Multi-agent collaborative scientific hypothesis automatic generation method

Abstract

The invention discloses a multi-agent collaborative scientific hypothesis automatic generation method, and belongs to the field of creative scientific research engineering, and the method comprises the steps: inputting a to-be-researched problem and research background information of the to-be-researched problem into a creative generation agent, so as to obtain a plurality of creative hypothesis texts of the to-be-researched problem; respectively inputting the plurality of creative hypothesis texts into a hypothesis testing agent to obtain evaluation data of each creative hypothesis text under a plurality of evaluation dimensions; and according to the evaluation data of each creative hypothesis text under the plurality of evaluation dimensions, determining a target creative hypothesis text corresponding to the to-be-researched problem from all creative hypothesis texts. Through cooperative work of the creative generation agent and the hypothesis test agent, the problems of insufficient innovation, one-sided evaluation result and incomplete reasoning chain in the automatic hypothesis process are improved, and the overall working efficiency and decision quality of researchers in complex scientific research tasks are improved.

Description

Technical Field

[0001] This application belongs to the field of creative scientific research engineering, and specifically relates to a method, device, equipment and storage medium for the automatic generation of scientific hypotheses through multi-agent collaboration. Background Technology

[0002] In scientific research, researchers typically need to propose innovative scientific hypotheses around specific research questions and conduct comprehensive evaluations of these hypotheses in terms of feasibility, rationality, and potential risks to determine the most valuable target hypothesis. As the complexity of research problems continues to increase, researchers often need to process large amounts of interdisciplinary background information and generate and screen multiple candidate hypotheses within a limited timeframe. Therefore, there is a greater demand for intelligent and systematic support for the hypothesis generation and verification process.

[0003] In existing technologies, rule-based automated methods are commonly used to assist researchers in generating scientific hypotheses. For example, some methods generate several candidate hypotheses by inputting the research question into a language model, or by using knowledge graphs to perform association analysis on the input question to assist in hypothesis generation. Meanwhile, other technologies perform preliminary screening of candidate hypotheses by performing simple text similarity calculations, rule-based logical verification, or scoring on a single evaluation index.

[0004] However, when dealing with complex scientific research problems, such technical solutions are prone to problems such as insufficient originality in generated hypotheses, one-sided evaluation results, and incomplete reasoning chains, making it difficult to effectively determine the most suitable target creative hypothesis text. As a result, researchers still need to invest a lot of manual effort in screening and judging, which is inefficient. Summary of the Invention

[0005] This application aims to provide a method, apparatus, device, and storage medium for the automatic generation of scientific hypotheses through multi-agent collaboration, which at least solves the problems of inefficiency in the automatic generation and screening process of scientific hypotheses caused by insufficient innovativeness, one-sided hypothesis evaluation results, and incomplete reasoning chains.

[0006] In a first aspect, embodiments of this application disclose a method for automatically generating scientific hypotheses through multi-agent collaboration, comprising: The research question and its background information are input into the creative generation agent to obtain multiple creative hypothesis texts for the research question. The multiple creative hypothesis texts are input into the hypothesis testing agent to obtain evaluation data for each creative hypothesis text under multiple evaluation dimensions. Based on the evaluation data of each of the creative hypothesis texts under multiple evaluation dimensions, the target creative hypothesis text corresponding to the research question is determined from all the creative hypothesis texts.

[0007] Secondly, embodiments of this application also disclose an automatic scientific hypothesis generation device for multi-agent collaboration, comprising: The idea generation module is used to input the research question and its research background information into the idea generation agent to obtain multiple creative hypothesis texts for the research question. The creative evaluation module is used to input multiple creative hypothesis texts into the hypothesis testing agent to obtain evaluation data for each creative hypothesis text under multiple evaluation dimensions. The creative screening module is used to determine the target creative hypothesis text corresponding to the research question from all the creative hypothesis texts based on the evaluation data of each creative hypothesis text under multiple evaluation dimensions.

[0008] Thirdly, embodiments of this application also disclose an electronic device, including a processor and a memory, wherein the memory stores a program or instructions that can run on the processor, and the program or instructions, when executed by the processor, implement the steps of the method described in the first aspect.

[0009] Fourthly, embodiments of this application also disclose a readable storage medium storing a program or instructions that, when executed by a processor, implement the steps of the method described in the first aspect.

[0010] In summary, in this embodiment, by inputting the research question and its background information into the creative generation agent, the system can generate multiple creative hypothesis texts with differentiated perspectives, supported by structured domain knowledge. This significantly enhances the diversity and potential innovativeness of the generated hypotheses, thereby reducing the risk of insufficient innovativeness due to a single generation approach. Furthermore, by inputting these multiple creative hypothesis texts into the hypothesis testing agent and evaluating each text from multiple dimensions based on structured domain knowledge, the system can comprehensively analyze the hypotheses from multiple angles, improving the comprehensiveness and reliability of the evaluation results and reducing the uncertainty caused by biased evaluations or incomplete reasoning chains. Finally, based on evaluation data across multiple dimensions, the system determines the target creative hypothesis text from all the creative hypothesis texts. This allows the system to systematically screen candidate hypotheses based on multi-agent collaborative reasoning, improving the scientific rigor and rationality of the target hypothesis selection with the support of multi-dimensional evaluation results. This reduces the workload of researchers in the manual screening process, increases hypothesis screening efficiency, and reduces bias caused by the subjectivity of human judgment. Therefore, the method based on the embodiments of this application, through the collaborative work of the creative generation agent and the hypothesis testing agent, makes the hypothesis generation process more diversified, the hypothesis testing process more systematic, and the hypothesis screening process more scientific, thereby improving the problems of insufficient innovation, one-sided evaluation results, and incomplete reasoning chains in the automatic hypothesis process, and improving the overall work efficiency and decision-making quality of researchers in complex scientific research tasks. Attached Figure Description

[0011] In the attached diagram: Figure 1 This is a flowchart illustrating the steps of an automatic scientific hypothesis generation method for multi-agent collaboration provided in an embodiment of this application. Figure 2 This is a flowchart of another method for automatically generating scientific hypotheses through multi-agent collaboration provided in an embodiment of this application; Figure 3 This application provides an automated scientific discovery system. Figure 4 This application provides a complete scientific research discovery process in its embodiments; Figure 5 This is a block diagram of a multi-agent collaborative scientific hypothesis automatic generation device provided in an embodiment of this application; Figure 6 This is a block diagram of an electronic device provided in one embodiment of this application. Detailed Implementation

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0014] like Figure 1 The image shows an embodiment of the present application that provides a method for automatically generating scientific hypotheses through multi-agent collaboration.

[0015] The method may include the following steps: Step 101: Input the research question and its research background information into the creative agent to obtain multiple creative hypothesis texts for the research question.

[0016] In some embodiments of this application, to ensure that the subsequently generated creative hypothesis texts cover the semantic and background scope of the research problem, the research problem and its background information are input into the creative generation agent to obtain multiple creative hypothesis texts for the research problem. During this process, the creative generation agent utilizes structured domain knowledge to perform semantic parsing and relational expansion on the input content. This structured domain knowledge is used to represent entities, attributes, and their relationships within the domain in graph structures or other forms, supporting the creative generation agent in deriving hypotheses based on knowledge associations. This allows for diverse perspectives and knowledge support when generating creative hypothesis texts, thereby enhancing the diversity and potential innovativeness of the generated hypotheses and reducing the risk of insufficient innovativeness due to a single generation approach.

[0017] In a specific example, the system receives the research question input by the researcher: "Does a certain cytokine affect the immunosuppressive mechanism in the tumor microenvironment?", along with relevant research background information, including the cytokine's known functions, related signaling pathways, and potential mechanisms of action mentioned in existing literature. This question and background information can then be input into the creative agent, which further supplements the background information based on a literature review generation module. This allows the creative agent to generate multiple creative hypothesis texts based on a richer knowledge base. This will output multiple creative hypothesis texts surrounding the cytokine's mechanism of action, such as different hypotheses regarding its impact on immune cell migration, metabolic regulation, or signaling pathway interactions, providing diverse candidate hypothesis inputs for the subsequent hypothesis testing phase.

[0018] Step 102: Input multiple creative hypothesis texts into the hypothesis testing agent to obtain evaluation data for each creative hypothesis text under multiple evaluation dimensions.

[0019] In some embodiments of this application, to provide structured and comparable evaluation data for subsequent hypothesis screening, multiple creative hypothesis texts are input into the hypothesis testing agent to obtain evaluation data for each creative hypothesis text across multiple evaluation dimensions. During this process, the hypothesis testing agent can perform semantic reasoning, logical analysis, and multi-faceted argumentation on the creative hypothesis texts based on structured domain knowledge, reasoning models, or external professional tools. Structured domain knowledge is used to represent entities, attributes, and their relationships within the domain, supporting the hypothesis testing agent in performing semantic association judgments and logical support analysis during the evaluation process. This yields evaluation data covering multiple dimensions such as feasibility, rationality, potential risks, innovativeness, and credibility, thereby improving the comprehensiveness and reliability of the hypothesis evaluation results and reducing the uncertainty caused by biased evaluations or incomplete reasoning chains.

