Education experiment synthesis student intelligent agent construction method and system and electronic equipment

Abstract

The application discloses a kind of education experiment synthesis student intelligent agent construction method, system and electronic equipment, it is related to artificial intelligence education simulation field.The method includes: initialization configuration covers the multi-dimensional feature parameter set of cognitive ability, personality characteristics and knowledge state, and builds simulation environment including private reasoning channel;Double closed loop mechanism is introduced in simulation interaction: use in for consistency check closed loop, based on multi-dimensional feature matching or reward model, real-time correction "role deviation" behavior;With state dynamic evolution closed loop, combine positive cognitive gain, forgetting curve and "inherent characteristics-transient state" separation model, real-time update agent knowledge and psychological state;Synchronous generation includes the structured data set of behavior traceability label.The application solves the problem of behavior homogeneity and unexplainable in education simulation, and can generate high reliable and attribution ability scientific research data.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of educational artificial intelligence (AIED) and digital education simulation, specifically involving a method, system, and electronic device for constructing a synthetic student intelligent agent for educational experiments. Background Technology

[0002] With the rapid development and deepening application of Large Language Model (LLM) in the field of education, agent-based virtual classroom simulation, teaching dialogue interaction, learning process analysis and teaching strategy evaluation have become one of the research hotspots in the intersection of educational technology and artificial intelligence. Related research shows that the LLM-driven multi-agent framework can simulate the classroom role interaction process to a certain extent and support experimental analysis of teaching interaction[1]. At the same time, review literature on "opportunities and challenges of generative AI / LLM in education" also points out that LLM can be used for content generation, learning interaction and personalized support, but there are still limitations in terms of reliability and robustness[2].

[0003] In the aforementioned research and application scenarios, educational researchers often need a large amount of controllable, repeatable, and scalable student interaction data to support various types of educational experiments and algorithm research.

[0004] Existing LLM educational simulation systems typically adopt an interaction paradigm of “teacher agent-student agent”: the LLM drives the teacher to generate instructions for explanation, questioning or feedback; the LLM drives the student to generate answers, questions or responses based on prompts, and forms an interaction log in multiple rounds of dialogue. Some studies also use “generative student agents” for learning process simulation and data generation [3]. The output of these systems is often mainly dialogue text or interaction records, which are used for subsequent teaching analysis, model training or strategy evaluation.

[0005] The aforementioned existing technologies can generate a certain scale of teaching interaction data at a low cost through automated dialogue generation without the need to organize real classroom participation, and facilitate rapid prototype verification and experimental reproduction, thus possessing certain research and engineering value [1].

[0006] However, existing educational simulation systems based on large language models still have some objective shortcomings in practical applications, which may limit their effectiveness and credibility in educational research and experiments. These shortcomings are mainly reflected in the following aspects:

[0007] The problem of homogenization of student behavior: Multiple student agents often rely on the same or highly similar prompt words to constrain them, resulting in similar answering styles, expression methods and error patterns, which makes it difficult to reflect the differences among students in the real classroom.

[0008] Insufficient interpretability and controllability of character modeling: Some systems only use coarse-grained labels or implicit prompts to control capabilities, without explicitly modeling dimensions such as cognitive ability, personality traits and knowledge status, making the generation basis opaque and difficult to support refined experimental design and result analysis;

[0009] Consistency in multi-round interactions is difficult to maintain: as the number of rounds increases, phenomena such as sudden changes in knowledge level, abnormal reasoning depth, or language style drift may occur, thereby affecting the stability and realism of simulation data;

[0010] Insufficient structured data and metadata in generated data: Most systems focus on preserving the original dialogue text, lacking structured records and annotations of generation motivations, internal state changes, or policy triggering processes, which affects their reusability in model training, behavior analysis, and scientific research replication.

[0011] Therefore, for virtual education experiment scenarios, how to obtain high-quality simulation interaction data that better meets the needs of education experiments and can be used for training and evaluation (including requirements for diversity, stability and reusability) remains one of the issues that the field continues to focus on [2]. Summary of the Invention

[0012] The purpose of this invention is to provide a method, system, and electronic device for constructing synthetic student agents for educational experiments, in order to solve problems such as severe homogenization of student agent behavior, insufficient interpretability of character design, and difficulty in maintaining consistency in multi-round interactions in existing educational simulation systems.

[0013] The objective of this invention is achieved through the following technical solution:

[0014] A method for constructing a synthetic student intelligent agent for educational experiments, the method comprising the following steps:

[0015] Step S1: Construct a parameterized model of the synthetic student agent. According to the preset experimental requirements, initialize and configure the multidimensional feature parameter set of each synthetic student agent. The multidimensional feature parameter set includes at least cognitive ability parameters, personality feature parameters, and knowledge state parameters.

[0016] Step S2: Build a simulation interaction environment based on a large language model, instantiate at least one teacher agent and multiple synthetic student agents configured in step S1 in the environment, and establish a communication topology network between the agents.

[0017] Step S3: Execute the teaching simulation interaction process. The teacher agent generates teaching instructions based on the teaching objectives. The synthetic student agent receives the teaching instructions and the environmental context, and performs reasoning based on the multi-dimensional feature parameter set to generate candidate mimicry response behaviors.

[0018] Step S4: Perform behavior consistency verification. Before the synthetic student agent generates candidate mimicry response behaviors and outputs them externally, the behavior consistency evaluation module calculates the matching degree of the candidate mimicry response behaviors based on the multidimensional feature parameter set of the synthetic student agent at the current moment. If the matching degree is lower than the preset threshold, a correction mechanism is triggered.

[0019] Step S5: Generate and output a structured educational dataset, and drive the dynamic evolution closed loop of the synthesized student agent parameters based on data feedback.

[0020] Specifically, this includes: recording dialogue logs, behavioral decision sequences, and state change data in real time during simulation interaction, and constructing a structured educational experiment dataset based on this data; extracting feedback features based on the educational experiment dataset, and feeding these feedback features back to step S1 to dynamically update the multi-dimensional feature parameter set of the corresponding synthetic student agent, thereby forming a simulation data-driven model parameter evolution closed loop; generating a structured educational dataset containing interpretable attribution labels during this process, and synchronously executing the dynamic state evolution of the synthetic student agent;

[0021] The generation of a structured educational dataset containing interpretable attribution labels includes:

[0022] Behavioral attribution annotation: Generate metadata tags for each interaction record to annotate the generation motivation behind the response behavior; the generation motivation covers the knowledge association dimension based on the probability of knowledge mastery and cognitive bias, the personality-driven dimension driven by inherent characteristics or instantaneous state, and the strategy simulation dimension including guessing or avoidance strategies based on low mastery.

