Three-stage optimization method for improving problem-solving ability of large language model
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XUZHOU NORMAL UNIVERSITY
- Filing Date
- 2026-07-07
- Publication Date
- 2026-08-07
AI Technical Summary
[0007]现有方法在多智能体协作、自动提示词优化、反馈修正和路径搜索等方面取得了进展，但其关注重点多集中于局部提示改写、单轮搜索或单一评估流程
本发明通过基于检索增强辅助分类的题目类型识别、基于多智能体协作的局部反馈迭代、以及基于全局生成的提示词优化，形成提升大语言模型解题效果的方法，能够有效提升大语言模型解题效果。
Smart Images
Figure CN122529007A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of artificial intelligence and natural language processing technology, specifically involving a three-stage optimization method to improve the problem-solving ability of large language models. Background Technology
[0002] The accuracy of large language models in solving mathematical problems is influenced by factors such as problem understanding, input instruction design, and solution strategy selection. Problem understanding determines whether the model can correctly identify the problem type, quantitative relationships, and solution objective; input instruction design determines whether the model can organize the intermediate reasoning process as expected; and solution strategy selection determines the solution path and result stability of the model under different difficulties and conditions. The degree of matching among these three factors directly affects the correctness, stability, and generalization ability of the model's output. To improve the overall accuracy of large language models in mathematical problem-solving tasks, existing research has proposed various technical methods.
[0003] Multi-agent collaboration: In their paper "AutoGen: Enabling Next-Gen LLM Applications via Multi-AgentConversation" published at COLM 2024 (Conference on Large Language Models), Qingyun Wu et al. proposed AutoGen, which completes tasks through the interaction between multiple conversational agents and allows developers to define the behavior patterns of different agents.
[0004] Automatic prompt optimization: In their paper "Large Language Models Are Human-Level Prompt Engineers" published at ICLR 2023 (International Conference on Learning Representations), Yongchao Zhou et al. proposed Automatic Prompt Engineer, which models instruction generation and selection as a search problem and optimizes the prompt effect through candidate instruction generation and scoring selection.
[0005] Feedback Correction: Aman Madaan et al. proposed Self-Refine in their paper "Self-Refine: Iterative Refinement with Self-Feedback" published in Advances in Neural Information Processing Systems 36 (NeurIPS 2023). Self-Refine improves the initial output of the model through an iterative process of "generation-feedback-regeneration".
[0006] Path search: Shunyu Yao et al. proposed Tree of Thoughts in their paper "Tree of Thoughts: Deliberate Problem Solving with Large Language Models" published at NeurIPS 2023 (the 37th Conference on Neural Information Processing Systems). Tree of Thoughts extends single-path reasoning to multi-path exploration, self-evaluation, and backtracking, thereby enhancing planning and search capabilities in complex problems.
[0007] Existing methods have made progress in areas such as multi-agent collaboration, automatic suggestion optimization, feedback correction, and path search, but their focus is mostly on local suggestion rewriting, single-round search, or single evaluation processes. A review of automatic suggestion optimization also points out that this field still faces challenges such as diverse optimization targets, high search and evaluation costs, and insufficient robustness and generality. A review of multi-agent systems indicates that current methods still face challenges in coordination, planning, and stable deployment. Summary of the Invention
[0008] The purpose of this invention is to provide a three-stage optimization method to improve the problem-solving ability of large language models, which can effectively improve the problem-solving performance of large language models.
[0009] To achieve the above objectives, this invention provides a three-stage optimization method for improving the problem-solving ability of large language models, comprising the following steps: S1, in the first stage, the dataset is normalized for type labels and divided into training set, validation set and unseen test set. A question knowledge base containing question text and standard type labels is constructed. The question text is vectorized to obtain question vectors. The question knowledge base is retrieved based on vector distance to obtain auxiliary classification features, which are then fused with the question text features to construct a question type recognition model and identify the standard type labels. S2, in the second stage, optimize the subject area based on the standard type labels identified in S1, perform global generation through multi-agent collaboration, and obtain the optimal prompt word for this stage; In the third stage, based on the optimal prompt words in S2, multi-agent collaborative local feedback iterative optimization is performed to form a prompt word library for question types, thereby improving the problem-solving performance of the large language model.
