A smart learning method and system based on a large language model

Abstract

The present invention discloses a smart learning method and system based on a large language model, which relates to the field of generative artificial intelligence and smart education technology. The method includes: guiding users to conduct in-depth interactions through a human-intelligence interaction module to generate a question-and-answer collection containing the wisdom of multiple parties; the teacher user purifies the question-and-answer collection through a question-and-answer archiving unit to obtain high-quality data. The system uses these complete interaction processes to automatically extract knowledge from the conversation to dynamically update a multi-level knowledge base composed of a personalized knowledge graph and a global knowledge graph; and simultaneously uses the archived high-quality data to fine-tune the model of an AI agent based on a cognitive framework and a reasoning framework, forming a closed loop of dual evolution of the knowledge base and model parameters. The present invention can significantly improve the professional capabilities of AI in vertical fields and effectively cultivate students' high-level learning abilities.

Description

Technical Field

[0001] The present invention relates to the field of generative artificial intelligence and smart education technology, and in particular to a smart learning method and system based on a large language model. Background Art

[0002] With the rapid development of generative AI technologies, represented by large language models, their application in education has become increasingly widespread, giving rise to a variety of intelligent teaching systems and intelligent tutoring tools. These existing technologies have significantly improved the efficiency of knowledge acquisition and can provide students with instant question-and-answer services.

[0003] However, existing intelligent teaching systems still face several deep-seated pain points in their application. First, the AI ​​models in existing systems are typically static, with their knowledge bases and capabilities fixed after deployment. They are unable to learn from continuous user interaction and achieve self-optimization and evolution, resulting in difficulty in continuously improving the professionalism and accuracy of their answers. Second, when dealing with complex problems that require deep understanding and logical connections, existing systems often rely on shallow information retrieval based on keyword matching. They lack effective multi-step reasoning and global analysis capabilities, making it difficult to guide students into in-depth exploration and prone to inaccurate or logically incoherent "model illusions." Finally, most systems focus on one-way "instilling" standard answers into students, lacking effective mechanisms to systematically cultivate students' critical thinking, question-posing skills, and proactive inquiry-based "learning" abilities, failing to fully realize the potential of human-intelligence collaboration.

[0004] Therefore, how to design a new intelligent teaching paradigm that can enable AI models to evolve on their own, have deep reasoning capabilities, and effectively cultivate students' high-order thinking abilities is a technical problem that needs to be urgently solved in this field. Summary of the Invention

[0005] In order to overcome the above problems existing in the prior art, the present invention proposes an intelligent learning method and system based on a large language model.

[0006] The technical solution adopted by the present invention to solve the technical problem is: a smart learning method based on a large language model, comprising the following steps:

[0007] Step 1: The AI ​​agent constructs a multi-level knowledge graph based on the course teaching materials in advance;

[0008] Step 2: The student user asks the AI ​​agent an initial question about a specific knowledge point in the course. The AI ​​agent generates an initial answer based on the multi-level knowledge graph obtained in step 1.

[0009] Step 3: The student user evaluates the initial answer obtained in step 2 and submits an improved answer, forming an initial conversation record with context;

[0010] In step 4, based on the initial conversation record in step 3, the student user enters the iterative question-and-answer phase. Through guided conversation, the AI ​​agent tracks and evaluates the quality of the student user's questions and generates a competency score for them.

[0011] Step 5: Guide other users to conduct secondary evaluation and content improvement on the initial conversation record obtained in step 3 and the iterative learning results obtained in step 4, forming a question and answer collection that contains the wisdom of multiple users;

[0012] Step 6: Automatically extract structured knowledge from the conversation records, iterative learning results, and question-answer sets obtained in steps 3, 4, and 5, and dynamically update the multi-level knowledge graph described in step 1 based on the attributes of the structured knowledge;

[0013] Step 7: The teacher user screens, refines, and archives the question-answer set obtained in step 5 to obtain high-quality question-answer data pairs that can be used for model training.

[0014] Step 8: Train the AI ​​agent based on the high-quality question-answer data pairs obtained in step 7 to obtain an evolved AI agent.

