Multi-agent dynamic task planning and cooperation method and system for assisting teaching process

Abstract

The invention discloses a multi-agent dynamic task planning and cooperation method and system for assisting a teaching process. The multi-agent dynamic task planning and cooperation method for assisting the teaching process comprises the steps of obtaining task demand information; through multi-agent task planning and path design, a structured task planning graph and an Agent cooperation path graph are generated according to task demand information; and generating an executable final task planning scheme according to the structured task planning graph and the Agent cooperation path graph. By introducing a multi-agent dynamic cooperation mechanism, the system can dynamically disassemble a complex comprehensive learning task and distribute the task to sub-agents with different specialities for parallel processing according to autonomous planning of a large language model. The expert cooperation mode effectively overcomes the limitation of a single agent in the aspects of professional skills and parallel processing, so that the system can flexibly meet the diversified learning requirements of high scene coupling degree and complex process, and output a deeper and more complete result with a rigorous structure.

Description

Technical Field

[0001] This application relates to the field of multi-agent collaborative technology, specifically to a multi-agent dynamic task planning and collaboration method and system for assisting the teaching process. Background Technology

[0002] The greatest potential of AI agents lies in their ability to "autonomously complete tasks." In the past, AI was largely seen as a "production tool," assisting people in completing various tasks. Now, with the development of AI agents, AI is gradually evolving from a production tool into a "productivity force" itself. Essentially, an AI agent is an intelligent system composed of autonomy and action, which can be figuratively described as a collaborative structure of "brain + hand." The "brain" must not only be able to think autonomously but also interact with the environment and dynamically adjust its behavioral strategies according to changes in the environment; the "hand" needs to directly complete tasks according to the instructions of the "brain" (e.g., Deep Research) and can also use external tools (e.g., tool calling). Its behavior is no longer a static response but a complete cycle including planning, execution, and adjustment, thus achieving a true task closure loop.

[0003] In the AI ​​Agent model, when faced with tasks that assist in the teaching process, the capabilities of a single intelligent agent are increasingly limited, making it difficult to effectively solve problems such as high scene coupling and difficulty in parallel processing. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-agent dynamic task planning and collaboration method to assist in the teaching process, at least solving one of the above-mentioned technical problems.

[0005] One aspect of the present invention provides a multi-agent dynamic task planning and collaboration method for assisting teaching processes, the multi-agent dynamic task planning and collaboration method for assisting teaching processes comprising: Obtain task requirement information; Through multi-agent task planning and path design, a structured task planning diagram and an agent collaboration path diagram are generated based on task requirement information. An executable final task planning scheme is generated based on the structured task planning diagram and the agent collaboration path diagram.

[0006] Optionally, obtaining task requirement information includes: Obtain natural language task information input by the user; Convert natural language task information into structured task data; Generate task feature vectors based on structured task data; Generate collaborative mode decision results based on task feature vectors; Task requirement information is generated based on structured task data, task feature vectors, and collaborative mode decision results.

[0007] Optionally, the conversion of natural language task information into structured task data includes: Obtain a pre-defined knowledge base for the teaching domain; The input natural language task text is preprocessed to obtain the preprocessed task text encoding; The task text encoding is compared with a pre-set knowledge base for the teaching domain to obtain a set of relevant search results; Obtain the trained BERT-NER algorithm; The text to be input is obtained by concatenating the natural language task text with the relevant search result set. The text information to be input is fed into the trained BERT-NER algorithm to obtain structured task data.

[0008] Optionally, generating the task feature vector from the structured task data includes: Obtain a pre-defined evaluation index system for the complexity of teaching tasks; Extract multiple indicator features from structured task data; The characteristics of each indicator are normalized to obtain the normalized indicator characteristics. Obtain a trained lightweight multilayer perceptron model; The normalized index features are input into a trained lightweight multilayer perceptron model to obtain a task complexity score. Obtain a pre-defined library of rules for classifying teaching task constraints; Structured task data is classified into constraint types and scored for constraint strength using a pre-defined teaching task constraint classification rule base, thereby obtaining constraint strength vectors and constraint classification information; A task feature vector is generated based on structured task data, constraint strength vector, constraint classification information, and task complexity score.

[0009] Optionally, generating collaborative mode decision results based on task feature vectors includes: Obtain the collaboration mode; Generate a high-order feature interaction map based on the task feature vector and the collaborative mode; A dynamic potential weight matrix is ​​generated based on the high-order feature interaction graph. Evolutionary robustness adaptation score of collaborative mode is generated based on high-order feature interaction map and dynamic potential weight matrix. Candidate results for collaborative patterns are generated based on the evolutionary robustness adaptation score of collaborative patterns, high-order feature interaction maps, and dynamic potential weight matrices.

[0010] Optionally, generating a high-order feature interaction map based on the task feature vector includes: Each feature vector in the task feature vector is treated as a feature node, and the initial weight of the node is set to the corresponding feature value. For any k-order feature combination, calculate the interaction gain value G based on the knowledge base K, retain only the beneficial interactions with G≥0.1, construct the interaction edges between feature nodes, and the edge weight is the gain value G; The attention aggregation strategy encodes 3rd to 5th order high-order interaction information, generates interaction enhancement feature values ​​for each feature node, and thus generates a high-order feature interaction graph. The high-order feature interaction graph includes multiple feature nodes and multiple interaction edges, and labels the interaction enhancement feature values ​​and edge weights of each node. Optionally, generating the dynamic potential weight matrix based on the high-order feature interaction map includes: The initial dynamic potential weight matrix W is obtained from the high-order feature interaction map; The dynamic potential energy weight matrix is ​​generated based on the initial dynamic potential energy weight matrix W.

[0011] Optionally, the evolutionary robustness adaptation score of the cooperative mode generated based on the high-order feature interaction map and dynamic potential weight matrix includes: Obtain the trained LSTM prediction model and the trained generative adversarial network; The task feature vector is input into the trained LSTM prediction model to obtain the probability distribution of scene feature vector changes; Virtual interference scenarios are generated using generative adversarial networks; Generate an adaptation score for the virtual interference scenario corresponding to each collaborative mode; Calculate the resource coupling degree for each collaboration mode; The evolutionary robustness adaptation score of each collaborative mode is generated based on the probability distribution of scene feature vector changes, the adaptation score of virtual interference scene, and resource coupling degree.

[0012] Optionally, the step of generating candidate results for cooperative modes based on the evolutionary robustness adaptation score of the cooperative mode, the high-order feature interaction map, and the dynamic potential weight matrix includes: Based on the high-order feature interaction graph and collaborative patterns, feature-pattern adaptation bipartite graph and pattern-scene community graph are generated; The feature-pattern adaptation bipartite graph and the pattern-scenario community graph are fused to obtain collaborative patterns that meet preset conditions as candidate results for collaborative patterns.

[0013] This application also provides a multi-agent dynamic task planning and collaboration system for assisting teaching processes, the multi-agent dynamic task planning and collaboration system for assisting teaching processes includes: A task requirement information acquisition module, wherein the task requirement information acquisition module is used to acquire task requirement information; The graph generation module is used to generate a structured task planning graph and an agent collaboration path graph based on task requirement information through multi-agent task planning and path design. The final task planning scheme generation module is used to generate an executable final task planning scheme based on the structured task planning diagram and the Agent collaboration path diagram.