[0010] As a further aspect of the present invention: S1 specifically includes: S1.1 Based on the preset standard type set, alias merging table and type data space, calculate the mapping of the normalized question text in the standard type space; establish the correspondence between the question text and the standard type labels, and form a question knowledge base; S1.2, Based on the question knowledge base, retrieve the Top-K similar questions from the training set, and generate auxiliary classification features based on the standard type labels and similarity scores of the similar questions; the auxiliary classification features include the frequency of occurrence of each standard type and the similarity score of each standard type of question; S1.3, based on the standard type space, extract character-level TF-IDF features and word-level TF-IDF features from the question text; train two support vector machine models, SVM1 and SVM2, as a joint augmentation classification model; the training of SVM1 integrates character-level fused TF-IDF features and auxiliary classification features; the training of SVM2 integrates word-level fused TF-IDF features, character-level TF-IDF features, and auxiliary classification features; SVM1 and SVM2 compete on the validation set, and the model with the higher accuracy is selected as the final question type recognition model.
[0011] As a further aspect of the present invention: S2 specifically includes: S2.1, the planning agent, teacher agent, and student agent work together based on the identified standard type labels to generate prompt words specific to the standard type label as candidate prompt words; S2.2, evaluate the accuracy of candidate prompt words on the validation set, and filter the second-stage candidate results based on the dynamic threshold corresponding to the baseline without prompt words. t The dynamic threshold for the second stage of the round is: ; in, A 0( c The ) represents the accuracy of the no-cues baseline on the validation set. γ Represents the desired gain parameter. β The attenuation parameter is represented by `clip(·)`, which represents the cutoff function that limits the threshold to between 0 and 1. When the accuracy of the candidate prompt words reaches the dynamic threshold, the second stage ends and the optimal prompt word is output; when the accuracy of the candidate prompt words does not reach the dynamic threshold, the iteration continues, and the second stage ends when the maximum number of iterations is reached, and the optimal prompt word in the second stage is output.
[0012] As a further aspect of the present invention: S2.1 specifically includes: S2.1.1, the planning agent generates optimization steps for prompt words specific to the identified standard type label; the teacher agent generates corresponding constraint guidance information for each prompt word optimization step; the student agent generates candidate prompt words based on the prompt word optimization steps and the corresponding constraint guidance information. S2.1.2 For the standard type labels to be identified, multiple student agents generate candidate prompt words in parallel. Each student agent independently tests and improves based only on the candidate prompt words of the previous round, and different student agents do not share intermediate rewriting results, forming mutually independent candidate prompt word evolution paths.
[0013] As a further aspect of the present invention: S3 specifically includes: S3.1, the review agent obtains the evaluation results of the optimal prompt words in the second stage on the validation set and generates suggested prompt words, including: retaining effective content, correcting failure reasons, and supplementing reasoning steps; S3.2, the student agent generates new prompts based on the suggested prompts, evaluates the prompts on the validation set, and retains the one with higher accuracy as the current optimal prompt; S3.3 When the number of iterations between the review agent and the student agent reaches the maximum number of iterations, or when the accuracy of the generated prompt words does not improve for several consecutive rounds, output the prompt words to be filtered, filter and verify to determine the final prompt words corresponding to each standard type label, and form a prompt word library for question types.
[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention forms a method to improve the problem-solving performance of large language models by using retrieval-enhanced auxiliary classification for question type identification, local feedback iteration based on multi-agent collaboration, and optimization based on globally generated prompt words. This method can effectively improve the problem-solving performance of large language models.
