Multi-agent discussion method, apparatus, device, and medium

Abstract

The application provides a multi-agent discussion method, device, equipment and medium, the method comprises the following steps: generating a decision factor set containing a preset number of nodes according to discussion topic information, performing semantic clustering, hierarchical labeling and causal relationship analysis on each node in the decision factor set, and constructing a directed acyclic influence graph; performing semantic retrieval on the vector database according to the discussion topic information, and performing node retrieval on the directed acyclic influence graph according to the innovation ability of the agent; generating a speech prompt word according to the results of double retrieval, the six-dimensional prompt of the agent, the role setting and the discussion context, and calling a large language model to obtain the speech proposal of each agent; determining the comprehensive utility score of the speech proposal of each agent to determine the speech agent of the current round from each agent and call the speech of the corresponding speech agent. The application overcomes the homogenization defect of the agent role, ensures that the speech content is both creative and consistent with the business logic, and improves the efficiency and quality of group discussion.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device, and medium for multi-agent discussion. Background Technology

[0002] With the rapid development of artificial intelligence technology, human-machine collaboration has become an important paradigm for improving decision-making quality and innovation efficiency in education, training, and enterprise management. Traditional offline or remote group discussions are often limited by the negative effects of group dynamics, resulting in problems such as productivity bottlenecks, social loafing, and a lack of information structure, making it difficult to cope with the challenges of complex decision-making scenarios. To overcome these limitations, intelligent agent technology based on Large Language Models (LLM) has been introduced into group collaboration, aiming to enhance human thinking and decision-making abilities through algorithms and promote the transformation of collaboration models from purely manual interaction to human-machine collaborative intelligence.

[0003] Currently, most LLM-based agent group collaboration technologies adopt a multi-agent debate framework based on a large language model as a solution. This type of technology typically involves pre-setting multiple artificial intelligence (AI) agents with different roles or identities, and using the generation capabilities of LLM to simulate the discussion process of a human group. Its specific implementation logic generally includes: setting system prompts to assign specific perspectives or tasks to the agents, and adopting a turn-based dialogue management mechanism to enable multiple rounds of viewpoint interaction and argumentation between agents or between agents and humans. By continuously iterating the above interaction process, the system uses the reasoning capabilities of the large model to critique, refute, or correct the generated viewpoints, in an attempt to converge the truth or optimize decision-making schemes through the collision of viewpoints among multiple agents, in order to achieve a better decision result than that of a single agent or individual thinking.

[0004] However, while existing multi-agent debate frameworks have broadened the scope of group discussions to some extent, they lack refined modeling of role decision-making styles, thought processes, and innovation preferences. This leads to severe role homogenization among agents, resulting in convergent discussion content and difficulty in replicating the diverse intellectual exchanges found in real human groups. Furthermore, these debate frameworks often remain at the superficial stages of semantic generation and creative brainstorming, lacking a deep understanding of specific business logic and causal relationships. They also generally employ a passive, turn-based interaction model, lacking real-time perception and dynamic intervention capabilities regarding the pace and atmosphere of the discussion. Consequently, the system is prone to generating "illusionary" suggestions detached from real-world constraints and cannot simulate the natural interruptions and talk-stealing mechanisms found in real human communication, thus limiting the depth and practicality of human-machine collaboration. Summary of the Invention

[0005] This invention provides a multi-agent discussion method, apparatus, device, and medium to address the shortcomings of existing multi-agent debate frameworks, which suffer from homogenized roles and a lack of deep understanding of business logic, leading to similar and superficial discussion content. It achieves differentiated information acquisition based on different innovation preferences, reproducing the diverse intellectual collisions of real human groups. This results in generated proposals exhibiting significant differences in decision-making, reasoning, and control, overcoming the homogenization of agent roles and ensuring that the content is both creative and consistent with business logic, thereby significantly improving the accuracy of simulating real human thought processes.

[0006] This invention provides a multi-agent discussion method, comprising: generating a decision element set containing a preset number of nodes based on discussion topic information; performing semantic clustering, hierarchical labeling, and causal relationship analysis on each node in the decision element set to construct a directed acyclic influence graph; performing semantic retrieval on a vector database based on discussion topic information, and performing node retrieval on the directed acyclic influence graph based on previously acquired agent innovation capabilities; wherein the vector knowledge base is constructed prior to target domain documents; generating speaking prompts and calling a large language model based on the results of the dual retrieval and previously acquired agent six-dimensional prompts, role settings, and discussion context to obtain speaking proposals for each agent; determining the comprehensive utility score of each agent's speaking proposal, and determining the speaking agent for the current round based on the comprehensive utility score and calling the corresponding speaking agent to speak.

[0007] According to a multi-agent discussion method provided by the present invention, semantic retrieval is performed on a vector database based on discussion topic information, and node retrieval is performed on a directed acyclic influence graph based on previously acquired agent innovation capabilities. The method includes: constructing a hybrid query vector based on discussion topic information, combining speech summaries within a target number of rounds with previously acquired and currently acquired nodes of interest; performing semantic retrieval on the vector database based on the hybrid query vector to obtain knowledge retrieval results including external document fragments; determining corresponding node retrieval strategies based on previously acquired agent innovation capabilities and a strategy selection mechanism, and performing node retrieval on the directed acyclic influence graph based on the node retrieval strategies to obtain node retrieval results; wherein the strategy selection mechanism is configured to: increase the retrieval scope or retrieval weight of corresponding associative nodes in response to improvements in agent innovation capabilities; and obtaining a dual retrieval result based on the knowledge retrieval results and the node retrieval results.

[0008] According to the multi-agent discussion method provided by the present invention, the agent six-dimensional prompts are generated based on a pre-constructed agent six-dimensional capability parameterization model and the level configuration of each dimension; based on the results of dual retrieval and the pre-acquired agent six-dimensional prompts, role settings, and discussion context, speaking prompt words are generated and a large language model is invoked to obtain the speaking proposals of each agent, including: constructing speaking prompt words according to a preset format based on the results of dual retrieval and the pre-acquired agent six-dimensional prompts, role settings, and discussion context; inputting the speaking prompt words into the large language model, and performing the following generation logic for each agent based on the large language model: determining the speaking perspective of the corresponding agent according to the role settings, and combining the six-dimensional capability prompts. The system constrains decision-making and innovation capabilities to determine the number of target nodes to extract from the dual-search results, and adjusts the selection bias of associative nodes according to the innovation capability level. Based on the constraints of the reasoning capability dimension in the six-dimensional capability prompts, causal analysis is performed on the selected arguments to construct a logical chain, and the discussion atmosphere is analyzed based on the discussion context. Combined with the constraints of the conflict coordination or thought stimulation dimensions in the six-dimensional capability prompts, the type of speaking intent is determined. Based on the speaking intent type and logical chain, and combined with the discussion control dimension, the rhythm and direction are controlled to generate text content. The citation information of target nodes during the generation process and the confidence level of the generated text content are encapsulated into data to output a structured speaking proposal.

