An intelligent learning method and related device

Abstract

This invention discloses an intelligent learning companion method and related devices, relating to the field of intelligent education technology. The intelligent learning companion method includes: acquiring learner interaction data during the learning process and determining the current dynamic cognitive state based on the interaction data; generating multiple candidate teaching actions based on the current dynamic cognitive state and the current learning task; simulating the learner's reactions under each candidate teaching action based on a learner digital twin model aligned with the current dynamic cognitive state; determining the learning benefit index and teaching risk index of each candidate teaching action based on the reaction data corresponding to each candidate teaching action; and selecting a target teaching action and outputting it to the target learner based on the learning benefit index and teaching risk index of each candidate teaching action. This invention can proactively perceive the learner's cognitive state and proactively plan teaching strategies during the learning companion process, truly meeting the needs for continuous companionship and proactive guidance in one-on-one learning companionship.

Description

Technical Field

[0001] This invention relates to the field of intelligent education technology, and in particular to an intelligent learning companion method and related devices. Background Technology

[0002] With the development of large language models and intelligent agent technology, intelligent education products are now capable of performing tasks such as answering questions, explaining test questions, automatic grading, generating learning suggestions, recommending materials, and providing Q&A support. Large language models can generate relatively complete natural language explanations based on learners' questions, and can also generate examples, analyses, summaries, and exercises based on textbooks, questions, knowledge points, and learners' answers.

[0003] Current intelligent education solutions include intelligent homework answering and explanation solutions based on large language models. These solutions utilize the generative capabilities of large language models to directly output answer explanations or learning suggestions. Their working mode is essentially a passive response mode of "learners ask questions, and the system answers." Current intelligent education solutions also include solutions based on learning path planning and resource recommendation. These solutions typically recommend the next knowledge point, practice questions, course videos, or materials to learners based on course knowledge graphs, learning order, learning objectives, resource difficulty, and learner mastery. Their focus is on what to learn next and what resources to recommend.

[0004] In the K-12 education sector, one-on-one tutoring is gaining increasing attention; however, current smart education solutions are not suitable for one-on-one tutoring scenarios. Summary of the Invention

[0005] In view of this, the present invention provides an intelligent learning companion method and related device to solve the problem that existing intelligent education solutions are not applicable to one-on-one learning companion scenarios. The technical solution is as follows:

[0006] The first aspect of this invention provides an intelligent learning companion method, comprising:

[0007] Acquire interaction data of the target learner during the learning process, and determine the current dynamic cognitive state of the target learner based on the interaction data;

[0008] Based on the current dynamic cognitive state and the current learning task, multiple candidate teaching actions are generated;

[0009] Obtain a learner digital twin model that aligns with the current dynamic cognitive state, and simulate the target learner's response under each of the candidate instructional actions based on the learner digital twin model to obtain the predicted learner response data corresponding to each of the candidate instructional actions.

[0010] Based on the predicted learner response data corresponding to each of the candidate teaching actions, the learning benefit indicators and teaching risk indicators of each candidate teaching action are determined.

[0011] Based on the learning benefit indicators and teaching risk indicators of each candidate teaching action, a target teaching action is selected from the candidate teaching actions and output to the target learner.

[0012] In one possible implementation, the intelligent learning companion method further includes:

[0013] After outputting the target teaching action, the reaction data of the target learner under the target teaching action is collected to obtain the real learner reaction data corresponding to the target teaching action;

[0014] Determine the deviation between the actual learner response data corresponding to the target instruction action and the predicted learner response data corresponding to the target instruction action;

[0015] Based on the deviation, the current dynamic cognitive state and the learner's digital twin model are updated, and then the process of generating multiple candidate teaching actions and subsequent steps based on the current dynamic cognitive state and the current learning task is executed until the learning companionship termination condition is met.

[0016] In one possible implementation, the current dynamic cognitive state includes a first cognitive state, or includes both a first cognitive state and a second cognitive state.

[0017] The first cognitive state includes some or all of the following states: knowledge mastery status, typical misconceptions, and current learning obstacles;

[0018] The second cognitive state includes some or all of the following states: cognitive load, risk of forgetting, learning speed, expression preference, risk of emotional frustration, zone of proximal development, cue dependence, grade level matching, textbook version matching, and curriculum standard matching.

[0019] In one possible implementation, the intelligent learning companion method further includes:

[0020] Obtain teaching constraint information, wherein the teaching constraint information is used to impose at least one of the following constraints: age-appropriateness constraint, beyond-syllabus constraint, prompt intensity constraint, and learning load constraint;

[0021] The process of generating multiple candidate teaching actions based on the current dynamic cognitive state and the current learning task includes:

[0022] Based on the current dynamic cognitive state and the current learning task, multiple candidate teaching actions are generated under the constraints of the teaching constraint information;

[0023] The simulation of the target learner's response to each of the candidate instructional actions based on the learner's digital twin model includes:

[0024] Based on the learner digital twin model, and under the constraints of the teaching constraint information, the model simulates the target learner's response to each of the candidate teaching actions.

[0025] In one possible implementation, each of the candidate teaching actions is one of the following teaching actions:

[0026] Follow-up questions, gentle hints, step-by-step explanations, example demonstrations, analogical explanations, supplementary prior knowledge, reduced difficulty exercises, variation exercises, clarification of error causes, summarization, and requiring learners to continue trying.

[0027] In one possible implementation, each of the candidate teaching actions includes some or all of the following attributes:

[0028] Action type, target knowledge point, cue intensity, difficulty level, teaching intention, applicable conditions, prohibition conditions, and expected learner response.

[0029] In one possible implementation, the step of selecting a target teaching action from the candidate teaching actions and outputting it to the target learner based on the learning benefit index and teaching risk index of each candidate teaching action includes:

[0030] Candidate teaching actions whose teaching risk indicators exceed the preset risk threshold will be excluded.

[0031] Based on the learning benefit indicators and teaching risk indicators of the remaining candidate teaching actions, a target teaching action is selected from the remaining candidate teaching actions and output to the target learner.

[0032] In one possible implementation, the step of selecting a target teaching action from the remaining candidate teaching actions and outputting it to the target learner based on the learning benefit indicators and teaching risk indicators of the remaining candidate teaching actions includes:

[0033] Based on the learning benefit indicators and teaching risk indicators of the remaining candidate teaching actions, and combined with the matching degree between the remaining candidate teaching actions and the teaching constraint information, the comprehensive score of the remaining candidate teaching actions is determined.

[0034] Based on the combined scores of the remaining candidate teaching actions, a target teaching action is selected from the remaining candidate teaching actions and output to the target learner.

[0035] In one possible implementation, the intelligent learning companion method further includes:

[0036] Simulations are performed on each of the candidate teaching actions based on a virtual learner group to obtain simulation result sets corresponding to each candidate teaching action. The virtual learner group includes multiple virtual learners constructed based on multiple learner samples similar to the target learner. The multiple virtual learners have different learning characteristics. The simulation result sets contain the response data of the multiple virtual learners under the corresponding candidate teaching actions.

[0037] Based on the simulation result sets corresponding to each of the candidate teaching actions, the stability index of each candidate teaching action is determined.

[0038] The step of selecting a target teaching action from among the candidate teaching actions and outputting it to the target learner based on the learning benefit indicators and teaching risk indicators of each candidate teaching action includes:

[0039] Based on the learning benefit indicators and teaching risk indicators of each candidate teaching action, and in conjunction with the stability indicators of each candidate teaching action, a target teaching action is selected from the candidate teaching actions and output to the target learner.

[0040] In one possible implementation, the learning benefit metrics include at least one of the following metrics: mastery of target knowledge points, probability of overcoming current obstacles, probability of independent completion, probability of correcting misconceptions, degree of supplementation of prior knowledge, and probability of transfer and application.

[0041] The teaching risk indicators are used to assess the risk status of at least one of the following risk dimensions: risk of over-prompting, risk of excessive difficulty, risk of reinforcing misconceptions, risk of excessive cognitive load, risk of frustration, risk of answer dependence, risk of deviating from learning objectives, and risk of exceeding teaching constraints.

[0042] A second aspect of this application provides an electronic device, including at least one processor and a memory connected to the processor, wherein:

[0043] The memory is used to store computer programs;

[0044] The processor is used to execute the computer program to enable the electronic device to implement any of the above-mentioned intelligent learning companion methods.

