Method and system for constructing manuscript review agent embedded in gradient evaluation system

Abstract

The invention provides a manuscript review agent construction method and system embedded with a gradient evaluation system, and relates to the technical field of agents, and the method comprises the steps: pre-storing a constructed evaluation dimension knowledge base in a memory layer; activating a progressive review model in the plan layer; performing fine adjustment on a first reference model in the multi-layer progressive reference model through a predetermined tool set in the tool layer to obtain a first review model; storing the first manuscript review model to an execution layer, and performing manuscript review execution on the target paper through the execution layer to obtain a target manuscript review opinion; and the activation feedback layer analyzes the target manuscript review suggestions to obtain target modification suggestions. According to the method and the device, the technical problems that in the prior art, manuscript evaluation mainly depends on artificial experience judgment, so that a manuscript evaluation result has high subjectivity, machine manuscript evaluation is mostly limited to form review, and deep content evaluation is difficult to carry out can be solved; and objectivity and systematicness of manuscript review evaluation are improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent agent technology, and in particular to a method and system for constructing an intelligent review agent embedded in a gradient evaluation system. Background Art

[0002] Traditional review methods rely primarily on expert reviewers to evaluate papers based on their personal knowledge and experience. While expert reviewers typically possess extensive domain knowledge and experience, this process is subject to significant subjectivity and inconsistency. In practice, different experts may use significantly different criteria for the same paper, and review cycles are generally long, leading to a backlog of review tasks. This heavy reliance on human experience makes it difficult to ensure a standardized and consistent review process, impacting both quality and efficiency.

[0003] With the development of machine-assisted review technology, it has been applied to the formal review level of papers, such as grammar checking, format specifications, citation format, etc. However, the existing machine review still fails to achieve an in-depth assessment of the content of the paper and can only deal with surface-level issues. It lacks in-depth analysis of the academic value, innovation, research methods and results analysis of the paper.

[0004] In summary, there are technical problems in the existing technology: the review evaluation mainly relies on manual experience and judgment, resulting in a strong subjectivity in the review results, and the machine review is mostly limited to formal review, making it difficult to conduct in-depth content evaluation. Summary of the Invention

[0005] The present application provides a method and system for constructing an intelligent reviewer embedded in a gradient evaluation system, which is used to solve the technical problems in the prior art that the review evaluation mainly relies on manual experience and judgment, resulting in strong subjectivity in the review results, and that machine review is mostly limited to formal review and is difficult to conduct in-depth content evaluation.

[0006] In view of the above problems, this application provides a method and system for constructing a review agent embedded in a gradient evaluation system.

[0007] In the first aspect, the present application provides a method for constructing a review agent embedded in a gradient evaluation system, wherein the method for constructing a review agent embedded in a gradient evaluation system is implemented by a review agent construction system embedded in a gradient evaluation system, wherein the method for constructing a review agent embedded in a gradient evaluation system includes: pre-storing the constructed evaluation dimension knowledge base to the memory layer; activating the progressive review model in the planning layer, wherein the progressive review model includes a multi-layer progressive benchmark model; fine-tuning the first benchmark model in the multi-layer progressive benchmark model through a predetermined tool set in the tool layer to obtain a first review model; storing the first review model to the execution layer, and performing review on the target paper through the execution layer to obtain a target review opinion; activating the feedback layer to analyze the target review opinion to obtain a target modification suggestion.

[0008] Optionally, a review evaluation dimension set for review papers is formed; an academic evaluation dimension set for academic papers is formed; and the evaluation dimension knowledge base is constructed based on a first correspondence between the review papers and the review evaluation dimension set, and a second correspondence between the academic papers and the academic evaluation dimension set; wherein the review evaluation dimension set includes at least topic clarity, author research foundation, innovative differences, literature completeness and writing standardization, and the academic evaluation dimension set includes at least the frontier and innovative evaluation of the topic, the integrity of the basic structure of the paper, the standardization of the experimental part and the academic standardization.

[0009] Optionally, the multi-layer progressive benchmark model includes an 8B model, a 32B model and a 235B model.

[0010] Optionally, the first benchmark model is adaptively fine-tuned to obtain the first manuscript review model, which includes: if the first benchmark model is the 8B model, constructing a basic format check instruction fine-tuning dataset in combination with the predetermined tool set; fine-tuning the basic format check instruction dataset and fine-tuning the first benchmark model using the QLoRA method to obtain the first manuscript review model; if the first benchmark model is the 32B model, constructing an experimental plan evaluation instruction fine-tuning dataset in combination with the predetermined tool set; fine-tuning the experimental plan evaluation instruction dataset and fine-tuning the first benchmark model using the QLoRA method to obtain the first manuscript review model; if the first benchmark model is the 235B model, constructing an innovative analysis instruction fine-tuning dataset in combination with the predetermined tool set; fine-tuning the innovative analysis instruction dataset and fine-tuning the first benchmark model using the QLoRA method to obtain the first manuscript review model.

[0011] Optionally, the 8B model is used to evaluate predetermined basic indicators of the target paper, wherein the predetermined basic indicators include at least format standardization, chapter completeness and citation standardization.

[0012] Optionally, the 32B model is used to evaluate predetermined experimental indicators of the target paper, wherein the predetermined experimental indicators at least include the rationality of the experimental plan and the standardization of data analysis.

[0013] Optionally, the 235B model is used to evaluate predetermined innovation indicators of the target paper.

[0014] Optionally, the target type of the target paper is obtained; the predetermined innovation indicators are adjusted in combination with the target type to obtain target innovation indicators, which include: if the target type is a review type, the predetermined innovation indicators include a first predetermined interdisciplinary fusion indicator and a first predetermined industry pain point solution indicator; if the target type is an academic type, the predetermined innovation indicators include a second predetermined interdisciplinary fusion indicator and a second predetermined industry pain point solution indicator; wherein, the first predetermined interdisciplinary fusion indicator includes theoretical integration degree, method reference innovation, and interdisciplinary research trend insight, and the first predetermined industry pain point solution indicator includes accurate identification of pain points, feasibility of solutions, and industry influence assessment; wherein, the second predetermined interdisciplinary fusion indicator includes disciplinary cross-depth, interdisciplinary team collaboration, and interdisciplinary achievement transformation, and the second predetermined industry pain point solution indicator includes targeted research on pain points, technological breakthroughs and application prospects, and industry collaborative innovation mechanisms.

