Multi-agent collaborative reasoning system and method

Abstract

This invention discloses a multi-agent collaborative reasoning system and method. The system includes: a dynamic policy coordination module for analyzing the original problem input by the user and determining initial policy parameters; a multi-round reasoning module for sending the original problem and initial policy parameters to a pool of agents, obtaining reasoning data from each agent, performing multiple rounds of interaction until round control conditions are met, and determining collaborative reasoning data based on the latest reasoning data; obtaining updated reasoning data from each agent during each round of interaction; a parameter optimization module for calculating trajectory scores on the reasoning data from the previous round for each agent, determining the current interaction mode and current interaction parameters, and updating the trajectory score weights; and a communication optimization module for processing the reasoning data from each agent and sending it to other agents. This invention enables multi-agent collaboration that organically integrates efficient interaction, deep reasoning evaluation, and optimal result selection.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of artificial intelligence technology, and in particular to a multi-agent collaborative reasoning system and method. Background Technology

[0002] The rapid development of artificial intelligence technology has gradually revealed the limitations of single models in handling complex reasoning, creative tasks, and decision support. To overcome performance bottlenecks and simulate the intelligent mode of human group collaboration, multi-agent collaborative problem-solving has become a cutting-edge research direction in the field of artificial intelligence.

[0003] Traditional methods for solving complex problems often rely on a single large model or simple result integration. Their capabilities are limited by the inherent biases of the model itself and knowledge limitations. Furthermore, they lack in-depth analytical mechanisms in controversial scenarios, which can easily lead to errors or mediocre results.

[0004] To overcome the shortcomings of a single model, various collaborative solution schemes based on multi-agent debate have emerged in existing technologies, which can be mainly divided into the following two categories: The first category is the classic multi-agent debate framework based on consensus. This type of scheme currently dominates the field. Its core lies in enabling multiple agents to engage in multiple rounds of dialogue or exchange of opinions around the same issue, with reaching a consensus as the termination condition for the debate, ultimately outputting a unified conclusion accepted by the majority of agents. This scheme improves the diversity of thinking to some extent through group interaction and achieves better performance than single models in some tasks. However, this scheme is limited by its majority-rule decision-making mechanism, which suffers from inefficiency and "majority dominance." Specifically, agents holding correct but non-mainstream opinions may be ignored due to this mechanism, causing the group to eventually converge to a generally accepted but not optimal conclusion. The essence of this mechanism is to equate the correctness of the conclusion with its degree of support, while ignoring the evaluation of the quality of the reasoning process, making it difficult to identify and adopt logically rigorous but initially low-support correct reasoning answers.

[0005] The second category comprises simplified debate or voting schemes adopted in pursuit of efficiency. To reduce the overhead of lengthy debates in classic frameworks, simplified strategies such as single-round voting, weighted averaging of outputs, or directly selecting the highest confidence result have emerged. These schemes aim to quickly integrate group opinions with extremely low communication costs, but they are prone to insufficient exploration, lack of deep reasoning, and quality degradation due to simple fusion. Single-round or minimally simplistic interactions fail to achieve sufficient collision and iteration of viewpoints, and the differences between agents cannot be effectively identified and integrated, thus losing the potential to improve the quality of reasoning answers through debate. Furthermore, these schemes typically only integrate the final outputs of each agent, completely ignoring the intermediate reasoning process, resulting in poor decision interpretability and failure to benefit from high-quality reasoning trajectories. Simultaneously, voting or averaging strategies struggle to distinguish whether conclusions stem from rigorous reasoning or random generation; such mechanical fusion methods are prone to "group convergence" effects, failing to guarantee that the output result is superior to the performance of the best individual agent.

[0006] Existing technical solutions suffer from an irreconcilable conflict between consensus-building mechanisms and core objectives such as the accuracy and efficiency of reasoning answers. Classical debate frameworks sacrifice efficiency in pursuit of consensus and may introduce systemic biases; simplified solutions abandon deep interaction to improve efficiency, resulting in limited output quality. Neither approach achieves a collaborative mechanism that efficiently utilizes collective intelligence and accurately identifies and adopts the optimal reasoning process without requiring mandatory consensus.

[0007] Therefore, there is an urgent need for a new multi-agent collaborative framework that can fundamentally streamline the relationship between consensus and decision-making, and organically integrate efficient interaction, deep reasoning evaluation, and optimal result selection to solve the above-mentioned technical problems.

[0008] The background section described above is merely a description made by the inventor based on their understanding, and the above content should not be regarded as evidence of prior art disclosed before the filing date of this application. Summary of the Invention

[0009] This invention provides a multi-agent collaborative reasoning system to achieve efficient interaction, deep reasoning evaluation, and optimal result selection in multi-agent collaboration. The system includes: The dynamic strategy coordination module is used to analyze the original question input by the user and determine the initial strategy parameters, which include the interaction mode, interaction parameters, trajectory scoring weights and round control conditions. The multi-round inference module is used to send the original problem and initial policy parameters to the agent pool, obtain the inference data output by each agent, perform multiple rounds of interaction until the round control conditions are met, and determine the collaborative inference data based on the latest inference data output by the agents. In each round of interaction, the policy parameters are updated through the parameter optimization module, and the inference data is processed and sent through the communication optimization module, and the updated inference data of each agent is obtained. The parameter optimization module is used to calculate the trajectory score of the inference data output by each agent in the previous round according to the trajectory score weight, determine the current interaction mode and the current interaction parameters based on the trajectory score, and update the trajectory score weight. The communication optimization module is used to process the inference data of each agent according to the current interaction mode; and to send the processed inference data, the current interaction mode, and the current interaction parameters of each agent to other agents. The agent pool is used to store agents for artificial intelligence models.

[0010] Another aspect of the present invention provides a multi-agent cooperative reasoning method to achieve multi-agent cooperation that organically integrates efficient interaction, deep reasoning evaluation, and optimal result selection. The method includes: Analyze the original question input by the user to determine the initial strategy parameters, which include the interaction mode, interaction parameters, trajectory scoring weights, and round control conditions; The original problem and initial policy parameters are sent to the pool of agents to obtain the inference data output by each agent. The process involves multiple rounds of interaction until the round control conditions are met. During each round, a trajectory score is calculated for the inference data output by each agent in the previous round, based on the trajectory score weights. The current interaction mode and parameters are determined according to the trajectory score, and the trajectory score weights are updated. The inference data of each agent is then processed according to the current interaction mode. The processed inference data, the current interaction mode, and the current interaction parameters of each agent are sent to other agents. Finally, updated inference data for each agent is obtained. Based on the latest inference data output by the agent, determine the collaborative inference data.

[0011] This invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described multi-agent cooperative reasoning method.

[0012] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned multi-agent cooperative reasoning method.

[0013] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described multi-agent cooperative reasoning method.

