A method and system for evaluating and training metacognition of a meta-merchant

Abstract

The application provides a meta-commerce MQ evaluation and meta-cognition training method and system, belonging to the technical field of meta-cognition evaluation and training. The method comprises the following steps: obtaining interaction data with a user terminal; preprocessing and test secondary feature extraction are performed on the interaction data to obtain test secondary intermediate features, and a 6+1-dimensional meta-commerce index is calculated; the 6+1-dimensional meta-commerce index is weighted and fused to obtain a meta-commerce total score, and the user is classified to obtain a user meta-cognition portrait; according to the user meta-cognition portrait, a personalized training game combination is matched from a preset training game library to obtain a personalized training game; based on the personalized training game, answer behavior data of the user is obtained; according to the answer behavior data, a feedback signal is determined, and the meta-cognition sub-ability corresponding to the user meta-cognition portrait is coupled in a test secondary manner to obtain a test secondary feedback result; according to the test secondary feedback result, the 6+1-dimensional meta-commerce index and the user meta-cognition portrait are updated, and a user level result and achievement data are determined.

Description

Technical Field

[0001] This application relates to the field of metacognitive assessment and training technology, and in particular to a meta-quotient MQ assessment and metacognitive training method and system. Background Technology

[0002] Metacognition is an individual's ability to self-monitor and regulate cognitive behavior, and it is widely used in learning and training, human-computer collaboration, and decision-making. Existing metacognitive assessment systems are often one-dimensional, focusing only on single dimensions such as confidence calibration and memory, lacking standardized quantitative indicators that encompass all dimensions of metacognitive monitoring and control. This makes it impossible to form a unified comprehensive quantitative score for metacognition and to accurately characterize individual user metacognitive weaknesses. Furthermore, traditional metacognitive assessments often employ coarse-grained, overall statistical methods, failing to capture instantaneous metacognitive fluctuations in a single task trial, resulting in low assessment accuracy. Existing metacognitive training models are mostly general-purpose cognitive training, failing to achieve personalized adaptation based on user metacognitive deficit profiles. This results in poor training targeting, limited corrective effects, and training feedback that simply judges the correctness of answers, failing to couple with sub-level metacognitive ability weaknesses and thus unable to achieve root-cause correction. In addition, existing technologies are only based on single-person cognitive training scenarios and are not adapted to decision-making biases in AI-human-computer collaboration scenarios, unable to dynamically calibrate users' undesirable decision-making behaviors such as blindly following AI or overconfidence. Existing products such as CogniFit and Lumosity only train basic attention and memory, lacking confidence calibration and error detection metacognitive abilities for the AI ​​era, and their training transfer effects are weak. Summary of the Invention

[0003] To address the aforementioned technical issues, this application provides a meta-quotient MQ assessment and metacognitive training method and system.

[0004] A first aspect of this application provides a method for meta-quotient MQ assessment and metacognitive training, including: In response to evaluation commands, acquire interaction data with the user terminal; The interactive data is preprocessed and trial-level features are extracted to obtain trial-level intermediate features; Based on the intermediate features of the trial, a 6+1 dimensional meta-quotient index is calculated; the 6+1 dimensional meta-quotient index includes metacognitive sensitivity, confidence calibration, uncertainty recognition, error detection, strategy control, meta-memory, and meta-time perception; the 6+1 dimensional meta-quotient index is weighted and fused according to preset non-negative weights to obtain the total meta-quotient score; Based on the total meta-common score, users are categorized to obtain a user metacognitive profile; Based on the user metacognitive profile, a personalized training game combination is matched from a preset training game library to obtain a personalized training game; based on the personalized training game, the user's answer behavior data is obtained. Based on the response behavior data, a feedback signal is determined, and the feedback signal is coupled with the metacognitive sub-capabilities corresponding to the user metacognitive profile at the trial level to obtain the trial level feedback result; based on the trial level feedback result, the 6+1 dimensional meta-quotient index is updated; and based on the updated 6+1 dimensional meta-quotient index, the user metacognitive profile is updated, the user level result and achievement data are determined. Based on the updated user metacognitive profile, the user's decision-making in the AI ​​collaboration process is calibrated to obtain calibration results and user responses; The interaction data, the intermediate features of the trial, the 6+1 dimensional meta-quotient index, the trial feedback results and the updated user metacognitive profile, the calibration results and user responses, and the achievement data are stored in the database respectively.

[0005] A second aspect of this application provides a meta-quotient MQ assessment and metacognitive training system, comprising: The interactive data acquisition module is used to respond to evaluation commands and acquire interactive data with the user terminal. The feature extraction module is used to preprocess the interactive data and extract trial-level features to obtain trial-level intermediate features; The assessment calculation module is used to calculate a 6+1 dimensional meta-quotient index based on the intermediate features of the trial. The 6+1 dimensional meta-quotient index includes metacognitive sensitivity, confidence calibration, uncertainty recognition, error detection, strategy control, meta-memory, and meta-time perception. The 6+1 dimensional meta-quotient index is weighted and fused according to preset non-negative weights to obtain the total meta-quotient score. The user profile classification module is used to classify users based on their total meta-quotient score to obtain a user metacognitive profile. The game recommendation module is used to match personalized training game combinations from a preset training game library based on the user metacognitive profile to obtain personalized training games; and to obtain user response behavior data based on the personalized training games. The feedback coupling module is used to determine a feedback signal based on the response behavior data, and to perform trial-level coupling between the feedback signal and the metacognitive sub-capabilities corresponding to the user metacognitive profile to obtain trial-level feedback results; to update the 6+1 dimensional meta-quotient index based on the trial-level feedback results; and to update the user metacognitive profile, determine the user level results and achievement data based on the updated 6+1 dimensional meta-quotient index. The decision calibration module is used to calibrate the user's decisions during the AI ​​collaboration process based on the updated user metacognitive profile, and obtain calibration results and user responses. The data storage module is used to store the interaction data, the intermediate features of the trial and the 6+1 dimensional meta-quotient index, the trial feedback results and the updated user metacognitive profile, the calibration results and user responses, and the achievement data into the database respectively.

[0006] 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 meta-quotient MQ assessment and metacognitive training method.

[0007] 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 meta-quotient MQ assessment and metacognitive training method.

[0008] The beneficial effects of the metacognitive ability assessment and metacognitive training method and system provided in this application are as follows: This application improves the comprehensiveness and objectivity of the quantitative assessment of metacognitive ability by collecting interactive data and weighted fusion of 6+1 dimensional metacognitive indicators. It matches exclusive training games according to the user's metacognitive profile and couples the feedback signals with the metacognitive sub-ability trial sub-levels to complete dynamic iterative training. Simultaneously, it actively outputs decision calibration intervention in AI collaboration scenarios based on the updated user metacognitive profile. Moreover, the interactive data, trial sub-level intermediate features, 6+1 dimensional metacognitive indicators, trial sub-level feedback results, updated user metacognitive profiles, calibration results and user responses, and achievement data are uniformly stored in the system. This solves the defects of existing metacognitive tools, such as single assessment dimensions, lack of personalized training adaptation, inability to quantify and iterate training effects in real time, lack of metacognitive risk intervention in AI collaboration scenarios, and inability to completely trace and retain behavioral data. This application realizes the integrated and automated operation of metacognitive assessment, training, risk intervention, and data auditing, and continuously iterates and optimizes various sub-abilities of user metacognition, thereby reducing the risk of decision bias caused by users blindly trusting AI output from the root. Attached Figure Description

[0009] Figure 1 A flowchart illustrating the meta-quotient MQ assessment and metacognitive training method provided in an embodiment of this application; Figure 2 This is a structural block diagram of a meta-quotient MQ assessment and metacognitive training system 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

[0010] 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.

[0011] 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.

[0012] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating a meta-quotient MQ assessment and metacognitive training method provided in an embodiment of this application. The method includes: S101: In response to the evaluation command, acquire the interaction data with the user terminal.

[0013] In this embodiment, the assessment instruction represents the operation signal that triggers the Yuanshang MQ assessment process. It refers to the instruction triggered by the user actively clicking the assessment entry on the user terminal or by the teacher / parent issuing the assessment task in the backend. The user terminal refers to the electronic device that carries the assessment interaction interface. It refers to any front-end carrier that can communicate with the platform, such as the student's WeChat mini program, the parent's WeChat mini program, the teacher's WeChat mini program, the elderly's WeChat mini program, a computer web browser, or a mobile app. The interaction data is used to record the original behavioral information generated by the user in the entire process of completing the objective behavioral carrier task. It refers to the complete set of data generated by the user in the process of answering the objective behavioral task, including the first correct answer, the first confidence score, the first reaction time, the help-seeking behavior, the answer modification mark, and the first error prediction intensity.

[0014] S102: Preprocess the interactive data and extract trial-level features to obtain trial-level intermediate features.

[0015] In this embodiment, preprocessing refers to the pre-processing operations of cleaning, correcting, and standardizing the interactive data to eliminate the interference of invalid, abnormal, and missing content in the original data on subsequent calculations; trial level refers to the data granularity with a single assessment task as the smallest processing unit, which refers to the independent data slice corresponding to a single answer to a question. For example, all the data generated by a complete answer to a logic reasoning question is a trial unit; trial level intermediate features refer to the standardized derivative index set generated after preprocessing and single-trial dimension calculation. These are intermediate quantitative results that support the calculation of the 6+1 dimensional meta-quotient index, such as: single-trial confidence bias, answering efficiency, error prediction matching degree, help trigger marker, strategy adjustment range, and other quantitative values.

[0016] For example, the interaction data is read and the operation is performed in two main stages. Specifically, the first stage performs preprocessing. First, all trials are traversed to verify the validity of each data entry. Missing fields (such as no confidence score, no reaction time) are marked as missing and filled with neutral default values. Abnormal values ​​are truncated, and extreme time-consuming data exceeding the upper and lower thresholds of reaction time are corrected to interval boundary values. The upper and lower thresholds of reaction time and the threshold for judging missing fields are both statistically determined based on user evaluation samples. Specifically, the first reaction time of normal trials is collected and fitted with a normal distribution, and the mean plus or minus three times the standard deviation is taken as the upper and lower thresholds of reaction time. The threshold for judging missing fields is the critical value of the number of missing key fields in a single trial of interaction data. If the number of missing fields in a single trial reaches the threshold, it is judged as an invalid trial and neutral default values ​​are filled in. Boolean actions such as asking for help and modifying answers are uniformly converted to 0 / 1 values. After cleaning, the interactive dataset is obtained. In the second stage, trial-level feature extraction is performed. The interactive data is traversed row by row with each trial as the loop unit. For each group of trial raw fields, derived quantitative indicators are calculated in batches, including the difference between confidence and accuracy in a single trial, the offset of reaction time for a single question relative to the overall mean, the difference between the strength of error prediction and the actual correct or incorrect match, the trigger weight of asking for help, and the accuracy gain brought by modifying answers. After all trials are calculated, all derived indicators corresponding to each question are associated and packaged by user, session, and trial ID to generate structured trial-level intermediate features that can be directly used for meta-quotient indicator calculation.

[0017] S103: Calculate the 6+1 dimensional meta-quotient index based on the intermediate features of the trial. The 6+1 dimensional meta-quotient index includes metacognitive sensitivity, confidence calibration, uncertainty recognition, error detection, strategy control, meta-memory, and meta-time perception. The 6+1 dimensional meta-quotient index is weighted and fused according to the preset non-negative weights to obtain the total meta-quotient score.

[0018] In this embodiment, the 6+1 dimensional meta-quotient index refers to a standardized scoring dimension system of seven metacognitive aspects for users, used to represent the cognitive level of users' self-monitoring and self-regulation. The 6+1 dimensional meta-quotient index includes metacognitive sensitivity, confidence calibration, uncertainty recognition, error detection, strategy control, meta-memory, and meta-time perception. Among them, metacognitive sensitivity represents the user's basic ability to distinguish between the correctness of their own answers and match their confidence levels, referring to the correlation mapping score between the confidence score sequence and the correctness sequence of answers; confidence calibration represents the degree of fit between the user's subjective confidence and objective answer accuracy, used to judge whether there is an overconfidence or over-inferiority bias; uncertainty recognition represents the user's behavioral capture dimension of identifying their own knowledge blind spots and actively choosing to seek help / postpone decisions; error detection represents the user's predictive ability to anticipate their own incorrect answers and identify answer loopholes; strategy control represents the user's ability to flexibly switch between fast thinking, slow thinking, and help-seeking strategies; meta-memory represents the user's predictive matching dimension of their own memory capacity and memory accuracy; and meta-time perception represents the matching dimension between the user's estimated answering time and the actual reaction time. The preset non-negative weights represent the contribution coefficients assigned to the seven meta-quotient indicators. These weights are constrained to be greater than or equal to 0, and the sum of all weights equals 1. This is used to distinguish the importance of each dimension to the overall ability. The total meta-quotient score represents the user's comprehensive metacognitive quantitative score obtained after weighted fusion of the 6+1 dimension meta-quotient indicators.