[0023] Standardized encapsulation and adaptation: The dialogue logs, behavioral decision sequences, state evolution data and the aforementioned metadata tags are uniformly encapsulated into standardized time-series data, which can be used to support the training of downstream educational diagnostic algorithms, fine-tuning of instruction for large teaching models, or as an evaluation benchmark for virtual educational environments.

[0024] The execution of dynamic state evolution involves establishing a dual update mechanism for knowledge and personality to drive agent state transitions, including:

[0025] Knowledge state transition: After the interaction ends, the dialogue log is parsed and the knowledge state parameters are updated using reinforcement learning algorithms or state transition matrices. This includes a positive cognitive gain mechanism, which adds the probability of mastering knowledge points based on a preset learning rate parameter when effective teaching scaffolding or feedback is detected; and a reverse time decay mechanism, which introduces a memory forgetting curve function to automatically reduce the probability of mastering knowledge points when they are not activated within a preset window.

[0026] The static inherent feature and dynamic instantaneous state coupled model update: The personality parameters are divided into inherent feature parameters and instantaneous state parameters; the relative stability of the inherent feature parameters configured in the initial configuration is maintained within the simulation cycle; at the same time, based on the interaction duration, task difficulty and the emotional tone of the teacher's feedback, the instantaneous state vector containing fatigue, motivation level and emotional valence is calculated in real time, so that the agent exhibits dynamic characteristics of inattention or enhanced learning motivation.

[0027] Furthermore, the parameter construction of the multidimensional feature parameter set in step S1 includes:

[0028] The construction of the cognitive ability parameters involves: quantifying the agent's thinking depth using a multi-level cognitive classification standard; the cognitive classification standard covers at least six dimensions: memory, comprehension, application, analysis, evaluation, and creation; and assigning a cognitive ability vector to each synthetic student agent. The vector Each dimension corresponds to a cognitive dimension's ability weight, used to constrain the reasoning depth and logical complexity of the large language model when generating responses; simultaneously, a cognitive fluctuation factor is set. It is used to simulate the fluctuation of students' cognitive abilities under different levels of fatigue or attention, so that the generated cognitive performance has random perturbation characteristics.

[0029] The construction of the personality trait parameters: A personality parameter vector is constructed based on the Big Five personality theory. The vector It includes numerical values ​​for five dimensions: openness, conscientiousness, extraversion, agreeableness, and neuroticism; and establishes a mapping database between personality traits and language styles, wherein the mapping database is defined differently. The value range corresponds to vocabulary selection preferences, sentence structure features, and emotional polarity; during the initialization phase, cue word engineering techniques are used to vectorize the personality parameters. The system instructions are converted into natural language descriptions and embedded into the context window of the synthetic student agent to guide the agent's behavior style and expression patterns;

[0030] The construction of the knowledge state parameters involves: importing a knowledge graph of a specific discipline and defining a set of knowledge points. ; Construct a knowledge mastery probability vector for each synthetic student agent. , of which elements This indicates that the agent has correctly grasped the knowledge points. The probability; further define the set of erroneous concepts, and associate specific subsets of erroneous concepts with specific types of synthetic student agents to simulate typical learning obstacles or cognitive biases; the knowledge state parameters are used to determine the accuracy rate and the type of errors generated when the agent answers academic questions through probability sampling during the simulation process.

[0031] Furthermore, step S2, which involves building a simulation interactive environment based on a large language model, includes:

[0032] Configure the global environment controller: used to manage the simulation clock, round number and interaction sequence to ensure synchronous cooperation among agents;

[0033] Construct a simulation base model: Load a pre-tuned large-scale education model with a corpus of education vertical domains as the base model, or attach an external education knowledge base through retrieval enhancement generation technology to provide professional background knowledge for the intelligent agent;

[0034] Define the interaction protocol: The protocol specifies the message transmission format between the teacher agent and the student agent, and the format includes at least the sender ID, receiver ID, message type, message content and timestamp;

[0035] Set up a private reasoning channel: This is used to synthesize student agents to perform implicit reasoning before generating public responses, and to record the reasoning process in a log but not send it to the teacher agent, so as to realize the internal recording of the agent's thinking path;

[0036] Configuring teacher agent strategies:

[0037] Teaching style configuration: Supports the instantiation of multiple teaching styles, including Socratic questioning, direct instruction, scaffolded guidance, and emotional support;

[0038] Dynamic strategy adjustment: The teacher agent has a teaching strategy library and can dynamically adjust subsequent teaching instructions based on the real-time feedback from the synthesized student agent;

[0039] Teaching task guidance: Before the simulation begins, the teaching syllabus and expected teaching objectives are input into the teacher agent as guiding criteria for the interaction process.

[0040] Furthermore, step S3, executing the teaching simulation interaction process, specifically includes the following sub-steps:

[0041] S3-1, Global Context Construction: Extract the knowledge intent and emotional polarity from the teacher's instructions, and combine them with the historical dialogue sequence and the state of the simulation environment to construct global context information that includes teaching context and social context;

[0042] S3-2, Multidimensional parameter instruction mapping: The multidimensional feature parameter set is dynamically transformed into behavioral constraint prompts for the large language model; wherein, cognitive ability parameters are mapped to the reasoning step size and logical rigor constraints of the response content, personality feature parameters are mapped to the vocabulary preference and emotional expression tendency of the mimicry response, and knowledge state parameters are mapped to the degree of mastery of specific knowledge points and the expected error distribution.

[0043] S3-3, Controlled Mimicry Reasoning Generation: The global context information and behavioral constraint prompts are semantically fused and input into the base large language model. Through restricted sampling strategies or thought chain guidance, candidate mimicry response behaviors that conform to preset student characteristics are generated.

[0044] S3-4, Candidate Behavior Sequence Output: Extract and structure the text responses, action intentions, and potential psychological states in the candidate mimicry response behaviors, and output them as data to be verified to step S4.

