Collaborative personalized learning system and method based on large model

Abstract

The invention discloses a collaborative personalized learning system and method based on a large model, and relates to the technical field of collaborative learning, and the system collects and fuses multi-modal data, generates a student state vector, plans a personalized learning path based on the state and a knowledge graph, and generates multi-modal explanation content according with the student style. The method comprises the following steps: extracting a student reasoning path, identifying thinking deviation, analyzing task performance, generating a path and content adjustment suggestion, identifying a motivation state, triggering an intervention strategy, receiving a teacher strategy, and realizing path and style co-construction. The multi-modal content of the matched style is generated, adaptive intervention is achieved in combination with inference analysis and emotion adjustment, and the learning efficiency and the teaching response intelligent level are improved.

Description

Technical Field

[0001] The present invention relates to the field of collaborative learning technology, and in particular to a collaborative personalized learning system and method based on a large model. Background Art

[0002] With breakthroughs in natural language understanding, multimodal perception, and task generation achieved by large-scale AI models, educational technology is moving from "content push" to a deeper intelligence phase characterized by "cognitive modeling" and "strategic intervention." However, traditional systems based on single recommendation engines or static teaching strategies struggle to adapt in real time to dynamic changes in students' cognitive levels, emotional states, learning styles, and social interactions, and lack the ability to understand, reason about, and guide the learning process itself. This situation urgently requires a system architecture with stronger task awareness, autonomous planning, and collaborative feedback capabilities. Meanwhile, while existing large-scale models possess strong generalization capabilities, they still face challenges in real-world educational scenarios, such as weak structural drive, slow real-time feedback, and shallow human-computer collaboration. Therefore, combining the language understanding advantages of large-scale models with multi-agent systems with task autonomy to build an understandable, interventionable, and evolvable intelligent teaching system has become a key issue in the next phase of "AI + Education."

[0003] At present, the Chinese patent application number CN202210642928.4 discloses an intelligent online teaching system, including a login subsystem, an intelligent lesson preparation subsystem, a micro-class teaching subsystem, a mobile learning subsystem, a question bank subsystem, a data analysis subsystem, an online test subsystem, a database subsystem, a human-computer interaction subsystem and a communication subsystem; in this invention, parents, teachers and students can all communicate through the communication subsystem; the intelligent lesson preparation subsystem can facilitate teachers to prepare lessons online; the shared database module can realize communication and learning between teachers; online teaching is realized by setting up teaching modules; the teaching module logged in by the teacher account synchronously imports files to the teaching module corresponding to the student account for unified classroom management; the mobile learning subsystem facilitates students' self-learning management; the combination of the online test subsystem and the data analysis subsystem can not only facilitate online testing and online marking, but also enable teachers, students and parents to better understand students' learning situation.

[0004] The above technologies are difficult to solve the limitations in cognitive perception, path planning, content generation, thinking modeling, evaluation feedback, emotional regulation and teacher co-construction, and it is difficult to achieve a dynamic, adaptive and explainable personalized intelligent learning closed loop for individual differences. Summary of the Invention

[0005] The technical problem solved by the present invention is that the existing technology is difficult to solve the limitations in cognitive perception, path planning, content generation, thinking modeling, evaluation feedback, emotional regulation and teacher co-construction, and it is difficult to achieve a dynamic, adaptive and explainable personalized intelligent learning closed loop for individual differences.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] A collaborative personalized learning system based on a large model, including a student cognitive state perception module, a personalized path planning module, a teaching content generation module, a reasoning chain construction module, a cognitive assessment feedback module, an emotion regulation intervention module, and a teacher collaborative intervention module:

[0008] The student cognitive state perception module is used to collect and fuse multimodal data to generate a student state vector;

[0009] The personalized path planning module is used to plan a personalized learning path based on the state and knowledge graph;

[0010] The teaching content generation module is used to generate multimodal teaching content that suits the students' style;

[0011] The reasoning chain building module is used to extract students' reasoning paths and identify thinking deviations;

[0012] The cognitive assessment feedback module is used to analyze task performance and generate path and content adjustment suggestions;

[0013] The emotion regulation intervention module is used to identify motivational states and trigger intervention strategies;

[0014] The teacher collaborative intervention module is used to receive teacher strategies and realize the co-construction of paths and styles.

[0015] Preferably, the student cognitive state perception module includes a multimodal data acquisition unit, a feature fusion unit, an emotional state estimation unit and a state vector generation unit;

[0016] The multimodal data acquisition unit is used to collect students' visual channel data, behavioral channel data, voice channel data and system-side data. The visual channel data includes camera image stream, eye movement trajectory and facial micro-expressions; the behavioral channel data includes click logs, answering time and path jump sequence; the voice channel data includes speaking speed, pauses and intonation changes; the system-side data includes completed tasks, error statistics and prerequisite node coverage;

[0017] The feature fusion unit is used to perform multi-channel embedding processing on the visual channel data, the behavioral channel data, the voice channel data and the system side data to generate a modality representation vector, wherein the modality representation vector includes:

[0018]

[0019] K t =f GNN ;

[0020] in, is the visual image vector, f ViT is the visual channel data corresponding to the visual image vector, is the behavior sequence vector, f AudioBERT is the behavior channel data corresponding to the behavior sequence vector, K t is the speech stream vector, f GNN is the voice channel data corresponding to the voice stream vector, K t is the knowledge graph state vector, f GNN is the system-side data corresponding to the knowledge graph state vector;

[0021] The modal representation vectors are nested and fused through the multimodal alignment layer, and the cross-modal fusion module based on the attention mechanism is used to complete the weighted combination and output high-dimensional fusion features. The mathematical expression of the high-dimensional fusion features is:

[0022]

[0023] Among them, h t is the high-dimensional fusion feature, || is the vector splicing operation;

[0024] Based on the behavior channel data, the corresponding short-term behavior change trend is modeled through a sliding time window and weighted average method, and the output is the behavior trend feature;

[0025] The emotional state estimation unit is used to predict and obtain the emotional four-dimensional index based on the visual channel data and the voice channel data based on the convolutional neural network and the long short-term memory network. The emotional four-dimensional index includes: pleasure e val , awakening degree aro , sense of control dom and polarity e pol ,The four-dimensional index of emotion is normalized into the emotion vector;

[0026] The state vector generation unit is used to construct a standard student state vector based on the high-dimensional fusion feature, the knowledge graph state vector and the emotion vector. The mathematical expression of the standard student state vector is:

[0027] S t ={h t ,K t ,E t}∈R d ;

[0028] Among them, St is the standard student state vector, R d Behavioral trend characteristics;

[0029] Write the standard student state vector to the shared state pool.

