A method and system for adaptive teaching material generation based on a large model
By constructing learner models and personalized teaching paths, combining them with large language models to generate teaching materials, and conducting multi-level verification, the problem of dynamically tracking learner status in online education has been solved. This has enabled the generation of highly adaptable teaching materials, improving learning efficiency and effectiveness.
Patent Information
- Application Number
- CN202611083202.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-21
- Publication Date
- 2026-11-10
AI Technical Summary
Existing online education platforms cannot dynamically track learners' status, resulting in generated teaching materials that cannot adapt to learners' personalized needs and fail to meet the "personalized teaching" requirements.
We construct learner models to update learners' knowledge mastery, cognitive level, and learning preference characteristics in real time. Combined with the teaching syllabus and knowledge graph, we generate personalized teaching paths and use a large language model to generate teaching materials for multi-level verification and optimization feedback.
It enables the dynamic and personalized generation of teaching materials, improves learning efficiency and effectiveness, ensures the systematic and logical coherence of teaching content, and constructs a continuously optimized adaptive learning loop.
Smart Images

Figure CN123068528A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of online education technology, and in particular to an adaptive teaching material generation method and system based on a large model. Background Technology
[0002] With the popularization of online education, how to provide personalized learning experiences for learners with different backgrounds and learning paces has become a core issue in the field of educational technology. Traditional online education platforms typically use a "static" teaching material library, meaning that all learners receive pre-written and fixed textbooks, exercises, and explanations. While this approach achieves widespread distribution of educational resources, it cannot address the "personalized" teaching needs of individual learners. Learners encounter different difficulties, show different interests, and have different weaknesses in their knowledge. Using uniform materials often results in some learners not being challenged enough, while others cannot digest the material.
[0003] In recent years, large language models (hereinafter referred to as large models) have brought new possibilities to intelligent teaching due to their powerful content generation and understanding capabilities. Existing technologies have attempted to utilize large models to generate teaching questions or provide answers. However, most of these solutions are passive and one-off interactive. They typically generate an answer to a learner's current question, lacking tracking and understanding of the learner's long-term learning process. For example, when a learner asks "What is calculus?", a large model might generate a standard definition. But if it's a beginner, they might need more relatable analogies; and if it's a student who has already learned the basic concepts but is stuck on limit problems, they would need more in-depth mathematical derivations. Existing methods cannot perceive these subtle differences, resulting in generated content that, while correct, may not be "age-appropriate" or "suitable for the learner's needs."
[0004] Therefore, there is an urgent need for a method that can dynamically track learners' status and proactively plan and generate a complete set of adaptive teaching materials with teaching logic and coherence based on their current knowledge structure and learning goals. Summary of the Invention
[0005] The purpose of this invention is to provide a method and system for generating adaptive teaching materials based on a large model, which can dynamically track the learner's status and, based on their current knowledge structure and learning objectives, proactively plan and generate a complete set of adaptive teaching materials with teaching logic and coherence.
[0006] To achieve the above objectives, in a first aspect, the present invention provides an adaptive teaching material generation method based on a large model, comprising the following steps: Construct and update learner models in real time, wherein the learner model includes learner knowledge mastery vector, cognitive level label and learning preference features; Based on the updated learner model, combined with the pre-set syllabus and knowledge graph, personalized teaching objectives and achievement paths are generated, and the teaching paths consist of teaching activities. Based on the current teaching activities in the teaching path and the key state information extracted from the learner model, teaching prompt words are constructed, and the teaching prompt words are input into the large language model to generate initial teaching materials; The initial teaching materials are verified for knowledge accuracy, consistency of teaching strategies, and student suitability. If the verification fails, optimization feedback is generated to guide the large language model to make corrections until the verification passes.
[0007] This includes building and updating the learner model in real time, including: The system collects learners' basic information and historical learning data in a legal and compliant manner, and collects learners' interactive behaviors in real time through front-end tracking and back-end logging. These interactive behaviors include operational behaviors, input content, and physiological behavioral characteristics. The collected interactive behaviors are transformed into structured feature tuples and input into the diagnostic model. The diagnostic model integrates Bayesian knowledge tracing and item response theory to dynamically update the learner's knowledge mastery vector. Meanwhile, the diagnostic model dynamically adjusts learners' cognitive level labels by analyzing their cognitive requirement levels for problem-solving, and extracts and updates their learning preference features by clustering analysis of learners' historical behavioral patterns.
[0008] Based on the updated learner model, and combined with a pre-set syllabus and knowledge graph, personalized learning objectives and achievement paths are generated, including: The learner model is analyzed in depth to generate a learning status diagnostic report, which includes a list of weak knowledge points, a list of knowledge points in the zone of proximal development, and a list of knowledge points in which cognitive depth needs to be improved. According to the preset priority rules, a set of candidate teaching objectives is generated from the diagnostic report. The priority rules are, in order, remedial objectives, developmental objectives, and in-depth objectives. Each teaching objective includes the target knowledge point, the expected cognitive level, and the teaching type. After selecting a teaching objective, a sequential teaching path consisting of multiple teaching activities is gradually generated based on the path planning model of deep reinforcement learning, combined with the current learner status and the relationship between knowledge points in the knowledge graph. As learners execute the instructional path, subsequent instructional activities are dynamically adjusted based on real-time feedback to achieve adaptive optimization of the path.
[0009] Among these, key state information is extracted from the learner model, including: Based on the type of current teaching activity, target knowledge points, and expected cognitive level, extract qualitative descriptions of mastery, cognitive level labels, and historical error patterns directly related to the target knowledge points from the learner model. Simultaneously, learners' stable preference characteristics and performance summaries from prior instructional activities are extracted; The extracted raw data is converted into natural language descriptions to form student profile paragraphs, which are then used to construct subsequent prompt words.
[0010] The construction of teaching prompts further includes: The structured dynamic filling method is used to combine five modules in sequence: role and scene setting, teaching task definition, student profile description, teaching strategy guidance and output format requirements into complete teaching prompt words; The teaching strategy guidance is derived from the strategies determined in the teaching path planning, and the output format requirements are preset according to the type of teaching activity; Based on the current teaching tasks and extracted key status information, the content of each module is dynamically filled in to ensure that the teaching prompts generated each time are targeted and clear instructions.
[0011] The teaching prompts are input into a large language model to generate initial teaching materials, including: A large language model fine-tuned for teaching scenarios is used as the generation base. The corpus used for fine-tuning includes classroom transcripts, textbooks and supplementary materials, and ideal answers annotated by experts. During the generation process, the generation results are controlled by setting temperature parameters, maximum generation length, and stopping sequence. The output text is then post-processed to extract the parts that meet the format requirements and generate the initial teaching materials.
[0012] The initial teaching materials are subjected to knowledge accuracy verification, teaching strategy consistency verification, and student suitability verification, including: The basic format verification uses a rule engine to check whether the material contains the necessary components; the knowledge accuracy verification combines entity links and large model-assisted judgment to ensure that the knowledge is expressed correctly; the teaching strategy consistency verification and student suitability verification call a multi-task evaluation model to output a matching score and specific comments.
[0013] The multi-task evaluation model is based on a general model and fine-tuned through a multi-task learning approach, simultaneously learning the accuracy of knowledge, the matching degree of teaching strategies, and the student suitability. During inference, the model is guided to perform the corresponding evaluation task through the set prompt template, and the scores and optimization suggestions for each dimension are output.
[0014] The process of generating optimization feedback and guiding the large language model to make corrections includes: The scores and comments from each level of verification are compiled into a comprehensive quality report. If the report determines that the material fails, a list of optimization suggestions is generated in order of priority. The list of optimization suggestions is encapsulated as optimization feedback and returned to the construction of teaching prompts along with the original teaching task and learner model snapshot. It is then added to the new teaching prompts to guide the large language model to generate revised materials, and can be repeated multiple times until the verification is successful.