[0045] Using the above technical solution, the intelligent learning companion method provided by this invention determines the current dynamic cognitive state of the target learner by acquiring the target learner's interaction data, and generates multiple candidate teaching actions based on the dynamic cognitive state and the current learning task. On this basis, based on the learner's digital twin model aligned with the current dynamic cognitive state, each candidate teaching action is simulated and verified to obtain the predicted learner response data corresponding to each candidate teaching action. Then, based on the predicted learner response data corresponding to each candidate teaching action, the learning benefit index and teaching risk index of each candidate teaching action are determined. Finally, based on the learning benefit index and teaching risk index of each candidate teaching action, the target teaching action is selected and output to the target learner. This invention proactively senses the learner's learning status and plans teaching strategies during the learning process, rather than passively waiting for the learner to ask questions before generating answers. This truly meets the need for continuous companionship and proactive guidance in one-on-one learning. Furthermore, this invention uses a learner digital twin model aligned with the learner's current dynamic cognitive state to simulate and verify each candidate teaching action. This allows for the prediction of the learner's reaction to each candidate action, thereby assessing the learning benefits and teaching risks of each action. Based on the assessment results, the invention selects the optimal target teaching action for output. This process ensures that the output target teaching action pursues learning benefits while maintaining controllable teaching risks, effectively improving the scientific nature, safety, and adaptability of teaching decisions in one-on-one learning scenarios. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0047] Figure 1 This is a flowchart illustrating the intelligent learning companion method provided in an embodiment of the present invention;

[0048] Figure 2 This is a schematic diagram of the closed-loop verification of teaching action simulation provided in an embodiment of the present invention;

[0049] Figure 3 This is a schematic diagram of the learning companion intelligent agent system provided in an embodiment of the present invention. Detailed Implementation

[0050] The embodiments of the present invention will now be described with reference to the accompanying drawings. The terminology used in the embodiments section is for illustrative purposes only and is not intended to limit the scope of the invention.

[0051] The embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art will recognize that, with technological advancements and the emergence of new scenarios, the technical solutions provided by the embodiments of the present invention are equally applicable to similar technical problems.

[0052] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms are interchangeable where appropriate; this is merely a way of distinguishing objects with the same attributes in the description of embodiments of the invention. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion, so that a process, method, system, product, or apparatus that comprises a series of units is not necessarily limited to those units, but may include other units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0053] In the process of realizing this case, the inventors discovered that, unlike standardized classroom teaching or one-way Q&A tutoring, one-on-one tutoring emphasizes continuous companionship with learners throughout the teaching process, real-time perception of learners' status changes, and dynamic adjustment of teaching strategies based on learners' immediate reactions. Existing intelligent education solutions are not suitable for one-on-one tutoring scenarios.

[0054] In view of this, the present invention proposes an intelligent tutoring method suitable for one-on-one tutoring scenarios. The intelligent tutoring method provided by the present invention will be described below through the following embodiments.

[0055] Please see Figure 1 The diagram illustrates a flowchart of an intelligent learning companion method provided by an embodiment of the present invention. This intelligent learning companion method may include:

[0056] Step S101: Obtain the interaction data of the target learner during the learning process, and determine the current dynamic cognitive state of the target learner based on the interaction data.

[0057] The interaction data of the target learner during the learning process is a multi-source information set reflecting the learner's behavior and thought process during the learning activities. In this embodiment, the interaction data may include, but is not limited to, some or all of the following: answer data (such as the answer results during the question-solving process, the answer process trajectory (such as step breakdown, draft content, etc.)), dialogue content, pause information (such as the pause time during the answer process), modification records, error annotations, question information, textbook chapter information, and historical learning trajectory. The above data constitutes the basic raw materials for characterizing the learner's real-time state.

[0058] This embodiment, after obtaining multi-source interactive data, constructs the current dynamic cognitive state of the target learner through fusion analysis. Unlike static learner profiles used only for display or resource recommendation, the current dynamic cognitive state of the target learner is a cognitive state oriented towards teaching decisions, used to characterize the target learner's cognitive situation in the current learning task.

[0059] In one possible implementation, the current dynamic cognitive state of the target learner may include a first cognitive state, which may include, but is not limited to, some or all of the following states (preferably all of them): knowledge mastery state, typical misconceptions, and current learning bottlenecks.

[0060] Among them, knowledge mastery status is used to characterize the degree to which the target learner has mastered each knowledge point, typical misconceptions are used to characterize the typical error types and their causes in the learning process, such as conceptual confusion misconceptions, step omission misconceptions, and omissions in reading the question, etc., and current learning bottlenecks are used to characterize the specific obstacles encountered by the target learner in the current learning task.

[0061] For example, when it is detected that a learner makes unit conversion errors repeatedly in the same type of question and repeatedly asks "Why do we need to unify units?" in the dialogue, the current dynamic cognitive state can be determined by analyzing the interaction data as follows: there is a misunderstanding of the concept of units and the current bottleneck lies in understanding the necessity of unifying units.

[0062] The first cognitive state is the core cognitive state for instructional decision-making. In another possible implementation, the current dynamic cognitive state of the target learner can include not only the first cognitive state but also a second cognitive state. The second cognitive state extends and supplements the core cognitive state, and when data is sufficient, it can further enrich the dimensions of the learner's state, making instructional decisions more precise.

[0063] The second cognitive state may include, but is not limited to, some or all of the following states (preferably all of them): cognitive load, risk of forgetting, learning speed, expression preference, risk of emotional frustration, range of the zone of proximal development, degree of cue dependence, grade level matching, textbook version matching, and curriculum standard matching.

[0064] Among them, the matching status of grade level, textbook version, and curriculum standard respectively represent the degree of conformity or gap between the learner's current cognitive level and the requirements of the current grade level, the knowledge system of the textbook version used, and the national curriculum standard. These are used to determine whether teaching actions need to supplement prior knowledge or whether they may exceed the curriculum.

[0065] There are several ways to determine the current dynamic cognitive state based on interaction data. One possible approach is to use a rule-matching algorithm to determine the current dynamic cognitive state. Specifically, several cognitive state determination rules can be preset, each rule corresponding to one or more triggering conditions for a cognitive state. Based on these rules, the interaction data is transformed into the learner's state in dimensions such as knowledge mastery, typical misconceptions, and current learning difficulties. Another possible approach is to use a trained cognitive diagnostic model to map the interaction data to a preset cognitive state space. The confidence level of each state dimension is then calculated using probability, thereby determining the learner's state in dimensions such as knowledge mastery, typical misconceptions, and current learning difficulties.

[0066] For example, the current dynamic cognitive state of the target learner can be represented as:

[0067] (1).

[0068] in, This represents the learner's dynamic cognitive state during the t-th round of interaction. Indicates the state of knowledge mastery. This indicates a typical misconception. Indicates the current learning bottleneck. Indicates cognitive load status. This indicates a learning mood or a state of risk of frustration. Indicates the extent of the zone of most recent development. It represents the historical learning trajectory.

[0069] By constructing the current dynamic cognitive state of the target learner, we can accurately identify the actual difficulties and cognitive level of the target learner in the current learning task, providing precise input for the generation of subsequent candidate teaching actions for teaching decision-making.

[0070] Step S102: Generate multiple candidate teaching actions based on the current dynamic cognitive state and the current learning task.

[0071] In this embodiment, the candidate teaching actions are specific teaching interventions generated based on the current dynamic cognitive state and current learning task of the target learner, which are examples of executable teaching strategies.

[0072] This embodiment generates multiple candidate teaching actions based on the current learning task (such as a test question being answered or a knowledge point being learned) and the aforementioned determined current dynamic cognitive state. These candidate teaching actions have different teaching intentions, difficulty levels, and prompting intensities. The multiple candidate teaching actions can be represented as follows:

[0073] (2).

[0074] in, Represents the set of candidate teaching actions. Let represent the i-th candidate teaching action, and n be the number of candidate teaching actions.

[0075] Multiple candidate teaching actions are specific teaching intervention methods that can be selected. Each candidate teaching action may be, but is not limited to, one of the following teaching actions: follow-up questioning, gentle prompting, step-by-step explanation, example demonstration, analogy explanation, supplementing prior knowledge, reducing the difficulty of exercises, variation exercises, clarifying the causes of errors, summarizing and generalizing, and requiring learners to continue trying, etc.