[0020] In a specific example, after receiving multiple creative hypothesis texts regarding "whether a certain cytokine affects the immunosuppressive mechanism of the tumor microenvironment," the system inputs these hypothesis texts into a hypothesis testing agent. The hypothesis testing agent can then perform semantic reasoning based on structured domain knowledge and can also invoke external analysis tools (such as statistical analysis tools for analyzing gene expression data) to assist in verifying the biological mechanisms involved in the hypotheses. This generates evaluation data for each creative hypothesis text, encompassing multiple dimensions such as feasibility, rationality, potential risks, innovativeness, and credibility, providing structured quantitative evidence for subsequent hypothesis screening steps.

[0021] Step 103: Based on the evaluation data of each creative hypothesis text under multiple evaluation dimensions, determine the target creative hypothesis text corresponding to the research question from all creative hypothesis texts.

[0022] In some embodiments of this application, in order to determine the target creative hypothesis text that best meets the needs of the research problem from among the candidate hypotheses, and thus provide researchers with hypotheses with higher research value, the target creative hypothesis text corresponding to the research problem is determined from all creative hypothesis texts based on evaluation data of each creative hypothesis text under multiple evaluation dimensions. Multiple evaluation dimensions are used to quantitatively describe the creative hypothesis texts from different perspectives such as feasibility, rationality, potential risks, innovativeness, and credibility, to support the system in making comprehensive judgments from multiple perspectives during the screening process. In this way, candidate hypotheses are systematically screened with the support of multi-dimensional evaluation data, thereby improving the scientific rigor and rationality of the target hypothesis selection and reducing biases caused by the subjectivity of human judgment.

[0023] In a specific example, the system has obtained evaluation data from multiple creative hypothesis texts regarding whether a certain cytokine affects the immunosuppressive mechanism of the tumor microenvironment. These evaluation data cover multiple evaluation dimensions, including feasibility, rationality, potential risks, novelty, and credibility. At this point, a scoring weighting strategy can be used to assign different weights to different evaluation dimensions; for example, assigning higher weights to novelty and credibility, and lower weights to potential risks, to reflect the research task's focus on different evaluation dimensions. Finally, based on the weighted comprehensive score, the system selects the target creative hypothesis text that best meets the research needs from all the creative hypothesis texts, providing a more valuable hypothesis output for subsequent research.

[0024] In summary, in this embodiment, by inputting the research question and its background information into the creative generation agent, the system can generate multiple creative hypothesis texts with differentiated perspectives, supported by structured domain knowledge. This significantly enhances the diversity and potential innovativeness of the generated hypotheses, thereby reducing the risk of insufficient innovativeness due to a single generation approach. Furthermore, by inputting these multiple creative hypothesis texts into the hypothesis testing agent and evaluating each text from multiple dimensions based on structured domain knowledge, the system can comprehensively analyze the hypotheses from multiple angles, improving the comprehensiveness and reliability of the evaluation results and reducing the uncertainty caused by biased evaluations or incomplete reasoning chains. Finally, based on evaluation data across multiple dimensions, the system determines the target creative hypothesis text from all the creative hypothesis texts. This allows the system to systematically screen candidate hypotheses based on multi-agent collaborative reasoning, improving the scientific rigor and rationality of the target hypothesis selection with the support of multi-dimensional evaluation results. This reduces the workload of researchers in the manual screening process, increases hypothesis screening efficiency, and reduces bias caused by the subjectivity of human judgment. Therefore, the method based on the embodiments of this application, through the collaborative work of the creative generation agent and the hypothesis testing agent, makes the hypothesis generation process more diversified, the hypothesis testing process more systematic, and the hypothesis screening process more scientific, thereby improving the problems of insufficient innovation, one-sided evaluation results, and incomplete reasoning chains in the automatic hypothesis process, and improving the overall work efficiency and decision-making quality of researchers in complex scientific research tasks.

[0025] Figure 2 This is another method for automatically generating scientific hypotheses through multi-agent collaboration provided in the embodiments of this application.

[0026] The method may include the following steps: Step 201: Input the research question and its research background information into the creative agent to obtain multiple creative hypothesis texts for the research question.

[0027] The method shown in this step has been explained in step 101 and will not be repeated here.

[0028] Optionally, step 201 includes the following sub-steps: Sub-step 2011 involves performing a knowledge graph association analysis on the research problem and its research background information to obtain structured domain knowledge about the research problem.

[0029] In some embodiments of this application, to provide a semantic foundation and knowledge support for subsequent innovation point discovery and hypothesis generation, it is necessary to perform knowledge graph association analysis on the research problem and its research background information to obtain structured domain knowledge of the research problem. Knowledge graph association analysis is used to semantically parse and expand the input information based on a graph structure of entities, attributes, and their relationships, enabling the system to identify key concepts, potential causal chains, and related domain knowledge involved in the research problem. This allows the system to obtain structured domain knowledge built around the research problem, thereby improving the semantic accuracy and logical support of subsequent innovation point identification and hypothesis generation.

[0030] In a specific example, the system receives the research question input by the researcher: "Does a certain cytokine affect the immunosuppressive mechanism in the tumor microenvironment?", along with relevant research background information, such as the known functions of the cytokine, related signaling pathways, immune cell types, and their interactions. The system then utilizes a knowledge graph association analysis module to extract entities, identify relationships, and semantically link the information. This is combined with a literature review generation module to supplement the information with the latest research progress related to the cytokine, enabling the structured domain knowledge to cover a more comprehensive range of knowledge. The output will then contain structured domain knowledge encompassing cytokines, immune cells, signaling pathways, and their interrelationships, providing a semantic foundation for subsequently identifying candidate innovation datasets.

[0031] Sub-step 2012: Based on the research problem, determine the candidate innovation point dataset for the research problem from structured domain knowledge.

[0032] The candidate innovation point dataset contains at least one innovation point description text for the research problem.

[0033] In some embodiments of this application, in order to identify potential innovations related to the research problem from already structured domain knowledge, so as to provide innovative directions with research value for subsequent hypothesis generation, it is necessary to determine a candidate innovation point dataset for the research problem from the structured domain knowledge. This candidate innovation point dataset contains at least one innovation point descriptive text for the research problem. The candidate innovation point descriptive text is used to characterize key concepts, potential causal relationships, or insufficiently explored mechanisms that may lead to new research ideas in natural language or structured form. This enables the system to construct diverse and creative hypothesis texts based on these innovation points during the hypothesis generation stage. Ultimately, the system obtains a candidate innovation point dataset built around the research problem, thereby improving the directionality and innovativeness of subsequent hypothesis generation and reducing the risk of limited hypothesis generation scope due to a lack of innovation point identification.

[0034] In a specific example, the system has acquired structured domain knowledge regarding "whether a certain cytokine affects the immunosuppressive mechanism in the tumor microenvironment," which includes entities and relationships such as cytokines, immune cell types, signaling pathways, and their interactions. At this point, based on an adaptive innovation point expansion strategy, the system identifies potential innovation points related to this cytokine from the structured domain knowledge, such as metabolic pathways it may regulate, immune cell migration behavior it may affect, or signaling interaction mechanisms it may participate in. This information is then used to generate corresponding innovation point description texts. This process outputs a candidate innovation point dataset containing multiple innovation point description texts, providing an input foundation for subsequently generating multiple creative hypothesis texts with innovative directions.

[0035] Sub-step 2013 involves performing semantic consistency detection and / or logical conflict detection on the description text of each innovation point in the candidate innovation point dataset, and updating the candidate innovation point dataset based on the detection results of each innovation point description text.

[0036] In some embodiments of this application, to eliminate semantically incoherent or logically contradictory innovation points and thus ensure that the subsequently generated creative hypothesis texts have higher semantic rationality and logical reliability, semantic consistency detection and / or logical conflict detection are performed on the description text of each innovation point in the candidate innovation point dataset, and the candidate innovation point dataset is updated based on the detection results of each innovation point description text. Semantic consistency detection is used to determine whether the semantic expression within the innovation point description text is coherent and consistent with structured domain knowledge, while logical conflict detection is used to identify whether the innovation point description text contains contradictory causal relationships or logical chains that do not conform to domain knowledge. This enables the system to obtain a filtered and corrected candidate innovation point dataset, thereby improving the semantic accuracy and logical support of subsequent creative hypothesis generation and reducing hypothesis generation bias caused by poor innovation point quality.

[0037] In a specific example, the system has obtained a dataset of candidate innovative points regarding "whether a certain cytokine affects the immunosuppressive mechanism of the tumor microenvironment." This dataset contains multiple descriptive texts for innovative points, such as "this cytokine enhances the migration ability of immunosuppressive cells" or "this cytokine inhibits the migration ability of immunosuppressive cells." The system then uses a semantic consistency detection module to determine whether each innovative point description is consistent with entity relationships in structured domain knowledge, and a logical conflict detection module to identify any logical chains that contradict known biological mechanisms. This updates the candidate innovative point dataset, for example, by retaining semantically coherent and logically sound innovative point descriptions and removing or correcting texts with semantic inconsistencies or logical conflicts, thus providing higher-quality innovative point input for subsequent generation of creative hypothesis texts.