[0014] This application's multi-agent dynamic task planning and collaboration method for assisting teaching processes introduces a multi-agent dynamic collaboration mechanism. Based on the autonomous planning of a large language model, the system can dynamically decompose complex, comprehensive learning tasks (such as project-based learning and in-depth research report writing) and assign them to sub-agents with different specialties (such as search, encoding, analysis, and report generation agents) for parallel processing. This "expert collaboration" model effectively overcomes the limitations of a single agent in terms of knowledge breadth, professional skills, and parallel processing, enabling the system to flexibly respond to diverse learning needs with high scenario coupling and complex processes, outputting more in-depth, complete, and structurally rigorous results. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating a multi-agent dynamic task planning and collaboration method for an auxiliary teaching process according to an embodiment of this application. Figure 2 yes Figure 1 The diagram shows the interface interaction in the multi-agent dynamic task planning and collaboration method of the auxiliary teaching process. Figure 3 yes Figure 1 The diagram illustrates the task requirement description in the multi-agent dynamic task planning and collaboration method of the auxiliary teaching process. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be described in more detail below with reference to the accompanying drawings. In the drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The described embodiments are some, but not all, embodiments of this application. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application. The embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0017] like Figure 1 The multi-agent dynamic task planning and collaboration method for the auxiliary teaching process shown includes: Obtain task requirement information; Through multi-agent task planning and path design, a structured task planning diagram and an agent collaboration path diagram are generated based on task requirement information. An executable final task planning scheme is generated based on the structured task planning diagram and the agent collaboration path diagram.

[0018] In this embodiment, obtaining task requirement information includes: Obtain natural language task information input by the user; for example, complex natural language teaching tasks submitted by the user (such as "develop a midterm exam review plan"). Obtain a pre-defined knowledge base for the teaching domain (including SLO assessment standard documents, historical teaching task cases, and constraint templates, which have been segmented and preprocessed at the paragraph level); store the knowledge base for the teaching domain using a vector database; convert each segment into a dense vector of fixed dimensions using a pre-trained language model; normalize the vectors and then build a retrieval index to improve subsequent query efficiency. The input natural language task text is preprocessed to obtain the preprocessed task text encoding; specifically, the input natural language task text is preprocessed (special characters are removed and the text format is unified), and the task text is converted into task vectors with the same vector dimension as the knowledge base preprocessing through the same pre-trained language model, ensuring vector space alignment. The task text encoding is compared with a pre-defined teaching domain knowledge base to obtain a relevant search result set. Specifically, the similarity between the task vector and the knowledge base vector is calculated using the cosine similarity algorithm. The results are sorted in descending order of similarity score, and the top 5 relevant results are selected (including the unique case identifier, original paragraph, similarity score, and category: SLO standard / task case / constraint template). Invalid results with a similarity score below 0.6 are filtered out. Then, the filtering tool of JYD-Tool is used to remove duplicate cases and templates that do not match the task scenario (such as filtering out elementary school teaching cases corresponding to high school physics tasks).

[0019] Obtain the trained BERT-NER algorithm; in this embodiment, the NER model adopts the BIO annotation system and defines four types of core entity labels for teaching tasks, namely, task target (labeled as B-Target), time constraint (labeled as B-Time), output format (labeled as B-Format), and teaching scenario (labeled as B-Scene), covering the core key information types in teaching tasks; The text to be input is obtained by concatenating the natural language task text with the relevant search result set (the original paragraphs in the Top-5 relevant search result set); The text information to be input is fed into the trained BERT-NER algorithm to obtain structured task data. Specifically, the concatenated text is input into the fine-tuned NER model, which outputs an entity label sequence. The label sequence is parsed by a decoding algorithm to extract four types of key entities. The extracted entities are deduplicated (if the same task objective appears repeatedly, only one is retained), and the relevant information of the entity source text is recorded to obtain a structured entity set (JSON format, containing core fields such as unique task identifier, task objective, constraint array, and scene label set) and an entity extraction report (including the number of entities of each type, average confidence, and description of unidentified entity types).

[0020] Generate task feature vectors based on structured task data; In this embodiment, generating a task feature vector based on structured task data includes: A pre-set teaching task complexity evaluation index system (pre-set 3 core input dimensions: number of task sub-modules, number of constraints, and complexity level of scene labels, generated based on Multi-Agent collaborative scene samples).

[0021] Three types of input indicators are extracted from the structured task data: the number of task sub-modules (e.g., the midterm exam review plan is broken down into four modules: information collection, key point analysis, time planning, and output, counted as 4); the number of constraints (the total number of time, outcome, and teaching objective constraints in the constraint array, such as completing the midterm exam 10 days before the exam + web-based output + meeting the SLO assessment criteria, counted as 3); and the complexity level of the scenario labels (according to the preset rules of the teaching scenario feature library, each label in the scenario label set is mapped to a level from 1 to 5, with basic scenarios = 1, regular teaching scenarios = 2, comprehensive application scenarios = 3, in-depth research scenarios = 4, and cross-domain integration scenarios = 5, such as high school physics project-based learning + SLO assessment, counted as 4). Indicator normalization: The Min-Max normalization method is used to map the values ​​of the three types of indicators to the 0-1 range; Complexity modeling: A lightweight multilayer perceptron (MLP) model is adopted. The input layer consists of 3 normalized indicators, the hidden layer has 16 neurons (using the ReLU activation function), and the output layer has 1 neuron (using the Sigmoid activation function). The output is a complexity score in the range of 0-1 (0 = low complexity, 1 = high complexity). The model is trained and optimized based on 1000+ teaching task samples (covering SLO assessment, project-based learning, deep report writing, etc.) to ensure that the score is highly correlated with the actual task complexity, thereby obtaining the task complexity score (in the range [0,1], with 2 decimal places, such as 0.78).

[0022] Obtain a pre-defined library of teaching task constraint classification rules (pre-defined 3 core constraint types: time constraint, outcome constraint, and teaching objective constraint, containing 12 classification rules, adaptable to teaching scenarios such as SLO assessment and formative assessment).

