Learning performance prediction method and system for large language model and multi-agent collaboration

Abstract

The application belongs to the technical field of knowledge tracking, and particularly relates to a learning performance prediction method and system of a large language model and a multi-agent collaboration, a graph construction agent constructs a knowledge point relationship network; a statistical analysis agent combines a network operation knowledge tracking model to output statistical prediction probability and knowledge mastery degree, calculates a question passing rate, maps a time interval into a forgetting label, and quantifies a question related evidence score; a dynamic parameter adjustment agent adjusts a subsequent deduction temperature according to the evidence score; a logical deduction agent combines a last moment cognitive image, the knowledge mastery degree, the forgetting label and a question feature to perform recursive reasoning, outputs a logical prediction probability and updates the image; and a decision fusion agent adaptively weights and fuses double-track probabilities based on the evidence score, and outputs a final performance prediction result. The application dynamically tracks the knowledge mastery state of students by combining a knowledge tracking model and a large language model, and improves the accuracy of student learning performance prediction under data sparsity and cold start.

Description

Technical Field

[0001] This invention belongs to the field of student knowledge tracking technology, and relates to a method and system for predicting learning performance through large language models and multi-agent collaboration. Background Technology

[0002] In the digital age of education, student cognitive assessment is receiving increasing attention. Cognitive assessment aims to uncover students' underlying cognitive structures by analyzing their historical response sequences. It can not only assess students' learning levels and psychological characteristics but also provide decision support for personalized teaching, such as recommending learning resources. How to accurately assess students' knowledge mastery and predict their future learning performance has become a research hotspot and core task in intelligent education. Currently, cognitive diagnostics and knowledge tracking technologies are key means to achieve this goal. However, existing cognitive assessment-based learning performance prediction methods still face significant challenges in practical applications.

[0003] In related technologies, predicting student learning performance mainly employs the following methods: Deep learning-based knowledge tracking technology uses recurrent neural networks or attention mechanisms to model students' historical answer sequences to predict future performance. However, this method relies on massive amounts of dense interaction data. When faced with newly enrolled students or cold-start scenarios with minimal interaction, the model fails to converge, resulting in a sharp drop in prediction accuracy. Alternatively, neurocognitive diagnostic techniques based on multidimensional continuous vectors model students' cognitive states as multidimensional continuous vectors for different knowledge points, providing a fine-grained description of students' learning states. These models are typically static diagnoses, lacking support for temporal state evolution and forgetting mechanisms. Furthermore, with sparse data, the models exhibit severe overfitting, leading to a significant decline in prediction performance.

[0004] Regarding the aforementioned technologies, existing student performance prediction methods are mostly limited to cold-start scenarios or sparse answer records. Due to a lack of sufficient evidence, traditional models often degenerate into random guessing. Furthermore, current research has not yet addressed the use of large language models to mine and utilize implicit knowledge point relationships, thus affecting the accuracy of predicting learners' knowledge mastery and consequently limiting the effectiveness of performance prediction. Finally, existing research largely involves the simple superposition of single models, lacking a dynamic student performance prediction framework that can adaptively coordinate statistical data and large model logical reasoning in cold-start or sparse data scenarios to avoid model overfitting or non-convergence. Summary of the Invention

[0005] Objective: The technical problem this invention aims to solve is to provide a method and system for predicting learning performance using a large language model and multi-agent collaboration. Based on a multi-agent collaborative framework, it utilizes offline large language model knowledge enhancement for graph reasoning and combines a dual-track statistical analysis mechanism to obtain more accurate multi-dimensional knowledge mastery, implicit logical connections, and temporal forgetting features. This provides the system with high-confidence statistical and semantic benchmarks, enabling high-precision and interpretable dynamic learning performance prediction.

[0006] Technical Solution: To achieve the above objectives, this invention provides a method for predicting the learning performance of a large language model and multi-agent collaboration, comprising the following steps: S1. The acquired students' historical answer records are timestamped and serialized to construct a standardized interaction time series, and the system is restricted to only acquiring interaction records before the target time. S2. Based on the knowledge point name identifiers of all questions in the dataset, construct an agent through a graph and call a large language model to extract the logical relationship of knowledge points and calculate the association strength, thereby obtaining the potential relationship network of knowledge points. S3. Based on the standardized interaction time series and combined with the knowledge point relationship network, a statistical analysis agent performs three-track feature processing. A Bayesian knowledge tracing model with graph and temporal enhancement is run on the numerical track to output statistical prediction probabilities and knowledge mastery levels. The pass rate of the target questions is calculated based on the remaining student interaction data. On the semantic track, the response time interval is calculated and mapped to obtain the natural language forgetting label. On the association quantification track, the association weight of historical questions and the time decay effect are taken into account, and the question association evidence score is calculated and output. S4. The temperature parameters of the large language model are adaptively adjusted based on the evidence score associated with the question by a dynamically tuned intelligent agent to perform logical deduction. S5. Based on the cognitive profile, knowledge mastery, target question features, and natural language forgetting labels of the previous moment, a recursive dynamic semantic memory stream based on the Markov state transition hypothesis is constructed through a logical deduction agent, and the logical prediction probability and the updated cognitive profile at the current moment are output under the control of the temperature parameter. S6. An adaptive gating fusion mechanism based on question-related evidence scores is adopted. The statistical prediction probability and the logical prediction probability are adaptively weighted and fused by the decision fusion agent to obtain the final learning performance prediction probability.