[0038] Sub-step 2014 generates a creative hypothesis text based on the description text of each innovation point in the innovation point dataset.

[0039] In some embodiments of this application, to provide candidate hypotheses with clear research directions and logical foundations for the subsequent hypothesis testing stage, a creative hypothesis text is generated based on the description text of each innovation point in the innovation point dataset. The creative hypothesis text is used to express potential scientific hypotheses surrounding the innovation point in natural language or a structured form, enabling the system to perform reasoning analysis and multi-dimensional evaluation of these hypotheses in subsequent steps. The structured representation can be used to organize key entities, attributes, and their relationships in the creative hypothesis text in the form of graph structures or semantic vectors to support the execution of subsequent knowledge graph reasoning analysis. This yields creative hypothesis texts corresponding to each innovation point, thereby improving the directionality and logical coherence of the hypothesis generation process and reducing the risk of unstable hypothesis generation quality due to the ineffective utilization of innovation points.

[0040] In a specific example, the system has obtained a dataset of candidate innovative points regarding "whether a certain cytokine affects the immunosuppressive mechanism of the tumor microenvironment." This dataset contains multiple descriptive texts of innovative points, such as "this cytokine may regulate the migration behavior of immunosuppressive cells" or "this cytokine may affect tumor-related metabolic pathways." At this point, corresponding creative hypothesis texts can be generated based on each innovative point description text. A structured representation method is then used to semantically organize the key entities in the hypothesis (such as cytokines, immune cells, and signaling pathways) and their potential relationships, thereby outputting multiple creative hypothesis texts, such as "this cytokine affects the immunosuppressive mechanism of the tumor microenvironment by regulating the migration behavior of immunosuppressive cells." This provides clear hypothetical input for subsequent knowledge graph reasoning analysis and argumentation.

[0041] Step 202: Based on the structured domain knowledge of the creative hypothesis text, perform knowledge graph reasoning analysis on each creative hypothesis text to obtain the reasoning analysis results for each creative hypothesis text.

[0042] The reasoning analysis results are used to characterize the degree of semantic association and / or logical support between each creative hypothesis text and structured domain knowledge.

[0043] In some embodiments of this application, to facilitate the identification of the correlation and support level between hypotheses and domain knowledge, thereby providing a basis for subsequent multi-dimensional evaluation, it is necessary to perform knowledge graph reasoning analysis on each creative hypothesis text based on structured domain knowledge to obtain the reasoning analysis results for each creative hypothesis text. Structured domain knowledge is used to represent entities, attributes, and their relationships within the domain in the form of graph structures, tree structures, or semantic vectors, enabling knowledge graph reasoning analysis to determine the correlation between the hypothesis text and domain knowledge based on node relationships, semantic paths, and logical chains. Ultimately, reasoning analysis results are obtained to characterize the degree of semantic correlation and / or logical support between each creative hypothesis text and structured domain knowledge, thereby improving the logical integrity and knowledge support strength of the subsequent evaluation process.

[0044] In a specific example, after receiving multiple creative hypothetical texts regarding "whether a certain cytokine affects the immunosuppressive mechanism of the tumor microenvironment," the system performs correlation analysis on these hypothetical texts with structured domain knowledge. In this example, structured domain knowledge can be represented by a graph structure, containing entities and relationships such as cytokines, immune cell types, signaling pathways, and their interactions. The knowledge graph reasoning module can then be used to perform path searching and semantic matching on key entities in the hypothetical texts to determine whether the proposed mechanisms of action have potential logical chains within the knowledge graph. Finally, the system outputs the reasoning analysis results for each creative hypothetical text; for example, one hypothesis may have multiple supporting paths in the knowledge graph, while another hypothesis may only have weakly related paths, thus providing structured logical support for subsequent multi-dimensional evaluation.

[0045] Step 203: Conduct debate arguments between multiple debate agent agents with different expert identities for each creative hypothesis text to obtain the debate argument results for each creative hypothesis text.

[0046] Among them, the results of the debate and argumentation are used to characterize at least one of the following conclusions of the creative hypothesis text: feasibility conclusion, reasonable conclusion, potential risk conclusion, innovative conclusion, and credibility conclusion.

[0047] In some embodiments of this application, to identify the supporting and opposing points of a hypothesis under different knowledge positions, thereby providing more discriminative evidence for subsequent multi-dimensional evaluation, each creative hypothesis text is subjected to debate and argumentation among multiple debate agent agents with different expert identities to obtain the debate and argumentation results for each creative hypothesis text. Multiple different expert identities are used to simulate reasoning methods under different knowledge backgrounds, research positions, or professional preferences, enabling the debate agent agents to argue the hypothesis from different perspectives such as support, opposition, or neutrality. The debate and argumentation results are used to characterize at least one of the following conclusions regarding the feasibility, rationality, potential risks, innovation, and credibility of the creative hypothesis text, supporting multi-dimensional quantitative descriptions of the hypothesis in subsequent evaluation stages. This allows the system to obtain argumentation results covering multiple reasoning perspectives, thereby improving the comprehensiveness and logical support of the hypothesis evaluation process and reducing the risk of bias caused by a single evaluation perspective.

[0048] In a specific example, after receiving multiple creative hypothesis texts regarding "whether a certain cytokine affects the immunosuppressive mechanism of the tumor microenvironment," the system inputs each hypothesis text into multiple debate agents with different expert identities. At this point, based on a dynamic configuration strategy of expert identities, immunology experts, molecular biology experts, and clinical oncology experts can be randomly selected from a candidate expert identity pool as debate participants. This allows debate agents with different expert identities to argue the hypothesis texts from the perspectives of support, opposition, or a comprehensive judgment. This will output the debate argumentation results for each creative hypothesis text. For example, a hypothesis may be highly feasible from the perspective of immunology experts, but pose potential risks from the perspective of clinical experts, thus providing discriminatory evidence for subsequent multi-dimensional evaluation.

[0049] Optionally, step 203 includes the following sub-steps: Sub-step 2031: Input the creative hypothesis text into at least one debate agent with the identity of a first expert to obtain at least one first argument result of the creative hypothesis text.

[0050] Each of the first argument results is used to characterize the data in the structured domain knowledge that supports the creative hypothesis text.

[0051] In some embodiments of this application, to meticulously identify evidence or logical chains within structured domain knowledge that positively support the hypothesis, thereby providing foundational supporting argument data for subsequent comprehensive debate, the creative hypothesis text is input into a debate agent with at least one first expert identity to obtain at least one first argument result for the creative hypothesis text. The first expert identity simulates a professional role holding a supporting stance or having positive preferences in the relevant domain, enabling the debate agent to generate supporting arguments based on related entities, causal paths, or logical supporting relationships within the structured domain knowledge. The first argument result is used to characterize data on the supporting attributes of the creative hypothesis text within the structured domain knowledge, such as the quantity, strength, or completeness of supporting evidence or logical chains. This allows for argument results from a supportive perspective, thereby enhancing the completeness and depth of the subsequent multi-expert debate process.

[0052] In a specific example, after receiving a creative hypothesis text regarding "whether a certain cytokine affects the immunosuppressive mechanism of the tumor microenvironment," the system inputs this hypothesis text into a debate agent with at least one first-level expert identity, such as an immunology expert. During this step, the system uses a dynamic expert identity configuration strategy to select an expert identity relevant to the supporting analysis from a candidate expert identity database. This allows the debate agent to extract supporting evidence from structured domain knowledge, such as reports in the literature that the cytokine can regulate the migration behavior of immunosuppressive cells or affect the activity of related signaling pathways. This will enable the system to output at least one first-level argument, such as "this cytokine has been shown to be related to the function of immunosuppressive cells in multiple studies," thus providing a supporting data foundation for subsequent opposing arguments and comprehensive debate.

[0053] Sub-step 2032: Input the creative hypothesis text into at least one debate agent with the identity of a second expert to obtain at least one second argument result of the creative hypothesis text.

[0054] Each second argument result is used to characterize the data on the opposing properties of the creative hypothesis text in the structured domain knowledge.

[0055] In some embodiments of this application, to meticulously identify evidence in structured domain knowledge that has a potential refutation effect or logical conflict on the hypothesis, thereby providing a basis for reverse argumentation in subsequent comprehensive debates, the creative hypothesis text is input into a debate agent with at least one second expert identity to obtain at least one second argumentation result for the creative hypothesis text. The second expert identity simulates a professional role in the relevant domain who holds an opposing stance or has a critical analytical tendency, enabling the debate agent to generate opposing arguments based on reverse evidence, contradictory relationships, or potential risk factors in the structured domain knowledge. The second argumentation result is used to characterize data on the opposing attributes of the creative hypothesis text in the structured domain knowledge, such as the quantity, strength, or degree of logical conflict of opposing evidence. This allows for the acquisition of argumentation results from an opposing perspective, thereby improving the balance and depth of argumentation in subsequent multi-expert debates and reducing the risk of bias caused by a single supporting analysis.

