A Large-Scale Ecosystem Online Evolutionary Learning Method Based on End-to-Cloud Collaboration and Its Application

Abstract

This invention provides an online evolutionary learning method and application for a large-scale model ecosystem based on edge-cloud collaboration, and applies it to the e-commerce field, achieving efficient training and deployment of e-commerce models. The method employs an edge-cloud collaborative architecture, deploying a large language model in the cloud and using small to medium-sized language models on the edge. The two work collaboratively to achieve efficient training and inference of the large-scale model ecosystem. During training, the cloud model acts as a teacher model, generating pseudo-labels to supplement the dataset and guide the optimization training of the edge model. Simultaneously, a multi-model collaborative self-training mechanism is introduced, where multiple models play different roles and discuss with each other, autonomously generating training labels to further optimize the edge model. Meanwhile, user feedback information is incorporated into the prompts of the cloud model for contextual learning, thereby continuously optimizing model performance. This invention applies this method to e-commerce scenarios, successfully building and deploying an e-commerce model ecosystem, realizing intelligent development in this field.

Description

Technical Field

[0001] This invention relates to the field of e-commerce technology, specifically to a method and application of online evolutionary learning for a large-scale model ecosystem based on end-to-end cloud collaboration. Background Technology

[0002] With the rapid development of artificial intelligence technology, language models are increasingly widely used in the field of natural language processing, especially demonstrating powerful capabilities in tasks such as customer service, content generation, and data analysis. However, existing customer service systems still face many challenges, particularly in efficiently responding to changing user needs and ensuring data security. Taking the e-commerce sector as an example, customer service systems on e-commerce platforms undertake multiple tasks, including providing consultation and services, answering questions, and handling complaints. These tasks not only require rapid response but also a high degree of contextual understanding and personalized service capabilities.

[0003] Currently, mainstream customer service solutions can be mainly divided into the following categories:

[0004] Human customer service: Relies on a large number of people to handle customer inquiries and complaints, and can provide personalized services, but its scalability is limited due to high costs and difficulty in dealing with a large number of inquiries during peak periods.

[0005] Rule-based automated customer service systems use preset rules and keywords to match and answer questions. While they can handle some simple customer questions, they struggle to address complex and diverse customer needs, and their intelligence and accuracy remain limited.

[0006] Systems based on traditional machine learning or simple deep learning: Some systems use machine learning algorithms such as decision trees and support vector machines, or models such as recurrent neural networks. Although these methods have improved the level of intelligence to some extent, they are still insufficient when faced with complex scenarios.

[0007] In recent years, the emergence of large-scale models (such as GPT-3 and Gemini) has brought tremendous progress to the field of natural language processing. Large language models, pre-trained on massive amounts of data, possess the ability to handle complex language tasks. According to the scaling law of large models, the final performance of the model is related to the computational cost, the number of model parameters, and the data size. When the other two factors are not constrained, the model performance exhibits a power-law relationship with each factor. As the scale of model parameters increases, the performance of large models gradually improves, but the computational requirements also increase significantly, making it difficult to run them directly on edge devices with limited computing power.

[0008] To address this issue, cloud computing has been introduced. Taking e-commerce as an example, large models can be deployed in the cloud, and edge devices can process customer inquiries by calling the inference capabilities of the cloud models. However, while this cloud-based solution improves the availability of large models, it also introduces risks of communication latency and privacy breaches, especially when handling sensitive e-commerce customer data, which poses significant security vulnerabilities.

[0009] Against this backdrop, edge-cloud collaboration technology is gradually becoming a major technological development trend in customer service systems. Edge-cloud collaboration is a method of collaboratively completing tasks in a distributed computing environment by coordinating computing resources in the cloud and at the edge. This technology combines the advantages of cloud computing and edge computing, enabling models of different scales to be deployed in the cloud and on the edge respectively for collaborative inference and training. In e-commerce scenarios, edge devices can deploy lightweight models to handle simple or privacy-related data, while the cloud handles complex tasks. Through this collaborative approach, the computing resources of both the cloud and the edge can be fully utilized, improving customer service response speed and reducing the risk of privacy leaks.