[0015] Secondly, the present invention provides an adaptive teaching material generation system based on a large model, applied to an adaptive teaching material generation method based on a large model as provided in the first aspect, comprising: The model building and updating module is used to build and update the learner model in real time. The learner model includes the learner's knowledge mastery vector, cognitive level label, and learning preference features. The teaching strategy planning module is used to generate personalized teaching objectives and achievement paths based on the updated learner model and in combination with the preset teaching syllabus and knowledge graph. The teaching path consists of teaching activities. The teaching material generation module is used to construct teaching prompt words based on the current teaching activities in the teaching path and the key state information extracted from the learner model, and input the teaching prompt words into the large language model to generate initial teaching materials. The teaching logic verification module is used to verify the accuracy of knowledge, the consistency of teaching strategies, and the suitability of students in the initial teaching materials. If the verification fails, optimization feedback is generated and the large language model is guided to make corrections until the verification passes.
[0016] This invention discloses a method and system for generating adaptive teaching materials based on a large model. It constructs and updates a multi-dimensional learner model in real time, incorporating knowledge mastery, cognitive level, and learning preferences. Based on the teaching syllabus and knowledge graph, a personalized, sequential teaching path is generated through a teaching strategy planner. Structured prompts are constructed according to the teaching path and learner key states to guide the large language model in generating initial teaching materials. A teaching logic checker performs multi-level consistency checks on the materials, generating optimization feedback to guide model correction when checks fail. Finally, the materials are pushed to learners, and feedback is collected for model updates. This invention achieves dynamic and personalized generation of teaching materials, significantly improving learning efficiency and effectiveness, ensuring the systematic and logical coherence of teaching content, and constructing a continuously optimizing adaptive learning closed loop. Attached Figure Description
[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.
[0018] Figure 1 This is a schematic diagram illustrating the steps of an adaptive teaching material generation method based on a large model according to the first embodiment of the present invention.
[0019] Figure 2 This is a simplified flowchart illustrating an adaptive teaching material generation method based on a large model provided by the present invention.
[0020] Figure 3 This is a complete flowchart of an adaptive teaching material generation method based on a large model provided by the present invention.
[0021] Figure 4 This is a structural principle diagram of an adaptive teaching material generation system based on a large model, according to the second embodiment of the present invention.
[0022] Figure 5 This is a schematic diagram of the electronic device of the present invention.
[0023] In the diagram: 101 - Model building and updating module, 102 - Teaching strategy planning module, 103 - Teaching material generation module, 104 - Teaching logic verification module. Detailed Implementation
[0024] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.
[0025] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.
[0026] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."
[0027] The first embodiment of this application is as follows: Please see Figures 1-3 This invention provides an adaptive teaching material generation method based on a large model, comprising the following steps: S1. Construct and update the learner model in real time. The learner model includes the learner's knowledge mastery vector, cognitive level label, and learning preference features.
[0028] Specifically, to ensure learners' privacy and security and comply with relevant laws and regulations (such as the Personal Information Protection Act and GDPR), the system must clearly and conspicuously display the Privacy Policy and User Agreement to learners or their guardians before collecting any data, and obtain explicit authorization and consent before data collection can proceed. The collected data is strictly limited to what is necessary for teaching and learning scenarios.
[0029] Basic Information Acquisition: Before learners register for the first time or begin learning, necessary static information is collected through a guided interface, including: grade level, subject preferences, and learning goals (such as exam preparation, interest development, etc.). This information is actively filled in by learners, who are allowed to view and modify it at any time.
[0030] Historical Learning Data Acquisition: For users with existing learning records, the system, after obtaining authorization, retrieves their past learning behavior data from the local database or secure cloud storage. This data has been anonymized, retaining only features relevant to learning analysis, such as historical answer records, completed learning units, and the number of practice sessions and accuracy rates for each knowledge point. All data is stored using encryption and with strict access permissions to prevent data leakage.
[0031] Real-time interaction data collection is crucial for dynamically updating the learner model. The system comprehensively captures various behaviors during the learning process through front-end event tracking, back-end logging, and optional multimodal awareness technology. Specific data collection methods are as follows: User behavior data collection: Using JavaScript tracking technology, the system listens for and records every click, page scroll, option selection, answer submission, video playback / pause / fast forward / rewind, document page turning, and other events on the learner's web interface (e.g., web, app). These events, along with their timestamps, are encapsulated into logs and sent to the server in real-time or near real-time. For example, when a learner clicks on an option on a practice question interface, the system records the option's ID, the time of selection, and the time taken from clicking to submission.
[0032] Input Content Collection: For open-ended questions or scenarios requiring text input, the system records the learner's text input after obtaining explicit authorization (e.g., displaying a message next to the input box stating "We will analyze your input to provide more precise guidance"). The system performs real-time anonymization of the text, such as removing personal identifying information like names and contact details, retaining only semantic features related to the knowledge content for subsequent diagnostic analysis.
[0033] Physiological and Behavioral Feature Collection: With additional authorization and device support, the system can capture learners' facial expressions and head postures at low frequency via a camera (user-activated), using computer vision models to analyze their attention span (e.g., whether they frequently leave the screen, whether they close their eyes, etc.). Simultaneously, it can collect ambient volume via a microphone (user-activated) to help determine if the learning environment is noisy. All image and audio data undergo real-time feature extraction on the local device; only anonymized analysis results (e.g., "attention score," "ambient noise level") are uploaded, without uploading the original audio and video data, to maximize privacy protection.
[0034] The learner model is designed as a data structure that evolves dynamically over time, containing three core dimensions to comprehensively characterize the learner's current state: Knowledge Mastery Vector: This is a vector space based on a knowledge graph, where each dimension corresponds to a knowledge point of the smallest granularity (such as "the discriminant of a quadratic equation" or "the inverse of the Pythagorean theorem"). The value for each knowledge point is a real number between 0 and 1, representing the learner's probability of mastering that knowledge point. The initial value can be set based on historical data or preset prior knowledge (such as the average level of the grade).
[0035] Cognitive Level Labels: Based on Bloom's Taxonomy of Educational Objectives, learners' cognitive levels are divided into six levels: memorization, comprehension, application, analysis, evaluation, and creation. The system maintains a label for the current cognitive level for each knowledge point or overall learning progress. For example, for "Newton's Second Law," a learner may have reached the "application" level (able to solve problems using the formula) but not yet the "analysis" level (able to decompose complex force scenarios). This label is automatically inferred by analyzing the type and complexity of the problems the learner solves.
[0036] Learning Preference Features: By analyzing learners' historical interaction behaviors using clustering algorithms, personalized learning style features are extracted, such as "preferring text-based explanations to videos," "preferring case studies to abstract definitions," and "preferring challenging problems to basic exercises." These features are stored in the form of tag clouds or feature vectors to guide the style selection for subsequent content generation.
[0037] Whenever new interaction data arrives, the system triggers an update to the learner model, and its workflow is as follows: Data preprocessing and feature extraction: The raw interaction logs are transformed into feature tuples that can be understood by the diagnostic model. For example, a question-answering event is transformed into structured information such as (knowledge point ID, whether it is correct, answering time, whether help was requested, and question difficulty level).
[0038] Knowledge Mastery Update: For the knowledge mastery vector, the diagnostic model employs an improved Bayesian knowledge tracing algorithm. When a learner answers correctly on a knowledge point, the model appropriately increases the mastery probability of that knowledge point based on the question's difficulty and the answering time (short-term correct answers may include guesswork); when the answer is incorrect, the probability is decreased. Simultaneously, due to the dependencies between knowledge points (such as the "predecessor-successor" relationship in a knowledge graph), an update to a knowledge point also affects the mastery probability of its adjacent knowledge points through pre-defined propagation rules. For example, if the mastery of the "quadratic formula" increases, the mastery of its predecessor, the "completing the square method," may also be slightly increased because the two are highly correlated.