[0009] According to a multi-agent discussion method provided by the present invention, the comprehensive utility score of each agent's speaking proposal is determined, and the speaking agent for the current round is determined from among the agents based on the comprehensive utility score. This includes: obtaining the current discussion stage information and determining the stage intention weight; wherein the stage information includes a divergent stage, a debate stage, and a convergent stage, and the stage intention weight is configured to gradually increase as the discussion progresses from focusing on divergence to focusing on convergence; determining the capability matching degree of each agent based on the six-dimensional capability parameters of each agent; and determining the speaking quality and balance factor of the speaking proposal, combined with global information, based on the stage intention weight, capability matching degree, speaking quality, and balance factor of the speaking proposal. The urgency of the topic extracted from the pool determines the initial utility score. The speaking quality of the speaking proposal is based on the logical coherence of the corresponding speaking proposal and its matching degree with the results of the dual retrieval. The balance factor is based on the dispersion of the historical speaking frequency distribution of each agent before the current round. The urgency of the topic is determined based on the frequency of mention of the discussion topic within the preset rounds. A random perturbation factor is introduced into the initial utility score, and a cooling-off penalty is imposed on agents who speak continuously according to the historical speaking records to obtain a corrected comprehensive utility score. The agent with the highest comprehensive utility score is selected as the speaking agent for the current round.

[0010] According to a multi-agent discussion method provided by the present invention, the method determines the comprehensive utility score of each agent's speaking proposal and determines the speaking agent for the current round based on the comprehensive utility score. The method further includes: determining whether the current discussion interaction mode is an artificial intelligence (AI) discussion mode or a human-computer collaborative interaction mode; when in AI discussion mode, determining the comprehensive utility score of each agent's speaking proposal and determining the speaking agent for the current round based on the comprehensive utility score, and updating the continuous speaking count of the target speaking agent and the number of silent rounds of the remaining agents; or, when in human-computer collaborative interaction mode, based on a parallel-started task monitoring process, when a user-triggered interaction signal is detected, sending a global cancellation signal to each agent to forcibly interrupt the proposal generation task and identifying the user as the speaking subject for the current round; when no user-triggered interaction signal is detected, waiting for all speaking agents determined based on the comprehensive utility score of each agent's speaking proposal to complete the proposal generation task, and comparing the proposal generation time of each speaking agent with a preset intent trigger time threshold of the human-computer interaction task to determine the speaking subject for the current round.

[0011] According to a multi-agent discussion method provided by the present invention, after determining the speaking agent for the current round from among the agents based on a comprehensive utility score, the method includes: monitoring the progress of the current round of discussion, and triggering a topic deviation detection step when a preset round cycle condition is met; the topic deviation detection step is used to: evaluate the semantic relevance between the current discussion content and the discussion topic information after the determined speaking agent has spoken; when the semantic relevance is determined to be lower than a preset relevance threshold, select candidate agents whose discussion control ability meets preset requirements based on the six-dimensional capability parameters of each agent, and generate a correction instruction for the discussion topic information; send the correction instruction to the candidate agent to control the candidate agent to generate a speaking proposal containing a guiding regression topic in the next round.

[0012] According to a multi-agent discussion method provided by the present invention, a decision element set containing a preset number of nodes is generated based on discussion topic information. Semantic clustering, hierarchical labeling, and causal relationship analysis are performed on each node in the decision element set to construct a directed acyclic influence graph. The method includes: generating a decision element set containing a preset number of nodes based on discussion topic information and a preset node generation prompt template; using a node generation model to generate a decision element set containing a preset number of nodes; using the node generation model again to perform semantic clustering on the nodes in the decision element set to generate a topic classification system of the target number of nodes; using the generated topic classification systems to assign a topic identifier and abstract level to each node in each topic classification system using multi-level abstract labeling technology; using a batch pairwise decision algorithm to identify causal relationships between nodes based on each topic classification system; and constructing a directed acyclic influence graph based on the causal relationships between nodes and all nodes in the topic classification systems after assigning their topics and abstract levels.

[0013] This invention also provides a multi-agent discussion device, comprising: a graph construction module, which generates a set of decision elements containing a preset number of nodes based on discussion topic information, and performs semantic clustering, hierarchical labeling, and causal relationship analysis on each node in the decision element set to construct a directed acyclic influence graph; a dual retrieval module, which performs semantic retrieval on a vector database based on discussion topic information, and performs node retrieval on the directed acyclic influence graph based on previously acquired agent innovation capabilities; wherein the vector knowledge base is previously constructed based on target domain documents; a proposal prediction module, which generates speaking prompts and calls a large language model based on the results of the dual retrieval and previously acquired agent six-dimensional prompts, role settings, and discussion context to obtain speaking proposals from each agent; and a speaking module, which determines the comprehensive utility score of each agent's speaking proposal, determines the speaking agent for the current round based on the comprehensive utility score, and calls the corresponding speaking agent to speak.

[0014] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the multi-agent discussion method as described above.

[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the multi-agent discussion method as described above.

[0016] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the multi-agent discussion method as described above.