[0007] Furthermore, in step S1, the process of timestamp sorting and serialization of the acquired students' historical answer records specifically includes: acquiring raw interaction data containing student identifier, question identifier, answer score and answer timestamp; strictly sorting the historical answer records of the same student into a time series according to the order of timestamps; when predicting the learning performance at a target time, restricting the system to only acquire and process interaction records before that target time, forming a standardized interaction time series that prevents future data leakage.

[0008] Furthermore, in step S2, the specific calculation method for obtaining the potential relationship network of knowledge points by constructing the agent through the graph is as follows: (1) Knowledge logical relationship extraction: In the scenario where the specific question text content is missing, the abstract names of all knowledge points involved in the dataset are used as prompt words to input the pre-trained large language model; the implicit teaching logical relationship between knowledge points is extracted, and the logical relationship includes at least the prior relationship representing the order of learning paths and the similarity relationship representing the transferability of problem-solving thinking. (2) Generation of latent relation network: Based on the extracted logical relations and their association strength weights output by the large language model, a directed weighted graph is constructed. ,in The set of knowledge points in the dataset is represented by N, where N is the total number of knowledge points and E represents the semantic association edges between knowledge points. Each edge contains the relation type and weight, generating the potential relation network of the knowledge points, which provides implicit basis for cross-knowledge point state synchronization and logical reasoning.

[0009] Furthermore, in step S3, the feature processing on the numerical orbit, semantic orbit, and correlation quantization orbit is performed by the statistical analysis agent. The specific calculation method is as follows: (1) On the numerical track, construct a graph-enhanced Bayesian knowledge tracing (BKT) model. Initialize the prior mastery probability of each knowledge point. guessing rate With error rate Based on the time-series answer records, receive the mastery level of the target knowledge point k from the previous time step. If it's a cold start, the initial knowledge mastery level is set to 0.5; Introducing a time decay function to simulate the forgetting mechanism: If students forget within a time interval... Without targeted practice, their mastery level declined exponentially. ,in To determine the degree attenuation coefficient; After the time prediction is completed, the system analyzes the target knowledge points based on the actual answers. Update its mastery level and fine-tune its related knowledge points. The probability of mastery; based on the updated mastery level of the target knowledge points. Calculate the statistical prediction probability for the target question. The calculation formula is as follows: ; Furthermore, the pass rate for the target questions is calculated based on the remaining student interaction data: ; in, This represents the number of students among the remaining students who answered the target question correctly. This represents the total number of students who have completed the target question; (2) On the semantic track, identify the core knowledge points tested by the target question, and retrieve the timestamp of the student's last answer to a question containing the core knowledge points in the historical interaction sequence. Calculate time With the time interval ; Using a pre-defined piecewise mapping function Based on the time thresholds τ1 and τ2 set by the Ebbinghaus forgetting curve theory, the absolute values ​​of consecutive time intervals are... Transformed into natural language forgetting labels that are easy for large language models to understand : like If ≤τ1, then Mapped to recent semantic labels representing periods of active memory; If τ1 < If ≤τ2, then The mapping is to a medium-term semantic label representing the short-term recession period; like >τ2, then Mapped to dormant semantic labels representing the risk of severe forgetting; like If it does not exist, then The mapping is to the initial semantic label representing the first contact. (3) Traverse time on the correlated quantization orbit. Calculate the question association evidence score for all records in the previous historical interaction sequence that are related to the target question. : ; in, For a moment A collection of previous history questions; The time interval between each historical record and the current moment; The preset time decay coefficient; Weighting coefficients for the questions: If the historical questions and the target questions contain exactly the same core knowledge points, then... Take the first weight value If it contains some of the same knowledge points, then Take the second weight value If there are no identical knowledge points but there are implicit logical connections in the potential relationship network of knowledge points, then Take the third weight value ;in If there are no shared knowledge points and no implicit logical connections exist within the potential relationship network of knowledge points, Set the weight value to 0.

[0010] Furthermore, in step S4, the temperature parameter of the large language model is adaptively adjusted by the dynamically tuned intelligent agent based on the question-related evidence score. The specific calculation method is as follows: The dynamically tuned agent employs an adaptive temperature adjustment algorithm driven by question-related evidence scores, which adjusts the question-related evidence scores... As the sole driving variable, the inference temperature of the real-time mapping large language model is... : ; in, To explore the upper temperature limit, For a strict lower limit temperature, This is the sensitivity coefficient. When At lower levels, Approaching This triggers the exploration mode of the large language model to make cross-knowledge point associations; when At higher levels, Approaching This forces large language models to perform deterministic inferences.