[0076] Follow-up questions are used to guide learners to think more deeply when their answers are incomplete or biased. This helps learners discover flaws in their reasoning and is suitable for situations where learners' answers are close to the correct answer but are not clearly expressed.

[0077] Hints are a few guiding messages provided when learners get stuck (such as rereading keywords in the question, reviewing relevant concept definitions, or using diagrams to help them overcome the current obstacle, rather than directly giving them the answer). They are suitable for scenarios where learners' thinking is basically correct but they have temporarily forgotten a key point.

[0078] Step-by-step explanation breaks down a complete problem-solving path or reasoning process into several consecutive logical steps and presents them sequentially. This aims to reduce the cognitive processing load of a single session and help learners build a complete problem-solving logic framework. It is suitable for scenarios where learners lack an overall thought process or where there is a logical break between steps.

[0079] The example demonstration focuses on selecting a question with the same knowledge structure as the current learning task but a different context and providing a complete solution demonstration. It uses an analogy transfer mechanism to help learners understand abstract problem-solving rules and is suitable for scenarios where learners lack a cognitive framework for new question types.

[0080] Analogical explanation maps complex concepts to intuitive understanding by constructing familiar situations (such as real-life examples or known models) that are structurally or logically similar to the current difficulty. It is suitable for scenarios where abstract concepts are difficult to understand.

[0081] Prerequisite knowledge supplementation is used to fill in the gaps in the learner's prior knowledge. It is applicable to scenarios where learners cannot understand the current content due to a lack of prior knowledge.

[0082] Lowering the difficulty level involves providing similar questions of lower difficulty for the target learner to practice. This is suitable for situations where the target learner's current ability does not match the difficulty of the questions.

[0083] Variation exercises provide different variations of the same knowledge point (changing the non-essential attributes of the questions while retaining the core knowledge structure), training learners' ability to flexibly apply knowledge and resist interference. They are suitable for scenarios where learners have initially mastered the knowledge but lack the ability to transfer it.

[0084] Error clarification differs from simple answer correction. Instead, it focuses on the typical misconceptions or erroneous thinking patterns revealed by the target learner during the answering process, conducts targeted attribution analysis, and clearly points out the error type, error cause, and correct approach. It is applicable to scenarios where the target learner has fixed misconceptions.

[0085] Summarizing and generalizing involves systematically summarizing the current knowledge point or problem-solving method after completing the learning or problem-solving of a certain knowledge point. It is suitable for scenarios where the target learner has basically mastered the knowledge and needs to form a knowledge structure.

[0086] Asking learners to continue trying encourages them to attempt to answer independently again without providing additional guidance or hints. This approach aims to cultivate learners' willpower and independent problem-solving habits, and is suitable for scenarios where learners are close to solving the problem independently and only require minimal encouragement.

[0087] For example, when a learner makes a mistake in writing an equation for a word problem, three candidate actions can be generated simultaneously: first, ask a follow-up question that focuses on the meaning of the problem; second, provide a half-step hint and guidance to help the learner find the quantitative relationship; and third, directly explain the complete solution step by step.

[0088] Each candidate instructional action generated based on the current dynamic cognitive state and the current learning task may include some or all of the following attribute information: action type, target knowledge point, prompt intensity, difficulty level, instructional intent, applicable conditions, prohibition conditions, and expected learner response.

[0089] Among them, the action type is used to identify what kind of teaching action the candidate teaching action belongs to, such as whether it is a follow-up question, a light prompt, or a step-by-step explanation; the target knowledge point is used to indicate the knowledge point targeted by the candidate teaching action to ensure the relevance of the teaching content; the prompt intensity is used to indicate the strength of the prompt information in the candidate teaching action, such as whether it only provides directional guidance or directly presents key steps, which directly affects the degree of space occupied for learners' independent thinking; the difficulty level is used to indicate the cognitive difficulty level of the candidate teaching action; the teaching intention is used to describe the teaching purpose of the candidate teaching action, such as whether it is to correct a specific misconception or to consolidate a certain prior knowledge point; the applicable conditions are used to indicate the learner's state conditions to which the candidate teaching action is applicable; the prohibitive conditions are used to indicate the learner's state conditions to which the candidate teaching action is not applicable; and the expected learner response describes the expected response of the target learner after receiving the candidate teaching action.

[0090] There are several ways to generate multiple candidate teaching actions based on the current dynamic cognitive state and the current learning task. One possible implementation involves calling a large language model to generate multiple candidate teaching actions based on the current dynamic cognitive state and the current learning task. Specifically, the text description of the current dynamic cognitive state and the current learning task, along with a preset teaching action template, can be organized into prompt words and input into the large language model. The large language model then generates multiple candidate teaching actions with different teaching intentions, difficulty levels, and prompt intensities, and outputs attribute information for each candidate teaching action, including action type, target knowledge point, prompt intensity, difficulty level, teaching intention, applicable conditions, prohibition conditions, and expected learner response. It should be noted that this embodiment is not limited to generating multiple candidate teaching actions based on a large language model. Another possible implementation involves using a rule-based retrieval method to generate multiple candidate teaching actions. Specifically, based on the current dynamic cognitive state, an applicable action type is matched in a preset teaching strategy rule base. For example, when the checkpoint is "missed step," "light prompt" or "step-by-step explanation" is matched, and candidate teaching actions are generated based on the matching results. By generating multiple candidate teaching actions, a diverse range of choices is provided for subsequent optimal decision-making.

[0091] Step S103: Obtain a learner digital twin model that is aligned with the current dynamic cognitive state, and simulate the target learner's response under each candidate instructional action based on the learner digital twin model to obtain the predicted learner response data corresponding to each candidate instructional action.

[0092] A learner digital twin model is a simulation model built in digital space that is highly aligned with the current state of the target learner. It can also be called a shadow learner. This model is not a generalized virtual learner role, but is specifically designed to simulate the behavioral responses of a particular target learner in the current dynamic cognitive state.

[0093] The process of obtaining a learner digital twin model aligned with the current dynamic cognitive state includes: if there is no learner digital twin model corresponding to the target learner, then construct a learner digital twin model based on the target learner's current dynamic cognitive state to obtain a learner digital twin model aligned with the current dynamic cognitive state; if there is already a learner digital twin model corresponding to the target learner, then update the learner digital twin model based on the target learner's current dynamic cognitive state to obtain a learner digital twin model aligned with the current dynamic cognitive state.

[0094] Learner digital twin models can be implemented using any one or more of the following: large language model, knowledge tracking model, rule model, cognitive diagnostic model, and intelligent agent. These implementation schemes can even be encapsulated into standardized Skill modules for use.

[0095] After obtaining a learner digital twin model aligned with the current dynamic cognitive state, the learner's response to each candidate instructional action is simulated based on this model. Specifically, each candidate instructional action and the learner's current dynamic cognitive state are input into the learner digital twin model to simulate the learner's response after receiving the candidate instructional action. This simulates one or more of the following: correct progress, partial understanding, continued hesitation, new misunderstandings, emotional frustration, requesting answers, and giving up. This yields predicted learner response data for each candidate instructional action, which characterizes the specific learning outcomes and psychological state changes that the learner may experience after receiving the corresponding candidate instructional action.

[0096] Step S104: Based on the predicted learner response data corresponding to each candidate teaching action, determine the learning benefit index and teaching risk index for each candidate teaching action.

[0097] After obtaining the predicted learner response data for each candidate teaching action, for each candidate teaching action, the learning benefit target and teaching risk indicator are determined based on the current dynamic cognitive state of the target learner and the predicted learner response data corresponding to the candidate teaching action.

[0098] The learning benefit index of any candidate teaching action is used to characterize the positive teaching effect that the candidate teaching action may bring. It may include, but is not limited to, at least one of the following indicators: mastery of target knowledge points, probability of resolving current obstacles, probability of independent completion, probability of correcting misconceptions, degree of supplementation of prior knowledge, and probability of transfer and application.