[0023] Based on the rule engine, classification rules are triggered to accurately determine each constraint: time constraints (e.g., completion 10 days before the midterm exam, first draft output within 3 days, no explicit time requirement); outcome constraints (e.g., web page format, structured report, data visualization charts, PPT-compatible format); and learning objective constraints (e.g., meeting SLO assessment standards, covering core requirements of project-based learning, adapting to formative assessment processes, and satisfying summative assessment indicators). Each type of constraint is scored from 1 to 5 points based on its level of strictness (1 = lenient, 2 = relatively lenient, 3 = moderate, 4 = relatively strict, 5 = strict). The scoring rules are based on a pre-set database of teaching scenario features: Time constraints (completion before a fixed date = 5 points, completion within a specified period = 4 points, no specific requirements = 1 point); Outcome constraints (multi-format adaptation = 4 points, single fixed format = 5 points, no special requirements = 2 points); Teaching objective constraints (requires verification through SLO assessment = 5 points, adapts to a specific assessment process = 4 points, only meets basic teaching objectives = 2 points). Vector normalization: The Min-Max normalization method is used to map the scores of the three types of constraints to the interval [0,1], generating a 3-dimensional constraint strength vector. The final vector format includes the time constraint strength score, the outcome constraint strength score, and the teaching objective constraint strength score, thus generating a 3-dimensional constraint strength vector (each dimension value is in the interval [0,1], and two decimal places are retained, such as (0.80, 0.60, 0.90)). Scene Dimension Mapping: The scene tags are semantically aligned with the 8 secondary sub-dimensions of the teaching scene feature library (containing 3 primary adaptation dimensions and 8 secondary sub-dimensions; primary dimensions: real-time requirements, content depth requirements, and process fixation requirements, adapting to the characteristics of three types of collaborative modes) (based on word vector similarity calculation, similarity ≥ 0.7 is considered a match), and mapped to the 3 primary dimensions: real-time requirements (secondary sub-dimensions: immediate response, short cycle completion, no real-time requirements); content depth requirements (secondary sub-dimensions: basic understanding, in-depth analysis, comprehensive application); process fixation requirements (secondary sub-dimensions: fixed process, flexible process). In this embodiment, the scene tag and teaching scene feature library includes the following: First-level dimension: Real-time requirements: It has three secondary sub-dimensions, which are the core measures of the intensity of the timeliness requirements for completing teaching tasks: Real-time response: This refers to tasks that need to be completed within minutes, with results directly serving the real-time teaching scenario. Typical examples include real-time Q&A in class, statistics on the accuracy of in-class assignments, and instant summarization of quiz scores. Short-cycle completion: This refers to tasks that need to be completed within hours or days, with clear short-term time constraints. Typical examples include developing a review plan for the three days before an exam, summarizing and correcting daily assignments, and making minor adjustments to classroom teaching materials for the next day. No real-time requirements: This means that the task has no strict time limit and can be completed within a few days or weeks, with a greater emphasis on the quality of the outcome. Typical examples include writing semester teaching outcome reports, designing cross-unit project-based learning programs, and compiling school-based curriculum textbooks.

[0024] First-level dimension: Content depth requirements: It has three secondary sub-dimensions, which are the core measures of the depth of knowledge mining and application required by the teaching tasks: Basic understanding: This refers to the task focusing on the basic review, memorization, and simple application of knowledge points, without requiring complex analysis. Typical examples include compiling lists of textbook knowledge points, arranging sets of basic practice questions, and designing outlines for memorizing formulas and theorems. In-depth analysis refers to tasks that require multi-dimensional breakdown of teaching data, student learning information, and other data to extract hidden patterns or core issues. Typical examples include analyzing the reasons for midterm exam score losses, identifying students' weak knowledge points, and analyzing differences in teaching effectiveness among classes. Integrated application: This refers to tasks that integrate knowledge points from multiple disciplines and modules, designing comprehensive teaching content or programs. Typical examples include the design of interdisciplinary project-based learning tasks, the planning of integrated practical activities, and the development of thematic review programs that integrate multiple knowledge points.

[0025] III. Primary Dimension: Requirements for Fixed Processes It has two secondary sub-dimensions, which primarily measure the standardization of the steps involved in executing teaching tasks: Fixed procedures: These refer to tasks executed with clearly defined, standardized steps and operational procedures that cannot be arbitrarily adjusted and must be strictly followed. Typical examples include unit test creation based on the teaching syllabus, generation of standard chemistry learning reports, and homework grading using fixed templates. Flexible processes: These refer to tasks executed without fixed steps, where the execution logic can be dynamically adjusted based on actual learning situations and teaching needs, emphasizing personalization and innovation. Typical examples include personalized learning path planning, innovative teaching activity design, and the development of tutoring programs for students with special needs.

[0026] Dimensional score calculation: Based on the mapping results, supplementary scores are added according to task complexity and constraint strength: Real-time requirement score (no real-time requirement = 0.2, short cycle completion = 0.6, immediate response = 0.9); Content depth requirement score (basic understanding = 0.3, in-depth analysis = 0.7, comprehensive application = 0.8, academic research = 1.0; if the task complexity score ≥ 0.7, add an additional 0.1); Process fixity requirement score (flexible process = 0.3, semi-fixed process = 0.6, fixed process = 0.9; if the result constraint strength score ≥ 0.8, add an additional 0.1). The scores of the three primary dimensions are integrated into a 3D scene adaptation vector, which is formatted as real-time requirement score, content depth requirement score, and process fixed requirement score. The values ​​of each dimension are all in the range of [0,1], thus obtaining the 3D scene adaptation vector (each dimension value is in the range of [0,1], and is retained to two decimal places, such as (0.20,0.80,0.90)).

[0027] The above features are combined to form the final task feature vector (1-dimensional task complexity score + 3-dimensional constraint strength vector + 3-dimensional scene adaptation vector).

[0028] Generate collaborative mode decision results based on task feature vectors; The step of generating collaborative mode decision results based on task feature vectors includes: In this embodiment, the collaboration modes include free configuration and combination, workflow orchestration and handover, and Plan-and-Execute (P&E) collaboration. Generate a high-order feature interaction map based on the task feature vector and the collaborative mode; In this embodiment, generating a high-order feature interaction map based on the task feature vector includes: Each feature vector in the task feature vector is treated as a feature node, and the initial weight of the node is set to the corresponding feature value. Specifically, the 7-dimensional task feature vector V = [c, t1, t2, t3, s1, s2, s3] (c = task complexity score, t1 = time constraint strength, t2 = outcome constraint strength, t3 = teaching objective constraint strength, s1 = real-time requirement score, s2 = content depth requirement score, s3 = process fixity requirement score); each dimension of the 7-dimensional feature vector V is regarded as a feature node, and the initial weight of the node is set to the corresponding feature value (e.g., if c = 0.8, then the weight of the node is 0.8). Obtain a pre-defined high-level interactive knowledge base K for teaching tasks (containing 5000+ characteristic interactive cases of teaching tasks, labeled as beneficial / ineffective interactions, such as high complexity c+ high content depth s2 as beneficial interactions in P&E mode). For any k-order feature combination, calculate the interaction gain value G based on the knowledge base K, retain only the beneficial interactions with G≥0.1, construct the interaction edges between feature nodes, and the edge weight is the gain value G; G = (Number of successful cases in the history of this combination / Total number of cases in this combination) − 0.5. Only beneficial interactions with G ≥ 0.1 are retained, and interaction edges between feature nodes are constructed. Edge weight = gain value G. Encode 3rd to 5th order high-order interaction information using an attention aggregation strategy to generate interaction-enhanced feature values ​​for each feature node. , ( Let w be the weight of the interaction edge between feature nodes i and j. For example, if the weight of the interaction edge between nodes c and s2 is w=0.3, then... This generates a high-order feature interaction graph, which includes multiple feature nodes and multiple interaction edges, and labels the interaction enhancement feature values ​​and edge weights of each node.