[0030] Preferably, the personalized path planning module includes a perception agent construction unit, a learning path planning unit and a personalized content generation unit;

[0031] The perceptual agent building unit is used to establish an agent group, and the agent group includes:

[0032] The student state perception agent is used to collect and encode the student's knowledge mastery status, behavior patterns and emotional variables, and output a standard state vector S t , and push perception events when the state mutates;

[0033] The learning path planning agent is used to perform path search and dynamic planning on the directed acyclic graph task graph based on the student's current cognitive map and the prerequisite structure graph to generate a task sequence Q t And maintain the task execution context;

[0034] A teaching content generation agent, which is used to receive the task plan and style vector, generate multimodal teaching content based on the student profile through a hybrid expert model and a retrieval-enhanced generation model, and automatically match the visual, language, or interactive teaching style;

[0035] Reasoning guides the agent, which is used to guide the agent around the task q i Constructing students' individual reasoning paths T path , complete the process of thinking chain construction, hypothesis verification and error reflection;

[0036] A cognitive evaluation agent, which is used to construct a cognitive evaluation graph, explainability indicators, and error cause location based on task behavior records and reasoning path results, and output them to the path planning and teacher agents;

[0037] An emotion regulation agent, which is used to monitor multimodal data and behavioral trend characteristics in real time, detect trends from overload to cooling down to abandonment, and adjust task intensity or trigger incentive events;

[0038] Teacher collaborative agent, which is used to provide natural language interaction, policy parameter adjustment interface and convert it into policy offset vector Δπ teacher ;

[0039] The communication mechanism between the agent groups includes:

[0040] Shared state pool, which is used to store all agents' state variables in real time through the shared state pool storage platform. The shared state pool supports state caching, long-term trajectory maintenance and historical query;

[0041] Intent event bus, which is used to register the intent type of each agent with the bus. Events are transmitted in the form of asynchronous messages, supporting priority scheduling, conflict decoupling, and event tracing;

[0042] The learning path planning unit is used to construct a directed acyclic graph knowledge graph G based on the course structure and prerequisite dependencies. K =(V,E), where each node v i ∈V represents a fine-grained learning unit, and edge e ij ∈E represents the pre-dependence of knowledge, and the student’s current knowledge coverage weight function κ(v i ,S t ), the student's current knowledge coverage weight function represents the degree of mastery of the knowledge unit;

[0043] Modeling a task sequence decision process that constructs a state set based on the student's current learning trajectory and the local subgraph structure in the directed acyclic graph knowledge graph;

[0044] According to the knowledge mastery and precondition relationship in different states in the state set, the next learning task is selected from the optional task nodes as the action set;

[0045] The task transfer relationship is determined based on the changes in the knowledge graph structure and the student's cognitive state, and a task reward mechanism is constructed based on knowledge gain, learning interest, emotional stability, and the time cost required to complete the task;

[0046] A graph-structure-related reinforcement learning strategy optimization method is used to train the task sequence decision process and generate an integrated path.

[0047] The mathematical expression of the objective function of the optimization method of the task sequence decision process is:

[0048]

[0049] Among them, S is the current learning trajectory and the local subgraph structure in the directed acyclic graph knowledge graph, a t is the action set, is the reward discount factor, α, β, γ and δ are the strategy adjustment weights, ΔK t The improvement in knowledge mastery estimated by the knowledge tracking model, d task is the task difficulty, d student is the degree of matching of students’ current abilities, Msocial is the community impact factor, C load To estimate the task cognitive load by combining the four-dimensional indicators of emotion and behavioral trend characteristics;

[0050] If the cognitive assessment agent detects a failure in task understanding, serious skipping of questions, and a downward trend in mastery, it triggers a replanning mechanism, recalculates the local path map, inserts a backtracking task, or reduces the difficulty of the fine-tuning strategy.

[0051] Preferably, the personalized content generation unit is used to generate multimodal teaching content based on the integration path and in combination with the student's current state vector and cognitive preference vector. The multimodal teaching content includes text explanation, image assistance, dynamic animation and experimental interaction script, and is embedded with explanation logic structure, anchor point association and cognitive load adjustment mechanism.

[0052] The generation process of the multimodal teaching content is as follows:

[0053] A hybrid architecture is generated based on the hybrid expert model and the retrieval-enhanced generative model, wherein the hybrid architecture includes:

[0054] Visual experts for generating SVG animations, spatial compositions, geometric diagrams, and physical force diagrams;

[0055] Language experts, said language experts are used to generate structured text explanations, analogies, metaphors, and comparative arguments;

[0056] Operation experts, used for building interactive experiment scripts, code generation, and virtual evaluator configuration;

[0057] A language model collaboration core that provides unified semantic construction, context alignment, and cross-expert style scheduling capabilities.

[0058] After receiving the task node and the student's current state vector, the cognitive style vector is parsed. The cognitive style vector is a personality preference indicator extracted from the student's long-term behavioral data, and the hybrid architecture weight is assigned through the gating function.

[0059] Taking the task node as input, the system performs vector semantic recall from the knowledge retrieval database through a retrieval model based on a dual-tower structure to obtain the most relevant knowledge item set, which is then fed into the large model generator for imperative content generation and outputs the generated content.

[0060] Add semantic tags and trackable anchors to generated content.

[0061] Preferably, the reasoning chain building module includes a thinking path extraction unit, a path alignment analysis unit and a logic deviation identification unit;

[0062] The thinking path extraction unit is used to extract the task node qi As the starting point, the directed acyclic graph knowledge graph and the behavioral trend characteristics of the current student cognitive state are called to construct a task target path tree, which is the shortest effective explanation path of the task node on the knowledge graph;

[0063] During task execution, the reasoning-guided agent constructs the student's actual reasoning path in real time. The student's actual reasoning path extracts behavioral fragments, reasoning intermediaries, and explicit statements from the answering behavior sequence. The thinking chain generator maps the structured path diagram to extract explicit reasoning steps from the answering behavior and operation log.

[0064] The path alignment analysis unit is based on a structured attention graph network to model student behavior sequences, extract paths, and align sequential nodes;

[0065] Compare the students' actual reasoning path with the task target path tree to detect whether the following structural deviations exist:

[0066] Skipping, wherein a preset necessary node is skipped;

[0067] Wrong chain, where the wrong chain refers to invalid or incorrect reasoning in the student's actual reasoning path;

[0068] Weak links, where the weak links are nodes with insufficient attention;

[0069] The logical deviation identification unit is used to encode the structural deviation into a cognitive anomaly vector and transmit it to the cognitive assessment agent for evaluation and weighting. If the cognitive anomaly vector exceeds a preset tolerance threshold, the emotion regulation agent is triggered to intervene and the teacher collaborative agent prompts the student to insert a review task.

[0070] The student's actual reasoning path is rendered as a visual structure diagram, which includes steps, risk nodes and missed solution links, and the visual structure diagram is embedded in the student terminal interface and teacher dashboard.

[0071] Preferably, the cognitive assessment feedback module includes a state change analysis unit, an error attribution modeling unit and a feedback strategy generation unit;

[0072] The state change analysis unit is used to obtain the student's actual reasoning path, behavioral trend characteristics in the current student's cognitive state, historical task trajectory, actual behavior data and task target path tree. The actual behavior data includes answer options, time distribution, revocation frequency and interface focus. Based on deep knowledge tracking and graph structure reinforcement modeling, it constructs a mastery state vector and predicts the mastery evolution trajectory of knowledge points in the future path. The mastery evolution trajectory is used to update the directed acyclic graph knowledge graph;

[0073] The error attribution modeling unit is used to identify high-contribution nodes in the student's actual reasoning path based on a gradient-weighted class activation mapping variant, quantify the student's actual reasoning path based on a concept activation vector test, and output a quantitative result, wherein the quantitative result includes reasoning completeness, unit standardization, and graphic comprehension; calculate and analyze the marginal influence of each input feature in the current task output based on Shapley additive interpretation, output a knowledge radar chart, a concept heat map, and a strategy error analysis chart; input the strategy error analysis chart into the cognitive evaluation agent for error cause classification, output the error cause type, wherein the error cause type includes step skipping error, misconception mixing, and path circular thinking, and output an evaluation result;

[0074] The feedback strategy generation unit outputs path adjustment prompts and explanation reconstruction suggestions based on the evaluation results.