[0010] However, existing edge-cloud collaborative systems still have some shortcomings. For example, in e-commerce customer service, collaborative optimization between cloud-based and edge-based models is often not achieved. Cloud-based models cannot adequately guide the optimization training of edge-based models, and edge-based models cannot provide feedback to cloud-based models based on user feedback. Furthermore, edge-based models still rely heavily on manually labeled data for supervision during training, but labeled data for specific domains is scarce, making model training difficult and costly. In addition, most large models rely on offline updates, making it difficult to quickly adapt to the rapidly changing customer needs and market environment in e-commerce. These problems limit the response speed and accuracy of e-commerce platform customer service systems when facing complex and ever-changing customer demands.

[0011] Online evolutionary learning, as an adaptive learning method, can continuously optimize models based on user feedback and environmental changes, enhancing their personalized service capabilities. For example, in e-commerce scenarios, online evolutionary learning can help large language models continuously evolve based on real-time user feedback, thereby providing more accurate and personalized responses and contributing to improved customer service experiences in e-commerce.

[0012] In summary, existing customer service systems and large-scale model deployment solutions still fall short when dealing with complex scenarios. However, an edge-cloud collaborative architecture can not only efficiently coordinate cloud and edge resources, but also continuously optimize models in a dynamic environment by introducing multi-model self-training mechanisms and online evolutionary learning, providing more personalized services. Applying this technical solution to the e-commerce field is expected to improve the intelligence level and service quality of e-commerce customer service systems. Summary of the Invention

[0013] To address the shortcomings of existing technologies, this invention provides a method and application for online evolutionary learning of a large model ecosystem based on edge-cloud collaboration. When applied to the e-commerce field, it designs a multi-model collaborative discussion mechanism and introduces multiple role models to effectively simulate the human team collaborative discussion process. This enables the models to autonomously generate training data during the writing process, reducing the reliance on manually labeled data. At the same time, by introducing human feedback and strategies such as online fine-tuning and contextual learning, it effectively realizes the online evolution of the models on both sides.

[0014] To achieve the above objectives, the present invention provides the following technical solution: a method and application for online evolutionary learning of a large-scale model ecosystem based on edge-cloud collaboration, specifically including the following steps.

[0015] S1. Multi-model role allocation

[0016] Multiple small edge models and one large cloud model are assigned different roles to facilitate effective discussion and collaboration during training. The large cloud model acts as the teacher model, providing prior input and constraints for multi-model discussions. The roles of the small edge models are assigned as follows:

[0017] Questioner Model: Focuses on asking industry-related questions, simulating the way real users ask questions.

[0018] Respondent Model: Responsible for generating responses based on industry knowledge and task objectives, improving the accuracy and professionalism of the replies.

[0019] S2. Cloud-based large model prior input and constraint support

[0020] During each round of collaborative discussion, the large cloud model supports the training of the smaller model by providing prior inputs and constraints. Specifically, in each round of dialogue, the large cloud model has two roles: generating scene priors and evaluating the quality of the generated content.

[0021] Scenario prior generation: Provides core industry knowledge, such as the best answer patterns in professional services, techniques for personalized services, and requirements for handling privacy data, so that small models have a unified standard reference during training.

[0022] Quality assessment: When the small model generates answers, the large model performs quality assessment and feedback on these answers. This constraint guides the small model to generate content that better meets industry needs, ensuring the consistency and accuracy of the models.

[0023] S3. Multi-model collaborative discussion

[0024] During the collaboration process, the models on each edge discuss and generate questions and answers for each other, forming multi-turn dialogues. This discussion mechanism enables the models to learn and understand various scenario requirements through interaction.

[0025] Prior input: Before the questioner model and the answerer model generate questions and answers, the scenario prior input of the large model is given to the questioner and answerer models to provide a prior knowledge base for the training of the small model on the edge, thereby improving the quality of question and answer generation.

[0026] Question generation: The questioner model generates questions based on industry needs and task scenarios.

[0027] Answer generation: The responder model answers the questions generated by the questioner model by providing detailed solutions, analyzing scenario requirements, and improving service quality.

[0028] Quality verification: After each round of question and answer, the dialogue generated by the edge model is evaluated and corrected using a large model, which accelerates the convergence of multi-model training on the edge.