[0014] In this embodiment of the invention, a dynamic strategy coordination module is used to analyze the original question input by the user and determine the initial strategy parameters, which include interaction mode, interaction parameters, trajectory scoring weights, and round control conditions. A multi-round inference module is used to send the original question and the initial strategy parameters to the agent pool, obtain the inference data output by each agent, execute multiple rounds of interaction until the round control conditions are met, and determine collaborative inference data based on the latest inference data output by the agents. During each round of interaction, the strategy parameters are updated through the parameter optimization module, and the inference data is processed and sent through the communication optimization module, obtaining the updated inference data for each agent. The parameter optimization module is used to calculate the trajectory score of the inference data output by each agent in the previous round according to the trajectory scoring weights, determine the current interaction mode and current interaction parameters based on the trajectory scores, and update the trajectory scoring weights. The communication optimization module is used to process the inference data of each agent according to the current interaction mode and send the processed inference data, the current interaction mode, and the current interaction parameters of each agent to other agents. The agent pool is used to store the agents of the artificial intelligence model. Compared to existing technologies where there is an irreconcilable conflict between consensus formation mechanisms and core objectives such as the accuracy and efficiency of reasoning answers, this approach abandons the mandatory requirement for consensus in multi-agent collaborative reasoning. This avoids the herd mentality and majority rule problems caused by forced consensus, ensuring that agents' viewpoints can evolve independently during collaborative reasoning without requiring consensus as the termination condition. It achieves systematic control over the multi-agent collaborative reasoning process, from initial problem analysis and initial strategy parameter setting to the entire process of multi-round interactive reasoning, forming a standardized and regulated collaborative reasoning process to ensure the orderliness of the reasoning process. It achieves dynamic adaptation and optimization of strategy parameters. Through a parameter optimization module, it updates trajectory scoring weights, interaction modes, and interaction parameters in real time during multiple rounds of interaction, ensuring that strategy parameters fit the actual reasoning situation of each round of agent interaction, avoiding the limitations of fixed parameters on reasoning effectiveness, and improving the adaptability of collaborative reasoning. It improves the efficiency of reasoning data interaction between multiple agents. Through a communication optimization module, it performs targeted processing of reasoning data for different interaction modes, reducing meaningless reasoning data transmission, lowering communication redundancy, and ensuring the efficiency and relevance of reasoning information transmission between agents. This system enables full-process trajectory scoring and quantitative evaluation of agent reasoning data. Based on trajectory scoring, it guides strategy parameter updates, shifting the decision-making basis for multi-agent collaborative reasoning from a purely result-oriented approach to a combination of process and outcome. This accurately identifies the actual contribution of each agent in collaborative reasoning, improving the rationality of the reasoning results. It also ensures the efficiency of multi-agent collaborative reasoning by using preset round control conditions as the termination criteria, avoiding meaningless and lengthy arguments between agents and ensuring that collaborative reasoning proceeds within a controllable range, thus improving the overall timeliness of the system's reasoning decisions.This approach achieves unified collaborative scheduling of heterogeneous intelligent agents. By integrating agents from different AI models through an agent pool, it fully leverages the complementary knowledge and varying reasoning abilities of these agents, promoting the exchange of ideas and divergent thinking among them, thus laying the foundation for extracting high-quality collaborative reasoning data. Furthermore, it ensures that the policy parameters for multi-agent collaborative reasoning are more closely aligned with the characteristics of the original problem. Through a dynamic policy coordination module, it determines the initial policy parameters based on a targeted analysis of the original problem, avoiding the insufficient adaptability of generalized parameters to specific problem reasoning and improving the initial accuracy of collaborative reasoning. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present 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 only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 This is a schematic diagram of a multi-agent collaborative reasoning system in an embodiment of the present invention; Figure 2 This is another schematic diagram of the multi-agent collaborative reasoning system in an embodiment of the present invention; Figure 3 This is a detailed execution logic diagram of the multi-agent collaborative reasoning system in an embodiment of the present invention; Figure 4 This is a detailed execution logic diagram of the interaction between multiple intelligent agents in an embodiment of the present invention; Figure 5 This is a flowchart of the multi-agent cooperative reasoning method in an embodiment of the present invention; Figure 6 This is a schematic diagram of a computer device in an embodiment of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0017] This invention addresses the technical problems in existing multi-agent collaborative frameworks, such as the herding effect caused by mandatory consensus requirements, neglect of process quality due to simple voting, low interaction efficiency, and rigid and non-scalable systems. It provides a multi-agent collaborative reasoning system and method that fundamentally eliminates the mandatory requirement for consensus. By introducing a trajectory-based decision-making mechanism and modular process design, it systematically overcomes the shortcomings of existing technologies. The solutions proposed in this invention can achieve the following objectives: 1. To address the conflict between mandatory consensus and optimal solutions, this invention introduces problem classification. At the initial stage of collaboration, problems are categorized and preprocessed based on task characteristics, providing different adaptive weights. Simultaneously, it guides agents with differentiated initial policy parameters, forming initial stances and discussion directions. This allows minority opinions to be legitimately retained and evolved throughout the decision-making process, avoiding herd mentality and majority dominance. It ensures that the final output is selected based on the optimality of the reasoning process rather than the convergence of conclusions, thus improving the correctness of the decision.

[0018] 2. To achieve a paradigm shift from outcome-based voting to process-based evaluation, this invention constructs a comprehensive quantitative evaluation system by designing a dynamic policy coordination module and a parameter optimization module that runs through the online multi-agent collaborative reasoning stage. This system dynamically configures evaluation weights through the parameter optimization module and continuously evaluates the quality of each agent's reasoning answers, the strength of the reasoning process, the effectiveness of critical analysis, and the rationality of position evolution throughout the entire collaborative process. Based on the overall quality of the reasoning trajectory, it generates the final output, achieving a precise measurement of the individual agent's contribution and the quality of their interaction.

[0019] 3. To achieve an effective balance between in-depth analysis and decision-making efficiency, this embodiment of the invention adjusts interaction parameters in real time during the core interaction process of the online multi-agent collaborative reasoning stage through a parameter optimization module. Simultaneously, a feedback mechanism archives the entire process data of this collaboration and feeds it back to the dynamic strategy coordination module and the parameter optimization module, forming a system-level closed-loop adaptive optimization. This effectively avoids ineffective and lengthy debates while ensuring sufficient intellectual exchange, thereby improving the overall resource utilization and decision-making timeliness of the system.

[0020] 4. To construct a flexible, transparent, and scalable general collaboration framework, this embodiment of the invention adopts a modular decoupling design, explicitly dividing the system into multiple modules. This design enables the system to be compatible with various heterogeneous underlying models and supports the pluggable replacement of scoring and interaction strategies. When introducing new agents or new evaluation criteria, there is no need to refactor the core logic of the system, thereby significantly enhancing the framework's frontier adaptability and scalability, providing complete and auditable interaction trajectories, and improving the interpretability of the decision-making process and the credibility of the system.

[0021] In summary, the purpose of this invention is to provide an efficient, robust, fair, and interpretable multi-agent collaborative reasoning system and method. This solution, without mandating consensus, incentivizes high-quality reasoning and critical interaction through modular process design, full-process quality assessment, and closed-loop adaptive optimization. This allows for the accurate extraction of optimal solutions from swarm intelligence, providing an innovative technological infrastructure for complex problem-solving, creative tasks, and critical decision support.

[0022] Figure 1 This is a schematic diagram of a multi-agent cooperative reasoning system in an embodiment of the present invention, such as... Figure 1 As shown, the system includes: The dynamic strategy coordination module 101 is used to analyze the original question input by the user and determine the initial strategy parameters, which include the interaction mode, interaction parameters, trajectory scoring weights and round control conditions. The multi-round inference module 102 is used to send the original problem and initial policy parameters to the agent pool, obtain the inference data output by each agent, perform multi-round interactions until the round control conditions are met, and determine the collaborative inference data based on the latest inference data output by the agents. In each round of interaction, the policy parameters are updated through the parameter optimization module, and the inference data is processed and sent through the communication optimization module, and the updated inference data of each agent is obtained. The parameter optimization module 103 is used to calculate the trajectory score of the inference data output by each agent in the previous round according to the trajectory score weight, determine the current interaction mode and the current interaction parameters according to the trajectory score, and update the trajectory score weight. The communication optimization module 104 is used to process the inference data of each agent according to the current interaction mode; and to send the processed inference data, the current interaction mode and the current interaction parameters of each agent to other agents. Agent pool 105 is used to store agents for artificial intelligence models.

