Method for dynamically generating and pushing personalized auxiliary teaching content based on AI large model

Abstract

The application discloses an AI large model-based personalized auxiliary teaching content dynamic generation and pushing method, and belongs to the technical field of data processing, and comprises the following steps: acquiring the conditional probability distribution of an AI large model output layer and constructing a benchmark probability distribution vector, acquiring a generation probability distribution feature vector in a subsequent generation round and performing distribution similarity analysis to determine a generation probability distribution drift state; performing controlled adjustment on the generation parameter set in the drift state, constructing a stable window in a continuous non-drift state, and dynamically updating the benchmark probability distribution vector. Teaching content is generated based on the adjusted generation parameter set and is pushed in batches, generation stability monitoring is performed in the pushing process, pushing constraint adjustment is performed according to the shock state and the convergence state, and finally, the converged generation parameter set is saved as an initial generation strategy corresponding to a teaching task, thereby solving the problem of implicit degradation of the teaching content generation quality in the prior art.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method for dynamically generating and pushing personalized auxiliary teaching content based on AI large models. Background Technology

[0002] With the rapid development of AI big data models, they have been applied in numerous fields, such as personalized learning assistance. Online education platforms can use AI big data models to recommend personalized supplementary teaching content to learners, helping them to better engage in self-study. In personalized learning assistance scenarios, the generation of teaching content based on AI big data models is usually not a one-time event, but rather a continuous and multi-round process that dynamically generates corresponding teaching content units as the teaching progresses, learning feedback is received, and teaching tasks are broken down. These units are then pushed to learning terminals in batches according to a preset rhythm. In this continuous generation and push scenario, the AI ​​big data model continuously superimposes previous generation results as context input in multiple generation rounds, and the conditional probability distribution of its model output layer will continuously evolve with the accumulation of context.

[0003] However, most existing technologies only provide static or empirical settings for a single generation process, lacking a systematic monitoring and constraint mechanism for the changing trends of the conditional probability distribution within the model across multiple generation cycles. Especially in continuous call scenarios, even if a single generation result does not show obvious anomalies, the conditional probability distribution within the model may still experience implicit drift over multiple generation cycles, gradually deviating from its intended path. Such implicit drifts are often difficult to detect in a timely manner, and once the drift accumulates to a certain extent, it will lead to a gradual degradation of the generated content in terms of style consistency, difficulty gradient, expression structure, and teaching adaptability. Furthermore, in application scenarios combined with teaching content delivery mechanisms, frequent or unconstrained adjustments to generation parameters may cause engineering problems such as generation process oscillations, convergence failures, and disordered delivery rhythms, thereby affecting the stability and controllability of the overall teaching experience.

[0004] For example, Chinese invention patent CN114741583B discloses an information recommendation method, apparatus, and electronic device based on artificial intelligence, which includes: acquiring ranking features of information to be recommended; performing mapping processing on the ranking features based on multiple recommendation tasks to obtain a first feature mapping result corresponding to each recommendation task; performing mapping processing on the ranking features based on multiple feature dimensions to obtain a second feature mapping result corresponding to each feature dimension; taking any recommendation task as the target recommendation task, performing attention processing based on the first feature mapping result and the second feature mapping result corresponding to each feature dimension to obtain a recommendation index for the target recommendation task corresponding to the information to be recommended; and performing a recommendation operation based on the recommendation index for each recommendation task corresponding to the information to be recommended.

[0005] For example, Chinese invention patent CN114329175B discloses an information recommendation method, system, storage medium, and server, which includes: outputting an information recommendation list corresponding to an information access request through a pre-trained information recommendation model. During the pre-training process, the information recommendation model can select a portion of sample information through a teacher filtering submodule and a student filtering submodule to filter out training samples that interfere with the training of the teacher and student submodules, making the training of the information recommendation model more accurate. Furthermore, since the teacher and student submodules process less sample information, the training time for the information recommendation model is less.

[0006] The above-mentioned technology has at least the following technical problems:

[0007] Existing methods for generating personalized teaching content based on large AI models can lead to imperceptible probability distribution drift in application scenarios with multiple rounds of generation and delivery. This causes the generation strategy to gradually deviate without explicit abnormal feedback, which can easily lead to oscillations in the generation process, convergence failures, and loss of control over the delivery rhythm, resulting in a hidden degradation in the quality of generated teaching content. Summary of the Invention

[0008] To address the issue of implicit degradation in the quality of generated teaching content in existing technologies, this invention provides a method for dynamically generating and pushing personalized auxiliary teaching content based on a large AI model. The technical solution is as follows:

[0009] A method for dynamically generating and pushing personalized auxiliary teaching content based on an AI large-scale model includes the following steps: initializing and configuring the AI ​​large-scale model and executing the first round of content generation; collecting the conditional probability distribution of the output layer of the AI ​​large-scale model during the first round of generation and using it as the baseline probability distribution vector; summarizing the conditional probability distribution of the model output layer of the AI ​​large-scale model during subsequent rounds of content generation, and judging it in conjunction with the baseline probability distribution vector to obtain the generation probability distribution drift state of the current round, which includes drift state and non-drift state; when the generation probability distribution drift state is a drift state, the controlled adjustment processing of the generation parameter set is triggered; otherwise, the controlled adjustment processing is not executed; when it is detected that the generation process of consecutive preset rounds is in a non-drift state, A stable window is constructed, and the baseline probability distribution vector is dynamically updated within the stable window until the updated baseline probability distribution vector converges, thus obtaining the adjusted generation parameter set. Otherwise, the update of the baseline probability distribution vector is not performed. The generation parameter set includes randomness parameters and / or candidate space constraint parameters and / or repetition penalty parameters. Based on the adjusted generation parameter set, a large AI model is invoked to dynamically generate personalized auxiliary teaching content units, which are then pushed in batches according to preset push rules, while recording the corresponding push execution logs. During the push process, generation stability monitoring is performed to obtain stability monitoring results. When the stability monitoring results are in an oscillating state and / or a non-converged state, the push constraints are adjusted; otherwise, the adjusted generation parameter set is updated to the initial generation strategy of the corresponding teaching task.

[0010] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0011] 1. The method for dynamic generation and delivery of personalized auxiliary teaching content based on AI large model provided by this invention collects the conditional probability distribution of the output layer of AI large model and constructs a benchmark probability distribution vector. In subsequent generation processes, the similarity analysis is performed based on the feature vector of the generated probability distribution and the benchmark probability distribution vector to determine the drift state of the generated probability distribution in real time. In the drift state, the controlled adjustment of the generation parameter set is triggered, thereby realizing continuous control of the overall probability distribution stability of the teaching content generation process and effectively solving the problem of implicit degradation of the quality of teaching content generation in the prior art.

[0012] 2. This invention constructs a stable window when the probability distribution drift state is in a non-drift state for consecutive preset rounds, and performs a moving average processing on the feature vector of the generated probability distribution within the stable window to dynamically update the benchmark probability distribution vector. This allows the benchmark probability distribution vector to gradually and adaptively evolve with the generation process, thereby realizing the dynamic calibration of the generation probability distribution judgment benchmark, avoiding the misjudgment problem caused by a fixed benchmark, and improving the accuracy and robustness of the generation probability distribution drift judgment.

