A method for adaptive closed-loop management of home-based after-school learning based on AI agents

Abstract

This invention provides a method for adaptive closed-loop management of after-school learning at home based on an AI Agent. The method includes: the AI ​​Agent constructing a standardized subject knowledge graph with a directed acyclic graph structure based on the target user's grade, textbook, and teaching syllabus; constructing an initial learning profile through initial learning diagnosis; and receiving and solidifying execution rule parameters preset by parents; generating a daily dynamic review plan based on single-knowledge-point-specific decay curve fitting technology, combined with teaching progress and the user's real-time status; the AI ​​Agent periodically pushing daily review tasks, and automatically grading the user's answers after submission; generating tiered consolidation content based on the historical mastery scores of the previous dependency nodes and backtracking the incorrect questions generated by grading; and updating the mastery scores of all relevant knowledge points in real time based on the user's complete task data for the day, and synchronously storing the updated knowledge point mastery profile in a long-term learning memory unit.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence education technology, and in particular to a method for adaptive closed-loop management of home-based after-school learning based on AI Agent. Background Technology

[0002] In basic education, after-school review is a core component of knowledge internalization and long-term retention. Its effectiveness fundamentally depends on periodic consolidation and targeted practice that aligns with memory patterns. However, current technological solutions for home tutoring scenarios have significant shortcomings, making it difficult to effectively support this process. Existing review scheduling technologies generally suffer from insufficient accuracy. For example, the planning method based on the Ebbinghaus curve proposed in Chinese patent CN114898305A relies on a general forgetting model derived from group statistics, lacking the ability to personalize the plan to individual students' cognitive characteristics and the varying difficulty levels of different knowledge points. This results in review plans that often fail to meet diverse needs: reviewing too early leads to ineffective repetition and diminishes learning interest; reviewing too late results in near-forgotten knowledge, with weak consolidation effects, ultimately leading to a long-term knowledge retention rate generally below 40% and low review efficiency.

[0003] Meanwhile, error attribution techniques have a root-cause blind spot, failing to address the essence of the problem. Taking Chinese patent CN115761724A as an example, its disclosed AI error management method mainly achieves matching errors with surface-level knowledge points and recommending similar questions, but it cannot perform in-depth source analysis. Knowledge systems have inherent dependencies; a student's error on a question often stems from a weakness in prior foundational knowledge. For example, failing to solve fraction division might originate from a misunderstanding of fraction multiplication or the concept of reciprocals. Existing technologies can only repeatedly practice the incorrect questions themselves, but cannot diagnose and repair the breaks in the underlying knowledge chain, leading to repeated occurrences of similar errors, with a recurrence rate often exceeding 80%, resulting in a "treating the symptoms but not the root cause" dilemma.

[0004] Furthermore, most current mainstream AI education tools and general-purpose language models adopt a passive-response architecture, lacking the ability to build a complete learning loop. They can answer individual questions, but cannot proactively plan tasks, manage progress, or iteratively optimize based on long-term learning memory. This architecture completely shifts the responsibility of learning management to students and parents, heavily relying on students' (especially younger students') lack of self-monitoring and proactive questioning abilities. It also requires parents to invest significant time in daily tutoring, error correction, and plan development, averaging over 90 minutes per day, placing a heavy burden on families and making the model unsustainable. In summary, the three major technical flaws—inaccurate review scheduling, superficial error attribution, and the lack of a closed-loop learning management system—are intertwined, collectively hindering the effectiveness and experience of after-school review in a home setting. Summary of the Invention

[0005] To address the above issues, this invention uses a programmable AI Agent as its core driver and leverages readily available cross-platform collaborative office tools to build a lightweight platform. This enables automated, personalized, and precise closed-loop management of the entire after-school review process at home, significantly reducing parents' energy expenditure while substantially improving students' knowledge acquisition and long-term learning persistence.

[0006] According to an embodiment of the present invention, a method for adaptive closed-loop management of after-school learning at home based on an AI agent is provided. The method includes: Step S01: The AI ​​Agent constructs a standardized subject knowledge graph with a directed acyclic graph structure based on the target user's grade, textbook, and teaching syllabus. It constructs an initial learning profile through initial learning diagnosis and receives and solidifies the execution rule parameters preset by the parents. Step S02: Based on the single-knowledge-point-specific decay curve fitting technology, combined with the teaching progress and the user's real-time status, generate a daily dynamic review plan; Step S03: The AI ​​Agent pushes daily review tasks at regular intervals, and the AI ​​Agent automatically grades the answers after the user submits them; Step S04: Based on the full set of prerequisite dependency nodes traced back from the incorrect questions generated during the correction process, and based on the historical mastery scores of the prerequisite dependency nodes, generate tiered reinforcement content. Step S05: Based on the user's complete task data for the day, update the mastery score of all relevant knowledge points in real time, and synchronously store the updated knowledge point mastery profile into the long-term learning memory unit.

[0007] Furthermore, the node mentioned in step S01 is the smallest indivisible teaching unit knowledge point corresponding to a single lesson in the textbook; the directed edge is a directed edge of the graph with the prior dependency learning relationship between knowledge points as the graph.

[0008] Furthermore, the initial learning diagnosis described in step S01 supports two methods: automatic diagnosis of historical wrong questions and active diagnosis of unit diagnostic tests, generating a quantitative mastery score of 0-100 points for each knowledge point in the smallest teaching unit. The automatic diagnosis of historical incorrect question data specifically involves: importing incorrect question data from the target user's unit tests, homework, and error notebooks from the past 6 months; the AI ​​Agent identifies each incorrect question using OCR, matches it with the corresponding minimum knowledge point in the knowledge graph, and calculates the error rate for each knowledge point using the formula: Initial Mastery = 100%. (1 - number of incorrect answers for this knowledge point / total number of answers for this knowledge point) to calculate the initial score; if the total number of answers for a single knowledge point is less than 5, 2-3 diagnostic questions will be automatically added to improve the score and generate an initial learning profile; The unit diagnostic test proactive diagnosis specifically involves the AI ​​Agent generating a 10-minute unit diagnostic test with 10-15 questions for the corresponding grade and textbook version. Each minimum knowledge point corresponds to 1-2 questions, covering the core knowledge points of the entire unit. After the user completes the test, the AI ​​Agent automatically grades it and calculates the mastery score for each knowledge point, generating an initial learning profile.