[0017] The multi-agent discussion method, apparatus, device, and medium provided by this invention construct a directed acyclic influence graph containing hierarchical annotations and causal relationships to establish a structured business logic framework for the discussion. This avoids a lack of deep understanding of specific business logic and effectively prevents agents from making "illusionary" suggestions that are detached from actual constraints. It ensures the logical rigor and professional depth of subsequent discussion content. Furthermore, it employs a dual retrieval mechanism combining a vector database and the directed acyclic influence graph. This introduces a broad external knowledge base while utilizing a deep causal logic network, avoiding the narrowness and lack of breadth in discussion content. It also breaks the limitations of homogeneous agent roles, enabling differentiated information acquisition based on different innovation preferences and reproducibility. This approach simulates the diverse intellectual exchanges within a real human group and further refines the constraints on the large language model using six-dimensional capability cues. This results in significant differences in the generated proposals across dimensions such as decision-making, reasoning, and control of the discussion. Simultaneously, it generates structured proposals based on search results, ensuring that the content is both creative and logically sound. This significantly improves the accuracy of simulating real human thought processes. Furthermore, it dynamically selects speaking agents based on comprehensive utility scores, rather than simply rotating them mechanically. This empowers the system to autonomously allocate speaking rights based on discussion quality and the current situation, avoiding a lack of dynamic perception of the discussion's pace and atmosphere. This ensures that the most valuable viewpoints are presented promptly, improving the efficiency and quality of group discussions. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the multi-agent discussion method provided by the present invention; Figure 2 This is a schematic diagram of the architecture of the multi-agent discussion method provided by the present invention; Figure 3 This is a schematic diagram of the structure of the multi-agent discussion device provided by the present invention; Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0021] Figure 1 This is a flowchart illustrating the multi-agent discussion method provided by the present invention, as shown below. Figure 1 As shown, the method includes the following: S11. Generate a set of decision elements containing a preset number of nodes based on the discussion topic information, and perform semantic clustering, hierarchical labeling and causal relationship analysis on each node in the set of decision elements to construct a directed acyclic influence graph. S12, semantic retrieval of the vector database based on the discussion topic information, and node retrieval of the directed acyclic influence graph based on the previously acquired agent innovation capabilities; wherein, the vector knowledge base is constructed in advance based on the target domain documents; S13. Based on the results of the dual search and the previously obtained six-dimensional prompts, role settings and discussion context of the agents, generate speaking prompts and call the large language model to obtain the speaking proposals of each agent. S14, determine the comprehensive utility score of each agent's speech proposal, determine the speaking agent for the current round from among the agents based on the comprehensive utility score, and call the corresponding speaking agent to speak.

[0022] It should be noted that the surface is combined with the specific details. Figure 2 The multi-agent discussion method of the present invention is described.

[0023] Step S11: Generate a set of decision elements containing a preset number of nodes based on the discussion topic information, and perform semantic clustering, hierarchical labeling and causal relationship analysis on each node in the set of decision elements to construct a directed acyclic influence graph.

[0024] In this embodiment, reference Figure 2The process involves generating a set of decision elements containing a predetermined number of nodes based on discussion topic information, and performing semantic clustering, hierarchical labeling, and causal relationship analysis on each node in the decision element set to construct a directed acyclic influence graph. This includes: generating a set of decision elements containing a predetermined number of nodes based on discussion topic information and a preset node generation prompt template; using a node generation model to generate a set of decision elements containing a predetermined number of nodes; using the node generation model again to perform semantic clustering on the nodes in the decision element set to generate a topic classification system of the target number of nodes; using the generated topic classification systems to assign a topic identifier and abstraction level to each node in each topic classification system using multi-level abstract labeling technology; using a batch pairwise decision algorithm to identify causal relationships between nodes based on each topic classification system; and constructing a directed acyclic influence graph based on the causal relationships between nodes and all nodes in the topic classification systems after assigning their topics and abstraction levels.

[0025] It should be added that the node generation model can be selected according to actual design needs, such as the large language model GPT-4o, etc., without further limitation here. The decision element set includes a preset number of nodes, and each node includes an index, name, and type. By generating a structured decision element set containing indexes, names, and types, the initial deconstruction and discretization of complex discussion topics can be achieved, solving the problem that the discussion content is superficial and lacks substantial business support due to the lack of specific business element analysis. This provides basic data units for the subsequent construction of a discussion framework with in-depth logic. The types of nodes include information nodes, decision nodes, and utility nodes. The output decision element set can be output according to the actual design format. After generating the decision element set containing the preset number of nodes, the output decision element set is format-validated, and a decision element regeneration mechanism is triggered based on the inconsistency in the validation.

[0026] In addition, the topic classification system includes technology, regulations, operations, etc. By generating a topic classification system through semantic clustering of nodes, the discrete decision-making elements can be systematically classified, which helps to clarify the dimensions and scope of the discussion, improve the agent's ability to control the macro structure of the discussion topic, and avoid generating chaotic content.

[0027] Furthermore, based on the generated topic classification systems, the node generation model is invoked again to utilize multi-level abstract annotation technology. This includes: processing the generated topic classification systems in batches, and using the node generation model to assign a topic identifier ID and abstract level to each node. The number of nodes in each batch can be configured according to actual design requirements, such as processing 10 nodes per batch. No further limitation is made here. The abstract levels include first-level macro nodes, second-level meso nodes, and third-level micro nodes to form a pyramid-shaped knowledge structure.

[0028] It should be noted that by using multi-level abstract annotation technology, nodes are given topic identifiers and abstract levels, enabling refined modeling of node business attributes. This hierarchical structure allows agents to understand problems from different levels of abstraction (such as the strategic and execution levels), overcoming the shortcomings of existing agents' single-minded thinking and lack of depth, and providing structured support for generating more hierarchical and logical statements.

[0029] It is worth noting that when assigning abstraction levels to each node, the proportion of nodes assigned to the first-level macro nodes is 10%-15%, the proportion of nodes assigned to the second-level meso nodes is 25%-35%, and the proportion of nodes assigned to the third-level micro nodes is 50%-65%. When the distribution ratio of each abstraction level exceeds the corresponding node proportion range, the process of re-labeling the deviating nodes is triggered. For details, please refer to the above text, which will not be repeated here.

[0030] Furthermore, based on the aforementioned topic classification systems, a batch pairwise determination algorithm is used to identify causal relationships between nodes, including: grouping all nodes according to a preset number of nodes based on each topic classification system to obtain multiple node subgroups; traversing all combinations of node subgroups to obtain a set of node pairs whose causal relationship needs to be determined, the set of node pairs includes intra-group node pairs within the same node subgroup and inter-group node pairs between different node subgroups; inputting the set of node pairs whose causal relationship needs to be determined and the previously constructed causal determination prompt words containing the set of node pairs into the large language model to perform causal direction determination for each pair of nodes; wherein, causal direction determination includes identifying whether there is a direct causal relationship between node pairs "from A to B", "from B to A", or "bidirectional"; obtaining the causal determination result output by the large language model, and filtering node pairs with causal relationships based on the causal determination result, extracting the corresponding causal direction edges to construct a directed acyclic influence graph.

