A graduation thesis review method and system based on task decomposition and collaborative decision-making
By constructing a graduation thesis review system based on task decomposition and collaborative decision-making, and utilizing large language models and multimodal large models, the efficiency and accuracy problems of traditional manual review are solved, achieving efficient and interpretable automatic review of graduation theses and improving the consistency and reliability of review results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING UNIV OF POSTS & TELECOMM
- Filing Date
- 2026-02-12
- Publication Date
- 2026-05-29
AI Technical Summary
Traditional manual review of graduation theses suffers from high knowledge update burden, unstable review efficiency, and large subjective cognitive biases. The application of existing digital technologies in graduation thesis review has significant limitations.
A task-based decomposition and collaborative decision-making approach is adopted to construct a thesis review system using a large language model and a vision-language multimodal model. The system includes a thesis parsing agent, a review agent, an optimization agent, and a finalization agent, thereby achieving automated decomposition and collaborative decision-making evaluation of the thesis.
The generation of review results that are highly consistent with expert reviews improves the accuracy and credibility of the review process, enhances the interpretability and scalability of the review process, and increases review efficiency.
Smart Images
Figure CN122113899A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a text processing method, specifically, to a method and system for reviewing graduation theses based on task decomposition and collaborative decision-making. This invention belongs to the field of artificial intelligence technology. Background Technology
[0002] The graduation thesis is a comprehensive assessment of students' mastery of basic professional knowledge and their ability to apply it. It serves as a means to cultivate students' ability to connect theory with practice, comprehensively apply their knowledge, and improve their analytical and problem-solving skills. It is also the final stage of students' academic assessment. Thesis review refers to the academic evaluation of the thesis by reviewers before the defense, scoring it according to pre-established scoring criteria and determining whether the student is eligible to participate in the defense. In this crucial stage of thesis review, the traditional manual review model faces many pressing problems. Research shows that, in the context of the current popularization and large-scale development of higher education, manual review suffers from high knowledge update burdens, unstable review efficiency, and significant subjective cognitive biases. These problems seriously affect the quality and efficiency of thesis review work.
[0003] In recent years, the field of automated academic paper review has undergone a revolutionary transformation with the rapid development of Large Language Models (LLM) and Multimodal Large Models (MLLM). However, despite the significant practical value and application potential of digital intelligence technologies, represented by Large Language Models, in empowering paper review, existing research still has limitations. Most studies focus on exploring the application of generative artificial intelligence technologies in academic texts of moderate length, such as journal or conference papers, with few studies discussing the capabilities of next-generation artificial intelligence technologies in tasks such as graduation theses.
[0004] In view of this, this invention addresses the real pain points and practical needs of graduation thesis review in actual educational scenarios, exploring a graduation thesis review method and system with task decomposition and collaborative decision-making capabilities. Specifically, this invention aims to construct a graduation thesis review method and system with task decomposition and collaborative decision-making capabilities by integrating key technologies such as prompt word engineering, tool invocation, and model context protocols, based on the core capabilities of large language models and visual-language multimodal large models. Summary of the Invention
[0005] For the reasons mentioned above, the purpose of this invention is to provide an efficient, reliable, and interpretable method and system for reviewing graduation theses based on task decomposition and collaborative decision-making.