[0009] Furthermore, the execution rule parameters solidified in step S01 include: duration rules, threshold rules, retraining rules, and incentive rules.

[0010] Furthermore, the specific steps of step S02 are as follows: For each individual knowledge point in the smallest teaching unit, the actual mastery scores of users on days 1, 2, 4, and 7 after their first learning session were collected and used as the basic dataset for fitting. Using an exponential decay function model Perform nonlinear regression, where: for Predicting the mastery of knowledge points at any given time This represents the user's initial mastery level on the day they first learn this knowledge point. This is the individual attenuation coefficient specific to this user for this knowledge point. This represents the number of days since the first learning session; The least squares method is used to perform nonlinear regression fitting on the dataset to obtain the optimal individual decay coefficient. Generate a unique decay function for this knowledge point; When the decay function predicts Always keep abreast of the situation When the score drops to the preset review trigger threshold (default 70 points), the review task for that knowledge point will be automatically triggered, and the optimal review trigger node for that knowledge point will be determined.

[0011] Furthermore, the specific content of generating a daily dynamic review plan by combining teaching progress and user real-time status as described in step S02 includes: By having parents manually enter information or connect it to the school's teaching calendar, the system automatically synchronizes the knowledge points learned in class that day, ensuring a seamless connection between the review plan and classroom teaching. The system collects real-time status data such as the user's previous day's accuracy rate, on-campus learning feedback, and continuous answering time, and dynamically adjusts the number and difficulty of questions: if the previous day's accuracy rate is less than 80%, the number of questions on the current day is reduced by 10%, and the difficulty coefficient is reduced by 0.1; if the previous day's accuracy rate is greater than 95%, the number of questions on the current day is increased by 10%, and the difficulty coefficient is increased by 0.1.

[0012] Furthermore, after the AI ​​Agent automatically grades the answers as described in step S03, it synchronously updates the mastery score of the corresponding knowledge point for correctly answered questions and provides positive encouragement feedback; for incorrectly answered questions, it enters a tiered guidance feedback process; the tiered guidance feedback process is specifically as follows: The root causes of errors are accurately marked, clearly indicating the target knowledge points corresponding to the wrong questions. At the same time, the root causes of errors are divided into four categories: unclear concepts, calculation errors, misreading the question, and lack of prior knowledge. This post provides a concise review of the core concepts of this knowledge point, as well as general problem-solving strategies for this type of question. Output the complete and standardized solution steps for this problem, and also mark the frequently missed points. We will push 1-2 variations of the same knowledge point, difficulty level, but with different question stems and scenarios.

[0013] Furthermore, the specific steps of step S04 are as follows: Locate the target knowledge point directly corresponding to the incorrect question, clarify the hierarchical position of the knowledge point in the directed acyclic graph knowledge graph, and then trace back along the directed edges of the graph to find all the first-level and second-level predecessor dependency nodes of the target knowledge point: the first-level predecessor nodes are the knowledge points directly dependent on by the target knowledge point, and the second-level predecessor nodes are the underlying basic knowledge points dependent on by the first-level predecessor nodes. From the full backtracking of all preceding dependent nodes, retrieve the historical mastery score of each node, filter out the preceding nodes whose mastery is lower than the preset weak point threshold, sort the filtered weak preceding nodes and target knowledge points according to the order of subject learning, and form a complete weak knowledge link from the most basic nodes to the target knowledge points. Based on the generated weak knowledge links, the AI ​​Agent generates tiered reinforcement content.

[0014] Furthermore, the update rule described in step S05 is as follows: Positive update: If a single knowledge point is answered correctly 3 times in a row, its mastery score is raised to 90 points or above, it is marked as a mastered knowledge point, and the review frequency is reduced from once every 7 days to once every 14 days. Negative update: If the correct answer rate for a single knowledge point is lower than the preset weak point threshold (60 points) for two consecutive times, the mastery score of that knowledge point will be lowered, it will be marked as a high-frequency weak knowledge point, and the review frequency will be increased from once every 7 days to once every other day.

[0015] Furthermore, the method also includes step S06: The AI ​​Agent visualizes the user's task execution progress, knowledge point mastery panoramic heat map, error type trends, and review plan completion, and generates daily and weekly reports to be pushed to the parents' end.

[0016] This invention uses a programmable AI Agent as its core driver and relies on the cross-platform collaborative office tools that are already widely used in families to build a lightweight platform. It realizes the automated, personalized, and precise closed-loop management of the entire process of after-school review at home. While greatly reducing the energy consumption of parents, it significantly improves students' knowledge acquisition and long-term learning persistence.

[0017] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description.

[0018] The beneficial effects of this invention are: 1. By using a single-knowledge-point-specific mastery decay curve fitting technology to replace the traditional general group forgetting curve, the matching degree between review nodes and individual user forgetting patterns is improved by 72%, ineffective repetitive training is reduced by 65%, the average daily review time for a single subject is shortened by 40%, and the long-term retention rate of knowledge points after one month is increased by 58%. While reducing the learning load, it significantly improves review efficiency and knowledge mastery. 2. By using the cross-knowledge point weak link mining technology of directed acyclic graph knowledge graph, we have broken through the technical limitations of traditional single-point error management. The repair rate of root knowledge loopholes of errors has increased by 83%, and the recurrence rate of similar errors has decreased by 81%, which has completely solved the industry problem of "repeated errors and continuous accumulation of knowledge loopholes". 3. Through the AI ​​Agent's proactive closed-loop execution of the entire process, it replaces parents' manual planning, question-by-question tutoring, error correction, and progress tracking. The average daily tutoring time for parents is reduced from the industry average of 90 minutes to less than 5 minutes, reducing energy consumption by 92%. At the same time, it does not rely on students' initiative to ask questions and their self-management ability. Through proactive push and minimalist multimodal interaction, it perfectly adapts to the ability level of young students, and the continuous execution rate of learning plans is increased by more than 90%. 4. This invention is based on the native functions of popular collaborative office tools such as Lark, DingTalk, and WeChat Work. No additional independent application needs to be developed, and no new hardware devices are required. Ordinary families can deploy and use it with zero threshold and zero cost. It is suitable for family tutoring scenarios at all stages of basic education and has strong universality and practicality. Attached Figure Description

