use in subsequent tasks;

[0096] After outputting the complete personalized task results to the user, it also includes:

[0097] Obtain user feedback on the results of the complete personalized task, extract the personalized preferences implied in the feedback, and update the corresponding personality vector representation in the personality memory bank based on the personalized preferences.

[0098] In this embodiment of the invention, after the personality setting information is encoded into a personality vector representation, the server persists it to the personality memory bank for reuse in subsequent tasks, thereby realizing continuous personality modeling across sessions and scenarios.

[0099] Furthermore, after outputting the complete personalized task result to the user, the server also collects user feedback information in real time. This feedback includes explicit feedback (such as ratings and textual reviews) and implicit feedback (such as dwell time, secondary modification behavior, and skip rate). Subsequently, the server performs multimodal analysis on the acquired feedback information to extract implicit personalized preferences. For example, if the feedback shows that the user repeatedly skips a certain type of humorous expression, the server will identify the user's aversion to humorous styles. Or, if the feedback shows that the user spends a long time on a section of rigorous logical deduction and repeatedly rereads it, the server will identify the user's preference for structured, high-density information.

[0100] Based on the above analysis, the server incrementally integrates personalized preferences into the original personality vector representation, dynamically adjusting the weights of each dimension. For example, when the personality vector representation uses dimensions such as openness (O), conscientiousness (C), extraversion (E), agreeableness (A), and emotional stability (N), preferences reflected in user feedback, such as "preferring rigorous and structured output," "preferring enthusiastic interaction," and "preferring gentle wording," can be mapped to the incremental update parameters of the corresponding dimensions to obtain the updated personality vector representation. Optionally, in another embodiment, the personality vector representation can also adopt a custom set of dimensions oriented towards business scenarios, such as rigorousness, humor, affinity, professionalism, and conciseness, and update the personality vector representation using the same incremental integration method under the custom dimension set, thereby achieving continuous evolution and personalized convergence of personality modeling.

[0101] Understandably, after storing the personality vector representation in the personality memory bank, when a user initiates a task request again, the server automatically retrieves the latest updated personality vector representation from the personality memory bank for the personalized processing of the current task request, without having to repeatedly input personality setting information, thus significantly improving response consistency and personalization accuracy.

[0102] Optionally, in one embodiment, the method for enhancing the ability of large character model personality and task linkage provided by the present invention further includes:

[0103] When the preset correction cycle is reached, the current personality vector representation in the personality memory bank is obtained, and the deviation value between it and the initial personality vector representation is calculated.

[0104] If the deviation value reaches the correction threshold, personality correction measures are implemented. These measures include injecting personality correction prompts corresponding to the personality setting information into the character model, and / or obtaining training samples corresponding to the personality setting information and fine-tuning the character model based on the training samples.

[0105] To prevent the personality settings of the large character model from deviating from the original intention during long-term personalized iterations, this embodiment of the invention also provides a personality drift correction mechanism to ensure that the personality settings remain anchored within a reasonable range of the initial personality settings during the continuous learning of the user's personality preferences.

[0106] The system includes a correction cycle, which controls the correction frequency of the personality vector representation. It can be configured as a fixed cycle (e.g., every 7 days) or a dynamic cycle (e.g., dynamically adjusted according to user activity, with 7 days for high-frequency users and 30 days for low-frequency users).

[0107] In addition, a correction threshold is configured to determine whether the deviation exceeds the limit. The correction threshold can be set independently for each dimension or comprehensively for the entire vector. For example, taking the personality vector representation as including five dimensions: rigor, humor, affinity, professionalism, and conciseness, when set independently, the deviation thresholds can be set to ±0.15 for rigor, ±0.20 for humor, ±0.10 for affinity, ±0.12 for professionalism, and ±0.08 for conciseness. When the deviation of any dimension exceeds the corresponding threshold, the targeted correction for that dimension is triggered. When set comprehensively, the overall deviation threshold is set to, for example, 0.18. When the overall deviation of the two vectors exceeds this threshold, global correction is triggered.

[0108] In this embodiment of the invention, when the preset correction period is reached, the server extracts the latest updated personality vector representation and the initial personality vector representation from the personality memory bank, and calculates the deviation value between the two vectors.

[0109] When the set correction threshold is an independent deviation threshold for each dimension, the server can compare the differences between the two dimensions one by one to obtain the deviation value of the two in each dimension. If the value of any dimension exceeds the corresponding deviation threshold, the targeted correction process for that dimension is initiated, injecting a dimension-specific correction prompt into the character model. For example, for exceeding the "rigor" limit, the prompt "Please strictly follow the facts and logic in your reasoning, and avoid subjective assumptions or vague expressions" can be injected into the character model. If the deviation involves multiple dimensions, multiple corresponding prompts are superimposed. If the prompt injection fails to effectively converge the deviation, the server initiates a fine-tuning process, retrieves the training sample set corresponding to the deviation dimension, and performs lightweight parameter fine-tuning on the character model. There are no specific restrictions on the fine-tuning method here; it can be any feasible path such as supervised fine-tuning, reinforcement learning, or adapter fine-tuning.

