Multi-agent-based scientific research idea generation method

Abstract

The invention discloses a scientific research thought generation method based on multiple agents, and belongs to the technical field of agent application. The invention aims to solve the problems that the existing multi-agent scientific research cannot simulate the random sequence of scientific discussion and does not reflect the diversity of the agents. Comprising the steps that each scientist generates a scientific research theme according to theme prompt words, and a leader determines a final scientific research theme; under the scientific research theme, each scientist generates an initial thought according to a knowledge base of the scientist and a first prompt word, and other scientists modify the initial thought according to a second prompt word to obtain a modified thought; the leader summarizes all the revised ideas to obtain revising suggestions, the revising suggestions are returned to the initial scientists for reflection, and the final idea of each scientist is obtained; and the innovation of the final idea of each scientist is checked to determine the final scientific research idea. The method is used for automatically generating scientific research ideas.

Description

Technical Field

[0001] The present invention relates to a method for generating scientific research ideas based on multi - agents, belonging to the technical field of agent applications. Background Art

[0002] The emergence of large language models (LLMs) represented by ChatGPT and LLaMA has revolutionized the capabilities of conversational agents, catalyzing a boom in research on LLM - based agents. These agents are complex artificial intelligence systems that simulate specific individuals by leveraging the advanced capabilities of LLMs, such as following commands and social intelligence. This enables them to effectively assist humans or perform tasks autonomously.

[0003] Research shows that single LLM - based agents can be applied to different fields such as gaming, healthcare, and professional environments. These agents are used as game characters, medical advisors, or collaborative colleagues, attracting users through personalized role - playing and handling predefined tasks.

[0004] However, single agents often struggle when faced with complex tasks. To address this limitation, multi - agent systems based on LLMs are increasingly being adopted to tackle complex challenges. These systems typically employ two main strategies: The first strategy involves pre - defining multiple agent roles and assigning specific tasks to each role, which is built on top of the single - agent framework. The second strategy uses dynamic role - definition, allowing agents to determine tasks and roles based on factors such as memory and environmental cues, thus providing greater flexibility. This dynamic approach improves the adaptability and efficiency of multi - agent systems in complex scenarios.

[0005] The pioneering work of artificial intelligence scientists is an important milestone that explores the potential of using LLMs to simulate the scientific research process, positioning artificial intelligence as a partner of human scientists. This approach enables agents to participate in open - ended scientific discovery, thus assisting and in some cases replicating moments of human creativity and serendipitous innovation. On this basis, VIRSCI introduced a novel multi - agent framework enhanced by reinforcement learning, further promoting the generation of scientific ideas. Compared with single - agent systems, by adopting teams of agents, VIRSCI can more accurately simulate the real - world research environment, facilitating the generation of richer and more diverse ideas while promoting critical evaluation and innovation.

[0006] The Agent Laboratory further emphasizes the advantages of multi-agent systems in scientific research, highlighting their ability to replicate the collaborative dynamics among scientists. Such systems can not only improve research efficiency but also drive innovation by enabling agents to conduct experiments, analyze data, and collectively evaluate new concepts, thereby stimulating creative thinking. In recent years, this multi-agent paradigm, characterized by its ability to enable innovative scientific discoveries, has received extensive attention in various disciplines. For example, SciAgents and AtomAgents integrate domain-specific knowledge bases, such as materials science, to support the discovery of new materials. Similarly, ProtAgent utilizes interdisciplinary knowledge to design entirely new proteins, demonstrating the versatility and transformative potential of multi-agent systems in scientific research.

[0007] Although existing methods have demonstrated their superiority in automated scientific research, their design of the sequential discussion process fails to accurately simulate the random order of scientific discussions. Specifically, traditional methods downplay the importance of interaction and communication among agents during the discussion process, and the roles within the multi-agent team are not fully represented. On the other hand, past methods have ignored the necessity of agent diversity when simulating real research teams and have failed to capture the different backgrounds and perspectives of researchers presented in real scientific collaborations. Summary of the Invention

[0008] In view of the problem that existing multi-agent scientific research cannot simulate the random order of scientific discussions and does not reflect agent diversity, the present invention provides a method for generating scientific research ideas based on multi-agents.