[0013] 3. This invention performs generation stability monitoring during the batch delivery of teaching content, determines whether the generation process is in an oscillating or convergent state based on the generation parameter set, and analyzes the oscillation intensity value and divergence intensity value when an oscillating state and / or non-convergent state are detected, thereby realizing the analysis of the stability of the generation parameter set adjustment behavior and avoiding the problem of frequent reverse adjustment of the generation parameter set during the delivery stage.

[0014] 4. This invention obtains the number of generation parameter sets that can be adjusted within a single cycle, the adjustment range limit, and the push interval by jointly analyzing the oscillation intensity value and the divergence intensity value. During the push process, a push constraint condition is introduced to prohibit the execution of opposite adjustment operations on the same generation parameter set within two consecutive generation cycles. This achieves coordinated control of the generation parameter set adjustment intensity and the push rhythm. Furthermore, after the generation process reaches a convergence state, the current generation parameter set is solidified as the initial generation strategy for the corresponding teaching task, improving the consistency and reusability of personalized auxiliary teaching content in the long-term generation and push process. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating the method for dynamically generating and pushing personalized auxiliary teaching content based on a large AI model, as provided in this application embodiment;

[0017] Figure 2 A flowchart illustrating the overall steps of the method for dynamically generating and pushing personalized auxiliary teaching content based on a large AI model, as provided in the embodiments of this application.

[0018] Figure 3 A flowchart illustrating the generation probability distribution drift determination and controlled adjustment process of the method for dynamically generating and pushing personalized auxiliary teaching content based on a large AI model, as provided in the embodiments of this application.

[0019] Figure 4 A schematic diagram illustrating the generation probability distribution drift of the method for dynamically generating and pushing personalized auxiliary teaching content based on a large AI model, provided in the embodiments of this application.

[0020] Figure 5 This diagram illustrates the controlled evolution of generation stability of the method for dynamically generating and pushing personalized auxiliary teaching content based on a large AI model, as provided in the embodiments of this application. Detailed Implementation

[0021] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0022] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0023] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0024] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0025] like Figure 1 As shown, Figure 1The flowchart illustrates a method for dynamically generating and pushing personalized auxiliary teaching content based on an AI large-scale model, as provided in this application embodiment. This method includes the following steps: initializing and configuring the AI ​​large-scale model and performing the first round of content generation; collecting the conditional probability distribution of the AI ​​large-scale model's output layer during the first round of generation and using it as a baseline probability distribution vector; summarizing the conditional probability distribution of the AI ​​large-scale model's output layer during subsequent rounds of content generation and combining it with the baseline probability distribution vector to determine the current round's generation probability distribution drift state, which includes a drift state and a non-drift state; when the generation probability distribution drift state is a drift state... If the generation parameter set is not controlled, a controlled adjustment process is triggered; otherwise, the controlled adjustment process is not executed. When it is detected that the generation process for consecutive preset rounds is in a non-drift state, a stable window is constructed, and the baseline probability distribution vector is dynamically updated within the stable window until the updated baseline probability distribution vector converges. Convergence specifically means that within the stable window, as multiple rounds of content generation and probability distribution collection continue, the difference between the continuously updated baseline probability distribution vectors gradually decreases and remains below the preset distribution change threshold for consecutive preset rounds, thus obtaining the adjusted generation parameter set. Otherwise, the update of the baseline probability distribution vector is not executed. The generation parameter set includes randomness. Parameters and / or candidate space constraint parameters and / or repetition penalty parameters; based on the adjusted generation parameter set, the AI ​​large model is invoked to dynamically generate personalized auxiliary teaching content units, and these units are pushed in batches according to preset push rules (the preset push rules include rules on the number of content units pushed in a single push, push time interval rules, and push order rules), while recording the corresponding push execution logs; during the push process, generation stability monitoring is performed to obtain stability monitoring results. When the stability monitoring results are in an oscillating state and / or a non-converged state, the push constraints are adjusted; otherwise, the adjusted generation parameter set is updated to the initial generation strategy of the corresponding teaching task, where personalized auxiliary teaching content units are generated. A teaching content unit refers to the smallest independently pushable and usable teaching content carrier generated by the AI ​​model in a single generation call, based on an adjusted generation parameter set, for a specific teaching object or teaching scenario. In the specific generation process, the system uses the above information and the adjusted generation parameter set as model input based on the teaching objectives of the current teaching task, calls the AI ​​model to execute content generation, and outputs structured teaching content results. For example, the content unit can be a practice question and its explanation that matches the current knowledge point. The complete result output by each model call constitutes a teaching content unit, which serves as the basic object for subsequent batch push and stability monitoring.The push execution log is used to record the entire process of teaching content generation and push in a traceable manner. It includes the following information: the current push cycle identifier and the corresponding generation cycle identifier, the generation parameter set identifier used in this push, the actual number of teaching content units pushed and their content unit identifiers, the push timestamp and the time interval information between adjacent pushes, and the corresponding generation stability monitoring result identifier after the push (such as stable state, oscillation state or non-convergence state).

[0026] In this embodiment, as Figure 2 As shown, Figure 2 This document presents a flowchart illustrating the overall steps of a method for dynamically generating and pushing personalized auxiliary teaching content based on an AI large-scale model, as provided in this embodiment. The method involves initializing and configuring the AI ​​large-scale model and executing the first round of content generation. This generates a conditional probability distribution, collects the conditional probability distribution of the model's output layer, and constructs a baseline probability distribution vector. The initialization configuration of the AI ​​large-scale model involves setting an initial set of generation parameters for the model without introducing historical adjustment effects, and fixing the corresponding candidate space constraints and randomness control strategies. Specifically, this includes setting the randomness parameters to a preset baseline value range, setting the candidate space constraint parameters and repetition penalty parameters to system default initial values, and loading basic prompts for the corresponding teaching tasks. After completing the initialization configuration, the AI ​​large-scale model is invoked to execute the first round of content generation. During the generation process, the conditional probability distribution of the model's output layer at each generation position is collected, and this conditional probability distribution is then summarized and encoded to construct a baseline probability distribution vector for subsequent comparative analysis. In subsequent rounds of content generation, the conditional probability distribution of the current round is continuously collected to form a generation probability distribution feature vector, which is then compared with the baseline probability distribution vector to determine the drift state of the generation probability distribution. When a drift state is detected, controlled adjustment is performed on the generated parameter set. When the state remains non-drifting for a consecutive preset number of rounds, a stable window is constructed and the baseline probability distribution vector is dynamically updated. Based on the adjusted generated parameter set, personalized auxiliary teaching content units are dynamically generated and pushed out in batches according to preset push rules, while monitoring the generation stability during the push process. When the generation process is detected to be in an oscillating state and / or a non-converged state, push constraint adjustments are performed. When the generation process is determined to have reached a convergent state, the current generated parameter set is updated to the initial generation strategy of the corresponding teaching task, thus forming a closed-loop generation, monitoring, adjustment, and stabilization control process.

[0027] It should be noted that one generation process corresponds to one generation parameter set configuration. With this parameter set configuration remaining unchanged, the generated teaching content is broken down and distributed across multiple push cycles for delivery. The generation process refers to a continuous content generation process executed by the AI ​​large-scale model under the same generation parameter set constraints. A push cycle refers to a time interval for batch delivery of the generated teaching content units after the generation process has ended. Teaching content generated in the same generation process can correspond to one or more consecutive push cycles. The push process refers to the process of dividing the teaching content units generated in the same generation process into multiple push cycles according to a preset push strategy and sequentially delivering the content.

