A teaching and research system and method based on a large language model agent

Abstract

The present application relates to the field of large language model agents, and more particularly to a teaching and research system and method based on a large language model agent, which method comprises: determining at least two teaching and research teacher expert agents participating in teaching and research discussion among a plurality of teacher expert agents according to the similarity calculation of the collected classroom data and the tendency feature vector; determining a teaching and research fusion prompt word according to a personal database, a topic vector and the classroom data; generating a teaching and research discussion text of the teaching and research teacher expert agent through a teaching and research discussion model based on a large language model architecture; and updating the topic vector to enable the at least two teaching and research teacher expert agents to carry out teaching and research discussion through the teaching and research discussion text. The present application realizes the digital transformation of teaching and research work, assists the teaching and research work of teachers, and improves the teaching ability of teachers.

Description

Technical Field

[0001] This invention relates to the field of large language model intelligent agents, and in particular to a teaching and research system and method based on large language model intelligent agents. Background Technology

[0002] Agent technology, as a modeling method in computer simulation, focuses on individual entities called agents to simulate the interaction between individuals and complex environments. Generally, agent simulation comprises three basic components: First, the agent, representing an individual, entity, or element in the modeled system; each agent has its own set of attributes, behaviors, and decision-making processes. Second, the environment, which is the space where agents operate and interact, including virtual space, physical space, and any external factors affecting agent behavior, such as weather conditions, economic changes, and natural disasters; agents may be constrained or influenced by the environment, and their interactions can also affect the environment itself. Third, interaction, where agents interact directly with each other or indirectly with the environment through predefined mechanisms.

[0003] In agent simulation, each agent possesses unique characteristics, including a perception module and a reasoning module, enabling it to make autonomous decisions and take actions based on context and environment. The environment in which the agent exists is used to introduce conditions, stimulate competition, define boundaries, and provide corresponding resources that influence the agent's behavior; the goal of the interaction is to reflect real-world behavior in the simulation based on predefined or adaptive rules.

[0004] Intelligent agents possess four fundamental attributes: autonomy, responsiveness, social competence, and initiative. The social competence attribute of intelligent agents has drawn greater attention to scenarios involving interactions between multiple agents, helping researchers better understand the complexities of interaction and communication within social movements, economic transformations, and cultural trends, in order to explore and design more effective communication strategies and systems.

[0005] In the field of education, agent technology has not yet achieved influential applications. The few existing applications are mostly concentrated in the single-agent domain, such as dialogues with large model clients. For example, the paper "An Empirical Study on Human-Computer Collaboration for Solving Complex Learning Problems Based on Multi-Agent Systems" by Zhai Xuesong, Ji Shuang, Jiao Lizhen, et al., discloses an empirical study from the learner's perspective on multi-agent systems. It finds that in a multi-agent environment, learners can spontaneously employ multi-dimensional questioning strategies to efficiently solve complex learning problems.

[0006] However, there is still no published literature on the application of multi-agent systems in the field of education and research.

[0007] Therefore, how to enable teacher expert agents to conduct teaching and research discussions to assist teachers' teaching and research work is a technical problem that needs to be solved. Summary of the Invention

[0008] Therefore, this invention provides a teaching and research system and method based on a large language model intelligent agent. By performing vector operations on multi-source data, intelligent agent features are constructed. Multiple intelligent agent features are then used in teaching and research discussions through a large language model, realizing the digital transformation of teaching and research work, assisting teachers in their teaching and research work, and improving their teaching abilities.

[0009] To achieve the above objectives, this invention proposes a teaching and research method based on a large language model intelligent agent, comprising:

[0010] Text mining algorithms are used to generate teacher preference feature vectors and / or expert preference feature vectors for multiple teacher-expert agents from a database of expert and teacher profiles.

[0011] Based on the collected classroom data, the similarity between the teacher preference feature vector and / or the expert preference feature vector, at least two teaching and research teacher expert agents participating in the teaching and research discussion are determined from among multiple teacher expert agents.

[0012] Based on the personal data database, the teacher preference feature vector, the expert preference feature vector, the topic vector, and the classroom data, prompt words for integrating teaching and research are determined;

[0013] The teaching and research integration prompts are used to generate the teaching and research discussion text of the teaching and research teacher expert agent through a teaching and research discussion model based on a large language model architecture;

[0014] The topic vector is updated based on the teaching and research discussion text to generate an iterative topic vector. The teaching and research integration prompt words and the teaching and research discussion text are regenerated based on the iterative topic vector, so that at least two teaching and research teacher expert agents can conduct teaching and research discussions through the teaching and research discussion text.

[0015] Furthermore, the classroom data includes classroom sampling data and classroom analysis data, and the process of determining the teaching and research teacher expert agent includes:

[0016] The classroom sampling data is subjected to vector transformation decomposition to generate a set of classroom sampling vectors;

[0017] The teacher preference feature vector and the expert preference feature vector are respectively weighted with the classroom sampling vector set using cosine similarity to generate a discussion guidance vector;

[0018] Based on the discussion guidance vector, text fragments from the personal database, and classroom analysis data, semantic similarity decoding is performed to generate a classroom similarity ranking of multiple teacher expert agents. The teaching and research teacher expert agent is then determined from among the multiple teacher expert agents according to the classroom similarity ranking.

[0019] Furthermore, the process of generating the discussion guide vector includes:

[0020] The teacher preference feature vector and the expert preference feature vector are respectively compared with the classroom sampling vector set through cosine similarity calculation and normalization calculation to generate similarity weights;

[0021] The teacher preference feature vector and the expert preference feature vector are weighted and calculated with the similarity weight to generate the discussion guidance vector.