[0031] It's worth noting that the batch pairwise decision algorithm accurately identifies causal relationships between nodes, uncovering the implicit logical transmission chains between business elements. This addresses the lack of deep understanding of specific business logic, enabling the agent to reason based on deep causal logic rather than simple semantic associations, effectively reducing the risk of generating "illusionary" suggestions detached from practical constraints. Furthermore, the directed acyclic influence graph is persistently stored in a preset data structure format, serving as a cognitive foundation to limit the agent's divergent thinking. The preset structured format can be configured similarly to the output format of the aforementioned decision element set, such as JSON, etc., without further limitations here.

[0032] Furthermore, by integrating causal relationships and hierarchical attributes to construct a directed acyclic influence graph, a directed acyclic influence graph with both logical depth and structural hierarchy is formed. This graph not only simulates the constraints and transmission relationships of elements in the real decision-making process, but also provides a rigorous logical boundary for subsequent retrieval and reasoning, ensuring that the intelligent agent maintains the logical consistency and business applicability of the discussion content while engaging in diverse thinking collisions.

[0033] Step S12 involves performing semantic retrieval on the vector database based on the discussion topic information, and node retrieval on the directed acyclic influence graph based on the previously acquired agent innovation capabilities; wherein, the vector knowledge base was previously constructed based on target domain documents.

[0034] It should be added that before performing semantic retrieval of the vector database based on the discussion topic information, and before performing node retrieval of the directed acyclic influence graph based on the previously acquired agent innovation capabilities, the following steps are required: constructing a vector knowledge base, specifically: acquiring target domain documents, including academic literature, industry reports, and policies and regulations; performing standardized preprocessing on the target domain documents, dividing the standardized preprocessed target domain documents into text blocks of fixed length, and calling the embedding model (BGE) to convert the text blocks into high-dimensional vectors, storing them in the vector database (FAISS), and establishing an index.

[0035] Furthermore, the target domain documents undergo standardization preprocessing, including text cleaning and removal of irrelevant formatting tags.

[0036] In this embodiment, semantic retrieval of the vector database is performed based on discussion topic information, and node retrieval of the directed acyclic influence graph is performed based on the previously acquired agent innovation capabilities. This includes: constructing a hybrid query vector based on discussion topic information, combining the speech summaries within the target number of rounds with previously acquired and currently acquired nodes of interest; performing semantic retrieval of the vector database based on the hybrid query vector to obtain knowledge retrieval results containing external document fragments; determining the corresponding node retrieval strategy based on the previously acquired agent innovation capabilities and a strategy selection mechanism, and performing node retrieval of the directed acyclic influence graph based on the node retrieval strategy to obtain node retrieval results; wherein, the strategy selection mechanism is configured to: increase the retrieval scope or retrieval weight of the corresponding associative nodes in response to the improvement of agent innovation capabilities; and obtaining a dual retrieval result based on the knowledge retrieval results and the node retrieval results.

[0037] It should be noted that the strategy selection mechanism dynamically adjusts the retrieval strategy based on the agent's innovation capability level. Low innovation level focuses only on strongly relevant nodes, while high innovation level mixes non-directly connected associative nodes to simulate the leaps in human thinking. For example, low innovation level retrieves 100% strongly relevant nodes, medium innovation level mixes 80% relevant nodes and 20% associative nodes, and high innovation level mixes 60% relevant nodes and 40% associative nodes. The specific design can be tailored to actual design needs and is not further limited here.

[0038] Furthermore, by integrating discussion topic information, historical summaries, and nodes of interest to construct multi-dimensional hybrid query vectors, this approach captures the semantic thread of the discussion while also encompassing contextual coherence and the current focus of the discussion. This enables precise expression of query intent, laying the foundation for subsequent accurate retrieval. The semantic retrieval capabilities of the vector database are utilized to acquire external document fragments, addressing the lack of specific business logic and deep understanding in existing technologies. By introducing a real and extensive external knowledge base, the information boundaries of the discussion are effectively broadened, avoiding "illusionary" suggestions generated by agents relying solely on internal parameters. Additionally, by utilizing node relationships in a directed acyclic influence graph and combining them with innovation capabilities for dynamic strategy adjustments, the limitations of homogenized agent roles and convergent thinking are overcome. Through differentiated control of the retrieval scope and weights, refined modeling of the thinking logic and innovation preferences of different agents is achieved, enabling agents to reproduce the diverse intellectual collisions within real human groups.

[0039] In an optional embodiment, before obtaining the speaking proposals of each agent, the process includes: for any agent, generating a six-dimensional prompt based on the level configuration of each dimension and a pre-established six-dimensional capability parameterization model; wherein, the six-dimensional capability parameters include decision-making, innovation, reasoning, thought stimulation, conflict coordination, and discussion control capability parameters, and the decision-making capability level is mapped to the scale of graph nodes that the agent needs to cover in a single thought, thereby distinguishing between "deliberate" and "intuitive" thinking modes at the underlying logic level. For example, when the decision-making capability level is high, the number of graph nodes covered in a single thought is 10; when the decision-making capability level is medium, the number of graph nodes covered in a single thought is 6; and when the decision-making capability level is low, the number of graph nodes covered in a single thought is 2.

[0040] Step S13: Based on the results of the dual search and the previously obtained six-dimensional prompts, role settings, and discussion context of the agents, generate speaking prompts and call the large language model to obtain the speaking proposals of each agent.

[0041] In this embodiment, the agent's six-dimensional prompts are generated based on a pre-constructed parameterized model of the agent's six-dimensional capabilities and the level configuration of each dimension. Based on the results of the dual search and the previously acquired agent six-dimensional prompts, role settings, and discussion context, speaking prompts are generated and a large language model is invoked to obtain speaking proposals for each agent. This includes: constructing speaking prompts according to a preset format based on the results of the dual search and the previously acquired agent six-dimensional prompts, role settings, and discussion context; inputting the speaking prompts into the large language model to execute the following generation logic for each agent based on the large language model: determining the corresponding agent's speaking perspective based on the role settings, and combining the decision capabilities in the six-dimensional capability prompts. Constraints on the dimensions of competence and innovation ability determine the number of target nodes to be extracted from the results of dual retrieval, and adjust the selection tendency of associative nodes according to the innovation ability level; based on the constraints of the reasoning ability dimension in the six-dimensional ability prompts, causal analysis is performed on the selected arguments to construct a logical chain, and the discussion atmosphere is analyzed based on the discussion context, and combined with the constraints of the conflict coordination or thinking stimulation dimension in the six-dimensional ability prompts, the speaking intention type is determined; based on the speaking intention type and logical chain, combined with the discussion control dimension, the rhythm and direction are controlled, text content is generated, and the reference information of the target nodes in the generation process and the confidence level of the generated text content are data encapsulated to output a structured speaking proposal.