[0015] In step 9, the evolved AI agent obtained in step 8 and the updated multi-level knowledge graph in step 6 are applied to respond to new questions raised by subsequent student users.

[0016] In the aforementioned intelligent learning method based on a large language model, the multi-level knowledge graph in step 1 includes summary nodes, which are composed of a global knowledge graph and at least one personalized knowledge graph specific to a student user;

[0017] The step 1 specifically includes:

[0018] Step 1.1: Use natural language processing technology to extract entities, relationships between entities, and attribute information of entities from the course teaching materials to build a global knowledge graph;

[0019] Step 1.2: Initialize an empty personalized knowledge graph for each student user in the system;

[0020] Step 1.3: Identify closely related entity groups in the global knowledge graph and use the large language model to summarize the text content contained in the entity groups, pre-generating a summary node for each entity group;

[0021] Step 1.4: Add the summary node generated in step 1.3 and its connection relationship with each entity in the entity group to the global knowledge graph to form an enhanced global knowledge graph.

[0022] In the above-mentioned intelligent question-and-answer method based on a large language model, in step 2, the student user conducts preliminary question-and-answering through the Q&A module. The AI ​​agent adopts a cognitive framework based on graphical thinking, calls a graph-enhanced retrieval-enhanced generation reasoning framework for multi-hop question-and-answering, and generates an initial answer based on a multi-level knowledge graph.

[0023] The above-mentioned intelligent learning method based on a large language model, the cognitive framework of graphical thinking works specifically as follows:

[0024] Step a: decompose the initial problem into one or more initial thinking nodes to form the starting point of the reasoning graph;

[0025] Step b, expanding the thinking nodes in the reasoning graph to generate one or more subsequent thinking nodes with dependency relationships;

[0026] Step c: Evaluate and aggregate different reasoning paths in the reasoning graph, and prune invalid paths according to preset rules;

[0027] In step d, iteratively execute steps b and c until the optimal reasoning path is obtained and integrated to form the final initial answer.

[0028] The above-mentioned intelligent learning method based on a large language model, the working mode of the reasoning framework is specifically as follows:

[0029] Step A: When receiving the student user's initial question, a two-way information retrieval is performed, which simultaneously searches the global knowledge graph to obtain general knowledge and searches the student user's personalized knowledge graph to obtain personal related information;

[0030] Step B: Perform multi-hop reasoning on the information obtained from the dual-path retrieval in Step A on the multi-level knowledge graph. The multi-hop reasoning includes dynamic question decomposition and multi-hop neighborhood expansion to obtain node information that is deeply related to the initial question.

[0031] In step C, all node information obtained in step B is integrated and reordered to generate multiple candidate answers. The candidate answers are cross-validated using a self-consistency verification mechanism to screen and integrate the answers with the highest confidence.

[0032] In the aforementioned intelligent learning method based on a large language model, step 4 specifically includes:

[0033] In step 4.1, the AI ​​agent conducts guided conversations with student users through heuristic questions, identifying and guiding the conversation type of the student users, so that they can gradually evolve from instructional or sharing to inquiry, questioning, and debate.

[0034] In step 4.2, during the guidance process, the AI ​​agent tracks and analyzes the changes in the depth and breadth of the questions asked by the student users in real time to evaluate the quality of their questions;

[0035] In step 4.3, the AI ​​agent generates an ability score for the student user based on the changes in the quality of the questions asked.

[0036] In the aforementioned intelligent learning method based on a large language model, step 6 is specifically as follows:

[0037] Step 6.1: After the conversation records, iterative learning results, and question-answer sets are generated, the system automatically calls the information extraction model to extract the newly generated structured knowledge including entities, relationships, and user preferences;

[0038] Step 6.2: Determine the attributes of the structured knowledge extracted in step 6.1. If it is personal information of a specific user, add it to the personalized knowledge graph of the user; if it is objective facts in a general field, add it to the global knowledge graph after verification.

