MAML-based generative engine optimization content policy meta-learning method and system

Abstract

The embodiment of the application provides a kind of based on MAML's generative engine optimization content strategy meta-learning method and system, belongs to artificial intelligence and big language model application technical field. Including: collect the GEO data of multiple historical big model versions, GEO data is divided into support set and query set according to set proportion to construct meta-training task set;With the content feature in GEO data, query feature as input, with content optimization action as output, construct GEO strategy model and loss function;With GEO strategy model and loss function, with the minimum loss of any meta task in support set after gradient update for a set number of times on query set as target, execute MAML meta-training;Based on the optimal initial parameter, quickly adapt new model or new version.The application is especially suitable for big language model high-frequency version update, cross-model migration scene, learn optimal initial strategy parameter through MAML framework, realize that a small amount of sample and a small amount of gradient update can quickly adapt new model / new version.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and large language model application technology, specifically to a generative engine-based method and system for optimizing content strategy meta-learning based on MAML. Background Technology

[0002] Currently, the iteration speed of Large Language Models (LLMs) continues to accelerate, with iteration cycles shortened to weekly or even shorter. Each version update adjusts and optimizes the model's retrieval logic, ranking preferences, and content generation rules. These changes in underlying logic and core parameters directly lead to a significant decline in the effectiveness of optimized Generative Engine Optimization (GEO) data content, making it unable to continuously meet user retrieval needs and business optimization goals. This poses a significant challenge to the adaptation of GEO content strategies.

[0003] Existing GEO adaptation solutions have at least the following drawbacks: passive re-optimization, requiring data to be collected and training from scratch for each update, taking several days to weeks and incurring high costs; rule-based adaptation, relying on manual experience rules, resulting in poor generalization ability and inability to cope with changes in the underlying logic of the model; slow convergence speed of online learning, poor performance in the early stages of new version launches, and lack of historical pattern reuse capability.

[0004] Therefore, a solution is needed to at least partially address the aforementioned technical deficiencies. Summary of the Invention

[0005] This invention provides a generative engine optimization content strategy meta-learning method and system based on MAML, which is particularly suitable for scenarios involving high-frequency version updates and cross-model transfer of large language models. By learning the optimal initial strategy parameters through the MAML framework, it can quickly adapt to new models / versions with a small number of samples and a small number of gradient updates, thereby at least partially solving the problems of existing technologies.

[0006] The purpose of this invention is to provide a meta-learning method for content strategy optimization in a generative engine based on MAML. This meta-learning method includes: collecting GEO data from multiple historical large model versions and dividing the GEO data of each historical large model version into a support set and a query set according to a set ratio to construct a meta-training task set; using the content features and query features in the GEO data as input and the content optimization action as output, constructing a GEO strategy model and a loss function; using the GEO strategy model and the loss function, with the goal of minimizing the loss on the query set after a set number of gradient updates for any meta-task in the support set, performing MAML meta-training to obtain optimal initial parameters; and quickly adapting to new models or new versions based on the optimal initial parameters.

[0007] On the other hand, the present invention also provides a meta-learning system for content strategy optimization based on MAML in generative engines. The meta-learning system includes: a data acquisition device for collecting GEO data from multiple historical large model versions and dividing the GEO data of each historical large model version into a support set and a query set according to a set ratio to construct a meta-training task set; a model building device for constructing a GEO strategy model and a loss function using the content features and query features in the GEO data as input and the content optimization action as output; a meta-training device for performing MAML meta-training using the GEO strategy model and the loss function, with the goal of minimizing the loss on the query set after any meta-task in the support set is updated by a set number of gradients, to obtain optimal initial parameters; and a model adaptation device for quickly adapting to new models or new versions based on the optimal initial parameters.

[0008] On the other hand, the present invention provides a machine-readable storage medium storing instructions for causing a machine to execute: the content strategy meta-learning method based on MAML-based generative engine optimization described above.

[0009] On the other hand, the present invention provides a processor for running a program, wherein the program is executed to perform: the content strategy meta-learning method for optimizing a generative engine based on MAML as described above.

[0010] On the other hand, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements: the content strategy meta-learning method for optimizing a generative engine based on MAML as described above.

[0011] Through the above technical solutions, this invention proposes a content policy meta-learning method and system based on MAML for Generative Engine Optimization (GEO), involving content policy adaptation technology for generative engine optimization, which can at least partially solve the technical defects in the prior art. This invention is particularly suitable for scenarios involving high-frequency version updates and cross-model transfer of large language models. By learning the optimal initial policy parameters through the MAML framework, it enables rapid adaptation to new models / versions with a small number of samples and a small number of gradient updates.

[0012] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0013] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 The flowchart illustrates a content strategy meta-learning method based on MAML generative engine optimization according to an embodiment of this application. Figure 2 This is a schematic diagram illustrating the overall process according to an embodiment of this application; Figure 3 This is a schematic diagram of the inner / outer loop of MAML meta-training according to an embodiment of this application; Figure 4 This is a comparison chart of MAML and traditional fine-tuning parameter updates according to embodiments of this application; Figure 5 This is a graph illustrating the relationship between meta-generalization error and sample size according to an embodiment of this application; Figure 6 This is a flowchart illustrating a new version rapid adaptation process according to an embodiment of this application; Figure 7 This is a schematic diagram illustrating the structure of a MAML-based generative engine-optimized content strategy meta-learning system according to an embodiment of this application; Figure 8 This is a schematic diagram of the internal structure of a computer device provided in an embodiment of this application, wherein A01-processor; A02-network interface; A03-internal memory; A04-display screen; A05-input device; A06-non-volatile storage medium; B01-operating system; B02-computer program. Detailed Implementation

[0014] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0015] It should be noted that the acquisition, transmission, storage, use, and processing of data in the technical solution of this application all comply with relevant laws and regulations. In the embodiments of this application, certain existing industry solutions such as software, components, and models may be mentioned. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solution of this application, and do not imply that the applicant has already used or necessarily used such solutions.