[0028] Furthermore, the drift state of the generation probability distribution in the current round is obtained. Specifically, the conditional probability distribution of the output layer of the AI ​​large model is sampled and statistically summarized to obtain the feature vector of the generation probability distribution in the current round. The feature vector of the generation probability distribution is compared with the baseline probability distribution vector to obtain the distribution distance index, and the main probability peak displacement and probability entropy change are obtained simultaneously. The distribution distance index, the main probability peak displacement, and the probability entropy change are compared with the preset stable reference threshold set in the database. If any one of them exceeds the corresponding stable reference threshold set, the generation probability distribution in the current round is determined to be in a drift state; otherwise, it is in a non-drift state. The stable reference threshold set includes the distribution distance threshold, the main probability peak displacement threshold, and the probability entropy change threshold.

[0029] In this embodiment, during each round of teaching content generation, the AI ​​large-scale model outputs teaching content units step by step according to a predetermined set of generation parameters. During this generation process, the AI ​​large-scale model's output layer provides a corresponding conditional probability distribution for each candidate output unit at each generation step. Specifically, the data collection method is as follows: during the model inference phase, the conditional probability distribution generated by the model output layer is read, and the conditional probability distribution corresponding to each generation step in the current round is recorded; after completing the content generation for the current round, the conditional probability distributions corresponding to each generation step within that round are uniformly summarized to form a conditional probability distribution sampling result for subsequent analysis.

[0030] After sampling and summarizing the conditional probability distribution of the current round, statistical processing is performed on the conditional probability distribution collected within the round to extract distribution features that reflect the overall generation state of the round. Specifically, the conditional probability distribution of each generation step within the current round is statistically integrated to extract statistical features reflecting the degree of concentration of the probability distribution, the dominant output position, and uncertainty changes. These statistical features include probability distribution concentration features, dominant output position features, and uncertainty change features, which are combined in a preset order: probability distribution concentration features, dominant output position features, and uncertainty change features, forming a fixed-dimensional data structure. This yields a generation probability distribution feature vector that characterizes the overall characteristics of the generation behavior in the current round. The probability distribution concentration feature characterizes whether the probability is concentrated on a small number of output candidates; this feature is obtained by calculating the cumulative probability proportion of high-probability intervals (e.g., 0.7 and above) in the statistical conditional probability distribution. The dominant output position feature characterizes the output position that occupies a dominant position in the probability distribution; this is obtained by determining the position of the output candidate with the highest probability value in the candidate sequence. The uncertainty change feature characterizes the dispersion of the probability distribution; this is obtained by calculating the information entropy of the conditional probability distribution.

[0031] The distribution distance index is obtained by comparing the current generation probability distribution feature vector with the baseline probability distribution vector as two discrete probability distributions. Existing probability distribution difference calculation methods are used, such as calculating the overall difference between the two (specifically, calculating the absolute value of the probability difference between the current generation probability distribution feature vector and the baseline probability distribution vector in each dimension, and summing the difference results of all dimensions to obtain the distribution distance index). The result is used as the distribution distance index to reflect the degree of deviation of the current generation probability distribution from the baseline probability distribution.

[0032] The method for obtaining the peak displacement of the principal probability is as follows: In the current round's generated probability distribution feature vector and the baseline probability distribution vector, determine the output position with the highest probability value. By comparing the positions of the principal probability centers of the two, subtract them to obtain the peak displacement of the principal probability. This displacement is used to characterize whether the high-probability output center has shifted during the generation process. The method for determining the output position with the highest probability value is as follows: During the current round's generation process, sample the conditional probability distribution generated by the AI ​​large model output layer at each generation step. Then, based on the summarized generation probability distribution, sort the output candidates according to their probability magnitude. In the sorting results, select the output candidate with the highest probability value and record its position index in the corresponding candidate set. This position index serves as the dominant output position for the current round.

[0033] The method for obtaining the probability entropy change is as follows: Calculate the uncertainty of the probability distribution for both the current generation probability distribution feature vector and the baseline probability distribution vector, and then perform a difference operation on the two results to obtain the probability entropy change. This reflects the trend of overall randomness in the current generation process relative to the baseline state. The uncertainty of the probability distribution can be quantified using the information entropy calculation method. Specifically, based on their respective probability distributions, calculate the entropy value of their probability values ​​to obtain the information entropy value of the current generation probability distribution and the information entropy value of the baseline probability distribution; then perform a difference operation on the two to obtain the probability entropy change, which characterizes the change of randomness in the current generation process relative to the baseline generation state.

[0034] By determining the drift state of the generated probability distribution, the stability of generation can be monitored directly at the probability distribution level of the AI ​​large model generation process. This allows the system to identify potential deviation trends in the generation process before obvious anomalies appear in the generation quality, and provides an objective basis for the controlled adjustment of subsequent generation parameter sets. This avoids the generation strategy gradually deviating from the optimal generation region during continuous generation, and improves the stability and controllability of personalized auxiliary teaching content in the multi-round generation and delivery process.

[0035] Furthermore, the controlled adjustment of the generated parameter set is specifically implemented as follows: when the generated probability distribution is in a drift state, the drift direction is obtained, and the generated parameter set is adjusted in a directional manner based on the drift direction; when the probability entropy change is detected to be greater than the probability entropy change threshold, the range of randomness parameters is reduced; when the principal probability peak displacement is detected to be greater than the preset principal probability peak displacement threshold, the range of candidate space constraint parameters is tightened; when the distribution distance index is detected to be greater than the distribution distance threshold, the intensity of the repetition penalty parameter is increased.

[0036] In this embodiment, as Figure 3 As shown, Figure 3 The flowchart for determining and controlling the generation probability distribution drift in the AI-based large-scale model-based method for dynamically generating and pushing personalized auxiliary teaching content provided in this application embodiment is as follows: During the generation probability distribution drift determination process, a generation probability distribution feature vector for the current round is constructed, and the distribution distance index, principal probability peak displacement, and probability entropy change of this feature vector relative to the baseline probability distribution vector are calculated. These indices are compared with a preset set of stable reference thresholds. If any index exceeds the corresponding threshold, the generation probability distribution for the current round is determined to be in a drift state; otherwise, it is determined to be in a non-drift state. When a drift state is determined, the drift direction is obtained, and the generation parameter set is directionally and controllably adjusted based on the drift direction to suppress abnormal shifts in the generation distribution and avoid disordered fluctuations in the generation process.

[0037] The drift direction of the drift state is obtained by comparing the feature vector of the generated probability distribution obtained in the current generation process with the feature vector of the baseline probability distribution. The determination criteria include the change in probability entropy, the displacement of the principal probability peak, and the changing trend of the distribution distance index. When the change in probability entropy shows an increasing trend, the drift direction is determined to be "drifting towards the direction of high uncertainty"; when the principal probability peak undergoes a significant displacement in the output space, the drift direction is determined to be "drifting towards the direction of output centroid shift"; when the distribution distance index continues to increase, the drift direction is determined to be "drifting towards the direction of deviation from the overall distribution structure".