[0045] Furthermore, the specific steps for performing behavioral consistency verification in step S4 include:

[0046] S4-1, Construct a consistency evaluator: Deploy an evaluation component based on a large language model or a pre-trained discriminator network as a consistency evaluation module to perform real-time auditing of the output of the synthetic student agent;

[0047] S4-2, Input multidimensional verification data: Input the candidate mimicry response text generated by the synthesized student agent, the global context information of the current simulation environment, and the multidimensional feature parameter set of the agent at the current moment as input vectors into the consistency evaluation module;

[0048] S4-3, Multi-dimensional Consistency Scoring: Perform parallel verification across three dimensions:

[0049] Knowledge consistency verification: Evaluate whether the correctness of the response content and the error type are consistent with the knowledge mastery probability distribution and the preset set of error concepts of the agent;

[0050] Cognitive consistency verification: Evaluate whether the logical reasoning depth and language organization complexity of the response content are consistent with the cognitive ability level of the agent;

[0051] Personality consistency verification: Evaluate whether the tone, emotional polarity, and word choice of the response content are consistent with the personality trait parameters set for the agent;

[0052] S4-4, Judgment and Interception Strategy: Calculate a weighted comprehensive score based on the scores of the above three dimensions. If the comprehensive score is higher than the preset consistency threshold, the candidate mimicry response is deemed to have passed the verification and is allowed to be output externally; otherwise, it is deemed to be a role deviation behavior and the candidate response is intercepted.

[0053] S4-5, Adaptive Correction Mechanism: When an interception is triggered, a feedback signal containing a description of the specific violation type is generated and sent back to the synthetic student agent; at the same time, by dynamically adjusting the generation hyperparameters, including the sampling temperature or nuclear sampling threshold, and adding parameter reinforcement instructions to the prompt words, the agent is required to regenerate the response; if the number of consecutive regenerations reaches the preset limit and still fails to pass the verification, the rule-based backup library is called to generate a fallback response that conforms to the current parameter settings and output it externally.

[0054] A synthetic student agent construction and simulation interaction system for virtual educational experiments includes:

[0055] Parameterized configuration module: It can provide a visual interactive interface or a structured configuration file interface to receive a set of multi-dimensional feature parameters of the synthetic student intelligent agent set by the user. The set of parameters covers cognitive ability parameters, personality feature parameters and knowledge state parameters, and serializes and stores the parameters to form an intelligent agent profile library.

[0056] Intelligent agent instance generation module: used to read data from the parameterized configuration module, call the large language model interface to instantiate multiple independent synthetic student intelligent agent objects in parallel, and instantiate at least one teacher intelligent agent object at the same time, and establish communication handles between intelligent agents;

[0057] Simulation Interaction Execution Engine: As the runtime environment of the virtual classroom, it drives the simulation clock, manages message distribution, context state maintenance and dialogue turn flow among all agents, and supports synchronous or asynchronous teaching interaction modes.

[0058] Group dynamics simulation unit: integrated into the engine, used to support lateral interaction between synthetic student agents; the lateral interaction is controlled by preset social relationship parameters, which define the familiarity, social influence and willingness to cooperate among individual students;

[0059] Behavior consistency verification and correction module: Embedded in the output pipeline of the simulation interaction execution engine, it is used to intercept and audit the candidate mimicry responses of the synthesized student agent in real time before they are broadcast to the outside world, to ensure that the generated content conforms to the parameter set settings of the agent at the current moment; in the group interaction mode, this module also has a group behavior verification function, which is configured to verify the sociological rationality of group behavior, including identifying and evaluating the simulation realism of group conformity behavior or social loafing behavior;

[0060] Full-domain data acquisition and analysis module: used to capture the data stream of the entire simulation process without loss, and generate a structured experimental report containing multi-round interactive text, thought chain reasoning logs and state parameter evolution curves;

[0061] Real-time status monitoring terminal: Used to render and display the fluctuation charts of psychological state indicators and the heat map of knowledge mastery of each synthetic student agent in real time.

[0062] Furthermore, the behavior consistency verification and correction module includes a reward model unit and a policy optimization unit built based on reinforcement learning from human feedback, wherein:

[0063] Reward Model Unit: The reward model unit is trained using pairwise preference data or ranking data labeled by real education experts, thereby gaining the ability to infer and score the synthetic student agent's response in real time on three dimensions: cognitive logic consistency, knowledge mastery deviation, and language style realism.

[0064] The strategy optimization unit is configured to use the score of the reward model unit as a reward signal, and use a proximal policy optimization algorithm or a weighted rejection sampling strategy to adjust the cue word weights or decoding parameters in the model generation process in a targeted manner to maximize the output probability that conforms to the distribution of real student behavior.

[0065] The present invention also relates to a computer electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above method.

[0066] The beneficial effects of this invention are as follows:

[0067] By using three-dimensional parametric modeling of the cognitive abilities, personality traits, and knowledge status of synthetic student agents, different students can exhibit clear and controllable differences in their depth of thinking, language expression style, and knowledge mastery. This enables the controllable configuration of individual student differences and effectively alleviates the homogenization problem in existing educational simulations, where student agents exhibit highly consistent behavioral patterns and convergent answer styles and error types.

[0068] By introducing a closed-loop mechanism for behavioral consistency verification and correction during the simulation interaction process, the matching degree between the generated content of the synthesized student agent and the predetermined character parameters and knowledge state is evaluated and constrained in real time before the content is output. Samples that do not conform to the settings are intercepted and corrected in a timely manner, thereby significantly improving the stability of simulation interaction and the overall quality of generated data in multi-round teaching dialogue.

[0069] By recording the simulation interaction process in a structured log and automatically attaching metadata tags reflecting the motivation and internal state of the generated student responses, the interpretability and reusability of the synthesized educational data at the cognitive and behavioral levels are improved. This makes the data more suitable for applications such as educational diagnostic analysis, evaluation of the effectiveness of teaching strategies, and training, fine-tuning, and evaluation of large-scale educational models.

[0070] By supporting group interaction mechanisms between synthetic student agents and introducing social relationship parameters for constraints, the system can simulate mutual influence, conformity behavior, and collaborative relationships among students, thereby extending to classroom group dynamics research scenarios and covering more complex and diverse virtual education experiment needs.

[0071] By using automated simulation interaction and data generation methods based on large language models, large-scale, high-quality educational experimental data can be produced with low manpower and time costs and high generation efficiency without the need for real classroom participation. This effectively reduces the organizational costs, time costs, and compliance risks required for real classroom experiments. Attached Figure Description

[0072] Figure 1 This is a schematic diagram of the overall process of the method of the present invention;

[0073] Figure 2 This is a system architecture diagram of the present invention;

[0074] Figure 3 This is a schematic diagram of the electronic device of the present invention. Detailed Implementation

[0075] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings and examples. It should be noted that the following examples are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. Without departing from the principles of the present invention, those skilled in the art can make various modifications or substitutions to the embodiments, all of which should fall within the scope of protection of the present invention.