[0016] To better illustrate the essential features of this invention, a brief introduction to its technical background is provided first. Existing technologies for content adaptation schemes in generative engine optimization (GEO) for large model iterations generally suffer from the following significant drawbacks, making it difficult to meet the efficient adaptation requirements of actual business applications: 1. Passive re-optimization mode is inefficient: After each major model version update, the relevant technical solutions need to re-collect a large amount of GEO sample data and retrain the adaptation strategy model from scratch. The entire training process takes several days to several weeks, which not only consumes a lot of computing resources, but also generates high manpower and time costs. Moreover, the long adaptation period will cause the GEO content effect to remain in a decaying state for a long time, affecting the business experience. 2. Weak generalization ability of rule-based adaptation: Existing adaptation solutions mostly rely on empirical rules summarized by humans. The optimization logic and adjustment standards of GEO content are set manually. This method is greatly affected by subjective factors and cannot cope with fundamental changes in the underlying logic of large models. When the model's retrieval rules and ranking preferences are significantly adjusted, the original empirical rules will become completely ineffective, resulting in extremely poor generalization ability. 3. Poor adaptation effect of online learning: Some solutions adopt online learning for adaptation, but its convergence speed is slow. In the early stage of the launch of the new version of the large model, the adaptation strategy cannot quickly converge to the optimal state, resulting in poor optimization effect of GEO content. Moreover, this method lacks the ability to reuse historical adaptation patterns and cannot use the optimization experience of previous versions to improve the adaptation efficiency and effect of new scenarios.

[0017] To address the aforementioned limitations of existing technologies and overcome the problems of low adaptation efficiency, high cost, and weak generalization ability of GEO content strategies during large model iterations and cross-model transfers, this invention first provides a generative engine optimization content strategy meta-learning method 100 based on Model-Agnostic Meta-Learning (MAML). The overall approach can be found in [reference needed]. Figure 2 As shown: Step S1 constructs the meta-training task set; Step S2 constructs the policy model and loss function; Step S3 performs MAML meta-training (learning the optimal initial parameters θ*); Step S4 quickly adapts to the new version / new model.

[0018] Through the above technical solution, the present invention can accurately solve the following core technical problems existing in the process of large model version iteration and cross-model migration of the current GEO technology: The existing technology suffers from several drawbacks. First, it fails to reuse historical experience. It cannot extract general adaptation patterns from the GEO optimization process of multiple historical large model versions. Each version update or cross-model migration requires the adaptation strategy to be trained almost from scratch, failing to leverage the value of historical data. Second, the adaptation cost is too high. Adapting to new versions of large models requires a large sample size and numerous model gradient updates, resulting in excessively long convergence cycles and hindering rapid adaptation. Third, its generalization ability is weak. Existing strategies can only adapt to a single model or a specific version, leading to poor adaptation performance during cross-model and cross-version migrations and an inability to address the underlying logical differences between different models. Fourth, the adaptation effect lacks quantitative evidence. Existing solutions rely heavily on subjective judgment to evaluate adaptation effectiveness, lacking scientific quantitative evaluation standards and easily influenced by human experience, resulting in unstable adaptation results.

[0019] Through in-depth research, the applicant discovered that the core advantage of MAML lies in its ability to learn universally optimal initial parameters across multiple tasks, enabling new tasks to achieve optimal performance with only a small number of gradient updates. However, to date, no technology has been introduced into the scenario of rapid cross-version and cross-model adaptation of GEO content strategies. In response, this invention innovatively introduces MAML technology into the field of GEO content strategy adaptation. Combining the characteristics of large-scale model iteration with the features of GEO data, a dedicated meta-training process and strategy model are designed. This aims to solve the core problems of low efficiency, high sample consumption, long recovery cycle, and weak generalization ability in existing technologies for GEO content strategy adaptation, achieving rapid and efficient adaptation of large models across versions and models.

[0020] Specifically, such as Figure 1 As shown, the MAML-based generative engine optimization content strategy meta-learning method 100 of the present invention may include steps S110-S140, and the specific implementation of each step will be described in detail in the following content.

[0021] Step S110: Collect GEO data from multiple historical large model versions, and divide the GEO data of each historical large model version into a support set and a query set according to a set ratio to construct a meta-training task set.

[0022] This step is the basic preparatory step of the meta-learning method of this invention. Its purpose is to construct a meta-training task set to provide a standardized and diversified training sample foundation for subsequent meta-training. The core is to construct a task set that is suitable for the meta-learning scenario by hierarchically dividing and selecting samples from multiple versions of GEO data, taking into account both sample representativeness and task diversity, and laying the foundation for improving the generalization ability of the model in the future.