[0022] Furthermore, the process of generating a ranking of classroom similarity for intelligent agents includes:

[0023] The cosine similarity between the embedding vector of the text segment and the discussion guidance vector is calculated based on classroom analysis data to generate classroom topic similarity parameters;

[0024] The distance metric is calculated between the embedding vector of the text fragment and the discussion guidance vector to generate classroom context similarity parameters;

[0025] Based on the comparison results of the classroom topic similarity parameters and classroom situation similarity parameters of multiple teacher expert agents, a classroom similarity ranking of agents is generated.

[0026] In particular, the fusion of multi-source data ensures that the discussion texts are both rooted in front-line teaching practice and possess theoretical depth. By introducing classroom topic similarity and context similarity parameters, and using intelligent agents to rank classroom similarity, quantitative screening of teaching and research resources is achieved. This data-driven decision-making approach reduces the subjectivity of traditional teaching and research methods that rely on experience-based judgment, making the choices made by teaching and research participants more objective and interpretable.

[0027] Furthermore, the process of generating prompts for integrating teaching and research includes:

[0028] The teacher preference feature vector, the expert preference feature vector, and the topic vector are passed through a convolutional layer to generate an agent stance assignment vector;

[0029] The agent's stance assignment vector and the classroom data are subjected to the data selection constraints to perform stance assignment and similarity calculation, so as to generate a classroom data selection vector;

[0030] The personal database is retrieved based on the agent's position allocation vector and the topic vector to generate teaching and research integration prompts.

[0031] Furthermore, the process of generating the agent's position assignment vector includes:

[0032] The teacher preference feature vector, the expert preference feature vector, and the topic vector are mapped through the convolution operation of the convolutional layer to generate mapped features;

[0033] The mapped features are passed through the softmax function of the convolutional layer to generate an agent position assignment vector.

[0034] Furthermore, the process of generating classroom data selection vectors includes:

[0035] The agent's position assignment vector is used to generate a position vector through the first fully connected layer;

[0036] The stance vector and the classroom data are aligned using a cosine similarity layer to generate an alignment vector;

[0037] The alignment vector and the position vector are weighted and fused for optimization to generate a classroom data selection vector.

[0038] Furthermore, the process of updating the topic vector to generate iterative topic vectors includes:

[0039] The encoded vector of the teaching and research discussion text is mapped to the topic vector to be updated through a second fully connected layer to generate gating weights;

[0040] The topic vectors to be updated are weighted and fused based on the gating weights to generate inherited topic vectors;

[0041] The encoded vector of the teaching and research discussion text and the topic vector to be updated are passed through a multilayer perceptron model to generate an updated vector. The updated vector and the gating weights are then weighted and fused to generate an updated topic vector.

[0042] Based on the inherited topic vector, the updated topic vector, and the historical topic vector, an iterative topic vector is generated.

[0043] Furthermore, the process of generating teaching and research discussion texts includes:

[0044] Based on the comparison between the model parameters of the teaching and research discussion model and the agent's position allocation vector, it is determined whether the state transition function of the teaching and research discussion model is a rebuttal state;

[0045] Based on the consistency between the teaching and research discussion text of the opposing teaching and research teacher expert agent and the classroom data, it is determined whether the state transition function is a concession state.

[0046] In particular, by mapping teacher / expert preference vectors and topic vectors through convolutional layers and combining this with the softmax function to generate agent stance assignment vectors, the quantification and precise division of teaching and research stances are achieved. This mechanism avoids the problem of stance ambiguity in traditional prompt word generation, enabling teaching and research integration prompt words to accurately anchor the agent's role positioning and ensuring a high degree of matching between the discussion text and the agent's attributes.

[0047] The present invention also provides a teaching and research system based on a large language model intelligent agent, wherein the teaching and research system applies the teaching and research method based on a large language model intelligent agent.

[0048] Compared with the prior art, the beneficial effects of the present invention are that it constructs intelligent agent features through vector operations of multi-source data, and conducts teaching and research discussions on multiple intelligent agent features through a large language model, thereby realizing the digital transformation of teaching and research work, assisting teachers in their teaching and research work, and improving their teaching abilities.

[0049] In particular, this invention ensures that the discussion texts are both rooted in front-line teaching practice and possess theoretical depth through multi-source data fusion. It introduces classroom topic similarity parameters and context similarity parameters, and achieves quantitative screening of teaching and research resources through intelligent agent-based classroom similarity ranking. This data-driven decision-making approach reduces the subjectivity of traditional teaching and research methods that rely on experience-based judgment, making the selection of teaching and research participants more objective and interpretable.

[0050] In particular, this invention maps teacher / expert preference vectors and topic vectors through convolutional layers and generates agent stance assignment vectors using the softmax function, achieving the quantification and precise division of teaching and research stances. This mechanism avoids the problem of ambiguous stances in traditional prompt word generation, enabling teaching and research integration prompt words to accurately anchor the agent's role positioning and ensuring a high degree of matching between discussion text and agent attributes. Attached Figure Description

[0051] Figure 1 This is a flowchart illustrating the teaching and research method based on a large language model intelligent agent according to an embodiment of the present invention;

[0052] Figure 2 This is a flowchart illustrating the process of determining the teaching and research teacher expert intelligent agent in the teaching and research method based on a large language model intelligent agent according to an embodiment of the present invention.

