A meta-cognitive digital avatar agent method and system for individual users

Abstract

The application provides a metacognition digital avatar intelligent agent method and system for individual users, belonging to the technical field of metacognition digital avatars, which comprises: collecting individual interaction data to generate structured interaction events, storing the events in layers of event log layer, short-term episodic memory layer, long-term semantic memory layer and fact repository layer, and outputting the search results of the layered storage; generating auditable character hypothesis entries based on the content of the layered storage; generating candidate output results of digital avatars based on a main generative model, evaluating the results based on a guard model in terms of multi-dimensional metacognition features and calculating a comprehensive score; making a gating decision, if the comprehensive score is greater than or equal to a first threshold value and the risk level of the candidate output result is less than a second threshold value, then the candidate output result is allowed to be output, otherwise a restricted strategy is triggered; writing back to the layered storage and the auditable character hypothesis entries through a closed-loop writing channel, generating a strategy effect record, and updating the late personalized strategy selection.

Description

Technical Field

[0001] This application relates to the field of metacognitive digital avatar technology, and in particular to a metacognitive digital avatar intelligent agent method and system for individual users. Background Technology

[0002] While current large-scale model-driven personal digital avatars can achieve basic functions such as conversational companionship and schedule planning, existing technologies suffer from multiple shortcomings. First, user interaction data is mostly stored in an unordered manner, lacking a four-layer hierarchical memory architecture. There is no multi-version retention or conflict verification mechanism for facts, and retrieval relies solely on keywords, making it impossible to trace the source of evidence. Second, user profiles are output from a model black box, lacking a standardized and auditable item structure. Fields such as confidence level, counterexamples, authorization, and rebuttal markers are missing, and Bayesian dynamic updates are not supported, preventing users from verifying or refuting erroneous user characteristics. Third, generation and security verification are not decoupled; the main model directly outputs content, lacking an independent guardian model for multi-dimensional meta-cognitive assessment. Relying solely on simple keyword risk control fails to quantify indicators such as evidence coverage, memory conflict, and overconfidence, and there is no unified scoring standard for overall credibility. Finally, the output release judgment logic is simplistic, lacking dual threshold gates for comprehensive scoring and risk level. High-illusion, high-risk content is directly output without any layered restriction strategies. There is also a lack of a complete data loop. User feedback and strategy adoption results cannot be fed back to update memory and user assumptions, and cannot iteratively optimize subsequent personalized outputs. The feedback process also suffers from illusions and safety biases caused by changes in the main model weights. At the same time, existing technologies have not formed a human-machine collaborative metacognitive system. The processes of memory, profiling, generation, verification, and feedback are fragmented and cannot simultaneously address the needs for output authenticity, security, and long-term personalized adaptation.

[0003] Therefore, there is an urgent need for a metacognitive digital avatar intelligent agent method and system for individual users. Summary of the Invention

[0004] To address the aforementioned technical issues, this application provides a method and system for metacognitive digital avatar intelligent agents for individual users.

[0005] A first aspect of this application provides a metacognitive digital avatar intelligent agent method for individual users, comprising: Collect multimodal personal interaction data with user authorization and generate unified structured interaction events with timestamps; The unified structured interactive events are stored in layers according to the event log layer, short-term context memory layer, long-term semantic memory layer and fact version library layer, and the retrieval results of the layered storage are output. Based on the content of the hierarchical storage, auditable person hypothesis entries are generated. Each auditable person hypothesis entry includes a hypothesis statement, source of evidence, confidence level, last update time, counterexample record, applicable scenario, authorization status, and rebuttal flag field. The confidence level is updated using a Bayesian update mechanism. Based on the preset master generation model, candidate output results for digital clones are generated by combining the hierarchically stored retrieval results with the auditable person hypothesis entries; The guard model, which is independent of the preset master generation model, performs multidimensional metacognitive feature evaluation on the candidate output results and calculates a comprehensive score. Gating decisions are made based on the comprehensive score and the risk level of the candidate output results. If the comprehensive score is greater than or equal to the first threshold and the risk level of the candidate output results is less than the second threshold, the candidate output results are allowed to be output; otherwise, a restricted policy is triggered to obtain the evaluation result of the guard model. Obtain user feedback and strategy adoption results, and write back the evaluation results of the guard model, user feedback, and strategy adoption results to the hierarchical storage and the auditable person hypothesis entries through a closed-loop write-back channel, and generate strategy effect records for updating later personalized strategy selection; wherein the closed-loop write-back does not modify the parameter weights of the main generation model.

[0006] A second aspect of this application provides a metacognitive digital avatar intelligent agent system for individual users, comprising: The personal interaction module is used to collect multimodal personal interaction data with user authorization and generate unified structured interaction events with timestamps. The hierarchical storage module is used to store the unified structured interactive events in layers according to the event log layer, short-term context memory layer, long-term semantic memory layer and fact version library layer, and output the retrieval results of the hierarchical storage. The Person Assumption Module is used to generate auditable person assumption entries based on the hierarchically stored content. The auditable person assumption entries include the assumption statement, source of evidence, confidence level, last update time, counterexample record, applicable scenario, authorization status, and rebuttal flag fields; wherein the confidence level is updated using a Bayesian update mechanism. The candidate output module is used to generate candidate output results for digital clones based on the preset master generation model, through the retrieval results stored in the hierarchical storage and the auditable person hypothesis entries; The guard evaluation module is used to evaluate the candidate output results based on a guard model that is independent of the preset master generation model and to calculate a comprehensive score. The gating decision module is used to make gating decisions based on the comprehensive score and the risk level of the candidate output result. If the comprehensive score is greater than or equal to a first threshold and the risk level of the candidate output result is less than a second threshold, the candidate output result is allowed to be output; otherwise, a restricted policy is triggered to obtain the evaluation result of the guard model. The closed-loop write-back module is used to obtain user feedback and strategy adoption results, and write back the evaluation results of the guard model, user feedback and strategy adoption results to the hierarchical storage and the auditable person hypothesis entries through the closed-loop write-back channel, and generate strategy effect records for updating later personalized strategy selection; wherein the closed-loop write-back does not modify the parameter weights of the main generation model.

[0007] A third aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the above-described metacognitive digital avatar intelligent agent method for individual users.

[0008] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described metacognitive digital avatar intelligent agent method for individual users.

[0009] The beneficial effects of the metacognitive digital avatar intelligent agent method and system for individual users provided in this application are as follows: This application solves the shortcomings of existing digital avatars, such as fragmented memory, black box profiles, uncontrolled output illusions, and security risks in iteration, by using a hierarchical four-layer memory storage, auditable person hypotheses, dual-path verification of independent guard models, and a complete closed-loop link that does not change the weights of the main model. The hierarchical storage uniformly manages multimodal interaction data and can output retrieval results, providing a complete source of evidence for personalized generation and security verification; the independent guard model is completely isolated from the weights and inference graph of the main generation model, realizing mandatory gating before output and intercepting high-risk content lacking evidence support; the closed-loop write-back only updates the structured memory and hypothesis entries without changing the weights of the large model, avoiding the risk of catastrophic forgetting and backdoor implantation caused by online fine-tuning, and the entire update is traceable and auditable, taking into account long-term personalized adaptation, output security and controllable privacy, and can adapt to long-term human-machine collaboration in all scenarios such as learning, companionship, and elderly care. Attached Figure Description

[0010] Figure 1 A flowchart illustrating a metacognitive digital avatar intelligent agent method for individual users provided in an embodiment of this application; Figure 2This is a structural block diagram of a metacognitive digital avatar intelligent agent system for individual users provided in an embodiment of this application; Figure 3 This is a schematic block diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0011] 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 this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0012] To make the purpose, technical solution, and advantages of this application clearer, the following will be described in conjunction with the appendix. Figure 1 -Appendix Figure 3 The following is an explanation using specific examples.

[0013] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating a method for a metacognitive digital avatar intelligent agent for individual users, provided in an embodiment of this application. The method includes: S101: Collect multimodal personal interaction data with user authorization and generate unified structured interaction events with timestamps.

[0014] In this embodiment, user authorization is the user's active and explicit granting of permission to the system to collect and store their own interaction data. It is divided into long-term authorization, one-time authorization, and scenario-based authorization. Without authorization, the system is prohibited from collecting any personal interaction information. Users can view and withdraw authorization at any time, which is a prerequisite for privacy compliance.

[0015] In this embodiment, data collection involves acquiring raw data generated by user human-computer interaction in real-time or offline via terminal hardware and software interfaces, including three methods: real-time streaming acquisition, offline local caching, and online batch acquisition. Multimodal refers to various types of interactive information on different carriers, including text, voice, images, video, wearable sensor data, schedules and tasks, cognitive response records, confidence scores, and AI dialogue logs.

[0016] In this embodiment, personal interaction data refers to the raw, unprocessed material generated during the entire process of interaction between the user and the digital avatar, mobile phone / speaker / robot terminal. The data is in a messy format and lacks a unified standard, serving as the original data source for generating interaction events.

[0017] In this embodiment, the timestamp is a standard time marker accurate to the second, bound to each interaction record, used for time-series sorting, factual version time decay calculation, audit backtracking, and behavioral time-series analysis. Unified structuring involves standardizing and encapsulating multimodal heterogeneous raw data. All interaction records share a fixed set of field templates; the fields are fixed, and the machine can uniformly parse and store them in the database, eliminating various data format differences. Standard fields include a unique event ID, timestamp, modality type, interaction content summary, scene tag, authorization flag, and initial sentiment screening result.

[0018] In this embodiment, the interaction event is the smallest standardized human-computer interaction data unit. One complete human-computer interaction corresponds to one interaction event, which is the bottom-level input for the entire process of hierarchical memory, character assumption, and guard evaluation.

[0019] S102: Store the unified structured interactive events in layers according to the event log layer, short-term context memory layer, long-term semantic memory layer and fact version library layer, and output the retrieval results of the layered storage.

[0020] In this embodiment, the unified structured interaction event is the smallest human-computer interaction data unit with a precise timestamp after authorized collection, standardized encapsulation, and all multimodal data such as text / voice / image / wearable devices are unified with a fixed set of fields to eliminate format differences, and it is the input data source of this embodiment.

[0021] In this embodiment, the hierarchical storage does not use a single database to store all interactive data. Instead, it divides the data into four independent storage partitions according to the data lifecycle, information type, and retention value. The data is written in categories and indexed in partitions to achieve differentiated management and accurate retrieval.