[0023] The meta-training task set constructed in this step supports cross-version migration of the same model and cross-platform migration of different models, thus having a wider range of applications. Specifically, it is divided into two application scenarios: Scenario 1: Migration across versions of the same model. In this case, the multiple historical large model versions are different iterations of the same model, such as the V1, V2 and V3 versions of the large language model launched by the same vendor. In this scenario, the meta-task focuses on extracting the version adaptation rules of the same model. Scenario 2: Cross-platform migration of different models. In this scenario, the multiple historical large model versions are different versions of the model, such as the large language model V2 version of vendor A, the large language model V1 version of vendor B, and the large language model V3 version of vendor C. In this scenario, the meta-task focuses on extracting the commonalities of GEO adaptation rules of different models.

[0024] Specifically, let N be the total number of historical large model versions (N is a positive integer, and N≥3, to ensure task diversity and improve model generalization ability). GEO data for N historical large model versions are collected, and each historical large model version corresponds to a meta-task, that is, the v-th historical large model version corresponds to the v-th meta-task. (v is a positive integer, and 1≤v≤N).

[0025] Then, it is necessary to divide the support set and query set for each meta-task. The corresponding GEO data is divided into support sets according to a set ratio. (e.g., 10-50 samples) and query set This ultimately forms a complete meta-training task set. Through extensive experimental verification, the ratio of the support set to the query set is set to 3:7-6:4. This ratio range ensures that the support set provides sufficient task features for model adaptation while effectively verifying the model's generalization ability through the query set. The preferred ratio is 4:6 or 5:5.

[0026] Furthermore, the GEO data is the core data for generative engine optimization, specifically containing three types of core information: content features, effect tags, and the v-th meta-task. The corresponding query features; whereby the task data corresponding to each meta-task includes: query statement (i.e., the query requirements input by the user into the generative engine, used to represent the direction of user needs), content features x (i.e., the core features of the content output by the generative engine, including but not limited to content semantics, entity distribution, structural hierarchy, etc.), effect tags y, etc. Among them, effect tags are used to quantify the optimization effect of the content output by the generative engine, specifically including ranking, first recommendation probability, and citation strength. The ranking represents the sorting position of the content in the query results, the first recommendation probability represents the probability that the content is recommended to the user first, and the citation strength represents the frequency and weight of the content being cited by other content.

[0027] To further improve the generalization ability of the meta-trained model, this step involves adjusting the support set. A hard sample mining strategy can be adopted: Specifically, samples with large fluctuations in performance across versions (i.e., samples with significantly different output performance in different historical large model versions under the same type of query requirements) are selected as support set samples. By focusing on these representative hard samples, the model can learn the optimization rules across versions and models more accurately, thereby improving the model's generalization ability in subsequent adaptation to new models and new versions.

[0028] Meanwhile, this step provides an alternative to support set samples: support set Active learning sampling can be used to replace hard sample mining. The active learning algorithm selects the samples with the largest information gain and the best representation of task features as the support set, ensuring the representativeness and effectiveness of the samples, avoiding the sample bias problem that may occur in the hard sample mining process, and adapting to the needs of different data distribution scenarios.

[0029] This step can achieve the following technical effects: 1. The constructed meta-training task set takes into account both cross-version scenarios of the same model and cross-platform scenarios of different models, thus expanding the applicability of meta-learning methods. 2. By mining difficult samples or actively learning to sample, the representativeness of the support set samples is improved, providing a high-quality sample foundation for subsequent meta-training; 3. It supports a reasonable ratio of training set to query set and sample size control, which ensures training effect, avoids sample redundancy, and improves the efficiency of subsequent meta-training. 4. The core components of GEO data were clarified, ensuring the integrity and standardization of the data, and providing a unified standard for subsequent model training and performance evaluation.

[0030] In summary, this step proposes a method for constructing a meta-training task set that balances diversity, representativeness, and standardization. By collecting, rationally dividing, and selecting samples from multiple versions of GEO data, it provides a standardized task foundation for the subsequent meta-training of GEO strategy models. At the same time, through sample substitution schemes and scenario adaptation design, it enhances the flexibility and applicability of the solution.

[0031] Step S120: Using the content features and query features in the GEO data as input and the content optimization action as output, construct the GEO strategy model and loss function.

[0032] This step is to build the policy model and loss function. The core is to clarify the input-output relationship of the GEO policy model, the model architecture selection, and the quantitative design of the loss function. This ensures that the model can accurately learn the mapping relationship between content features, query features, and content optimization actions. At the same time, the quantitative loss enables the model training effect to be evaluated and reproducible.

[0033] First, define the GEO strategy model. The model takes content features and query features from GEO data as joint inputs and outputs content optimization actions to optimize the content of the generative engine. Here, θ represents all learnable initial parameters of the GEO strategy model. The initial parameters θ are randomly initialized to provide a basis for parameter optimization in subsequent meta-training.

[0034] The content optimization actions are specific operations used to improve the output effect of the generative engine. Its core can include: entity density adjustment (i.e., adjusting the distribution density of core entities in the output content to avoid semantic redundancy due to excessively dense entities, or loss of core information due to excessively sparse entities), structural modification (i.e., adjusting the paragraph structure and logical hierarchy of the output content to improve the readability and logic of the content), and evidence type optimization (i.e., adjusting the types of evidence used to support viewpoints in the output content, such as prioritizing authoritative data sources and highly credible evidence to improve the persuasiveness of the content).

[0035] Regarding the architecture selection for the GEO strategy model, this step provides several adaptation options. The model can adopt any one of the following: MLP (Multi-Layer Perceptron), Transformer, or LSTM (Long Short-Term Memory). Among them, the MLP architecture has a simple structure and high computational efficiency, making it suitable for GEO data with low feature dimensionality; the Transformer architecture has powerful feature extraction and association modeling capabilities, making it suitable for scenarios with more complex content and query features; and the LSTM architecture excels at processing sequence features, making it suitable for GEO data with temporal characteristics (such as model version data at different time points).