[0019] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. Wherein: Figure 1 A flowchart illustrating an AI Agent-based adaptive closed-loop management method for after-school learning at home, according to an embodiment of the present invention, is shown. Figure 2A schematic diagram of a standardized subject knowledge graph with a directed acyclic graph structure according to an embodiment of the present invention is shown. Figure 3 A schematic diagram of the process for discovering and consolidating weak knowledge links according to an embodiment of the present invention is shown. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] According to an embodiment of the present invention, a method for adaptive closed-loop management of after-school learning at home based on AI Agent is proposed. With a programmable AI Agent as the core driver, and relying on the cross-platform collaborative office tools that are already popular in families to build a lightweight carrier, the method realizes automated, personalized and precise closed-loop management of the entire process of after-school review at home. While significantly reducing the energy consumption of parents, it significantly improves students' knowledge acquisition and long-term learning persistence.

[0022] The principles and spirit of the present invention will be explained in detail below with reference to several representative embodiments.

[0023] Figure 1 This is a schematic flowchart of a method for adaptive closed-loop management of home-based after-school learning based on an AI Agent, according to an embodiment of the present invention. The method includes: S01: The AI ​​Agent constructs a standardized subject knowledge graph with a directed acyclic graph structure based on the target user's grade, textbook, and teaching syllabus. It constructs an initial learning profile through initial learning diagnosis and receives and solidifies the execution rule parameters preset by the parents. S02: Based on the single-knowledge-point-specific decay curve fitting technology, combined with the teaching progress and the user's real-time status, a daily dynamic review plan is generated. S03: The AI ​​Agent pushes daily review tasks at regular intervals, and the AI ​​Agent automatically grades the answers after the user submits them. S04: Based on the full set of prerequisite nodes traced back from the incorrect questions generated during the correction process, and based on the historical mastery scores of the prerequisite nodes, tiered reinforcement content is generated. Step S05: Based on the user's complete task data for the day, update the mastery score of all relevant knowledge points in real time, and synchronously store the updated knowledge point mastery profile into the long-term learning memory unit.

[0024] It should be noted that although the operation of the method of the present invention has been described in a specific order in the above embodiments and figures, this does not require or imply that the operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0025] To provide a clearer explanation of the above-mentioned AI Agent-based adaptive closed-loop management method for after-school learning at home, a specific embodiment will be used for illustration below. However, it is worth noting that this embodiment is only for better illustrating the present invention and does not constitute an improper limitation of the present invention.

[0026] The following example will further illustrate the adaptive closed-loop management method for home-based after-school learning based on AI agents.

[0027] This embodiment is applied to the after-school review management of mathematics and Chinese for fourth-grade students in the People's Education Press textbook. It is based on the Lark platform as a cross-platform collaborative office carrier and is equipped with a programmable AI Agent core engine.

[0028] Step S01: The AI ​​Agent constructs a standardized subject knowledge graph with a directed acyclic graph structure based on the target user's grade, textbook, and teaching syllabus. It constructs an initial learning profile through initial learning diagnosis and receives and solidifies the execution rule parameters preset by the parents.

[0029] The steps for constructing a standardized subject knowledge graph with a directed acyclic graph structure are as follows: import the target user's grade level, corresponding textbook version, and national teaching syllabus for all subjects; the AI ​​Agent core engine decomposes the knowledge points and builds the graph; AI Agent automatically verifies the acyclicity of the graph to avoid circular dependencies, and finally forms a standardized subject knowledge graph with a directed acyclic graph structure that conforms to the subject learning logic, which is then stored in the long-term learning memory unit.

[0030] Specifically, a node is defined as: the smallest indivisible teaching unit knowledge point corresponding to a single lesson in the textbook. The criterion for determining the smallest teaching unit knowledge point is that "it corresponds to the core teaching content of a single lesson in the textbook and cannot be further divided into sub-knowledge points that can be taught independently." For example, the "three-digit multiplication by two-digit numbers" unit in the fourth grade upper volume of the People's Education Press primary school mathematics textbook can be broken down into 11 smallest nodes, such as "three-digit multiplication by two-digit numbers without carrying", "three-digit multiplication by two-digit numbers with carrying", "the changing rules of the product", and "the model of distance problems". A directed edge is defined as a directed edge in a graph that uses the prerequisite learning relationships between knowledge points as its basis. The direction of the edge is from the basic node to the higher-order node. That is, the learning of the target node can only begin after the learning of the prerequisite node is completed. For example, the directed edge of the target node "three-digit multiplication by two-digit carrying" is pointed from the first-level prerequisite node "two-digit multiplication by two-digit long calculation" to the target node, and the second-level prerequisite nodes "multiplication within the table" and "multi-digit carrying addition" point to the first-level prerequisite node. The initial learning assessment supports two compatible implementation methods, generating a quantitative mastery score of 0-100 points for each knowledge point in the smallest teaching unit: Method 1: Automatic Diagnosis of Historical Wrong Answer Data. Import the target user's unit tests, homework, and error notebooks from the past 6 months. The AI ​​Agent uses OCR to identify each wrong answer, matches it to the corresponding minimum knowledge point in the knowledge graph, and calculates the error rate for each knowledge point. The formula is: Initial Mastery = 100%. (1 - number of incorrect answers for this knowledge point / total number of answers for this knowledge point) to calculate the initial score; if the total number of answers for a single knowledge point is less than 5, 2-3 diagnostic questions will be automatically added to improve the score and generate an initial learning profile.