[0039] Cognitive Level Updates: The diagnostic model dynamically adjusts cognitive level labels by analyzing the highest cognitive requirements of problems solved by learners recently. For example, if the system repeatedly detects that a learner has successfully answered questions requiring the "analysis" level (such as distinguishing between different concepts), the label for that knowledge point or overall cognitive level will be upgraded to "analysis." Conversely, if a learner frequently makes mistakes on questions requiring the "application" level, the label will be temporarily maintained at the "understanding" or "application" level.
[0040] Learning Preference Adjustment: The diagnostic model periodically (e.g., weekly) performs offline clustering analysis on learners' behavioral logs to identify stable behavioral patterns. For example, if it is found that learners consistently skip text and images when encountering concept explanations and repeatedly watch case video clips, the "preference for case derivation" characteristic will be strengthened; if it is found that they consistently choose more difficult questions under adaptive difficulty recommendations, the "preference for challenge" characteristic will be enhanced. These preference characteristics are used to adjust the content presentation method in subsequent teaching planning.
[0041] Through this mechanism, the learner model achieves continuous and dynamic tracking of learner states. After each interaction, the model incorporates new evidence, updates its understanding of the learner, and provides the most accurate input for the next step of instructional planning.
[0042] S2. Based on the updated learner model, combined with the preset teaching syllabus and knowledge graph, generate personalized teaching objectives and achievement paths, wherein the teaching paths consist of teaching activities.
[0043] Specifically, after completing the real-time update of the learner model, the system enters the instructional planning stage. The core of this step lies in transforming the abstract concept of "personalized instruction" into actionable instructional decisions, namely, determining "what to teach" and "how to teach." Unlike common rule-based knowledge point recommendations (such as directly recommending the knowledge points with the lowest mastery) or fixed-order course delivery in existing technologies, this invention deeply integrates the teaching syllabus, knowledge graph, and learner multi-dimensional states, generating personalized teaching objectives and achievement paths through intelligent decision-making.
[0044] Before executing the planning task, the following three types of core information need to be loaded and understood: 1. Dynamic Learner Model: This is the basis for planning and contains the most comprehensive description of the learner's current state. Specifically, it includes: Knowledge mastery vector: the probability of mastering each knowledge point, used to identify weak links and strengths.
[0045] Cognitive level labels: The Bloom's hierarchy of knowledge points or the whole set of knowledge points (memory, comprehension, application, analysis, evaluation, creation) are used to determine the learner's current depth of understanding.
[0046] Learning preference characteristics: learners' inclinations toward content presentation methods, exercise difficulty, and explanation style, such as "preferring case-driven learning" and "preferring challenging problems".
[0047] 2. Structured Knowledge Graph: This is a directed graph structure where nodes represent the smallest granularity of knowledge points, and edges define various semantic relationships between these knowledge points, mainly including: Prerequisite relationship: To learn knowledge point B, you must first master knowledge point A (e.g., "linear equation in one variable" is a prerequisite for "quadratic equation in one variable").
[0048] Containment relationship: Knowledge point A is composed of multiple sub-knowledge points B and C (e.g., "function" contains "linear function" and "quadratic function").
[0049] Similarity / Relationship: There is a logical correlation between knowledge points (such as "Pythagorean theorem" and "square root" are often used together).
[0050] Knowledge graphs provide planners with logical constraints on "what to learn".
[0051] 3. Standardized Teaching Syllabus: The teaching syllabus is a macro-level plan for the entire subject's teaching activities. It is usually pre-designed by educational experts and specifies the following: Teaching module division: The subject knowledge is divided into several teaching units (such as "basics of algebra", "introduction to geometry" and "functions"), and a recommended learning order for the units is specified.
[0052] Learning objectives for each module: Define the minimum cognitive level that learners should achieve by the end of each unit (e.g., the "Functions" unit requires reaching the "Applications" level).
[0053] Core knowledge points and key difficulties: Mark the core knowledge points that must be mastered in each module, as well as the generally recognized difficulties.
[0054] The syllabus provides the planner with a macro-level goal orientation: "Why should we learn this?"
[0055] First, a deep analysis of the learner model is performed to generate a "Learning Status Diagnostic Report." The report identifies two key areas by comparing the knowledge mastery vector with the prior relationships in the knowledge graph: Weak knowledge points set: Knowledge points whose mastery level is below a preset threshold (e.g., 0.6) are the ones that need to be addressed first.
[0056] The zone of proximal development (ZPD) refers to the knowledge points that learners have not yet mastered, but whose prerequisite knowledge points have all reached a high level of mastery (e.g., >0.8). These knowledge points are the most likely and most achievable "stretch-up" goals that learners should currently begin to learn.
[0057] At the same time, the planner also combines cognitive level labels to determine whether learners have insufficient "depth of understanding" in certain knowledge points. For example, if a learner's mastery of a knowledge point meets the standard (e.g., 0.9), but remains at the "memory" level for a long time and has never practiced "application" or "analysis" level questions, the planner will mark it as "cognitive depth needs to be improved".
[0058] This process is divided into three levels: Analysis of the degree of mastery Weakness identification: The diagnostic analyzer traverses all knowledge points in the knowledge graph and marks knowledge points with a mastery level below a preset threshold (e.g., 0.6) as "weak knowledge points". Meanwhile, knowledge points with a mastery level between 0.6 and 0.8 are marked as "knowledge points that need to be strengthened".
[0059] Knowledge Gap Tracing: By combining the prerequisite relationships of the knowledge graph, the analyzer traces the source of weaknesses. For example, if the "quadratic formula" is weak, it traces back to the mastery of its prerequisite knowledge points "completing the square" and "square roots". If the prerequisite knowledge points are also weak, the diagnosis is "basic deficiency"; if the prerequisite knowledge points are well mastered, the diagnosis is "conceptual comprehension difficulties" or "insufficient application ability".
[0060] Cognitive Level Analysis Insufficient Cognitive Depth Identification: The diagnostic analyzer retrieves the cognitive level label associated with each knowledge point and compares it with the expected level for that knowledge point in the syllabus. If a knowledge point is mastered to the standard, but its highest attainable level remains stagnant at a lower level (such as "memorization"), while the syllabus requires reaching the "application" level, it is marked as "insufficient cognitive depth".
[0061] Learning potential assessment: The analyzer assesses the learner's overall performance at various cognitive levels. For example, if the learner demonstrates strong "analysis" ability in multiple knowledge points, it is diagnosed that the learner has high learning potential and can be appropriately accelerated or challenged in subsequent planning.
[0062] Learning Preference Analysis Preference stability assessment: The analyzer performs time-series analysis on the learned preference characteristics to determine which preferences are stable and consistent (e.g., derived from preference cases over a continuous month) and which are temporary fluctuations. Stable preferences will serve as an important reference for planning, while temporary fluctuations may be caused by specific situations and will not be included in the core planning at this time.
[0063] Based on the module division and core knowledge point annotations in the teaching syllabus, an ordered set of candidate learning objectives is generated. Each learning objective is a structured triple: Teaching objectives = {target knowledge points, expected cognitive level, teaching type} Candidate targets are generated according to the following priority rules (from high to low): Remedial Goals: The planner first checks the "List of Weak Knowledge Points" in the diagnostic report. If weak knowledge points exist within the current teaching module (located according to the syllabus), the core knowledge points with the lowest mastery level are selected as the primary goals. The teaching type is set to "Remedial Teaching," and the expected cognitive level is set to the lowest cognitive level required by the module. For example, if a learner is currently in the "Functions" module and their mastery level of "linear function graphs" is only 0.3, while the module requires reaching the "Application" level, then the goals would be: {linear function graphs, application, remedial teaching}.