[0036] Meanwhile, this step provides model alternatives: the policy model can be replaced by lightweight Transformer or CNN (Convolutional Neural Network) instead of MLP / LSTM. Lightweight Transformer retains the core feature extraction capability of Transformer, but reduces computational overhead by simplifying the network structure, making it suitable for resource-constrained scenarios. CNN is good at extracting local features and is suitable for scenarios where content features contain a lot of local semantic information, thereby achieving adaptation to different feature dimensions and different resource scenarios.

[0037] Secondly, a loss function needs to be defined, which can be set as follows: , used to characterize the v-th meta-task In the initial parameters θ, the dataset D (i.e., the metatask) The task loss on the support set and query set data is used to measure the deviation between the prediction effect of the GEO strategy model and the actual optimization needs, and to provide direction for model parameter updates.

[0038] in, The error function or quantization loss is used to measure the optimization actions of the predicted content. With the effect label The deviation between them. Specifically, when using an error function, it is the error function between the predicted optimized action and the optimal action (such as mean squared error, cross-entropy error). The optimal action is the content optimization action that can make the effect label reach the best (such as the highest ranking, the highest first recommendation probability, the strongest citation strength). When using quantized loss, it is a quantized loss based on ranking and first recommendation rate (such as ranking loss, accuracy loss). The quantized index directly measures the degree to which the optimized action output by the model improves the effect of the generative engine.

[0039] Crucially, the loss function and model parameter update process in this step adopt a fully quantitative design, that is, all loss calculations and parameter updates are achieved through quantitative indicators, and there is no fuzzy qualitative evaluation. This ensures that the model training process and adaptation effect are assessable and reproducible, and avoids performance deviations caused by inconsistent evaluation standards.

[0040] The technical effects that can be achieved by this step are: 1. The constructed GEO strategy model clarifies the input-output relationship and provides multiple model architectures and alternatives to adapt to different feature dimensions and resource scenarios, thereby improving the flexibility of the solution; 2. The defined content optimization actions are highly targeted and can directly affect the content optimization of the generative engine, ensuring that the model output has practical application value; 3. The fully quantized loss function design enables the model training effect to be evaluated and reproduced, providing a clear evaluation standard for subsequent parameter optimization in meta-training; 4. The diverse design of the model architecture balances computational efficiency and training effectiveness, making it adaptable to different practical application scenarios.

[0041] In summary, this step proposes a highly adaptable GEO strategy model and loss function construction scheme with clear evaluation criteria. By clarifying the model input and output, providing diverse architecture options, and designing a fully quantized loss function, it provides a core model foundation and evaluation basis for subsequent MAML meta-training, ensuring that the meta-training process can be carried out accurately and efficiently.

[0042] Step S130: Using the GEO policy model and the loss function, with the goal of minimizing the loss on the query set after a set number of gradient updates for any meta-task in the support set, perform MAML meta-training to obtain the optimal initial parameters.

[0043] Step S130, which involves performing MAML meta-training, is a crucial step in this invention. Its core objective is to minimize the query set loss after a small number of gradient updates for any meta-task. This is achieved through the inner and outer loop gradient update mechanisms of MAML, learning universally optimal initial parameters. This allows for the reuse of historical adaptation patterns, laying the foundation for rapid adaptation in the future.

[0044] The goal of learning the optimal initial parameters is to minimize the loss on the query set after a small number of gradient updates for any meta-task. Essentially, this involves ensuring the initial parameters have strong generalization ability, enabling them to quickly adapt to different meta-tasks. Through in-depth research, the applicant discovered that, referring to… Figure 4 Comparison chart of MAML and traditional fine-tuning parameter updates and Figure 5 The graph shows the relationship between meta-generalization error and sample size. The core advantage of MAML lies in its ability to learn universally optimal initial parameters across multiple tasks, allowing new tasks to achieve optimal performance with only a small number of gradient updates. This technology has been applied in multiple fields and its effectiveness has been verified. However, to date, no technology has been introduced into the cross-version, cross-model rapid adaptation scenario of GEO content strategies, failing to address the technical pain points unique to the GEO domain, such as low adaptation efficiency, large sample consumption, and long effect recovery cycle.

[0045] The MAML meta-training involves iterating the single iteration process multiple times until the initial parameter θ converges. For example, the convergence criterion is that the parameter change is less than a set threshold for three consecutive iterations, or the average loss on the query set is less than a set loss threshold. The MAML meta-training can be performed using either the first-order MAML algorithm (FOMAML) or the Reptile algorithm: the first-order MAML algorithm omits the calculation of the second derivative, significantly reducing computational overhead and improving meta-training efficiency while ensuring training effectiveness; the Reptile algorithm approximates the optimal initial parameters through multiple iterations, resulting in a more stable training process and adaptability to scenarios with large fluctuations in data distribution.

[0046] Meanwhile, this step provides alternative algorithms: MAML can be replaced with meta-learning algorithms such as ProtoNets (prototype networks) and ANIL (Almost No Inner Loop). The overall training process remains unchanged after the replacement, only the inner and outer loop gradient update mechanisms of MAML are replaced with the parameter update mechanisms of the corresponding algorithms, adapting to different meta-learning scenario requirements and further improving the flexibility of the solution.