[0015] Optionally, a target entity set of the target paper is extracted, and a target knowledge graph is constructed based on the target entity set; an arbitrary paper is extracted, and the arbitrary paper corresponds to an arbitrary knowledge graph; when the similarity between the target knowledge graph and the arbitrary knowledge graph reaches a predetermined similarity threshold, the arbitrary paper is added to an originality evaluation reference list; and the target paper is evaluated for originality according to the originality evaluation reference list.

[0016] In the second aspect, the present application also provides a review agent construction system embedded in a gradient evaluation system, which is used to execute a review agent construction method embedded in a gradient evaluation system as described in the first aspect, wherein the review agent construction system embedded in a gradient evaluation system includes: an evaluation pre-storage module, which is used to pre-store the constructed evaluation dimension knowledge base to the memory layer; a model activation module, which is used to activate the progressive review model in the planning layer, wherein the progressive review model includes a multi-layer progressive benchmark model; a model fine-tuning module, which is used to fine-tune the first benchmark model in the multi-layer progressive benchmark model through a predetermined tool set in the tool layer to obtain a first review model; an execution review module, which is used to store the first review model to the execution layer, and perform review on the target paper through the execution layer to obtain a target review opinion; an opinion analysis module, which is used to activate the feedback layer to analyze the target review opinion to obtain a target modification suggestion.

[0017] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0018] By pre-storing the constructed evaluation dimension knowledge base in the memory layer; activating the progressive review model in the planning layer, wherein the progressive review model includes a multi-layer progressive benchmark model; fine-tuning the first benchmark model in the multi-layer progressive benchmark model through the predetermined tool set in the tool layer to obtain the first review model; storing the first review model in the execution layer, and performing the review on the target paper through the execution layer to obtain the target review opinion; activating the feedback layer to analyze the target review opinion to obtain the target modification suggestion. In other words, through the multi-layer progressive review model and standardized evaluation system, the paper is reviewed according to unified standards, and specific modification suggestions are provided based on the review results, which improves the objectivity and systematicness of the review evaluation and enhances the interpretability of the review evaluation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in this application or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely exemplary, and a person of ordinary skill in the art can obtain other drawings based on the provided drawings without creative work.

[0020] Figure 1 This is a flowchart of a method for constructing a review agent embedded in a gradient evaluation system for this application.

[0021] Figure 2 This is a structural diagram of a review agent construction system embedded in a gradient evaluation system for this application.

[0022] Explanation of the reference numerals: evaluation pre-storage module 11 , model activation module 12 , model fine-tuning module 13 , execution review module 14 , opinion analysis module 15 . DETAILED DESCRIPTION

[0023] This application provides a method and system for constructing a review agent embedded in a gradient evaluation system to solve the technical problems in the existing technology, such as the fact that the review evaluation mainly relies on manual experience judgment, resulting in a strong subjectivity in the review results, and the machine review is mostly limited to formal review, making it difficult to conduct in-depth content evaluation. Through a multi-layer progressive review model and a standardized evaluation system, papers are reviewed according to unified standards, and specific modification suggestions are provided based on the review results, which improves the objectivity and systematicness of the review evaluation and enhances the interpretability of the review evaluation results.

[0024] Example 1, as Figure 1 As shown, the present application provides a method for constructing a review agent embedded in a gradient evaluation system, wherein the method for constructing a review agent embedded in a gradient evaluation system is applied to a system for constructing a review agent embedded in a gradient evaluation system, and the method for constructing a review agent embedded in a gradient evaluation system specifically includes the following steps:

[0025] S100: Pre-store the constructed evaluation dimension knowledge base into the memory layer.

[0026] Furthermore, the present application S100 includes:

[0027] A review evaluation dimension set for review papers is formed; an academic evaluation dimension set for academic papers is formed; and the evaluation dimension knowledge base is constructed based on a first correspondence between the review papers and the review evaluation dimension set, and a second correspondence between the academic papers and the academic evaluation dimension set; wherein, the review evaluation dimension set includes at least the clarity of the subject, the author's research foundation, the difference in innovation, the completeness of the literature, and the standardization of writing; and the academic evaluation dimension set includes at least the evaluation of the cutting-edge and innovative nature of the topic, the integrity of the basic structure of the paper, the standardization of the experimental part, and the academic standardization.

[0028] Specifically, the review agent's memory layer stores a knowledge base of evaluation dimensions for review and academic papers. This includes a set of review evaluation dimensions for review papers and a set of academic evaluation dimensions for academic papers. The evaluation dimensions for review papers focus on five core aspects: topic clarity, author research foundation, innovative differentiation, completeness of the literature, and standard writing. Topic clarity requires that the review paper have a clear structural framework and readability, that is, whether it has a clear theme and whether it can effectively organize the literature to help readers understand the core issues in the field; the author's research basis is supported by his or her relevant achievements in the field, that is, the author's research experience and background in the relevant field are evaluated to ensure that the review has sufficient academic depth and authority; innovation and difference require comparative analysis with published reviews on the same topic, that is, comparison with similar published reviews to judge the innovation and difference of this review; literature completeness emphasizes the rationality of the search scope, time span and selection rules, that is, evaluation of whether the cited literature is extensive and comprehensive, especially the proportion of literature in the past 3-5 years should exceed 50%, and cover authoritative literature in the field; writing standardization requires in-depth analysis of the logical chain of "research subject-method theory-problem solving-future direction" to avoid simple abstract piling up.

[0029] In the evaluation dimension of academic papers, a full-process evaluation system has been established, from topic selection to data verification. First, attention is paid to the cutting-edge and innovative nature of the topic, avoiding outdated issues or outdated methods, and judging whether it can promote further research in the field; second, the basic structural integrity of the paper is verified to see whether it meets the conventional requirements of academic papers, ensuring the logical integrity of elements such as the introduction, research status, research framework, experimental process, data analysis, and conclusions; the focus is on evaluating the standardization of the experimental part, including the reliability of the data source, the completeness of the experimental steps, the rationality of the comparative experiments, and the necessity of the ablation experiments; at the same time, attention is paid to academic standardization, and the reference annotations of formulas, methods, and diagrams are strictly reviewed to ensure the accuracy of the expression of innovative points and the consistency of the data.

