Automatic scientific research method and system

Abstract

The invention discloses an automatic scientific research method and system, and the method comprises the steps: generating a paper outline based on the obtained existing research literature through the first function of a scientific research agent; based on the obtained existing research literature and the thesis outline, generating an experimental design scheme of the thesis by using a second function of a scientific research agent; obtaining an experiment result according to the experiment design scheme of the thesis; on the basis of the obtained existing research literature, the thesis outline, the experiment design scheme of the thesis and the experiment result, generating a complete thesis by utilizing a third function of a scientific research intelligent agent; and optimizing the complete paper according to the review opinions to obtain a new paper meeting a preset standard. According to the method, through deep integration of a large language model, semantic retrieval, code generation, simulation review and other technologies, the whole process independence from topic selection to paper completion is realized, the scientific research efficiency and the result quality are remarkably improved, and a foundation is laid for a scientific research normal form driven by artificial intelligence.

Description

Technical Field

[0001] This application relates to the field of computers, and particularly to an automated scientific research method and system. Background Art

[0002] Since the rise of computer science in the late 1970s and early 1980s, people's desire for automated scientific research has been increasing day by day, hoping to accelerate knowledge growth and solve human challenges with the help of machine intelligence. The emergence of large language models (LLMs) has brought dawn to the scientific research field, but also faces many problems.

[0003] Although existing research has tried to use LLMs to assist scientific research, there are obvious limitations. In the application mode, the "prompt engineering" - style interaction is mostly adopted, regarding LLMs only as information retrieval or text generation auxiliary tools, and failing to fully explore their potential for reasoning, planning, and iterative improvement, making it difficult to simulate the core cognitive processes of human scientific research. In key scientific research links, such as conceiving cutting - edge topics, designing innovative experiments, executing code and iterating, and writing high - quality papers, artificial intelligence in the existing technology still cannot replace humans. Summary of the Invention

[0004] The purpose of the embodiments of this application is to provide an automated scientific research method and system, and this method can realize the full - process autonomy from topic selection to the completion of the paper.

[0005] To achieve the above purpose, the embodiments of this application provide an automated scientific research method, including:

[0006] Based on the obtained existing research literature, using the first function of the scientific research intelligent agent to generate a paper outline; wherein, the first function of the scientific research intelligent agent is obtained by performing first supervised learning training on the first large language model using the first data set;

[0007] Based on the obtained existing research literature and the paper outline, using the second function of the scientific research intelligent agent to generate an experimental design scheme for the paper; wherein, the second function of the scientific research intelligent agent is obtained by performing second supervised learning training on the first large language model using the second data set;

[0008] Obtain experimental results according to the experimental design scheme of the paper;

[0009] Based on the obtained existing research literature, the paper outline, the experimental design scheme of the paper, and the experimental results, using the third function of the scientific research intelligent agent to generate a complete paper; wherein, the third function of the scientific research intelligent agent is obtained by performing third supervised learning training on the first large language model using the third data set;

[0010] Optimize the complete paper according to the review opinions to obtain a new paper that meets the preset standards.

[0011] Optionally, based on the acquired existing research literature, use the first function of the scientific research agent to generate a paper outline, including:

[0012] Construct a knowledge graph based on the acquired existing research literature in the relevant field; among them, the knowledge representation form of the knowledge graph adopts the triple "paper - reference - reference literature";

[0013] Obtain classic papers and frontier papers in the relevant field in the knowledge graph;

[0014] Input the classic papers and frontier papers into the scientific research agent, and use the first function of the scientific research agent to generate a paper outline.

[0015] Optionally, the construction of the domain knowledge graph includes:

[0016] Perform semantic parsing on the acquired research topic based on a semantic-driven retrieval strategy;

[0017] Obtain relevant papers according to the semantic parsing results;

[0018] Integrate the acquired relevant papers, their reference papers and reference relationships to obtain triples;

[0019] Use the triples to construct a domain knowledge graph.

[0020] Optionally, use the first dataset to perform the first supervised learning training on the first large language model to obtain the first function of the scientific research agent, including:

[0021] Construct the first dataset; among them, the first dataset includes first training samples, the first training samples include first input features and first output labels, the first input features adopt the reference literature of the paper to be generated, and the first output labels adopt the outline of the paper to be generated;

[0022] Determine the first loss function and the first optimizer; among them, the first loss function is used to measure the difference between the model prediction result and the first output label, and the first optimizer is used to adjust the model parameters to minimize the loss function;

[0023] Input the first input features into the first large language model to obtain the first-stage prediction result;

[0024] Calculate the loss value between the first-stage prediction result and the first output label, calculate the gradient of the loss function with respect to the model parameters through the backpropagation algorithm, and the first optimizer updates the model parameters according to the gradient to gradually reduce the loss value until the loss value converges or reaches the preset number of training rounds, completing the first supervised learning training of the first large language model to obtain the first function of the scientific research agent.

[0025] Optionally, perform second supervised learning training on the first large language model using a second data set to obtain a second function of the scientific research agent, including:

[0026] Construct a second data set; wherein, the second data set includes second training samples, the second training samples include second input features and second output labels, the second input features adopt the references of the paper to be generated and the outline of the paper to be generated, and the second output labels adopt the experimental design scheme of the paper to be generated;

[0027] Determine a second loss function and a second optimizer; wherein, the second loss function is used to measure the difference between the model prediction result and the second output label, and the second optimizer is used to adjust the model parameters to minimize the loss function;

[0028] Input the second input features into the second large language model to obtain a second stage prediction result;

[0029] Calculate the loss value between the second stage prediction result and the second output label, calculate the gradient of the loss function with respect to the model parameters through the backpropagation algorithm, and the second optimizer updates the model parameters according to the gradient to gradually reduce the loss value until the loss value converges or reaches the preset number of training rounds, completing the second supervised learning training of the first large language model to obtain the second function of the scientific research agent.

[0030] Optionally, perform third supervised learning training on the first large language model using a third data set to obtain a third function of the scientific research agent, including:

[0031] Construct a third data set; wherein, the third data set includes third training samples, the third training samples include third input features and third output labels, the third input features adopt the references of the paper to be generated, the outline of the paper to be generated, the experimental design scheme of the paper to be generated and the experimental results, and the third output labels adopt the complete paper to be generated;

[0032] Determine a third loss function and a third optimizer; wherein, the third loss function is used to measure the difference between the model prediction result and the third output label, and the third optimizer is used to adjust the model parameters to minimize the loss function;

[0033] Input the third input features into the first large language model to obtain a third stage prediction result;

[0034] Calculate the loss value between the third stage prediction result and the third output label, calculate the gradient of the loss function with respect to the model parameters through the backpropagation algorithm, and the third optimizer updates the model parameters according to the gradient to gradually reduce the loss value until the loss value converges or reaches the preset number of training rounds, completing the third supervised learning training of the first large language model to obtain the third function of the scientific research agent.

[0035] Optionally, obtain experimental results according to the experimental design scheme of the paper, including:

[0036] Input the experimental design scheme of the paper into the existing code domain large language model;

[0037] The code domain large language model modifies the preset baseline code according to the experimental design scheme to obtain the modified code;

[0038] Use the code interpreter to run the modified code to generate experimental results;

[0039] Or,

[0040] Execute the experimental design scheme of the paper in a manual operation manner to obtain the corresponding experimental results;

[0041] Process the experimental results according to the preset format and input the processed experimental results into the scientific research intelligent agent.

[0042] Optionally, optimize the complete paper according to the review opinions to obtain a new paper that meets the preset standards, including:

[0043] Input the generated complete paper into the review model to output the review opinions of the paper;

[0044] The scientific research intelligent agent iteratively optimizes the complete paper according to the review opinions until a new paper that meets the preset standards is obtained;

[0045] Among them, the review model is obtained by performing supervised learning training on the second large language model using the review data set.