[0042] It should be noted that by introducing six-dimensional ability prompts encompassing decision-making, innovation, and reasoning, a refined role constraint mechanism is constructed to address the existing problem of lacking refined modeling of role decision-making styles, thinking logic, and innovation preferences. This lays the foundation for breaking the homogenization of agent roles and achieving personalized speech. Furthermore, by utilizing constraints in decision-making and innovation capabilities to dynamically filter arguments, different agents can focus on different aspects when faced with the same information, thus overcoming the current deficiency of convergent discussion content among agents and effectively replicating the diverse intellectual collisions in real human groups. Simultaneously, causal relationships are constructed through reasoning ability constraints. By establishing logical chains and dynamically adjusting speaking intentions based on the discussion atmosphere, the intelligent agent moves beyond superficial semantic generation to possess a deep understanding of specific business logic and an awareness of the discussion situation. This significantly improves the logicality and relevance of the speech. By generating speeches based on speaking intention types and logical chains, combined with the dimension of discussion control, and by encapsulating citation information and confidence levels to output structured proposals, the intelligent agent is not only given the ability to actively control the pace and direction of the discussion, simulating the dynamic mechanisms in real human communication, but also provides quantifiable and traceable data support for subsequent scoring and evaluation, enhancing the system's practicality.

[0043] It should be added that the six-dimensional ability prompts include constraints on the dimensions of decision-making ability, innovation ability, reasoning ability, thinking stimulation ability, conflict coordination ability, and discussion control ability.

[0044] In addition, based on the discussion context analysis of the discussion atmosphere, and combined with the constraints of the conflict coordination or thought stimulation dimensions, the types of speaking intentions are determined, including: when the discussion atmosphere is a stalemate, the conflict coordination dimension is used to generate compromise or summary intentions; when the discussion atmosphere is a dull discussion, the thought stimulation dimension is used to generate questioning or new viewpoint intentions.

[0045] Step S14: Determine the comprehensive utility score of each agent's speech proposal, and based on the comprehensive utility score, determine the speaking agent for the current round from among the agents and call the corresponding speaking agent to speak.

[0046] In this embodiment, the comprehensive utility score of each agent's speaking proposal is determined, and the speaking agent for the current round is determined from among the agents based on the comprehensive utility score. This includes: obtaining the current discussion stage information and determining the stage intention weight; wherein, the stage information includes the divergence stage, the debate stage, and the convergence stage, and the stage intention weight is configured to gradually increase as the discussion progresses from focusing on divergence to focusing on convergence; determining the ability matching degree of each agent based on the six-dimensional ability parameters of each agent; and combining the stage intention weight, ability matching degree, speaking quality of the speaking proposal, and balance factor with the topics extracted from the global information pool. The urgency factor is used to determine the initial utility score. The quality of the speaking proposal is based on the logical coherence of the corresponding speaking proposal and its matching degree with the results of the dual retrieval. The balance factor is based on the dispersion of the historical speaking frequency distribution of each agent before the current round. The urgency of the topic is determined based on the frequency of mention of the discussion topic within the preset rounds. A random perturbation factor is introduced into the initial utility score, and a cooling-off penalty is imposed on agents who speak continuously according to the historical speaking records to obtain a corrected comprehensive utility score. The agent with the highest comprehensive utility score is selected as the speaking agent for the current round.

[0047] It should be noted that by identifying discussion stages and dynamically adjusting intent weights, an evaluation orientation adapted to the discussion process is established. This ensures that the system encourages innovative divergence in the early stages of discussion and focuses on logical convergence in the later stages. This addresses the problem of existing agents lacking real-time perception and dynamic intervention capabilities regarding the discussion pace, making the discussion process more consistent with the natural evolution of human group decision-making. Based on six-dimensional capability parameters, the system evaluates the agent's suitability for the current situation, enabling in-depth utilization of different roles' capabilities and avoiding a one-size-fits-all evaluation approach. This ensures that selected speaking agents possess the core capabilities required to handle the current topic, thereby improving the professionalism of the speech content. Furthermore, the system comprehensively calculates scores based on multiple dimensions, including speech quality, balance factors, and topic urgency, addressing the limitations of existing debate frameworks due to their singular evaluation criteria. The problem of inconsistent content quality and uneven participation caused by the first round of interactions is effectively addressed by introducing a balance factor and urgency. This effectively stimulates diversified participation within the group, simulates the dynamic balance between equal opportunity and focus of attention in real human discussions, introduces a random perturbation factor into the initial utility score to simulate the uncertainty of real discussions, and uses a cooling-off penalty mechanism to suppress the monopolistic tendencies of dominant agents. This overcomes the mechanical rigidity and lack of dynamic mechanisms in the existing passive turn-based interaction mode, making the allocation of speaking rights more flexible, natural, and fair. The final speaker is determined based on the corrected comprehensive utility score, ensuring that the optimal solution can stand out. At the same time, it takes into account the dynamic balance and randomness of the system, achieving intelligent, efficient, and intuitive allocation of speaking rights.

[0048] It should be added that when determining the ability matching degree of each agent, the weight of the decision-making ability dimension is configured to be higher than that of other dimensions. For example, the decision-making ability weight independently accounts for 60%, and no further restrictions are made here. The urgency of the topic can be determined based on the distribution of intentions in the preset rounds of speaking to determine the atmosphere. For example, when the proportion of rebuttal is greater than or equal to 40%, it is judged as tense. The specific configuration can be based on actual design requirements and prior experience, and no further restrictions are made here. In addition, the comprehensive utility score of each agent's speaking proposal is determined. The agent for speaking in the current round is determined from the agents based on the comprehensive utility score. This can be implemented using a meta-agent. The specific configuration can be based on actual design requirements, and no further restrictions are made here.

[0049] In one optional embodiment, determining the overall utility score of each agent's speaking proposal and determining the speaking agent for the current round based on the overall utility score further includes: determining whether the interaction mode of the current discussion is an artificial intelligence (AI) discussion mode or a human-computer collaborative interaction mode; when in AI discussion mode, determining the overall utility score of each agent's speaking proposal and determining the speaking agent for the current round based on the overall utility score, and updating the continuous speaking count of the target speaking agent and the number of silent rounds of the remaining agents.