[0030] Based on the evaluation dimension sets for review papers and academic papers, we establish a first correspondence and a second correspondence between the two dimension sets. The first correspondence is used to match review papers with the review evaluation dimension set, while the second correspondence is used to match academic papers with the academic evaluation dimension set. By combining the correspondences between the evaluation dimension sets for review papers and academic papers, we form a complete evaluation dimension knowledge base that includes all dimensions and standards used to review and academic papers, enabling paper evaluation based on these standards.

[0031] The constructed evaluation dimension knowledge base is stored in the memory layer, which serves as the long-term data storage area of ​​the review agent. This ensures that the evaluation criteria and knowledge base are readily available during the review process. By building and storing the evaluation dimension knowledge base, papers can be comprehensively reviewed according to predetermined criteria, reducing subjective bias in manual review and improving review consistency. The evaluation dimension knowledge base goes beyond surface inspections and can delve deeper into various aspects of the paper, such as structure, content innovation, and experimental design, enhancing the depth and accuracy of the review.

[0032] S200: Activate a progressive review model in a planning layer, wherein the progressive review model includes a multi-layer progressive benchmark model.

[0033] Furthermore, the present application S200 includes:

[0034] The multi-layer progressive benchmark model includes an 8B model, a 32B model and a 235B model.

[0035] Specifically, within the review agent, the planning layer is responsible for coordinating and scheduling the review process. By setting review strategies and processes, it ensures the smooth progress of each step. When the review process begins, the progressive review model within the planning layer is activated. This progressive review model is divided into multiple levels, each using a different benchmark model to gradually review the paper. Each benchmark model focuses on different aspects of the paper, deepening the review layer to ensure comprehensiveness.

[0036] The multi-layer progressive benchmark model is the core component of the progressive review model. It includes multiple model levels, each responsible for different dimensions and depths of review, including the 8B model, the 32B model, and the 235B model. The 8B model is mainly used to conduct preliminary checks on the basic format, structure, and content of the paper. It checks the external specifications of the paper, such as the completeness of the chapters and whether the format meets the requirements. The 32B model is a mid-level review model, which focuses on evaluating technical content such as the experimental plan and data analysis standardization in the paper, and checks whether the technical implementation of the paper in academic research is reasonable. The 235B model is a deep review model that deeply analyzes key elements such as the innovation, academic contribution, and research limitations of the paper to evaluate the academic value and contribution of the paper.

[0037] In the initial review phase, papers are categorized according to their attributes (e.g., review type, research type, etc.) and intelligently matched to corresponding evaluation dimension templates. Different types of papers require different evaluation criteria, which helps to improve the targeted nature of the review. In the shallow review phase, the 8B model is deployed to complete preliminary checks on basic formatting and chapter completeness. In the mid-level review phase, the 32B model focuses on evaluating technical content such as the rationality of the experimental plan and the standardization of data analysis. In the deep review phase, the 235B model is used to conduct an in-depth analysis of key elements such as the paper's innovation, academic contribution, and research limitations. Next, the verification and filtering phase focuses on identifying and filtering inappropriate content to ensure the professionalism and standardization of the review comments. Inappropriate comments are checked to ensure that the generated review comments are constructive and academically sound. Finally, in the comprehensive evaluation phase, the 235B model integrates all review results from the previous phases to generate a multi-dimensional quantitative score. Based on the scores for each dimension, graded revision suggestions are generated. This comprehensive evaluation ensures that the review comments are comprehensive and objective, and that specific revision suggestions are provided to address specific issues in the paper.

[0038] The progressive review model ensures that papers are comprehensively evaluated from multiple perspectives through a step-by-step, in-depth review process, reducing the omissions and inefficiencies common in manual review. Each benchmark model (8B, 32B, and 235B) focuses on different review dimensions, ensuring accurate evaluation of all aspects of a paper, from format and technical content to academic contributions.

[0039] S300: Fine-tune the first benchmark model in the multi-layer progressive benchmark model using a predetermined tool set in the tool layer to obtain a first manuscript review model.

[0040] Furthermore, the present application S300 includes:

[0041] The first benchmark model is adaptively fine-tuned to obtain the first manuscript review model, which includes: if the first benchmark model is the 8B model, a basic format check instruction fine-tuning dataset is constructed in combination with the predetermined tool set; according to the basic format check instruction fine-tuning dataset, the first benchmark model is fine-tuned using the QLoRA method to obtain the first manuscript review model; if the first benchmark model is the 32B model, an experimental scheme evaluation instruction fine-tuning dataset is constructed in combination with the predetermined tool set; according to the experimental scheme evaluation instruction fine-tuning dataset, the first benchmark model is fine-tuned using the QLoRA method to obtain the first manuscript review model; if the first benchmark model is the 235B model, an innovative analysis instruction fine-tuning dataset is constructed in combination with the predetermined tool set; according to the innovative analysis instruction fine-tuning dataset, the first benchmark model is fine-tuned using the QLoRA method to obtain the first manuscript review model.

[0042] The 8B model is used to evaluate predetermined basic indicators of the target paper, wherein the predetermined basic indicators at least include format standardization, chapter completeness and citation standardization.

[0043] The 32B model is used to evaluate the predetermined experimental indicators of the target paper, wherein the predetermined experimental indicators at least include the rationality of the experimental plan and the standardization of data analysis.

[0044] The 235B model is used to evaluate the predetermined innovation indicators of the target paper.

[0045] Specifically, the tool layer provides specific technical tools and methods to support tasks such as fine-tuning, execution, and analysis of the review agent. This layer includes a predefined tool set to help the review agent optimize and adjust to ensure efficient review. This predefined tool set is a collection of specialized tools for model fine-tuning and execution, including algorithms for formatting checks, data analysis methods, and text processing tools, ensuring the model can efficiently complete its assigned tasks.

[0046] When the first benchmark model is the 8B model, the predefined tool set in the tool layer is used to fine-tune the dataset and the 8B model in conjunction with basic formatting instructions to ensure that the 8B model can better handle issues related to paper formatting, such as chapter completeness and proper citation formatting. The 8B model is the first benchmark model in the review process, responsible for basic formatting checks, primarily assessing basic standards such as formatting standardization, chapter completeness, and citation standardization. It serves as the initial assessment model in the review process, ensuring that the paper's format meets basic academic requirements.