[0053] Figure 3 This is a schematic diagram of the process for generating teaching and research integration prompt words in the teaching and research method based on a large language model intelligent agent according to an embodiment of the present invention;

[0054] Figure 4 This is a schematic diagram illustrating the technical framework and effect of the teaching and research method based on a large language model intelligent agent according to an embodiment of the present invention;

[0055] Figure 5This is a schematic diagram illustrating the implementation effect of the teaching and research system based on a large language model intelligent agent according to an embodiment of the present invention. Detailed Implementation

[0056] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0057] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0058] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0059] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0060] like Figures 1 to 5 As shown, this invention provides a teaching and research system and method based on a large language model intelligent agent. By performing vector operations on multi-source data, intelligent agent features are constructed. Multiple intelligent agent features are then used in teaching and research discussions through a large language model, realizing the digital transformation of teaching and research work, assisting teachers in their teaching and research work, and improving their teaching abilities.

[0061] like Figure 1 As shown, this embodiment proposes a teaching and research method based on a large language model intelligent agent, including:

[0062] Text mining algorithms are used to generate teacher preference feature vectors and / or expert preference feature vectors for multiple teacher-expert agents from a database of expert and teacher profiles.

[0063] Based on the collected classroom data, the similarity between the teacher preference feature vector and / or the expert preference feature vector, at least two teaching and research teacher expert agents participating in the teaching and research discussion are determined from among multiple teacher expert agents.

[0064] Based on the personal data database, the teacher preference feature vector, the expert preference feature vector, the topic vector, and the classroom data, prompt words for integrating teaching and research are determined;

[0065] The teaching and research integration prompts are used to generate the teaching and research discussion text of the teaching and research teacher expert agent through a teaching and research discussion model based on a large language model architecture;

[0066] The topic vector is updated based on the teaching and research discussion text to generate an iterative topic vector. The teaching and research integration prompt and the teaching and research discussion text are then regenerated based on the iterative topic vector, so that at least two teaching and research teacher expert agents can conduct teaching and research discussions through the teaching and research discussion text.

[0067] Specifically, the teacher-expert agent is constructed using the open-source model Llama2, and is configured with teacher preference feature vectors or expert preference feature vectors for teaching and research discussions. The teacher preference feature vectors reflect the teaching style characteristics in the teacher's personal database, and the expert preference feature vectors reflect the teaching style characteristics in the expert's personal database. In this embodiment, the teaching and research discussions can be conducted using two or more teacher preference feature vectors or two or more expert preference feature vectors, or multiple teacher preference feature vectors and the aforementioned expert preference feature vectors. The only difference between the teacher preference feature vectors and expert preference feature vectors is the personal data corresponding to their respective agent avatars. For example, the type dimensions of the teacher preference feature vectors and expert preference feature vectors include active components, realistic components, philosophical components, and insightful components.

[0068] Specifically, the topic vector is a set variable in teaching and research discussions.

[0069] Specifically, the personal database includes personal publications and working documents. The database is processed by a BERT encoder to generate encoded features. These encoded features are then processed by a teacher preference decoder to generate teacher preference feature vectors, and by an expert preference decoder to generate expert preference feature vectors. Both the teacher preference decoder and the expert preference decoder employ an architecture consisting of a first linear layer, a GELU activation function, a normalization layer (LayerNorm), and a second linear layer. The difference lies in their fine-tuning using a domain-specific vocabulary from HuggingFace Transformers.

[0070] In this embodiment, as Figure 2 As shown, the classroom data includes classroom sampling data and classroom analysis data. The process of determining the teaching and research teacher expert agent includes:

[0071] The classroom sampling data is subjected to vector transformation decomposition to generate a set of classroom sampling vectors;

[0072] The teacher preference feature vector and the expert preference feature vector are respectively weighted with the classroom sampling vector set using cosine similarity to generate a discussion guidance vector;

[0073] Based on the discussion guidance vector, text fragments from the personal database, and classroom analysis data, semantic similarity decoding is performed to generate a classroom similarity ranking of multiple teacher expert agents. The teaching and research teacher expert agent is then determined from among the multiple teacher expert agents according to the classroom similarity ranking.

[0074] Specifically, the classroom sampling data refers to the classroom transcript, which is analyzed and quantified through a CSMS (Classroom Automation Observation System) to generate a set of classroom sampling vectors, including semantic content vectors, teaching style vectors, and strategy vectors. Preferably, the CSMS classroom automation observation system includes time-slicing processing, a knowledge graph, and a neural network-based word vector mapping. The word vector mapping is a pre-trained model that transforms the word segments mapped to the knowledge graph into vectors to calculate similarity with the tendency feature vectors.

[0075] In this embodiment, the process of generating the discussion guidance vector includes:

[0076] The teacher preference feature vector and the expert preference feature vector are respectively compared with the classroom sampling vector set through cosine similarity calculation and normalization calculation to generate similarity weights;

[0077] The teacher preference feature vector and the expert preference feature vector are weighted and calculated with the similarity weight to generate the discussion guidance vector.

[0078] Specifically, the process of generating similarity weights can be represented as:

[0079] ;

[0080] In the formula, Represents the similarity weight. This represents the k-th component of vector c in the classroom sampling vector set and the teacher preference feature vector / expert preference feature vector. The cosine similarity is calculated, and the denominator is used for normalization.

[0081] More specifically, the cosine similarity in the above formula is calculated as follows: ,in , Let L2 norm (Euclidean norm) be the L2 norm of the vector in the classroom sampling vector set and the s-th component of the teacher preference feature vector / expert preference feature vector, respectively.

[0082] Specifically, the process of generating the discussion guide vector can be represented as:

[0083] ;

[0084] In the formula, Represents the similarity weight. This indicates a discussion of the guiding vector. This represents the k-th teacher preference feature vector / expert preference feature vector.

[0085] In this embodiment, the process of generating the intelligent agent classroom similarity ranking includes:

[0086] The cosine similarity between the embedding vector of the text segment and the discussion guidance vector is calculated based on classroom analysis data to generate classroom topic similarity parameters;

[0087] The distance metric is calculated between the embedding vector of the text fragment and the discussion guidance vector to generate classroom context similarity parameters;

[0088] Based on the comparison results of the classroom topic similarity parameters and classroom situation similarity parameters of multiple teacher expert agents, a classroom similarity ranking of agents is generated.

