An AI agent-driven personalized science education system

Abstract

The application relates to an AI agent-driven personalized scientific education system, and belongs to the technical field of information science. The education system runs a main agent, a user twin agent, a learning planning platform and a graph structure database. The main agent is used for collecting the knowledge state and learning behavior of a user, constructing a user model, and training one or more user twin agents based on the user model. The user twin agent is used for simulating the learning behavior and result of the user under different learning paths. The learning planning platform generates a plurality of candidate learning paths in combination with a knowledge graph stored in the graph structure database, so as to be simulated by the user twin agent. The main agent selects an optimal path according to the simulation result and implements the optimal path into the personalized teaching process of the user. The technical scheme realizes high-adaptation-degree learning path recommendation and dynamic adjustment, and improves the intelligent level of individualized teaching.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of information science, and specifically relates to an individualized scientific education system based on AI agent driving. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, individualized education has gradually become an important research direction of intelligent education system. Traditional teaching systems often rely on unified course arrangement and fixed learning path, which is difficult to dynamically adjust teaching strategies according to individual differences, resulting in limited learning efficiency and learning experience. In recent years, intelligent education platforms based on big data analysis, knowledge graph and natural language processing have initially realized the analysis and content recommendation of user behavior, but still mainly rely on static rules or simple user portraits, lacking dynamic simulation and adaptive optimization ability. At the same time, the existing system lacks predictable modeling means for "user learning process" itself, and cannot effectively evaluate the adaptability and learning effect of a specific user for a learning path before implementing teaching. Therefore, an education system capable of simulating user learning behavior, dynamically optimizing learning path and having high-level intelligent decision-making ability is needed to realize truly individualized teaching.

[0003] After consulting relevant public technologies, the technical solution with publication number TWM431397U proposes a teaching platform system, which generates teaching animation content dynamically according to the real-time state feedback of students, reduces the learning pressure in the learning process and improves the teaching effect; the technical solution with publication number US20140295398A1 proposes an interactive teaching system, which sets multiple interactive components in the system, so that the teaching can understand the learning state of students in real time in remote online teaching and adjust the teaching method; the technical solution with publication number CN117830044B proposes an interactive teaching data management system and method based on cloud computing, which can analyze the mastery degree of students on calligraphy skills through a series of performances of students in the calligraphy teaching process.

[0004] The above technical solutions all propose various application digital technology assisted teaching planning or teaching process; but there are few related teaching technical solutions applying artificial intelligence agent technology at present.

[0005] The foregoing discussion of the background art is intended only to facilitate an understanding of the present application. It is not admitted that any of the materials referred to in this discussion are part of the common general knowledge of the person skilled in the art. SUMMARY

[0006] The application aims to provide an AI agent-driven personalized science education system; the education system runs a main agent, a user twin agent, a learning planning platform and a graph structure database. The main agent is used to collect the knowledge state and learning behavior of the user, construct a user model, and train one or more user twin agents based on the user model. The user twin agent is used to simulate the learning behavior and results of the user under different learning paths. The learning planning platform combines the user model and the knowledge graph stored in the graph structure database to generate multiple candidate learning paths for the user twin agent to simulate. The main agent selects the optimal path according to the simulation results and implements it into the user's personalized teaching process. This technical solution realizes high-adaptation learning path recommendation and dynamic adjustment, and improves the intelligent level of individualized teaching.

[0007] The application adopts the following technical solution: an AI agent-driven personalized science education system, the education system comprises;

[0008] The main agent is configured to interact with the user, collect the knowledge state, behavior characteristics and learning preferences of the user, and construct a user model;

[0009] One or more user twin agents are trained and generated by the main agent based on the user model, and are used to simulate the learning behavior and learning results of the user under a specified learning path;

[0010] The learning planning platform is configured to receive the input of the main agent and / or the user twin agent, and output a multi-stage personalized learning path and learning resources;

[0011] The learning planning platform comprises a graph structure-based database; the database provides content data support for the learning path; the main agent obtains content data support for the learning path from the learning planning platform based on the user model, formulates multiple learning paths, drives one or more user twin agents to simulate learning, and selects one or more learning paths based on the simulation learning results and implements them into the user's personalized teaching.

[0012] Preferably, the main agent comprises the following functional modules:

[0013] The large language model module is configured to perform natural language understanding and generation, realize language interaction and content expression with the user;

[0014] The context modeling and memory module is configured to record the historical interaction information, intention context and behavior preference of the user, and maintain the user state;

[0015] a graph interface module configured to connect the database, perform graph data retrieval, semantic query and learning path information extraction;

[0016] a reasoning and evaluation module configured to analyze, judge and optimize the current learning state based on the user model, database information and task target;

[0017] an interaction management module for controlling the dialogue process, information confirmation, feedback triggering and prompt delivery between the main agent and the user.

[0018] Preferably, the user model is constructed by the main agent, and the construction steps include:

[0019] collecting behavior data of the user in the learning process, the behavior data including the user's learning duration, answer accuracy, reaction time, error type and operation path, etc.;

[0020] based on the behavior data, analyzing the user's mastery level and cognitive preference at different knowledge points;

[0021] combining the user's historical learning records and current learning tasks, constructing a multi-dimensional vectorized representation representing the user's knowledge state, behavior characteristics and learning tendency;

[0022] storing the multi-dimensional vectorized representation as a user model for initializing the user twin agent and planning the personalized learning path.

[0023] Preferably, the main agent includes determining the number of user twin agents to be generated based on a plurality of consideration conditions, and the steps in this process include:

[0024] S100: analyzing the behavior characteristics in the user model to identify the volatility of the user in various learning behaviors and learning results;

[0025] S200: calculating an uncertainty score based on the volatility of performance, representing the consistency of the user's learning behavior;

[0026] S300: determining the number of user twin agents to be generated based on the uncertainty score;

[0027] Through the above consideration conditions, the generated user twin agents cover different behavior patterns that the user may have.

[0028] Preferably, the graph structure-based database includes one or more knowledge graphs; the knowledge graph is used to represent the structured knowledge relationship of learning content, and the composition of the knowledge graph includes:

[0029] A plurality of nodes, each node is used to represent an entity, the entity includes a knowledge point, a learning resource, a learning goal, a user state or a skill module, etc.;

[0030] A plurality of edges, each edge is used to represent a semantic association relationship between two nodes, the association relationship includes one of a prerequisite relationship, a dependency relationship, a similarity relationship, a feedback relationship or a teaching order relationship;

[0031] Each node and edge is attached with attribute information, the attribute information includes a label, context metadata, importance weight or vector embedding, used to support semantic retrieval and path planning.

[0032] Preferably, the main agent is further configured to formulate a plurality of candidate learning paths based on the user model and the database, comprising the following steps:

[0033] Identify the node corresponding to the current knowledge state of the user and the node corresponding to the target knowledge point in the database;

[0034] Based on the prerequisite relationship, semantic association degree and teaching order between each knowledge point in the database, a plurality of feasible paths from the current node to the target node are constructed.

[0035] The beneficial effects obtained by the present application are:

[0036] 1. The technical solution constructs a user model through a main agent, and derives a plurality of user twin agents, which can perform multi-path simulation test before teaching implementation, select the optimal path for application to the user, significantly improve the matching degree between the teaching scheme and the user learning characteristics, and realize fine and differentiated teaching strategies.

[0037] 2. The main agent of the technical solution continuously collects user behavior data during actual user learning, and dynamically adjusts the user model, so that the subsequent twin agent can more accurately simulate the user state, and the system can optimize the teaching content and rhythm in real time according to the change of the user learning state, thereby enhancing the responsiveness and adaptability of the system.

[0038] 3. The graph structure database used in the technical solution has the expression ability of knowledge point prerequisite relationship, semantic label and attribute weight, and in combination with the graph interface module and the reasoning module in the main agent, semantic understanding and rapid construction of complex knowledge paths can be realized, thereby improving the intelligent level of content scheduling and the logical integrity of path construction.

[0039] 4. The software and hardware parts of the education system of the technical solution adopt modular design, and each working module and component of the hardware part and the instructions, parameters and algorithms of the software part in the system can be conveniently replaced and / or upgraded in the later stage, thereby reducing the construction cost and maintenance cost of the system. BRIEF DESCRIPTION OF DRAWINGS

[0040] The present application can be further understood from the following description in conjunction with the accompanying drawings. The components in the drawings are not necessarily drawn to scale, emphasis instead being placed upon illustrating the principles of the embodiments. In different views, the same reference numerals designate corresponding parts.

