An intelligent question and answer system and method based on multi-agent collaborative reasoning

Abstract

This invention discloses an intelligent question-answering system and method based on multi-agent collaborative reasoning. First, by introducing a scheduler module and a modular expert pool, it achieves precise decomposition of complex tasks and dynamic invocation of professional intelligent agents, improving the adaptability to multimodal and cross-domain problems. Second, it adopts a collaborative reasoning mechanism that combines parallelism and pipelined approaches, ensuring processing efficiency while supporting the structured construction of causal chains. Third, it uses a verification committee module to cross-validate reasoning conclusions and resolve conflicts, replacing the traditional self-reflection mode with a multi-expert consensus mechanism, significantly enhancing the system's fault resistance and reasoning robustness. Finally, through a confidence-weighted synthesis and streaming output mechanism, it generates natural, fluent, and interpretable answers with reasoning tracing, significantly improving user trust and interactive experience.

Description

Technical Field

[0001] This invention relates to the field of intelligent education technology, specifically to an intelligent question-answering system and method based on multi-agent collaborative reasoning. Background Technology

[0002] With the development of artificial intelligence technology, especially the widespread application of large language models in natural language understanding and generation, question-answering systems based on monolithic models have been widely used in fields such as education tutoring, technical support, and intelligent customer service.

[0003] In existing technologies, mainstream methods rely on a single model to improve reasoning accuracy through self-reflection or chain-of-thought (CoT) mechanisms, such as multiple sampling, consensus voting, or post-validation to enhance output reliability. However, these methods are essentially still within the "monopolistic enhancement" paradigm, where the reasoning process is confined within the same model, lacking cross-validation capabilities from external perspectives. When faced with tasks involving multimodal inputs (such as text-image combination problems), cross-domain knowledge fusion, or complex causal inference, they are prone to erroneous conclusions due to model illusions, knowledge blind spots, or logical breaks. Furthermore, existing systems generally lack fine-grained interpretability support for the reasoning process, making it difficult for users to trace the answer generation path, thus limiting their application in high-confidence scenarios. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention aims to provide an intelligent question-answering system and method based on multi-agent collaborative reasoning.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: An intelligent question-answering system based on multi-agent collaborative reasoning includes a user input interface, a scheduler module, an expert pool, a collaborative reasoning engine, a verification committee module, a synthesizer module, and an interpretable output interface. The user input interface is used to receive question requests that contain text and / or images; The scheduler module integrates a lightweight semantic analysis model to perform multimodal semantic parsing on the question request, extract keywords, graphic information and implicit assumptions, and identify the task type and required capabilities. Then, based on the preset expert capability index table, it matches the required combination of expert agents from the expert pool and generates an expert agent invocation priority sequence. The expert pool contains multiple heterogeneous expert agents, specifically including image understanding expert agents, logic deduction expert agents, common sense verification expert agents, code execution expert agents, and several domain expert agents with the ability to process knowledge in a specified domain. Image understanding expert agents are used to transform image information into a machine-readable structured description language when a problem contains image information. The logical deduction expert agent is used to perform formal logical reasoning, deduction and induction analysis based on the information of the problem, to ensure that the reasoning process conforms to logical equivalence or implication relationship, and serves as an important basis for the verification committee module to compare whether there is a logical conflict in the reasoning conclusions of the expert agent. The capabilities of the code execution expert agent include automatically generating executable code scripts based on reasoning needs and running them in a secure sandbox environment to obtain accurate numerical results or structured data; Domain expert agents are used to handle problems within a relevant domain. They have the ability to model and analyze problems in the relevant domain, and can call upon relevant domain knowledge to verify the feasibility of the reasoning process, ensuring that the generated reasoning chain does not violate the constraints of relevant domain knowledge. The collaborative reasoning engine supports parallel and pipelined task scheduling mechanisms, which are used to schedule the corresponding expert agents for parallel and / or pipelined collaborative reasoning based on the priority sequence of expert agent invocation. In pipelined collaborative reasoning, the reasoning process follows a causal dependency chain. The output of the preceding expert agent constitutes the reasoning premise of the subsequent expert agent. The output of the expert agent generates structured intermediate results through code execution and is transmitted to the subsequent expert agent through a unified intermediate representation protocol. The structured intermediate results are timestamped, which facilitates dependency management by the collaborative reasoning engine. When the current task has multiple parallel reasoning sequences, the expert agent that outputs the reasoning conclusion submits its reasoning conclusion to the verification committee module. The verification committee module is used to collect the reasoning conclusions submitted by each expert agent. It uses a consensus model or a multi-expert voting mechanism to verify the consistency of the reasoning conclusions submitted by each expert agent. When logical conflicts or abnormal conclusions are detected, the review process is triggered and common sense verification expert agents are dynamically introduced to participate in decision correction. The synthesizer module is equipped with a confidence-weighted fusion algorithm, which is used to weight and integrate the reasoning conclusions of each expert agent that have passed the verification committee module, based on the historical accuracy of each expert agent and the semantic integrity of its current output, and generate a streaming output answer in natural language form. The interpretable output interface is used to present the final answer to the user, while providing the identity of the expert agents involved in the reasoning, a summary of the reasoning path at each stage, and an explanation of the conclusion-reaching mechanism.