[0076] Example 1: Overall Process of Synthetic Student Agent Construction and Classroom Simulation Interaction

[0077] This embodiment provides a method for constructing and simulating synthetic student agents for virtual educational experiments. This method is driven by a Large Language Model (LLM) and aims to generate a high-fidelity, controllable, and dynamically evolving virtual student group. For example... Figure 1 As shown, the method in this embodiment mainly includes steps S1 to S5.

[0078] Step S1: Construct a parameterized model of the synthetic student agent

[0079] Based on preset experimental requirements, taking the examination of students at different cognitive levels' responses to a certain teaching method as an example, the system initializes and configures the multidimensional feature parameter set for each synthetic student agent. During initialization, cognitive ability, personality traits, and knowledge state parameters are directly configured to meet the experimental requirements for examining students at different cognitive levels' responses to the teaching method.

[0080] When constructing the parameterized model, the system uses the following dimensions and constraints to ensure the controllability and realism of the agent's behavior. These dimensions are also important references for measuring the agent's performance in subsequent simulation experiments:

[0081] Quantitative Standards: Using multi-level standards such as Bloom's Taxonomy of Educational Objectives, the depth of thinking is quantified into six dimensions: memory, comprehension, application, analysis, evaluation, and creation.

[0082] Vector configuration: Assigning cognitive ability vectors to each agent ,in Represents the capability weights of the corresponding dimensions, such as This represents students with high memory but low analytical skills. This vector is used in the S3 phase to constrain the length of the reasoning chain and the logical complexity when the LLM generates text.

[0083] Random perturbation: Setting a cognitive fluctuation factor It is used to simulate the fluctuations in students' cognitive abilities caused by inattention during the simulation process, so that the output performance has more realistic randomness.

[0084] Based on the above quantification standards and configuration logic, this invention constructs a multi-dimensional feature parameter set for the synthetic student agent through the following specific process:

[0085] The cognitive ability parameter construction system employs a multi-level cognitive classification standard to quantify the agent's thinking depth, specifically including three parts: dimension definition, vector allocation, and a random perturbation mechanism. The dimension definition uses Bloom's Taxonomy of Educational Objectives as the cognitive classification standard, covering at least six dimensions: Remember, Understand, Apply, Analyze, Evaluate, and Create. The vector allocation assigns a cognitive ability vector to each synthetic student agent. : Each dimension corresponds to a cognitive dimension's ability weight, serving as a hard constraint during response generation to limit the reasoning depth and logical complexity of the large language model. The random perturbation mechanism sets a cognitive fluctuation factor. It is used to simulate the fluctuation of students' cognitive performance under different fatigue or attention states, and its actual cognitive ability is reflected. satisfy ,in To make the agent's output exhibit nondeterministic behavior that conforms to human characteristics, it is a random variable that follows a preset distribution.

[0086] The construction of the personality trait parameters includes two parts: vector configuration and Prompt mapping. The vector configuration is based on the Big Five personality theory to construct parameter vectors. It includes normalized values ​​for five dimensions: openness, conscientiousness, extraversion, agreeableness, and neuroticism. The Prompt mapping refers to establishing a "parameter-language style" mapping library. For example, when extraversion > 0.8, it maps to "actively asking questions, using interjections, and using short sentences"; when agreeableness < 0.3, it maps to "a tendency to refute and an increased probability of using a questioning tone." During the initialization phase, the system utilizes prompt word engineering to... The system instructions are translated into natural language descriptions and embedded in the long-term memory window of the intelligent agent.

[0087] The construction of the knowledge state parameters includes two parts: probability vector construction and incorrect concept implantation. The probability vector construction process includes importing a knowledge graph of a specific discipline, defining a knowledge point set K, and constructing a knowledge mastery probability vector. , of which elements This indicates the grasp of probability. The implantation of associated erroneous concepts refers to defining a set of erroneous concepts; for example, for "fraction addition," associating it with the typical error of "adding numerators and denominators." Based on the agent type, the system binds a specific subset of erroneous concepts to it, which is used in phase S3 to simulate real "wrong answer" behavior through probability sampling.

[0088] Step S2: Build a simulation interactive environment based on a large language model

[0089] Instantiate the environment controller, teacher agent, and multiple synthetic student agents in the simulation execution engine, and establish a communication topology.

[0090] Environment and model configuration

[0091] Controller: Configure the global environment controller to manage the simulation clock, number of interaction rounds, and speaking order.

[0092] Base Model: Load an LLM from the education vertical as a base, and use RAG technology to mount external textbooks or question banks to provide the agent with accurate subject background knowledge.

[0093] Interaction protocol: Defines a standardized message format This ensures that the data flow is structured.

[0094] Teacher agent configuration

[0095] Teaching style: Supports instantiation as a "Socratic questioner" or a "direct lecturer".

[0096] Dynamic strategy adjustment: The teacher agent has a built-in strategy library that can dynamically switch strategies based on students' real-time feedback.

[0097] Private reasoning channel

[0098] A private inference channel is opened for each synthetic student agent. Before generating a public response, the agent first generates an "inner monologue," which is recorded in the log but not sent to the teacher, and is used for subsequent personality state analysis.

[0099] Step S3: Execute the teaching simulation interaction process

[0100] After the teacher agent generates instructions based on the learning objectives, the synthesized student agent performs the following sub-steps to generate a response:

[0101] Full-domain context construction: The system extracts the knowledge intent and emotional polarity in the teacher's instructions, and combines them with the historical dialogue sequence to construct a full-domain context that includes both teaching and social contexts.

[0102] Multidimensional parameter instruction mapping:

[0103] cognitive parameters Transform it into an inference step size constraint;

[0104] personality parameters Transformed into a tone constraint;

[0105] knowledge parameters Transform it into a right / wrong constraint.

[0106] Controlled mimicry reasoning generation: The above context and constraint instructions are fused and input into the LLM. Candidate mimicry responses that conform to the "persona" are generated through restricted sampling strategies or thought chain guidance.

[0107] Candidate sequence output: Extract the text response, action intent, and potential psychological state from the candidate response to prepare for the verification stage.

[0108] Step S4: Behavioral Consistency Verification and Correction

[0109] Before responding to external output, perform a strict consistency closed-loop check:

[0110] Consistency evaluator: Utilizes a trained discriminator model or RLHF reward model, inputting candidate responses. Context and the current parameter set .