[0038] When the probability entropy change is detected to be greater than the probability entropy change threshold, the range of randomness parameters is adjusted by gradually shrinking them, i.e., the upper limit of randomness parameters is gradually reduced in multiple generation rounds according to a preset adjustment step size; when the principal probability peak displacement is detected to be greater than the preset principal probability peak displacement threshold, the range of candidate space constraint parameters is gradually tightened, and the generation probability distribution is guided to regress to a stable region by gradually reducing the candidate output space; when the distribution distance index is detected to be greater than the distribution distance threshold, the intensity of the repetition penalty parameter is gradually increased according to a preset penalty step size. The randomness parameters include the sampling temperature parameter and the random sampling perturbation coefficient. The sampling temperature parameter is used to scale the conditional probability distribution of the model output layer, and its value is used to adjust the flatness of the probability distribution. The random sampling perturbation coefficient is used to introduce controlled random perturbation during the sampling stage, and its value represents the weight of random perturbation in the original conditional probability distribution. The candidate space constraint parameters include the upper limit parameter for the number of candidates and the cumulative probability coverage threshold. The repetition penalty parameters include the repetition penalty coefficient and the historical context penalty weight. The historical context penalty weight is used to apply a penalty factor to the candidate output probability corresponding to the generated content. The larger the value, the stronger the suppression of repeated candidates. It is used to apply additional penalties to the candidate outputs that have appeared in the historical generated sequence, and its value represents the weight of the influence of historical repetition on the current sampling.

[0039] like Figure 4 As shown, Figure 4 This is a schematic diagram of the probability distribution drift of the dynamic generation and push method for personalized auxiliary teaching content based on AI large model provided in the embodiments of this application. The horizontal axis represents the candidate output sequence number, the vertical axis represents the conditional probability value, and the baseline probability distribution vector corresponds to the reference distribution formed within the stable window. By analyzing and comparing the feature vector of the current generation probability distribution and the baseline probability distribution vector, the purpose is to calculate the overall offset, peak position change and morphological difference between the two curves, which are used to calculate and determine the distribution distance index, the displacement of the main probability peak and the change of probability entropy.

[0040] Furthermore, the baseline probability distribution vector is dynamically updated, specifically including: when the generation probability distribution drift state of consecutive preset rounds is detected to be in a non-drift state, a stable window is constructed, and within the stable window, the generation probability distribution feature vector of the corresponding round is subjected to a moving average processing to obtain the updated baseline probability distribution vector, thereby realizing the dynamic update of the baseline probability distribution vector; the updated baseline probability distribution vector replaces the original baseline probability distribution vector and is used for drift determination in subsequent rounds.

[0041] In this embodiment, a stable window is constructed by taking the set of generation rounds that most recently and consecutively satisfy the non-drift state determination as the time range of the stable window. The size of the stable window is directly determined by the number of "continuous preset rounds", that is, the length of the stable window is equal to the preset number of consecutive non-drift rounds.

[0042] Perform a moving average process to obtain the updated baseline probability distribution vector. Specifically, within the stable window, align the generation probability distribution feature vectors of each generation round according to their corresponding dimensions, accumulate the probability feature values ​​of the same dimension within the window, and then divide by the number of rounds within the stable window to obtain the average probability feature values ​​of each dimension. Finally, combine these to form the updated baseline probability distribution vector.

[0043] By performing a moving average on the feature vector of the generated probability distribution while it remains in a non-drift state throughout the continuous generation probability distribution drift, and obtaining an updated baseline probability distribution vector, the baseline probability distribution vector can be gradually and adaptively adjusted with the long-term stable changes in the generation process. This avoids the reference distortion problem caused by a long-term fixed baseline distribution. Furthermore, allowing the baseline probability distribution vector to be updated only when the generated probability distribution is in a non-drift state effectively prevents the introduction of transient distribution features into the baseline distribution during abnormal fluctuations, drastic parameter adjustments, or unstable content phases in the generation process.

[0044] Furthermore, batch push is carried out. The specific method is as follows: determine the corresponding push frequency level and the number of pushes per batch based on the current generation parameter set; push the generated teaching content units in batches according to the preset push interval, where the preset push interval is the initial push interval stored in the database; obtain the push execution cooldown time and compare it with the preset minimum cooldown time; if the push execution cooldown time is less than the minimum cooldown time, then push cooling is carried out based on the minimum cooldown time.

[0045] In this embodiment, the push execution cooldown time is used to characterize the actual time interval between two adjacent batch pushes. Specifically, it is obtained by recording the start time of the current batch push and the completion time of the previous batch push during each batch push of teaching content unit, and calculating the actual interval between adjacent push batches based on the time difference between the two. The time difference is used as the push execution cooldown time in the current push cycle.

[0046] Push cooling based on minimum cooldown time refers to imposing a delay constraint on subsequent push operations when the current push execution cooldown time is detected to be less than the preset minimum cooldown time. Specifically, the push operation for triggering the next batch of teaching content units is paused, and a delay waiting mechanism is introduced into the push scheduling module so that the trigger time of the next push is at least at the minimum cooldown time interval relative to the completion time of the previous batch of pushes. During this delay waiting period, no new push instructions are executed, and only the generated parameter set and push status remain unchanged until the cumulative waiting time reaches the minimum cooldown time, at which point the subsequent batch push process is allowed to resume.

[0047] The corresponding push frequency level and single-batch push quantity are determined based on the current generated parameter set. The specific method is as follows: First, obtain the value status of each generated parameter in the current generated parameter set, including the randomness parameter, the candidate space constraint parameter, and the duplication penalty parameter; match the generated parameter set with the preset push strategy mapping table in the system. This mapping table has pre-established rules for push frequency levels and single-batch push quantities corresponding to different stability levels of the generated parameter sets. For example, when both the randomness parameter and the candidate space constraint parameter converge (convergence means that the parameter is less than the corresponding preset threshold within a consecutive preset number of periods, such as the randomness parameter being less than the randomness parameter threshold), it is matched to the high-frequency, small-batch push level; when the generated parameter set is still in an adjustment or restriction state, it is matched to the low-frequency, small-batch, or medium-batch push level, thereby determining the push frequency level and single-batch push quantity.

[0048] By determining the push frequency level and the number of pushes per batch based on the generated parameter set, the push rhythm of teaching content units can be kept consistent with the stability of the current generation process. This avoids high-frequency, large-batch pushes when the generation parameters are not yet stable or are still frequently adjusted, thereby reducing the impact of unstable generated content on teaching effectiveness. Simultaneously, by analyzing the push execution cooldown time and comparing it with the preset minimum cooldown time, the minimum cooldown time is forcibly adopted for push cooling when the cooldown time is insufficient. This prevents problems such as concentrated system resource consumption, excessive user load, or incomplete effective feedback adjustment of generation parameters caused by excessively short intervals between adjacent pushes. While ensuring the continuity of teaching content, adaptive constraints on the push rhythm are achieved, improving the stability, security, and consistency of the overall teaching experience.