[0006] Furthermore, the verification committee module specifically includes a conflict detection unit, an arbitration triggering unit, and a self-repair unit; The conflict detection unit is used to determine whether the reasoning conclusions output by each expert agent satisfy logical equivalence or implication relations; The arbitration triggering unit is used to automatically call the common sense verification expert agent or external knowledge base for third-party verification when the difference in confidence of the reasoning conclusions output by each expert agent exceeds the threshold or when contradictory conclusions occur. The self-repair unit is used to backtrack the erroneous reasoning node of the expert agent with reasoning errors based on the arbitration result output by the arbitration trigger unit, and drive the expert agent to recalculate and obtain a new reasoning conclusion.

[0007] Furthermore, the synthesizer module maintains a dynamically updated capability score table for each expert agent. The capability score table is used to record the accuracy, response speed and user feedback of the corresponding expert agent in historical tasks. When performing fusion using the confidence-weighted fusion algorithm, the reasoning conclusions of expert agents with high capability scores are given higher weight coefficients, while low-weight but novel reasoning conclusions are retained as supplementary explanations.

[0008] Furthermore, the expert agents in the expert pool are deployed using a modular microservice architecture, supporting hot-swappable expansion. Newly added expert agents can be automatically connected to the system by registering capability tags and inference interfaces.

[0009] Furthermore, the system also includes a training management module, which is used for: Record the deviation between the actual output of each expert agent and the final output answer in each task; Regularly perform incremental training or parameter fine-tuning on each expert agent to achieve individual capability evolution; By leveraging a federated learning framework, joint optimization of expert agents across devices can be achieved while protecting data privacy.

[0010] The present invention also provides a multi-agent cooperative reasoning method utilizing the above-described system, comprising the following steps: S1. Receive user-submitted question requests through the user input interface; S2. The scheduler model uses a lightweight semantic analysis model to perform multimodal semantic parsing on the question request, extract keywords, graphic information and implicit assumptions, and identify the task type and required capabilities. Then, according to the preset expert capability index table, it matches the required combination of expert agents from the expert pool and generates an expert agent invocation priority sequence. S3. Start the collaborative reasoning process: The collaborative reasoning engine drives the activated expert agents to perform parallel and / or pipelined collaborative reasoning operations based on the expert agent combination and expert agent invocation priority sequence selected in step S2. S4. When the current task has multiple parallel reasoning sequences, the expert agent that outputs the reasoning conclusion submits its own reasoning conclusion to the verification committee module. The verification committee module uses a consensus model or a multi-expert voting mechanism to verify the consistency of the reasoning conclusions submitted by each expert agent. When a logical conflict or abnormal conclusion is detected, the review process is triggered and common sense verification expert agents are dynamically introduced to participate in decision correction. S5. After verification by the verification committee module, the synthesizer module assigns weights to the reasoning conclusions output by each expert agent based on their historical performance, and fuses the reasoning conclusions of each expert agent through a confidence-weighted fusion algorithm to generate a coherent, readable answer with an explanation path. S6. The final answer is returned in a streaming manner through an interpretable output interface. Users can choose to display a reasoning tracing view to show the identity of the expert agents involved in the reasoning, a summary of the reasoning path at each stage, and an explanation of the conclusion achievement mechanism.