[0023] Depend on Figure 1It can be seen that, compared with existing technologies where there is an irreconcilable conflict between consensus formation mechanisms and the accuracy and efficiency of reasoning answers, the system proposed in this invention abandons the mandatory requirement for consensus in multi-agent collaborative reasoning, avoiding the herd effect and majority dominance problem caused by mandatory consensus. It ensures that the viewpoints of agents can evolve independently during collaborative reasoning, without requiring consensus as the termination condition for reasoning. It achieves systematic control over the multi-agent collaborative reasoning process, from initial problem analysis to initial strategy parameter setting, and then to the entire process of multi-round interactive reasoning, forming a standardized and regulated collaborative reasoning process, ensuring the orderliness of the reasoning process. It achieves dynamic adaptation and optimization of strategy parameters, updating trajectory scoring weights, interaction modes, and interaction parameters in real time during multiple rounds of interaction through a parameter optimization module. This allows the strategy parameters to fit the actual reasoning situation of each agent in each round, avoiding the limitation of fixed parameters on reasoning effects and improving the adaptability of collaborative reasoning. To improve the efficiency of inference data interaction among multiple agents, a communication optimization module is used to process inference data according to different interaction modes, reducing meaningless data transmission and communication redundancy, and ensuring the efficiency and relevance of inference information transmission between agents. Full-process trajectory scoring and quantitative evaluation of agent inference data is implemented. Based on trajectory scoring, strategy parameter updates are guided, shifting the decision-making basis for multi-agent collaborative inference from a purely result-oriented approach to a combination of process and result. This accurately identifies the actual contribution of each agent in collaborative inference, improving the rationality of the collaborative inference answers. To ensure the efficiency of multi-agent collaborative inference, preset round control conditions are used as the basis for inference termination, avoiding meaningless and lengthy arguments between agents and ensuring that collaborative inference proceeds within a controllable range, thus improving the overall timeliness of the system's inference decisions. Unified collaborative scheduling of heterogeneous agents is achieved by integrating agents from different artificial intelligence models through an agent pool. This fully leverages the complementary knowledge and different reasoning abilities of heterogeneous agents, promoting the collision of viewpoints and divergent thinking among agents, laying the foundation for extracting high-quality collaborative inference data. The strategy parameters for multi-agent collaborative reasoning are made more closely aligned with the characteristics of the original problem. The initial strategy parameters are determined through targeted analysis of the original problem by the dynamic strategy coordination module, avoiding the problem of insufficient adaptability of generalized parameters to reasoning for specific problems and improving the initial accuracy of collaborative reasoning.

[0024] In this embodiment of the invention, the agent pool is the foundation for realizing diverse collective intelligence. Each agent possesses independent reasoning and critical examination capabilities: based on its own independent knowledge base and reasoning ability, it analyzes the original question under preset interaction parameters, generating independent reasoning data containing the reasoned answer and the complete reasoning process; it can receive processed reasoning data from other agents, critically examine their viewpoints, argumentation logic, and reasoning processes, and generate feedback. The heterogeneity among agents, whether stemming from differences in model architecture, capability scale, or different task parsing and construction strategies for the same architecture model, provides the necessary conditions for realizing the collision of viewpoints, knowledge complementarity, and divergent thinking. Each agent instance is encapsulated as an independent microservice, providing a unified request-response interface, capable of receiving questions and context information, and returning answers, arguments, or critical opinions containing the reasoning process.

[0025] In this embodiment, each agent is used for: After receiving the original question and initial policy parameters, inference data is generated based on the interaction mode and interaction parameters in the original question and initial policy parameters. After receiving inference data processed by other intelligent agents, the current interaction mode, and the current interaction parameters, the system corrects its own inference process based on the inference data processed by other intelligent agents, the current interaction mode, and the current interaction parameters, obtains the corrected inference process and the corresponding inference answer, and forms updated inference data.

[0026] In practice, after receiving the user's original question and the initial strategy parameters issued by the system, the intelligent agent combines its own knowledge base, reasoning model and reasoning logic, and independently completes reasoning analysis based on the interaction mode and interaction parameters in the original question and initial strategy parameters, generates reasoning data containing reasoning answers and complete reasoning processes, and feeds the reasoning data back to the multi-round reasoning module. In each round of multi-round interaction, after receiving the inference data processed by other agents, the current interaction mode, and the current interaction parameters sent by the communication optimization module, the agent first analyzes the inference data processed by other agents, focusing on extracting the inference process (including logical deduction steps, evidence, analytical perspective, critical viewpoints, and other core content). Then, combined with the current interaction mode and current interaction parameters (such as anti-conformity strength, debate configuration parameters, and the number of agents configured), the agent makes targeted adjustments to the inference process in its own inference data. Specifically, this includes identifying and filling gaps in its own inference logic, supplementing and improving the evidence, optimizing and expanding the inference perspective, correcting or deleting unreasonable deduction steps, and verifying and optimizing the inference logic in conjunction with the critical viewpoints of other agents. After completing the adjustment of its own inference process, the agent synchronously updates the inference answer in the inference data (if the adjustment of the inference process affects the conclusion), forming updated complete inference data. This updated inference data is then fed back to the multi-round inference module as the inference result of this round of interaction, participating in subsequent trajectory scoring and system parameter optimization.

[0027] In this embodiment, the dynamic policy coordination module includes: The task classification unit is used to extract keywords from the original question and obtain at least one task category label based on the keywords and a preset classification strategy. The keywords include topic words, key verbs, and domain identifiers. The task complexity assessment unit is used to determine the complexity level of the original problem based on at least one task category label and according to a preset assessment dimension. The strategy parameter optimization unit is used to calculate the initial strategy parameters based on at least one task category label and complexity level of the original problem.

[0028] For example, key verbs include explaining, designing, predicting, and comparing, while domain identifiers involve specific categories such as law, medicine, or physics. The prediction classification strategy can be a pre-trained classification model, a task category label rule base, clustering algorithms, matching algorithms, etc., without limitation, and can generate one or more task category labels, such as multi-constraint optimization, open-ended creative generation, or fact-checking and causal reasoning. The task category label rule base stores all predefined task category labels.

[0029] In the embodiments, the preset evaluation dimensions include, but are not limited to, the sentence structure of the original problem, the number of constraints, the fuzzy vocabulary, and the implicit range of the solution space; The task complexity evaluation unit is used for: For each task category label, determine the complexity value for each preset evaluation dimension; For each preset evaluation dimension, the complexity values ​​of all task category labels in that preset evaluation dimension are fused to obtain a fused value; Based on the fusion value of each preset evaluation dimension, the complexity level of the original problem in each preset evaluation dimension is determined.

[0030] Sentence structure involves the number and interrelationships of different decision variables; fuzzy terms include "possible" and "best"; and the implicit scope of the solution space includes its size and degree of openness. The following strategy can be used when calculating the fusion value: If there are multiple task category labels, a fusion strategy is used to calculate the overall complexity score: Strategy 1 (commonly used): Take the maximum value and use the sub-complexity of the most complex label as the overall complexity; Strategy 2: Weighted summation, set weights, and perform weighted summation; Strategy 3: Calculate the average value.