[0011] Furthermore, in step S5, the construction of a recursive dynamic semantic memory stream based on the Markov state transition hypothesis through a logical deduction agent, and the output of the logical prediction probability and the cognitive profile updated at the current moment under the control of the temperature parameter, specifically includes: (1) Constructing contextual input: at time When making predictions, the input context of a large language model is strictly limited to: its own context. The student cognitive profile is generated in real time and knowledge mastery Target question characteristics, natural language forgetting labels The characteristics of the target questions include the knowledge points associated with the target questions and the pass rate of the questions. ; (2) Execution of rule-guided thought chain reasoning: at the reasoning temperature T t Under the constraints of [unclear], the large language model is guided to perform thought chain reasoning. The large language model is first based on forgetting labels. and knowledge mastery right The semantic cognitive profile is calibrated and used to guide the large language model to consider the question pass rate, outputting the logical prediction probability for the target question. When making the first prediction, the output question pass rate is the logistic prediction probability; (3) Iterative update of cognitive profile: Based on the inference logic of this time, the large language model outputs the changes in cognitive state as the updated cognitive profile at the current moment. The aforementioned As a closed-loop propagation of the student's state variable to time... In the prediction task of +1, we achieve long sequence learning state tracking with low resource consumption.

[0012] Furthermore, in step S6, the adaptive gating fusion mechanism based on question-related evidence scores adaptively weights and fuses the statistical prediction probability and the logical prediction probability through a decision fusion agent. The specific calculation method is as follows: (1) Generate confidence weights: The decision fusion agent will associate the evidence related to the question. As the sole variable driving the system to determine the confidence weights, the sigmoid gating function is used to map them to the confidence weights of the statistical model. : ; in, This is a preset cognitive switching threshold. When much smaller (i.e., during a cold start or after a long period of no review) confidence weight When the value approaches 0, the system's control is handed over to the logical reasoning of the large language model; when Much larger At that time, confidence weight Approaching 1, the system's dominance is handed over to the Bayesian knowledge tracing model, which combines graph and temporal reinforcement; simultaneously, to preserve the logical correction space of the large language model, [the following is omitted as the text is incomplete and requires further context]. Set upper safety limit ; (2) Weighted probability synthesis: using the calculated dynamic confidence weights A linear synthesis is performed on the dual-track prediction probabilities to obtain the final learning performance prediction probability. : ;

[0013] A dynamic learning performance prediction system based on a large language model and multi-agent collaboration includes: (1) Data time-series processing module, used to sort and serialize the acquired students' historical answer records by timestamp, and construct a standardized interactive time series; (2) The graph construction agent module is used to extract the implicit teaching logic relationship and association strength weight between knowledge points by calling the large language model based on the knowledge point name identifiers of all questions in the dataset, construct a directed weighted graph, and generate a potential relationship network of knowledge points; (3) Statistical analysis intelligent agent module, used to perform three-track feature processing based on the standardized interaction time series, run the graph and time-series augmented Bayesian knowledge tracing model on the numerical track, output statistical prediction probability and its corresponding knowledge mastery degree, and calculate the pass rate of the target question based on the remaining student interaction data. On the semantic track, the time interval between historical answers and the target time is calculated, and the forgetting curve is combined to segment and map it into natural language forgetting labels; on the association quantification track, the question association evidence score is calculated based on the knowledge point association weight and the time decay effect. (4) Dynamic parameter tuning agent module, which is used to dynamically adjust the temperature parameters of the large language model for logical deduction based on the evidence score associated with the current question and the adaptive reasoning temperature control algorithm. (5) Logical reasoning agent module, used to construct a recursive dynamic semantic memory stream based on the Markov state transition hypothesis. Under the control of the temperature parameter, it takes the cognitive profile generated by itself at the previous moment, the knowledge mastery, the target question information and the natural language forgetting label as the input context of the large language model. Through the execution of rule-guided thinking chain reasoning, it outputs the logical prediction probability and the cognitive profile updated at the current moment. (6) Decision fusion intelligent agent module, which uses the topic-related evidence score as the only driving variable, maps it to the confidence weight of the statistical model through the Sigmoid gating function, and performs adaptive weighted fusion of the statistical prediction probability and the logical prediction probability to obtain the final learning performance prediction probability.

[0014] A terminal device includes a memory and a processor. The memory stores a computer program that can run on the processor. When the processor loads and executes the computer program, it employs a large language model and a multi-agent collaborative learning performance prediction method.

[0015] A computer-readable storage medium storing a computer program that, when loaded and executed by a processor, employs a large language model and a multi-agent collaborative learning performance prediction method.

[0016] The beneficial effects of this invention are: (1) Compared with traditional knowledge tracing, this application utilizes a large language model and multi-agent collaborative mechanism to reconstruct the semantic network of knowledge points in textless scenarios through offline graph construction. Combined with spatiotemporal effective evidence accumulation and dynamic temperature adjustment, it can mine the correlation between knowledge points and the information on changes in students' potential cognitive states in learning data, and obtain more accurate information on learners' mastery of knowledge points. This provides learners and smart teaching platforms with a transparent and interpretable student performance prediction framework, and provides a reliable decision-making basis for subsequent personalized path recommendation and precise teaching intervention. (2) Existing knowledge tracing models tend to degenerate into random guessing when faced with newly enrolled students or entirely new knowledge points due to a lack of sufficient response records. This application constructs a latent relationship network of knowledge points, providing implicit logical anchors for cross-knowledge point transfer when interaction records are sparse or lack specific question texts. At the same time, it deeply integrates the semantic reasoning of traditional knowledge tracing models with that of large language models, and combines it with the adaptive temperature adjustment mechanism of the statistical analysis agent's dynamic parameter tuning agent, enabling the system to adaptively predict student learning performance based on historical data density, thus greatly alleviating the problems of data sparsity and cold start. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the learning performance prediction method of the large language model and multi-agent collaboration of the present invention. Detailed Implementation