[0019] Specifically, the calculation method of the six-dimensional meta-quotient index includes: before calculation, the system prunes or truncates the reaction time (RT) according to the preset upper and lower quantiles or preset maximum and minimum thresholds, and marks or removes missing values, invalid inputs and data that exceed the defined domain; the metacognitive sensitivity is obtained by linearly mapping the correlation coefficient between the user's normalized confidence score p and the correctness of the answer y∈{0,1} to the interval [0,100], that is, sensitivity=clip((corr(p,y)+1) / 2×100,0,100), where corr(p,y) is the Pearson correlation coefficient, which is used to represent the degree of linear correlation between the two sets of sequences, and the value range is [-1,1]. `clip(a,0,100)` is the numerical clipping function `clip`, where `a = (corr(p,y) + 1) / 2 × 100` is the metacognitive sensitivity score. `clip(a,0,100)` restricts the metacognitive sensitivity score to 0-100, taking the boundary value directly for scores exceeding the limit. Specifically, when `p` or `y` is a constant sequence or `corr(p,y)` is undefined, `corr(p,y)` is set to 0 or `sensitivity` is set to a preset neutral value. The confidence calibration is calculated by subtracting the absolute value of the difference between the user's first average confidence score mean(p) and the first average accuracy mean(y) from 1, and then linearly mapping it to the interval [0,100]. That is, calibration = clip((1-|mean(p)-mean(y)|)×100,0,100). It can be further given by the expected calibration error ECE: ECE = Σ_m(|B_m| / N)·|acc(B_m)-conf(B_m)|, where acc(B_m) is the average accuracy of the first sample in bin B_m, conf(B_m) is the average confidence value of the first sample in bin B_m, Σ is the summation operation, N is the total number of valid trials in the current window, and B_m is the sample set divided according to the confidence interval.Uncertainty is calculated by the proportion of users actively seeking help, choosing uncertain options, or delaying decisions in trials with low confidence or incorrect answers. The formula is: uncertainty = clip(sum((u+h)·I_hard) / sum(I_hard)×100,0,100), where u is the uncertainty choice, h is the help-seeking flag, and I_hard is the set of trials with a first confidence score less than 0.65 or incorrect answers. When sum(I_hard) = 0, uncertainty is set to zero by default or marked as a trial without difficulty and processed according to preset rules during weighted fusion. Error awareness is determined by comparing the user's error premonition intensity sequence with the actual error label sequence. The correlation coefficient and the mean of the prediction accuracy of the knowledge blind spot are linearly mapped to the interval [0,100]. When the correlation coefficient is undefined, a preset neutral value is used or only the prediction accuracy component of the knowledge blind spot is used. The strategy_control is linearly mapped to the interval [0,100] based on the accuracy gain of the trial group with slow thinking, asking for help or modifying answers and the non-slow thinking group. When any comparison group is empty, a preset neutral value is used or it is temporarily excluded from the window weighting. The metamemory is linearly mapped to the interval [0,100] based on the correlation coefficient between confidence score and recall correctness in the memory task subset. When the number of trials of the memory task is less than the preset minimum sample size, it is marked as temporarily missing and renormalized according to the weight in the total score calculation. The preset neutral value is a fixed score pre-set by the system to represent a user's metacognitive performance as moderate, without significant bias. It is set by taking the median of each dimension from a sample of normal users' six-dimensional meta-quotient indicators as the basic neutral score, and then verifying through ablation experiments that this value minimizes the deviation from the total meta-quotient score in scenarios with missing samples, ultimately obtaining the preset neutral value. The preset rule processing is a pre-defined standardized weighted processing logic for scenarios where sum(I_hard) = 0 (no difficult trials). This is used to address the problem of uncertainty identification indicators failing to participate in the meta-quotient total score fusion when there are no low-confidence or incorrect question samples. It is set by distinguishing between two branch rules: one is to temporarily reset the weight of that dimension to zero and exclude it from the current window weighted summation; the other is to use the average uncertainty identification score of the user population to replace the current empty window indicator. The preset minimum sample size is a critical threshold for determining whether a subset of memory task data has statistical validity. When the total number of memory-related response trials is less than the preset minimum sample size, it is determined that the sample is insufficient and the meta-memory indicator label is temporarily missing. The setting method uses ROC curves to compare the accuracy of the meta-memory indicator under different trial numbers based on the silhouette coefficient, and selects the minimum number of trials where the indicator discrimination tends to be stable as the preset minimum sample size.

[0020] Further, the seventh dimension, meta-time perception (time_perception), is calculated by comparing the relative error between the user's predicted or actual reaction time before or after answering the question and the actual reaction time. That is, time_perception = clip((1-min(|predicted time - actual time| / max(actual time, ε), 1.0)) × 100, 0, 100), where ε is a preset positive number to prevent division by zero. When the actual time is missing or abnormal, the meta-time perception is processed according to the reaction time clipping and missing handling rules to obtain an independent 6+1 dimension meta-quotient index.

[0021] The 6+1 dimensional meta-quotient indicators are weighted and fused according to preset non-negative weights to obtain the total meta-quotient score. Seven preset non-negative weights correspond to the metacognitive sensitivity, confidence calibration, uncertainty recognition, error detection, strategy control, meta-memory, and meta-time perception of the 6+1 dimensional complete meta-quotient indicators. Standardized assessment samples were first collected from three groups: students, working professionals, and the elderly. User-generated AI collaborative error decisions and training ability improvement were used as labels. Two parallel fitting methods, Ridge Regression and ROC-AUC maximization, were used to output initial weight sets. All weights were constrained to be greater than or equal to 0, and the sum of the seven weights was equal to 1. Subsequently, multiple ablation experiments were conducted, fixing the weights in one dimension while keeping other variables unchanged. The discriminative power of the total meta-quotient score and the accuracy of the user profile classification were compared before and after weight adjustment. Redundant weights with extremely low contribution after fitting were removed and slightly corrected. Finally, three sets of differentiated fixed weights were generated for students, working professionals, and the elderly, respectively, to obtain the non-negative weights.

[0022] In another specific embodiment, the total meta-quotient score (MQ_total) is calculated. Specifically, it is calculated from at least six dimensions of indicators according to the following weighted fusion formula and mapped to a 0-200 score scale: MQ_total = clip((w_1·sensitivity + w_2·calibration + w_3·uncertainty + w_4·error_awareness + w_5·strategy_control + w_6·metamemory) × 2, 0, 200); where w_1 to w_6 are non-negative weights and Σw_i = 1; the weights are preset to w_1, w_2, w_i ... 3, w_4, w_5, and w_6 are 0.22, 0.20, 0.18, 0.16, 0.16, and 0.08, respectively. They can also be reset based on the offline labeled data of the user group using calibration methods such as logistic regression, ridge regression, or ROC-AUC maximization. When a certain dimension is marked as temporarily missing due to insufficient sample size or outlier handling, the non-negative weights of the remaining effective dimensions are renormalized before calculating MQ_total. ×2 is the scaling factor for linearly mapping the fused values ​​of 0-100 dimensions to a 0-200 score. The meta-time perception (time_perception) is presented independently as an additional sub-item and is not included in the 200-point total score of MQ_total.

[0023] S104: Based on the total meta-quotient score, users are categorized to obtain user metacognitive profiles.

[0024] In this embodiment, "user" refers to the main user of this MQ assessment and training platform, including students, working professionals, elderly anti-fraud users, teachers, parents, and other user groups. "User metacognitive profile" refers to a standardized identity label after classifying the user's overall metacognitive ability and shortcomings, which can be any one of the following seven labels: high metacognitive type, overconfident type, uncertainty-sensitive type, weak error perception type, rigid strategy type, weak self-monitoring type, and growth type.

[0025] Specifically, the process first reads the total meta-quotient score and 6+1-dimensional meta-quotient indicators synchronously output from memory, then retrieves the pre-stored complete profile classification and judgment rules. This includes: performing multi-condition joint verification according to the priority order in the rules, prioritizing the verification of profile judgment conditions corresponding to high comprehensive capabilities, synchronously matching each meta-quotient sub-indicator with its corresponding sub-indicator threshold, and identifying whether the user has one or more weak meta-cognitive dimensions. During the verification process, the problem of multi-label conflicts is avoided. When multiple profile judgment conditions are met simultaneously, the unique matching profile type is locked according to the feature priority judgment rule. After completing the full condition traversal verification, a unique corresponding profile label is generated as the identifier of the user's metacognitive profile. At the same time, the user's current strength and weakness status of the 6+1-dimensional meta-quotient indicators and the comprehensive meta-quotient total score are bound to form a complete profile data package. The generated user metacognitive profile data package is synchronously pushed to the game recommendation module for matching personalized training games and simultaneously written to the database.

[0026] For example, a user with a total meta-quotient score greater than or equal to 160 and at least the minimum value of the six-dimensional meta-quotient indicators greater than or equal to 60 is classified as having high metacognitive ability; if the confidence-calibrated meta-quotient is less than 55 and the metacognitive sensitivity meta-quotient is greater than or equal to 55, they are classified as overconfident; if the uncertainty recognition meta-quotient is less than 45, they are classified as having low uncertainty tolerance; if the error detection meta-quotient is less than 45, they are classified as having weak error detection; if the strategy control meta-quotient is less than 45, they are classified as having rigid strategies; if the metacognitive sensitivity is less than 45 and does not meet any of the above criteria, they are classified as having weak self-monitoring; and an intermediate state that does not meet any of the above criteria but has some monitoring ability is classified as having growth potential. The classification results are used to drive personalized recommendations in the training game library module, recommending different training game combinations, difficulty strategies, feedback styles, and risk warning frequencies for different user profile categories.

[0027] S105: Based on the user's metacognitive profile, match personalized training game combinations from the preset training game library to obtain personalized training games; based on the personalized training games, obtain the user's answer behavior data.

[0028] In this embodiment, the pre-defined training game library refers to the set of all training game resources that the system stores in advance and categorizes hierarchically. It refers to a resource pool containing thirteen categories of training games, including a core layer, an AI-era extension layer, and a scene-specific layer. For example, it uniformly stores all game configurations, interaction logic, and task materials for games such as "Confidence Radar" and "Minnan Dialect Anti-Fraud Master" in memory and the database. A training game refers to an independently encapsulated gamified training unit within the training game library, corresponding to at least one metacognitive sub-capability dimension, used to specifically train a user's single or multiple metacognitive weaknesses. A training game combination refers to a set of training content formed by combining multiple different training games to address the same user's metacognitive weaknesses, used to simultaneously compensate for multiple weak dimensions in a user. A personalized training game refers to a customized game training scheme that, after selection, configuration of difficulty, feedback style, and reminder frequency, is exclusively adapted to the current user's metacognitive state. Response behavior data refers to the set of all operation records generated by the user during each trial interaction in a personalized training game. For example, records such as the second correct answer, second confidence score, combo count, and second error prediction intensity generated by the user for each question in a single round of the Confidence Radar game. The second indicator, "Correctness of Answer," represents whether the user's chosen answer in the training game matches the standard answer, distinguishing between correct and incorrect answers. The second "Confidence Score" represents the confidence score mapped after the user selects from the three button levels in the training trials, characterizing the user's subjective level of certainty. The "Success Count" represents the cumulative count of multiple consecutive correct answers by the user. The second "Indicator of Error" represents the numerical value indicating the degree to which the user pre-judged their own prediction of error in the training trials.

[0029] In this embodiment, the user's metacognitive profile type tags are obtained, which are: high metacognitive type, overconfident type, uncertainty-sensitive type, weak error perception type, rigid strategy type, weak self-monitoring type, and growth type. Based on the type tags, one of the target metacognitive sub-ability dimensions corresponding to the type tags is queried from a preset profile-game mapping rule table. The metacognitive sub-ability tags of each training game in the training game library are obtained, and the training game library is filtered according to the target metacognitive sub-ability dimensions and the metacognitive sub-ability tags of each training game to obtain games to be recommended. Based on the type tags, the initial difficulty level, feedback style preference, and risk warning frequency corresponding to the type tags are queried from a preset difficulty mapping table. Personalized training games are generated based on the games to be recommended and the initial difficulty level, feedback style preference, and risk warning frequency.

[0030] In this embodiment, the user terminal receives and renders complete personalized training game resources, loading all question materials, interactive controls, difficulty parameters, and feedback rules within the game. In each trial, the user completes interactive operations such as reading the question, selecting the answer, clicking on confidence levels, and optionally seeking help / marking errors. The terminal synchronously captures the behavioral information corresponding to each step of the operation and records behavioral fields such as the second correct answer, second confidence score, current combo count, second error prediction intensity, answer time, and strategy selection markers, obtaining a single piece of answer behavior data. If network communication is normal, it is immediately uploaded to the feedback coupling module through the real-time data feedback channel. If abnormal scenarios such as network disconnection or weak signal are detected, the current trial's answer behavior data is stored in a local temporary cache, and all data in the cache is automatically resent after the network is restored. The server continuously receives the single-trial answer behavior data uploaded by the front end.

[0031] S106: Based on the response behavior data, determine the feedback signal, and couple the feedback signal with the metacognitive sub-capabilities corresponding to the user's metacognitive profile at the trial level to obtain the trial level feedback result; based on the trial level feedback result, update the 6+1 dimensional metacognitive quotient; and based on the updated 6+1 dimensional metacognitive quotient, update the user's metacognitive profile, determine the user's level result and achievement data.