[0041] BRIEF DESCRIPTION OF DRAWINGS: 10 - user; 100 - master agent; 110 - large language model module; 120 - context modeling and memory module; 130 - interaction management module; 140 - graph interface module; 150 - reasoning and evaluation module; 165 - learning planning platform; 170 - user model; 200 - user twin agent; 230 - database;

[0042] Figure 1 a schematic diagram of the framework of the educational system described in the present application;

[0043] Figure 2 a schematic diagram of the database of the graph structure described in the embodiment of the present application;

[0044] Figure 3 a schematic diagram of the architecture of the master agent described in the embodiment of the present application;

[0045] Figure 4 a schematic diagram of the composition of the user twin agent described in the embodiment of the present application;

[0046] Figure 5 a schematic diagram of the framework of the computer system used by the educational system described in the embodiment of the present application. DETAILED DESCRIPTION

[0047] In order to make the technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the embodiments thereof; it should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application. For those skilled in the art, other systems, methods and / or features of the embodiments will become apparent after reading the following detailed description. All such additional systems, methods, features and advantages are intended to be included within the scope of the present application. They are included within the scope of the present application and are protected by the appended claims. Additional features of the disclosed embodiments are described in the following detailed description, and will be apparent to one of ordinary skill in the art from the following detailed description.

[0048] The same or similar reference numerals in the drawings of the embodiments of the present application correspond to the same or similar components; in the description of the present application, it is understood that if the orientations or positional relationships indicated by the terms "upper", "lower", "left", "right" and the like are based on the orientations or positional relationships shown in the drawings, they are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or components referred to must have a particular orientation. The orientation and operation are constructed in a particular orientation, so the positional relationship described in the drawings is only used for exemplary illustration, and cannot be understood as a limitation on the present patent. For those skilled in the art, the specific meanings of the above terms can be understood according to the specific circumstances.

[0049] Embodiment one: illustratively, an AI agent-driven personalized scientific education system, the education system comprising;

[0050] A main agent configured to interact with a user, collect the user's knowledge state, behavior characteristics and learning preferences, and build a user model;

[0051] One or more user twin agents generated by the main agent based on the user model for simulating the user's learning behavior and learning results under a specified learning path;

[0052] A learning planning platform configured to receive the input of the main agent and / or the user twin agent, and output a multi-stage personalized learning path and learning resources;

[0053] Wherein, the learning planning platform comprises a graph structure-based database running on it; the database provides content data support for learning paths; the main agent formulates multiple learning paths based on the content data support of the learning paths provided by the database, and drives one or more user twin agents to simulate learning, and selects one or more learning paths based on the results of the simulation learning to implement personalized teaching to the user.

[0054] Preferably, the main agent comprises the following functional modules:

[0055] A large language model module configured to perform natural language understanding and generation, realize language interaction and content expression with the user;

[0056] A context modeling and memory module configured to record the user's historical interaction information, intent context and behavior preference, and maintain the user state;

[0057] A graph interface module configured to connect the database, perform graph data retrieval, semantic query and learning path related information extraction;

[0058] The reasoning and evaluation module is configured to analyze, judge and optimize the path of the current learning state based on the user model, database information and task target.

[0059] The interaction management module is used to control the dialogue process, information confirmation, feedback triggering and prompt delivery between the main agent and the user.

[0060] Preferably, the user model is constructed by the main agent, and the construction steps include:

[0061] Behavior data of the user in the learning process is collected, including the learning duration, answer accuracy, reaction time, error type and operation path of the user;

[0062] Based on the behavior data, the user's mastery level and cognitive preference at different knowledge points are analyzed;

[0063] Combined with the user's historical learning records and current learning tasks, a multi-dimensional vectorized representation representing the user's knowledge state, behavior characteristics and learning tendency is constructed;

[0064] The multi-dimensional vectorized representation is stored as a user model, which is used to drive the initialization of the user twin agent and the planning of the personalized learning path.

[0065] Preferably, the main agent includes determining the number of user twin agents to be generated based on a plurality of consideration conditions, and the steps in this process include:

[0066] S100: Analyzing the behavior characteristics in the user model to identify the volatility of the user in various learning behaviors and learning results;

[0067] S200: Calculate the uncertainty score based on the volatility of performance, which represents the consistency of the user's learning behavior;

[0068] S300: Determine the number of user twin agents to be generated based on the uncertainty score;

[0069] Through the above consideration conditions, the generated user twin agent covers different behavior patterns that the user may have.

[0070] Preferably, the database based on the graph structure includes one or more knowledge graphs; the knowledge graph is used to represent the structured knowledge relationship of the learning content, and the composition of the knowledge graph includes:

[0071] A plurality of nodes, each node representing an entity, including knowledge points, learning resources, learning goals, user states or skill modules, etc.

[0072] Multiple edges, each edge representing a semantic association between two nodes, the association including one of the following: prerequisite relationship, dependency relationship, similarity relationship, feedback relationship, or teaching sequence relationship;

[0073] Each node and edge is associated with attribute information, including labels, contextual metadata, importance weights, or vector embeddings, which are used to support semantic retrieval and path planning.

[0074] Preferably, the main intelligent agent is further configured to formulate multiple candidate learning paths based on the user model and the database, including the following steps:

[0075] Identify the node corresponding to the user's current knowledge state and the node corresponding to the target knowledge point in the database;

[0076] Based on the prerequisite relationships, semantic relevance, and teaching order among the knowledge points in the database, multiple feasible paths are constructed from the current node to the target node.

[0077] As attached Figure 1 The diagram illustrates an exemplary implementation of the educational system. The educational system operates within a computer system. This computer system may include one or more processors and memory, and runs an operating system and one or more application programs. The various components of the computer system can be physically, communicatively, and / or operationally interconnected to enable inter-component communication.

[0078] Preferably, the operating system and the application program are used to perform various functions of this execution scheme, including running one or more pre-trained AI-based main agents, and also training other agents. For example, the computer system further includes a learning planning platform 165. The learning planning platform 165 can receive input from one or more main agents and, in response to these inputs, generate corresponding learning plans and output corresponding learning content based on those plans. Furthermore, the output of the learning planning platform 165 can be fed back to the user through the computer system's output device, or the output of the learning planning platform 165 can also be used as new input to be reintroduced into the agents, achieving a dynamic closed loop of information flow.

[0079] The computer system can receive user input via an input device and provide that input to the learning planning platform 165. The computer system's output can be sent to the user's client device via a user interface.

[0080] The main agent, i.e., the artificial intelligence agent, in the technical solution specifically refers to an artificial intelligence software program for driving the personalized scientific teaching interaction process. Through training, the agent can have the following functional characteristics:

[0081] environmental perception capability, capable of collecting and identifying the learning state of the user, including behavior data (clicks, stays, browsing, etc.), voice tone, facial expression, interaction preferences, etc.;

[0082] knowledge modeling capability, based on the existing knowledge graph and learning process data of the user, constructing a user portrait based on the user, generating a personalized cognitive model of the user, and finally generating a user twin agent capable of simulating the learning and cognitive ability and feedback ability of the user;

[0083] task planning capability, i.e., dynamically generating recommended learning paths and learning tasks according to the teaching goals or the guiding learning plan of the learning planning platform, in combination with the learning state and interests of the user and / or the user twin agent of the user;

[0084] natural language interaction capability, i.e., conducting question and answer, explanation, and guided dialogue with the user through natural language to provide interactive with humans;

[0085] multimodal fusion capability, i.e., supporting content generation and output in combination with multiple forms such as images, texts, animations, and voices.

[0086] For example, to achieve deeper semantic understanding and efficient learning path recommendation, the education system further constructs a graph-based database 230 for the learning planning platform 165, for storing multi-dimensional information such as user characteristics, knowledge content, learning behavior, and ability graph, and forming the underlying knowledge graph of the education system.

[0087] Specifically, the database 230 adopts a large-scale scalable graph model structure, and organizes various education data in the form of nodes and edges. The node represents an entity, such as a knowledge point, a learning resource, a user state, etc., and the edge represents the association relationship between entities, such as a prerequisite relationship, a feedback relationship, a similarity relationship, etc., and attribute information such as context metadata, semantic labels or vector embedding can be attached to the node and the edge. The database 230 is a NoSQL database type, which stores data in the form of nodes (Nodes), edges (Edges), and properties (Properties), representing entities, relationships between entities, and attribute information, respectively. Therefore, the database 230 can efficiently store, retrieve, and analyze complex and highly interconnected data structures.

[0088] At the same time, the database 230 is connected to the learning planning platform 165, the user interaction platform 170, and the user twin agent 180, and is responsible for storing and managing the data of the education system. Figure 2As shown, user 10 can also be represented as a node within database 230. For convenience of presentation, this view only shows a region within database 230, and other users can be included in this view if the view is panned or zoomed. In this illustrative view, user 10 includes an association with other nodes, for example, including an association with a node 20 registered as a software developer.