[0099] It should be noted that the target knowledge point mastery gain is used to characterize the improvement in the mastery of the target knowledge point that the candidate teaching action may bring; the current obstacle removal probability is used to characterize the probability that the candidate teaching action may remove the learner's current learning obstacle; the independent completion probability is used to characterize the probability that the learner can independently complete the same type of task after the candidate teaching action; the misconception correction probability is used to characterize the probability that the candidate teaching action may correct the learner's typical misconceptions; the degree of prior knowledge completion is used to characterize the degree that the candidate teaching action may complete the learner's missing prior knowledge; and the transfer application probability is used to characterize the probability that the learner can transfer and apply the knowledge to a new situation after the candidate teaching action.

[0100] The teaching risk index for any candidate teaching action is used to assess the potential negative effects or risks that the candidate teaching action may cause. Specifically, the teaching risk index can be used to assess the risk situation of at least one of the following risk dimensions: risk of over-cueing, risk of excessive difficulty, risk of reinforcing misconceptions, risk of excessive cognitive load, risk of frustration, risk of answer dependence, risk of deviating from learning objectives, and risk of exceeding teaching constraints.

[0101] It should be noted that the risks of over-prompting are: the risk that excessive prompting may deprive learners of opportunities for independent thinking; the risk of excessive difficulty is the risk that the candidate teaching action may exceed the learner's current cognitive level; the risk of reinforcing misconceptions is the risk that the candidate teaching action may reinforce the learner's misunderstandings by failing to identify misconceptions beforehand; the risk of excessive cognitive load is the risk that the candidate teaching action may overload the learner's cognitive resources; the risk of frustration is the risk that the candidate teaching action may cause learners to feel frustrated; the risk of answer dependence is the risk that the candidate teaching action may cause learners to develop a habit of directly asking for answers; the risk of deviating from learning objectives is the risk that the candidate teaching action may deviate from the current learning objectives; and the risk of exceeding teaching constraints is the risk that the candidate teaching action may exceed the constraints of the grade level, textbook version, or curriculum standards.

[0102] Step S105: Based on the learning benefit indicators and teaching risk indicators of each candidate teaching action, select the target teaching action from the candidate teaching actions and output it to the target learner.

[0103] Targeted instructional actions can be the most suitable instructional intervention strategies for the target learners, selected from multiple candidate instructional actions.

[0104] Specifically, based on the learning benefit indicators and teaching risk indicators of each candidate teaching action, a comprehensive evaluation is conducted on each candidate teaching action, and the target teaching action is selected and output to the target learner. Assuming that the target learner's current dynamic cognitive state is the same as the dynamic cognitive state at the t-th round of interaction in the one-on-one tutoring process, the selected target teaching action will be output to the target learner as the teaching action in the (t+1)-th round of interaction.

[0105] By using a screening mechanism based on both benefit and risk assessment, inappropriate teaching actions (such as overly difficult explanations that may lead to frustration or excessive prompting that deprives learners of the opportunity to think) can be avoided from being directly delivered to the target learners, thereby ensuring that each teaching intervention is as appropriate as possible to the current cognitive state of the target learners.

[0106] The intelligent learning companion method provided in this invention first acquires multi-source interactive data and constructs a current dynamic cognitive state oriented towards teaching decisions. This accurately identifies the learner's actual difficulties and cognitive level in the current learning task, providing a precise basis for generating subsequent teaching actions. Second, by generating multiple candidate teaching actions based on the current dynamic cognitive state and the current learning task, different teaching intervention strategies can be proactively planned in each round of interaction. The differences in teaching intent, difficulty level, and prompt intensity among the multiple candidate teaching actions provide a diverse selection space for subsequent optimal decision-making. Furthermore, by acquiring a digital twin model of the learner aligned with the current dynamic cognitive state... The model simulates and verifies each candidate teaching action, enabling pre-prediction and evaluation of the teaching action before its actual output. It can predict the learner's possible reaction after receiving each candidate teaching action. Finally, based on the predicted learner reaction data, the learning benefit index and teaching risk index of each candidate teaching action are determined. Based on the above evaluation results, the target teaching action is selected for output, thus achieving a balance between pursuing positive teaching effects and avoiding teaching risks. This ensures that the output target teaching action can bring high learning benefits to the target learner while ensuring that teaching risks are controllable, thereby effectively improving the scientific nature, safety, and adaptability of teaching decisions in one-on-one tutoring scenarios.

[0107] Considering that teaching actions in K-12 teaching scenarios (referring to basic education scenarios from kindergarten to high school, covering teaching activities across all stages of preschool education, primary school, junior high school, and high school) need to meet various teaching standards such as age-appropriateness, exceeding the curriculum requirements, prompting intensity, and learning load, in order to avoid generating teaching actions that exceed teaching objectives or are inconsistent with learners' cognitive development levels, in some embodiments of this application, the intelligent learning companion method may further include:

[0108] Before generating multiple candidate teaching actions based on the current dynamic cognitive state and current learning task of the target learner, teaching constraint information is obtained. Then, when generating multiple candidate teaching actions, multiple candidate teaching actions are generated under the constraints of the teaching constraint information based on the current dynamic cognitive state and current learning task.

[0109] Teaching constraint information may include, but is not limited to, some or all of the following: the target learner's grade level, textbook version, curriculum standards, current chapter, learning objectives, and preset boundary conditions. The preset boundary conditions can be personalized teaching restrictions set by teachers or parents based on their understanding of the target learner and educational goals. These conditions differ from relatively fixed constraints such as grade level and textbook version, and are more dynamic and individualized.

[0110] Instructional constraints typically encompass one or more of the following: age-appropriate constraints, out-of-syllabus constraints, cue intensity constraints, and learning load constraints. Age-appropriate constraints ensure that the content and methods of instruction align with the learner's cognitive development level for their age group. Out-of-syllabus constraints ensure that the instruction does not exceed the knowledge scope defined by the current grade level, textbook version, and national curriculum standards. Cue intensity constraints control the level of assistance provided to learners, avoiding excessive or insufficient prompting. Learning load constraints limit the amount of information presented and the cognitive pressure within a single interaction or unit of time, preventing cognitive fatigue or frustration caused by information overload.

[0111] Based on the information obtained about teaching constraints, this information can be introduced into the process of generating candidate teaching actions to ensure that each candidate teaching action is within the boundaries of the teaching standards.

[0112] Through this method, each candidate teaching action generated is personalized to address the learner's specific problems, and a compliance verification mechanism is built into the technical implementation. This approach allows the teaching actions to avoid compliance risks at the generation stage, avoiding the computational overhead and efficiency losses caused by secondary screening of a large number of unqualified actions in subsequent steps.

[0113] The multiple candidate teaching actions generated in the above manner together constitute a safe solution space that satisfies the teaching constraints, providing a high-quality input basis for subsequent digital twin simulation verification, benefit and risk assessment, and optimal output selection, effectively improving the efficiency and reliability of the overall decision-making process.

[0114] In addition to applying teaching constraint information during the candidate teaching action generation stage, teaching constraint information can also be applied during the simulation verification stage of the candidate teaching actions.

[0115] Specifically, after generating multiple candidate instructional actions, a learner digital twin model, aligned with the target learner's current dynamic cognitive state, simulates the target learner's responses to each candidate instructional action under the constraints of instructional constraints. More specifically, the candidate instructional actions, the current dynamic cognitive state, and the instructional constraints are all input into the learner digital twin model, which then simulates the target learner's responses to each candidate instructional action under the constraints of the instructional constraints.

[0116] target learners in candidate instructional actions The following reaction can be represented as:

[0117] (3).

[0118] Wherein, F represents the learner's digital twin model, which can be implemented by a large language model, a knowledge tracing model, a cognitive diagnostic model, a rule model, or a combination thereof. This includes the current topic, teaching constraints (such as learning objectives, textbook content, curriculum standards, etc.), and contextual constraints. Indicates candidate teaching actions The corresponding predictive learner response data.

[0119] As can be seen from the above, teaching constraint information runs through the two core stages of candidate teaching action generation and simulation verification in the intelligent learning companion method (teaching constraint information sets the boundary conditions that must be followed for the generation and simulation of candidate teaching actions). This dual constraint mechanism internalizes external teaching norms into the logical boundaries of the method's operation, ensuring that before any teaching action is output, it has undergone multiple checks on age-appropriateness, compliance, and cognitive load. This makes the generated teaching actions not only meet the learner's personalized cognitive needs but also strictly follow the professional standards of K-12 education, thereby improving the practicality and credibility of the intelligent learning companion method.