[0011] Furthermore, in step S4, if there is an irreconcilable logical contradiction in the reasoning conclusions of two or more expert agents, the verification committee module initiates a deep tracing mechanism to re-evaluate whether the initial assumptions in the question request are valid, and returns the ambiguous information to the user requesting clarification.

[0012] The beneficial effects of this invention are as follows: First, by introducing a scheduler module and a modular expert pool, this invention achieves precise decomposition of complex tasks and dynamic invocation of specialized intelligent agents, enhancing its adaptability to multimodal and cross-domain problems. Second, by employing a collaborative reasoning mechanism that combines parallelism and pipelined approaches, it ensures processing efficiency while supporting the structured construction of causal chains. Third, by using a verification committee module for cross-validation and conflict resolution of reasoning conclusions, and by replacing the traditional self-reflection mode with a multi-expert consensus mechanism, it significantly enhances the system's fault resistance and reasoning robustness. Finally, through a confidence-weighted synthesis and streaming output mechanism, it generates natural, fluent, and interpretable answers with reasoning tracing, significantly improving user trust and interactive experience.

[0013] Overall, this invention breaks through the limitations of "monolithic enhancement" and constructs an open, collaborative, and transparent next-generation intelligent question-answering architecture, which is suitable for educational scenarios. Detailed Implementation

[0014] The present invention will be further described below. It should be noted that this embodiment is based on the present technical solution and provides detailed implementation methods and specific operation processes, but the protection scope of the present invention is not limited to this embodiment.

[0015] Example 1 This embodiment provides an intelligent question-answering system based on multi-agent collaborative reasoning, including a user input interface, a scheduler module, an expert pool, a collaborative reasoning engine, a verification committee module, a synthesizer module, and an interpretable output interface; The user input interface is used to receive question requests that contain text and / or images; The scheduler module integrates a lightweight semantic analysis model to perform multimodal semantic parsing on the question request, extract keywords, graphic information and implicit assumptions, and identify the task type and required capabilities. Then, based on the preset expert capability index table, it matches the required combination of expert agents from the expert pool and generates an expert agent invocation priority sequence. The expert pool contains multiple heterogeneous expert agents, specifically including image understanding expert agents, logic deduction expert agents, common sense verification expert agents, code execution expert agents, and several domain expert agents with the ability to process knowledge in a specified domain (such as physical law expert agents).

[0016] Image understanding expert agents are used to transform various image information, such as circuit diagrams, topology diagrams, or engineering drawings, into machine-readable structured description languages ​​when a problem contains image information. The logical deduction expert agent is used to perform formal logical reasoning, deduction and induction analysis based on the information of the problem, to ensure that the reasoning process conforms to logical equivalence or implication relationship, and serves as an important basis for the verification committee module to compare whether there is a logical conflict in the reasoning conclusions of the expert agent. The capabilities of the code execution expert agent include automatically generating executable code scripts based on reasoning needs and running them in a secure sandbox environment to obtain accurate numerical results or structured data; Domain expert agents are used to handle problems within a relevant domain. They have the ability to model and analyze problems in the relevant domain, and can call upon relevant domain knowledge to verify the feasibility of the reasoning process, ensuring that the generated reasoning chain does not violate the constraints of relevant domain knowledge. The collaborative reasoning engine supports parallel and pipelined task scheduling mechanisms, which are used to schedule the corresponding expert agents for parallel and / or pipelined collaborative reasoning based on the priority sequence of expert agent invocation. In pipelined collaborative reasoning, the reasoning process follows a causal dependency chain. The output of the preceding expert agent constitutes the reasoning premise of the subsequent expert agent. The output of the expert agent generates structured intermediate results through code execution and is transmitted to the subsequent expert agent through a unified intermediate representation protocol. The structured intermediate results are timestamped, which facilitates dependency management by the collaborative reasoning engine. When the current task has multiple parallel reasoning sequences, the expert agent that outputs the reasoning conclusion submits its reasoning conclusion to the verification committee module. The verification committee module is used to collect the reasoning conclusions submitted by each expert agent. It uses a consensus model or a multi-expert voting mechanism to verify the consistency of the reasoning conclusions submitted by each expert agent. When logical conflicts or abnormal conclusions are detected, the review process is triggered and common sense verification expert agents are dynamically introduced to participate in decision correction. The synthesizer module is equipped with a confidence-weighted fusion algorithm, which is used to weight and integrate the reasoning conclusions of each expert agent that have passed the verification committee module, based on the historical accuracy of each expert agent and the semantic integrity of its current output, and generate a streaming output answer in natural language form. The interpretable output interface is used to present the final answer to the user, while providing the identity of the expert agents involved in the reasoning, a summary of the reasoning path at each stage, and an explanation of the conclusion-reaching mechanism.