[0089] Illustratively, the logic for the establishment and maintenance of database 230 itself includes one or more software developer nodes 20. Illustratively, the connection between developer node 20 and user 10 can be expressed as “Founded by”. And, the software developer node 20 is connected to multiple skill or professional direction forming “Sub-Field”, such as cloud computing node 21, full-stack JavaScript (JS) development node 22, mobile development node 23, backend development node 24, database development node 25, and front-end development node 26.

[0090] Further, connection paths about scientific learning are also illustratively established in database 230. Illustratively, for example, biology node 30, astronomy node 40, etc. For biology node 3, its sub-fields also establish astrophysics node 31, astrometry node 32, planetary geography node 33, etc.

[0091] And preferably, the technical personnel of the educational system described in the technical solution can create interactive and information-rich visual data representations through data visualization tools. These tools usually support multiple chart types and provide rich customization options to enhance the clarity and aesthetics of graphical expression.

[0092] And, based on the graph structure of database 230, multi-level and multi-relationship data between logical elements such as “user→age stage→learning goal” or “concept→knowledge point→learning path” can be flexibly described. Facilitating semantic retrieval and path recommendation, such as recommending to the user closely related sub-fields that have not yet been covered based on the current learning content of user 10. When the graph model structure-based database 230 is used in combination with large language models, intelligent agents can understand and use the context and structure in the graph to generate content or learning path recommendations that better meet the needs of learners.

[0093] In an exemplary implementation, the behavior of the main agent 100 is based on a Large Language Model (LLM) and a generative AI architecture. The LLM provides powerful natural language understanding and generation capabilities, enabling it to learn language knowledge from massive corpora and perform complex language reasoning. Building upon this, the main agent, combining graph-structured natural language processing (NLP) and knowledge retrieval capabilities from graph-structured databases, can efficiently acquire, understand, and reason about educational domain expertise prepared by designers or obtained through the main agent's own searches.

[0094] For example, when a user raises and discusses a scientific question, the main agent 100 can extract key concepts from the question using NLP methods, locate corresponding nodes in the knowledge graph, and obtain relevant facts or premise knowledge. Then, with the help of LLM, it can generate a fluent and easy-to-understand explanation and response from this structured knowledge. This "large model + knowledge graph" approach fully leverages the advantages of both: large language models excel at language understanding and generation, while knowledge graphs excel at accurately storing and reasoning about domain knowledge. Their integration can construct a more comprehensive and refined educational knowledge system. This combination improves the customizability and accuracy of science education content planning, providing personalized learning paths and resource recommendations while reducing the risk of factual bias or "illusions" in responses from large language models under low-constraint conditions.

[0095] For example, see attached Figure 3 As shown, the main intelligent agent 100 constructs and implements the following functional modules through software architecture or modular system design; these functional modules cooperate to support the operation of the main intelligent agent 100. Preferably, the multiple functional modules include:

[0096] The large language model module 110, as the core language model, is the foundation for the main intelligent agent 100 to perform dialogue and content generation. The large language model module 110, after pre-training and specialized optimization, undertakes the main natural language processing tasks. The large language model module is configured to convert the user's natural language input into a structured internal representation, enabling the main intelligent agent to understand the user's intent and the meaning of the expressed content, and to generate corresponding natural language responses based on internal decisions. The large language model module 110 is the foundation for the main intelligent agent's dialogue and content generation; its powerful language understanding and generation capabilities ensure the fluency and contextual coherence of the interaction with the user.

[0097] The context modeling and memory module 120 is configured to be responsible for the modeling and long-term memory of the dialogue state and the user learning ability background. When the main agent interacts with the user in real time, the user model 170 is continuously maintained by the context modeling and memory module 120, including the user's knowledge mastery, learning preferences, historical questions, and learning content given by the system. The context modeling and memory module 120 can continuously update the user portrait based on the dialogue history and the progress information obtained from the learning planning platform 165. Illustratively, the context modeling and memory module 120 records the knowledge points, links, and fields in which the user performs weakly or skillfully, and stores the previous dialogue content for subsequent reference. Based on the generation and thinking of context memory, the main agent maintains the sensitivity to the user state in long dialogue, thereby generating personalized interaction for different objects. Moreover, the context modeling and memory module 120 combines with the database 230 for memory storage, and through the context retrieval technology, relevant historical information is injected into the running context window of the large language model module 110, thereby realizing the consistency and personalization of long dialogue.

[0098] The interaction management module 130 is configured to be responsible for controlling the process and strategy of the main agent communicating with the user. The interaction management module 130 determines the interaction behavior of the main agent based on the recognition of the user's intention and the current dialogue state. Illustratively, it includes determining whether the current behavior is to answer the user's question, prompt guidance, ask a new question, or enter the explanation link, etc. The interaction management module has built-in strategy logic for educational dialogue, illustratively, for example, in the form of Socrates-style questioning strategy, error correction strategy, and incentive feedback mechanism, etc. to organize dialogue strategies to ensure the effectiveness of the interaction and comply with the teaching law. Illustratively, when the user asks a question, the interaction management module generates an answer by calling the large language model module 110 and the context modeling and memory module 120; when the user solves a problem or practices, it calls the reasoning and evaluation module 150 to judge right or wrong and give feedback; if it detects that the user is confused or deviates from the topic, it decides whether to insert a prompt, adjust the topic, or seek help from the learning planning platform to re-plan. The interaction management module ensures the coherence and teaching purpose of the dialogue process: it avoids blind response in single-round dialogue, and gradually guides the user to achieve the learning goal through multi-round dialogue.

[0099] The graph interface module 140 is configured as an interface for the main agent to interact with the database 230. The graph interface module converts natural language queries or internal decision requirements into retrieval operations on the graph database, and parses and filters the retrieval results for use by other modules. For example, when the large language model module 110 understands that the user's question involves the concept of "photosynthesis", the graph interface module constructs a corresponding graph database query, such as a SPARQL query or an API request, to retrieve the definition, attributes, or relationships with other concepts of the "photosynthesis" concept node. An example of the execution code is as follows:

[0100] PREFIX edu:<http: / / example.org / edu#>

[0101] SELECT?subprocess?label

[0102] WHERE {

[0103] edu:Photosynthesis edu:hasSubProcess?subprocess.

[0104] ?subprocess rdfs:label?label.

[0105] }

[0106] In the above code segment,

[0107] edu:Photosynthesis refers to the URI of the "photosynthesis" concept in the graph structure-based database 230;

[0108] edu:hasSubProcess refers to a custom "contains sub-process" predicate;

[0109] After the code is executed, the query returns all "sub-process" nodes connected to the "photosynthesis" node and their labels (Chinese names or terms). For example, the following list can be returned, in which the corresponding node path is returned with the http address:

[0110] subprocess label http: / / example.org / edu#LightAbsorption absorbing light energy http: / / example.org / edu#ElectronTransfer electron transfer http: / / example.org / edu#CarbonFixation carbon fixation

[0111] In combination with the structured results, the large language model module 110 generates a natural language answer, for example: "Photosynthesis mainly includes the following basic processes: first, absorbing light energy, second, electron transfer process, and finally, carbon fixation. These steps together complete the process of converting light energy into chemical energy."

[0112] For example, when the interaction management module 130 needs to arrange the next stage learning topic for the user 10, it can request the graph interface module 140 to obtain a list of unmastered nodes adjacent to the current knowledge point. The graph interface module 130 is also responsible for calling the graph structure NLP algorithm to add important information in the dialogue to the user's knowledge graph. For example, the user's mastered or misunderstood knowledge points can be marked in the graph. Through this interface, the main agent 100 realizes efficient retrieval and updating of domain knowledge, so that it has a basis for answering and decision-making.

[0113] The reasoning and evaluation module 150 is configured to perform: (1) complex reasoning calculation; (2) user learning effect evaluation. In some embodiments, when multi-step deductive reasoning or mathematical calculation is needed, the reasoning and evaluation module 150 assists the large language model module 110 to complete chain reasoning, such as decomposing problems, step-by-step solving, etc. The reasoning and evaluation module 150 can be regarded as a tool or Chain-of-Thought mechanism of the large language model. In other embodiments, for content that needs to be logically verified, the reasoning and evaluation module 150 can use a rule engine or external calculation library for verification, enhancing the accuracy of the agent's answers.

[0114] In other embodiments, the main agent 100 quantitatively analyzes and judges the user's learning performance through the reasoning and evaluation module 150. For example, for the user's exercise answers, the reasoning and evaluation module 150 can compare with the standard answers or judge true or false according to the graph knowledge point relationship; for the user's learning process data, the reasoning and evaluation module 150 performs statistics on completion rate, accuracy, reaction time, etc. Based on these evaluation results, the reasoning and evaluation module 150 generates annotations and / or information for updating the user model. For example, annotations such as "mastered" or "to be consolidated" are marked on the knowledge point nodes, and decision-making basis is provided for the interaction management module 130 and the learning planning platform 165. The reasoning and evaluation module 150 can also be configured to generate a brief learning report or a confidence score for the user and the teacher to refer to, so as to realize the measurability and interpretability of the teaching process.