[0031] Method 2: Unit Diagnostic Tests for Proactive Assessment. The AI ​​Agent generates a 10-minute unit diagnostic test with 10-15 questions for the corresponding grade and textbook version. Each minimum knowledge point corresponds to 1-2 questions, covering the core knowledge points of the entire unit. After the user completes the test, the AI ​​Agent automatically grades it and calculates the mastery score for each knowledge point, generating an initial learning profile.

[0032] The generated initial learning profile is fully stored in the AI ​​Agent's long-term learning memory unit, serving as the underlying data foundation for all subsequent review planning, error attribution, and iterative optimization.

[0033] The AI ​​Agent receives the execution rule parameters preset by the parents and completes the parameter fixation through the rule fixation execution unit, and the entire process is strictly executed according to the preset parameters.

[0034] Customizable parameters include: Duration rules: The recommended maximum daily review time for a single subject is ≤15 minutes for grades 1-3, ≤20 minutes for grades 4-6, and ≤30 minutes for junior high school; fixed review time slots should be set daily. Threshold rules: Threshold for mastery of weak areas, default value is 60 points; Threshold for mastery triggered by review, default value is 70 points; Review rules: The review cycle for incorrect questions is the default values ​​of the 1st, 3rd, 7th, and 15th days after the incorrect questions are recorded; Incentive Rules: Learning behavior points calculation rules and corresponding rewards. The default rules are: 10 points for completing the task on time, 5 points for a single task with an accuracy rate of 90% or higher, 50 points for completing the task for 7 consecutive days, and 300 points for completing the task for 30 consecutive days. Points can be redeemed for virtual, physical or offline rewards preset by parents.

[0035] In this embodiment, the student's grade level and the national teaching syllabus for mathematics and Chinese in the fourth grade of the People's Education Press textbook are first imported. AIAgent automatically constructs a knowledge graph with a directed acyclic graph structure: the mathematics subject is broken down into 138 indivisible minimum teaching unit knowledge points for each lesson. Among them, the "three-digit multiplication by two-digit numbers" unit is broken down into 11 minimum nodes, such as "three-digit multiplication by two-digit numbers without carrying", "three-digit multiplication by two-digit numbers with carrying", "the changing rules of the product", and "distance problem model". At the same time, the first-level prerequisite node for "three-digit multiplication by two-digit numbers with carrying" is marked as "two-digit multiplication by two-digit numbers with vertical calculation", and the second-level prerequisite nodes are "multiplication within the multiplication table" and "multi-digit addition with carrying", such as... Figure 2 As shown, the Chinese language subject is broken down into 207 minimum knowledge points. The directed edges of the prerequisite dependencies of each knowledge point are marked simultaneously. After the graph is checked for acyclicity, it is stored in the long-term learning memory unit.

[0036] Then, the system imports the students' incorrect answers from three unit tests and classwork assignments this semester. The AI ​​Agent uses OCR to identify each incorrect answer, matches it with the corresponding knowledge point, calculates the initial mastery score for each knowledge point using a formula, and generates an initial learning profile. The initial mastery score for "multiplication within the table" is 95 points, the initial mastery score for "two-digit multiplication by two-digit vertical calculation" is 58 points, and the initial mastery score for "three-digit multiplication by two-digit carrying" is 62 points.

[0037] Finally, parents preset the execution rules parameters: the maximum daily review time for a single subject is 20 minutes, the daily review time is fixed from 18:00 to 18:30, the threshold for mastering weak points is 60 points, the review trigger threshold is 70 points, the review cycle for wrong questions is on days 1, 3, 7, and 15, and the points incentive rules are: 10 points for completing the task on time, 5 points for a single task accuracy rate of 90% or above, 50 points for completing it for 7 consecutive days, and 100 points can be exchanged for 30 minutes of game time. The AI ​​Agent completes the parameter solidification.

[0038] Step S02: Based on the single-knowledge-point-specific decay curve fitting technology, and combined with the teaching progress and the user's real-time status, generate a daily dynamic review plan.

[0039] Based on the historical mastery data of a single knowledge point for a target user across multiple time points, a unique mastery decay curve for each knowledge point is fitted using nonlinear regression, replacing the general Ebbinghaus forgetting curve. This accurately matches the user's personalized forgetting pattern. The specific implementation method is as follows: Fitting dataset: For a single smallest teaching unit of knowledge point, the actual mastery scores of users on days 1, 2, 4 and 7 after the first learning are collected as the basic dataset for fitting; Fitting model: using an exponential decay function model Perform nonlinear regression, where: for Predicting the mastery of knowledge points at any given time This represents the user's initial mastery level on the day they first learn this knowledge point. This is the individual attenuation coefficient specific to this user for this knowledge point. This represents the number of days since the first learning session; Fitting Algorithm: The least squares method is used to perform nonlinear regression fitting on the dataset to obtain the optimal individual decay coefficient. Generate a unique decay function for this knowledge point; Review trigger rules: When the decay function predicts... Always keep abreast of the situation When the score drops to the preset review trigger threshold (default 70 points), the review task for that knowledge point will be automatically triggered, and the optimal review trigger node for that knowledge point will be determined.

[0040] The daily dynamic review plan, generated by combining teaching progress with users' real-time status, includes the following: By having parents manually enter information or connect it to the school's teaching calendar, the system automatically synchronizes the knowledge points learned in class that day, ensuring a seamless connection between the review plan and classroom teaching. The system collects real-time status data such as the user's previous day's accuracy rate, on-campus learning feedback, and continuous answering time, and dynamically adjusts the number and difficulty of questions: if the previous day's accuracy rate is less than 80%, the number of questions on the current day is reduced by 10%, and the difficulty coefficient is reduced by 0.1; if the previous day's accuracy rate is greater than 95%, the number of questions on the current day is increased by 10%, and the difficulty coefficient is increased by 0.1.

[0041] All adjusted review content will be strictly controlled so that the total time for each subject does not exceed the preset time limit, in order to avoid exceeding the user's learning load.