[0009] A method for generating scientific research ideas based on multi-agents according to the present invention includes:

[0010] Forming a scientist team consisting of n agents including a leader and n - 1 scientists, and configuring different knowledge bases and prompting words for each scientist;

[0011] Determining the topic: Each scientist generates a scientific research topic respectively according to the topic prompting words, and the leader determines the final scientific research topic;

[0012] Generating the final idea of each scientist: Under the scientific research topic, each scientist generates an initial idea according to their own knowledge base and the first prompting word, and other scientists modify the initial idea according to the second prompting word respectively to obtain the modified idea; the leader then summarizes all the modified ideas to obtain modification opinions and returns them to the initial scientist for reflection to obtain the final idea of each scientist;

[0013] Conduct innovative inspections to determine the final scientific research ideas: Select multiple papers related to the final ideas from the paper database composed of all scientists' papers according to the cooperation relationships among scientists as references and incorporate them into the prompt library; evenly distribute the references to the scientists; each scientist assigns scores to all the final ideas based on the assigned references and the prompt words updated in combination with the references according to novelty, feasibility, and potential impact to form innovative opinions; after the leader collects all the innovative opinions, an opinion report is formed; all scientists vote on all the final ideas according to the opinion report to determine the final scientific research ideas.

[0014] According to the multi-agent-based scientific research idea generation method of the present invention, let A1 represent the leader, and A2,..., A n represent scientists;

[0015] In the process of generating the final ideas of each scientist, the initial idea generated by the i-th scientist is denoted as I i :

[0016] I i = generate(K i , P1),

[0017] where K i represents the knowledge base configured by the i-th scientist, P1 represents the first prompt word, generate represents the generation function, and i = 2, 3,..., n.

[0018] According to the multi-agent-based scientific research idea generation method of the present invention, other scientists modify I i according to the second prompt word to obtain the modified idea I' ij :

[0019] I' ij = revise(K j , P2, I i ),

[0020] j represents the number of other scientists except the i-th scientist; revise represents the modification function, K j represents the knowledge base configured by the j-th scientist, and P2 represents the second prompt word;

[0021] The leader summarizes all the modified ideas to obtain the modification opinions S i :

[0022] S i = g({I' i,2 , I' i,3 ,..., I' i,i-1 , I' i,i+1 ,..., I'i,n , K1)

[0023] Wherein, g represents a function for modifying the set of opinions, and K1 represents the knowledge base of the leader.

[0024] According to the multi-agent-based scientific research idea generation method of the present invention, the i-th scientist obtains the final idea I i according to the modified opinion S i ':

[0025] I i ' = Reflect(I i , S i ),

[0026] Wherein, Reflect represents a reflection function.

[0027] According to the multi-agent-based scientific research idea generation method of the present invention, the method for selecting reference documents includes:[[]]

[0028] Using the adjacency matrix M ij to represent the paper cooperation relationship between the i-th scientist and the j-th scientist. If the i-th scientist and the j-th scientist have paper cooperation, then M ij = 1; otherwise, M ij = 0;

[0029] According to the adjacency matrix M ij and the Euclidean distance between all the final ideas and each paper, multiple papers of scientists who are related to all the final ideas and have paper cooperation relationships are selected as reference documents in the paper database.

[0030] According to the multi-agent-based scientific research idea generation method of the present invention, knowledge bases are configured for each scientist according to a preset degree of difference, so that the background knowledge of different knowledge bases has an overlapping part and the scientist team has diversity;

[0031] The diversity of the scientist team is expressed as Diversity:

[0032]

[0033] Wherein, Distance(K i , K j ) represents the Euclidean distance between the knowledge base K i and the knowledge base K j .

[0034] According to the multi-agent-based scientific research idea generation method of the present invention, the updated prompt word for the i-th scientist in the innovation check is expressed as P' i :

[0035] P'i = P i ∪ {Paper m | Paper m ∈ Top-k(D(I, Paper)},

[0036] where P i is the original prompt corresponding to P', i Paper m is the m-th paper in Top-k, and Top-k D(I, Paper) is the k references with the smallest Euclidean distance from the final idea set I calculated by the nearest neighbor search algorithm Faiss; m = 1, 2, 3,..., k; I = {I'2, I'3,..., I' n}}, and Paper represents the paper database.

[0037] According to the multi-agent-based scientific research idea generation method of the present invention, the voting results of all scientists are statistically scored using the weighted Borda voting counting method to obtain the ranking of all final ideas, and the final idea with the highest score is used as the final scientific research idea.

[0038] According to the multi-agent-based scientific research idea generation method of the present invention, the statistical score of the i-th final idea using the weighted Borda voting counting method is expressed as B i :

[0039]

[0040] where r ji represents the rank assigned by the j-th scientist to I', i and c ji represents the trust score given by the j-th scientist to I'. i

[0041] Advantages of the present invention: Compared with the existing methods, the method of the present invention is closer to the real scientific research environment. Instead of simply taking turns to output opinions by agents, more discussion links are added. By increasing the interaction between agents, more perfect ideas can be generated. At the same time, the method of the present invention fully stimulates the creativity of agents with different knowledge backgrounds, and can help scientists think from more perspectives whether the ideas they generate are innovative and whether they can have an impact in more fields.