[0049] Furthermore, stability monitoring is performed to obtain stability monitoring results. Specifically, the adjustment direction, number of adjustments, and adjustment period of the generation parameter set within each push cycle are obtained and statistically analyzed. When it is detected that a certain generation parameter set repeatedly switches between forward and reverse directions within adjacent push cycles, and / or continuously triggers generation parameter set adjustments exceeding the corresponding preset number of times, the generation process corresponding to the generation parameter set adjustment across adjacent push cycles is determined to be in an oscillating state; otherwise, it is in a non-oscillating state. When the generation parameter set remains unchanged or its change amplitude is lower than the corresponding change amplitude threshold within a preset number of consecutive cycles, the generation process is determined to be in a convergent state; otherwise, the generation process is determined to be in a non-convergent state, thus obtaining the stability monitoring results. The stability monitoring results include oscillating and non-convergent states, non-oscillating and non-convergent states, and non-oscillating and convergent states.

[0050] In this embodiment, the adjustment direction, number of adjustments, and adjustment period of the generated parameter set are all obtained by recording and statistically analyzing the actual adjustment behavior of the generated parameter set within the push cycle. Specifically, in each push cycle, the parameter values ​​of each generated parameter in the generated parameter set at the start and end of the cycle are compared. When the parameter value changes relative to the previous push cycle, it is determined that the generated parameter has undergone one adjustment in the current push cycle. If the parameter value increases relative to the previous push cycle, it is recorded as a positive adjustment; if the parameter value decreases, it is recorded as a negative adjustment, thereby determining the adjustment direction of the generated parameter set within the push cycle. If the generated parameter set is triggered for controlled adjustment multiple times within the same push cycle, the number of adjustment events occurring within that cycle is accumulated to obtain the corresponding number of adjustments. Simultaneously, using the push cycle as the basic time unit, the number of push cycles in which continuous adjustment behavior occurs is counted to characterize the adjustment period of the generated parameter set.

[0051] By determining whether there are repeated forward and reverse switching or excessive triggering of adjustments in the generated parameter set within adjacent push cycles, high-frequency disturbances and adjustment imbalances in the generation process can be identified in a timely manner. This avoids further amplifying the adjustment amplitude or frequency before the generated parameters have stabilized, preventing fluctuations in the quality of the generated content. By analyzing whether the generated parameter set remains unchanged or only undergoes small unidirectional changes within continuous cycles, it is possible to determine whether the generation process has truly entered the convergence phase. This provides a basis for subsequent decisions on whether to solidify the generated parameter set, remove push constraints, and update the initial generation strategy. By decoupling the determination of oscillation and convergence states, the system can identify both the intermediate state of "still adjusting but not unstable" and the stable state of "stable convergence." Ultimately, this achieves synergistic optimization between generated parameter adjustment, content generation, and push control, improving the overall stability, continuity, and controllability of the personalized teaching content generation and push process.

[0052] Furthermore, the push constraints are adjusted as follows: when the stability monitoring result is in an oscillation state, the oscillation intensity value is analyzed; when the stability monitoring result is in a non-converged state, the divergence intensity value is analyzed; based on the joint analysis of the oscillation intensity value and the divergence intensity value, the number of generation parameter sets allowed to be adjusted within a single cycle, the adjustment range limit, and the push interval are obtained respectively, and push constraints are added during the push process; the push constraints prohibit the execution of opposite adjustment operations on the same generation parameter set within two consecutive generation cycles; when it is determined that the generation process has reached a convergence state, the current generation parameter set is saved as the initial generation strategy for the corresponding teaching task.

[0053] In this embodiment, the oscillation intensity value characterizes the repetitiveness and severity of the parameter set adjustment, while the divergence intensity value characterizes the expanding trend of the generated distribution continuously deviating from the stable reference state. Joint analysis of the oscillation intensity and divergence intensity values ​​is used to dynamically determine the number of generated parameter sets allowed for adjustment within a single cycle, the adjustment range limit, and the push interval. This allows the push constraint to adaptively tighten or loosen according to the degree of generation instability, suppressing over-adjustment and reducing the risk of parameter oscillation while still retaining necessary adjustment space to guide the generation process back to the stable range. By introducing a push constraint condition that prohibits performing opposite-direction adjustment operations on the same generated parameter set within two consecutive generation cycles, the high-frequency feedback loop of "positive and negative back-and-forth correction" is effectively cut off, preventing the generated parameters from falling into repeated fluctuations in a short period, thus improving the convergence efficiency of the generation process from a mechanistic perspective. Once the generation process is determined to have reached convergence, the current set of generation parameters is saved as the initial generation strategy for the corresponding teaching task. This allows the verified and stable parameter configurations to be directly reused in subsequent similar teaching tasks, reducing the number of frequent trials and adjustments during the cold start phase. Overall, this achieves a simultaneous improvement in generation stability, push controllability, and consistency of teaching content quality.

[0054] Furthermore, the oscillation intensity value is obtained by: acquiring the oscillation response parameters of the parameter set generated per unit time; normalizing the oscillation response parameters and a preset oscillation response reference set, and then weighting and superimposing them to obtain the oscillation intensity value; the oscillation response parameters include the number of reverse adjustments, the rate of change of adjustment amplitude, and the frequency of adjustment direction switching; the oscillation response reference set includes the influence value of the number of reverse adjustments per unit, the reference value of the rate of change of adjustment amplitude, and the reference value of the frequency of adjustment direction switching.

[0055] In this embodiment, the number of reverse adjustments refers to the cumulative number of times the adjustment direction of the same generated parameter set switches from forward adjustment to reverse adjustment, or vice versa, within a unit time or a push cycle. Specifically, it is obtained by performing time-series statistics on the adjustment directions of the generated parameter set recorded in the push execution log; when two adjacent adjustment directions have opposite signs, it is counted as one reverse adjustment. The rate of change in adjustment amplitude refers to the degree of relative change in adjustment amplitude between two adjacent generated parameter set adjustment operations. It is obtained by calculating the ratio of the difference in adjustment amplitude of the same generated parameter within adjacent adjustment cycles to the adjustment amplitude of the previous cycle, used to characterize whether the adjustment amplitude rapidly increases or decreases. The frequency of adjustment direction switching refers to the frequency of changes in the adjustment direction of the generated parameter set within a unit time. It is obtained by statistically analyzing the number of times the adjustment direction of the generated parameter set switches within a unit time window and dividing by the corresponding time length, used to characterize the frequency of adjustment behavior.

[0056] During the dynamic adjustment of the generated parameter set, there is a clear coupling relationship between the number of reverse adjustments, the rate of change of adjustment amplitude, and the frequency of adjustment direction switching: when the frequency of adjustment direction switching increases, it usually means that the generated parameters frequently change the adjustment direction in adjacent cycles, which directly leads to an increase in the number of reverse adjustments; the accumulation of the number of reverse adjustments is often accompanied by the back-and-forth correction of the adjustment amplitude in the positive and negative directions, which widens the difference in adjustment amplitude between adjacent adjustment cycles, thereby pushing up the rate of change of adjustment amplitude; when the rate of change of adjustment amplitude is too large, it is easy to trigger overcorrection, prompting the system to reverse adjust again in subsequent cycles to "correct the deviation", thereby further increasing the frequency of adjustment direction switching and the number of reverse adjustments, ultimately forming an oscillating feedback loop at the parameter adjustment level.