[0111] Three-dimensional scoring mechanism:

[0112] Knowledge Consistency: Check whether the correctness / incorrectness of the answer is consistent with... The probability matches.

[0113] Cognitive consistency: Check if the logical depth exceeds the limit. Restrictions.

[0114] Personality consistency: Check if the tone of voice is consistent. .

[0115] Interception and Correction: Calculate the weighted total score. If it falls below a preset threshold, it is judged as "role deviation" and intercepted. Trigger correction mechanism:

[0116] Parameter-enhanced regeneration: Add specific criticisms to the prompts, requiring LLM rewriting.

[0117] fallback strategy: If continuous regeneration occurs If the simulation fails again, a rule-based backup library is invoked to generate a fallback response that conforms to the settings, ensuring that the simulation is not interrupted.

[0118] Step S5: Dynamic Evolution and Data Generation

[0119] This step updates the agent's state and outputs the final data:

[0120] Perform dynamic state evolution:

[0121] Knowledge state transition: Analyzing the results of this round of interaction. If the student receives a valid explanation from the teacher, the system proceeds according to the preset learning rate. Perform positive cognitive gains: Conversely, if a knowledge point has not been accessed for a long time, reverse time decay is performed according to the forgetting curve function.

[0122] Model update coupling static inherent features and dynamic instantaneous state: distinguishing between "inherent features" and "instantaneous state". Maintaining... The instantaneous state vector remains unchanged, but is updated in real time based on the interaction duration and task difficulty. For example, repeatedly answering incorrectly can lead to increased feelings of frustration, which in turn can affect the next round of S3 reasoning.

[0123] Closed-loop feedback: The updated and The data is then fed back to steps S1 / S3 as a baseline for the next round of interaction, thus achieving a closed loop in the evolution of model parameters.

[0124] Generate structured causal data:

[0125] Behavioral attribution annotation: Generate metadata tags for each record to explain "why it is said this way".

[0126] Knowledge association dimension: [Error generated based on 40% mastery rate] / [Triggering misconceptions: adding denominators].

[0127] Personality-dominant dimension: [Driven by high neuroticism: exhibiting anxiety].

[0128] Strategy simulation dimension: [Trigger avoidance strategy: changing the topic].

[0129] Data encapsulation: The dialogue logs, decision sequences, state evolution curves, and the aforementioned attribution labels are encapsulated into standardized time-series data for subsequent instruction fine-tuning (SFT) of the large-scale educational model or training of educational diagnostic algorithms.

[0130] Example 2: The dynamic evolution mechanism of students' knowledge state and personality state

[0131] This embodiment details how the system updates the internal state of the synthesized student agent in real time based on multi-round interaction data in the simulated interactive closed loop of step S5. This process aims to simulate the knowledge consolidation, natural forgetting, and fluidity of psychological states of real learners under teaching intervention.

[0132] Dynamic transfer of knowledge state

[0133] In this embodiment, the system maintains a dynamically updated knowledge acquisition probability vector. The update of this vector follows a two-way adjustment mechanism of "positive gain" and "reverse decay".

[0134] Positive cognitive gain mechanism:

[0135] Triggering conditions: When the synthetic student agent receives effective instructional scaffolding, correctness feedback, or detailed explanations from the teacher agent during interaction, or when the student agent successfully completes a deep reasoning task.

[0136] Update logic: The system updates based on preset learning rate parameters. Calculate the cognitive gain value. This gain value depends not only on the current interaction round but also on the upper limit of the agent's cognitive ability. For example, regarding a certain knowledge point Its update formula can be expressed as:

[0137]

[0138] The specific meanings of the parameters are as follows: The updated state indicates the synthetic student agent's understanding of a specific knowledge point after the current round of interaction. The probability of mastering; Initial state: This refers to the agent's understanding of the knowledge points before the start of this round of interaction. The original probability of control; The learning rate is a preset hyperparameter used to control how quickly an agent acquires new knowledge. Interaction depth factor: Used to quantify the effectiveness of this round of instructional interaction. Its value depends on the quality of the instructional scaffolding received by the agent, the level of detail in the teacher's feedback, or the complexity of the task in which the student participates in reasoning. Cognitive gain space, this term serves as a weighting term, ensuring that as the probability of mastery increases... As the probability approaches 100%, the probability increment from a single interaction gradually decreases. This design aims to simulate the diminishing marginal effect in the real learning process, meaning that the higher the level of knowledge mastery, the greater the difficulty of further improvement.

[0139] Error reinforcement simulation: If the synthetic student agent reveals a specific misconception in its answer, the system will lock the mastery probability of the relevant knowledge point, or fine-tune the probability distribution in the "error solidification" mode, so that it reproduces the error with a higher probability in subsequent interactions, until it receives clear corrective feedback.

[0140] Reverse time decay mechanism:

[0141] Triggering condition: The system introduces a time decay function based on the Ebbinghaus forgetting curve to monitor each knowledge point in real time. unvisited time interval .

[0142] Update logic: When a knowledge point is not activated within a preset time window, the system automatically performs a decay operation.

[0143] in, This is the forgetting rate coefficient, which can be correlated with the agent's "memory" cognitive parameters. This mechanism ensures the long-term dynamic realism of the simulation process, forcing the teacher agent to adopt "review strategies" to maintain the students' knowledge status.

[0144] Construction and updating of the coupling model between static inherent characteristics and dynamic instantaneous state

[0145] To address the problem of rigid agent personalities or unreasonable emotional fluctuations in traditional simulations, this embodiment introduces a two-layer modeling mechanism that separates "inherent features" and "instantaneous states".

[0146] Inherent Feature Layer: Maintain the personality parameter vector configured in step S1. (e.g., the Big Five personality traits) exhibit relative stability within a single simulation experiment cycle. This ensures that the core persona of the intelligent agent will not experience Out-of-Distribution (OOD) drift due to short-term interactions.

[0147] Transient state layer:

[0148] Vector definition: Establishing an instantaneous state vector independent of the personality vector. This vector contains at least four dimensions: [fatigue level, learning motivation, anxiety level, and frustration].

[0149] Real-time fine-tuning logic: The system calculates the state increment in real time based on the context features of the current interaction.

[0150] Fatigue level update: based on interaction rounds The fatigue level increases non-linearly and is weighted by the task difficulty. When fatigue exceeds a threshold, the agent will tend to output short sentences or trigger "attention lapse" behavior when generating responses.