[0056] In a specific example, after receiving a creative hypothesis text regarding "whether a certain cytokine affects the immunosuppressive mechanism of the tumor microenvironment," the system inputs this hypothesis text into a debate agent with at least one second expert identity, such as a clinical oncology expert. At this point, a dynamic configuration strategy based on the expert identity can be used to select an expert identity with critical analysis capabilities from a candidate expert identity database. This allows the debate agent to extract opposing evidence from structured domain knowledge, such as the cytokine not showing a function related to immunosuppression in some studies, or its mechanism of action containing unresolved logical contradictions. This will enable the system to output at least one second argument, such as "this cytokine has not shown evidence of being related to immunosuppressive cell function in some studies," thus providing a basis for opposing data in subsequent comprehensive debate.

[0057] Sub-step 2033: Input the first argument result and the second argument result together into the debate agent under the identity of the third expert to obtain the debate argument result of the creative hypothesis text.

[0058] In some embodiments of this application, to generate a debate argument result that reflects the overall credibility and logical balance of the creative hypothesis text, the first argument result and the second argument result are jointly input into a debate agent with a third expert identity to obtain the debate argument result of the creative hypothesis text. The third expert identity simulates a professional role with comprehensive judgment capabilities in the relevant field, enabling the debate agent to perform balanced reasoning and comprehensive judgment on the hypothesis text based on structured domain knowledge and the logical relationships between supporting and opposing evidence. The debate argument result is used to characterize the comprehensive conclusions of the creative hypothesis text regarding feasibility, rationality, potential risks, innovativeness, and credibility. This allows the system to obtain a comprehensive argument result that integrates supporting and opposing perspectives, thereby improving the logical integrity and judgment reliability of the subsequent evaluation process and reducing the risk of bias caused by a single perspective.

[0059] In a specific example, the system has obtained the first argument (e.g., supporting evidence from an immunology expert) and the second argument (e.g., opposing evidence from a clinical oncology expert) for a creative hypothesis text regarding "whether a certain cytokine affects the immunosuppressive mechanism of the tumor microenvironment." These two types of argument results can then be input into a debate agent with a third-party expert identity, such as a molecular biology expert. This agent can then comprehensively analyze the supporting and opposing evidence based on structured domain knowledge, identifying the logical relationships and potential conflicts between them. The final output of the creative hypothesis text's debate argument results, such as "this cytokine may affect the tumor microenvironment by regulating the migration behavior of immunosuppressive cells, but its mechanism of action still needs further verification," will provide comprehensive evidence for subsequent multi-dimensional evaluations.

[0060] Optionally, in some embodiments of this application, the debate agent intelligence under the identity of the first expert and the identity of the second expert can also conduct multiple rounds of debate interaction. In this case, the above process can also be supplemented with the following additional steps: Sub-step 2034 updates the creative hypothesis text of the debate agent under the first expert identity with at least one second argument result, and returns to sub-step 2031.

[0061] In some embodiments of this application, to enable the supporting debate agent to conduct more thorough arguments based on new opposing information in subsequent rounds, thereby improving the depth and logical integrity of the debate process, the creative hypothesis text input to the debate agent under the first expert identity can be updated using at least one second argument result, and the process can return to the previous steps to continue executing the supporting arguments. The second argument result is used to characterize the data on opposing attributes of the creative hypothesis text in the structured domain knowledge. By integrating this opposing information into the creative hypothesis text, the supporting debate agent can respond to or supplement the arguments against opposing viewpoints in the next round of debate. This allows the system to gradually improve the argument chain of the creative hypothesis text in multiple rounds of debate, thereby enhancing the dynamism and reasoning depth of the debate process and reducing the argumentation bias caused by insufficient information in a single round of debate.

[0062] In a specific example, the system has obtained a second argument regarding the creative hypothesis text on whether a certain cytokine affects the immunosuppressive mechanism of the tumor microenvironment. For instance, if the debate agent, acting as a clinical oncology expert, points out that the cytokine has not shown a function related to the immunosuppressive mechanism in some studies, this opposing argument can be integrated into the creative hypothesis text. For example, the information "some studies have not observed a direct association between this cytokine and the immunosuppressive mechanism" can be added to the text. The updated creative hypothesis text is then re-entered into the debate agent, acting as the first expert. This allows the supporting debate agent to supplement the opposing arguments or present new supporting evidence in the next round of debate, thereby driving the debate process in a more in-depth direction.

[0063] Sub-step 2035 updates the creative hypothesis text of the debate agent under the input second expert identity with at least one first argument result, and returns to sub-step 2032.

[0064] In some embodiments of this application, to enable the opposing debate agent to conduct more thorough reverse arguments based on new supporting information in subsequent rounds, thereby improving the balance and reasoning depth of the debate process, the creative hypothesis text input to the debate agent under the second expert identity is updated through at least one first argument result, and the process returns to the previous steps to continue executing the opposing argument. The first argument result is used to characterize the data in the structured domain knowledge that supports the creative hypothesis text. By integrating this supporting information into the creative hypothesis text, the opposing debate agent can respond to, question, or propose new opposing evidence against supporting viewpoints in the next round of debate. In this way, the system can gradually enrich the argumentation content of the creative hypothesis text in multiple rounds of debate, thereby improving the dynamism and logical integrity of the debate process and reducing the argumentation bias caused by insufficient information in a single round of debate.

[0065] In a specific example, the system has obtained the first argument for a creative hypothesis text regarding "whether a certain cytokine affects the immunosuppressive mechanism of the tumor microenvironment," such as supporting evidence provided by a debate agent acting as an immunology expert, indicating that the cytokine has been observed to be associated with the migration behavior of immunosuppressive cells in multiple studies. This supporting argument can then be integrated into the creative hypothesis text, for example, by adding the information "existing studies have shown that this cytokine is associated with the migration behavior of immunosuppressive cells." The updated creative hypothesis text is then re-entered into the debate agent acting as a second expert. At this point, the system enables the opposing debate agent to provide counter-arguments against the supporting viewpoints in the next round of debate, such as pointing out that the association may be limited by experimental conditions or that alternative explanations exist, thereby driving the debate process in a more comprehensive direction.

[0066] Optionally, in some embodiments of this application, multiple debate agent intelligent agents are preset to serve as candidates for the first expert identity, the second expert identity, and the third expert identity. To further ensure the fairness of the reasoning process, the following additional steps can be added before sub-step 2031: Sub-step 2030: Randomly select at least one debate agent with the identity of a first expert, at least one debate agent with the identity of a second expert, and one debate agent with the identity of a third expert from among multiple debate agent agents.

[0067] In some embodiments of this application, to reduce reasoning bias caused by fixed expert identity combinations and thus improve the fairness and diversity of the overall debate process, at least one debate agent with a first expert identity, at least one debate agent with a second expert identity, and one debate agent with a third expert identity can be randomly selected from multiple debate agent agents. Multiple debate agent agents are used to simulate expert roles with different knowledge backgrounds, research positions, or reasoning methods. The random selection mechanism dynamically determines the expert identity combinations participating in the debate before each debate, enabling the debate process to cover a wider range of professional perspectives. This allows the system to obtain more diverse reasoning paths and argumentation angles in subsequent debate and argumentation stages, thereby improving the objectivity and reliability of the debate results and reducing systematic bias caused by fixed expert identities.

[0068] In a specific example, the system pre-defines multiple debate agents that can act as debate participants, such as immunology experts, molecular biology experts, clinical oncology experts, statistics experts, and bioinformatics experts. A dynamic expert identity configuration strategy can be used to randomly select an immunology expert as the first expert, a clinical oncology expert as the second, and a molecular biology expert as the third. In the subsequent debate and argumentation phase, these three expert identities are used to conduct supporting arguments, opposing arguments, and comprehensive arguments, respectively, thus creating a more diverse and equitable debate and reasoning process.

[0069] In the embodiments of this application, the debate agent model can employ currently publicly available and widely used general-purpose large language models, such as Generative Pre-trained Transformer (GPT), Bidirectional Encoder Representations from Transformers (BERT), Text-to-Text Transfer Transformer (T5), DeepSeek Model (DeepSeek), Copilot Model (Copilot), and Gemini Multimodal Generative Model (Gemini) as basic model instances. All of these models can simulate reasoning styles, argumentation methods, and knowledge preferences under different expert identities through instruction fine-tuning, role setting, expert identity embedding, or multi-agent collaboration.