[0022] In this embodiment, the event log layer is the first layer of a four-layer storage system, serving as the original interaction pipeline storage area. It retains all original fields of each original interaction event without merging or deletion, only performing chronological archiving; this is used for full auditing, behavior tracing, and retrieval of original evidence. The event log layer stores all original interaction information, including the complete event ID, timestamp, original multimodal summary, scene, initial sentiment screening, authorization marker, and single response record.

[0023] In this embodiment, the short-term context memory layer is the second layer of the four-layer storage, storing temporary context data within a single session and a short period. It only retains temporary dialogues, to-do lists, temporary preferences, and short-term emotions within a recent session. After the session ends / times out, these are automatically archived to the long-term layer, avoiding long-term occupation of high-frequency retrieval space. The short-term context memory layer stores the current chat context, temporary plans for the day, short-term user needs, and temporary data from a single training session.

[0024] In this embodiment, the long-term semantic memory layer is the third layer of the four-layer storage, serving as a stable, session-based storage area for user-defined information. Stable and long-term valid user facts are extracted from massive event logs and short-term memory, deduplicated, and integrated into semantic knowledge units for long-term persistent storage and frequent retrieval. The long-term semantic memory layer stores information that does not easily change, such as user habits, long-term goals, allergies / chronic disease constraints, stable preferences, long-term learning plans, and fixed routines.

[0025] In this embodiment, the fact version repository layer is the fourth layer of a four-layer storage system. It is specifically used to store multiple versions of the same fact with different statements, at different times, and with different confidence levels. It is equipped with version arbitration calculation logic to distinguish between active and valid versions and historical conflicting versions. It is specifically used to verify memory conflicts and trace changes in source information. The storage content of the fact version repository layer includes multiple different versions of the same fact, the confidence level of each version, the generation time, counterexample associations, and retrieval frequency.

[0026] In this embodiment, the retrieval results of the hierarchical storage are provided to the outside world after the hierarchical storage is completed. The unified retrieval interface is provided to the outside world according to the four-layer storage joint retrieval, and all relevant data matched by the four layers are integrated to form a standardized retrieval data package. The internal contents include: credible and valid facts (active version), conflict and opposition historical versions, short-term context, and original event log evidence, which are the inputs for subsequent generation of character hypotheses and guard model evaluation.

[0027] S103: Generate auditable person hypothesis entries based on hierarchical storage content. The auditable person hypothesis entries include hypothesis statement, source of evidence, confidence level, last update time, counterexample record, applicable scenario, authorization status, and rebuttal flag fields; among which, the confidence level is updated using a Bayesian update mechanism.

[0028] In this embodiment, the content of the layered storage refers to all data within the four-layer storage system: the original interaction flow of the event log layer, the short-term contextual memory of the session context, the long-term semantic memory of stable user characteristics, and the multi-version fact records of the fact version library; it is the source of all evidence for generating hypotheses, including positive materials supporting the hypotheses and opposing materials that contradict the hypotheses, which can provide a complete and traceable chain of evidence for each hypotheses and avoid unfounded inferences.

[0029] In this embodiment, the auditable profile is assumed to be a standardized, structured data unit used to summarize and describe a user's long-term behavior, cognitive habits, and preference characteristics. "Auditable" means that the entire record, every modification, supporting evidence, and contradictory cases can be queried and traced. Users can view, modify, refute, and deactivate the profile, eliminating the possibility of a black-box profile. Each profile is the smallest independently stored unit.

[0030] In this embodiment, the hypothetical statements are the core text of the entries, using standardized statements to summarize the user's stable behavior, cognitive characteristics, lifestyle habits, and needs and preferences. For example: the user frequently exhibits overconfidence in answering questions; the user is suited to low-intensity memorization and review at night.

[0031] In this embodiment, the source of evidence is the unique set of all original interaction event numbers that support the validity of this hypothesis. It can be redirected to the hierarchical storage event log to retrieve the original dialogue, response, and sensor records, thus enabling evidence tracing.

[0032] In this embodiment, the confidence level is a value ranging from 0 to 1. The quantification system determines the credibility of the hypothesis in relation to the user's actual situation. The larger the value, the stronger the reliability. The main generator model and the guard model will prioritize calling the high-confidence hypothesis to make personalized strategies.

[0033] In this embodiment, the most recent update time is a timestamp accurate to the second, recording the confidence level of this hypothesis, counterexamples, and the last change time of any field in the statement, for use in timeliness filtering, audit tracing, and distinguishing between old and new features.

[0034] In this embodiment, the counterexample record is a set of interaction event IDs that contradict the conclusion of this hypothesis; it is directly retrieved when the guard model calculates the memory conflict degree, and is used to identify output contradictions and lower unreasonable high confidence levels.

[0035] In this embodiment, the applicable scenario is limited to business scenario tags (learning, workplace, health care, fraud prevention for the elderly, family companionship, etc.) where this assumption takes effect. The strategy is only calculated when the scenario is matched to prevent incorrect inference across scenarios.

[0036] In this embodiment, the authorization status is a binary switch flag (authorized / revoked), which is controlled by the user. If the authorization is revoked, the system will prohibit the retrieval of the hypothetical response, plan, and suggestion throughout the entire process.

[0037] In this embodiment, the rebuttal flag is a failure status flag (not refuted / rebutted); when the confidence level is less than the preset rebuttal threshold, or when the user manually refutes the item, it is set to refuted, and after being marked, the item will no longer participate in the personalized strategy calculation.

[0038] In this embodiment, the Bayesian update mechanism iteratively updates the confidence quantification rule through log-likelihood increment and Sigmoid function, with the formula: p←σ(logit(p)±δ); adding positive evidence increases confidence, while adding contradictory counterexamples decreases confidence; only the confidence field within an entry is modified, without altering the weights of the master generator model, and all update operations are logged for auditability.

[0039] Where p is the current confidence level of the hypothesis entry, σ() is the Sigmoid function, logit(p)=log(p / (1-p)), and δ is the log-likelihood increment of a single piece of evidence, with positive evidence taking a positive sign and negative examples taking a negative sign; when p is less than the preset rebuttal threshold or is actively rebutted by the user, the hypothesis entry is marked as a rebuttal and is no longer used for policy selection; the update only applies to the confidence field of the structured hypothesis entry and does not modify the parameter weights of the main generative model, thus distinguishing it from the online model weight update of reinforcement learning.

[0040] The preset rebuttal threshold is determined in three ways: First, a manual static preset method, which sets a global basic threshold of 0.3 based on general human-computer interaction business experience. If the confidence level is less than this value, the hypothesis is considered invalid, which is suitable for most common daily companionship and learning scenarios. Second, an offline dataset calibration method, which collects a large number of labeled hypothesis samples, iterates through different thresholds, and selects the optimal F1 score as the threshold for specific scenarios, suitable for highly rigorous scenarios such as health care and fraud prevention for the elderly. Finally, an online adaptive adjustment method, which dynamically raises or lowers the threshold based on the user's real-time metacognitive quotient, tightening the judgment criteria for overconfident users and appropriately relaxing the screening conditions for users with excellent cognitive levels. At the same time, a double fallback rule is added: even if the confidence level does not reach the threshold, if the user manually initiates a rebuttal or a batch of contradictory counterexamples appear in a short period of time, the item is also marked as rebuttal.

[0041] S104: Based on the preset master generation model, candidate output results for digital clones are generated through hierarchically stored retrieval results and auditable person hypothesis entries.

[0042] In this embodiment, the preset main generation model is a pre-deployed large language basic model. As the core carrier of generated content, it is only responsible for producing preliminary content based on input materials, has no independent verification capability, is completely isolated from the independent guard model, and does not participate in risk or credibility review throughout the process.

[0043] Specifically, the pre-defined main generation model is divided into four layers: The first layer is the multi-source information input layer, which loads user interaction questions, hierarchically stored retrieval data, auditable person assumption entries, and scene tags; the second layer is the context fusion encoding layer, which uniformly encodes and splices together all input information to construct a complete reasoning context that binds user facts and personalized characteristics; the third layer is the logical reasoning generation layer, which performs semantic inference and content organization through context fusion to generate original response content that adapts to user needs; the fourth layer is the candidate output distribution layer, which outputs candidate output results that have not undergone security verification, which are transmitted unidirectionally to the guard model for evaluation, without a direct external output channel.

[0044] The pre-built master generator model adopts a standard Transformer generative architecture, performing only the single function of content generation. It is completely physically isolated from the guard model, and the two do not share parameter tensors, word embeddings, or inference caches. This model only has built-in basic modules for text encoding and sequence generation, and does not include review units such as evidence verification, risk assessment, or conflict detection. It also lacks autonomous gating and policy adjustment logic. The model size is unlimited, and any pre-trained model of any scale can be used to generate a large model. It only serves as a pre-processing material generation unit, providing the original text to be evaluated for subsequent metacognitive verification, and does not participate in the output release decision throughout the entire process.

[0045] The preset main generation model inputs and outputs include three types of data: real-time user interaction questions and scene identifiers, retrieval results from a four-layer hierarchical storage system, and auditable person hypothesis entries carrying confidence levels and behavioral characteristics. The model output is the candidate output for digital avatars, which is raw generated text without verified credibility or security risk screening. This output serves only as input material for the six detection sub-components of the guard model and does not have external publishing rights. Whether it can be displayed externally is entirely controlled by the independent guard model's gating decision.

[0046] In this embodiment, the retrieval results stored in a hierarchical manner are standardized data packets output after the retrieval is performed, including original evidence from event logs, short-term session context, long-term stable user characteristics, and all data from multiple versions of facts, providing complete user behavior data for the main model.

[0047] In this embodiment, the auditable person hypothesis entry is a valid user feature entry that has generated and completed confidence updates and has not been marked with a rebuttal state, providing personalized constraints such as user habits, cognitive biases, and applicable scenarios.

[0048] In this embodiment, the main generation model reads two types of input materials, integrates user features and the current session context, and infers and produces raw output content that has not undergone security review. The digital avatar is a metacognitive intelligent companion and decision-making aid intelligent agent carrier for single individuals, running on terminals such as mobile phones, speakers, and companion robots.

[0049] In this embodiment, the candidate output results are the original content generated by the main model that has not yet undergone gating verification by the guard model. These contents may contain hallucinations or high-risk expressions and cannot be directly displayed. The candidate output results include candidate responses, candidate suggestions, or candidate proxy actions. Candidate responses are dialogue responses in text or voice format. Candidate suggestions are guidance content such as study plans, schedules, health reminders, and career solutions. Candidate proxy actions are control commands issued to the accompanying robots and equipment, such as timed reminders and training start commands.

