Multi-agent-assisted question-setting methods, systems, devices, and media for educational testing

By leveraging multi-agent collaborative work and a centralized task scheduling engine, the problem of insufficient subject adaptability and weak cross-question type generalization ability in existing intelligent question-setting solutions has been solved. This enables the generation of standardized test questions that are adaptable to multiple subjects and multiple question types, while meeting the data confidentiality and information technology compliance requirements of the educational examination environment, thereby improving the efficiency and quality of question setting.

CN123118044APending Publication Date: 2026-11-17QICHEN GUANGZHOU ELECTRONICS TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611316570.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-27
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing intelligent question-setting solutions suffer from insufficient subject adaptability, weak cross-question type generalization ability, fragmented tools, difficulty in forming a closed-loop workflow, and inability to meet the data confidentiality and information technology compliance requirements of educational examination environments.

Method used

Employing a multi-agent collaborative approach, this system dynamically acquires subject knowledge fragments by combining hierarchical structured prompt word templates with a subject knowledge vector database. It then uses a sample question vector database to recall question-setting ideas and structural paradigms. A centralized task scheduling engine orchestrates the processes of image generation, question proofreading, AI trial testing, image and text plagiarism checking, and difficulty analysis in a finite state machine manner, forming an end-to-end closed-loop iterative mechanism. A large language model is deployed offline in a private environment to meet the data confidentiality requirements of the examination venue.

Benefits of technology

It enables the generation of standardized test questions that are adaptable to multiple disciplines and question types. The tool is deeply embedded in the test-setting business process, meets the data confidentiality and information technology compliance requirements of the test venue environment, and improves the efficiency and quality of test setting.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118044A_ABST
    Figure CN123118044A_ABST
Patent Text Reader

Abstract

This invention discloses a multi-agent-assisted question-setting method, system, device, and medium for educational examinations, aiming to solve the problems of insufficient cross-disciplinary adaptability, weak targeting of question type generation, and poor coordination of tools throughout the entire question-setting process in traditional intelligent question-setting. The method receives a question-setting task containing the subject, question type, test points, and difficulty level. It then retrieves a subject knowledge vector database to inject subject context, recalls sample questions, and generates a draft test question. A centralized task scheduling engine then uses a finite state machine to sequentially schedule agents for image matching, proofreading, trial testing, plagiarism checking, and difficulty analysis for multi-dimensional verification. If verification fails, it iterates back until successful, finally completing the test paper assembly and generating scoring criteria, forming an end-to-end question-setting closed loop. This invention significantly improves the targeting and accuracy of test question generation and the efficiency of the entire question-setting process.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and educational informatization technology, and specifically relates to a multi-agent-assisted question-setting method, system, electronic device and storage medium for educational examinations. Background Technology

[0002] The quality of educational examination questions is directly related to educational equity and the quality of talent cultivation. Provincial and higher-level educational examination institutions generally face real challenges such as a surge in the workload of question setting, a shortage of subject matter experts, stringent requirements for examination room confidentiality, and low efficiency in the entire question setting process. Under the traditional manual question setting model, question setters are responsible for the entire chain of work, including material collection, question writing, multiple rounds of review and proofreading, and test paper assembly and scoring, which is difficult to meet the needs of large-scale, multi-disciplinary, and high-frequency examinations.

[0003] The development of large language model technology has provided technical support for the intelligent upgrading of educational examination question setting. However, existing intelligent question setting solutions have the following shortcomings: First, they mostly focus on the generation of questions using a single model, resulting in insufficient subject adaptability and weak cross-question type generalization ability; second, tools such as image matching, plagiarism checking, trial testing, proofreading, difficulty analysis, and test paper generation are not deeply embedded in the question setting business process, and there are gaps between the tools, making it difficult to form a closed-loop workflow; third, most solutions are based on cloud models and do not meet the rigid requirements of physical isolation, data confidentiality, and information technology compliance in the educational examination environment.

[0004] Therefore, there is an urgent need for an intelligent question-setting method that can simultaneously achieve multi-disciplinary adaptation, multi-question standardization, closed-loop verification throughout the entire process, and secure deployment of the examination venue. Summary of the Invention

[0005] The main objective of this invention is to overcome the shortcomings and deficiencies of the prior art and provide a multi-agent-assisted question-setting method, system, device, and medium for educational examinations.

[0006] To achieve the first objective mentioned above, the first aspect of the present invention discloses a multi-agent-assisted question-setting method for educational examinations, the question-setting method comprising the following steps:

[0007] S1. Receive a question-setting task request, and parse the question-setting task request to obtain subject information, question type information, test point scope and target difficulty;

[0008] S2. Based on subject information and the scope of test points, construct prompt words carrying subject knowledge: Using a hierarchical structured prompt word template, the scope of test points is vectorized to obtain a query vector. The query vector is then used to perform semantic similarity retrieval on the subject knowledge vector database to obtain the top k subject knowledge fragments that are semantically most similar to the scope of test points. These subject knowledge fragments are then injected into the hierarchical structured prompt word template to obtain prompt words carrying subject knowledge. The hierarchical structured prompt word template is a four-layer prompt word structure from top to bottom, including a basic role layer, a subject knowledge layer, a proposition constraint layer, and an output format layer. The subject knowledge vector database is obtained by vectorizing and storing multi-subject knowledge nodes and their corresponding knowledge texts in a hierarchical knowledge graph, which is constructed based on curriculum standards and examination syllabi. The subject knowledge fragments correspond to multi-subject knowledge nodes and are used to represent the concepts, rules, formulas, examples, or relationships of the subject knowledge node.

[0009] S3. Generate a draft of the test questions based on the question type information and prompt words. The process is as follows: Using the question type information and the scope of test points as query conditions, perform semantic similarity retrieval on the sample question vector database to recall several sample questions of the same question type and test points. Inject the question-setting ideas and structural paradigms of the sample questions into the question-setting constraint layer of the prompt words as a reference paradigm. Call the large language model to generate the draft of the test questions. The sample question vector database is a vector database obtained by vectorizing and encoding the stems, options, answers, explanations, and question type tags of past exam questions and standard sample questions. The sample questions are labeled question data used to provide question-setting ideas, structural paradigms, and semantic duplication benchmarks.