[0064] Developmental Goals: If there are no weak knowledge points in the current module, the planner switches to the "Zone of Proximal Development List." A knowledge point is selected from the list, and based on the module order of the syllabus, priority is given to the knowledge point that is closest to the current module and meets the syllabus's progress requirements. The teaching type is set to "New Knowledge Instruction," and the expected cognitive level is set to the initial level that this knowledge point should typically reach in a new lesson (e.g., "Understanding" or "Application"). For example, if the current module is completed, the next module in the syllabus is "Quadratic Functions," and the prerequisites for "Definition of Quadratic Functions" have been met, then the generated goal is: {Definition of Quadratic Functions, Understanding, New Knowledge Instruction}.
[0065] In-depth objectives: If there are no weaknesses or new knowledge points in the zone of proximal development (i.e., learners have already completed the syllabus ahead of schedule), the planner selects knowledge points from the "List of Knowledge Points to be Improved in Cognitive Depth." Priority is given to knowledge points closely related to current learning interests or future learning, raising their expected cognitive level by one level (e.g., from "understanding" to "application"). The teaching type is set to "In-depth Practice." For example, if the learner has a good grasp of "linear function graphs" but is only at the "memorization" level, the generated objective would be: {linear function graphs, application, in-depth practice}.
[0066] The candidate goal set typically contains 1-3 goals, ordered by priority. The planner can merge multiple goals into a single comprehensive goal (e.g., merging two related weaknesses into a single comprehensive exercise) or retain them as multiple independent goals for subsequent path planning selection.
[0067] Once a learning objective is selected, the planner needs to further plan how to achieve that objective, that is, generate a learning path—a sequence of ordered learning activities. Each learning activity corresponds to a learning behavior, such as: "concept explanation", "case demonstration", "correction and error identification", "differentiated exercises", "micro-project exploration", etc.
[0068] The planner employs a path planning model based on a graph neural network, which has been pre-trained on a large amount of historical teaching sequence data. Its workflow is as follows: Construct a teaching action space: predefine all possible teaching activity types and associate a set of attributes for each activity, such as the required time, suitable cognitive level, applicable preference type, etc.
[0069] State encoding: The information of the current learner model, teaching objectives, and relevant subgraphs in the knowledge graph is encoded into a high-dimensional state vector through a graph neural network.
[0070] Path Generation: The path planner treats path generation as a sequence decision problem. It employs a generative model based on the Transformer architecture, which has been pre-trained on massive amounts of historical teaching sequence data (high-quality teaching paths annotated by experts or automatically recorded by the system) and optimized through reinforcement learning to maximize learners' knowledge acquisition and satisfaction after completing the path.
[0071] The generation process proceeds step by step: Step 1: Initialize the path. The planner uses the current learning state and the goal as the starting context and inputs it into the model.
[0072] Step 2: Iterative prediction of the next activity. In each step, the model predicts the next most suitable activity from the instructional activity library based on the currently accumulated context (the generated sequence of activities, simulated or actual feedback from learners on these activities). The prediction considers the following factors: Instructional suitability: Is the activity suitable for the learner's current cognitive level (e.g., activities requiring "analysis" should not be recommended for learners whose cognitive level is only "memorization").
[0073] Goal-oriented: Does the activity help advance the achievement of the goal (e.g., for "remedial teaching", the model may first arrange diagnostic questions and then explain the misconceptions).
[0074] Preference consistency: Whether the activity style matches the learner's preference characteristics (e.g., for learners who prefer case studies, the model will tend to choose case demonstrations rather than pure concept explanations).
[0075] Path coherence: A reasonable teaching rhythm should be formed between activities, such as explanation before practice, and easy before difficult, avoiding logical jumps. The model uses a self-attention mechanism to capture long-range dependencies in the sequence to ensure coherence.
[0076] Step 3: Termination Judgment. At each step, the model outputs a "termination probability." When this probability exceeds a threshold, generation stops, and the current path is considered complete. Typically, a comprehensive exercise or evaluation is placed at the end of the path to verify the achievement of the objective.
[0077] The learning path is not static. As learners actually perform activities within the path, the system collects their performance data in real time (such as accuracy, time spent, and requests for help), and sends this feedback to the path planner. The planner then reassesses the current learning status based on this data and may trigger path adjustments. Advance planning: If a learner performs far better than expected in a particular exercise (e.g., answering a difficult question correctly on the first try), the planner can skip subsequent basic exercises of the same type and move directly to the next stage, shortening the learning path.
[0078] Temporary scaffolding insertion: If a learner encounters difficulties in an activity (such as answering incorrectly multiple times), the planner can insert a "scaffolded explanation" activity into the path in real time, providing additional hints or decomposed steps to help the learner overcome obstacles before continuing the original path.
[0079] Goal adjustment: If a learner shows strong interest or confusion about a certain related knowledge point during the path execution, the planner can dynamically generate a branch goal, allowing the learner to temporarily deviate from the original path to explore, and then return to the main line.
[0080] The final generated teaching path is a structured sequence, where each element is an instance of a teaching activity, including the activity type, specific content generation prompts (for use by the three main models in steps), estimated duration, and difficulty level. For example, for the objective {quadratic function graph, understanding, new knowledge instruction}, the possible generated path is: [Contextual Introduction]: Play a short video of parabolic motion to arouse interest.
[0081] [Concept Explanation]: Define quadratic functions and their graphs using everyday language.
[0082] [Case Demonstration]: Demonstrates the generation process of the y=x² graph using dynamic drawing software.
[0083] [Correct / Incorrect Analysis]: Present several common error images and ask learners to identify and explain the reasons.
[0084] [Tiered Exercises]: Provides a set of progressive exercises ranging from plotting points and determining the direction of openings to simple applications.
[0085] [Summary and Conclusion]: Guide learners to summarize the key features of the graph of a quadratic function in their own words.
[0086] Furthermore, the planner supports multi-objective fusion. When the diagnostic report shows that a learner has multiple weaknesses, the planner can try to generate a comprehensive learning objective that covers multiple weaknesses (e.g., designing a comprehensive inquiry question that involves two weak knowledge points), or generate a branching path that includes multiple parallel learning tasks, allowing learners to choose which objective to tackle first based on their interests.
[0087] The knowledge graph serves as the knowledge framework of the instructional plan, defining the inherent logical relationships between knowledge points. This system employs a semi-automated, continuously evolving construction method to ensure that the graph conforms to subject-specific standards while dynamically adapting to feedback from teaching practice.
[0088] The construction of the initial map includes: Experts define the core framework: Based on national curriculum standards, mainstream textbook catalogs, and teaching syllabi, a team of subject education experts manually draws the hierarchical structure and prerequisite relationships of knowledge points. Experts first determine the smallest granularity of knowledge points in the subject (e.g., "discriminant of a quadratic equation"), then construct "inclusion relationships" (e.g., "quadratic equation" includes "quadratic formula," "discriminant," and "Vieta's formulas") and "prerequisite relationships" (e.g., learning "quadratic formula" requires mastering "completing the square"). This stage forms version 1.0 of the knowledge graph, ensuring the accuracy of the basic framework.