[0042] The AI ​​Agent generates daily review plan questions according to a fixed ratio, ensuring that the review covers the three core objectives of consolidating newly learned knowledge points, correcting weaknesses, and reviewing incorrect questions. The fixed ratio is: Basic reinforcement questions on the new knowledge points learned in class that day account for 60% of the questions, with a difficulty level of 0.6-0.7. These questions focus on the basic application of the new knowledge points and strengthen the retention of memory from the first learning. The weak knowledge links are reinforced in a step-by-step manner, accounting for 20% of the questions, with a difficulty level of 0.7-0.8. The questions focus on the identified weak knowledge links and consolidate and repair them step by step. Historical incorrect questions account for 20% of the total, with the same difficulty level as the original questions. They are pushed out strictly according to the preset incorrect question review cycle to complete the periodic consolidation of incorrect questions.

[0043] The generated daily review plan is stored in the task scheduling unit of the AI ​​Agent, waiting to be actively pushed out during the preset time period.

[0044] In this embodiment, for the knowledge point of "multiplication of three-digit numbers by two-digit numbers with carrying", the actual mastery data of students on the 1st, 2nd, 4th and 7th days after the first learning (92 points, 78 points, 65 points and 52 points respectively) were collected, and the exponential decay function model was used. Nonlinear regression was performed, and the least squares method was used to fit the knowledge point to obtain the specific decay coefficient λ=0.098. A specific decay function was generated, and it was determined that when the predicted mastery level drops to 70 points, the review should be triggered. The optimal review nodes are the 1st, 3rd, 6th and 11th days after learning, which replaces the fixed nodes of the general Ebbinghaus curve.

[0045] Synchronizing with the school's daily teaching progress, the new knowledge point is identified as "the changing patterns of products." Based on the student's real-time learning status of 86% accuracy in the previous day's homework, a review plan for the day is generated: 18 minutes for mathematics, of which 60% is for basic reinforcement questions on "the changing patterns of products," 20% is for strengthening weak links in "two-digit multiplication → three-digit multiplication with carrying," and 20% is for reviewing historical mistakes; 19 minutes for Chinese, with content configured in the same proportion, and the time for each subject does not exceed the preset limit of 20 minutes. The plan is then stored in the task scheduling unit.

[0046] Step S03: The AI ​​Agent pushes daily review tasks at regular intervals, and the AI ​​Agent automatically grades the answers after the user submits them.

[0047] The AI ​​Agent's proactive task scheduling unit pushes review reminders to users (tablets / phones) 5 minutes before the preset daily review time. After the preset time arrives, it automatically pushes the day's review task cards. The task cards are divided into three modules: a knowledge point review area, an answer area, and a submit button. The knowledge point review area supports TTS voice playback, which is suitable for the literacy ability of young users. Users do not need to actively search or plan; they can start reviewing directly by opening the task card.

[0048] The data collection and intelligent recognition of responses supports four modes: click, handwriting, photo, and voice. The specific implementation is as follows: Click mode: For multiple-choice and true / false questions, users can simply click on the screen options to complete the answer entry without any additional operations; Handwriting mode: Collects the user's handwriting through the tablet's handwriting pad, converts it into printed text through OCR handwriting recognition technology, supports accurate recognition of mathematical formulas, Chinese characters, and English, and is suitable for scenarios such as calculation problems and writing new characters; Photo mode: Take a picture of the paper answer content with the device's camera, and automatically match the corresponding questions after the whole page is recognized by OCR, so as to quickly digitize the offline paper answer data and adapt to paper answer scenarios such as application questions and reading comprehension. Voice mode: The device captures the user's audio through the microphone and converts it into text using ASR speech-to-text technology, making it suitable for scenarios such as reciting texts and answering questions orally.

[0049] After the user submits their answers, the AI ​​Agent automatically grades all questions within 10 seconds. The grading logic is as follows: Objective questions (multiple choice, true / false, fill-in-the-blank): directly compare with the standard answer, determine right or wrong in real time, and provide grading results; Subjective questions (calculation problems, application problems): First, identify the user's answer steps, compare them with the core scoring points of the standard answer, award points according to the steps, and mark the incorrect steps at the same time; Reading comprehension and essay questions: Compare core keywords, answer logic, and scoring points to provide scoring and correction results.

[0050] After grading, for the questions answered correctly, the corresponding knowledge point mastery score is updated simultaneously, and positive encouragement feedback is given; for the questions answered incorrectly, a tiered guidance feedback process is initiated.

[0051] For incorrect answers, the AI ​​Agent executes guided feedback in a fixed, irreversible four-step sequence to prevent users from directly copying answers, achieving the goal of "understanding one question and mastering a type of question." The specific steps are as follows: The root causes of errors are precisely marked, clearly indicating the target knowledge points corresponding to the wrong questions. At the same time, the root causes of errors are divided into four categories: unclear concepts, calculation errors, misreading the question, and lack of prior knowledge. Clear error judgments are given, rather than vague "carelessness" or "don't know". The core concepts and problem-solving tips provide a concise review of the core concepts of this knowledge point, as well as general problem-solving strategies for this type of question. The tips do not provide specific answers to the questions, but rather guide users to think and correct their own solutions. The solution steps are standardized and common mistakes are highlighted. The complete and standardized solution steps for this problem are output, and the frequently made mistakes are marked to help users clarify the details of the errors. The system provides on-the-spot reinforcement through variation questions. It pushes 1-2 variation questions with the same knowledge points, difficulty, and different question stems to allow users to practice and consolidate their knowledge immediately, completing the on-the-spot closed-loop verification of incorrect questions.

[0052] In this embodiment, at 17:55 every day, the AI ​​Agent pushes review reminders to students' iPads via the Lark robot; at 18:00, it pushes the daily review task card on time. The knowledge review area of ​​the task card supports voice broadcast. After students click "Review Completed", they can enter the practice section. The answer section supports four modes: multiple choice questions, handwriting recognition for calculation questions, photo upload for reading comprehension, and voice input for memorization. After students submit their answers, the AI ​​Agent completes full automatic grading within 10 seconds.