[0032] In this embodiment, the feedback signal represents an integrated incentive output carrier that combines three types of information: score value, visual performance, and sound effect performance. It is used to display a single-trial training evaluation to the user, such as bonus points, golden particle animation, and prompting sound effects. The feedback type represents seven differentiated evaluation identifiers based on the user's answer behavior data: flow feedback type, combo feedback type, perfect calibration feedback type, honesty reward feedback type, self-awareness feedback type, normal correctness feedback type, and neutral feedback type. For example, an honesty reward feedback type is matched when the user has low confidence and answers incorrectly. The metacognitive sub-ability represents the seven core cognitive regulation abilities corresponding to the 6+1 dimensional meta-quotient indicators, which are the training target dimensions that need to be specifically strengthened in this trial. Trial level represents the data granularity with a single game question as an independent calculation and evaluation unit, used to achieve independent evaluation of a single question and binding it to the ability. Trial level coupling represents the operational logic that establishes a binding association between the feedback signal generated in a single trial and the metacognitive sub-ability that needs to be strengthened in the current trial, achieving a one-to-one correspondence between feedback incentives and dimensional training objectives. Trial feedback results represent a complete structured record of the output after feedback signal generation and sub-capability coupling, including feedback type, signal parameters, bound metacognitive dimensions, and single-trial capability gain. Updates represent the iterative correction and dynamic refresh of original indicator values ​​based on the dimensional gain increment carried by the trial feedback results, reflecting capability changes brought about by a single training session. User level results represent capability stage identifiers based on cumulative training points, used to distinguish the user's overall metacognitive growth stage. Achievement data represents a complete set of records of badges unlocked, unlock snapshots, and unlock times after a user triggers specified training or assessment events, used to retain user growth milestones. For example, the user's overall metacognitive growth stages include L1 (Introduction), L2 (Awareness), L3 (Calibration), L4 (Reflection), L5 (Stability), L6 (Advanced), L7 (Optimization), L8 (Critique), L9 (Advanced), and L10 (Excellence). Badges include, but are not limited to, "Beginner's Guide," "Confidence Calibration Master," "Error Hunter," "Calm Brake," "Memory Craftsman," "Strategy Master," "AI Criterionist," and "Fraud Prevention Guardian." The unlocking conditions for each stage of a user's overall metacognitive growth (L1-L10) are as follows: Default initial level, completion of basic assessment, first trigger of calibration-type feedback, completion of review / debriefing tasks, stable performance in key dimensions, completion of multiple training games, formation of a consistent training habit, achievement of AI criterion task targets, continuous achievement of multi-dimensional indicators, and entry into high-level metacognition or reaching the comprehensive threshold.

[0033] S107: Based on the updated user metacognitive profile, calibrate the user's decisions during the AI ​​collaboration process to obtain calibration results and user responses; the calibration results include the output records of metacognitive prompts, and the user responses include the user's subsequent decision-making behavior based on the metacognitive prompts.

[0034] In this embodiment, the updated user metacognitive profile represents a standardized data package that has been refreshed through training trials and includes the latest 6+1-dimensional meta-quotient items, cognitive bias types, and weak metacognitive dimensions. This data serves as the basis for AI collaborative decision calibration. The AI ​​collaboration process represents the entire interactive scenario where the user completes tasks using AI tools. This is the trigger scenario for decision calibration, such as students using AI to complete assignments, professionals using AI to generate solutions, or elderly users answering AI-simulated fraudulent calls. Decision calibration refers to intervening and correcting the user's blind following of AI's cognitive guidance process based on the user's inherent cognitive biases when AI provides reference conclusions. This aims to reduce judgment errors caused by users readily accepting erroneous AI outputs. The calibration result represents a complete archived data set recording the entire calibration process. This data includes the output records of metacognitive prompts, basic scenario information, and risk level markers, used to trace the intervention process. Metacognitive prompts refer to the guiding text and display materials generated based on risk levels, used to remind users to carefully consider the AI's output. The output record of metacognitive prompt information represents a complete set of archived fields for the entire process of prompt delivery, used for auditing and traceability. For example, it records the prompt delivery time, scenario type, prompt intensity, text content, and risk level. User response represents the set of all actions taken by the user after receiving the metacognitive prompt information.

[0035] In this embodiment, historical user interaction data and updated user metacognitive profiles are acquired, and the user's overconfidence index in each domain is calculated according to preset domain segmentation rules. When a user is detected to be at a preset key node in collaboration with the AI ​​system, the risk level of the current collaboration scenario is determined based on the updated user metacognitive profile and overconfidence index. If the risk level is greater than or equal to a preset risk threshold, corresponding metacognitive prompts are generated based on the risk level and overconfidence index, and pushed to the corresponding user terminal for display. The metacognitive prompts are used to assist users in making judgments and decisions during AI collaboration. The output records of the metacognitive prompts and the user's decision-making behavior based on the metacognitive prompts are recorded to obtain calibration results and user responses.

[0036] S108: Store the interaction data, trial intermediate features and 6+1 dimensional meta-quotient indicators, trial feedback results and updated user metacognitive profiles, calibration results and user responses, and achievement data into the database respectively.

[0037] In this embodiment, the database serves as a structured storage medium for all process data. Internally, it is divided into multiple independent data tables for hierarchical archiving of different data types, such as `trials`, `sessions`, `ai_decision_logs`, and `achievements` tables. The storage function represents the operation of writing various data types into the corresponding database tables after format validation, field standardization, and association binding, enabling long-term data retention and cross-session traceability. The association binding of various data types is implemented using association keys, which represent unique identifier fields used to link different data tables, ensuring that multiple data types can be queried together. Examples include `user_id` (user identifier) ​​and `session_id` (session identifier).

[0038] In this embodiment, each type of data is first independently verified according to preset data verification rules. These rules include checking the integrity of key fields, the legality of numerical value ranges, and the standardization of character formats. Incomplete or abnormal data is diverted to temporary storage in the abnormal log, while qualified data enters the standardization process. Subsequently, a unified association key is bound to all qualified data, using user_id and session_id as cross-table concatenation identifiers, and matching the corresponding dedicated data table in the database according to the data type. For example, interactive data is written to the trials table, secondary intermediate features, 6+1 dimensional meta-quotient indicators, and updated user metacognitive profiles are written to the sessions table, secondary feedback results are synchronously associated with trial_id and written to the trials sub-field, calibration results and user responses are uniformly stored in the ai_decision_logs table, and achievement data is stored in the achievements table. After completion, the module performs the write operation, and triggers the database entry receipt verification upon completion. Specifically, the system checks the write success flag returned by the database; if the write fails, the corresponding data is stored in the retry cache queue; if the write succeeds, it means that all data storage operations are completed. At this time, an audit log for this storage operation is generated, recording the storage time, number of data entries, and number of successes / failures, thus completing the unified persistent storage of all types of data.

[0039] The `trials` table stores the interaction data, single-trial feedback results, and `trial_id` associated fields for each training and evaluation trial. It is the smallest granularity table recording response behavior, and its fields include at least `trial_id`, `session_id`, `user_id`, `timestamp`, `module`, `task_type`, `game_id`, `question`, `correct_answer`, `user_answer`, `is_correct`, `confidence`, `reaction_time`, `used_hint`, `revised`, `uncertainty_choice`, `strategy`, `error_awareness`, `points`, `predicted_correct`, `predicted_time`, `ai_advice`, `crowd_answer`, `domain`, `combo_count`, `flow_triggered`, and `fe`. `edback_type`; where `session_id` is a foreign key pointing to the `sessions` table; `trial_id`: a unique number for a single question / training trial, serving as the primary key of the `trials` table, binding the feedback results and answer behavior data of a single trial to achieve accurate tracking of single question data within a session; `session` represents a complete assessment / training session. This table stores summary data for the entire session, including intermediate features at the trial level, 6+1 dimensional meta-quotient indicators, and updated user metacognitive profiles. Each complete user process is distinguished by `session_id`, and the fields include at least `session_id`, `user_id`, `timestamp`, `mq_total`, `sensitivity`, `calibration`, `uncertainty`, `error_awareness`, `strategy_control`, `metamemory`, `time_perception`, `profile`, `n_trials`, and `domain_breakdown`.`ai_decision_logs`: `ai` refers to AI collaborative interaction, `decision` represents the AI ​​decision node, and `logs` is a log table specifically storing calibration results generated by key AI collaborative nodes and a full record of user responses to metacognitive prompts. Fields include at least `log_id`, `session_id`, `user_id`, `timestamp`, `scenario`, `domain`, `ai_recommendation`, `user_decision`, `nudge_type`, `nudge_text`, `risk_score`, `verification_status`, and `outcome`. `session_id` is a unique identifier for a single training / evaluation session, used to distinguish multiple different evaluation records for the same user, and it is linked to the `sessions` main table and the `trials` sub-table. `achievements` refers to user achievements. The achievements table stores all data related to unlocking badges, points, levels, and training achievements. Fields include at least `achievement_id`, `user_id`, `timestamp`, `badge`, `description`, `unlocked_level`, `criteria_snapshot`, and `source_session_id`. `user_id` is a globally unique identifier for each user, serving as the core cross-table join key, linking all sessions, trials, logs, and achievements of the same user. `source_session_id` is a nullable field used to optionally record the source session of an achievement; it is not a necessary condition for achievement creation or existence. Therefore, `sessions` and `trials`, and `sessions` and `ai_decision_logs` have a 1:M relationship, while `achievements` and `sessions` have a user-level or optional source relationship. A 1:M relationship means the relationship between the `sessions` table and the `trials` table is one to many; for example, one `sessions` record can correspond to multiple `trials` records, but each `trial` can only belong to one session.

[0040] In one specific embodiment, the method is implemented in Python, the server-side framework uses Streamlit or an equivalent web service framework, data persistence uses SQLite, PostgreSQL, MongoDB or an equivalent database, and the client uses a web browser or WeChat mini-program for rendering. The visual effects of the real-time feedback mechanism are implemented using HTML / CSS keyframe animations or equivalent front-end animation components, such as particle-burst, msg-pop, flow-pulse, and combo-bounce animations, and are injected through a front-end component interface. The visual effects of the real-time feedback mechanism include feedback signals and secondary feedback results.

[0041] For example, the WeChat Mini Program includes four types of interfaces: student, parent, teacher, and senior citizen. The student interface facilitates interaction between the game recommendation module and the feedback coupling module. The parent interface is used for data storage, including viewing reports on changes in the student's total meta-quotient score and six-dimensional sub-items, setting training duration, and revoking authorization for objective behavioral data. The teacher interface displays the distribution of the class's total meta-quotient score and provides statistical views of user metacognitive profiles. The senior citizen interface features a simplified interface with large font size, slow speech, and high error tolerance, offering simplified assessments, anti-fraud training / simplified training, anti-fraud reminders, meta-quotient growth reports, user metacognitive profile viewing, and large font mode / voice broadcast functions. The simplified assessment includes low-question, low-time-pressure confidence calibration and suspiciousness judgment tasks with voice broadcast capabilities. The senior citizen interface defaults to an AI-based anti-fraud mode and writes simplified assessment and anti-fraud training results into the `trials`, `sessions`, and `ai_decision_logs` tables to form user metacognitive profiles and security reminders. The WeChat Mini Program interacts with each module via cloud APIs.

[0042] This embodiment also includes: performing sliding window calculations on at least six-dimensional meta-quotient indicators according to a preset time to form time series data of long-term growth curves and presenting them to users in a visual manner; at the same time, the long-term growth curves aggregate the user's historical meta-quotient profile evolution data into a metacognitive data flywheel, which is used to optimize the weight calibration of the assessment calculation module, the boundary of metacognitive profile classification, and the difficulty adaptive curve of the training game library; the long-term growth curves are visually presented as at least the main curve of the total meta-quotient score (thick solid line) and at least three sub-curves (confidence calibration, error detection, and strategy control) superimposed on each other.

[0043] As can be seen from the above, this application improves the comprehensiveness and objectivity of the quantitative assessment of metacognitive ability by collecting interactive data and weighted fusion of 6+1 dimensional meta-quotient indicators. It matches exclusive training games according to the user's metacognitive profile and couples the feedback signals with the metacognitive sub-capabilities to complete dynamic iterative training. Simultaneously, it actively outputs decision calibration intervention in AI collaboration scenarios based on the updated user metacognitive profile. Moreover, the interactive data, intermediate features of the trial, 6+1 dimensional meta-quotient indicators, trial feedback results, updated user metacognitive profile, calibration results and user responses, and achievement data are all uniformly stored in the process. This solves the shortcomings of existing metacognitive tools, such as single assessment dimensions, lack of personalized training adaptation, inability to quantify and iterate training effects in real time, lack of metacognitive risk intervention in AI collaboration scenarios, and inability to completely trace and retain behavioral data. This application realizes the integrated and automated operation of metacognitive assessment, training, risk intervention, and data auditing, continuously iteratively optimizing various sub-capabilities of user metacognition, and fundamentally reducing the risk of decision bias caused by users blindly trusting AI output.

[0044] In one embodiment of this application, in response to an evaluation command, acquiring interaction data with a user terminal includes: In response to the evaluation command, objective behavior tasks are selected from the preset objective behavior carrier task library according to the preset extraction rules, and the objective behavior tasks are pushed to the user terminal for display. Among them, the objective behavior carrier task library includes perception judgment task, logical reasoning task, AI suggestion adoption task, memory challenge task, color-word conflict task, visual search task, inhibition control task, and fuzzy uncertainty task; By acquiring users' responses to objective behavioral tasks, interactive data is generated. This interactive data includes the user's first correctness of response to the objective behavioral task, first confidence score, first reaction time, help-seeking behavior, answer modification markers, and first error prediction intensity.

[0045] In this embodiment, the current user's identity, assessment access permissions, authorization status of minors' personal information, and simplified assessment function switch for the elderly are first verified. If the verification fails, an assessment blocking prompt is returned to the user terminal, and all subsequent task push processes are terminated. If all verification items are compliant, the preset extraction rules in the local cache or database are read. Based on the total number of tasks limited by the rules, the distribution ratio of question types in the eight categories of objective behavior tasks, the difficulty gradient of questions, the proportion of erroneous AI conclusions in AI suggestion adoption tasks, and the multi-option ratio constraints of fuzzy and uncertain tasks, objective behavior tasks that match the needs of this assessment are selected from the preset objective behavior carrier task library. After the selection is completed, the question stem, answer control, three confidence selection buttons, help entry, error prediction mark component and other front-end rendering materials are uniformly packaged and pushed to the user terminal through the cloud communication interface to complete the page rendering and display.