[0070] Regarding training data, to enable the debate agent to be applicable to cross-domain creative hypothesis generation and argumentation tasks, it can be trained using public datasets of various types and disciplines. For example, general encyclopedic corpora (such as Wikipedia text sets), cross-domain scientific paper abstract sets (such as arXiv multidisciplinary abstract sets), social science and engineering technology knowledge bases (such as DBpedia, Wikidata), general reasoning and debate datasets (such as MultiNLI, SNLI, DebateSum), open-domain question-answering datasets (such as Natural Questions, HotpotQA), and cross-industry technical reports, policy documents, and engineering case libraries can be used. Through combined training with the above-mentioned multi-source heterogeneous data, the model can acquire capabilities such as cross-domain knowledge extraction, logical reasoning, supporting and opposing argument generation, and multi-round debate interaction, thereby adapting to the domain-independent creative hypothesis generation and evaluation requirements of this application.

[0071] It should be noted that the specific training methods and implementation details of the corresponding models have been fully disclosed and applied in the existing technology, and will not be repeated in this application.

[0072] Step 204: Determine the evaluation data of each creative hypothesis text under multiple evaluation dimensions based on the reasoning analysis results and debate arguments of each creative hypothesis text.

[0073] In some embodiments of this application, in order to generate evaluation data that reflects the quality of creative hypothesis texts from multiple perspectives, thereby providing a quantifiable and comparable basis for subsequent hypothesis screening, evaluation data for each creative hypothesis text is determined under multiple evaluation dimensions based on its own reasoning analysis results and debate argumentation results. The reasoning analysis results characterize the degree of semantic association and / or logical support between the creative hypothesis text and structured domain knowledge, while the debate argumentation results characterize the argumentation conclusions of the creative hypothesis text in terms of feasibility, rationality, potential risks, innovativeness, and credibility. Multiple evaluation dimensions are used to quantitatively describe the creative hypothesis texts from different angles, supporting the system in making multi-dimensional comprehensive judgments during subsequent screening. The evaluation data obtained in this way, covering multiple dimensions such as semantic association, logical support, feasibility, rationality, potential risks, innovativeness, and credibility, will improve the comprehensiveness and discriminatory power of the hypothesis evaluation process and reduce the risk of bias caused by a single evaluation perspective.

[0074] In a specific example, the system has obtained the reasoning analysis results and debate results of multiple creative hypothetical texts regarding "whether a certain cytokine affects the immunosuppressive mechanism of the tumor microenvironment." At this point, the semantic relevance and logical support reflected in the reasoning analysis results can be integrated with the conclusions regarding feasibility, rationality, potential risks, innovativeness, and credibility reflected in the debate results. Based on an expanded evaluation dimension strategy, quantitative data under multiple evaluation dimensions can be generated for each creative hypothetical text. For example, the system can generate evaluation data such as "semantic relevance 0.82," "feasibility score 0.75," "innovation score 0.91," and "potential risk score 0.40" for a given hypothesis. This will output structured evaluation data covering multiple evaluation dimensions, providing a comparable quantitative basis for subsequently determining the target creative hypothetical text.

[0075] Step 205: Based on the evaluation data of each creative hypothesis text under multiple evaluation dimensions, determine the target creative hypothesis text corresponding to the research question from all creative hypothesis texts.

[0076] The method shown in this step has been explained in step 103 and will not be repeated here.

[0077] Optionally, step 205 includes the following sub-steps: Sub-step 2051 involves comparing the evaluation data of each pair of creative hypothesis texts under different evaluation dimensions to determine the comparative evaluation results between each pair of creative hypothesis texts, and determining the relative evaluation value between each pair of creative hypothesis texts based on the comparative evaluation results between each pair of creative hypothesis texts.

[0078] In some embodiments of this application, to provide quantifiable comparative criteria for subsequent ranking and screening, the evaluation data of each pair of creative hypothesis texts under different evaluation dimensions can be compared pairwise to determine the comparative evaluation results between each pair of creative hypothesis texts. Based on these comparative evaluation results, the relative evaluation value between each pair of creative hypothesis texts can be determined. The comparative evaluation results characterize the differences between two hypotheses in evaluation dimensions such as feasibility, rationality, potential risk, innovativeness, and credibility. The relative evaluation value quantifies these differences into comparable numerical values ​​to support the subsequent ranking process. This allows the system to obtain relative evaluation values ​​covering all hypotheses, thereby improving the comparability and scientific rigor of the subsequent screening process and reducing the risk of bias caused by a single dimension or single score.

[0079] In a specific example, the system has obtained multi-dimensional evaluation data for several creative hypotheses regarding "whether a certain cytokine affects the immunosuppressive mechanism of the tumor microenvironment." Each hypothesis has a feasibility score, a reasonableness score, an innovation score, a potential risk score, and a credibility score. During this step, the system uses a scoring weighting strategy to compare the scores of each pair of hypotheses across all evaluation dimensions. For example, it compares the differences in scores between hypothesis A and hypothesis B in the innovation dimension and the feasibility dimension, integrating these differences into a comparative evaluation result. This generates a relative evaluation value for each pair of hypotheses, such as "the overall relative evaluation value of hypothesis A relative to hypothesis B is +0.18," thus providing a quantifiable basis for comparison in subsequent ranking steps.

[0080] Optionally, in order to determine the relative evaluation value between each pair of creative hypothesis texts based on the comparative evaluation results between each pair of creative hypothesis texts, sub-step 2051 includes the following sub-steps: Sub-step 20511: Based on the initial embedding evaluation values ​​of the two creative hypothesis texts, and using a preset learning rate parameter, the differences between the two creative hypothesis texts under different evaluation dimensions, as represented by the comparative evaluation results, are characterized to obtain the differential features between the two creative hypothesis texts.

[0081] In some embodiments of this application, to quantify the differences between creative hypothesis texts across different evaluation dimensions and provide calculable differential features for subsequent dynamic updates of embedded evaluation values, the initial embedded evaluation values ​​of two creative hypothesis texts can be determined, and the differences between the two creative hypothesis texts across different evaluation dimensions, as represented by the comparative evaluation results, can be characterized using a preset learning rate parameter to obtain differential features between the two creative hypothesis texts. The construction of these differential features can utilize a modeling mechanism similar to the Expected Score Differential (ESD) in the Expected Loss Optimization Rating System (ELO). This involves comparing the multi-dimensional differences reflected in the evaluation results (e.g., differences in feasibility, innovativeness, rationality, potential risk, credibility, etc.) and combining this with the learning rate parameter to map these differences into feature vectors that can be used to update the embedded evaluation values. The learning rate parameter controls the magnitude of the influence of the differential features on the update of the embedded evaluation values, thereby avoiding instability caused by excessively rapid updates or insufficient discriminative power caused by excessively slow updates. This allows the system to obtain the differentiating features between the two creative hypothesis texts, providing structured input for subsequent adjustment of the embedded evaluation value.

[0082] In a specific example, the system has obtained comparative evaluation results of hypotheses A and B across multiple evaluation dimensions. For instance, hypothesis A is higher than hypothesis B in the innovativeness dimension but lower than hypothesis B in the potential risk dimension. The initial embedding evaluation values ​​for hypotheses A and B can be determined first, for example, 0.62 and 0.55 respectively. Subsequently, based on a preset learning rate parameter (e.g., 0.1), the system characterizes the multi-dimensional differences reflected in the comparative evaluation results, for example, assigning "innovativeness difference +0.12" and "potential risk difference" as features. The difference, such as 0.08", is mapped to a differential feature vector. The system can then output differential features used to update the embedded evaluation values, such as a vector containing multi-dimensional differences [+0.12, ...]. [0.08,…], thus providing the basic input for the dynamic adjustment of the embedded evaluation value in the next sub-step.

[0083] Sub-step 20512: Adjust the embedding evaluation values ​​of the two creative hypothesis texts according to the differential characteristics, and determine the difference between the two adjusted embedding evaluation values ​​as the relative evaluation value between the two creative hypothesis texts.

[0084] In some embodiments of this application, to ensure that the embedded evaluation values ​​reflect the true differences between the two ideas under multi-dimensional evaluation, the embedded evaluation values ​​of the two creative hypothesis texts are adjusted according to the differentiation features, and the difference between the adjusted embedded evaluation values ​​is determined as the relative evaluation value between the two creative hypothesis texts. This adjustment process can also employ a dynamic update strategy similar to the ELO scoring system, that is, increasing or decreasing the embedded evaluation values ​​of the two hypotheses based on the "strength of the comparison results" represented by the differentiation features. For example, when the differentiation features indicate that one hypothesis is significantly better than another hypothesis in multiple evaluation dimensions, its embedded evaluation value will be positively adjusted according to the learning rate parameter, while the embedded evaluation value of the other hypothesis will be negatively adjusted accordingly. Through this dynamic update method, the embedded evaluation values ​​not only reflect the results of a single comparison but also gradually converge into a stable relative evaluation index in multiple rounds of comparison, enabling the system to obtain the relative evaluation value between the two creative hypothesis texts, providing a quantifiable and comparable input basis for subsequent ranking steps.

