Data enhancement method and system based on multi-agent self-evolution and hybrid evaluation

Abstract

The invention discloses a data enhancement method and system based on multi-agent self-evolution and hybrid evaluation, and relates to the technical field of artificial intelligence, and the method comprises the steps: employing a teacher model as a multi-agent cooperation system, and carrying out the interaction of a reasoning agent, an evaluation agent, a reflection agent and a memory management agent through a reasoning agent, an evaluation agent, a reflection agent and a memory management agent; iteratively generating a high-quality reasoning data set with a traceable reasoning path and a self-verification label, and performing multi-task supervision fine tuning on a learning model; a diversity sampling strategy based on reinforcement learning is adopted to generate multiple groups of outputs, and a weak point data set is screened by using a consistency score; and for the weak point data set, the multi-view instruction rewriting agent performs diversity rewriting on the instruction and then returns to perform deep reasoning distillation to generate new enhanced data and combine the new enhanced data to high-quality reasoning data. According to the method, multiple agents are arranged, so that the model is iteratively synthesized, evaluated, reflected and modified, the obtained enhanced data can be directly used, a manual auditing step is omitted, and the synthesis efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a data augmentation method and system based on multi-agent self-evolution and hybrid evaluation. Background Technology

[0002] Currently, medical language models have emerged in the medical field for medical health question answering, medical diagnosis, and medical knowledge discrimination. These models are mostly general-purpose, utilizing general medical knowledge to enhance their breadth of knowledge across general medicine. While the data volume is large, the depth is insufficient, and they lack specificity. Their answers in vertical domains are inaccurate, making practical application difficult. This is because they do not fully utilize data from vertical domains and lack high-quality data specific to particular scenarios. Due to the complexity, specialization, and scarcity of data in the medical field, training models using publicly available datasets currently does not yield ideal results. Accurate and specialized datasets need to be collected and processed from hospitals, which is inherently difficult, time-consuming, costly to process, and often involves large but limited datasets. If these processed datasets are directly used for training, fine-tuning with a small sample size is unlikely to achieve the desired results.

[0003] Pre-trained large language models have demonstrated significant performance on various benchmark datasets and in real-world applications. However, in practical applications, to achieve outputs consistent with human expectations, high-quality prompts and few-shot learning examples are often required. The ideal solution is to retrain the pre-trained large language model to fine-tune its performance in specific vertical domain tasks. However, obtaining a sufficient amount of high-quality training data becomes a pressing problem when fine-tuning domain-specific models. Unlike data acquisition in general knowledge domains, data resources in vertical domains are scarce. In the field of liver cancer, data privacy is strong, acquisition is difficult, and much of the data is unstructured, making post-processing complex. The number of medical records is limited, and existing data annotation work is scarce and of inconsistent quality. Manually collecting, cleaning, and labeling this data is not only costly but also extremely time-consuming. Therefore, how to effectively process and expand the data under limited conditions to create a larger dataset for model fine-tuning, thereby improving the model's performance in domain-specific tasks, has become an important research topic. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a data augmentation method and system based on multi-agent self-evolution and hybrid evaluation, which solves the problem of insufficient high-quality data in large-scale clinical diagnostic models for liver cancer on specific tasks.

[0005] The first aspect of this invention provides a data augmentation method based on multi-agent self-evolution and hybrid evaluation, comprising the following steps: S1: Utilize the teacher model as a multi-agent collaborative system, including reasoning agent, evaluation agent, reflection agent and memory management agent, to iteratively generate a high-quality reasoning dataset with traceable reasoning paths and self-verifying labels; S2: Use the high-quality inference dataset to perform multi-task supervised fine-tuning of the student model, including inference chain generation, error correction, and missing inference completion tasks; S3: Employ a diversity sampling strategy based on reinforcement learning to dynamically adjust temperature sampling and kernel sampling parameters, generate multiple sets of outputs, and use consistency scores to filter weak point datasets. S4: For the aforementioned weak point dataset, after rewriting the instructions in a variety of ways by the multi-view instruction rewriting agent, return to S1 for deep inference distillation to generate new augmented data and combine it with high-quality inference data. Repeat S2 to S4 until the student model reaches the preset performance threshold on the cross-domain test set.