[0006] To achieve the above objectives, the present invention adopts the following technical solution: a graduation thesis review method based on task decomposition and collaborative decision-making, which includes the following steps: S1: Receive and parse the graduation thesis D, and generate a structured document tree H; Receive the graduation thesis D and input it into the graduation thesis parsing agent; The graduation thesis parsing agent parses the graduation thesis, extracts text data from the thesis, performs fine-grained information extraction based on VLM multimodal understanding for non-text data, and locates, segments and recognizes image regions to generate editable and structured text descriptions for image data. After the overall parsing is completed, all information is integrated into Markdown format and a structured document tree H is output. S2: Review the structured document tree H based on the graduation thesis review prompt framework F, and generate the graduation thesis review result R; Construct a graduation thesis review prompt framework F specifically for graduation thesis review tasks. It includes review tasks, review indicators, scoring criteria, output format, and available tools. Based on the constructed graduation thesis review prompt framework, graduation thesis review prompt words are constructed; The constructed graduation thesis review prompts and the structured document tree H generated in step S1 are input into the graduation thesis review intelligent agent; The thesis review agent decomposes the review task according to the review indicators, generates a review plan containing n steps, and completes each step in the review plan one by one. After completing all steps, the thesis review result R is obtained. S3: Optimize the generated graduation thesis review result R to generate the optimized graduation thesis review result R′; Extract the title and abstract of the graduation thesis D, generate embedding vectors using an embedding model, and recall the academic material D′ with the highest semantic relevance by calculating the cosine similarity between the embedding vectors and the documents in the knowledge base, including relevant paragraphs, theoretical frameworks or typical writing paradigms. The recalled academic materials D′, graduation thesis D, and graduation thesis review results R are organized together into a multimodal context window; The intelligent agent for optimizing the graduation thesis review results extracts key information based on context window input and performs multi-level content revision and expansion on the graduation thesis review results R, including correcting judgments that do not conform to existing academic standards, identifying weaknesses in the argumentation logic, strengthening the correspondence between the comments and the original text and reference examples, and generating the optimized graduation thesis review results R′. S4: Self-correct the optimized graduation thesis review results R′ and output the final version of the graduation thesis review report R″; Construct a review framework S specifically for the evaluation results of graduation theses, including structural integrity, language standardization, scoring accuracy, review objectivity, and review guidance; Based on the constructed framework S for reviewing graduation thesis review results, construct prompts for reviewing graduation thesis review results; The constructed graduation thesis review result review prompts and the optimized graduation thesis review result R′ generated in step S3 are input into the graduation thesis review result finalization agent; The graduation thesis review result finalization agent reviews the optimized graduation thesis review result R′ according to the review framework S and generates internal review feedback G; the optimized graduation thesis review result R′ is revised according to the internal review feedback G until no new problems are found in the review stage, or the preset iteration limit is reached; Output the graduation thesis review report R″.
[0007] Preferably, the thesis parsing agent adopts a large vision-language model; the thesis review agent, the thesis review result optimization agent, and the thesis review result finalization agent adopt a large language model; the embedding model is one of BGE-M3, Qwen3-Embedding-8B, and BCE-Embedding-base_v1.
[0008] Preferably, the large visual-language model includes Qwen2-5-vl-32b-32k, PaddleOCR-VL-1.5, and DeepSeek-OCR; The large language model includes Deepseek-r1-671b-64k-thinking, QwQ-32b-128k-thinking, Qwen3-32b-128k-thinking, and Qwen2-5-32b-128k.
[0009] Preferably, the document type of the graduation thesis is .docx, .doc, or .pdf.
[0010] Preferably, the graduation thesis review prompt framework F specifically includes: Review Task T: providing a clear and accurate task description; Review Indicators N: providing thesis review indicators and detailed explanations to ensure the model accurately understands the indicator content; Scoring Standard C: providing indicator values and scoring basis; Output Format M: providing output examples and standardizing the output format; Available Tools I: prompting available tools and their uses, guiding the model to determine when to use which tool.
[0011] Preferably, the review framework specifically for the evaluation results of graduation theses includes the following: Structural integrity: Review whether the review report covers all the components of the standard template, including: basic information about the paper, review details, review results, and specific suggestions for improvement; Language standardization: Review the academic standardization of the language used, including whether the terminology is used accurately, whether the sentences are clear and fluent, whether the logic is coherent, and whether there are any grammatical errors or inconsistent formats. Accuracy of scoring: Review whether the scoring process of the review report is rigorous and reliable. Check whether the scores for each item are strictly based on the established scoring criteria, whether the scores for each item are logically consistent with the corresponding comments, and whether the total score is calculated accurately. Objectivity of review: whether the review comments are based on the quality and evidence of the paper itself, whether there are subjective assumptions or emotional expressions, and whether objective evidence is provided when pointing out shortcomings; Review guidance: Review whether the overall review conclusion is specific and clear, whether the review comments are targeted, and whether the improvement suggestions are actionable.