[0057] By normalizing and superimposing the number of reverse adjustments, the rate of change of adjustment amplitude, and the frequency of adjustment direction switching, the oscillation intensity value can be obtained, which can reflect whether there are problems such as high-frequency repeated corrections, over-adjustment, or imbalance in adjustment rhythm during the generation process. The oscillation intensity value provides a clear quantitative basis for subsequent push constraint adjustments, and can adaptively limit the number of adjustments, adjustment amplitude, and push interval of the generation parameter set according to the level of oscillation risk. This avoids the generation process from being in an unstable oscillation range during the push phase, improves the convergence efficiency of generation parameters, and ultimately ensures the stability and consistency of personalized auxiliary teaching content during continuous generation and push.

[0058] The oscillation intensity value is obtained as follows: Multiply the number of reverse adjustments by the impact value of each unit of reverse adjustment to obtain the reverse adjustment number processing value. Divide the rate of change of adjustment amplitude and the frequency of adjustment direction switching by the corresponding control values ​​for the rate of change of adjustment amplitude and the frequency of adjustment direction switching, respectively, to obtain the adjustment amplitude processing value and the adjustment direction switching frequency processing value. Then, weight these values ​​(reverse adjustment number processing value, adjustment amplitude processing value, and adjustment direction switching frequency processing value) according to their respective oscillation intensity weighting coefficients and sum them to obtain the oscillation intensity value. The oscillation intensity weighting coefficients include the reverse adjustment number weighting coefficient, the adjustment amplitude weighting coefficient, and the adjustment direction switching frequency weighting coefficient.

[0059] The weighting coefficients for the number of reverse adjustments, the adjustment amplitude, and the adjustment direction switching frequency can be obtained in the following ways: For example, within a historical push cycle or offline simulation cycle, the number of reverse adjustments, the rate of change of adjustment amplitude, and the adjustment direction switching frequency are treated as three types of oscillation factors. In each push cycle, the corresponding generation stability degradation characterization quantity is recorded synchronously. This quantity includes the deviation of the generation distribution feature vector from the stable reference distribution vector, the number of failed regressions in the stable interval, and the reduction ratio of effective generation rounds. For each type of oscillation factor, first, the change in its value within adjacent push cycles is calculated, then the change in the generation stability degradation characterization quantity within the corresponding cycle is calculated. Based on multi-cycle samples, the average influence intensity of the oscillation factor change on the generation stability degradation change is statistically analyzed. The average influence intensity can be quantified by the mean of the product of changes, yielding the reverse adjustment influence value, the amplitude change influence value, and the direction switching influence value. The three influence values ​​are normalized so that the sum of the normalized values ​​is 1. The normalized influence values ​​are then determined as the weighting coefficients for the number of reverse adjustments, the adjustment amplitude, and the adjustment direction switching frequency, respectively.

[0060] Furthermore, the divergence intensity value is obtained as follows: Within each push cycle, the generated distribution feature vector corresponding to the current generation process is obtained, and its deviation is calculated against the corresponding stable reference distribution vector to obtain the distribution deviation value. Based on the changes in the distribution deviation value within adjacent push cycles, the deviation growth value is analyzed to characterize whether the deviation of the generated distribution feature vector from the stable reference distribution vector shows a continuously expanding trend within adjacent push cycles. When it is detected that the distribution deviation value is continuously greater than the corresponding stable threshold within consecutive preset push cycles, and / or the deviation growth value is continuously positive, the corresponding divergence response parameters are extracted. The divergence response parameters are normalized and weighted with a preset divergence response reference set to obtain the divergence intensity value. The divergence response parameters include the distribution deviation amplitude value, the deviation growth rate, and the number of regression failures in the stable interval. The divergence response reference set includes the distribution deviation amplitude value reference value, the deviation growth rate reference value, and the impact value of the number of regression failures in the stable interval.

[0061] In this embodiment, the stable reference distribution vector refers to the reference probability distribution representation used to characterize the conditional probability distribution of the AI ​​large model output layer when the generation process is in a stable state. When it is detected that the generation probability distribution drift state is non-drift state within a consecutive preset push period, and the generation stability monitoring result is non-oscillating and converged, the generation distribution feature vector obtained within the corresponding push period is taken as a stable sample; the stable sample is processed by moving average within the stable window to obtain the stable reference distribution vector.

[0062] The specific method for obtaining the divergence intensity value is as follows: divide the divergence response parameter by the corresponding divergence response control set to obtain each divergence response processing value. Multiply each divergence response processing value by the corresponding divergence response weight set and then add them together to obtain the divergence intensity value. The divergence response weight set includes the distribution deviation magnitude weight, the deviation growth rate weight, and the number of regression failures in the stable interval weight.

[0063] The weights for the distribution deviation magnitude, deviation growth rate, and number of failed regressions in the stable interval are calculated as follows: In historical generation and push records, the changes in the values ​​of the distribution deviation magnitude, deviation growth rate, and number of failed regressions in the stable interval are statistically analyzed within different push periods, and the changes in the generation stability degradation index within the corresponding period are also statistically analyzed. For each type of divergent response parameter, the correlation index between its value change and the change in the generation stability degradation index is calculated, and the correlation index is used as the original weight value for that type of divergent response parameter. Subsequently, each original weight value is normalized so that the sum of the values ​​of the distribution deviation magnitude weight, deviation growth rate weight, and number of failed regressions in the stable interval is 1. The specific method for obtaining the correlation index is as follows: In the historical generation and push records, multiple push periods are selected, and the changes in the values ​​of the corresponding divergent response parameters and the changes in the generation stability degradation index are obtained in each push period. Statistical analysis is performed on the sequence of changes in the values ​​of the divergent response parameters and the sequence of changes in the generation stability degradation index. After standardizing the two change sequences, the synchronous change ratio in the same push period is calculated, thereby obtaining the corresponding correlation index.

[0064] The specific process for generating the distribution feature vector is as follows: During the content generation process of the current round, the conditional probability distributions obtained by the output layer of the AI ​​large model in each decoding step are sampled and summarized to form the overall conditional probability distribution of this round; Based on the overall conditional probability distribution, the candidate output items are sorted from high to low according to their probability values, and the candidate output items ranked in the top preset order are selected and their cumulative probability values ​​are calculated to obtain the concentration feature used to characterize the concentration of the probability distribution; Next, the information entropy of the overall conditional probability distribution is calculated to obtain the entropy feature used to characterize the uncertainty of the probability distribution; The candidate output item with the largest probability value in the overall conditional probability distribution is determined, and the index position of the candidate output item in the candidate output sequence is obtained to obtain the peak position feature used to characterize the dominant output position; The concentration feature, entropy feature, and peak position feature are vectorized and combined in a preset order to form the generated distribution feature vector used to characterize the overall probability distribution state of the current generation process.

[0065] Distribution deviation is a numerical indicator used to quantify the degree of deviation of the currently generated distribution feature vector from the stable reference distribution vector. Specifically, it is obtained by calculating the probability distribution difference between the currently generated distribution feature vector and a pre-determined stable reference distribution vector within the same push period. Using existing probability distribution distance measurement methods (such as calculations based on distribution distance indicators), a numerical value is obtained to characterize the overall degree of difference between the two; this value is the distribution deviation. A larger value indicates that the currently generated distribution deviates more from the stable reference distribution; a smaller value indicates that the generated distribution is closer to the stable reference range.