[0089] Specifically, the process of generating classroom topic similarity parameters can be represented as follows:

[0090] ,

[0091] In the formula, This represents the similarity parameter between classroom topics. , , Let represent the i-th embedding vector of the text segment, the discussion guidance vector, and the classroom analysis data, respectively. This indicates the calculation of cosine similarity. , For two different subject projection vectors This represents the element-wise product (Hadamard product).

[0092] Specifically, the process of generating classroom context similarity parameters can be represented as follows:

[0093] ;

[0094] ;

[0095] In the formula, Represents the embedding vector of a text segment Or the discussion guide vector Perform context-aware mapping. Represents the context projection matrix. This represents vector concatenation. Represents the distance metric of the mapping. Indicates adaptive bandwidth. This represents the similarity parameter of the classroom context.

[0096] Specifically, the similarity parameters of classroom topics and classroom situations are weighted, summed, and normalized to generate a comprehensive score for the intelligent agents. Based on the comprehensive score, multiple teacher expert agents are ranked to generate a ranking of classroom similarity among the intelligent agents.

[0097] In particular, the fusion of multi-source data ensures that the discussion texts are both rooted in front-line teaching practice and possess theoretical depth. By introducing classroom topic similarity and context similarity parameters, and using intelligent agents to rank classroom similarity, quantitative screening of teaching and research resources is achieved. This data-driven decision-making approach reduces the subjectivity of traditional teaching and research methods that rely on experience-based judgment, making the choices made by teaching and research participants more objective and interpretable.

[0098] Furthermore, such as Figure 3 As shown, the process of generating prompts for integrating teaching and research includes:

[0099] The teacher preference feature vector, the expert preference feature vector, and the topic vector are passed through a convolutional layer to generate an agent stance assignment vector;

[0100] The agent's stance assignment vector and the classroom data are subjected to the data selection constraints to perform stance assignment and similarity calculation, so as to generate a classroom data selection vector;

[0101] The personal database is retrieved based on the agent's position allocation vector and the topic vector to generate teaching and research integration prompts.

[0102] In this embodiment, the process of generating teaching and research integration prompts includes:

[0103] The position assignment vector is mapped onto the educational knowledge graph to generate an educational graph mapping path;

[0104] The educational graph mapping paths are sorted using a path sorting algorithm based on the teacher preference feature vector and / or expert preference feature vector to generate teaching and research integration prompt words.

[0105] Specifically, the process of generating teaching and research integration prompts using the path sorting algorithm can be represented as follows:

[0106] ;

[0107] In the formula, This represents the agent's position assignment vector. Represents the mapping matrix. This indicates the optimization of the parameters s of the mapping matrix. Indicate teaching strategies, This indicates that scoring is based on educational knowledge graph paths and teacher preference feature vectors and / or expert preference feature vectors.

[0108] Represents the overall path of the education graph. Represents the path mapping matrix. This indicates that a knowledge graph path has been selected. Within the selected knowledge graph path, suggest keywords related to the integration of teaching and research.

[0109] Furthermore, the process of generating the agent's position assignment vector includes:

[0110] The teacher preference feature vector, the expert preference feature vector, and the topic vector are mapped through the convolution operation of the convolutional layer to generate mapped features;

[0111] The mapped features are passed through the softmax function of the convolutional layer to generate an agent position assignment vector.

[0112] Specifically, the process of generating the agent's position assignment vector can be represented as:

[0113] ;

[0114] In the formula, The vector represents the agent's position assignment vector, and softmax represents the softmax function. Denotes a convolutional layer, where , For the learnable position weight matrix and learnable position bias term of the convolution operation of the convolutional layer, , These represent the teacher preference feature vector, the expert preference feature vector, and the topic vector, respectively.

[0115] Furthermore, the process of generating classroom data selection vectors includes:

[0116] The agent's position assignment vector is used to generate a position vector through the first fully connected layer;

[0117] The stance vector and the classroom data are aligned using a cosine similarity layer to generate an alignment vector;

[0118] The alignment vector and the position vector are weighted and fused for optimization to generate a classroom data selection vector.

[0119] Specifically, the process of generating the position vector can be represented as:

[0120] ;

[0121] In the formula, Represents the position vector. express Activation function , Let represent the weight matrix and bias vector, respectively, and p represent the agent's position assignment vector.

[0122] Specifically, the process of generating alignment vectors using cosine similarity alignment can be represented as:

[0123] ;

[0124] In the formula, This represents the i-th alignment vector. This indicates the calculation of cosine similarity. Represents the position vector. This represents the i-th vector in the set of classroom sampling vectors for classroom data. This represents the L2 norm (Euclidean norm). This indicates the minimum value to prevent division by zero.

[0125] Specifically, the process of generating classroom data selection vectors can be represented as follows:

[0126] ;

[0127] In the formula, This indicates that the classroom data selection vector is used. This represents the optimization function, which is obtained through... Please solve. Indicates the weighted weight. This represents the i-th alignment vector. Represents the position vector. This represents the correction factor. This represents the information content index.

[0128] More specifically, the calculation process of the information content index can be expressed as follows:

[0129] ;

[0130] In the formula, Indicates the information content index. Elements representing a subset of classroom data, Indicates a given teaching and research topic, This represents the conditional probability calculated based on topic similarity. This represents the prior probability.