[0018] Methods for predicting the learning performance of large language models and multi-agent collaborative learning, such as... Figure 1 As shown, the present invention includes: S1. Data time-series processing yields a standardized interaction time series, restricting the system to only acquire interaction records before the predicted target time. In dynamic knowledge tracing tasks, learners' cognitive states evolve non-linearly over time, so it is essential to strictly ensure the temporal order of data input. S11. The system first retrieves raw log data containing student interaction behavior from the online education platform or learning management system. Each interaction record is structured and parsed into a four-tuple. These represent the student identifier, question identifier, answer score, and answer timestamp, respectively. S12, Targeting a specific student All its historical answer records are sorted by timestamp. The order of events is strictly ordered in ascending order to form a standardized interaction time series. . S13. When the system is triggered by an instruction, it needs to predict the student's time. When processing the target question, the data preprocessing module will perform a truncation operation on the sequence, forcibly masking the time points. All subsequent interaction data, including the time window, are restricted to access by the system and subsequent agents. A subset of sequences within.

[0019] S2. Constructing the potential relationships of knowledge points for intelligent agents using graph construction; In many real-world educational scenarios, due to privacy concerns or data availability limitations, systems often only obtain abstract knowledge point names, not specific question texts. To bridge this semantic gap, the system deploys a graph-based intelligent agent during the offline initialization phase. Specifically, this step includes: S21. In scenarios where specific question text content is missing, input all knowledge point names involved in the dataset as prompt words into the pre-trained large language model. S22. Large-Scale Model Pre-Training Knowledge Extraction: The system extracts a list of knowledge point names corresponding to the entire question bank and constructs structured prompts, which are then input into the pre-trained large-scale language model. The large-scale language model is required to output the implicit teaching logic relationship between any two knowledge points based on its inherent subject-specific general knowledge. The logical relationship is primarily extracted in two categories: first, prerequisite and pre-learning relationships, representing the sequential order of learning paths; and second, similarity of thought relationships, representing the lateral transferability of problem-solving skills and cognitive logic. S23. Directed Weighted Graph Generation: Analyze the text output of a large language model and construct a latent relational network graph of knowledge points. The set of nodes Includes all core knowledge points, edge set Includes connections between knowledge points. For nodes that are logically related... and Edge weights are assigned based on the association strength evaluated by the large language model. .

[0020] S3, Statistical Analysis of Three-Track Feature Processing of Intelligent Agents; The pure numerical values ​​output by traditional knowledge tracing models are difficult for large language models to understand directly, while relying solely on large language models lacks sensitivity to absolute time values ​​and objective statistics. Therefore, a statistical analysis agent performs three-track feature processing on a subset of the input sequence, considering numerical, semantic, and relational quantization tracks: S31. Construct a Bayesian knowledge tracing model with graph and temporal enhancement on the numerical orbit: Traditional BKT models treat knowledge points as isolated Hidden Markov chains and ignore the forgetting effect over time. This embodiment significantly improves upon these models: First, initialize the parameters by setting the prior mastery probability for each knowledge point. guessing rate With error rate ; Secondly, the time series and graph are updated in a linked manner. As the time series inputs the answer records, the target knowledge points from the previous time step are received. degree of mastery If it's a cold start, the initial knowledge mastery level is set to 0.5; A time decay function is introduced to simulate the forgetting mechanism: if students do not engage in targeted practice within a time interval, their mastery decays exponentially, i.e. ,in To determine the degree attenuation coefficient. At the target time Output statistical prediction probability based on the latest level of knowledge. : ; Once the time prediction is complete, the system will receive the student's time. The actual answer to the target question This triggers an update to knowledge mastery. First, update the knowledge points using the standard Bayesian posterior formula. Current level of control And calculate the increase in mastery caused by this response. ; Based on the potential relationship network of the knowledge points, the system retrieves information related to... Knowledge points related to the existence of directly related edges And utilize the associated edge weights As a step size adjustment factor, synchronization The level of mastery has been slightly updated: ; in, To update and adjust parameters accordingly; Furthermore, the pass rate for the target questions is calculated based on the remaining student interaction data: ; in, This represents the number of students among the remaining students who answered the target question correctly. This represents the total number of students who have completed the target question; Finally, Pass rate of questions at any time and knowledge mastery Extracted as an objective statistical benchmark for large-scale model reasoning in subsequent logical deduction agents. S32. Perform temporal feature mapping on the semantic track; Large language models lack intuitive understanding of time differences in numerical form, therefore this embodiment adopts temporal semantic engineering. Determine the time The core knowledge points tested in the target questions The system in the sequence A reverse search reveals that the student's most recent answer contained this knowledge point. Record timestamp Calculate the absolute value of the time interval. Subsequently, a piecewise mapping function based on the Ebbinghaus forgetting curve was used. , will the value Transform into natural language forgetting labels In a specific implementation scenario, time thresholds are set as τ1 = 3 days and τ2 = 15 days: (1) If If ≤τ1, then The mapping is "[Memory state: recent active period, knowledge points fully consolidated]"; (2) If τ1 < If ≤τ2, then The mapping is "[Memory state: short-term decline, with some signs of forgetting]"; (3) If >τ2, then Mapped to "[Memory state: Forgetting period, facing serious risk of forgetting]"; (4) If If it does not exist, i.e., in a cold start state, then The mapping is "[Memory state: initial state, first time encountering this test point]". S33. Calculate the relevant evidence score for the question on the correlation quantification track; Statistical analysis is performed on all records in the historical interaction sequence before time t that are related to the knowledge points of the target question, taking into account spatial association weights and time decay effects, to quantify the comprehensive score of the currently available evidence: ; in, For a moment A collection of previous history questions; The time interval between each historical record and the current moment; The preset time decay coefficient; Spatial association weights are determined based on the potential relationship network of knowledge points: if a historical question and a target question contain exactly the same core knowledge points, then... Take the highest weight value (e.g., 1.0); if it contains some of the same knowledge points, then Take the second weight value (e.g., 0.3); if there are no identical knowledge points but there are implicit logical connections in the potential relationship network of knowledge points, then the third weight is taken. value (For example, 0.2); If there are no identical knowledge points and no implicit logical connections exist in the potential relationship network of knowledge points, Set the weight value to 0.