[0047] Combined with the predetermined tool set, a basic format check instruction fine-tuning dataset is constructed, which contains various data items related to the basic format of the paper, such as chapter title format, reference format, paragraph typesetting, chart specifications, etc. The 8B model is fine-tuned through the basic format check instruction fine-tuning dataset combined with the QLoRA method. QLoRA is an efficient large language model fine-tuning method that significantly reduces memory usage through 4-bit quantization and low-rank adaptation technology. The core idea is to quantize the pre-trained model parameters into 4-bit format for storage, and at the same time add a low-rank adaptation matrix for parameter update during the training process. The hyperparameter settings are as follows: batch_size: int = 20, micro_batch_size: int = 2, num_epochs: int = 3, learning_rate: float = 1e -5, lora_r: int = 8, lora_alpha: int = 16, lora_dropout: float = 0.05. For the i-th layer of the model, QLoRA first converts the original weight matrix W i Quantized to 4-bit format, two projection matrices A are defined at the same time i and B i , the dimensions are (d, r) and (r, d) respectively, where d is the hidden layer dimension of the model, r is the projection dimension, and r<<d.

[0048] During forward propagation, QLoRA dequantizes the quantized weights and performs calculations, and adds a correction term based on the projection matrix. Assume that the original forward calculation of layer i can be expressed as: h i =f i (x i ), usually f i is the i-th layer in the representation model, x i is the data input to this layer, h i is the output of the i-th layer, which can then be used as the input of the next layer, or the output of the entire model. In QLoRA, the modified forward calculation formula is: in, Indicates dequantizing the 4-bit weight to a floating point number, Δ i =A i B i f i (x i ) represents the low-rank correction term introduced by QLoRA. The optimization goal of QLoRA is to minimize the loss function of the corrected model on the new task: Among them, θ 4bit represents the original model parameters after quantization, represents all the projection matrices introduced by QLoRA, D is the training dataset of the new task, and l is the task-related loss function. During the optimization process, only update While keeping θ 4bit constant.

[0049] Compared to LoRA, QLoRA further reduces memory usage through 4-bit quantization, enabling efficient fine-tuning of large models even with limited video memory. Furthermore, by maintaining the low-rank update feature, the number of additional parameters introduced by QLoRA is still far smaller than the original model.

[0050] QLoRA utilizes quantization and low-rank adaptation techniques to make the fine-tuning process more efficient, reducing computational resource consumption while maintaining model performance. During fine-tuning, the model adjusts parameters to more accurately assess the basic formatting of papers in subsequent review tasks. After adaptive fine-tuning, the 8B model becomes the first-in-line review model, equipped with enhanced basic formatting capabilities, capable of more accurately detecting essential standards such as formatting compliance, section completeness, and citation integrity.

[0051] All instruction fine-tuning datasets (i.e., the basic format check instruction fine-tuning dataset, the experimental plan evaluation instruction fine-tuning dataset, and the innovative analysis instruction fine-tuning dataset) use the triple form of {instruction, input, output}. The instruction portion includes a clear task definition and supplements the specific description of the evaluation dimensions, providing clear evaluation boundaries and core indicators. In the input field, the dataset contains a large number of paragraph samples from scientific papers, covering various writing characteristics in academic papers. In the output field, different evaluation tasks use different output methods. For example, the basic format check uses a structured scoring table ("format_score": 8.5, "structure_score": 7.8) as its output to ensure the quantification of the evaluation; the experimental plan evaluation task uses the output form of "Evaluation Dimension: Detailed Analysis" to achieve multi-angle solution evaluation; the output of the innovative analysis task adopts a multi-dimensional evaluation form, covering three levels of theoretical innovation, methodological innovation, and application innovation, comprehensively reflecting the innovative characteristics of the paper.

[0052] For example, for the basic format check instruction fine-tuning dataset, instruction: "Evaluate the format and structure of the following paper sections, assess whether they meet academic writing standards, and output scores for format and structure"; input: "The experiment was conducted using standard laboratory equipment, the data was collected within six months, and statistical methods were used for analysis"; output: {"format_score":8.5,"structure_score":7.8; Suggestion: Add specific details of the equipment and formulate the statistical methods used}. The results of fine-tuning the basic format check are as follows: the accuracy of the citation format is 0.934, the precision is 0.928, the recall is 0.925, and the F1 is 0.926; the accuracy of the paragraph structure is 0.958, the precision is 0.955, the recall is 0.952, and the F1 is 0.953; the accuracy of the reference format is 0.927, the precision is 0.922, the recall is 0.918, and the F1 is 0.920; and the accuracy of the figure format is 0.952, the precision is 0.948, the recall is 0.945, and the F1 is 0.946.

[0053] The 32B model is used to evaluate the experimental part of the target paper, including the rationality of the experimental plan in the paper and the standardization of data analysis. For example, whether the experimental design is reasonable, whether the data collection is scientific, whether the experimental steps are complete, and whether the data analysis is standardized. It is an important tool for mid-level review.

[0054] When the first benchmark model is the 32B model, a predefined toolset is used to construct a dataset for fine-tuning experimental plan evaluation instructions. This dataset contains content related to the paper's experimental plan, primarily focusing on the rationality of the experimental design, the standardization of experimental procedures, and data collection and analysis methods. Using this dataset, the 32B model is fine-tuned using the QLoRA method. Similar to the previous steps, QLoRA enables the model to be optimized for experimental plan evaluation tasks using fewer computing resources. This fine-tuning process makes the model more suitable for evaluating experimental rationality and data analysis standardization, improving its accuracy and efficiency.

[0055] After fine-tuning, the 32B model becomes the first-review model. This model has been optimized based on the experimental protocol evaluation dataset, enabling in-depth analysis and evaluation of the experimental portion of a paper during the subsequent review process. In actual review tasks, the first-review model is applied to the experimental portion of a paper, assessing the rationality of the experimental design and compliance with standard data analysis procedures. The first-review model checks the completeness of the experimental procedures, the rigor of the data analysis, and the adequacy of the experimental conclusions in the paper's argument.