[0120] As mentioned in the above embodiments, after obtaining the predicted learner response data corresponding to each candidate teaching action, the learning benefit index and teaching risk index of each candidate teaching action are determined based on the predicted learner response data corresponding to each candidate teaching action.

[0121] The learning benefit index of any candidate teaching action is used to characterize the learning benefits that the candidate teaching action may bring, such as improved knowledge mastery, removal of obstacles, correction of misunderstandings, and improvement of independent completion ability.

[0122] For any candidate teaching action Based on this candidate teaching action Corresponding predictive learner response data At the same time, it takes into account the current cognitive state of the target learner's actions. and the above Determine the candidate teaching action Learning benefit metrics.

[0123] Candidate teaching actions The learning benefit metric can be expressed as:

[0124] (4).

[0125] in, Indicates candidate teaching actions Learning benefit metrics This represents the function for evaluating learning benefits.

[0126] The teaching risk index for any candidate teaching action is used to characterize the teaching risks that the candidate teaching action may bring, such as excessive prompting or excessive difficulty.

[0127] For any candidate teaching action Based on this candidate teaching action Corresponding predictive learner response data At the same time, it takes into account the current cognitive state of the target learner's actions. The candidate teaching action and the above The teaching risk indicators for the candidate teaching action are determined.

[0128] Candidate teaching actions The teaching risk indicators can be expressed as:

[0129] (5).

[0130] in, Indicates candidate teaching actions Teaching risk indicators This represents a teaching risk assessment function.

[0131] After obtaining the learning benefit indicators and teaching risk indicators of each candidate teaching action through the above methods, the target teaching action is selected from the candidate teaching actions based on the learning benefit indicators and teaching risk indicators of each candidate teaching action, and then the target teaching action is output to the target learner.

[0132] In one possible implementation, the process of selecting the target teaching action from among the candidate teaching actions based on the learning benefit indicators and teaching risk indicators of each candidate teaching action may include:

[0133] Step a1: Exclude candidate teaching actions whose teaching risk indicators exceed the preset risk threshold.

[0134] Preset risk thresholds serve as quantitative boundaries for ensuring teaching safety and optimizing the learning experience, defining the upper limit of acceptable risk for teaching actions. These thresholds can be pre-defined by education experts based on K-12 teaching standards, dynamically configured by teachers or administrators according to specific teaching scenarios and individual learner differences, or automatically determined through statistical learning based on safe teaching cases from historical teaching data.

[0135] This embodiment compares the teaching risk indicators of each candidate teaching action with a preset risk threshold. If the teaching risk indicator of a candidate teaching action exceeds the preset risk threshold, it indicates that the candidate teaching action carries a high risk. If it is delivered to learners, it may cause negative learning experiences (such as frustration, cognitive overload, etc.) or teaching accidents (such as teaching beyond the syllabus, excessive prompting, etc.). Therefore, it is excluded and not considered. Through the above risk exclusion strategy, it can be ensured that all candidate teaching actions participating in the subsequent selection competition have met the basic qualification standards in terms of teaching safety.

[0136] Step a2: Select the target teaching action from the remaining candidate teaching actions based on the learning benefit indicators and teaching risk indicators of each candidate teaching action.

[0137] The remaining candidate teaching actions are acceptable in terms of teaching risk. Based on this, this embodiment comprehensively considers the learning benefit indicators and teaching risk indicators of the remaining candidate teaching actions, conducts a comprehensive evaluation of the remaining candidate teaching actions, and selects the target teaching action from them.

[0138] In one possible approach, the implementation of step a2 may include:

[0139] Step a21: Based on the learning benefit indicators and teaching risk indicators of the remaining candidate teaching actions, and in combination with the matching degree between the remaining candidate teaching actions and the teaching constraint information, determine the comprehensive score of the remaining candidate teaching actions.

[0140] For any of the remaining candidate teaching actions In one possible implementation, The overall score can be calculated using the following formula:

[0141] (6).

[0142] in, Indicates candidate teaching actions The overall score, Indicates candidate teaching actions Learning benefit metrics Indicates candidate teaching actions Teaching risk indicators Indicates candidate teaching actions The degree of matching with teaching constraint information, with α, β, and γ as weighting coefficients, is used to balance the relative importance of learning benefit indicators, teaching risk indicators, and the degree of matching with teaching constraint information in the overall score.

[0143] Step a22: Select the target teaching action from the remaining candidate teaching actions based on the comprehensive score of each candidate teaching action.

[0144] Preferably, the candidate teaching action with the highest overall score among the remaining candidate teaching actions can be determined as the target teaching action, which can be represented as:

[0145] (7).

[0146] The target teaching actions selected through the above process are the most suitable teaching intervention strategies for the target learners in the current dynamic cognitive state. They can maximize the positive effects of teaching while ensuring teaching safety.

[0147] It should be noted that this embodiment is not limited to selecting the target teaching action from the remaining candidate teaching actions using steps a21 to a22. In another possible implementation, the comprehensive score of the remaining candidate teaching actions can be determined based solely on the learning benefit index and teaching risk index of the remaining candidate teaching actions, and then the target teaching action can be selected from the remaining candidate teaching actions based on the comprehensive score of the remaining candidate teaching actions.

[0148] This embodiment is not limited to the method described in steps a1 to a2 for selecting the target teaching action. In another possible implementation, the process of selecting the target teaching action from the candidate teaching actions may include:

[0149] Step b1: Simulate each candidate teaching action based on a virtual learner group to obtain the simulation result set corresponding to each candidate teaching action, and determine the stability index of each candidate teaching action based on the simulation result set corresponding to each candidate teaching action.

[0150] The virtual learner group includes multiple virtual learners constructed based on multiple learner samples similar to the target learner. These virtual learners have different learning characteristics (such as different cognitive levels, different types of misconceptions, different learning motivations, and different levels of cue dependence). The simulation result set corresponding to any candidate teaching action contains the response data of multiple virtual learners under that candidate teaching action.

[0151] A virtual learner group is a collection of virtual learners with different learning characteristics. Essentially, it is a digital mapping of a similar learning group to the target learner, rather than a repetition of a single object. In constructing this group, multiple learner samples similar to the target learner are first obtained. Selection criteria may include, for example, the same grade level or similar mastery of knowledge points. Then, for each selected learner sample, a corresponding virtual learner is constructed based on its sample data. These virtual learners have different learning characteristics; for example, some virtual learners have strong logical reasoning abilities but are prone to calculation errors, some are highly dependent on prompts, while others have a strong desire for independent exploration.

[0152] Each candidate instructional action generated in the preceding steps is input into a batch simulation test of each virtual learner within the virtual learner group. The simulation process simulates the possible reactions of different virtual learners when they receive the same instructional action, and these reaction data constitute the simulation result set. For each candidate instructional action, its corresponding simulation result set contains the reaction data of all virtual learners in that group after receiving the action. Through this group-level simulation, the perspective can be expanded from that of a single target learner to that of similar groups, capturing the performance fluctuations of candidate instructional actions under different individual characteristics.

[0153] After obtaining the simulation result sets corresponding to each candidate teaching action, a stability index for each candidate teaching action is determined based on these sets. The stability index of any candidate teaching action characterizes the consistency of its effectiveness across different virtual learners. In one implementation, the stability index can be determined based on the dispersion of response data from each virtual learner under the same candidate teaching action. The lower the dispersion of response data, the more consistent the performance of the candidate teaching action across different virtual learners, and the higher the stability index. Conversely, the higher the dispersion of response data, the greater the difference in performance across different learners, and the lower the stability index.

[0154] Step b2: Based on the learning benefit indicators and teaching risk indicators of each candidate teaching action, and in conjunction with the stability indicators of each candidate teaching action, select the target teaching action from among the candidate teaching actions.

[0155] After obtaining the stability indicators of each candidate teaching action, these indicators are used as a new decision-making dimension and comprehensively weighed with learning benefit indicators and teaching risk indicators.