[0010] S4. The centralized task scheduling engine sequentially schedules the image matching agent, proofreading agent, trial agent, plagiarism detection agent, and difficulty analysis agent in a finite state machine manner to perform image generation status verification, question proofreading status verification, AI trial status verification, image and text plagiarism detection status verification, and difficulty analysis status verification, and generates verification results for each. The centralized task scheduling engine is a task orchestration and execution component that models the question verification task as a finite state machine. The finite state machine includes the sequentially executed image generation status verification, question proofreading status verification, AI trial status verification, image and text plagiarism detection status verification, and difficulty analysis status verification. Preset dependencies and jump conditions are configured between each status verification.

[0011] S5. Determine whether the draft test question has passed all 5 checks based on the verification results of the 5 agents: If the verification result output by any functional agent is "fail", the centralized task scheduling engine will send back the reason for failure and the modification instructions, regenerate the draft test question and iterate until all checks are passed.

[0012] S6. Generate test papers and scoring criteria from the first draft of the test questions that have passed all verifications: The test paper generation agent performs multi-objective optimization from the candidate test question pool according to the test point range and the target difficulty to obtain the test paper.

[0013] Furthermore, the candidate question pool consists of existing questions from the original question bank and initial draft questions. The purpose of setting up the candidate question pool is to compare and select between the reviewed questions from the original question bank and the newly generated and verified questions from this question-setting process within a unified data structure. The candidate question pool records the subject, question type, test point, difficulty, source, similarity, verification status, and version identifier for each question. This allows the test-setting agent to simultaneously evaluate knowledge point coverage, difficulty gradient, question type ratio, and originality within the same constraint space, avoiding duplicate testing or difficulty imbalance caused by the separation of the original question bank and newly generated questions.

[0014] Furthermore, the centralized task scheduling engine invokes the corresponding functional agents to perform the following operations under each state verification:

[0015] During the image generation status verification, the image generation agent generates corresponding imagery for the test questions based on the imagery requirements described in the initial draft of the test questions, and outputs test questions with images.

[0016] During the test question proofreading status verification, the test question proofreading agent performs multi-dimensional proofreading of the initial test question in terms of political, scientific and fairness aspects, and outputs the proofreading results;

[0017] In the AI ​​trial state verification, the AI ​​trial agent simulates the examinee to answer the draft of the test questions, verifies the rationality and correctness of the answers to the draft of the test questions, and outputs the trial results;

[0018] In the image and text plagiarism check status verification, the image and text plagiarism check agent performs semantic-level similarity comparison between the draft test question and the sample questions in the sample question vector database and the existing test questions in the original question bank, and outputs the plagiarism check results;

[0019] During the difficulty analysis state verification, the difficulty analysis agent predicts the difficulty coefficient of the first draft of the test questions based on the trial results and the target difficulty, and outputs the difficulty analysis results.

[0020] A centralized task scheduling engine is employed, which encapsulates the originally independent tools for image matching, proofreading, trial testing, plagiarism checking, and difficulty analysis into state nodes with clearly defined inputs, outputs, and pass / fail conditions. The scheduling engine only calls the next state after the previous state produces an output that satisfies the dependency conditions, and records the task identifier, question version, state identifier, verification conclusion, and modification instructions at each call. This ensures that the multi-agent collaborative process has a definite execution order, a traceable data flow, and a repeatable failure recovery path.

[0021] Furthermore, the preset dependencies and transition conditions configured between each state check in the finite state machine are as follows:

[0022] The image generation status verification depends on the initial draft of the test questions in the test question generation status; the test question proofreading status verification depends on the test questions with images output by the image generation status verification; the AI ​​trial status verification depends on the test questions after the test question proofreading status verification has passed; the image and text plagiarism check status verification depends on the test questions after the AI ​​trial status verification has passed, as well as the sample question vector database and the original question bank; the difficulty analysis status verification depends on the trial results output by the AI ​​trial status verification and the target difficulty.

[0023] In the image generation, question proofreading, AI trial, image and text plagiarism checking, and difficulty analysis states, when the functional agent corresponding to the current state has completed its execution and the output result meets the preset passing conditions, the finite state machine jumps to the next state verification. When any state verification outputs a failing result, the finite state machine jumps back to the question generation state. The centralized task scheduling engine transmits the reason for failure, the corresponding state identifier, and the modification instruction to the agent corresponding to the question generation state, regenerates the initial draft of the question, and executes the subsequent verification states sequentially starting from the image generation state verification.

[0024] The aforementioned dependencies and jump conditions address the issues of uncertain execution order and inaccurate feedback of failure reasons when multiple tools are chained together. Each verification status returns either "pass" or "fail"; when "fail" is returned, the failure field, failure reason, and modification instructions for the question generation status are also output. The scheduling engine uses the status value to determine the next node, uses the status identifier to determine the question version that needs to be reprocessed, and uses the modification instructions to constrain the regeneration of content, thereby avoiding unconditional duplicate generation and making the processing path under the same input and the same set of status rules reproducible.

[0025] Furthermore, the four levels in the hierarchical structured prompt template each perform the following functions:

[0026] The basic role layer defines the subject identity and proposition role of each intelligent agent; wherein, the subjects include Chinese, mathematics, English, physics, chemistry, biology, politics, history and geography;

[0027] The subject knowledge layer carries subject knowledge fragments injected by semantic retrieval;

[0028] The question constraint layer sets the question type specifications, test point requirements, and difficulty constraints;

[0029] The output format layer specifies the output structure and field format of the test questions.

[0030] Decomposing the prompts into a basic role layer, a subject knowledge layer, a proposition constraint layer, and an output format layer decouples the relatively stable subject identity, dynamically retrieved knowledge, task-dependent proposition constraints, and output fields for downstream programs to parse. During actual execution, only the layers relevant to the current task are updated, reusing the others. This reduces prompt rewriting during cross-subject switching, prevents the model from missing constraints due to the mixing of knowledge text, question type rules, and output formats, and facilitates subsequent validation by allowing the agent to check each structured field.