[0046] After the user terminal completes the rendering of the objective behavioral task page, it will continuously monitor all user interface interactions to capture the user's complete response behavior for a single objective behavioral task. This includes the entire process of answer selection, confidence level clicks, help button triggering, answer modification, and error prediction checkmark selection. Each user operation will record the corresponding raw parameters. The terminal uses a single objective behavioral task as the smallest collection unit, binding and encapsulating all response behavior parameters captured within the same trial to generate standardized interaction data. If the user exits the assessment session midway, the currently generated interaction data will be temporarily cached and the collection record will be saved when the user re-enters the assessment page.

[0047] The pre-built objective behavior task library is a standardized question bank resource pool that can be iteratively updated over a long period of time. The objective behavior task library includes eight categories of objective behavior tasks: perception judgment tasks, logical reasoning tasks, AI suggestion adoption tasks, memory challenge tasks, color-word conflict tasks, visual search tasks, inhibition control tasks, and fuzzy uncertainty tasks. Each category of tasks is independently divided into difficulty levels and scenario tags, and supports adding, modifying, and removing question materials in the backend. The pre-built extraction rules support multi-dimensional custom configuration and can distinguish between four scenarios: student self-testing, teacher batch assessment, simplified assessment for the elderly, and parent monitoring assessment, setting differentiated extraction logic. The extraction ratio and question limit of various objective behavior tasks in different scenarios can be adjusted independently. At the same time, the rules have a built-in anomaly filtering mechanism that will remove expired or logically flawed task materials. For example, in the scenario of batch assessment for teachers, the total number of tasks can be set to 60, the proportion of logical reasoning tasks to 30%, the proportion of difficult questions to 40%, and the proportion of erroneous AI conclusions in the AI ​​suggestion adoption task to 60%. In the scenario of simplified assessment for the elderly, the total number of tasks can be modified to 30, the proportion of suppression and control tasks can be reduced, only easy and normal level questions can be retained, and the number of multiple options for fuzzy and uncertain tasks can be reduced. The multi-dimensional parameters of the two scenarios are independent of each other, forming two completely differentiated task extraction logics.

[0048] Specifically, the tasks include: perceptual micro-judgment tasks, such as comparing the mean or variance of two sets of numerical sequences; logical reasoning judgment tasks, such as propositional logic, relational transitivity, conditional inference, and counting problems; AI suggestion adoption tasks, such as providing a suggestion on the truth or falsehood of a factual proposition with a certain probability of being wrong, and the user deciding whether to adopt it; memory challenge tasks, such as recalling a word list and predicting the number of words that can be remembered; color-word conflict tasks, such as displaying Chinese characters in various colors and presenting them in font colors that are inconsistent with or consistent with the meaning of the characters, requiring the user to report the font colors while ignoring the meanings; and a visual search task using a Schulte table, such as clicking the numbers 1 to n in sequence in an n×n grid. 2The Go / No-Go suppression control task includes displaying a Go or No-Go signal and requiring the user to press the button when Going and suppressing the button when No-Going; the fuzzy uncertainty task includes at least three options for the task context, such as moving forward, calling for help, and going backward, to collect the user's uncertainty recognition behavior.

[0049] This embodiment also includes a three-speed button selection interface for collecting confidence scores for objective behavioral carrier tasks. The three speeds correspond to fast answering (I am very sure), medium answering (I think so), and slow answering (I am not so sure), and the three speeds are mapped to preset confidence values, for example, fast answering, medium answering, and slow answering are mapped to 90, 60, and 30. The three-speed selection interface is used to replace continuous slider input, reducing the interference of continuous slider input on the user's main task answering process.

[0050] As can be seen from the above, this embodiment limits eight types of standardized objective behavioral carrier tasks as interactive data collection carriers, collecting multi-dimensional interactive data including first answer correctness, first confidence score, first reaction time, help-seeking behavior, answer modification marking, and first error premonition intensity. This can improve the completeness of capturing the details of metacognitive behavior throughout the user's answering process, avoid the problems of data distortion and dimension loss in subjective questionnaire-style assessments, and provide sufficient, objective, and standardized underlying behavioral data sources for the subsequent calculation of 6+1-dimensional meta-quotient indicators, effectively improving the objectivity, robustness, and interpretability of meta-quotient assessment results.

[0051] In one embodiment of this application, a personalized training game is obtained by matching a combination of personalized training games from a preset training game library based on the user's metacognitive profile, including: Obtain user metacognitive profile type tags, which are: high metacognitive type, overconfident type, uncertainty-sensitive type, weak error awareness type, rigid strategy type, weak self-monitoring type, and growth type; Based on the type label, retrieve one of the target metacognitive sub-ability dimensions corresponding to the type label from the preset profile-game mapping rule table; Obtain the metacognitive sub-ability tags of each training game in the training game library, and filter the training game library according to the target metacognitive sub-ability dimensions and the metacognitive sub-ability tags of each training game to obtain the games to be recommended. Based on the type label, the initial difficulty level, feedback style preference, and risk warning frequency corresponding to the type label are queried from the preset difficulty mapping table. The initial difficulty level includes easy, normal, and hard levels. The feedback style preference includes incentive feedback style and neutral feedback style. The risk warning frequency includes high frequency, medium frequency, and low frequency. Personalized training games are generated based on the games to be recommended, the initial difficulty level, feedback style preferences, and the frequency of risk warnings.

[0052] In this embodiment, the user's metacognitive profile is first extracted into its own unique type tags. These type tags include seven categories: high metacognitive type, overconfident type, uncertainty-sensitive type, weak error awareness type, rigid strategy type, weak self-monitoring type, and growth type. If the verification finds that the type tag is invalid, the default fallback matching logic is triggered, and the matching rules corresponding to the growth type are used for subsequent calculations. After the verification is successful, the valid type tag is used as the retrieval primary key for subsequent queries of the preset profile-game mapping rule table.

[0053] In this embodiment, a preset profile-game mapping rule table stored locally is retrieved for retrieval. This preset profile-game mapping rule table includes a one-to-one binding mapping between each type label and the corresponding target metacognitive sub-capability dimension. Each mapping record only outputs the target metacognitive sub-capability dimension. During retrieval, a primary key matching mode is used to return the mapping data corresponding to the current type label, and the target metacognitive sub-capability dimension is extracted as the matching standard for subsequent game selection and training.

[0054] In this embodiment, a preset training game library is read, and the metacognitive sub-capability tags bound to each training game in the library are extracted synchronously. The retrieved target metacognitive sub-capability dimension is used as the filtering keyword, and all training games and their corresponding metacognitive sub-capability tags are traversed to perform matching comparisons. The comparison rule is that as long as any metacognitive sub-capability tag carried by a training game is consistent with the target metacognitive sub-capability dimension, the game is determined to meet the matching condition and is included in the set of games to be recommended. After the traversal is completed, all training games whose tags are not related to the target metacognitive sub-capability dimension are automatically removed, and the games that are successfully matched are taken as games to be recommended.

[0055] In this embodiment, after locking down the set of games to be recommended, the user's metacognitive profile type tags are used as search keys to retrieve a preset difficulty mapping table to complete the parameter query. This difficulty mapping table has three sets of parameters configured independently for each type tag: initial difficulty level, feedback style preference, and risk warning frequency. The initial difficulty level is divided into three categories: easy, normal, and hard, used to control the total number of questions, question complexity, and time limit in the game. The feedback style preference is divided into two categories: motivating feedback style and neutral feedback style, used to adjust the weight of the seven feedback types and the intensity of visual rendering. The risk warning frequency is divided into three levels: high frequency, medium frequency, and low frequency, used to control the pop-up interval and triggering conditions of metacognitive risk pop-ups in the game. Based on the games to be recommended and the initial difficulty level, feedback style preference, and risk warning frequency, personalized training games are generated.

[0056] As can be seen from the above, this embodiment matches the corresponding target metacognitive sub-capability dimensions based on seven types of user metacognitive profile tags. It completes the selection of recommended games based on the training games in the training game library, and simultaneously matches the initial difficulty level, feedback style preference, and risk warning frequency of the user profile to generate personalized training games. This abandons the uniform and undifferentiated training content push mode, improves the accuracy of matching corresponding training materials to the user's metacognitive weakness dimensions, and completes personalized adaptation of the training environment from three levels: difficulty, feedback, and risk warning. This effectively shortens the training cycle of the user's metacognitive weaknesses and further improves the matching degree between training content and user cognitive deficiencies.

[0057] In one embodiment of this application, the training game library includes core layer training games, AI era extension layer training games, and scene specialization layer training games. The core layer training games include Confidence Radar Game, Uncertainty Maze Game, Error Detective Game, Strategy Switching Master Game, and Memory Investor Game. The AI ​​era extension layer training games include Time Perception Challenge Game, AI Controversial Judge Game, Decision Review Game, Knowledge Blind Spot Radar Game, and Collective Intelligence PK Game. The scene specialization layer training games include Telepathy Game, Thinking Adventure RPG Game, and Minnan Dialect Anti-Fraud Master Game. Each training game corresponds to at least one metacognitive sub-ability dimension.

[0058] As one possible implementation, the training game library includes core layer training games, AI era extension layer training games, and scenario-specific layer training games. The core layer training games include Confidence Radar Game, Uncertainty Maze Game, Error Detective Game, Strategy Switching Master Game, and Memory Investor Game. The AI ​​era extension layer training games include Time Perception Challenge Game, AI Controversial Judge Game, Decision Review Game, Knowledge Blind Spot Radar Game, and Collective Intelligence PK Game. The scenario-specific layer training games include Telepathy Game, Thinking Adventure RPG Game, and Minnan Dialect Anti-Fraud Master Game. Each training game corresponds to at least one metacognitive sub-ability dimension.

[0059] Specifically, the Confidence Radar game is used to train confidence calibration capabilities. Users input confidence scores via three speed settings when answering questions in perception judgment or logical reasoning tasks. Radar accuracy is calculated based on the degree of agreement between the user's confidence scores from the most recent N1 trials and the correctness of their answers, and is presented visually. N1 is a fixed threshold number of historical consecutive trials used in the Confidence Radar game to statistically calculate radar accuracy. It represents the sample window size for extracting the user's recent N1 rounds of confidence and correct / incorrect data to perform confidence calibration index calculations. Its setting method is based on ablation experiments comparing the matching error between radar accuracy and the actual confidence calibration dimension score under different trial windows, selecting the trial value with the smallest error as the fixed value of N1. Simultaneously, two sets of differentiated N1 parameters are configured for students and the elderly to adapt to the different fluctuation characteristics of users' answering behavior. The Uncertainty Maze game is used to train uncertainty recognition capabilities, providing at least one option in fuzzy tasks: moving forward or seeking help. The game offers three response options: call for help, back off, etc. For example, when a user correctly chooses to ask for help or back off in a low-confidence or uncertain situation, the brightness of the maze map points increases and reward gems drop. The "Error Detective Game" is used to train error detection ability. Users are required to complete N2 questions consecutively and mark the questions they suspect are wrong without checking the correct answers. The detective accuracy is calculated based on the degree of match between the user's marking results and the actual number of errors, and badges are unlocked accordingly. N2 is a fixed threshold for the number of questions that need to be answered and then self-marked as suspected errors in a single training unit of the Error Detective Game. It is the boundary of the statistical window for judging the user's error detection performance. The setting method is to compare the accuracy and stability of the user's prediction of errors in multiple question intervals through offline user tests, and select the number of questions that can ensure sufficient samples without causing user operation burden as N2. The value of N2 is appropriately lowered to suit the elderly and young students in the simplified assessment mode.

[0060] The Time Perception Challenge game trains meta-time perception abilities. Users are required to estimate the number of seconds they took after completing a task, without seeing their actual reaction time. A score is given based on the relative error between the estimate and the actual reaction time. The AI ​​Controversy Judge game trains the reliability of AI outputs. Multiple factual judgments from AI are presented, some of which are intentionally mixed in with errors. The user's task is to identify which AI judgments are incorrect. The Decision Review game trains reflective and retrospective metacognitive abilities. After completing a certain number of training sessions... The system presents the user's most recent N3 decisions in a replay format, requiring the user to mark the trials where confidence and performance misaligned without viewing the correct answers. A "Knowledge Blind Spot Radar" game trains the user's willingness to seek help and their ability to recognize knowledge boundaries. Users are required to make a binary prediction before answering each question, such as "Will I answer correctly?", and are scored based on the accuracy of their binary predictions. A "Collective Intelligence PK" game trains relative calibration and judgment independence. In each trial, in addition to the user's answer, the AI's output and the answers from a simulated peer group are presented. The user's score is based on the relative advantage of their judgment being closer to the truth. N3 represents the number of the user's most recent trials used for replaying and marking trials where confidence and performance misaligned in the decision replay game. It is the window size for selecting decision samples for replay. The method for setting N3 is based on ablation experiments using historical user training data. By comparing the accuracy of user confidence misalignment marking under different replay windows, the trial value with the best recognition effect is fixed as N3. Different parameters are configured for students and the elderly. The "certain quantity" refers to the number of training questions / trials the user needs to complete before the replay stage of the decision-making replay game is triggered. After reaching this quantity, the user's most recent N3 decisions are replayed. The specific value of this "certain quantity" is determined through expert experience. N1, N2, and N3 are all greater than or equal to 2.