[0050] It should be noted that by identifying and matching different interaction patterns, the system is given the ability to flexibly adapt to different application scenarios, which solves the problem of existing multi-agent discussion systems having single functions and being unable to switch operating mechanisms according to user needs. This improves the system's versatility and practicality. In the AI ​​discussion mode, a comprehensive utility score mechanism is used to ensure that the selected speakers have high-quality proposals. At the same time, by updating the consecutive speaking count and the number of silent rounds, the speaking balance among agents is maintained quantitatively. This solves the problem of imbalanced participation caused by existing agents easily leading to "one-man show" or long-term silence. This makes the group discussion process more fair, orderly and sustainable, and avoids the homogenization of thinking caused by the monopoly of individual agents.

[0051] In another optional embodiment, determining the comprehensive utility score of each agent's speaking proposal and determining the speaking agent for the current round based on the comprehensive utility score further includes: when in human-computer collaborative interaction mode, based on the parallel-started task monitoring process, when it is determined that an interaction signal triggered by the user is detected, sending a global cancellation signal to each agent to forcibly interrupt the proposal generation task and identifying the user as the speaking subject for the current round; when it is determined that no interaction signal triggered by the user is detected, waiting for all speaking agents determined based on the comprehensive utility score of each agent's speaking proposal to complete the proposal generation task, and comparing the proposal generation time of each speaking agent with the preset intent trigger time threshold of the human-computer interaction task to determine the speaking subject for the current round.

[0052] It should be noted that the task monitoring process includes human-computer interaction tasks for user input and proposal generation tasks for proposal generation.

[0053] Furthermore, by configuring flexible interaction mode switching, it meets the user's participation needs in different scenarios, solves the problem of existing frameworks having single functions and being unable to adapt to complex collaborative environments, and provides basic architectural support for deep human-machine collaboration. Through parallel listening and global cancellation mechanisms, it grants human users the highest priority speaking rights, allowing users to interrupt the AI's generation process at any time. This solves the problem of existing passive turn-based interaction modes lacking real-time perception and dynamic intervention capabilities, achieving truly natural "speaking up" and instant interaction, significantly improving the immersiveness and user experience of human-machine collaboration. When the user does not actively intervene, it compares the generation time with the time threshold to achieve dynamic competitive allocation of speaking rights. This not only preserves the efficiency of AI discussion but also simulates the impact of reaction speed on speaking opportunities in real communication, breaking the limitations of mechanical turn-taking and making the communication process smoother and closer to the dynamic mechanism of real human social interaction.

[0054] In one optional embodiment, after invoking the corresponding speaking agent to speak, the process includes: determining the atmosphere state based on the intention distribution of the preset rounds of speaking; determining the topic relevance based on the speaking agent's speech and discussion topic; and writing the speaking agent's speech, discussion topic, corresponding atmosphere state, and topic relevance into the global information pool.

[0055] It should be added that after determining the topic relevance, the following steps are taken: based on the topic relevance and combined with the three-stage evolution strategy, the generation of the speech proposals of each agent in the next round is optimized; the three-stage evolution strategy includes the divergence stage, the debate stage, and the convergence stage.

[0056] Furthermore, when the topic relevance is 0-50%, it corresponds to the divergent stage; when the topic relevance is 50%-80%, it corresponds to the debate stage; and when the topic relevance is 80%-100%, it corresponds to the convergent stage.

[0057] In an optional embodiment, after determining the speaking agent for the current round from among the agents based on the comprehensive utility score, the process includes: monitoring the progress of the current round of discussion, and triggering a topic deviation detection step when a preset round cycle condition is met; the topic deviation detection step is used to: evaluate the semantic relevance between the current discussion content and the discussion topic information after the determined speaking agent has spoken; when it is determined that the semantic relevance is lower than a preset relevance threshold, select candidate agents whose discussion control capabilities meet preset requirements based on the six-dimensional capability parameters of each agent, and generate a correction instruction for the discussion topic information; send the correction instruction to the candidate agent to control the candidate agent to generate a speaking proposal containing a guiding regression topic in the next round.

[0058] It should be noted that by periodically monitoring the progress of each round and triggering a detection mechanism, continuous global monitoring of the discussion process is achieved. This addresses the lack of real-time perception of the discussion rhythm in existing technologies, ensuring that the system can promptly detect and correct deviations in the discussion, maintaining the stability of the discussion process. By evaluating semantic relevance, it can quantitatively identify whether the discussion has gone off-topic, thus preventing the system from falling into meaningless divergence under the lack of constraints. This solves the problem of existing technologies easily generating suggestions that are out of practical constraints, ensuring that the discussion always revolves around the core objective. When the semantic relevance is lower than the preset relevance threshold, the six-dimensional capability parameters are used to accurately select agents with high field control capabilities to correct deviations, achieving refined and intelligent intervention strategies. This avoids logical confusion caused by blind intervention, ensuring that the correction operation is performed by the most appropriate role, enhancing the realism of simulating a "moderator" or "leader" guiding the discussion back on track in a real group. Furthermore, by sending correction instructions to specific agents, dynamic guidance of subsequent discussion content is achieved, solving the problem of the lack of dynamic intervention capabilities in existing technologies. This ensures that the system can automatically correct deviations in the discussion path, improving the autonomy of the multi-agent system and the effectiveness of the final conclusion.

[0059] In summary, this invention constructs a directed acyclic influence graph (DAG) containing hierarchical annotations and causal relationships to establish a structured business logic framework for discussions. This avoids a lack of deep understanding of specific business logic and effectively prevents agents from making "illusionary" suggestions that are detached from real-world constraints. It ensures the logical rigor and professional depth of subsequent discussions. Furthermore, it employs a dual retrieval mechanism combining a vector database and the DAG, leveraging a broad external knowledge base and a deep causal logic network to avoid monotonous and narrow discussions. This breaks the limitations of homogeneous agent roles, enabling differentiated information acquisition based on different innovation preferences and recreating a realistic human group dynamic. The system fosters diverse thinking and further utilizes six-dimensional capability cues to refine the constraints on the large language model, resulting in significant differences in the generated proposals across dimensions such as decision-making, reasoning, and control. Simultaneously, it generates structured proposals based on search results, ensuring that the content is both creative and logically sound. This significantly improves the accuracy of simulating real human thought processes. Furthermore, it dynamically selects speaking agents based on comprehensive utility scores, rather than simply rotating them mechanically. This empowers the system to autonomously allocate speaking rights based on discussion quality and the current situation, avoiding a lack of dynamic perception of the discussion's pace and atmosphere. This ensures that the most valuable viewpoints are presented promptly, improving the efficiency and quality of group discussions.