[0029] In this embodiment, The interaction enhancement feature value of the i-th feature node; Let be the initial feature value of the i-th feature node; j is the index of the adjacent feature node that has a beneficial interaction with the i-th feature node; The interaction edge weights between the i-th feature node and the j-th adjacent node; Let be the initial feature value of the j-th adjacent feature node.

[0030] In this embodiment, the high-order feature interaction graph is a high-order feature interaction graph with 7 nodes and m interaction edges. (V) N For the feature node set, E I (a set of beneficial interaction edges), labeled with the interaction enhancement feature values ​​of each node. With edge weight And a 7-dimensional interactive enhancement feature vector V′=[Vc′,Vt1′,Vt2′,Vt3′,Vs1′,Vs2′,Vs3′] (each dimension value ∈ [0,1.5]).

[0031] A dynamic potential weight matrix is ​​generated based on the high-order feature interaction graph. In this embodiment, obtaining the high-order feature interaction map to generate the dynamic potential weight matrix includes: Obtain the preset collaborative mode potential energy parameter library P (including the basic potential energy thresholds and interaction sensitivity coefficients of the three modes). (Set as needed, for example, pattern-based fault tolerance rating: the higher the fault tolerance, the better.) Smaller values: P&E collaboration (low fault tolerance, high requirement for feature matching): λ3=0.3 - Workflow orchestration (medium fault tolerance): λ2=0.5 - Free configuration (high fault tolerance, flexible adaptation): λ1=0.7, Evolutionary fitness function, based on the experience of education experts and historical data, such as P&E mode λ3=0.3); Obtain the priority coefficient α for the teaching scenario ([0.8, 1.2], α=1.1 for SLO assessment scenario, α=0.9 for basic review scenario). Define the dynamic potential weight matrix W as a 7×3 matrix (rows = 7 feature nodes, columns = 3 cooperative patterns), matrix elements Represents the dynamic potential weight of the i-th feature node for the j-th pattern (j=1=free configuration, j=2=workflow orchestration, j=3=P&E collaboration). Potential energy function calculation: Characteristic potential energy function of the j-th mode: In this embodiment, x is the interaction enhancement feature value. That is, the value of the i-th feature node after interaction enhancement (e.g. It is an interactive enhancement value of complexity features. It is the interactive enhancement value of content depth features). Let J be the characteristic potential function of the j-th mode; Calculate the Euclidean distance: ;in, The optimal feature value for the j-th type of pattern (taken from the potential energy parameter library P) is defined. In this embodiment, similar pattern tasks with "execution success rate ≥ 95% and result quality score ≥ 90 points" are selected from historical task data. The mean of the 7-dimensional feature vectors of these high-quality tasks is calculated, and then fine-tuned by education experts to finally determine the optimal feature value. . Enhance feature values ​​for the current interaction With the optimal eigenvalue of the pattern The one-dimensional Euclidean distance (calculated only for the i-th feature dimension); Let be the value of the i-th dimension in the optimal feature vector of the j-th cooperative mode; Calculate higher-order interactive product terms: Iterate through the adjacent interaction nodes of each feature node, retaining only the beneficial interaction nodes corresponding to the pattern for computation, to obtain... ;in, The higher-order interaction amplification factor is based on the degree of dependence of the pattern on higher-order feature interactions: P&E collaboration: γ3=1.2; Workflow orchestration (medium dependence): γ2=1.0; Free configuration (low dependence): γ1=0.8; Calculate each Then through the normalization factor (satisfy After normalization, the final dynamic potential energy weight matrix W is obtained: ; in, Let be the dynamic potential weight of the i-th feature node for the j-th cooperative mode; This is the priority coefficient for the teaching scenario (e.g., α=1.1 for the SLO assessment scenario). is the weight normalization factor for the j-th type of cooperative mode; Let be the potential energy function value of the j-th pattern for the i-th feature node; Let be the interaction sensitivity coefficient for the j-th type of pattern; This involves performing a product operation on all beneficial interaction nodes of the i-th feature node; k is the index of the feature node that has a beneficial interaction with the i-th feature node; This represents the higher-order interaction amplification factor for the j-th type of pattern; Finally, a 7×3 dimensional dynamic potential energy weight matrix W is obtained (containing each...) Specific values, such as Indicates "content depth" The weight for the P&E model is 0.28.

[0032] Evolutionary robustness adaptation score of collaborative mode is generated based on high-order feature interaction map and dynamic potential weight matrix. Obtain the preset system resource-scene coupling matrix C (3×3 dimension, core definition: This represents the coupling strength between the j-th type of collaborative mode and the k-th type of system core resources. j=1,2,3 correspond to the three types of collaborative modes, and k=1,2,3 correspond to the three types of system core resources: k1=MCP Server communication load, k2=sub-agent (plan / search / code, etc.) utilization rate, k3=tool library (JYD-Tool / API, etc.) call response latency; Acquisition method: based on the tool layer + agent layer + MCP Server architecture, through the statistical analysis of resource consumption data from 10,000+ historical tasks, combined with the calibration by experts in the education field, a fixed matrix is ​​generated and supports dynamic updates). Obtain the scene evolution dynamic dataset E (records the scene change trajectory of 1000+ teaching tasks, including the feature vector change sequence of the initial, execution and final scenes, and stored according to 3 types of patterns). Obtain the priority coefficient α of the teaching scenario; Obtain the evolutionary fitness function for the three modes (P&E Collaboration) Emphasis on content depth and stability, workflow orchestration Emphasis on process stability and flexibility Emphasis on flexibility and stability); Real-time system resource data R (R1=MCP Server current communication load rate, R2=sub-agent average occupancy rate, R3=tool library average call response latency rate, all values ​​in the range [0,1], collected in real time through the system background).

[0033] Computational scene evolution adaptability Based on an LSTM prediction model (trained separately for three modes, with training data consisting of the trajectories of the corresponding modes in the scene evolution dataset E), the input is the current feature vector V, and the output is the probability distribution P(ΔV) of the scene feature vector change. Solve ( (where j is the evolution fitness function of the j-th pattern). Calculate the adversarial robustness score Five virtual disturbance scenarios are generated based on the sensitivity points of three modes using a Generative Adversarial Network (GAN). (For example, the disturbance scenario for the P&E mode is "abrupt change in content depth requirements," and the disturbance scenario for workflow orchestration is "disruption of process fixity.") The adaptation score of the mode under each disturbance scenario is calculated. Take the average ; In this embodiment, the adaptation score for each interference scenario is obtained in the following way: For each virtual interference scenario, substitute the evolutionary robustness adaptation score calculation logic: Based on feature vectors of interference scene Calculate the scene evolution adaptability Evol(j) (the tolerance of scene changes under the interference predicted by the LSTM model). Calculate the resource coupling degree Couple(j) under the interference scenario; Substitute the formula snippet: , This is the basic adaptation score for interference scenarios; ; The robustness adjustment coefficients for interference scenarios are (weak interference = 1.0, medium interference = 0.9, strong interference = 0.8). The score for adaptation to the virtual interference scenario; Computational resource coupling : The initial matrix is ​​shown in Table 1 below:

[0034] Then substitute into the formula to calculate: ; The real-time utilization rate of system resources, ∈[0,1]; R1 represents the coupling strength between the j-th pattern and the k-th system resources; R2 represents the current communication load rate of the MCP Server; R3 represents the average occupancy rate of the sub-agents; and R4 represents the average call response latency rate of the tool library. Final fit score calculation: The evolutionary robustness fit score of the three types of patterns is calculated using the following formula. (∈[0,2.5]): ; in For robustness weighting coefficients, Resource coupling penalty coefficients (P&E mode η3=0.4, μ3=0.4; workflow orchestration η2=0.3, μ2=0.3; free configuration η1=0.2, μ1=0.2): Ultimately, the evolutionary robustness adaptation scores of the three collaborative modes were obtained: S1 (free configuration), S2 (workflow orchestration), and S3 (P&E collaboration). Candidate results for collaborative patterns are generated based on the evolutionary robustness adaptation score of collaborative patterns, high-order feature interaction maps, and dynamic potential weight matrices.

