Class learning condition analysis system and method based on AI large model

Abstract

The invention provides a class learning condition analysis system and method based on an AI large model, and the system comprises a multi-person single-question analysis subsystem which is used for carrying out the error cause analysis and the number of persons through employing different analysis strategies according to different question types, and carrying out the error cause analysis through employing different preset strategies according to the question setting modes of different subjects. Obtaining a first analysis report; the multi-person multi-question knowledge point analysis subsystem is used for aggregating question information for different knowledge points and obtaining a second analysis report based on a multi-modal LLM model; and the multi-person and multi-question ability analysis subsystem is used for aggregating question information for different abilities and obtaining a third analysis report based on a multi-modal LLM model. The first analysis report, the second analysis report and the third analysis report output under the multi-person single question, the multi-person multi-question knowledge points and the multi-person single question ability are integrated, and a final learning condition analysis report is given.

Description

Technical Field

[0001] This invention belongs to the technical fields of smart education and AI teacher test paper grading systems, specifically involving a class learning analysis system and method based on an AI large model. Background Technology

[0002] Existing learning analysis systems typically utilize big data and artificial intelligence technologies to collect students' learning behavior data in real time from multiple channels such as learning management systems (LMS), online course platforms, and smart classroom devices. They then employ machine learning and data mining algorithms to analyze the data, presenting the results in charts and graphs to provide educators with insights into students' learning status and offering personalized learning resource recommendations based on the analysis results.

[0003] However, existing learning analysis systems mainly have the following five problems: 1. Limited analytical dimensions: Most systems only provide simple score statistics, lacking in-depth analysis of knowledge points and ability dimensions; 2. Disconnect between individual and group analysis: It is difficult to effectively aggregate individual error analysis into class-level learning insights; 3. Weak interdisciplinary analytical ability: unable to establish a unified learning assessment system across multiple disciplines such as mathematics, physics, chemistry, and biology; 4. Poor real-time performance: The analysis results are delayed and difficult to be fed back into the teaching process in a timely manner; 5. Lack of predictive analysis: It is impossible to predict the overall learning trend and risk points of the class based on the existing learning situation.

[0004] Therefore, there is an urgent need for a new multidisciplinary class learning analysis system to solve the above problems. Summary of the Invention

[0005] This invention provides a class learning analysis system and method based on an AI large-scale model, aiming to solve the following problems: 1. Enable multi-student single-question analysis, analyzing the differences in problem-solving approaches and error patterns among different students in a class for specific questions; 2. Construct a knowledge point map for multiple students and multiple questions, and aggregate and analyze the overall mastery of specific knowledge points across questions; 3. Establish an ability dimension assessment system to evaluate the overall level of the class from the dimensions of understanding, application, analysis, and synthesis; 4. Support cross-disciplinary unified analysis, establishing a comparable analytical indicator system among the four subjects of mathematics, physics, chemistry, and biology.

[0006] To achieve the above objectives, the present invention provides the following solution: A class learning analysis system based on an AI large model, the system comprising: a multi-person single-question analysis subsystem, a multi-person multi-question knowledge point analysis subsystem, and a multi-person multi-question ability analysis subsystem; The multi-person single-question analysis subsystem is used to perform error analysis and number matching based on different question types and different pre-set strategies based on the question format of different subjects, and to obtain the first analysis report. The multi-user, multi-question knowledge point analysis subsystem is used to aggregate question information for different knowledge points and obtain a second analysis report based on a multimodal LLM model. The multi-person, multi-question ability analysis subsystem is used to aggregate question information for different abilities and obtain a third analysis report based on a multimodal LLM model.

[0007] Preferably, the multi-person single-question analysis subsystem employs different analysis strategies for error cause analysis and number of participants based on different question types. Simultaneously, it uses different preset strategies for error cause analysis based on the question format of different subjects. The process of obtaining the first analysis report includes: The multi-person single-question analysis subsystem determines whether all students have answered the input question correctly. If all students have answered correctly, it directly outputs the error reason; otherwise, it determines the type of the input question. For multiple-choice and fill-in-the-blank questions, preset error reasons are generated directly based on the input question and stored in the MongoDB database. For open-ended questions, a multimodal LLM model is used to perform error reason analysis and generation in real time to obtain the error reason analysis and number of respondents matching results. Based on the question-setting methods of different subjects, different pre-set strategies are used for analysis and processing to obtain subject analysis results; Based on the analysis of the reasons for errors in the questions, the results of matching the number of participants, and the results of subject analysis, the first analysis report was obtained.

[0008] Preferably, the multi-user, multi-question knowledge point analysis subsystem aggregates question information for different knowledge points and obtains a second analysis report based on a multimodal LLM model, including the following process: Based on the test paper information, summarize the knowledge points contained in all questions, and aggregate the question information according to the knowledge points to obtain the knowledge point questions; Collect all student OCR answers for each knowledge point question, dynamically insert prompt information, and truncate the prompt information according to the token length; Input the truncated prompt information into the multimodal LLM model, request the multimodal LLM model to output the analysis results, and format the analysis results to obtain the second analysis report.

[0009] Preferably, the process by which the multi-person, multi-question ability analysis subsystem aggregates question information for different abilities and obtains a third analysis report based on a multimodal LLM model includes: Based on the test paper information, summarize the abilities contained in all questions, and aggregate the question information according to the abilities to obtain the ability questions; Collect all student OCR errors for each ability question, dynamically insert prompt information, and truncate the prompt information according to the token length; Input the truncated prompt information into the multimodal LLM model, request the multimodal LLM model to output the analysis results, and format the analysis results to obtain the third analysis report.

[0010] This invention also provides a method for analyzing class learning performance based on an AI large-scale model. The method is implemented using the aforementioned system and includes: Based on the multi-person single-question analysis subsystem, different analysis strategies are adopted for error analysis and number matching according to different question types. At the same time, different preset strategies are adopted for error analysis according to the question setting methods of different subjects to obtain the first analysis report. Based on the multi-person, multi-question knowledge point analysis subsystem, question information is aggregated for different knowledge points, and a second analysis report is obtained based on the multimodal LLM model; Based on the multi-person, multi-question ability analysis subsystem, question information is aggregated for different abilities, and a third analysis report is obtained based on the multimodal LLM model.

[0011] Preferably, based on a multi-person single-question analysis subsystem, different analysis strategies are adopted for error cause analysis and number matching according to different question types. Simultaneously, different preset strategies are used for error cause analysis according to the question format of different subjects. The method for obtaining the first analysis report includes: The multi-person single-question analysis subsystem determines whether all students have answered the input question correctly. If all students have answered correctly, it directly outputs the error reason; otherwise, it determines the type of the input question. For multiple-choice and fill-in-the-blank questions, preset error reasons are generated directly based on the input question and stored in the MongoDB database. For open-ended questions, a multimodal LLM model is used to perform error reason analysis and generation in real time to obtain the error reason analysis and number of respondents matching results. Based on the question-setting methods of different subjects, different pre-set strategies are used for analysis and processing to obtain subject analysis results; Based on the analysis of the reasons for errors in the questions, the results of matching the number of participants, and the results of subject analysis, the first analysis report was obtained.

[0012] Preferably, the method for obtaining a second analysis report based on a multi-user, multi-question knowledge point analysis subsystem, which aggregates question information for different knowledge points, and based on a multimodal LLM model, includes: Based on the test paper information, summarize the knowledge points contained in all questions, and aggregate the question information according to the knowledge points to obtain the knowledge point questions; Collect all student OCR answers for each knowledge point question, dynamically insert prompt information, and truncate the prompt information according to the token length; Input the truncated prompt information into the multimodal LLM model, request the multimodal LLM model to output the analysis results, and format the analysis results to obtain the second analysis report.