[0047] Specifically, you can refer to Figure 3 The diagram illustrates the inner / outer loop of MAML meta-training. The single iteration process is detailed below, and its core consists of two stages: the inner loop (task adaptation) and the outer loop (meta-parameter update). Step S131, Task Sampling: Randomly select batch tasks from the meta-training task set. .

[0048] For example, the batch task size can be 2-8 meta-tasks. This batch size has been experimentally verified to ensure the diversity of training while avoiding the computational pressure caused by excessively large batches.

[0049] Step S132, Inner Loop (Task Adaptation): The learning rate of the inner loop can be set. For each meta-task In its support set The gradient update is performed a predetermined number of times using the above formula to obtain the meta-task. Adaptation parameters : .

[0050] For example, the value of α ranges from 0.001 to 0.01, and is used to control the gradient update step size of a single meta-task, avoiding parameter oscillations caused by an excessively large step size and slow training caused by an excessively small step size. The number of iterations is set to K (K is a positive integer), meaning that gradient updates are performed K times for each meta-task, achieving rapid adaptation of the meta-task.

[0051] Step S133, Outer Loop (Meta-Update): Set the outer loop learning rate. Based on this meta-task Adaptation parameters In query set The cumulative loss is used to perform a meta-update on the initial parameter θ using the following formula: .

[0052] For example, the value of β ranges from 0.0001 to 0.001, which is less than the inner loop learning rate α. It is used to control the update step size of the initial parameter θ to ensure the universality of the initial parameter.

[0053] The above single-iteration process can be repeated until the initial parameters θ converge, ultimately yielding the optimal initial parameters. .

[0054] In addition, the set number of times is K. Before performing MAML meta-training, the meta-learning method further includes: determining the range of values ​​for K based on the version differences of the multiple historical large model versions, so as to achieve dynamic adaptation of the gradient update number and improve training effect and efficiency.

[0055] The version difference metric is used to quantify the differences between different historical large model versions. Specifically, it is determined by calculating the similarity of content features and effect labels in the GEO data of different versions (the lower the similarity, the greater the version difference). A version difference threshold γ is set (γ ranges from 0.3 to 0.5 and can be adjusted according to the actual model type). When the version difference metric is less than the threshold γ, it indicates that the differences between different model versions are small, and adaptation can be achieved without excessive gradient updates; therefore, the value of K ranges from 1 to 5. When the version difference metric is greater than the threshold γ, it indicates that the differences between different model versions are large, requiring more gradient updates to achieve adaptation; therefore, the value of K ranges from 5 to 10. This design achieves dynamic adjustment of the adaptation steps: K = 5-10 when the version difference is large, ensuring adaptation effect; K = 1-5 when the version difference is small, improving training efficiency.

[0056] The core advantage of this step is that the number of adaptation samples is reduced to 10-50, and the gradient update takes 1-5 steps (when the version difference is small). Compared with the traditional method of training from scratch, which requires a large number of samples and several days of training time, the adaptation time is shortened from days to minutes, which greatly improves the model adaptation efficiency.

[0057] The technical effects that can be achieved by this step are: 1. Through MAML meta-training and various alternative algorithms, the optimal initial parameters with strong generalization ability are learned, enabling the reuse of historical adaptation patterns and providing a foundation for subsequent rapid adaptation; 2. The reasonable setting of the inner and outer loop gradient update mechanism and the learning rate ensures the stability and efficiency of meta-training and avoids problems such as parameter oscillation or slow training. 3. The dynamic adaptation design of gradient update count K adjusts the K value according to the version difference, balancing the adaptation effect and training efficiency. 4. Significantly reduced the number of adaptation samples and gradient updates, shortening the adaptation time from days to minutes, significantly improving training efficiency and reducing computational costs; 5. The provision of algorithm alternatives further expands the applicability of the solution and adapts it to different meta-learning scenarios.

[0058] In summary, this step proposes an efficient and flexible MAML meta-training scheme. Through inner and outer loop gradient update mechanisms, dynamically adapted gradient update times, and diverse algorithm selection, it efficiently learns the optimal initial parameters, realizes the reuse of historical adaptation patterns, significantly improves training efficiency, and provides core parameter support for the rapid adaptation of subsequent new models and versions.

[0059] Step S140: Quickly adapt the new model or new version based on the optimal initial parameters.

[0060] Step S140 is designed for rapid adaptation to new versions / models. Its core function is to leverage the strong generalization ability of the optimal initial parameters θ* and quickly obtain the parameters for adapting to the new model / version through gradient updates on a small number of samples. This eliminates the need for training from scratch, significantly shortening the adaptation cycle and reducing adaptation costs.

[0061] First, you can refer to Figure 6 The new version quick adaptation flowchart is as follows. The adaptation process is divided into the following core steps. The process is simple and efficient, as detailed below.

[0062] Step S141: Collect sample data from the new model or the new version as a new support set. ; The new support set The sample size is 10-50, which is consistent with the sample size of the support set in step S110. This eliminates the need to collect a massive number of samples, further reducing the adaptation cost and adaptation cycle.

[0063] Step S142, from the optimal initial parameters Starting from the support set Perform gradient updates a predetermined number of times to obtain the adaptation parameters. ; The value of the set number of iterations K still follows the dynamic adaptation rule in step S130, that is, the value of K is determined to be between 1 and 10 based on the difference between the new model / version and the historical large model version, to ensure the adaptation effect. The specific gradient update formula is as described in step S132. The inner loop learning rate α remains consistent with that in the meta-training stage and does not need to be readjusted, further improving the adaptation efficiency.