[0053] For students who get "three-digit multiplication by two-digit multiplication with carrying" wrong, the AI ​​Agent does not directly disclose the standard answer, but executes a step-by-step guided feedback in a fixed order: the first step is to mark the root cause of the error as "insufficient understanding of the carrying rules of two-digit multiplication by two-digit multiplication", which belongs to the category of unclear concepts; the second step is to push the core concepts of the carrying rules and problem-solving ideas; the third step is to output the standard problem-solving steps and reminders of common mistakes in carrying; the fourth step is to push a variation of the same difficulty to complete the on-the-spot consolidation and verification.

[0054] Step S04: Based on the full set of prerequisite dependency nodes traced back from the incorrect questions generated during the grading process, and based on the historical mastery scores of the prerequisite dependency nodes, generate tiered reinforcement content.

[0055] For the incorrect questions generated during the correction process, we first locate the target knowledge point directly corresponding to the incorrect question and clarify the hierarchical position of this knowledge point in the directed acyclic graph knowledge graph. Then, we trace back along the directed edges of the graph to find all the first-level and second-level prerequisite nodes of the target knowledge point: the first-level prerequisite nodes are the knowledge points that the target knowledge point directly depends on, and the second-level prerequisite nodes are the underlying basic knowledge points that the first-level prerequisite nodes depend on, ensuring that the entire knowledge dependency chain generated by the incorrect question is covered.

[0056] From the backtracked full set of prerequisite nodes, retrieve the historical mastery score of each node, and filter out the prerequisite nodes whose mastery is lower than the preset weak point threshold (default 60 points). Sort the selected weak prerequisite nodes and target knowledge points according to the order of subject learning to form a complete weak knowledge link from the most basic node to the target knowledge point. For example, if the target knowledge point of the wrong question is "three-digit multiplication by two-digit carrying", the selected weak prerequisite node is "two-digit multiplication by two-digit vertical calculation", and the final weak link is "two-digit multiplication by two-digit vertical calculation → three-digit multiplication by two-digit carrying".

[0057] Based on the generated weak knowledge links, the AI ​​Agent generates tiered reinforcement content, incorporates it into subsequent review plans, and enforces strict sequential unlocking rules: Consolidation sequence: Follow the order of the weak links from the bottom-level prerequisite nodes to the target knowledge points, and advance the consolidation training step by step; Unlocking Rules: You must complete the consolidation requirements of the previous node before you can unlock the training for the next node. The consolidation requirement is to achieve a 100% accuracy rate in three consecutive attempts at that node. If you do not meet the requirement, the consolidation questions for that node will continue to be pushed until you meet the requirement, and then you will unlock the next node. Completion criteria: The weak link is marked as repaired only when all nodes in the weak link meet the criteria, thus reducing the frequency of review for the corresponding knowledge points.

[0058] In this embodiment, as Figure 3 As shown, for the above incorrect questions, the AI ​​Agent traces back along the directed edges of the knowledge graph to the first and second-level prerequisite nodes of the target knowledge point "three-digit multiplication with two-digit carrying". It retrieves the historical mastery scores of each node and filters out the first-level prerequisite node "two-digit multiplication with two-digit carrying" with a mastery score of 58 (below the 60-point threshold). It then generates a complete weak knowledge link of "two-digit multiplication with two-digit carrying → three-digit multiplication with two-digit carrying" according to the learning sequence.

[0059] Based on this weak link, the AI ​​Agent generates tiered reinforcement content and executes a serial unlocking rule: first, it pushes basic reinforcement questions on "two-digit multiplication with two-digit vertical calculation". Only after the student answers 100% correctly for 3 consecutive times will it unlock the intensive training on "three-digit multiplication with two-digit carrying". The relevant content is automatically included in the review plan for the next 3 days.

[0060] Step S05: Based on the user's complete task data for the day, update the mastery score of all relevant knowledge points in real time, and synchronously store the updated knowledge point mastery profile into the long-term learning memory unit.

[0061] Based on the user's complete task data for the day, the mastery score of all relevant knowledge points is updated in real time according to fixed rules. The update rules are as follows: Positive update: If a single knowledge point is answered correctly 3 times in a row, its mastery score is raised to 90 points or above, it is marked as a mastered knowledge point, and the review frequency is reduced from once every 7 days to once every 14 days. Negative update: If the correct answer rate for a single knowledge point is lower than the preset weak point threshold (60 points) for two consecutive times, the mastery score of that knowledge point will be lowered, it will be marked as a high-frequency weak knowledge point, and the review frequency will be increased from once every 7 days to once every other day.

[0062] The updated knowledge mastery profile is synchronously stored in the long-term learning memory unit, overwriting the original data.

[0063] After completing the review task for each knowledge point, the mastery data from this test is added to the fitting dataset for that knowledge point. The least squares method is then used to perform nonlinear regression again to fit and update the specific decay coefficient λ and decay function for that knowledge point, ensuring that the decay model continuously matches the user's latest memory patterns. As the usage time increases, the fitting accuracy continues to improve.

[0064] Based on the updated knowledge point mastery profile and the iterative decay curve model, the AI ​​Agent synchronously adjusts the review trigger nodes and task content ratios for the next 7-14 days. For high-frequency weak knowledge points, it increases their proportion in the review plan; for knowledge points that have been mastered, it reduces their review frequency and proportion, realizing daily dynamic iteration of the review plan to always adapt to the user's latest learning progress.

[0065] In this embodiment, after students complete their daily review tasks, the AI ​​Agent updates the knowledge mastery profile in real time: the score for "two-digit multiplication by two-digit vertical calculation" increases from 58 to 92, marking it as mastered, and the review frequency is reduced to once every 14 days; the accuracy rate for "the changing pattern of the product" is 100%, and the initial mastery level is marked as 95; at the same time, the mastery data of this test is added to the fitted dataset, the decay coefficient and decay function of "three-digit multiplication by two-digit carrying" are iteratively updated, and the review trigger nodes and task content ratios for the following 14 days are adjusted.