[0012] A graduation thesis review system based on task decomposition and collaborative decision-making includes a graduation thesis parsing agent, a graduation thesis review agent, a graduation thesis review result optimization agent, and a graduation thesis review result finalization agent. The thesis parsing agent receives the thesis D, parses it, extracts text data from the thesis, performs fine-grained information extraction based on VLM multimodal understanding for non-text data, and locates, segments, and identifies image regions to generate editable and structured text descriptions for image data. After the overall parsing is completed, all information is integrated into Markdown format and outputs a structured document tree H. The thesis review agent receives the structured document tree H output by the thesis parsing agent, as well as the pre-built thesis review prompt framework F and thesis review prompt words; decomposes the review task according to the review indicators in the thesis review prompt framework, generates a review plan containing n steps, and completes the steps in the review plan one by one; after executing all steps, the thesis review result R is obtained. The graduation thesis review result optimization agent: receives the graduation thesis review result R generated by the graduation thesis review agent; based on the multimodal context window, performs key information extraction, and performs multi-level content revision and expansion on the graduation thesis review result R, including correcting judgments that do not conform to existing academic standards, identifying weaknesses in the argumentation logic, strengthening the correspondence between the comments and the original text and reference examples, and generating the optimized graduation thesis review result R′; The thesis review result finalization agent receives the optimized thesis review result R′ generated by the thesis review result optimization agent, as well as a pre-constructed review framework S and review prompts for the thesis review result; reviews the optimized thesis review result R′ according to the review framework S and generates internal review feedback G; revises the optimized thesis review result R′ according to the internal review feedback G until no new problems are found in the review stage, or the preset iteration limit is reached; and outputs the thesis review report R″.
[0013] This invention effectively integrates large language models and multimodal large models through a multi-agent architecture, overcoming the limitations of single models in professional thesis review tasks. It significantly reduces illusions in the model's reasoning process, generating review results highly consistent with expert reviews, thus improving the accuracy and credibility of the review. By systematically integrating key technologies such as prompt word engineering, tool invocation, and context protocols, this invention achieves automated decomposition and collaborative decision-making evaluation for long text tasks such as graduation theses. This not only enhances the interpretability and scalability of the review process but also provides efficient and stable technical support for practical review scenarios. This invention introduces generative artificial intelligence into the complex long text analysis scenario of thesis review, expanding its application boundaries in academic evaluation and potentially inspiring new research directions and promoting the deep integration of artificial intelligence and higher education evaluation systems. Attached Figure Description
[0014] Figure 1 This is a flowchart of the graduation thesis review method of the present invention; Figure 2 This invention provides a flowchart for generating a structured document tree from a graduation thesis. Figure 3 This invention provides a flowchart for generating review results for graduation theses. Figure 4 This is a flowchart illustrating the optimization process for review results in this invention. Figure 5 This is a flowchart illustrating the process of correcting and outputting a paper review report based on the optimized review results, as described in this invention. Detailed Implementation
[0015] The structure and features of the present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that various modifications can be made to the embodiments disclosed herein; therefore, the embodiments disclosed in this specification should not be considered as limitations on the present invention, but merely as examples to make the features of the present invention readily apparent.
[0016] This invention provides a thesis review method with task decomposition and collaborative decision-making capabilities. Four agents—a thesis parsing agent, a thesis review agent, a thesis review result optimization agent, and a thesis review result finalization agent—collaborate to complete an end-to-end review mechanism from the receipt of the thesis to the generation of the review report. Based on a multi-agent collaborative architecture, this invention provides an automatic, quantitative, and interpretable comprehensive evaluation of the thesis from three core dimensions: research significance, academic level, and writing quality.