[0046] Optionally, the supervised learning training of the second large language model using the review data set to obtain the review model includes:

[0047] Construct a review data set; among them, the review data set includes review training samples, and the review training samples include review input features and corresponding review output labels. Among them, the review input features are the obtained existing papers, and the review output labels include multiple review opinions with the first review weight and review opinions with the second review weight;

[0048] Determine the fourth loss function and the fourth optimizer; among them, the fourth loss function is used to measure the difference between the model prediction result and the review output label, and the fourth optimizer is used to adjust the model parameters to minimize the loss function;

[0049] Input the review input features into the second large language model to obtain the second prediction result;

[0050] Calculate the loss value between the second prediction result and the review output label, calculate the gradient of the loss function with respect to the model parameters through the backpropagation algorithm, and the second optimizer updates the model parameters according to the gradient to gradually reduce the loss value until the loss value converges or reaches the preset number of training rounds, completing the supervised learning training of the second large language model to obtain the review model.

[0051] An embodiment of the present application also provides an automated scientific research system, including:

[0052] An academic search engine for obtaining existing research literature from a preprint platform;

[0053] A scientific research agent configured to generate a paper outline based on the obtained existing research literature; generate an experimental design plan for the paper based on the obtained existing research literature and the paper outline;

[0054] An experimental result generation module configured to generate experimental results according to the experimental design plan of the paper;

[0055] The scientific research agent is further configured to generate a complete paper based on the obtained existing research literature, the paper outline, the experimental design plan of the paper, and the experimental results; wherein, the scientific research agent is obtained by performing supervised learning training on a first large language model;

[0056] A review model configured to review the complete paper to obtain review opinions; wherein, the review model is obtained by performing supervised learning training on a second large language model using a review data set;

[0057] The scientific research agent is further configured to optimize the complete paper according to the review opinions to obtain a new paper that meets the preset standards.

[0058] The automated scientific research method and system provided by the embodiments of the present application construct a knowledge graph based on the obtained existing research literature in the relevant field, conduct thinking and knowledge fusion based on the relatively relevant topic literatures in the knowledge graph to generate a structured paper outline, and automatically write the paper content in LaTeX format; the scientific research agent generates a detailed experimental design plan based on the obtained existing research literature and the paper outline, supports both the automatic experiment mode and the manual-assisted experiment mode, and integrates the experimental results into the paper; the review model conducts multi-dimensional evaluation on the paper and provides specific review opinions to guide the scientific research agent to make targeted modifications and improvements, forming a closed-loop optimization process. The automated scientific research method and system provided by the present application realize the full process autonomy from topic selection to paper completion by deeply integrating technologies such as large language models, semantic retrieval, code generation, and simulation review, significantly improving the scientific research efficiency and the quality of research results, and laying a foundation for the AI-driven scientific research paradigm. Description of the Drawings

[0059] Figure 1 Flow chart of the automated scientific research method according to the embodiment of the present application;

[0060] Figure 2 For the embodiment of the present application Figure 1 Flow chart of step S100 in

[0061] Figure 3 For the embodiment of the present application Figure 2 Flow chart of step S110 in

[0062] Figure 4 Flow chart of the first function of the scientific research agent obtained by training in the automated scientific research method according to the embodiment of the present application;

[0063] Figure 5 Flow chart of the second function of the scientific research agent obtained by training in the automated scientific research method according to the embodiment of the present application;

[0064] Figure 6 For the embodiment of the present application Figure 1 Flow chart of an embodiment of step S300 in

[0065] Figure 7 Flow chart of the third function of the scientific research agent obtained by training in the automated scientific research method according to the embodiment of the present application;

[0066] Figure 8 For the embodiment of the present application Figure 1 Flow chart of step S500 in

[0067] Figure 9 Flow chart of the review model obtained by training in the automated scientific research method according to the embodiment of the present application;

[0068] Figure 10 Structural block diagram of the automated scientific research system according to the embodiment of the present application. Detailed implementation manners

[0069] The various solutions and features of the present application are described herein with reference to the accompanying drawings. It should be understood that various modifications can be made to the embodiments applied herein. Therefore, the above description should not be construed as a limitation, but only as an example of the embodiments. Those skilled in the art will think of other modifications within the scope and spirit of the present application.

[0070] Currently, existing artificial intelligence-driven research tools perform poorly in key dimensions for ensuring research quality. Specifically, these tools are relatively weak in understanding the complexity of research questions, ensuring the feasibility of research methods, evaluating the reliability of research results, and enhancing the value of research contributions. They often struggle to penetrate the essence of problems, and the solutions they provide lack both theoretical basis and practical testing, with significantly insufficient innovation. At the same time, these tools also have difficulty effectively handling the iterative feedback necessary for complex research tasks.

[0071] Although some research attempts to optimize specific research processes using large language models (LLMs), due to the lack of a complete closed-loop system, it is impossible to organically integrate various research steps. Therefore, the existing technology is still in the initial stage of research automation, and there is still a large gap from achieving an intelligent and efficient research process.

[0072] The automated research method and system provided in this application aim to tap the potential of open-source large language models and construct an automated research method and system that can mimic human research reasoning to achieve the automation of the entire research process and provide new paradigms and tools for research work.

[0073] An automated research method provided in an embodiment of this application can achieve the automation and intelligence of the entire research life cycle by integrating key links such as literature retrieval, outline generation, experimental design, real experiment operation, paper writing, peer review, and paper optimization, and significantly improve the efficiency and output quality of research work, thereby promoting the intelligent development process in the research field and providing new solutions to solve complex research problems.

[0074] The following will combine with the accompanying drawings to provide a detailed description of this automated research method. Figure 1 is a flowchart of the automated research method according to an embodiment of this application. As Figure 1 shown, this method includes the following steps:

[0075] S100. Based on the obtained existing research literature, use the first function of the research intelligent agent to generate a paper outline; wherein, the first function of the research intelligent agent is obtained by performing first supervised learning training on the first large language model using the first data set.

[0076] Among them, a large number of existing research literature related to the research topic can be collected through academic databases, professional books, conference proceedings in related fields, etc. These literatures are the basis and starting point for subsequent research, enabling researchers to understand information such as existing research directions, achievements, gaps, and controversial points in this field.

[0077] The scientific research intelligent agent is built based on the first large language model, which can be an open-source large language model. Such open-source large language models need to include at least 1B (billion) learnable parameters; during the unsupervised pre-training process, 10T (trillion) Tokens (basic units in text processing) are used; when processing text, it can process a context with a length of at least 64,000 Tokens at one time, that is, the context window size of such models is above 64,000. The larger context window makes the model more advantageous in processing long texts. It can better capture long-distance dependencies in the text, understand complex discourse structures and semantic information. For example, when processing long texts such as long academic papers and novels, the larger context window can make the abstracts or continued writings generated by the model more accurate and coherent. Specifically, the first large language model can adopt advanced open-source large language models such as Mistral and Qwen 2.5.

[0078] By learning a large amount of academic text data, the scientific research intelligent agent masters the structural characteristics, language expression habits of academic papers, and logical laws in different research fields, so as to be able to generate a paper outline based on the input literature information. Among them, the paper outline includes parts such as research motivation, research objectives, related work, research methods, experimental design, experimental result analysis, and conclusions, and the content of each part is organized in a structured form. For example, it can identify the content of parts such as research background, purpose, methods, main findings, and conclusions in different literatures, and based on this information, sort out a paper outline framework suitable for the current research topic in a manner that conforms to academic norms and logic.