[0061] Strategy Switch Master game, designed to train strategy control capability, which provides users with a fixed quota of three strategies, namely fast thinking, slow thinking and help-seeking, in a set of tasks with increasing difficulty. Under the condition of limited strategy quotas, users select the optimal strategy for each question, and the score is given according to the matching degree between the strategy and the task difficulty. When the strategy matching degree is greater than a preset threshold, a "divine one move" special effect is triggered, wherein the preset threshold is a score critical value in the Strategy Switch Master game that distinguishes the quality of the user's strategy matching, and is used to determine whether to trigger the "divine one move" visual special effect; the setting method thereof is based on quartile segmentation of the distribution of user game strategy matching scores, the upper quartile is selected as the preset threshold, and the threshold can be fine-tuned up or down according to the type of the user's metacognitive profile, for overconfident users, the threshold is appropriately raised to increase the training constraint intensity; Memory Investor game, designed to train metamemory capability, which requires users to place a bet prediction on the number of items they can remember before a memory challenge task, and rewards or deducts points according to the gap between the predicted value and the actual value after the recall is completed, for example, a gold coin rain and the title of Investment Master are triggered when the prediction error is less than or equal to 1. Telepathy game, designed to train others' metacognitive capability, which requires users to estimate the percentage of one hundred peers who can answer each question correctly when answering each question, and scores according to the proximity between the user's estimated value and the correct rate of the simulated group; Thinking Adventure RPG game, designed for immersive comprehensive metacognitive capability training, in a section of RPG adventure narrative, users need to complete metacognitive judgments including knowledge blind spot prediction, AI oracle verification and strategy selection at at least three plot nodes along the way, and after completing the whole chapter, the Meta-cognition Quotient Hero Medal and the next chapter of the story are unlocked; Minnan Dialect Anti-Fraud Master game, designed to train users' capability to judge the suspiciousness of AI voice fraud, which plays or displays simulated call content, requires users to judge whether the call is suspicious, performs weighted matching according to the keyword sets of the call content and five preset fraud modes, and provides clue analysis.

[0062] In this embodiment, the five fraud modes include emergency money request from relatives, fake public security and procuratorial organs, AI phishing customer service, winning notification, and fake hospital. The keywords of the five fraud modes include, the keywords of emergency money request from relatives include but are not limited to "I am your son" (Minnan dialect variant), emergency, change number, transfer, your son, your daughter, your grandson, an accident happened, hit someone, urgently need; the keywords of fake public security and procuratorial organs include but are not limited to police, procuratorate, involved in a case, safety account, confidential investigation, wanted, summon, suspect; the keywords of AI phishing customer service include but are not limited to AI customer service, account abnormality, verify password, SMS verification code, balance; the keywords of winning notification include but are not limited to congratulations on winning a prize, free claim, time limit, bonus, lucky; the keywords of fake hospital include but are not limited to medical insurance account frozen, serious illness, account activation, medical subsidy.

[0063] The weighting method involves using historical real dialect fraud text samples obtained through statistics, training a preset text classification model to obtain the contribution of various keywords to the differentiation of fraud behavior as the initial weight, then manually verifying and adjusting the dialect vocabulary recognition deviation, and finally determining the exclusive k coefficient for each type of pattern for the weighted sum of the keyword hit scores in the phone content. After completing the fraud suspicion calculation, the keyword hit items of various types are output simultaneously as clue analysis content. For example, the weights for "relatives urgently requesting money," "impersonating public security, procuratorate, and court officials," "AI customer service phishing," "prize winning notifications," and "impersonating hospitals" are 9, 10, 8, 7, and 8, respectively. The fraud suspicion score is calculated as: likelihood = min(Σ_i(matched_i·weight_i) / 30×100, 99), where matched_i represents the number of keywords hit in the i-th fraud pattern; weight_i is the weight coefficient corresponding to the i-th fraud pattern (corresponding to 9, 10, 8, 7, and 8, respectively); Σ_i(matched_i·weight_i) represents the sum of the number of keyword hits and their corresponding weights for all fraud patterns, i.e., the weighted total hit score. For example, when likelihood is greater than or equal to 60, a highly suspicious warning is output; when likelihood is greater than or equal to 25 and less than 60, a suspicious prompt is output; and when likelihood is less than 25, a normal label is output.

[0064] The text classification model employs a lightweight rule-based and shallow statistical classification fusion architecture, consisting of a rule matching submodule and a shallow statistical fitting submodule connected in series. First, it performs precise and fuzzy text rule matching based on a pre-defined dictionary of five types of fraudulent Minnan (Southern Fujian) keywords, quickly extracting the original hit scores for each pattern. Then, the original scores are fed into a shallow statistical fitting module based on least squares, iteratively generating independent weighted k-coefficients for each pattern based on manually annotated fraudulent corpus, supplemented by manual dialect bias correction. The model's input is simulated telephone text, including mixed Mandarin and Minnan variants. Before input, character cleaning and dialect word segmentation are performed uniformly. The input includes a matching five-element keyword dictionary set, corresponding to five types of fraud: relatives urgently requesting money, impersonating public security, procuratorate, and court officials, AI customer service phishing, prize-winning notifications, and impersonating hospitals. The model output is divided into two parts. The first output is the original matching score of each of the five fraud patterns. The second output is the independent weighted coefficients k1, k2, k3, k4, and k5 after model fitting and manual adjustment. The output k coefficients are used for weighted summation of the suspiciousness of the call content. At the same time, each matched keyword entry is output as clue analysis material.

[0065] As can be seen from the above, this embodiment constructs three types of training games in a layered manner, which are used to specifically train the capabilities of seven meta-intelligence dimensions and AI collaboration, anti-fraud scenarios, etc. Through ablation experiments, offline user testing, and expert experience, the N1, N2, and N3 window thresholds and strategy matching thresholds are scientifically set to balance sample sufficiency and operational burden, and to be suitable for different groups of people such as students and the elderly. The Minnan dialect anti-fraud master trains a lightweight mixed text classification model based on real dialect fraud data, calculates suspiciousness by weighted keywords and provides graded warnings, while taking into account the accuracy of dialect recognition. Each game is equipped with exclusive visual incentives and badge unlocking mechanisms, which make the training of metacognitive shortcomings concrete and gamified. This not only achieves precise reinforcement of each meta-intelligence dimension, but also improves the comprehensiveness of AI human-computer collaboration and elderly dialect anti-fraud scenarios, and further improves the pertinence, fun and scenario application value of metacognitive training.

[0066] In one embodiment of this application, a personalized training game is generated based on the game to be recommended, the initial difficulty level, feedback style preferences, and the frequency of risk warnings, including: Based on the type tags of the user's metacognitive profile and the metacognitive sub-capability tags of the game to be recommended, calculate the sub-capability coverage vector of the game to be recommended. Each element in the sub-capability coverage vector represents the training coverage intensity of the game to be recommended for each dimension of the 6+1 dimensional meta-quotient index. Based on the initial difficulty level and sub-ability coverage vector, an initial configuration parameter set for the game to be recommended is generated. The initial configuration parameter set includes question difficulty parameters, time pressure parameters, and available hint parameters. Based on feedback style preferences and the sub-capability coverage vector of the game to be recommended, the probability distribution of target occurrence for each feedback type is determined through the feedback policy generation function. The feedback policy generation function determines the overall feedback tendency based on feedback style preferences and adjusts the occurrence probability of each feedback type according to the coverage strength of each dimension in the sub-capability coverage vector, so that the distribution of feedback types matches the training target of the game to be recommended. The initial configuration parameter set is associated with the probability distribution of target occurrence for each feedback type to generate a personalized training game.

[0067] In this embodiment, firstly, based on the type tags of the user's metacognitive profile and the metacognitive sub-capability tags of the game to be recommended, if any game to be recommended is missing a metacognitive sub-capability tag, it is automatically marked and supplemented with basic general coverage weights. When calculating the sub-capability coverage vector, the 6+1-dimensional meta-quotient index is used as the vector dimension benchmark. Each dimension is traversed in turn. Based on the reinforcement priority corresponding to the type tags of the user's metacognitive profile and the matching degree of the metacognitive sub-capability tags of the game to be recommended, a sub-capability coverage vector is generated. Each independent element inside the sub-capability coverage vector corresponds to the training coverage intensity of a single dimension in the 6+1-dimensional meta-quotient index. The coverage intensity value fluctuates positively with the matching degree between the game and the profile's shortcomings.

[0068] Secondly, the initial difficulty level is used as the basic control benchmark. Based on the training coverage intensity of each dimension in the sub-ability coverage vector of the game to be recommended, the constraint values ​​are distributed differently to generate the initial configuration parameter set in batches. The initial configuration parameter set includes three types of parameters: question difficulty parameter, time pressure parameter, and hint availability parameter. Among them, the question difficulty parameter sets the basic question gradient according to the initial difficulty level and increases the proportion of corresponding question types in the high coverage intensity dimension of the sub-ability coverage vector. The time pressure parameter matches the upper limit of the answering time according to the level, while appropriately relaxing the answering time limit for high reinforcement dimensions. The hint availability parameter controls the number of times and triggering conditions of help-seeking interactive functions. All parameters are generated by the initial difficulty level and the sub-ability coverage vector.

[0069] Subsequently, feedback style preference and sub-capability coverage vectors are extracted, and both types of data are uniformly input into the feedback strategy generation function for computation. The feedback strategy generation function consists of two layers of computational logic. The first layer prioritizes locking the overall feedback tendency based on feedback style preference. Incentive feedback styles will increase the base probability of positive feedback, while neutral feedback styles will balance the base proportion of various types of feedback. The second layer reads the training coverage intensity corresponding to each dimension in the sub-capability coverage vector. For metacognitive dimensions with high coverage intensity that need to be strengthened, the probability of the feedback type appearing in that dimension is adjusted upwards to complete the probability shift adjustment. After two layers of computation, a standardized target appearance probability distribution for each feedback type is output, which fits the training target of the game to be recommended.

[0070] Finally, the initial configuration parameter set is used as the underlying operating constraint of the game to be recommended, and the probability distribution of the occurrence of each feedback type is used as the incentive control rule in the game to be recommended. The two sets of data are bound one-to-one to form a complete set of game operation rules. After the game to be recommended completes the parameter association combination, a traceability identifier is uniformly attached, and the type label of the user metacognitive profile and the original value of the sub-capability coverage vector used in the generation process are recorded. Finally, complete personalized training games are batch packaged and generated. The packaged personalized training games include difficulty constraints, time limits, prompt permissions, and feedback probabilities, and can be directly pushed to the user terminal for loading and running.

[0071] As can be seen from the above, this embodiment quantifies the training intensity of the game on each meta-quotient dimension by calculating the sub-capability coverage vector of the game to be recommended. Based on this, three initial configuration parameters are generated for the adaptation profile: question difficulty, time pressure, and available hints. Then, the probability distribution of the occurrence of feedback types is dynamically adjusted according to the feedback strategy generation function and the sub-capability coverage vector. Finally, the initial configuration parameter group is combined with the target occurrence probability distribution of each feedback type to generate a personalized training game. This achieves simultaneous matching of the user's metacognitive weaknesses in three dimensions: training difficulty, task pressure, and incentive feedback. This allows training resources to be tilted towards the weak metacognitive dimensions, avoiding the problems of unbalanced training difficulty and disconnect between feedback incentives and training objectives, and further enhancing the pertinence and efficiency of personalized training.

[0072] In one embodiment of this application, after generating a personalized training game based on the game to be recommended, the initial difficulty level, feedback style preferences, and the frequency of risk warnings, the method further includes: Obtain the user's second confidence score and second response accuracy in personalized training, and determine whether anchor point trials are needed; If an anchor trial needs to be inserted, the third confidence score and the third correctness of the response for the anchor trial are obtained, and the metacognitive heartbeat signal is calculated; the metacognitive heartbeat signal includes the current confidence calibration deviation and the current reaction time offset; Metacognitive heartbeat signals are used as adjustment parameters and fed back to the difficulty adjustment module of the personalized training game to adjust the question difficulty parameters and time pressure parameters.

[0073] In this embodiment, the second confidence score and second answer correctness generated by the user in each trial within the personalized training are collected. The distribution of the second confidence score and second answer correctness in the current window is compared with the standard matching interval corresponding to the user's metacognitive profile. The system comprehensively judges whether there is a shift in the current user's confidence stability and answer matching degree, thereby determining whether anchor trial needs to be inserted. The upper and lower thresholds of the standard matching interval are generated by clustering the confidence-correctness matching samples of training trials of the same user metacognitive profile type. A separate matching interval threshold is defined for each type of user metacognitive profile. The continuous trial judgment threshold is a preset critical value for the number of continuous trials. This preset number of continuous trials is determined by controlling different variables of the number of continuous trials separately, keeping all other experimental conditions unchanged, and comparing the recognition accuracy of metacognitive fluctuations under different continuous lengths. The number of trials corresponding to the peak recognition accuracy is taken as the preset critical value for the number of continuous trials. Specifically, the judgment logic distinguishes between two trigger scenarios: one is when multiple consecutive trials reach the consecutive trial judgment threshold and a long-term misalignment occurs between confidence and correctness; the other is when the current training difficulty and user fit continuously deviate from the standard matching range. If either scenario is met, it is determined that an anchor point trial needs to be inserted.

[0074] When the determination result indicates that an anchor trial needs to be inserted, the anchor trial is inserted into the current task flow of personalized training, and the third confidence score and third answer correctness data generated after the user answers the anchor trial are collected simultaneously. A metacognitive heartbeat signal is calculated based on the third confidence score and the third answer correctness data. This metacognitive heartbeat signal is composed of the current confidence calibration deviation and the current reaction time offset. The current confidence calibration deviation is calculated based on the third confidence score and the third answer correctness, and is used to quantify the deviation between the user's subjective confidence and objective answer results. When calculating the current reaction time offset, the third answer correctness is used as the classification label to match the corresponding historical answer duration benchmark. The current reaction time offset is used to characterize the degree of fluctuation in the user's answer rhythm relative to its own stable benchmark. After each of these two indicators completes standardized numerical mapping, they are combined to generate a complete metacognitive heartbeat signal, which serves as a quantitative carrier representing the user's real-time metacognitive fluctuations.