[0017] like Figure 1 As shown, the thesis review method of this invention, which possesses task decomposition and collaborative decision-making capabilities, is as follows: S1: Receive and parse the graduation thesis D, and generate a structured document tree H.
[0018] like Figure 2 As shown, the specific method is: S1.1: Receive one graduation thesis D.
[0019] S1.2: Check if the graduation thesis is an allowed document type (allowed document types include: .docx, .doc, .pdf). If not, resubmit another graduation thesis; if yes, proceed to the next step.
[0020] S1.3: Input the received graduation thesis D into the graduation thesis parsing intelligent agent.
[0021] S1.4: The graduation thesis parsing agent parses the graduation thesis and outputs a structured document tree H.
[0022] For the text data contained in the graduation thesis, the graduation thesis parsing agent calls the pre-built document extractor to extract the text data.
[0023] For non-text data within the graduation thesis, the thesis parsing agent is guided by system prompts and performs fine-grained information extraction based on VLM multimodal understanding capabilities.
[0024] For the image data contained in the graduation thesis, the graduation thesis parsing agent locates, segments and recognizes the image regions, and generates editable and structured text descriptions.
[0025] After the overall parsing is completed, the thesis parsing agent integrates all the information into Markdown format and outputs a structured thesis tree H.
[0026] The above analytical process can be expressed as Formula 1, where: Represents the final generated structured document tree; D represents the received graduation thesis, f parse This represents an analytic function.
[0027] (Formula 1) S2: Review the structured document tree H based on the graduation thesis review prompt framework F, and generate the graduation thesis review result R.
[0028] like Figure 3 As shown, the specific method is as follows: S2.1: Construct a graduation thesis review prompt framework F specifically for graduation thesis review tasks.
[0029] As shown in Table 1, the graduation thesis review prompt framework F includes review task T, review indicators N, scoring criteria C, output format M, and available tools I.
[0030] Table 1. Graduation Thesis Review Guidelines S2.2: Construct graduation thesis review prompts based on the graduation thesis review prompt framework built in step S2.1.
[0031] S2.3: Input the structured document tree H generated in step S1 and the graduation thesis review prompts constructed in S2.2 into the graduation thesis review agent.
[0032] S2.4: The thesis review agent decomposes the review task T according to the review index N, and generates a review plan P containing n steps.
[0033] This process can be represented by Equation 2, where, This refers to the planning strategy of the intelligent agent used to review graduation theses.
[0034] (Formula 2) S2.5: The graduation thesis review agent completes the plan step by step. The steps in the process can be represented by Formula 3, where, This refers to the planning strategy of the intelligent agent used to review graduation theses. For the first... The steps involved in executing E i It will depend on both the review task T and the complete plan, and will use available tools I according to the scoring criteria C, and follow the output format M to execute the process. After all the steps are completed, the review result R of the graduation thesis will be obtained.
[0035] (Formula 3) S3: Optimize the generated graduation thesis review result R to generate the optimized graduation thesis review result R′.
[0036] like Figure 4 As shown, the specific steps are as follows: S3.1: Extract the title and abstract of the graduation thesis D, generate an embedding vector using an embedding model, and recall the academic material D′ with the highest semantic relevance by calculating the cosine similarity between the embedding vector and the documents in the knowledge base, including relevant paragraphs, theoretical frameworks or typical writing paradigms.
[0037] The above retrieval process can be represented by Formula 4: (Formula 4) Where Q represents the title / abstract of the paper being queried, and K represents the document in the knowledge base. This is an embedded function.
[0038] S3.2: Organize the recalled academic materials D′, graduation thesis D, and graduation thesis review results R into a multimodal context window.
[0039] S3.3: The intelligent agent for optimizing the graduation thesis review results extracts key information based on the context window input and performs multi-level content revision and expansion on the graduation thesis review results R, including correcting judgments that do not conform to existing academic standards, identifying weaknesses in the argumentation logic, strengthening the correspondence between the comments and the original text and reference examples, and generating the optimized graduation thesis review results R′.