[0056] For example, for the experimental protocol evaluation instruction fine-tuning the dataset, instruction: "Analyze the experimental design and methods of the following study parts to evaluate their completeness, validity and scientific rigor"; input: "1000 protein samples were collected and a deep learning model was applied, trained on 80% of the data and tested on the remaining 20%. Cross-validation was performed using 5-fold validation"; output: methodological completeness: {highly clear description of data splitting and validation methods; sample_size_validity: sufficient, 1000 samples are statistically significant; validation method: strong, cross-validation is used correctly; recommendation: consider adding control group details}. The results of fine-tuning the experimental plan evaluation showed that the accuracy of the experimental design completeness was 0.943, the precision was 0.935, the recall was 0.928, and the F1 was 0.931; the accuracy of the experimental data scale was 0.952, the precision was 0.945, the recall was 0.938, and the F1 was 0.941; the accuracy of the comprehensiveness of the evaluation indicators was 0.937, the precision was 0.925, the recall was 0.91, and the F1 was 0.922; and the accuracy of the model theoretical rationality was 0.958, the precision was 0.953, the recall was 0.947, and the F1 was 0.950.

[0057] The 235B model is designed to conduct an in-depth analysis of the innovation of target papers. Unlike the 8B and 32B models, which primarily focus on format and experimental sections, the 235B model focuses on the paper's innovation, assessing its degree of innovation in theory, method, technology, or academic contribution.

[0058] When the first baseline model is the 235B model, a dataset of innovation analysis instructions is constructed based on the characteristics of the innovative parts of the paper. This dataset includes examples of innovative content in the paper, such as descriptions of new theories, new methods, or technological breakthroughs in the paper. The 235B model is fine-tuned using QLoRA technology. Through QLoRA, the model can quickly adapt to innovation analysis tasks with limited computing resources, improving its performance in innovation assessment. The fine-tuned model will focus more on evaluating the innovative content of the paper, such as novel theoretical frameworks, innovative research methods, and technological innovations. After fine-tuning, the 235B model is transformed into a first-line review model, capable of performing in-depth analysis of the paper's innovation and evaluating whether the paper has made innovative contributions to academia or industry.

[0059] For example, for the innovative analysis instruction fine-tuning dataset, instruction: "Analyze the innovative aspects of the following research content. Evaluate theoretical innovation, methodological innovation and practical application value"; input: "A new hybrid converter architecture is introduced that combines quantum computing principles with traditional deep learning. Compared with the most advanced models, this method reduces the computational complexity from O(n 2 ) to O(nlogn), while improving the accuracy by 15%, and successfully applied"; output: {Theoretical Innovation: Score: 9.2, Analysis: Innovative fusion of quantum principles and Transformer architecture, Progress: Major theoretical breakthrough in computing efficiency; Methodological Innovation: Score: 8.8, Analysis: Unique hybrid architecture design, Improvement: Achieved a 15% accuracy improvement while reducing complexity; Application Innovation: Score: 8.5, Analysis: Successfully applied in practical scenarios, Impact: Demonstrated practical application value; Comprehensive Evaluation: High innovation value in all aspects}.

[0060] By fine-tuning each baseline model, we ensure that each model focuses on its specific task, improving the accuracy of paper reviews. Automated fine-tuning and review processes can improve review efficiency, reduce manual intervention, and make the review process more efficient, standardized, and systematic.

[0061] Furthermore, the present application further comprises the following steps:

[0062] Obtain the target type of the target paper; adjust the predetermined innovation indicators in combination with the target type to obtain target innovation indicators, which include: if the target type is a review type, the predetermined innovation indicators include a first predetermined interdisciplinary integration indicator and a first predetermined industry pain point solution indicator; if the target type is an academic type, the predetermined innovation indicators include a second predetermined interdisciplinary integration indicator and a second predetermined industry pain point solution indicator; wherein, the first predetermined interdisciplinary integration indicator includes theoretical integration degree, method reference innovation, and interdisciplinary research trend insight, and the first predetermined industry pain point solution indicator includes accurate pain point identification, solution feasibility, and industry influence evaluation; wherein, the second predetermined interdisciplinary integration indicator includes interdisciplinary depth, interdisciplinary team collaboration, and interdisciplinary achievement transformation, and the second predetermined industry pain point solution indicator includes targeted pain point research, technological breakthroughs and application prospects, and industry collaborative innovation mechanism.

[0063] Specifically, the target type of the target paper is determined, categorized into review-type and academic-type papers, and the predetermined innovation indicators are adjusted based on the target type to obtain the target innovation indicators. If the paper is a review-type paper, the predetermined innovation indicators will include the first predetermined interdisciplinary integration indicator and the first predetermined industry pain point solution indicator; if the paper is an academic-type paper, the predetermined innovation indicators will include the second predetermined interdisciplinary integration indicator and the second predetermined industry pain point solution indicator.

[0064] Among them, the first predetermined interdisciplinary integration indicator of the predetermined innovation indicators of review-type papers includes theoretical integration (evaluating whether the paper can effectively integrate theories and knowledge from different disciplines), methodological reference innovation (whether the paper innovatively references and integrates methods from different disciplines), and interdisciplinary research trend insights (whether the paper can identify and foresee the latest development trends in interdisciplinary research). The first predetermined industry pain point solution indicator of the predetermined innovation indicators of review-type papers includes accurate pain point identification (whether the paper can accurately identify and define the key pain points in the industry), solution feasibility (whether the solution proposed in the paper is feasible and can effectively solve industry pain points), and industry impact assessment (assessing the actual impact and driving role of the research results of the paper on the industry).

[0065] The second predetermined interdisciplinary integration indicator of the predetermined innovation indicators of academic papers includes the depth of interdisciplinary intersection (assessing the depth of intersection and degree of integration between different disciplines in the paper), interdisciplinary team collaboration (assessing the interdisciplinary team collaboration in the paper and whether the team's collaborative work can promote research innovation), and interdisciplinary results transformation (assessing whether the research results in the paper can be transformed and applied between different disciplines). The second predetermined industry pain point solution indicator of the predetermined innovation indicators of academic papers includes targeted pain point research (whether the paper focuses on the core issues of the industry and proposes targeted research and solutions), technological breakthroughs and application prospects (whether the paper demonstrates technological breakthroughs and predicts its application prospects), and industry collaborative innovation mechanisms (whether the paper can propose collaborative innovation mechanisms within the industry to promote industry cooperation and technological innovation).

[0066] Based on the target paper type, the predetermined innovation index is adjusted to obtain the target innovation index, ensuring that the evaluation criteria for paper innovation are consistent with the paper type. By automatically adjusting the innovation index based on the paper type, the paper's innovation is accurately assessed, avoiding the generalization or inapplicability of the evaluation criteria.

[0067] S400: Storing the first review model in an execution layer, and performing a review on a target paper through the execution layer to obtain a target review opinion.