[0015] This invention utilizes techniques such as type recognition, multi-agent systems, large language models, and prompt word optimization to provide methodological support for improving the problem-solving capabilities of large language models. Attached Figure Description
[0016] Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 This is a diagram showing the type and category processing of the present invention; Figure 3 This is the second stage of the global generation of prompt word optimization process in this invention; Figure 4 This is the local feedback iteration diagram for the third stage of the present invention. Detailed Implementation
[0017] The present invention will be further illustrated by the following examples.
[0018] A three-stage optimization method for improving the problem-solving ability of large language models includes the following steps: S1, in the first stage, the dataset is normalized for type labels and divided into training set, validation set and unseen test set. A question knowledge base containing question text and standard type labels is constructed. The question text is vectorized to obtain question vectors. The question knowledge base is retrieved based on vector distance to obtain auxiliary classification features, which are then fused with the question text features to construct a question type recognition model and identify the standard type labels of the questions.
[0019] Furthermore, S1 specifically includes: S1.1 Based on the preset standard type set, alias merging table and type data space, calculate the mapping of the normalized question text in the standard type space; establish the correspondence between the question text and the standard type labels, and form a question knowledge base; S1.2, Based on the question knowledge base, retrieve the Top-K similar questions from the training set, and generate auxiliary classification features based on the standard type labels and similarity scores of the similar questions; the auxiliary classification features include the frequency of occurrence of each standard type and the similarity score of each standard type of question; S1.3, based on the standard type space, extract character-level TF-IDF features and word-level TF-IDF features from the question text; train two support vector machine models, SVM1 and SVM2, as a joint augmentation classification model; the training of SVM1 integrates character-level fused TF-IDF features and auxiliary classification features; the training of SVM2 integrates word-level fused TF-IDF features, character-level TF-IDF features, and auxiliary classification features; SVM1 and SVM2 compete on the validation set, and the model with the higher accuracy is selected as the final question type recognition model.
[0020] S2, in the second stage, optimizes the subject area based on the standard type labels identified in S1, performs global generation through multi-agent collaboration, and obtains the optimal prompt word for this stage.
[0021] Furthermore, S2 specifically includes: S2.1, the planning agent, teacher agent, and student agent work together based on the identified standard type labels to generate prompt words specific to the standard type label as candidate prompt words; S2.1 specifically includes: S2.1.1, the planning agent generates optimization steps for prompt words specific to the identified standard type label; the teacher agent generates corresponding constraint guidance information for each prompt word optimization step; the student agent generates candidate prompt words based on the prompt word optimization steps and the corresponding constraint guidance information. S2.1.2, For the identified question type, multiple student agents generate candidate prompts in parallel. Each student agent independently tests and improves based on the candidate prompts from the previous round, and different student agents do not share intermediate rewriting results, forming mutually independent candidate prompt evolution paths. S2.2, evaluate the accuracy of candidate prompt words on the validation set, and filter the second-stage candidate results based on the dynamic threshold corresponding to the baseline without prompt words. t The dynamic threshold for the second stage of the round is: ; in, A 0(c) represents the accuracy of the no-catch word baseline on the validation set, γ represents the expected gain parameter, β represents the decay parameter, and clip(·) represents the cutoff function that limits the threshold to between 0 and 1. When the accuracy of the candidate prompt words reaches the dynamic threshold, the second stage ends and the optimal prompt word is output; when the accuracy of the candidate prompt words does not reach the dynamic threshold, the iteration continues, and the second stage ends when the maximum number of iterations is reached, and the optimal prompt word in the second stage is output.
[0022] In the third stage, based on the optimal prompt words in S2, multi-agent collaborative local feedback iterative optimization is performed to form a prompt word library for question types, thereby improving the problem-solving performance of the large language model.
[0023] Furthermore, S3 specifically includes: S3.1, the review agent obtains the evaluation results on the validation set based on the optimal prompt words in the second stage, and generates suggested prompt words, including: retaining effective content, correcting failure reasons, and supplementing reasoning steps; S3.2, the student agent generates new prompts based on the suggested prompts, evaluates the prompts on the validation set, and retains the one with higher accuracy as the current optimal prompt; S3.3 When the number of iterations between the review agent and the student agent reaches the maximum number of iterations, or when the accuracy of the generated prompt words does not improve for several consecutive rounds, output the prompt words to be filtered, filter and verify the prompt words to be filtered, determine the final prompt words corresponding to each standard type label, and form a prompt word library for question types.