[0075] Metacognitive heartbeat signals are encapsulated into standardized adjustment parameter packages and fed back to the built-in difficulty adjustment module in personalized training. The difficulty adjustment module has built-in bidirectional parameter adjustment logic, corresponding to two types of adjustable configuration items: question difficulty parameters and time pressure parameters. Based on the sign and magnitude of the current confidence calibration deviation, the difficulty adjustment module adjusts the question difficulty parameters accordingly: if the deviation value is higher, it means that the user is more overconfident, so the proportion of conflicting question types between the anchor point trial and subsequent regular questions is increased, and the question difficulty parameters are increased; if the deviation is lower, it means that the user is more cautious, so the question difficulty parameters are appropriately decreased. At the same time, the time pressure parameters are adjusted synchronously according to the current reaction time deviation. When the answering time deviates significantly from the baseline, the answering time limit for a single question is relaxed or tightened accordingly to balance the training rhythm.

[0076] As can be seen from the above, this embodiment determines whether to insert anchor point trials by collecting user confidence and response performance during the training process. By using anchor point trials to calculate metacognitive heartbeat signals including confidence calibration deviation and reaction time offset, and then adjusting the game difficulty and time pressure parameters in reverse, dynamic adaptive parameter tuning of the training process is achieved. It can capture confidence deviation and abnormal response rhythm in real time and adjust the task intensity in a timely manner, avoiding the problems of training being too difficult and frustrating or too easy and inefficient due to fixed difficulty, maintaining a stable and sustainable training flow, and further ensuring the dynamic adaptability of metacognitive training.

[0077] In one embodiment of this application, a feedback signal is determined based on response behavior data, and the feedback signal is coupled to the metacognitive sub-capabilities corresponding to the user's metacognitive profile at the trial level to obtain a trial-level feedback result, including: Based on the second correctness of response, second confidence score, number of hits, and intensity of second error premonition in the response behavior data, the feedback type is determined; the feedback types include flow feedback type, hit feedback type, perfect calibration feedback type, honesty reward feedback type, self-awareness feedback type, normal correctness feedback type, and neutral feedback type. Based on the feedback type, determine the corresponding score, visual presentation method, and sound effect presentation method to obtain the feedback signal; The feedback signal is coupled with the metacognitive sub-capabilities corresponding to the user's metacognitive profile in a trial-level coupling to obtain the trial-level feedback result.

[0078] In this embodiment, the combination of the second correctness of the answer, the second confidence score, the number of hits, and the intensity of the second error prediction is compared item by item in the answer behavior data. Only the feedback type that meets all the corresponding conditions is matched. The feedback types include seven categories: flow feedback type, hit feedback type, perfect calibration feedback type, honesty reward feedback type, self-awareness feedback type, normal correct feedback type and neutral feedback type.

[0079] After determining the feedback type, the pre-set feedback material configuration library is retrieved to obtain the corresponding score, visual presentation method, and sound effect presentation method for that feedback type. The pre-set feedback material configuration library is a standardized resource database that stores all the output parameters associated with each feedback type. The library uses the feedback type as the unique search key and binds a complete set of material parameters for each type of feedback, including the score, visual presentation method, and sound effect presentation method. It is also divided into incentive, prompt, and general material sections for categorized management. The setup method is as follows: first, differentiated score levels are defined for each of the seven feedback types and the corresponding score values ​​are entered; then, exclusive pop-ups, particle animations, text color schemes, highlighting effects, and other visual materials are matched for each type of feedback and the visual parameters are entered; finally, sound effect parameters are bound to incentive, prompt, and general audio clips that match the mood. After all the materials are entered, a one-to-one mapping index between the feedback type and the three types of parameters is established.

[0080] The score is a quantified score specific to the corresponding feedback type, which can be accumulated and included in the user's level results and achievement data unlocking progress; the visual presentation includes front-end rendering materials such as pop-up styles, particle animations, text color schemes, and highlighting effects; the sound effect presentation is a dedicated audio clip adapted to the emotion of the feedback, distinguishing between incentive-type, prompt-type, and normal sound effects. The score value obtained by matching the current feedback type, the visual presentation method, and the sound effect presentation method are integrated and encapsulated to obtain the feedback signal, which is used to send to the user's terminal to complete the visual and auditory display.

[0081] The feedback signal is coupled with the metacognitive sub-capabilities corresponding to the user's metacognitive profile at the trial level to obtain the trial-level feedback result. The coupling operation is performed at the independent granularity of a single training trial. Specifically, the trial-level coupling operation establishes a strong binding association between the feedback signal generated in the current trial and the metacognitive sub-capabilities corresponding to the matched user metacognitive profile, simultaneously annotating the short-term gain magnitude of this feedback on that metacognitive sub-capability and recording the binding source relationship. After completing the trial-level coupling binding operation, all structured information, including the current feedback type, complete feedback signal, bound metacognitive sub-capabilities, dimensional short-term gain, and current trial identifier, is integrated to obtain the trial-level feedback result. The generated trial-level feedback result is transmitted via two links: one is pushed to the user terminal for rendering and displaying the feedback signal composed of scores, animations, and sound effects; the other is cached in memory for subsequent computational steps to update the 6+1 dimensional meta-quotient index and refresh the user metacognitive profile.

[0082] For example, the second correctness of the answer includes whether the user answers correctly or incorrectly, and the method for judging the feedback type includes: (a) flow feedback type: the user answers correctly and the second confidence score is greater than or equal to 0.7 and the number of hits is greater than or equal to 4, which is visually presented as an animation of purple to magenta two-color particles bursting outward; (b) Combo Feedback Type: If the user answers correctly and the combo count is greater than or equal to 3 but does not meet the flow condition, it is visually presented as an orange-red combo number bouncing animation; (c) Perfect Calibration Feedback Type: When the user answers correctly and the second confidence score is greater than or equal to 0.8 but does not meet the combo condition, it is visually presented as a dazzling animation of golden particles; (d) Honest Reward Feedback Type: When a user answers incorrectly and their second confidence score is less than or equal to 0.3, the feedback is visually presented in green text with the message "Although you answered incorrectly, you already felt it - metacognition + X". Honest reward feedback regards the accuracy of confidence calibration as a positive reflection of metacognitive ability rather than a punishment for failure. Based on metacognitive calibration theory, honest reward feedback considers low confidence + incorrect answer as a reflection of metacognitive ability where the confidence score is highly matched with the actual performance. The score for rewarding is calculated as base + 5·(1-p2) + 2·(e / 100), where base = 0, p (second confidence score) is less than or equal to 0.3, to encourage users to honestly express low confidence when uncertain rather than blindly having high confidence; e is the intensity of the second error premonition, which is filled in by the user when answering.

[0083] (e) Self-awareness feedback type: If the user answers incorrectly and the second error prediction intensity is greater than 60 but does not meet the conditions for honesty reward feedback type, it is visually presented as blue text with a detective awareness badge animation; (f) Normal correct feedback type: The user answers correctly but does not meet any of the above bonus conditions; (g) Neutral feedback type: Other situations that do not belong to any of the above types.

[0084] Among them, the honesty reward feedback provides positive rewards for low confidence and incorrect answers, and regards the accuracy of confidence calibration as a positive manifestation of metacognitive ability rather than a punishment for failure to answer, forming a feedback strategy that is different from simply rewarding the correctness of the answer.

[0085] As can be seen from the above, this embodiment classifies seven types of differentiated feedback based on multi-dimensional parameters of the answer behavior data, and provides corresponding three-dimensional integrated feedback signals of scores, visuals, and sound effects. The feedback signals are then coupled with the user's corresponding metacognitive sub-ability test sub-test. This breaks through the traditional cognitive training's single feedback logic of awarding rewards based solely on the correctness of answers. It can distinguish and capture differentiated metacognitive behaviors such as user confidence matching, error prediction, and continuous stable answers, and provide targeted incentives. In a single training session, real-time evaluation and training reinforcement of metacognitive sub-abilities are completed simultaneously, achieving a deep binding between feedback incentives and metacognitive ability improvement, effectively enhancing user training participation and metacognitive calibration training effects.

[0086] In one embodiment of this application, the user's decisions during the AI ​​collaboration process are calibrated based on the updated user metacognitive profile to obtain calibration results and user responses, including: Obtain users' historical interaction data and updated user metacognitive profiles, and calculate users' overconfidence index in each domain according to preset domain segmentation rules; When a user is detected to be at a preset key node in collaboration with the AI ​​system, the risk level of the current collaboration scenario is determined based on the updated user metacognitive profile and overconfidence index. If the risk level is greater than or equal to the preset risk threshold, then based on the risk level and the overconfidence index, a corresponding metacognitive prompt will be generated and pushed to the corresponding user terminal for display. The metacognitive prompt is used to assist users in making judgments and decisions during AI collaboration. The output of metacognitive prompts and the decision-making behavior of users based on metacognitive prompts are recorded to obtain calibration results and user responses.

[0087] In this embodiment, the user's historical interaction data and updated user metacognitive profile are first acquired. Based on preset domain segmentation rules, the user's overconfidence index in each domain is calculated. These domain segmentation rules define domain boundaries according to different AI collaboration scenarios such as learning, workplace, and anti-fraud for the elderly. Based on the user's historical interaction data and records of correct and incorrect answers and subjective confidence in each domain, the user's overconfidence index in each domain is quantitatively calculated. Normalization logic is used in the quantification process to ensure that the overconfidence indices obtained from different domains are comparable. The overconfidence index is then categorized into high and low deviation thresholds. These thresholds are determined by quartile segmentation based on paired samples of historical confidence and accuracy in each domain. The upper quartile serves as the high overconfidence threshold, and the lower quartile as the low deviation threshold, used for subsequent risk level weighting calculations. After calculation, the overconfidence index in each domain is bound and stored with the corresponding domain identifier.

[0088] This embodiment monitors user interaction with AI and identifies whether the user has entered a preset key node for collaborative interaction with AI. The preset key nodes include interactive nodes such as confirming and submitting AI-generated content, confirming and executing AI-provided decision-making solutions, and confirming key operations guided by AI voice. Once the corresponding node is detected and triggered, the overconfidence index of the current conversation matching domain and the updated user metacognitive profile are retrieved. Based on the user metacognitive profile and the overconfidence index, the risk level of the current collaborative scenario is determined.

[0089] The risk level of the current collaboration scenario is compared with a preset risk threshold. Only when the risk level is greater than or equal to the preset risk threshold is the metacognitive prompt information generation process executed. The preset risk threshold is calculated using ROC curve analysis on a historical AI collaboration risk case annotation dataset, with the optimal value combining risk identification precision and recall as the preset risk threshold. Simultaneously, scenario-specific sub-thresholds are configured for students, workplaces, and the elderly, with these sub-thresholds fine-tuned based on the proportion of historical misjudgment samples in each scenario. The generation process also considers two core parameters: the risk level of the current collaboration scenario and the overconfidence index of the corresponding domain. The risk level determines the overall intensity level of the prompt, while the overconfidence index adjusts the emphasis of deviation warnings within the prompt content. Based on the intensity level, corresponding materials from a preset prompt template library are matched, and personalized filling is completed based on the updated user metacognitive profile's age, dialect preferences, and historical prompt response records to generate metacognitive prompt information adapted to the current scenario. After content rendering and material encapsulation, the metacognitive prompt information is pushed to the corresponding user terminal for display via a communication link. The metacognitive prompt information provides objective verification suggestions from the perspective of subjective cognitive bias to assist users in judgment and decision-making during AI collaboration. If the risk level is less than the preset risk threshold, the prompt generation and push process is skipped, and no prompt content is sent to the user's terminal. The preset prompt template library is a resource collection that centrally stores standardized basic materials for metacognitive prompts in different scenarios. It stores the corresponding prompt content structure, visual style parameters, and interactive guidance methods based on scenario type identifiers, serving as the underlying source of materials for generating personalized prompt content. The setup method involves creating independent material folders for three major scenarios: students, workplace, and the elderly. For each scenario, a prompt text framework, matching color scheme, font, and interactive operation logic adapted to the target audience's comprehension ability is written. All basic templates are entered, and a one-to-one mapping index between scenario identifiers and templates is established.

[0090] In this embodiment, after the metacognitive prompt information is pushed to the user terminal and displayed, two independent recording and collection links are simultaneously initiated. The first link retains the output record of the metacognitive prompt information corresponding to this distribution operation, including the prompt distribution time, current collaboration scenario identifier, risk level, overconfidence index, prompt strength, complete prompt text, and display parameters. The second link continuously monitors the user terminal's interactive operations, capturing all user actions after receiving the metacognitive prompt information, collecting various decision-making behaviors made by the user based on the metacognitive prompt information, and distinguishing different behavior types such as ignoring the prompt and directly confirming the AI ​​solution, reviewing the prompt and re-verifying the AI ​​content, and terminating the current AI collaboration operation. The output record of the metacognitive prompt information and various user decision-making behaviors are structured, integrated, and encapsulated to obtain calibration results and standardized user responses.

[0091] As can be seen from the above, this embodiment calculates the overconfidence index by domain based on the user's historical interaction data and the updated user metacognitive profile. At key nodes of AI collaboration, the risk level of the scenario is determined based on the updated user metacognitive profile and the overconfidence index. Based on the risk level, corresponding metacognitive prompts are generated to assist the user in verifying AI decisions. At the same time, the output of metacognitive prompts and the user's decision-making behavior are recorded to form calibration results and user responses. This fills the gap in existing AI collaboration tools that lack personalized risk reminders based on the user's inherent cognitive biases. It specifically corrects the cognitive defects of users' overconfidence and blind acceptance of AI conclusions, and reduces the risk of misjudgment in AI-assisted decision-making in scenarios such as learning, workplace, and anti-fraud for the elderly.