[0068] Specifically, the first draft model is stored in the execution layer. The execution layer not only stores the model itself, but also provides an execution environment for the model to ensure that the model can perform relevant tasks according to the needs of the paper. The execution layer will activate and review the paper to be reviewed through the model. The execution layer first uses the text classification model to intelligently identify the paper type, divide the article into different categories such as review type and research type, and automatically matches the corresponding evaluation dimension template accordingly. After completing the type identification and dimension matching, the gradient enhancement evaluation process begins. During the initialization execution phase, the 8B model is responsible for basic evaluation tasks, mainly reviewing the format standardization, chapter completeness and citation specifications of the paper to ensure that the paper meets basic academic standards.

[0069] Next, the mid-level implementation phase begins, where the 32B model focuses on the technical content and experimental design of the paper. This phase primarily assesses the rationale of the experimental plan, including the completeness of experimental procedures, standardization of data collection, and the design of control groups. Data analysis methods are also reviewed, assessing the adequacy of the data set and the appropriateness of the statistical methods, and verifying the reliability of the experimental results. For example, in the evaluation of medical papers, the 32B model focuses on key elements such as the clinical trial design, sample size calculation, and randomization methods.

[0070] During the deep execution phase, the 235B model is used to conduct an in-depth analysis of the paper's innovation and academic value. The model evaluates the paper from three perspectives: theoretical innovation, methodological innovation, and applied innovation, thoroughly analyzing the paper's theoretical breakthroughs, technological innovations, and practical application value. For example, for computer science papers proposing new algorithms, the 235B model assesses the algorithm's theoretical foundation, performance improvements, and application scenarios. It also considers the field's development trends to determine the research's cutting-edge nature and impact.

[0071] During the comprehensive evaluation phase, the 235B model is re-implemented to integrate the previous evaluation results and generate structured review comments. This not only generates quantitative scoring results but also provides targeted revision suggestions based on the paper type and research field. For example, for experimental research papers, the 235B model provides specific improvement suggestions in areas such as optimizing experimental design, improving data analysis methods, and refining innovative points.

[0072] After the executive layer analyzes the target paper through the model, it will generate target review opinions, including detailed evaluation of each part of the paper and specific modification suggestions, such as corrections to formatting issues, optimization suggestions for experimental design, and improvement suggestions for innovative expression.

[0073] Furthermore, the present application S400 includes:

[0074] Extract the target entity set of the target paper and construct a target knowledge graph based on the target entity set; extract any paper, and the arbitrary paper corresponds to any knowledge graph; when the similarity between the target knowledge graph and the arbitrary knowledge graph reaches a predetermined similarity threshold, add the arbitrary paper to the originality evaluation reference list; and perform originality evaluation on the target paper according to the originality evaluation reference list.

[0075] Specifically, we extract the target entity set of the target paper, including research areas, methods, experimental data, conclusions, key figures, and so on. These entities are extracted using natural language processing (NLP) and information extraction techniques. We then construct a target knowledge graph using the extracted target entity set, where nodes identify the extracted entities and edges represent relationships between entities, helping us understand the structure and relevance of the paper's content.

[0076] Extract any paper, and any paper corresponds to any knowledge graph, that is, extract the target entity set for any selected paper and construct an arbitrary knowledge graph of the paper, including the important entities in the paper and the relationships between them.

[0077] A graph comparison algorithm (such as a similarity measurement algorithm) is used to calculate the similarity between the target knowledge graph and any other knowledge graph. If the similarity reaches a predetermined threshold, this indicates that the research content or methods between the target paper and any other paper are highly similar. When the knowledge graph similarity between the target paper and any other paper reaches a predetermined threshold, the paper is added to the originality evaluation reference list and is considered as one of the references when evaluating the target paper for originality.

[0078] Based on the originality evaluation reference list, the target paper is analyzed for originality. By comparing the target paper with the reference papers in the list, any over-reliance on existing research, duplication of content, or suspicion of plagiarism can be identified. Originality evaluation assesses whether the target paper is sufficiently novel or unique to avoid plagiarism or over-reliance on existing research. The final evaluation results include the target paper's originality score and provide revision suggestions, such as adding new research results and optimizing the presentation of innovative points. By constructing a knowledge graph and calculating similarity, originality evaluation is automated, eliminating the need for human intervention and improving review efficiency and accuracy.

[0079] S500: Activate the feedback layer to analyze the target review opinions and obtain target modification suggestions.

[0080] Specifically, the feedback layer has established a comprehensive evaluation and feedback system, including a multi-dimensional scoring system, graded revision suggestions, and a feedback mechanism for review results at each level. Targeted revision suggestions are specific revision suggestions given by the feedback layer based on reviewer comments. They help guide authors to make targeted improvements to their papers based on the review comments, ensuring that the papers meet the requirements of academic journals or conferences.

[0081] The multi-dimensional scoring system uses a five-star scale to quantitatively evaluate papers across dimensions such as innovation, methodological reliability, experimental standardization, data reliability, and application value. For example, the innovation score examines the level of theoretical innovation, the degree of methodological innovation, and the value of applied innovation; the methodological reliability score focuses on the rationality of the research method, the completeness of the technical route, and the standardization of the experimental design.

[0082] In terms of graded revision suggestions, revision suggestions are divided into three levels according to the severity of the problem: critical revisions, important improvement suggestions, and general improvement suggestions. Critical revisions address core issues of the paper, such as serious flaws in experimental design and lack of innovation; important improvement suggestions focus on significant issues in methodology and data analysis; and general improvement suggestions focus on detailed issues such as formatting standards and optimized presentation.

[0083] Feedback on the results of each level of review is collected and organized separately for the shallow review, mid-level review, deep review, and comprehensive evaluation. Feedback on the shallow review focuses on improvements to formatting and compliance issues; feedback on the mid-level review focuses on optimizing experimental protocols and data analysis; and feedback on the deep review focuses on enhancing innovation and academic value. Finally, the comprehensive evaluation feedback integrates improvements at each level into the final evaluation report.

[0084] By collecting and organizing feedback from shallow, mid-level, deep, and comprehensive reviews, we track and analyze review comments layer by layer. The results of all levels of review are summarized into a final evaluation report that includes scoring results, revision suggestions, and detailed feedback from each level of review, helping to tailor the paper.