[0039] A large language model-based intelligent learning system, using the aforementioned large language model-based intelligent learning method, includes an AI agent, a human-intelligence interaction module, and a teaching management module. The core of the AI ​​agent is a graph-based cognitive framework, which is responsible for invoking a graph-enhanced, retrieval-enhanced generative reasoning framework for multi-hop question answering to perform deep answer generation, and is responsible for the dynamic update of the knowledge graph and the self-evolution of the learning model.

[0040] The human intelligence interaction module is used to perform preliminary learning, iterative learning and collaborative improvement functions;

[0041] The teaching management module includes a question and answer archiving unit, which filters and archives the question and answer data generated during the human-intelligence interaction process.

[0042] In the aforementioned intelligent learning system based on a large language model, the human-intelligence interaction module includes a question-and-answer unit, a question-and-answer unit, and a collaborative improvement unit. Student users conduct preliminary learning through the question-and-answer unit, evaluate the initial answers generated by the AI ​​agent in the question-and-answer unit, and submit improved answers. Student users complete iterative learning through the question-and-answer unit.

[0043] In the question-and-answer unit, a student user scores an improved answer submitted by another student user and makes further suggestions for improvement;

[0044] In the collaborative improvement unit, all answer versions around the same initial question are displayed to multiple users, and multiple users are allowed to like any answer version or submit a new answer.

[0045] The beneficial effects of the present invention are: (1) by introducing the GoT cognitive framework, the AI ​​agent is given more flexible and powerful global planning and complex reasoning capabilities, enabling it to better organize its thinking process to solve open-ended, multi-step problems.

[0046] (2) Through the GRAG-MH reasoning framework and the pKG / gKG dual-layer knowledge graph structure, the accuracy and logical depth of AI responses in vertical fields are significantly enhanced, the risk of "model hallucination" is reduced, and highly personalized responses can be provided.

[0047] (3) By introducing an advanced interactive process that includes preliminary questioning and iterative questioning, as well as a "double self-evolution" closed loop of knowledge base and model parameters, it not only effectively cultivates students' critical thinking and high-level questioning ability, but also achieves the continuous iteration and enhancement of AI teaching capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 Schematic diagram of the overall process of the method of the present invention;

[0049] Figure 2 It is a structural diagram of the functional modules of the system of the present invention;

[0050] Figure 3 A schematic diagram of the workflow of the cognitive framework and its calling reasoning framework in the present invention;

[0051] Figure 4 Schematic diagram of the structure of the multi-level knowledge graph in the present invention. DETAILED DESCRIPTION

[0052] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0053] The core idea of ​​the intelligent learning method based on a large language model proposed in this invention is to build a closed-loop ecosystem with dual evolution of "knowledge base" and "model parameters". Figure 1 As shown, this method runs through the entire process of knowledge pre-construction, human-intelligence interaction, dynamic knowledge growth, data purification, model evolution, and intelligent feedback. The methodology of this invention clearly divides the user's learning process into two progressive stages: initial learning and iterative learning. It aims to achieve the continuous improvement of AI capabilities and the simultaneous cultivation of students' higher-order thinking through systematic process design.

[0054] A smart learning method based on a large language model specifically includes the following steps:

[0055] In step 1, the AI ​​agent constructs a multi-level knowledge graph based on the course teaching materials, which includes summary nodes and consists of a global knowledge graph (gKG) and at least one personalized knowledge graph (pKG) specific to the student user. This step is the intelligent foundation of the entire method. Figure 4 As shown, the multi-level knowledge graph is a unique knowledge base structure, which includes two knowledge graphs with different granularity and attributes, one is the global knowledge graph (gKG), and the other is the personalized knowledge graph (pKG).

[0056] Step 1 specifically includes:

[0057] Step 1.1, extracting entities, relationships between entities, and attribute information of entities from the teaching materials of the course using natural language processing technology to construct the global knowledge graph;

[0058] Step 1.2: Initialize an empty personalized knowledge graph (pKG) for each student user in the system;

[0059] Step 1.3: Identify closely related entity groups in the global knowledge graph (gKG), summarize the textual content of the entity groups using a large language model, and pre-generate a summary node for each entity group;

[0060] Step 1.4, adding the summary node generated in step 1.3 and the connection relationship between it and each entity in the entity group to the global knowledge graph (gKG) to form an enhanced global knowledge graph.