[0031] The complexity level is output as a multi-dimensional vector, thus quantifying the challenge of the original problem.

[0032] When the policy parameter optimization unit calculates the interaction parameters used for agent interaction, it can use built-in optimization algorithms. These algorithms include rule-based mapping, meta-learning initialization, or lightweight Bayesian optimization. These algorithms calculate and generate a complete set of initial policy parameters "tailored" to the current task. The policy parameters are used for subsequent adaptive trajectory scoring, with round control conditions such as maximum number of rounds, group consensus, or trajectory score stabilization. Interaction modes include, but are not limited to, debate, brainstorming, and deliberation, which are used by multiple agents in multi-round dynamic collaborative interactions.

[0033] The dynamic strategy coordination module analyzes the original problem layer by layer to form an accurate task profile. Then, based on this profile, a highly adapted initial strategy parameter is configured before collaborative reasoning begins, thus laying the foundation for efficient and targeted collaboration.

[0034] After obtaining the initial policy parameters, the multi-round inference module sends the original problem and the initial policy parameters to the agent pool to obtain the inference data output by each agent, which is used for subsequent multi-round iterations.

[0035] During multiple iterations, the parameter optimization module needs to continuously update the parameter set, such as interaction parameters and trajectory scoring weights. Its general update logic can be expressed as:

[0036] in, For the optimized parameter set of the (k+1)th round, For the parameter set of the k-th round, This is the status data for the dynamic collaborative interaction of the rounds already executed. This is the archived full-process data. The following describes the update process of interactive parameters and trajectory scoring weights in the parameter set.

[0037] In this embodiment, the parameter optimization module is used for: According to the trajectory scoring weights corresponding to different trajectory analysis dimensions, a weighted sum is calculated for the inference data output by each agent. The inference data includes the inference answer and the inference process. The trajectory analysis dimensions include, but are not limited to, the quality of the inference answer, the strength of the inference process, the critical effectiveness, and the rationality of the evolution of the position. The weighted sum is multiplied by the scoring factor of the current round to obtain the trajectory score of the inference data output by each agent.

[0038] Trajectory score weights are implicitly passed to the agent, allowing it to adjust its reasoning and answer generation focus based on these weights when independently analyzing problems. The initial values ​​of the trajectory score weights are included in the initial policy parameters, and subsequent weights are updated over multiple iterations.

[0039] The quality of the reasoning answer is used to evaluate the quality of the reasoning result; the strength of the reasoning process is used to evaluate the logical rigor, relevance of evidence, and depth of reasoning provided by the reasoning process; critical effectiveness is used to evaluate the accuracy of identifying loopholes or logical flaws in the reasoning process, as well as the constructive value of the proposed rebuttals or improvement suggestions; and the rationality of the position evolution is used to evaluate whether the basis for revising the reasoning answer is sufficient and whether the reasoning process is logical. Each trajectory analysis dimension has a corresponding trajectory scoring weight. For example, for highly complex creative tasks, and for tasks categorized as open-ended creative generation, critical effectiveness and rationality of position evolution will have higher trajectory scoring weights.

[0040] For example, in highly complex creative tasks, higher trajectory scoring weights are given to the effectiveness of criticism and the rationality of the evolution of positions. When generating reasoning answers and reasoning processes, the agent will deliberately strengthen the innovativeness of viewpoints and the iterability of reasoning, rather than just pursuing the absolute correctness of reasoning answers.

[0041] Interaction modes are used to constrain the format, level of detail, and content boundaries of the reasoning process and the reasoning answers generated by the agent.

[0042] For example, if the interaction mode is brainstorming, the reasoning process generated by the agent will focus on creative derivation logic, and the reasoning answer will be more open-ended, without the need for overly rigorous argumentation loops, which conforms to the "divergent creativity" mode rule; if the interaction mode is debate, the reasoning process generated by the agent will focus on logical loops and evidence support, and the reasoning answer will clearly state the viewpoint and position, and the reasoning process will reserve logical nodes that can be criticized and refuted, which conforms to the "debate and analysis" mode rule; if the interaction mode is deliberation, the reasoning process generated by the agent will focus on the trade-off analysis of multiple constraints, and the reasoning answer will include a comparison of different solutions, which conforms to the "prudent decision-making" mode rule.

[0043] Round control conditions (such as maximum rounds and score change thresholds that trigger deep reflection) allow the agent to establish expectations for reasoning iterations in advance when independently generating initial insights, reserving reasonable space for correction and iteration during the reasoning process, rather than generating absolute reasoning answers that are "uncorrectable and have no possibility of iteration".

[0044] For example, if the maximum number of rounds is large, the agent will deliberately retain some incomplete reasoning nodes when generating the first round of reasoning, reserving space for subsequent iterations. For example, if a low threshold for triggering deep reflection is set, the agent will be more cautious in choosing viewpoints when generating reasoning answers, and the reasoning process will pay more attention to the rigor of logic, avoiding triggering deep reflection due to small score changes and overturning its own core viewpoints.

[0045] The scoring factor was used when calculating the trajectory score weight.

[0046] In this embodiment, the calculation formula for the scoring factor is as follows:

[0047] in, Let be the scoring factor for the k-th round. The base value of the rating factor. For the current round, The descent velocity coefficient, This represents the maximum number of interaction rounds.

[0048] The scoring factor, acting as a scoring decay coefficient, controls the rate at which the influence of historical trajectory scores on the current trajectory score decays. Replacing a fixed decay factor, it becomes an adjustable parameter, allowing for more flexible agent management. When the task requires long-term memory and continuous contribution evaluation, the system may increase the scoring factor. This results in a slow decay of historical trajectory scores, giving greater weight to the agent's long-term performance. When the task emphasizes recent performance and rapid adaptation, the system may lower the score. This gives greater weight to trajectory scores from recent rounds, making the trajectory scores more reflective of the agent's current capability status.

[0049] In this embodiment, the parameter optimization module is used for: Based on the inference data output by each agent in the previous round and the updated inference data, analyze the position change of each agent, and determine the position consistency of each agent based on the position change. Calculate the average positional consistency of all agents to obtain the average positional consistency; Calculate the difference between the average position consistency and the previous target position consistency, and determine the current target position consistency based on the difference; Based on the average stance consistency and the current target stance consistency, the trajectory score weights are updated using the following formula:

[0050] in, The updated trajectory score weights for round k+1. The trajectory score weight for the k-th round is... For learning rate, To maintain consistency in average stance, To maintain consistency with current goals and positions.

[0051] The above-mentioned trajectory scoring weight update steps are applied to any trajectory analysis dimension, including the quality of the reasoning answer, the strength of the reasoning process, the effectiveness of the critique, and the rationality of the evolution of the stance.

[0052] The update strategy for trajectory score weights is as follows: If the system detects that premature or excessively frequent changes in stance lead to wavering opinions and difficulty in forming consensus, it may increase the weight of increasing the penalty for changing the answer and decrease the weight of decreasing the reward for the new answer to encourage more deliberate changes in stance. Conversely, if the system detects that the agent is too stubborn and refuses to accept reasonable rebuttals, resulting in rigid thinking, it may decrease the weight of increasing the penalty for changing the answer and increase the weight of decreasing the reward for the new answer to incentivize reasonable evolution in the face of strong arguments. In the later stages of the interaction, to promote consensus formation, the trajectory score weight for maintaining the initial reasoning answer may be appropriately increased to affirm the behavior of adhering to a viewpoint that remains stable after multiple rounds of debate.