[0013] Preferably, the method for obtaining a third analysis report based on a multi-person, multi-question ability analysis subsystem, which aggregates question information for different abilities, and based on a multimodal LLM model, includes: Based on the test paper information, summarize the abilities contained in all questions, and aggregate the question information according to the abilities to obtain the ability questions; Collect all student OCR errors for each ability question, dynamically insert prompt information, and truncate the prompt information according to the token length; Input the truncated prompt information into the multimodal LLM model, request the multimodal LLM model to output the analysis results, and format the analysis results to obtain the third analysis report.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention employs different approaches for different subjects and question types, inputting the solution processes of all students who made mistakes into a multimodal LLM model. Through carefully designed prompt information, it analyzes common errors and key points of deduction among students. By tagging predicted questions with knowledge point labels and then aggregating different questions by knowledge point, it analyzes students' mastery of knowledge points from a knowledge point perspective. Furthermore, by prompting the multimodal LLM model to tag questions with ability labels and then aggregating different questions by ability, it analyzes the overall ability radar of students in the class from an ability dimension. Since the question types differ slightly across subjects, different analysis logics are used for different subjects, achieving personalized processing of class learning analysis. Attached Figure Description

[0015] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of a class learning analysis system module based on an AI large model, according to an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the process of dynamically cropping prompt information according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the class learning situation analysis method based on AI large model according to an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] Example 1 like Figure 1 As shown, the present invention provides a class learning analysis system based on an AI large model, including: a multi-person single-question analysis subsystem, a multi-person multi-question knowledge point analysis subsystem, and a multi-person multi-question ability analysis subsystem; The multi-person single-question analysis subsystem is used to perform error analysis and number matching based on different question types and different pre-set strategies based on the question format of different subjects, and to obtain the first analysis report. The multi-user, multi-question knowledge point analysis subsystem is used to aggregate question information for different knowledge points and obtain a second analysis report based on a multimodal LLM model. The multi-person, multi-question ability analysis subsystem is used to aggregate question information for different abilities and obtain a third analysis report based on a multimodal LLM model.

[0020] The specific implementation process of this invention is as follows: The multi-user single-question analysis subsystem includes: The question type determination unit is used to automatically identify the question type based on the input question type field, which includes multiple choice questions, fill-in-the-blank questions, and problem-solving questions; The "All Correct" judgment unit is used to determine whether all students have answered the question correctly. If all students have answered correctly, the error reason will be output directly. Subject-specific adaptation units are used to employ corresponding pre-defined analysis strategies based on the question formats of different subjects.

[0021] The multi-person single-question analysis subsystem employs different analysis strategies for error analysis and participant matching based on different question types. Simultaneously, it uses different pre-set strategies for error analysis based on the question format of different subjects. The process of obtaining the first analysis report includes: The multi-person single-question analysis subsystem determines whether all students have answered the input question correctly. If all students have answered correctly, it directly outputs the error reason; otherwise, it determines the type of the input question. For multiple-choice and fill-in-the-blank questions, preset error reasons are generated directly based on the input question and stored in the MongoDB database. For open-ended questions, a multimodal LLM model is used to perform error reason analysis and generation in real time to obtain the error reason analysis and number of respondents matching results. Based on the question-setting methods of different subjects, different pre-set strategies are used for analysis and processing to obtain subject analysis results; Based on the analysis of the reasons for errors in the questions, the results of matching the number of participants, and the results of subject analysis, the first analysis report was obtained.

[0022] Furthermore, for multiple-person single-question analysis, the subsystem also includes the following for handling multiple-choice questions: The multiple-choice question analysis unit is used to analyze the reasons for errors in the three incorrect options other than the correct one in a multiple-choice question. The multiple-choice question analysis unit is used to analyze the reasons for errors in all four options of a multiple-choice question. It analyzes the reasons for missing correct options and the reasons for incorrect options. The error cause storage unit is used to store the analyzed error causes into the MongoDB database using the error options as the query key.

[0023] Furthermore, for the processing of fill-in-the-blank questions, the multi-user single-question analysis subsystem also includes: The error answer prediction unit is used to predict possible incorrect answers from students as preset error reasons using a multimodal LLM model; The "Catch-Up Error Reason" section is used to cover errors caused by insufficient understanding of the relevant knowledge points. The matching judgment unit is used to enable a fallback error cause when a student's OCR answer does not match the predicted incorrect answer.

[0024] Furthermore, for handling problem-solving questions, the multi-user single-question analysis subsystem also includes: The real-time analysis unit is used to analyze the causes of errors in students' problem-solving process in real time using a multimodal LLM model; The scoring correlation unit is used to determine the severity of errors based on students' scores: when the score is ≥90% of the full score, the focus is on detailed errors; when the score is 50% ≤ and <90%, the analysis focuses on errors in key steps and knowledge gaps; when the score is <50%, the systemic lack of basic concepts and problem-solving methods is identified. The student matching unit is used to assign students to the corresponding error category using a multimodal LLM model.

[0025] Furthermore, for problem-solving questions in different subjects, the multi-user single-question analysis subsystem also includes: The math problem-solving analysis unit is used for real-time analysis of student answer images input into a multimodal LLM model; The physics experiment question analysis unit is used to pre-determine the common error reasons for each blank and the summary error reasons at the question-wide level; The analysis unit for chemistry experiment questions uses the same processing logic as that for physics experiment questions. The biological materials question analysis unit is used to directly output common errors for each blank, without needing to summarize the error causes.

[0026] Furthermore, the multi-user single-question analysis subsystem also includes: The format repair unit is used to repair the format of the output of the multimodal LLM model using a trained LaTeX repair model, so as to ensure the format accuracy of the output results. The marginal error processing unit is used to process students who cannot be matched with any candidate error type, merge them into the marginal error category, and count the number of students with marginal errors.

[0027] Specifically, the multi-person single-question analysis subsystem adopts different analysis strategies according to different question types, including preset logic and real-time error statistics for multiple-choice and fill-in-the-blank questions, multimodal LLM model analysis for problem-solving questions, and matching and filtering strategies for student errors. At the same time, the question-setting methods for different subjects will be slightly different, and the forms of error analysis will also be slightly different.

[0028] 1) First, the multi-student single-question analysis subsystem determines whether all students have answered the question correctly. If all students have answered correctly, it directly outputs an empty error reason; otherwise, it determines the type of the input question (when constructing the input structured data, the multi-student single-question analysis subsystem includes a "question type" field, specifying the question type as choice, fill-in-the-blank, or subjective problem-solving; it directly extracts the question type field and performs subsequent logical operations based on these three types). For choice and fill-in-the-blank questions, since there are no specific details of the problem-solving process, students' error reasons are often relatively fixed. Therefore, it does not consider the specific answers of students, but directly generates preset error reasons based on the question information and stores them in the MongoDB database. For problem-solving questions, which contain various specific problem-solving steps, a multimodal LLM model (in this embodiment, the Doubao 1.6 Thinking multimodal model is used, calling the API interface) is used for real-time error reason analysis and generation. Then, for different subjects, the question examination methods will also have certain differences. The multi-student single-question analysis subsystem will use different preset strategies for analysis and processing. The main process of the entire multi-student single-question analysis subsystem is as follows: First, the options in multiple-choice questions are fixed, and the reasons for each incorrect option are roughly the same. Therefore, the reasons for students choosing each option incorrectly are likely to be consistent. Thus, the multimodal LLM model can directly provide the reasons for each incorrect option in multiple-choice questions based on the question stem, answer, explanation, and image hints. These reasons are used as preset reasons and stored in a MongoDB database using the incorrect options as queries. When analyzing a class, the preset reasons in the database are directly queried based on the students' answer options, and the number of students choosing each incorrect option is statistically analyzed. This provides the error distribution for subsequent learning analysis. For single-choice questions, the reasons for the three incorrect options other than the correct one need to be analyzed. For multiple-choice questions, the reasons for the errors of all four options need to be analyzed. The correct option is analyzed to see why it was missed, and the incorrect options are analyzed to see why they were chosen incorrectly.