[0064] Step S143, using the adaptation parameters Optimize the content of the new model or the new version to complete the adaptation.

[0065] That is, by outputting corresponding content optimization actions (entity density adjustment, structural modification, evidence type optimization) through the GEO strategy model, the corresponding actions are applied to the new model / new version of the generative engine to complete the entire adaptation process, so that the new model / new version can quickly achieve the optimal output effect.

[0066] Meanwhile, this step provides an alternative deployment solution: adopting a deployment mode of "training meta-parameters in the cloud and performing rapid adaptation on the device side". The meta-training process of the optimal initial parameters θ* is completed in the cloud (utilizing the powerful computing resources of the cloud to improve the efficiency of meta-training), while the rapid adaptation process of the new model / new version is performed on the device side (such as local server, terminal device). There is no need to upload the device side data to the cloud, which reduces the computing pressure on the cloud and meets the needs of low-latency scenarios (such as real-time adaptation and local deployment scenarios), further improving the practicality of the solution.

[0067] The technical effects that can be achieved by this step are: 1. Based on the optimal initial parameters θ*, only 10-50 samples and 1-10 gradient updates are needed to complete the adaptation of the new model / version, reducing the adaptation time from days to minutes, greatly improving the adaptation efficiency and reducing the adaptation cost; 2. The adaptation process is simple, requiring no redesign of the model architecture and loss function, enabling the reuse of historical adaptation patterns and reducing the difficulty of adaptation; 3. The provision of alternative deployment solutions adapts to different scenario requirements such as low latency and local deployment, improving the practicality and flexibility of the solutions; 4. Adaptation parameters It can accurately adapt to the features of new models / versions, ensure content optimization effects, and improve the output quality of generative engines.

[0068] In summary, this step proposes an efficient and convenient solution for rapid adaptation of new models / versions. By reusing the optimal initial parameters obtained from meta-training and combining gradient updates with a small number of samples, rapid adaptation of generative engines is achieved, significantly shortening the adaptation cycle and reducing adaptation costs. At the same time, by deploying alternative solutions, the applicable scenarios of the solution are further expanded, enhancing its practical application value.

[0069] Overall, this invention is a meta-learning method for content strategy optimization based on MAML in generative engines. Compared with existing technologies, it overcomes the drawbacks of traditional generative engine optimization, such as needing to train from scratch, requiring a large number of samples, having a long adaptation period, and having a narrow scope of application. It has the following significant advantages: 1. Innovation: This is the first time that the MAML meta-learning method has been applied to the field of GEO (Generative Engine Optimization), which enables the reuse of historical adaptation patterns and avoids the dilemma of needing to train from scratch for each new model and version in traditional methods, thus greatly reducing the difficulty and cost of adaptation. 2. High efficiency: The number of adapted samples is reduced to 10-50, and gradient updates only require 1-5 steps (when the version difference is small), which shortens the traditional adaptation time from days to minutes, significantly improving the adaptation efficiency of the generative engine and meeting the needs of rapid iteration. 3. Wide applicability: Supports cross-version migration of the same model and cross-platform migration of different models, adapts to different architectures, different deployment platforms and different versions of generative large models, with a wider range of applications and greater flexibility; 4. Reproducibility: The loss function and parameter update process are fully quantified, and all training and adaptation processes have clear quantitative indicators, ensuring that the adaptation effect is assessable and reproducible, avoiding the effect bias caused by qualitative evaluation, and improving the reliability of the solution.

[0070] On the other hand, the present invention also provides a MAML-based generative engine optimization content strategy meta-learning system 200, specifically, as follows: Figure 7 As shown, the learning system 200 may include the following four functional devices: The data acquisition device 210 is used to collect GEO data from multiple historical large model versions and divide the GEO data of each historical large model version into a support set and a query set according to a set ratio to construct a meta-training task set. The model building device 220 is used to construct a GEO strategy model and a loss function by taking the content features and query features in the GEO data as inputs and the content optimization action as outputs. Meta-training device 230 is used to perform MAML meta-training using the GEO policy model and the loss function, with the objective of minimizing the loss on the query set after a predetermined number of gradient updates for any meta-task in the support set, to obtain optimal initial parameters; and The model adaptation device 240 is used to quickly adapt a new model or new version based on the optimal initial parameters.

[0071] As can be seen, this invention provides an optimization system with a clear structure, well-defined division of labor, and coordinated operation of various devices. It can accurately correspond to the entire process of the aforementioned MAML-based generative engine optimization content strategy meta-learning method, transforming the technical logic of the method into functional modules that can be practically deployed and executed. It achieves the integration of meta-training task construction, strategy model building, optimal parameter learning, and rapid adaptation of new models, effectively solving the problems of low adaptation efficiency, large sample requirements, and narrow applicability of traditional generative engine optimization systems, ensuring the effective implementation of the method and the effective realization of its technical effects.

[0072] Other advantages and benefits can be found in the description of the above optimization methods, and will not be repeated here.

[0073] Example To better illustrate the technical effects of the present invention, specific embodiments are provided. It should be understood that these embodiments are only used to further describe the present invention in detail and should not be construed as limiting the scope of protection of the present invention; those skilled in the art can make some non-essential improvements and adjustments to the present invention based on its technical content, and such improvements and adjustments should all fall within the protection scope of the claims of the present invention.