[0021] S4. Adaptive parameter adjustment of a dynamically tuned intelligent agent; In dynamic sequence prediction, the uncertainty of the system regarding the student's cognitive state is not monotonically decreasing; when related exercises are forgotten due to the passage of time, the uncertainty increases again. Therefore, the dynamically tuned agent abandons the reliance solely on the time step. The monotonic adjustment strategy proposes a method based on question-related evidence. Cognitive entropy dynamic adjustment algorithm: Specifically, the dynamically tuned agent employs an adaptive temperature-tuning algorithm driven by question-related evidence scores, adjusting the generated temperature parameters of the large language model. and Establish a nonlinear inverse proportional mapping relationship to link the questions with evidence. As the sole driving variable, the inference temperature of the real-time mapping large language model is... : ; in, To preset the upper limit temperature, To preset the lower limit temperature, This is the sensitivity coefficient; when At lower levels (such as during a cold start, or when students have not been exposed to this type of question for a long time). Approaching This triggers the exploration mode of the large language model, utilizing external general knowledge to make cross-knowledge point associations; when When the level is high (e.g., due to recent intensive practice). Approaching This forces large language models to strictly follow recent factual profiles for reasoning.

[0022] S5, recursive dynamic semantic memory stream of logical deduction agent; To address the issues of context window overflow and forgetting that arise when large language models process extremely long learning trajectories, the logical inference agent abandons the traditional full history splicing method and innovatively introduces the idea of ​​recursive dynamic semantic memory flow based on the Markov state transition hypothesis. S51. Constructing contextual input: at time... When making predictions, the input contextual cue words for the large language model are highly compressed and strictly limited to four elements: First, the cognitive profile of the previous moment. : A highly generalized natural language text generated by the large language model at time t−1, which records the student's learning and cognitive level up to the previous step; Secondly, knowledge mastery : Receives the objective knowledge mastery value of the statistical analysis agent; Third, target question characteristic information: including the knowledge points related to the question and the pass rate of the question. ; Fourth, Natural Language Forgetting Labels : Received from statistical analysis agent. S52, Execution rule-guided thought chain reasoning: at the current temperature Under control, the large language model is guided by the system's preset prompt word framework and performs reasoning according to predetermined logic: if the prediction is made on the first attempt, the large language model is required to directly output the question pass rate as the logical prediction probability. ; If it is not the first prediction, the large language model is required to first read the natural language forgetting labels. Calibrate according to label First, the student profiles recorded in the database; second, the calibrated semantic cognitive profiles are compared with the input knowledge mastery levels. Cross-comparison and calibration are performed; logical deduction is conducted in conjunction with the knowledge points being tested, and the results are calibrated using the pass rate of the questions to output the probability of logical prediction. . S53. Generate and transmit updated cognitive profiles: Based on the logical conclusions of this deduction, the large language model semantically modifies the profile from the previous time step, generating an updated cognitive profile for the current time step. (For example, a new diagnostic annotation could be added: "Although this student has a good foundation in algebra, their performance has shown signs of decline due to a lack of practice.") It will be used as a short-memory text variable, flowing through the closed loop to the prediction task at time t+1, thus completing dynamic knowledge tracking.