[0075] Preferably, the emotion regulation intervention module includes an exponential function construction unit and a micro-intervention unit:

[0076] The index function construction unit is used to construct a dynamic motivation index function based on rational emotive behavior therapy, real-time emotion recognition, cognitive fatigue modeling, behavioral stagnation detection and community incentive mechanism. The construction process of the dynamic motivation index function is as follows:

[0077] Physiological and behavioral features are extracted through a multimodal emotion perception network. The physiological and behavioral features include:

[0078] Facial expression parameters, including the degree of activation of specific facial muscle groups, brow tension, and eyelid movement;

[0079] Eye movement trajectory features, including gaze drift, number of fixations, and pupil dilation rate;

[0080] Operation behavior indicators, including task dwell time, continuous click frequency, and operation interruption times;

[0081] Audio emotion indicators, including abnormal speaking speed, falling intonation, and long pauses;

[0082] Background social signals, including current class learning popularity and social leaderboard performance;

[0083] The micro-intervention unit is used to construct a representation vector of the student's current emotional state and behavioral motivation state based on physiological and behavioral characteristics, input the vector into the motivation sensor to calculate the multi-factor motivation function, and obtain the multi-factor motivation value. If the multi-factor motivation value is greater than a preset high motivation threshold, an additional challenge task is added. If the multi-factor motivation value is less than a preset low motivation threshold, a micro-intervention strategy is triggered. If the multi-factor motivation value is between the low motivation threshold and the high motivation threshold, the task path is maintained unchanged and the enhanced monitoring state is entered;

[0084] The micro-intervention strategy includes content structure adjustment, external incentive intervention and emotional guidance feedback. The content structure adjustment is used to split the current task into small knowledge units and reduce cognitive density. The external incentive intervention is used to push short-term community incentive events. The emotional guidance feedback is used to call an affinity text generation model to generate text with an empathetic language style.

[0085] Preferably, the teacher collaborative intervention module includes a strategy injection unit, a strategy co-construction unit and an auxiliary reasoning unit;

[0086] The strategy injection unit is used to adjust the strategy through a natural language interface. The strategy adjustment includes modifying the task difficulty weight, rearranging the order of knowledge nodes, specifying the explanation style, or inserting teaching intervention points. The strategy adjustment is parsed into a strategy adjustment vector, and the original path generation strategy is updated in a targeted manner through a strategy co-evolution mechanism.

[0087] The strategy co-construction unit is used to input the visual structure diagram, knowledge radar diagram, concept heat map and strategy error analysis diagram into the teacher terminal through the three-dimensional dashboard. The teacher terminal performs the following operations: viewing the knowledge breakpoints and score loss frequency bands within the group, judging the overall learning sentiment trend of students, adjusting the search priority of the path generator for a certain knowledge area, configuring the reward threshold and community incentive frequency and content style switching logic, converting the operation path into the teacher operation trajectory encoding vector, and inputting it into the strategy distillation network for fusion with the preset reinforcement learning strategy;

[0088] The auxiliary reasoning unit is used to generate a candidate strategy set and send a teacher participation path voting signal when strategy oscillation occurs in the system path planning or the content module falls into style conflict.

[0089] Preferably, the data consistency transmission between modules is achieved through a shared state pool, and the task triggering and feedback response are decoupled with the help of an asynchronous intent message bus, wherein the standard student state vector is automatically updated and broadcast after each task is completed.

[0090] A collaborative personalized learning method based on a large model includes the following steps:

[0091] Step S1, collect and fuse multimodal data to generate student state vector;

[0092] Step S2: planning a personalized learning path based on the state and knowledge graph;

[0093] Step S3, generating multimodal explanation content that suits the students' style;

[0094] Step S4, extracting students’ reasoning paths and identifying thinking deviations;

[0095] Step S5, analyzing task performance and generating path and content adjustment suggestions;

[0096] Step S6, identifying motivational states and triggering intervention strategies;

[0097] Step S7, receiving teacher strategies and implementing path and style co-construction.

[0098] Beneficial effects of the present invention: Based on a large model and a multi-agent collaborative mechanism, the present invention accurately perceives student status, dynamically plans personalized learning paths, generates multimodal content with matching styles, and combines reasoning analysis and emotion regulation to achieve adaptive intervention, thereby improving learning efficiency and the level of intelligent teaching response. BRIEF DESCRIPTION OF THE DRAWINGS

[0099] Figure 1 A schematic diagram of the basic process of a collaborative personalized learning system based on a large model provided by one embodiment of the present invention;

[0100] Figure 2 A flowchart of the steps of a collaborative personalized learning method based on a large model is provided for one embodiment of the present invention. DETAILED DESCRIPTION

[0101] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, but not all of the embodiments.

[0102] Example 1, refer to Figure 1 , provides a collaborative personalized learning system based on a large model, including a student cognitive state perception module, a personalized path planning module, a teaching content generation module, a reasoning chain construction module, a cognitive evaluation feedback module, an emotion regulation intervention module, and a teacher collaborative intervention module:

[0103] The student cognitive state perception module is used to collect and fuse multimodal data to generate student state vectors.

[0104] The personalized path planning module is used to plan personalized learning paths based on state and knowledge graphs.

[0105] The teaching content generation module is used to generate multimodal teaching content that suits students' style.

[0106] The reasoning chain building module is used to extract students' reasoning paths and identify thinking deviations.

[0107] The cognitive assessment feedback module is used to analyze task performance and generate path and content adjustment suggestions.

[0108] The emotion regulation intervention module is used to identify motivational states and trigger intervention strategies.

[0109] The teacher collaborative intervention module is used to receive teacher strategies and achieve co-construction of paths and styles.

[0110] The student cognitive state perception module includes a multimodal data acquisition unit, a feature fusion unit, an emotional state estimation unit and a state vector generation unit.

[0111] The multimodal data acquisition unit is used to collect students' visual channel data, behavioral channel data, voice channel data and system-side data. The visual channel data includes camera image stream, eye movement trajectory and facial micro-expressions. The behavioral channel data includes click log, answering time and path jump sequence. The voice channel data includes speaking speed, pauses and intonation changes. The system-side data includes completed tasks, error statistics and prerequisite node coverage.

[0112] The multimodal data acquisition unit realizes the comprehensive collection of students' visual, voice, behavioral and system operation data during the learning process, ensuring the integrity and multi-dimensional coverage of state modeling.

[0113] The feature fusion unit is used to perform multi-channel embedding processing on the visual channel data, behavioral channel data, voice channel data and system side data to generate a modal representation vector. The modal representation vector includes:

[0114]

[0115] K t =f GNN ;

[0116] in, is the visual image vector, f ViT is the visual channel data corresponding to the visual image vector, is the behavior sequence vector, f AudioBERT is the behavior channel data corresponding to the behavior sequence vector, K t is the speech stream vector, f GNN is the voice channel data corresponding to the voice stream vector, K t is the knowledge graph state vector, f GNN It is the system-side data corresponding to the knowledge graph state vector.