[0031] Furthermore, the semantic similarity retrieval process in step S2 is as follows:

[0032] The scope of test points is vectorized to obtain the query vector;

[0033] Calculate the cosine similarity between the query vector and the knowledge node vectors in the subject knowledge vector database;

[0034] Sort the knowledge nodes in descending order of similarity, take the subject knowledge fragments corresponding to the top k knowledge nodes, and concatenate them to the subject knowledge layer of the hierarchical structured prompt word template.

[0035] Vector semantic retrieval, rather than relying solely on keyword matching, is employed because the same test point may be expressed differently in curriculum standards, examination syllabi, and question-setting tasks. By querying the cosine similarity between vectors and knowledge node vectors, semantically corresponding concepts, rules, formulas, examples, and relationships can be recalled. Then, selecting the top k segments based on similarity achieves a balance between knowledge coverage and model context length, ensuring that generated test questions are constrained by the knowledge of the current test point, rather than solely relying on the inherent knowledge of a large language model.

[0036] Furthermore, in step S3, the recalled sample questions also serve as the benchmark data for semantic similarity comparison by the image-text plagiarism detection agent in step S4.

[0037] The same set of sample questions is reused in both the generation and deduplication phases: in the generation phase, only the question structure, question format, and test point organization are extracted as reference paradigms; in the deduplication phase, the semantics of the question stem, the logic of the options, and the problem-solving approach of the new test questions are compared with those of the sample questions. Using a set of source samples ensures that "reference standards" and "avoiding repetition" are based on a consistent comparison benchmark; when the similarity exceeds a threshold, the scheduling engine sends back modification instructions that require changing the context, numerical relationships, or question paths, in order to balance the standardization of question types and the originality of the test questions.

[0038] Furthermore, the finite state machine modeled by the centralized task scheduling engine includes a task receiving state, a subject knowledge injection state, a question generation state, an image generation state, a question proofreading state, an AI trial state, an image and text plagiarism check state, a difficulty analysis state, a test paper assembly state, a scoring standard generation state, and a completion state. The task receiving state is used to parse the question-setting task request to obtain subject information, question type information, test point scope, and target difficulty. The subject knowledge injection state is used to retrieve the subject knowledge vector database to obtain subject knowledge fragments and generate question prompts. The question generation state is used to generate test questions based on the question prompts and recalled sample questions. The test paper is divided into several stages: Draft Questions; Image Generation Stage (for generating images based on the draft); Question Proofreading Stage (for outputting proofreading results regarding political correctness, scientific rigor, and fairness); AI Trial Stage (for outputting trial results regarding the reasonableness and correctness of answers); Image / Text Plagiarism Check Stage (for outputting similarity comparison results between the draft and benchmark data); Difficulty Analysis Stage (for outputting difficulty analysis results based on trial results and target difficulty); Test Paper Assembly Stage (for multi-objective optimization of the candidate question pool to obtain the test paper); Scoring Criteria Generation Stage (for generating tiered scoring rules for subjective questions in the test paper); and Completion Stage (for outputting the test paper and its accompanying scoring criteria).

[0039] Integrating task reception, knowledge injection, question generation, five types of verification, test paper assembly, scoring criterion generation, and completion status into a single finite state machine ensures that the entire process from question entry to test paper output has clearly defined states. Each state receives the structured output of the previous state and generates the data objects required for the next state; the test paper assembly state only receives candidate questions that have passed all verifications, and the scoring criterion generation state only generates scoring points and stratification rules for subjective questions in the test paper assembly results. This prevents unverified questions from bypassing the quality control process and directly entering the final test paper.

[0040] The transition conditions between states in the finite state machine are as follows: After the task receiving state completes the parsing of the task request, it jumps to the subject knowledge injection state; after the subject knowledge injection state generates the prompt words, it jumps to the test question generation state; after the test question generation state generates the first draft of the test question, it jumps sequentially to the image generation state, test question proofreading state, AI trial state, image and text plagiarism check state, and difficulty analysis state; after the difficulty analysis state passes, it jumps to the test paper assembly state; after the test paper assembly state obtains the test paper, it jumps to the scoring standard generation state; after the scoring standard generation state generates the matching scoring standard, it jumps to the completion state; when any verification state outputs a verification result that fails, it jumps back to the test question generation state, and the centralized task scheduling engine returns the reason for failure, the corresponding state identifier, and the modification instruction, and calls the large language model again to generate the first draft of the test question and execute the subsequent state verification.

[0041] Furthermore, the question-setting method generates a test paper after performing multi-objective optimization and test paper generation. At the same time, the scoring standard generation agent generates tiered scoring rules for the subjective questions in the test paper and outputs the corresponding scoring standards for the test paper.

[0042] Multi-objective optimized test paper generation addresses the issue of individual questions being of acceptable quality but resulting in question combinations that suffer from repetitive knowledge points, discontinuous difficulty levels, or unsuitable question type ratios. The test paper generation agent uses test point coverage, target difficulty distribution, number of question types, total score, and repetition constraints as objectives or limitations to select questions from a candidate question pool. The scoring criteria generation agent then breaks down the scoring points, weights, step-by-step scoring, and acceptable wording based on the standard answers for subjective questions and the question's intent. Finally, the test paper, answer key, and scoring criteria are simultaneously output, creating a complete deliverable for actual exam review and use.

[0043] To achieve the second objective mentioned above, a second aspect of this invention discloses a multi-agent-assisted question-setting system for educational examinations, used to execute the aforementioned multi-agent-assisted question-setting method for educational examinations, the question-setting system comprising:

[0044] The question-setting task request receiving module is used to receive question-setting task requests and parse the question-setting task requests to obtain subject information, question type information, test point scope and target difficulty.