[0085] In a specific example, the system has obtained the differential features between hypothesis A and hypothesis B, for example, the differential feature vector is [+0.12, [0.08,…] indicates that hypothesis A is superior to hypothesis B in the dimension of innovativeness, but weaker than hypothesis B in the dimension of potential risk. In this case, the embedding evaluation values ​​of the two hypotheses can be dynamically adjusted based on the learning rate parameter (e.g., 0.1): if the initial embedding evaluation value of hypothesis A is 0.62, its updated embedding evaluation value may become 0.62 + 0.1 × (overall difference) = 0.63; if the initial embedding evaluation value of hypothesis B is 0.55, its updated embedding evaluation value may become 0.55. 0.1 × (Comprehensive Difference) = 0.54. The final system will then calculate the difference between the two updated embedded evaluation values ​​(0.63). The value of 0.54 (0.09) is determined as the relative evaluation value between hypothesis A and hypothesis B, thus providing a structured basis for comparison in the subsequent ranking steps.

[0086] Sub-step 2052: Based on the relative evaluation value between every two creative hypothesis texts, sort all the creative hypothesis texts and determine the target creative hypothesis text from the sorted creative hypothesis texts.

[0087] In some embodiments of this application, in order to ultimately determine the target creative hypothesis text that best meets the needs of the research problem, all creative hypothesis texts can be sorted based on the relative evaluation value between every two creative hypothesis texts, and the target creative hypothesis text can be determined from the sorted creative hypothesis texts. The relative evaluation value is used to quantify the comprehensive differences between different hypotheses, enabling the sorting process to make a comprehensive judgment based on multiple evaluation dimensions such as feasibility, rationality, potential risks, innovativeness, and credibility. In this way, the system determines the target creative hypothesis text with the support of multi-dimensional evaluation data and relative comparison results, thereby improving the scientificity and rationality of the hypothesis screening process and reducing the bias caused by the subjectivity of human judgment.

[0088] In a specific example, the system has obtained relative evaluation values ​​among multiple creative hypothesis texts regarding "whether a certain cytokine affects the immunosuppressive mechanism of the tumor microenvironment." For instance, the relative evaluation value of hypothesis A relative to hypothesis B is +0.18, the relative evaluation value of hypothesis A relative to hypothesis C is 0.25, and the relative evaluation value of hypothesis B relative to hypothesis C is... 0.12, etc. Then, based on a scoring weighting strategy, these relative evaluation values ​​can be integrated into a comprehensive ranking index for each hypothesis, and all creative hypothesis texts can be ranked accordingly. Finally, the creative hypothesis text with the best overall performance is determined from the ranking results. For example, if hypothesis A ranks highest, the system will identify it as the target creative hypothesis text, providing the final hypothesis output for subsequent research tasks.

[0089] It should be further pointed out that the creative hypothesis texts generated by the system are essentially "scientific hypotheses" in scientific research activities. Their role is to provide researchers with ideas for exploration, potential mechanism inferences, or research directions worth verifying. Scientific hypotheses themselves are not equivalent to scientific conclusions; they are only used to assist researchers in conducting preliminary exploratory analyses. Their nature is that of scientific hypotheses awaiting verification, rather than conclusive information that can be directly used for decision-making, diagnosis, clinical application, engineering deployment, or policy formulation. They still require verification through subsequent stages such as experimental design, data collection, statistical analysis, and peer review. Therefore, the technical solution provided in this application aims to improve the efficiency of the hypothesis conception and screening stage, enabling researchers to more quickly identify hypotheses with potential value, rather than replacing experimental verification or guaranteeing the truthfulness of hypotheses.

[0090] Furthermore, since any generation process based on a large model may result in inaccuracies or missing evidence, this application does not attempt to address such "data illusion" problems. The creative hypothesis texts output by the system should be considered research inspirations, not directly credible scientific evidence. Researchers still need to independently evaluate the hypotheses using real data and professional judgment. The technical positioning of this application is to assist in the early conceptualization stage of the research process, rather than to construct a complete experimental science loop.

[0091] like Figure 3 The diagram illustrates the automated scientific discovery system M provided in this application, which operates through a two-way interactive closed-loop research mechanism between researchers and the system. Researchers first input the research question, and the automated scientific discovery system M then initiates its internal processes for idea generation, hypothesis testing, and hypothesis ranking. A cyclical reasoning structure is formed among the three agents: the idea generation agent M1, the hypothesis testing agent M2, and the hypothesis ranking agent M3 sequentially execute tasks, continuously feeding back the results to the idea generation agent M1, thus achieving multi-round iterative optimization of the creative hypothesis. After multiple rounds of iteration, the system finally outputs the target creative hypothesis, which is then verified and scientifically judged by the researchers. During system operation, researchers can also input new information into the automated scientific discovery system M based on experimental observations or research feedback, enabling the system to further revise or iterate the creative hypothesis according to the latest research progress, thereby forming a continuous evolution mechanism of "problem—hypothesis—verification—feedback".

[0092] To enhance reasoning capabilities, knowledge coverage, and cross-round memory, the automated scientific discovery system M interacts bidirectionally with the external calling module N. External calling module N comprises a Large Language Model (LLM), a Retrieval Enhancement Model (RAG), a Knowledge Graph Model (KG), and a memory module Mo, providing the system with semantic generation capabilities, external knowledge retrieval capabilities, structured causal reasoning capabilities, and cross-round contextual memory capabilities. Through collaboration with external calling module N, the automated scientific discovery system M can continuously accumulate information, deepen its arguments, and enhance the scientific value of its innovative hypotheses across multiple rounds of reasoning.

[0093] Within the internal workflow of the automated scientific discovery system M, the idea generation agent M1 first uses a deep research mechanism to autonomously plan, repeatedly search, and analyze to acquire the technical background context relevant to the research question. This generated research review serves as domain knowledge, and further, through a cyclical process of "idea iteration—analysis and identification—idea generation—idea verification," it constructs preliminary idea hypotheses. Subsequently, the hypothesis testing agent M2 performs literature retrieval, scientific debate, and multi-dimensional evaluation of the idea hypothesis, including indicators such as novelty, feasibility, and practicality, thus forming a preliminary scientific judgment of the hypothesis. Next, the hypothesis ranking agent M3 automatically ranks multiple candidate hypotheses based on a preset ranking strategy, triggering scientific debate again when necessary to ensure the ranking results have sufficient logical support. The ranking results are fed back to the idea generation agent M1 for the next round of idea iteration, enabling the system to continuously optimize hypothesis quality.

[0094] Through the above methods, the scientific discovery automation system M realizes a complete scientific research closed loop from problem input, idea generation, hypothesis testing, hypothesis ranking to experimenter verification, and can continuously iterate under the drive of experimental feedback, thereby improving the efficiency and systematicness of the "hypothesis-verification" link in scientific research activities.

[0095] like Figure 4 As shown, it is Figure 3 A complete scientific discovery process under the system architecture: The user first submits a research question to the automated scientific discovery system M via step R1. The system's internal thought-generating agent M1, hypothesis-testing agent M2, hypothesis-ranking agent M3, and memory module Mo then collaborate to conduct multiple rounds of reasoning and iteration on the question. After completing one round of reasoning, the system returns the generated target thought hypothesis to the user via step R3. If the user obtains new observations or feedback in subsequent research, they can input them back into the system via step R2, allowing the system to continue adjusting and optimizing the hypothesis based on the latest information.

[0096] Regarding knowledge input, the system supports two sources of literature: users can upload local literature (S1.1) or obtain literature from external data sources (S1.2). After entering the system, all literature first undergoes parsing (S2). The parsed content is used to construct a knowledge graph (S3.1) and is also vectorized (S3.2). The vectorized data is stored in the external calling module N (S3.3), enabling it to be called by the retrieval enhancement model RAG and the knowledge graph model KG.

[0097] Furthermore, during the system's reasoning process, the hypothesis testing agent M2 initiates a retrieval request (S4) to the external calling module N to obtain literature evidence, counterexamples, or background knowledge related to the current hypothesis. The large language model, retrieval enhancement model RAG, knowledge graph model KG, and memory module Mo in the external calling module N return the corresponding knowledge content based on the retrieval request, allowing the three agents within the system to continue reasoning, debating, and ranking.

[0098] In summary, in this embodiment, by inputting the research question and its background information into the creative generation agent, the system can generate multiple creative hypothesis texts with differentiated perspectives, supported by structured domain knowledge. This significantly enhances the diversity and potential innovativeness of the generated hypotheses, thereby reducing the risk of insufficient innovativeness due to a single generation approach. Furthermore, by inputting these multiple creative hypothesis texts into the hypothesis testing agent and evaluating each text from multiple dimensions based on structured domain knowledge, the system can comprehensively analyze the hypotheses from multiple angles, improving the comprehensiveness and reliability of the evaluation results and reducing the uncertainty caused by biased evaluations or incomplete reasoning chains. Finally, based on evaluation data across multiple dimensions, the system determines the target creative hypothesis text from all the creative hypothesis texts. This allows the system to systematically screen candidate hypotheses based on multi-agent collaborative reasoning, improving the scientific rigor and rationality of the target hypothesis selection with the support of multi-dimensional evaluation results. This reduces the workload of researchers in the manual screening process, increases hypothesis screening efficiency, and reduces bias caused by the subjectivity of human judgment. Therefore, the method based on the embodiments of this application, through the collaborative work of the creative generation agent and the hypothesis testing agent, makes the hypothesis generation process more diversified, the hypothesis testing process more systematic, and the hypothesis screening process more scientific, thereby improving the problems of insufficient innovation, one-sided evaluation results, and incomplete reasoning chains in the automatic hypothesis process, and improving the overall work efficiency and decision-making quality of researchers in complex scientific research tasks.