[0042] The method of the present invention alleviates the problems of redundant and repetitive answers by reducing the dependence on long conversation histories. It improves the quality of the final output by emphasizing the exchange of different knowledge perspectives. The method of the present invention more accurately reflects the collaborative dynamics of research teams in the real world, where a designated leader synthesizes the contributions of multiple experts. It can not only improve the efficiency of knowledge exchange but also tap potential to generate more innovative and influential scientific ideas in an automated research environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a comparison diagram between the traditional method and the method of the present invention in the idea generation stage and the innovation check stage; where (a) corresponds to the traditional method and (b) corresponds to the method of the present invention;

[0044] Figure 2 It is the innovative opinions given by each scientist and the final scientific research ideas determined when agents using the same LLM in the verification experiment obtain 8 references and 2 references respectively; in the figure, Decision Made represents the innovation check stage, where the idea with the highest innovation is selected, and Idea represents the idea;

[0045] Figure 3 It is a schematic diagram showing the contrast relationship between team diversity and the dissimilarity of generated scientific research ideas to historical papers and the dissimilarity to contemporary papers;

[0046] Figure 4 It is a schematic diagram showing the influence of team diversity on contemporary influence and overall innovation;

[0047] Figure 5 It is a comparison diagram of the effects of the method of the present invention and VIRSCI when setting 8 agents. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0048] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0049] It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0050] Next, the present invention will be further described in conjunction with the accompanying drawings, but it is not a limitation of the present invention.

[0051] DETAILED DESCRIPTION OF THE EMBODIMENTS I. In combination with Figure 1 As shown, the present invention provides a multi-agent-based scientific research idea generation method, including,

[0052] A scientific team consisting of n agents, including a leader and n - 1 scientists, is configured with different knowledge bases and prompt words for each scientist;

[0053] Determine the theme: Each scientist generates a scientific research theme according to the theme prompt words respectively, and the leader determines the final scientific research theme;

[0054] Generate the final ideas of each scientist: Under the scientific research theme, each scientist generates an initial idea based on their own knowledge base and the first prompt word, and other scientists modify the initial idea according to the second prompt word respectively to obtain the modified ideas; The leader then summarizes all the modified ideas to obtain modification opinions and returns them to the initial scientist for reflection to obtain the final ideas of each scientist;

[0055] Conduct an innovation check to determine the final scientific research idea: Select multiple papers related to the final idea from the paper database composed of all scientists' papers as references according to the cooperation relationship between scientists and incorporate them into the prompt library; Distribute the references evenly to the scientists; Each scientist assigns scores to all the final ideas according to the assigned references and the prompt words updated in combination with the references, considering novelty, feasibility, and potential impact, to form innovation opinions; After the leader collects all the innovation opinions, an opinion report is formed; All scientists vote on all the final ideas according to the opinion report to determine the final scientific research idea.

[0056] In actual use, abstract generation can be carried out as needed. The specific method is: One scientist generates an initial abstract based on the final scientific research idea, and then other scientists modify it in turn. Finally, after being modified by the leader, the final scientific research abstract is obtained.

[0057] In this embodiment, the agent is implemented based on a large language model.

[0058] The scientific team composed of n agents in this embodiment is denoted as S team =<Agent1,Agent2,...,Agent n >>, and the agents communicate and collaborate in a pre - defined workflow to finally generate a scientific research idea. The scientific research idea = f(S team ).

[0059] Dynamic Knowledge Interaction DKE:

[0060] In traditional multi-agent systems, the common approach is to assign different roles to the same large language model-based agents and deploy them in a fixed workflow, where opinions are sequentially output based on the conversation history. This method proves effective when all agents share the same knowledge background. However, when agents are assigned to different automated scientific research knowledge bases, this rigid knowledge exchange mechanism severely deviates from the dynamic interactions in real-world scientific research teams. Additionally, in the absence of a structured knowledge construction process, the conversation history often becomes overly long as the number of participants and the conversation progress increase. This may lead to agents making the same responses despite being assigned roles, especially when the underlying large language model does not have a large enough number of parameters.

[0061] To address this issue, this embodiment draws inspiration from the review processes adopted by major artificial intelligence conferences, particularly the Area Chair (AC) and Program Committee (PC) models, and proposes a Dynamic Knowledge Exchange (DKE) method that fundamentally redefines the interaction pattern among agents. Different from the traditional sequential communication method, this embodiment introduces a leader role, similar to the AC in conference reviews, while other agents assume roles similar to the PC. Specifically, each ordinary scientist has their own unique knowledge base and prompt, and is responsible for generating and presenting their ideas to the leader. The leader, leveraging its knowledge and functional expertise, aggregates and synthesizes the results of all team members into a cohesive summary.