[0089] Automatic Corpus Extraction and Expansion: Text materials such as textbook chapters, teaching reference books, and authoritative online course handouts undergo corpus cleaning and natural language processing. Pre-trained language models (such as BERT) are used to perform entity recognition on the text, extracting candidate knowledge points. Then, a relation extraction model is used to identify logical connections expressed in sentences (such as "depends on," "is the basis of," "includes," etc.), generating candidate relations. These candidate relations are compared and fused with expert-constructed skeletons, and after expert review, they are added to the graph, thereby enriching the graph's relation density, for example, discovering implicit relationships between "discriminative expressions" and "intersections of images and the x-axis."
[0090] Knowledge graphs are not static. As the system is applied in actual teaching, a large amount of learner learning path data will accumulate. The system uses a data-driven approach to dynamically correct the knowledge graph: Learning by association strength: By analyzing the learning sequences of a large number of learners, the system statistically analyzes the correlation between the frequency of learning two knowledge points sequentially and the improvement in mastery. If the data shows that after learning knowledge point A, a large number of learners experience a significant improvement in their learning of knowledge point B, and A and B were not originally directly related in the expert graph, the system will generate a "strongly correlated" candidate edge and submit it to experts for review. This mechanism can uncover implicit associations that experts might overlook.
[0091] Relationship weight adjustment: In prerequisite relationships, the system dynamically adjusts the "necessity" of the relationship based on the learner's learning progress. For example, if data shows that a significant proportion of learners who do not master the "completing the square method" can still master the "quadratic formula," then the necessity weight of the "completing the square method" as a prerequisite will be reduced, thus providing a more flexible learning path in the planning.
[0092] Through the above methods, the knowledge graph becomes a dynamic knowledge network that both follows subject-specific norms and integrates real-world teaching principles, providing precise constraints and guidance for subsequent planning.
[0093] After obtaining the updated learner model, the system needs a diagnostic analyzer to deeply interpret it, transforming the raw multidimensional data into actionable diagnostic conclusions.
[0094] S3. Based on the current teaching activities in the teaching path and the key state information extracted from the learner model, construct teaching prompt words and input the teaching prompt words into the large language model to generate initial teaching materials.
[0095] Specifically, key state information refers to the portion of data selected from the learner model that is most instructive for generating the current teaching activity. The extraction process is not a simple copy of the entire model, but rather a targeted extraction and semantic transformation based on the type of the current teaching activity, the target knowledge points, and the expected cognitive level. Each time initial teaching materials are generated, it follows the explicit instructions from step 2, namely the specific teaching activity to be generated. These instructions include the following elements: Activity types: such as "concept explanation", "case demonstration", "tiered exercises", etc.
[0096] Target knowledge point: For example, "the formula for finding the root of a quadratic equation in one variable".
[0097] Expected cognitive level: such as "understanding" or "application".
[0098] Teaching strategy suggestions: such as "starting from common mistakes" and "using analogical reasoning" (these strategies may originate from decisions made during path planning).
[0099] Based on the above factors, the system extracts three types of key information from the learner model: 1. Status directly related to the target knowledge point: Knowledge Mastery: Obtain the current value of the knowledge point in the knowledge mastery vector and convert it into a qualitative description. For example, if the mastery is 0.3, it can be described as "weak mastery, low accuracy rate in recent practice"; if it is 0.8, it can be described as "basic mastery, but occasional mistakes".
[0100] Cognitive Level Label: This indicates the highest cognitive level the student has currently achieved for this knowledge point. For example, if the label is "Memory," it means the student can only recall the definition; if it is "Understanding," it means the student can explain the meaning. This information is used to determine the student's current depth of understanding.
[0101] Historical Error Patterns: Retrieves typical errors related to the knowledge point from the learner's recent answer records stored in the learner model. For example, it can extract errors such as "This student often forgets to simplify the coefficients to their simplest form when simplifying the quadratic formula" or "Often confuses the discriminant with the quadratic formula." These errors are stored in the form of natural language fragments and can be directly referenced.
[0102] 2. States related to the learner's overall characteristics: Learning preference characteristics: Extract stable preference labels, such as "preferring case-driven learning", "preferring visual aids", and "preferring challenging problems", and transform them into guiding statements, such as "This student prefers to understand abstract concepts through concrete examples".
[0103] Current mood or focus (if available): If the system collects focus or mood information through optional modules (such as "current focus is low"), it can also be used as status information to adjust the length or interest of the content.
[0104] 3. States related to the context of the instructional path: Performance of preceding activities: If the current activity is a subsequent activity in the path (e.g., an exercise after an explanation), the system will extract a summary of the learner's performance in previous activities, such as "In the previous concept explanation, the student showed confusion about the 'discriminative' part and clicked for help twice." This helps to specifically reinforce weak points when generating exercises.
[0105] The extracted raw data (such as numerical values and labels) needs to be converted into natural language descriptions that a large language model can understand. This involves filling the structured data into fluent sentences using pre-defined templates. For example: Mastery level 0.3 → "This student has a weak grasp of the 'quadratic formula for solving quadratic equations,' and the accuracy rate of related exercises recently is only about 30%." Preference for "case-driven" learning → "This student prefers to understand concepts through concrete examples rather than abstract definitions." Error pattern "Forget to simplify coefficients" → "A common mistake this student made in solving problems was forgetting to cancel out the common factors in the result after substituting into the quadratic formula." These semantic descriptions are combined into a short “student profile” paragraph, which serves as an important component of subsequent prompts.
[0106] Teaching prompts serve as a bridge connecting learners' states with the large language model, and their quality directly determines the effectiveness of the generated content. This system employs a "structured + dynamic filling" prompt construction method to ensure that each generation accurately reflects the teaching intent and individual differences.
[0107] Each teaching activity's prompts consist of the following five modules, combined in sequence: Module 1: Character and Scene Setting Objective: To enable the model to clearly understand its own identity and task context.
[0108] Content: For example, "You are an experienced junior high school math teacher, providing one-on-one tutoring to a student. Please tailor teaching materials to his needs based on the student's situation and teaching requirements." Module Two: Clear Teaching Objectives Objective: To provide specific instructions generated this time, including activity type, knowledge points, and objectives.
[0109] Content: For example, "Please generate a [conceptual explanation] material about the 'quadratic formula for solving quadratic equations'. Your explanation should aim to help students reach the [understanding] level, that is, be able to explain the derivation process and applicable conditions of the formula in their own words." Module 3: Student Profile Description Objective: To embed the key state information extracted in the first step into natural language, so that the model can "see" the student's current state.
[0110] Content: For example, "This student's specific situation is as follows: He has a weak grasp of the quadratic formula (accuracy rate of about 30%), and is currently only at the stage of memorizing the formula, without understanding its derivation logic. In previous exercises, his common mistake was forgetting to simplify the coefficients after substituting them into the formula. In addition, he prefers to understand abstract concepts through specific examples." Module Four: Guidance on Teaching Strategies Objective: To integrate the strategies defined in the teaching path planning and guide the model to adopt specific teaching methods.
[0111] Content: For example, "In your explanation, please adopt the strategy of 'starting from mistakes': first, show an example of a common mistake, then analyze the root cause of the mistake, and then gradually derive the correct formula. At the same time, please combine this with a 'case-driven' style, using at least two real-life examples to illustrate the application scenarios of the formula." Module 5: Output Format Requirements Objective: To standardize the format of generated content to facilitate subsequent processing and presentation.
[0112] Content: For example, “Please organize your explanation in the following format: 1. A typical error case; 2. Error analysis and heuristic questions; 3. Step-by-step derivation of the formula (with accompanying text description); 4. Two real-life application examples; 5. Summary of key points. Please use a friendly and encouraging tone throughout the text, and avoid using overly complex mathematical terminology.” The five modules mentioned above are not fixed templates, but are dynamically generated based on the type of teaching activity and the students' situation. For example: For the "differentiated practice" activity, Module 4 may become "Please design a set of progressive practice questions, including basic questions (directly applying formulas), variant questions (requiring transformation and application), and extension questions (integrating with other knowledge)." Module 5 requires "Each question must be accompanied by a detailed explanation, which should provide hints for students' possible mistakes."