[0117] The modal representation vectors are nested and fused through the multimodal alignment layer, and the cross-modal fusion module based on the attention mechanism is used to complete the weighted combination and output high-dimensional fusion features. The mathematical expression of the high-dimensional fusion features is:

[0118]

[0119] Among them, h tis the high-dimensional fusion feature, and || is the vector concatenation operation.

[0120] Based on the behavioral channel data, the corresponding short-term behavioral change trend is modeled through a sliding time window and weighted average method, and the output is the behavioral trend feature.

[0121] The feature fusion unit uniformly encodes and fuses data of different modalities through embedding and alignment mechanisms, extracts key high-dimensional features, and models the recent behavioral change trends of students.

[0122] The emotional state estimation unit is used to predict and obtain the emotional four-dimensional index based on the visual channel data and the voice channel data based on the convolutional neural network and the long short-term memory network. The emotional four-dimensional index includes: pleasure e val , awakening degree aro , sense of control dom and polarity e pol ,The four-dimensional sentiment index is normalized into a sentiment vector.

[0123] The emotional state estimation unit uses visual and speech features to accurately predict students' emotional states during the learning process, forming a standardized emotional vector for judging participation and emotional fluctuations.

[0124] The state vector generation unit is used to construct a standard student state vector based on high-dimensional fusion features, knowledge graph state vector and emotion vector. The mathematical expression of the standard student state vector is:

[0125] S t ={h t ,K t ,E t}∈R d ;

[0126] Among them, S t is the standard student state vector, R d Behavioral trend characteristics.

[0127] Write the standard student state vector to the shared state pool.

[0128] The state vector generation unit comprehensively integrates features, knowledge graph states and emotion vectors to construct a unified standard student state vector and writes it into a shared pool to achieve consistent data sharing between modules.

[0129] The student cognitive state perception module realizes comprehensive perception and integrated analysis of students' multimodal learning behaviors, dynamically generates standard state vectors that reflect their cognitive, emotional and behavioral trends, and provides an accurate basis for personalized learning path planning and content generation.

[0130] The personalized path planning module includes a perception agent construction unit, a learning path planning unit and a personalized content generation unit.

[0131] The perceptual agent building unit is used to build an agent group, which includes:

[0132] The student state perception agent is used to collect and encode students’ knowledge mastery status, behavior patterns and emotional variables, and output a standard state vector S t , and push perception events when the state changes.

[0133] Learning path planning agent, which is used to perform path search and dynamic planning on the directed acyclic graph task graph based on the student's current cognitive map and the prerequisite structure graph, and generate the task sequence Q t And maintain the task execution context.

[0134] The teaching content generation agent is used to receive task plans and style vectors, generate multimodal explanation content based on student portraits through a hybrid expert model and a retrieval-enhanced generation model, and automatically match visual, language or interactive explanation styles.

[0135] Reasoning guides the agent, which is used to guide the agent around the task q i Constructing students' individual reasoning paths T path , complete the process of thinking chain construction, hypothesis verification and error reflection.

[0136] The cognitive evaluation agent is used to construct cognitive evaluation graphs, explainability indicators, and error location based on task behavior records and reasoning path results, and output them to the path planning and teacher agents;

[0137] Emotion regulation agent, the emotion regulation agent is used to monitor multimodal data and behavioral trend characteristics in real time, detect trends from overload to cold withdrawal to abandonment, adjust task intensity or trigger incentive events.

[0138] Teacher collaborative agent, which is used to provide natural language interaction, policy parameter adjustment interface and convert it into policy offset vector Δπ teacher .

[0139] The communication mechanisms between groups of agents include:

[0140] The shared state pool is used to store state variables of all agents in real time through the shared state pool storage platform. The shared state pool supports state caching, long-term trajectory maintenance and historical query.

[0141] Intent event bus, the intent event bus is used to register the intent type of each agent to the bus. Events are transmitted in the form of asynchronous messages, supporting priority scheduling, conflict decoupling and event tracing.

[0142] The perception agent construction unit constructs a multi-role agent group, which divides the work of handling functions such as student status perception, path planning, content generation, reasoning guidance, emotion regulation and teacher collaboration, to achieve intelligent collaboration and dynamic scheduling in the task flow. The student status perception agent perceives and encodes the student's knowledge status, behavior patterns and emotional variables in real time, and pushes perception events when the status suddenly changes to guide path or content adjustment.

[0143] The learning path planning unit is used to construct a directed acyclic graph knowledge graph G based on the course structure and prerequisite dependencies. K =(V,E), where each node v i ∈V represents a fine-grained learning unit, and edge e ij ∈E represents the pre-dependence of knowledge, and the student’s current knowledge coverage weight function κ(v i ,S t ), the student’s current knowledge coverage weight function represents the degree of mastery of the knowledge unit.

[0144] Modeling the task sequence decision process, the task sequence decision process constructs a state set based on the student's current learning trajectory and the local subgraph structure in the directed acyclic graph knowledge graph.

[0145] According to the knowledge mastery and precondition relationship in different states in the state set, the next learning task is selected from the optional task node as the action set.

[0146] The task transfer relationship is determined jointly according to the changes in the knowledge graph structure and the students' cognitive state, and a task reward mechanism is constructed based on knowledge gain, learning interest, emotional stability and the time cost required to complete the task.

[0147] A graph-structure-related reinforcement learning strategy optimization method is used to train the task sequence decision process and generate an integrated path.

[0148] The mathematical expression of the objective function of the optimization method of the task sequence decision process is:

[0149]

[0150] Among them, S is the current learning trajectory and the local subgraph structure in the directed acyclic graph knowledge graph, a t is the action set, is the reward discount factor, α, β, γ and δ are the strategy adjustment weights, ΔK t The improvement in knowledge mastery estimated by the knowledge tracking model, d task is the task difficulty, d student is the degree of matching of students’ current abilities, M social is the community impact factor, C loadTo estimate the task cognitive load by combining the four-dimensional indicators of emotion and behavioral trend characteristics.

[0151] If the cognitive assessment agent detects a failure in task understanding, serious skipping of questions, and a downward trend in mastery, it triggers a replanning mechanism, recalculates the local path map, inserts a backtracking task, or reduces the difficulty of the fine-tuning strategy.

[0152] The learning path planning unit constructs a knowledge graph based on a directed acyclic graph, integrates students' knowledge coverage weights, models task sequence decisions through reinforcement learning, and outputs an optimized path.

[0153] The personalized content generation unit is used to generate multimodal teaching content based on the integration path and combined with the student's current state vector and cognitive preference vector. The multimodal teaching content includes text explanation, image assistance, dynamic animation and experimental interaction scripts, and is embedded with the explanation logic structure, anchor point association and cognitive load adjustment mechanism.

[0154] The generation process of multimodal teaching content is as follows:

[0155] A hybrid architecture is generated based on the hybrid expert model and the retrieval-enhanced generative model. The hybrid architecture includes:

[0156] Visual Expert, Visual Expert is used to generate SVG animations, spatial compositions, geometric diagrams, and physical force diagrams.

[0157] Language experts are used to generate structured text explanations, analogies, metaphors, and comparative arguments.

[0158] Operations Expert, which is used to build interactive experiment scripts, code generation, and virtual evaluator configuration.

[0159] The language large model collaborative core is used to provide unified semantic construction, context alignment and cross-expert style scheduling capabilities.