[0045] The question prompt word construction module constructs question prompt words carrying subject knowledge based on subject information and the scope of test points. It employs a hierarchical structured prompt word template, vectorizing the test point scope to obtain a query vector. This query vector is then used to perform semantic similarity retrieval on the subject knowledge vector database, identifying the top k subject knowledge fragments most semantically similar to the test point scope. These subject knowledge fragments are then injected into the hierarchical structured prompt word template to obtain question prompt words carrying subject knowledge. The hierarchical structured prompt word template is a four-layer prompt word structure from top to bottom, comprising a basic role layer, a subject knowledge layer, a question constraint layer, and an output format layer. The subject knowledge vector database is obtained by vectorizing and storing multi-subject knowledge nodes and their corresponding knowledge texts in a hierarchical knowledge graph, constructed based on curriculum standards and examination syllabi. Each subject knowledge fragment corresponds to a multi-subject knowledge node and represents the concepts, rules, formulas, examples, or relationships of that subject knowledge node.

[0046] The test draft generation module generates test drafts based on question type information and prompt words. The process is as follows: Using question type information and test point scope as query conditions, semantic similarity is performed on the sample question vector database to retrieve several sample questions of the same question type and test point. The question-setting ideas and structural paradigms of the sample questions are injected into the question-setting constraint layer of the prompt words as a reference paradigm. The large language model is then called to generate the test draft. The sample question vector database is a vector database obtained by vectorizing and encoding the stems, options, answers, explanations, and question type tags of past exam questions and standard sample questions. The sample questions are labeled question data used to provide question-setting ideas, structural paradigms, and semantic duplication benchmarks.

[0047] The verification result generation module, through a centralized task scheduling engine using a finite state machine, sequentially schedules the image matching agent, proofreading agent, trial agent, plagiarism detection agent, and difficulty analysis agent to perform image generation status verification, question proofreading status verification, AI trial status verification, image and text plagiarism detection status verification, and difficulty analysis status verification, and generates verification results for each. The centralized task scheduling engine is a task orchestration and execution component that models the question verification task as a finite state machine. This finite state machine includes sequentially executed image generation status verification, question proofreading status verification, AI trial status verification, image and text plagiarism detection status verification, and difficulty analysis status verification, with preset dependencies and jump conditions configured between each status verification.

[0048] The verification result judgment module determines whether the draft test question has passed all five verifications based on the verification results of the five intelligent agents: if the verification result output by any functional intelligent agent is "fail", the centralized task scheduling engine will send back the reason for failure and the modification instructions, regenerate the draft test question and iterate until all verifications are passed.

[0049] The test paper generation module generates test papers and scoring criteria from the initial drafts of test questions that have passed all verifications: the test paper generation agent performs multi-objective optimization from the candidate test question pool according to the test point range and the target difficulty to obtain the test paper.

[0050] To achieve the third objective mentioned above, a third aspect of the present invention discloses an electronic device, including a processor and a memory for storing a processor-executable program, wherein when the processor executes the program stored in the memory, it implements the aforementioned multi-agent-assisted question-setting method for educational examinations.

[0051] To achieve the fourth objective mentioned above, the fourth aspect of this invention discloses a storage medium storing a program, characterized in that, when the program is executed by a processor, it implements the aforementioned multi-agent-assisted question-setting method for educational examinations.

[0052] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0053] (1) This invention adopts a subject knowledge injection method that combines hierarchical structured prompt word templates with subject knowledge vector database. It dynamically obtains the subject knowledge fragments most relevant to the current test point through vectorized semantic retrieval and injects them into the prompt word context, so that the same question framework can adapt to the knowledge system and question requirements of different subjects, thus solving the problem of insufficient subject adaptability in the existing single model scheme.

[0054] (2) The present invention adopts a sample question vector database recall mechanism. By storing past exam questions and standard sample questions in vector form and recalling them by semantic similarity according to question type and test point, the question-setting ideas and structural paradigms of the recalled sample questions are used as references to inject into the generation process, so that the test question generation can follow the normative requirements of each question type, and solves the problem of weak cross-question type generalization ability of the existing scheme.

[0055] (3) The present invention adopts a centralized task scheduling engine to arrange five verification links in the form of a finite state machine: image generation, test question proofreading, AI trial, image and text plagiarism detection and difficulty analysis. When the verification fails, the modification instruction is sent back and the state machine is rolled back to regenerate the test question, forming an end-to-end closed-loop iteration mechanism. This enables tools such as image generation, proofreading, and plagiarism detection to be deeply embedded in the test question business node and achieve collaborative linkage, solving the problem of fragmented toolchain and lack of closed-loop feedback in the existing solution.

[0056] (4) Due to the adoption of offline private deployment, local loading of large language models and national cryptographic algorithms for encrypted storage and transmission, the entire question-setting process runs in a physically isolated environment, which meets the rigid requirements of educational examination venues in terms of data confidentiality and information technology innovation compliance. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is the overall flowchart of the multi-agent-assisted question-setting method for educational examinations disclosed in this invention;

[0059] Figure 2 This is a schematic diagram of the centralized task scheduling engine orchestrating the states of various functional agents in this invention.

[0060] Figure 3 This is a schematic diagram of the examination paper generated in Embodiment 1 of the present invention;

[0061] Figure 4 This is a structural block diagram of the multi-agent-assisted question-setting system for educational examinations disclosed in this invention;

[0062] Figure 5 This is a structural block diagram of the electronic device in an embodiment of the present invention. Detailed Implementation

[0063] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present application without creative effort are within the scope of protection of the present application.

[0064] The question-setting method is deployed offline in a private environment within the educational examination venue. In the embodiments described below, the large language models called in steps S2 to S6 are all locally loaded large language models. All question-setting data and operation logs generated during the execution of the question-setting method of this invention are stored in the local storage medium within the venue, without connecting to external networks or calling cloud interfaces. The question-setting materials, test questions, and operation records are encrypted and stored and transmitted using national cryptographic algorithms. Access control is implemented in a hierarchical manner for subject secretaries, question-setting experts, and question-reviewing experts according to their roles.

[0065] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application can be combined with other embodiments.

[0066] Example 1

[0067] This embodiment 1 discloses a complete implementation method for generating a thematic test paper on modern Chinese text reading.