[0029] S4. Generation and Online Evolution of Co-training Data

[0030] During the discussion and collaboration process, through multiple rounds of dialogue, the responses generated by the smaller models are constrained and validated by the larger model, labeled as high-quality training data, and added to the new dataset. The generated collaborative data and efficient parameter fine-tuning techniques (such as low-rank adaptive fine-tuning or prefix fine-tuning) are used to train the smaller models, enabling them to achieve optimized performance on specific tasks. Through a regular edge-cloud collaborative multi-model discussion and training mechanism, cloud and edge models achieve continuous online optimization. The edge and cloud models learn from and collaborate with each other, continuously improving their ability to respond to industry scenarios, achieving adaptive learning and performance optimization. The online-evolving edge smaller models will have stronger responsiveness, able to adaptively handle specific industry service tasks based on collaboratively trained data, and achieve higher-quality responses.

[0031] Preferably, in step S2, the adaptive and contextual learning capabilities of the large model are used to perform contextual modeling on the customer service dialogue dataset in a specific domain, enabling it to handle specific complex problems in that domain scenario and to assume the prior and constraint role of collaboration with the small model.

[0032] Preferably, in step S3, one model in the edge model acts as the questioner, primarily responsible for raising industry-specific questions, while the other model is responsible for responding to the questions raised by the questioner. Information is exchanged through discussions involving the mutual generation of questions and answers. Features of the dialogue data within the domain are autonomously mined, and high-quality pseudo-labels are generated to guide the optimized training of the edge model. The cloud model participates in and evaluates the discussions among the smaller models, using this as a constraint to correct the pseudo-labels it generates, further improving the quality of information.

[0033] Preferably, in step S4, multiple edge-side small models collaboratively achieve online evolutionary learning through discussion. The system automatically collects feedback from multi-model interactions and integrates this feedback information into the training data of the edge-side models, forming a continuously updated data source. During collaboration, the edge-side small models adaptively adjust based on the prior knowledge and quality assessment of the large model, achieving self-optimization through multiple rounds of discussion and pseudo-label generation without human intervention. The cloud-based large model uses the dataset collected during the discussion process as contextual cue words for contextual learning, continuously optimizing its generation capabilities and achieving continuous adaptive optimization.

[0034] This invention provides a method and application for online evolutionary learning of a large-scale model ecosystem based on edge-cloud collaboration. It has the following beneficial effects:

[0035] 1. This invention provides a method and application for online evolutionary learning of a large-scale model ecosystem based on edge-cloud collaboration. Through an edge-cloud collaborative architecture, it innovatively solves the problem of insufficient computing power and inability to handle complex tasks on edge devices. In traditional edge model training, a large amount of high-quality manually labeled data is often required as training corpus, but obtaining this labeled data is costly and inefficient. This invention introduces a technology for generating pseudo-labels from cloud models, automatically generating high-quality pseudo-labels based on existing data. This not only greatly expands the training corpus of edge models but also significantly reduces reliance on manually labeled data, improving training efficiency. In traditional edge models, limited computing power and storage resources prevent them from handling complex tasks during inference. This invention proposes a novel inference mechanism through the collaborative work of cloud and edge models. During inference, the cloud model provides high-quality supplementary inference information to the edge model, enabling the edge model to not only integrate the output information of the cloud model but also combine its own local knowledge to provide users with more accurate and personalized responses, significantly improving inference ability and accuracy.

[0036] 2. This invention provides a method and application for online evolutionary learning of large-scale models based on edge-cloud collaboration. By introducing a multi-model collaborative discussion mechanism and realizing online evolutionary learning of cloud and edge models, it significantly improves the system's adaptability and self-optimization capabilities. In traditional model training, edge models often rely on a single data source and training method. In addition, traditional large-scale model updates often rely on offline data, which cannot respond to users' personalized needs in a timely manner. Based on edge-cloud collaboration, this invention enables cloud models and several edge models to generate new training labels through collaborative discussion and information feedback, achieving self-training and optimization. With the help of real-time user feedback, the model can evolve online, continuously adjusting weights or parameters to gradually adapt to users' personalized needs and changing usage habits, forming a dynamically self-evolving intelligent customer service system. This mechanism greatly reduces the dependence on manually labeled data and improves the system's adaptability and intelligence level in complex scenarios.