[0050] S105: Based on a guardian model independent of the preset master generator model, perform multidimensional metacognitive feature evaluation on candidate output results and calculate a comprehensive score.

[0051] In this embodiment, the guard model and the main generation model are two completely independent inference units. They are not coupled and do not share model structures or parameters, thus decoupling the generation and verification processes. The guard model is a discriminant model responsible for content credibility and risk compliance verification. It does not generate dialogues / suggestions, but only scores the candidate outputs produced by the main model and determines whether to allow them, acting as a security gate.

[0052] Specifically, the guardian model, independent of the main generative model, includes six sub-components: evidence checker, emotion checker, risk classifier, conflict detector, independent reviewer, and strategy selector. The evidence checker outputs evidence coverage ∈ [0,1], the emotion checker outputs emotion matching ∈ [0,1], the overconfidence checker outputs overconfidence ∈ [0,1], the risk classifier outputs risk level ∈ [0,1], the conflict detector outputs memory conflict ∈ [0,1], and the cognitive gain estimator outputs cognitive gain ∈ [0,1].

[0053] The guard model is hierarchically divided into: Raw Input Layer: Receives two types of isolated inputs: candidate output text from the main generator model and retrieval results from hierarchical storage and fact version database retrieval, with no intermediate activation from the main generator model and KV cache input; Six-Dimensional Feature Extraction Sub-component Layer: Deploys in parallel an evidence checker, sentiment checker, overconfidence checker, risk classifier, conflict detector, and cognitive gain estimator, outputting six sets of 0-1 standardized metacognitive features: evidence coverage, sentiment matching, overconfidence, risk level, memory conflict, and cognitive gain; Comprehensive scoring fusion. The fusion layer: Through a linear weighted fusion formula, five sets of normalized weighted coefficients and independent conflict penalty coefficients λ are configured to complete reverse correction and conflict deduction calculations, and output a comprehensive score Q; the weights support offline annotation set calibration and online incremental fine-tuning, and can also replace nonlinear, tree model, and Bayesian fusion schemes; Gated decision output layer: The comprehensive score is compared with the first threshold and the risk second threshold to complete the joint judgment, and output three types of decision instructions: pass, reject, and downgrade. The complete six-dimensional features, comprehensive score, and risk level are simultaneously output as the guard assessment results and sent to the output execution module and closed-loop write-back channel.

[0054] The guard model is a discriminative review architecture completely independent of the main generator model. It is built from six functional sub-components, a feature fusion unit, and a gating decision unit. In engineering, it meets three layers of physical isolation constraints: independent parameter weights, not sharing the main model's word embeddings and tensors, and using a lightweight pre-trained discriminative model or a rule-based, Bayesian, or decision tree-based architecture without neural networks; an independent inference graph, receiving only the main model's final text and hierarchical memory retrieval data, isolating all intermediate computational information from the main model; and independent control, with the output gate decision having mandatory force, preventing the main generator model from bypassing the guard to directly publish content. The pipeline executes sequentially by default: "Main model generates candidates - guard multi-dimensional evaluation - guard gating decision - module publishing output." In low-latency scenarios, speculative verification can be initiated in parallel, but the final publishing authority still belongs to the guard model. The model weights and fusion parameters do not rely on manually fixed values. Instead, they are based on offline labeled samples with binary pass / reject labels. Batch calibration is completed using convex optimization algorithms such as logistic regression, LBFGS, and Adam. After going live, historical decision records and user feedback are continuously collected to achieve online incremental parameter tuning. Various nonlinear fusion, rule review, and other equivalent solutions all belong to the same implementation method.

[0055] The guard model's input includes only two types of data: candidate responses, candidate suggestions, and candidate agent actions output by the main generator model after inference; and credible evidence, conflicting facts, and long-term user interaction records retrieved from hierarchical storage and fact version databases. It does not incorporate the main model's attention, intermediate layer features, or key-value caches. The output consists of two parts: first, gating control instructions, including allowing output and triggering various restricted strategies; second, standardized guard evaluation results, carrying six-dimensional metacognitive features, a comprehensive score, risk level, and execution strategy type. Compared to generator-verification architectures such as AlphaProof, OpenAI Critic, and Self-Refine, the guard model is not limited to single-output scoring. It can link four layers of hierarchical memory, auditable user hypothesis entries, and a closed-loop write-back channel to build a long-term personalized metacognitive closed loop. Simultaneously, the closed-loop write-back only updates the database's structured fields and user hypotheses, without altering any parameters or weights of the main generator model. This is fundamentally different from reinforcement learning schemes like RLHF, PPO, and DPO that rely on online fine-tuning of model weights, avoiding the hallucination and emotional bias transmission problems caused by coupling with the main model.

[0056] In this embodiment, the multidimensional metacognitive feature assessment quantifies the rationality and safety of candidate outputs from six dimensions: evidence coverage, memory conflict, emotional fit, overconfidence index, risk level, and cognitive benefit. The comprehensive score is a single quantitative score calculated by integrating the six-dimensional metacognitive features, which serves as the core criterion for gating and allowing access.

[0057] In this embodiment, the guard model and the preset main generator model do not share parameter weights or inference graphs, and the guard model has mandatory gating authority at the release gate, using only candidate output results and hierarchically stored retrieval results as input. Not sharing parameter weights means that the guard and main generator models have their own independent parameter matrices, and there is no shared weight layer; updates to one will not affect the other. Not sharing the inference graph means that their inference calculation processes and network forward computation graphs are completely separate, with separate scheduling and computation, and no reuse of computing power or computational logic. Using only the candidate output results and hierarchically stored retrieval results as input means that the guard model only receives two types of input: first, the candidate content to be reviewed generated by the main model; and second, all historical facts and multiple versions of evidence of the user obtained from hierarchical storage retrieval, with no other external input data.

[0058] In this embodiment, the candidate output results are the unapproved original content generated by the main generation model, including candidate responses, candidate suggestions, and candidate proxy actions. The hierarchically stored retrieval results are the output four-layer memory joint retrieval data packets, including original interactions, long-term features, and multiple versions of facts, used to calculate conflict and sufficiency of evidence.

[0059] S106: Make a gating decision based on the comprehensive score and the risk level of the candidate output results. If the comprehensive score is greater than or equal to the first threshold and the risk level of the candidate output results is less than the second threshold, the candidate output results are allowed to be output. Otherwise, the restricted policy is triggered to obtain the evaluation result of the guard model.

[0060] In this embodiment, the comprehensive score is a single quantitative score calculated by the guard model by integrating six-dimensional metacognitive features. The higher the score, the more sufficient, conflict-free, and user-suitable the candidate output evidence is, and the lower the risk is. It serves as the basis for credibility judgment.

[0061] In this embodiment, the risk level of the candidate output results is divided into four levels based on whether there is high-risk information such as self-harm, medical treatment, missing minors, elderly people, and fraud in the candidate content, which are low, medium, high and extremely high risk in order from low to high.

[0062] In this embodiment, the gating decision is the pass / interception judgment logic of the guard model. As a security gate for the output of digital clone content, the content can only be displayed to the outside world if both conditions are met at the same time; otherwise, intervention measures are executed.

[0063] In this embodiment, the first threshold is the release threshold for the comprehensive score, used to define the acceptable standard for content credibility; a score of 100 or higher indicates that the content is credible. The second threshold is the release threshold for the risk level; only low-risk content below this level is allowed to be directly output, while high-risk content is blocked.

[0064] The first threshold is determined by a static preset method based on human experience. Combined with the default weight of the linear weighted fusion formula and the general evaluation standard of the six-dimensional metacognitive index, it is uniformly set to 0.55. This is suitable for all common scenarios such as learning, caregiving, and health reminders. A comprehensive score greater than or equal to 0.55 indicates that the content evidence is sufficient, there are no obvious factual conflicts, and it meets the needs of users.

[0065] The second threshold is obtained by collecting a large number of high and low risk candidate output samples offline, traversing the risk boundaries at all levels and selecting the lowest false negative rate as the second threshold for health care scenarios; it can also be obtained online for manual review and feedback to dynamically adjust the threshold, and automatically lower the second threshold for use by minors, allowing only the lowest level of risk content to be output directly.

[0066] In this embodiment, allowing candidate output results to be output is considered qualified, and the response, suggestion, or device instruction is directly pushed and displayed to the user or issued to the robot.

[0067] In this embodiment, the restricted strategies include at least one of the following: re-retrieval, reducing assertion strength, clarifying inquiry, strategy switching, transferring to human intervention, or conservative output. Restricted strategies are intervention measures initiated when content credibility is insufficient or risk exceeds limits, used to correct, replace, or transfer content for review. Specifically, re-retrieval involves re-retrieving hierarchical memory to supplement more evidence and regenerate candidate content; reducing assertion strength weakens the definitive statements in the original text and removes absolute conclusions; clarifying inquiry proactively asks the user questions to supplement missing information; strategy switching involves changing to a less adaptable and more conservative interaction strategy; transferring to human intervention pushes the request to human customer service / caregiver for review; and conservative output provides a neutral, non-conclusive, and conservative response.

[0068] In this embodiment, the evaluation result of the guard model is to output a complete record after the gating decision is completed, including metacognitive scores of each dimension, comprehensive score, risk level, whether to allow passage, and the type of restricted strategy executed, which is used for subsequent closed-loop write-back and retention.

[0069] S107: Obtain user feedback and strategy adoption results, and write back the evaluation results of the guard model, user feedback and strategy adoption results to the hierarchical storage and auditable person assumption entries through the closed-loop write-back channel, and generate strategy effect records for updating the later personalized strategy selection; the closed-loop write-back does not modify the parameter weights of the main generation model.

[0070] In this embodiment, user feedback refers to the positive / negative evaluations given by users regarding the output content of the digital clone, including explicit feedback (likes, rebuttals, content corrections, negative suggestions) and implicit feedback (whether to execute the recommendation plan, whether to ignore reminders, whether to ask questions again). The strategy adoption result records whether the user accepts the output strategy selected by the guard, categorized as adoption or rejection, such as whether to adopt conservative output, whether to accept clarification questions, and whether to transfer to human intervention. The evaluation result of the guard model is the complete review data generated in the gating decision-making process, including six-dimensional metacognitive feature scores, comprehensive score, risk level, whether to allow access, and the type of restricted strategy executed.

[0071] In this embodiment, the closed-loop write-back channel is an independent data return transmission link used to synchronously write various feedback data after interaction to the underlying storage and the character hypothesis module, completely isolated from the main generation model inference link. Write-back synchronously updates the feedback and evaluation data to the designated storage unit, modifying only the structured text and numerical fields, without involving changes to the model weights.