[0035] In this embodiment, generating candidate results for cooperative modes based on the evolutionary robustness adaptation score of the cooperative mode, the high-order feature interaction map, and the dynamic potential weight matrix includes: Constructing a feature-pattern adaptation bipartite graph U consists of 7 feature nodes (with interactively enhanced feature values). V represents three types of collaborative modes (with adaptation scores). And document hard constraint tags, for example, P&E collaboration: s2≥0.7 and s1≤0.5; workflow orchestration: s3≥0.6; free configuration: s3≤0.4), edge E B Weight = Dynamic potential weight The maximum weight matching is solved by the KM algorithm to obtain preliminary candidate patterns. If a candidate pattern does not meet the hard constraints, it is directly eliminated. Build Pattern - Scenario Community Graph GC=(M,S,E) C ): Node M represents three modes, node S represents five historical scenario communities, and edge E C Weight = Historical success rate of the pattern in the community of this scenario (successful cases are selected based on pattern characteristics). By maximizing the modularity of the community, the pattern with the highest fit with the current scenario is found. Dual-graph verification and fusion: If the maximum weight matching result of the bipartite graph is consistent with the pattern with the highest community fit and satisfies the document hard constraints, it is directly used as a candidate pattern; if they are inconsistent, calculate the comprehensive score of matching weight × 0.6 + community fit × 0.4, select the pattern with the highest comprehensive score and satisfying the hard constraints, and finally obtain the collaborative pattern candidate results (including pattern identifier, comprehensive score, dual-graph matching details, and annotation of the document hard constraints that are satisfied).

[0036] The process of generating task requirement information based on structured task data, task feature vectors, and collaborative mode decision results includes: The template rendering algorithm is used to fill the structured task data, feature vectors, and scene adaptation analysis results into a preset template to generate a natural language report. Generate collaborative mode planning and configuration scheme: Based on the selected mode, integrate Agent collaboration rules, tool call thresholds, and configuration parameter details to generate a structured configuration file that can be directly used for subsequent planning; Initialize a dedicated workspace for task planning: Through a permission mapping algorithm, bind access permissions to pre-set pluggable sub-agents (plan, search, code, etc.) and tool libraries (JYD-Tool, API, etc.) to generate a unique identifier for the workspace.

[0037] In this embodiment, the task requirement information includes a task requirement analysis report (including core objectives, constraints, scenario labels, and feature vector score details); a collaborative mode planning and configuration scheme (including selected mode, agent collaboration rules, tool call thresholds, and configuration parameter set); and a dedicated task planning workspace (including workspace ID, list of pluggable sub-agents, and list of tool library access permissions).

[0038] In this embodiment, through multi-agent task planning and path design, a structured task planning graph and an agent collaboration path graph are generated based on task requirement information, including: Structured task data is transformed into formal parameters (such as PDDL format) that can be directly executed by the agent, providing a unified execution standard for task decomposition and allocation, and avoiding the duplication of processing of the original task.

[0039] Specifically, it directly extracts the core information from structured task data without repeatedly parsing natural language: Task goal mapping: Transform task goals into PDDL goal states (e.g., converting the generation of a high school physics project-based learning SLO assessment report into generate_report(report_type="SLO",subject="high school physics")). Constraint mapping: Convert constraints into PDDL prerequisites (e.g., complete 10 days before the midterm exam → time_constraint(end_time="2024-11-10")). Scene adaptation mapping: Based on scene tags and collaboration modes, supplement execution parameters (e.g., P&E collaboration mapping is task_type="plan / execute").

[0040] Formal validation: The rule engine matches the task decomposition rule base to verify the completeness of the PDDL description (if the data source is missing in the structured data, "Teaching Knowledge Base API" is added based on the scenario tag) to ensure that it is unambiguous and executable.

[0041] Collaborative pattern adaptation: Adjusting formal parameters: Free configuration mode: Supplement Agent transfer relationship parameters (e.g., transfer_rule(AgentA→AgentB)); Workflow orchestration mode: Supplement node sequence parameters (e.g., node_sequence(start node → data acquisition agent → analysis agent → end node)); P&E Collaborative Mode: Clearly separate planning parameters (e.g., plan_task="decompose subtasks, allocate resources") and execution parameters (e.g., execute_task="data acquisition, report writing"). Final Output: Execution layer standardized parameter table (including PDDL formal description, collaboration mode adaptation parameters, and constraint details, with no duplicate fields); Formal transformation verification report (with supplementary items and basis, and related structured data).

[0042] The formalized task is recursively broken down into atomic subtasks, with clear dependencies, adapted to a defined collaboration pattern, and free from circular dependencies and redundant subtasks.

[0043] Granularity Decision: Determining Subtask Granularity Based on Task Complexity Score c High complexity (c≥0.7): Decompose into a three-level structure of "parent task → child task → atomic task" (e.g., "generate SLO evaluation report" → "data collection → analysis → writing → export" → "call API → data cleaning → knowledge point matching → report formatting"). Low complexity (c<0.7): Decompose into a two-level structure of "parent task → atomic task" (e.g., "formulate basic review plan" → "organize knowledge points → formulate daily schedule → generate web page version").

[0044] Collaborative model-oriented decomposition: Flexible configuration mode: It is divided according to the logic of "expert collaboration", and each sub-task corresponds to a type of special agent capability (such as "data collection" corresponding to search agent, and "report writing" corresponding to report agent). Workflow orchestration mode: It is divided according to "fixed process" and the sub-tasks are matched with workflow nodes in sequence (such as "start node → data collection sub-task → analysis sub-task → writing sub-task → end node"). P&E Collaboration Model: Strictly separate "planning sub-tasks" (such as "sub-task decomposition, resource allocation, and schedule planning") and "execution sub-tasks" (such as "data collection, content generation, and result verification"), which correspond to Plan Agent and Execute Agent respectively.

[0045] Dependency modeling: Based on the constraint strength vector (such as time constraint strength t1), hard dependencies (sequential execution) and soft dependencies (parallel execution) are identified. Visualize dependencies using a directed acyclic graph (DAG) and verify the absence of circular dependencies through topological sorting (e.g., "data cleaning" can only be performed after "data acquisition" is completed, marked as "subtask A → subtask B").