[0113] If a student's current level of concentration is low, Module 3 may add a sentence such as "This student's current level of concentration is low. Please try to make the language more vivid and interesting, and use metaphors appropriately." Module 4 may suggest "Keep the explanation length within 200 words."
[0114] The entire process of constructing the prompt words begins with obtaining the teaching task in step two, extracting key states from the learner model, matching the corresponding teaching strategies from the strategy library, and then concatenating the modules in a fixed order to form the final prompt word text.
[0115] The constructed teaching prompts are fed into a large language model, which is fine-tuned for specific teaching scenarios to generate high-quality content that conforms to teaching standards.
[0116] This system uses a general-purpose large language model (such as the GPT series, LLaMA, etc.) as its foundation and performs supervised fine-tuning on it using a self-built teaching corpus. The teaching corpus includes: High-quality lesson plans and classroom transcripts: Award-winning lesson plans and transcripts of renowned teachers' classes obtained from public channels.
[0117] Textbooks and supplementary teaching materials: Authorized textbook chapters, exercise solutions, and concept explanations.
[0118] Highly rated answers from online education platforms: These are detailed explanations provided by teachers to students in educational Q&A communities.
[0119] "Ideal Answers" Annotated by Experts: Sample texts written by subject matter experts for common teaching tasks.
[0120] The goal of fine-tuning is to teach the model how to provide heuristic instruction from a teacher's perspective, how to adjust language complexity for students with different cognitive levels, and how to naturally incorporate teaching techniques such as examples, analogies, and questioning. The fine-tuned model outperforms the general-purpose model in terms of the accuracy, educational value, and accessibility of the generated content.
[0121] After inputting the prompt words into the model, the generation quality can be controlled by setting the following parameters: Temperature parameter: Set to medium to low (e.g., 0.7) to maintain content diversity while avoiding excessive randomness that could lead to deviation from the theme.
[0122] Maximum generated length: Preset based on activity type (e.g., concept explanation no more than 500 words, exercise question analysis no more than 300 words) to prevent content from being too long.
[0123] Stop sequence: Set specific stop words (such as "in conclusion", "next") to ensure that the generated content ends naturally after the requirements are met.
[0124] The model outputs the raw text. The system performs initial post-processing: removing extra spaces and line breaks, checking for sensitive or inappropriate content, and extracting parts that conform to the output format requirements (e.g., marking error cases, derivations, and instances separately). If the model output does not fully comply with the format requirements, the system can regenerate it or correct it using rules.
[0125] S4. Verify the accuracy of knowledge, the consistency of teaching strategies, and the suitability of students in the initial teaching materials. If the verification fails, generate optimization feedback and guide the large language model to make corrections until the verification passes.
[0126] Specifically, before executing the verification task, the following four types of information need to be received as evaluation criteria: Initial teaching materials: The specific text content generated in step three, such as a concept explanation, a set of exercises and their solutions.
[0127] Original teaching task: A description of the current activity from the teaching path in step two, including activity type, target knowledge points, expected cognitive level, and teaching strategy guidance. This is the benchmark for verification.
[0128] Current learner model snapshot: Key information about the learner's state used when generating the material, such as mastery of the target knowledge points, historical error patterns, and learning preferences. Used to assess whether the material is truly suitable for the student.
[0129] Knowledge graph and syllabus excerpt: Local knowledge graphs related to the target knowledge point (precedence and succession relationships, inclusion relationships) and the basic requirements for this knowledge point in the syllabus (such as whether it is an important or difficult point).
[0130] The core objective of validation is to ensure that the generated teaching materials meet preset standards in four aspects: accuracy of knowledge, relevance to teaching objectives, rationality of teaching strategies, and suitability for students. If any aspect is deficient, the validater must output clear feedback for optimization.
[0131] The verification process is not a one-time overall assessment, but rather a hierarchical and progressive process, with each level focusing on a different quality dimension. The four levels are: basic format verification, knowledge accuracy verification, teaching strategy consistency verification, and student suitability verification.
[0132] First layer: Basic format validation This layer is a fast screening layer, using a rule engine for automated checks to ensure materials meet basic structural requirements. The rule base predefines standard format templates for various teaching activities. Conceptual explanations: should include a definition, key points, and at least one example.
[0133] Case demonstration type: should include case background, step-by-step analysis, and conclusion summary.
[0134] Tiered practice exercises: should include questions, answers, and explanations, with the explanations detailing the problem-solving process.
[0135] Error identification category: should include error presentation, error cause analysis, and correct demonstration.
[0136] The validator parses the structure of the initial teaching materials and checks whether they contain all necessary components. For example, if the teaching task requires generating "concept explanation" materials, but the generated content only contains definitions without examples, it is determined that the basic format is not passed, and the feedback is "Missing example section, please add a specific example to help with understanding".
[0137] Second layer: Knowledge accuracy verification This layer ensures that the knowledge points in the material are presented correctly and consistent with the authoritative definitions in the knowledge graph. The validator uses a combination of two methods: Entity Linking and Knowledge Graph Comparison: Key terms in the material (such as "quadratic equation formula" and "discriminant") are linked to standard knowledge point names in the knowledge graph. Descriptions of these terms in the material are then extracted and compared semantically with authoritative definitions or attributes stored in the knowledge graph. If a significant discrepancy exists in the description (e.g., stating the quadratic equation as "x equals negative b plus or minus the square root of b squared minus 4ac divided by 2," omitting the overall nature of the denominator 2a), it is considered a knowledge error.
[0138] Larger Model-Assisted Judgment: For more complex knowledge relationships (such as the logical correctness of the derivation process), a dedicated judgment model for fact-checking is invoked (also based on the larger model, but focusing on logical reasoning and factual judgment). The reasoning paragraphs in the material are compared with the derivation steps in the standard textbook, and the judgment model outputs whether there are logical jumps or errors. For example, if the material jumps directly from the general formula to the final formula when deriving the quadratic formula, omitting the completing the square step, the judgment model can judge it as "incomplete derivation, which is not conducive to student understanding."
[0139] Third layer: Consistency verification of teaching strategies This layer assesses whether the material follows the instructional strategies specified in the instructional task. This requires a deeper level of semantic understanding and therefore relies primarily on a strategy evaluation model fine-tuned for the instructional strategies. During training, this model learns the characteristic features of various instructional strategies: "Start with mistakes" strategy: Does the material first present a typical error case and then perform an attribution analysis? "Analogy teaching" strategy: Does the material use appropriate analogies to explain abstract concepts? The "heuristic questioning" strategy: Does the material contain questions that guide students to think, rather than directly imparting conclusions? The strategy evaluation model receives strategy labels (such as "starting from errors") from the teaching task and initial teaching materials, and outputs a matching score (such as a value between 0 and 1) and specific comments. For example, if the task requires "starting from errors", but the material directly presents the correct derivation at the beginning, the model may give a low score and provide feedback such as "The strategy of starting from errors is not reflected. It is recommended to present a common error case first, and then introduce the correct explanation."
[0140] Fourth layer: Student suitability verification This layer is the core of the validation process, assessing whether the material truly matches the learner's individual needs. The validater compares the initial teaching materials with the "student profile" extracted in the first step, also performed by the suitability assessment model. This model is trained to identify the matching relationship between the material content and learner characteristics. Regarding weaknesses: Does the material focus on explaining concepts that students struggle with? Does it specifically address past error patterns? For example, if students frequently make the mistake of "forgetting to simplify," does the material emphasize the simplification steps in the relevant sections? Matching cognitive level: Does the language complexity and conceptual depth of the material match the students' cognitive level? For example, for students who are only at the "memory" level, the material should not introduce complex variation analysis too early.