[0068] This embodiment is executed in the offline environment within the examination venue. The question-setting method adopts a three-layer architecture: "base layer—intelligent agent layer—business application layer." The base layer includes a local digital resource library, a local question-setting model inference framework, and a locally loaded large language model; the intelligent agent layer includes subject-specific intelligent agents, question-type-specific intelligent agents, and functional tool intelligent agents; the business application layer provides interfaces for question initiation, task monitoring, quality verification, test paper generation, and result export. All question-setting materials, model inputs and outputs, and status logs are stored in the examination venue's internal storage medium.

[0069] Step S1: Receive and parse the task assignment request. The subject secretary submits the following tasks on the task assignment interface. The task scheduling engine generates a task identifier and writes each field into the task object, as shown in Table 1:

[0070] Table 1. Values ​​of each field

[0071]

[0072] Step S2 involves constructing prompts carrying subject-specific knowledge. The subject-specific AI agent encodes "the role of character emotional changes and detailed descriptions" into a query vector, retrieving three knowledge segments from the knowledge vector database: character emotional clue identification, the expressive role of detailed descriptions, and the design of distractors for reading comprehension multiple-choice questions. The similarity of these three segments in the existing parameterized records is 0.88, 0.83, and 0.79, respectively, all higher than 0.75. The system writes these segments into a four-layer prompt template: the basic role layer is set as "Eighth-Grade Chinese Language Question-Setting Expert"; the subject knowledge layer contains the aforementioned knowledge segments; the question constraint layer specifies the material length, question type, test points, difficulty, and originality requirements; and the output format layer specifies the output of the question stem, options, answer, explanation, difficulty label, and scoring points.

[0073] Step S3: Retrieve sample questions and generate a draft test. The question-type specialized agent retrieves sample questions with the same test points by combining the query conditions of question type and test point. It only extracts the structural paradigm of "reading material - emotional clue question - detail function question - subjective inquiry question", without copying the characters, context, and text. Based on this, the large language model generates the original reading material "The Lights of the Old Platform" and 6 candidate test questions, and outputs the standard answers, explanations, and test point tags for each.

[0074] The reading material generated in Example 1 is as follows:

[0075] The lights of the old platform

[0076] When I was little, I didn't like going to the old train station where my father worked. There was no waiting hall, only a row of wooden benches, whitened by the wind. Every evening, my father would carry his toolbox and wipe the dim yellow light on the east side of the platform. I always thought he was too slow and would often urge him to go home across the tracks.

[0077] Years later, I returned from out of town. The old platform was no longer in use. It was raining heavily, and my father still habitually walked to the lamppost and wiped the water droplets off the lampshade with his sleeve. He said that in the past, students from the neighborhood would take the last bus, and on rainy days they couldn't see the steps clearly. The brighter the light, the more stable their steps would be. I noticed for the first time that there was an old scar on the web of his right hand, left from repairing the lamp many years ago.

[0078] On the day the station was dismantled, my father handed over the keys, but his hand lingered on the lamppost for a long time. When I got home, I found a thin notebook in his old toolbox; the last page simply read: "On rainy days, light the east-side lamp first." As dusk deepened outside the window, I suddenly remembered those evenings waiting for my father, and I understood why he always came home late.

[0079] Step S4: The centralized task scheduling engine executes all states according to the finite state machine. The state execution record of this task is shown in Table 2 below:

[0080] Table 2. Task Status Runtime Record Table

[0081]

[0082] Step S5: Handling the failed verification branch. To illustrate the rollback mechanism of the finite state machine, the branch execution is performed using the policy-related multiple-choice question review records recorded in the R&D materials: In the initial draft, option A wrote the first National Ecological Day as 2022, and option D described carbon emission trading as "operating only in the pilot phase." The question checking agent outputs "failed," and the failure status is marked as "question checking status." The reason for the failure is that the time node and policy implementation status are incorrectly described. The modification instruction is "verify the National Ecological Day time node and rewrite the carbon emission trading implementation status." The scheduling engine jumps from the question checking status back to the question generation status, carrying the above status identifier, reason, and instruction to generate a new version; the new version starts executing subsequent verifications again from the image generation status. This branch proves that the failure result does not stop at the prompt message, but can drive the question version update and the re-execution of subsequent states.

[0083] Step S6: Perform multi-objective test paper generation and generate scoring criteria. The test paper generation agent selects 4 questions from the candidate question pool based on the constraints of 2 multiple-choice questions, 2 subjective questions, a total score of 20 points, and coverage of all three test points. The final output is as follows: Figure 3 The exam paper shown.

[0084] Example 2

[0085] This embodiment discloses an implementation method under different subjects, question types, and search parameters.

[0086] This embodiment uses the same task reception, knowledge injection, question generation, status verification, failure rollback, test paper assembly, and scoring standard generation process as Embodiment 1, and the same content will not be repeated. The difference is that the question task is set as a high school mathematics fill-in-the-blank question, the test point is "function monotonicity and derivative determination", and the target difficulty is high; the number of knowledge fragments retrieved is adjusted from the first 3 to the first 5, and the similarity threshold is adjusted from 0.75 to 0.70, so that relevant knowledge such as parameter discussion, domain constraints, and error-prone conditions can enter the context of the prompt words.

[0087] In an exemplary run, the test question generation state first outputs: "If the function f(x)=x³-3x²+ax+1 is monotonically increasing over the real number domain, then the value range of the real number a is ______." The AI trial agent calculates f′(x)=3x²-6x+a=3(x-1)²+a-3, and obtains the answer a≥3; the difficulty analysis state determines that this question only requires judging the minimum value of the derivative function, outputs the difficulty level "medium", which is inconsistent with the target difficulty level "high", and thus outputs "not passed". The scheduling engine returns the difficulty analysis state identifier, the reason of being too easy, and the modification instruction of "adding parameter discussion and constraint of extreme points within the interval", and jumps back to the test question generation state. The regenerated test question is: "If the function f(x)=x³-3x²+ax+1 has exactly two extreme points within the interval (0,2), then the value range of the real number a is ______." The AI trial agent sets f′(x)=0, and obtains two roots x=1±√(1-a / 3); the condition for both two distinct roots to lie within (0,2) is 0<a<3, so the standard answer is (0,3). After re-executing proofreading, trial solving, duplicate checking and difficulty analysis, the maximum semantic similarity is lower than 0.75 and the difficulty reaches the target level, then the new version of the test question, its answer, analysis and test point label are written into the candidate test question pool.