[0062] Furthermore, let A1 denote the leader, and A2,...,A n denote the scientists; A = {A1, A2,...,A n} denote the set of agents; thus, each scientist A i (i ≥ 2) will generate an idea according to their respective knowledge base K i and the given prompt P1.

[0063] In the process of generating the final idea of each scientist, the initial idea generated by the i-th scientist is denoted as I i :

[0064] I i = generate(K i , P1),

[0065] where K i represents the knowledge base configured by the i-th scientist, P1 represents the first prompt word, generate represents the generation function, and i = 2, 3, ……, n.

[0066] Then, other scientists modify I i according to the second prompt word to obtain the modified idea I′ij :

[0067] I′ ij = revise(K j , P2, I i ),

[0068] where j represents the number of other scientists except the i-th scientist; revise represents the modification function, K j represents the knowledge base configured by the j-th scientist, and P2 represents the second prompt;

[0069] The leader A1 summarizes all the modified ideas to obtain the modification opinion S i :

[0070] S i = g({I′ i,2 , I′ i,3 ,..., I′ i,i-1 , I′ i,i+1 ,..., I′ i,n}, K1),

[0071] where g represents the function of aggregating modification opinions, and K1 represents the knowledge base of the leader.

[0072] Compared with traditional methods, the DKE method has several advantages. First, it alleviates the problems of redundancy and repetitive answers by reducing the dependence on long conversation histories. Second, it improves the quality of the final output by emphasizing the exchange of different knowledge perspectives. Third, it more accurately reflects the collaborative dynamics of research teams in the real world, where a designated leader usually synthesizes the contributions of multiple experts. This method can not only improve the efficiency of knowledge exchange but also tap potential to generate more innovative and influential scientific ideas in an automated research environment.

[0073] The i-th scientist obtains the final idea I i ′ according to the modification opinion S i :

[0074] I i ′ = Reflect(I i , S i ),

[0075] where Reflect represents the reflection function.

[0076] Double Diversity Review DDR:

[0077] To enable the multi-agent system to simulate a real scientific research environment, this embodiment introduces the Dual Diversity Review (DDR) mechanism, which combines agents with different knowledge backgrounds with dynamically customized prompts. This approach ensures that the system can not only simulate the heterogeneity of research teams in the real world but also promote innovation by integrating different viewpoints and knowledge sources.

[0078] Before starting automated scientific research, the research environment must be carefully configured. This configuration consists of two main parts: scientists and papers. The papers used for automated research are stored in a dataset of past papers. Author information is extracted from these papers, and the collaboration relationships between authors are represented using an adjacency matrix. The published papers and collaboration network are embedded into the agents as background knowledge to construct the scientists for automated research.

[0079] Furthermore, the method for selecting reference papers includes:

[0080] Using the adjacency matrix M ij to represent the paper collaboration relationship between the i-th scientist and the j-th scientist. If the i-th scientist and the j-th scientist have collaborated on a paper, then M ij = 1; otherwise, M ij = 0;

[0081] Based on the adjacency matrix M ij and the Euclidean distances between all final ideas and each paper, multiple papers of scientists who are relevant to all final ideas and have paper collaboration relationships are selected as reference papers in the paper database.

[0082] In this embodiment, to maximize the diversity of viewpoints, scientists with partially overlapping knowledge backgrounds are selectively constructed. This ensures that during the conversation process, the agents can introduce a wider range of innovative ideas.

[0083] Configure a knowledge base for each scientist according to a preset degree of difference, so that the background knowledge of different knowledge bases has overlapping parts and the scientist team has diversity;

[0084] Quantify the diversity of the scientist team as Diversity:

[0085]

[0086] where Distance(K i , K j ) represents the Euclidean distance between knowledge base K i and knowledge base K j , that is, the degree of difference between the two.

[0087] During the creative generation process, the DDR mechanism continuously improves the prompts by integrating the most relevant papers in the database. This is achieved by leveraging the Faiss library, which can effectively calculate the Euclidean distance between the ideas generated by scientists and the papers in the database.

[0088] To ensure the selection of the most relevant papers, the nearest neighbor search function of Faiss is used to find the top few papers with the smallest Euclidean distance. These papers are then incorporated into the prompt library, enabling scientists to combine their background knowledge with the insights from the most relevant literature to generate new ideas.