[0029] Similarly, fill-in-the-blank questions do not have detailed solution processes, and the test points are generally relatively fixed. Students' incorrect answers tend to converge, meaning that most students' calculated numbers will lean towards a certain candidate range. Therefore, fill-in-the-blank questions can also adopt the same pre-set logic as multiple-choice questions. By using a multimodal LLM model to provide students with possible incorrect answers as pre-set error reasons, and then using insufficient knowledge of the question as a fallback error reason, when the student's OCR does not match the candidate incorrect answer inferred by the multimodal LLM model, a fallback pre-set error reason is given. Finally, the number of people with each type of error reason is counted to provide the error reason distribution for fill-in-the-blank questions for subsequent analysis.

[0030] For open-ended questions, which involve a detailed explanation of the student's solution process, the system needs to analyze each step of the student's answer and identify specific reasons for errors. The system will provide the multimodal LLM model with question information and images, question scores, knowledge points, and all students' incorrect answers. The system will also prompt the multimodal LLM model to analyze common error reasons among students in real time. The input prompt information is as follows: Task Description: "Analyze the common errors made by multiple students in a physics problem and provide corresponding teaching suggestions. Note that if the output involves mathematical formulas, ensure that each output is in "LaTeX" format, such as $x_{1}$, $x_{2}$ should be x_{1}, x_{2}, and $-\infty$ should be -\infty."

[0031] The specific steps are as follows: 1.1) Read and understand the given question, question type, question score, question answer, detailed explanation of the question, knowledge points that the question may involve, the main types of errors made by students, and each student's answer (including the student's score for this question).

[0032] 1.2) Compare the answers of all students with the answers to the question, and point out the common reasons for the errors of multiple students in a question (student answers may be blank or only some steps are written). The students' answers may be garbled due to OCR recognition errors. Pay attention to distinguishing between blank student answers and OCR recognition errors.

[0033] 1.3) Based on all the information provided, including the answers and detailed explanations, identify the common causes of errors for each major error type. Pay special attention to ensuring the given causes are concise and specific, avoiding vagueness. The error descriptions should be detailed and feasible, not abstract, to ensure practical reference value. Avoid repeating causes for the same question, especially semantic repetition, ensuring diversity and accuracy. Furthermore, prioritize logical errors related to knowledge points, followed by logical errors at the calculation level, i.e., knowledge point error type > calculation error type. Control the number of causes, ranking them according to the priority of knowledge-level errors > calculation-level errors.

[0034] 1.4) Provide an analysis report that details the common causes of students' errors and provides specific and feasible teaching suggestions for each error cause. The error type (type field) returned in the results is selected from the preset "Student Main Error Types" and is unique. The error cause description is based on the students' problem-solving approach and is expressed rigorously.

[0035] 1.5) Mathematical Formatting: All mathematical and physical formulas, symbols, and expressions appearing in the question content must be transcribed in LaTeX format and enclosed in a single dollar sign $.

[0036] 1.6) Analyze the reasons for student errors reasonably based on the question scores and student scores; for multiple-choice questions (which may include multiple selections), locate the options and analyze the common reasons for students' incorrect selections, multiple selections, and omissions. The analysis of incorrect options should correspond to the answers of the incorrect options, and the results of the analysis should be consistent with the students' incorrect selections; for problem-solving questions, if the student's score is high, there should be few errors, and if the score is low, there should be many errors; pay special attention to the question type, and provide reasonable error types according to the question format. For multiple-choice and fill-in-the-blank questions, do not provide error types such as "lack of detailed steps." In addition, multiple-choice questions may include single-choice and multiple-choice questions, so avoid simply providing meaningless error information such as "incorrect selection" or "omission of the correct option." Also, ensure that the description of the error reasons under all questions is relevant to the question content and the knowledge points being tested.

[0037] 1.7) Score Correlation Analysis: Specifically, the severity of errors is determined based on student scores. A score of ≥90% (out of 5): Focus on minor errors (such as calculation mistakes / omission of units, etc.). 50% ≤ Score < 90%: Analyze key steps for errors and knowledge gaps; Score <50%: Systematic lack of understanding of basic concepts and problem-solving methods.

[0038] 1.8) The output follows a preset return format.

[0039] 1.9) Preset return format requirement: Output results in JSON format.

[0040] Students' answers are extracted using an OCR model and used to construct input data. The system then considers the question scores and students' specific scores to determine the error reasons. If most students lose few points, the errors may be concentrated on some details, and the multimodal LLM model will provide relatively few error reasons. Conversely, if most students do not answer or lose many points, the multimodal LLM model will provide more detailed error reasons such as "xx sub-question not answered".

[0041] In addition, the output of the multimodal LLM model may have some issues with LaTeX format and garbled characters. The system will use the trained LaTeX to repair the model, and will process each field of the output JSON through LaTeX to repair the model, so as to ensure the accuracy of the output format. 2) Matching the number of students to the causes of errors in open-ended questions is relatively complex. Unlike multiple-choice and fill-in-the-blank questions, it's not possible to simply look up specific fields based on students' fixed options and answers and assign them to specific error causes. Since different students' answer processes vary greatly, it's necessary to analyze each student individually. Therefore, this embodiment will again use prompt information to guide the multimodal LLM model to assign students to appropriate error cause results. The prompt information for matching the number of students is as follows: Task Description: "Based on the correct answer provided by the user, the student's response, the response analysis, and a set of possible error types, determine which type of error the student's response belongs to."

[0042] Require: Simply output the sequence number corresponding to the error type; no explanation is required. If no error type can be matched for the student, output -1.

[0043] Input: Question, explanation, correct answer, student answers (uploaded by users), answer analysis, and list of error types.

[0044] Output: Only output the sequence number corresponding to the error type; do not output any other unnecessary information. Note that if a student only failed to answer a sub-question, or only wrote the sub-question number, "Solution:", or listed the formula but did not perform calculations or derivations, and the "error type" to be judged is because the student did answer but made a mistake, do not classify this student under the "error type" to be judged.

[0045] The multi-person, single-question analysis subsystem provides the multimodal LLM model with question information and images of students' answer sheets, facilitating the model's observation of students' specific reasoning details and summarizing common error causes. In the single-person, single-question scenario, the multimodal LLM model first provides a preliminary error cause analysis for the student as a reference for error matching. Then, a list of candidate error types and their explanations are input. If a student's answer belongs to a certain sub-error type, the corresponding index is output; if the student's error does not belong to any candidate error type, -1 is output to indicate that it does not exist. The error similarity list provided in this embodiment represents typical error causes summarized by the multimodal LLM model in the first step based on all students' incorrect answers. Therefore, some students' errors may belong to rare, marginal error causes. It is normal for these students not to be included in the matching of each error type (outputs of -1 will be merged into the marginal error cause students and the number of marginal error cause students will be given).

[0046] 3) The question format will vary depending on the subject (for math problem-solving, we need to see the student's detailed steps. Different students have different solutions. We use a multimodal LLM model to input student answer images and prompts with specific analysis requirements for real-time analysis. The processing logic for math problem-solving is basically the same as that for physics. This example only provides analysis services for all science exam papers and does not involve humanities subjects such as Chinese and English). All four subjects, including math, physics, chemistry, and biology, have multiple-choice questions, including both single-choice and multiple-answer questions. Therefore, the processing method is basically the same for all four subjects.