[0046] Decomposition and verification: Ensure that atomic tasks meet the following requirements: single responsibility (adapting to the capabilities of one agent); quantifiable results (such as "cleaned data table" or "knowledge point list"); and compliance with the hard constraints of the collaborative mode (such as planning sub-tasks in the P&E mode not involving specific execution operations).

[0047] Finally, obtain the task hierarchical decomposition table (including task ID, level, name, type, output requirements, dependencies, and associated collaboration modes); DAG dependency graph (annotating hard / soft dependencies, collaboration mode adaptation nodes, and supporting visualization); and decomposition verification report (including adjustment records and associated feature vector scores).

[0048] Assign atomic subtasks to sub-agents with appropriate capabilities (plan / search / code, etc.) to clarify the boundaries of collaboration.

[0049] In this embodiment, the system has a pre-built intelligent agent role library (dedicated to teaching scenarios, pluggable, containing core capabilities and tool permissions):

[0050] Fit calculation: Based on the dynamic potential weight matrix W, the fit between the subtask and the agent is calculated. Fit S= ×0.6 + Agent's historical success rate for similar tasks / total number of executions ×0.4 (The pattern weights for the feature dimensions corresponding to the sub-tasks). Only agents with a fitness S≥0.6 are retained in the candidate list.

[0051] Collaborative mode-oriented allocation: Flexible configuration mode: Subtasks are assigned based on the user-defined Agent handover relationship and adaptability (e.g., "Data collection Agent → Analysis Agent → Content generation Agent"). Workflow orchestration mode: Subtasks are assigned to the corresponding Agent nodes according to the DAG node order (e.g., “Node 1 → Data Acquisition Agent, Node 2 → Analysis Agent”) to ensure that the process is fixed and controllable. P&E Collaboration Mode: Automatically assigns "planning subtasks" to the Plan Agent and "execution subtasks" to the Execute Agent, and allocates resources according to preset collaboration logic (e.g., the Plan Agent is responsible for prioritizing subtasks, and the Execute Agent is responsible for parallel execution).

[0052] Allocation verification and optimization: Verification rules: No single agent undertakes more than 5 parallel subtasks; a backup agent is configured for the core subtask (fitness S≥0.8); and the permission constraints of the collaborative mode are met (e.g., the Plan Agent has no tool execution permissions). Optimization and adjustment: If the candidate list does not have a compatible Agent, enable pluggable extension.

[0053] Finally, obtain the subtask-agent allocation table (including task ID, agent name, suitability, collaboration mode, substitute agent, and associated with the weight matrix of step 1); agent collaboration boundary specification (clearly defining the task scope, communication objects, and tool usage permissions of each agent); and allocation verification report (including adjustment records and associated collaboration mode constraints).

[0054] Generate a well-defined and executable sequence of subtasks, annotating serial / parallel relationships, timelines, and resource requirements: Based on the topological sorting algorithm, the DAG dependency graph is traversed to output a loop-free execution sequence (such as "data collection → data cleaning → [knowledge point matching | accuracy calculation] → report writing → format export"). Adjust the sequence in conjunction with the collaborative mode: Free configuration mode: Mark "dynamic handover node" (e.g., "After data collection is completed, hand over to the analysis agent, and subsequent agents can be adjusted according to the execution status"). Workflow orchestration mode: Generates a sequence according to a fixed node order, and marks "non-adjustable nodes" (such as "start node → data acquisition → analysis → writing → end node", supporting dialog interface for debugging the entire process); P&E Collaborative Mode: Generates a two-stage sequence of "Plan → Execution" (e.g., "Plan Agent completes sub-task decomposition → Execute Agent performs parallel data collection and analysis → Plan Agent verifies progress → Execute Agent completes writing").

[0055] Timeline planning: Based on the time constraint strength t1 in step 1, assign the estimated start / end time to each subtask: High time constraints (t1≥0.8): Compress the parallel execution interval and reserve 10% buffer time; Low time constraint (t1<0.5): Optimize resource allocation and prioritize the quality of results.

[0056] Resource pre-allocation: Based on the real-time status of system resources, tool permissions are assigned to each subtask (e.g., "pre-allocate search tool quota once for data acquisition subtask"). To avoid resource conflicts: when parallel subtasks compete for the same resource, it is allocated according to task priority (calculated based on feature vectors), and low-priority tasks are delayed in execution.

[0057] Path visualization: Generates Gantt charts (labeling subtask timelines, execution agents, and dependencies) and path flowcharts, supporting manual fine-tuning (only the order of non-core nodes can be adjusted in workflow orchestration mode).

[0058] In this embodiment, the structured task planning graph includes a DAG dependency graph and a Gantt chart. The DAG dependency graph visualizes the subtask hierarchy and dependencies (which is the core structure of task planning), while the Gantt chart annotates the timeline and execution order (which are the implementation details of the plan). In this embodiment, the Agent collaboration path diagram includes a path flowchart, which clearly marks the executing Agent of each subtask and the handover relationship between Agents (dynamic handover in the free configuration mode / fixed nodes in the workflow mode / planning-execution collaboration in the PE mode).

[0059] In this embodiment, generating an executable final task planning scheme based on the structured task planning graph and the Agent collaboration path graph includes: Extracted from the structured task planning diagram: Subtask basic information: ID, name, and deliverable requirements for all subtasks (e.g., "Subtask 001: Data Acquisition, Deliverable = Cleaned Data Table"). Execution constraint information: hard / soft dependencies between subtasks (e.g., "subtask 001 completes → subtask 002 starts" is a hard dependency), and the estimated start / end time of each subtask (from the Gantt chart); Extracted from the Agent collaboration path graph: Agent allocation information: The name of the execution agent corresponding to each subtask (e.g., “Subtask 001 → Data Collection Agent”); Collaboration relationship information: handover nodes between agents (e.g., "After the data collection agent completes the process, it is handed over to the analysis agent"), and collaboration mode adaptation rules (e.g., the collaboration boundary of "Plan Agent → Execute Agent" in P&E mode).

[0060] Finally, obtain the basic information list (subtask ID, name, result, dependency relationship, estimated time) and the Agent association list (subtask ID → execution Agent, collaboration transfer relationship).

[0061] The basic information list and the Agent association list are merged by sub-task ID to form a unified task-Agent association table, ensuring that the Agent, time, and dependency relationship of each sub-task correspond one-to-one. Adaptation and collaboration mode rules: Free configuration mode: marked "The transfer agent can be dynamically adjusted" (e.g., "Subtask 002 can be transferred to the analysis agent or the standby analysis agent according to the execution status"). Workflow orchestration mode: marked "fixed execution order, cannot be adjusted" (e.g., "subtasks 001→002→003, strictly execute in this order"). P&E Collaboration Mode: Mark the "Planning / Execution Boundary" (e.g., "After the Plan Agent completes subtask 000 (task decomposition), all execution-type subtasks are handled by the Execute Agent") to obtain the task-Agent relationship table (including full subtask information + Agent information + collaboration rules).