[0156] In one possible implementation, candidate teaching actions whose teaching risk indicators exceed a preset risk threshold can be excluded first. Then, based on the learning benefit indicators and teaching risk indicators of the remaining candidate teaching actions, combined with the matching degree of each remaining candidate teaching action with the teaching constraint information and the stability indicator of each candidate teaching action, a comprehensive score for each remaining candidate teaching action can be determined (for example, a weighted comprehensive calculation can be performed on the learning benefit indicators, teaching risk indicators, stability indicators, and matching degree with the teaching constraint information to obtain a comprehensive score for each candidate teaching action). Finally, based on the comprehensive scores of the remaining candidate teaching actions, the target teaching action is selected from the remaining candidate teaching actions. It should be noted that candidate teaching actions with higher learning benefit indicators, lower teaching risk indicators, higher stability indicators, and higher matching degree with the teaching constraint information have higher comprehensive scores and are more likely to be selected as the target teaching action.

[0157] In another possible implementation, candidate teaching actions with teaching risk indicators exceeding a preset risk threshold and stability indicators below a preset stability threshold can be excluded first. This ensures that the remaining candidate teaching actions meet the basic requirements of both controllable teaching risk and stable effect. Then, based on the learning benefit indicators and teaching risk indicators of the remaining candidate teaching actions, combined with their stability indicators and their matching degree with the teaching constraint information, the comprehensive score of the remaining candidate teaching actions is determined. Finally, the target teaching action is selected from the remaining candidate teaching actions based on their comprehensive scores.

[0158] By introducing stability indicators, the consistency of candidate instructional actions across virtual learners with different learning characteristics is incorporated into the decision-making process. This effectively reduces the risk of overfitting instructional actions to a single learner's instantaneous state (i.e., overfitting to a particular instantaneous state while ignoring dynamic changes in the state). This ensures that the final selected target instructional action exhibits good consistency across various cognitive states the target learner may be in, thereby enhancing the robustness of instructional decisions. The virtual learner group simulation and individual digital twin simulation form a complementary verification mechanism. Individual digital twin simulation ensures the applicability of instructional actions to the current learner, while virtual learner group simulation ensures the stability and safety of instructional actions across different cognitive states. The combination of these two methods effectively avoids instructional action bias caused by over-reliance on specific states.

[0159] In some embodiments of this application, the intelligent learning companion method, in addition to steps S101 to S105 described above, may also include:

[0160] Step S106: After outputting the target teaching action, continuously collect the target learner's response data under the target teaching action to obtain the real learner response data corresponding to the target teaching action.

[0161] After the target instruction action is output to the target learner, a data collection process for the execution effect of the instruction action is initiated, that is, to capture the actual behavioral response of the target learner after receiving the target instruction action, thereby obtaining the actual learner response data corresponding to the target instruction action.

[0162] Step S107: Determine the deviation between the actual learner response data corresponding to the target instruction action and the predicted learner response data corresponding to the target instruction action.

[0163] After obtaining the actual learner response data corresponding to the target instructional action, the deviation between this data and the predicted learner response data corresponding to the target instructional action is determined. This deviation can be expressed as:

[0164] (8).

[0165] in, Indicates deviation, This represents real learner response data corresponding to the target instructional action. d represents the predicted learner response data corresponding to the target instruction action, and d represents the difference measure between the actual learner response data and the predicted learner response data.

[0166] Step S108: Update the current dynamic cognitive state of the target learner and the learner's digital twin model based on the deviation.

[0167] After identifying the bias, the learner's current dynamic cognitive state is updated based on the bias. For example, if the actual response is "continue to get stuck," while the predicted response is "progress correctly," the corresponding knowledge mastery level in the current dynamic cognitive state is adjusted downwards, and the stuck state is corrected. The updated dynamic cognitive state can be represented as:

[0168] (9).

[0169] Meanwhile, the learner digital twin model is calibrated based on the bias, that is, the internal parameter configuration or strategy logic of the learner digital twin model is adjusted so that its next simulation prediction is closer to the actual response of the target learner.

[0170] The current dynamic cognitive state and learner digital twin model, after deviation correction, will be directly used to drive the next round of teaching decisions. That is, after updating the current dynamic cognitive state and learner digital twin model, the process returns to step S102, which generates multiple candidate teaching actions and subsequent steps based on the current dynamic cognitive state and current learning task, until the learning companionship termination conditions are met (such as completion of the learning task, learner voluntarily withdrawing, learning time being reached, etc.). This forms a closed-loop optimization mechanism for one-on-one learning companionship teaching actions.

[0171] As can be seen from the above process, the generation of teaching actions is no longer based on static historical profiles or generalized group models, but rather adapts with each actual response of the learner. That is, the teaching actions are always dynamically adjusted around the learner's current cognitive state, and the teaching decisions are always highly aligned with the learner's actual cognitive level.

[0172] Please see Figure 2 This diagram illustrates the closed-loop simulation verification of instructional actions, where the current dynamic cognitive state of the target learner is... ,based on Multiple candidate teaching actions are generated based on the current learning task, including candidate teaching action A (follow-up question), candidate teaching action B (complete explanation), and candidate teaching action C (supplementing prior knowledge). These candidate teaching actions are then input into the current dynamic cognitive state. The aligned learner digital twin model was used for simulation verification to predict the target learner's reaction to each candidate instructional action, obtaining predicted learner response data for candidate instructional action A (follow-up question), candidate instructional action B (complete explanation), and candidate instructional action C (supplementing prior knowledge). Based on this, the learning benefit and instructional risk indicators of the three candidate instructional actions were determined. Finally, based on the determined learning benefit and instructional risk indicators, the target instructional action was selected from the three candidate instructional actions and output to the target learner. Assuming that candidate action A has a high learning benefit and low instructional risk, candidate action B carries the risk of over-prompting, and candidate action C carries the risk of being too slow, candidate instructional action A was ultimately selected as the target instructional action output to the target learner. Subsequently, the target learner's response data under the target instructional action was continuously collected to obtain the actual learner response data corresponding to the target instructional action. Then, the deviation between the actual learner response data corresponding to the target instructional action and the predicted learner response data corresponding to the target instructional action was determined, and the current dynamic cognitive state of the target learner was analyzed based on this deviation. Update and get At the same time, the learner's digital twin model is corrected based on this deviation.

[0173] This invention transforms the selection of teaching actions in one-on-one tutoring into a closed-loop optimization process involving candidate teaching action generation, digital twin simulation, benefit-risk assessment, and real-world feedback calibration. Through this closed loop, the teaching decision-making process of this invention more closely resembles that of a real teacher. For example, for learners who are close to understanding, a gentle prompting may be chosen instead of a complete explanation; for learners with significant gaps in prior knowledge, supplementing prior knowledge may be chosen instead of continuing to ask follow-up questions; and for learners with fixed misconceptions, clarifying the causes of errors may be prioritized instead of repeating calculation steps.

[0174] The intelligent learning companion method provided by this invention has the following advantages:

[0175] Firstly, it upgrades from responsive Q&A to proactive teaching decision-making. Existing large language model Q&A solutions adopt a passive response mode of "learners ask questions, the system answers," lacking proactive perception and planning capabilities. In contrast, this invention proactively constructs a dynamic cognitive state by acquiring learners' interaction data and generates multiple candidate teaching actions based on this cognitive state. This enables teaching behavior to make autonomous decisions and dynamically adjust according to the learner's real-time state, thereby achieving continuous learner state tracking and adaptive teaching strategy planning in one-on-one tutoring, truly meeting the deep-seated need for proactive guidance and continuous companionship in tutoring scenarios.

[0176] Secondly, it upgrades from content delivery to personalized action-level decision-making. The core of existing learning path planning schemes lies in identifying weaknesses and pushing standardized exercises or video resources. Its decision-making granularity remains at the macro level of "what to learn." This invention, however, refines teaching decisions to the action level of "how to teach." By generating different types of candidate teaching actions such as follow-up questions, subtle hints, and step-by-step explanations, it transforms teaching intervention from resource delivery to precise action-level decisions that adapt to the learner's current cognitive state, achieving true time-based and individualized instruction.

[0177] Third, it possesses pre-output simulation verification capabilities. This invention acquires a learner's digital twin model aligned with their current dynamic cognitive state, performs simulation verification on each candidate teaching action, predicts the learner's potential reactions after receiving each candidate teaching action, and quantitatively assesses the learning benefits and teaching risks of each candidate teaching action. This allows for the identification and elimination of teaching actions with risks such as excessive prompting, inappropriate difficulty, reinforcement of misconceptions, or cognitive overload before the action is output, intercepting the negative effects of teaching intervention before the output stage. This avoids the consumption of learners' attention, time, and learning motivation by inappropriate teaching from the source, significantly improving the safety and learning experience of the learning process.