[0047] For mathematics, fill-in-the-blank questions are mostly single-blank. Even when there are multiple-blank questions, the knowledge points tested are generally concentrated, and the number of blanks is not too large. Therefore, a pre-set format can be used for analysis, which is not much different from fill-in-the-blank questions. Examples of prompt information and output for multiple-choice and fill-in-the-blank questions are as follows: Task description: "Analyze the error type of each option in a math multiple-choice question, and for multiple-answer questions, also analyze the reasons why the correct options were not selected."

[0048] The specific steps are as follows: 3.1) Read and understand the given questions, answers, and explanations.

[0049] 3.2) Analyze each incorrect option, think about why the wrong option was selected or the correct option was missed, and then associate it with the most likely error type.

[0050] 3.3) The error type must be specific and cannot be vague such as "option error" or "selection error". Only the most likely error type needs to be given.

[0051] 3.4) For multiple-choice questions, analyze each option. For correct options, analyze why students might miss them, and for incorrect options, analyze why students might choose them incorrectly. For single-choice questions, only analyze the incorrect options.

[0052] 3.5) Mathematical Formatting: All mathematical formulas, symbols, and expressions appearing in the question content (e.g., variable $x$, set $A$, formula $x^2 - 3x - 10 \le 0$, symbol $\cap$, $\ge$, $\pi$, etc.) must be transcribed in "LaTeX" format and enclosed in a single dollar sign $.

[0053] 3.6) Error types must be described in Chinese; English is strictly prohibited.

[0054] Answer format: Option: A; Error type: Error type (no need to specify whether it is an incorrect selection or an omission, just give the type); Description: A description of the error type; Recommendation: Recommendations for different error types.

[0055] For multiple-choice questions, analyze options A, B, C, and D. For correct options, analyze why you missed selecting them; for incorrect options, analyze why you chose the wrong ones.

[0056] For multiple-choice questions, only the three incorrect options will be analyzed.

[0057] For example, if the question asks "Which of the following statements is correct?" and the answer is "AC", it is a multiple choice question. You need to analyze the four options A, B, C, and D. For A and C, analyze why you missed selecting them, and for B and D, analyze why you selected them incorrectly.

[0058] If the question asks "Which of the following statements is incorrect ( ) or which statement is wrong ( )", and the answer is "B", it is necessary to analyze why the student incorrectly chose ACD.

[0059] It is also necessary to pay attention to the description of some terms. For example, the concept of "derivative" should not be expressed as "calculus"; avoid using advanced vocabulary. For example, before learning derivatives, avoid using the term "extreme value" and instead use more basic expressions such as "maximum value", "minimum value", and "range of values"; when it comes to concepts such as "percentile" and "decile", do not simply use "quantile" to describe them.

[0060] The default processing method for multiple-choice questions is as follows: If the correct option for this question is D, the possible causes and pitfalls of the other three incorrect options are analyzed using a multimodal LLM model. When making an online request, the error cause analysis of the corresponding incorrect option in the pre_error_analysis field is directly matched according to the student's options. This ensures that the error cause analysis of the same incorrect option for a student is consistent.

[0061] Similarly, the pre-set processing method for fill-in-the-blank questions is as follows: Since fill-in-the-blank questions do not have detailed solution processes, and most students tend to make similar mistakes, such as calculation errors, such as calculating 1 as -1, the multimodal LLM model will first infer the possible wrong answers that students may make based on the question stem, answer analysis, and knowledge points as the pre-set error reasons. Online, based on the OCR results of the student's answer image, the corresponding error reason analysis will be directly matched. At this time, there will be a situation where no match is found, so a fallback error reason will be pre-set. When none of the possible wrong answers given by the multimodal LLM model match, the fallback error reason will be given directly, and the analysis given for the fallback error reason will be given in the form of insufficient mastery of the knowledge points involved in the question.

[0062] In physics, fill-in-the-blank questions are generally experimental questions, and they are all in the form of multiple-blank questions. Moreover, the examination methods include both multiple-choice and fill-in-the-blank questions. Therefore, it is not possible to adopt a pre-set analysis method. This invention uses a multimodal LLM model to perform real-time analysis by prompting information.

[0063] In addition, the knowledge points tested in physics experimental questions are generally quite concentrated. The specific knowledge points for each blank may be slightly different, but the overall test is basically an experimental topic. Therefore, a common error reason for each blank will be preset, as well as a summary error reason at the question level, which will be placed first in the output result, while the error reasons for each blank will be arranged from the second position to the last.

[0064] For chemistry and biology, fill-in-the-blank questions are generally experimental procedure questions or material combination questions, and are also in the form of multi-blank fill-in-the-blank questions. Chemistry usually involves one experimental topic, and the knowledge points tested in each blank are roughly the same. Therefore, the same processing logic as physics experimental questions is adopted. However, for biology material questions, the knowledge tested in each blank is relatively scattered. Generally, each blank can stand alone as a question, and the contextual connection is not so strong. Therefore, biology does not need a summary error factor. The multimodal LLM model can directly output the common error factor for each blank.

[0065] The multi-user, multi-question knowledge point analysis subsystem includes: The knowledge point aggregation unit is used to summarize all the knowledge points contained in all the questions based on the question information of the input test paper, and to aggregate all the questions containing that knowledge point by each knowledge point; The dynamic pruning unit is used to dynamically truncate the prompt information according to the token length, and to progressively eliminate students' answering processes from low to high score loss. Analysis restriction unit is used to restrict the multimodal LLM model to analyze only the current knowledge point, and to input other knowledge points in the test paper as knowledge points that are not allowed to be mentioned into the model.

[0066] The multi-user, multi-question knowledge point analysis subsystem aggregates question information for different knowledge points and, based on a multimodal LLM model, obtains a second analysis report. The process includes: Based on the test paper information, summarize the knowledge points contained in all questions, and aggregate the question information according to the knowledge points to obtain the knowledge point questions; Collect all student OCR answers for each knowledge point question, dynamically insert prompt information, and truncate the prompt information according to the token length; Input the truncated prompt information into the multimodal LLM model, request the multimodal LLM model to output the analysis results, and format the analysis results to obtain the second analysis report.

[0067] Furthermore, the dynamic clipping unit includes: The length judgment subunit is used to determine whether the total character length of the concatenated prompt information exceeds the limit; The progressive elimination sub-unit is used to eliminate the last student's answer in order of increasing student error rate; The loop checking sub-unit is used to continuously check the input length until the maximum number of requested input tokens of the multimodal LLM model is met.

[0068] Furthermore, the analysis limiting units include: The knowledge point constraint sub-unit is used to prompt the multimodal LLM model to analyze only the current knowledge point; The student matching subunit is used to require the multimodal LLM model to provide a specific student sequence for each error type; The format specification subunit is used to require that all mathematical formulas, symbols, and expressions must be transcribed in LaTeX format.

[0069] Specifically: 1) First, the test questions are aggregated based on knowledge points (directly summarizing the union of all knowledge points contained in the questions based on the input test information, and then counting the questions containing each knowledge point based on the total number of knowledge points in the summary. For example, if a test paper has three questions testing "solving triangles", then the error analysis for the knowledge point of "solving triangles" will start from these three questions. The subsystem will also provide the multimodal LLM model with the incorrect student answers for each of these three questions, and summarize them so that the multimodal LLM model can analyze the common knowledge point errors, such as whether the problem is due to a lack of mastery of the triangle-solving method or a calculation error). All question information under each knowledge point is summarized, and then all student error OCR answers for multiple questions under each knowledge point are collected. Since a knowledge point may involve multiple questions, and a question may involve multiple students who made mistakes, inserting all the content into the prompt information may cause the token to be too long, resulting in the input exceeding the maximum request input token amount of the multimodal LLM model. Therefore, this invention dynamically prunes the input prompt information.