[0115] In summary, the above-mentioned functional modules closely cooperate in the system architecture to ensure dialogue understanding and personalization with the large language model module 110 and the context modeling and memory module 120, the dialogue strategy and overall teaching process are taken care of by the interaction management module 130, the graph interface module 140 provides authoritative knowledge support, and the reasoning and evaluation module 150 ensures logical correctness and effect tracking.

[0116] In exemplary embodiments, the main intelligent agent 100 needs to establish a user model 170 for each user. The user model is an exemplary structured expression of information about the user's learning ability, knowledge state, behavior characteristics, and preference parameters, etc. Preferably, these user data are represented in a relational data structure or graph form, but the user model does not have execution functions, but is used as basic data by the main intelligent agent or other modules for judgment, decision-making, and matching.

[0117] Preferably, the content described by the user model includes one or more of the following:

[0118] (1) Current knowledge mastery level, which can be scored for each knowledge point to numerically represent the mastery level;

[0119] (2) Learning ability indicators, such as memory retention rate, reaction time, problem solving accuracy, etc.

[0120] (3) Behavior patterns, such as learning time period, content preference, problem completion strategy, etc.

[0121] (4) Learning goals and constraints, including exam time, target level, etc.

[0122] The above only exemplarily proposes several indicators related to the user's learning field, and is not a restrictive description of the technical solution. All indicators that can be quantified as parameters can be used as parameter items of the user model 100.

[0123] Preferably, the user model includes attribute tables, graph database node attribute sets, JSON document structures, etc. for expression and storage.

[0124] In exemplary embodiments, the user twin intelligent agent 200 refers to an intelligent agentized simulated learner constructed based on the user model of each user, i.e., a digital mapping of the user in the computer system. The user twin intelligent agent 200 has the ability to execute and "participate" in the process of learning content output by the learning planning platform, to simulate learning task execution, and to predict the user's learning behavior through execution results, reflect the user's state changes, i.e., has the functions of behavior simulation, path testing, and feedback.

[0125] And, since the user twin agent 200 itself is an artificial intelligence-based agent, it can have the ability of perception, reasoning, and task execution to perform "simulated learning" in the computer system as a user mapping, simulate the user to implement various learning paths, learning tasks, and estimate the user's mastery curve. At the same time, the behavior results based on the user twin agent 200 can optimize the learning path of the real user in reverse. And in the process of interaction between the main agent 100 and the user, the main agent 100 updates the user model in real time, so that the ability and state of the user twin agent 200 can reflect the real situation of the user as much as possible, thereby realizing digital twinning.

[0126] The main working mechanism of the main agent 100 will be described in detail below.

[0127] For example, the main agent 100 includes a closed-loop mechanism based on "perception-decision-action" execution, which dynamically updates the user model 170 and adjusts the teaching strategy in the process of continuous interaction with the user.

[0128] Among them, the user model 170 is a structured digital expression of the user's learning ability, knowledge state, behavior characteristics, and preference parameters and other information of the education system. Details will be described later.

[0129] The closed-loop mechanism can include the following execution modes:

[0130] Preferably, in the process of interaction between the main agent 100 and the user 10, the main agent 100 can update the user model 170 through multiple trigger conditions. For example, when the user answers a question, the reasoning and evaluation module 150 will determine the correctness and estimate the user's mastery of the relevant knowledge points. For example, if the user 10 answers the question correctly and quickly, it is considered that the user 10 has a high degree of mastery of the relevant knowledge points, and if it is wrong, it is recorded as a weak item. The context modeling and memory module 120 is responsible for writing the evaluation results into the user model and updating the mastery state of the knowledge point. The main agent 100 can "remember" where the user has difficulties based on the subsequent dialogue, and consciously strengthen the explanation or practice of these weak items. This running mechanism ensures that the user model gradually approaches the real situation: as the interaction increases, the user model 170 can more and more accurately reflect the user's knowledge structure and learning behavior.

[0131] Preferably, the main agent 100 includes identifying the user's intention and determining the real situation of the user deviates from the user model 170. Among them, the interaction management module 130 uses the natural language understanding ability of the large language model to analyze the intention of each input of the user in real time. For example, it is evaluated whether the user's new question is expressing confusion or deviating from the current learning topic. Combined with the user model 170, the main agent 100 can also detect the real deviation of the user model 170, that is, the situation where the predicted user state does not match the actual behavior. For example, the system expects the user to have mastered a certain concept, but the user continues to answer related questions incorrectly, indicating that the user model 170 has a real deviation. Or, the user's question exceeds the current learning plan, indicating that his interest or demand has changed. The main agent compares the expected and actual interaction performance of the user model 170 to determine whether there is a cognitive bias, interest shift, etc., and then adjusts the strategy accordingly. This process is similar to a teacher revising teaching assumptions based on classroom feedback to ensure that the teaching plan aligns with the user's actual situation.

[0132] Preferably, the main agent 100 includes decision-making on the timing of reinitiating the dialogue. For example, when the main agent 100 detects that the user has not responded for a long time, is obviously confused, or the topic is interrupted, it will actively reinitiate the dialogue. For example, if the user does not respond to the previous content for a long time, the main agent 100 can choose to send a prompt or explain again in a different way to restart the communication. If the user continues to show misunderstanding, the main agent 100 may suspend the current explanation and ask questions to find out the reason for the misunderstanding, forming a new dialogue round to clarify the concept. This dialogue restart mechanism ensures the continuity and effectiveness of the interaction, preventing the user from stopping learning in silence or misunderstanding.

[0133] Preferably, the main agent 100 decides whether to adjust the current learning task or the overall learning path based on the learning effect data provided by the reasoning and evaluation module 150 and the state of the user model 170. Preferably, the main agent 100 triggers learning strategy adjustment when the following situations occur. For example, there is a deviation between the difficulty of the learning path and the actual ability of the user, which is reflected in that if the user performs much better than expected in the current task, for example, completes the practice quickly and correctly multiple times, it means that the current learning task is too easy, and the main agent 100 can skip redundant practice or introduce higher-level content in advance; on the contrary, if the user repeatedly fails, the difficulty is too high, and the main agent 100 should go back to supplement the prerequisite knowledge or reduce the difficulty of the task. In some embodiments, when the learning plan output by the learning planning platform 165 does not match the user's actual ability assumption, for example, the learning method is not suitable or the interest point is inconsistent, the main agent 100 can interact with the learning planning platform 165 to request to re-plan the learning path to better meet the needs of the user. In some embodiments, the user shows strong interest in other topics or proposes new learning goals during the implementation of a learning plan, the main agent 100 first identifies the real intention behind the deviation behavior, confirms the new learning goal, and can interrupt the current task and cooperate with the learning planning platform 165 to reconstruct the plan. Through these mechanisms, the main agent 100 can adaptively adjust the teaching process, neither give up difficult points too early nor repeat easy points too much, dynamically optimize the next step of action according to user feedback, so that personalized learning remains effective and targeted.

[0134] During the entire operation process, the main agent 100 always operates in a closed loop based on monitoring feedback, that is, based on the cycle of "perceiving user behavior -> updating user model -> deciding strategy adjustment -> executing interactive action -> perceiving feedback again". This cycle allows the behavior of the agent to be constantly self-corrected and optimized to ensure that the teaching process always revolves around the real-time needs and long-term goals of the user.

[0135] Further, in an exemplary embodiment, the main agent 100 and the learning planning platform 165 work together through an explicit interface protocol to dynamically develop and update personalized learning paths. Preferably, the interaction mechanism between the main agent 100 and the learning planning platform 165 can be divided into the following stages:

[0136] Initial learning path acquisition: When a new learning session starts or the user establishes a learning goal, the main agent sends the user's key information to the learning planning platform 165, requesting the generation of an initial learning planning scheme. For example, the information sent includes the existing knowledge level assessment in the user model 170, learning preferences, target subjects and ranges, and other limited requirements, such as examination dates, etc. The learning planning platform 165 accordingly calls its internal course planning algorithm and rule base, combined with the knowledge graph of the learning field, analyzes and mines the user's learning needs, matches appropriate knowledge points and resources, and formulates a customized learning path scheme. For example, the learning planning platform 165 can identify a number of core concepts and skills that the user needs to master, generate a learning unit sequence in accordance with the prerequisite relationship, and each unit is accompanied by recommended learning materials and exercises. This initial scheme, together with priority order, estimated time, etc. Metadata will be returned to the main agent.