[0074] This embodiment requires execution based on the optimization method described in steps S110-S140 above. The overall approach can be referred to... Figure 2 The overall flowchart is divided into two specific implementations, corresponding to cross-version adaptation of the same model and cross-platform migration of different models, respectively, covering the core application scenarios of the present invention and intuitively verifying the technical effects, efficiency and practicality of the present invention.

[0075] Example 1: Cross-version adaptation of the same model (DeepSeek V3.5 adaptation) Task Construction: Four historical versions of DeepSeek V2, DeepSeek V2.5, DeepSeek V3, and DeepSeek V3.1 were used as four meta-tasks. Each meta-task collected 30 GEO data points, which were divided into 15 support sets (using hard sample mining to select samples with large fluctuations in performance across versions) and 15 query sets in a 1:1 ratio to construct the meta-training task set. Model setup: A 2-layer MLP was selected as the GEO strategy model, with 80 input dimensions (30 for content features and 50 for query features) and output dimensions corresponding to 10 types of content optimization actions (covering sub-actions such as entity density adjustment, structural modification, and evidence type optimization). Meta-training: Inner loop learning rate α=0.005, gradient update times K=1 (because the difference between the four historical versions is less than the set threshold γ=0.4), outer loop learning rate β=0.0005, the initial parameters converge after 5000 iterations, and the optimal initial parameters θ* are obtained; Adaptation effect: For the new version of DeepSeek V3.5, only 15 samples were collected as the new support set. Starting from the optimal initial parameters θ*, a 1-step gradient update was performed to complete the adaptation. After adaptation, the first recommendation rate of the new model recovered to 41%, and the overall adaptation time was only 1.5 hours. In contrast, the traditional optimization method requires the collection of 500+ samples, takes 3 days to train, and the final first recommendation rate is only 20%, which fully verifies the efficiency of the present invention.

[0076] Example 2: Cross-model transfer adaptation (DeepSeek → Wenxin Yiyan) Meta-training preparation: Using GEO data from multiple historical versions of DeepSeek (V2-V3.1) as meta-training tasks, construct the meta-training task set, build the GEO policy model (using the Transformer architecture) according to the aforementioned steps, and perform MAML meta-training to learn and obtain the optimal initial parameters θ*. Adaptation process: 20 GEO samples from the Wenxin Yiyan model are collected as a new support set (active learning sampling is used to ensure sample representativeness). The inner loop learning rate is set to α=0.005, and a two-step gradient update is performed to obtain the adaptation parameters θ for the Wenxin Yiyan model. ne w'; Adaptation effect: Compared with traditional cross-model adaptation methods, the sample efficiency of this embodiment is improved by 15 times. There is no need to train the model from scratch. After adaptation, the model quickly reaches the commercially viable effect, which verifies the practicality and adaptability of the present invention in cross-platform migration scenarios of different models.

[0077] In summary, the beneficial effects of the present invention are as follows: 1. Deeply integrate the MAML framework with the GEO content strategy to innovatively construct a cross-version, cross-model meta-learning paradigm, enabling the reuse of historical adaptation patterns; 2. Design a dedicated two-layer gradient update mechanism (inner loop task adaptation, outer loop meta-parameter update) adapted to the GEO scenario to achieve rapid adaptation of a small number of samples (10-50); 3. Employing a fully quantized loss function design, the entire process of model training and parameter updates is calculable and quantifiable, avoiding performance bias caused by subjective judgment; 4. It supports two adaptation scenarios simultaneously: iteration with the same model and cross-platform adaptation with different models, expanding the scope of application and improving the flexibility and practicality of the solution.

[0078] The key points of this invention are as follows: 1. A full-process method for GEO content strategy meta-learning based on MAML, including all technical processes and parameter settings of the aforementioned steps S110-S140; 2. Support set / query set partitioning rules for GEO tasks (3:7-6:4 ratio, 10-50 support set samples) and hard sample selection rules, and active learning sampling alternatives; 3. MAML inner / outer loop parameter update formulas, learning rate settings (inner loop α=0.001-0.01, outer loop β=0.0001-0.001) and algorithm selection (first-order MAML, Reptile, etc.) adapted to GEO scenarios. 4. The system modules for executing this method include the functional definitions and collaborative working methods of the data acquisition device 210, model building device 220, meta-training device 230, and model adaptation device 240, as well as alternative cloud-edge deployment schemes.

[0079] In addition, embodiments of the present invention also provide a machine-readable storage medium storing instructions for causing a machine to execute: a content strategy meta-learning method for optimizing a generative engine based on MAML as described above.

[0080] This invention provides a processor for running a program, wherein the program is executed to perform: a content strategy meta-learning method for optimizing a generative engine based on MAML as described above.

[0081] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 8As shown in the figure, the computer device may include a processor A01, a network interface A02, a display screen A04, an input device A05, and a memory (not shown) connected via a system bus. The processor A01 provides computing and control capabilities. The memory may include internal memory A03 and a non-volatile storage medium A06. The non-volatile storage medium A06 stores an operating system B01 and a computer program B02. The internal memory A03 provides an environment for the operation of the operating system B01 and the computer program B02 stored in the non-volatile storage medium A06. The network interface A02 is used to communicate with external terminals via a network connection. When the computer program is executed by the processor A01, it implements the MAML-based generative engine optimization content strategy meta-learning method described above. The display screen A04 may be a liquid crystal display (LCD) or an e-ink display. The input device A05 may be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0082] Those skilled in the art will understand that Figure 8 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0083] In one embodiment, the MAML-based generative engine optimization content strategy meta-learning method provided in this application can be implemented as a computer program, which can be implemented in, for example... Figure 8 The method runs on the computer device shown. The computer device's memory can store various program modules that constitute the MAML-based generative engine-optimized content strategy meta-learning method. The computer program, composed of these program modules, causes the processor to execute the various steps in the MAML-based generative engine-optimized content strategy meta-learning method of the various embodiments of this application described in this specification.