[0053] In this embodiment, the interaction parameters include debate configuration parameters and trigger threshold parameters for different interaction modes; The parameter optimization module is used to update the debate configuration parameters and the trigger threshold parameters for different interaction modes based on the dynamic collaborative interaction status data of the executed rounds and the archived full-process data.

[0054] In specific implementation, the dynamic collaborative interaction status data of the executed rounds includes the dynamic collaborative interaction status data of the rounds already executed in this multi-round interaction, specifically including trajectory scoring, group interaction status indicators (such as the diversity of group viewpoints, consistency of positions, variance of score distribution, resolution of core conflicts, etc., which are key evidence to judge whether the interaction is in-depth and whether group thinking has emerged); and the characteristics of the agent's response content in this round (such as whether there is excessive debate, convergence of viewpoints, or the proportion of constructive reasoning, etc.).

[0055] The archived end-to-end data includes previously archived dynamic collaborative interaction status data in the end-to-end database, which serves as a long-term / empirical basis for parameter optimization.

[0056] Adjusting the debate configuration parameter R: If, after a preset number of rounds, the variance of the trajectory scores of all agents remains large, or key conflicts remain unresolved, the value of R in subsequent stages may be dynamically increased to allow for more in-depth specialized debates. Conversely, if a high degree of consensus has been reached early on, R can be reduced to save computational resources. The specific steps include: extracting the trajectory score distribution data of all agents in the current round, recording the conflict points of the core reasoning viewpoints, and the configuration parameter R value of the previous round of debate; calculating the variance of the trajectory scores, the larger the variance value, the higher the degree of group disagreement; simultaneously counting the number of unresolved core conflict points; determining whether the current round has reached the preset debate depth assessment round, if not, keeping the R value unchanged; if it has, proceeding to the next step; if the score variance value is greater than the preset variance threshold and the number of unresolved core conflict points is greater than the preset value: increasing the R value by a preset fixed step size, allowing for more in-depth specialized debates in the future; if the score variance value is less than the preset variance threshold and all core conflict points have been resolved: decreasing the R value by a preset fixed step size, reducing meaningless deep reasoning; if none of the above conditions are met: keeping the R value unchanged. Comparing the adjusted R value with the preset value range, if it exceeds the range, taking the boundary value; determining the configuration parameter R value for the (k+1)th round of debate and passing it to the communication optimization module.

[0057] The steps for updating the trigger threshold parameters for different interaction modes may include: extracting group interaction status indicators such as the degree of group divergence, consistency of positions, variance of score distribution, and resolution of core conflicts in the current round; analyzing the matching degree between the status indicators and the current interaction mode; if the interaction efficiency is low in the current mode (e.g., there is no core conflict in the debate mode but the debate continues), then adjusting the trigger threshold of the corresponding mode; fine-tuning the trigger thresholds of the three interaction modes of debate, brainstorming, and deliberation in small steps (e.g., increasing the trigger threshold of the debate mode and decreasing the trigger threshold of the brainstorming mode); determining the adjusted interaction mode trigger thresholds as the effective thresholds for the next round and passing them to the dynamic strategy coordination module as the basis for interaction mode decision-making.

[0058] In this embodiment, the interaction parameters also include anti-conformity strength, which is used to control the severity of the agent's criticism of other agents; The parameter optimization module is used for: Calculate the difference in reasoning answers among agents based on the trajectory scores of all agents; The population diversity value is obtained by aggregating and calculating the differences between all reasoning answers. Based on the group diversity value and the pre-set target opinion diversity value, the anti-conformity strength is calculated using the following formula:

[0059] in, The anti-herd strength in round k+1, Let the anti-conformity strength be the strength in round k. For coefficients, To preset the target opinion diversity value, This represents the population diversity value.

[0060] The aforementioned anti-conformity strength controls the severity with which an agent criticizes other agents. If the system detects signs of "groupthink," it may increase the anti-conformity strength to encourage stronger critical scrutiny. If the interaction devolves into excessive debate and lacks constructive contention, it may decrease the anti-conformity strength, prompting agents to focus more on the constructive aspects of the argument. This dynamic adjustment can be linked to indicators of group opinion diversity.

[0061] The parameter optimization module dynamically adjusts the aforementioned interaction parameters and trajectory scoring weights based on the dynamic collaborative interaction status data of the executed rounds and the archived full-process data, using online learning or optimization algorithms. Its optimization goal is to ensure that the evaluation system and scoring results can most effectively drive the group towards improving decision-making quality or efficiency.

[0062] In this embodiment, the interaction parameters also include the number of intelligent agents configured; The parameter optimization module is used for: After obtaining the trajectory scores of the inference data output by all agents participating in the trajectory score calculation, calculate the variance of all trajectory scores. Based on the variance, determine the number of agents configured to participate in the trajectory scoring calculation; Based on the number of configured agents, a set of agents is selected from the agent pool to participate in trajectory scoring calculation; The communication optimization module is used to send the inference data processed by each agent, the current interaction mode, and the current interaction parameters to the agents in the agent set.

[0063] In this embodiment, the communication optimization module is used for: If the current interaction mode is debate mode, extract key information from the reasoning data of each agent to obtain the processed reasoning data of each agent. If the current interaction mode is brainstorming mode, perform lightweight semantic compression on the inference data to obtain the processed inference data for each agent. If the current interaction mode is deliberation mode, the core decision-making basis is extracted from the reasoning data to obtain the processed reasoning data for each agent.

[0064] In practical implementation, the communication optimization module is primarily responsible for optimizing the bandwidth efficiency and response latency of information exchange between agents. Based on the current interaction mode, this module implements targeted information processing strategies and manages the concurrent sending, receiving, and asynchronous processing mechanisms of inference data to minimize overall system latency caused by communication. If the interaction mode is debate, key information is extracted from the inference data, preserving complete reasoning and critical logic without excessive compression. If the interaction mode is brainstorming, lightweight semantic compression is applied to the inference data, prioritizing the integrity of viewpoints and transmission speed. If the current interaction mode is deliberation, core decision-making criteria are extracted from the inference data, including problem constraint analysis, comparison of advantages and disadvantages of multiple solutions, feasibility analysis of solutions, key decision-making indicators / data support, and multi-dimensional trade-off logic. Excessive compression is avoided, focusing on ensuring the complete logic of the agent's comparative analysis of different solutions, constraint trade-offs, and decision-making reasoning, ensuring other agents can clearly understand the deliberation process. Redundant expressions unrelated to solution comparison and constraint trade-offs, such as creative divergent ideas and unfounded argument details, are removed from the inference data.

[0065] The communication optimization module can initiate concurrent message sending, receiving, and asynchronous processing mechanisms, monitor the transmission status in real time, avoid data blocking, and minimize communication latency; it records the data processing mode, the amount of data processed, the transmission latency, and the targeted distribution results as basic data for offline learning.

[0066] In this embodiment, the multi-round inference module is used for: Calculate the trajectory score of the inference data output by each agent when the round control conditions are met, and use the inference data with the highest trajectory score as the collaborative inference data of the multi-agent system.

[0067] The interaction ends when the round control conditions are met. The trajectory score is calculated according to the final trajectory score weight of each agent, and the inference data with the highest trajectory score is used as the collaborative inference data of the multi-agents. The inference answer in the collaborative inference data is the final viewpoint after multiple rounds of evolution, and is determined as the output solution.