[0137] For example, in the actual teaching interaction process, the main agent 100 judges whether the original plan needs to be adjusted according to the user's performance. When the reasoning and evaluation module 150 finds that the user has mastered some knowledge points in advance or has repeatedly encountered difficulties in a certain link, the main agent will send a mid-course adjustment request to the learning planning platform. The request contains the updated user model 170 to add the mastered / weak information in real time, and update the current teaching progress. After receiving these data, the learning planning platform 165 recalculates and optimizes the path through the algorithm, including skipping mastered content, inserting previously supplemented content, or changing the number of exercises, etc., and then sends the modified path increment to the main agent. The main agent 100 accordingly connects the learning and switches the teaching strategy. For example, if the platform returns an indication that "the user has not mastered the premise Y of X, and it is recommended to supplement Y", the main agent will pause the current topic in the dialogue, change to a teaching unit that introduces Y, and then return to the original topic at an appropriate time.

[0138] Preferably, the interaction between the main agent 100 and the learning planning platform 165 is a continuous closed loop rather than one-time. After each major learning unit is completed, the main agent 100 summarizes the user's 10 test results, error types, and other learning feedback in that unit and sends them to the learning planning platform 165. The learning planning platform 165 adjusts the difficulty and pace of the subsequent path based on the latest data. For example, if the user performs well in a certain unit, the platform may reduce the number of subsequent related exercises; otherwise, it may increase the consolidation section. In this way, the learning path is calibrated and optimized once every time a stage is completed. The main agent 100 synchronizes with the learning planning platform 165 periodically to ensure that subsequent teaching decisions are always based on the latest and most accurate user model and data analysis.

[0139] To achieve the above synergy, a high-efficiency communication interface and protocol are preferred. The main agent and the learning planning platform 165 usually communicate through an internal API or message queue, using a unified data format to describe user models and learning paths. For example, a JSON format file can be used, which internally describes the user ID, knowledge point ID, state, recommended operation, etc. of the object. The learning planning platform 165 processes requests by querying the database 230 to understand knowledge point relationships or find available resources, and then makes a comprehensive decision. The above communication processing process can be transparent to the user, and the main agent 100 presents the final result to the user 10. The main agent 100 ensures that the adjustments recommended by the learning planning platform 165 are naturally integrated into the human-computer dialogue, such as guiding the user into the new recommended learning activity through dialogue.

[0140] Through the above interaction mechanism, the learning planning platform 165, as the core processing department of the overall learning planning relative to the main agent 100, is responsible for deep learning path calculation and macro planning; while the main agent 100 as the front-end execution department, implements these plans in specific teaching dialogues, and the two closely cooperate.

[0141] Further, in the exemplary embodiment, the main agent 100 makes full use of the domain knowledge graph stored in the database 230 to achieve accurate acquisition of teaching content and instant answering through a collaborative retrieval mechanism. Preferably, the collaboration mode of the main agent 100 and the database 230 includes:

[0142] The database 230 provides knowledge query and answer support for the main agent 100. When the user asks a question related to scientific knowledge or needs to obtain detailed information about a concept in the dialogue, the main agent 100 will call the graph interface module 140 to generate a corresponding retrieval request to the database 230. For example, when the user is learning the principle of photosynthesis, the main agent 100 extracts the keyword "photosynthesis" from the learning topic and locates the concept node in the knowledge graph, then extracts its definition, related reaction formula and upstream and downstream knowledge. The structured results returned by the database 230 are summarized and arranged by the main agent 100, and a coherent explanation answer is generated to the user through the large language model module. This process ensures the accuracy and reliability of the answer content, because the information comes directly from the knowledge base that has been organized and verified, rather than relying entirely on model memory. In addition, the main agent 100 can also expand related knowledge points for the user based on the association relationship provided by the database 230, such as mentioning the respiration related to photosynthesis for analogy, to enrich the coherence and systematicness of the teaching content.

[0143] Further, the graph structure of the database 230 supports the main agent 100 to perform multi-hop knowledge reasoning and association retrieval. For example, when a user raises a complex question that requires the integration of multiple concepts, such as an interdisciplinary question or a question that requires derivation, the main agent 100 can perform multi-step queries in the database 230. For example, first find the concept nodes involved in the question, and then trace the relationship paths between them layer by layer. For example, the current learning topic is "battery performance changes with the seasons", the main agent 100 can retrieve the "battery performance" node, and then find the relationship related to temperature, and reason out the chain that low temperature affects the chemical reaction rate to cause the power to drop. In each step of retrieval, the agent uses Graph-based NLP technology to convert the intermediate result into a new query constraint, and constantly expands a subgraph on the knowledge graph of the database 230, and finally finds the complete relationship link required by the question answer. Through such collaborative retrieval, the main agent 100 can analyze and decompose a deep learning topic that needs to cross multiple knowledge points, and implicitly integrate the reasoning process into the answer.

[0144] Further, the main agent 100 can acquire prior knowledge from the database 230 to generate personalized learning paths for users. When planning personalized learning content, the main agent 100 includes calling the node graph chain in the database 230 about the user to obtain the knowledge set related to the user. For example, for a weak knowledge point A marked in the user model 170, the agent can query the graph to find the prerequisite knowledge B and C of A, and the advanced application D, to determine the teaching order, such as first fill in B, C, then continue A, and finally extend to D. For example, when selecting exercises for users, the agent can retrieve a list of questions or case resources marked with the same knowledge point in the knowledge graph according to the current learning node, and then select appropriate difficulty items to push to the user. The database 230 not only supports the provision of knowledge for the question to be answered, but also serves as an index of teaching resources, and the main agent 100 can rely on the database 230 to quickly locate lecture fragments, video teaching resources, examples, etc. related to the current learning content, avoiding blind search, and greatly improving the accuracy and efficiency of personalized recommendation.

[0145] In some embodiments, the main agent 100 includes necessary maintenance of the database 230. In addition to obtaining knowledge from the database 230, the main agent 100 also updates the knowledge graph information of the database 230 under specified circumstances, forming a co-evolution mechanism. Preferably, after the user completes a knowledge evaluation, the main agent 100 can write the results back to the user's user model 170, marking the mastery level. Preferably, if the system introduces new teaching content or finds that the knowledge graph needs to be corrected and supplemented, the main agent 100 or the main agent 100 updates the database 230 in cooperation with other application programs to ensure that the graph knowledge base 230 is always authoritative and up-to-date. These updates ensure the quality of subsequent searches and the complete description of the user's portrait.

[0146] Further, in exemplary embodiments, the construction of the main agent 100 relies on the training paradigm of large language models and is customized for pre-training, fine-tuning, and continuous optimization in specific educational scenarios. Preferably, the overall generation and training process of the main agent 100 can be divided into the following stages:

[0147] Initial pre-training stage: In this stage, the core large language model of the system is pre-trained unsupervisedly using massive general corpus. For example, by training GPT series models on encyclopedic articles, textbooks, popular science materials, and Internet texts, the model learns the grammar and semantic knowledge of the language as well as extensive common sense and scientific knowledge. This process usually produces a pre-trained model with hundreds of millions or even tens of billions of parameters, with strong natural language understanding and generation capabilities, but has not yet been optimized for specific educational tasks. The pre-training stage endows the model with general intelligence and knowledge storage, providing a good starting point for downstream tasks.

[0148] Special fine-tuning stage: After obtaining the pre-trained model, the system fine-tunes the pre-trained model supervisedly to adapt to the specific role of the "educational agent" and form an agent with certain practical ability. Preferably, a large amount of domain data and interaction data will be used to further train the model in this stage, including: scientific education materials and question banks to enhance the depth of the model's professional knowledge; historical teaching dialogue corpus to simulate teacher-student dialogue and enable the model to learn teaching dialogue strategies and tone; and knowledge graph triple conversion text data to enable the agent to convert structured knowledge into explanatory language, etc.

[0149] Preferably, during the fine-tuning process, the pre-trained model is guided to learn how to explain in a hierarchical manner according to the user's questions, how to ask questions to guide students to think, when to provide hints or clues, and other teaching strategies, so that the output of the model is closer to the expected behavior pattern of the intelligent tutor. In addition, the method of human feedback reinforcement learning (RLHF) can also be used: the model's answer quality is scored and fed back by education experts, and the model parameters are further fine-tuned to make the answers both accurate and meet the teaching syllabus and ethical requirements. After fine-tuning, the main intelligent agent 100 is obtained, and the large language model module 110 of the main intelligent agent 100 is optimized in terms of knowledge accuracy, educational tone, and interaction strategy, and can better perform personalized teaching in a dialogue scenario.