[0085] In summary, the method for constructing a review agent embedded in a gradient evaluation system provided by this application has the following technical effects:

[0086] By pre-storing the constructed evaluation dimension knowledge base in the memory layer; activating the progressive review model in the planning layer, wherein the progressive review model includes a multi-layer progressive benchmark model; fine-tuning the first benchmark model in the multi-layer progressive benchmark model through the predetermined tool set in the tool layer to obtain the first review model; storing the first review model in the execution layer, and performing the review on the target paper through the execution layer to obtain the target review opinion; activating the feedback layer to analyze the target review opinion to obtain the target modification suggestion. In other words, through the multi-layer progressive review model and standardized evaluation system, the paper is reviewed according to unified standards, and specific modification suggestions are provided based on the review results, which improves the objectivity and systematicness of the review evaluation and enhances the interpretability of the review evaluation results.

[0087] Example 2: Based on the same inventive concept as the method for constructing a manuscript review agent embedded in a gradient evaluation system in the aforementioned Example 1, this application also provides a manuscript review agent construction system embedded in a gradient evaluation system, please refer to the attached Figure 2 The review agent construction system embedded in the gradient evaluation system includes:

[0088] An evaluation pre-storage module 11 is used to pre-store the constructed evaluation dimension knowledge base in the memory layer; a model activation module 12 is used to activate the progressive review model in the planning layer, wherein the progressive review model includes a multi-layer progressive benchmark model; a model fine-tuning module 13 is used to fine-tune the first benchmark model in the multi-layer progressive benchmark model through a predetermined tool set in the tool layer to obtain a first review model; an execution review module 14 is used to store the first review model in the execution layer, and perform a review on the target paper through the execution layer to obtain a target review opinion; an opinion analysis module 15 is used to activate the feedback layer to analyze the target review opinion to obtain a target modification suggestion.

[0089] Furthermore, the evaluation pre-stored module 11 in the review agent construction system embedded in the gradient evaluation system is also used for:

[0090] A review evaluation dimension set for review papers is formed; an academic evaluation dimension set for academic papers is formed; and the evaluation dimension knowledge base is constructed based on a first correspondence between the review papers and the review evaluation dimension set, and a second correspondence between the academic papers and the academic evaluation dimension set; wherein, the review evaluation dimension set includes at least the clarity of the subject, the author's research foundation, the difference in innovation, the completeness of the literature, and the standardization of writing; and the academic evaluation dimension set includes at least the evaluation of the cutting-edge and innovative nature of the topic, the integrity of the basic structure of the paper, the standardization of the experimental part, and the academic standardization.

[0091] Furthermore, the model activation module 12 in the review agent construction system embedded in the gradient evaluation system is also used for:

[0092] The multi-layer progressive benchmark model includes an 8B model, a 32B model and a 235B model.

[0093] Furthermore, the model fine-tuning module 13 in the review agent construction system embedded in the gradient evaluation system is also used to:

[0094] The first benchmark model is adaptively fine-tuned to obtain the first manuscript review model, which includes: if the first benchmark model is the 8B model, a basic format check instruction fine-tuning dataset is constructed in combination with the predetermined tool set; according to the basic format check instruction fine-tuning dataset, the first benchmark model is fine-tuned using the QLoRA method to obtain the first manuscript review model; if the first benchmark model is the 32B model, an experimental scheme evaluation instruction fine-tuning dataset is constructed in combination with the predetermined tool set; according to the experimental scheme evaluation instruction fine-tuning dataset, the first benchmark model is fine-tuned using the QLoRA method to obtain the first manuscript review model; if the first benchmark model is the 235B model, an innovative analysis instruction fine-tuning dataset is constructed in combination with the predetermined tool set; according to the innovative analysis instruction fine-tuning dataset, the first benchmark model is fine-tuned using the QLoRA method to obtain the first manuscript review model.

[0095] Furthermore, the model fine-tuning module 13 in the review agent construction system embedded in the gradient evaluation system is also used to:

[0096] The 8B model is used to evaluate predetermined basic indicators of the target paper, wherein the predetermined basic indicators at least include format standardization, chapter completeness and citation standardization.

[0097] Furthermore, the model fine-tuning module 13 in the review agent construction system embedded in the gradient evaluation system is also used to:

[0098] The 32B model is used to evaluate the predetermined experimental indicators of the target paper, wherein the predetermined experimental indicators at least include the rationality of the experimental plan and the standardization of data analysis.

[0099] Furthermore, the model fine-tuning module 13 in the review agent construction system embedded in the gradient evaluation system is also used to:

[0100] The 235B model is used to evaluate the predetermined innovation indicators of the target paper.

[0101] Furthermore, the model fine-tuning module 13 in the review agent construction system embedded in the gradient evaluation system is also used to:

[0102] Obtain the target type of the target paper; adjust the predetermined innovation indicators in combination with the target type to obtain target innovation indicators, which include: if the target type is a review type, the predetermined innovation indicators include a first predetermined interdisciplinary integration indicator and a first predetermined industry pain point solution indicator; if the target type is an academic type, the predetermined innovation indicators include a second predetermined interdisciplinary integration indicator and a second predetermined industry pain point solution indicator; wherein, the first predetermined interdisciplinary integration indicator includes theoretical integration degree, method reference innovation, and interdisciplinary research trend insight, and the first predetermined industry pain point solution indicator includes accurate pain point identification, solution feasibility, and industry influence evaluation; wherein, the second predetermined interdisciplinary integration indicator includes interdisciplinary depth, interdisciplinary team collaboration, and interdisciplinary achievement transformation, and the second predetermined industry pain point solution indicator includes targeted pain point research, technological breakthroughs and application prospects, and industry collaborative innovation mechanism.

[0103] Furthermore, the execution review module 14 in the review agent construction system embedded in the gradient evaluation system is also used for:

[0104] Extract the target entity set of the target paper and construct a target knowledge graph based on the target entity set; extract any paper, and the arbitrary paper corresponds to any knowledge graph; when the similarity between the target knowledge graph and the arbitrary knowledge graph reaches a predetermined similarity threshold, add the arbitrary paper to the originality evaluation reference list; and perform originality evaluation on the target paper according to the originality evaluation reference list.

[0105] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. Figure 1 The method for constructing a review agent embedded in a gradient evaluation system and the specific examples in Example 1 are also applicable to a review agent construction system embedded in a gradient evaluation system in this embodiment. Through the above detailed description of the method for constructing a review agent embedded in a gradient evaluation system, those skilled in the art can clearly understand the review agent construction system embedded in a gradient evaluation system in this embodiment, so for the sake of brevity of the specification, it will not be described in detail here.