[0151] Motivation and anxiety update: modulated based on the emotional tone of teacher feedback.

[0152] Example 1 (Positive Feedback): When receiving "encouraging feedback," the "anxiety level" decreases, and the "learning motivation" increases. This, through the Prompt injection, makes the agent show a higher willingness to try in the next round, daring to answer even if its knowledge is not high.

[0153] Example 2 (Negative Feedback): When repeatedly answering incorrectly and receiving "harsh criticism," the agent experiences a significant increase in "frustration" and "anxiety." This may lead the agent to trigger "learned helplessness" behaviors, such as refusing to answer or directly seeking help.

[0154] State update algorithm implementation

[0155] The process of updating the aforementioned knowledge and psychological state can be implemented in one of the following two ways:

[0156] Rule-driven engine: Based on a predefined state transition matrix, it is suitable for teaching attribution research with extremely high interpretability requirements.

[0157] Reinforcement learning adaptation: Treating teachers' teaching behavior as the Environment and students' internal states as the State, the Q-Learnin algorithm is used to dynamically adjust the state parameters to maximize the fit to the behavioral distribution of the real student dataset.

[0158] Example 3: Group Dynamics Simulation and Lateral Interaction

[0159] To construct complex teaching experiment scenarios, this embodiment uses a group interaction mode to simulate the mutual influence and group behavior characteristics among students in a real classroom environment.

[0160] Horizontal interaction mechanism based on social relations

[0161] In group interaction mode, such as group discussions or free speech by the whole class, in addition to maintaining vertical interaction between teachers and students, the system also establishes horizontal communication connections between synthetic student agents.

[0162] Social Relationship Definition: The system presets or randomly generates social relationship parameters to quantify the "familiarity" and "social influence" between individual students.

[0163] Interaction Logic: When a student agent speaks, other agents connected to it receive that speech as new context input. During this process, the system determines the weight of information based on the aforementioned parameters: the opinions of students with high influence are more likely to be noticed or cited by other students, thus simulating the "opinion leader" phenomenon in the classroom.

[0164] Simulation of typical group psychological phenomena

[0165] The system has built-in swarm dynamics rules, supporting the spontaneous emergence or guided generation of the following two typical behaviors in a multi-agent environment:

[0166] Conformity: When a mainstream viewpoint (even if incorrect) emerges in the classroom, the system adjusts its generation probability based on the agent's "acceptability" and "knowledge confidence" parameters. For example, it may induce students with lower confidence to abandon their original correct viewpoints and instead conform to the group's opinion.

[0167] Social loafing: In large group collaborative tasks, the system will dynamically reduce the participation of some agents according to the group size, causing them to exhibit "free-riding" characteristics, such as speaking briefly and only expressing agreement without contributing substantive content.

[0168] Consistency verification of group behavior

[0169] To prevent simulation distortion, the system performs logical verification of group behavior at the output:

[0170] Reasonableness assessment: The verification module will evaluate in real time whether the agent's changes in opinion are consistent with its persona.

[0171] Passing scenario: If a student who is set to have "easygoing personality and average grades" exhibits conformity behavior, the system judges it as a reasonable simulation and allows output.

[0172] Interception scenario: If a student who is set to be "independent and academically excellent" blindly follows the crowd without reason, the system will determine it as role deviation (OOD), trigger the correction mechanism, and force the student to maintain an independent viewpoint.

[0173] Feedback Correction: For intercepted unreasonable group behavior, the system will strengthen the instructions "maintain independent thinking" or "actively participate" in the prompt, requiring the agent to regenerate a response that conforms to its parameter settings.

[0174] Example 4: System Architecture and Hardware Implementation

[0175] This embodiment details the system logic architecture and electronic equipment used to execute the above-described method for constructing and simulating synthetic student agents.

[0176] System Logical Architecture

[0177] like Figure 2 As shown, the simulation system of this invention is composed of the following six core functional modules coupled together:

[0178] Parameterized Configuration Module: This module provides a visual interactive interface (GUI) or a structured configuration file interface. Users can use this module to set a multi-dimensional set of feature parameters for the synthesized student agent (covering cognitive ability, personality traits, and knowledge state). The module is responsible for serializing these configurations to build a reusable agent profile library.

[0179] The agent instance generation module reads data from the profile database, calls the Large Language Model (LLM) API interface to instantiate multiple independent synthetic student agent objects in parallel, and simultaneously loads the teacher agent configuration. This module is also responsible for establishing communication handles between agents and initializing memory space for subsequent message passing.

[0180] Simulation Interaction Execution Engine: Serving as the runtime environment for the virtual classroom, this engine is responsible for driving the simulation clock and managing message distribution, context maintenance, and dialogue turn flow among all agents. Specifically, this engine integrates a swarm dynamics simulation unit, used to calculate and process lateral interaction requests between student agents based on preset social relationship parameters in swarm interaction mode.

[0181] Behavioral Consistency Verification and Correction Module: This module is embedded in the output pipeline of the simulation engine and consists of two parts: individual basic verification and group rationality verification. First, basic verification: Real-time interception of candidate responses to evaluate their matching degree with the current parameter set; RLHF enhancement: Built-in reward model unit based on human feedback reinforcement learning. This model is trained using pairwise preference data labeled by educational experts and can output reward signals from three dimensions: logical consistency, knowledge bias, and style fidelity. It also uses the PPO algorithm or rejection sampling strategy to optimize the generated results. Second, group verification: In multi-person mode, it is specifically responsible for verifying the rationality of conformity or social loafing behavior.

[0182] The full-domain data acquisition and analysis module is used to capture the data flow of the entire simulation process without loss. This module not only records explicit dialogue text, but also extracts implicit thought chain (CoT) inference logs and state parameter evolution curves, and finally generates a structured experimental report containing behavioral causal metadata tags.

[0183] Real-time status monitoring terminal: Used for front-end visualization. Supports real-time rendering of psychological state index fluctuation graphs and knowledge mastery heatmaps for each synthesized student, allowing experimental personnel to monitor the simulation process in real time.

[0184] The computer electronic device of the present invention includes: a memory and a processor; wherein the memory includes 64G of RAM and more than 512G of storage space, used to store large-scale context data and state matrices generated during computer programs and simulations; the processor includes at least an 8-core central processing unit (CPU) used to execute the computer program, drive the inference calculation of the large language model and the operation of simulation logic, thereby realizing the above method.