[0150] Preferably, after the main intelligent agent 100 is actually deployed and runs, a mechanism for continuous learning and optimization is set up to adapt to changing user data and needs. On the one hand, small-scale learnable parameters or context adaptation are used to individualize responses. For example, the education system can maintain a dedicated preference vector or prompt template for each user, and incorporate this part into the user model 170. When generating an answer, the user's preferences are embedded in the context, so that the main intelligent agent 100 understands the style and known areas of a particular user. In this way, the main model parameters of the main intelligent agent 100 can be customized for individual users without changing the global parameters.

[0151] On the other hand, in some embodiments, continuous optimization of the global model of the main intelligent agent 100 is included, for example, the system collects a large number of real user interaction logs for periodic retraining or fine-tuning of the model. This includes analyzing logs to find that the model repeatedly makes mistakes or ineffective responses on certain types of questions, and then the development team adds corresponding training data (or corrects the knowledge graph) based on this to fine-tune the model again to correct the bias; or when new knowledge updates appear in the scientific field, the system will integrate the new knowledge into the training data to ensure that the model keeps pace with the times. Continuous optimization also includes performance tuning (such as distilling small models to speed up responses) and security tuning (such as strengthening the filtering response to bad questions and answers). In addition, the main intelligent agent may cooperate with an adaptive learning algorithm during operation: adjusting the dialogue strategy parameters according to real-time user feedback (this is more like training at the strategy level, for example, adjusting the algorithm weights for adjusting the difficulty of the questions, not directly changing the LLM parameters).

[0152] In summary, the continuous optimization mechanism ensures that the main intelligent agent becomes more intelligent as it is used: it can accumulate global experience to improve overall teaching ability, and it can also be adapted to individuals for long-term user growth.

[0153] Embodiment Two: This embodiment should be understood as at least including all the features of any one of the preceding embodiments, and further improving on the basis thereof.

[0154] For example, as shown in the accompanyingFigure 4 As shown, the user twin agent 200 is schematically illustrated to explain the principle of its construction. The user twin agent 200 is generated based on the existing user model 170. The user model 170 includes a large number of multi-aspect data description of the user 10, including:

[0155] Knowledge state, which refers to the user's mastery level at each knowledge point or skill field, which can be represented as a proficiency vector or probability distribution of knowledge concepts.

[0156] Behavioral characteristics, which refer to the typical behavior patterns of the user in the learning process, such as problem-solving speed, habitual error types, learning duration distribution, etc.

[0157] Learning preferences, which refer to the user's preferences for learning content and methods, such as preference for image-based explanation or text-based explanation, preference for self-directed exploration or guided teaching, and tendency for exercise difficulty, etc.

[0158] Further, in some embodiments, one or more user twin agents 200 are trained using the master agent 100. The goal of the training is to make the twin agent's response in the dynamic learning environment consistent with the real user, i.e., to fit the learning performance of the real user. Preferably, the training process includes the following stages:

[0159] First, the initial parameters of the user twin agent 200 are set. Exemplarily, based on the information of the user model 170, the initial parameters of the user twin agent 200 are extracted from the user model 170 by parameter copying or mapping. In other words, when creating the user twin agent 200, the knowledge mastery of the user 10 is filled into the corresponding internal representation of the user twin agent 200, such as the initial value of the weight of the neural network or the weight of the knowledge graph node, to convert the behavior characteristics of the user 10 into the behavior strategy parameters of the user twin agent 200, such as decision threshold, reaction delay distribution, etc., and to convert the learning preferences into the tendency configuration of the twin agent in the simulation to select learning resources or paths. After this initialization, the twin agent should be highly consistent with the user in the known characteristics at the starting state, laying a foundation for subsequent simulation training.

[0160] Then, the user twin agent 200 is driven by the master agent to perform simulation learning tasks, which can include performing a series of learning tasks or exercise questions, preferably, which can be the learning tasks or exercise questions that have been performed by the real user. The user twin agent 200 responds to each task based on its current training level, such as the actual time used to answer the question, the number of correct or incorrect answers, etc.

[0161] Further, the fitting parameter update of the user twin agent 200 is performed. The host agent 100 compares the simulation result of the user twin agent 200 with the expected user performance, calculates the error signal, and adjusts the internal parameters of the user twin agent 200 accordingly, so that the output of the twin agent is closer to the real behavior of the mapped user. In some embodiments, such adjustment can adopt the method of supervised learning or reinforcement learning. Preferably, for continuous numerical differences (such as completion time error), gradient descent can be used to minimize mean square error; for matching of discrete decision-making such as multiple-choice questions, a reward function can be designed to encourage the decision of the twin to be consistent with the user's choice. The host agent 100 iteratively performs the "simulation-evaluation-update" cycle to gradually reduce the deviation between the prediction of the twin and the actual performance of the user.

[0162] In an exemplary embodiment, when the user twin agent 200 is trained, that is, it has a high probability of being able to simulate the user's effect feedback after performing a learning plan at the digital level. Therefore, the host agent 100 uses the trained user twin agent 200 to run simulation on multiple candidate learning paths to evaluate the possible evolution of the user's knowledge state under different teaching strategies. The education system can use the host agent 100 in a virtual environment to instruct one or more user twin agents 200 of the user to sequentially or synchronously perform different learning content sequences or paths, and record the change trajectory of their knowledge mastery and forgetting.

[0163] For each candidate learning path, to ensure that the simulation of different learning paths is carried out under the same initial state, the host agent 100 will take a snapshot of the state of the user twin agent 200 and derive multiple copies; then drive the user twin agent 200 to complete each learning unit or exercise according to the order defined by the learning path. After completing each learning unit, record the change in the degree of mastery of the relevant knowledge points of the user twin agent 200.

[0164] Exemplarily, in an embodiment of simulation, a forgetting curve model can be introduced to simulate the decay of knowledge over time during the path learning process. In some embodiments, a decay function can be applied to the knowledge parameters of the user twin agent 200 according to the Ebbinghaus forgetting curve theory. Exemplarily, let Δt be the interval time between two learning, introduce a factor to simulate the forgetting effect. Among them, At represents the time interval between the last learning and the current time of the related knowledge point; λ represents the forgetting rate constant, the larger the value, the faster the forgetting, which can be set according to the actual situation of the user in the user model 170; and finally f(At) represents the proportion of proficiency retained by the knowledge point under the condition of time interval At. That is, the user twin agent 200 in the simulation decreases the corresponding mastery degree according to the forgetting function every time interval, which approximately simulates the phenomenon that human memory decreases over time.

[0165] After completing the entire path, record the change curve of the knowledge mastery degree of the user twin agent 200 over time (or over learning steps) and the final knowledge level and skill mastery under the path. The main agent 100 can compare the results of different paths, analyze the knowledge consolidation program and forgetting program of each learning path, and preferentially apply the learning path with higher memory efficiency.

[0166] Similarly, in some embodiments, the user twin agent 200 can be used to simulate the mastery efficiency of knowledge, such as comparing the time spent or the computational cost of different learning paths for each knowledge point to reach the same proficiency. Shorter time or lower computational cost indicates that the implementation of the learning path produces lower load and relatively light stress.

[0167] Preferably, the knowledge change and forgetting process is modeled as an explicit mathematical function or algorithm process. For example, the proficiency of the user twin agent 200 for each knowledge point is recorded using a data structure, which decreases according to the forgetting curve formula with time steps, and increases the corresponding proficiency according to the learning gain model when a learning event occurs. This multi-path simulation provides a basis for teaching path optimization and provides rich data for evaluating the consistency of the dynamic behavior of the twin body model with the actual situation of the user.

[0168] Further, in exemplary embodiments, it is necessary to verify the degree of realism of the user simulated by the user twin agent 200, so the education system includes a matching degree evaluation mechanism. Periodic learning effect evaluation can be performed with the user to obtain the real learning effect of the user, and the predicted output of the user twin agent 200 is compared with the actual performance of the real user under the same learning conditions, and the difference between the two is quantified.

[0169] For example, key comparison indicators include:

[0170] Task completion time difference: compare the difference or ratio between the time taken by the user twin agent 200 to complete a learning task and the actual time taken by the user 10. The smaller the time difference, the more accurate the simulation of the user's pace by the twin agent.

[0171] Answer accuracy deviation: Compare the difference between the simulated accuracy of the user twin agent 200 on a set of practice questions and the real accuracy of the user 10. This can be refined to the accuracy deviation of each knowledge point, which is used to check whether the mapping of the user twin agent 200 to the user's 10 strong and weak points is in place.