[0079] Specifically, the first dataset is a data set collected and sorted specifically for training the function of the scientific research intelligent agent to generate a paper outline. It includes a large number of samples related to academic papers, and these samples cover different types of papers in various subject fields. Each sample includes a complete paper and its corresponding outline, or key information extracted from the paper and the corresponding outline fragments. For example, the first dataset includes research papers in the computer field, which detail research backgrounds, experimental methods, result discussions, etc., and at the same time correspond to a well-organized paper outline.

[0080] S200. Based on the obtained existing research literature and the said paper outline, use the second function of the scientific research intelligent agent to generate an experimental design plan for the paper; among them, the second function of the scientific research intelligent agent is obtained by performing second supervised learning training on the first large language model using the second dataset.

[0081] It should be noted that the first large language model itself has powerful language understanding and generation capabilities, but for specific tasks in the scientific research field, further optimization is required. Through the second supervised learning training, using the second dataset closely related to scientific research experimental design, the first large language model learns the patterns, rules, and key elements of different types of scientific research experimental design. For example, in chemical experimental design, it learns how to select appropriate chemical reaction conditions, reactant ratios, and product detection methods, etc.; in social science experimental design, it learns how to design questionnaires, select samples, and control interference factors in the experiment, etc. After such training, the second function of the scientific research intelligent agent can generate a highly targeted experimental design plan according to the input research-related information.

[0082] For another example, in the field of machine learning in computer science, the experimental design plan can include experimental settings and statistical metrics. Among them, when designing the experiment, for several key aspects such as network architecture, optimization strategy, and regularization method, different specific methods are selected respectively and combined together to form different experimental conditions. For example, experimental setting A is "network architecture (AlexNet) + optimization strategy (Adam) + regularization method (L2 regularization)", and experimental setting B is "network architecture (ResNet) + optimization strategy (SGD) + regularization method (Dropout)". By comparing the performance of the model under these different combinations, the influence of each factor and their interactions on the performance of the machine learning model can be deeply understood, so as to find the optimal model configuration.

[0083] The second dataset includes a large amount of sample data related to scientific research experimental design, and these data can come from actual scientific research papers, experimental reports, scientific research project documents, etc. Each sample in the dataset usually includes a description of the experimental purpose and the corresponding experimental design plan, etc. By learning these sample data, the first large language model can master the patterns and rules of scientific research experimental design, so that when receiving research-related information subsequently, it can generate a reasonable experimental design plan.

[0084] For each experimental setup, the first large language model is given the corresponding required statistical metrics. These statistical metrics are used to accurately evaluate the experimental results and help researchers judge the performance of the model under different experimental setups. For example, in an image classification task, common statistical metrics may include accuracy, which is the ratio of the number of correctly classified samples to the total number of samples, and is used to intuitively reflect the classification ability of the first large language model; precision, recall, and F1 value. These metrics are more effective for evaluating model performance when dealing with imbalanced datasets. Precision is used to measure the proportion of samples predicted as positive that are actually positive, recall is used to measure the proportion of samples that are actually positive and are correctly predicted as positive, and the F1 value is the harmonic mean of precision and recall. In addition, the statistical metrics can also include the loss function value, which is used to measure the difference between the model's predicted value and the true value. The smaller the loss function value, the better the model's prediction effect.

[0085] The experimental design plan can be in JSON format. The JSON format is concise, clear, easy to read, write, and parse, which is convenient for researchers to exchange and share data between different platforms and tools.

[0086] S300. Obtain the experimental results according to the experimental design plan of the paper.

[0087] Specifically, for an experimental design plan with characteristics such as rich quantifiable data, a standardized and programmable experimental process, a clear and predictable goal, or relying on highly automated and integrable instrument and equipment, an automatic experiment mode can be adopted to automatically execute and feedback the experimental results. For example, according to the chip function requirements, artificial intelligence can be used for circuit design and layout optimization. Automated manufacturing equipment produces chips, and a test platform detects performance parameters such as chip power consumption and computing speed. Artificial intelligence analyzes the test data and feedbacks whether the chip meets the standards and the improvement direction.

[0088] For an experimental design plan with characteristics such as emphasizing subjective judgment and experience, focusing on individual interaction and human factors, or having a flexible and changeable experimental process, such as experimental design plans in humanities and social sciences, art creation, and clinical diagnosis, an artificial-assisted experiment mode can be adopted to complete the corresponding experiment and obtain the experimental results.

[0089] S400. Based on the obtained existing research literature, paper outline, experimental design plan of the paper, and experimental results, use the third function of the scientific research intelligent agent to generate a complete paper; where the third function of the scientific research intelligent agent is obtained by performing third supervised learning training on the first large language model using a third dataset.

[0090] Among them, the scientific research agent gradually generates each chapter of the paper according to the structure of the paper outline, in the order of abstract, introduction, experimental method, experimental setup, experimental results, and experimental conclusion. After generating the experimental setup section, it interrupts and waits for the experimental results completed according to the experimental setup. After obtaining the experimental results, it can continue to write and output the experimental results and experimental conclusion sections. Specifically, the generated complete paper can be in LaTeX format, which is convenient for paper typesetting, citation management, and formula display, thus realizing the automation and standardization of paper writing.

[0091] S500. Optimize the complete paper according to the review comments to obtain a new paper that meets the preset standards.

[0092] Among them, the review comments are obtained by comprehensively evaluating the complete paper from multiple dimensions such as innovation, scientificity, feasibility, integrity, and expression quality. The review comments can include evaluation scores and modification suggestions.

[0093] For example, the modification suggestion can be "This paper proposes a new deep learning-based image super-resolution reconstruction method, and the selected topic has a certain degree of innovation. However, the description of the method part of the paper is not clear enough, and there are also some problems in the experimental design. For example, the influence of different noise levels is not considered. It is recommended that the author supplement more experimental results and provide a more detailed description of the method." According to this modification suggestion, the corresponding part of the paper can be optimized targeted.

[0094] For another example, if the review comments point out that the innovation of the paper is insufficient, the scientific research agent can conduct a new literature search to find a new research direction; if the review comments point out that there are defects in the experimental design, the scientific research agent can modify the experimental plan or provide a more detailed description of the experimental steps; if the review comments point out that the expression quality of the paper is not good, the scientific research agent can polish and modify the language of the paper. This iterative optimization process will continue until the generated paper meets the expected quality standards or publication requirements. In this way, generating the paper and review optimization form a complete closed loop, which can continuously improve the quality and innovation of the generated paper.

[0095] It can be understood that, according to the review comments, an existing paper optimization model can also be used to optimize the generated complete paper, so as to obtain a new paper that meets the preset standards.

[0096] In one embodiment, as Figure 2 shown, in the above step S100, based on the obtained existing research literature, use the first function of the scientific research agent to generate a paper outline, including:

[0097] S110. Construct a knowledge graph based on the existing research literature in the relevant field; among them, the knowledge representation form of the knowledge graph adopts the triple "paper - reference - reference literature".

[0098] S120. Obtain classic papers and frontier papers in the relevant field from the knowledge graph.

[0099] S130. Input the classic papers and frontier papers into the scientific research intelligent agent, and use the first function of the scientific research intelligent agent to generate a paper outline.

[0100] After constructing the domain knowledge graph of the research topic input by the user, the scientific research intelligent agent can identify classic papers and frontier papers in the relevant field according to the publication year of the papers. Classic papers usually refer to those papers that were published earlier but have a very high number of citations and are of fundamental significance, while frontier papers are recently published papers that are closely related to the research topic.

[0101] In the embodiment of the present application, as Figure 3 shown, in the above step S110, constructing the domain knowledge graph includes:

[0102] S1110. Perform semantic analysis on the obtained research topic based on a semantic-driven retrieval strategy;

[0103] S1120. Obtain relevant papers according to the semantic analysis result;

[0104] S1130. Integrate the obtained relevant papers, their reference papers, and reference relationships to obtain triples.