[0006] In this scheme, in step S1, a teacher model is used as a multi-agent collaborative system, including a reasoning agent, an evaluation agent, a reflection agent, and a memory management agent. A high-quality reasoning dataset with traceable reasoning paths and self-verification labels is generated iteratively, including: The reasoning agent receives the original question input, identifies key features in the question based on the attention mechanism, decomposes the question into diagnostic sub-tasks based on the key features, and uses the attention mechanism to determine the importance of each key feature. Based on the importance, a hierarchical reasoning structure is generated and deduced step by step according to medical logic, generating a reasoning chain and answer containing detailed reasoning steps. The reasoning chain and answer are evaluated in multiple dimensions by an evaluation agent. Evaluation indicators are obtained based on factual accuracy, logical coherence, and clinical rationality. Evaluation results are calculated and output based on the evaluation indicators. The correctness of the reasoning chain and answer is judged based on the evaluation results. A historical error pattern library is constructed. The reflective agent identifies and analyzes logical loopholes or evidence gaps in the reasoning chain based on the historical error pattern library, predicts potential risk points through historical error patterns, and generates feedback guidance and preventive guidance suggestions. The memory management agent records the complete reasoning history of local sessions, maintains a historical error pattern library, retrieves processing experience of similar historical medical records, provides real-time alerts for repeated errors, and integrates the latest clinical knowledge to optimize the historical error pattern library. The system acquires evaluation scores during the iteration process of the reasoning agent, evaluation agent, reflection agent, and memory management agent. When the evaluation score reaches a threshold, it outputs a reasoning chain and answer that includes a complete iteration history, specific content of modifications, and reasons. It also adds quality score labels, uncertainty labels, and verification status labels to generate high-quality reasoning data.

[0007] In this scheme, step S2 involves using the high-quality inference dataset to perform multi-task supervised fine-tuning of the student model through a dynamic course learning strategy, including: Acquire a high-quality inference dataset and perform data formatting processing, organizing each data point into instruction-response pairs; The dynamic curriculum learning strategy uses formatted data to fine-tune student models. Initially, the model learns the complete inference chain, training it to learn standard diagnostic procedures and reasoning patterns, establish logical connections between medical concepts, and master professional terminology and assessment standards. The intermediate learning process involves gradually hiding some reasoning steps, first concealing the conclusion, then gradually reducing hints in the intermediate reasoning steps, and finally dynamically selecting the hidden content. This trains students' ability to autonomously complete the reasoning chain, apply medical knowledge, and adapt to clinical situations. In the later stages, only questions and answers are provided for learning, enabling students to internalize the complete diagnostic reasoning logic of the model, form an automated clinical thinking mode, and have the generalization ability to handle rare cases; The performance of the student model is obtained during the fine-tuning training process. The performance level of the model is characterized by the accuracy rate. The course progress is adjusted in a personalized manner according to the performance level. Multi-task learning is used to combine direct answer generation tasks, intermediate reasoning completion tasks, and error step correction tasks to optimize the reasoning ability of the student model.

[0008] In this scheme, in step S3, a reinforcement learning-based diversity sampling strategy is adopted to dynamically adjust the temperature sampling and kernel sampling parameters and generate multiple sets of outputs, specifically: For each question in the high-quality inference dataset, we perform reinforcement learning-based diversity sampling, evaluate the model performance status based on the student model's current accuracy level, output diversity, and error type distribution characteristics, and analyze the current question input and model performance status. In each sampling, a reinforcement learning policy network generates temperature parameters T and Top-p parameters, generating N different outputs, each including diverse inference chains and corresponding answers; The Jaccard similarity between different groups of outputs is calculated to obtain the word overlap rate as a surface diversity index. The cosine similarity matrix between different groups of outputs is calculated to obtain the embedding space distance as a semantic diversity index. Furthermore, the consistency of different groups of outputs with medical logic is verified as a logical rationality index. The surface diversity index, semantic diversity index, and logical rationality index are used to perform diversity quantification analysis. When the diverse quantitative analysis results of N different outputs meet the preset standards, consistency screening is performed.

[0009] In this scheme, step S3 involves using a consistency evaluator that combines rules and models to calculate consistency scores and filter the weak point dataset, including: A consistency evaluator is constructed based on a rule module and a model module. In the rule module, a precise fact check is performed based on a medical knowledge base. For each question, the standard answer generated by the teacher model and N different outputs generated by the student model are obtained. Key medical facts are extracted from the standard answer. The system checks whether the N different outputs generated by the student model completely cover the facts and obtains the fact accuracy score. In the model module, a logic and rationality assessment is performed, a decision tree for the standard answer is drawn, the deviation of the reasoning path of the N different outputs generated by the student model is compared, a logical coherence score is obtained, and knowledge errors, logical errors and omission errors are marked to obtain a decision rationality score. The comprehensive module uses a weighted scoring method to obtain a consistency score based on the accuracy of facts, logical coherence, and reasonableness of decision-making. Response samples with consistency scores less than a preset threshold are selected to generate a weakness dataset, and each weak sample in the weakness dataset is labeled with defects.

[0010] In this scheme, in step S4, for the vulnerability dataset, the instruction rewriting agent performs diverse rewriting of the instructions, including: A vulnerability dataset is acquired and grouped according to the defect type. An instruction rewriting agent is constructed, and rewriting requirements are randomly selected to rewrite the grouped vulnerability samples. The rewriting requirements include: changing the perspective of instruction questioning, replacing clinical word meanings, introducing additional information, expanding clinical context, and converting between Chinese and English. The rewritten question instructions are subjected to semantic consistency checks. A sentence embedding model is used to calculate the semantic similarity with the original instructions. If the similarity is lower than the threshold, the instructions are rewritten. The rewritten question instructions are then pre-screened by a teacher model.