[0141] Alignment with learning preferences: Does the presentation style of the materials align with students' preferences? For example, for students who prefer case-driven learning, do the materials contain enough concrete examples? The suitability assessment model outputs a comprehensive suitability score, along with specific mismatches. For example, it might provide feedback such as: "The material emphasizes the relationship between the discriminant and the intersection of the image, but the student currently only needs to understand the formula itself, which may be beyond the scope of the curriculum; in addition, the material is mainly text-based, while the student prefers case-driven learning, so it is recommended to add two real-life application examples." The evaluation models used in the above verification process (including the knowledge accuracy evaluation model, strategy evaluation model, and suitability evaluation model) are not multiple independent models, but a unified multi-task evaluation model. This model is fine-tuned based on the general large model through multi-task learning, as specifically implemented below: 1. Construction of training data for multi-task learning A team of subject education experts conducted multi-dimensional joint annotation of a large number of history teaching material samples. Each sample (i.e., one teaching material) simultaneously received multiple tags: Knowledge accuracy label: A three-level rating system is used (accurate, mostly accurate but with omissions, and incorrect), with the location of errors marked.
[0142] Strategy matching degree label: For each preset teaching strategy (such as "starting from mistakes", "analogy teaching", "heuristic questioning"), mark whether the material reflects the strategy and the degree of reflection (strong, medium, weak).
[0143] Student fit labeling: Given a simulated student profile (including mastery, cognitive level, preferences, and historical errors), experts assess the fit between the material and the profile (high, medium, or low) and write brief fit comments.
[0144] Each sample therefore corresponds to multiple learning tasks: a knowledge accuracy classification task, multiple policy existence judgment tasks, an adaptive regression or classification task, and a comment generation task.
[0145] Model architecture and training methods Based on the general large-scale Transformer architecture, keeping the main body unchanged, the output layer is modified as follows: The shared underlying semantic understanding layer (shared by all tasks) enables the model to learn common semantic representations across different tasks.
[0146] Multiple task-specific output heads are set at the top level, each responsible for a specific prediction task. For example, one head outputs the probability distribution of knowledge accuracy, multiple heads output the matching score of each policy respectively, and another head generates an adaptation evaluation through a text decoder.
[0147] During training, the loss functions of multiple tasks are weighted and summed to jointly optimize the model parameters. This multi-task joint training approach allows the model to consider multiple evaluation criteria simultaneously when learning and updating parameters. Different tasks promote each other and share knowledge, thereby improving the generalization ability and accuracy of each individual task.
[0148] In practical applications, the multi-task evaluation model activates corresponding evaluation capabilities through specific prompt templates and outputs structured scores and optimization suggestions. The specific implementation is as follows: Task activation mechanism When the instructional logic verifier needs to evaluate an initial set of instructional materials, it constructs a corresponding evaluation prompt template based on the current verification level to be performed (such as knowledge accuracy verification, instructional strategy consistency verification, or student suitability verification). Each template contains three core elements: Task instructions: Clearly state the type of evaluation task the model needs to perform. For example, "As a middle school math teacher, evaluate the accuracy of knowledge in the following teaching materials" or "Determine whether the material uses a 'starting from error' teaching strategy and explain your reasoning."
[0149] Input materials: Attach the complete initial teaching materials to be evaluated to the prompt words.
[0150] Output format requirements: Specify the structure of the model output, which typically requires a quantitative score (e.g., 0-10 points or four levels: Excellent / Good / Average / Poor) and optimization suggestions in text form. For example, "Please output in the following format: Knowledge accuracy score: [Score]; Problems: [Description]; Optimization suggestions: [Specific modification suggestions]".
[0151] Although the model learns multiple tasks during training, it does not need to explicitly switch between different output heads during inference. Through carefully designed cue templates, the model can automatically activate the corresponding task's knowledge representation based on the instruction semantics. This is because, during multi-task fine-tuning, the model learns to associate different instruction patterns with different task objectives. For example: When the prompt "assess the accuracy of knowledge" appears, the model tends to invoke the parameter patterns related to fact-checking learned during training.
[0152] When the question "judgment of teaching strategies" appears, the model activates representations related to pedagogical analysis.
[0153] When the "Assess Student Fit" prompt appears along with a student profile, the model invokes its personalized matching capabilities.
[0154] The model's raw output is a natural language text containing scores and optimization suggestions. The teaching logic validator extracts the following from the output text using regular expressions or simple keyword matching: Quantitative scoring: such as "Knowledge accuracy score: 8 points", which is converted into a numerical value and used for threshold judgment.
[0155] Qualitative conclusion: such as "the material is basically accurate, but there are leaps in the derivation steps".
[0156] Optimization suggestions: For example, "It is suggested to add intermediate steps of the completing the square method in the formula derivation section to help students understand it."
[0157] The extracted information is encapsulated into structured evaluation objects and ultimately summarized in the comprehensive quality report. If a certain verification level requires more refined multi-dimensional evaluation (e.g., the consistency of teaching strategies requires simultaneous evaluation of multiple strategy labels), the model will also output multiple scoring items at once according to the prompt template requirements. This prompt template-based inference control method allows a unified model to flexibly handle various different evaluation needs without the need to deploy a separate model for each task, greatly simplifying the system architecture. Furthermore, because the output format is strictly constrained, subsequent automated processing can reliably parse the model output, achieving complete automation of the verification process.
[0158] After completing the above four levels of verification, the instructional logic verifier will summarize the scores and comments from each level to form a comprehensive quality report. The report includes: Overall conclusion: Pass / Fail (generally requires all levels to score above the preset threshold and no fatal errors).
[0159] Scores and comments at each level: For example, basic format: pass; knowledge accuracy: 0.9, no errors; teaching strategy consistency: 0.6, not fully reflected "starting from errors"; student fit: 0.7, ignored historical error patterns.
[0160] Optimization suggestion list: Specific modification suggestions sorted by priority, for example: Please insert a common mistake the student made, "forgetting to simplify coefficients," at the beginning of your explanation, and analyze the reasons for the mistake.
[0161] Add a real-life example to explain the application of the quadratic formula in practical problems.
[0162] This paper simplifies the discussion on the relationship between the discriminant and the image, focusing instead on the derivation and application of the formula itself.
[0163] If the comprehensive report determines that the material fails, the report is packaged into optimization feedback and returned to step three along with the original teaching task and learner model snapshot. The prompt word assembler in step three uses the optimization feedback as new input (e.g., appending a sentence like "Please modify the material based on the following feedback:..." after module four "Teaching Strategy Guidance") to guide the large language model in generating a revised version. This process can be repeated multiple times until validation passes or the maximum number of iterations is reached.
[0164] The validated final teaching materials are presented to learners through graphical interfaces, audio, and other means. The learner's learning process (such as reading time, answering exercises, and click behavior) will be collected in real time as new interactive data and fed back to the learner model in step one for the next round of status updates, thus forming a continuously optimized and dynamically adaptive learning loop.
[0165] The second embodiment of this application is as follows: Please see Figure 4 This invention provides an adaptive teaching material generation system based on a large model, applied to an adaptive teaching material generation method based on a large model as provided in the first embodiment, comprising: The model building and updating module is used to build and update the learner model in real time. The learner model includes the learner's knowledge mastery vector, cognitive level label, and learning preference features. The teaching strategy planning module is used to generate personalized teaching objectives and achievement paths based on the updated learner model and in combination with the preset teaching syllabus and knowledge graph. The teaching path consists of teaching activities. The teaching material generation module is used to construct teaching prompt words based on the current teaching activities in the teaching path and the key state information extracted from the learner model, and input the teaching prompt words into the large language model to generate initial teaching materials. The teaching logic verification module is used to verify the accuracy of knowledge, the consistency of teaching strategies, and the suitability of students in the initial teaching materials. If the verification fails, optimization feedback is generated and the large language model is guided to make corrections until the verification passes.