[0066] The deviation growth value is obtained by performing time series analysis on the distribution deviation values ​​obtained in adjacent push cycles, calculating the difference between the distribution deviation value of the current push cycle and the distribution deviation value of the previous push cycle, which is the deviation growth value. When this change value is positive and continues to appear, it indicates that the degree of deviation of the generated distribution from the stable reference distribution is constantly expanding in the time dimension, thus reflecting the divergence trend of the generated distribution.

[0067] The distribution deviation magnitude refers to the degree of deviation of the generated distribution feature vector from the stable reference distribution vector within a single push cycle; its specific value is the distribution deviation value itself. The deviation growth rate refers to the growth rate of the distribution deviation value in adjacent push cycles, which is obtained by calculating the difference between the distribution deviation values ​​in adjacent cycles. The number of failed regressions in the stable interval refers to the cumulative number of times the generated distribution deviation value exceeds the corresponding stable threshold multiple times in consecutive push cycles and fails to fall back to the stable interval within a preset period. It is obtained by statistically analyzing whether the distribution deviation value enters the stable threshold interval based on the push cycle logs and counting the cases of unsuccessful regression.

[0068] By introducing a divergence intensity value at the push cycle scale, it is possible to quantitatively identify whether the generation process continuously deviates from the stable reference distribution. This allows the adjustment strategy of the generation parameter set to be restricted for the generation state of "long-term deviation and continuous expansion," thereby effectively preventing the generation strategy from gradually slipping out of the stable generation region during continuous push and improving the consistency and controllability of personalized auxiliary teaching content in the long-term generation and push process.

[0069] Furthermore, the number of generated parameter sets that can be adjusted within a single period, the limit on adjustment range, and the push interval are obtained. Specifically, the following method is used: based on the oscillation intensity value and the divergence intensity value, the corresponding unit oscillation intensity value influence coefficient and the unit divergence intensity value influence coefficient are introduced, multiplied, and then added to obtain the joint influence intensity value; based on the joint influence intensity value, mapping and matching are performed with the corresponding mapping sets, thereby obtaining the limit on the number of generated parameter sets that can be adjusted within a single period, the limit on adjustment range, and the push interval.

[0070] In this embodiment, the joint impact intensity value is mapped and matched with the corresponding mapping set to obtain the number of restricted generation parameter sets, the restricted adjustment range, and the push interval that can be adjusted within a single period. Specifically, the following method is used: obtain the preset joint impact intensity value ranges in the database and the number of restricted generation parameter sets, the restricted adjustment range, and the push interval that can be adjusted within a single period corresponding to each joint impact intensity value range, and match them with the current joint impact intensity value. If the current joint impact intensity value is within a certain preset joint impact intensity value range, then obtain the number of restricted generation parameter sets, the restricted adjustment range, and the push interval that can be adjusted within a single period corresponding to that range as the number of restricted generation parameter sets, the restricted adjustment range, and the push interval that can be adjusted within a single period corresponding to the current joint impact intensity value.

[0071] By matching the joint influence strength value with a preset mapping set, the abstract generation stability risk can be directly mapped into executable control strategy parameters, including the number of generation parameter sets that can be adjusted within a single cycle, the limit on the adjustment range, and the push interval. This transforms generation control from empirical adjustment to fine-grained regulation based on stability strength levels, achieving adaptive constraints on the adjustment density and rhythm of generation parameters. This prevents frequent and significant modifications to the generation strategy under oscillation or divergent trends from causing system-level instability. Furthermore, it can automatically relax constraints as stability gradually recovers, improving generation and push efficiency.

[0072] The method for obtaining the influence coefficient of unit oscillation intensity value is as follows: During the historical generation process, multiple continuous generation cycles are divided according to a preset time window. In each generation cycle, the corresponding oscillation intensity value and the actual adjustment behavior data of the generated parameter set are recorded synchronously. The actual adjustment behavior data of the generated parameter set includes the number of parameter adjustment triggers, the number of parameter rollbacks, or the magnitude of generation distribution shift within a unit time. Based on the range of oscillation intensity values, the historical oscillation intensity values ​​are divided into intervals. Within each oscillation intensity interval, the mean of the number of parameter adjustment triggers, the number of rollbacks, or the magnitude of generation distribution shift in the corresponding generation cycle is calculated. The mean is used as the oscillation influence statistics corresponding to the oscillation intensity interval. Furthermore, each oscillation influence statistics is normalized to obtain the standardized influence degree of the unit oscillation intensity value change on the stability of the generated parameter adjustment, thereby determining the influence coefficient of unit oscillation intensity value.

[0073] The specific method for obtaining the influence coefficient of unit divergence intensity value is as follows: During the historical generation process or offline simulation generation process, the generation process is segmented according to a preset generation cycle, and the corresponding divergence intensity value and the state change data of the generated distribution within each generation cycle are obtained; the state change data of the generated distribution includes at least the change in the entropy of the generated distribution, the dominant probability shift, or the change in the size of the effective candidate set; a divergence intensity interval division rule is constructed based on the range of historical divergence intensity values, and within each divergence intensity interval, the mean of the change in the entropy of the generated distribution, the dominant probability shift, or the change in the size of the effective candidate set in the corresponding generation cycle is calculated; the mean is used as the divergence influence statistical value corresponding to the divergence intensity interval, and each divergence influence statistical value is normalized to obtain the standardized influence degree of the change in unit divergence intensity value on the divergence trend of the generated distribution, thereby determining the influence coefficient of unit divergence intensity value.

[0074] like Figure 5 As shown, Figure 5 This diagram illustrates the controlled evolution of generation stability in the AI-based large-scale model-based method for dynamically generating and pushing personalized auxiliary teaching content, as provided in this embodiment. The horizontal axis represents the generation cycle number, and the vertical axis represents the intensity value calculated within the corresponding cycle. The oscillation intensity value represents the oscillation intensity value obtained statistically based on the overall generation and adjustment behavior within each generation cycle, used to characterize the degree of adjustment fluctuation of the system within that cycle. The divergence intensity value represents the divergence intensity value calculated based on the distribution of multiple generation results within each generation cycle, used to characterize the diffusion trend of the output result distribution.

[0075] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0076] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0077] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0078] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0079] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0080] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for dynamically generating and pushing personalized auxiliary teaching content based on AI large-scale models, characterized in that, Includes the following steps: The AI ​​large model is initialized and configured, and the first round of content generation is performed. The conditional probability distribution of the output layer of the AI ​​large model is collected during the first round of generation and used as the baseline probability distribution vector. The conditional probability distribution of the AI ​​large model's output layer is summarized in the subsequent content generation process, and combined with the baseline probability distribution vector for judgment to obtain the generation probability distribution drift state of the current round. The generation probability distribution drift state includes drift state and non-drift state. When the generated probability distribution is in a drift state, the controlled adjustment process of the generated parameter set is triggered; otherwise, the controlled adjustment process is not executed. The controlled adjustment process for generating the parameter set is specifically as follows: When the probability distribution is in a drift state, the drift direction is obtained, and the generation parameter set in the generation process is adjusted based on the drift direction. Specifically, when the detected change in probability entropy exceeds the threshold, the range of values ​​for the randomness parameter is reduced. When the detected principal probability peak displacement is greater than the preset principal probability peak displacement threshold, the range of the candidate space constraint parameters is tightened. When the distribution distance index is detected to be greater than the distribution distance threshold, the strength of the repetition penalty parameter is increased; When it is detected that the generation process of consecutive preset rounds is in a non-drift state, a stable window is constructed, and the baseline probability distribution vector is dynamically updated within the stable window until the updated baseline probability distribution vector converges, and the adjusted generation parameter set is obtained. Otherwise, the update of the baseline probability distribution vector is not performed. The generation parameter set includes randomness parameters and / or candidate space constraint parameters and / or repetition penalty parameters. Based on the teaching objectives of the current teaching task and the adjusted set of generated parameters, the AI ​​large model is called to dynamically generate personalized auxiliary teaching content units, namely, practice questions and explanations that match the current knowledge points, and they are pushed in batches according to the preset push rules, while recording the corresponding push execution logs. During the push process, stability monitoring is performed to obtain the stability monitoring results. When the stability monitoring results are in an oscillating state and / or a non-converged state, the push constraints are adjusted; otherwise, the adjusted generation parameter set is updated to the initial generation strategy of the corresponding teaching task.