[0011] In this scheme, deep inference distillation and re-execution includes: The rewritten problem instructions are input into the reasoning agent, evaluation agent, reflection agent, and memory management agent to strengthen the verification mechanism for controversial conclusions and force verification steps to be inserted into the original weak points, thereby generating new reasoning chains and answers. Record historical error patterns. If the similarity between the current case and the historical error pattern is greater than a preset similarity threshold, then based on the historical error pattern, a similar risk is indicated, and new enhanced data is generated from the reasoning chain and answer with similar risk labels. Stratified sampling is performed on the base data of the new augmented data and high-quality inference dataset, and data merging is performed by dynamically weighting using reinforcement learning agents to dynamically expand the high-quality inference dataset. Security enhancements are performed on several hospital data sources using a federated learning mechanism.

[0012] A second aspect of the present invention provides a data augmentation system based on multi-agent self-evolution and hybrid evaluation, the system comprising: a deep inference distillation module, a data fine-tuning module, a verification sampling and evaluation module, and an iterative augmentation module; The deep reasoning distillation module utilizes a teacher model as a multi-agent collaborative system, including a reasoning agent, an evaluation agent, a reflection agent, and a memory management agent, and iteratively generates a high-quality reasoning dataset with traceable reasoning paths and self-verifying labels. The data fine-tuning module uses the high-quality inference dataset to perform multi-task supervised fine-tuning of the student model, including inference chain generation, error correction, and missing inference completion tasks. The inspection sampling and evaluation module adopts a diverse sampling strategy based on reinforcement learning, dynamically adjusts temperature sampling and kernel sampling parameters, generates multiple sets of outputs, and uses consistency scores to filter the weak point dataset. The iterative enhancement module targets the weak point dataset by rewriting instructions in a variety of ways through a multi-perspective instruction rewriting agent, dynamically expanding the high-quality inference dataset, and continuously optimizing model performance until the model reaches a preset performance threshold on a cross-domain test set.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention addresses the problem of how to augment data under limited conditions when large-scale model training requires high-quality data but the original data is insufficient. Unlike traditional data augmentation methods, this invention proposes a data augmentation framework that uses multi-agent collaborative guidance and verification sampling iteration. This framework can iteratively generate data while simultaneously verifying data reliability and correcting the data.

[0014] (2) The synthesis method proposed in this invention constructs a deep distillation module. By setting up multiple agents, the model can iteratively synthesize, evaluate, reflect and modify itself. The resulting enhanced data can be used directly, the diversity is expanded, the reliability is verified, the manual review steps are saved, and the synthesis efficiency is improved.

[0015] (3) This invention proposes a sampling iteration method for testing, which uses augmented data synthesized from a large model. In addition to training the model, it can also be used to screen for model weaknesses. After training, the small model uses distilled data for inference testing, uses diversity sampling to evaluate the model effect, and screens out the current weaknesses of the student model. Then, the teacher model rewrites the weak data and uses the deep inference distillation method again to expand the weak data of the student model. Repeated iterations improve the performance of the fine-tuned model. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0017] Figure 1 A flowchart of a data augmentation method based on multi-agent self-evolution and hybrid evaluation is shown; Figure 2 A flowchart illustrating the generation of high-quality medical inference data based on multi-agent collaboration is shown. Figure 3 A flowchart of the iterative student model optimization based on knowledge distillation and reinforcement learning is shown; Figure 4 The flowchart illustrates the construction of a consistency evaluator based on the rule module and the model module. Figure 5 A block diagram of a data augmentation system based on multi-agent self-evolution and hybrid evaluation is shown. Detailed Implementation

[0018] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0019] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0020] Figure 1 A flowchart of a data augmentation method based on multi-agent self-evolution and hybrid evaluation is shown.

[0021] S1: Utilize the teacher model as a multi-agent collaborative system, including reasoning agent, evaluation agent, reflection agent and memory management agent, to iteratively generate a high-quality reasoning dataset with traceable reasoning paths and self-verifying labels; S2: Use the high-quality inference dataset to perform multi-task supervised fine-tuning of the student model, including inference chain generation, error correction, and missing inference completion tasks; S3: Employ a diversity sampling strategy based on reinforcement learning to dynamically adjust temperature sampling and kernel sampling parameters, generate multiple sets of outputs, and use consistency scores to filter weak point datasets. S4: For the aforementioned weak point dataset, after rewriting the instructions in a variety of ways by the multi-view instruction rewriting agent, return to S1 for deep inference distillation to generate new augmented data and combine it with high-quality inference data. Repeat S2 to S4 until the student model reaches the preset performance threshold on the cross-domain test set.