[0037] 3. This invention provides a method and application for online evolutionary learning of a large model ecosystem based on edge-cloud collaboration. Addressing the stringent requirements for user privacy and data security in e-commerce scenarios, this invention employs a local privacy discrimination mechanism. When processing user requests, the edge model can autonomously identify sensitive information and complete inference locally, avoiding the uploading of private data to the cloud and ensuring data security. This mechanism not only effectively enhances data security but also meets the high standards for data privacy protection in e-commerce scenarios, further improving user trust in the system and user experience. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating the large-scale model ecosystem online evolutionary learning method based on edge-cloud collaboration of the present invention.

[0039] Figure 2 This is a flowchart illustrating the multi-model collaborative discussion process of the present invention;

[0040] Figure 3 This is a flowchart illustrating the collaborative reasoning and online evolutionary learning process of the model in this invention.

[0041] Figure 4 This is a schematic diagram illustrating the application process of the large-scale model ecosystem based on edge-cloud collaboration of the present invention in the e-commerce field.

[0042] in, Figure 1 In the diagram, the icons on the left represent databases, the icons on the right represent users, and the arrows represent the flow and transmission of information. Figure 2 and Figure 3 In the diagram, each arrow represents the relationship and interaction process between the modules; Figure 4 In the diagram, each arrow represents the process of information flow and data transmission. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] like Figure 1-4 As shown, this embodiment of the invention provides a method and application for online evolutionary learning of a large-scale model ecosystem based on edge-cloud collaboration, specifically including the following steps.

[0045] S1. Cloud-based large-scale model contextual learning and adjustment

[0046] Leveraging the powerful contextual learning capabilities of cloud-based large models and employing prompt word engineering, these models are adapted to be suitable for specific domains. Through the adaptive capabilities of these large models, contextual modeling is performed on customer service dialogue datasets in specific domains, enabling them to handle specific complex problems in those domain scenarios.

[0047] Specifically, taking the e-commerce industry as an example, this study demonstrates the application of an online evolutionary learning method based on edge-cloud collaboration to a real-world industry. In the experiment, the Gemini 1.5 Pro model was selected as the cloud-based large model, and the ChatGLM3-6B model as the edge model. Following the steps in S1, the cloud-based large model used the following prompts as input: "Assume you are an e-commerce customer service representative with extensive e-commerce expertise, excellent service awareness, and the ability to properly handle customer needs. Next, please act as an e-commerce customer service representative and handle customer issues. Below are examples of user questions and customer service responses for your reference; please reply according to the customer service model in the examples." Then, 500 actual multi-turn dialogues from the Taobao e-commerce platform were selected as supplementary prompts to further enrich the model's scenarios. For example:

[0048] User: 'Could you send me a clip?'

[0049] Customer service: "I'm sorry, sir, we don't have clips here, only mouth openers."

[0050] User: 'Are there any discounts available?'

[0051] Customer service: "Sir / Madam, if you buy more, you can use a coupon. Please verify your shipping address."

[0052] User: 'Yes.'

[0053] Customer service: Okay, sir / madam, we'll ship your order tomorrow. The courier picked it up at 4 PM and it will be dispatched tomorrow.

[0054] By collecting real dialogues from e-commerce platforms and combining them with user data input into a large cloud model, multi-turn dialogues were simulated to generate a hybrid dataset consisting of real e-commerce dialogues and model-generated dialogues.

[0055] S2. Supervision Information Generation and Dataset Construction

[0056] After completing the contextual learning of the large cloud model, the cloud model generates supervised information for training the edge model. At the same time, the supervised information generated by the large cloud model is used as the generated dataset and mixed with the natural dataset for subsequent model training.

[0057] Specifically, the ChatGLM3-6B model was used as the client-side model, and it was fine-tuned using a mixed dataset. Low-rank adaptive fine-tuning (LoRA) and other methods were employed to train the client-side model, gradually adapting it into a model specifically designed for e-commerce customer service tasks.

[0058] S3. Training of edge models and multi-model collaboration

[0059] Based on high-quality supervised information generated in the cloud and supplemented industry-specific dialogue datasets, this study trains edge models using efficient parameter fine-tuning methods. Simultaneously, a multi-model collaborative discussion mechanism is introduced, where edge models and other smaller models play different roles and engage in mutual discussion. Specifically, one model acts as the questioner, primarily responsible for raising industry-specific questions, while another model is responsible for responding to the questioner's questions. Information is exchanged through discussions involving the mutual generation of questions and answers, autonomously mining features from the domain-specific dialogue data and generating high-quality pseudo-labels to guide the optimized training of edge models. The cloud model participates in and evaluates the discussions among smaller models, using these as constraints to correct the generated pseudo-labels, further improving the quality of information.