[0105] S1140. Use the triples to construct the domain knowledge graph.

[0106] Specifically, the Semantic Scholar API can be used as a literature search tool to find papers containing specific vocabulary; a general large language model, such as QWen-2.5-Instruct 7B, can be used to semantically parse the obtained research topics and deeply understand the semantic information behind the research topics input by the user. For example, after obtaining the user input of "Grokking phenomenon", the scientific research agent parses it into a series of related keywords, such as "deep learning generalization", "model complexity", "overfitting", "mutation learning", etc., and uses these keywords for preliminary retrieval, not only retrieving the latest papers in the field of Grokking, but also tracing back to related classic literature, such as research on the relationship between deep learning generalization ability, model complexity and data set size, thereby constructing a complete knowledge context. In order to ensure the comprehensiveness of the retrieval results, the scientific research agent automatically sets multiple time nodes, for example, retrieving the literature of the past ten years and the past year respectively. Such a time span setting enables the system to capture both the classic literature in the field and the latest research progress.

[0107] In addition, according to the semantic information of the topic, Semantic Scholar API can perform multi-level, recursive literature retrieval. For example, not only papers directly related to keywords such as "image super-resolution reconstruction" and "deep learning" are retrieved, but also higher-level concepts such as "computer vision" and "image processing" and other related classic literature and latest research results. At the same time, the references cited by these papers can be recursively retrieved, and the title, abstract, author, publication year, citation relationship and other information of these documents can be extracted. In order to more comprehensively understand the knowledge structure of the research field, this application can perform deep semantic analysis on the title, abstract, introduction and conclusion of each document, identify key concepts, research methods and core conclusions, and use the text summary capability of LLMs to extract the core content of each document into a concise summary. On this basis, the scientific research agent constructs a research topic knowledge graph, which represents the citation relationship between documents, the association relationship between key concepts, and the evolution relationship between research methods in the form of a graph, thereby forming a comprehensive background knowledge understanding of the research topic and storing it in a structured form suitable for language model processing.

[0108] A knowledge graph describes concepts, entities and the relationships between them in the objective world in a structured form. Triples are a common representation form in a knowledge graph. A triple is usually expressed as (head entity, relation, tail entity), and through this form, a certain association between two entities can be concisely expressed. In the embodiments of the present application, the head entity in the triple can be the existing papers in the obtained region of interest, the tail entity can be the reference literature, and the relationship between the head entity and the tail entity is a reference relationship. Through a large number of triples, a huge knowledge network in the field of interest can be constructed, and it can also help to understand the user's query intention and provide more accurate search results.

[0109] Each node in the constructed domain knowledge graph represents a paper, and the unique identifier of the paper can be used as the unique ID of the node. The edges in the knowledge graph represent the reference or citation relationship between papers. For example, if paper A references or cites paper B, then there is a directed edge in the knowledge graph from the node of paper A to the node of paper B.

[0110] In one embodiment, in the above step S100, as Figure 4 shown, the first large language model is trained by the first supervised learning using the first data set to obtain the first function of the scientific research agent, including:

[0111] S1150. Construct the first data set; wherein, the first data set includes first training samples, the first training samples include first input features and first output labels, the first input features adopt the reference literature of the paper to be generated, and the first output labels adopt the outline of the paper to be generated.

[0112] Among them, the first data set may further include first validation samples and / or first test samples. Among them, the first validation samples are used to adjust the model hyperparameters during the training process to avoid overfitting; the first test samples are used to evaluate the performance of the model after the training is completed.

[0113] Specifically, the reference literature of the paper to be generated is stored in a bib file. A bib file is a plain text file used to store literature citation information, and its file extension is.bib. It plays a key role in academic writing and literature management, especially when used in combination with the LaTeX document writing system.

[0114] S1160. Determine the first loss function and the first optimizer; wherein, the first loss function is used to measure the difference between the model prediction result and the first output label, and the first optimizer is used to adjust the model parameters to minimize the loss function.

[0115] It is understandable that the first loss function is used to measure the difference between the model's prediction result and the true label (the first output label in this step). This quantitative representation of the difference can provide a clear goal for the training of the model, that is, by adjusting the model parameters, the value of the loss function is made as small as possible. Specifically, the loss function adopted in the embodiments of this application can be cross-entropy loss or negative log-likelihood loss, etc.

[0116] For the task of generating text, it is usually desired that the probability distribution of the next word predicted by the model matches the word that appears at that position in the true text. For example, when generating a certain part of a paper outline, the model predicts the probabilities of different expressions such as "research methods" and "experimental results", and the cross-entropy loss can effectively measure the difference between these predicted probabilities and the words used in the true text. It is sensitive to changes in the probability distribution and helps the model learn the correct pattern of text generation.

[0117] Minimizing the negative log-likelihood loss is equivalent to maximizing the probability that the model generates the true text. This prompts the model to learn the parameters that can generate the correct text sequence with a high probability. For example, when generating a body paragraph of a paper, the model adjusts the parameters by minimizing the negative log-likelihood loss to generate content similar to the true body.

[0118] The role of the first optimizer is to continuously reduce the value of the loss function by adjusting the model's parameters, so that the model gradually learns the patterns and rules in the data and improves the prediction accuracy. In the embodiments of this application, the first optimizer uses the LoRA-GA (Low Rank Gradient Approximation) optimization algorithm to train the model.

[0119] The LoRA-GA optimization algorithm combines LoRA (Low Rank Adaptation) and GA (Gradient Approximation). During the LoRA fine-tuning process of the model, by using the method of gradient approximation, the gradient of the first step is approximated to the gradient during the full-parameter fine-tuning of the model, which effectively improves the training speed. This method helps to improve the training efficiency of the model in the LoRA fine-tuning stage and can achieve an effect similar to full-parameter fine-tuning.

[0120] S1170: Input the first input feature into the first large language model to obtain the first-stage prediction result.

[0121] S1180: Calculate the loss value between the first-stage prediction result and the first output label, calculate the gradient of the loss function with respect to the model parameters through the backpropagation algorithm, and the first optimizer updates the model parameters according to the gradient to gradually reduce the loss value until the loss value converges or reaches the preset number of training rounds, completing the first supervised learning training of the first large language model to obtain the first function of the scientific research intelligent agent.

[0122] In supervised learning, the model generates prediction results based on input features. For example, in a text generation task, given a text prompt as input, the model generates a subsequent text as the prediction result. The "first output label" is the true and expected output corresponding to the input features, which is the pre-labeled standard answer.

[0123] The role of the loss function is to measure the degree of difference between the prediction result and the true label. By inputting the prediction result of the first stage and the first output label into the selected loss function, a specific loss value can be calculated, which reflects the size of the prediction error of the current model in this round of training. The larger the loss value, the greater the gap between the prediction result of the model and the first output label; conversely, the smaller the loss value, the closer the model's prediction is to the real situation.

[0124] The backpropagation algorithm is the core method for calculating gradients in deep learning. The gradient can be understood as the rate of change of a function at a certain point. In deep learning, the gradient of the loss function with respect to the model parameters represents how the value of the loss function changes with a small change in the model parameters. For example, if the gradient of a certain model parameter is positive, it means that increasing the value of this parameter will increase the loss function, so the value of this parameter should be decreased when updating the parameters; conversely, if the gradient is negative, the value of this parameter should be increased.

[0125] In each round of training, the first optimizer updates the parameters of the first large language model with an appropriate step size according to the calculated gradients, so that the value of the loss function gradually decreases, making the prediction result of the first large language model closer and closer to the first output label.