[0089] Denote the updated prompt word for the i-th scientist in the innovation check as P′ i :

[0090] P′ i = P i ∪{Parper m |Paper m ∈Top-k(D(I,Paper)},

[0091] where P i is the original prompt word corresponding to P′ i Paper m is the m-th paper in Top-k, Top-kD(I,Paper) are the k references with the smallest Euclidean distance from the final idea set I calculated by the nearest neighbor search algorithm Faiss; m = 1, 2, 3, ……, k; I = {I′2,I′3,...,I′ n}, Paper represents the paper database.

[0092] This method can ensure the diversity of prompts and improve creativity and influence through dynamic updates and continuous optimization, rather than being static or requiring manual intervention. By establishing a set of general rules, the system autonomously improves the prompts through conversations with other scientists and exploration of relevant literature. This process is identical to the way researchers obtain new ideas in the real world, making the DDR mechanism both efficient and highly realistic.

[0093] Furthermore, the weighted Borda voting counting method is used to count the scores of the voting results of all scientists to obtain the ranking of all final ideas, and the final scientific research idea is the one with the highest score.

[0094] The statistical score of the i-th final idea using the weighted Borda voting counting method is denoted as B i :

[0095]

[0096] where r jiIndicates the level assigned by the j-th scientist to I′ i and c ji Indicates the trust score given by the j-th scientist to I′ i .

[0097] After setting up the scientist team in this embodiment, the workflow for generating scientific research ideas is divided into four steps, namely theme discussion, generating the final ideas of each scientist, and innovative checking to determine the final scientific research ideas. Among them, generating the final ideas of each scientist and innovative checking are the two most critical stages in determining the innovation of ideas.

[0098] In the theme discussion step, following the operation of VIRSCI, each agent in the team generates themes according to the preset prompt words, combined with its own background knowledge and the sequential conversation history of the team, and the team will select a final theme according to the probability distribution.

[0099] Stage of generating the final ideas of each scientist: Existing methods do not well simulate real scientific research discussions in the idea generation stage. They adopt a method similar to the previous theme discussion step, that is, the agents take turns to output ideas through preset prompt words, and each agent optimizes or generates new ideas by combining its own background and conversation history. Finally, the team selects the few ideas that are most likely to be innovative according to the scores. Compared with real scientific research teams in reality, this method lacks discussions among scientists, that is, the so-called brainstorming. In order to make full use of the advantages brought by the diversity of agent backgrounds, this embodiment adopts a dynamic knowledge interaction method, specifically: Each Scientist i After generating I i , it will be passed to the scientists except the leader. After that, these scientists who receive I i will modify I i according to their own backgrounds and knowledge; each scientist participating in the modification will submit the modified ideas to the leader for summary. The leader synthesizes the modified opinions of each scientist and his own knowledge to generate a modified opinion and returns it to the Scientist i who generated I i . The initial scientist reflects according to the modified opinion returned by the leader and I i , and finally generates a final version of the idea I i ′. The method proposed in this embodiment is closer to the real scientific research environment compared with the existing methods. It is no longer that the agents simply take turns to output views, but more discussion links are added. By increasing the interaction between agents, more perfect ideas are generated. At the same time, it also fully stimulates the creativity of agents with different knowledge backgrounds, helping scientists think from more perspectives whether the ideas they generate are innovative and whether they can have an impact in more fields.

[0100] Checking for innovation: Through experiments, it has been found that different scientists sometimes make exactly the same choices for the same idea and give almost the same ideas and reasoning processes. As Figure 2 shown, this situation was alleviated after deleting some references. Therefore, it is considered that the long prompt words and the relatively low-performance language model led to approximate answers. This phenomenon is formally expressed as the redundancy factor R, which quantifies the similarity of the responses, and the similarity of the responses is proportional to the number D of different reference documents:

[0101]

[0102] where α is a constant representing the rate at which the response converges as the number of reference documents increases. As the number of documents decreases, D approaches zero, indicating lower similarity of the responses, that is, the more documents, the more similar the responses.

[0103] After reducing the number of reference articles, it was observed that R increased, resulting in more diverse responses. Based on this observation, it is hypothesized that reducing the input context while maintaining the total number of documents can increase the diversity of the responses without affecting the quality of the novelty assessment. In this regard, in addition to using the DKE method to build the interaction mode between agents in the team, the DDR method is also needed. The specific approach is to evenly distribute the reference documents to each scientist. Each scientist generates a response based on the reference documents and the ideas generated in the previous stage. Finally, the leader aggregates the generated responses to generate content regarding the innovation of the ideas. The weighted Borda count method is used to rank the ideas according to their novelty, feasibility, and potential impact. Let I represent the set of ideas, where each idea I′ will be assigned a

[0104] score according to the voting procedure. Let V i represent the vote of the i-th scientist, where V i is a tuple composed of r ji and c ji . The idea with the highest score is selected as the final choice, which reflects the team's consensus on the novelty and feasibility of the ideas.