[0024] The overall process of implementing this invention is as follows: Figure 1As shown, the process includes a training and optimization phase and an invocation phase. In the training and optimization phase, the questions are preprocessed, including question normalization and type standardization. The first phase involves question type identification, where retrieval-enhanced auxiliary classification features are constructed based on a question knowledge base, and a question type identification model is trained to obtain standard type labels for the input questions. The second phase involves global prompt word generation, where the planning agent, teacher agent, and student agent collaboratively generate prompt words. The planning agent generates specific prompt word optimization steps based on the standard type labels of the identified questions, the teacher agent provides constraint guidance information, and the student agent generates candidate prompt words. The interaction ends when the accuracy reaches a dynamic threshold or the maximum number of iterations, and the optimal prompt word for this phase is output. In the third phase, the review agent and student agent engage in local iterative interaction. The review agent generates executable suggestions based on the optimal prompt words output in the second phase, the student agent rewrites the prompt words based on the executable suggestions, and the review agent generates the next round of modification suggestions based on the accuracy verified by the target agent. After generating prompt words to be selected, the final prompt words corresponding to each question type are determined through screening and verification, forming a question type prompt word library. During the invocation phase, after receiving the user's input of the question to be solved, the system first calls the question type recognition model to identify the type and outputs the standard type label of the question. Then, based on the standard type label, it selects the corresponding prompt word from the question type prompt word library and inputs the selected prompt word and the question to be solved into the large language model, and finally outputs the question result.
[0025] Specifically, the first phase involves standardizing the labeled question dataset, normalizing the type labels, constructing a question knowledge base, and training a question type recognition model. Let the labeled question dataset be: ; in, x i Indicates the first i The question text of the problem. Indicates the first i The answer field for this question, Indicates the first i The original type label of the question.
[0026] The question text and answer fields are standardized, including removing invalid characters, unifying capitalization, unifying number and symbol formats, removing extra spaces, and standardizing the answer expression format, resulting in the processed question dataset: ; in, This represents the standardized question text. This represents the normalized answer field.
[0027] Based on the preset standard type set and alias merging table, the original type labels are normalized into standard type labels: ; in, Φ () represents a type label mapping function. C Represents the set of standard types under a unified type space. Show the first i The standard type labels for the questions after normalization are shown in Table 1. The results of type label normalization and examples of question text feature words corresponding to each standard type are shown in Table 1.
[0028] Table 1. Examples of Type Label Normalization and Title Text Features
[0029] After completing the type label normalization, the normalized labeled question dataset will be... D Divided into training set, validation set, and unseen test set: ; in, Represents the training set, Represents the validation set. This indicates that no test set is available, and all three sets are pairwise disjoint. The training set is used for building the question knowledge base and training the question type recognition model; the validation set is used for selecting the type recognition model, generating prompts, and optimizing prompts; the absence of a test set means it is not used in training or prompt optimization, but only for final performance evaluation.
[0030] Furthermore, the validation set is divided into a first validation subset and a second validation subset: ; in, This represents the first verification subset, used for the global generation and initial screening of candidate prompts in the second stage; This represents the second verification subset, used for the third stage of local iterations based on review feedback and the final candidate prompt word selection. and They do not overlap.
[0031] Establish a correspondence between question texts and standard type tags, and form a question knowledge base: ; in, K This represents the question knowledge base. This indicates the normalized question text. The problem vector obtained by vectorization.
[0032] For any question Vectorize it asv and in the question knowledge base K The search engine retrieves the top-K similar questions to obtain a set of similar questions: ; in, This represents the similar question texts retrieved through the search. This indicates a standard type label for similar questions. Represents the current question vector v Similar question vectors Similarity score between them.