[0022] It should be noted that the core objective of this framework is to distill long thought chain data from the teacher model through multi-round stimulus guidance, constructing a multi-level reasoning question-answer pair dataset, including traceable long reasoning thought chains, self-reflection thought chains, and self-assessment datasets. Student models undergo supervised fine-tuning followed by iterative enhancement through diversity sampling, gradually developing multi-step reasoning capabilities and improving performance on vertical and cross-domain tasks. A consistency evaluator based on a combination of rules and models reduces the risk of illusion through key fact matching rules, enhancing the reliability and interpretability of the evaluation.

[0023] like Figure 2 As shown, a multi-agent system is constructed using a teacher model to generate a structured medical reasoning chain, including a reasoning agent, an evaluation agent, a reflection agent, and a memory management agent. Through multiple iterations, a high-quality reasoning dataset with traceable reasoning paths and self-verification labels is generated. The reasoning agent receives the original question input and identifies key features in the question based on an attention mechanism, including symptom feature extraction (identifying key medical features in the chief complaint and physical signs), laboratory indicator correlation analysis (establishing medical logical connections between indicators), and image feature parsing (extracting key diagnostic elements from the image report). Based on these key features, the question is decomposed into diagnostic sub-tasks, and the attention mechanism is used to determine the importance of each key feature. A hierarchical reasoning structure is generated according to the importance, deducing step-by-step according to medical logic. For example, the first layer assesses basic indicators, the second layer integrates key evidence, and the third layer outputs a comprehensive diagnostic conclusion. A reasoning chain and answer containing detailed reasoning steps are generated, including diagnostic basis weight scores and uncertainty annotations.

[0024] The reasoning chain and answer are evaluated from multiple dimensions by an evaluation agent. Evaluation indicators are obtained based on factual accuracy, logical coherence, and clinical rationality. The preferred indicator for factual accuracy is the correctness of medical knowledge; for logical coherence, it is the completeness of the reasoning chain; and for clinical rationality, it is the feasibility of the treatment plan. Fact-checking is performed by comparing the latest clinical guidelines and medical knowledge bases; logical verification is conducted by checking the causal rationality of the reasoning chain; and clinical verification is performed by simulating real-world clinical application scenarios. Based on the evaluation indicators, an evaluation score is calculated and output, along with the specific locations of defects. The correctness of the reasoning chain and answer is determined based on the evaluation results.

[0025] A historical error pattern library is constructed. Based on this library, a reflective agent identifies and analyzes logical loopholes or evidentiary gaps in the reasoning chain. The historical error patterns are used to predict potential risk points and generate feedback guidance and preventive guidance suggestions, including missing medical knowledge points, suggestions for optimizing reasoning paths, and real-world factors that need to be added.

[0026] The system utilizes a memory management agent to record the complete reasoning history of local sessions, maintain a historical error pattern database, retrieve processing experience from similar historical medical records, provide real-time alerts for recurring errors, and integrate the latest clinical knowledge to optimize the historical error pattern database. It acquires evaluation scores from the iterative processes of the reasoning agent, evaluation agent, reflection agent, and memory management agent. When the evaluation score reaches a threshold, it outputs a reasoning chain and answer containing a complete iterative history, specific modifications, and reasons, and adds quality score labels, uncertainty labels, and verification status labels to generate high-quality reasoning data. Through deep collaboration and iterative optimization among multiple agents, it generates high-quality medical reasoning data with traceable reasoning paths and self-verifying labels, ensuring that the output results conform to medical standards and have clinical applicability, providing a reliable data foundation.

[0027] It should be noted that, as Figure 3 As shown, high-quality medical inference data is used. Fine-tuning student models (e.g., PMC-LLaMA, Meditron-7B, etc.) yields supervised fine-tuned student models. Through a progressive training strategy, student models are guided from relying on explicit reasoning chains to internalizing medical reasoning logic, and then the learning effectiveness of the student models is evaluated. This is achieved through a reinforcement learning-based diversity sampling strategy and consistency scoring. Evaluate the different feedback from the student model, and based on the differences between the student model and the teacher model, further strengthen and expand the weak points of the student model.

[0028] High-quality inference datasets are acquired and formatted, with each data point organized into instruction-response pairs. A dynamic learning strategy is employed to fine-tune student models using this formatted data. Initially, complete inference chains are used for learning, training student models to learn standard diagnostic procedures and reasoning patterns, establish logical connections between medical concepts, and master professional terminology and evaluation standards. In the intermediate stage, some reasoning steps are progressively hidden, first concealing the conclusions, gradually reducing hints for intermediate reasoning steps, and finally dynamically selecting hidden content to train student models to autonomously complete inference chains, apply medical knowledge, and adapt to clinical situations. In the later stage, only questions and answers are provided for learning, enabling student models to internalize complete diagnostic reasoning logic, form automated clinical thinking patterns, and possess the ability to generalize to rare cases. During the fine-tuning training of student models, performance is assessed. Accuracy is used to characterize the model's performance level, and the course progress is adjusted accordingly. For example, an accuracy rate greater than 90% indicates excellent performance, accelerating progress to the next stage; an accuracy rate of 70%-19% indicates performance needs improvement, extending the current training stage. By utilizing multi-task learning to combine direct answer generation tasks, intermediate reasoning completion tasks, and error step correction tasks, the reasoning ability of student models can be optimized.