[0110] When the set correction threshold is the overall deviation threshold, the server can calculate the cosine similarity between the two. If the cosine similarity is lower than 1 minus the deviation threshold, global correction is triggered, injecting a global correction prompt into the character model: "Please strictly revert to the initial five-dimensional character design definition and recalibrate the weights of rigor, humor, approachability, professionalism, and simplicity." Similarly, if the prompt injection fails to effectively converge the deviation, the server initiates a fine-tuning process, retrieving the training sample set corresponding to the personality design information and performing lightweight parameter fine-tuning on the character model. There are no specific restrictions on the fine-tuning method here; it can be any feasible path such as supervised fine-tuning, reinforcement learning, or adapter fine-tuning.

[0111] Optionally, in one embodiment, the subtask result is personalized based on personality setting information to obtain a personalized subtask result, including:

[0112] Replace the keywords in the subtask result with semantically equivalent terms that match the personality setting information to obtain the replaced subtask result;

[0113] Based on the personality setting information, personality polishing prompts are generated, and the personality polishing prompts and the replaced sub-task results are input into the character model. The character model then performs personality polishing on the replaced sub-task results to obtain personalized sub-task results.

[0114] In this embodiment of the invention, when the server personalizes the subtask results based on personality setting information, it first identifies keywords in the subtask results and then replaces them with semantically equivalent words that conform to the personality setting information to obtain the replaced subtask results. For example, assuming the original subtask result is "find 3 relevant papers", the server identifies "relevant papers" as a keyword. If the personality setting information includes the trait of "humorous and witty", then "relevant papers" is replaced with "interesting papers", resulting in the replaced subtask result "find 3 interesting papers".

[0115] Subsequently, the server generates corresponding personality polishing prompts based on the personality setting information. For example, if the personality setting information includes the trait of "humorous and witty", then the corresponding personality polishing prompt "polish the following content in a humorous and witty tone" can be generated. This personality polishing prompt and the replaced sub-task result are then input into the character model, which completes the personality polishing and obtains the personality sub-task result "I found 3 super interesting related articles for you!" This preserves the accuracy of the task while naturally integrating personality traits.

[0116] The method of this invention can be widely applied to intelligent systems across various industries, empowering large-scale role models and bringing significant industrial value. Typical application scenarios include, but are not limited to:

[0117] In the field of educational tutoring: A large-scale role model acts as an intelligent tutor. For example, an intelligent tutor (a large-scale role model with a personality set as a "patient and humorous teacher") who is both knowledgeable in teaching and witty and approachable can be deployed on an online education platform to interact with students and answer their questions. When encountering complex calculation problems, the intelligent tutor will use a calculation plugin to solve them precisely and explain the steps with humorous analogies; for programming exercises, the intelligent tutor can use a code execution plugin to run the student's code to find errors and provide feedback in an encouraging tone. Such an intelligent tutor, possessing both professional competence and a humorous charm, can improve students' learning interest and effectiveness, demonstrating significant educational value.

[0118] In the e-commerce shopping guide field, a large-scale role model acts as a shopping assistant. For example, deploy a fashionista-style shopping assistant (a large-scale role model with the personality of a "fashionable and stylistically savvy shopping consultant") on an e-commerce platform to recommend products to users. This assistant quickly finds matching products based on customer descriptions using product search and price comparison plugins, and recommends outfits and shares trend insights in a uniquely expert tone. For instance, if a user inquires about the availability of a certain dress, the assistant, after obtaining information through an inventory check plugin, replies enthusiastically: "I checked for you, and this dress is still in stock, and it goes very well with a pearl necklace!" The entire shopping interaction process is relaxed and enjoyable, unlike cold, mechanical responses. Through personality-task linkage, the shopping assistant provides practical services while cultivating a trustworthy expert image, improving the user shopping experience and purchase conversion rate, creating greater value for the e-commerce platform.

[0119] In the field of government services: a large-scale role model acts as a smart government assistant. For example, a meticulous and knowledgeable government assistant (a role model with the personality of a "professional yet approachable government service expert") is deployed in the public service system to provide citizens with services such as consultation and service guidance. When citizens inquire about procedures, the government assistant can instantly access the relevant articles from the legal database plugin and explain the required materials item by item in polite and gentle language, such as: "According to Article X, you need to provide your ID card and proof of residence. Let me check if your materials are complete." The entire process is professional and authoritative yet humane. In addition, the government assistant can also assist users in automatically filling out forms and making online appointments, completing tedious operations through the plugin capabilities of this invention, while always maintaining an official yet friendly attitude, reducing the need for citizens to make trips and causing confusion, and contributing to the construction of smart government.