[0131] Furthermore, the process of updating the topic vector to generate iterative topic vectors includes:

[0132] The encoded vector of the teaching and research discussion text is mapped to the topic vector to be updated through a second fully connected layer to generate gating weights;

[0133] The topic vectors to be updated are weighted and fused based on the gating weights to generate inherited topic vectors;

[0134] The encoded vector of the teaching and research discussion text and the topic vector to be updated are passed through a multilayer perceptron model to generate an updated vector. The updated vector and the gating weights are then weighted and fused to generate an updated topic vector.

[0135] Based on the inherited topic vector, the updated topic vector, and the historical topic vector, an iterative topic vector is generated.

[0136] Specifically, the process of generating iterative topic vectors can be represented as:

[0137] ;

[0138] In the formula, This represents the gate weight in the t-th iteration. This represents the sigmoid activation function. , These represent the learnable weight matrix and bias term of the second fully connected layer, respectively. The encoding vector representing the teaching and research discussion text in the t-th iteration. And topic vectors to be updated , This represents element-wise multiplication. This represents the inherited topic vector used to update the topic vector in the (t+1)th iteration. Let represent the updated topic vector used in the (t+1)th iteration to update the topic vector. This represents a multilayer perceptron model. LayerNorm represents the iterative topic vector, and LayerNorm represents layer normalization. This indicates the historical retention weight, preferably 0.6. Let represent the historical topic vector before the t-th iteration. Preferably, the topic vector is determined by weighting the relevance of the historical data to the current topic through an attention mechanism.

[0139] Specifically, the difference between the agent's position assignment vector and the teaching and research discussion text is quantified using KL divergence (Kullback-Leibler Divergence) or JS divergence (Jensen-Shannon Divergence) to construct the reward function.

[0140] Specifically, the process of adjusting the model parameters of the teaching and research discussion model based on the large language model architecture can be represented as follows:

[0141] ;

[0142] In the formula, Represents the optimal model parameters. Indicates the search for the objective function Maximize model parameters , This represents the expectation operator, which calculates the probability-weighted average of the generated output teaching and research discussion text V. This represents the reward function, which quantifies the quality of the teaching and research discussion text V. The output text V, representing the teaching and research discussion, is derived from the model parameters. The results of sampling in the teaching and research discussion model LLM of the large language model architecture.

[0143] Specifically, the process of adjusting the topic vector can be represented as follows: in, This represents the adjusted iterative topic vector. Represents the state transition function. These represent the topic vector and the teaching and research discussion text of the opposing teaching and research teacher expert agent, respectively.

[0144] Furthermore, the process of generating teaching and research discussion texts includes:

[0145] Based on the comparison between the model parameters of the teaching and research discussion model and the agent's position allocation vector, it is determined whether the state transition function of the teaching and research discussion model is a rebuttal state;

[0146] Based on the consistency between the teaching and research discussion text of the opposing teaching and research teacher expert agent and the classroom data, it is determined whether the state transition function is a concession state.

[0147] In particular, by mapping teacher / expert preference vectors and topic vectors through convolutional layers and combining this with the softmax function to generate agent stance assignment vectors, the quantification and precise division of teaching and research stances are achieved. This mechanism avoids the problem of stance ambiguity in traditional prompt word generation, enabling teaching and research integration prompt words to accurately anchor the agent's role positioning and ensuring a high degree of matching between the discussion text and the agent's attributes.

[0148] Specifically, the state transition function can be expressed as:

[0149] ;

[0150] In the formula, Represents the state transition function. These represent the topic vector and the research discussion text of the opposing research teachers / experts, respectively. Indicates model parameters, This indicates an adjustment of the weight, preferably 0.7. This refers to the teaching and research discussion text generated by the opposing teaching and research teacher expert agents through the judge's large language model. The support probability score derived from classroom data C The similarity threshold is preferably 0.2. This indicates the teaching and research discussion texts for the current cycle and the teaching and research discussion texts from the cycle two weeks prior. This indicates the threshold for triggering episode termination, preferably 0.2. Indicates a state of rebuttal. This indicates a concession.

[0151] Specifically, the rebuttal status is ,in Represents the optimization function. Indicates the number of times data is referenced. This represents the weight matrix of the data type.

[0152] Specifically, the concession state is ,in Represents the optimization function. , These represent the agent's position assignment vectors for the current cycle and the previous cycle, respectively.

[0153] In this embodiment, the process of generating teacher preference feature vectors and / or expert preference feature vectors includes:

[0154] The personal database is passed through the BERT layer of a text mining model to generate an input representation;

[0155] The input representation is passed through the encoding layer of a text mining model to generate initial encoded text;

[0156] The initial encoded text is passed through the subject-controlled multilayer perceptron layer of the text mining model to generate subject knowledge encoded text;

[0157] The initial encoded text is passed through multiple mapping layers of a text mining model to generate initial encoded text gating weights;

[0158] The subject knowledge encoded text and the initial encoded text are weighted and fused based on the gating weights of the initial encoded text to generate the encoded teacher preference feature vector and / or expert preference feature vector.

[0159] Specifically, the process of generating teacher preference feature vectors and / or expert preference feature vectors can be represented as follows:

[0160] ;

[0161] In the formula, This represents the 0th and 1st inputs in the personal database. express layer, This represents the original input text. This represents the initial encoded text. Let k represent the l-th Transformer module built based on a feedforward neural network (FFN), and k represent the subject knowledge encoded text. LN represents the multi-layer sensory mechanism for subject regulation, and LN represents the normalization layer. This indicates element-wise multiplication. This represents the disciplinary regulation characteristics, where g represents the initial gating weight of the encoded text. express function, express function, Representing four different mapping matrices, This represents the teacher preference feature vector and / or expert preference feature vector.