[0088] The two sets of parameterized operation results recorded in the existing manuscript are shown in Table 3 below. The first set includes 20 medium-difficulty multiple-choice tasks for Chinese, of which 15 tasks pass at the first attempt and 5 tasks pass after one rollback; the second set includes 20 high-difficulty fill-in-the-blank tasks for mathematics, of which 13 tasks pass at the first attempt, 6 tasks pass after one rollback, and 1 task passes after two rollbacks. The average number of rollbacks calculated according to "total number of rollbacks ÷ number of tasks" is 0.25 times and 0.40 times respectively.

[0089] Table 3. Operation results of two sets of parameters

[0090]

[0091] The above two sets of results are used to illustrate that the same finite state machine process can adapt to different subjects, question types, difficulty levels and retrieval parameters, and can perform one or multiple rollbacks on unpassed tasks. Since the subject, question type and target difficulty of the two sets of tasks change simultaneously, the results are not used alone to prove the effect difference caused by a certain retrieval parameter; if it is necessary to evaluate the independent influence of the number of top-k knowledge fragments or the similarity threshold, a supplementary control test should be performed by changing only one parameter under the same task set, the same model version and the same operation environment.

[0092] Embodiment 3

[0093] Reference Figure 4This embodiment provides a multi-agent-assisted question-setting system for educational examinations. The system includes, in sequence, a question-setting task request receiving module 401, a question prompt word construction module 402, a test draft generation module 403, a verification result generation module 404, a verification result judgment module 405, and a test paper assembly and generation module 406.

[0094] The question-setting task request receiving module 401 is used to receive question-setting task requests and parse and obtain subject information, question type information, test point scope and target difficulty from the question-setting task requests;

[0095] The prompt word construction module 402 constructs prompt words carrying subject knowledge based on subject information and the scope of test points. It employs a hierarchical structured prompt word template, vectorizing the scope of test points to obtain a query vector. This query vector is then used to perform semantic similarity retrieval on the subject knowledge vector database, identifying the top k subject knowledge fragments most semantically similar to the scope of test points. These subject knowledge fragments are then injected into the hierarchical structured prompt word template to obtain prompt words carrying subject knowledge. The hierarchical structured prompt word template is a four-layer structure from top to bottom, comprising a basic role layer, a subject knowledge layer, a prompt constraint layer, and an output format layer. The subject knowledge vector database is obtained by vectorizing and storing multi-subject knowledge nodes and their corresponding knowledge texts in a hierarchical knowledge graph, which is constructed based on curriculum standards and examination syllabi. Each subject knowledge fragment corresponds to a multi-subject knowledge node and represents the concept, rules, formulas, examples, or relationships of that subject knowledge node.

[0096] The test draft generation module 403 generates a test draft based on question type information and prompt words. The process is as follows: using question type information and test point range as query conditions, semantic similarity is performed on the sample question vector database to recall several sample questions of the same question type and test point. The question-setting ideas and structural paradigms of the sample questions are injected into the question-setting constraint layer of the prompt words as a reference paradigm. The large language model is then called to generate the test draft. The sample question vector database is a vector database obtained by vectorizing and encoding the stems, options, answers, explanations, and question type tags of past exam questions and standard sample questions. The sample questions are labeled question data used to provide question-setting ideas, structural paradigms, and semantic duplication benchmarks.

[0097] The verification result generation module 404, through a centralized task scheduling engine, sequentially schedules the image matching agent, proofreading agent, trial agent, plagiarism detection agent, and difficulty analysis agent in a finite state machine manner to perform image generation status verification, question proofreading status verification, AI trial status verification, image and text plagiarism detection status verification, and difficulty analysis status verification, and generates verification results for each. The centralized task scheduling engine is a task orchestration and execution component that models the question verification task as a finite state machine. The finite state machine includes sequentially executed image generation status verification, question proofreading status verification, AI trial status verification, image and text plagiarism detection status verification, and difficulty analysis status verification, with preset dependencies and jump conditions configured between each status verification.

[0098] The verification result judgment module 405 determines whether the draft test question has passed all five verifications based on the verification results of the five intelligent agents: if the verification result output by any functional intelligent agent is a failure, the centralized task scheduling engine will send back the reason for failure and the modification instruction, regenerate the draft test question and iterate until all verifications are passed.

[0099] The test paper generation module 406 generates test papers and scoring criteria from the first drafts of test questions that have passed all verifications: the test paper generation agent performs multi-objective optimization from the candidate test question pool according to the test point range and the target difficulty to obtain the test paper.

[0100] The content of the above method embodiments is applicable to this system embodiment, and the specific functions implemented in this system embodiment are the same as those in the above method embodiments.

[0101] Example 4

[0102] This embodiment provides an electronic device, which can be a computer, such as... Figure 5 As shown, the processor 502, memory, input device 503, display 504, and network interface 505 are connected via system bus 501. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium 506 and an internal memory 507. The non-volatile storage medium 506 stores the operating system, computer programs, and database. The internal memory 507 provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. When the processor 502 executes the computer programs stored in the memory, it implements the multi-agent-assisted question-setting method for educational examinations proposed in Embodiment 1. The multi-agent-assisted question-setting method for educational examinations includes steps S1 to S6 disclosed in Embodiment 1.

[0103] Example 5

[0104] This embodiment provides a storage medium, which is a computer-readable storage medium, storing a computer program. When the computer program is executed by a processor, it implements the multi-agent-assisted question-setting method for educational examinations proposed in Embodiment 1 above. The multi-agent-assisted question-setting method for educational examinations includes steps S1 to S6 disclosed in Embodiment 1.

[0105] The storage medium described in this embodiment can be a disk, optical disk, computer memory, random access memory (RAM), USB flash drive, portable hard drive, etc.