[0068] In this embodiment, the multi-round inference module is used for: A decision tracing report is generated based on user questions, task category tags and corresponding complexity levels, initial strategy parameters, and historical interaction status data.

[0069] The decision traceability report can include data from the entire process, that is, data involved in all the above steps. This data is standardized and packaged, and stored as a complete experience in the full-process database.

[0070] Figure 2 This is another schematic diagram of a multi-agent cooperative reasoning system in an embodiment of the present invention. In this embodiment, the system further includes a feedback mechanism module 201, used for: Collect all data from each inference process and archive it into the full-process database; Based on the full-process data in the full-process database, the dynamic strategy coordination module and parameter optimization module are driven to perform offline learning.

[0071] After the multi-round reasoning module generates a decision tracing report, asynchronous offline learning and system evolution are carried out. This is the core closed loop for the system to achieve continuous self-improvement. The offline learning process is driven by coordinating the full-process data in the full-process database.

[0072] First, the accumulated full-process data is used to retrain and fine-tune the classification models and algorithms in the task classification unit and task complexity evaluation unit within the dynamic strategy coordination center, aiming to improve the initial analytical accuracy and evaluation accuracy of the classification models and algorithms for various new problems.

[0073] Next, the optimization algorithm for the policy parameter optimization unit is trained, using the entire process data as training samples and inputting them into the optimization algorithm, such as a neural network or Bayesian optimization model for meta-learning. By learning the mapping relationship between the task profile and the validated interaction parameter configuration, the quality of initial interaction parameter recommendations for future new tasks is improved.

[0074] Next, the parameter optimization module is tuned. Using offline reinforcement learning or global optimization methods, the underlying strategy of the parameter optimization function in this module is iteratively updated to make its online dynamic adjustment more accurate and efficient, thereby better guiding the collaboration of swarm intelligence.

[0075] Finally, the system iteration and deployment are completed. After completing the offline learning and optimization described above, new version parameters or models of the system are generated. After verification and testing, these updates are smoothly deployed to the online system, enabling the entire system to have better initial configuration and dynamic coordination capabilities in subsequent tasks, achieving continuous performance evolution based on experience.

[0076] In the multi-agent collaborative reasoning system described in this invention, the multi-round dynamic collaborative interaction and optimization process is the core of its continuous performance improvement. This process involves not only the exchange and iteration of viewpoints among agents, but more importantly, the dynamic adjustment of a series of key parameters within the system. These adjustments are driven by the parameter optimization function in the parameter optimization module, aiming to enable the evaluation system and group interaction mode to adapt to task progress and agent performance in real time, thereby most effectively driving the group to evolve towards improving decision-making quality or efficiency.

[0077] The multi-agent collaborative inference system employs a layered, heterogeneous computing architecture. The core computing modules, including a dynamic policy coordination module, a parameter optimization module, and a large number of general-purpose agents, are preferably deployed in a centralized data center equipped with high-performance GPU server clusters to meet its high-intensity parallel computing and model inference requirements. To reduce interaction latency and handle domain-specific tasks, some lightweight or specialized agents can be deployed on edge computing nodes closer to the data source. All functional modules are containerized and decoupled through a microservices-based architecture, with modules exchanging data efficiently and reliably through well-defined interfaces. The entire system is preferably scheduled, managed, and elastically scaled using a container orchestration platform such as Kubernetes to ensure high availability and scalability.

[0078] Figure 3 The detailed execution logic diagram of the multi-agent cooperative reasoning system in this embodiment of the invention includes: Original Question Reception: The system obtains the original question to be reasoned from the user input and passes it to the dynamic strategy coordination module; Initial strategy parameter determination: The dynamic strategy coordination module performs a structured analysis of the original problem, generates task category labels and complexity levels, and generates initial strategy parameters that include interaction modes, interaction parameters, trajectory scoring weights, and round control conditions. Agent pool initialization and distribution: The multi-round inference module sends the original problem and initial policy parameters to all heterogeneous agents in the agent pool; First-round inference data generation: Each agent in the agent pool independently analyzes the original problem based on an independent knowledge base and initial policy parameters, generates first-round inference data containing the inference answer and inference process, and submits it to the multi-round inference module through the communication optimization module; Multi-round interaction trigger judgment: The multi-round reasoning module checks whether the round control conditions are met. If not, the multi-round interaction process is triggered; if met, it directly enters the reasoning answer selection stage. Inter-round policy parameter update: This includes trajectory scoring and policy parameter update. The parameter optimization module performs full-dimensional trajectory scoring on the inference data output by all agents in the previous round according to the current trajectory score weight. Based on the scoring results, it determines the current interaction mode and the current interaction parameters, and updates the trajectory score weight. Inference data processing: The communication optimization module receives the current interaction mode from the parameter optimization module and performs pattern adaptation processing on the inference data of each agent in the previous round (such as key information extraction, semantic compression, etc.). Inference data targeted distribution: The communication optimization module sends the processed inference data, the current interaction mode, and the current interaction parameters to other agents in the agent pool; Agent inference data update: The agent receives information distributed by the communication optimization module, critically examines the inference data of other agents, combines it with its own original inference logic, updates and generates a new round of inference data, and submits it to the multi-round inference module again through the communication optimization module. Collaborative inference data selection: The multi-round inference module selects the inference data with the highest score as collaborative inference data based on the final trajectory scores of all agents output by the parameter optimization module. Output: The multi-round inference module outputs the collaborative inference data as the final solution of the system.

[0079] Offline learning feedback: The feedback mechanism module generates full-process data to enable offline learning of the models and algorithms in the parameter optimization module and dynamic strategy coordination module, thereby achieving system iteration and deployment.

[0080] Figure 4 This is a detailed execution logic diagram of the interaction between multiple intelligent agents in an embodiment of the present invention, including: Initialize N agents; Each agent generates initial inference data; Trajectory scoring is performed on each initial inference data based on the trajectory scoring weights in the initial policy parameters; Update rating factors; Analyze the changes in the stance of each agent; Update the trajectory score weights; Perform iterative loops to determine whether the round control conditions are met; If not, collect the inference data of all agents from the previous round; Analyze and adjust the reasoning data of each agent; Obtain the reasoning data for the current round; Calculate the trajectory score for the current round, and then enter the loop iteration again; If so, find the highest trajectory score; Output the collaborative inference data corresponding to the highest trajectory score.

[0081] This invention also provides a multi-agent cooperative reasoning method, as described in the following embodiments. Since the principle behind this method is similar to that of the multi-agent cooperative reasoning system, its implementation can be found in the implementation of the multi-agent cooperative reasoning system; repeated details will not be elaborated further.

[0082] Figure 5 This is a flowchart of the multi-agent cooperative reasoning method in an embodiment of the present invention, such as... Figure 5 As shown, the method includes: Step 501: Analyze the original question input by the user and determine the initial strategy parameters, which include the interaction mode, interaction parameters, trajectory scoring weights, and round control conditions. Step 502: Send the original problem and initial policy parameters to the agent pool to obtain the inference data output by each agent; Step 503: Perform multiple rounds of interaction until the round control conditions are met. During each round of interaction, a trajectory score is calculated on the inference data output by each agent in the previous round, based on the trajectory score weight. The current interaction mode and current interaction parameters are determined based on the trajectory score, and the trajectory score weight is updated. The inference data of each agent is processed according to the current interaction mode. The processed inference data, current interaction mode, and current interaction parameters of each agent are sent to other agents. Updated inference data for each agent is obtained. Step 504: Determine collaborative reasoning data based on the latest reasoning data output by the agent.