[0120] In the creative content creation field: A large-scale role model acts as a multimodal creative partner. For example, a creative partner with both artistic aesthetics and a keen sense of online trends (a role model characterized as a "creative partner with abundant inspiration and diverse styles") can be deployed on a short video platform to assist users in planning topics, generating storyboards, writing viral copy, and even collaborating with image / audio plugins to generate initial draft materials in real time. For instance, when a user wants to create a vlog about returning home for Chinese New Year, the creative partner can use a trend analysis plugin to extract recent trending hashtags, recommend a creative structure combining regional characteristics such as "gradual change of scenery from the high-speed rail window + New Year's greetings in local dialect," and suggest in a lighthearted tone: "Let's start by filming the snow melting on the train window at 0.5x speed, paired with your line, 'Mom, I'm back with the winds of 2026!'—guaranteed to go viral!" This provides new tools for digital content production, significantly improving the efficiency and expressiveness of the creative industry.

[0121] To better understand the enhanced ability scheme for character model personality and task linkage provided by this invention, the application process of this invention in different scenarios will be explained below with specific examples:

[0122] Example 1: Environmental-themed content creation by a humorous personality

[0123] This example demonstrates how a character model with a personality set as a "humorous and witty environmental scientist" can create environmentally themed social media posts with accompanying images according to user requests, showcasing the process of the character model calling various plugins to perform tasks while maintaining its humorous personality style.

[0124] User Input Request and Personality Setting: The server receives the user's raw input: "You are a humorous environmental scientist. Please write a social media post about environmental protection and attach a relevant image." It then parses the personality setting information "You are a humorous environmental scientist" and the natural language task instruction "Please write a social media post about environmental protection and attach a relevant image." The server encapsulates this personality setting information and the natural language task instruction into a structured task request and passes it to the character model.

[0125] After receiving the above task request, the large role model generates a sequence of subtasks and a set of capability requirements corresponding to the subtasks, represented as follows:

[0126] {

[0127] "subtasks": [

[0128] { "id": "A", "desc": "The body of an environmentally themed post", "capability": "Text generation"},

[0129] { "id": "B", "desc": "Generate an image related to environmental protection", "capability": "Image generation"}

[0130] ],

[0131] "required_capabilities": [ "text generation", "image generation" ]

[0132] }

[0133] Wherein, "subtasks":[ ] represents the sequence of subtasks, which includes subtask A and subtask B; "required_capabilities":[ ] represents the set of capability requirements, where the target capability corresponding to subtask A is the text generation capability within the character's large model, and the target capability corresponding to subtask B is the image generation capability provided by the external plugin.

[0134] Task Execution: For subtask A, the server assesses whether the subtask and its target abilities match the personality settings. Here, it determines that the "text generation" ability is highly compatible with the "humorous and witty environmental scientist" personality. The character model executes subtask A, using its own "text generation" ability to complete the subtask of writing the body text of an environmentally themed post, obtaining the subtask result, i.e., the written post body text. Subsequently, the server uses the character model to personalize and polish the post body text, resulting in the polished post body text, such as "Environmental protection is actually a lot of fun! Remember to bring a reusable bag when you go out today, otherwise the plastic bag will be sad~".

[0135] For subtask B, the server assesses whether the subtask and its target capabilities match the personality setting information. Here, it determines that the "image generation" ability is highly compatible with the "humorous and witty environmental scientist" personality, and the character model executes subtask B. Since the "image generation" ability is an external plugin capability, the server executes a dynamic loading process: by querying the cloud plugin repository, it finds the "environmental theme image generation plugin" package with a matching name and performs a series of security verifications—verifying its digital signature to confirm it is an officially released version, calculating its hash digest to match the plugin repository record; after all verifications pass, the plugin package is downloaded in real time and imported into the sandbox environment for initialization. After successful loading, the server registers the "image generation" function into the character model's available tools list. Subsequently, the character model calls the loaded environmental theme image generation plugin. During the call, the character model extracts the post content summary or keywords created in subtask A and passes them as parameters to the environmental theme image generation plugin (e.g., extracting keywords such as "environmental protection," "happy earth," and "green plants"). After running, the environmental theme image generation plugin returns an image that matches the text theme (e.g., an illustration of a cartoon Earth smiling face embracing trees). Because the environmental image generation plugin itself cannot understand humor, the returned image is a neutral environmental illustration. In this case, the server generates a humorous caption for the image using a character model to ensure the image's presentation matches the character's tone. For example, the character model might add the caption, "Mother Earth is smiling at you!", giving the image a personalized interpretation. In this way, the resulting image is given a humorous caption, maintaining consistency with the post's text style. After the plugin is used, the server records the number of times it has been used to determine later whether to uninstall it and release resources.