[0062] Generate an executable final task planning scheme based on a fixed template, as shown in the example below:

[0063] Execution order: Subtasks are executed in order of their IDs. Hard-dependent subtasks must wait for their predecessors to complete, while soft-dependent subtasks can be executed in parallel (e.g., "subtasks 003 and 004 can be executed simultaneously"). Transfer rules: Transfer according to the nodes marked on the Agent collaboration path icon. When transferring, the results of subtasks must be synchronized (e.g., "After the data collection Agent is completed, synchronize the data table to the working directory of the analysis Agent"). Collaborative mode supplement: In the P&E mode, the Plan Agent monitors the execution progress throughout the process without participating in the actual execution.

[0064] The step of generating a high-order feature interaction map based on the task feature vector includes: Each feature vector in the task feature vector is treated as a feature node, and the initial weight of the node is set to the corresponding feature value. For any k-order feature combination, calculate the interaction gain value G based on the knowledge base K, retain only the beneficial interactions with G≥0.1, construct the interaction edges between feature nodes, and the edge weight is the gain value G; The high-order interaction information of the 3rd to 5th order is encoded by an attention aggregation strategy to generate the interaction enhancement feature value of each feature node, thereby generating a high-order feature interaction graph. The high-order feature interaction graph includes multiple feature nodes and multiple interaction edges, and the interaction enhancement feature value and edge weight of each node are labeled.

[0065] This application has the following advantages: This application achieves precise adaptation to teaching scenarios through a high-order feature interaction and dynamic potential weight decision system: This application takes each dimension of the 7-dimensional task feature vector as a node, calculates the interaction gain value G of the k-order feature combination, encodes the 3-5 order high-order interaction information to generate enhanced feature values, and captures the linkage effect of features such as high complexity + high content depth, fixed process + strong result constraint. This application integrates a triple calculation logic of potential energy function, Euclidean distance, and higher-order interactive product terms. The potential energy function amplifies the contribution of high-value features, the Euclidean distance measures the fit between the current feature and the optimal pattern value, and the higher-order interactive product terms enhance the gain of beneficial feature combinations. Finally, the optimal pattern feature values, calibrated by experts in the teaching field, are used... (Average of historical high-quality tasks + expert fine-tuning) and interaction sensitivity coefficient (Dynamically adjusted based on mode tolerance differentiation settings)

[0066] The high-order feature interaction graph of this application breaks through the limitations of traditional single feature decision-making, so that pattern selection no longer considers each feature in isolation, but is based on the synergistic effect between features, thereby improving the accuracy of adaptation from the root. The potential function in this application is designed to address the strong adaptation requirements of high feature values ​​in teaching tasks (such as the sensitivity of P&E mode to high content depth features). It amplifies the weight of high-value features through a quadratic function; Euclidean distance ensures that features fit the optimal standard of the mode and avoids adaptation bias; and higher-order interactive product terms integrate the feature linkage effect into the weight calculation. The combination of these three elements achieves a refined matching of "feature-mode".

[0067] This application first retrieves the knowledge base of the teaching domain (including SLO evaluation criteria and historical teaching cases) through cosine similarity, obtains the Top-5 relevant results and concatenates them with the natural language task text, and then inputs them into a fine-tuned BERT-NER model with custom entity labels (task goal / B-Time / B-Format / B-Scene) to accurately extract teaching-specific entities; This application integrates a 1D complexity score, a 3D constraint strength vector, and a 3D scene adaptation vector. The complexity score is calculated by training a lightweight MLP model using a 3D index of the number of task sub-modules, the number of constraints, and the complexity of scene labels. The scene adaptation vector is generated based on semantic alignment of a teaching scene feature library (3 primary dimensions and 8 secondary sub-dimensions), which perfectly matches the attributes of the teaching task.

[0068] The teaching knowledge base in this application provides domain context support, which solves the problem that general NER is not good at recognizing teaching-specific terms such as formative assessment and SLO standards; text concatenation allows the model to obtain more domain-related information; This application constructs a triple robust system consisting of scenario evolution prediction, adversarial interference verification, and resource coupling degree calculation: Based on the LSTM model, the probability distribution of scene feature vector changes is predicted to generate scene evolution adaptability and predict the impact of scene changes on the pattern in advance. Five virtual interference scenarios are generated by GAN based on the sensitivity points of three types of modes (such as "deep mutation of content" in P&E mode and "process break" in workflow mode), and the adaptation score of each mode is calculated. The resource coupling degree is calculated based on the system resource-scenario coupling matrix C (3×3 dimension, historical data statistics + expert calibration) to avoid the risk of resource conflict. Design a dual-graph fusion decision engine: construct a feature-pattern adaptation bipartite graph (KM algorithm to solve maximum weight matching) and a pattern-scenario community graph (maximize community modularity), and use dual verification to select the optimal pattern.

[0069] The LSTM model is trained based on the evolution trajectory of 1000+ teaching task scenarios, and can accurately predict scenario changes such as "low real-time task → increased content depth requirements", allowing the model decision to adapt to the evolution trend in advance. The targeted interference scenarios generated by GAN cover the core risk points of teaching scenarios. The most resistant mode is selected by the adaptation score to avoid task interruption due to "deep changes in content" or "process break". Resource coupling calculation quantifies the coupling strength between different modes and MCP Server load, Agent utilization, and tool latency, thus avoiding problems such as "multiple agents calling tools simultaneously" and "highly coupled modes consuming too many resources" in advance. The core of this invention is to build an application development framework that supports LLM+RAG and Multi-Agent. The system structure includes three core modules: (1) [Free configuration and matching] Supports zero-code multi-agent collaboration. Users can easily set up transfer relationships to realize an "expert collaboration system".

[0070] (2) [Workflow Orchestration and Handover] Based on the Agent nodes in the workflow, the workflow canvas is used to orchestrate Agents through a fixed process, ensuring the stability and controllability of the task execution process. After creating a Multi-Agent mode application, you only need to adjust the Agent collaboration mode to workflow orchestration to debug the entire workflow in the dialog interface.

[0071] (3) **Plan-and-Execute (P&E) Collaboration:** This mode explicitly decomposes tasks into two independent agents: Plan and Execute. Compared to other collaboration methods, P&E collaboration pre-defines specific agent roles, collaboration logic, and memory management mechanisms, currently offering only a limited number of configuration options for user adjustment. This mode incorporates best practices for multi-agent collaboration. Through meticulous planning and efficient multi-agent collaboration, it can generate results with rigorous structure and rich details. It is particularly suitable for scenarios with low real-time requirements but high requirements for content depth and completeness, such as generating in-depth analysis reports.

[0072] By introducing a multi-agent dynamic collaboration mechanism, the system can dynamically decompose complex and comprehensive learning tasks (such as project-based learning and in-depth research report writing) according to the autonomous planning of the large language model and assign them to sub-agents with different specialties (such as search, encoding, analysis, and report generation agents) for parallel processing. This "expert collaboration" model effectively overcomes the limitations of a single agent in terms of knowledge breadth, professional skills, and parallel processing, enabling the system to flexibly respond to diverse learning needs with high scenario coupling and complex processes, and output more in-depth, complete, and structurally rigorous results.