2. The method for dynamically generating and pushing personalized auxiliary teaching content based on AI large model as described in claim 1, characterized in that: The method for obtaining the generation probability distribution drift state of the current round is as follows: After sampling and statistically summarizing the conditional probability distribution of the output layer of the large AI model, the feature vector of the generation probability distribution of the current round is obtained. The generated probability distribution feature vector is compared with the baseline probability distribution vector to obtain the distribution distance index, and the comparison is performed simultaneously to obtain the displacement of the main probability peak and the change of probability entropy. The distribution distance index, the displacement of the main probability peak, and the change in probability entropy are compared with the preset stable reference threshold set in the database. If any one of them exceeds the corresponding stable reference threshold set, the generation probability distribution of the current round is determined to be in a drift state; otherwise, it is in a non-drift state. The stable reference threshold set includes the distribution distance threshold, the principal probability peak displacement threshold, and the probability entropy change threshold.

3. The method for dynamically generating and pushing personalized auxiliary teaching content based on AI large model as described in claim 2, characterized in that: The dynamic updating of the baseline probability distribution vector specifically includes: When it is detected that the generation probability distribution drift state of consecutive preset rounds is in a non-drift state, a stable window is constructed, and within the stable window, the generation probability distribution feature vector of the corresponding round is subjected to a moving average process to obtain an updated baseline probability distribution vector, thereby realizing the dynamic update of the baseline probability distribution vector. The updated baseline probability distribution vector replaces the original baseline probability distribution vector and is used for drift determination in subsequent rounds.

4. The method for dynamically generating and pushing personalized auxiliary teaching content based on AI large model as described in claim 1, characterized in that: The specific method for batch push is as follows: Determine the corresponding push frequency level and the number of pushes per batch based on the current generated parameter set. The generated teaching content units are pushed out in batches according to the preset push interval. The push execution cooldown time is obtained and compared with the preset minimum cooldown time. If the push execution cooldown time is less than the minimum cooldown time, the push is cooled down based on the minimum cooldown time.

5. The method for dynamically generating and pushing personalized auxiliary teaching content based on AI large model as described in claim 1, characterized in that: The process of generating stability monitoring results is as follows: Obtain the adjustment direction, number of adjustments, and adjustment period of the parameter set generated in each push cycle, and perform statistics; When it is detected that a certain generation parameter set repeatedly switches between forward and reverse directions within an adjacent push cycle, and / or the generation parameter set adjustment is triggered continuously more than the corresponding preset number of times, it is determined that the generation process corresponding to the generation parameter set adjustment across adjacent push cycles is in an oscillation state; otherwise, it is in a non-oscillation state. When the generated parameter set remains unchanged or changes less than the corresponding change threshold within a preset number of consecutive periods, the generation process is determined to be in a convergent state; otherwise, the generation process is determined to be in a non-convergent state, thus obtaining the stability monitoring result. The stability monitoring results include oscillating and non-converging states, non-oscillating and non-converging states, and non-oscillating and convergent states.

6. The method for dynamically generating and pushing personalized auxiliary teaching content based on AI large model as described in claim 1, characterized in that: The specific method for adjusting the push constraints is as follows: When the stability monitoring result indicates an oscillation state, the oscillation intensity value is obtained through analysis; When the stability monitoring result is a non-converged state, the divergence intensity value is obtained through analysis; Based on the joint analysis of oscillation intensity value and divergence intensity value, the number of generation parameter sets allowed to be adjusted within a single cycle, the limit of adjustment amplitude and the push interval are obtained respectively, and push constraints are added during the push process; The push constraint condition prohibits performing opposite adjustment operations on the same set of generation parameters within two consecutive generation cycles; Once the generation process is determined to have reached convergence, the current set of generation parameters is saved as the initial generation strategy for the corresponding teaching task.

7. The method for dynamically generating and pushing personalized auxiliary teaching content based on AI large model as described in claim 6, characterized in that: The specific method for obtaining the oscillation intensity value is as follows: Obtain the oscillation response parameters of the parameter set generated per unit time, and then perform weighted superposition based on the oscillation response parameters and the preset oscillation response reference set after normalization to obtain the oscillation intensity value; The oscillation response parameters include the number of reverse adjustments, the rate of change of adjustment amplitude, and the frequency of adjustment direction switching. The oscillation response reference set includes the influence value of the number of reverse adjustments per unit, the reference value of the rate of change of adjustment amplitude, and the reference value of the frequency of adjustment direction switching.

8. The method for dynamically generating and pushing personalized auxiliary teaching content based on AI large model as described in claim 6, characterized in that: The specific method for obtaining the divergence intensity value is as follows: Within each push cycle, obtain the generation distribution feature vector corresponding to the current generation process, and calculate the deviation between it and the corresponding stable reference distribution vector to obtain the distribution deviation value. Based on the changes in distribution deviation within adjacent push cycles, the deviation growth value is analyzed to characterize whether the degree of deviation of the generated distribution feature vector from the stable reference distribution vector shows a continuously expanding trend within adjacent push cycles. When the distribution deviation value is detected to be continuously greater than the corresponding stable threshold within a continuous preset push period, and / or the deviation growth value is continuously positive, the corresponding divergent response parameters are extracted. The divergent response parameters are normalized and weighted and superimposed with a preset divergent response reference set to obtain the divergent intensity value; The divergent response parameters include the magnitude of the distribution deviation, the rate of deviation growth, and the number of regression failures in the stable interval. The divergent response control set includes control values ​​for the magnitude of distribution deviation, control values ​​for the rate of deviation growth, and the impact value of the number of regression failures in the stable interval.

9. The method for dynamically generating and pushing personalized auxiliary teaching content based on AI large model as described in claim 6, characterized in that: The specific method for obtaining the number of generated parameter sets allowed to be adjusted within a single period, the limit on the adjustment range, and the push interval is as follows: Based on the oscillation intensity value and the divergence intensity value, respectively, the corresponding unit oscillation intensity value influence coefficient and unit divergence intensity value influence coefficient are introduced, multiplied and then added together to obtain the joint influence intensity value; Based on the joint influence intensity value, mapping and matching are performed with the corresponding mapping set, thereby obtaining the number of limit generation parameter sets, the limit adjustment range, and the push interval that can be adjusted within a single period.

Images