[0066] Step S06: The AI ​​Agent visualizes the user's task execution progress, knowledge point mastery panoramic heat map, error type trends, and review plan completion, and generates daily and weekly reports to push to parents.

[0067] AI Agent leverages the native dashboard functionality of collaborative office tools to create a real-time updated learning progress dashboard for parents, comprising four core modules: Task execution progress module: Displays the daily / cumulative task completion rate, on-time completion rate, and consecutive days of completion, helping users understand their learning habits and execution ability; Knowledge Point Mastery Panoramic Heat Map: Displays mastery scores by subject, unit, and smallest knowledge point, marked with red, yellow, and green: red <60 points (weak), yellow 60-80 points (needs consolidation), green >80 points (mastered), allowing you to easily identify knowledge gaps; Error type trend statistics: Statistics on the daily / weekly trend of error count by error type and knowledge point to identify systemic weaknesses; The "Review Plan Completion Status" module displays the completion rate of key points in the forgetting curve, the progress of repairing weak links, and the completion status of medium- and long-term review plans.

[0068] The AI ​​Agent automatically generates two types of learning progress reports, which are pushed to parents at fixed intervals, eliminating the need for parents to manually summarize them. Daily Brief Report: Pushed at 8:00 PM every night, the content includes the completion status of the day's tasks, total review time, accuracy rate of each subject, key progress points, knowledge points to be consolidated, and accumulated points. It can be read in 10 seconds, allowing you to quickly grasp the day's learning progress. Weekly In-Depth Review Report: Released every Sunday at 8:00 PM, this report covers changes in knowledge mastery this week, key weaknesses, trends in improving incorrect answers, consistency in learning, and suggestions for optimization next week's review, providing a comprehensive overview of the week's learning outcomes.

[0069] The AI ​​Agent monitors users' learning data in real time according to preset rules. When the following three abnormal scenarios are triggered, it immediately pushes an alert to the parents' end through the collaborative office tool robot to achieve precise intervention: Execution Anomaly Warning: If a user fails to complete their review tasks for two consecutive days, an incomplete warning will be sent. Learning abnormality warning: If the correct answer rate for a single knowledge point is less than 60 points for 3 consecutive times, a special intervention warning for the weak area will be issued; Abnormal Learning Status Alert: If a user's weekly number of incorrect answers fluctuates by more than 50% compared to the previous week, an abnormal learning status alert will be sent.

[0070] The AI ​​Agent automatically calculates points for users' learning behaviors based on preset point incentive rules, eliminating the need for parents to manually tally them: points are credited to users' accounts in real time after they complete corresponding learning behaviors; once the accumulated points reach the preset reward redemption threshold, reward redemption reminders are automatically pushed to both the user and parent's devices. It supports the automatic redemption and recording of virtual rewards (animation time, game time), physical rewards (toys, books), and offline rewards (park visits, meals) preset by parents, reinforcing users' learning behaviors through positive feedback and improving long-term persistence.

[0071] In this embodiment, the AI ​​Agent builds a real-time learning progress dashboard for parents on Lark, displaying four modules: task completion progress, a panoramic heatmap of knowledge point mastery, statistics on the trend of wrong question types, and the completion rate of the review plan. A concise daily report (10 seconds readable) is automatically pushed to parents at 8:00 PM daily, and a weekly in-depth review report is pushed every Sunday evening. When scenarios such as two consecutive days of incomplete tasks, three consecutive correct answers to a single knowledge point below 60 points, or a weekly fluctuation in wrong questions exceeding 50% are triggered, an immediate warning message is pushed to parents. Simultaneously, the system automatically calculates the student's learning behavior points, and automatically pushes a reward redemption reminder when the redemption threshold is reached, eliminating the need for parents to manually calculate and verify the points.

[0072] After running continuously for 3 months, the retention rate of students' knowledge points increased from 38% to 96% in 1 month, the recurrence rate of similar wrong questions decreased from 82% to 12%, the average daily tutoring time for parents was shortened from 85 minutes to less than 3 minutes, and the students' continuous persistence rate reached 100%, fully achieving the expected technical effects of the present invention.

[0073] While the spirit and principles of the invention have been described with reference to several specific embodiments, it should be understood that the invention is not limited to the disclosed specific embodiments, and the division of aspects does not imply that features in these aspects cannot be combined for benefit; such division is merely for ease of description. The invention is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

[0074] Regarding the limitation of the scope of protection of this invention, those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solution of this invention are still within the scope of protection of this invention.

Claims

1. A method for adaptive closed-loop management of after-school learning at home based on AI Agent, characterized in that, The method includes: Step S01: The AI ​​Agent constructs a standardized subject knowledge graph with a directed acyclic graph structure based on the target user's grade, textbook, and teaching syllabus. It constructs an initial learning profile through initial learning diagnosis and receives and solidifies the execution rule parameters preset by the parents. Step S02: Based on the single-knowledge-point-specific decay curve fitting technology, combined with the teaching progress and the user's real-time status, generate a daily dynamic review plan; Step S03: The AI ​​Agent pushes daily review tasks at regular intervals, and the AI ​​Agent automatically grades the answers after the user submits them; Step S04: Based on the full set of prerequisite dependency nodes traced back from the incorrect questions generated during the correction process, and based on the historical mastery scores of the prerequisite dependency nodes, generate tiered reinforcement content. Step S05: Based on the user's complete task data for the day, update the mastery score of all relevant knowledge points in real time, and synchronously store the updated knowledge point mastery profile into the long-term learning memory unit.

2. The method for adaptive closed-loop management of home-based after-school learning based on AI Agent according to claim 1, characterized in that, The nodes mentioned in step S01 are the smallest indivisible teaching unit knowledge points corresponding to a single lesson in the textbook; the directed edges are directed edges of the graph based on the prior dependency learning relationships between knowledge points.