[0073] This invention innovatively provides three configurable collaborative modes (free transfer, workflow orchestration, and Plan-and-Execute) to meet the needs of different teaching scenarios. Users can achieve dynamic and flexible agent collaboration through "free transfer," or orchestrate fixed processes through a visual "workflow canvas" to ensure the stability and controllability of key learning paths. Furthermore, they can directly adopt the "P&E Collaboration" mode with built-in best practices to obtain high-quality output. This multi-mode design significantly lowers the barrier to building complex AI applications, enabling educators to quickly customize intelligent auxiliary systems that meet specific teaching objectives without coding.

[0074] The system adopts a modular design, treating sub-agents (such as plan, search, and code) and external tools as pluggable components. This architecture allows new functionalities (such as new specialized agents or tools) to be easily integrated into the system in a "plug-and-play" manner without modifying the core framework, greatly enhancing the system's scalability and maintainability. This not only reserves space for future integration of more advanced AI capabilities but also facilitates the construction of an educational tool ecosystem around the core system, continuously evolving to meet the long-term needs of educational informatization development.

[0075] This application also provides a multi-agent dynamic task planning and collaboration system for assisting teaching processes, the multi-agent dynamic task planning and collaboration system for assisting teaching processes includes: A task requirement information acquisition module, wherein the task requirement information acquisition module is used to acquire task requirement information; The graph generation module is used to generate a structured task planning graph and an agent collaboration path graph based on task requirement information through multi-agent task planning and path design. The final task planning scheme generation module is used to generate an executable final task planning scheme based on the structured task planning diagram and the Agent collaboration path diagram.

[0076] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A multi-agent dynamic task planning and collaboration method for assisting the teaching process, characterized in that, The multi-agent dynamic task planning and collaboration method for the auxiliary teaching process includes: Obtain task requirement information; Through multi-agent task planning and path design, a structured task planning diagram and an agent collaboration path diagram are generated based on task requirement information. An executable final task planning scheme is generated based on the structured task planning diagram and the agent collaboration path diagram.

2. The multi-agent dynamic task planning and collaboration method for assisting teaching processes as described in claim 1, characterized in that, The acquisition of task requirement information includes: Obtain natural language task information input by the user; Convert natural language task information into structured task data; Generate task feature vectors based on structured task data; Generate collaborative mode decision results based on task feature vectors; Task requirement information is generated based on structured task data, task feature vectors, and collaborative mode decision results.

3. The multi-agent dynamic task planning and collaboration method for assisting teaching processes as described in claim 2, characterized in that, The process of converting natural language task information into structured task data includes: Obtain a pre-defined knowledge base for the teaching domain; The input natural language task text is preprocessed to obtain the preprocessed task text encoding; The task text encoding is compared with a pre-set knowledge base for the teaching domain to obtain a set of relevant search results; Obtain the trained BERT-NER algorithm; The text to be input is obtained by concatenating the natural language task text with the relevant search result set. The text information to be input is fed into the trained BERT-NER algorithm to obtain structured task data.

4. The multi-agent dynamic task planning and collaboration method for assisting teaching processes as described in claim 3, characterized in that, The generation of task feature vectors from structured task data includes: Obtain a pre-defined evaluation index system for the complexity of teaching tasks; Extract multiple indicator features from structured task data; The characteristics of each indicator are normalized to obtain the normalized indicator characteristics. Obtain a trained lightweight multilayer perceptron model; The normalized index features are input into a trained lightweight multilayer perceptron model to obtain a task complexity score. Obtain a pre-defined library of rules for classifying teaching task constraints; Structured task data is classified into constraint types and scored for constraint strength using a pre-defined teaching task constraint classification rule base, thereby obtaining constraint strength vectors and constraint classification information; A task feature vector is generated based on structured task data, constraint strength vector, constraint classification information, and task complexity score.

5. The multi-agent dynamic task planning and collaboration method for assisting teaching processes as described in claim 4, characterized in that, The step of generating collaborative mode decision results based on task feature vectors includes: Obtain the collaboration mode; Generate a high-order feature interaction map based on the task feature vector and the collaborative mode; A dynamic potential weight matrix is ​​generated based on the high-order feature interaction graph. Evolutionary robustness adaptation score of collaborative mode is generated based on high-order feature interaction map and dynamic potential weight matrix. Candidate results for collaborative patterns are generated based on the evolutionary robustness adaptation score of collaborative patterns, high-order feature interaction maps, and dynamic potential weight matrices.

6. The multi-agent dynamic task planning and collaboration method for assisting teaching processes as described in claim 5, characterized in that, The step of generating a high-order feature interaction map based on the task feature vector includes: Each feature vector in the task feature vector is treated as a feature node, and the initial weight of the node is set to the corresponding feature value. For any k-order feature combination, calculate the interaction gain value G based on the knowledge base K, retain only the beneficial interactions with G≥0.1, construct the interaction edges between feature nodes, and the edge weight is the gain value G; The high-order interaction information of the 3rd to 5th order is encoded by an attention aggregation strategy to generate the interaction enhancement feature value of each feature node, thereby generating a high-order feature interaction graph. The high-order feature interaction graph includes multiple feature nodes and multiple interaction edges, and the interaction enhancement feature value and edge weight of each node are labeled.

7. The multi-agent dynamic task planning and collaboration method for assisting teaching processes as described in claim 6, characterized in that, The step of generating the dynamic potential weight matrix based on the high-order feature interaction graph includes: The initial dynamic potential weight matrix W is obtained from the high-order feature interaction map; The dynamic potential energy weight matrix is ​​generated based on the initial dynamic potential energy weight matrix W.

8. The multi-agent dynamic task planning and collaboration method for assisting teaching processes as described in claim 7, characterized in that, The evolutionary robustness adaptation score of the cooperative mode generated based on the high-order feature interaction map and dynamic potential weight matrix includes: Obtain the trained LSTM prediction model and the trained generative adversarial network; The task feature vector is input into the trained LSTM prediction model to obtain the probability distribution of scene feature vector changes; Virtual interference scenarios are generated using generative adversarial networks; Generate an adaptation score for the virtual interference scenario corresponding to each collaborative mode; Calculate the resource coupling degree for each collaborative mode; The evolutionary robustness adaptation score of each collaborative mode is generated based on the probability distribution of scene feature vector changes, the adaptation score of virtual interference scene, and resource coupling degree.

9. The multi-agent dynamic task planning and collaboration method for assisting teaching processes as described in claim 8, characterized in that, The generation of candidate cooperative modes based on the evolutionary robustness adaptation score of the cooperative mode, the high-order feature interaction map, and the dynamic potential weight matrix includes: Based on the high-order feature interaction graph and collaborative patterns, feature-pattern adaptation bipartite graph and pattern-scene community graph are generated; The feature-pattern adaptation bipartite graph and the pattern-scenario community graph are fused to obtain collaborative patterns that meet preset conditions as candidate results for collaborative patterns.

10. A multi-agent dynamic task planning and collaboration system for assisting the teaching process, characterized in that, The multi-agent dynamic task planning and collaboration system for the auxiliary teaching process includes: A task requirement information acquisition module, wherein the task requirement information acquisition module is used to acquire task requirement information; The graph generation module is used to generate a structured task planning graph and an agent collaboration path graph based on task requirement information through multi-agent task planning and path design. The final task planning scheme generation module is used to generate an executable final task planning scheme based on the structured task planning diagram and the Agent collaboration path diagram.

Images