[0072] In this embodiment, the hierarchical storage is a four-layer memory storage system (event log layer, short-term contextual memory layer, long-term semantic memory layer, and fact version repository layer), which writes back newly added interactive feedback events and supplements fact versions. The auditable person hypothesis entries are generated user characteristic entries, and during write-back, the confidence level of the entries is updated using a Bayesian mechanism based on positive and negative user feedback, and counterexample records are supplemented.

[0073] In this embodiment, the strategy effect record is a newly created independent data table that records each round's interaction strategy, guard score, user adoption status, and feedback quality, serving as a reference for subsequent personalized strategy scheduling. The later personalized strategy selection involves referencing historical strategy effect records and updated user assumptions when generating candidate outputs and selecting interaction strategies, to match responses and solutions that better suit user preferences and risk tolerance.

[0074] In this embodiment, the parameter weights of the main generator model are not modified. The entire closed-loop write-back process only modifies the structured data in the database and does not make any fine-tuning or updates to the network parameters and weights of the main generator model, thus avoiding the risk of model forgetting or backdoors.

[0075] As can be seen from the above, this application solves the shortcomings of existing digital avatars, such as fragmented memory, black-box portraits, uncontrolled output illusions, and security risks in iterations, through a complete chain of hierarchical four-layer memory storage, auditable character assumptions, dual-path verification of independent guard models, and closed-loop without changing the weights of the main model. Hierarchical storage uniformly manages multimodal interaction data and can output retrieval results, providing a complete source of evidence for personalized generation and security verification; the independent guard model is completely isolated from the weights and inference graph of the main generation model, realizing mandatory gating before output and intercepting high-risk content lacking evidence support; the closed-loop write-back only updates the structured memory and assumption entries without changing the weights of the large model, avoiding the risk of catastrophic forgetting and backdoor implantation caused by online fine-tuning, and the entire update is traceable and auditable, taking into account long-term personalized adaptation, output security and controllable privacy, and can adapt to long-term human-machine collaboration in all scenarios such as learning, companionship, and elderly care.

[0076] In one embodiment of this application, unified structured interaction events are stored hierarchically according to an event log layer, a short-term context memory layer, a long-term semantic memory layer, and a fact version database layer, and the hierarchical storage retrieval results are output, including: In the fact version repository layer, when faced with multiple versions of the same fact, the arbitration score of each version is calculated; The version with the highest arbitration score is designated as the active version, while the remaining versions are retained as historical versions for audit backtracking.

[0077] In this embodiment, the unified structured interaction event is a standardized human-computer interaction data packet with fixed fields and timestamps, which is the raw input data stored in layers. The event log layer, short-term context memory layer, long-term semantic memory layer, and fact version library layer form a four-layer hierarchical storage architecture: the event log layer stores the full raw flow; the short-term context memory stores the temporary context of the current session; the long-term semantic memory stores stable long-term user characteristics; and the fact version library layer specifically stores multiple different descriptive versions of the same objective fact.

[0078] In this embodiment, hierarchical storage archives interactive events independently in four layers based on data lifecycle and information attributes, with indexes built for each layer to support cross-layer joint queries. The retrieval result of hierarchical storage is a data package integrated after joint retrieval of the four layers, including the original logs, short-term context, long-term stable features, and all versions of facts, for subsequent use in hypothesis generation and guard assessment.

[0079] In this embodiment, multiple versions of the same fact refer to multiple descriptions of the same objective user information generated at different times and in different interaction scenarios, which may differ or even contradict each other. The arbitration score is a quantitative score calculated separately for each version of the fact, which is derived by combining version confidence, generation time, and recent retrieval frequency, and is used to judge the credibility and timeliness of the version.

[0080] Specifically, when the hierarchical memory and fact version library module faces multiple versions of the same fact, it calculates the version arbitration score using the following formula: score_i = α·c_i·exp(-Δt_i / τ) + β·recency_i, where c_i is the confidence level of the i-th version, Δt_i is the interval between the version and the current time, τ is the time decay constant, α and β are non-negative weights, and recency_i is defined as the frequency of the version being actually called in the most recent N retrievals (N is the preset window length, preferably N=20 in this embodiment); and the version with the highest score is taken as the active version, while the other versions are retained as historical versions for auditing, backtracking, and counterexample queries.

[0081] In this embodiment, the active version is the factual version with the highest arbitration score and the best credibility and timeliness. The system defaults to using this version as the benchmark for generating character hypotheses and reasoning about candidate content. Historical versions are all other factual versions besides the active version. They are not used as the benchmark for reasoning, but they are completely and persistently stored and cannot be deleted. Audit backtracking allows for the retrieval of all historical versions to view the entire process of factual changes and contradiction records when subsequently checking for memory conflicts, user corrections, and content tracing, achieving full-process auditability.

[0082] As can be seen from the above, this embodiment, through a version arbitration mechanism of the limited fact version library, designs a quantitative scoring system to screen active versions for multiple version conflicts of the same fact, while retaining the remaining historical versions for auditing and backtracking. This effectively solves the shortcomings of existing digital human memory inconsistencies, loss of old data, and inability to trace the source of conflicts. By comprehensively considering confidence, timeliness, and retrieval frequency through arbitration scores to screen valid facts, the timeliness and credibility of hierarchical storage and retrieval results are ensured. The permanent retention of historical versions forms a complete chain of evidence, allowing users and the platform to retrieve past fact records at any time for error correction and verification. This improves the interpretability and compliance auditing capabilities of the memory system, providing a complete data source of opposing versions for the guard model to calculate memory conflict degree, thereby reducing the probability of the digital clone outputting contradictory factual content from the bottom layer.

[0083] In one embodiment of this application, a multidimensional metacognitive feature evaluation of candidate output results is performed based on a guard model independent of a preset master generator model, including: The multidimensional cognitive features used by the Guardian Model to assess include at least evidence coverage, memory conflict, emotion fit, overconfidence index, risk level, and cognitive benefit. Among them, the candidate output results and the evidence marked as credible in the hierarchical storage retrieval results are mapped to the same semantic vector space, and the evidence coverage is calculated by taking the maximum semantic similarity or the average of the top k similarities. The candidate output results and the fact versions marked as refutation, low confidence, and semantic opposition in the hierarchical storage retrieval results are mapped to the same semantic vector space, and the memory conflict degree is calculated by taking the maximum semantic similarity. The user's historical dominant emotion vector and candidate output emotion tendency vector are obtained and mapped to a unified emotion label space. The emotion matching degree is calculated by cosine similarity. Obtain the assertion strength and corresponding evidence confidence in the candidate output results, calculate the difference or ratio, and obtain the overconfidence index. Risk keywords, user status, and scenario sensitivity are obtained to jointly determine the risk level and classify it into four risk levels: Level 1, Level 2, Level 3, and Level 4. Among them, Level 1 is lower than Level 2, Level 2 is lower than Level 3, and Level 3 is lower than Level 4. The expected gains of candidate outputs on users' learning goals, planning goals, and emotion regulation goals are estimated to obtain cognitive gains.

[0084] In this embodiment, credible evidence consists of user facts and interaction events marked as active versions, high-confidence, and undisputed in hierarchical storage. The same semantic vector space is a unified sentence embedding representation space (e.g., the Sentence-BERT vector space), which transforms candidate outputs and historical evidence into equal-dimensional vectors for semantic similarity calculation. Semantic similarity is a quantitative value measuring the degree of matching between candidate outputs and historical evidence content, ranging from 0 to 1. The average of the top k similarities is calculated by averaging the similarities of the top k most similar credible evidence pieces to the candidate output, serving as coverage; alternatively, the maximum similarity can be directly used. Evidence coverage characterizes how much historical credible evidence supports the candidate output; a higher value indicates more evidence and a lower probability of illusion.

[0085] In this embodiment, the refutation-state, low-confidence, and semantically contradictory fact versions represent historical fact records in the fact version library that have low arbitration scores, are marked with refutation, and contradict the candidate output viewpoint. The memory conflict degree is the maximum semantic similarity between the candidate output and existing conflicting facts; the larger the value, the more serious the contradiction between the current output and the user's historical facts.

[0086] In this embodiment, the user's historical dominant emotion vector is extracted from short-term episodic memory and interactive events, and mapped to a standardized emotion vector (dimensions such as joy / anger / sadness / fear / calmness, etc.). The candidate output emotion tendency vector is generated by the guard model recognizing the inherent emotion guidance of the response and generating an emotion vector of the same dimension. The unified emotion label space fixes the discrete / continuous emotion representation dimensions, unifying the emotion vector format for both the user and the output. Cosine similarity is calculated by taking the cosine of the angle between two sets of emotion vectors to measure the degree of emotional fit; a higher value indicates better empathy matching. Emotion matching degree quantifies whether the clone's output emotion matches the user's current emotional state.

[0087] In this embodiment, assertion strength is the degree of certainty in the candidate output, ranging from 0 to 1, with higher strength indicating absolute conclusions. Evidence confidence is the average confidence value of all credible evidence supporting this output. Difference / ratio calculation: Overconfidence index = Assertion strength - Evidence confidence or Overconfidence index = Assertion strength / Evidence confidence; a larger value indicates that the model draws conclusions without evidence, and the more severe the overconfidence.

[0088] In this embodiment, the risk keywords are a high-risk sensitive word library including medical medication, self-harm, fraud, danger to minors, and missing elderly. User status represents high-risk user identity tags such as elderly, minors, chronic disease patients, and those experiencing emotional breakdowns. Scenario sensitivity represents highly sensitive scenario markers such as medical care, financial management, and elderly care. The joint judgment is a weighted comprehensive score based on keywords, user identity, and scenario to classify risk into four levels. The four risk levels are: Level 1 (low risk), Level 2 (medium risk), Level 3 (high risk), and Level 4 (extremely high risk), with higher numbers indicating higher risk.

[0089] In this embodiment, the learning objective / planning objective / emotion regulation objective is the user's long-term training, schedule planning, and emotional soothing needs recorded by the system. Expected gain is the estimated degree to which the candidate output can help the user achieve their goal, ranging from 0 to 1, with higher values ​​indicating greater value to the user. Cognitive gain quantifies the positive value of this suggestion / response for improving the user's cognition, completing tasks, and relieving emotional distress.