[0061] In step 2, the student user conducts preliminary learning through the "You Ask, You Answer" unit, and asks the AI ​​agent initial questions around the specific knowledge points of the course. The AI ​​agent adopts a cognitive framework based on Graph of Thoughts (GoT), calls the Graph-enhanced Retrieval-Augmented Generation for Multi-Hop QA (GRAG-MH) reasoning framework, and generates an initial answer based on the multi-level knowledge graph obtained in step 1, such as Figure 3 This step is the initial human-intelligence interaction, which aims to help students understand and master basic knowledge points.

[0062] The GoT cognitive framework works by:

[0063] Step 2.1, decomposing the initial problem into one or more initial thinking nodes to form the starting point of the reasoning graph;

[0064] Step 2.2, expanding the thought nodes in the reasoning graph to generate one or more subsequent thought nodes with dependency relationships;

[0065] Step 2.3, evaluating and aggregating different reasoning paths in the reasoning graph, and pruning invalid paths according to preset rules;

[0066] Step 2.4, iteratively execute steps 2.2 and 2.3 until the optimal reasoning path is obtained and integrated to form the final initial answer.

[0067] The specific working method of the GRAG-MH reasoning framework includes:

[0068] Step 2.5: When receiving the initial question from a specific student user, the global knowledge graph (gKG) is searched in parallel to obtain general knowledge, and the personalized knowledge graph (pKG) of the specific student user is searched to obtain personal related information, thereby performing a dual-path information retrieval.

[0069] Step 2.6, performing multi-hop reasoning on the information obtained from the dual-path retrieval in step 2.5 on the multi-level knowledge graph, wherein the multi-hop reasoning includes dynamic problem decomposition and multi-hop neighborhood expansion to obtain node information that is deeply related to the initial problem;

[0070] In step 2.7, all node information obtained in step 2.6 is integrated and reordered to generate multiple candidate answers, and a self-consistency verification mechanism is used to cross-validate the candidate answers to screen and integrate the answer with the highest confidence.

[0071] In step 3, based on the initial answer from step 2, the student user evaluates it in the "You Ask, You Answer" section and submits an improved answer, forming an initial conversation record with context. This step is a key step in guiding students in critical thinking training and is the first manifestation of "human-intelligence collaboration."

[0072] In step 4, based on the initial conversation record from step 3, the student user enters an iterative question-and-answer session. Through guided conversation, the AI ​​agent tracks and evaluates the quality of the student user's questions and generates a competency score. This step is a core component of the present invention designed to cultivate students' higher-order thinking skills. Its goal is not to provide answers, but to train students' thinking processes.

[0073] In a preferred embodiment of the present invention, step 4 specifically includes:

[0074] In step 4.1, the AI ​​agent does not directly provide a final answer. Instead, it conducts a guided conversation with the student user by asking heuristic questions or invoking external tools for visualization assistance. During this guidance process, a core task of the AI ​​agent is to identify and guide the student user's conversation type, gradually evolving from more basic instructional or sharing-based conversations to more advanced inquiry-based, questioning-based, and debate-based conversations, thereby systematically training the depth and breadth of their thinking.

[0075] In step 4.2, during the guidance process, the AI ​​agent tracks the quality of the questions asked by the student user in real time by analyzing the changes in the questions asked. For example, it analyzes whether the questions become more specific from broad, or whether they change from descriptive questions to analytical or creative questions.

[0076] In step 4.3, the AI ​​agent generates an ability score for the student user based on the change trajectory of the question quality tracked in step 4.2. The score can be used as a quantitative indicator of the growth of the student's higher-order thinking ability.

[0077] In step 5, based on the initial conversation records from step 3 and the iterative learning results from step 4, other users are guided to conduct secondary evaluation and content improvement on the Q&A content through the "He Asks, You Answer" and "Collaborative Improvement" modules, thus forming a Q&A collection that incorporates the wisdom of multiple users. This step aims to further refine and improve knowledge from multiple perspectives by leveraging collective wisdom.