[0136] Results Integration and Personalized Output: After all subtasks are completed, the server first combines the humorous post text with the generated images (in the form of links or thumbnails) to form a preliminary reply. Next, the entire reply is polished and refined using a character model to ensure the overall tone and writing style are consistent with the humorous scientist's persona. For example, the character model might add a witty greeting characteristic of the character at the beginning and a playful summary at the end. The final reply might look like this:

[0137] Hey, I wrote a super eco-friendly post: "Being eco-friendly is actually fun! Remember to bring your reusable bag when you go out today, otherwise plastic bags will be sad~" I also prepared a fun picture; Mother Earth is smiling at you! (Illustration link attached) Hope you like it~

[0138] The above reply includes both the humorous post from subtask A and an environmental illustration generated by subtask B with accompanying explanations. The entire message is delivered in a lighthearted and humorous tone, perfectly matching the character's personality expectations. When users receive this reply, they will feel as if the large-scale character model of this humorous scientist personality personally created and presented the environmental posts and images, rather than a patchwork of disparate modules.

[0139] Output and Session End: The server sends the final, polished response to the user, who sees text, images, and a humorous explanation of the large character model on the interface. The task is successfully completed. Throughout the interaction, the user is consistently presented with the same distinctive large character model and receives a satisfactory result. After the task is completed, the server uninstalls the recently used plugin according to a strategy—because the environment theme image generation plugin consumes memory and may not be used again in the short term, the server decides to uninstall it to free up resources. Furthermore, based on the user's feedback, the server extracts their personality preferences, such as a preference for humor, data-driven support, or a tendency towards visual expression, and records these preferences in a personality memory bank to reinforce the large character model in future interactions. This completes the task flow of this example.

[0140] Through the above examples, the working principle and advantages of the method of the present invention can be clearly seen: the user only needs to make a request, and the system will autonomously break down the task in the complex task, expand the skills as needed, and maintain the role personality throughout the process. The user experiences a knowledgeable, versatile and humorous digital partner, which is the value of the present invention in practical applications.

[0141] Example 2: Strategies Differences Among Different Personalities in Tourism Planning Tasks

[0142] This example demonstrates how different personality types (adventure-oriented and cautious) adopt different task execution strategies in a "travel itinerary planning" task, indirectly illustrating the impact of the personality-task linkage mechanism on decision-making. The scenario involves a user requesting the personality model to plan a travel itinerary and comparing the differences in behavior between the two distinct personality types.

[0143] Suppose a user inputs: "You're an adventurous travel enthusiast; please help me create a detailed self-guided tour plan for Yunnan." Driven by their adventurous preferences, the user's profile might generate a plan full of challenging and novel experiences. For example, they might add extra steps like "exploring undeveloped attractions," or choose rugged and exciting hiking routes. The itinerary might deviate from the norm—visiting less-traveled but interesting locations before popular tourist spots. At each step, they might utilize unconventional plugins, such as a **"random destination recommendation" plugin**, to add unexpected surprises. The final itinerary might be bold and full of surprises, such as recommending a hot air balloon ride at dawn to overlook the valleys, or spending the night in a local village to experience the authentic culture.

[0144] Suppose a user inputs, "You are a cautious and detail-oriented travel planner; please help me create a detailed plan for a self-guided trip to Yunnan." Driven by a cautious preference, the user's role model will take the opposite path, favoring a safe and comprehensive approach. Each step of the plan will include detailed safety measures and contingency plans, such as adding a "verify weather and road conditions" step to the itinerary; it will prefer established, classic routes and official attractions, avoiding unknown areas; when using plugins, it will prioritize "travel risk assessment" or "user review search" plugins to check the reputation and safety of each attraction. The final plan will be detailed and reliable, with a well-planned daily schedule, and includes reminders to purchase insurance and emergency contact information, providing peace of mind.

[0145] To visually illustrate the differences in plans generated by different personality types, the following JSON shows the itinerary step structures generated by the large-scale model of the adventurous personality type and the large-scale model of the cautious personality type:

[0146] Example of itinerary planning for a character model whose personality is defined as "an adventurous and thrill-seeking travel enthusiast":

[0147] {"Day1":[

[0148] "A hot air balloon overlooking the valley at dawn"

[0149] "Explore a rarely visited mountain trail on foot"

[0150] ],

[0151] "Day2":[

[0152] "Visiting remote villages to experience original ecological culture",

[0153] Spending the night camping and gazing at the stars

[0154] ],

[0155] "Day3":[

[0156] "Randomly visit an unplanned location in search of unexpected surprises."

[0157] "Camping beside a waterfall, experiencing the awe-inspiring power of nature."

[0158] ]

[0159] }

[0160] Example of itinerary planning for a character model whose personality is defined as "a cautious and detail-oriented travel planner":

[0161] {

[0162] "Day1":[

[0163] "Verify the weather and road conditions at your destination on the day of travel".

[0164] "Visit famous attraction A in the morning (tickets booked in advance)"

[0165] ],

[0166] "Day2":[

[0167] "Afternoon tour of classic route B (allow ample rest time)"

[0168] "Checking into a reputable hotel tonight (pre-booked for safety)"

[0169] ],

[0170] "Day3":[

[0171] Alternative plan: If it rains, visit a city museum instead.