[0023] S6. Adaptive gating fusion of decision fusion agents based on question-related evidence scores; Because large language models and traditional statistical models each have their own advantages in different scenarios, the decision fusion agent is responsible for... and The final decision is made between them. This embodiment employs an adaptive gating fusion mechanism based on the issue-related evidence score to adaptively weight and fuse the statistical prediction probability and the logical prediction probability; S61. Generate Confidence Weights: To avoid interference from statistical models in scenarios with very few samples or long learning intervals, this embodiment will use the question-related evidence scores calculated in step S3. As the sole variable determining the probability confidence weights, the confidence weights of the statistical model are calculated using a variant of the sigmoid function. : ; in, The preset cognitive switching threshold (e.g., set to 3.0). When much smaller (i.e., during a cold start or after a long period of no review) confidence weight When the value approaches 0, the system's control is handed over to the logical reasoning of the large language model; when Much larger At that time, confidence weight When the value approaches 1, the system's dominance is handed over to the Bayesian knowledge tracing model with graph and temporal enhancement. S62. To preserve the logical correction space of the large language model, for Set upper safety limit ,Right now This is to ensure that at least a certain proportion of the weight is reserved for the large language model so that it can play its corrective role. S63, Weighted Probability Synthesis: Finally, the decision fusion agent utilizes the calculated dynamic confidence weights. A linear synthesis is performed on the dual-track prediction probabilities to obtain the final learning performance prediction probability. : ;

[0024] A large language model and multi-agent collaborative learning performance prediction system includes:

[0025] The data time-series processing module is used to acquire students' historical interaction logs in real time, perform strict ascending sorting based on the timestamp field of each log, generate a standardized time series of student interactions, and use it as the original data source for downstream modules.

[0026] The graph construction agent module is used to construct implicit knowledge networks. Upon receiving a set of knowledge points containing only abstract names, this module constructs specific structured cue words and invokes a large language model to extract the latent relationships between the knowledge points. Subsequently, the module transforms the natural language relationships and their confidence levels output by the large language model into nodes, edges, and weights in a directed weighted graph, generating and solidifying the latent relationship network of the knowledge points.

[0027] The statistical analysis agent module is used to perform feature processing based on the standardized interaction time series, run a graph-enhanced and time-series-enhanced Bayesian knowledge tracing model on a numerical track, output statistical prediction probabilities and their corresponding knowledge mastery levels, and calculate the pass rate of the target questions based on the remaining student interaction data. On the semantic track, the time interval between historical answers and the current target time is calculated and mapped to a natural language forgetting label using the Ebbinghaus forgetting curve. On the correlation quantification track, the spatial correlation weight and time decay effect of historical records are considered together to calculate and output the question correlation evidence score that represents the current evidence density.

[0028] The dynamic parameter tuning agent module is used to receive the question-related evidence score calculated by the statistical analysis agent in real time, and dynamically calculate the inference temperature parameter of the large language model through the exponential decay mapping function.

[0029] The logical reasoning agent module is used to construct a recursive dynamic semantic memory stream based on the Markov state transition hypothesis. Under the control of the temperature parameter, it shields the input of the entire historical sequence and only uses the text-type cognitive profile generated in the previous moment, the knowledge mastery level, the target question features, and the natural language forgetting label as the input context of the large language model. Through the execution of rule-guided thought chain reasoning, it synchronously outputs the logical prediction probability and the cognitive profile updated at the current moment.

[0030] The decision fusion intelligent agent module is used to implement an adaptive gating fusion mechanism based on the question-related evidence score. It maps the question-related evidence score to the confidence weight of the statistical model through the Sigmoid gating function and performs truncation restriction within a set safety upper bound. Finally, it performs adaptive linear weighted fusion of the statistical prediction probability and the logical prediction probability based on the confidence weight, and outputs the student's final learning performance prediction probability at that moment.

[0031] This application also discloses a terminal device, including a memory and a processor. The memory stores a computer program that can run on the processor. When the processor loads and executes the computer program, it adopts a learning performance prediction method that combines a large language model with multi-agent collaboration.

[0032] The terminal device can be a computer device such as a desktop computer, a laptop computer, or a cloud server. The terminal device includes, but is not limited to, a processor and a memory. For example, the terminal device may also include input / output devices, network access devices, and buses.

[0033] The processor can be a central processing unit (CPU). Of course, depending on the actual use, it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc., and this application does not limit it.

[0034] The memory can be an internal storage unit of the terminal device, such as a hard disk or RAM of the terminal device, or an external storage device of the terminal device, such as a plug-in hard disk, smart memory card (SMC), secure digital card (SD), or flash memory card (FC) equipped on the terminal device. Furthermore, the memory can be a combination of internal storage units and external storage devices of the terminal device. The memory is used to store computer programs and other programs and data required by the terminal device. The memory can also be used to temporarily store data that has been output or will be output. This application does not limit this.

[0035] In this terminal device, the large language model and the multi-agent collaborative learning performance prediction method described in the above embodiments are stored in the memory of the terminal device and loaded and executed on the processor of the terminal device for convenient use.

[0036] This application also discloses a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it employs the learning performance prediction method of large language model and multi-agent collaboration described in the above embodiments.

[0037] The computer program can be stored in a computer-readable medium. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or certain middleware. The computer-readable medium includes any entity or device capable of carrying computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the computer-readable medium includes, but is not limited to, the above-mentioned components.

[0038] The large language model and multi-agent collaborative learning performance prediction method described in the above embodiments are stored in the computer-readable storage medium and loaded and executed on the processor to facilitate the storage and application of the above method.

[0039] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of protection of this application is limited to these examples; under the concept of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of this application as described above, which are not provided in detail for the sake of brevity.

[0040] One or more embodiments in this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments in this application should be included within the protection scope of this application.