[0106] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0107] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A multi-agent-assisted question-setting method for educational examinations, characterized in that, The proposition method includes the following steps: S1. Receive a question-setting task request, and parse the question-setting task request to obtain subject information, question type information, test point scope and target difficulty; S2. Based on subject information and the scope of test points, construct prompt words carrying subject knowledge: Using a hierarchical structured prompt word template, the scope of test points is vectorized to obtain a query vector. The query vector is then used to perform semantic similarity retrieval on the subject knowledge vector database to obtain the top k subject knowledge fragments that are semantically most similar to the scope of test points. These subject knowledge fragments are then injected into the hierarchical structured prompt word template to obtain prompt words carrying subject knowledge. The hierarchical structured prompt word template is a four-layer prompt word structure from top to bottom, including a basic role layer, a subject knowledge layer, a proposition constraint layer, and an output format layer. The subject knowledge vector database is obtained by vectorizing and storing multi-subject knowledge nodes and their corresponding knowledge texts in a hierarchical knowledge graph, which is constructed based on curriculum standards and examination syllabi. The subject knowledge fragments correspond to multi-subject knowledge nodes and are used to represent the concepts, rules, formulas, examples, or relationships of the subject knowledge node. S3. Generate a draft of the test questions based on the question type information and prompt words. The process is as follows: Using the question type information and the scope of test points as query conditions, perform semantic similarity retrieval on the sample question vector database to recall several sample questions of the same question type and test points. Inject the question-setting ideas and structural paradigms of the sample questions into the question-setting constraint layer of the prompt words as a reference paradigm. Call the large language model to generate the draft of the test questions. The sample question vector database is a vector database obtained by vectorizing and encoding the stems, options, answers, explanations, and question type tags of past exam questions and standard sample questions. The sample questions are labeled question data used to provide question-setting ideas, structural paradigms, and semantic duplication benchmarks. S4. The centralized task scheduling engine sequentially schedules the image matching agent, proofreading agent, trial agent, plagiarism detection agent, and difficulty analysis agent in a finite state machine manner to perform image generation status verification, question proofreading status verification, AI trial status verification, image and text plagiarism detection status verification, and difficulty analysis status verification, and generates verification results for each. The centralized task scheduling engine is a task orchestration and execution component that models the question verification task as a finite state machine. The finite state machine includes the sequentially executed image generation status verification, question proofreading status verification, AI trial status verification, image and text plagiarism detection status verification, and difficulty analysis status verification. Preset dependencies and jump conditions are configured between each status verification. S5. Determine whether the draft test question has passed all 5 checks based on the verification results of the 5 agents: If the verification result output by any functional agent is "fail", the centralized task scheduling engine will send back the reason for failure and the modification instructions, regenerate the draft test question and iterate until all checks are passed. S6. Generate test papers and scoring criteria from the first draft of the test questions that have passed all verifications: The test paper generation agent performs multi-objective optimization from the candidate test question pool according to the test point range and the target difficulty to obtain the test paper.

2. The multi-agent-assisted question-setting method for educational examinations according to claim 1, characterized in that, The candidate question pool consists of existing questions and draft questions from the original question bank; in step S3, the recalled sample questions also serve as the benchmark data for semantic similarity comparison by the image and text plagiarism detection agent in step S4.

3. The multi-agent-assisted question-setting method for educational examinations according to claim 1, characterized in that, The centralized task scheduling engine calls the corresponding functional agent to perform the following operations under each state verification: During the image generation status verification, the image generation agent generates corresponding imagery for the test questions based on the imagery requirements described in the initial draft of the test questions, and outputs test questions with images. During the test question proofreading status verification, the test question proofreading agent performs multi-dimensional proofreading of the initial test question in terms of political, scientific and fairness aspects, and outputs the proofreading results; In the AI ​​trial state verification, the AI ​​trial agent simulates the examinee to answer the draft of the test questions, verifies the rationality and correctness of the answers to the draft of the test questions, and outputs the trial results; In the image and text plagiarism check status verification, the image and text plagiarism check agent performs semantic-level similarity comparison between the draft test question and the sample questions in the sample question vector database and the existing test questions in the original question bank, and outputs the plagiarism check results; During the difficulty analysis state verification, the difficulty analysis agent predicts the difficulty coefficient of the first draft of the test questions based on the trial results and the target difficulty, and outputs the difficulty analysis results.

4. The multi-agent-assisted question-setting method for educational examinations according to claim 1, characterized in that, The preset dependencies and transition conditions configured between each state check in the finite state machine are as follows: The image generation status verification depends on the initial draft of the test questions in the test question generation status; the test question proofreading status verification depends on the test questions with images output by the image generation status verification; the AI ​​trial status verification depends on the test questions after the test question proofreading status verification has passed; the image and text plagiarism check status verification depends on the test questions after the AI ​​trial status verification has passed, as well as the sample question vector database and the original question bank; the difficulty analysis status verification depends on the trial results output by the AI ​​trial status verification and the target difficulty. In the image generation, question proofreading, AI trial, image and text plagiarism checking, and difficulty analysis states, when the functional agent corresponding to the current state has completed its execution and the output result meets the preset passing conditions, the finite state machine jumps to the next state verification. When any state verification outputs a failing result, the finite state machine jumps back to the question generation state. The centralized task scheduling engine transmits the reason for failure, the corresponding state identifier, and the modification instruction to the agent corresponding to the question generation state, regenerates the initial draft of the question, and executes the subsequent verification states sequentially starting from the image generation state verification.

5. The multi-agent-assisted question-setting method for educational examinations according to claim 1, characterized in that, The four levels of the hierarchical structured prompt template each perform the following functions: The basic role layer defines the subject identity and proposition role of each intelligent agent; wherein, the subjects include Chinese, mathematics, English, physics, chemistry, biology, politics, history and geography; The subject knowledge layer carries subject knowledge fragments injected by semantic retrieval; The question constraint layer sets the question type specifications, test point requirements, and difficulty constraints; The output format layer specifies the output structure and field format of the test questions.