[0083] In this embodiment, the original question input by the user is analyzed to determine the initial policy parameters, including: Extract keywords from the original question, and based on the keywords and a preset classification strategy, obtain at least one task category label. The keywords include topic words, key verbs, and domain identifiers. Based on at least one task category label, the complexity level of the original problem is determined according to a preset evaluation dimension; Calculate the initial strategy parameters based on at least one task category label and complexity level of the original problem.

[0084] In the embodiments, the preset evaluation dimensions include, but are not limited to, the sentence structure of the original problem, the number of constraints, the fuzzy vocabulary, and the implicit range of the solution space; Based on at least one task category label, the complexity level of the original problem is determined according to preset evaluation dimensions, including: For each task category label, determine the complexity value for each preset evaluation dimension; For each preset evaluation dimension, the complexity values ​​of all task category labels in that preset evaluation dimension are fused to obtain a fused value; Based on the fusion value of each preset evaluation dimension, the complexity level of the original problem in each preset evaluation dimension is determined.

[0085] In this embodiment, a trajectory score is calculated for the inference data output by each agent in the previous round, according to the trajectory score weight, including: According to the trajectory scoring weights corresponding to different trajectory analysis dimensions, a weighted sum is calculated for the inference data output by each agent. The inference data includes the inference answer and the inference process. The trajectory analysis dimensions include, but are not limited to, the quality of the inference answer, the strength of the inference process, the critical effectiveness, and the rationality of the evolution of the position. The weighted sum is multiplied by the scoring factor of the current round to obtain the trajectory score of the inference data output by each agent.

[0086] In this embodiment, the calculation formula for the scoring factor is as follows:

[0087] in, Let be the scoring factor for the k-th round. The base value of the rating factor. For the current round, The descent velocity coefficient, This represents the maximum number of interaction rounds.

[0088] In this embodiment, updating the trajectory scoring weights includes: Based on the inference data output by each agent in the previous round and the updated inference data, analyze the position change of each agent, and determine the position consistency of each agent based on the position change. Calculate the average positional consistency of all agents to obtain the average positional consistency; Calculate the difference between the average position consistency and the previous target position consistency, and determine the current target position consistency based on the difference; Based on the average stance consistency and the current target stance consistency, the trajectory score weights are updated using the following formula:

[0089] in, The updated trajectory score weights for round k+1. The trajectory score weight for the k-th round is... For learning rate, To maintain consistency in average stance, To maintain consistency with current goals and positions.

[0090] In this embodiment, the interaction parameters include debate configuration parameters and trigger threshold parameters for different interaction modes; The current interaction mode and current interaction parameters are determined based on the trajectory score, including: Based on the dynamic collaborative interaction status data of the executed rounds and the archived full-process data, update the debate configuration parameters and the trigger threshold parameters for different interaction modes.

[0091] In this embodiment, the interaction parameter further includes anti-conformity strength; The current interaction mode and current interaction parameters are determined based on the trajectory score, including: Calculate the difference in reasoning answers among agents based on the trajectory scores of all agents; The population diversity value is obtained by aggregating and calculating the differences between all reasoning answers. Based on the group diversity value and the pre-set target opinion diversity value, the anti-conformity strength is calculated using the following formula:

[0092] in, The anti-herd strength in round k+1, Let the anti-conformity strength be the strength in round k. For coefficients, To preset the target opinion diversity value, This represents the population diversity value.

[0093] In this embodiment, the interaction parameters also include the number of intelligent agents configured; The current interaction mode and current interaction parameters are determined based on the trajectory score, including: After obtaining the trajectory scores of the inference data output by all agents participating in the trajectory score calculation, calculate the variance of all trajectory scores. The method further includes: Based on the variance, determine the number of agents configured to participate in the trajectory scoring calculation; Based on the number of configured agents, a set of agents is selected from the agent pool to participate in trajectory scoring calculation; Each agent sends its processed inference data, current interaction mode, and current interaction parameters to the agents in the agent set.

[0094] In this embodiment, the inference data of each agent is processed according to the current interaction mode, including: If the current interaction mode is debate mode, extract key information from the reasoning data of each agent to obtain the processed reasoning data of each agent. If the current interaction mode is brainstorming mode, perform lightweight semantic compression on the inference data to obtain the processed inference data for each agent. If the current interaction mode is deliberation mode, the core decision-making basis is extracted from the reasoning data to obtain the processed reasoning data for each agent.

[0095] In this embodiment, collaborative reasoning data is determined based on the latest reasoning data output by the agent, including: Calculate the trajectory score of the inference data output by each agent when the round control conditions are met, and use the inference data with the highest trajectory score as the collaborative inference data of the multi-agent system.

[0096] In this embodiment, the method further includes: A decision tracing report is generated based on user questions, task category tags and corresponding complexity levels, initial strategy parameters, and historical interaction status data.

[0097] In this embodiment, the method further includes: Collect all data from each inference process and archive it into the full-process database; Based on the full-process data in the full-process database, the dynamic strategy coordination module and parameter optimization module are driven to perform offline learning.

[0098] This invention also provides a computer device. Figure 6 This is a schematic diagram of a computer device in an embodiment of the present invention. The computer device 600 includes a memory 610, a processor 620, and a computer program 630 stored in the memory 610 and executable on the processor 620. When the processor 620 executes the computer program 630, it implements the above-mentioned multi-agent cooperative reasoning method.

[0099] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned multi-agent cooperative reasoning method.

[0100] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described multi-agent cooperative reasoning method.

[0101] The system and method proposed in the embodiments of the present invention have the following beneficial effects: 1. The system of the present invention can effectively identify and adopt solutions based on high-quality reasoning, especially in scenarios where there are disagreements on the answers. It ensures that the final output decision has the best internal logical quality by using multi-dimensional trajectory scoring rather than simple voting.

[0102] 2. By introducing task-driven dynamic initial strategy parameter configuration and real-time interactive parameter optimization mechanism, the system significantly improves the utilization efficiency of heterogeneous computing resources and the overall decision-making speed of collaborative reasoning, while avoiding meaningless interactive consumption.

[0103] 3. The modular and microservice design of the system supports flexible expansion and convenient adaptation of intelligent agents, evaluation dimensions and interaction modes. Its full-process trajectory recording and traceability report generation functions greatly enhance the transparency and interpretability of complex decision-making processes.

[0104] 4. This framework provides general, self-evolving infrastructure support for a variety of complex scenarios that require in-depth analysis and collective wisdom, such as collaborative code generation, academic seminar assistance, complex task planning, and risk assessment.

[0105] 5. The system's adaptive optimization closed loop can dynamically adjust the trajectory scoring weights and interaction modes based on real-time interaction effects, enabling the system to have online learning capabilities and thus maintain a highly efficient group collaboration state when facing tasks with different characteristics.

[0106] 6. The communication optimization module, through intelligent information processing, significantly reduces the bandwidth requirements and latency of large-scale communication between multiple agents while ensuring the integrity of core semantics, enabling the system to operate efficiently in a distributed deployment environment.

[0107] 7. The offline learning design of the feedback mechanism module constitutes the system's continuous evolution engine, which can continuously extract experience from historical tasks, optimize strategy parameters, and enable the system performance to systematically improve with the number of uses, thus possessing long-term vitality.

[0108] 8. The system’s tolerance and utilization of the heterogeneity of agents effectively promotes the collision and complementarity between different knowledge backgrounds and reasoning styles, which can stimulate innovative solutions that are difficult to generate with a single model and broaden the boundaries of problem solving.