[0029] Quality monitoring is implemented at each stage to ensure training quality. The initial stage monitors the completeness of the reasoning chain reproduction and the accuracy of medical facts; the mid-stage monitors the consistency of reasoning logic and the accuracy of key node completion; and the later stage monitors the final diagnostic accuracy and the safety of clinical decisions. Through a dynamic learning strategy, student models can systematically internalize clinical reasoning abilities, progressing from initial reliance on explicit guidance to the formation of independent clinical thinking, thus ensuring both learning effectiveness and medical safety.

[0030] It should be noted that by dynamically optimizing sampling parameters through reinforcement learning, output diversity is maximized while ensuring medical accuracy, thereby improving the model's adaptability in complex clinical scenarios. For each question in the high-quality inference dataset, reinforcement learning-based diversity sampling is performed. The model performance is evaluated based on the student model's current accuracy level, output diversity, and error type distribution characteristics, and a state space is set. An action space is set based on the adjustment range of temperature parameters, the adjustment range of kernel sampling parameters, and the combination strategy of temperature and kernel sampling parameters. The combination strategy includes: low temperature and low kernel sampling to ensure accuracy, and high temperature and high kernel sampling to promote diversity. A reward function is constructed based on a multi-objective reward system. The reward dimensions include accuracy reward, diversity reward, and penalty term. The accuracy reward is the matching degree between the model output and the standard answer, the diversity reward is the negative value of the similarity between the model outputs, and the penalty term is the number of medical errors. In the policy training phase of reinforcement learning, historical interaction data is used to initialize the experience replay buffer, the initial policy network is trained on the validation set, and the sampling parameters are ensured not to cause medical errors. In online learning, the current input and model state are analyzed, the policy network outputs temperature and kernel sampling parameters, N sets of candidate outputs are generated using the sampling parameters, the quality and diversity of each set of outputs are evaluated according to the reward function, and finally the policy network is updated based on the reward.

[0031] Analyzing the current problem input and model performance, in each sampling, a reinforcement learning policy network generates temperature parameters T and Top-p parameters, generating N different outputs, each including diverse inference chains and corresponding answers. The Jaccard similarity between different output groups is calculated to obtain word overlap rate as a surface diversity indicator, and the cosine similarity matrix between different output groups is calculated to obtain embedding space distance as a semantic diversity indicator. Furthermore, the conformity of different output groups to medical logic is verified as a logical rationality indicator. Diversity quantification analysis is performed using the surface diversity indicator, semantic diversity indicator, and logical rationality indicator. When the diversity quantification analysis results of the N different output groups meet preset standards, consistency screening is performed.

[0032] like Figure 4 As shown, a consistency evaluator is constructed based on the rule module and the model module. By combining the precision of the rule engine with the flexibility of model evaluation, a comprehensive and reliable consistency assessment of student model outputs is achieved. The rule module performs precise fact checking based on a medical knowledge base. For each question, it obtains the standard answer generated by the teacher model and N different outputs generated by the student model. Key medical facts are extracted from the standard answer, including: disease entities, diagnostic criteria, numerical information, and treatment terms. The system checks whether the N different outputs generated by the student model fully cover these facts, obtaining a fact accuracy score. During the coverage check, key medical facts must appear completely; numerical information is allowed an error range of ±5%; and diagnostic criteria must fully comply with guidelines. The matching degree is then combined with preset weights for the key medical facts to obtain the fact accuracy score.

[0033] In the model module, logical and reasonableness assessments are performed, a decision tree for the standard answer is drawn, and the deviation of the reasoning paths of N different outputs generated by the student model is compared to obtain a logical coherence score. Factual errors (such as misuse of the Child-Pugh standard), logical errors (such as causal inversion), and omission errors (such as ignoring portal vein invasion) are labeled. The number of errors is mapped to a 0-1 interval to obtain a decision reasonableness score. Using a comprehensive module, a weighted scoring method is applied based on the factual accuracy score, logical coherence score, and decision reasonableness score to obtain a consistency score. Answer samples with consistency scores below a preset threshold are selected to generate a weakness dataset, and each weak sample in the weakness dataset is labeled with defects. For the weakness dataset, an instruction rewriting agent is used to rewrite the instructions in a variety of ways. The goal of the instruction rewriting agent is to increase the complexity and diversity of the instructions, rewriting a completely new question instruction. The new instruction differs from the original instruction in expression and structure, but maintains semantic and data details, without losing key data and features. Randomly select from the following requirements: 1. Change the perspective of the instruction question: By changing the way the question is phrased or the sentence structure, construct the question using rhetorical questions, hypotheses, etc., so that the answerer needs to analyze the core of the problem, rather than just directly calculating the numerical value; 2. Clinical word substitution: Use synonyms to change the expression; 3. Introduce additional information: Introduce additional conditions or assumptions into the instruction; 4. Expand the clinical context: Add real-life scenarios to make the instruction more relevant to real life or complex situations; 5. Chinese-English conversion: Switch between Chinese and English languages ​​in the instruction.