[0099] refer to Figure 5 This application illustrates a multi-agent collaborative automatic scientific hypothesis generation device 30 provided in an embodiment of the present application, comprising: The idea generation module 301 is used to input the research question and its research background information into the idea generation agent to obtain multiple creative hypothesis texts for the research question. The creative evaluation module 302 is used to input multiple creative hypothesis texts into the hypothesis testing agent to obtain evaluation data for each creative hypothesis text under multiple evaluation dimensions. The creative screening module 303 is used to determine the target creative hypothesis text corresponding to the research question from all creative hypothesis texts based on the evaluation data of each creative hypothesis text under multiple evaluation dimensions.

[0100] Optionally, the creative generation module 301 includes: The domain knowledge submodule is used to perform knowledge graph association analysis on the research problem and its research background information in order to obtain structured domain knowledge of the research problem. The Innovation Ideas submodule is used to determine a candidate innovation point dataset for the research problem from structured domain knowledge; the candidate innovation point dataset contains at least one innovation point description text for the research problem. The Creative Text submodule is used to generate a creative hypothesis text based on the description text of each innovation point in the innovation point dataset.

[0101] Optionally, the multi-agent collaborative scientific hypothesis automatic generation device 30 also includes: The self-checking module is used to perform semantic consistency detection and / or logical conflict detection on the description text of each innovation point in the candidate innovation point dataset, and update the candidate innovation point dataset based on the detection results of each innovation point description text.

[0102] Optional, the creative evaluation module 302 includes: The reasoning analysis submodule is used to perform knowledge graph reasoning analysis on each creative hypothesis text based on the structured domain knowledge of the creative hypothesis text, and obtain the reasoning analysis results for each creative hypothesis text. The reasoning analysis results are used to characterize the degree of semantic association and / or logical support between each creative hypothesis text and the structured domain knowledge. The expert debate submodule is used to conduct debate arguments between multiple debate agent agents with different expert identities for each creative hypothesis text, so as to obtain the debate argument results for each creative hypothesis text. The debate argument results are used to characterize at least one of the following conclusions of the creative hypothesis text: feasibility conclusion, reasonable conclusion, potential risk conclusion, innovative conclusion, and credibility conclusion. The multidimensional evaluation submodule is used to determine the evaluation data of each creative hypothesis text under multiple evaluation dimensions based on the reasoning analysis results and debate arguments of each creative hypothesis text.

[0103] Optionally, the expert debate submodule includes: The affirmative argumentation unit is used to input the creative hypothesis text into a debate agent agent with at least one first expert identity to obtain at least one first argumentation result of the creative hypothesis text; each first argumentation result is used to characterize the data of the supporting attributes of the creative hypothesis text in the structured domain knowledge. The opposing argument unit is used to input the creative hypothesis text into a debate agent agent with at least one second expert identity to obtain at least one second argument result for the creative hypothesis text; each second argument result is used to characterize the data in the structured domain knowledge that has opposing attributes to the creative hypothesis text. The conclusion unit is used to input the first and second argument results into the debate agent under the identity of a third expert, so as to obtain the debate argument results of the creative hypothesis text.

[0104] Optionally, the multi-agent collaborative scientific hypothesis automatic generation device 30 also includes: The opposing feedback module is used to update the creative hypothesis text of the debate agent under the identity of the first expert by at least one second argument result, and return to the step of inputting the creative hypothesis text into the debate agent under the identity of at least one first expert to obtain at least one first argument result of the creative hypothesis text; The positive feedback module is used to update the creative hypothesis text of the debate agent under the second expert identity through at least one first argument result, and return to the step of inputting the creative hypothesis text into the debate agent under at least one second expert identity to obtain at least one second argument result of the creative hypothesis text.

[0105] Optionally, the multi-agent collaborative scientific hypothesis automatic generation device 30 also includes: The identity selection module is used to randomly select at least one debate agent with a first expert identity, at least one debate agent with a second expert identity, and one debate agent with a third expert identity from multiple debate agent agents.

[0106] Optionally, the creative filtering module 303 includes: The relative evaluation submodule is used to determine the comparative evaluation results between each pair of creative hypothesis texts by comparing the evaluation data of each pair of creative hypothesis texts under different evaluation dimensions, and to determine the relative evaluation value between each pair of creative hypothesis texts based on the comparative evaluation results. The Creative Screening submodule is used to sort all creative hypothesis texts based on the relative evaluation value between every two creative hypothesis texts, and to determine the target creative hypothesis text from the sorted creative hypothesis texts.

[0107] Optional, the relative evaluation submodule includes: The differential feature unit is used to determine the initial embedding evaluation values ​​of the two creative hypothesis texts and to characterize the differences between the two creative hypothesis texts under different evaluation dimensions by comparing the evaluation results through a preset learning rate parameter, so as to obtain the differential features between the two creative hypothesis texts. The relative evaluation unit is used to adjust the embedding evaluation values ​​of the two creative hypothesis texts respectively according to the differentiated features, and the difference between the two adjusted embedding evaluation values ​​is determined as the relative evaluation value between the two creative hypothesis texts.

[0108] In summary, in this embodiment, by inputting the research question and its background information into the creative generation agent, the system can generate multiple creative hypothesis texts with differentiated perspectives, supported by structured domain knowledge. This significantly enhances the diversity and potential innovativeness of the generated hypotheses, thereby reducing the risk of insufficient innovativeness due to a single generation approach. Furthermore, by inputting these multiple creative hypothesis texts into the hypothesis testing agent and evaluating each text from multiple dimensions based on structured domain knowledge, the system can comprehensively analyze the hypotheses from multiple angles, improving the comprehensiveness and reliability of the evaluation results and reducing the uncertainty caused by biased evaluations or incomplete reasoning chains. Finally, based on evaluation data across multiple dimensions, the system determines the target creative hypothesis text from all the creative hypothesis texts. This allows the system to systematically screen candidate hypotheses based on multi-agent collaborative reasoning, improving the scientific rigor and rationality of the target hypothesis selection with the support of multi-dimensional evaluation results. This reduces the workload of researchers in the manual screening process, increases hypothesis screening efficiency, and reduces bias caused by the subjectivity of human judgment. Therefore, the method based on the embodiments of this application, through the collaborative work of the creative generation agent and the hypothesis testing agent, makes the hypothesis generation process more diversified, the hypothesis testing process more systematic, and the hypothesis screening process more scientific, thereby improving the problems of insufficient innovation, one-sided evaluation results, and incomplete reasoning chains in the automatic hypothesis process, and improving the overall work efficiency and decision-making quality of researchers in complex scientific research tasks.

[0109] Reference Figure 6The electronic device 500 may include one or more of the following components: processing component 502, memory 504, power supply component 506, multimedia component 508, audio component 510, input / output (I / O) interface 512, sensor component 514, and communication component 516.

[0110] Processing component 502 typically controls the overall operation of electronic device 500, such as operations associated with display, telephone calls, data communication, camera operation, and recording. Processing component 502 may include one or more processors 520 to execute instructions to perform all or part of the steps of the methods described above. Furthermore, processing component 502 may include one or more modules to facilitate interaction between processing component 502 and other components. For example, processing component 502 may include a multimedia module to facilitate interaction between multimedia component 508 and processing component 502.

[0111] Memory 504 is used to store various types of data to support the operation of electronic device 500. Examples of this data include instructions for any application or method operating on electronic device 500, contact data, phonebook data, messages, pictures, multimedia, etc. Memory 504 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0112] Power supply component 506 provides power to various components of electronic device 500. Power supply component 506 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to electronic device 500.

[0113] Multimedia component 508 includes an interface that provides an output interface between electronic device 500 and user. In some embodiments, the interface may include a liquid crystal display (LCD) and a touch panel (TP). If the interface includes a touch panel, the interface may be implemented as a touchscreen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touches, swipes, and gestures on the touch panel. The touch sensors may not only sense the boundaries of touch or swipe actions but also detect the duration and pressure associated with the touch or swipe operation. In some embodiments, multimedia component 508 includes a front-facing camera and / or a rear-facing camera. When electronic device 500 is in an operating mode, such as shooting mode or multimedia mode, the front-facing camera and / or rear-facing camera may receive external multimedia data. Each front-facing camera and rear-facing camera may be a fixed optical lens system or have focal length and optical zoom capabilities.

[0114] Audio component 510 is used to output and / or input audio signals. For example, audio component 510 includes a microphone (MIC) used to receive external audio signals when electronic device 500 is in an operating mode, such as call mode, recording mode, and voice recognition mode. The received audio signals may be further stored in memory 504 or transmitted via communication component 516. In some embodiments, audio component 510 also includes a speaker for outputting audio signals.