Claims

1. A method for constructing a synthetic student intelligent agent for educational experiments, characterized in that, The method includes the following steps: Step S1: Construct a parameterized model of the synthetic student agent. According to the preset experimental requirements, initialize and configure the multidimensional feature parameter set of each synthetic student agent. The multidimensional feature parameter set includes at least cognitive ability parameters, personality feature parameters, and knowledge state parameters. Step S2: Build a simulation interaction environment based on a large language model, construct at least one teacher agent and multiple synthetic student agents based on the configuration in step S1 in the environment, and establish a communication topology network between the agents. Step S3: Execute the teaching simulation interaction process. The teacher agent generates teaching instructions based on the teaching objectives. The synthetic student agent receives the teaching instructions and the environmental context, and performs reasoning based on the multi-dimensional feature parameter set to generate candidate mimicry response behaviors. Step S4: Perform behavior consistency verification. Before the synthetic student agent generates candidate mimicry response behaviors and outputs them externally, the behavior consistency evaluation module calculates the matching degree of the candidate mimicry response behaviors based on the multidimensional feature parameter set of the synthetic student agent at the current moment. If the matching degree is lower than the preset threshold, a correction mechanism is triggered. Step S5: Generate and output a structured educational dataset, and drive the dynamic evolution closed loop of the synthesized student agent parameters based on data feedback.

2. The method for constructing a synthetic student agent for educational experiments according to claim 1, characterized in that, The parameter construction process of the multidimensional feature parameter set in step S1 is as follows: The process of constructing the cognitive ability parameters includes: quantifying the agent's thinking depth using a multi-level cognitive classification standard; the cognitive classification standard covers at least six dimensions: memory, comprehension, application, analysis, evaluation, and creation; and assigning a cognitive ability vector to each synthetic student agent. The vector Each dimension corresponds to a cognitive dimension's ability weight, used to constrain the reasoning depth and logical complexity of the large language model when generating responses; simultaneously, a cognitive fluctuation factor is set. It is used to simulate the fluctuation of students' cognitive abilities under different levels of fatigue or attention, so that the generated cognitive performance has random perturbation characteristics. The process of constructing the personality trait parameters includes: constructing a personality parameter vector based on the Big Five personality theory. The vector It includes numerical values ​​for five dimensions: openness, conscientiousness, extraversion, agreeableness, and neuroticism; and establishes a mapping database between personality traits and language styles, wherein the mapping database is defined differently. The value range corresponds to vocabulary selection preferences, sentence structure features, and emotional polarity; during the initialization phase, cue word engineering techniques are used to vectorize the personality parameters. The system instructions are converted into natural language descriptions and embedded into the context window of the synthetic student agent to guide the agent's behavior style and expression patterns; The process of constructing the knowledge state parameters includes: importing a knowledge graph of a specific discipline and defining a set of knowledge points. ; Construct a knowledge mastery probability vector for each synthetic student agent. , of which elements This indicates that the agent has correctly grasped the knowledge points. The probability; further define the set of erroneous concepts, and associate specific subsets of erroneous concepts with specific types of synthetic student agents to simulate typical learning obstacles or cognitive biases; the knowledge state parameters are used to determine the accuracy rate and the type of errors generated when the agent answers academic questions through probability sampling during the simulation process.

3. The method for constructing a synthetic student intelligent agent for educational experiments according to claim 1, characterized in that, In step S2, the construction of the simulation interaction environment based on the large language model includes: Configure the global environment controller: used to manage the simulation clock, round number and interaction sequence to ensure synchronous cooperation among agents; Construct a simulation base model: Load a large education model that has been fine-tuned by a corpus of education vertical domains as the base model, or attach an external education knowledge base through retrieval enhancement generation technology to provide professional background knowledge for the intelligent agent; Define the interaction protocol: The protocol specifies the message transmission format between the teacher agent and the student agent, and the format includes at least the sender ID, receiver ID, message type, message content and timestamp; Set up a private reasoning channel: This is used to synthesize student agents to perform implicit reasoning before generating public responses, and to record the reasoning process in a log but not send it to the teacher agent, so as to realize the internal recording of the agent's thinking path; Configuring teacher agent strategies: Teaching style configuration: Supports the instantiation of multiple teaching styles, including Socratic questioning, direct instruction, scaffolded guidance, and emotional support; Dynamic strategy adjustment: The teacher agent has a teaching strategy library and can dynamically adjust subsequent teaching instructions based on the real-time feedback from the synthesized student agent; Teaching task guidance: Before the simulation begins, the teaching syllabus and expected teaching objectives are input into the teacher agent as guiding criteria for the interaction process.

4. The method for constructing a synthetic student intelligent agent for educational experiments according to claim 1, characterized in that, Step S3, which executes the teaching simulation interaction process, specifically includes the following sub-steps: S3-1, Global Context Construction: Extract the knowledge intent and emotional polarity from the teacher's instructions, and combine them with the historical dialogue sequence and the state of the simulation environment to construct global context information that includes teaching context and social context; S3-2, Multidimensional parameter instruction mapping: The multidimensional feature parameter set is dynamically transformed into behavioral constraint prompts for the large language model; wherein, cognitive ability parameters are mapped to the reasoning step size and logical rigor constraints of the response content, personality feature parameters are mapped to the vocabulary preference and emotional expression tendency of the mimicry response, and knowledge state parameters are mapped to the degree of mastery of specific knowledge points and the expected error distribution. S3-3, Controlled Mimicry Reasoning Generation: The global context information and behavioral constraint prompts are semantically fused and input into the base large language model. Through restricted sampling strategies or thought chain guidance, candidate mimicry response behaviors that conform to preset student characteristics are generated. S3-4, Candidate Behavior Sequence Output: Extract and structure the text responses, action intentions, and potential psychological states in the candidate mimicry response behaviors, and output them as data to be verified to step S4.