[0070] like Figure 2 As shown, when the input prompt information is concatenated, the total character length of the input is judged. If it exceeds the limit, the last student's answer is removed. Then, the input character length is checked again. If it still does not meet the limit, the process of removing the last student's answer continues until the input length meets the limit requirement. (In this embodiment, the question stem, answer, explanation, and related knowledge points and scores are not deleted. Only the answer process of students with excessive scores is deleted. In this embodiment, the scores of all students who answered all questions related to the knowledge point incorrectly are sorted. Students with fewer points may only have some minor mistakes. Therefore, they will not receive much attention and analysis from the multimodal LLM model when analyzing typical knowledge point deficiencies. Therefore, this invention deletes the answer process of students with fewer points first, and then removes them progressively from low to high score until the length of the prompt information reaches a reasonable token range acceptable to the multimodal LLM model.) Then, the multimodal LLM model is requested to perform analysis.

[0071] 2) The prompt information for inputting the multimodal LLM model is as follows: Task Description: "Analyze the typical errors and underlying causes of students' mistakes on a specific knowledge point. Based on multiple questions corresponding to a given knowledge point (the knowledge point to be analyzed) and the answers of multiple students for each question, analyze the underlying causes of students' errors from the perspective of the given knowledge point and summarize their typical error patterns. Provide corresponding teaching suggestions for each type of error pattern. Note that if the output involves mathematical formulas, ensure that each output is in "LaTeX" format."

[0072] The specific steps are as follows: 2.1) Read and understand the multiple questions on the given knowledge points, including the question types, question scores, question answers, detailed explanations, and each student's answers (including student scores).

[0073] 2.2) Compare all students' answers with the question answers, and point out all possible reasons for students' errors in this knowledge point (student answers may be blank or only some steps are written). Other knowledge points should not be included. Students' answers may be garbled due to OCR recognition errors. Pay attention to distinguishing between blank student answers and OCR recognition errors.

[0074] 2.3) Based on all the information provided in the question analysis and detailed explanation, identify the main error types and common causes of errors for each "knowledge point to be analyzed." Note that the error types given must be relevant to the knowledge point, such as errors in determining the range, ignoring radicals, or limitations on the denominator, rather than calculation errors, logical breaks, multiple-choice questions / missing options, or incomplete analysis of multiple-choice options—errors related to ability. Pay special attention to ensuring that the given error reasons are concise and specific, not too general. The error descriptions should be detailed and feasible, not vague, to ensure practical reference value. Avoid repeating error reasons, especially semantic repetition, ensuring diversity and accuracy. Furthermore, the descriptions of error reasons must be relevant to the knowledge point and the content tested in the questions related to that knowledge point.

[0075] 2.4) Provide an analysis report that details the common reasons for student errors in the knowledge points to be analyzed. You can expand on the reasons appropriately, but do not mention other knowledge points that are not to be analyzed. If there are multiple common reasons, you should give suggestions for such error situations for each error reason. In addition, provide a specific student sequence for each error type, such as ["Zhang San", "Li Si", "Wang Wu"], and ensure that the listed student sequences are all students who have appeared in the input data.

[0076] 2.5) Mathematical Formatting: All mathematical formulas, symbols, and expressions appearing in the question content must be transcribed in LaTeX format and enclosed in a single dollar sign $.

[0077] 2.6) Analyze the reasons for student errors based on the question scores and student scores. For multiple-choice questions (which may include multiple selections), locate the options and analyze the common reasons for students' incorrect, multiple, and omissions. For problem-solving questions, if the student's score is high, there should be few errors; if the score is low, there should be many errors. Pay special attention to the question type and provide reasonable error types based on the question format. For multiple-choice and fill-in-the-blank questions, do not provide error types such as "lack of detailed steps." When describing the reasons for errors, be precise and avoid using vocabulary beyond the syllabus.

[0078] 2.7) Error types must be described in Chinese; English is strictly prohibited.

[0079] Analysis requirements: Knowledge points that are not allowed to be mentioned cannot be mentioned; the focus must be on the knowledge point to be analyzed. If some students' mistakes cannot be explained by the knowledge point to be analyzed, it means that they do not belong to the knowledge point to be analyzed. The cause of the mistake must be focused on the knowledge point to be analyzed. Even if the given question contains other knowledge points, no forced analysis will be performed. All mistakes that do not belong to this knowledge point will be ignored.

[0080] Restrictions: When listing the students corresponding to each error, ensure that the student actually falls under this error category. If a student's error does not fall within the given error categories, the student may be omitted from the list.

[0081] Since a question may involve multiple knowledge points, and each aggregated question under a knowledge point belongs to that knowledge point, the question itself may contain other knowledge points. To prevent the multimodal LLM model from unexpectedly outputting other knowledge points involved in the example questions when analyzing the error causes of the current knowledge point, the analysis of the multimodal LLM model is conditionally restricted. The multimodal LLM model is only allowed to analyze the knowledge point to be analyzed, and knowledge points that are not allowed to be mentioned are not permitted. This subsystem will prompt the multimodal LLM model that the knowledge point to be analyzed is the provided current knowledge point. Simultaneously, the system will summarize all knowledge points involved in all questions of the requested test paper, and provide the multimodal LLM model with all knowledge points tested in this test paper except for the currently analyzed knowledge point, indicating that these are knowledge points that are not permitted to be mentioned. This restricts the multimodal LLM model from forcibly analyzing other knowledge points contained in the questions involving the current knowledge point. Furthermore, the system will statistically analyze all knowledge points involved in the test paper, classifying all non-requested knowledge points as not permitted to be mentioned and inputting them into the multimodal LLM model. The error causes output by the multimodal LLM model are formatted and the model automatically maps students to the output error types. Each error cause is accompanied by the student ID belonging to that error type, providing end-to-end error type, description suggestions, and matching students.

[0082] The multi-user, multi-question ability analysis subsystem includes: The ability aggregation unit is used to summarize the abilities contained in all questions based on the test information of the input test paper, and aggregate all questions containing that ability for each ability. The subject competency database stores the definitions and explanations of competencies and qualities in different subjects, and retrieves the corresponding subject's prompt information based on the requested subject ID. The capability matching unit is used to enable the multimodal LLM model to match the specific student ID for each error type.

[0083] The multi-user, multi-question ability analysis subsystem aggregates question information for different abilities and, based on a multimodal LLM model, obtains a third-party analysis report. The process includes: Based on the test paper information, summarize the abilities contained in all questions, and aggregate the question information according to the abilities to obtain the ability questions; Collect all student OCR errors for each ability question, dynamically insert prompt information, and truncate the prompt information according to the token length; Input the truncated prompt information into the multimodal LLM model, request the multimodal LLM model to output the analysis results, and format the analysis results to obtain the third analysis report.

[0084] Furthermore, the system provided by this invention supports interdisciplinary unified analysis, establishing a comparable analytical index system across the four disciplines of mathematics, physics, chemistry, and biology, wherein: Mathematical competencies include: mathematical abstraction, logical reasoning, mathematical modeling, intuitive imagination, mathematical operations, and data analysis. The competencies and qualities in physics include: physical concepts, scientific thinking, scientific inquiry, scientific attitude and responsibility; The competencies and qualities required in chemistry include: macroscopic identification, microscopic analysis, and the concept of change; The competencies and qualities in biology include: life concepts, scientific inquiry, and social responsibility.