[0172] Learning curve similarity: Compare the learning curve of the user twin agent 200 simulation with the actual learning curve of the user 10, which can be the trajectory of the improvement of knowledge and ability with the number of exercises. For example, the dynamic time warping (DTW) distance, Pearson correlation coefficient, etc. can be used to quantify the similarity of the two curves to evaluate the fitting degree of the twin agent in the learning progress pace.

[0173] Further, for the above indicators, a comprehensive fitting score or loss function L is defined to quantify the overall matching degree. For example, the loss function L is designed as follows:

[0174] ;

[0175] In the above formula, the task time difference At and the accuracy difference acc are weighted and summed, and the smaller the value, the closer the twin agent and the user's performance. Then the loss can be converted to get the matching score S, for example, the matching score S can be taken as:

[0176] S = 1 / (1 + L),

[0177] The matching score S is normalized to 0-1, and 1 represents complete matching. In the training and evaluation stage, the main agent 100 calculates the matching score S to monitor the fidelity of the user twin agent 200. When the user twin agent 200 is accurate enough, each deviation index is small, and the matching score is close to 1, indicating that the prediction of the twin agent can almost coincide with the real data of the user.

[0178] Further, in an exemplary embodiment, when the matching score S of the user twin agent 200 is found to be lower than the preset threshold during the training process or normal operation, the main agent 100 will start the corresponding retraining or parameter correction mechanism to improve the mapping accuracy of the user twin agent 200.

[0179] Preferably, the main intelligent agent 100 performs retraining on the user twin intelligent agent 200 through an error feedback driving mode. Specifically, the main intelligent agent can collect the difference data between the simulation results of the current user twin intelligent agent 200 and the actual performance of the user, and use it as a new training sample to perform targeted additional training on the user twin intelligent agent 200. For example, if it is found that the user twin intelligent agent 200 systematically exceeds the learning rate of the user and the correct rate of examination on a certain learning content, the model can be adjusted using these cases to correct the simulation to the real performance on such topics.

[0180] In addition, in the preferred embodiment, the model parameters of the user twin intelligent agent 200 can be adjusted based on the details of the bias. For example, when the analysis finds that the knowledge decay is too slow or too fast due to improper setting of the forgetting rate of the user twin intelligent agent 200, the parameter λ of the forgetting curve can be corrected; or if the response sensitivity of the user twin intelligent agent 200 to a certain behavior characteristic of the user is too low, the feature weight of the characteristic in the user twin intelligent agent 200 can be increased, and if necessary, a new feature dimension can be introduced to enrich the input of the twin agent, so that the model can more comprehensively fit the user.

[0181] Further, in the exemplary embodiment, the main intelligent agent 100 includes multiple user twin intelligent agents 200 of the same user, which simulate learning on multiple candidate learning paths to predict the mastery efficiency, forgetting rate and learning cost of the user under different paths, and finally filter the optimal learning scheme for the real user. To ensure the balance between representativeness and computational efficiency of the simulation, the system needs to reasonably set the number of user twin intelligent agents 200 Nt and the number of candidate learning paths Np, and evaluate and filter the path results based on statistical methods.

[0182] For example, the following steps are included:

[0183] Step one, set the number of user twin intelligent agents Nt. The user twin intelligent agent is used to simulate the behavior and cognitive characteristics of the user 10 in the learning process. Considering the differences in attention, stability, consistency of knowledge mastery, etc. between different users, the system introduces a user uncertainty score U ∈ [0, 1] to evaluate the degree of behavior fluctuation. In some embodiments, the uncertainty score U is calculated based on the user's historical answer variance, learning time variation, preference fluctuation, etc.

[0184] According to the score, the main intelligent agent 100 dynamically determines the number of derived user twin intelligent agents 200, and the calculation method is as follows:

[0185] ;

[0186] In the above formula, Nt baseNt is the number of user twin agents for a user, γ is an uncertainty amplification factor, which can be set to 10 or other suitable values, and the final Nt value is calculated by rounding up the product of both γ and U.

[0187] The calculation strategy ensures that fewer user twin agents are generated for users with stable behavior to save computing resources, and the number of user twin agents is increased for users with greater behavior fluctuations to cover their potential behavior distribution, improving the reliability of simulation prediction.

[0188] Step two, set the number of candidate learning paths Np. The number of candidate learning paths Np is mainly determined based on the prerequisite path structure between the user's current knowledge state and the target knowledge point on the knowledge graph, and the allowed computing budget of the system.

[0189] In some embodiments, the main agent 100 first forms a path candidate set based on all possible prerequisite paths between the user's current knowledge node and the target node in the learning planning platform and the database 230. Then, to control complexity, the system applies restriction rules to the path candidate set, such as limiting the length of the learning path step, the length of the learning time, the learning difficulty, filtering repeated nodes or irrelevant branches. Then, combined with the simulation budget B max , the number of acceptable paths is calculated:

[0190] ;

[0191] Where B max refers to the total computing power consumption of the system in completing the process of once learning path selection, for example, B max can be the total simulation number, set to 100, and the number of user twin agents Nt is 5, then the number of learning paths allocated to each user twin agent 200 is 20.

[0192] In other embodiments, B max may refer to the total computing power or total computing time, etc.

[0193] Step three, performance analysis of the simulation results of all learning paths. For example, the main agent 100 drives all user twin agents to complete the simulation of each candidate learning path, and records the key indicators in the path execution process. Preferably, these key indicators can include:

[0194] M ij , refers to the final knowledge mastery index of the jth user twin agent under the ith learning path;

[0195] F ij, refers to the knowledge forgetting index simulated by the j-th user twin agent under the i-th learning path;

[0196] T ij , refers to the time taken for the j-th user twin agent to complete the entire learning path under the i-th learning path.

[0197] The above three key indicators are dimensionless values ​​after normalizing the original values.

[0198] Furthermore, calculate the average performance score for each learning path:

[0199] ;

[0200] In the above formula, η and ε are the forgetting penalty coefficient and the time penalty coefficient, respectively, to balance the weights of different evaluation dimensions. The specific values ​​of the two can be set by relevant technical personnel.

[0201] Preferably, the stability of the learning path results is then evaluated. The fluctuation of each path's performance under each user's twin agent is statistically analyzed, and the performance score of the i-th learning path is calculated. i Variance Var i .

[0202] Finally, the main agent 100 can determine the learning path's performance score based on all candidate learning paths. i and the variance of its score Var i A comprehensive selection process should be conducted. The selection strategy could begin by considering performance scores. i A higher learning path, while allowing variance Var within an acceptable range. i Slightly higher volatility can be reflected in choosing a learning path with high volatility but relatively high returns. Other strategies prioritize low volatility learning paths to pursue stable learning outcomes.

[0203] Example 3: This example should be understood as including at least all the features of any of the foregoing examples, and further improving upon them.

[0204] Further details are attached. Figure 5 The illustration shows an exemplary implementation of the computer system used in the education system; the computer system 500 can be applied to the data storage, calculation, and result output processes of each working module in the identification and judgment system.

[0205] By way of example, computer system 500 includes a bus 502 or other communication mechanism for communicating information, and a processor 504 coupled with bus 502 for processing information. Processor 504 can be, for example, a general purpose microprocessor.

[0206] Computer system 500 also includes a main memory 506, such as a random access memory (RAM), cache and / or other dynamic storage devices, coupled to bus 502 for storing information and instructions to be executed by processor 504. Main memory 506 also can be used for storing temporary variables or other intermediate information during execution of instructions to be executed by processor 504. Such instructions can be stored or executed in order to implement the operations described herein with regard to the management system.

[0207] Computer system 500 further can include a read only memory (ROM) 508 or other static storage device coupled to bus 502 for storing static information and instructions for processor 504. A storage device 510, such as a magnetic disk, optical disk, or USB drive (flash drive), etc., can be coupled to bus 502 for storing information and instructions.

[0208] Still further, a user input device 514, such as a keyboard, mouse or touch input, can be coupled to bus 502 for communicating information and commands, such as in the form of alphanumeric and other input, to processor 504. A display 512, such as a liquid crystal display (LCD), can also be coupled to bus 502 for displaying information to the user, such as in the form of graphical user interface (GUI) elements.

[0209] A preferred manner of interacting with the management system can be through a cursor control device 516, such as a computer mouse, or similar control / navigational mechanism.

[0210] Further, computer system 500 can also include a network interface device 518 coupled to bus 502; network interface device 518 can include, for example, a wired network adapter, a wireless network adapter, a switching chip, a router, a switch, etc.

[0211] Generally, the terms "engine," "component," "system," "database," and the like, as used herein, can refer to logical divisions of hardware or firmware embodied in circuitry, or to collections of software instructions, possibly with entry and exit points, written in programming languages such as Java, C or C++, and which can be compiled and linked into executable programs, installed in dynamic link libraries, or written in interpreted languages such as BASIC, Perl, or Python; it will be understood that software components can call each other and / or be called by other components or by themselves, and / or can be called in response to detected events or interrupts.