[0162] It is understood that in this embodiment, a personal database is established by experts or teachers, which includes personal works and working documents. For these personal works and working documents, text mining and other algorithms are used to generate educational tendency feature vectors for the teacher or expert's intelligent agent. These vectors are used to define the intelligent agent's characteristics and identity, and are applied to each speech of the intelligent agent. After data collection and analysis for a lesson, the system automatically selects several teacher-expert intelligent agents according to certain rules, forming a virtual teaching and research discussion area with the teacher-expert intelligent agent of the instructor for that lesson. Within the virtual discussion area, the speaking order is arranged according to the physical characteristics and educational tendency feature vectors of the teacher-expert intelligent agents. The teacher and expert intelligent agents speak in turn, discussing the situation of the lesson. For each expert or teacher's intelligent agent's speech, classroom recording data, fragments related to the topic vector in the personal database, and the speech text of previous intelligent agents in the discussion are selected as context, fused with prompt words, input into a large language model, and the speech text of each teacher or expert is output. After all speeches are finished, voiceprint cloning is performed based on the characteristics of each teacher-expert intelligent agent, driving the corresponding digital human image of the teacher-expert intelligent agent, and the entire teaching and research discussion text is sequentially formed into multimedia output.

[0163] Therefore, this embodiment generates the thinking tendency characteristics of teachers or experts through text mining: thinking characteristics include mindset, knowledge-action thinking, complexity preference, etc., and trains a proprietary neural network model. It selects segments related to the topic vector from classroom transcript data, classroom analysis data, and expert or teacher databases: Classroom transcript data, classroom analysis data, and expert or teacher databases are pre-segmented into segments, and the feature vector of each segment, as well as the feature vector of the topic vector, are calculated. Appropriate segments are selected using cosine similarity.

[0164] This embodiment also provides a teaching and research system based on a large language model intelligent agent, which applies the teaching and research method based on a large language model intelligent agent.

[0165] Understandably, the scarce resources that primary and secondary school teachers most desire when conducting teaching and research activities are twofold: guidance from senior experts and in-depth discussions with outstanding peers. The teacher expert intelligent agents in this embodiment can be used to create a virtual teaching and research environment, simulating the interaction and intellectual exchange between teachers and professional groups, providing valuable reference and guidance for real-world teaching and research exchanges. Imagine that each teacher or expert has a corresponding teacher intelligent agent avatar. For example, teacher A corresponds to the digital avatar teacher expert intelligent agent A, and experts B and C correspond to their respective digital avatar teacher expert intelligent agents B and C. These different teacher expert intelligent agents A, B, and C interact around a set topic (i.e., the environment, such as classroom recordings, big data analysis reports on classroom teaching characteristics, etc.). This agent simulation process essentially encompasses all the key elements of teaching and research activities.

[0166] Currently, intelligent agents based on large language models can exhibit social attributes such as interactive communication, emotional understanding, and role-playing in effectively simulating human social interactions. These social attributes enable teachers to receive more open, diverse, and targeted opinions and suggestions in a timely manner, practice and improve their teaching skills independently in a safe environment, and gain more respect and understanding, which precisely meets the supply of the most scarce resources for primary and secondary school teaching and research activities.

[0167] The architecture of the teaching and research system based on a large language model intelligent agent in this embodiment is as follows: Figure 4 and 5As shown, the unique attributes of the intelligent agents participating in teaching and research have a significant impact on dialogue. These unique attributes can include subject, grade, professional identity, thinking tendency, and voice characteristics. Identity information such as subject, grade, and professional identity can be obtained by collecting teacher information or classroom types through the system, while voice characteristics are obtained through voiceprint analysis and language usage features. The most challenging aspect is analyzing the teacher's thinking tendency characteristics; the difficulty lies in representing these characteristics in a way that is valuable in an educational sense. The knowledge-action model is a method to solve this problem. This embodiment combines the knowledge-action model to categorize thinking characteristics into proactive, realistic, philosophical, and perceptual types, and provides a quantitative and calculable path for analyzing these thinking characteristics.

[0168] Multi-agent technology based on large language models is a rapidly developing field, with many technical frameworks already emerging. Based on the aforementioned needs for teaching and research exchanges, multi-agent technology based on large language models particularly needs to focus on the following aspects: First, it requires establishing a unified dialogue framework in a multi-agent text environment with strong dialogue management capabilities. This is because teaching and research primarily focuses on dialogue and is sensitive to the orderly nature of communication. Second, agents need to possess dialogue capabilities; any agent within the environment should be able to send and receive messages from other agents to initiate or continue dialogue. This is determined by the freedom inherent in teaching and research activities.

[0169] As an analysis of the specific implementation results of this embodiment, such as Figure 5 As shown, the experimental process of the teacher's intelligent agent avatar conducting teaching research dialogue is as follows: After the teaching research subject uploads a lesson, the system first performs big data analysis of classroom teaching, and then automatically organizes a teaching research exchange involving multiple intelligent agents. The feedback from the teaching research exchange is automatically generated and output in the form of text and the corresponding agent's voice imitation. After a large number of experiments, the system has obtained relatively stable and similar results.

[0170] like Figure 5 As shown, this embodiment is used to conduct an experiment. The experimental process includes: Teacher Tang uploads the classroom data of the lesson "Inequalities" to the system. The system first completes the classroom data analysis. After the analysis is completed, the intelligent agent of the digital avatar is triggered to conduct teaching and research. The system takes Teacher Tang as the main body and automatically invites the digital avatars of three experts, namely Dr. Li, an education theory expert, Dr. Liu, a technology analysis expert, and Dr. Xu, a subject teaching and research expert, to conduct teaching and research discussions. The teaching and research discussion text of this embodiment is generated. Based on the teaching and research discussion text, the final output audio and text of teaching and research exchange feedback are generated.