[0060] The multi-agent discussion device provided by the present invention is described below. The multi-agent discussion device described below can be referred to in correspondence with the multi-agent discussion method described above.

[0061] Figure 3A schematic diagram of a multi-agent discussion device is shown, the device comprising: The graph construction module 31 generates a set of decision elements containing a preset number of nodes based on the discussion topic information, and performs semantic clustering, hierarchical labeling and causal relationship analysis on each node in the set of decision elements to construct a directed acyclic influence graph. The dual retrieval module 32 performs semantic retrieval of the vector database based on the discussion topic information, and node retrieval of the directed acyclic influence graph based on the previously acquired agent innovation capabilities; wherein, the vector knowledge base is constructed in advance based on the target domain files; The proposal prediction module 33 generates speaking prompts and calls the large language model based on the results of the dual search and the previously acquired six-dimensional prompts, role settings and discussion context of the agents to obtain the speaking proposals of each agent. The speaking module 34 determines the comprehensive utility score of each agent's speaking proposal, and based on the comprehensive utility score, determines the speaking agent for the current round from among the agents and calls the corresponding speaking agent to speak.

[0062] It should be noted that the specific principles of the embodiments of the present invention are the same as those of the method embodiments described above. For details, please refer to the method embodiments above. More detailed explanations will not be repeated here.

[0063] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4 As shown, the electronic device may include: a processor 410, a communications interface 420, a memory 430, and a communications bus 440, wherein the processor 410, the communications interface 420, and the memory 430 communicate with each other through the communications bus 440. The processor 410 can call logical instructions in the memory 430 to execute a multi-agent discussion method, which includes: generating a set of decision elements containing a preset number of nodes based on the discussion topic information, and performing semantic clustering, hierarchical labeling, and causal relationship analysis on each node in the decision element set to construct a directed acyclic influence graph; performing semantic retrieval on a vector database based on the discussion topic information, and performing node retrieval on the directed acyclic influence graph based on the previously acquired agent innovation capabilities; wherein, the vector knowledge base is constructed in advance based on target domain files; generating speaking prompts and calling a large language model based on the results of the dual retrieval and the previously acquired agent six-dimensional prompts, role settings, and discussion context to obtain speaking proposals for each agent; determining the comprehensive utility score of each agent's speaking proposal, and determining the speaking agent for the current round from among the agents based on the comprehensive utility score and calling the corresponding speaking agent to speak.

[0064] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0065] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the multi-agent discussion method provided by the above methods. The method includes: generating a set of decision elements containing a preset number of nodes based on discussion topic information, and performing semantic clustering, hierarchical labeling, and causal relationship analysis on each node in the set of decision elements to construct a directed acyclic influence graph; performing semantic retrieval on a vector database based on discussion topic information, and performing node retrieval on the directed acyclic influence graph based on previously acquired agent innovation capabilities; wherein the vector knowledge base is constructed based on target domain files; generating speaking prompts and calling a large language model based on the results of the dual retrieval and previously acquired agent six-dimensional prompts, role settings, and discussion context to obtain speaking proposals for each agent; determining the comprehensive utility score of each agent's speaking proposal, and determining the speaking agent for the current round from among the agents based on the comprehensive utility score and calling the corresponding speaking agent to speak.

[0066] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the multi-agent discussion method provided by the above methods. The method includes: generating a set of decision elements containing a preset number of nodes based on discussion topic information, and performing semantic clustering, hierarchical labeling, and causal relationship analysis on each node in the set of decision elements to construct a directed acyclic influence graph; performing semantic retrieval on a vector database based on discussion topic information, and performing node retrieval on the directed acyclic influence graph based on previously acquired agent innovation capabilities; wherein the vector knowledge base is constructed prior to target domain files; generating speaking prompts and calling a large language model based on the results of the dual retrieval and previously acquired agent six-dimensional prompts, role settings, and discussion context to obtain speaking proposals for each agent; determining the comprehensive utility score of each agent's speaking proposal, and determining the speaking agent for the current round from among the agents based on the comprehensive utility score and calling the corresponding speaking agent to speak.

[0067] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0068] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0069] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-agent discussion method, characterized in that, include: Based on the discussion topic information, a set of decision elements containing a preset number of nodes is generated, and semantic clustering, hierarchical labeling and causal relationship analysis are performed on each node in the set of decision elements to construct a directed acyclic influence graph. Semantic retrieval of the vector database is performed based on the discussion topic information, and node retrieval of the directed acyclic influence graph is performed based on the previously acquired agent innovation capabilities; wherein, the vector knowledge base is constructed in advance based on target domain documents; Based on the results of the dual search, as well as the previously obtained six-dimensional prompts, role settings, and discussion context of the agents, speaking prompts are generated and the large language model is invoked to obtain the speaking proposals of each agent. Determine the overall utility score of each agent's speech proposal, and based on the overall utility score, determine the speaking agent for the current round from among the agents and call the corresponding speaking agent to speak.

2. The multi-agent discussion method according to claim 1, characterized in that, Semantic retrieval of the vector database based on the aforementioned discussion topic information, and node retrieval of the directed acyclic influence graph based on the previously acquired agent innovation capabilities, including: Based on the discussion topic information, and combining the speech summaries within the target number of rounds with the previously obtained and currently focused nodes, a hybrid query vector is constructed; Based on the hybrid query vector, semantic retrieval is performed on the vector database to obtain knowledge retrieval results containing external document fragments; Based on the previously acquired agent innovation capabilities and combined with the strategy selection mechanism, a corresponding node retrieval strategy is determined, and node retrieval is performed on the directed acyclic influence graph based on the node retrieval strategy to obtain the node retrieval results. The strategy selection mechanism is configured to: in response to the improvement of the agent's innovation capability, increase the retrieval range or retrieval weight of the corresponding associative node; Based on the knowledge retrieval results and the node retrieval results, a dual retrieval result is obtained.