[0085] Specifically: 1) For multiple students and multiple questions covering the same knowledge points, first, aggregate the questions related to each ability (using the same method as knowledge point aggregation). Then, insert prompt information and check its length. If the length is not met, dynamically remove the last student's answer. If the length is met, input it into the multimodal LLM model for analysis. In this embodiment, all abilities and competencies involved in different subjects, along with their explanations and definitions, will be provided and requested to the multimodal LLM model for reference and analysis. The input prompt information is as follows: Task Description: "Analyze students' typical errors, weaknesses, and underlying causes in a certain competency. Based on multiple questions corresponding to a given competency and the answers of multiple students for each question, analyze the underlying causes of students' weaknesses in that competency and summarize their typical errors and weaknesses. Provide corresponding teaching suggestions for each type of error. Note that if the output involves mathematical formulas, ensure that each output is in "LaTeX" format."

[0086] The specific steps are as follows: 1.1) Read and understand the given multiple questions related to the relevant abilities and qualities, including the question types, question scores, question answers, detailed explanations, and each student's answers (including student scores).

[0087] 1.2) Compare all students' answers with the question answers, and point out all possible reasons for students' errors, typical error methods, and weak points under this ability level (students' answers may be blank or they may only write a few steps). Students' answers may be garbled due to OCR recognition errors. Pay attention to distinguishing between blank student answers and OCR recognition errors.

[0088] 1.3) Based on all the information provided in the question analysis and detailed explanation, especially the information on the students' abilities tested in the question, analyze the reasons for the students' deficiencies and the types of weaknesses in this ability. Note that the error types given must be related to the ability and must be directly related to the ability itself, such as calculation errors, careless errors, etc., rather than errors related to knowledge points such as errors in judging the value range, ignoring the square root, or the restrictions of the denominator. "Pay special attention to the reasons for the errors given. They should be concise and specific, not too general, and the error descriptions should be detailed and feasible, not vague, to ensure that they have practical reference value. The reasons for the errors given for the same question should not be repeated, especially the semantic repetition, to ensure diversity and accuracy."

[0089] 1.4) Provide an analysis report that details the common causes and methods of student errors under this competency, as well as weaknesses. The report can be expanded upon appropriately. The summary of error causes should be concise and highlight key points. If there are multiple common weaknesses or error methods, suggestions should be given for each error cause. Furthermore, provide a specific sequence of weak students for this competency, such as ["Zhang San", "Li Si", "Wang Wu"], ensuring that the listed student sequences are all students who appeared in the input data.

[0090] 1.5) Mathematical Formatting: All mathematical formulas, symbols, and expressions appearing in the question content must be transcribed in LaTeX format and enclosed in a single dollar sign $.

[0091] 1.6) Analyze the reasons for students' errors based on the question scores and students' scores. For multiple-choice questions (which may include multiple selections), locate the options and analyze the common reasons for students' incorrect, multiple, and missing selections. For problem-solving questions, if the student's score is high, the lack of ability should be minimal; if the score is low, the lack of ability should be significant. Pay special attention to the question type and provide reasonable error types based on the question format. Do not provide error types such as "lack of detailed steps" for multiple-choice and fill-in-the-blank questions.

[0092] 1.7) Score Correlation Analysis: Specifically, the severity of errors is determined based on student scores. A score of ≥90% (out of 5): Focus on minor errors (such as calculation mistakes / omission of units, etc.). 50% ≤ Score < 90%: Analyze key steps for errors and knowledge gaps; Score <50%: Systematic lack of understanding of basic concepts and problem-solving methods.

[0093] In this embodiment, competency and skills include the following six aspects and explanations: Different disciplines have different competency and skills requirements in this embodiment. Each discipline has its own independent prompt information. When a request is made, the prompt information of the corresponding discipline will be retrieved based on the requested discipline ID. For example, the competency and skills requirements for physics are as follows: Physical concepts: Fundamental understandings of matter, motion and interaction, energy, etc., formed from a physics perspective; the refinement and sublimation of physical concepts and laws in the mind; the foundation for explaining natural phenomena and solving practical problems from a physics perspective. It should include elements such as the concept of matter, the concept of motion and interaction, and the concept of energy.

[0094] Scientific thinking is a way of understanding the essential attributes, inherent laws, and interrelationships of objective things from a physics perspective; it is the abstract generalization process of constructing physical models based on empirical facts; it is the specific application of methods such as analysis, synthesis, reasoning, and argumentation in the scientific field; it is the ability and character to question and criticize different viewpoints and conclusions based on factual evidence and scientific reasoning, to test and revise them, and ultimately to put forward creative insights. It mainly includes elements such as physical model construction, scientific reasoning, scientific argumentation, questioning, and innovation.

[0095] Scientific inquiry encompasses the ability to formulate physical questions based on observation and experimentation, develop conjectures and hypotheses, design experiments and formulate plans, acquire and process information, draw conclusions based on evidence and provide explanations, and communicate, evaluate, and reflect on the process and results of scientific inquiry. It primarily includes elements such as questions, evidence, explanations, and communication.

[0096] Scientific attitude and responsibility: Based on an understanding of the nature of science and the relationship between science, technology, society, and the environment, this gradually develops into an intrinsic motivation to explore nature, a rigorous, serious, pragmatic, and persistent scientific attitude, and a sense of responsibility to abide by ethical standards, protect the environment, and promote sustainable development. It mainly includes elements such as the nature of science, scientific attitude, and social responsibility.

[0097] Mathematical skills and literacy are as follows: Mathematical abstraction: the ability to understand mathematical research objects through abstraction of quantitative relationships and spatial forms. This mainly includes: abstracting mathematical concepts and relationships from quantities and quantitative relationships, and from graphs and graph relationships; abstracting general rules and structures from the specific context of things, and representing them using mathematical language.

[0098] Logical reasoning: The ability to deduce other propositions from facts and propositions according to rules. It mainly includes: reasoning from the specific to the general, with induction and analogy as the main forms of reasoning; and reasoning from the general to the specific, with deduction as the main form of reasoning.

[0099] Mathematical modeling: The ability to abstract real-world problems mathematically, express them using mathematical language, and construct mathematical models to solve them. This primarily includes: identifying and posing problems from a mathematical perspective in real-world situations; analyzing problems; establishing mathematical models; determining parameters; calculating and solving; verifying results; improving mathematical models; and ultimately solving practical problems.

[0100] Intuitive imagination: The ability to perceive the form and changes of things through geometric intuition and spatial imagination, and to understand and solve mathematical problems using spatial forms, especially graphics. This mainly includes: recognizing the positional relationships, morphological changes, and laws of motion of things through spatial forms; describing and analyzing mathematical problems using graphics; establishing connections between form and number; constructing intuitive models of mathematical problems; and exploring problem-solving strategies.

[0101] Mathematical operations: The ability to solve mathematical problems based on a clear understanding of the objects of operation and according to the rules of operation. This mainly includes: understanding the objects of operation, mastering the rules of operation, exploring operational thinking, selecting operational methods, designing operational procedures, and obtaining the results.

[0102] Data analysis involves acquiring data about a research subject, using mathematical methods to organize, analyze, and infer information from that data, and developing knowledge about the research subject. This primarily includes: collecting data, organizing data, extracting information, building data analysis models, making inferences, and drawing conclusions.

[0103] Error types must be described in Chinese; English is strictly prohibited.

[0104] 2) Format the output of the multimodal LLM model to provide specific deficiencies and suggestions for each capability in the form of 1. 2. 3., and let the multimodal LLM model match the specific student ID for each error type.

[0105] Furthermore, the system provided by this invention also includes: The report integration module is used to integrate the first analysis report, the second analysis report, and the third analysis report to generate the final learning analysis report; The statistical indicators module is used to generate statistical indicators for class learning assessment, including the accuracy rate of summarizing common mistakes and the matching rate of students.

[0106] By combining the first, second, and third analysis reports based on the output of multiple students working on single questions, multiple students working on multiple questions covering various knowledge points, and multiple students working on single questions demonstrating their abilities, a final learning analysis report and some statistical indicators are provided to facilitate subsequent assessment of the overall learning situation of the class.