[0166] Regarding the system in the above embodiments, the specific manner in which each module performs its operations has been described in detail in the embodiments related to the method, and will not be elaborated upon here.
[0167] For the system embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. 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 application according to actual needs. Those skilled in the art can understand and implement this without creative effort.
[0168] Accordingly, this application also provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; and, when the one or more programs are executed by the one or more processors, causing the one or more processors to implement the adaptive teaching material generation method based on a large model as described above. Figure 5 The diagram shown is a hardware structure diagram of any device with data processing capabilities, used in an embodiment of the present invention to provide an adaptive teaching material generation system based on a large model. (Except for...) Figure 5 In addition to the processor, memory, and network interface shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.
[0169] Accordingly, this application also provides a computer-readable storage medium storing computer instructions, which, when executed by a processor, implement the aforementioned method for generating adaptive teaching materials based on a large model. The computer-readable storage medium can be an internal storage unit of any data-processing device as described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data-processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data-processing device, and can also be used to temporarily store data that has been output or will be output.
[0170] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.
[0171] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.
Claims
1. A method for generating adaptive teaching materials based on a large model, characterized in that, Includes the following steps: Construct and update learner models in real time, wherein the learner model includes learner knowledge mastery vector, cognitive level label and learning preference features; Based on the updated learner model, combined with the pre-set syllabus and knowledge graph, personalized teaching objectives and achievement paths are generated, and the teaching paths consist of teaching activities. Based on the current teaching activities in the teaching path and the key state information extracted from the learner model, teaching prompt words are constructed, and the teaching prompt words are input into the large language model to generate initial teaching materials; The initial teaching materials are verified for knowledge accuracy, consistency of teaching strategies, and student suitability. If the verification fails, optimization feedback is generated to guide the large language model to make corrections until the verification passes.
2. The adaptive teaching material generation method based on a large model as described in claim 1, characterized in that, Build and update learner models in real time, including: The system collects learners' basic information and historical learning data in a legal and compliant manner, and collects learners' interactive behaviors in real time through front-end tracking and back-end logging. These interactive behaviors include operational behaviors, input content, and physiological behavioral characteristics. The collected interactive behaviors are transformed into structured feature tuples and input into the diagnostic model. The diagnostic model integrates Bayesian knowledge tracing and item response theory to dynamically update the learner's knowledge mastery vector. Meanwhile, the diagnostic model dynamically adjusts learners' cognitive level labels by analyzing their cognitive requirement levels for problem-solving, and extracts and updates their learning preference features by clustering analysis of learners' historical behavioral patterns.
3. The adaptive teaching material generation method based on a large model as described in claim 1, characterized in that, Based on the updated learner model, and combined with the pre-set syllabus and knowledge graph, personalized learning objectives and achievement paths are generated, including: The learner model is analyzed in depth to generate a learning status diagnostic report, which includes a list of weak knowledge points, a list of knowledge points in the zone of proximal development, and a list of knowledge points in which cognitive depth needs to be improved. According to the preset priority rules, a set of candidate teaching objectives is generated from the diagnostic report. The priority rules are, in order, remedial objectives, developmental objectives, and in-depth objectives. Each teaching objective includes the target knowledge point, the expected cognitive level, and the teaching type. After selecting a teaching objective, a sequential teaching path consisting of multiple teaching activities is gradually generated based on the path planning model of deep reinforcement learning, combined with the current learner status and the relationship between knowledge points in the knowledge graph. As learners execute the instructional path, subsequent instructional activities are dynamically adjusted based on real-time feedback to achieve adaptive optimization of the path.
4. The adaptive teaching material generation method based on a large model as described in claim 1, characterized in that, Extract key state information from the learner model, including: Based on the type of current teaching activity, target knowledge points, and expected cognitive level, extract qualitative descriptions of mastery, cognitive level labels, and historical error patterns directly related to the target knowledge points from the learner model. Simultaneously, learners' stable preference characteristics and performance summaries from prior instructional activities are extracted; The extracted raw data is converted into natural language descriptions to form student profile paragraphs, which are then used to construct subsequent prompt words.
5. The adaptive teaching material generation method based on a large model as described in claim 1, characterized in that, Constructing teaching prompts, further including: The structured dynamic filling method is used to combine five modules in sequence: role and scene setting, teaching task definition, student profile description, teaching strategy guidance and output format requirements into complete teaching prompt words; The teaching strategy guidance is derived from the strategies determined in the teaching path planning, and the output format requirements are preset according to the type of teaching activity; Based on the current teaching tasks and extracted key status information, the content of each module is dynamically filled in to ensure that the teaching prompts generated each time are targeted and clear instructions.
6. The adaptive teaching material generation method based on a large model as described in claim 1, characterized in that, The teaching prompts are input into the large language model to generate initial teaching materials, including: A large language model fine-tuned for teaching scenarios is used as the generation base. The corpus used for fine-tuning includes classroom transcripts, textbooks and supplementary materials, and ideal answers annotated by experts. During the generation process, the generation results are controlled by setting temperature parameters, maximum generation length, and stopping sequence. The output text is then post-processed to extract the parts that meet the format requirements and generate the initial teaching materials.
7. The adaptive teaching material generation method based on a large model as described in claim 1, characterized in that, The initial teaching materials are subjected to knowledge accuracy verification, teaching strategy consistency verification, and student suitability verification, including: The basic format verification uses a rule engine to check whether the material contains the necessary components; the knowledge accuracy verification combines entity links and large model-assisted judgment to ensure that the knowledge is expressed correctly; the teaching strategy consistency verification and student suitability verification call a multi-task evaluation model to output a matching score and specific comments.
8. The adaptive teaching material generation method based on a large model as described in claim 7, characterized in that, The multi-task evaluation model is based on a general model and fine-tuned through a multi-task learning approach, simultaneously learning the accuracy of knowledge, the matching degree of teaching strategies, and the student suitability. During inference, the model is guided to perform the corresponding evaluation task through the set prompt template, and the scores and optimization suggestions for each dimension are output.
9. The adaptive teaching material generation method based on a large model as described in claim 1, characterized in that, Generate optimization feedback and guide the large language model to make corrections, including: The scores and comments from each level of verification are compiled into a comprehensive quality report. If the report determines that the material fails, a list of optimization suggestions is generated in order of priority. The list of optimization suggestions is encapsulated as optimization feedback and returned to the construction of teaching prompts along with the original teaching task and learner model snapshot. It is then added to the new teaching prompts to guide the large language model to generate revised materials, and can be repeated multiple times until the verification is successful.
10. A large-model-based adaptive teaching material generation system, applied to the large-model-based adaptive teaching material generation method as described in claim 1, characterized in that, include: The model building and updating module is used to build and update the learner model in real time. The learner model includes the learner's knowledge mastery vector, cognitive level label, and learning preference features. The teaching strategy planning module is used to generate personalized teaching objectives and achievement paths based on the updated learner model and in combination with the preset teaching syllabus and knowledge graph. The teaching path consists of teaching activities. The teaching material generation module is used to construct teaching prompt words based on the current teaching activities in the teaching path and the key state information extracted from the learner model, and input the teaching prompt words into the large language model to generate initial teaching materials. The teaching logic verification module is used to verify the accuracy of knowledge, the consistency of teaching strategies, and the suitability of students in the initial teaching materials. If the verification fails, optimization feedback is generated and the large language model is guided to make corrections until the verification passes.