[0171] like Figure 5 As shown, the output of the teaching and research discussion text is excerpted and simplified as follows:

[0172] Teacher Tang (as a teacher's AI): Hello everyone, I am Teacher Tang. It is a great honor to discuss my experiences and insights in teaching junior high school mathematics with you all. In this lesson, I mainly covered the following: 1. Definition of Inequalities: Introduced the definition of an inequality, which is an expression connecting sizes using inequality signs. 2. How to Set Up Inequalities: Explained how to set up inequalities based on keywords in a context, including explicit and implicit inequalities. 3. Solutions to Inequalities: Explained that the solution to an inequality is the value of the unknown that makes the inequality true, and distinguished between the concepts of the solution and the solution set. 4. Representation of the Solution Set of Inequalities: Introduced the use of a number line to represent the solution set of inequalities, including steps such as drawing the number line, finding key points, plotting points, and drawing directions. 5. Summary: Finally, summarized the entire lesson, emphasizing the definition of inequalities, how to set up inequalities, the concepts of solutions and solution sets, and the methods for representing the solution set. I hope to receive valuable feedback from everyone, especially on how to better cater to the different learning styles of students so that my classes can inspire the potential of every student.

[0173] Dr. Li (as the first expert agent): Hello, Teacher Tang. Based on your analysis report, you demonstrated a philosophical teaching style in your classes, emphasizing abstract concepts and logical reasoning, which helps students systematically master knowledge. Teacher Tang introduced new knowledge through real-life examples, allowing students to better understand abstract mathematical concepts, and emphasized knowledge transfer and application, which helps cultivate students' knowledge transfer and innovation abilities. In conclusion, Teacher Tang demonstrated a high level of teaching skill and artistry, which is worthy of learning and emulation.

[0174] Teacher Tang (as a teacher agent): Thank you, Dr. Li, for your evaluation and summary. I believe that sharing these ideas can indeed help students master abstract concepts while enhancing their practical skills and creative thinking. Next, I would like to ask the experts for their valuable opinions and suggestions regarding any shortcomings or areas for improvement in my teaching.

[0175] Dr. Liu (as a second expert agent): Hello, Professor Tang. Based on your teaching analysis report, I have summarized the following potential shortcomings: Uneven application of the four "whats": The report shows that teachers focus too much on "what is" and "why," and use "how" and "if" less frequently, which may limit students' practical abilities and innovative thinking. Insufficient student speaking time: The total speaking time for students was only 4 minutes and 13 seconds, with an average of only 4.77 seconds per session, which may limit students' proactive thinking and expression abilities. The above is my summary of potential shortcomings in Professor Tang's teaching, and I hope it will be helpful to you.

[0176] Teacher Tang (as a teacher agent): Thank you, Dr. Liu, for your insightful summary. I agree with the importance of classroom interaction. I will try to create more opportunities for student participation in the classroom, and I will guide students to think more about "how" and "what if," combining theoretical knowledge with practical application. Dr. Xu, what suggestions do you have for my teaching?

[0177] Dr. Xu (as a third expert agent): Hello, Teacher Tang. Regarding question types, teachers can increase open-ended questions to further stimulate students' thinking. From the perspective of classroom interaction, teachers can further enhance students' active role by increasing student discussions and collaborative learning. The above is intended to help teachers further improve their teaching skills.

[0178] Teacher Tang (as a teacher agent): Thank you all for your participation. That concludes our teaching and research activity for today.

[0179] The above experimental results demonstrate the rational thinking and practical value of digitalized teaching and research. It has achieved considerable similarity to actual teaching and research exchange activities. The authenticity and logic of the relevant topics have achieved the expected goals. Because the voice accurately imitates the participants, it generates a good sense of immersion, proving that the design of digitalized teaching and research is scientific and feasible.

[0180] Understandably, the application of intelligent digital avatar technology in teaching and research scenarios is an emerging and promising field. Beyond the application scenarios demonstrated in the experiment, there are many more potential applications.

[0181] First, digital teaching research can serve as a personalized advisor for teachers. The intelligent digital teaching research system can match teachers with senior expert AI agents and outstanding peer AI agents that match their different thinking styles, based on their individual characteristics, subject background, and classroom teaching goals. This provides teachers with classroom observation and feedback suggestions from multiple angles and levels, which not only helps teachers more accurately grasp the overall situation of classroom teaching and learning but also assists them in rapidly improving their digital teaching research capabilities.

[0182] Secondly, digitalized teaching research can innovate the content and form of on-site teaching research. Intelligent digitalized teaching research can be directly integrated into on-site teaching research, using senior expert intelligent agents and outstanding peer intelligent agents to exchange ideas and spark discussions with participating experts, teaching researchers, and subject teachers, thus promoting the breadth and depth of teaching research discussions.

[0183] Third, digital avatars can enrich the content and forms of remote teaching research interactions. In remote teaching research, digital avatars can overcome geographical limitations, providing teachers in different regions with diverse and real-time teaching research support and resource sharing, bridging communication barriers caused by distance, and are of great value in addressing the weakness of teaching research in underdeveloped areas. Furthermore, more application scenarios can be further explored in subsequent teaching research activities.