[0126] "Loss value convergence" means that after multiple rounds of training, the loss value no longer has an obvious downward trend and tends to be stable. This usually indicates that the first large language model has learned the patterns and rules in the training samples and reached a good state. The "preset number of training rounds" is an upper limit of the number of training times preset before the start of training. Due to the complexity of the data or the characteristics of the model, sometimes the loss value may not converge completely or the convergence speed is very slow. In this case, to avoid indefinite training, a preset number of training rounds is set, and when the training reaches this number of rounds, the training process stops. When the loss value converges or reaches the preset number of training rounds, it is considered that the supervised learning training of the first large language model is completed. At this time, the model has adjusted its own parameters by learning a large number of input-output data pairs and has certain prediction and generalization abilities. The trained first large language model can be used to predict new input data to complete various natural language processing tasks.

[0127] The first function of the trained scientific research agent can provide three inference modes, including exploration mode, rigorous mode, and rapid iteration mode, to meet different needs of users.

[0128] For example, if the user selects the exploration mode, the scientific research agent will generate multiple potentially innovative research directions based on the current knowledge graph and conduct a preliminary evaluation of the feasibility and expected influence of each direction. For example, for the topic of "Research on Image Super-Resolution Reconstruction Method Based on Deep Learning", the scientific research agent will propose the following research directions: (1) Explore new network structures to improve the accuracy and efficiency of image super-resolution reconstruction; (2) Study how to use unsupervised learning methods to improve the generalization ability of the model; (3) Combine the attention mechanism to enhance the model's ability to reconstruct local details of images. For each research direction, the scientific research agent generates a preliminary research outline and evaluates its feasibility, such as whether a large amount of data or computing resources are required, and whether there are relevant open-source code libraries.

[0129] If the user selects the rigorous mode, the scientific research agent will generate a logically rigorous and well-argued paper outline based on existing research results and provide detailed evidence support for each argument in the outline.

[0130] If the user selects the rapid iteration mode, the scientific research agent generates a brief paper outline and provides multiple optional experimental design schemes to support the user in quickly verifying the core concept.

[0131] In one embodiment, in the above step S200, as Figure 5 shown, the first large language model is trained with the second dataset through second supervised learning to obtain the second function of the scientific research agent, including:

[0132] S210. Construct the second dataset; wherein, the second dataset includes second training samples, the second training samples include second input features and second output labels, the second input features use the references of the paper to be generated and the outline of the paper to be generated, and the second output labels use the experimental design scheme of the paper to be generated.

[0133] Specifically, both the references of the paper to be generated and the outline of the paper to be generated are stored in a bib file.

[0134] S220. Determine the second loss function and the second optimizer; wherein, the second loss function is used to measure the difference between the model prediction result and the second output label, and the second optimizer is used to adjust the model parameters to minimize the loss function.

[0135] It is understandable that the second loss function is used to measure the difference between the model prediction result and the true label (the second output label in this step). The second optimizer uses the LoRA-GA (Low-Rank Gradient Approximation) optimization algorithm to train the model.

[0136] S230. Input the second input feature into the second large language model to obtain the second-stage prediction result.

[0137] S240. Calculate the loss value between the second-stage prediction result and the second output label, calculate the gradient of the loss function with respect to the model parameters through the backpropagation algorithm, and the second optimizer updates the model parameters according to the gradient to gradually reduce the loss value until the loss value converges or reaches the preset number of training epochs, completing the second supervised learning training of the first large language model to obtain the second function of the scientific research agent.

[0138] In one embodiment, in the above step S300, the experimental result can be obtained according to the experimental design scheme of the paper through the automatic experiment mode or the manual-assisted experiment mode.

[0139] Among them, through the automatic experiment mode, the experimental result is obtained according to the experimental design scheme of the paper, as Figure 6 shown, including:

[0140] S310. Input the experimental design scheme of the paper into the large language model in the existing code field; specifically, the large language model in the existing code field can adopt models such as Deepseek-v3 or GPT-o1 or GPT-o3. This type of large language model has been trained in a large number of code tasks and can complete the modification work of the experimental code in the scientific research process to meet the given experimental requirements.

[0141] S320. The large language model in the code field modifies the preset baseline code according to the experimental design scheme to obtain the modified code;

[0142] Exemplarily, the experimental design scheme of the paper can be used as a collaborative instruction, and both the collaborative instruction and the preset baseline code are input into the Deepseek-v3 model. The Deepseek-v3 model performs change operations on several lines of code in the preset baseline code according to the collaborative instruction to obtain the modified code.

[0143] Based on reading the code library such as the experimental code experiment.py, the large language model in the existing code field provides the code part that needs to be modified, such as adding new functions, deleting unnecessary code blocks, writing log information, or modifying the hyperparameters used in the original code, etc., to obtain the modified code.

[0144] S330. Run the modified code using a code interpreter to generate experimental results.

[0145] Among them, a code interpreter is a tool or system that can analyze and understand code and elaborate on the functions, logic, execution process, etc. of the code in a natural language form that is easy for humans to understand. Exemplarily, the code interpreter can be a python interpreter.

[0146] In addition, through the manual-assisted experiment mode, obtain experimental results according to the experimental design scheme of the paper, including:

[0147] S340. Execute the experimental design scheme of the paper in a manual operation manner to obtain corresponding experimental results.

[0148] Exemplarily, the scientific research intelligent agent can generate a detailed experimental procedure manual according to the experimental design scheme and provide auxiliary tools for data preprocessing, statistical analysis, and result visualization, so as to support users in conducting some experiments that cannot be fully replaced by AI at present. For example, certain specific biological or chemical experiments.

[0149] The above flexible experimental design and execution mechanism enable the scientific research intelligent agent to handle various types of scientific research tasks and ensure the reliability of experimental results.

[0150] S350. Process the experimental results according to a preset format and input the processed experimental results into the scientific research intelligent agent.

[0151] Specifically, the experimental results can be in JSON format. These result data will not only be integrated into the experimental chapter of the paper to provide data support for the argument, but also serve as an important basis for subsequent experimental scheme adjustment and paper content optimization.

[0152] In one embodiment, in the above step S400, as Figure 7 shown, perform third supervised learning training on the first large language model using the third data set to obtain the third function of the scientific research intelligent agent, including:

[0153] S410. Construct a third data set; among them, the third data set includes third training samples, the third training samples include third input features and third output labels, the third input features adopt the references of the paper to be generated, the outline of the paper to be generated, the experimental design scheme of the paper to be generated, and experimental results, and the third output labels adopt the complete paper to be generated.

[0154] S420. Determine a third loss function and a third optimizer; among them, the third loss function is used to measure the difference between the model prediction result and the third output label, and the third optimizer is used to adjust the model parameters to minimize the loss function.

[0155] S430. Input the third input feature into the first large language model to obtain the prediction result of the third stage.

[0156] S440. Calculate the loss value between the prediction result of the third stage and the third output label, calculate the gradient of the loss function with respect to the model parameters through the backpropagation algorithm, and the third optimizer updates the model parameters according to the gradient to gradually reduce the loss value until the loss value converges or reaches the preset number of training rounds, completing the third supervised learning training of the first large language model to obtain the third function of the scientific research intelligent agent.

[0157] In one embodiment, in the above step S500, as Figure 8 shown, optimize the complete paper according to the review opinions to obtain a new paper that meets the preset standards, including:

[0158] S510. Input the generated complete paper into the review model to output the review opinions of the paper.

[0159] Among them, the review model is obtained by performing supervised learning training on the second large language model using the review data set.

[0160] Specifically, the evaluation score in the review opinion is a quantitative summary of the quality of the initial complete paper and is the result of multi-dimensional evaluation based on a series of preset standards. The review is carried out from dimensions such as the reliability, presentation, and contribution of the initial complete paper. The range of the evaluation score can be set from 1 to 10 points. Different weights can be assigned to each dimension according to actual needs, and finally a comprehensive evaluation score is calculated.