[0160] After receiving the task node and the student's current state vector, the cognitive style vector is parsed. The cognitive style vector is a personality preference indicator extracted from the student's long-term behavioral data, and the hybrid architecture weight is assigned through the gating function.

[0161] Taking the task node as input, vector semantic recall is performed from the knowledge retrieval library through a retrieval model based on a dual-tower structure to obtain the most relevant knowledge item set, which is then sent to the large model generator to perform imperative content generation and output the generated content.

[0162] Add semantic tags and trackable anchors to generated content.

[0163] The personalized content generation unit combines students' cognitive preferences and task plans to generate multimodal teaching content with clear structure and matching style, control cognitive load and provide explanation anchor points.

[0164] The personalized path planning module realizes personalized learning path generation and content scheduling based on student status, cognitive style and knowledge structure. Through multi-agent collaboration and reinforcement learning optimization mechanism, it dynamically adapts learning tasks and explanation styles to improve individualized teaching accuracy and learning continuity.

[0165] The reasoning chain construction module includes a thinking path extraction unit, a path alignment analysis unit and a logical deviation identification unit.

[0166] The thinking path extraction unit is used to extract the task node q i Taking the directed acyclic graph knowledge graph and the behavioral trend characteristics in the current student cognitive state as the starting point, the task target path tree is constructed. The task target path tree is the shortest effective explanation path of the task node on the knowledge graph.

[0167] During the task execution process, the reasoning-guided intelligent agent constructs the student's actual reasoning path in real time. The student's actual reasoning path extracts behavior fragments, reasoning intermediaries and explicit expressions from the question-answering behavior sequence, and maps it into a structural path diagram through the thinking chain generator to extract explicit reasoning steps from the question-answering behavior and operation log.

[0168] The thinking path extraction unit constructs a task target path tree and extracts key behavioral fragments and thinking nodes in the students' actual reasoning process to form a structured student reasoning path diagram.

[0169] The path alignment analysis unit is based on a structured attention graph network to model student behavior sequences, extract paths, and align sequential nodes.

[0170] Compare the students' actual reasoning path with the task target path tree to detect whether the following structural deviations exist:

[0171] Skip step: Skip the preset necessary nodes.

[0172] Wrong chain,wrong chain means that there is invalid or incorrect reasoning in the student's actual reasoning path.

[0173] Weak links: weak links refer to nodes that do not receive enough attention.

[0174] The path alignment analysis unit aligns the node sequence of students' actual paths and target paths based on the structured attention network to identify the structural differences and cognitive gaps between the paths.

[0175] The logical deviation identification unit is used to encode structural deviations into cognitive anomaly vectors and transmit them to the cognitive evaluation agent for evaluation and weighting. If the cognitive anomaly vector exceeds the preset tolerance threshold, the emotion regulation agent intervention is triggered and the teacher collaborative agent prompts the insertion of a review task.

[0176] The student's actual reasoning path is rendered as a visual structure diagram, which includes steps, risk nodes, and missed links. The visual structure diagram is embedded in the student terminal interface and teacher dashboard.

[0177] The logical deviation identification unit detects structural deviations such as skipped steps, wrong links, and weak links in students' reasoning paths, generates cognitive anomaly vectors, and uses them as the basis for intervention and evaluation; at the same time, it outputs a visual reasoning structure diagram to facilitate teacher understanding and feedback.

[0178] The reasoning chain construction module is used to construct and analyze students' reasoning paths during task execution, identify thinking deviations by comparing them with the target path, generate structured feedback and cognitive abnormality indicators, assist in evaluation and intervention decisions, and improve the system's interpretability and diagnostic capabilities.

[0179] The cognitive assessment feedback module includes a state change analysis unit, an error attribution modeling unit and a feedback strategy generation unit.

[0180] The state change analysis unit is used to obtain the student's actual reasoning path, behavioral trend characteristics in the current student's cognitive state, historical task trajectory, actual behavior data and task target path tree. The actual behavior data includes answer options, time distribution, cancellation frequency and interface focus. Based on deep knowledge tracking and graph structure reinforcement modeling, a mastery state vector is constructed to predict the mastery evolution trajectory of knowledge points in the future path. The mastery evolution trajectory is used to update the directed acyclic graph knowledge graph.

[0181] The state change analysis unit integrates students' reasoning paths, behavioral data, and historical task trajectories, and through knowledge tracking and graph structure modeling, dynamically predicts the evolution path of knowledge mastery and updates the knowledge graph status.

[0182] The error attribution modeling unit is used to identify high-contribution nodes in students' actual reasoning paths based on gradient-weighted class activation mapping variants, quantify students' actual reasoning paths based on concept activation vector tests, and output quantitative results. The quantitative results include reasoning completeness, unit standardization, and graphic understanding. Based on Shapley additive interpretation, the marginal influence of each input feature in the current task output is calculated and analyzed, and a knowledge radar chart, a concept heat map, and a strategy error analysis chart are output. The strategy error analysis chart is input into the cognitive evaluation agent for error cause classification, and the error cause type is output. The error cause type includes skipping errors, misconception mixing, and path circular thinking, and the evaluation results are output.

[0183] The error attribution modeling unit uses a variety of interpretable analysis methods to identify key nodes and feature influences in students' reasoning paths, outputs multidimensional analysis maps and classifies error types, thereby improving the granularity and accuracy of error analysis.

[0184] The feedback strategy generation unit outputs path adjustment prompts and explanation reconstruction suggestions based on the evaluation results.

[0185] The feedback strategy generation unit outputs strategy recommendations for path optimization and explanation content adjustment based on the evaluation and error cause results, thereby enabling the system to adaptively respond to the learning process.

[0186] The cognitive assessment feedback module comprehensively evaluates changes in students' knowledge mastery, reasoning quality, and error types during the learning process. It generates diagnostic results and strategy feedback based on multi-source data, supports personalized path adjustment and teaching content optimization, and improves the accuracy and intelligence of learning closed-loop feedback.

[0187] The emotion regulation intervention module includes an exponential function construction unit and a micro-intervention unit:

[0188] The index function construction unit is used to construct a dynamic motivation index function based on rational emotive behavior therapy, real-time emotion recognition, cognitive fatigue modeling, behavioral stagnation detection, and community incentive mechanism. The construction process of the dynamic motivation index function is as follows:

[0189] Physiological and behavioral features are extracted through a multimodal emotion perception network. Physiological and behavioral features include:

[0190] Facial expression parameters,Facial expression parameters include the degree of activation of specific facial muscle groups, brow tension, and eyelid movement.

[0191] Eye movement trajectory features,Eye movement trajectory features include gaze drift, number of fixations, and pupil dilation rate.

[0192] Operation behavior indicators, which include task dwell time, continuous click frequency, and operation interruption times.

[0193] Audio emotion indicators include abnormal speaking speed, falling intonation, and long pauses.

[0194] Background social signals include current class learning enthusiasm and social ranking performance.

[0195] The index function construction unit integrates rational emotional behavior models, emotion recognition, fatigue and behavioral stagnation modeling, and community dynamics to construct a dynamic motivation index function that reflects students' current motivational status and is used to determine the timing of intervention.