[0184] In this embodiment, agent features are constructed through vector operations on multi-source data. These features are then used in teaching and research discussions via a large language model, enabling the digital transformation of teaching and research work, assisting teachers in their research efforts, and improving their teaching abilities. Multi-source data fusion ensures that the discussion text is both rooted in frontline teaching practice and possesses theoretical depth. Classroom topic similarity and context similarity parameters are introduced, and teaching and research resources are quantitatively screened through agent classroom similarity ranking. This data-driven decision-making approach reduces the subjectivity of traditional teaching and research methods that rely on experience-based judgment, making the selection of participants more objective and interpretable. By mapping teacher / expert preference vectors and topic vectors through convolutional layers and combining this with a softmax function to generate agent stance assignment vectors, the quantification and precise division of teaching and research stances are achieved. This mechanism avoids the ambiguity of stances in traditional prompt word generation, enabling teaching and research fusion prompt words to accurately anchor the role positioning of agents and ensure a high degree of matching between discussion texts and agent attributes.

[0185] Those skilled in the art will recognize that the modules and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0186] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0187] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A teaching and research method based on a large language model agent, characterized in that, The method comprises the following steps: a personal database of experts and / or teachers is processed by a text mining algorithm to generate a teacher tendency feature vector and / or an expert tendency feature vector of a plurality of teacher expert agents; at least two teacher expert agents participating in a teaching research discussion are determined from the plurality of teacher expert agents according to a similarity calculation of the collected classroom data, the teacher tendency feature vector and / or the expert tendency feature vector; a teaching research fusion prompt word is determined according to the personal database, the teacher tendency feature vector and / or the expert tendency feature vector, a topic vector or an iterative topic vector, and the classroom data; the teaching research fusion prompt word is processed by a teaching research discussion model based on a large language model architecture to generate a teaching research discussion text of the teaching research teacher expert agent; the topic vector is updated based on the teaching research discussion text to generate an iterative topic vector, and the teaching research fusion prompt word and the teaching research discussion text are cyclically and iteratively generated based on the iterative topic vector, so that at least two teaching research teacher expert agents perform a teaching research discussion through the teaching research discussion text; the classroom data comprises classroom sampling data and classroom analysis data, and the process of determining the teaching research teacher expert agent comprises: the classroom sampling data is decomposed by vector conversion to generate a classroom sampling vector set; The teacher tendency feature vector and / or the expert tendency feature vector is calculated with cosine similarity and normalized calculation with the classroom sampling vector set to generate similarity weight, wherein the cosine similarity calculation is: , wherein cosine similarity, respectively represent the s-th component of the vector in the classroom sampling vector set, the teacher tendency feature vector / expert tendency feature vector, respectively represent the corresponding L2 norm thereof; the teacher tendency feature vector and / or the expert tendency feature vector are weighted and calculated with the similarity weight to generate a discussion guide vector; the process of generating a classroom theme similarity parameter comprises: ； wherein, denotes a classroom theme similarity parameter, denotes the i-th embedding vector of a text snippet of the personal library, the discussion guide vector, the classroom analytics data, respectively, and cos denotes a cosine similarity computation, is a theme projection vector, denotes an element-wise multiplication; the process of generating a classroom context similarity parameter comprises: ； wherein, represents an embedding vector for a text snippet of a personal profile or the discussion guide vector performing context-aware mapping, represents a context projection matrix, represents vector concatenation, represents a distance metric of the mapping, represents an adaptive bandwidth, represents a classroom context similarity parameter; an agent classroom similarity ranking is generated according to a comparison result of the classroom theme similarity parameter and the classroom context similarity parameter of the plurality of teacher expert agents, and the teaching research teacher expert agent is determined from the plurality of teacher expert agents according to the agent classroom similarity ranking.

2. The large language model based agent teaching and research method according to claim 1, characterized in that, The process of generating a teaching research fusion prompt word comprises: the teacher tendency feature vector and / or the expert tendency feature vector and the topic vector are processed by a convolution layer to generate an agent position assignment vector; the agent position assignment vector and the classroom data are subjected to position assignment and similarity calculation to generate a classroom data selection vector; the personal database is retrieved based on the agent position assignment vector, the topic vector or the iterative topic vector to generate a teaching research fusion prompt word.

3. The teaching and research method of a large language model-based intelligent agent according to claim 2, characterized in that, The process of generating an agent position assignment vector comprises: the teacher tendency feature vector and / or the expert tendency feature vector and the topic vector or the iterative topic vector are mapped by a convolution operation of a convolution layer to generate a mapping feature; the mapping feature is processed by a softmax function of the convolution layer to generate an agent position assignment vector.

4. The teaching and research method of a large language model-based intelligent agent according to claim 2, characterized in that, The process of generating a classroom data selection vector comprises: the agent position assignment vector is processed by a first full connection layer to generate a position vector; the position vector and the classroom data are aligned by a cosine similarity layer to generate an alignment vector; the alignment vector and the position vector are weighted and fused to optimize to generate a classroom data selection vector.

5. The large language model based agent teaching and research method according to claim 2, characterized in that, The process of updating the topic vector to generate an iteration topic vector includes: mapping the encoding vector of the teaching research paper and the topic vector to be updated through a second fully connected layer to generate a gating weight; weighting and fusing the topic vector to be updated based on the gating weight to generate an inherited topic vector; mapping the encoding vector of the teaching research paper and the topic vector to be updated through a multi-layer perception model to generate an update vector, and weighting and fusing the update vector and the gating weight to generate an updated topic vector; generating an iteration topic vector based on the inherited topic vector, the updated topic vector and a historical topic vector.

6. The large language model based agent teaching and research method according to any one of claims 1 to 5, characterized in that, The process of generating a teaching research paper includes: determining whether the state transition function of the teaching research discussion model is a rebuttal state based on a comparison of the model parameters of the teaching research discussion model and the agent stance assignment vector; determining whether the state transition function is a concession state based on the consistency of the teaching research paper of the opponent teaching research teacher expert agent and the classroom data.

7. A teaching and research system based on a large language model agent, characterized in that, The teaching research system applies the teaching research method based on the large language model agent according to any one of claims 1 to 6.

Images