[0107] Because multiple-choice and fill-in-the-blank questions are pre-analyzed, they can achieve a high accuracy and matching rate. The problem-solving questions are slightly worse than the multiple-choice and fill-in-the-blank questions. Tables 1 and 2 are the analysis report indicators of a class learning analysis system based on an AI big data model in a pilot exam at a high school.

[0108] Table 1. Indicators from the Analysis Report of Grade 11 Classes in a Pilot Exam at a High School Table 2. Indicators from the Analysis Report of Grade 11 Classes in a Pilot Exam at a High School In summary, this invention employs different approaches for different subjects and question types, inputting the solution processes of all students who made mistakes into a multimodal LLM model. Through carefully designed prompt information, it analyzes common errors and key points of deduction among students. By tagging predicted questions with knowledge point labels and then aggregating different questions by knowledge point, it analyzes students' mastery of knowledge points from a knowledge point perspective. Furthermore, by prompting the multimodal LLM model to tag questions with ability labels and then aggregating different questions by ability, it analyzes the overall ability radar of students in the class from an ability dimension. The question types differ slightly across subjects, and different analysis logics are used for different subjects, achieving personalized processing of class learning analysis.

[0109] Example 2 like Figure 3 As shown, based on the same inventive concept, this invention also provides a method for analyzing class learning based on an AI large-scale model, implemented using the system described in the foregoing embodiments. The method includes: Based on the multi-person single-question analysis subsystem, different analysis strategies are adopted for error analysis and number matching according to different question types. At the same time, different preset strategies are adopted for error analysis according to the question setting methods of different subjects to obtain the first analysis report. Based on the multi-person, multi-question knowledge point analysis subsystem, question information is aggregated for different knowledge points, and a second analysis report is obtained based on the multimodal LLM model; Based on the multi-person, multi-question ability analysis subsystem, question information is aggregated for different abilities, and a third analysis report is obtained based on the multimodal LLM model; By combining the first, second, and third analysis reports, a final learning situation analysis report is generated, and a detailed analysis of the class's learning situation is carried out based on the final learning situation analysis report.

[0110] Furthermore, based on the multi-person single-question analysis subsystem, different analysis strategies are adopted for error cause analysis and number matching according to different question types. At the same time, different preset strategies are adopted for error cause analysis according to the question setting methods of different subjects. The methods for obtaining the first analysis report include: The multi-person single-question analysis subsystem determines whether all students have answered the input question correctly. If all students have answered correctly, it directly outputs the error reason; otherwise, it determines the type of the input question. For multiple-choice and fill-in-the-blank questions, preset error reasons are generated directly based on the input question and stored in the MongoDB database. For open-ended questions, a multimodal LLM model is used to perform error reason analysis and generation in real time to obtain the error reason analysis and number of respondents matching results. Based on the question-setting methods of different subjects, different pre-set strategies are used for analysis and processing to obtain subject analysis results; Based on the analysis of the reasons for errors in the questions, the results of matching the number of participants, and the results of subject analysis, the first analysis report was obtained.

[0111] Furthermore, based on the multi-user, multi-question knowledge point analysis subsystem, the method for aggregating question information for different knowledge points and obtaining a second analysis report based on a multimodal LLM model includes: Based on the test paper information, summarize the knowledge points contained in all questions, and aggregate the question information according to the knowledge points to obtain the knowledge point questions; Collect all student OCR answers for each knowledge point question, dynamically insert prompt information, and truncate the prompt information according to the token length; Input the truncated prompt information into the multimodal LLM model, request the multimodal LLM model to output the analysis results, and format the analysis results to obtain the second analysis report.

[0112] Furthermore, based on the multi-user, multi-question ability analysis subsystem, methods for aggregating question information for different abilities and obtaining a third-party analysis report based on a multimodal LLM model include: Based on the test paper information, summarize the abilities contained in all questions, and aggregate the question information according to the abilities to obtain the ability questions; Collect all student OCR errors for each ability question, dynamically insert prompt information, and truncate the prompt information according to the token length; Input the truncated prompt information into the multimodal LLM model, request the multimodal LLM model to output the analysis results, and format the analysis results to obtain the third analysis report.

[0113] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A class learning analysis system based on an AI large-scale model, characterized in that, The system includes: a multi-person single-question analysis subsystem, a multi-person multi-question knowledge point analysis subsystem, and a multi-person multi-question ability analysis subsystem; The multi-person single-question analysis subsystem is used to perform error analysis and number matching based on different question types and different pre-set strategies based on the question format of different subjects, and to obtain the first analysis report. The multi-user, multi-question knowledge point analysis subsystem is used to aggregate question information for different knowledge points and obtain a second analysis report based on a multimodal LLM model. The multi-person, multi-question ability analysis subsystem is used to aggregate question information for different abilities and obtain a third analysis report based on a multimodal LLM model.

2. The system according to claim 1, characterized in that, The multi-user single-question analysis subsystem includes: The question type determination unit is used to automatically identify the question type based on the input question type field, which includes multiple choice questions, fill-in-the-blank questions, and problem-solving questions; The "All Correct" judgment unit is used to determine whether all students have answered the question correctly. If all students have answered correctly, the error reason will be output directly. Subject-specific adaptation units are used to employ corresponding pre-defined analysis strategies based on the question formats of different subjects.

3. The system according to claim 2, characterized in that, The multi-person single-question analysis subsystem employs different analysis strategies for error analysis and participant matching based on different question types. Simultaneously, it uses different pre-set strategies for error analysis based on the question format of different subjects. The process of obtaining the first analysis report includes: The multi-person single-question analysis subsystem determines whether all students have answered the input question correctly. If all students have answered correctly, it directly outputs the error reason; otherwise, it determines the type of the input question. For multiple-choice and fill-in-the-blank questions, preset error reasons are generated directly based on the input question and stored in the MongoDB database. For open-ended questions, a multimodal LLM model is used to perform error reason analysis and generation in real time to obtain the error reason analysis and number of respondents matching results. Based on the question-setting methods of different subjects, different pre-set strategies are used for analysis and processing to obtain subject analysis results; Based on the analysis of the reasons for errors in the questions, the results of matching the number of participants, and the results of subject analysis, the first analysis report was obtained.

4. The system according to claim 2, characterized in that, For processing multiple-choice questions, the multi-person single-question analysis subsystem also includes: The multiple-choice question analysis unit is used to analyze the reasons for errors in the three incorrect options other than the correct one in a multiple-choice question. The multiple-choice question analysis unit is used to analyze the reasons for errors in all four options of a multiple-choice question. It analyzes the reasons for missing correct options and the reasons for incorrect options. The error cause storage unit is used to store the analyzed error causes into the MongoDB database using the error options as the query key.

5. The system according to claim 2, characterized in that, For handling fill-in-the-blank questions, the multi-user single-question analysis subsystem also includes: The error answer prediction unit is used to predict possible incorrect answers from students as preset error reasons using a multimodal LLM model; The "Catch-Up Error Reason" section is used to cover errors caused by insufficient understanding of the relevant knowledge points. The matching judgment unit is used to enable a fallback error cause when a student's OCR answer does not match the predicted incorrect answer.

6. The system according to claim 2, characterized in that, For handling problem-solving questions, the multi-user single-question analysis subsystem also includes: The real-time analysis unit is used to analyze the causes of errors in students' problem-solving process in real time using a multimodal LLM model; The scoring correlation unit is used to determine the severity of errors based on students' scores: when the score is ≥90% of the full score, the focus is on detailed errors; when the score is 50% ≤ and <90%, the analysis focuses on errors in key steps and knowledge gaps; when the score is <50%, the systemic lack of basic concepts and problem-solving methods is identified. The student matching unit is used to assign students to the corresponding error category using a multimodal LLM model.