[0078] Step 5 specifically includes:

[0079] Step 5.1: In the "He Asks, You Answer" module, a student user scores the improved answer submitted by another student user in step 3 and provides further improvement suggestions;

[0080] In step 5.2, in the "Collaborative Improvement" module, all versions of answers to the same initial question are displayed to multiple users, and multiple users are allowed to like any answer version or submit a new answer to pool collective wisdom.

[0081] Step 6: Automatically extract newly generated structured knowledge such as entities, relationships, or user preferences from the conversation records, iterative learning results, and question-and-answer sets obtained in steps 3, 4, and 5.

[0082] In step 7, based on the attributes of the structured knowledge extracted in step 6, the multi-level knowledge graph described in step 1 is dynamically updated. If the information is personal information of a specific user, it is added to the user's personalized knowledge graph (pKG). If the information is objective facts in a general field, it is added to the global knowledge graph (gKG) after verification. Steps 6 and 7 together constitute the dynamic knowledge base growth mechanism of this invention.

[0083] In step 8, the teacher user uses the "Q&A Archiving" module to filter, purify, and archive the Q&A collection generated in step 5 to obtain high-quality Q&A data pairs that can be used for model training. Teacher users filter the Q&A collection based on user ratings or likes, and can edit and optimize the filtered Q&A data pairs before archiving. This step represents the expert-controlled phase in the method, ensuring a high-quality data closed loop.

[0084] In step 9, based on the high-quality question-answer data obtained in step 8, the model of the AI ​​agent is trained using model fine-tuning techniques to obtain an evolved AI agent. This step constitutes the evolutionary path of the model parameters in the "double self-evolution closed loop" of the present invention.

[0085] In step 10, the evolved AI agent from step 9 and the updated multi-level knowledge graph from step 7 are applied to respond to new questions raised by subsequent student users, thus forming a closed loop of dual evolution of the knowledge base and model parameters. This step is the key to completing the entire self-evolution closed loop and realizing its value.

[0086] like Figure 2 As shown, to implement the above method, the present invention provides an intelligent learning system based on a large language model, which includes at least one AI agent, a human-intelligence interaction module and a teaching management module.

[0087] The AI ​​agent is the intelligent core of this system, and it undertakes the main intelligent processing tasks in this invention.

[0088] The core of the AI ​​agent is the graphical thinking cognitive framework, which is responsible for calling the graph-enhanced retrieval-enhanced generative reasoning framework for multi-hop question answering to perform deep answer generation, and is responsible for the dynamic update of the knowledge graph and the self-evolution of the question-learning model.

[0089] The human-intelligence interaction module is the user-facing interface of this system, used to perform preliminary learning, iterative learning, and collaborative improvement functions.

[0090] The human-intelligence interaction module includes a question-and-answer unit, a question-and-answer unit, and a collaborative improvement unit. Student users conduct preliminary learning through the question-and-answer unit, evaluate the initial answers generated by the AI ​​agent in the question-and-answer unit, and submit improved answers. Student users complete iterative learning through the question-and-answer unit.

[0091] In the question-and-answer unit, a student user scores an improved answer submitted by another student user and makes further suggestions for improvement.

[0092] In the collaborative improvement unit, all answer versions around the same initial question are displayed to multiple users, and multiple users are allowed to like any answer version or submit a new answer.

[0093] The teaching management module includes a question and answer archiving unit, which is mainly aimed at teacher users and provides functions such as question and answer archiving. It is used to screen and purify the massive data generated by the human-intelligence interaction module, which is a key link in ensuring the quality of data used for model evolution.

[0094] When the system is running, part of the conversation records and data generated by the human-intelligence interaction module is used to dynamically update the knowledge base of the AI ​​agent module, and the other part is used to train the model parameters of the AI ​​agent module after being purified by the teaching management module. The evolved AI agent module then provides users with better services through the human-intelligence interaction module, thus forming a dual self-evolution closed loop at the system level.