[0034] A vulnerability dataset is acquired and grouped according to defect type. An instruction rewriting agent is constructed, and rewriting requirements are randomly selected to rewrite the grouped vulnerability samples. The rewriting requirements include: changing the perspective of instruction questioning, replacing clinical words, introducing additional information, expanding clinical context, and converting between Chinese and English. The semantic consistency of the rewritten instruction is checked, and the semantic similarity with the original instruction is calculated using a sentence embedding model. If it is lower than the threshold, it is rewritten. The rewritten instruction is pre-screened by a teacher model.

[0035] The rewritten problem instructions are input into the reasoning agent, evaluation agent, and reflection agent to strengthen the verification mechanism for controversial conclusions. This includes retrieving the latest guidelines, simulating clinical expert consultation processes, and cross-validating multimodal data. Verification steps are forcibly inserted into previously weak points; for example, when calculating the Child-Pugh score, the calculation formula must be output simultaneously to generate new reasoning chains and answers. Historical error patterns are recorded; if the similarity between the current case and the historical error pattern exceeds a preset similarity threshold, a similarity risk is indicated based on the historical error pattern, and new enhanced data is generated for the reasoning chains and answers labeled with similarity risk. According to the method, stratified sampling is performed on the base data of the new augmented data and the high-quality inference dataset. For example, 30% of the original data is retained in the base data to ensure coverage, and 70% of the augmented data focuses on the weak links in the original dataset. The data is merged using dynamic weighting using a reinforcement learning agent to dynamically expand the high-quality inference dataset. In the reinforcement learning agent, the state space is constructed based on the frequency distribution of each error type of the current model, the severity of each error type, and the overall performance index of the model. In the action space, a sampling weight is assigned to each type of error data. The higher the weight, the greater the probability that the sample corresponding to that type of error will be sampled in the next round of augmentation. A reward function is set to evaluate the quality of the action and guide the agent's learning. The reward function includes the improvement of the model's main performance index (such as accuracy) on the retained test set after taking the action, a diversity reward term for the improved model on all types of errors after augmentation, and a penalty term for the deterioration of the improved model's performance on high-risk errors (such as incorrect recommendation of surgery).

[0036] The cloud server initializes the global model and reinforcement learning agent, distributing the global model to all hospital nodes participating in federated learning. Hospital nodes evaluate model performance using their latest local data, analyzing model errors locally, identifying high-frequency error types and their distribution, and encrypting and uploading model parameter updates, performance metrics, and error distributions to the cloud. The cloud server uses algorithms such as FedAvg to aggregate model updates uploaded by each hospital, forming a new global model and summarizing error distributions to create a global error type heatmap. The agent observes the current global state (global error heatmap and model performance) and generates an action (global sampling weight strategy) based on the policy. This global sampling weight strategy is then distributed to each hospital node. Upon receiving the sampling weight strategy, each hospital performs priority sampling, local data augmentation, and local model training locally.

[0037] Figure 5 A block diagram of a data augmentation system based on multi-agent self-evolution and hybrid evaluation is shown.

[0038] A second aspect of the present invention provides a data augmentation system based on multi-agent self-evolution and hybrid evaluation, the system comprising: a deep inference distillation module, a data fine-tuning module, a verification sampling and evaluation module, and an iterative augmentation module; The deep reasoning distillation module utilizes a teacher model as a multi-agent collaborative system, including a reasoning agent, an evaluation agent, a reflection agent, and a memory management agent, and iteratively generates a high-quality reasoning dataset with traceable reasoning paths and self-verifying labels. The data fine-tuning module uses the high-quality inference dataset to perform multi-task supervised fine-tuning of the student model, including inference chain generation, error correction, and missing inference completion tasks. The inspection sampling and evaluation module adopts a diverse sampling strategy based on reinforcement learning, dynamically adjusts temperature sampling and kernel sampling parameters, generates multiple sets of outputs, and uses consistency scores to filter the weak point dataset. The iterative enhancement module targets the weak point dataset by rewriting instructions in a variety of ways through a multi-perspective instruction rewriting agent, dynamically expanding the high-quality inference dataset, and continuously optimizing model performance until the model reaches a preset performance threshold on a cross-domain test set.