[0196] The micro-intervention unit is used to construct a representation vector of the student's current emotional state and behavioral motivation state based on physiological and behavioral characteristics, input the multi-factor motivation function into the motivation sensor, and obtain the multi-factor motivation value. If the multi-factor motivation value is greater than the preset high motivation threshold, an additional challenge task is added. If the multi-factor motivation value is less than the preset low motivation threshold, the micro-intervention strategy is triggered. If the multi-factor motivation value is between the low motivation threshold and the high motivation threshold, the task path is maintained unchanged and the enhanced monitoring state is entered.

[0197] Micro-intervention strategies include content structure adjustment, external incentive intervention and emotional guidance feedback. Content structure adjustment is used to split the current task into small knowledge units and reduce cognitive density. External incentive intervention is used to push short-term community incentive events. Emotional guidance feedback is used to call an affinity text generation model to generate text with an empathetic language style. The affinity text generation model is based on a language generation model, integrating the learner's current emotional state and behavioral characteristics, and outputting text with a comforting, motivational or empathetic language style to support motivational intervention and cognitive recovery.

[0198] The micro-intervention unit calculates motivation values ​​based on students' emotional and behavioral characteristics, and intelligently decides whether to add challenging tasks, trigger micro-interventions, or maintain the current path. It also implements fine-grained interventions through content adjustments, motivational events, or friendly language feedback.

[0199] The emotion regulation intervention module realizes dynamic monitoring and intervention of students' emotional state and motivation level during the learning process. Based on multimodal perception and behavioral modeling, it triggers timely and adjustable intervention mechanisms to prevent learning interruptions caused by emotional fluctuations, and improve learning continuity and the intelligence level of psychological support.

[0200] The teacher collaborative intervention module includes strategy injection unit, strategy co-construction unit and assisted reasoning unit.

[0201] The strategy injection unit is used to adjust the strategy through the natural language interface. The strategy adjustment includes modifying the task difficulty weight, rearranging the order of knowledge nodes, specifying the explanation style or inserting teaching intervention points. The strategy adjustment is parsed into a strategy adjustment vector, and the original path generation strategy is updated in a targeted manner through the strategy co-evolution mechanism.

[0202] The strategy injection unit supports teachers to flexibly adjust teaching strategies through a natural language interface, converting adjustment operations into vector form and injecting them into the system to guide the directional evolution of task paths and explanation styles.

[0203] The strategy co-construction unit is used to input the visual structure diagram, knowledge radar diagram, concept heat map and strategy error analysis diagram into the teacher terminal through the three-dimensional dashboard. The teacher terminal performs the following operations: checking the knowledge breakpoints and score loss frequency bands within the group, judging the overall learning sentiment trend of students, adjusting the search priority of the path generator for a certain knowledge area, configuring the reward threshold and community incentive frequency and content style switching logic, converting the operation path into the teacher operation trajectory encoding vector, and inputting it into the strategy distillation network to merge it with the preset reinforcement learning strategy.

[0204] The strategy co-construction unit provides a multi-dimensional visual map and a three-dimensional dashboard interactive interface to assist teachers in judging the learning status of student groups and configuring personalized strategies. Its operation trajectory will be used as a decision vector to participate in system strategy optimization.

[0205] The auxiliary reasoning unit is used to generate a set of candidate strategies and send a voting signal for the teacher to participate in the path when strategy oscillation occurs in the system path planning or the content module falls into style conflict.

[0206] When there is a strategy conflict or content style fluctuation, the auxiliary reasoning unit generates candidate strategies for teachers to choose from, supports a path voting mechanism, and realizes strategy correction and style coordination under human-computer co-construction.

[0207] The teacher collaborative intervention module realizes deep collaboration between teachers and intelligent teaching systems. Through strategy input and path co-construction mechanism, it enhances the human guidance and controllability of path planning and content generation process, and improves the accuracy, interpretability and adaptability of teaching intervention.

[0208] Consistent data transmission between modules is achieved through a shared state pool, and task triggering and feedback response are decoupled with the help of an asynchronous intent message bus. The standard student state vector is automatically updated and broadcast after each task is completed.

[0209] Example 2, refer to Figure 2 , provides a collaborative personalized learning method based on a large model, including the following steps:

[0210] Step S1: Collect and fuse multimodal data to generate a student state vector.

[0211] Step S2: Plan a personalized learning path based on the state and knowledge graph.

[0212] Step S3: Generate multimodal explanation content that suits the student's style.

[0213] Step S4: extract students' reasoning paths and identify thinking deviations.

[0214] Step S5: Analyze task performance and generate path and content adjustment suggestions.

[0215] Step S6, identifying motivational states and triggering intervention strategies.

[0216] Step S7, receiving teacher strategies and implementing path and style co-construction.

[0217] The present invention constructs a high-dimensional state vector through a student state-aware agent, covering core variables such as knowledge mastery, emotional state, and behavioral trends. This breaks away from the constraints of traditional label-based learning and achieves holistic cognitive and emotional modeling, from whether the answer is correct to why the error occurred, whether there is overload, and whether there is burnout. This provides a unified perception basis for subsequent path planning, explanation generation, and strategy intervention. The present invention realizes real-time prediction of cognitive evolution trends and path adaptive construction. It can not only dynamically adjust the difficulty and sequence of tasks, but also support teachers to manually intervene in path strategies through the strategy interface, constructing a hybrid optimal path generation mechanism of machine learning + human experience. The content generation agent dynamically dispatches different expert modules according to the student's cognitive style, historical performance, and task type, generating teaching content with matching style, adaptive rhythm, and clear expression. It not only evaluates the results, but also the process. Through reasoning, it guides the agent to explicitly construct the implicit reasoning path in the student task. Combined with the interpretable model of the evaluation agent, it performs graph-level tracing of cognitive breaks, logical jumps, and error chain structures, realizing a leap-forward cognitive diagnosis from answering behavior to thinking structure. The emotion index function introduced by the emotion regulation agent can identify the downward trend of students' emotional state before they experience overload, burnout or cognitive frustration, and automatically trigger lightweight strategic micro-interventions, effectively improving learning continuity, behavioral stability and task completion. Through the strategy interface and knowledge transfer channel designed by the teacher collaborative agent, teachers can not only visually control the status of student groups and individuals, but also directly participate in strategy generation processes such as path reconstruction, task priority setting, and explanation style adjustment. The teacher's behavior is encoded into a vector and integrated into the path strategy network, truly realizing the strategy co-construction and model co-education mechanism of teacher experience + AI evolution. The present invention supports edge deployment, and the key model supports quantitative compression and low-resource operation. The multi-agent architecture has natural scalability, supports cross-disciplinary teaching module access, agent role expansion and teaching strategy sharing, and can be widely adapted to various educational scenarios such as K12, vocational education, higher education and individual self-study.

[0218] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a fully hardware embodiment, a fully software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium may be implemented by any type of volatile or non-volatile storage device, or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0219] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A collaborative personalized learning system based on a large model, characterized by: It includes student cognitive state perception module, personalized path planning module, teaching content generation module, reasoning chain construction module, cognitive assessment feedback module, emotion regulation intervention module and teacher collaborative intervention module: The student cognitive state perception module is used to collect and fuse multimodal data to generate a student state vector; The personalized path planning module is used to plan a personalized learning path based on the state and knowledge graph; The teaching content generation module is used to generate multimodal teaching content that suits the students' style; The reasoning chain building module is used to extract students' reasoning paths and identify thinking deviations; The cognitive assessment feedback module is used to analyze task performance and generate path and content adjustment suggestions; The emotion regulation intervention module is used to identify motivational states and trigger intervention strategies; The teacher collaborative intervention module is used to receive teacher strategies and realize the co-construction of paths and styles.