[0095] Example 1

[0096] This embodiment will take the course theme of "Artificial Intelligence Ethics" as an example to illustrate the application process of the method of the present invention in detail.

[0097] First, execute step 1 to build a knowledge graph.

[0098] Execute steps 2 and 3 (initial questioning). Student A asks, "In an emergency, should a self-driving car prioritize protecting passengers or pedestrians?" The AI ​​agent gives an initial answer, and Student A evaluates and submits an improved answer, forming an initial conversation record.

[0099] In step 4 (iterative questioning), the AI ​​agent doesn't directly answer, but instead asks, "This is a classic ethical dilemma. What perspectives do you think can be used to analyze this issue?" (guided dialogue, exploratory). After Student A answers, the AI ​​agent follows up with, "Excellent! Do you think the principle of 'collectivism' applies here? Why?" (questioning). Through multiple rounds of dialogue, the quality of Student A's questions is tracked and ultimately a competency score is generated.

[0100] Execute step 5. The results of the preliminary and iterative learning will enter the "He Asks, You Answer" and "Collaborative Improvement" modules for secondary optimization by other users.

[0101] Then execute steps 6-10 to complete the closed loop of knowledge extraction, knowledge base update, teacher archiving, model fine-tuning and intelligent feedback.

[0102] Example 2

[0103] This embodiment takes the knowledge point of "TCP three-way handshake" in the "Computer Network" course as an example to illustrate the application of the method of the present invention under the guidance of a teacher.

[0104] Teachers noticed that students generally had difficulty understanding the specific functions and state changes of each handshake in the "TCP three-way handshake." To optimize teaching, teachers decided to use the system of the present invention to create high-quality teaching resources and enhance the tutoring capabilities of AI agents.

[0105] First, the teacher retrieved previous student interactions on the TCP three-way handshake from the Q&A archive. He found that while there was extensive discussion, there was a lack of high-quality answers that could visualize the abstract concept.

[0106] The teacher then used the human-intelligence interaction module of the present invention, playing the role of a student, to ask a question (step 2): "Please use a vivid metaphor to explain the TCP three-way handshake process." The AI ​​agent initially responded with an analogy based on a phone call (step 3). The teacher, however, felt the analogy was inaccurate and submitted a revised answer, revising it to a more appropriate metaphor: "Confirming a delivery address."

[0107] Afterwards, the teacher will directly archive the high-quality question-answer data pair he or she created through the "Question and Answer Archiving" function (execute step 8).

[0108] Next, the teacher uses this and other selected high-quality data to fine-tune the AI ​​agent model (proceed to step 9).

[0109] Finally, in subsequent teaching, when other students asked questions about the "three-way handshake" again, the evolved AI agent was able to proactively give this vivid and accurate "confirming the delivery address" metaphor, significantly improving the teaching effect.

[0110] The above embodiments are merely exemplary embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art may make various modifications or equivalent substitutions to the present invention within the spirit and scope of protection of the present invention, and such modifications or equivalent substitutions shall also be deemed to fall within the scope of protection of the present invention.