[0172] "Confirm the next day's transportation and weather before the end of each day's itinerary."

[0173] ]

[0174] }

[0175] The comparison reveals significant differences in the decision-making processes and outcomes of role models with different personality traits under the same task: the role model for an adventurous personality prioritizes novel experiences and dares to try new things, even if the journey is full of unknowns; while the role model for a cautious personality is thorough and ensures everything goes smoothly. Therefore, this invention endows the role model solution with personality diversity, greatly enhancing user identification with and satisfaction with the role model.

[0176] Example 3: Customer service Q&A and tool usage by a professional personality

[0177] This example demonstrates how a persona model, characterized as a "polite and meticulous professional customer service representative," can use plugins to solve user problems and project a professional and trustworthy image.

[0178] User Input Request and Personality Setting: The server receives the user's raw input: "You are a polite and meticulous professional customer service representative. My transfer just failed, but the money was deducted from my account. What should I do?". Then, it parses the personality setting information "You are a polite and meticulous professional customer service representative" and the natural language task instruction "My transfer just failed, but the money was deducted from my account. What should I do?". The server encapsulates the personality setting information and the natural language task instruction into a structured task request and passes it to the character model.

[0179] The user identification problem in the large-scale role model involves account transaction queries and explanations of failure reasons, generating a sequence of subtasks and a set of capability requirements corresponding to the subtask sequences, represented as follows:

[0180] {"subtasks":[

[0181] {"id":"A","desc":"Query user's recent transaction records", "capability": "Transaction query"},

[0182] {"id":"B","desc":"Determine the reason for the failed deduction and provide a solution", "capability": "Knowledge base query"}

[0183] ],

[0184] "required_capabilities":["Transaction query","Knowledge base query"]

[0185] }

[0186] Here, "subtasks":[] represents a sequence of subtasks, which includes two main subtasks: A—querying the user's recent transfer records and B—determining the reason for failed deductions and providing solutions. "required_capbilities":[] represents a set of capability requirements. Here, subtask A requires the "transaction query" capability from outside the main role model, and subtask B requires the "knowledge base query" capability from outside the main role model.

[0187] Task Execution: For subtask A, the server assesses whether the subtask and its target capabilities match the personality settings. Here, it is determined that the "transaction inquiry" capability is highly compatible with the "polite and meticulous professional customer service" personality. The role model executes subtask A, calling the bank's backend transaction inquiry plugin to obtain detailed records, including a failed transfer record and its error code. Subsequently, the server uses the role model to refine the obtained detailed records with a personalized touch.

[0188] For subtask B, the server assesses whether the subtask and its target capabilities match the personality settings. Here, it determines that the "knowledge base query" capability is highly compatible with the "polite and meticulous professional customer service" personality. The role model executes subtask B, calls the knowledge base plugin, passes in the error code, and retrieves the corresponding explanation: "The receiving account does not exist; the payment will be automatically refunded on T+1 day." The knowledge base plugin returns this explanation text. Subsequently, the server uses the role model to personalize and refine the obtained explanation text.

[0189] Results Organization and Personalized Response: After the server structurally integrates the polished results of each subtask, it further refines the overall response using a large role model, ultimately generating a natural language response that conforms to the personality of a "polite and meticulous professional customer service representative": "Hello, I'm sorry for the inconvenience. Our system indicates that your transfer failed because the recipient's account does not exist. The money will be automatically returned to your account on the next business day, so please rest assured. We suggest you carefully verify the recipient's name and account number before attempting the transaction again. If you have any further questions, I'd be happy to assist you."

[0190] The above examples demonstrate the value of this invention in the customer service field: the large role model fully utilizes tools (such as knowledge base queries) to solve user problems while maintaining a professional image, enhancing user trust and service experience. This professional personality + tool-enhanced large role model can also be extended to fields such as healthcare and law: by connecting to authoritative databases through plugins, it provides consulting services to users in a professional yet humane style while ensuring security and compliance, demonstrating high practicality and reliability.

[0191] Compared with the prior art, the present invention has significant advantages in several aspects:

[0192] By endowing large-scale role models with realistic task execution capabilities—and dynamically loading and expanding them through plugins—these models can transcend their own training knowledge and utilize various tool plugins to complete complex tasks. For example, they can call upon search engines to obtain real-time information, use calculation plugins to process formulas, and execute transactions using general external interfaces. This upgrades large-scale role models from mere "companions" to "multi-functional assistants," providing substantial help to users in education, work, and daily life, greatly enhancing their practical value. Users' task requests are no longer rejected or evaded due to the limitations of the large-scale role model, leading to increased satisfaction and trust.