3. The multi-agent discussion method according to claim 2, characterized in that, The six-dimensional prompts for the agent are generated based on the previously constructed parameterized model of the agent's six-dimensional capabilities and the level configuration of each dimension; Based on the results of the dual search, as well as the previously acquired six-dimensional prompts, role settings, and discussion context of the agents, speaking prompts are generated and a large language model is invoked to obtain speaking proposals from each agent, including: Based on the results of the dual search and the previously obtained six-dimensional prompts of the agent, role settings and discussion context, construct speaking prompts according to a preset format; The spoken prompts are input into a large language model, and the following generation logic is executed for each agent based on the large language model: Based on the role settings, the speaking perspective of the corresponding intelligent agent is determined, and combined with the constraints of the decision-making ability and innovation ability dimensions in the six-dimensional ability prompts, the number of target nodes extracted from the results of the dual retrieval is determined, and the selection tendency of the associative nodes is adjusted according to the innovation ability level. Based on the constraints of the reasoning ability dimension in the six-dimensional ability prompts, causal analysis is performed on the selected arguments to construct a logical chain, and the discussion atmosphere is analyzed based on the discussion context. In combination with the constraints of the conflict coordination or thinking stimulation dimension in the six-dimensional ability prompts, the speaking intention type is determined. Based on the speaking intent type and the logical chain, and combined with the discussion control dimension, the rhythm and direction are controlled to generate text content. The reference information of the target node in the generation process and the confidence level of the generated text content are encapsulated into data to output a structured speaking proposal.

4. The multi-agent discussion method according to claim 1, characterized in that, Determine the overall utility score of each agent's speaking proposal, and determine the speaking agent for the current round from among the agents based on the overall utility score, including: Obtain the current discussion stage information and determine the stage intention weight; wherein, the stage information includes the divergence stage, the debate stage, and the convergence stage, and the stage intention weight is configured to gradually increase as the discussion progresses from focusing on divergence to focusing on convergence; Based on the six-dimensional capability parameters of each agent, the capability matching degree of each agent is determined. An initial utility score is determined based on the stage intent weight, the capability matching degree, the speaking quality of the speaking proposal, and the balance factor, combined with the topic urgency extracted from the global information pool. The speaking quality of the speaking proposal is based on the logical coherence of the corresponding speaking proposal and its matching degree with the results of the dual retrieval. The balance factor is based on the dispersion of the historical speaking frequency distribution of each agent before the current round. The topic urgency is determined based on the frequency of mention of the discussion topic within a preset round. A random perturbation factor is introduced into the initial utility score, and a cooling-off penalty is imposed on agents who speak continuously based on historical speaking records to obtain a corrected comprehensive utility score. The agent with the highest overall utility score is selected as the speaking agent for the current round.

5. The multi-agent discussion method according to claim 1, characterized in that, Determining the overall utility score of each agent's speaking proposal, and determining the speaking agent for the current round from among the agents based on the overall utility score, further includes: Determine whether the current discussion interaction mode is an artificial intelligence (AI) discussion mode or a human-computer collaborative interaction mode; When in the AI ​​discussion mode, determine the comprehensive utility score of each agent's speaking proposal, and based on the comprehensive utility score, determine the speaking agent for the current round from among the agents, and update the continuous speaking count of the target speaking agent and the number of silent rounds of the remaining agents; or, When in human-computer collaborative interaction mode, based on the parallel-started task monitoring process, when it is determined that the user-triggered interaction signal is detected, a global cancellation signal is sent to each of the intelligent agents to forcibly interrupt the proposal generation task, and the user is identified as the speaker of the current round. When it is determined that no user-triggered interaction signal is detected, wait for all speaking agents, determined based on the comprehensive utility score of each speaking agent's proposal, to complete the proposal generation task. Then, based on the proposal generation time of each speaking agent, and combined with the preset intent trigger time threshold of the human-computer interaction task, a speed comparison is performed to determine the speaking subject of the current round.

6. The multi-agent discussion method according to claim 1, characterized in that, After determining the speaking agent for the current round from among the agents based on the comprehensive utility score, the process includes: Monitor the progress of the current discussion round, and trigger the topic deviation detection step when the preset round cycle conditions are met; the topic deviation detection step is used for: After a given speaking agent speaks, the semantic relevance between the current discussion content and the discussion topic information is evaluated. When the semantic relevance is determined to be lower than a preset relevance threshold, candidate agents whose discussion control capabilities meet preset requirements are selected based on the six-dimensional capability parameters of each agent, and a correction instruction for the discussion topic information is generated. The correction instruction is sent to the candidate agent to control the candidate agent to generate a speech proposal containing the guiding regression topic in the next round.

7. The multi-agent discussion method according to claim 1, characterized in that, Based on the discussion topic information, a set of decision elements containing a preset number of nodes is generated. Semantic clustering, hierarchical labeling, and causal relationship analysis are then performed on each node in the set of decision elements to construct a directed acyclic influence graph, including: Based on the discussion topic information and the preset node generation prompt template, the node generation model is invoked to generate a set of decision elements containing a preset number of nodes; Based on the set of decision elements, the node generation model is invoked again to perform semantic clustering on the nodes in the set of decision elements, thereby generating a topic classification system for the target system quantity. Based on the generated topic classification systems, the node generation model is called again to use multi-level abstract annotation technology to assign a topic identifier and abstract level to each node in each topic classification system. Based on the aforementioned topic classification system, the causal relationship between nodes is identified using a batch pairwise determination algorithm; Based on the causal relationships between the nodes, and combining all nodes in the topic classification system after assigning their respective topics and abstract levels, a directed acyclic influence graph is constructed.

8. A multi-agent discussion device, characterized in that, include: The graph construction module generates a set of decision elements containing a preset number of nodes based on the discussion topic information, and performs semantic clustering, hierarchical labeling and causal relationship analysis on each node in the set of decision elements to construct a directed acyclic influence graph. The dual retrieval module performs semantic retrieval of the vector database based on the discussion topic information, and node retrieval of the directed acyclic influence graph based on the previously acquired agent innovation capabilities; wherein, the vector knowledge base is constructed prior to the target domain files; The proposal prediction module generates speaking prompts and calls the large language model based on the results of the dual search and the previously obtained six-dimensional prompts, role settings and discussion context of the agents to obtain the speaking proposals of each agent. The speaking module determines the comprehensive utility score of each agent's speaking proposal, and based on the comprehensive utility score, determines the speaking agent for the current round from among the agents and calls the corresponding speaking agent to speak.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the multi-agent discussion method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the multi-agent discussion method as described in any one of claims 1 to 7.

Images