[0084] In one embodiment, this application also provides a computer program product that, when executed on a data processing device, is adapted to execute a program that initializes the various steps of the above-described method.

[0085] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application 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.

[0086] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. 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... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0087] 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.

[0088] 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 functional steps specified in one or more boxes.

[0089] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0090] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0091] Computer-readable media include both permanent and non-permanent, removable and non-removable media that can store information by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0092] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0093] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A meta-learning method for optimizing content strategies based on MAML generative engines, characterized in that, The meta-learning methods include: Collect GEO data from multiple historical large model versions, and divide the GEO data of each historical large model version into a support set and a query set according to a set ratio to construct a meta-training task set; Using the content features and query features in the GEO data as input and the content optimization action as output, a GEO strategy model and loss function are constructed. Using the GEO policy model and the loss function, with the objective of minimizing the loss on the query set after a set number of gradient updates for any meta-task in the support set, MAML meta-training is performed to obtain the optimal initial parameters; and The new model or version is quickly adapted based on the optimal initial parameters.

2. The meta-learning method according to claim 1, characterized in that, The multiple historical large model versions are versions of different models, and the meta-learning method is applied to cross-platform transfer of different models; or The multiple historical large model versions are different versions of the same model, and the meta-learning method is applied to cross-version transfer of the same model.

3. The meta-learning method according to claim 1 or 2, characterized in that, Each version of the historical master model corresponds to a meta-task. ( , The total number of historical large model versions and ), The vth meta-task The corresponding support set is Query set is The meta-training task set is , The GEO data includes: content features Effect tags and the vth meta-task Corresponding query features .

4. The meta-learning method according to claim 3, characterized in that, The GEO strategy model is any one of the following: MLP, Transformer, CNN, LSTM; The GEO strategy model is set as follows: θ represents all learnable initial parameters of the GEO strategy model; The loss function is set as follows: Represents the v-th meta-task The task loss on initial parameters θ and dataset D. in, The error function or quantization loss is used to measure the optimization actions of the predicted content. With the effect label The deviation between them.

5. The meta-learning method according to claim 4, characterized in that, The content optimization actions include: entity density adjustment, structural modification, and evidence type optimization; The effect label y includes: ranking, probability of being recommended first, and intensity of citation.

6. The meta-learning method according to claim 4, characterized in that, The MAML meta-training includes multiple iterations of a single iteration process until the initial parameter θ converges, wherein the single iteration process includes: Randomly select batches of tasks from the meta-training task set. ; Set the inner loop learning rate For each meta-task In its support set The gradient update is performed a predetermined number of times using the above formula to obtain the meta-task. Adaptation parameters : ; Set the outer loop learning rate Based on this meta-task Adaptation parameters In query set The cumulative loss is used to perform a meta-update on the initial parameter θ using the following formula: 。 7. The meta-learning method according to claim 1, characterized in that, The rapid adaptation of the new model or version based on the optimal initial parameters includes: Collect sample data from the new model or the new version as a new support set. ; From optimal initial parameters Starting from the support set Perform gradient updates a predetermined number of times to obtain the adaptation parameters. ; Using the aforementioned adaptation parameters Optimize the content of the new model or the new version to complete the adaptation.

8. The meta-learning method according to any one of claims 1-7, characterized in that, The support set is obtained through hard sample mining or active learning sampling; The ratio of the support set to the query set is set to 3:7-6:4; The MAML meta-training is performed using a first-order MAML algorithm, ProtoNets, ANIL, or Reptile algorithm.

9. The meta-learning method according to any one of claims 1-8, characterized in that, The set number of iterations is K. Before performing MAML meta-training, the meta-learning method further includes: Based on the version differences among the multiple historical large model versions, the range of values ​​for K is determined. Specifically, when the version difference is less than the set threshold, the value of K ranges from 1 to 5, and when the version difference is greater than the set threshold, the value of K ranges from 5 to 10.

10. A meta-learning system for optimizing content strategies based on a generative engine using MAML, characterized in that, The meta-learning system includes: The data acquisition device is used to collect GEO data from multiple historical large model versions and divide the GEO data of each historical large model version into a support set and a query set according to a set ratio to construct a meta-training task set. The model building device is used to construct a GEO strategy model and a loss function by taking the content features and query features in the GEO data as input and the content optimization action as output. The meta-training apparatus is used to perform MAML meta-training using the GEO policy model and the loss function, with the objective of minimizing the loss on the query set after a predetermined number of gradient updates for any meta-task in the support set, to obtain optimal initial parameters; and A model adaptation device is used to quickly adapt a new model or version based on the optimal initial parameters.

11. A machine-readable storage medium, characterized in that, The machine-readable storage medium stores instructions for causing the machine to execute: the MAML-based generative engine optimization content strategy meta-learning method according to any one of claims 1-9.

12. A processor, characterized in that, Used to run a program, wherein the program is run to execute: a MAML-based generative engine optimization content strategy meta-learning method according to any one of claims 1-9.

13. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements: the MAML-based generative engine optimization content strategy meta-learning method according to any one of claims 1-9.

Images