5. The method for constructing a synthetic student intelligent agent for educational experiments according to claim 1, characterized in that, The specific steps for performing the behavior consistency verification in step S4 include: S4-1, Construct a consistency evaluator: Deploy an evaluation component based on a large language model or a pre-trained discriminator network as a consistency evaluation module to perform real-time auditing of the output of the synthetic student agent; S4-2, Input multidimensional verification data: Input the candidate mimicry response text generated by the synthesized student agent, the global context information of the current simulation environment, and the multidimensional feature parameter set of the agent at the current moment as input vectors into the consistency evaluation module; S4-3, Multi-dimensional Consistency Scoring: Perform parallel verification across three dimensions: Knowledge consistency verification: Evaluate whether the correctness of the response content and the error type are consistent with the knowledge mastery probability distribution and the preset set of error concepts of the agent; Cognitive consistency verification: Evaluate whether the logical reasoning depth and language organization complexity of the response content are consistent with the cognitive ability level of the agent; Personality consistency verification: Evaluate whether the tone, emotional polarity, and word choice of the response content are consistent with the personality trait parameters set for the agent; S4-4, Judgment and Interception Strategy: Calculate a weighted comprehensive score based on the scores of the above three dimensions. If the comprehensive score is higher than the preset consistency threshold, the candidate mimicry response is deemed to have passed the verification and is allowed to be output externally; otherwise, it is deemed to be a role deviation behavior and the candidate response is intercepted. S4-5, Adaptive Correction Mechanism: When an interception is triggered, a feedback signal containing a description of the specific violation type is generated and sent back to the synthetic student agent; at the same time, by dynamically adjusting the generation hyperparameters, including the sampling temperature or nuclear sampling threshold, and adding parameter reinforcement instructions to the prompt words, the agent is required to regenerate the response; if the number of consecutive regenerations reaches the preset limit and still fails to pass the verification, the rule-based backup library is called to generate a fallback response that conforms to the current parameter settings and output it externally.

6. The method for constructing a synthetic student intelligent agent for educational experiments according to claim 1, characterized in that, Step S5 specifically includes: recording dialogue logs, behavioral decision sequences, and state change data in real time during the simulation interaction, and constructing a structured educational experiment dataset based on this; extracting feedback features based on the educational experiment dataset, and transmitting the feedback features back to step S1 to dynamically update the multi-dimensional feature parameter set of the corresponding synthetic student agent, forming a closed loop of model parameter evolution driven by simulation data; generating a structured educational dataset containing interpretable attribution labels, and synchronously executing the dynamic state evolution of the synthetic student agent; The generation of a structured educational dataset containing interpretable attribution labels includes: Behavioral attribution annotation: Generate metadata tags for each interaction record to annotate the generation motivation behind the response behavior; the generation motivation covers: knowledge association dimension based on knowledge mastery probability and cognitive bias, personality-driven dimension driven by inherent characteristics or instantaneous state, and strategy simulation dimension including guessing or avoidance strategies based on low mastery. Standardized encapsulation and adaptation: The dialogue logs, behavioral decision sequences, state evolution data and the aforementioned metadata tags are uniformly encapsulated into standardized time-series data, which can be used to support the training of downstream educational diagnostic algorithms, fine-tuning of instruction for large teaching models, or as an evaluation benchmark for virtual educational environments. The execution of dynamic state evolution involves establishing a dual update mechanism for knowledge and personality to drive agent state transitions, including: Knowledge state transition: After the interaction ends, the dialogue log is parsed and the knowledge state parameters are updated using reinforcement learning algorithms or state transition matrices. This includes a positive cognitive gain mechanism, which adds the probability of mastering knowledge points based on a preset learning rate parameter when effective teaching scaffolding or feedback is detected; and a reverse time decay mechanism, which introduces a memory forgetting curve function to automatically reduce the probability of mastering knowledge points when they are not activated within a preset window. The static inherent feature and dynamic instantaneous state coupled model update: The personality parameters are divided into inherent feature parameters and instantaneous state parameters; the relative stability of the inherent feature parameters configured in the initial configuration is maintained within the simulation cycle; at the same time, based on the interaction duration, task difficulty and the emotional tone of the teacher's feedback, the instantaneous state vector containing fatigue, motivation level and emotional valence is calculated in real time, so that the agent exhibits dynamic characteristics of inattention or enhanced learning motivation.

7. A system for constructing synthetic student intelligent agents for educational experiments, characterized in that, include: Parameterized configuration module: It can provide a visual interactive interface or a structured configuration file interface to receive a set of multi-dimensional feature parameters of the synthetic student intelligent agent set by the user. The set of parameters covers cognitive ability parameters, personality feature parameters and knowledge state parameters, and serializes and stores the parameters to form an intelligent agent profile library. Intelligent agent instance generation module: used to read data from the parameterized configuration module, call the large language model interface to instantiate multiple independent synthetic student intelligent agent objects in parallel, and simultaneously construct at least one teacher intelligent agent object, and establish communication handles between intelligent agents; Simulation Interaction Execution Engine: As the runtime environment of the virtual classroom, it drives the simulation clock, manages message distribution, context state maintenance and dialogue turn flow among all agents, and supports synchronous or asynchronous teaching interaction modes. Group dynamics simulation unit: integrated into the engine, used to support lateral interaction between synthetic student agents; the lateral interaction is controlled by preset social relationship parameters, which define the familiarity, social influence and willingness to cooperate among individual students; Behavior consistency verification and correction module: Embedded in the output pipeline of the simulation interaction execution engine, it is used to intercept and audit the candidate mimicry response of the synthesized student agent in real time before it is broadcast to the outside world, to ensure that the generated content conforms to the parameter set settings of the agent at the current moment; In the group interaction mode, this module also has a group behavior verification function, which is configured to verify the sociological rationality of group behavior, including identifying and evaluating the simulation realism of group conformity behavior or social loafing behavior; Full-domain data acquisition and analysis module: Used to capture the data stream of the entire simulation process without loss, and generate a structured experimental report containing multi-round interactive text, thought chain reasoning logs and state parameter evolution curves; Real-time status monitoring terminal: used to render and display the fluctuation chart of psychological state indicators and the heat map of knowledge mastery of each synthetic student agent in real time.

8. A system for constructing synthetic student intelligent agents for educational experiments according to claim 7, characterized in that, The behavior consistency verification and correction module includes a reward model unit and a policy optimization unit built based on human feedback reinforcement learning: Reward Model Unit: The reward model unit is trained using pairwise preference data or ranking data labeled by real education experts, thereby gaining the ability to infer and score the synthetic student agent's response in real time on three dimensions: cognitive logic consistency, knowledge mastery deviation, and language style realism. The strategy optimization unit is configured to use the score of the reward model unit as a reward signal, and use a proximal policy optimization algorithm or a weighted rejection sampling strategy to adjust the cue word weights or decoding parameters in the model generation process in a targeted manner to maximize the output probability that conforms to the distribution of real student behavior.

9. A computer electronic device, characterized in that, include: Memory and processor; The memory stores a computer program, and the processor executes the computer program to implement the steps of the above method.

Images