7. The system according to claim 2, characterized in that, For problem-solving questions in different subjects, the multi-user single-question analysis subsystem also includes: The math problem-solving analysis unit is used for real-time analysis of student answer images input into a multimodal LLM model; The physics experiment question analysis unit is used to pre-determine the common error reasons for each blank and the summary error reasons at the question-wide level; The analysis unit for chemistry experiment questions uses the same processing logic as that for physics experiment questions. The biological materials question analysis unit is used to directly output common errors for each blank, without needing to summarize the error causes.

8. The system according to claim 2, characterized in that, The multi-user single-question analysis subsystem also includes: The format repair unit is used to repair the format of the output of the multimodal LLM model using a trained LaTeX repair model, so as to ensure the format accuracy of the output results. The marginal error processing unit is used to process students who cannot be matched with any candidate error type, merge them into the marginal error category, and count the number of students with marginal errors.

9. The system according to claim 1, characterized in that, The multi-user, multi-question knowledge point analysis subsystem includes: The knowledge point aggregation unit is used to summarize all the knowledge points contained in all the questions based on the question information of the input test paper, and to aggregate all the questions containing that knowledge point by each knowledge point; The dynamic pruning unit is used to dynamically truncate the prompt information according to the token length, and to progressively eliminate students' answering processes from low to high score loss. Analysis restriction unit is used to restrict the multimodal LLM model to analyze only the current knowledge point, and to input other knowledge points in the test paper as knowledge points that are not allowed to be mentioned into the model.

10. The system according to claim 9, characterized in that, The multi-user, multi-question knowledge point analysis subsystem aggregates question information for different knowledge points and, based on a multimodal LLM model, obtains a second analysis report. The process includes: Based on the test paper information, summarize the knowledge points contained in all questions, and aggregate the question information according to the knowledge points to obtain the knowledge point questions; Collect all student OCR answers for each knowledge point question, dynamically insert prompt information, and truncate the prompt information according to the token length; Input the truncated prompt information into the multimodal LLM model, request the multimodal LLM model to output the analysis results, and format the analysis results to obtain the second analysis report.

11. The system according to claim 9, characterized in that, The dynamic cropping unit includes: The length judgment subunit is used to determine whether the total character length of the concatenated prompt information exceeds the limit; The progressive elimination sub-unit is used to eliminate the last student's answer in order of increasing student error rate; The loop checking sub-unit is used to continuously check the input length until the maximum number of requested input tokens of the multimodal LLM model is met.

12. The system according to claim 9, characterized in that, The analysis restriction unit includes: The knowledge point constraint sub-unit is used to prompt the multimodal LLM model to analyze only the current knowledge point; The student matching subunit is used to require the multimodal LLM model to provide a specific student sequence for each error type; The format specification subunit is used to require that all mathematical formulas, symbols, and expressions must be transcribed in LaTeX format.

13. The system according to claim 1, characterized in that, The multi-user, multi-question ability analysis subsystem includes: The ability aggregation unit is used to summarize the abilities contained in all questions based on the test information of the input test paper, and aggregate all questions containing that ability for each ability. The subject competency database stores the definitions and explanations of competencies and qualities in different subjects, and retrieves the corresponding subject's prompt information based on the requested subject ID. The capability matching unit is used to enable the multimodal LLM model to match the specific student ID for each error type.

14. The system according to claim 13, characterized in that, The multi-user, multi-question ability analysis subsystem aggregates question information for different abilities and, based on a multimodal LLM model, obtains a third-party analysis report. The process includes: Based on the test paper information, summarize the abilities contained in all questions, and aggregate the question information according to the abilities to obtain the ability questions; Collect all student OCR errors for each ability question, dynamically insert prompt information, and truncate the prompt information according to the token length; Input the truncated prompt information into the multimodal LLM model, request the multimodal LLM model to output the analysis results, and format the analysis results to obtain the third analysis report.

15. The system according to claim 13, characterized in that, The system supports unified interdisciplinary analysis, establishing a comparable analytical index system across the four disciplines of mathematics, physics, chemistry, and biology, including: Mathematical competencies include: mathematical abstraction, logical reasoning, mathematical modeling, intuitive imagination, mathematical operations, and data analysis. The competencies and qualities in physics include: physical concepts, scientific thinking, scientific inquiry, scientific attitude and responsibility; The competencies and qualities required in chemistry include: macroscopic identification, microscopic analysis, and the concept of change; The competencies and qualities in biology include: life concepts, scientific inquiry, and social responsibility.

16. The system according to claim 1, characterized in that, The system also includes: The report integration module is used to integrate the first analysis report, the second analysis report, and the third analysis report to generate the final learning analysis report; The statistical indicators module is used to generate statistical indicators for class learning assessment, including the accuracy rate of summarizing common mistakes and the matching rate of students.

17. A method for analyzing class learning based on an AI large-scale model, wherein the method is implemented using the system described in any one of claims 1-16, characterized in that, The method includes: Based on the multi-person single-question analysis subsystem, different analysis strategies are adopted for error analysis and number matching according to different question types. At the same time, different preset strategies are adopted for error analysis according to the question setting methods of different subjects to obtain the first analysis report. Based on the multi-person, multi-question knowledge point analysis subsystem, question information is aggregated for different knowledge points, and a second analysis report is obtained based on the multimodal LLM model; Based on the multi-person, multi-question ability analysis subsystem, question information is aggregated for different abilities, and a third analysis report is obtained based on the multimodal LLM model; By combining the first, second, and third analysis reports, a final learning situation analysis report is generated, and a detailed analysis of the class's learning situation is carried out based on the final learning situation analysis report.

18. The method according to claim 17, characterized in that, Based on the multi-person single-question analysis subsystem, different analysis strategies are adopted for error analysis and number matching according to different question types. At the same time, different preset strategies are used for error analysis according to the question setting methods of different subjects. The methods for obtaining the first analysis report include: The multi-person single-question analysis subsystem determines whether all students have answered the input question correctly. If all students have answered correctly, it directly outputs the error reason; otherwise, it determines the type of the input question. For multiple-choice and fill-in-the-blank questions, preset error reasons are generated directly based on the input question and stored in the MongoDB database. For open-ended questions, a multimodal LLM model is used to perform error reason analysis and generation in real time to obtain the error reason analysis and number of respondents matching results. Based on the question-setting methods of different subjects, different pre-set strategies are used for analysis and processing to obtain subject analysis results; Based on the analysis of the reasons for errors in the questions, the results of matching the number of participants, and the results of subject analysis, the first analysis report was obtained.

19. The method according to claim 17, characterized in that, Based on a multi-user, multi-question knowledge point analysis subsystem, the method for aggregating question information for different knowledge points and obtaining a second analysis report based on a multimodal LLM model includes: Based on the test paper information, summarize the knowledge points contained in all questions, and aggregate the question information according to the knowledge points to obtain the knowledge point questions; Collect all student OCR answers for each knowledge point question, dynamically insert prompt information, and truncate the prompt information according to the token length; Input the truncated prompt information into the multimodal LLM model, request the multimodal LLM model to output the analysis results, and format the analysis results to obtain the second analysis report.

20. The method according to claim 17, characterized in that, Based on a multi-user, multi-question ability analysis subsystem, methods for aggregating question information for different abilities and obtaining a third-party analysis report based on a multimodal LLM model include: Based on the test paper information, summarize the abilities contained in all questions, and aggregate the question information according to the abilities to obtain the ability questions; Collect all student OCR errors for each ability question, dynamically insert prompt information, and truncate the prompt information according to the token length; Input the truncated prompt information into the multimodal LLM model, request the multimodal LLM model to output the analysis results, and format the analysis results to obtain the third analysis report.

Images