[0017] In this embodiment, the verification committee module specifically includes a conflict detection unit, an arbitration triggering unit, and a self-repair unit; The conflict detection unit is used to determine whether the reasoning conclusions output by each expert agent satisfy logical equivalence or implication relations; The arbitration triggering unit is used to automatically call the common sense verification expert agent or external knowledge base for third-party verification when the difference in confidence of the reasoning conclusions output by each expert agent exceeds the threshold or when contradictory conclusions occur. The self-repair unit is used to backtrack the erroneous reasoning node of the expert agent with reasoning errors based on the arbitration result output by the arbitration trigger unit, and drive the expert agent to recalculate and obtain a new reasoning conclusion.

[0018] In this embodiment, the synthesizer module maintains a dynamically updated capability score table for each expert agent. The capability score table is used to record the accuracy, response speed and user feedback of the corresponding expert agent in historical tasks. When performing fusion using the confidence-weighted fusion algorithm, the reasoning conclusions of expert agents with high capability scores are given higher weight coefficients, while low-weight but novel reasoning conclusions are retained as supplementary explanations.

[0019] In this embodiment, the expert agents in the expert pool are deployed using a modular microservice architecture, which supports hot-swappable expansion. New expert agents can be automatically connected to the system by registering capability tags and inference interfaces.

[0020] In this embodiment, the system further includes a training management module, which is used for: Record the deviation between the actual output of each expert agent and the final output answer in each task; Regularly perform incremental training or parameter fine-tuning on each expert agent to achieve individual capability evolution; By leveraging a federated learning framework, joint optimization of expert agents across devices can be achieved while protecting data privacy.

[0021] Example 2 This embodiment provides a multi-agent cooperative reasoning method using the system described in Embodiment 1, including the following steps: S1. Receive user-submitted question requests through the user input interface; S2. The scheduler model uses a lightweight semantic analysis model to perform multimodal semantic parsing on the question request, extract keywords, graphic information and implicit assumptions, and identify the task type and required capabilities. Then, according to the preset expert capability index table, it matches the required combination of expert agents from the expert pool and generates an expert agent invocation priority sequence. S3. Start the collaborative reasoning process: The collaborative reasoning engine drives the activated expert agents to perform parallel and / or pipelined collaborative reasoning operations based on the expert agent combination and expert agent invocation priority sequence selected in step S2. S4. In the parallel reasoning process, the expert agent that outputs the reasoning conclusion submits its reasoning conclusion to the verification committee module. The verification committee module uses a consensus model or a multi-expert voting mechanism to verify the consistency of the reasoning conclusions submitted by each expert agent. When a logical conflict or abnormal conclusion is detected, the review process is triggered and common sense verification expert agents are dynamically introduced to participate in decision correction. S5. After verification by the verification committee module, the synthesizer module assigns weights to the reasoning conclusions output by each expert agent based on their historical performance, and fuses the reasoning conclusions of each expert agent through a confidence-weighted fusion algorithm to generate a coherent, readable answer with an explanation path. S6. The final answer is returned in a streaming manner through the interpretable output interface. Users can choose to display the reasoning tracing view to show the identity of the expert agents involved in the reasoning, the summary of the reasoning path at each stage, and the explanation of the conclusion achievement mechanism.