Claims

1. A method for predicting learning performance using a large language model and multi-agent collaboration, characterized in that, The method includes the following steps: S1. Obtain students' historical interaction records and strictly sort them into a time series according to timestamps to ensure that the predicted time is accurate. When performing a function, the system only obtains the time. and the previous record sequence; S2. By calling the large language model through "graph construction agent", the abstract knowledge point names are transformed into a potential relationship network of knowledge points containing multi-dimensional teaching logical relationships and weights; S3. The time series is processed by a "statistical analysis agent" to perform three-track parallel feature processing: a Bayesian knowledge tracing model with graph and time-series enhancement is run on the numerical track to output statistical prediction probabilities and knowledge mastery levels, and the pass rate of the target questions is calculated based on the remaining student interaction data. ; Calculate the response time interval on the semantic track and map it to a natural language forgetting label; On the correlation quantification track, take into account the correlation weight of historical records and the time decay effect, calculate and output the question correlation evidence score; S4. The "dynamic parameter tuning agent" directly receives the question-related evidence score, and adaptively calculates and adjusts the temperature parameters of the large language model for logical deduction based on the question-related evidence score. S5. Construct a recursive dynamic semantic memory stream based on the Markov state transition hypothesis using a "logic deduction agent," and under the control of the temperature parameter, record the time... The student's cognitive profile, target question characteristics, knowledge mastery, and the aforementioned natural language forgetting label are input into the large language model, which simultaneously outputs the logical prediction probability and the updated time. Student cognitive profile; S6. Based on the question-related evidence score obtained in step S3, the "decision fusion agent" is used to dynamically weight and fuse the statistical prediction probability and the logical prediction probability using an adaptive gating fusion mechanism based on the question-related evidence score, and the final student learning performance prediction result is output.

2. The method according to claim 1, characterized in that, In step S1, the student's historical interaction records are obtained and strictly sorted into a time series according to timestamps to ensure that the predicted time is accurate. When performing a function, the system only obtains the time. and the preceding record sequence, including the following steps: S11. The system retrieves raw log data containing student interaction behavior from the online education platform or learning management system. Each interaction record is structured and parsed into a four-tuple. These represent the student identifier, question identifier, answer score, and answer timestamp, respectively. S12, Targeting a specific student All its historical answer records are sorted by timestamp. The order of events is strictly ordered in ascending order to form a standardized interaction time series. ; S13. When the system is triggered by an instruction, it needs to predict the student's time. When processing the target question, the data preprocessing module will perform a truncation operation on the sequence and mask the timing. and all subsequent interaction data.

3. The method according to claim 1, characterized in that, In step S2, the operation mode of the "graph construction agent" specifically includes: S21. In scenarios where specific question text content is missing, input all knowledge point names involved in the dataset as prompt words into the pre-trained large language model. S22. Mine out and extract the implicit logical relationships between knowledge points. The logical relationships include at least the prerequisite relationships that represent the pre-constraints of the learning path and the similarity relationships that represent the transferability of problem-solving thinking. S23. Construct a directed weighted graph, with knowledge points as nodes and the logical relationships and association strengths as weighted edges, to generate the potential relationship network of the knowledge points.

4. The method according to claim 1, characterized in that, In step S3, the processing of the "statistical analysis agent" specifically includes: S31. On the numerical trajectory, construct a Bayesian knowledge tracing model with graph and temporal enhancement, and set a prior mastery probability for each knowledge point. guessing rate With error rate ; Based on the sequential answer record, receive the target knowledge point from the previous moment. degree of mastery If it's a cold start, the initial knowledge mastery level is set to 0.5; Introducing a time decay function to simulate the forgetting mechanism: If students forget within a time interval... Without targeted practice, their mastery level declined exponentially. ,in To determine the degree attenuation coefficient; Extracting knowledge mastery This serves as an anchor point for the knowledge mastery of the intelligent agent in the next step of logical deduction. Output target time The statistical prediction probability, i.e. ; Once the time prediction is complete, the system will receive the student's time. The actual answer to the target question This triggers an update to knowledge mastery. First, update the knowledge points using the standard Bayesian posterior formula. Current level of control And calculate the increase in mastery caused by this response. ; Based on the potential relationship network of the knowledge points, the system retrieves information related to the knowledge points. Related knowledge points And utilize the associated edge weights As a step size adjustment factor, synchronization degree of mastery Minor updates: ； in, To update and adjust parameters accordingly; Furthermore, the pass rate for the target questions is calculated based on the remaining student interaction data: ； in, This represents the number of students among the remaining students who answered the target question correctly. This represents the total number of students who have completed the target question; Question pass rate and knowledge mastery Together, we provide an objective statistical benchmark for the inference of large language models in logical inference agents; S32. On the semantic track, identify the core knowledge points tested by the target question, and retrieve the timestamp of the student's last answer to a question containing that knowledge point from the historical sequence. and calculate the time interval. ; Using a pre-defined piecewise mapping function, combined with the threshold range of the Ebbinghaus forgetting curve theory, the numerical values ​​of consecutive time intervals are... The forgetting labels are mapped to natural language labels that are easy for large language models to understand. These labels include periods of active memory, short-term decay, severe forgetting, and first exposure. S33. Traverse the time intervals along the correlated quantization track. All records in the previous historical interaction sequence that have the same or related knowledge points as the target question are considered, taking into account the historical question association weights and time decay effects, to quantify the evidence score related to the current question. : ； in, For a moment A collection of previous history questions; The time interval between each historical record and the current moment; The preset time decay coefficient; The association weights are determined based on the potential relationship network of knowledge points: if a historical question and a target question contain exactly the same core knowledge points, then W... i Take the highest weight value If it contains some of the same knowledge points, then Take the second weight value If there are no identical knowledge points but there are implicit logical connections in the potential relationship network of knowledge points, then Take the third weight value If there are no identical knowledge points and no implicit logical connections in the potential relationship network of knowledge points, Set the weight value to 0.