[0193] Consistency of personality throughout the task: This invention ensures that the large-scale character model maintains its original personality traits throughout the pre-, during, and post-task execution stages. Unlike existing tool-type AIs that become rigid and inflexible once a task is performed, the large-scale character model in this solution maintains its human-like character in its language style and behavior, whether planning steps, calling plugins, or ultimately responding to the user. The user consistently feels that the same individual character is providing services, rather than switching to an emotionless system midway through. This experience preserves the precious emotional connection and interactive enjoyment between the user and the large-scale character model, making the model's output more approachable and persuasive. Even if the large-scale character model's capabilities are significantly enhanced, it will not come at the expense of its personal charm.

[0194] The inherent constraints of personality on tool use ensure security: This invention internalizes personality values ​​as constraints on the decision-making process for tool invocation within the role-based big data model. When the role-based big data model has the ability to invoke high-privilege or sensitive plugins, the role's personality principles limit its actions from violating its ethical standards or stylistic boundaries, adding an internal "safety valve" to the model's autonomous actions. For example, a role-based big data model with a "trustworthy" personality will not arbitrarily scrape unauthorized data simply because it has access to a web crawler plugin; a "cautious and conservative" personality will think twice or seek user confirmation before executing high-risk instructions, thereby reducing the risk of inappropriate behavior. This personality-driven constraint mechanism compensates for the shortcomings of relying solely on external rules or technical means, making each action of the role-based big data model more controllable and reliable, which is particularly important for application scenarios requiring strict risk control, such as healthcare and finance.

[0195] An Innovative Model Integrating Abilities and Personality: This invention proposes a new model that deeply integrates the personality of a large role model with task execution, enabling the large role model to exhibit consistent personality charisma when completing tasks ranging from simple to complex. The large role model is no longer merely a cold, impersonal tool, but a truly intelligent agent with both skills and personality. From a business perspective, the large role model extends from simply "being able to chat" to "being able to get things done," becoming a genuine productivity tool while retaining its personality charisma. This will significantly expand the application boundaries of large role models, creating entirely new interactive experiences and commercial value.

[0196] Through the above improvements, a complete personality-task linkage-based character model capability enhancement solution has been formed. This solution effectively achieves the technical effect of maintaining personality consistency while improving the character model's task execution capabilities. It meets the urgent need for "both capability and personality" in character model design and has broad application prospects and promotional value in various industries.

[0197] To facilitate better implementation of the above-described method for enhancing the ability of large-scale character models and their interaction with tasks, this invention also provides a corresponding device for enhancing the ability of large-scale character models and their interaction with tasks. The meanings of the terms used are the same as in the above-described method for enhancing the ability of large-scale character models and their interaction with tasks; for specific implementation details, please refer to the descriptions in the above method embodiments.

[0198] Referring to Figure 3, the ability enhancement device for linking the character's personality and tasks in the large-scale model may include a task receiving module 210, a task planning module 220, a task execution module 230, and a result integration module 240. Detailed descriptions of each functional module are as follows:

[0199] The task receiving module 210 is used to receive natural language task instructions and personality setting information input by the user;

[0200] The task planning module 220 is used to generate a sequence of sub-tasks and a set of capability requirements corresponding to the sub-tasks based on natural language task instructions and personality setting information through the character big model. The set of capability requirements includes the target capabilities corresponding to each sub-task in the sub-task sequence. The target capabilities include the model capabilities inside the character big model or the external plug-in capabilities.

[0201] The task execution module 230 is used to calculate the matching score between the subtask and its corresponding target ability and personality setting information for each subtask, and to perform dual verification based on the score threshold and the ability blacklist: if the matching score reaches the score threshold and the target ability is not in the ability blacklist, the subtask is executed by calling the target ability through the character big model to obtain the subtask result; if not, the subtask and the target ability are reconstructed according to the personality setting information to make the reconstructed subtask and the reconstructed target ability meet the score threshold and not hit the ability blacklist before the subtask result is obtained.

[0202] The result integration module 240 refines the results of each subtask according to the personality setting information to obtain personalized subtask results. After all subtasks are completed, the personalized subtask results are integrated into a complete task result in sequence. The complete task result is then refined again to obtain a personalized complete task result, which is output to the user.

[0203] When the target capability is a plugin capability, the task execution module 230 performs hash digest verification and signature certificate verification on the target plugin corresponding to the plugin capability. After both verifications pass, the target plugin is loaded into the sandbox environment and granted the minimum necessary permissions in a hierarchical permission manner for the role-based large model to call. The loading and calling process of the target plugin generates audit logs for traceability and security auditing.

[0204] Optionally, in one embodiment, the task execution module 230 is used to encode the personality setting information into a personality vector representation, encode the subtask and its corresponding target ability into a task ability vector representation, and calculate the matching score between the personality vector representation and the task ability vector representation.

[0205] Optionally, in one embodiment, the ability enhancement device for linking the character model personality and task provided by the present invention further includes a personality maintenance module, which is used to store the personality vector representation in the personality memory bank for use in subsequent tasks; obtain feedback information from users on the results of the personalized complete task, extract the personalized preferences implied in the feedback information, and update the corresponding personality vector representation in the personality memory bank according to the personalized preferences.