[0090] As can be seen from the above, this embodiment quantifies the calculation logic by limiting the six dimensions of metacognitive features of the guard model, comprehensively evaluating candidate outputs from the perspectives of evidence credibility, memory conflict, emotional fit, overconfidence, risk, and usage benefits, which differs from coarse-grained verification schemes that only score. Each feature is calculated in a unified semantic / emotional space through hierarchical storage and retrieval results, making the quantitative indicators objective; the risk level is hierarchically divided into four levels and can accurately identify high-risk scenarios such as self-harm, medical treatment, and minors, intercepting dangerous outputs in advance; emotion matching is based on users' historical interaction data to align emotional tendencies, avoiding empathy misalignment; multi-dimensional joint evaluation constructs a refined security verification system, reducing the probability of model hallucinations, emotional fragmentation, and high-risk inappropriate assertions, and improving the reliability and adaptability of the digital clone's output content.

[0091] In one embodiment of this application, calculating the comprehensive score includes: A linear weighted fusion method was used to fuse and calculate six multidimensional metacognitive features, namely evidence coverage, memory conflict, emotion fit, overconfidence index, risk level, and cognitive benefit, to obtain a comprehensive score. Five sets of non-negative weighted coefficients are configured, corresponding to evidence coverage, overconfidence correction, emotional fit, risk correction, and cognitive benefit, respectively, and the sum of all weighted coefficients is 1; at the same time, an independent conflict penalty coefficient is configured to deduct and correct the memory conflict degree. During the fusion calculation process, the overconfidence index and risk level are reversed to reduce the overall score of highly overconfident and high-risk content; the memory conflict degree is combined with the conflict penalty coefficient to deduct points, thereby weakening the overall score of candidate outputs with factual conflicts. The weighting coefficients are calibrated and optimized using an offline labeled dataset that includes feature vectors and pass / reject binary labels. During operation, the weighting coefficients can be incrementally calibrated online based on the historical gating decision records of the guard model and the final user feedback.

[0092] In this embodiment, the comprehensive score is a single quantitative score output by the guard model after fusing six-dimensional metacognitive features. It is the core basis for gating decision-making to determine whether the candidate output is credible and qualified.

[0093] In this embodiment, the linear weighted fusion method uses a linear formula to sum the weighted indicators of multiple dimensions. By allocating the influence ratio of each feature on the total score with fixed weights, the calculation logic is transparent and auditable.

[0094] In this embodiment, the six types of multidimensional metacognitive features include evidence coverage, memory conflict degree, emotion matching degree, overconfidence index, risk level, cognitive benefit, and six quantitative indicators calculated in advance by the guard model.

[0095] In this embodiment, the non-negative weighting coefficients correspond to the weights of evidence coverage, overconfidence correction, emotional fit, risk correction, and cognitive gain, respectively, and are w1, w2, w3, w4, and w5. The values ​​are greater than or equal to 0, and w1+w2+w3+w4+w5=1, which controls the contribution ratio of each positive indicator to the total score.

[0096] In this embodiment, the overconfidence correction term is obtained by subtracting the overconfidence index from 1, which transforms the higher the overconfidence score into a positive bonus term, thus achieving reverse correction.

[0097] In this embodiment, the risk correction item is obtained by subtracting the risk level from 1. The higher the risk, the smaller the value of this item, thus automatically reducing the score of high-risk content.

[0098] In this embodiment, the conflict penalty coefficient is an independent hyperparameter λ, which is always positive. It is used to deduct points for memory conflict, and the greater the conflict, the more points are deducted.

[0099] In this embodiment, the reverse correction process inverts and transforms negative indicators such as overconfidence index and risk level, which are worse the higher the value, into positive indicators that participate in the weighted summation, thereby automatically reducing the total score of high overconfidence and high-risk content.

[0100] In this embodiment, the superimposed conflict penalty coefficient deduction is based on the linear summation by subtracting "conflict penalty coefficient × memory conflict degree". The more serious the factual contradiction, the more the total score is deducted.

[0101] In this embodiment, the offline labeled dataset is a training set consisting of multiple samples. Each sample includes a set of six-dimensional feature vectors and is manually labeled with binary labels (pass / reject) for offline batch optimization of weight coefficients.

[0102] In this embodiment, the offline calibration optimization of weighted coefficients utilizes a labeled dataset to solve for the optimal five sets of weighted coefficients with the goal of maximizing metrics such as F1 and ROC-AUC, replacing purely manual experience-based assignment.

[0103] In this embodiment, the historical gating decision record of the guard model is the guard judgment log retained by the system for each round: including six-dimensional features, comprehensive score, pass / interception result, and triggered restriction policy.

[0104] In this embodiment, the final user feedback results are explicit and implicit feedback data such as the user's likes, rebuttals, corrections, acceptance / rejection of the output content.

[0105] In this embodiment, online incremental calibration involves continuously collecting new decision logs and user feedback during the system's online operation, iteratively fine-tuning the five sets of weighting coefficients in small steps, and adaptively adapting to the current user and scenario distribution without the need for large-scale offline re-labeling.

[0106] Specifically, the comprehensive assessment includes fusing multidimensional metacognitive features to obtain a comprehensive score Q, where Q is calculated by the following formula: Q=w1·Evidence Coverage + w2·(1-Overconfidence Index) + w3·Emotion Matching Degree + w4·(1-Risk Level) + w5·Cognitive Benefit - λ·Memory Conflict Degree. Where w1, w2, w3, w4, w5 are non-negative weights with a sum of 1, and λ is a conflict penalty coefficient.

[0107] It can be concluded from the above that this embodiment calculates the comprehensive score by adopting a linear weighted fusion method, and is equipped with weighting coefficient calibration, conflict point deduction, and reverse correction mechanisms, so as to realize the balanced assessment of multi-dimensional metacognitive features. The reverse weight reduction is performed on overconfident and high-risk content, and superimposed penalty points are deducted for factual conflicts, which accurately suppresses the scores of candidate outputs with insufficient credibility; the weights support offline labeling calibration and online incremental adaptive adjustment, which can adapt to the assessment standards of different user groups such as the elderly, students, and workplace people and various subdivided scenarios, and dynamically optimize the verification standards without modifying the model structure; the fusion logic is simple, discriminable, and logically transparent, which solves the problems that the black-box scoring of nonlinear models is difficult to debug and the reason for point deduction cannot be traced, and balances the assessment accuracy and system maintainability.

[0108] In an embodiment of the present application, a metacognitive digital twin agent method for individual users further includes metacognitive assessment and cognitive enhancement steps: Collect the correct marks and real-time confidence scores of the user's answers to cognitive tasks, and calculate the expected calibration error, Brier score, and metacognitive quotient; Generate adaptive cognitive training tasks based on retrospective memory training tasks, color conflict inhibition tasks, selective inhibition tasks, flanker interference inhibition tasks, and spaced repetition active recall tasks; Write back the task trial indicators and metacognitive quotient assessment results to the hierarchically stored and auditable persona assumption entries, and adjust the difficulty parameters of the cognitive paradigm according to the user's training performance.

[0109] In this embodiment, the metacognitive assessment and cognitive enhancement step is an independent functional process supporting the digital twin, which is divided into two major modules: the metacognitive assessment module quantifies the accuracy of the user's self-judgment, and the cognitive enhancement module pushes targeted training tasks to improve the user's cognitive bias and overconfidence problems.

[0110] In this embodiment, the cognitive task involves system-pushed judgment, selection, and memory-based test questions to assess the user's true cognitive level and self-judgment ability. The "correct" label is a tag for each cognitive question's answer, categorized as "correct" or "incorrect," objectively recording the correctness of the answer. The real-time confidence score is a 0-1 range self-judgment score given by the user after answering the questions, representing the user's subjective confidence in their answer. The expected calibration error (ECE) quantifies the deviation between the user's subjective confidence and objective answer accuracy; a higher value indicates a greater deviation in self-judgment and a more severe problem of overconfidence / underestimation. The Brier score measures the loss metric of confidence probability prediction accuracy; a lower score indicates a greater match between the user's subjective confidence and the actual answer result, and a more accurate metacognitive judgment. The metacognitive quotient (MQ) is a 0-1 comprehensive quantitative indicator obtained by normalizing the expected calibration error and the Brier score. As a core value for measuring the user's overall metacognitive ability, a higher score indicates a more objective and accurate self-cognitive judgment.

[0111] In this embodiment, the retrospective memory training task is a specialized cognitive task that trains users to retrieve past interactive memories and differentiate between versions of facts, thereby improving memory retrieval and fact-finding abilities. The color conflict suppression task is a derivative of the classic Stroop paradigm, training users to suppress intuitive erroneous judgments and reduce overconfident behavior that leads to hasty conclusions. The choice suppression task is a multi-option decision-making suppression training exercise, developing users' ability to postpone subjective and arbitrary judgments and make decisions based on multiple pieces of evidence. The flanking interference suppression task is a discrimination training exercise under the coexistence of interfering information, improving users' cognitive ability to filter irrelevant and conflicting information and focus on valid evidence. The spaced repetitive active recall task is a memory consolidation training designed based on the Ebbinghaus forgetting curve, actively retrieving memories at different times to enhance long-term semantic memory retention. The adaptive cognitive training task is a customized training exercise with dynamically matched difficulty and type based on the user's metacognitive quotient assessment results, rather than a uniform fixed question bank.

[0112] In this embodiment, the task trial index is a complete set of process data within a single training task, including the accuracy rate of answering questions, the time taken to answer, the confidence bias, and the number of failures to suppress. The difficulty parameter of the cognitive paradigm is an adjustable parameter that controls the difficulty level of the training task, including the intensity of question interference, the capacity of memory material, the answering time limit, and the confusion of answer options.

[0113] In this embodiment, writing back to hierarchical storage and auditable person hypothesis entries involves storing the complete set of data generated from the evaluation and training into the four-layer memory storage through a closed-loop channel, synchronously updating the confidence level and counterexample records of the corresponding person hypothesis entries, and improving the user cognitive bias-related features.

[0114] In this embodiment, adjusting the difficulty parameter based on the user's training performance means that if the user's metacognitive quotient is low and the training error rate is high, the difficulty is lowered; if the user consistently and stably completes the training and the metacognitive quotient improves, the task difficulty is gradually increased to achieve adaptive training.

[0115] As can be seen from the above, this embodiment, by adding a pure Chinese naming cognitive enhancement assessment process, quantifies the user's self-judgment accuracy through expected calibration error, Brill score, and metacognitive quotient, and generates adaptive training content through five types of standardized Chinese cognitive training tasks. This overcomes the shortcomings of existing digital avatars that only provide conversational companionship and cannot improve users' metacognitive abilities. The training indicators are written back to hierarchical storage and character hypotheses, integrating user cognitive performance into personalized strategy generation. The system can dynamically adjust the task difficulty according to the user's training level, achieving the integration of daily interaction and cognitive improvement; it can continuously improve users' overconfidence and judgment bias, simultaneously realizing the dual value of companionship and cognitive ability enhancement in educational and elderly cognitive care scenarios.