[0040] S4: Perform self-correction on the optimized graduation thesis review results R′ and output the final version of the graduation thesis review report R″.
[0041] like Figure 5 As shown, the specific steps are as follows: S4.1: Construct a review framework S specifically for the evaluation results of graduation theses, as shown in Table 2.
[0042] Table 2. Framework for Reviewing Graduation Thesis Evaluation Results S4.2: Based on the graduation thesis review result review framework S constructed in step S4.1, construct the graduation thesis review result review prompt words.
[0043] S4.3: Set the number of "review-revision" iterations, n.
[0044] S4.4: Input the optimized thesis review result R′ generated in step S3 and the thesis review result review prompt words constructed in S4.2 into the thesis review result finalization agent.
[0045] S4.5: The graduation thesis review result finalization agent reviews the optimized graduation thesis review result R′ and generates internal review feedback G.
[0046] S4.6: Finalizing the Graduation Thesis Review Results. Based on the internal review feedback G, the intelligent agent revises the optimized graduation thesis review results R' generated in step S3, and generates an improved graduation thesis review report. .
[0047] S4.7: Stores the "Review-Revision" events of S4.5 and S4.6.
[0048] S4.8: The above "review-revision" cycle can be repeated multiple times until no new problems are found in the review stage, or the preset maximum number of iterations is reached.
[0049] The review stage can be represented by Formula 5: (Formula 5) in, The review strategy for the intelligent agent is determined by the evaluation results of the graduation thesis. For the first Internal review feedback generated in the next iteration.
[0050] The revision stage can be represented by Formula 6: (Formula 6) in, The revision strategy for the intelligent agent is determined by the evaluation results of the graduation thesis. For the first The result of the next iteration.
[0051] S4.9: Output the graduation thesis review report R″.
[0052] To implement the above method, the graduation thesis review system based on task decomposition and collaborative decision-making provided by the present invention includes a graduation thesis parsing agent, a graduation thesis review agent, a graduation thesis review result optimization agent, and a graduation thesis review result finalization agent. The four agents work together to complete the process from receiving the graduation thesis to generating the graduation thesis review report.
[0053] The thesis parsing agent receives the thesis D, parses it, extracts text data from the thesis, performs fine-grained information extraction based on VLM multimodal understanding for non-text data, and locates, segments, and identifies image regions to generate editable and structured text descriptions for image data. After the overall parsing is completed, all information is integrated into Markdown format and outputs a structured document tree H.
[0054] The thesis review agent receives the structured document tree H output by the thesis parsing agent, as well as the pre-constructed thesis review prompt framework F and thesis review prompt words; decomposes the review task according to the review indicators in the thesis review prompt framework, generates a review plan containing n steps, and completes each step in the review plan; after executing all steps, the thesis review result R is obtained.
[0055] The thesis review result optimization agent receives the thesis review result R generated by the thesis review agent; based on the multimodal context window, it performs key information extraction and performs multi-level content revision and expansion on the thesis review result R, including correcting judgments that do not conform to existing academic standards, identifying weaknesses in the argumentation logic, strengthening the correspondence between the comments and the original text and reference examples, and generating the optimized thesis review result R′.
[0056] The thesis review result finalization agent receives the optimized thesis review result R′ generated by the thesis review result optimization agent, as well as a pre-constructed review framework S and review prompts for the thesis review result; reviews the optimized thesis review result R′ according to the review framework S and generates internal review feedback G; revises the optimized thesis review result R′ according to the internal review feedback G until no new problems are found in the review stage, or the preset iteration limit is reached; and outputs the thesis review report R″.
[0057] The thesis parsing agent adopts a large vision-language model; the thesis review agent, the thesis review result optimization agent, and the thesis review result finalization agent adopt a large language model.
[0058] The following example, using a specific graduation thesis, illustrates the method for reviewing graduation theses according to this invention: S1. Obtain and parse a graduation thesis D, and generate a structured document tree H.
[0059] S2. Review the structured document tree H according to the graduation thesis review prompt framework F, and generate the graduation thesis review result R.