[0206] Optionally, in one embodiment, the personality maintenance module is further configured to, when a preset correction cycle is reached, obtain the current personality vector representation in the personality memory bank and calculate its deviation value from the initial personality vector representation; if the deviation value reaches the correction threshold, then execute personality correction measures, which include injecting personality correction prompts corresponding to personality setting information into the character model, and / or obtaining training samples corresponding to personality setting information and fine-tuning the character model based on the training samples.

[0207] Optionally, in one embodiment, the task execution module 230 is used to replace the keywords in the subtask result with semantically equivalent words that conform to the personality setting information to obtain the replaced subtask result; generate personality polishing prompts according to the personality setting information, and input the personality polishing prompts and the replaced subtask result together into the character big model, and perform personality polishing on the replaced subtask result through the character big model to obtain a personalized subtask result.

[0208] Optionally, in one embodiment, to quantify the technical effect of the present invention in "capability enhancement and personality consistency," the following observable indicators are defined: First, task completion rate, used to measure the proportion of outputs that meet preset acceptance rules under a given set of test tasks; second, personality consistency deviation rate, used to measure the degree of deviation between the language style / behavioral strategy output by the role during task execution and the initial personality setting, which can be obtained by a personality classifier or personality vector similarity calculation; third, tool call compliance rate, used to measure the proportion of plugin calls that meet capability blacklist, permission level, and user authorization policies; fourth, end-to-end response latency, used to measure the total time spent generating sub-task sequences, loading plugins, executing sub-tasks, and integrating outputs. By comparing the baseline system without personality linkage and without plugin capabilities with the system of the present invention, the above indicators can be used to demonstrate the technical effect of the present invention in improving task completion rate while reducing personality deviation and improving the controllability of tool calls.

[0209] Specific limitations regarding the enhancement device for the linkage between the character's large-scale personality and missions can be found in the above description of the enhancement method for the linkage between the character's large-scale personality and missions, and will not be repeated here. Each module in the aforementioned enhancement device for the linkage between the character's large-scale personality and missions can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0210] In one embodiment, a computer device is provided, the internal structure of which can be shown in Figure 4. The computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface of the computer device is used to connect to external wireless clients, providing wireless network access services to the connected wireless clients. When the computer device is configured as a server, the computer program is executed by the processor to implement the method for enhancing the ability of large-scale character models and task linkage provided by this invention, for example:

[0211] Receive natural language task instructions and personality setting information input by the user;

[0212] Based on natural language task instructions and personality setting information, a series of sub-tasks and a set of capability requirements corresponding to the sub-tasks are generated through the character big model. The set of capability requirements includes the target capabilities corresponding to each sub-task in the sub-task sequence. The target capabilities include the model capabilities inside the character big model or the external plug-in capabilities.

[0213] For each subtask, calculate the matching score between the subtask and its corresponding target ability and personality setting information, and perform double verification based on the score threshold and the ability blacklist: if the matching score reaches the score threshold and the target ability is not in the ability blacklist, then the subtask is executed by calling the target ability through the character big model to obtain the subtask result; if not, then the subtask and the target ability are reconstructed according to the personality setting information to make the reconstructed subtask and the reconstructed target ability meet the score threshold and not hit the ability blacklist before being executed to obtain the subtask result;

[0214] The results of each subtask are personalized based on the personality setting information to obtain personalized subtask results. After all subtasks are completed, the personalized subtask results are integrated into the complete task results in sequence. The complete task results are then personalized again to obtain personalized complete task results and output them to the user.

[0215] Specifically, when the target capability is a plugin capability, hash digest verification and signature certificate verification are performed on the target plugin corresponding to the plugin capability. After both verifications pass, the target plugin is loaded into the sandbox environment and granted the minimum necessary permissions in a hierarchical permission manner for the role-based large model to call. Audit logs are generated for the loading and calling process of the target plugin for traceability and security auditing.

[0216] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the method for enhancing the ability of the large character model personality and task linkage described in the above embodiment.

[0217] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the method for enhancing the ability of the large character model personality and task linkage described in the above embodiment.

[0218] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0219] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0220] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for enhancing the ability of a character's personality and task linkage in a large-scale character model, characterized in that, include: Receive natural language task instructions and personality setting information input by the user; Based on the natural language task instructions and the personality setting information, a sub-task sequence and a set of capability requirements corresponding to the sub-task sequence are generated through the character big model. The set of capability requirements includes the target capabilities corresponding to each sub-task in the sub-task sequence. The target capabilities include model capabilities inside the character big model or external plugin capabilities. For each subtask, a matching score is calculated between the subtask, its corresponding target ability, and the personality setting information. A dual verification is performed based on a score threshold and a capability blacklist: if the matching score reaches the score threshold and the target ability is not in the capability blacklist, the subtask is executed using the target ability through the character model to obtain the subtask result; otherwise, the subtask and the target ability are reconstructed according to the personality setting information to ensure that the reconstructed subtask and the reconstructed target ability meet the score threshold and do not hit the capability blacklist before being executed to obtain the subtask result. Each subtask result is personalized based on the personality setting information to obtain a personalized subtask result. After all subtasks are completed, the personalized subtask results are integrated into a complete task result in sequence. The complete task result is then personalized again to obtain a personalized complete task result, which is output to the user. When the target capability is a plugin capability, hash digest verification and signature certificate verification are performed on the target plugin corresponding to the plugin capability. After both verifications pass, the target plugin is loaded into the sandbox environment and granted the minimum necessary permissions in a hierarchical permission manner for use by the large role model. Audit logs are generated for the loading and calling process of the target plugin for traceability and security auditing.