[0116] In one embodiment of this application, a metacognitive digital avatar intelligent agent method for individual users further includes: The cognitive paradigm library is used to assess and train users’ metacognitive abilities. The cognitive paradigm library includes at least one of the following: working memory training sub-module, attention conflict and inhibition control training sub-module, and active recall and confidence calibration training sub-module. Each trial records the stimulus, user response, reaction time, and confidence score, and the trial performance metrics are written back to the auditable person hypothesis entries via the paradigm-metacognitive coupler.

[0117] In this embodiment, the cognitive paradigm library is a resource library that stores various standardized cognitive assessment and training task templates. It serves as a material carrier for conducting metacognitive assessments and cognitive enhancement, and is internally divided into multiple independent sub-modules according to cognitive functions.

[0118] In this embodiment, the working memory training submodule is a set of training tasks built-in to train short-term information storage and synchronous information processing capabilities, which is used to improve the cognitive deficits of users with insufficient memory capacity and easy omission of key evidence.

[0119] In this embodiment, the attention conflict and inhibition control training submodule includes color conflict, flanking interference, and selection inhibition tasks to train users' ability to filter out distracting information, inhibit hasty intuitive judgments, and alleviate overconfidence problems.

[0120] In this embodiment, the active recall and confidence calibration training submodule guides users to objectively assess their own confidence in answering questions through the spaced recall question-answering paradigm, thereby reducing the deviation between subjective confidence and objective accuracy and calibrating the accuracy of metacognitive judgment.

[0121] In this embodiment, a trial is the minimum round of experimental procedures for a single complete training / assessment stimulus presentation and user response. A training task consists of multiple trials. A stimulus is the text, image, options, or other question material presented to the user within a single trial; it is the input content for the user to make a judgment. The user response is the result of the user's answer selection, text input, or other operations in response to the current stimulus, used to determine the correctness of the answer. Reaction time represents the length of time from stimulus presentation to the user completing the response, and it reflects the user's decision-making speed and information processing difficulty.

[0122] In this embodiment, the confidence score is a subjective assessment score (0-1 range) entered by the user after completing the test, representing their judgment of the correctness of their answer. The trial performance index is a quantitative data calculated by integrating the single-trial stimulus, user response, reaction time, and confidence score, including accuracy rate, confidence bias, and average reaction time.

[0123] In this embodiment, the paradigm-metacognitive coupler is an independent data transformation relay unit that transforms the original training trial indicators into structured feature data adapted to the hypothetical items of the user, thus establishing a data link between the cognitive training module and the user feature storage module. The write-back to the auditable hypothetical items is the cognitive performance data transformed by the coupler. Through a closed-loop channel, it updates the corresponding hypothesis items describing the user's memory, inhibitory ability, and metacognitive biases, updating the item confidence and supplementing behavioral evidence.

[0124] As can be seen from the above, this embodiment refines the three training sub-modules of the cognitive paradigm library, standardizes the collection of full behavioral data (stimulus, response, reaction time, and confidence score) in a single test, and updates the character hypothesis through the paradigm-metacognitive coupler feedback index, thus establishing a data link between cognitive training and personalized modeling of the digital avatar. The three training categories—working memory, inhibitory control, and confidence calibration—cover the user's core metacognitive weaknesses, and the full-dimensional behavioral data from a single test comprehensively portrays the user's cognitive habits. The training indicators are directly synchronized to the character hypothesis, allowing the digital avatar to dynamically adjust dialogue and task reminder strategies based on the user's attention, memory, and confidence calibration abilities. This ensures that the training effect is applied to daily interaction, avoiding a disconnect between cognitive training and actual use, and improving the overall human-computer collaborative adaptation.

[0125] In one embodiment of this application, a metacognitive digital avatar intelligent agent method for individual users further includes an AI collaborative decision calibration step: When users use AI to make learning, work, or life decisions, or to receive advice, the following actions are performed: Read the user's historical metacognitive profile, which includes domain-specific overconfidence index, metacognitive quotient score, and personal knowledge boundaries; Obtain the evidence coverage and risk level of the AI's current output; Obtain the degree of memory conflict between the current AI output and the historical fact versions in the hierarchical storage; The risk score output by AI is assessed and determined based on evidence coverage, risk level, and memory conflict. Based on the user's historical overconfidence index in the relevant field from the user's historical metacognitive profile, determine the user's adoption risk score; A joint risk assessment is conducted based on the risk score output by AI and the risk score adopted by users to obtain the joint risk assessment result. Based on the joint risk assessment results, personalized verification prompts are output. Personalized verification prompts include at least one of the following: verification suggestions, assistance prompts, cooling-off period prompts, assertion reduction prompts, refusal to execute prompts, or prompts to transfer to human agent. Collect user's final decision-making behavior and user feedback after responding to personalized verification prompts. Use the joint risk assessment results, user's final decision-making behavior and user feedback as calibration results, and write them back to the metacognitive profile, strategy recommendation and memory experience base through a closed-loop write-back channel.

[0126] In this embodiment, the AI ​​collaborative decision calibration step is a dedicated risk calibration process when the digital avatar assists users in making various choices in learning, work, and life. It combines the credibility of the avatar's output with the user's own cognitive biases to assess risks, proactively pushes differentiated verification reminders, and collects user decision-making behavior to complete data closed-loop updates.

[0127] In this embodiment, the user's historical metacognitive profile is a set of long-term cognitive characteristics stored in hierarchical storage and auditable user assumption entries. It is used to characterize the user's self-judgment bias and cognitive ability ceiling in different domains. The domain-specific overconfidence index is a domain-specific quantitative indicator, calculated separately for different scenarios such as learning, finance, health, and workplace. The higher the value, the more likely the user is to ignore evidence and blindly accept conclusions in that domain.

[0128] In this embodiment, the metacognitive quotient score is a unified quantitative score obtained by combining the expected calibration error of the user's previous assessments and the Brill score, representing the user's overall objective judgment ability. The personal knowledge boundary is a tiered storage system that records labels for the user's familiar and unfamiliar knowledge domains, distinguishing between the user's areas of expertise and cognitive blind spots, where users are more prone to misjudgment.

[0129] In this embodiment, the current AI output represents the candidate responses, suggestions, and proxy actions produced by the main generation model, i.e., the decision reference content provided by the digital clone. Evidence coverage is a metric calculated by the guard model, representing the degree to which the clone's output matches credible historical evidence; the lower the score, the higher the risk of hallucinations and unfounded conclusions.

[0130] In this embodiment, the risk level is a four-level risk classification based on sensitive keywords, user identity, and scenario sensitivity, which represents the inherent security risks of the content itself.

[0131] In this embodiment, the memory conflict degree is the semantic similarity between the clone output and the opposing, low-confidence fact version in the hierarchical storage. The higher the value, the more serious the contradiction between the output and the user's past facts.

[0132] In this embodiment, the AI ​​output risk score is a single score calculated by integrating evidence coverage, risk level, and memory conflict, quantifying the factual and security risks inherent in the content output by the AI ​​clone. The user adoption risk score is calculated by combining the user's overconfidence index in their current field, representing the probability that the user subjectively believes high-risk or contradictory conclusions; the higher the overconfidence, the larger the score. The joint risk assessment integrates the AI ​​output risk and the user's own cognitive bias risk, taking into account both the defects in the content itself and the user's susceptibility to misjudgment to obtain the overall risk level.

[0133] In this embodiment, the joint risk assessment result is the comprehensive risk level / score calculated through fusion, which serves as the basis for determining the corresponding verification prompts. Personalized verification prompts are differentiated reminder texts / interaction instructions matched according to the comprehensive risk level, and the sub-types include: Verification suggestion: prompting the user to check relevant historical evidence; Assistance prompt: suggesting consulting others or professionals; Cooling-off period prompt: delaying the execution of decisions and allowing time for reflection; Reduced assertion prompt: weakening absolute statements in the user's output; Refusal to execute prompt: directly discouraging high-risk operations; Transfer to human prompt: transferring to human review and guidance.

[0134] In this embodiment, the user's final decision-making behavior after responding to personalized verification prompts is the action the user takes upon seeing the prompt, categorized into behaviors such as adopting suggestions, postponing execution, directly rejecting, and initiating human assistance. The calibration result is a uniformly encapsulated structured data packet, including joint risk assessment results, the user's final decision-making behavior, and user text / click feedback.

[0135] In this embodiment, the closed-loop write-back channel is an independent data backflow link that does not modify the weights of the main generation model, only updating the structured data in storage. The metacognitive profile is a set of hypothetical entries recording user cognitive biases, metacognitive quotients, and knowledge boundaries. Strategy recommendation is a record of historical strategy effects stored in the system, used for subsequent verification and strategy output in matching and adaptation. The memory experience base is a long-term semantic memory and fact version library in hierarchical storage, archiving the full record of this decision-making interaction.

[0136] As can be seen from the above, this embodiment, by adding a complete AI collaborative decision-making calibration process, constructs a dual-dimensional joint assessment mechanism of user metacognitive profile, AI output risk, and user adoption risk. It pushes tiered verification prompts for user AI decision-making scenarios, addressing the shortcomings of users blindly accepting AI suggestions and making erroneous decisions due to overconfidence. It synchronously reads layered stored memory data to calculate memory conflict degree, comprehensively judges risk based on the user's historical overconfidence characteristics, and outputs differentiated tiered prompts such as cooling-off period, transfer to human intervention, and refusal to execute. The user's final decision and feedback are all looped back to update the metacognitive profile, continuously optimizing the accuracy of risk assessment. This forms a dual-layer risk protection in high-risk decision-making scenarios such as learning, workplace, elderly fraud prevention, and health care, reducing the probability of user decision-making errors.

[0137] In one embodiment of this application, a metacognitive digital avatar intelligent agent method for individual users further includes a confidence calibration and evaluation submodule: The correct label y∈{0,1} and confidence level p∈[0,1] used to record the user's answer in each user response or decision round, and to calculate the expected calibration error ECE, Brier score, and metacognitive quotient MQ: ECE=Σm ·|acc(Bm)-conf(Bm)|Brier=(1 / N)Σi(p_i-y_i) 2 MQ = (1-ECE)·Resolution_factor, where Bm is the sample set divided according to the confidence interval, N is the total number of samples, and Resolution_factor represents the distinguishability of confidence between correct and incorrect trials; ECE, Brier and MQ are used to represent the user's overconfidence, underestimation of self and metacognitive discrimination ability, and serve as inputs for character hypothesis items and adaptive training difficulty.