[0022] In this embodiment, in step S4, if there is an irreconcilable logical contradiction in the reasoning conclusions of two or more expert agents, the verification committee module initiates a deep tracing mechanism to re-evaluate whether the initial assumptions in the question request are valid, and returns the ambiguous information to the user requesting clarification.

[0023] Example 3 This embodiment provides an application example of Embodiments 1 and 2.

[0024] A middle school student submitted the following complex question through a smart learning assistant: "Based on the circuit diagram in the attached image, calculate how the current flowing through the bulb L changes when the resistance R1 increases. Please explain step by step." Complex question requests involve image input (circuit diagrams) and complex question-answering tasks requiring a combination of physics knowledge and logical reasoning. Traditional single AI models often exhibit comprehension biases or reasoning jumps when handling such multimodal complex problems. The system described in Embodiment 1 processes these complex question requests in the following specific process: Step 1: Task Analysis and Expert Agent Scheduling After a user submits a complex question request, the system receives the request, which includes both text and images, through the user input interface. The scheduler module activates a lightweight semantic analysis model to quickly identify that the question falls under the task category of "Junior High School Physics - Electricity - Dynamic Circuit Analysis," and determines that three core capabilities are required: ① Image understanding (interpreting circuit structure), ② Application of physical laws (Ohm's law and series / parallel rules), and ③ Causal logic deduction ("if-then" reasoning). The system then activates three corresponding expert agents from a pre-set expert pool: an image understanding expert agent, a physical law expert agent, and a logic deduction expert agent.

[0025] Physics expert agents are used to handle problems within the realm of physics. They possess the ability to model and analyze physical phenomena and can invoke classical physical laws (such as Newtonian mechanics, the laws of thermodynamics, and the principles of electromagnetism) to verify the physical feasibility of the reasoning process. When a physics expert agent is invoked, it receives a structured description language provided by a preceding image understanding expert, and calculates the causal relationships between physical quantities based on this language, ensuring that the generated reasoning chain does not violate fundamental physical constraints.

[0026] Step 2: Parallel and Pipeline Collaborative Reasoning The collaborative reasoning engine activates the workflow. First, the image understanding expert agent performs visual analysis on the uploaded circuit diagram, identifying the connection relationships between the power supply, switch, resistor R1, and bulb L, and outputs a structured description: "This is a series circuit, and the current path is: positive terminal → R1 → L → negative terminal." This result is simultaneously transmitted to the physics law expert agent and the logic deduction expert agent.

[0027] Next, the physics law expert agent, based on topological information, calls upon the knowledge base to extract the key formula: "Total current in a series circuit I = V / (R1 + R..." l The study concluded that "when R1 increases, the denominator increases, therefore the total current decreases." Meanwhile, logical reasoning experts construct causal chains: Premise 1: Increasing R1 increases the total resistance (based on the characteristics of a series circuit). Premise 2: With constant voltage, current is inversely proportional to total resistance (based on Ohm's Law). Conclusion: As the total current decreases, the current flowing through the bulb L also decreases (because the current is the same everywhere in a series circuit). Step 3: Intermediate result verification and conflict resolution The reasoning conclusions of the physics law expert agent and the logic deduction expert agent are submitted to the verification committee module. The verification committee module compares the two conclusions and finds them consistent and logically self-consistent, detecting no contradictions. If it is assumed that the logic deduction expert agent incorrectly believes that "a decrease in light bulb brightness means an increase in voltage," leading to the counterintuitive conclusion that "the current may increase," the verification committee will mark this anomaly and automatically introduce the common sense verification expert agent for arbitration. After consulting the basic physics common sense base, the common sense verification expert agent points out that "in a series circuit, the voltage distribution of components changes with resistance, but the current remains the same," thus correcting the erroneous reasoning path. The verification committee module then drives the logic deduction expert agent to recalculate, ensuring the final conclusion is correct.