[0106] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present application. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not limited to the embodiments shown herein, but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

[0107] Obviously, those skilled in the art may make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the present application and its equivalents, the present application is intended to include these modifications and variations.

Claims

1. A method for constructing a review agent embedded in a gradient evaluation system, characterized in that: include: Pre-store the constructed evaluation dimension knowledge base into the memory layer; Activate a progressive review model in the planning layer, wherein the progressive review model includes a multi-layer progressive benchmark model; Fine-tuning the first benchmark model in the multi-layer progressive benchmark model using a predetermined tool set in the tool layer to obtain a first manuscript review model; Storing the first review model in the execution layer, and performing a review on the target paper through the execution layer to obtain a target review opinion; The feedback layer is activated to analyze the target review opinions and obtain target modification suggestions.

2. A method for constructing a manuscript review agent embedded in a gradient evaluation system as claimed in claim 1, characterized in that: Before pre-storing the constructed evaluation dimension knowledge base into the memory layer, it includes: Construct a set of review evaluation dimensions for review papers; Construct a set of academic evaluation dimensions for academic papers; Constructing the evaluation dimension knowledge base according to the first correspondence between the review paper and the review evaluation dimension set, and the second correspondence between the academic paper and the academic evaluation dimension set; Among them, the review evaluation dimension set includes at least the clarity of the topic, the author's research foundation, innovative differences, literature completeness and writing standards, and the academic evaluation dimension set includes at least the evaluation of the cutting-edge and innovative nature of the topic, the integrity of the basic structure of the paper, the standardization of the experimental part and the academic standardization.

3. A method for constructing a manuscript review agent embedded in a gradient evaluation system as claimed in claim 1, characterized in that: The multi-layer progressive benchmark model includes an 8B model, a 32B model and a 235B model.

4. A method for constructing a manuscript review agent embedded in a gradient evaluation system as claimed in claim 3, characterized in that: Fine-tuning the first benchmark model in the multi-layer progressive benchmark model using a predetermined tool set in the tool layer to obtain a first manuscript review model includes: Adaptively fine-tuning the first benchmark model to obtain the first manuscript review model includes: If the first benchmark model is the 8B model, constructing a basic format check instruction fine-tuning dataset in combination with the predetermined tool set; Fine-tune the dataset according to the basic format check instruction, and fine-tune the first benchmark model using the QLoRA method to obtain the first manuscript review model; If the first benchmark model is the 32B model, constructing an experimental plan evaluation instruction fine-tuning dataset in combination with the predetermined tool set; Fine-tune the dataset according to the experimental plan evaluation instructions, and fine-tune the first benchmark model using the QLoRA method to obtain the first review model; If the first benchmark model is the 235B model, constructing an innovative analysis instruction fine-tuning dataset in combination with the predetermined tool set; The data set is fine-tuned according to the innovative analysis instructions, and the first benchmark model is fine-tuned using the QLoRA method to obtain the first review model.

5. A method for constructing a manuscript review agent embedded in a gradient evaluation system as claimed in claim 4, characterized in that: The 8B model is used to evaluate predetermined basic indicators of the target paper, wherein the predetermined basic indicators at least include format standardization, chapter completeness and citation standardization.

6. A method for constructing a manuscript review agent embedded in a gradient evaluation system as claimed in claim 4, characterized in that: The 32B model is used to evaluate the predetermined experimental indicators of the target paper, wherein the predetermined experimental indicators at least include the rationality of the experimental plan and the standardization of data analysis.

7. A method for constructing a manuscript review agent embedded in a gradient evaluation system as claimed in claim 4, characterized in that: The 235B model is used to evaluate the predetermined innovation indicators of the target paper.

8. A method for constructing a manuscript review agent embedded in a gradient evaluation system as claimed in claim 7, characterized in that: Obtaining the predetermined innovation indicator includes: Obtain the target type of the target paper; The predetermined innovation indicator is adjusted in combination with the target type to obtain a target innovation indicator, which includes: If the target type is a review type, the predetermined innovation indicators include a first predetermined interdisciplinary integration indicator and a first predetermined industry pain point solution indicator; If the target type is academic, the predetermined innovation indicator includes a second predetermined interdisciplinary integration indicator and a second predetermined industry pain point solution indicator; Among them, the first predetermined interdisciplinary integration indicator includes theoretical integration degree, method reference innovation, and interdisciplinary research trend insight; the first predetermined industry pain point solution indicator includes accurate identification of pain points, solution feasibility, and industry impact assessment; Among them, the second predetermined interdisciplinary integration indicators include the depth of interdisciplinary cross-disciplinary, interdisciplinary team collaboration, and interdisciplinary results transformation; the second predetermined industry pain point solution indicators include targeted research on pain points, technological breakthroughs and application prospects, and industry collaborative innovation mechanisms.

9. A method for constructing a manuscript review agent embedded in a gradient evaluation system as claimed in claim 1, characterized in that: After storing the first review model in the execution layer and reviewing the target paper through the execution layer to obtain the target review opinion, the method further includes: Extracting a target entity set of the target paper, and constructing a target knowledge graph based on the target entity set; Extract any paper, and the paper corresponds to any knowledge graph; When the similarity between the target knowledge graph and the arbitrary knowledge graph reaches a predetermined similarity threshold, the arbitrary paper is added to the originality evaluation reference list; The target paper is evaluated for originality according to the originality evaluation reference list.

10. A review agent construction system embedded in a gradient evaluation system, characterized by: The steps for implementing the method for constructing a manuscript review agent embedded in a gradient evaluation system as described in any one of claims 1 to 9, wherein the manuscript review agent construction system embedded in a gradient evaluation system comprises: Evaluation pre-storage module, used to pre-store the constructed evaluation dimension knowledge base into the memory layer; A model activation module, configured to activate a progressive review model in a planning layer, wherein the progressive review model includes a multi-layer progressive benchmark model; A model fine-tuning module, configured to fine-tune the first benchmark model in the multi-layer progressive benchmark model using a predetermined tool set in the tool layer to obtain a first manuscript review model; An execution review module, configured to store the first review model in an execution layer, and perform a review on a target paper through the execution layer to obtain a target review opinion; The opinion analysis module is used to activate the feedback layer to analyze the target review opinions and obtain target modification suggestions.

Images