[0161] The review model is constructed based on the second large language model. The second large language model can be advanced open-source large language models such as Mistral and Qwen2.5. Such open-source large language models need to include at least 1B (billion) learnable parameters; in the unsupervised pre-training process, 10T (trillion) Tokens (basic units in text processing) are used; when processing text, it can process a context with a length of at least 64000 Tokens at one time, that is, the context window size of such models is above 64000. On this basis, a more professional review model is obtained through additional training with the review data set, so as to be able to perform review tasks on the given academic papers. For this model, its input is the serialized text content of the paper, and the output is the review opinion including specific modification opinions and evaluation scores. This review data set includes a large amount of original text and review opinions of papers from public academic conferences. During the training process, the original text of the paper is given and the generation of review opinions is required.

[0162] The modification suggestions are more detailed and in-depth qualitative evaluations of the initial complete paper. For each part of the initial complete paper, a comprehensive analysis and elaboration can be carried out from two aspects of advantages and disadvantages, and modification suggestions can be given to facilitate targeted modification of the corresponding parts of the paper.

[0163] S520. The scientific research intelligent agent iteratively optimizes the complete paper according to the review opinions until a new paper that meets the preset criteria is obtained;

[0164] Among them, the problems existing in the initial complete paper in the review opinions can be prioritized. For key problems that seriously affect the scientificity and credibility of the paper, such as fundamental errors in research methods or major defects in logical arguments, they are the primary problems to be solved. For example, if the review opinion points out that there is a sample bias in the research method, resulting in the results not being generally representative, then the first thing to do during iterative optimization is to redesign the sample selection scheme and conduct data collection and analysis. For general problems, such as non-standard formats and language expression flaws, they can be centrally optimized after the key problems are solved.

[0165] The scientific research intelligent agent optimizes the complete paper in a targeted manner according to the review opinions. For example, if the review opinion points out that the paper lacks innovation, the scientific research intelligent agent re-conducts literature retrieval to find new research directions; if the review opinion points out that there are defects in the experimental design, the scientific research intelligent agent modifies the experimental design scheme or provides more detailed experimental procedure descriptions; if the review opinion points out that the expression quality of the paper is poor, the scientific research intelligent agent polishes and modifies the language of the paper. In this way, the scientific research intelligent agent and the review model execute a closed-loop iterative optimization process, enabling the generated paper to continuously improve in quality through multiple rounds of iteration and finally obtaining a new paper that meets the preset criteria.

[0166] In the paper writing stage, the scientific research intelligent agent gradually generates each chapter of the paper according to the structure of the outline. During the generation process, the scientific research intelligent agent makes full use of the previously constructed knowledge graph and the review opinions fed back during the iterative process to dynamically adjust the generation strategy. For example, when writing the introduction part, the scientific research intelligent agent combines the previously retrieved literature to outline the research background, current situation, and existing problems in the field of image super-resolution reconstruction, and clearly elaborates the research motivation and objectives of this paper. When writing the method part, the scientific research intelligent agent details the proposed deep learning-based image super-resolution reconstruction method, including the model architecture, training method, loss function, etc., and uses LaTeX syntax for typesetting to ensure the standardization and aesthetics of the paper format.

[0167] In one embodiment, in the above step S500, as Figure 9 shown, the second large language model is trained with supervised learning using the review data set to obtain the review model, including:

[0168] S530. Construct a review data set; wherein, the review data set includes review training samples, and the review training samples include review input features and corresponding review output labels. Among them, the review input features are the obtained existing papers, and the review output labels include review opinions with multiple first review weights and review opinions with second review weights.

[0169] S540. Determine a fourth loss function and a fourth optimizer; wherein, the fourth loss function is used to measure the difference between the model prediction result and the review output label, and the fourth optimizer is used to adjust the model parameters to minimize the loss function.

[0170] S550. Input the review input features into the second large language model to obtain a second prediction result.

[0171] S560. Calculate the loss value between the second prediction result and the review output label, calculate the gradient of the loss function with respect to the model parameters through the backpropagation algorithm, and the second optimizer updates the model parameters according to the gradient to gradually reduce the loss value until the loss value converges or reaches the preset number of training epochs, completing the supervised learning training of the second large language model to obtain the review model.

[0172] In the embodiment of the present application, the second large language model can adopt the Mistral-Large-2 model, and configure an 8xH100 80G GPU cluster, that is, an 8-piece NVIDIA H100 GPU computing cluster is used during training, and the video memory of each GPU is 80GB. Set training hyperparameters such as the learning rate, batch size, and number of training epochs.

[0173] The learning rate determines the step size of model parameter updates in each training iteration. A smaller learning rate means that the model parameter updates are relatively slow, the training process is more stable, but the convergence speed will be relatively slow; a larger learning rate will make the parameter update steps larger, which may cause the model to skip the optimal solution during training, unable to converge, and even the loss function value may diverge. For example, the learning rate can be set to 1e-5.

[0174] The batch size is set to 4x8. This means that in each training iteration, the model will process 4 batches of data at the same time, and each batch contains 8 samples. A larger batch size can utilize the parallel computing power of the GPU, improve the training efficiency, and more accurately reflect the overall distribution characteristics of the data when calculating the gradient, making the model converge more stably.

[0175] According to the specific dataset size, model complexity, and training effect, the number of training rounds can be set to 12 rounds. One epoch means that the model trains the entire review dataset once. Increasing the number of training rounds usually enables the model to better learn the features and patterns in the data. However, if the number of training rounds is too large, the model may overfit, that is, it performs well on the training set but its performance deteriorates on the test set or new data.

[0176] The automated scientific research method provided by this application has the following advantages:

[0177] First, it can achieve the intelligence and automation of the scientific research process.

[0178] After obtaining the research topic, a knowledge graph is constructed based on the existing research literature in the relevant field. Thinking and knowledge integration are carried out according to the relatively related topic literature in the knowledge graph to generate a structured paper outline, and the paper content is automatically written in LaTeX format.

[0179] Second, it has configurable experimental design and execution capabilities.

[0180] The scientific research agent can generate a detailed experimental design plan, support the automatic experiment mode (it can call external modules such as GPT-o1 to execute code) and the manual-assisted experiment mode, and integrate the results into the paper.

[0181] Finally, it integrates a simulated peer review and iterative optimization mechanism.

[0182] The review model evaluates the paper from multiple dimensions and provides specific review opinions to guide the scientific research agent to make targeted modifications and improvements, forming a closed-loop optimization process.

[0183] The automated scientific research method provided by this application realizes the full-process autonomy from topic selection to the completion of the paper by deeply integrating technologies such as large language models, semantic retrieval, code generation, and simulated review, significantly improving the scientific research efficiency and the quality of results, and laying a foundation for the AI-driven scientific research paradigm.

[0184] Based on the same concept, as Figure 10 shown, the embodiment of this application also provides an automated scientific research system, including:

[0185] An academic search engine 10, which is used to obtain existing research literature from preprint platforms.

[0186] A scientific research agent 20, which is configured to generate a paper outline based on the obtained existing research literature; and generate an experimental design plan for the paper based on the obtained existing research literature and the paper outline.

[0187] An experimental result generation module 30, which is configured to generate experimental results according to the experimental design plan of the paper.

[0188] The scientific research intelligent agent is also configured to generate a complete paper based on the obtained existing research literature, the paper outline, the experimental design scheme of the paper, and the experimental results; wherein, the scientific research intelligent agent is obtained by performing supervised learning training on a first large language model.

[0189] The review model 40 is configured to review the complete paper to obtain review opinions; wherein, the review model is obtained by performing supervised learning training on a second large language model using a review data set.