[0105] Compared with the traditional method, the method of the present invention greatly reduces the content of materials that each agent needs to input without reducing the total number of reference documents. This makes the generated responses more diverse because each scientist will focus on different subsets of the materials. The introduction of the weighted Borda count further ensures that the final choice will not be overly influenced by the ranking of any one scientist, and the higher the confidence in a certain ranking, the greater its weight in the final decision.

[0106] The method of the present invention helps to alleviate the homogenization problem caused by excessive reference materials or too long prompts, and improves the overall diversity and quality of novelty assessment.

[0107] Compared with the existing methods, without reducing the references, the present invention greatly shortens the context that each agent needs to refer to, increases the diversity of the content generated by the agent, and avoids the homogenization problem caused by too long prompt words and other content.

[0108] In the abstract generation stage, since the scientific research ideas of the scientist team have been generated, in order to make the quality of the generated abstract only affected by the innovativeness of the ideas, the same method as the existing method can be adopted. After a scientist generates an abstract based on the idea, the remaining scientists modify the abstract in sequence. Such an approach will not have too much impact on the result of generating the abstract, and is only related to the quality of the idea.

[0109] Verification experiment:

[0110] To comprehensively and fairly evaluate the performance of various methods in generating innovative ideas, models with exactly the same parameter scale and model architecture are selected for comparative analysis. The results are shown in Tables 1 to 4. Among them, HD represents the dissimilarity between the generated scientific research ideas and historical papers, CD represents the dissimilarity between the generated scientific research ideas and contemporary papers, CI represents the contemporary influence, and ON represents the overall innovativeness. HD is obtained by calculating the average squared Euclidean distance between the generated scientific research ideas and the historical paper database. CD is obtained by calculating the average squared Euclidean distance between the generated scientific research ideas and the contemporary paper database. CI is obtained by calculating the citation volume of the five papers in the contemporary paper database that are most similar to the scientific research ideas. All of these three need to be normalized by dividing by the average value. ON = HD * CI / CD.

[0111] Table 1 Comparison of the method of the present invention with the results of the state-of-the-art baseline experiments

[0112] Model HD↑ CD↓ CI↑ ON↑ AI-Scientist-LLAMA 8b 0.51 0.49 2.12 2.21 AI-Scientist-LLAMA 70b 0.53 0.48 2.11 2.33 VIRSCI-LLAMA 8b 0.43 0.42 3.29 3.40 VIRSCI-LLAMA 70b 0.44 0.40 3.36 3.70 IDVSCI-LLAMA 8b 0.40 0.39 4.38 4.49 IDVSCI-LLAMA 70b 0.40 0.39 4.38 4.49

[0113] Table 1 shows that the agent system of the present invention is superior to all simulated scientist systems in terms of CD, CI, and ON indicators.

[0114] In Table 1, IDVSCI represents the model of the present application, and Table 1 shows the effects of the model of the present application with different numbers of parameters, 8b and 70b, respectively.

[0115] Table 2 Comparison results of the method of the present invention with the method of deleting different modules

[0116] HD↑ CD↓ CI↑ ON↑ -Team Discuss 0.41 0.38 4.10 4.42 -Team Vote 0.38 0.38 4.42 4.42 IDVSCI 0.40 0.39 4.49 4.60

[0117] In Table 2, "-" represents the deleted module.

[0118] Table 3 Results generated by the method of the present invention in each round of discussion

[0119] Discussion Round HD↑ CD↓ CI↑ ON↑ 1 0.40 0.38 4.38 4.61 2 0.39 0.38 4.24 4.24 3 0.41 0.40 4.45 4.56 4 0.38 0.37 4.18 4.30 5 0.40 0.39 4.38 4.49

[0120] Table 3 reflects the influence of the number of times each step of the entire framework is carried out on the effect.

[0121] Table 4 Comparison of the effects of applying the method of the present invention to external teams and internal teams

[0122] HD↑ CD↓ CI↑ ON↑ ERSCI 0.40 0.39 4.04 4.14 The method of the present invention 0.40 0.39 4.38 4.49

[0123] As can be seen from Tables 1 to 4, the model based on the method of the present invention is always superior to the existing methods AI-Scientist and VIRSCI in the three key evaluation indicators CD, CI, and ON. These experimental evidences strongly indicate that, compared with other methods, the ideas generated by the model of the present invention have a higher level of innovation and greater scientific influence.