[0033] Type recognition processing such as Figure 2 As shown, auxiliary classification features are generated based on the standard type labels and similarity scores of the Top-K similar questions. For each standard type in the unified type space... c We count the number of times each item appears in the set of similar questions and sum the corresponding similarity scores. Therefore, the auxiliary classification feature is represented as follows: ; in, Represents the standard types in the unified type space. This represents the auxiliary classification features generated from the retrieval results of the question knowledge base; Indicates standard type Number of times it appears in Top-K similar questions; This indicates that among the Top-K similar questions, those belonging to the standard type... The sum of similarity scores.
[0034] For the training set Extract character-level TF-IDF features and word-level TF-IDF features respectively. (The question is incomplete.) The character-level TF-IDF features are Word-level TF-IDF features are .
[0035] auxiliary classification features By fusing character-level TF-IDF features and using the standard type labels of questions in the training set as supervision labels, a first linear support vector machine (SVM1) is trained to establish a character-level question type recognition model. ; By fusing character-level TF-IDF features, word-level TF-IDF features, and auxiliary classification features, and using the standard type labels of questions in the training set as supervision labels, a second linear support vector machine (SVM2) is trained to establish a joint character-level and word-level question type recognition model. ; In the validation set The recognition accuracy of the two models is calculated separately: ; ; The models with higher accuracy were selected as the final question type identification models. ; Output the question Standard type tags: ; in, The standard type label represents the question to be processed.
[0036] like Figure 3 As shown, the second stage is the global generation of prompt word optimization process. For any standard type c First, calculate the baseline of the no-prompt words in the first subset of the validation set. Accuracy on: ; in, Indicates the baseline without prompts. This represents the accuracy on the corresponding dataset. The first subset of the validation set The set of questions that belong to standard type c.
[0037] Candidate prompt words are generated collaboratively by the planning agent, teacher agent, and student agent. The planning agent, based on the prompt word optimization objective of the current standard type label, decomposes the prompt word optimization task into multiple execution steps, denoted as: ; in, Indicates the first m Each execution step.
[0038] The teacher agent generates step constraint information for each execution step, and multiple student agents update the prompts sequentially based on the step constraint information to obtain the result. k Round Step prompts: ; in, This indicates that the student agent performs a prompt word rewriting operation based on the guidance information from the teacher agent. The teacher's intelligent agent is targeting the first m The guidance information generated by each execution step is shown in Table 2.
[0039] Table 2. Correspondence between Standard Type Labels and Teacher Intelligent Agent Constraint Guidance Information
[0040] When the k After all steps in the round are completed, the candidate prompts for that round are obtained. It is in the first subset of the verification set. The accuracy rate is: ; Suppose the baseline of the no-catch word is in the first subset of the validation set. The accuracy rate is , No. k The dynamic threshold for the second stage of the round is: ; in, Represents the desired gain parameter. Indicates the attenuation parameter. This represents a cutoff function that limits the threshold to between 0 and 1.
[0041] When the candidate prompt word is in the first subset of the validation set The accuracy rate satisfies: ; If the second phase ends, the candidate suggestion word is output; otherwise, the next round of search continues. When the maximum number of rounds in the second phase is reached, the optimal suggestion word with the highest accuracy in the second phase is output. .
[0042] like Figure 4 The diagram shown illustrates the local feedback iteration in the third stage. For any standard type c, the optimal prompt word output in the second stage is... This serves as the initial prompt for the third phase. At the start of the third phase, the review agent uses this initial prompt and its position in the first subset of the validation set. accuracy Generate an executable suggestion; multiple student agents each generate a new candidate prompt word based on the executable suggestion. and in the first subset of the verification set The accuracy rates were evaluated separately as follows: ; When the new candidate suggestion word meets the following conditions: ; If the candidate suggestion is not selected, it is retained as a suggestion to be filtered; otherwise, it is counted as a failure. The review agent generates the next round of modification suggestions based on the candidate suggestion with the highest accuracy and the evaluation results of the target agent. Multiple student agents then generate new candidate suggestions based on the new modification suggestions. The maximum number of iterations is reached when the number of iterations between the review agent and the student agents reaches the maximum number of iterations. T or continuous accuracy y Stop iterating if the wheel is not promoted.