[0039] The third embodiment of this application also provides a computer-readable storage medium storing computer program code, which, when executed by the processor, causes the electronic device to perform the relevant method steps in the above method embodiments.

[0040] In the embodiments provided in this application, it should be understood that the disclosed methods and systems can be implemented in other ways. The system embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, and can be electrical, mechanical, or other forms. Furthermore, in the various embodiments of the present invention, all functional units can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0041] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or smart optical disks.

[0042] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A data augmentation method based on multi-agent self-evolution and hybrid evaluation, characterized in that, Includes the following steps: S1: Utilize the teacher model as a multi-agent collaborative system, including reasoning agent, evaluation agent, reflection agent and memory management agent, to iteratively generate a high-quality reasoning dataset with traceable reasoning paths and self-verifying labels; S2: Use the high-quality inference dataset to perform multi-task supervised fine-tuning of the student model, including inference chain generation, error correction, and missing inference completion tasks; S3: Employ a diversity sampling strategy based on reinforcement learning to dynamically adjust temperature sampling and kernel sampling parameters, generate multiple sets of outputs, and use consistency scores to filter weak point datasets. S4: For the aforementioned weak point dataset, after rewriting the instructions in a variety of ways by the multi-view instruction rewriting agent, return to S1 for deep inference distillation to generate new augmented data and combine it with high-quality inference data. Repeat S2 to S4 until the student model reaches the preset performance threshold on the cross-domain test set.

2. The data augmentation method based on multi-agent self-evolution and hybrid evaluation according to claim 1, characterized in that, In step S1, the teacher model is used as a multi-agent collaborative system, including a reasoning agent, an evaluation agent, a reflection agent, and a memory management agent. A high-quality reasoning dataset with traceable reasoning paths and self-verifying labels is generated iteratively, including: The reasoning agent receives the original question input, identifies key features in the question based on the attention mechanism, decomposes the question into diagnostic sub-tasks based on the key features, and uses the attention mechanism to determine the importance of each key feature. Based on the importance, a hierarchical reasoning structure is generated and deduced step by step according to medical logic, generating a reasoning chain and answer containing detailed reasoning steps. The reasoning chain and answer are evaluated in multiple dimensions by an evaluation agent. Evaluation indicators are obtained based on factual accuracy, logical coherence, and clinical rationality. Evaluation results are calculated and output based on the evaluation indicators. The correctness of the reasoning chain and answer is judged based on the evaluation results. A historical error pattern library is constructed. The reflective agent identifies and analyzes logical loopholes or evidence gaps in the reasoning chain based on the historical error pattern library, predicts potential risk points through historical error patterns, and generates feedback guidance and preventive guidance suggestions. The memory management agent records the complete reasoning history of local sessions, maintains a historical error pattern library, retrieves processing experience of similar historical medical records, provides real-time alerts for repeated errors, and integrates the latest clinical knowledge to optimize the historical error pattern library. The system acquires evaluation scores during the iteration process of the reasoning agent, evaluation agent, reflection agent, and memory management agent. When the evaluation score reaches a threshold, it outputs a reasoning chain and answer that includes a complete iteration history, specific content of modifications, and reasons. It also adds quality score labels, uncertainty labels, and verification status labels to generate high-quality reasoning data.

3. The data augmentation method based on multi-agent self-evolution and hybrid evaluation according to claim 1, characterized in that, In step S2, the student model is fine-tuned with multi-task supervision using the high-quality inference dataset through a dynamic course learning strategy, including: Acquire a high-quality inference dataset and perform data formatting processing, organizing each data point into instruction-response pairs; The dynamic curriculum learning strategy uses formatted data to fine-tune student models. Initially, the model learns the complete inference chain, training it to learn standard diagnostic procedures and reasoning patterns, establish logical connections between medical concepts, and master professional terminology and assessment standards. The intermediate learning process involves gradually hiding some reasoning steps, first concealing the conclusion, then gradually reducing hints in the intermediate reasoning steps, and finally dynamically selecting the hidden content. This trains students' ability to autonomously complete the reasoning chain, apply medical knowledge, and adapt to clinical situations. In the later stages, only questions and answers are provided for learning, enabling students to internalize the complete diagnostic reasoning logic of the model, form an automated clinical thinking mode, and have the generalization ability to handle rare cases; The performance of the student model is obtained during the fine-tuning training process. The performance level of the model is characterized by the accuracy rate. The course progress is adjusted in a personalized manner according to the performance level. Multi-task learning is used to combine direct answer generation tasks, intermediate reasoning completion tasks, and error step correction tasks to optimize the reasoning ability of the student model.