5. The method according to claim 1, characterized in that, In step S4, the "dynamic parameter tuning agent" adopts a dynamic temperature adjustment strategy based on the question-related evidence score, specifically including: An adaptive temperature adjustment algorithm driven by question-related evidence scores is adopted to adjust the temperature parameters generated by the large language model. and Establish a nonlinear inverse proportional mapping relationship to link the questions with evidence. As the sole driving variable, the dynamic output of the inference temperature of the large language model is determined. : ； in, To explore the upper temperature limit, For a strict lower limit temperature, η is the sensitivity coefficient; when At lower levels, Approaching This triggers the exploration mode of the large language model, utilizing external general knowledge to make cross-knowledge point associations; when At higher levels, Approaching This forces large language models to strictly follow recent factual profiles for reasoning.

6. The method according to claim 1, characterized in that, In step S5, the dynamic semantic memory stream constructed by the "logic deduction agent" specifically includes: S51, at time When making predictions, the input context of a large language model is limited to only the context in which it exists. The student cognitive profile is generated in real time and knowledge mastery Target question characteristics, natural language forgetting labels The target question characteristics include the knowledge points related to the question and the pass rate of the question. ; S52. Applying a rule-guided thinking chain mechanism, the large language model is forced to first verify whether the system is in a cold start state during reasoning. In the cold start state, the question pass rate is directly output as the logical prediction probability. In non-cold start scenarios, the attenuation penalty caused by the natural language forgetting label is considered first, followed by logical deduction based on the knowledge points being examined. The result is then calibrated using the question pass rate, and the logical prediction probability is output. ; S53. After outputting the logical prediction probability, the large language model outputs the state change of this reasoning as text, and outputs the student's cognitive profile at the current moment. It is passed as a state variable in a closed loop to time 1. The prediction task.

7. The method according to claim 1, characterized in that, In step S6, the adaptive weighted fusion of the "decision fusion intelligent agent" specifically includes: S61. Adaptive gating fusion mechanism based on question-related evidence scores: [This mechanism] integrates question-related evidence scores... As the sole variable determining the confidence weights, a variant of the Sigmoid function is used to calculate the confidence weights of the statistical model. The calculation formula is: ； in, The preset cognitive switching threshold; S62, to Set upper safety limit ,Right now This ensures that at least a certain percentage of the weight is reserved for the large language model, so that it can play a corrective role. S63. The decision fusion agent utilizes the calculated dynamic confidence weights. A linear synthesis is performed on the two prediction probabilities to obtain the final learning performance prediction probability. The calculation formula is: 。 8. A learning performance prediction system based on a large language model and multi-agent collaboration, characterized in that, include: The data time-series processing module is used to acquire students' historical answer records, sort the answer records by time into a standardized interaction time series, and ensure that the data is processed at the predicted time. When performing a function, the system only obtains the time. and the previous record sequence; The graph construction agent module is used to call large language models offline, extract the implicit teaching logic relationships and correlation strength between knowledge points, and construct a directed weighted knowledge point potential relationship network. The statistical analysis agent module is used to perform feature processing based on the standardized interaction time series, run a graph-enhanced and time-series-enhanced Bayesian knowledge tracing model on a numerical track, output statistical prediction probabilities and knowledge mastery levels, and calculate the pass rate of the target questions based on the remaining student interaction data. On the semantic track, the time interval between historical answers and the current target moment is calculated and mapped to a natural language forgetting label using the Ebbinghaus forgetting curve; on the association quantification track, the association weights of historical questions and the time decay effect are considered together to calculate and output the question association evidence score at the current moment. The dynamic parameter tuning agent module is used to adaptively adjust the inference temperature parameter of the large language model by using the evidence score associated with the question as the sole driving variable and through an exponential mapping function. The logical reasoning agent module is used to construct a recursive dynamic semantic memory stream based on the Markov state transition hypothesis. Under the control of the temperature parameter, it only takes the cognitive profile of the previous moment, the knowledge mastery, the target question features, and the natural language forgetting label as the input context of the large language model. Through rule-guided thought chain reasoning, it outputs the logical prediction probability and the cognitive profile updated at the current moment. The decision fusion intelligent agent module is used for an adaptive gating fusion mechanism based on the evidence score associated with the question. It adaptively weights and fuses statistical prediction probability and logical prediction probability to output the final student performance prediction.

9. A terminal device, comprising a memory and a processor, characterized in that, The memory stores a computer program that can run on a processor, and when the processor loads and executes the computer program, it employs the method described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is loaded and executed by the processor, it employs the method described in any one of claims 1 to 7.

Images