[0109] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0110] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0111] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0112] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0113] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-agent cooperative reasoning system, characterized in that, include: The dynamic strategy coordination module is used to analyze the original question input by the user and determine the initial strategy parameters, which include the interaction mode, interaction parameters, trajectory scoring weights, and round control conditions. The multi-round inference module is used to send the original problem and initial policy parameters to the agent pool, obtain the inference data output by each agent, perform multiple rounds of interaction until the round control conditions are met, and determine the collaborative inference data based on the latest inference data output by the agents. In each round of interaction, the policy parameters are updated through the parameter optimization module, and the inference data is processed and sent through the communication optimization module, and the updated inference data of each agent is obtained. The parameter optimization module is used to calculate the trajectory score of the inference data output by each agent in the previous round according to the trajectory score weight, determine the current interaction mode and the current interaction parameters based on the trajectory score, and update the trajectory score weight. The communication optimization module is used to process the inference data of each agent according to the current interaction mode; and to send the processed inference data, the current interaction mode, and the current interaction parameters of each agent to other agents. The agent pool is used to store agents for artificial intelligence models.

2. The system as described in claim 1, characterized in that, The dynamic strategy coordination module includes: The task classification unit is used to extract keywords from the original question and obtain at least one task category label based on the keywords and a preset classification strategy. The keywords include topic words, key verbs, and domain identifiers. The task complexity assessment unit is used to determine the complexity level of the original problem based on at least one task category label and according to a preset assessment dimension. The strategy parameter optimization unit is used to calculate the initial strategy parameters based on at least one task category label and complexity level of the original problem.

3. The system as described in claim 2, characterized in that, The preset evaluation dimensions include, but are not limited to, the sentence structure of the original question, the number of constraints, the fuzzy vocabulary, and the implicit scope of the solution space; The task complexity evaluation unit is used for: For each task category label, determine the complexity value for each preset evaluation dimension; For each preset evaluation dimension, the complexity values ​​of all task category labels in that preset evaluation dimension are fused to obtain a fused value; Based on the fusion value of each preset evaluation dimension, the complexity level of the original problem in each preset evaluation dimension is determined.

4. The system as described in claim 1, characterized in that, The parameter optimization module is used for: According to the trajectory scoring weights corresponding to different trajectory analysis dimensions, a weighted sum is calculated for the inference data output by each agent. The inference data includes the inference answer and the inference process. The trajectory analysis dimensions include, but are not limited to, the quality of the inference answer, the strength of the inference process, the critical effectiveness, and the rationality of the evolution of the position. The weighted sum is multiplied by the scoring factor of the current round to obtain the trajectory score of the inference data output by each agent.

5. The system as described in claim 4, characterized in that, The calculation formula for the rating factor is as follows: in, Let be the scoring factor for the k-th round. This is the base value for the rating factor. For the current round, The descent velocity coefficient, This represents the maximum number of interaction rounds.

6. The system as described in claim 1, characterized in that, The parameter optimization module is used for: Based on the inference data output by each agent in the previous round and the updated inference data, analyze the position change of each agent, and determine the position consistency of each agent based on the position change. Calculate the average positional consistency of all agents to obtain the average positional consistency; Calculate the difference between the average position consistency and the previous target position consistency, and determine the current target position consistency based on the difference; Based on the average stance consistency and the current target stance consistency, the trajectory score weights are updated using the following formula: in, The updated trajectory score weights for round k+1. The trajectory score weight for the k-th round is... For learning rate, To maintain consistency in average stance, To maintain consistency with current goals and positions.

7. The system as described in claim 1, characterized in that, The interaction parameters include debate configuration parameters and trigger threshold parameters for different interaction modes; The parameter optimization module is used to update the debate configuration parameters and the trigger threshold parameters for different interaction modes based on the dynamic collaborative interaction status data of the executed rounds and the archived full-process data.

8. The system as described in claim 1, characterized in that, The interaction parameters also include anti-conformity strength, which is used to control the severity of the agent's criticism of other agents. The parameter optimization module is used for: Calculate the difference in reasoning answers among agents based on the trajectory scores of all agents; The population diversity value is obtained by aggregating and calculating the differences between all reasoning answers. Based on the group diversity value and the pre-set target opinion diversity value, the anti-conformity strength is calculated using the following formula: in, The anti-conformity strength in round k+1, Let the anti-conformity strength be the value in round k. For coefficients, To preset the target opinion diversity value, This represents the population diversity value.

9. The system as described in claim 1, characterized in that, The interaction parameters also include the number of intelligent agents configured; The parameter optimization module is used for: After obtaining the trajectory scores of the inference data output by all agents participating in the trajectory score calculation, calculate the variance of all trajectory scores. Based on the variance, determine the number of agents configured to participate in the trajectory scoring calculation; Based on the number of configured agents, a set of agents is selected from the agent pool to participate in trajectory scoring calculation; The communication optimization module is used to send the inference data processed by each agent, the current interaction mode, and the current interaction parameters to the agents in the agent set.

10. The system as claimed in claim 1, characterized in that, The communication optimization module is used for: If the current interaction mode is debate mode, extract key information from the reasoning data of each agent to obtain the processed reasoning data of each agent. If the current interaction mode is brainstorming mode, perform lightweight semantic compression on the inference data to obtain the processed inference data for each agent. If the current interaction mode is deliberation mode, the core decision-making basis is extracted from the reasoning data to obtain the processed reasoning data for each agent.

11. The system as claimed in claim 1, characterized in that, Each agent is used for: After receiving the original question and initial policy parameters, inference data is generated based on the interaction mode and interaction parameters in the original question and initial policy parameters. After receiving inference data processed by other intelligent agents, the current interaction mode, and the current interaction parameters, the system corrects its own inference process based on the inference data processed by other intelligent agents, the current interaction mode, and the current interaction parameters, obtains the corrected inference process and the corresponding inference answer, and forms updated inference data.

12. The system as claimed in claim 1, characterized in that, Multi-turn inference module, used for Calculate the trajectory score of the inference data output by each agent when the round control conditions are met, and use the inference data with the highest trajectory score as the collaborative inference data of the multi-agent system.

13. The system as claimed in claim 1, characterized in that, Multi-turn inference module, used for A decision tracing report is generated based on user questions, task category tags and corresponding complexity levels, initial strategy parameters, and historical interaction status data.

14. The system as claimed in claim 1, characterized in that, It also includes a feedback mechanism module, used for: Collect all data from each inference process and archive it into the full-process database; Based on the full-process data in the full-process database, the dynamic strategy coordination module and parameter optimization module are driven to perform offline learning.

15. A multi-agent cooperative reasoning method, characterized in that, include: Analyze the original question input by the user to determine the initial strategy parameters, which include the interaction mode, interaction parameters, trajectory scoring weights, and round control conditions; The original problem and initial policy parameters are sent to the pool of agents to obtain the inference data output by each agent. The process involves multiple rounds of interaction until the round control conditions are met. During each round, a trajectory score is calculated for the inference data output by each agent in the previous round, based on the trajectory score weights. The current interaction mode and parameters are determined according to the trajectory score, and the trajectory score weights are updated. The inference data of each agent is then processed according to the current interaction mode. The processed inference data, current interaction mode, and current interaction parameters of each agent are sent to other agents. Finally, updated inference data for each agent is obtained. Based on the latest inference data output by the agent, determine the collaborative inference data.

16. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of claim 15.

17. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method of any one of claims 15.

18. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the method of any one of claims 15.

Images