[0190] The scientific research intelligent agent is also configured to optimize the complete paper according to the review opinions to obtain a new paper that meets the preset criteria.

[0191] In the embodiment of the present application, the scientific research intelligent agent has a first function, a second function, and a third function. Among them, the first function of the scientific research intelligent agent is obtained by performing first supervised learning training on a first large language model using a first data set, the second function of the scientific research intelligent agent is obtained by performing second supervised learning training on the first large language model using a second data set, and the third function of the scientific research intelligent agent is obtained by performing third supervised learning training on the first large language model using a third data set.

[0192] Specifically, the first data set includes first training samples, the first training samples include first input features and first output labels, the first input features use the references of the paper to be generated, and the first output labels use the outline of the paper to be generated.

[0193] The second data set includes second training samples, the second training samples include second input features and second output labels, the second input features use the references of the paper to be generated and the outline of the paper to be generated, and the second output labels use the experimental design scheme of the paper to be generated.

[0194] In the above embodiment, the third data set includes third training samples, the third training samples include third input features and third output labels, the third input features use the references of the paper to be generated, the outline of the paper to be generated, the experimental design scheme of the paper to be generated, and the experimental results, and the third output labels use the complete paper to be generated.

[0195] In the above embodiment, the step of performing first supervised learning training on the first large language model using the first data set to obtain the first function of the scientific research intelligent agent includes:

[0196] Construct a first data set;

[0197] Determine the first loss function and the first optimizer; wherein, the first loss function is used to measure the difference between the model prediction result and the first output label, and the first optimizer is used to adjust the model parameters to minimize the loss function;

[0198] Input the first input feature into the first large language model to obtain the first-stage prediction result;

[0199] Calculate the loss value between the first-stage prediction result and the first output label, calculate the gradient of the loss function with respect to the model parameters through the backpropagation algorithm, and the first optimizer updates the model parameters according to the gradient to gradually reduce the loss value until the loss value converges or reaches the preset number of training epochs, completing the first supervised learning training of the first large language model to obtain the first function of the scientific research intelligent agent.

[0200] In the above embodiment, the second supervised learning training of the first large language model using the second data set to obtain the second function of the scientific research intelligent agent includes:

[0201] Construct a second data set;

[0202] Determine the second loss function and the second optimizer; wherein, the second loss function is used to measure the difference between the model prediction result and the second output label, and the second optimizer is used to adjust the model parameters to minimize the loss function;

[0203] Input the second input feature into the second large language model to obtain the second-stage prediction result;

[0204] Calculate the loss value between the second-stage prediction result and the second output label, calculate the gradient of the loss function with respect to the model parameters through the backpropagation algorithm, and the second optimizer updates the model parameters according to the gradient to gradually reduce the loss value until the loss value converges or reaches the preset number of training epochs, completing the second supervised learning training of the first large language model to obtain the second function of the scientific research intelligent agent.

[0205] In the above embodiment, the third supervised learning training of the first large language model using the third data set to obtain the third function of the scientific research intelligent agent includes:

[0206] Construct a third data set; wherein, the third data set includes third training samples, the third training samples include third input features and third output labels, the third input features adopt the references of the paper to be generated, the outline of the paper to be generated, the experimental design scheme and experimental results of the paper to be generated, and the third output label adopts the complete paper to be generated;

[0207] Determine the third loss function and the third optimizer; wherein, the third loss function is used to measure the difference between the model prediction result and the third output label, and the third optimizer is used to adjust the model parameters to minimize the loss function;

[0208] Input the third input feature into the first large language model to obtain the third stage prediction result;

[0209] Calculate the loss value between the third stage prediction result and the third output label, calculate the gradient of the loss function with respect to the model parameters through the backpropagation algorithm, and the third optimizer updates the model parameters according to the gradient to gradually reduce the loss value until the loss value converges or reaches the preset number of training rounds, completing the third supervised learning training of the first large language model to obtain the third function of the scientific research intelligent agent.

[0210] In the above embodiment, obtaining the experimental results according to the experimental design scheme of the paper includes:

[0211] Input the experimental design scheme of the paper into the large language model in the existing code field;

[0212] The large language model in the code field modifies the preset baseline code according to the experimental design scheme to obtain the modified code;

[0213] Use the code interpreter to run the modified code to generate experimental results;

[0214] Or,

[0215] Execute the experimental design scheme of the paper in a manual operation manner to obtain the corresponding experimental results;

[0216] Process the experimental results in a preset format and input the processed experimental results into the scientific research intelligent agent.

[0217] In the above embodiment, optimizing the complete paper according to the review opinions to obtain a new paper that meets the preset standards includes:

[0218] Input the generated complete paper into the review model to output the review opinions of the paper;

[0219] The scientific research intelligent agent iteratively optimizes the complete paper according to the review opinions until a new paper that meets the preset standards is obtained;

[0220] Among them, the review model is obtained by performing supervised learning training on the second large language model using the review data set.

[0221] In the above embodiment, performing supervised learning training on the second large language model using the review data set to obtain the review model includes:

[0222] Construct a review data set; wherein, the review data set includes review training samples, and the review training samples include review input features and corresponding review output labels. Among them, the review input features are the obtained existing papers, and the review output labels include review opinions with multiple first review weights and review opinions with second review weights;

[0223] Determine a fourth loss function and a fourth optimizer; wherein, the fourth loss function is used to measure the difference between the model prediction result and the review output label, and the fourth optimizer is used to adjust the model parameters to minimize the loss function;

[0224] Input the review input features into the second large language model to obtain a second prediction result;

[0225] Calculate the loss value between the second prediction result and the review output label, calculate the gradient of the loss function with respect to the model parameters through the backpropagation algorithm, and the second optimizer updates the model parameters according to the gradient to gradually reduce the loss value until the loss value converges or reaches the preset number of training rounds, completing the supervised learning training of the second large language model to obtain the review model.

[0226] Based on the same concept, an embodiment of the present application further provides an electronic device, including a processor and a memory. An executable program is stored in the memory, and the processor executes the executable program to perform the steps of the method described above.

[0227] The storage medium in this embodiment may be included in an electronic device / system; it may also exist alone without being assembled into the electronic device / system. The above storage medium carries one or more programs, and when the above one or more programs are executed, the method according to the embodiment of the present application is implemented.

[0228] It should be understood that the first, second, third, fourth, and various digital numbers involved herein are only for the convenience of description and are not used to limit the scope of the present application.

[0229] It should also be understood that the term "and / or" herein is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character " / " herein generally represents an "or" relationship between the associated objects before and after.

[0230] In the implementation process, each step of the above method can be completed by the integrated logic circuit of the hardware in the processor or the instructions in the form of software. The steps of the method disclosed in combination with the embodiments of the present application can be directly embodied as being executed and completed by the hardware processor, or executed and completed by the combination of the hardware and software modules in the processor. The software module can be located in a mature storage medium in the art such as random access memory, flash memory, read-only memory, programmable read-only memory, or electrically erasable programmable memory, registers, etc. This storage medium is located in the memory, and the processor reads the information in the memory and combines its hardware to complete the steps of the above method. To avoid repetition, it will not be described in detail here.

[0231] In various embodiments of the present application, the sequence numbers of the above processes do not mean the order of execution. The order of execution of each process should be determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present application.

[0232] Those of ordinary skill in the art can realize that the various illustrative logical blocks (ILBs) and steps described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professionals can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present application.

[0233] In several embodiments provided by the present application, it should be understood that the disclosed automated scientific research method and system can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units (or modules) is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection can be through some interfaces. The indirect coupling or communication connection of the device or unit can be in an electrical, mechanical or other form.

[0234] The units described as separate components may or may not be physically separated. The components displayed as units may or may not be physical units, that is, they can be located in one place, or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0235] In addition, in each embodiment of the present application, each functional unit may be integrated into one processing unit, may exist physically alone for each unit, or two or more units may be integrated into one unit.