4. The data augmentation method based on multi-agent self-evolution and hybrid evaluation according to claim 1, characterized in that, In step S3, a reinforcement learning-based diversity sampling strategy is adopted to dynamically adjust the temperature sampling and kernel sampling parameters, generating multiple sets of outputs, specifically: For each question in the high-quality inference dataset, we perform reinforcement learning-based diversity sampling, evaluate the model performance status based on the student model's current accuracy level, output diversity, and error type distribution characteristics, and analyze the current question input and model performance status. In each sampling, a reinforcement learning policy network generates temperature parameters T and Top-p parameters, generating N different outputs, each including diverse inference chains and corresponding answers; The Jaccard similarity between different groups of outputs is calculated to obtain the word overlap rate as a surface diversity index. The cosine similarity matrix between different groups of outputs is calculated to obtain the embedding space distance as a semantic diversity index. Furthermore, the consistency of different groups of outputs with medical logic is verified as a logical rationality index. The surface diversity index, semantic diversity index, and logical rationality index are used to perform diversity quantification analysis. When the diverse quantitative analysis results of N different outputs meet the preset standards, consistency screening is performed.

5. The data augmentation method based on multi-agent self-evolution and hybrid evaluation according to claim 4, characterized in that, In step S3, a consistency score is calculated using a consistency evaluator that combines rules and models to filter the weak point dataset, including: A consistency evaluator is constructed based on a rule module and a model module. In the rule module, a precise fact check is performed based on a medical knowledge base. For each question, the standard answer generated by the teacher model and N different outputs generated by the student model are obtained. Key medical facts are extracted from the standard answer. The system checks whether the N different outputs generated by the student model completely cover the facts and obtains the fact accuracy score. In the model module, a logic and rationality assessment is performed, a decision tree for the standard answer is drawn, the deviation of the reasoning path of the N different outputs generated by the student model is compared, a logical coherence score is obtained, and knowledge errors, logical errors and omission errors are marked to obtain a decision rationality score. The comprehensive module uses a weighted scoring method to obtain a consistency score based on the accuracy of facts, logical coherence, and reasonableness of decision-making. Response samples with consistency scores less than a preset threshold are selected to generate a weakness dataset, and each weak sample in the weakness dataset is labeled with defects.

6. The data augmentation method based on multi-agent self-evolution and hybrid evaluation according to claim 1, characterized in that, In step S4, for the vulnerability dataset, the instructions are rewritten in a variety of ways by the instruction rewriting agent, including: A vulnerability dataset is acquired and grouped according to the defect type. An instruction rewriting agent is constructed, and rewriting requirements are randomly selected to rewrite the grouped vulnerability samples. The rewriting requirements include: changing the perspective of instruction questioning, replacing clinical word meanings, introducing additional information, expanding clinical context, and converting between Chinese and English. The rewritten question instructions are subjected to semantic consistency checks. A sentence embedding model is used to calculate the semantic similarity with the original instructions. If the similarity is lower than the threshold, the instructions are rewritten. The rewritten question instructions are then pre-screened by a teacher model.

7. The data augmentation method based on multi-agent self-evolution and hybrid evaluation according to claim 6, characterized in that, Deep inference distillation and re-execution includes: The rewritten problem instructions are input into the reasoning agent, evaluation agent, reflection agent, and memory management agent to strengthen the verification mechanism for controversial conclusions and force verification steps to be inserted into the original weak points, thereby generating new reasoning chains and answers. Record historical error patterns. If the similarity between the current case and the historical error pattern is greater than a preset similarity threshold, then based on the historical error pattern, a similar risk is indicated, and new enhanced data is generated from the reasoning chain and answer with similar risk labels. Stratified sampling is performed on the base data of the new augmented data and high-quality inference dataset, and data merging is performed by dynamically weighting using reinforcement learning agents to dynamically expand the high-quality inference dataset. Security enhancements are performed on several hospital data sources using a federated learning mechanism.

8. A data augmentation system based on multi-agent self-evolution and hybrid evaluation, characterized in that, For implementing the data augmentation method based on multi-agent self-evolution and hybrid evaluation as described in any one of claims 1-7, the system includes: a deep inference distillation module, a data fine-tuning module, a verification sampling and evaluation module, and an iterative enhancement module; The deep reasoning distillation module utilizes a teacher model as a multi-agent collaborative system, including a reasoning agent, an evaluation agent, a reflection agent, and a memory management agent, and iteratively generates a high-quality reasoning dataset with traceable reasoning paths and self-verifying labels. The data fine-tuning module uses the high-quality inference dataset to perform multi-task supervised fine-tuning of the student model, including inference chain generation, error correction, and missing inference completion tasks. The inspection sampling and evaluation module adopts a diverse sampling strategy based on reinforcement learning, dynamically adjusts temperature sampling and kernel sampling parameters, generates multiple sets of outputs, and uses consistency scores to filter the weak point dataset. The iterative enhancement module targets the weak point dataset by rewriting instructions in a variety of ways through a multi-perspective instruction rewriting agent, dynamically expanding the high-quality inference dataset, and continuously optimizing model performance until the model reaches a preset performance threshold on a cross-domain test set.

Images