[0178] Fourth, it possesses a risk-controllable teaching decision-making guarantee mechanism. This invention integrates teaching constraint information throughout the candidate teaching action generation, simulation verification, and action selection stages. By introducing teaching constraint information, it ensures that the final output teaching actions not only meet the learner's personalized needs but also strictly adhere to the teaching norms and requirements of the K-12 education scenario, including grade level, textbooks, curriculum standards, and the learner's current developmental level.

[0179] Fifth, it possesses a self-evolving closed-loop optimization capability. This invention compares real feedback with twin simulation prediction results, using the discrepancy between the two to continuously calibrate the dynamic cognitive state and the learner's digital twin model. This ensures that each round of interaction allows for learning from real feedback, and the teaching decision-making capability is continuously optimized as the learning process continues, forming a continuously evolving closed loop. This closed-loop mechanism means that the generation and selection of teaching actions are no longer one-time static decisions, but rather continuously self-correct and improve through ongoing interaction with the learner. Thus, the learning method possesses adaptive evolution capabilities, continuously improving the accuracy and reliability of teaching decisions as the learning process deepens.

[0180] Sixth, it possesses scalable group simulation enhancement verification capabilities. This invention can construct virtual learner groups based on similar learner samples, and perform batch simulations of candidate teaching actions to evaluate the stability of candidate teaching actions for virtual learners with different learning characteristics. This complements individual digital twin simulations, further enhancing the robustness of teaching action selection.

[0181] In summary, this invention effectively solves the technical problem that existing intelligent education solutions are not suitable for one-on-one tutoring scenarios by constructing a dynamic cognitive state oriented towards teaching decisions, generating multiple candidate teaching actions, verifying them through learner digital twin simulation and risk assessment, and continuously calibrating the cognitive state and learner digital twin model using real feedback. This transforms the teaching method from passive content generation to active teaching decision-making, significantly improving the teaching effectiveness, safety, and adaptability in one-on-one tutoring scenarios.

[0182] This invention also provides a learning companion intelligent agent system, such as... Figure 3 As shown, the learning companion intelligent agent system may include: an interactive data acquisition module 301, a learner dynamic cognitive state determination module 302, a candidate teaching action generation module 303, a learner digital twin simulation verification module 304, a teaching action evaluation and selection module 305, and a teaching action output module 306.

[0183] The interactive data acquisition module 301 is used to acquire interactive data of learners during the learning process.

[0184] The learner dynamic cognitive state determination module 302 is used to determine the current dynamic cognitive state for teaching decision-making based on interaction data.

[0185] The candidate teaching action generation module 303 is used to generate multiple candidate teaching actions based on the current dynamic cognitive state and the current learning task.

[0186] The learner digital twin simulation verification module 304 is used to acquire a learner digital twin model that is aligned with the current dynamic cognitive state, and to simulate the learner's response under each candidate teaching action based on the learner digital twin model, thereby obtaining the predicted learner response data corresponding to each candidate teaching action.

[0187] The instructional action assessment and selection module 305 is used to determine the learning benefit indicators and instructional risk indicators of each candidate instructional action based on the predicted learner response corresponding to each candidate instructional action, and to select the target instructional action from each candidate instructional action based on the learning benefit indicators and instructional risk indicators of each candidate instructional action.

[0188] The instructional action output module 306 is used to output the target instructional action to the target learner.

[0189] In one possible implementation, the learning companion intelligent agent system may further include: a feedback data acquisition module 307 and a learner state update and digital twin calibration module 308.

[0190] The feedback data acquisition module 307 is used to collect the target learner's reaction data under the target teaching action after the teaching action output module 306 outputs the target teaching action, so as to obtain the real learner's reaction data corresponding to the target teaching action.

[0191] The learner state update and digital twin calibration module 308 is used to determine the deviation between the predicted learner response data corresponding to the target teaching action and the real learner response data corresponding to the target teaching action. Based on the deviation, the current dynamic cognitive state and the learner digital twin model are updated so that the candidate teaching action generation module 303 can generate candidate teaching actions based on the updated dynamic cognitive state.

[0192] In one possible implementation, the current dynamic cognitive state includes a first cognitive state, or includes both a first cognitive state and a second cognitive state.

[0193] The first cognitive state includes some or all of the following states: knowledge mastery, typical misconceptions, and current learning obstacles.

[0194] The second cognitive state includes some or all of the following states: cognitive load, risk of forgetting, learning speed, expression preference, risk of emotional frustration, zone of proximal development, cue dependence, grade level matching, textbook version matching, and curriculum standard matching.

[0195] In one possible implementation, the learning companion intelligent agent system may also include: a teaching constraint module 300.

[0196] The teaching constraint module 300 is used to acquire teaching constraint information, wherein the teaching constraint information is used to perform at least one of the following constraints: age-appropriateness constraint, beyond-syllabus constraint, prompt intensity constraint, and learning load constraint.

[0197] When the candidate teaching action generation module 303 generates multiple candidate teaching actions based on the current dynamic cognitive state and the current learning task, it is specifically used for:

[0198] Based on the current dynamic cognitive state and the current learning task, and under the constraints of teaching constraints, multiple candidate teaching actions are generated.

[0199] The learner digital twin simulation verification module 304 is specifically used for simulating the target learner's responses to various candidate instructional actions based on the learner digital twin model, specifically for:

[0200] Based on the learner's digital twin model, and under the constraints of instructional information, the system simulates the target learner's response to each candidate instructional action.

[0201] In one possible implementation, each candidate teaching action is one of the following: follow-up questioning, gentle prompting, step-by-step explanation, example demonstration, analogy explanation, supplementation of prior knowledge, reduced difficulty exercises, variation exercises, clarification of error causes, summary and generalization, and requiring learners to continue trying.

[0202] In one possible implementation, each candidate instructional action includes some or all of the following attributes: action type, target knowledge point, cue intensity, difficulty level, instructional intent, applicable conditions, prohibition conditions, and expected learner response.

[0203] In one possible implementation, the instructional action evaluation and selection module 305, when selecting a target instructional action from among the candidate instructional actions and outputting it to the target learner based on the learning benefit indicators and instructional risk indicators of each candidate instructional action, is specifically used for:

[0204] Eliminate candidate teaching actions whose teaching risk indicators exceed the preset risk threshold.

[0205] Based on the learning benefit indicators and teaching risk indicators of the remaining candidate teaching actions, the target teaching action is selected from the remaining candidate teaching actions and output to the target learner.

[0206] In one possible implementation, when the instructional action evaluation and selection module 305 selects the target instructional action from the remaining candidate instructional actions and outputs it to the target learner based on the learning benefit indicators and instructional risk indicators of the remaining candidate instructional actions, it is specifically used for:

[0207] Based on the learning benefit indicators and teaching risk indicators of the remaining candidate teaching actions, and combined with the matching degree between the remaining candidate teaching actions and the teaching constraint information, the comprehensive score of the remaining candidate teaching actions is determined.

[0208] Based on the combined scores of the remaining candidate teaching actions, the target teaching action is selected from the remaining candidate teaching actions and output to the target learner.

[0209] In one possible implementation, the learning companion intelligent agent system may also include: a virtual learner group simulation module and a stability index determination module.

[0210] The virtual learner group simulation module is used to simulate each of the candidate teaching actions based on the virtual learner group, and obtain simulation result sets corresponding to each candidate teaching action. The virtual learner group includes multiple virtual learners constructed from multiple learner samples similar to the target learner. These virtual learners have different learning characteristics, and the simulation result sets contain the response data of each virtual learner under the corresponding candidate teaching actions.

[0211] The stability index determination module is used to determine the stability index of each candidate teaching action based on the simulation result set corresponding to each candidate teaching action.

[0212] The instructional action assessment and selection module 305, when selecting a target instructional action to output to the target learner based on the learning benefit indicators and instructional risk indicators of each candidate instructional action, is specifically used for:

[0213] Based on the learning benefit indicators and teaching risk indicators of each candidate teaching action, and combined with the stability indicators of each candidate teaching action, the target teaching action is selected from the candidate teaching actions and output to the target learner.