[0060] S3. Optimize R to generate the optimized thesis review result R′.
[0061] S4. Self-correct the optimized graduation thesis review results R′.
[0062] S5. Output the final version of the graduation thesis review report R″.
[0063] To verify the effectiveness of this invention, 121 graduation theses from a university were collected as a sample dataset. All samples underwent format standardization and content anonymization before being independently reviewed by both the present invention and experienced human experts.
[0064] In a preferred embodiment of the present invention, the thesis parsing agent adopts the Qwen2-5-vl-32b-32 vision-language large model; the thesis review agent and the thesis review result optimization agent adopt the Deepseek-r1-671b-64k-thinking large language model; the embedding model is BGE-M3; and the thesis review result finalization agent adopts the Qwen2-5-32b-128k large language model.
[0065] Of course, the thesis parsing agent can also use PaddleOCR-VL-1.5 or DeepSeek-OCR professional OCR models; the thesis review agent and the thesis review result optimization agent can also use QwQ-32b-128k-thinking or Qwen3-32b-128k-thinking large language models; the embedding model can also be Qwen3-Embedding-8B or BCE-Embedding-base_v1.
[0066] Evaluation indicators include the completion rate of review tasks, the consistency rate of review results, and the efficiency of review task completion.
[0067] A total of 103 valid review results were collected and included in the final consistency comparison and efficiency analysis.
[0068] The main test data and results are shown in Table 3: This invention successfully output review results for all 121 graduation theses, achieving a 100% completion rate; expert reviews were completed for 103 theses, achieving an 85.12% completion rate. Among the 103 valid comparison samples, the review conclusions of this invention were consistent with those of the expert reviews in 91 theses, resulting in a consistency rate of 88.35%. The average review time per thesis for this invention was 182.81 seconds (approximately 3.05 minutes), while the average review time for expert reviews was 20 minutes per thesis. Compared to expert reviews, this invention improved the review task completion efficiency by approximately 6.5 times.
[0069] Table 3 Main Test Results Experimental results show that the present invention possesses excellent task completion capabilities and processing efficiency in real-world scenarios. Its evaluation results are highly consistent with expert judgments, verifying the accuracy and reliability of the invention. Furthermore, the present invention significantly outperforms manual evaluation in terms of evaluation efficiency, demonstrating good engineering applicability and potential for widespread adoption.
[0070] Finally, it should be noted that the above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A graduation thesis review method based on task decomposition and collaborative decision-making, characterized in that: It includes the following steps: S1: Receive and parse graduation theses to generate a structured document tree; Receive graduation theses and input them into the graduation thesis parsing agent; The graduation thesis parsing agent parses the graduation thesis, extracts text data from the thesis, performs fine-grained information extraction based on VLM multimodal understanding for non-text data, and locates, segments and recognizes image regions to generate editable and structured text descriptions for image data. After the overall parsing is completed, all information is integrated into Markdown format and a structured document tree is output. S2: Review the structured document tree according to the graduation thesis review prompts framework and generate the graduation thesis review results; A framework for reviewing graduation theses specifically designed for thesis review tasks is constructed, which includes review tasks, review indicators, scoring criteria, output format, and available tools. Based on the constructed graduation thesis review prompt framework, graduation thesis review prompt words are constructed; The constructed graduation thesis review prompts and the structured document tree generated in step S1 are input into the graduation thesis review intelligent agent; The graduation thesis review agent decomposes the review task according to the review indicators and generates a review plan containing n steps. Then, complete each step in the review plan; after completing all steps, obtain the review results of the graduation thesis; S3: Optimize the generated graduation thesis review results to generate optimized graduation thesis review results; Extract the title and abstract of the graduation thesis, generate embedding vectors using an embedding model, and recall the academic materials with the highest semantic relevance by calculating the cosine similarity between the embedding vectors and the documents in the knowledge base, including relevant paragraphs, theoretical frameworks or typical writing paradigms. The recalled academic materials, graduation theses, and graduation thesis review results are organized together into a multimodal context window; The intelligent agent for optimizing the graduation thesis review results extracts key information based on context window input and performs multi-level content revision and expansion of the graduation thesis review results, including correcting judgments that do not conform to existing academic standards, identifying weaknesses in the argumentation logic, strengthening the correspondence between the comments and the original text and reference examples, and generating optimized graduation thesis review results. S4: Self-correct the optimized graduation thesis review results and output the final version of the graduation thesis review report; Construct a review framework specifically for the evaluation results of graduation theses, including structural integrity, language standardization, scoring accuracy, review objectivity, and review guidance; Based on the constructed framework for reviewing graduation thesis review results, prompts for reviewing graduation thesis review results are generated. The constructed graduation thesis review result review prompts and the optimized graduation thesis review results generated in step S3 are input into the graduation thesis review result finalization agent; The intelligent agent that determines the graduation thesis review results reviews the optimized graduation thesis review results according to the review framework and generates internal review feedback; it then revises the optimized graduation thesis review results based on the internal review feedback until no new problems are found in the review stage, or the preset iteration limit is reached. Output the graduation thesis review report.