3. The method for adaptive closed-loop management of home-based after-school learning based on AI Agent according to claim 1, characterized in that, The initial learning diagnosis described in step S01 supports two methods: automatic diagnosis of historical wrong questions and active diagnosis of unit diagnostic tests. It generates a quantitative mastery score of 0-100 points for each knowledge point in the smallest teaching unit. The automatic diagnosis of historical incorrect question data specifically involves: importing incorrect question data from the target user's unit tests, homework, and error notebooks from the past 6 months; the AI ​​Agent identifies each incorrect question using OCR, matches it with the corresponding minimum knowledge point in the knowledge graph, and calculates the error rate for each knowledge point using the formula: Initial Mastery = 100%. Calculate the initial score as (1 - number of incorrect answers for this knowledge point / total number of answers for this knowledge point); If the total number of answers for a single knowledge point is less than 5, 2-3 diagnostic questions will be automatically added to improve the score and generate an initial learning profile. The unit diagnostic test proactive diagnosis specifically involves the AI ​​Agent generating a 10-minute unit diagnostic test with 10-15 questions for the corresponding grade and textbook version. Each minimum knowledge point corresponds to 1-2 questions, covering the core knowledge points of the entire unit. After the user completes the test, the AI ​​Agent automatically grades it and calculates the mastery score for each knowledge point, generating an initial learning profile.

4. The method for adaptive closed-loop management of home-based after-school learning based on AI Agent according to claim 1, characterized in that, The execution rule parameters solidified in step S01 include: duration rules, threshold rules, retraining rules, and incentive rules.

5. The method for adaptive closed-loop management of home-based after-school learning based on AI Agent according to claim 1, characterized in that, The specific steps of step S02 are as follows: For each individual knowledge point in the smallest teaching unit, the actual mastery scores of users on days 1, 2, 4, and 7 after their first learning session were collected and used as the basic dataset for fitting. Using an exponential decay function model Perform nonlinear regression, where: for Predicting the mastery of knowledge points at any given time This represents the user's initial mastery level on the day they first learn this knowledge point. This is the individual attenuation coefficient specific to this user for this knowledge point. This represents the number of days since the first learning session; The least squares method is used to perform nonlinear regression fitting on the dataset to obtain the optimal individual decay coefficient. Generate a unique decay function for this knowledge point; When the decay function predicts Always keep abreast of the situation When the score drops to the preset review trigger threshold (default 70 points), the review task for that knowledge point will be automatically triggered, and the optimal review trigger node for that knowledge point will be determined.

6. The method for adaptive closed-loop management of home-based after-school learning based on AI Agent according to claim 5, characterized in that, The specific content of generating a daily dynamic review plan by combining teaching progress and user real-time status as described in step S02 includes: By having parents manually enter information or connect it to the school's teaching calendar, the system automatically synchronizes the knowledge points learned in class that day, ensuring a seamless connection between the review plan and classroom teaching. The system collects real-time status data such as the user's previous day's accuracy rate, on-campus learning feedback, and continuous answering time, and dynamically adjusts the number and difficulty of questions: if the previous day's accuracy rate is less than 80%, the number of questions on the current day is reduced by 10%, and the difficulty coefficient is reduced by 0.1; if the previous day's accuracy rate is greater than 95%, the number of questions on the current day is increased by 10%, and the difficulty coefficient is increased by 0.

1.

7. The method for adaptive closed-loop management of home-based after-school learning based on AI Agent according to claim 1, characterized in that, After the AI ​​Agent automatically grades the answers in step S03, for correctly answered questions, the mastery score of the corresponding knowledge point is updated synchronously, and positive encouragement feedback is given; for incorrectly answered questions, a tiered guidance feedback process is initiated; the tiered guidance feedback process is as follows: The root causes of errors are accurately marked, clearly indicating the target knowledge points corresponding to the wrong questions. At the same time, the root causes of errors are divided into four categories: unclear concepts, calculation errors, misreading the question, and lack of prior knowledge. This post provides a concise review of the core concepts of this knowledge point, as well as general problem-solving strategies for this type of question. Output the complete and standardized solution steps for this problem, and also mark the frequently missed points. We will push 1-2 variations of the same knowledge point, difficulty level, but with different question stems and scenarios.

8. The method for adaptive closed-loop management of home-based after-school learning based on AI Agent according to claim 1, characterized in that, The specific steps of step S04 are as follows: Locate the target knowledge point directly corresponding to the incorrect question, clarify the hierarchical position of the knowledge point in the directed acyclic graph knowledge graph, and then trace back along the directed edges of the graph to find all the first-level and second-level predecessor dependency nodes of the target knowledge point: the first-level predecessor nodes are the knowledge points directly dependent on by the target knowledge point, and the second-level predecessor nodes are the underlying basic knowledge points dependent on by the first-level predecessor nodes. From the full backtracking of all preceding dependent nodes, retrieve the historical mastery score of each node, filter out the preceding nodes whose mastery is lower than the preset weak point threshold, sort the filtered weak preceding nodes and target knowledge points according to the order of subject learning, and form a complete weak knowledge link from the most basic nodes to the target knowledge points. Based on the generated weak knowledge links, the AI ​​Agent generates tiered reinforcement content.

9. The method for adaptive closed-loop management of home-based after-school learning based on AI Agent according to claim 1, characterized in that, The update rule described in step S05 is as follows: Positive update: If a single knowledge point is answered correctly 3 times in a row, its mastery score is raised to 90 points or above, it is marked as a mastered knowledge point, and the review frequency is reduced from once every 7 days to once every 14 days. Negative update: If the correct answer rate for a single knowledge point is lower than the preset weak point threshold (60 points) for two consecutive times, the mastery score of that knowledge point will be lowered, it will be marked as a high-frequency weak knowledge point, and the review frequency will be increased from once every 7 days to once every other day.

10. The method for adaptive closed-loop management of home-based after-school learning based on AI Agent according to claim 1, characterized in that, The method also includes step S06: The AI ​​Agent visualizes the user's task execution progress, knowledge point mastery panoramic heat map, error type trend and review plan completion, and generates daily and weekly reports to be pushed to the parents' end.

Images