[0043] Subsequently, the retained prompts to be filtered will be placed in the second subset of the validation set. The system evaluates the suggestions and selects the most accurate suggestion as the final suggestion for the current standard type. When accuracy is the same, priority is given to candidate suggestions that are complete, have clear steps, and are shorter. The third stage ends, and the current best suggestion is selected. As the final prompt word.
[0044] Therefore, we obtain the question type hint vocabulary: ; Where B represents the question type hint vocabulary, This indicates the final prompt word corresponding to standard type c.
[0045] This invention ultimately generates a question type hint word library B corresponding to different question types. After the user inputs a question to be solved, the system identifies the question type. Based on the identification result, the system calls up the hint words corresponding to the question type. The system inputs the hint words and the question to be solved together into the large language model. The large language model completes condition extraction, reasoning analysis, and answer generation based on the hint words.
[0046] During the online call phase, a type prediction is performed on the input question to obtain standard type labels. And based on the standard type label, from the question type hint vocabulary. B After retrieving the corresponding prompt words, the large language model is invoked. Let the unseen test set be... Its standard type label is Based on the predicted standard type labels, select the corresponding prompt words from the question type prompt word library, and denot them as... .
[0047] During the evaluation, the baseline output and the prompt word output were as follows: ; ; in, This represents the function for solving large language models. This indicates the combination of the prompt word and the question text.
[0048] Baseline accuracy and prompt word accuracy are as follows: ; ; in, This indicates an indicator function that takes the value 1 when the condition inside the parentheses is true and takes the value 0 when the condition inside the parentheses is false. Indicates the first line below the no-prompt baseline i Output the answer to the question. Indicates the first [word] after the prompt word is invoked. i Output the answer to the question. Indicates the first i The standard answer to the question.
[0049] The absolute gain is: ; Finally, the output includes baseline accuracy, cue word accuracy, absolute gain, and accuracy metrics by prediction type.
[0050] Baseline accuracy refers to the accuracy rate obtained by the large language model directly solving questions without using optimized prompts or question-type-specific prompts, representing the model's original problem-solving ability. Prompt accuracy refers to the accuracy rate of the large language model on the same batch of questions after adding prompts generated by this method, measuring the improvement effect of prompts on the model's problem-solving performance. Absolute gain is the difference between prompt accuracy and baseline accuracy, i.e., "prompt accuracy - baseline accuracy", usually expressed as a percentage, used to intuitively reflect the actual improvement in accuracy. Accuracy indicators statistically analyzed by prediction type refer to first identifying the prediction type of each question based on the question type, and then statistically analyzing the model's problem-solving accuracy under different prediction types, such as the accuracy rates of probability questions, ratio questions, monetary calculation questions, and combinatorial counting questions, used to analyze the applicability and performance differences of this method on different question types.
[0051] Through the above methods, this invention can improve the adaptability of large language models to middle school mathematics problems; it can improve problems such as unstable reasoning paths, unclear calculation steps, and inaccurate answer generation in solving mathematics problems using large language models; it can enable large language models to generate more standardized, clear, and reliable problem-solving processes, thus assisting middle school mathematics teaching. Teachers can use this invention to obtain high-quality problem-solving processes, thereby improving the efficiency of lesson preparation, explanation, and Q&A, and reducing repetitive teaching burdens.