[0115] Input / output (I / O) interface 512 provides an interface between processing component 502 and peripheral interface modules, such as keyboards, click wheels, buttons, etc. These buttons may include, but are not limited to, home buttons, volume buttons, power buttons, and lock buttons.

[0116] Sensor assembly 514 includes one or more sensors for providing state assessments of various aspects of electronic device 500. For example, sensor assembly 514 may detect the on / off state of electronic device 500, the relative positioning of components such as the display and keypad of electronic device 500, changes in position of electronic device 500 or a component of electronic device 500, the presence or absence of user contact with electronic device 500, orientation or acceleration / deceleration of electronic device 500, and temperature changes of electronic device 500. Sensor assembly 514 may include a proximity sensor configured to detect the presence of nearby objects without any physical contact. Sensor assembly 514 may also include a light sensor, such as a CMOS or CCD image sensor, for use in imaging applications. In some embodiments, sensor assembly 514 may also include an accelerometer, gyroscope, magnetometer, pressure sensor, or temperature sensor.

[0117] Communication component 516 facilitates wired or wireless communication between electronic device 500 and other devices. Electronic device 500 can access wireless networks based on communication standards, such as WiFi, carrier networks (such as 2G, 3G, 4G, or 5G), or combinations thereof. In one exemplary embodiment, communication component 516 receives broadcast signals or broadcast-related information from an external broadcast management system via a broadcast channel. In one exemplary embodiment, communication component 516 also includes a near-field communication (NFC) module to facilitate short-range communication. For example, the NFC module may be implemented based on radio frequency identification (RFID) technology, Infrared Data Association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology, and other technologies.

[0118] In an exemplary embodiment, the electronic device 500 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to implement the methods provided in the embodiments of this application.

[0119] In an exemplary embodiment, a non-transitory computer-readable storage medium including instructions is also provided, such as a memory 504 including instructions, which can be executed by a processor 520 of an electronic device 500 to perform the above-described method. For example, the non-transitory storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc.

[0120] In an exemplary embodiment, the electronic device 500 may also be provided as a server, including a processing component 502, which further includes one or more processors, and memory resources represented by memory 504 for storing instructions, such as applications, that can be executed by the processing component 502. The applications stored in memory 504 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processing component 502 is configured to execute instructions to perform the methods provided in the embodiments of this application.

[0121] Electronic device 500 may also include a power supply component 506 configured to perform power management of electronic device 500, a wired or wireless communication component 516 configured to connect electronic device 500 to a network, and an input / output (I / O) interface 512. Electronic device 500 may operate on an operating system stored in memory 504, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™, or similar.

[0122] It should be noted that, for the sake of simplicity, the method embodiments of this application are all described as a series of actions. However, those skilled in the art should understand that the embodiments of this application are not limited to the described order of actions, because according to the embodiments of this application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily essential to the embodiments of this application.

[0123] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the application disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0124] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for automatically generating scientific hypotheses through multi-agent collaboration, characterized in that, include: The research question and its background information are input into the creative generation agent to obtain multiple creative hypothesis texts for the research question. The multiple creative hypothesis texts are input into the hypothesis testing agent to obtain evaluation data for each creative hypothesis text under multiple evaluation dimensions. Based on the evaluation data of each of the creative hypothesis texts under multiple evaluation dimensions, the target creative hypothesis text corresponding to the research question is determined from all the creative hypothesis texts.

2. The method for automatically generating scientific hypotheses through multi-agent collaboration as described in claim 1, characterized in that, The process involves inputting the research question and its background information into the creative agent to obtain multiple creative hypothesis texts for the research question, including: A knowledge graph association analysis is performed on the problem to be studied and its research background information to obtain structured domain knowledge about the problem to be studied. Based on the problem to be studied, a candidate innovation point dataset for the problem to be studied is determined from the structured domain knowledge; the candidate innovation point dataset contains descriptive text of at least one innovation point for the problem to be studied; A creative hypothesis text is generated based on the description text of each innovation point in the innovation point dataset.

3. The method for automatically generating scientific hypotheses through multi-agent collaboration as described in claim 1, characterized in that, The step of inputting multiple creative hypothesis texts into a hypothesis testing agent to obtain evaluation data for each creative hypothesis text across multiple evaluation dimensions includes: Based on the structured domain knowledge of the creative hypothesis texts, knowledge graph reasoning analysis is performed on each creative hypothesis text to obtain the reasoning analysis result for each creative hypothesis text; the reasoning analysis result is used to characterize the degree of semantic association and / or logical support between each creative hypothesis text and the structured domain knowledge. Each of the creative hypothesis texts is subjected to debate and argumentation among multiple debate agent agents with different expert identities to obtain the debate and argumentation results for each of the creative hypothesis texts; the debate and argumentation results are used to characterize at least one of the following conclusions of the creative hypothesis text: feasibility conclusion, reasonable conclusion, potential risk conclusion, innovative conclusion, and credibility conclusion; The evaluation data for each of the creative hypothetical texts is determined based on the reasoning analysis results and debate arguments of each text. The evaluation data for each text is determined under multiple evaluation dimensions.

4. The method for automatically generating scientific hypotheses through multi-agent collaboration as described in claim 3, characterized in that, The step of conducting debates among multiple expert agents for each of the creative hypothesis texts to obtain the debate results for each creative hypothesis text includes: The creative hypothesis text is input into a debate agent with at least one first expert identity to obtain at least one first argument result of the creative hypothesis text; each first argument result is used to characterize the data of the supporting attributes of the creative hypothesis text in the structured domain knowledge. The creative hypothesis text is input into a debate agent with at least one second expert identity to obtain at least one second argument result of the creative hypothesis text; each second argument result is used to characterize the data of the opposing attributes of the creative hypothesis text in the structured domain knowledge. The first argument result and the second argument result are input together into the debate agent intelligent agent under the identity of a third expert to obtain the debate argument result of the creative hypothesis text.

5. The method for automatically generating scientific hypotheses through multi-agent collaboration as described in claim 4, characterized in that, The method for automatically generating scientific hypotheses through multi-agent collaboration also includes: The creative hypothesis text input to the debate agent under the first expert identity is updated by at least one second argument result, and the process returns to the step of inputting the creative hypothesis text into at least one debate agent under the first expert identity to obtain at least one first argument result of the creative hypothesis text; The creative hypothesis text is updated by at least one of the first argument results and input into the debate agent under the second expert identity, and the process returns to the step of inputting the creative hypothesis text into at least one debate agent under the second expert identity to obtain at least one second argument result of the creative hypothesis text.

6. The method for automatically generating scientific hypotheses through multi-agent collaboration as described in claim 1, characterized in that, The step of determining the target creative hypothesis text corresponding to the research question from all the creative hypothesis texts based on the evaluation data of each of the creative hypothesis texts under multiple evaluation dimensions includes: By comparing the evaluation data of each pair of creative hypothesis texts under different evaluation dimensions, the comparative evaluation result between each pair of creative hypothesis texts is determined, and the relative evaluation value between each pair of creative hypothesis texts is determined based on the comparative evaluation result between each pair of creative hypothesis texts. Based on the relative evaluation value between every two of the creative hypothesis texts, all the creative hypothesis texts are sorted, and the target creative hypothesis text is determined from the sorted creative hypothesis texts.

7. The method for automatically generating scientific hypotheses through multi-agent collaboration as described in claim 6, characterized in that, The step of determining the relative evaluation value between each pair of creative hypothesis texts based on the comparative evaluation results between each pair of creative hypothesis texts includes: Based on the initial embedding evaluation values ​​of the two creative hypothesis texts, and by using a preset learning rate parameter, the difference features between the two creative hypothesis texts under different evaluation dimensions, as represented by the comparative evaluation results, are characterized to obtain the difference features between the two creative hypothesis texts. Based on the differentiated features, the embedding evaluation values ​​of the two creative hypothetical texts are adjusted respectively, and the difference between the two adjusted embedding evaluation values ​​is determined as the relative evaluation value between the two creative hypothetical texts.

8. A multi-agent collaborative automatic scientific hypothesis generation device, characterized in that, include: The idea generation module is used to input the research question and its research background information into the idea generation agent to obtain multiple creative hypothesis texts for the research question. The creative evaluation module is used to input multiple creative hypothesis texts into the hypothesis testing agent to obtain evaluation data for each creative hypothesis text under multiple evaluation dimensions. The creative screening module is used to determine the target creative hypothesis text corresponding to the research question from all the creative hypothesis texts based on the evaluation data of each creative hypothesis text under multiple evaluation dimensions.

9. A computer-readable storage medium, characterized in that, A computer program is stored on the computer-readable storage medium, which, when executed by a processor, implements the method for automatically generating scientific hypotheses through multi-agent collaboration as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the method for automatically generating scientific hypotheses through multi-agent collaboration as described in any one of claims 1 to 7.

Images