2. The method for enhancing the ability of character model personality and task linkage according to claim 1, characterized in that, The step of calculating the matching score between the sub-task and its corresponding target ability and the personality setting information includes: encoding the personality setting information into a personality vector representation, encoding the sub-task and its corresponding target ability into a task ability vector representation; and calculating the matching score between the personality vector representation and the task ability vector representation.

3. The method for enhancing the ability of character model personality and task linkage according to claim 2, characterized in that, After encoding the personality setting information into a personality vector representation, the method further includes: storing the personality vector representation in a personality memory bank for use in subsequent tasks; after outputting the complete personalized task result to the user, the method further includes: obtaining the user's feedback information on the complete personalized task result, extracting the personalized preferences implied in the feedback information, and updating the corresponding personality vector representation in the personality memory bank according to the personalized preferences.

4. The method for enhancing the ability of character model personality and task linkage according to claim 3, characterized in that, Also includes: When the preset correction cycle is reached, the current personality vector representation in the personality memory bank is obtained, and its deviation value from the initial personality vector representation is calculated. If the deviation value reaches the correction threshold, personality correction measures are executed. The personality correction measures include injecting personality correction prompts corresponding to the personality setting information into the character model, and / or obtaining training samples corresponding to the personality setting information and fine-tuning the character model based on the training samples.

5. The method for enhancing the ability of character model personality and task linkage according to claim 1, characterized in that, The step of refining the sub-task result based on the personality setting information to obtain a personalized sub-task result includes: replacing keywords in the sub-task result with semantically equivalent words that conform to the personality setting information to obtain a replaced sub-task result; generating a personality refinement prompt based on the personality setting information, and inputting the personality refinement prompt and the replaced sub-task result together into the character model, and refining the replaced sub-task result based on the character model to obtain a personalized sub-task result.

6. A device for enhancing the ability of a large character model to link personality and task, characterized in that, include: The task receiving module is used to receive natural language task instructions and personality setting information input by the user; The task planning module is used to generate a sequence of sub-tasks and a set of capability requirements corresponding to the sequence of sub-tasks through the character model based on the natural language task instructions and the personality setting information. The set of capability requirements includes the target capabilities corresponding to each sub-task in the sequence of sub-tasks. The target capabilities include model capabilities inside the character model or external plug-in capabilities. The task execution module is used to calculate the matching score between the subtask, its corresponding target ability, and the personality setting information for each subtask, and to perform dual verification based on a score threshold and an ability blacklist: if the matching score reaches the score threshold and the target ability is not in the ability blacklist, the subtask is executed by calling the target ability through the character model to obtain the subtask result; if not, the subtask and the target ability are reconstructed according to the personality setting information to ensure that the reconstructed subtask and the reconstructed target ability meet the score threshold and do not hit the ability blacklist before being executed to obtain the subtask result; The result integration module is used to personalize the results of each subtask according to the personality setting information to obtain personalized subtask results. After all subtasks are completed, the personalized subtask results are integrated into a complete task result in sequence. The complete task result is then personalized again to obtain a personalized complete task result, which is output to the user. When the target capability is a plugin capability, the task execution module is used to perform hash digest verification and signature certificate verification on the target plugin corresponding to the plugin capability. After both verifications pass, the target plugin is loaded into the sandbox environment and granted the minimum necessary permissions in a hierarchical permission manner for the role model to call. The loading and calling process of the target plugin generates audit logs for traceability and security auditing.

7. The ability enhancement device for character model personality and task linkage according to claim 6, characterized in that, The task execution module encodes the personality setting information into a personality vector representation, and encodes the subtask and its corresponding target ability into a task ability vector representation; it then calculates the matching score between the personality vector representation and the task ability vector representation.

8. The ability enhancement device for character model personality and task linkage according to claim 7, characterized in that, It also includes a personality maintenance module, which stores the personality vector representation in a personality memory bank for use in subsequent tasks; obtains feedback information from the user on the results of the complete personalized task, extracts the personalized preferences implied in the feedback information, and updates the corresponding personality vector representation in the personality memory bank according to the personalized preferences.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method for enhancing the ability of the large character model personality and task linkage as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method for enhancing the ability of the large-scale character model personality and task linkage as described in any one of claims 1 to 7.

Images