2. A collaborative personalized learning system based on a large model as claimed in claim 1, characterized in that: The student cognitive state perception module includes a multimodal data acquisition unit, a feature fusion unit, an emotional state estimation unit and a state vector generation unit; The multimodal data acquisition unit is used to collect students' visual channel data, behavioral channel data, voice channel data and system-side data. The visual channel data includes camera image stream, eye movement trajectory and facial micro-expressions; the behavioral channel data includes click logs, answering time and path jump sequence; the voice channel data includes speaking speed, pauses and intonation changes; the system-side data includes completed tasks, error statistics and prerequisite node coverage; The feature fusion unit is used to perform multi-channel embedding processing on the visual channel data, the behavioral channel data, the voice channel data and the system side data to generate a modality representation vector, wherein the modality representation vector includes: K t =f GNN ; in, is the visual image vector, f ViT is the visual channel data corresponding to the visual image vector, is the behavior sequence vector, f AudioBERT is the behavior channel data corresponding to the behavior sequence vector, K t is the speech stream vector, f GNN is the voice channel data corresponding to the voice stream vector, K t is the knowledge graph state vector, f GNN is the system-side data corresponding to the knowledge graph state vector; The modal representation vectors are nested and fused through the multimodal alignment layer, and the cross-modal fusion module based on the attention mechanism is used to complete the weighted combination and output high-dimensional fusion features. The mathematical expression of the high-dimensional fusion features is: Among them, h t is the high-dimensional fusion feature, || is the vector concatenation operation; Based on the behavior channel data, the corresponding short-term behavior change trend is modeled through a sliding time window and weighted average method, and the output is the behavior trend feature; The emotional state estimation unit is used to predict and obtain the emotional four-dimensional index based on the visual channel data and the voice channel data based on the convolutional neural network and the long short-term memory network. The emotional four-dimensional index includes: pleasure e val , awakening degree aro , sense of control dom and polarity e pol ,The four-dimensional index of emotion is normalized into the emotion vector; The state vector generation unit is used to construct a standard student state vector based on the high-dimensional fusion feature, the knowledge graph state vector and the emotion vector. The mathematical expression of the standard student state vector is: S t ={h t ,K t ,E t }∈R d ; Among them, S t is the standard student state vector, R d Behavioral trend characteristics; Write the standard student state vector to the shared state pool.

3. A collaborative personalized learning system based on a large model as claimed in claim 2, characterized in that: The personalized path planning module includes a perception agent construction unit, a learning path planning unit and a personalized content generation unit; The perceptual agent building unit is used to establish an agent group, and the agent group includes: The student state perception agent is used to collect and encode the student's knowledge mastery status, behavior patterns and emotional variables, and output a standard state vector S t , and push perception events when the state mutates; The learning path planning agent is used to perform path search and dynamic planning on the directed acyclic graph task graph based on the student's current cognitive map and the prerequisite structure graph to generate a task sequence Q t And maintain the task execution context; A teaching content generation agent, which is used to receive the task plan and style vector, generate multimodal teaching content based on the student profile through a hybrid expert model and a retrieval-enhanced generation model, and automatically match the visual, linguistic or interactive teaching style; Reasoning guides the agent, which is used to guide the agent around the task q i Constructing students' individual reasoning paths T path , complete the process of thinking chain construction, hypothesis verification and error reflection; A cognitive evaluation agent, which is used to construct a cognitive evaluation graph, explainability indicators, and error cause location based on task behavior records and reasoning path results, and output them to the path planning and teacher agents; An emotion regulation agent, which is used to monitor multimodal data and behavioral trend characteristics in real time, detect trends from overload to cooling down to abandonment, and adjust task intensity or trigger incentive events; Teacher collaborative agent, which is used to provide natural language interaction, policy parameter adjustment interface and convert it into policy offset vector Δπ teacher ; The communication mechanism between the agent groups includes: Shared state pool, which is used to store all agents' state variables in real time through the shared state pool storage platform. The shared state pool supports state caching, long-term trajectory maintenance and historical query; Intent event bus, which is used to register the intent type of each agent with the bus. Events are transmitted in the form of asynchronous messages, supporting priority scheduling, conflict decoupling, and event tracing; The learning path planning unit is used to construct a directed acyclic graph knowledge graph G based on the course structure and prerequisite dependencies. K =(V,E), where each node v i ∈V represents a fine-grained learning unit, and edge e ij ∈E represents the pre-dependence of knowledge, and the student’s current knowledge coverage weight function κ(v i ,S t ), the student's current knowledge coverage weight function represents the degree of mastery of the knowledge unit; Modeling a task sequence decision process that constructs a state set based on the student's current learning trajectory and the local subgraph structure in the directed acyclic graph knowledge graph; According to the knowledge mastery and precondition relationship in different states in the state set, the next learning task is selected from the optional task nodes as the action set; The task transfer relationship is determined based on the changes in the knowledge graph structure and the student's cognitive state, and a task reward mechanism is constructed based on knowledge gain, learning interest, emotional stability, and the time cost required to complete the task; A graph-structure-related reinforcement learning strategy optimization method is used to train the task sequence decision process and generate an integrated path. The mathematical expression of the objective function of the optimization method of the task sequence decision process is: Among them, S is the current learning trajectory and the local subgraph structure in the directed acyclic graph knowledge graph, a t is the action set, is the reward discount factor, α, β, γ and δ are the strategy adjustment weights, ΔK t The improvement in knowledge mastery estimated by the knowledge tracking model, d task is the task difficulty, d student is the degree of matching of students’ current abilities, M social is the community impact factor, C load To estimate the task cognitive load by combining the four-dimensional indicators of emotion and behavioral trend characteristics; If the cognitive assessment agent detects a failure in task understanding, serious skipping of questions, and a downward trend in mastery, it triggers a replanning mechanism, recalculates the local path map, inserts a backtracking task, or reduces the difficulty of the fine-tuning strategy.

4. A collaborative personalized learning system based on a large model as claimed in claim 3, characterized in that: The personalized content generation unit is used to generate multimodal teaching content based on the integration path and in combination with the student's current state vector and cognitive preference vector. The multimodal teaching content includes text explanation, image assistance, dynamic animation and experimental interaction script, and is embedded with the explanation logic structure, anchor point association and cognitive load adjustment mechanism; The generation process of the multimodal teaching content is as follows: A hybrid architecture is generated based on the hybrid expert model and the retrieval-enhanced generative model, wherein the hybrid architecture includes: Visual experts for generating SVG animations, spatial compositions, geometric diagrams, and physical force diagrams; Language experts, said language experts are used to generate structured text explanations, analogies, metaphors, and comparative arguments; Operation experts, used for building interactive experiment scripts, code generation, and virtual evaluator configuration; A language model collaboration core that provides unified semantic construction, context alignment, and cross-expert style scheduling capabilities. After receiving the task node and the student's current state vector, the cognitive style vector is parsed. The cognitive style vector is a personality preference indicator extracted from the student's long-term behavioral data, and the hybrid architecture weight is assigned through the gating function. Taking the task node as input, the system performs vector semantic recall from the knowledge retrieval database through a retrieval model based on a dual-tower structure to obtain the most relevant knowledge item set, which is then fed into the large model generator for imperative content generation and outputs the generated content. Add semantic tags and trackable anchors to generated content.