[0236] In the above embodiments, it may be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center by wire (such as coaxial cable, optical fiber, digital subscriber line) or wirelessly (such as infrared, wireless, microwave, etc.). The computer-readable storage medium may be any available medium that can be accessed by a computer, or a data storage device such as a server, data center, etc. that includes one or more integrated available media. The available medium may be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a DVD), or a semiconductor medium (for example, a solid-state drive), etc.

[0237] As described above, only the specific embodiments of the present application are provided, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed in the present application can easily think of changes or substitutions, which should all be covered by the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims

1. An automated scientific research method, characterized in that: include: Based on the acquired existing research literature, a paper outline is generated using a first function of the scientific research agent; wherein the first function of the scientific research agent is obtained by performing a first supervised learning training on a first language model using a first data set; Based on the acquired existing research literature and the paper outline, the experimental design scheme of the paper is generated using the second function of the scientific research agent; wherein the second function of the scientific research agent is obtained by performing a second supervised learning training on the first language model using the second data set; The experimental results were obtained according to the experimental design of the paper; Based on the acquired existing research literature, the paper outline, the experimental design plan of the paper and the experimental results, a complete paper is generated using the third function of the scientific research agent; wherein the third function of the scientific research agent is obtained by performing third supervised learning training on the first language model using a third data set; The complete paper is optimized according to the review comments to obtain a new paper that meets the preset standards.

2. The method according to claim 1, characterized in that Based on the existing research literature obtained, the first function of the scientific research agent is used to generate a paper outline, including: Construct a knowledge graph based on the existing research literature in the relevant field; the knowledge graph uses triples as the knowledge representation form; Obtain classic and cutting-edge papers in related fields in the knowledge graph; Input classic papers and cutting-edge papers into the scientific research agent, and use the first function of the scientific research agent to generate paper outlines.

3. The method according to claim 2, characterized in that The constructing of the domain knowledge graph includes: Semantic analysis of the acquired research topics based on semantically driven retrieval strategies; Get relevant papers based on semantic analysis results; Integrate the acquired relevant papers and their reference papers and reference relations to obtain triples; Use triples to build domain knowledge graphs.

4. The method according to claim 1, characterized in that: The first data set is used to perform the first supervised learning training on the first language model to obtain the first function of the scientific research agent, including: Constructing a first data set; wherein the first data set includes a first training sample, the first training sample includes a first input feature and a first output label, the first input feature uses a reference of a paper to be generated, and the first output label uses an outline of the paper to be generated; Determine a first loss function and a first optimizer; wherein the first loss function is used to measure the difference between the model prediction result and the first output label, and the first optimizer is used to adjust the model parameters to minimize the loss function; Input the first input feature into the first language model to obtain the first stage prediction result; Calculate the loss value of the first stage prediction result and the first output label, calculate the gradient of the loss function to the model parameters through the back propagation algorithm, and the first optimizer updates the model parameters according to the gradient to gradually reduce the loss value until the loss value converges or reaches a preset number of training rounds, thereby completing the first supervised learning training of the first large language model and obtaining the first function of the scientific research agent.

5. The method according to claim 1, characterized in that The second supervised learning training of the first language model is performed using the second data set to obtain the second function of the scientific research agent, including: Constructing a second data set; wherein the second data set includes a second training sample, the second training sample includes a second input feature and a second output label, the second input feature uses the references of the paper to be generated and the outline of the paper to be generated, and the second output label uses the experimental design scheme of the paper to be generated; Determine a second loss function and a second optimizer; wherein the second loss function is used to measure the difference between the model prediction result and the second output label, and the second optimizer is used to adjust the model parameters to minimize the loss function; Input the second input feature into the second largest language model to obtain the second stage prediction result; Calculate the loss value of the second stage prediction result and the second output label, calculate the gradient of the loss function to the model parameters through the back propagation algorithm, and the second optimizer updates the model parameters according to the gradient to gradually reduce the loss value until the loss value converges or reaches the preset number of training rounds, thereby completing the second supervised learning training of the first large language model and obtaining the second function of the scientific research agent.

6. The method according to claim 1, characterized in that The third data set is used to perform third supervised learning training on the first language model to obtain the third function of the scientific research agent, including: Constructing a third data set; wherein the third data set includes a third training sample, the third training sample includes a third input feature and a third output label, the third input feature adopts the references of the paper to be generated, the outline of the paper to be generated, the experimental design scheme and experimental results of the paper to be generated, and the third output label adopts the complete paper to be generated; Determine a third loss function and a third optimizer; wherein the third loss function is used to measure the difference between the model prediction result and the third output label, and the third optimizer is used to adjust the model parameters to minimize the loss function; Input the third input feature into the first language model to obtain the third stage prediction result; Calculate the loss value of the third stage prediction result and the third output label, calculate the gradient of the loss function to the model parameters through the back propagation algorithm, and the third optimizer updates the model parameters according to the gradient to gradually reduce the loss value until the loss value converges or reaches a preset number of training rounds, thereby completing the third supervised learning training of the first language model and obtaining the third function of the scientific research agent.

7. The method according to claim 1, characterized in that The experimental results were obtained according to the experimental design of the paper, including: Input the experimental design of the paper into the existing code domain language model; The code domain large language model modifies the preset baseline code according to the experimental design plan to obtain the modified code; Run the modified code using the code interpreter to generate experimental results; or, Execute the experimental design of the paper manually to obtain the corresponding experimental results; The experimental results are processed according to a preset format, and the processed experimental results are input into the scientific research agent.

8. The method according to claim 1, characterized in that The complete paper is optimized according to the review comments to obtain a new paper that meets the preset standards, including: Input the generated complete paper into the review model, and output the review opinions of the paper; The scientific research agent iteratively optimizes the complete paper according to the review opinions until a new paper that meets the preset standards is obtained; The review model is obtained by using the review data set to perform supervised learning training on the second largest language model.

9. The method according to claim 7, characterized in that: The method of using the review data set to perform supervised learning training on the second largest language model to obtain the review model includes: Constructing a review data set; wherein the review data set includes review training samples, and the review training samples include review input features and corresponding review output labels, wherein the review input features are obtained existing papers, and the review output labels include multiple review opinions of the first review weight and the review opinions of the second review weight; Determine a fourth loss function and a fourth optimizer; wherein the fourth loss function is used to measure the difference between the model prediction result and the review output label, and the fourth optimizer is used to adjust the model parameters to minimize the loss function; Input the review input features into the second largest language model to obtain a second prediction result; Calculate the loss value of the second prediction result and the review output label, calculate the gradient of the loss function to the model parameters through the back propagation algorithm, and the second optimizer updates the model parameters according to the gradient to gradually reduce the loss value until the loss value converges or reaches a preset number of training rounds, thereby completing the supervised learning training of the second largest language model and obtaining the review model.

10. An automated scientific research system, characterized in that: include: Academic search engines for obtaining existing research literature from preprint platforms; The scientific research agent is configured to generate a paper outline based on the acquired existing research literature; and generate an experimental design plan for the paper based on the acquired existing research literature and the paper outline; An experimental result generation module is configured to generate experimental results according to the experimental design scheme of the paper; The scientific research agent is further configured to generate a complete paper based on the acquired existing research literature, the paper outline, the experimental design scheme of the paper and the experimental results; wherein the scientific research agent is obtained by supervised learning training of the first language model; A review model is configured to review the complete paper and obtain review opinions; wherein the review model is obtained by using the review data set to perform supervised learning training on the second largest language model; The scientific research agent is also configured to optimize the complete paper based on the review opinions to obtain a new paper that meets preset standards.

Images