[0214] In one possible implementation, the learning benefit metrics include at least one of the following metrics: mastery of target knowledge points, probability of overcoming current obstacles, probability of independent completion, probability of correcting misconceptions, degree of supplementation of prior knowledge, and probability of transfer and application.

[0215] Teaching risk indicators are used to assess the risk status of at least one of the following risk dimensions: over-cue risk, excessive difficulty risk, reinforcement of misconceptions risk, excessive cognitive load risk, frustration risk, answer dependence risk, deviation from learning objectives risk, and exceeding teaching constraints risk.

[0216] The learning companion intelligent agent system provided in this invention can proactively perceive the learner's learning status and proactively plan teaching strategies during the learning companion process, rather than passively waiting for the learner to ask questions and then generating answers, thereby truly meeting the needs for continuous companionship and proactive guidance in one-on-one learning companionship.

[0217] This invention also provides an electronic device, which may include at least one processor and a memory connected to the processor.

[0218] The processor may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention; the memory may include high-speed RAM, and may also include non-volatile memory, such as at least one disk storage.

[0219] The memory is used to store computer programs, and the processor is used to execute the computer programs so that the electronic device can implement the intelligent learning companion method provided in the above embodiments.

[0220] This invention also provides a computer storage medium carrying one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can implement the intelligent learning companion method provided in the above embodiments.

[0221] This invention also provides a computer program product, including computer-readable instructions, which, when executed on an electronic device, enable the electronic device to implement the intelligent learning companion method provided in the above embodiments.

[0222] It should also be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and 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 modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the accompanying drawings of the device embodiments provided by this invention, the connection relationship between modules indicates that they have a communication connection, which can be implemented as one or more communication buses or signal lines.

[0223] Through the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus necessary general-purpose hardware, and of course, it can also be implemented by special-purpose hardware including application-specific integrated circuits, special-purpose CPUs, special-purpose memory, special-purpose components, etc. Generally, any function performed by a computer program can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be diverse, such as analog circuits, digital circuits, or special-purpose circuits. However, for the present invention, software program implementation is more often the preferred implementation method. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium, such as a computer floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk, or optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, training equipment, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0224] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.

[0225] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, training device, or data center to another website, computer, training device, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

Claims

1. An intelligent learning companion method, characterized in that, include: Acquire interaction data of the target learner during the learning process, and determine the current dynamic cognitive state of the target learner based on the interaction data; Based on the current dynamic cognitive state and the current learning task, multiple candidate teaching actions are generated; Obtain a learner digital twin model that aligns with the current dynamic cognitive state, and simulate the target learner's response under each of the candidate instructional actions based on the learner digital twin model to obtain the predicted learner response data corresponding to each of the candidate instructional actions. Based on the predicted learner response data corresponding to each of the candidate teaching actions, the learning benefit indicators and teaching risk indicators of each candidate teaching action are determined. Based on the learning benefit indicators and teaching risk indicators of each candidate teaching action, a target teaching action is selected from the candidate teaching actions and output to the target learner.

2. The intelligent learning companion method according to claim 1, characterized in that, Also includes: After outputting the target teaching action, the reaction data of the target learner under the target teaching action is collected to obtain the real learner reaction data corresponding to the target teaching action; Determine the deviation between the actual learner response data corresponding to the target instruction action and the predicted learner response data corresponding to the target instruction action; Based on the deviation, the current dynamic cognitive state and the learner's digital twin model are updated, and then the process of generating multiple candidate teaching actions and subsequent steps based on the current dynamic cognitive state and the current learning task is executed until the learning companionship termination condition is met.

3. The intelligent learning companion method according to claim 1 or 2, characterized in that, The current dynamic cognitive state includes the first cognitive state, or includes both the first and second cognitive states; The first cognitive state includes some or all of the following states: knowledge mastery status, typical misconceptions, and current learning obstacles; The second cognitive state includes some or all of the following states: cognitive load, risk of forgetting, learning speed, expression preference, risk of emotional frustration, zone of proximal development, cue dependence, grade level matching, textbook version matching, and curriculum standard matching.

4. The intelligent learning companion method according to claim 1 or 2, characterized in that, Also includes: Obtain teaching constraint information, wherein the teaching constraint information is used to impose at least one of the following constraints: age-appropriateness constraint, beyond-syllabus constraint, prompt intensity constraint, and learning load constraint; The process of generating multiple candidate teaching actions based on the current dynamic cognitive state and the current learning task includes: Based on the current dynamic cognitive state and the current learning task, multiple candidate teaching actions are generated under the constraints of the teaching constraint information; The simulation of the target learner's response to each of the candidate instructional actions based on the learner's digital twin model includes: Based on the learner digital twin model, and under the constraints of the teaching constraint information, the model simulates the target learner's response to each of the candidate teaching actions.

5. The intelligent learning companion method according to claim 1 or 2, characterized in that, Each of the candidate teaching actions is one of the following teaching actions: Follow-up questions, gentle hints, step-by-step explanations, example demonstrations, analogical explanations, supplementary prior knowledge, reduced difficulty exercises, variation exercises, clarification of error causes, summarization, and requiring learners to continue trying.

6. The intelligent learning companion method according to claim 1 or 2, characterized in that, Each of the candidate instruction actions includes some or all of the following attributes: Action type, target knowledge point, cue intensity, difficulty level, teaching intention, applicable conditions, prohibition conditions, and expected learner response.

7. The intelligent learning companion method according to claim 1 or 2, characterized in that, The step of selecting a target teaching action from among the candidate teaching actions and outputting it to the target learner based on the learning benefit indicators and teaching risk indicators of each candidate teaching action includes: Candidate teaching actions whose teaching risk indicators exceed the preset risk threshold will be excluded. Based on the learning benefit indicators and teaching risk indicators of the remaining candidate teaching actions, a target teaching action is selected from the remaining candidate teaching actions and output to the target learner.

8. The intelligent learning companion method according to claim 7, characterized in that, The step of selecting a target teaching action from the remaining candidate teaching actions and outputting it to the target learner based on the learning benefit indicators and teaching risk indicators of the remaining candidate teaching actions includes: Based on the learning benefit indicators and teaching risk indicators of the remaining candidate teaching actions, and combined with the matching degree between the remaining candidate teaching actions and the teaching constraint information, the comprehensive score of the remaining candidate teaching actions is determined. Based on the combined scores of the remaining candidate teaching actions, a target teaching action is selected from the remaining candidate teaching actions and output to the target learner.

9. The intelligent learning companion method according to claim 1 or 2, characterized in that, Also includes: Simulations are performed on each of the candidate teaching actions based on a virtual learner group to obtain simulation result sets corresponding to each candidate teaching action. The virtual learner group includes multiple virtual learners constructed based on multiple learner samples similar to the target learner. The multiple virtual learners have different learning characteristics. The simulation result sets contain the response data of the multiple virtual learners under the corresponding candidate teaching actions. Based on the simulation result sets corresponding to each of the candidate teaching actions, the stability index of each candidate teaching action is determined. The step of selecting a target teaching action from among the candidate teaching actions and outputting it to the target learner based on the learning benefit indicators and teaching risk indicators of each candidate teaching action includes: Based on the learning benefit indicators and teaching risk indicators of each candidate teaching action, and in conjunction with the stability indicators of each candidate teaching action, a target teaching action is selected from the candidate teaching actions and output to the target learner.

10. The intelligent learning companion method according to claim 1 or 2, characterized in that, The learning benefit metrics include at least one of the following metrics: mastery of target knowledge points, probability of overcoming current obstacles, probability of independent completion, probability of correcting misconceptions, degree of supplementation of prior knowledge, and probability of transfer and application. The teaching risk indicators are used to assess the risk status of at least one of the following risk dimensions: risk of over-prompting, risk of excessive difficulty, risk of reinforcing misconceptions, risk of excessive cognitive load, risk of frustration, risk of answer dependence, risk of deviating from learning objectives, and risk of exceeding teaching constraints.

11. An electronic device, characterized in that, It includes at least one processor and a memory connected to the processor, wherein: The memory is used to store computer programs; The processor is used to execute the computer program to enable the electronic device to implement the intelligent learning companion method as described in any one of claims 1 to 10.

Images