[0124] In addition, the experimental results also reveal an interesting phenomenon, namely the relationship between model parameters and the quality of creative generation. When the same method is adopted, the indicators of HD and CD show significant stability within different parameter ranges, indicating that the choice of models with different numbers of parameters has little impact on the basic creativity of the generated ideas. However, the analysis shows that language models with different parameter scales tend to generate content with different degrees of contemporary relevance and influence. Specifically, the ideas generated by larger models tend to be more closely combined with current scientific trends and show greater potential for direct scientific impact. This finding emphasizes the complex interaction between model capacity and the temporal relevance of the generated scientific viewpoints, providing valuable insights for future research in computational creativity and artificial intelligence-assisted scientific discovery.

[0125] To verify the effectiveness of each module in the method of the present invention, ablation experiments were conducted. As shown in Table 2, compared with the complete method, removing the internal discussion module in the creative generation stage results in a significant decrease in the CI value and CD value. Similarly, replacing the voting mechanism with a traditional method in the novelty inspection stage also leads to a moderate decrease in the CI value, although it is still superior to the variant without the discussion module. These results highlight that the internal discussion module enhances the influence of the generated ideas, while the voting mechanism reduces the dissimilarity with contemporary literature by dispersing the long text analysis.

[0126] The number of discussions is a key factor in the reasoning cost and deserves our attention. As shown in Table 3, the creativity and influence of the ideas generated by the research team did not change significantly with the number of discussion iterations. Specifically, the variance of the ON value in the 1 to 5 iterations was only 0.02. This observation indicates that the method of the present invention can achieve significant results with relatively few resources consumed.

[0127] Combined Figure 3 with Figure 4 shown, when the backgrounds of the scientists forming the research team are completely homogeneous, the ideas generated perform the worst in terms of innovation and influence. When the background diversity reaches 25%, the performance of the model reaches the best. At the 50% and 100% diversity levels, the effects are comparable but slightly inferior to the former. This is very consistent with the understanding of real-world research teams.

[0128] As shown in Table 4, with the assistance of an external team - external review experts (ERSCI), the results of applying the method of the present invention to generate scientific research ideas are inferior to those generated through internal collaborative discussions. Although both methods can obtain the same HD value and CD value, IDVSCI performs better in terms of CI. The main reason for this difference is that the communication between the external method and the idea proposer is limited, resulting in information distortion during transmission.

[0129] The experimental results of the VIRSCI team show that their multi-robot research team performs best when the team size is 4 and 8. In the previous comparison with the SOTA (best-performing) model, to avoid unfair comparison with methods such as AI-Scientist due to too large a team size, only the results when the team size is 4 were compared in this experiment. As Figure 5 can be seen, the method of the present invention still achieved the best performance when the team size was 8. However, compared with VIRSCI, the method of the present invention did not show a significant improvement when the number of members expanded from 4 to 8, indicating that the method of the present invention can achieve ideal results without incurring a large amount of additional costs.

[0130] Example:

[0131] Before conducting automatic scientific research during use, it is necessary to first set up the scientific research environment. The setup of the scientific research environment consists of two parts, namely scientists and articles. Among them, the articles used for automatic scientific research are stored in the Past Paper Dataset (Past Article Database). After that, the author information is extracted from the authors of the articles, and at the same time, the cooperation relationship between the authors is represented by an adjacency matrix. The articles published by the authors and the cooperation relationship are given to the agent as background knowledge to form the scientists for automatic scientific research. The Contemporary Paper Dataset (Contemporary Article Database) is used for subsequent calculations of similarity with the generated abstracts.

[0132] A framework developed using the Agentscope architecture is proposed. This architecture is a powerful platform designed specifically for building multi-agent systems driven by large language models. For performance evaluation, open-source large language models with different computing scales are adopted, namely Llama-3.1 with 8b and 70b parameters, accessed through the Ollama interface. Similar to the experimental steps established in VIRSCI, based on a team of four agents, iterative updates are carried out in five discussion rounds. To ensure statistical reliability, all performance metrics are from the average of 20 independent experiments.

[0133] To balance the number of references and avoid excessive or insufficient selection, based on experience, k = 8 is set for Top-kD(I,Paper) during the novelty checking phase.

[0134] Although the present invention has been described with reference to specific embodiments in this document, it should be understood that these embodiments are merely examples of the principles and applications of the present invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed, as long as they do not deviate from the spirit and scope of the present invention defined by the appended claims. It should be understood that different dependent claims and the features described in this document can be combined in a manner different from that described in the original claims. It should also be understood that the features described in connection with a single embodiment can be used in other described embodiments.