[0138] In this embodiment, the confidence calibration assessment submodule is an independent quantitative calculation unit in the metacognitive digital avatar. It is specifically designed to quantify the accuracy of the user's self-judgment by using the subjective confidence level and objective correctness of the user's answers when answering questions and making decisions, and output standardized metacognitive indicators.

[0139] In this embodiment, the answering or decision-making round is a complete interactive process in which a user completes a single cognitive assessment question or a single life / learning decision.

[0140] In this embodiment, the correct label y∈{0,1} represents a binary objective label, where y=1 indicates that the user's answer / decision result is correct, and y=0 indicates that the result is incorrect. The confidence score p∈[0,1] represents the subjective confidence score actively filled in by the user, ranging from 0 to 1. The higher the value, the higher the user's self-perception of the accuracy of this judgment.

[0141] In this embodiment, the expected calibration error (ECE) is an indicator that measures the overall deviation between the user's subjective confidence and the objective accuracy. The larger the value, the more serious the problem of overconfidence or underestimation by the user. The Brier score is a loss-based evaluation indicator that calculates the mean squared error between the confidence level and the true correctness. The lower the score, the closer the subjective confidence is to the objective result, and the more accurate the metacognitive judgment. The metacognitive quotient (MQ) is a comprehensive score obtained by combining the calibration error and confidence discrimination ability, and uniformly quantifies the user's overall metacognitive level.

[0142] In this embodiment, the confidence interval partitioning sample set Bm involves grouping all samples into segments with confidence levels of 0-0.1, 0.1-0.2, etc., with each group constituting set Bm. The total number of samples N is the total number of all answer and decision trials within the current statistical period. acc(Bm) is the average objective accuracy rate of the samples within group Bm. conf(Bm) is the average subjective confidence level of the samples within group Bm. The resolution_factor is a coefficient characterizing the user's ability to distinguish between correct and incorrect answers based on confidence level; this coefficient is larger when the confidence level of correct trials is significantly greater than that of incorrect trials.

[0143] In this embodiment, overconfidence is a cognitive bias where a user's subjective confidence level is consistently higher than their actual accuracy rate. Underestimation of self is a cognitive bias where a user's subjective confidence level is consistently lower than their actual accuracy rate. The input for the character hypothesis entries involves writing the expected calibration error (ECE), Brier score, and metacognitive quotient (MQ) into auditable character hypothesis entries that represent the user's cognitive bias, updating the entry confidence level and cognitive feature fields. The adaptive input for training difficulty uses these three indicators as criteria for adjusting the interference intensity, question complexity, and response time of the cognitive training task.

[0144] As can be seen from the above, this embodiment objectively calculates three standardized indicators—Expected Calibration Error (ECE), Brier score, and Metacognitive Quotient (MQ)—using a unified quantitative formula. This overcomes the limitations of subjective human judgment regarding user cognitive biases and can accurately distinguish between two typical metacognitive defects: overconfidence and underestimation of self. The indicators simultaneously affect user feature modeling and training difficulty adjustment, enabling dynamic updates to user cognitive profiles and adaptive matching of cognitive training tasks. The calculation logic is transparent and traceable, and all indicators can be stored in hierarchical storage for auditing and backtracking. This provides stable and quantitative user cognitive dimension data support for AI collaborative decision-making risk assessment and personalized verification of guard models, reducing the risk of users accepting erroneous AI suggestions due to their own judgment biases.

[0145] In one embodiment of this application, a metacognitive digital avatar intelligent agent method for individual users further includes: The three-layer architecture, consisting of a kernel layer, an evaluation and training layer, and a digital avatar application layer, forms a personal metacognitive profile based on 6+1 dimensional metacognitive indicators, ECE, Brier, and MQ, and jointly assesses the user's historical overconfidence risk and AI output risk in AI collaboration scenarios.

[0146] In this embodiment, a three-layer architecture consisting of a kernel layer, an assessment and training layer, and a digital avatar application layer is implemented. The kernel layer includes independent metacognitive guards, evidence verification, conflict detection, risk identification, and mandatory gating. The assessment and training layer includes meta-quotient (MQ) assessment, cognitive paradigm training, trial-level instant feedback, personal metacognitive profiles, long-term growth curves, and a paradigm-metacognitive coupler. Based on 6+1 dimensional metacognitive indicators, ECE, Brier, and MQ, a personal metacognitive profile is formed, and trial-level performance is written back to the user hypothesis, metacognitive profile, and strategy effect record through the paradigm-metacognitive coupler. The digital avatar application layer includes personal long-term memory, personal knowledge boundaries, life task proxy, learning and training suggestions, health care reminders, AI collaborative decision calibration, multi-terminal output execution, and authorization and audit management. The three layers are connected through a closed-loop write-back channel, enabling user behavior results in the application layer to be written back to the assessment and training layer and the kernel layer. The metacognitive profile in the assessment and training layer can serve as input for kernel layer gating and personalized output in the application layer.

[0147] In this embodiment, the 6+1 metacognitive index consists of 6 standardized metacognitive features output by the guardian model: evidence coverage (EC), emotion alignment (EA), overconfidence (OD), risk intensity (RS), memory conflict (MC), and cognitive gain (CB); and an additional dimension is the comprehensive score (Q). Together, they are called the 6+1 metacognitive index, which is used to represent the quality and credibility of the AI ​​output content.

[0148] In this embodiment, the personal metacognitive profile is a structured set of user cognitive features stored in the auditable person hypothesis entries. It is constructed by 6+1 dimensional metacognitive indicators, ECE, Brill score, and metacognitive quotient, and includes domain-specific overconfidence index, personal knowledge boundary, and long-term cognitive bias records.

[0149] In this embodiment, the AI ​​collaboration scenario is an interactive scenario in which users use their digital avatars to conduct learning planning, career plan development, life affairs decision-making, health consultation, and other activities that require AI to provide suggestions to assist in decision-making.

[0150] In this embodiment, the risk of user historical overconfidence is determined by the domain-specific overconfidence index within the personal metacognitive profile, representing the potential subjective cognitive risk of users blindly accepting AI conclusions in the current business domain.

[0151] In this embodiment, the AI ​​output risk is calculated by fusing 6+1 dimensional metacognitive indicators to obtain an AI output risk score, which represents the factual conflicts, lack of evidence, and security risks inherent in the content generated by the clone.

[0152] In this embodiment, the joint assessment simultaneously integrates the risks inherent in the AI ​​output and the risks of user cognitive bias to calculate the overall risk, outputting a joint risk assessment result and matching corresponding personalized verification prompts.

[0153] As can be seen from the above, this embodiment adopts a three-layered architecture of kernel layer, evaluation and training layer, and digital avatar application layer. This achieves decoupled layered deployment of underlying storage verification, user cognitive assessment, and front-end interactive business, allowing each module to iterate independently without interference, thus reducing system maintenance and functional expansion costs. By constructing a standardized personal metacognitive profile using multi-dimensional quantitative data such as 6+1 dimensional metacognitive indicators, expected calibration error, Brill score, and metacognitive quotient, it can completely retain users' long-term cognitive biases, judgment calibration abilities, and other characteristics, with auditable and traceable data. In AI collaborative decision-making scenarios, it simultaneously combines the user's historical overconfidence risk with the inherent risk of AI output for joint assessment. This generates dual hidden dangers of content facts, security flaws, and the user's own cognitive shortcomings, accurately identifying high-risk decision-making scenarios and matching differentiated verification prompts. This constructs a human-machine two-way metacognitive control closed loop, reducing various risks caused by users blindly accepting erroneous advice and improving the security and reliability of digital avatar-assisted decision-making.

[0154] Corresponding to the metacognitive digital avatar intelligent agent method for individual users in the above embodiment, Figure 2 This is a structural block diagram of a metacognitive digital avatar intelligent agent system for individual users, provided as an embodiment of this application. For ease of explanation, only the parts relevant to the embodiment of this application are shown. References Figure 2 The metacognitive digital avatar intelligent agent system 20 for individual users includes: a personal interaction module 21, a hierarchical storage module 22, a person hypothesis module 23, a candidate output module 24, a guard evaluation module 25, a gate control decision module 26, and a closed-loop write-back module 27.

[0155] Among them, the personal interaction module 21 is used to collect multimodal personal interaction data with user authorization and generate unified structured interaction events with timestamps; The hierarchical storage module 22 is used to store unified structured interactive events in layers according to the event log layer, short-term context memory layer, long-term semantic memory layer and fact version library layer, and output the retrieval results of the hierarchical storage. The Person Assumption Module 23 is used to generate auditable person assumption entries based on the hierarchically stored content. The auditable person assumption entries include the assumption statement, source of evidence, confidence level, last update time, counterexample record, applicable scenario, authorization status, and rebuttal flag fields; among which, the confidence level is updated using a Bayesian update mechanism. The candidate output module 24 is used to generate candidate output results for digital clones based on the preset master generation model, through the hierarchically stored retrieval results and auditable person hypothesis entries; The guard evaluation module 25 is used to evaluate the candidate output results based on the guard model, which is independent of the preset master generation model, and to calculate the comprehensive score. The gating decision module 26 is used to make gating decisions based on the comprehensive score and the risk level of the candidate output results. If the comprehensive score is greater than or equal to the first threshold and the risk level of the candidate output results is less than the second threshold, the candidate output results are allowed to be output; otherwise, the restricted policy is triggered to obtain the evaluation result of the guard model. The closed-loop write-back module 27 is used to obtain user feedback and strategy adoption results. It writes back the evaluation results of the guard model, user feedback and strategy adoption results to the hierarchical storage and auditable person hypothesis entries through the closed-loop write-back channel, and generates strategy effect records for updating the later personalized strategy selection. The closed-loop write-back does not modify the parameter weights of the main generation model.

[0156] See Figure 3 , Figure 3 This is a schematic block diagram of an electronic device provided according to an embodiment of this application. Figure 3 The electronic device 300 in this embodiment may include one or more processors 301, one or more input devices 302, one or more output devices 303, and one or more memories 304. The processors 301, input devices 302, output devices 303, and memories 304 communicate with each other via a communication bus 305. The memories 304 store computer programs, including program instructions. The processors 301 execute the program instructions stored in the memories 304. Specifically, the processors 301 are configured to invoke the program instructions to perform the functions of the modules in the aforementioned device embodiments, for example... Figure 2 The functions of the personal interaction module 21, hierarchical storage module 22, character hypothesis module 23, candidate output module 24, guard evaluation module 25, gate control decision module 26, and closed-loop write-back module 27 are shown.