[0092] In one embodiment of this application, the risk level of the current collaborative scenario is determined based on the user's metacognitive profile and overconfidence index, including: Obtain the scene feature vector of the current collaboration scenario. The scene feature vector includes task complexity features, decision consequence severity features, AI suggestion confidence features, and time pressure features. Based on the user's metacognitive profile, determine the user's weak metacognitive dimensions set and the corresponding weakness values; the weak metacognitive dimensions set is the set of dimensions in the 6+1 dimensional meta-quotient index that are less than the preset dimension threshold; The user's comprehensive metacognitive risk base is calculated based on the overconfidence index and the set of weak metacognitive dimensions. The comprehensive metacognitive risk base is obtained by weighting and summing the weakness values ​​of each weak dimension with the corresponding preset dimension weights, and then integrating them with the overconfidence index. Based on the scene feature vector, the scene risk contribution value is calculated through a preset risk level assessment function. The risk level assessment function calculates a weighted score based on the feature value and corresponding feature weight of each scene feature. Based on the comprehensive metacognitive risk base and the risk contribution value of the scenario, the risk level of the current collaboration scenario is determined through a preset risk level mapping matrix.

[0093] In this embodiment, firstly, based on the scene feature vector of the current collaborative scenario, multi-dimensional scene state information during the AI ​​collaboration process is captured. The scene feature vector includes task complexity features, decision consequence severity features, AI suggestion confidence features, and time pressure features. Among them, the task complexity feature is used to represent the knowledge difficulty, number of steps, and logical complexity of the current AI collaboration task; the decision consequence severity feature is used to determine the scope and degree of harm caused by the user's decision error; the AI ​​suggestion confidence feature is used to characterize the accuracy, stability, and model credibility of the AI ​​output content; and the time pressure feature is used to reflect the remaining time limit for the user to complete the decision and the urgency of the operation. These four types of features together constitute a complete quantitative basis for the scenario.

[0094] Based on the updated user metacognitive profile, the set of weak metacognitive dimensions and their corresponding weakness values ​​are determined. The weak metacognitive dimension set consists of dimensions in the 6+1 dimensional meta-quotient index that are below a preset threshold. Simultaneously, a corresponding weakness value is generated based on the difference between each dimension index and the preset threshold. This weakness value distinguishes the severity of deficiencies in different dimensions; a higher weakness value indicates a more prominent weakness in that metacognitive sub-capability and a greater likelihood of decision-making bias in AI collaboration. The preset threshold is determined by stratified sampling of standard assessment datasets from across the platform, calculating the median of the 6+1 dimensional meta-quotient index for three groups: students, professionals, and the elderly. This median is used as the general preset threshold. If a dimension index value is below the preset threshold, it is included in the weak metacognitive dimension set.

[0095] Secondly, the preset dimension weights are non-negative fixed coefficients pre-configured by the system. Different metacognitive sub-capability dimensions correspond to differentiated weights, which can reflect the priority of the impact of each weak dimension on AI decision-making bias. By weighting and summing the weakness values ​​of each weak dimension with the corresponding preset dimension weights, and then integrating them with the overconfidence index, a comprehensive metacognitive risk base that can characterize the user's individual cognitive biases and capability shortcomings is obtained. This comprehensive metacognitive risk base is used to characterize the user's inherent risk and is independent of external risks in the scenario.

[0096] This embodiment uses the constructed scene feature vector as input and calls a preset risk level assessment function to calculate the scene risk contribution value, thereby achieving a quantitative assessment of objective scene risk. The risk level assessment function takes the standardized scene feature vector as input, first standardizing and denoising each feature value to eliminate dimensional differences, then matching each feature with its preset exclusive feature weights to complete a weighted summation calculation, outputting a unique scene risk contribution value. This scene risk contribution value characterizes the inherent objective risk level of the current AI collaboration scenario. The standardized scene feature vector includes task complexity, severity of decision consequences, AI suggestion confidence, and time pressure. The preset exclusive feature weights are derived from statistical analysis of historical AI collaboration risk cases in the industry and joint assessment and calibration by experts in metacognition and human-computer interaction.

[0097] Specifically, based on the comprehensive metacognitive risk base and the scenario risk contribution value, the risk level of the current collaborative scenario is determined through a preset risk level mapping matrix. This preset risk level mapping matrix is ​​a two-dimensional association judgment matrix, with the horizontal axis corresponding to the grading intervals of the comprehensive metacognitive risk base and the vertical axis corresponding to the grading intervals of the scenario risk contribution value. The grading interval thresholds for both the comprehensive metacognitive risk base and the scenario risk contribution value are determined using the K-Means unsupervised clustering algorithm. Historical AI collaborative samples are binned and clustered to generate interval segmentation thresholds that uniformly distinguish between low, medium, and high risks. The segmentation point with the optimal clustering contour coefficient is used as the formal grading interval threshold and written into the risk level mapping matrix. Each intersection point within the matrix is ​​pre-bound with a unique corresponding standardized risk level. By accurately matching the grading intervals of the comprehensive metacognitive risk base and the scenario risk contribution value, the corresponding points in the matrix are located, resulting in a risk level suitable for the current user state and scenario environment.

[0098] As can be seen from the above, this embodiment collects scenario feature vectors of task complexity, severity of decision consequences, AI suggestion confidence, and time pressure. It also extracts a set of weak metacognitive dimensions and their degree of weakness based on the user's metacognitive profile, integrates the weighted scores of weak dimensions and the overconfidence index to calculate a comprehensive metacognitive risk base, and obtains the scenario risk contribution value through a risk level assessment function. Finally, it determines the risk level of the current collaborative scenario based on the risk level mapping matrix. This approach can simultaneously take into account both the user's inherent cognitive deficiencies and the dual risk factors of the external task environment, avoiding the one-sided and inaccurate risk assessment caused by relying solely on the overconfidence index for classification. It achieves refined and quantifiable classification of AI collaborative scenario risks, providing an accurate and reliable classification basis for subsequent differentiated push of metacognitive prompts, and further effectively reducing the probability of misjudgment and omission in high-risk decision-making scenarios.

[0099] In one embodiment of this application, before pushing the metacognitive prompt information to the corresponding user terminal for display, the method further includes: Obtain the scenario type identifier of the current collaboration scenario. The scenario type identifier includes AI-assisted homework scenario for students, AI decision-making suggestion scenario for the workplace, and AI voice interaction scenario for the elderly. Match the basic prompt template corresponding to the scene type identifier from the preset prompt template library. The basic prompt template includes the prompt content structure, visual style parameters and interaction guidance method. Obtain the user's personal information and / or the user's historical calibration records, and personalize the basic prompt template based on the personal information and / or historical calibration records to obtain personalized prompt content; personal information includes age information, educational background information and / or dialect preference information, and historical calibration records include user response behavior data to historical metacognitive prompt information; Based on the risk level, determine the intensity level of the personalized prompt content; the intensity levels include low intensity prompts, medium intensity prompts, and high intensity prompts; low intensity prompts correspond to displaying non-modal static prompt information on the user interface, medium intensity prompts correspond to displaying modal pop-ups on the user interface with visual highlighting effects, and high intensity prompts correspond to displaying modal pop-ups on the user interface with continuous visual emphasis animations and / or auditory reminder signals. Based on the prompt intensity level and scene type identifier, the target presentation parameters corresponding to the prompt intensity level and scene type identifier are queried from the preset presentation strategy table. The target presentation parameters include the display position, display size, background color, border style, animation effect and / or audio playback method of the prompt information. Based on the target presentation parameters, render metacognitive prompts and push them to the corresponding user terminals for display.

[0100] In this embodiment, the scene type identifier of the current collaboration scenario is first identified and obtained, which is used to lock the appropriate prompt resources and display logic. The scene type identifier includes student AI-assisted homework scenario, workplace AI decision-making suggestion scenario and elderly AI voice interaction scenario.

[0101] This embodiment categorizes AI collaboration tasks based on their business attributes, interaction methods, and user characteristics to achieve unique scene identifier matching. Different scene type identifiers correspond to independent prompt styles, content focuses, and interaction logic. The default match is the scene type identifier that best fits the current interaction behavior.

[0102] After identifying the scenario type, a pre-defined prompt template library is retrieved, and a basic prompt template corresponding to the scenario type is matched. The basic prompt template includes the prompt content structure, visual style parameters, and interaction guidance methods. Specifically, the prompt content structure defines the sentence structure of the prompt text, the arrangement logic of the risk explanation module, the decision suggestion module, and the closing guidance module; the visual style parameters pre-include the basic color scheme, font specifications, and control layout standards for the corresponding scenario; and the interaction guidance methods define the user's actionable path after the prompt is displayed, including standardized interaction logic such as viewing details, immediate verification, and ignoring the prompt. The pre-defined prompt template library is partitioned and stored according to the three scenario types. The basic prompt templates for each scenario are independent and not interchangeable, ensuring that the prompt style is highly adapted to the current AI collaboration scenario.

[0103] Further, the system acquires the user's personal information and / or historical calibration records. These two types of personalized data are used to dynamically and adaptively populate the matched basic prompt template, resulting in personalized prompt content. Personal information includes age, educational background, and / or dialect preference information, used to adapt to the user's cognitive comprehension ability and reading habits, achieving personalized adaptation in terms of text accessibility and expression style. Historical calibration records include user response behavior data to historical metacognitive prompts; based on the user's past behaviors such as ignoring, adopting, and verifying, the system adjusts the guidance focus and reminder intensity of the prompts. The population process does not alter the overall framework of the basic prompt template; it only replaces and modifies replaceable text, guiding phrases, and adaptation parameters within the template, transforming the standardized basic prompt template into exclusive personalized prompt content tailored to the user's characteristics.

[0104] This embodiment determines the corresponding prompt intensity level based on the risk level of the current collaboration scenario. Specifically, the system pre-processes the historical AI collaboration scenario risk level values ​​using the K-Means clustering algorithm in an unsupervised manner, dividing them into three non-overlapping risk value intervals and binding them to three prompt intensity levels: low intensity, medium intensity, and high intensity. During clustering, the user's effective reception of the prompts is used as the silhouette coefficient optimization index, and the interval segmentation threshold with the best clustering effect is taken as the judgment boundary. During runtime, the real-time calculated risk level value of the current collaboration scenario is directly compared with the three preset intervals in sequence. If the risk level falls into the lowest interval, a low intensity prompt is matched; if it falls into the middle interval, a medium intensity prompt is matched; and if it falls into the highest interval, a high intensity prompt is matched, thus completing the determination of the prompt intensity level. Low-intensity prompts correspond to displaying non-modal static prompts on the user interface, without pop-up forced interaction or animation / sound effects interference, providing only a lightweight, friendly reminder; medium-intensity prompts correspond to displaying modal pop-ups on the user interface with visual highlighting effects, forcing users to perceive the prompt content before continuing operation; high-intensity prompts correspond to displaying modal pop-ups on the user interface with continuous visual emphasis animations and / or auditory reminder signals, enhancing users' awareness of decision-making risks through multi-sensory reinforcement.

[0105] In this embodiment, the target presentation parameters that simultaneously adapt to the current prompt intensity level and scene type identifier are queried. The target presentation parameters include the display position, display size, background color, border style, animation effects, and / or audio playback method of the prompt information. The preset presentation strategy table includes a two-dimensional mapping relationship, and different scenes paired with different intensity levels can match a unique set of standardized presentation parameters.

[0106] Finally, the metacognitive prompt information is rendered using the target presentation parameters. The rendering process follows the constraints of the target presentation parameters, adapts uniformly to the current user terminal screen specifications and interaction mode, and after rendering verification, the metacognitive prompt information is accurately pushed to the corresponding user terminal for display.

[0107] As can be seen from the above, this embodiment first matches the corresponding basic prompt template according to the scene type identifier, then completes the personalized prompt content by filling the template with user personal information and historical calibration records, divides the prompt intensity into three levels by risk level and matches the corresponding target presentation parameters, and finally renders and pushes metacognitive prompt information according to the target presentation parameters. This achieves triple precision adaptation of prompt content, display style, reminder intensity and usage scenario, user individual characteristics, and risk level, avoiding the problems of weak prompt targeting, insufficient warning intensity or excessive disturbance to users caused by uniform templates and fixed display methods. It further improves the acceptance and importance of metacognitive prompt information to different groups of people, and fully leverages the calibration intervention role of metacognitive prompt information in AI collaborative decision-making.

[0108] Corresponding to the Meta-quotient MQ assessment and metacognitive training method in the above embodiments, Figure 2 This is a structural block diagram of a meta-quotient MQ assessment and metacognitive training system provided in one embodiment of this application. For ease of explanation, only the parts relevant to the embodiment of this application are shown. References Figure 2 The Meta-Commerce MQ Assessment and Metacognitive Training System 20 includes: an interactive acquisition module 21, a feature extraction module 22, an assessment calculation module 23, a profile classification module 24, a game recommendation module 25, a feedback coupling module 26, a decision calibration module 27, and a data storage module 28.