6. The multi-agent-assisted question-setting method for educational examinations according to claim 1, characterized in that, The semantic similarity retrieval process in step S2 is as follows: The scope of test points is vectorized to obtain the query vector; Calculate the cosine similarity between the query vector and the knowledge node vectors in the subject knowledge vector database; Sort the knowledge nodes in descending order of similarity, take the subject knowledge fragments corresponding to the top k knowledge nodes, and concatenate them to the subject knowledge layer of the hierarchical structured prompt word template.

7. The multi-agent-assisted question-setting method for educational examinations according to claim 1, characterized in that, The finite state machine modeled by the centralized task scheduling engine includes the following states: task receiving state, subject knowledge injection state, question generation state, image generation state, question proofreading state, AI trial state, image and text plagiarism check state, difficulty analysis state, test paper assembly state, scoring standard generation state, and completion state. Specifically, the task receiving state is used to parse the question-setting task request to obtain subject information, question type information, test point scope, and target difficulty; the subject knowledge injection state is used to retrieve the subject knowledge vector database to obtain subject knowledge fragments and generate question prompts; and the question generation state is used to generate a draft question based on the question prompts and recalled sample questions. The following states are used for different tasks: Image generation (to generate images for the test questions based on the initial draft); Question proofreading (to output proofreading results regarding political correctness, scientific rigor, and fairness); AI trial run (to output trial run results regarding the reasonableness and correctness of answers); Image and text plagiarism check (to output similarity comparison results between the initial draft and benchmark data); Difficulty analysis (to output difficulty analysis results based on trial run results and target difficulty); Test paper assembly (to obtain the test paper through multi-objective optimization from the candidate question pool); Scoring criteria generation (to generate tiered scoring rules for subjective questions in the test paper); and Completion (to output the test paper and its accompanying scoring criteria). The transition conditions between states in the finite state machine are as follows: After the task receiving state completes the parsing of the task request, it jumps to the subject knowledge injection state; after the subject knowledge injection state generates the prompt words, it jumps to the test question generation state; after the test question generation state generates the first draft of the test question, it jumps sequentially to the image generation state, test question proofreading state, AI trial state, image and text plagiarism check state, and difficulty analysis state; after the difficulty analysis state passes, it jumps to the test paper assembly state; after the test paper assembly state obtains the test paper, it jumps to the scoring standard generation state; after the scoring standard generation state generates the matching scoring standard, it jumps to the completion state; when any verification state outputs a verification result that fails, it jumps back to the test question generation state, and the centralized task scheduling engine returns the reason for failure, the corresponding state identifier, and the modification instruction, and calls the large language model again to generate the first draft of the test question and execute the subsequent state verification.

8. The multi-agent-assisted question-setting method for educational examinations according to claim 1, characterized in that, The question-setting method generates a test paper after performing multi-objective optimization and test paper generation. At the same time, the scoring standard generation agent generates tiered scoring rules for the subjective questions in the test paper and outputs the corresponding scoring standards for the test paper.

9. A multi-agent-assisted question-setting method for educational examinations, used to execute the multi-agent-assisted question-setting method for educational examinations as described in any one of claims 1 to 8, characterized in that, The propositional system includes: The question-setting task request receiving module is used to receive question-setting task requests and parse the question-setting task requests to obtain subject information, question type information, test point scope and target difficulty. The question prompt word construction module constructs question prompt words carrying subject knowledge based on subject information and the scope of test points. It employs a hierarchical structured prompt word template, vectorizing the test point scope to obtain a query vector. This query vector is then used to perform semantic similarity retrieval on the subject knowledge vector database, identifying the top k subject knowledge fragments most semantically similar to the test point scope. These subject knowledge fragments are then injected into the hierarchical structured prompt word template to obtain question prompt words carrying subject knowledge. The hierarchical structured prompt word template is a four-layer prompt word structure from top to bottom, comprising a basic role layer, a subject knowledge layer, a question constraint layer, and an output format layer. The subject knowledge vector database is obtained by vectorizing and storing multi-subject knowledge nodes and their corresponding knowledge texts in a hierarchical knowledge graph, constructed based on curriculum standards and examination syllabi. Each subject knowledge fragment corresponds to a multi-subject knowledge node and represents the concepts, rules, formulas, examples, or relationships of that subject knowledge node. The test draft generation module generates test drafts based on question type information and prompt words. The process is as follows: Using question type information and test point scope as query conditions, semantic similarity is performed on the sample question vector database to retrieve several sample questions of the same question type and test point. The question-setting ideas and structural paradigms of the sample questions are injected into the question-setting constraint layer of the prompt words as a reference paradigm. The large language model is then called to generate the test draft. The sample question vector database is a vector database obtained by vectorizing and encoding the stems, options, answers, explanations, and question type tags of past exam questions and standard sample questions. The sample questions are labeled question data used to provide question-setting ideas, structural paradigms, and semantic duplication benchmarks. The verification result generation module, through a centralized task scheduling engine using a finite state machine, sequentially schedules the image matching agent, proofreading agent, trial agent, plagiarism detection agent, and difficulty analysis agent to perform image generation status verification, question proofreading status verification, AI trial status verification, image and text plagiarism detection status verification, and difficulty analysis status verification, and generates verification results for each. The centralized task scheduling engine is a task orchestration and execution component that models the question verification task as a finite state machine. This finite state machine includes sequentially executed image generation status verification, question proofreading status verification, AI trial status verification, image and text plagiarism detection status verification, and difficulty analysis status verification, with preset dependencies and jump conditions configured between each status verification. The verification result judgment module determines whether the draft test question has passed all five verifications based on the verification results of the five intelligent agents: if the verification result output by any functional intelligent agent is "fail", the centralized task scheduling engine will send back the reason for failure and the modification instructions, regenerate the draft test question and iterate until all verifications are passed. The test paper generation module generates test papers and scoring criteria from the initial drafts of test questions that have passed all verifications: the test paper generation agent performs multi-objective optimization from the candidate test question pool according to the test point range and the target difficulty to obtain the test paper.

10. An electronic device comprising a processor and a memory for storing a processor-executable program, characterized in that, When the processor executes the program stored in the memory, it implements the multi-agent-assisted question-setting method for educational examinations as described in any one of claims 1 to 8.