5. A collaborative personalized learning system based on a large model as claimed in claim 4, characterized in that: The reasoning chain building module includes a thinking path extraction unit, a path alignment analysis unit and a logic deviation identification unit; The thinking path extraction unit is used to extract the task node q i As the starting point, the directed acyclic graph knowledge graph and the behavioral trend characteristics of the current student cognitive state are called to construct a task target path tree, which is the shortest effective explanation path of the task node on the knowledge graph; During task execution, the reasoning-guided agent constructs the student's actual reasoning path in real time. The student's actual reasoning path extracts behavioral fragments, reasoning intermediaries, and explicit statements from the answering behavior sequence. The thinking chain generator maps the structured path diagram to extract explicit reasoning steps from the answering behavior and operation log. The path alignment analysis unit is based on a structured attention graph network to model student behavior sequences, extract paths, and align sequential nodes; Compare the students' actual reasoning path with the task target path tree to detect whether the following structural deviations exist: Skipping, wherein a preset necessary node is skipped; Wrong chain, where the wrong chain refers to invalid or incorrect reasoning in the student's actual reasoning path; Weak links, where the weak links are nodes with insufficient attention; The logical deviation identification unit is used to encode the structural deviation into a cognitive anomaly vector and transmit it to the cognitive assessment agent for evaluation and weighting. If the cognitive anomaly vector exceeds a preset tolerance threshold, the emotion regulation agent is triggered to intervene and the teacher collaborative agent prompts the student to insert a review task. The student's actual reasoning path is rendered as a visual structure diagram, which includes steps, risk nodes and missed solution links, and the visual structure diagram is embedded in the student terminal interface and teacher dashboard.

6. A collaborative personalized learning system based on a large model as claimed in claim 5, characterized in that: The cognitive assessment feedback module includes a state change analysis unit, an error attribution modeling unit and a feedback strategy generation unit; The state change analysis unit is used to obtain the student's actual reasoning path, behavioral trend characteristics in the current student's cognitive state, historical task trajectory, actual behavior data and task target path tree. The actual behavior data includes answer options, time distribution, revocation frequency and interface focus. Based on deep knowledge tracking and graph structure reinforcement modeling, it constructs a mastery state vector and predicts the mastery evolution trajectory of knowledge points in the future path. The mastery evolution trajectory is used to update the directed acyclic graph knowledge graph; The error attribution modeling unit is used to identify high-contribution nodes in the student's actual reasoning path based on a gradient-weighted class activation mapping variant, quantify the student's actual reasoning path based on a concept activation vector test, and output a quantitative result, wherein the quantitative result includes reasoning completeness, unit standardization, and graphic comprehension; calculate and analyze the marginal influence of each input feature in the current task output based on Shapley additive interpretation, output a knowledge radar chart, a concept heat map, and a strategy error analysis chart; input the strategy error analysis chart into the cognitive evaluation agent for error cause classification, output the error cause type, wherein the error cause type includes step skipping error, misconception mixing, and path circular thinking, and output an evaluation result; The feedback strategy generation unit outputs path adjustment prompts and explanation reconstruction suggestions based on the evaluation results.

7. The collaborative personalized learning system based on a large model according to claim 6, characterized in that: The emotion regulation intervention module includes an exponential function construction unit and a micro-intervention unit: The index function construction unit is used to construct a dynamic motivation index function based on rational emotive behavior therapy, real-time emotion recognition, cognitive fatigue modeling, behavioral stagnation detection and community incentive mechanism. The construction process of the dynamic motivation index function is as follows: Physiological and behavioral features are extracted through a multimodal emotion perception network. The physiological and behavioral features include: Facial expression parameters, including the degree of activation of specific facial muscle groups, brow tension, and eyelid movement; Eye movement trajectory features, including gaze drift, number of fixations, and pupil dilation rate; Operation behavior indicators, including task dwell time, continuous click frequency, and operation interruption times; Audio emotion indicators, including abnormal speaking speed, falling intonation, and long pauses; Background social signals, including current class learning popularity and social leaderboard performance; The micro-intervention unit is used to construct a representation vector of the student's current emotional state and behavioral motivation state based on physiological and behavioral characteristics, input the vector into the motivation sensor to calculate the multi-factor motivation function, and obtain the multi-factor motivation value. If the multi-factor motivation value is greater than a preset high motivation threshold, an additional challenge task is added. If the multi-factor motivation value is less than a preset low motivation threshold, a micro-intervention strategy is triggered. If the multi-factor motivation value is between the low motivation threshold and the high motivation threshold, the task path is maintained unchanged and the enhanced monitoring state is entered; The micro-intervention strategy includes content structure adjustment, external incentive intervention and emotional guidance feedback. The content structure adjustment is used to split the current task into small knowledge units and reduce cognitive density. The external incentive intervention is used to push short-term community incentive events. The emotional guidance feedback is used to call an affinity text generation model to generate text with an empathetic language style.

8. The collaborative personalized learning system based on a large model according to claim 7, characterized in that: The teacher collaborative intervention module includes a strategy injection unit, a strategy co-construction unit and an auxiliary reasoning unit; The strategy injection unit is used to adjust the strategy through a natural language interface. The strategy adjustment includes modifying the task difficulty weight, rearranging the order of knowledge nodes, specifying the explanation style, or inserting teaching intervention points. The strategy adjustment is parsed into a strategy adjustment vector, and the original path generation strategy is updated in a targeted manner through a strategy co-evolution mechanism. The strategy co-construction unit is used to input the visual structure diagram, knowledge radar diagram, concept heat map and strategy error analysis diagram into the teacher terminal through the three-dimensional dashboard. The teacher terminal performs the following operations: viewing the knowledge breakpoints and score loss frequency bands within the group, judging the overall learning sentiment trend of students, adjusting the search priority of the path generator for a certain knowledge area, configuring the reward threshold and community incentive frequency and content style switching logic, converting the operation path into the teacher operation trajectory encoding vector, and inputting it into the strategy distillation network for fusion with the preset reinforcement learning strategy; The auxiliary reasoning unit is used to generate a candidate strategy set and send a teacher participation path voting signal when strategy oscillation occurs in the system path planning or the content module falls into style conflict.

9. The collaborative personalized learning system based on a large model according to claim 8, characterized in that: The method realizes data consistency transmission among modules through a shared state pool and decouples task triggering from feedback response with the help of an asynchronous intent message bus, wherein the standard student state vector is automatically updated and broadcast after each task is completed.

10. A collaborative personalized learning method based on a large model, applied to a collaborative personalized learning system based on a large model as claimed in any one of claims 1 to 9, characterized in that: The following steps are involved: Step S1, collect and fuse multimodal data to generate student state vector; Step S2: planning a personalized learning path based on the state and knowledge graph; Step S3, generating multimodal explanation content that suits the students' style; Step S4, extracting students’ reasoning paths and identifying thinking deviations; Step S5, analyzing task performance and generating path and content adjustment suggestions; Step S6, identifying motivational states and triggering intervention strategies; Step S7, receiving teacher strategies and implementing path and style co-construction.

Images