Claims

1. A smart learning method based on a large language model, characterized by: The steps include: Step 1: The AI ​​agent constructs a multi-level knowledge graph based on the course teaching materials in advance; Step 2: The student user asks the AI ​​agent an initial question about a specific knowledge point in the course. The AI ​​agent generates an initial answer based on the multi-level knowledge graph obtained in step 1. Step 3: The student user evaluates the initial answer obtained in step 2 and submits an improved answer, forming an initial conversation record with context; In step 4, based on the initial conversation record in step 3, the student user enters the iterative question-and-answer phase. Through guided conversation, the AI ​​agent tracks and evaluates the quality of the student user's questions and generates a competency score for them. Step 5: Guide other users to conduct secondary evaluation and content improvement on the initial conversation record obtained in step 3 and the iterative learning results obtained in step 4, forming a question and answer collection that contains the wisdom of multiple users; Step 6: Automatically extract structured knowledge from the conversation records, iterative learning results, and question-answer sets obtained in steps 3, 4, and 5, and dynamically update the multi-level knowledge graph described in step 1 based on the attributes of the structured knowledge; Step 7: The teacher user screens, refines, and archives the question-answer set obtained in step 5 to obtain high-quality question-answer data pairs that can be used for model training. Step 8: Train the AI ​​agent based on the high-quality question-answer data pairs obtained in step 7 to obtain an evolved AI agent. Step 9: Apply the evolved AI agent obtained in step 8 and the updated multi-level knowledge graph in step 6 to respond to new questions raised by subsequent student users; The multi-level knowledge graph in step 1 includes summary nodes, and is composed of a global knowledge graph and at least one personalized knowledge graph specific to a student user; The step 1 specifically includes: Step 1.1: Use natural language processing technology to extract entities, relationships between entities, and attribute information of entities from the course teaching materials to build a global knowledge graph; Step 1.2: Initialize an empty personalized knowledge graph for each student user in the system; Step 1.3: Identify closely related entity groups in the global knowledge graph and use the large language model to summarize the text content contained in the entity groups, pre-generating a summary node for each entity group; Step 1.4: Add the summary node generated in step 1.3 and its connection relationship with each entity in the entity group to the global knowledge graph to form an enhanced global knowledge graph.

2. The intelligent learning method based on a large language model according to claim 1, characterized in that: In the step 2, the student user conducts preliminary learning through the Q&A module. The AI ​​agent adopts a cognitive framework based on graphical thinking, calls a graph-enhanced retrieval-enhanced generation reasoning framework for multi-hop question and answer, and generates an initial answer based on a multi-level knowledge graph.

3. The intelligent learning method based on a large language model according to claim 2, characterized in that: The cognitive framework of graphical thinking works specifically as follows: Step a: decompose the initial problem into one or more initial thinking nodes to form the starting point of the reasoning graph; Step b, expanding the thinking nodes in the reasoning graph to generate one or more subsequent thinking nodes with dependency relationships; Step c: Evaluate and aggregate different reasoning paths in the reasoning graph, and prune invalid paths according to preset rules; In step d, iteratively execute steps b and c until the optimal reasoning path is obtained and integrated to form the final initial answer.

4. The intelligent learning method based on a large language model according to claim 2, characterized in that: The reasoning framework works as follows: Step A: When receiving the student user's initial question, a two-way information retrieval is performed, which simultaneously searches the global knowledge graph to obtain general knowledge and searches the student user's personalized knowledge graph to obtain personal related information; Step B: Perform multi-hop reasoning on the information obtained from the dual-path retrieval in Step A on the multi-level knowledge graph. The multi-hop reasoning includes dynamic question decomposition and multi-hop neighborhood expansion to obtain node information that is deeply related to the initial question. In step C, all node information obtained in step B is integrated and reordered to generate multiple candidate answers. The candidate answers are cross-validated using a self-consistency verification mechanism to screen and integrate the answers with the highest confidence.

5. The intelligent learning method based on a large language model according to claim 1, characterized in that: The step 4 specifically includes: In step 4.1, the AI ​​agent conducts guided conversations with student users through heuristic questions, identifying and guiding the conversation type of the student users, so that they can gradually evolve from instructional or sharing to inquiry, questioning, and debate. In step 4.2, during the guidance process, the AI ​​agent tracks and analyzes the changes in the depth and breadth of the questions asked by the student users in real time to evaluate the quality of their questions; In step 4.3, the AI ​​agent generates an ability score for the student user based on the changes in the quality of the questions asked.

6. The intelligent learning method based on a large language model according to claim 1, characterized in that: The step 6 is specifically as follows: Step 6.1: After the conversation records, iterative learning results, and question-answer sets are generated, the system automatically calls the information extraction model to extract the newly generated structured knowledge including entities, relationships, and user preferences; Step 6.2: Determine the attributes of the structured knowledge extracted in step 6.

1. If it is personal information of a specific user, add it to the personalized knowledge graph of the user; if it is objective facts in a general field, add it to the global knowledge graph after verification.

Images