Claims
1. A three-stage optimization method for improving the problem-solving ability of large language models, characterized in that, Includes the following steps: S1, in the first stage, the dataset is normalized for type labels and divided into training set, validation set and unseen test set. A question knowledge base containing question text and standard type labels is constructed. The question text is vectorized to obtain question vectors. The question knowledge base is retrieved based on vector distance to obtain auxiliary classification features, which are then fused with question text features to construct a question type recognition model and identify the standard type labels of the questions. S2, in the second stage, optimize the subject area based on the standard type labels identified in S1, perform global generation through multi-agent collaboration, and obtain the optimal prompt word for this stage; In the third stage, based on the optimal prompt words in S2, multi-agent collaborative local feedback iterative optimization is performed to form a prompt word library for question types, thereby improving the problem-solving performance of the large language model.
2. The three-stage optimization method for improving the problem-solving ability of large language models according to claim 1, characterized in that, S1 specifically includes: S1.1 Based on the preset standard type set, alias merging table and type data space, calculate the mapping of the normalized question text in the standard type space; establish the correspondence between the question text and the standard type labels, and form a question knowledge base; S1.2, Based on the question knowledge base, retrieve the Top-K similar questions from the training set, and generate auxiliary classification features based on the standard type labels and similarity scores of the similar questions; the auxiliary classification features include the frequency of occurrence of each standard type and the similarity score of each standard type of question; S1.3, based on the standard type space, extract character-level TF-IDF features and word-level TF-IDF features from the question text; train two support vector machine models, SVM1 and SVM2, as a joint augmentation classification model; the training of SVM1 integrates character-level fused TF-IDF features and auxiliary classification features; the training of SVM2 integrates word-level fused TF-IDF features, character-level TF-IDF features, and auxiliary classification features; SVM1 and SVM2 compete on the validation set, and the model with the higher accuracy is selected as the final question type recognition model.
3. The three-stage optimization method for improving the problem-solving ability of large language models according to claim 1, characterized in that, S2 specifically includes: S2.1, the planning agent, teacher agent, and student agent work together based on the identified standard type labels to generate prompt words specific to the standard type label as candidate prompt words; S2.2, evaluate the accuracy of candidate prompt words on the validation set, and filter the second-stage candidate results based on the dynamic threshold corresponding to the baseline without prompt words. t The dynamic threshold for the second stage of the round is: ； in, A 0( c The ) represents the accuracy of the no-cues baseline on the validation set. γ Represents the desired gain parameter. β The attenuation parameter is represented by `clip(·)`, which represents the cutoff function that limits the threshold to between 0 and 1. When the accuracy of the candidate prompt words reaches the dynamic threshold, the second stage ends and the optimal prompt word is output; when the accuracy of the candidate prompt words does not reach the dynamic threshold, the iteration continues, and the second stage ends when the maximum number of iterations is reached, and the optimal prompt word in the second stage is output.
4. The three-stage optimization method for improving the problem-solving ability of large language models according to claim 3, characterized in that, S2.1 specifically includes: S2.1.1, the planning agent generates optimization steps for prompt words specific to the identified standard type label; the teacher agent generates corresponding constraint guidance information for each prompt word optimization step; the student agent generates candidate prompt words based on the prompt word optimization steps and the corresponding constraint guidance information. S2.1.2 For the standard type labels to be identified, multiple student agents generate candidate prompt words in parallel. Each student agent independently tests and improves based only on the candidate prompt words of the previous round, and different student agents do not share intermediate rewriting results, forming mutually independent candidate prompt word evolution paths.
5. The three-stage optimization method for improving the problem-solving ability of large language models according to claim 1, characterized in that, S3 specifically includes: S3.1, the review agent obtains the evaluation results on the validation set based on the optimal prompt words in the second stage, and generates suggested prompt words, including: retaining effective content, correcting failure reasons, and supplementing reasoning steps; S3.2, the student agent generates new prompts based on the suggested prompts, evaluates the prompts on the validation set, and retains the one with higher accuracy as the current optimal prompt; S3.3 When the number of iterations between the review agent and the student agent reaches the maximum number of iterations, or when the accuracy of the generated prompt words does not improve for several consecutive rounds, output the prompt words to be filtered, filter and verify the prompt words to be filtered, determine the final prompt words corresponding to each standard type label, and form a prompt word library for question types.