[0060] Specifically, to further improve the training effect of the model, a multi-model collaborative discussion method was introduced. In implementation, two pre-trained models on the client side play the roles of a customer and an e-commerce customer service representative, respectively, engaging in dialogue to generate pseudo-labels. The client-side e-commerce model is the primary model trained and will be deployed in real-world environments for human interaction. The client-side customer model, based on the pre-trained model, learns from context through specific prompts. The prompts are: "Please play the role of a consumer (customer) in an e-commerce store, and I will play the role of an e-commerce customer service representative. You are having a multi-round e-commerce customer service dialogue with me. You can ask me various questions in the e-commerce field, including but not limited to product information, logistics, purchase return and exchange services, etc., and continue the dialogue based on the responses. Please imitate a human consumer, speak naturally and conversationally, and output the command 'Dialogue End' when you think the topic should end."

[0061] S4. Collaborative work of end-side models during the inference phase

[0062] The trained edge model is applied to real-world customer service scenarios. When the edge model receives a user's question, it first uses a local discrimination model to determine whether the question involves privacy. If the question involves privacy, the edge model answers it independently; if it does not involve privacy, the question is uploaded to the cloud model. The cloud model and the edge model work together to generate a response. The edge model then combines the cloud model's response with the cloud model's response to generate the final response.

[0063] Specifically, for cloud-based e-commerce models, the following prompts should be added: "Please evaluate the quality of e-commerce customer service responses from the customer's perspective. Generally, the criteria for judging the quality of e-commerce customer service responses include: whether the response is timely and effectively resolves the customer's problem; whether the response is accurate, polite, and professional; whether the communication is clear and friendly; and whether personalized service is provided. In addition, it is also necessary to consider whether the response is conversational, concise, efficient, and reflects the professional characteristics of e-commerce customer service."

[0064] The cloud-based large model, having learned from the aforementioned context, acts as a discriminative model to evaluate the quality of the dialogues. If a dialogue is deemed of high quality, it is retained in the training corpus and used to fine-tune the on-device model.

[0065] S5. Real-time feedback and online fine-tuning

[0066] When the edge model performs inference tasks, two different response schemes are generated at a certain frequency. After the user selects the better response, the answer is matched with the user's question and added to a new training set for online fine-tuning and optimization of the edge model. At the same time, the collected human feedback serves as contextual cue words for the cloud-based large model to learn context and continuously optimize its generation capabilities.

[0067] Specifically, when fine-tuning the edge model, mainstream efficient model parameter fine-tuning methods such as low-rank adaptive tuning (LoRA) and prefix-tuning are selected.

[0068] The finely tuned client-side e-commerce model is deployed to a real-world e-commerce scenario to handle user inquiries. When the client-side model receives a user's question, it first uses a local discrimination model to determine whether the question involves privacy information. If privacy is involved, the client-side model responds independently; otherwise, the question is uploaded to the cloud-based large model. After the cloud-based large model generates a response, the client-side model combines the output of the cloud model with the user's input to generate the final answer.

[0069] S6. Online Evolution and Model Iteration

[0070] Through regular human feedback and edge-cloud collaborative training, the cloud and edge models achieve continuous online optimization. By learning from and collaborating with each other, the edge and cloud models continuously improve their ability to respond to e-commerce customer service scenarios, achieving adaptive learning and performance optimization.

[0071] Specifically, to achieve personalized service, the system generates one or two different responses with a certain probability when generating replies. If two replies are generated, the user can choose the one they prefer. The user's selection is paired with the question and added to the training corpus for online fine-tuning of the on-device model. Simultaneously, these high-quality dialogues selected based on user feedback are also used to update the prompt words of the large cloud-based model, further improving the supervision and generation capabilities of the cloud model.

[0072] By continuously repeating this process, the edge model and the cloud model evolve online through ongoing interaction and feedback with users, thereby providing more personalized and accurate e-commerce customer service.