[0157] It should be understood that, in the embodiments of this application, the processor 301 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0158] Input device 302 may include a touchpad, a fingerprint sensor (for collecting the user's fingerprint information and fingerprint orientation information), a microphone, etc., and output device 303 may include a display (LCD, etc.), a speaker, etc.

[0159] The memory 304 may include read-only memory and random access memory, and provides instructions and data to the processor 301. A portion of the memory 304 may also include non-volatile random access memory. For example, the memory 304 may also store device type information.

[0160] In specific implementations, the processor 301, input device 302, and output device 303 described in the embodiments of this application can execute the implementation methods described in any embodiment of the metacognitive digital clone intelligent agent method for individual users provided in the embodiments of this application, or they can execute the implementation methods of the electronic devices described in the embodiments of this application, which will not be repeated here.

[0161] In another embodiment of this application, a computer-readable storage medium is provided. This computer-readable storage medium stores a computer program, which includes program instructions. When executed by a processor, the program instructions implement all or part of the processes in the methods described above. Alternatively, the computer program can instruct related hardware to complete the process. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include any entity or device capable of carrying computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0162] The computer-readable storage medium can be an internal storage unit of the electronic device in any of the foregoing embodiments, such as a hard disk or memory of the electronic device. The computer-readable storage medium can also be an external storage device of the electronic device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., provided on the electronic device. Furthermore, the computer-readable storage medium can include both internal and external storage units of the electronic device. The computer-readable storage medium is used to store computer programs and other programs and data required by the electronic device. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0163] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0164] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the electronic devices and units described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0165] In the several embodiments provided in this application, it should be understood that the disclosed electronic devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces or units, or it may be an electrical, mechanical, or other form of connection.

[0166] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of this application, depending on actual needs.

[0167] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0168] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A metacognitive digital avatar intelligent agent method for individual users, characterized in that, include: Collect multimodal personal interaction data with user authorization and generate unified structured interaction events with timestamps; The unified structured interactive events are stored in layers according to the event log layer, short-term context memory layer, long-term semantic memory layer and fact version library layer, and the retrieval results of the layered storage are output. Based on the content of the hierarchical storage, auditable person hypothesis entries are generated. Each auditable person hypothesis entry includes a hypothesis statement, source of evidence, confidence level, last update time, counterexample record, applicable scenario, authorization status, and rebuttal flag field. The confidence level is updated using a Bayesian update mechanism. Based on the preset master generation model, candidate output results for digital clones are generated by combining the hierarchically stored retrieval results with the auditable person hypothesis entries; The guard model, which is independent of the preset master generation model, performs multidimensional metacognitive feature evaluation on the candidate output results and calculates a comprehensive score. Gating decisions are made based on the comprehensive score and the risk level of the candidate output results. If the comprehensive score is greater than or equal to the first threshold and the risk level of the candidate output results is less than the second threshold, the candidate output results are allowed to be output; otherwise, a restricted policy is triggered to obtain the evaluation result of the guard model. Obtain user feedback and strategy adoption results, and write back the evaluation results of the guard model, user feedback, and strategy adoption results to the hierarchical storage and the auditable person hypothesis entries through a closed-loop write-back channel, and generate strategy effect records for updating later personalized strategy selection; wherein the closed-loop write-back does not modify the parameter weights of the main generation model.

2. The metacognitive digital avatar intelligent agent method for individual users according to claim 1, characterized in that, The unified structured interaction events are stored hierarchically according to an event log layer, a short-term context memory layer, a long-term semantic memory layer, and a fact version database layer. The resulting retrieval results are output as follows: In the fact version library layer, when faced with multiple versions of the same fact, the arbitration score for each version is calculated; The version with the highest arbitration score is designated as the active version, while the remaining versions are retained as historical versions for audit backtracking.

3. The metacognitive digital avatar intelligent agent method for individual users according to claim 1, characterized in that, The evaluation of the candidate output results using a guard model independent of the preset master generator model includes: The multidimensional cognitive features used by the guard model to evaluate include at least evidence coverage, memory conflict, emotion fit, overconfidence index, risk level, and cognitive benefit. Specifically, the candidate output results and the evidence marked as credible in the hierarchical storage retrieval results are mapped to the same semantic vector space, and the evidence coverage is calculated by taking the maximum semantic similarity or the average of the top k similarities. The candidate output results and the fact versions marked as refutation, low confidence, and semantic opposition in the hierarchical storage retrieval results are mapped to the same semantic vector space, and the memory conflict degree is calculated by taking the maximum semantic similarity. The user's historical dominant emotion vector and candidate output emotion tendency vector are obtained and mapped to a unified emotion label space. The emotion matching degree is calculated by cosine similarity. Obtain the assertion strength and corresponding evidence confidence from the candidate output results, calculate the difference or ratio, and obtain the overconfidence index; Risk keywords, user status, and scenario sensitivity are obtained to jointly determine the risk level and classify it into a first risk level, a second risk level, a third risk level, and a fourth risk level; wherein the first risk level is lower than the second risk level, the second risk level is lower than the third risk level, and the third risk level is lower than the fourth risk level. The expected gains of candidate outputs on users' learning goals, planning goals, and emotion regulation goals are estimated to obtain cognitive gains.

4. The metacognitive digital avatar intelligent agent method for individual users according to claim 3, characterized in that, The calculation of the comprehensive score includes: A linear weighted fusion method was used to fuse and calculate six multidimensional metacognitive features, namely evidence coverage, memory conflict, emotion fit, overconfidence index, risk level, and cognitive benefit, to obtain a comprehensive score. Five sets of non-negative weighted coefficients are configured, corresponding to the evidence coverage, overconfidence correction item, emotion matching degree, risk correction item, and cognitive benefit, respectively, and the sum of all weighted coefficients is 1; at the same time, an independent conflict penalty coefficient is configured to deduct and correct the memory conflict degree. During the fusion calculation process, the overconfidence index and risk level are reversed to reduce the overall score of highly overconfident and high-risk content; the memory conflict degree is deducted by adding the conflict penalty coefficient to weaken the overall score of candidate outputs with factual conflicts. The weighting coefficients are calibrated and optimized using an offline labeled dataset that includes feature vectors and pass / reject binary labels. During operation, the weighting coefficients can be incrementally calibrated online based on the historical gating decision records of the guard model and the final user feedback.

5. The metacognitive digital avatar intelligent agent method for individual users according to claim 1, characterized in that, It also includes metacognitive assessment and cognitive enhancement steps: Collect correct marks and real-time confidence scores of users' cognitive task responses, and calculate expected calibration error, Brill score, and metacognitive quotient; Adaptive cognitive training tasks are generated based on retrospective memory training tasks, color conflict suppression tasks, selection suppression tasks, flanking interference suppression tasks, and spaced repetition active recall tasks. The task trial metrics and metacognitive quotient assessment results are written back to hierarchical storage and auditable character hypothesis entries, and the difficulty parameters of the cognitive paradigm are adjusted based on the user's training performance.

6. The metacognitive digital avatar intelligent agent method for individual users according to claim 1, characterized in that, Also includes: The user's metacognitive ability is assessed and trained through a cognitive paradigm library, which includes at least one of the following: working memory training submodule, attention conflict and inhibition control training submodule, and active recall and confidence calibration training submodule. Each trial records the stimulus, user response, reaction time, and confidence score, and the trial performance metrics are written back to the auditable person hypothesis entries via a paradigm-metacognitive coupler.

7. The metacognitive digital avatar intelligent agent method for individual users according to claim 1, characterized in that, It also includes AI collaborative decision calibration steps: When users use AI to make learning, work, or life decisions, or to receive advice, the following actions are performed: Read the user's historical metacognitive profile, which includes the domain-specific overconfidence index, metacognitive quotient score, and personal knowledge boundary; Obtain the evidence coverage and risk level of the AI's current output; Obtain the degree of memory conflict between the current AI output and the historical fact versions in the hierarchical storage; The AI ​​output risk score is assessed and determined based on the evidence coverage, the risk level, and the memory conflict degree. Based on the user's historical overconfidence index in the corresponding field in the user's historical metacognitive profile, the user's adoption risk score is determined; A joint risk assessment is performed based on the AI-output risk score and the user adoption risk score to obtain the joint risk assessment result. Based on the joint risk assessment results, personalized verification prompts are output, including at least one of the following: verification suggestions, assistance prompts, cooling-off period prompts, assertion reduction prompts, refusal to execute prompts, or prompts to transfer to human verification. Collect the user's final decision-making behavior and user feedback after responding to the personalized verification prompts, and use the joint risk assessment results, the user's final decision-making behavior, and the user feedback as calibration results, and write them back to the metacognitive profile, strategy recommendation, and memory experience base through the closed-loop write-back channel.

8. A metacognitive digital avatar intelligent agent system for individual users, characterized in that, include: The personal interaction module is used to collect multimodal personal interaction data with user authorization and generate unified structured interaction events with timestamps. The hierarchical storage module is used to store the unified structured interactive events in layers according to the event log layer, short-term context memory layer, long-term semantic memory layer and fact version library layer, and output the retrieval results of the hierarchical storage. The Person Assumption Module is used to generate auditable person assumption entries based on the hierarchically stored content. The auditable person assumption entries include the assumption statement, source of evidence, confidence level, last update time, counterexample record, applicable scenario, authorization status, and rebuttal flag fields; wherein the confidence level is updated using a Bayesian update mechanism. The candidate output module is used to generate candidate output results for digital clones based on the preset master generation model, through the retrieval results stored in the hierarchical storage and the auditable person hypothesis entries; The guard evaluation module is used to evaluate the candidate output results based on a guard model that is independent of the preset master generation model and to calculate a comprehensive score. The gating decision module is used to make gating decisions based on the comprehensive score and the risk level of the candidate output result. If the comprehensive score is greater than or equal to a first threshold and the risk level of the candidate output result is less than a second threshold, the candidate output result is allowed to be output; otherwise, a restricted policy is triggered to obtain the evaluation result of the guard model. The closed-loop write-back module is used to obtain user feedback and strategy adoption results, and write back the evaluation results of the guard model, user feedback and strategy adoption results to the hierarchical storage and the auditable person hypothesis entries through the closed-loop write-back channel, and generate strategy effect records for updating later personalized strategy selection; wherein the closed-loop write-back does not modify the parameter weights of the main generation model.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method 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 steps of the method as described in any one of claims 1 to 7.

Images