[0109] Among them, the interactive acquisition module 21 is used to respond to the evaluation command and acquire the interactive data with the user terminal; Feature extraction module 22 is used to preprocess the interactive data and extract trial features to obtain trial intermediate features; The assessment calculation module 23 is used to calculate the 6+1-dimensional meta-quotient index based on the intermediate features of the trial. The 6+1-dimensional meta-quotient index includes metacognitive sensitivity, confidence calibration, uncertainty recognition, error detection, strategy control, meta-memory, and meta-time perception. The 6+1-dimensional meta-quotient index is weighted and fused according to the preset non-negative weights to obtain the total meta-quotient score. The user profile classification module 24 is used to classify users based on their total metacognitive score to obtain a user metacognitive profile. The game recommendation module 25 is used to match personalized training game combinations from a preset training game library based on the user's metacognitive profile to obtain personalized training games; and to obtain the user's answer behavior data based on the personalized training games. The feedback coupling module 26 is used to determine the feedback signal based on the answer behavior data, and to perform trial-level coupling between the feedback signal and the metacognitive sub-capabilities corresponding to the user's metacognitive profile to obtain the trial-level feedback result; to update the 6+1-dimensional meta-quotient index based on the trial-level feedback result; and to update the user's metacognitive profile, determine the user level result and achievement data based on the updated 6+1-dimensional meta-quotient index. The decision calibration module 27 is used to calibrate the user's decisions during the AI ​​collaboration process based on the updated user metacognitive profile, and obtain calibration results and user responses. The calibration results include the output records of metacognitive prompts, and the user responses include the user's subsequent decision-making behavior based on the metacognitive prompts. The data storage module 28 is used to store the interaction data, intermediate features of the trial and the 6+1 dimensional meta-quotient index, the trial feedback results and the updated user metacognitive profile, the calibration results and user responses, and the achievement data into the database respectively.

[0110] 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 interactive acquisition module 21, feature extraction module 22, evaluation calculation module 23, profile classification module 24, game recommendation module 25, feedback coupling module 26, decision calibration module 27, and data storage module 28 are shown to have the following functions.

[0111] It should be understood that, in the embodiments of this application, the processor 301 may be a central processing unit (CPU), but it may also 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.

[0112] 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.

[0113] 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.

[0114] 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 meta-quotient MQ assessment and metacognitive training method 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.

[0115] 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.

[0116] 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., equipped 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.

[0117] 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.

[0118] 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.

[0119] 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.

[0120] 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.

[0121] 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.

[0122] 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 meta-quotient MQ assessment and metacognitive training method, characterized in that, include: In response to evaluation commands, acquire interaction data with the user terminal; The interactive data is preprocessed and trial-level features are extracted to obtain trial-level intermediate features; Based on the intermediate features of the trial, a 6+1 dimensional meta-quotient index is calculated; the 6+1 dimensional meta-quotient index includes metacognitive sensitivity, confidence calibration, uncertainty recognition, error detection, strategy control, meta-memory, and meta-time perception. The 6+1 dimensional meta-quotient indicators are weighted and fused according to preset non-negative weights to obtain the total meta-quotient score; Based on the total meta-common score, users are categorized to obtain a user metacognitive profile; Based on the user metacognitive profile, a personalized training game combination is matched from a preset training game library to obtain a personalized training game. Based on the personalized training game, obtain the user's answer behavior data; Based on the response behavior data, a feedback signal is determined, and the feedback signal is coupled with the metacognitive sub-capabilities corresponding to the user metacognitive profile at the trial level to obtain the trial level feedback result; based on the trial level feedback result, the 6+1 dimensional meta-quotient index is updated; and based on the updated 6+1 dimensional meta-quotient index, the user metacognitive profile is updated, the user level result and achievement data are determined. Based on the updated user metacognitive profile, the user's decision-making in the AI ​​collaboration process is calibrated to obtain calibration results and user responses; The interaction data, the intermediate features of the trial, the 6+1 dimensional meta-quotient index, the trial feedback results and the updated user metacognitive profile, the calibration results and user responses, and the achievement data are stored in the database respectively.

2. The meta-quotient MQ assessment and metacognitive training method according to claim 1, characterized in that, The step of responding to the evaluation command and acquiring interaction data with the user terminal includes: In response to the evaluation command, an objective behavior task is selected from the preset objective behavior carrier task library according to the preset extraction rules, and the objective behavior task is pushed to the user terminal for display. The objective behavior carrier task library includes perception judgment task, logical reasoning task, AI suggestion adoption task, memory challenge task, color-word conflict task, visual search task, inhibition control task, and fuzzy uncertainty task. By acquiring the user's response behavior in the objective behavioral task, interactive data is generated. The interactive data includes the user's first correctness of response to the objective behavioral task, first confidence score, first reaction time, help-seeking behavior, answer modification mark, and first error prediction intensity.

3. The meta-quotient MQ assessment and metacognitive training method according to claim 1, characterized in that, The step of matching personalized training game combinations from a preset training game library based on the user metacognitive profile to obtain personalized training games includes: Obtain the type tags of the user metacognitive profile, which are high metacognitive type, overconfident type, uncertainty insensitive type, weak error perception type, rigid strategy type, weak self-monitoring type, and growth type; Based on the type label, query one of the target metacognitive sub-ability dimensions corresponding to the type label from the preset portrait-game mapping rule table; Obtain the metacognitive sub-ability tags of each training game in the training game library, and filter the training game library according to the target metacognitive sub-ability dimension and the metacognitive sub-ability tags of each training game to obtain games to be recommended; Based on the type label, the initial difficulty level, feedback style preference, and risk warning frequency corresponding to the type label are queried from the preset difficulty mapping table. The initial difficulty level includes easy, normal, and hard levels. The feedback style preference includes incentive feedback style and neutral feedback style. The risk warning frequency includes high frequency, medium frequency, and low frequency. The personalized training game is generated based on the game to be recommended, the initial difficulty level, feedback style preferences, and the frequency of risk warnings.

4. The meta-quotient MQ assessment and metacognitive training method according to claim 3, characterized in that, The training game library includes core-layer training games, AI-era extended-layer training games, and scenario-specific-layer training games. The core-layer training games include Confidence Radar Game, Uncertainty Maze Game, Error Detective Game, Strategy Switching Master Game, and Memory Investor Game. The AI-era extended-layer training games include Time Perception Challenge Game, AI Controversial Judge Game, Decision Review Game, Knowledge Blind Spot Radar Game, and Collective Intelligence PK Game. The scenario-specific-layer training games include Telepathy Game, Thinking Adventure RPG Game, and Minnan Dialect Anti-Fraud Master Game. Each training game corresponds to at least one metacognitive sub-capability dimension.

5. The meta-quotient MQ assessment and metacognitive training method according to claim 3, characterized in that, The process of generating the personalized training game based on the game to be recommended, the initial difficulty level, feedback style preferences, and the frequency of risk warnings includes: Based on the type label of the user metacognitive profile and the metacognitive sub-capability label of the game to be recommended, calculate the sub-capability coverage vector of the game to be recommended. Each element in the sub-capability coverage vector represents the training coverage intensity of the game to be recommended on each dimension of the 6+1 dimensional meta-quotient index. Based on the initial difficulty level and the sub-ability coverage vector, an initial configuration parameter set for the game to be recommended is generated. The initial configuration parameter set includes question difficulty parameters, time pressure parameters, and available hint parameters. Based on the feedback style preference and the sub-capability coverage vector of the game to be recommended, the target occurrence probability distribution of each feedback type is determined by the feedback strategy generation function. The feedback strategy generation function determines the overall feedback tendency based on the feedback style preference and adjusts the occurrence probability of each feedback type according to the coverage strength of each dimension in the sub-capability coverage vector, so that the feedback type distribution matches the training target of the game to be recommended. The initial configuration parameter set is associated with and combined with the target occurrence probability distribution of each feedback type to generate the personalized training game.

6. The meta-quotient MQ assessment and metacognitive training method according to claim 5, characterized in that, After generating the personalized training game based on the game to be recommended, the initial difficulty level, feedback style preferences, and risk warning frequency, the method further includes: Obtain the user's second confidence score and second response accuracy in personalized training, and determine whether anchor point trials are needed; If the anchor point trial needs to be inserted, the third confidence score and the third correctness of the response of the anchor point trial are obtained, and the metacognitive heartbeat signal is calculated; the metacognitive heartbeat signal includes the current confidence calibration deviation and the current reaction time offset. The metacognitive heartbeat signal is used as an adjustment parameter and fed back to the difficulty adjustment module of the personalized training game to adjust the question difficulty parameter and the time pressure parameter.

7. The meta-quotient MQ assessment and metacognitive training method according to claim 1, characterized in that, The step of determining a feedback signal based on the response behavior data and coupling the feedback signal with the metacognitive sub-capabilities corresponding to the user's metacognitive profile at the trial level to obtain the trial-level feedback result includes: The feedback type is determined based on the second correctness of the response, the second confidence score, the number of hits, and the intensity of the second error prediction in the response behavior data; the feedback type includes flow feedback type, hit feedback type, perfect calibration feedback type, honesty reward feedback type, self-awareness feedback type, normal correctness feedback type and neutral feedback type; Based on the feedback type, determine the corresponding score, visual presentation method, and sound effect presentation method to obtain the feedback signal; The feedback signal is coupled with the metacognitive sub-capabilities corresponding to the user's metacognitive profile in a trial-level coupling to obtain the trial-level feedback result.

8. The meta-quotient MQ assessment and metacognitive training method according to claim 1, characterized in that, The step of calibrating user decisions during AI collaboration based on the updated user metacognitive profile to obtain calibration results and user responses includes: Obtain users' historical interaction data and updated user metacognitive profiles, and calculate users' overconfidence index in each domain according to preset domain segmentation rules; When a user is detected to be at a preset key node in collaboration with the AI ​​system, the risk level of the current collaboration scenario is determined based on the user metacognitive profile and the overconfidence index. If the risk level is greater than or equal to the preset risk threshold, then based on the risk level and the overconfidence index, a corresponding metacognitive prompt is generated and pushed to the corresponding user terminal for display. The metacognitive prompt is used to assist users in making judgments and decisions during AI collaboration. The output of the metacognitive prompts and the decision-making behavior of the user based on the metacognitive prompts are recorded to obtain the calibration results and user responses.

9. The meta-quotient MQ assessment and metacognitive training method according to claim 1, characterized in that, The database includes four tables: trials, sessions, i_decision_logs, and achievements. The trials table records user interaction data and trial feedback results at the smallest granularity. The sessions table stores complete summary data of evaluation / training. The i_decision_logs table stores calibration results generated during AI collaboration and a full record of user responses to metacognitive prompts. The achievements table stores data related to user unlocking badges, points, levels, and training achievements. The sessions table and trials table, as well as the sessions table and ai_decision_logs table, are all 1:M related. The sessions_id field of the trials table serves as a foreign key pointing to the primary key of the sessions table. The ai_decision_logs table is related to the sessions table and the trials table through the sessions_id. The achievements table and sessions table are user-level related or optional source related, with user_id as the core cross-table related key to link all sessions, trials, logs, and achievements of the same user. The sources_session_id field of the achievements table is nullable and can be used to optionally record the source session of the achievement.

10. The meta-quotient MQ assessment and metacognitive training method according to claim 7, characterized in that, The honesty reward feedback type is used to distinguish it from gamified feedback mechanisms that simply reward correct answers. The reward score is calculated in the format of base + 5·(1-p2) + 2·(e / 100), where base = 0, p is a value less than or equal to 0.3 in the second confidence score, to encourage users to honestly express low confidence rather than blindly high confidence when they are unsure of the answer; e is the intensity of the second error prediction.

11. The meta-quotient MQ assessment and metacognitive training method according to claim 4, characterized in that, The Minnan dialect anti-fraud master game includes setting at least one fraud mode and a method for calculating fraud suspicion score, wherein each fraud mode includes at least one preset keyword; The fraud suspicion score is calculated as follows: likelihood=min(Σ_i(matched_i·weight_i) / 30×100,99) Where, likelihood represents the fraud suspicion score, matched_i represents the number of keywords hit in the i-th fraud pattern; Σ_i (matched_i·weight_i) represents the sum of the number of keyword hits in all fraud patterns × the corresponding weights, i.e., the weighted hit total score; weight_i is the weight coefficient corresponding to the i-th fraud pattern.

12. A meta-quotient MQ assessment and metacognitive training system, characterized in that, include: The interactive data acquisition module is used to respond to evaluation commands and acquire interactive data with the user terminal. The feature extraction module is used to preprocess the interactive data and extract trial-level features to obtain trial-level intermediate features; The assessment calculation module is used to calculate a 6+1 dimensional meta-quotient index based on the intermediate features of the trial; the 6+1 dimensional meta-quotient index includes metacognitive sensitivity, confidence calibration, uncertainty recognition, error detection, strategy control, meta-memory, and meta-time perception. The 6+1 dimensional meta-quotient indicators are weighted and fused according to preset non-negative weights to obtain the total meta-quotient score; The user profile classification module is used to classify users based on their total meta-quotient score to obtain a user metacognitive profile. The game recommendation module is used to match personalized training game combinations from a preset training game library based on the user metacognitive profile to obtain personalized training games. Based on the personalized training game, obtain the user's answer behavior data; The feedback coupling module is used to determine a feedback signal based on the response behavior data, and to perform trial-level coupling between the feedback signal and the metacognitive sub-capabilities corresponding to the user metacognitive profile to obtain trial-level feedback results; to update the 6+1 dimensional meta-quotient index based on the trial-level feedback results; and to update the user metacognitive profile, determine the user level results and achievement data based on the updated 6+1 dimensional meta-quotient index. The decision calibration module is used to calibrate the user's decisions during the AI ​​collaboration process based on the updated user metacognitive profile, and obtain calibration results and user responses. The data storage module is used to store the interaction data, the intermediate features of the trial and the 6+1 dimensional meta-quotient index, the trial feedback results and the updated user metacognitive profile, the calibration results and user responses, and the achievement data into the database respectively.

13. 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 11.

14. A meta-quotient MQ assessment and metacognitive training platform includes at least one processor, a memory, a communication interface, and a data storage module, wherein the memory stores program instructions executable by the processor, and the program instructions, when executed, implement the steps of the method as described in any one of claims 1 to 11.

15. 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 11.

Images