Claims

1. A multi-agent-based method for generating scientific research ideas, characterized in that including, comprising a scientist team consisting of n agents including a leader and n - 1 scientists, and configuring different knowledge bases and prompt words for each scientist; determining the theme: each scientist generates a scientific research theme respectively according to the theme prompt words, and the leader determines the final scientific research theme; generating the final ideas of each scientist: under the said scientific research theme, each scientist generates an initial idea according to their own knowledge base and the first prompt word, and other scientists modify the said initial idea respectively according to the second prompt word to obtain the modified ideas; the leader then summarizes all the modified ideas to obtain modification opinions, and returns them to the initial scientists for reflection to obtain the final ideas of each scientist; conducting an innovation check to determine the final scientific research idea: selecting multiple papers related to the final idea from the paper database composed of all scientists' papers as references according to the cooperation relationship among scientists, and incorporating them into the prompt library; evenly distributing the said references to the scientists; each scientist assigns scores to all the final ideas according to the assigned references and the updated prompt words combined with the references, in terms of novelty, feasibility and potential impact, to form innovation opinions; after the leader collects all the innovation opinions, an opinion report is formed; all scientists vote on all the final ideas according to the said opinion report to determine the final scientific research idea.

2. The multi-agent-based scientific research idea generation method according to claim 1, wherein Let A1 represent the leader, A2,..., A n represent scientists; In the process of generating the final ideas of each scientist, the initial idea generated by the i-th scientist is denoted as I i : I i = generate(K i , P1) where K i represents the knowledge base configured by the i-th scientist, P1 represents the first prompt word, generate represents the generation function, and i = 2, 3,..., n.

3. The multi-agent-based scientific research idea generation method according to claim 2, wherein Other scientists modify I according to the second prompt word i to obtain the modified idea I' ij : I′ ij = revise(K j , P2, I i ), j represents the number of other scientists except the i-th scientist; revise represents the modification function, K j represents the knowledge base configured by the j-th scientist, and P2 represents the second prompt word; the leader summarizes all the modified ideas to obtain modification opinions S2: S i = g({I′ i,2 , I′ i,3 ,..., I′ i,i-1 , I′ i,i+1 ,..., I′ i,n}, K1), where g represents the function of aggregating modification opinions, and K1 represents the knowledge base of the leader.

4. The multi-agent-based scientific research idea generation method according to claim 3, wherein The i-th scientist obtains the final idea I i ′ according to the modification opinion S i ′: I i ′ = Reflect(I i , S i ) where Reflect represents the reflection function.

5. The multi-agent-based scientific research idea generation method according to claim 4, wherein the method for selecting references includes: Use the adjacency matrix M ij to represent the paper collaboration relationship between the i-th scientist and the j-th scientist. If the i-th scientist and the j-th scientist have collaborated on a paper, then M ij = 1, otherwise M ij = 0; According to the adjacency matrix M ij Based on the Euclidean distances between all the final ideas and each paper, multiple papers of scientists who are related to all the final ideas and have paper cooperation relationships are selected as references in the paper database.

6. The multi-agent-based scientific research idea generation method according to claim 5, wherein configuring knowledge bases for each scientist according to a preset degree of difference, so that the background knowledge of different knowledge bases has an overlapping part and the scientist team has diversity; representing the diversity of the scientist team as Diversity: where Distance(K i ,K j ) represents the Euclidean distance between knowledge base K i and knowledge base K j .

7. The multi-agent-based scientific research idea generation method according to claim 6, wherein Denote the updated prompt for the \(i\)-th scientist in the innovation check as \(P'\). i : P′ i = P i ∪ {Paper m | Paper m ∈ Top-k(D(I, Paper)} where P i is the original prompt corresponding to P′ i , Paper m is the m-th paper in Top-k, and Top-kD(I, Paper) is the k references with the smallest Euclidean distance from the final idea set I calculated by the nearest neighbor search algorithm Faiss; m = 1, 2, 3,..., k; I = {I′2, I′3,..., I′ n}, and Paper represents the paper database.

8. The multi-agent-based scientific research idea generation method according to claim 7, wherein using the weighted Borda voting counting method to count the scores of the voting results of all scientists, obtaining the ranking of all the final ideas, and taking the final idea with the highest score as the final scientific research idea.

9. The multi-agent-based scientific research idea generation method according to claim 8, wherein The statistical score of the i-th final idea using the weighted Borda voting count method is denoted as B i : where r ji represents the level assigned by the j-th scientist to I′ i , and c ji represents the trust score given by the j-th scientist to I′ i .

Images