2. The thesis review method based on task decomposition and collaborative decision-making according to claim 1, characterized in that: The analytical agent for the graduation thesis adopts a large vision-language model. The thesis review agent, the thesis review result optimization agent, and the thesis review result finalization agent all adopt a large language model. The embedding model is one of BGE-M3, Qwen3-Embedding-8B, or BCE-Embedding-base_v1.
3. The thesis review method based on task decomposition and collaborative decision-making according to claim 2, characterized in that: The large-scale vision-language model includes Qwen2-5-vl-32b-32k, PaddleOCR-VL-1.5, and DeepSeek-OCR; The large language model includes Deepseek-r1-671b-64k-thinking, QwQ-32b-128k-thinking, Qwen3-32b-128k-thinking, and Qwen2-5-32b-128k.
4. The thesis review method based on task decomposition and collaborative decision-making according to claim 1, characterized in that: The document types for the graduation thesis are .docx, .doc, and .pdf.
5. The thesis review method based on task decomposition and collaborative decision-making according to claim 1, characterized in that: The graduation thesis review prompt framework specifically includes: Reviewing tasks: Provide a clear and accurate task description; Evaluation metrics: Provide paper evaluation metrics and detailed explanations to ensure the model accurately understands the metrics content; Scoring criteria: Provides indicator values and scoring basis; Output format: Provide output examples and standardize the output format; Available tools: Provides information on available tools and their uses, guiding the model to determine when and which tool to use.
6. The thesis review method based on task decomposition and collaborative decision-making according to claim 1, characterized in that: The specific content of the review framework specifically for graduation thesis evaluation results is as follows: Structural integrity: Review whether the review report covers all the components of the standard template, including: basic information about the paper, review details, review results, and specific suggestions for improvement; Language standardization: Review the academic standardization of the language used, including whether the terminology is used accurately, whether the sentences are clear and fluent, whether the logic is coherent, and whether there are any grammatical errors or inconsistent formats. Accuracy of scoring: Review whether the scoring process of the review report is rigorous and reliable. Check whether the scores for each item are strictly based on the established scoring criteria, whether the scores for each item are logically consistent with the corresponding comments, and whether the total score is calculated accurately. Objectivity of review: whether the review comments are based on the quality and evidence of the paper itself, whether there are subjective assumptions or emotional expressions, and whether objective evidence is provided when pointing out shortcomings; Review guidance: Review whether the overall review conclusion is specific and clear, whether the review comments are targeted, and whether the improvement suggestions are actionable.