[0073] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method and application for online evolutionary learning of a large-scale model ecosystem based on edge-cloud collaboration, characterized in that, Specifically, the following steps are included: S1. Multi-model role allocation Multiple edge-side small models and one cloud-based large model are assigned different roles to facilitate effective discussion and collaboration during training. The cloud-based large model acts as the teacher model, providing prior input and constraints for multi-model discussions. The roles of the edge-side small models are assigned as follows: Questioner Model: Focuses on asking industry-related questions, simulating the questioning style of real users. Respondent Model: Responsible for generating responses based on industry knowledge and task objectives, improving the accuracy and professionalism of the replies; S2. Cloud-based large model prior input and constraint support During each round of collaborative discussion, the cloud-based large model supports the training of the small model by providing prior inputs and constraints. Specifically, in each round of dialogue, the cloud-based large model has two roles: scene prior generation and generation quality evaluation. Scenario prior generation: Provides core knowledge of e-commerce services, including the best answer patterns in customer service, personalized service techniques, and privacy data processing requirements, so that small models have a unified standard reference during training; Quality assessment: When the small model generates answers, the large model performs quality assessment and feedback on these answers, and uses this constraint to guide the small model to generate content that better meets industry needs, ensuring the consistency and accuracy of the model; S3. Multi-model collaborative discussion During the collaboration process, the models on each edge discuss by generating questions and answers for each other, forming multi-turn dialogues. This discussion mechanism enables the models to learn and understand various scenario requirements through interaction: Prior input: Before the questioner model and the answerer model generate questions and answers, the scenario prior input of the large model is given to the questioner and answerer models to provide a prior knowledge base for the training of the small model on the edge, thereby improving the quality of question and answer generation; Question generation: The questioner model generates questions based on industry needs and task scenarios; Answer generation: The respondent model answers questions generated by the questioner model by providing detailed solutions, analyzing scenario requirements, and improving service quality. Quality verification: After each round of question and answer, the dialogue generated by the edge model is evaluated and corrected using a large model to accelerate the convergence of multi-model training on the edge. S4. Generation and Online Evolution of Co-training Data During the discussion and collaboration process, through multiple rounds of dialogue, the responses generated by the smaller models are constrained and verified by the larger model, labeled as high-quality training data, and added to the new dataset. The generated collaborative data and efficient parameter fine-tuning techniques are used to train the smaller models, enabling them to achieve optimized performance on specific tasks. Through regular edge-cloud collaborative multi-model discussion and training mechanisms, cloud and edge models achieve continuous online optimization. Edge and cloud models learn from and collaborate with each other, continuously improving their ability to respond to e-commerce customer service scenarios, achieving adaptive learning and performance optimization. The online evolving edge smaller models will have stronger response capabilities, able to adaptively process e-commerce customer service tasks based on collaboratively trained data, and achieve higher quality responses.

2. The online evolutionary learning method and application for a large-scale model ecosystem based on edge-cloud collaboration as described in claim 1, characterized in that: In S2, the adaptive and contextual learning capabilities of the large model are used to perform contextual modeling on the customer service dialogue dataset in a specific domain, enabling it to handle specific complex problems in that domain scenario and to assume the prior and constraint role of collaboration with the small model.

3. The online evolutionary learning method and application for a large-scale model ecosystem based on edge-cloud collaboration as described in claim 1, characterized in that: In S3, one model in the edge model acts as the questioner, responsible for raising specific questions within the industry, while the other model is responsible for responding to the questions raised by the questioner. Information is exchanged through discussions that generate questions and answers with each other. The model autonomously mines the features of the dialogue data within the domain and generates high-quality pseudo-labels to guide the optimization training of the edge model. The cloud model participates in and evaluates the discussions between the small and medium-sized models, and uses this as a constraint to correct the pseudo-labels it generates, further improving the quality of information.

4. The online evolutionary learning method and application for a large-scale model ecosystem based on edge-cloud collaboration as described in claim 1, characterized in that: In S4, multiple edge-side small models collaborate and discuss to achieve online evolutionary learning. The system automatically collects feedback from multi-model interactions and integrates the feedback information into the training data of the edge-side models, forming a continuously updated data source. The edge-side small models adaptively adjust based on the prior knowledge and quality assessment of the large model during collaboration. Through multiple rounds of discussion and pseudo-label generation, they achieve self-optimization without human intervention. The cloud-based large model uses the dataset collected during the discussion process as contextual cue words for contextual learning, continuously optimizing its generation capabilities and achieving continuous adaptive optimization.

Images