[0212] Software components configured to execute on a computing device can be provided on a computer- readable medium, such as an optical disc, a digital video disc, a flash drive, a magnetic disc or any other tangible medium, or as a digital download (and can initially be stored) in a compressed or installable format requiring installation, decompression or decryption before execution); such software code can be stored in whole or in part on a memory device of the executing computing device for execution by the computing device; software instructions can be embedded in firmware, such as an EPROM; it will also be appreciated that hardware components can be comprised of connected logic units (such as gates and flip-flops), and / or can be comprised of programmable units (such as programmable gate arrays or processors).

[0213] Computer system 500 includes a custom hard-wired logic, one or more ASICs or FPGAs, firmware and / or program logic that when in operation implement the techniques described herein, which are combinational logic that when implemented by the computer system 500 renders the computer system 500 a special purpose computing device.

[0214] In accordance with one or more embodiments, the techniques herein are performed by computer system 500 in response to processor 504 executing one or more sequences of one or more instructions contained in main memory 506; such instructions can be read into main memory 506 from another storage medium, such as storage device 510; execution of the sequences of instructions contained in main memory 506 causes processor 504 to perform the process steps described herein; in alternative embodiments, hard-wired circuitry can be used in place of or in combination with software instructions.

[0215] The term "non-transitory medium" and similar terms as used herein do not encompass paged, such as virtual, memory, cache, and / or BIOS.

[0216] Common forms of non-transitory media include, for example, a floppy disk, a flexible disk, a hard disk, a solid-state drive, a magnetic tape, or any other magnetic data storage medium, a CD-ROM, any other optical data storage medium, any physical medium with patterns of holes, a RAM, a PROM, and EPROM, a FLASH-EPROM, NVRAM, any other memory chip or cartridge, and a networked version of any of the above.

[0217] Non-transitory media differentiate from transmission media, which are involved with transmitting information between non-transitory media; transmission media include coaxial cables, copper wire, and fiber optic cables, including wires that comprise bus 502; transmission media can also take the form of acoustic or light waves, such as those generated during radio frequency and infrared data communications.

[0218] While this application has been described above with reference to various embodiments, it should be understood that many changes and modifications can be made without departing from the scope of this application. That is, the methods, systems, and devices discussed above are examples. Various configurations can be appropriately omitted, substituted, or added to various processes or components. For example, in alternative configurations, methods can be performed in a different order than those described, and / or various components can be added, omitted, and / or combined. Moreover, features described with respect to certain configurations can be combined in various other configurations, such as different aspects and elements of the configuration can be combined in a similar manner. Furthermore, the elements therein can be updated as the technology develops; that is, many elements are examples and do not limit the scope of this disclosure or the claims.

[0219] Specific details are provided in the specification to offer a thorough understanding of exemplary configurations, including implementations. However, configurations can be practiced without these specific details; for example, well-known circuits, processes, algorithms, structures, and techniques have been shown without unnecessary detail to avoid obscuring the configuration. This description provides only exemplary configurations and does not limit the scope, applicability, or configuration of the claims. Rather, the foregoing description of the configurations will provide those skilled in the art with an enabling description for implementing the described techniques. Various changes can be made to the function and arrangement of the elements without departing from the spirit or scope of this disclosure.

[0220] In summary, the above detailed description is intended to be illustrative rather than restrictive, and it should be understood that these embodiments are for illustrative purposes only and not for limiting the scope of protection of the invention. After reading the description of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent changes and modifications also fall within the scope defined by the claims of this invention.

Claims

1. An AI agent driven personalized science education system based on, The education system comprises: a main agent configured to interact with a user, collect the user's knowledge state, behavior characteristics and learning preferences, and build a user model; one or more user twin agents generated by the main agent based on the user model for simulating the user's learning behavior and learning results under a specified learning path; a learning planning platform configured to receive inputs of the main agent and / or the user twin agent, and output a multi-stage personalized learning path and learning resources; wherein the learning planning platform comprises a database running based on a graph structure; the database provides content data support for the learning path; the main agent formulates multiple learning paths based on the content data support of the learning path obtained from the learning planning platform based on the user model, and drives one or more user twin agents to perform simulation learning, and selects one or more learning paths based on the results of the simulation learning to implement personalized teaching to the user; the main agent comprises the following functional modules: a large language model module configured to perform natural language understanding and generation, realize language interaction and content expression with the user; a context modeling and memory module configured to record the user's historical interaction information, intention context and behavior preference, and maintain the user state; a graph interface module configured to connect the database, perform graph data retrieval, semantic query and learning path related information extraction; an inference and evaluation module configured to analyze, judge and path-optimize the current learning state based on the user model, database information and task target; an interaction management module for controlling the dialogue process, information confirmation, feedback triggering and prompt delivery between the main agent and the user; the user model is built by the main agent, and the building steps comprise: collecting behavior data of the user in the learning process, the behavior data including the user's learning duration, answer accuracy, reaction time, error type and operation path; based on the behavior data, analyzing the user's mastery level and cognitive preference at different knowledge points; combining the user's historical learning records and current learning tasks, building a multi-dimensional vectorized representation representing the user's knowledge state, behavior characteristics and learning tendency; storing the multi-dimensional vectorized representation as a user model for initializing the user twin agent and planning the personalized learning path; the main agent comprises determining the number of user twin agents to be generated based on multiple consideration conditions the steps in this process include: S100: analyzing the behavior characteristics in the user model to identify the user's volatility in various learning behaviors and learning results; S200: calculating an uncertainty score based on the volatility of performance, representing the consistency of the user's learning behavior; S300: determining the number of user twin agents to be generated based on the uncertainty score; Through the above consideration conditions, the generated user twin agents cover different behavior patterns that the user may have; the main agent is further configured to formulate multiple candidate learning paths based on the user model and the database, comprising the following steps: ​ Identify the node corresponding to the current knowledge state of the user and the node corresponding to the target knowledge point in the database; Based on the prerequisite relationship, semantic correlation and teaching order between each knowledge point in the database, multiple feasible paths from the current node to the target node are constructed; The main agent includes multiple user twin agents of the same user, which simulate learning on multiple candidate learning paths to predict the mastery efficiency, forgetting rate and learning cost of the user under different paths, and finally filter the optimal learning scheme for the real user; The main agent includes multiple user twin agents of the same user, which simulate learning on multiple candidate learning paths to predict the mastery efficiency, forgetting rate and learning cost of the user under different paths, and finally filter the optimal learning scheme for the real user, including the following steps: Step one, set the number of user twin agents Nt; The system introduces a user uncertainty score U∈[0,1] to evaluate the degree of behavior fluctuation; According to the score, the main agent dynamically determines the number of derived user twin agents, and the calculation method is: ； In the above formula, Nt base is the reference number of user twin agents; γ is an uncertainty amplification coefficient, and after the product of γ and U is rounded up, the final Nt value is calculated; Step two, set the number of candidate learning paths Np; ； B max the total computing power consumption of the system in completing the process of one learning path selection; Step three, performance analysis of the simulation results of all learning paths; The main agent drives all user twin agents to simulate each candidate learning path completely, and records the key indicators in the path execution process; These key indicators include: M ij , where Mij denotes the final knowledge mastery index of the jth user twin agent under the ith learning path. F ij , refers to the knowledge forgetting index of the jth user twin agent simulated under the ith learning path; T ij , which indicates the time taken by the jth user twin agent to complete the entire learning path under the ith learning path. The above three key indicators are dimensionless values after normalization processing of the original numerical values; Further, calculate the average performance score Score for each learning path: ； In the above formula, η and ε are respectively a forgetting penalty coefficient and a time penalty coefficient, to balance the weights of different evaluation dimensions, and the specific values of the two can be specifically set by a person skilled in the art; the performance score Score of the i-th learning path is calculated as i the variance Var i of the performance score Score of the i-th learning path. The master agent can make a comprehensive selection according to the performance score Score of all candidate learning paths i and the variance Var of its score i ​ 2. The educational system of claim 1, wherein, The database based on the graph structure includes one or more knowledge graphs; The knowledge graph is used to represent the structured knowledge relationship of learning content, and the composition of the knowledge graph includes: Multiple nodes, each node is used to represent an entity, and the entity includes knowledge points, learning resources, learning goals, user states or skill modules, etc. Multiple edges, each edge is used to represent the semantic association relationship between two nodes, and the association relationship includes one of prerequisite relationship, dependency relationship, similarity relationship, feedback relationship or teaching order relationship; Each node and edge is attached with attribute information, and the attribute information includes label, context metadata, importance weight or vector embedding, which is used to support semantic retrieval and path planning.

Images