7. A system for implementing the thesis review method based on task decomposition and collaborative decision-making as described in any one of claims 1-6, characterized in that: It includes a graduation thesis parsing agent, a graduation thesis review agent, a graduation thesis review result optimization agent, and a graduation thesis review result finalization agent; The thesis parsing agent receives the thesis, parses it, extracts text data, performs fine-grained information extraction based on VLM multimodal understanding for non-text data, and locates, segments, and identifies image regions to generate editable and structured text descriptions for image data. After the overall parsing is completed, all information is integrated into Markdown format and a structured document tree is output. The thesis review agent receives the structured document tree output by the thesis parsing agent, as well as the pre-built thesis review prompt framework and thesis review prompt words; decomposes the review task according to the review indicators in the thesis review prompt framework, generates a review plan containing n steps, and completes the steps in the review plan one by one; after executing all steps, the thesis review result is obtained. The graduation thesis review result optimization agent: receives the graduation thesis review results generated by the graduation thesis review agent; Based on a multimodal context window, key information is extracted, and the graduation thesis review results are revised and expanded at multiple levels. This includes correcting judgments that do not conform to existing academic standards, identifying weaknesses in the argumentation logic, strengthening the correspondence between the comments and the original text and reference examples, and generating optimized graduation thesis review results. The thesis review result finalization agent receives the optimized thesis review results generated by the thesis review result optimization agent, as well as a pre-constructed review framework and review prompts for the thesis review results; reviews the optimized thesis review results according to the review framework and generates internal review feedback; and revises the optimized thesis review results based on the internal review feedback until no new problems are found during the review stage, or the preset iteration limit is reached. Output the graduation thesis review report.
8. The graduation thesis review system based on task decomposition and collaborative decision-making according to claim 7, characterized in that: The graduation thesis review prompt framework includes: Reviewing tasks: Provide a clear and accurate task description; Evaluation metrics: Provide paper evaluation metrics and detailed explanations to ensure the model accurately understands the metrics content; Scoring criteria: Provides indicator values and scoring basis; Output format: Provide output examples and standardize the output format; Available tools: Provide information on available tools and their uses, guiding the model to determine when and which tool to use; The review framework specifically for graduation thesis evaluation results includes the following: Structural integrity: Review whether the review report covers all the components of the standard template, including: basic information about the paper, review details, review results, and specific suggestions for improvement; Language standardization: Review the academic standardization of the language used, including whether the terminology is used accurately, whether the sentences are clear and fluent, whether the logic is coherent, and whether there are any grammatical errors or inconsistent formats. Accuracy of scoring: Review whether the scoring process of the review report is rigorous and reliable. Check whether the scores for each item are strictly based on the established scoring criteria, whether the scores for each item are logically consistent with the corresponding comments, and whether the total score is calculated accurately. Objectivity of review: whether the review comments are based on the quality and evidence of the paper itself, whether there are subjective assumptions or emotional expressions, and whether objective evidence is provided when pointing out shortcomings; Review guidance: Review whether the overall review conclusion is specific and clear, whether the review comments are targeted, and whether the improvement suggestions are actionable.
9. The graduation thesis review system based on task decomposition and collaborative decision-making according to claim 8, characterized in that: The multimodal context window consists of the recalled academic materials, graduation thesis, and graduation thesis review results; The academic materials to be recalled are obtained by extracting the titles and abstracts of graduation theses, generating embedding vectors using an embedding model, and then calculating the cosine similarity between the embedding vectors and the documents in the knowledge base to recall the academic materials with the highest semantic relevance, including relevant paragraphs, theoretical frameworks, or typical writing paradigms.
10. The graduation thesis review system based on task decomposition and collaborative decision-making according to claim 7, characterized in that: The thesis parsing agent adopts a large vision-language model; the thesis review agent, the thesis review result optimization agent, and the thesis review result finalization agent adopt a large language model. The large-scale vision-language model includes Qwen2-5-vl-32b-32k, PaddleOCR-VL-1.5, and DeepSeek-OCR; The large language model includes Deepseek-r1-671b-64k-thinking, QwQ-32b-128k-thinking, Qwen3-32b-128k-thinking, and Qwen2-5-32b-128k. The embedding model is one of BGE-M3, Qwen3-Embedding-8B, or BCE-Embedding-base_v1.