[0028] Step 4: Confidence-weighted and interpretable streaming synthesis output The synthesizer module collects the outputs of the image understanding expert agent, the physics law expert agent, and the logic deduction expert agent, and assigns weights to their reasoning conclusions based on the historical accuracy of the physics law expert agent and the logic deduction expert agent: the physics law expert agent (95% historical accuracy) has the highest weight, followed by the logic deduction expert agent (90%). The synthesizer module generates natural language answers and begins streaming output through an interpretable output interface. "Based on the circuit diagram you provided, this is a series circuit..." "According to Ohm's law I = V / R, when R1 increases, the total resistance increases..." "Since the current is the same everywhere in a series circuit, the current flowing through the bulb L will also decrease." At the end of the answer, the synthesizer module can also add an interpretability note: "This conclusion was deduced by image understanding experts, physical law experts and logical deduction experts, and has been cross-validated to confirm its accuracy." The entire process is quick, responsive, logically rigorous, and yields reliable results. Users not only receive accurate answers but can also view, when needed, "which expert agents participated in the decision-making" and "how the reasoning was conducted at each step," greatly enhancing the transparency and trust in the learning process.

[0029] For those skilled in the art, various corresponding changes and modifications can be made based on the above technical solutions and concepts, and all such changes and modifications should be included within the protection scope of the claims of this invention.

Claims

1. An intelligent question and answer system based on multi-agent collaborative reasoning, characterized in that, It includes a user input interface, a scheduler module, an expert pool, a collaborative reasoning engine, a validation committee module, a synthesizer module, and an interpretable output interface; The user input interface is used to receive question requests that contain text and / or images; The scheduler module integrates a lightweight semantic analysis model to perform multimodal semantic parsing on the question request, extract keywords, graphic information and implicit assumptions, and identify the task type and required capabilities. Then, based on the preset expert capability index table, it matches the required combination of expert agents from the expert pool and generates an expert agent invocation priority sequence. The expert pool contains multiple heterogeneous expert agents, specifically including image understanding expert agents, logic deduction expert agents, common sense verification expert agents, code execution expert agents, and several domain expert agents with the ability to process knowledge in a specified domain. Image understanding expert agents are used to transform image information into a machine-readable structured description language when a problem contains image information. The logical deduction expert agent is used to perform formal logical reasoning, deduction and induction analysis based on the information of the problem, to ensure that the reasoning process conforms to logical equivalence or implication relationship, and serves as an important basis for the verification committee module to compare whether there is a logical conflict in the reasoning conclusions of the expert agent. The capabilities of the code execution expert agent include automatically generating executable code scripts based on reasoning needs and running them in a secure sandbox environment to obtain accurate numerical results or structured data; Domain expert agents are used to handle problems within a relevant domain. They have the ability to model and analyze problems in the relevant domain, and can call upon relevant domain knowledge to verify the feasibility of the reasoning process, ensuring that the generated reasoning chain does not violate the constraints of relevant domain knowledge. The collaborative reasoning engine supports parallel and pipelined task scheduling mechanisms, which are used to schedule the corresponding expert agents for parallel and / or pipelined collaborative reasoning based on the priority sequence of expert agent invocation. In pipelined collaborative reasoning, the reasoning process follows a causal dependency chain. The output of the preceding expert agent constitutes the reasoning premise of the subsequent expert agent. The output of the expert agent generates structured intermediate results through code execution and is transmitted to the subsequent expert agent through a unified intermediate representation protocol. The structured intermediate results are timestamped, which facilitates dependency management by the collaborative reasoning engine. When the current task has multiple parallel reasoning sequences, the expert agent that outputs the reasoning conclusion submits its reasoning conclusion to the verification committee module. The verification committee module is used to collect the reasoning conclusions submitted by each expert agent. It uses a consensus model or a multi-expert voting mechanism to verify the consistency of the reasoning conclusions submitted by each expert agent. When logical conflicts or abnormal conclusions are detected, the review process is triggered and common sense verification expert agents are dynamically introduced to participate in decision correction. The synthesizer module is equipped with a confidence-weighted fusion algorithm, which is used to weight and integrate the reasoning conclusions of each expert agent that have passed the verification committee module, based on the historical accuracy of each expert agent and the semantic integrity of its current output, and generate a streaming output answer in natural language form. The interpretable output interface is used to present the final answer to the user, while providing the identity of the expert agents involved in the reasoning, a summary of the reasoning path at each stage, and an explanation of the conclusion-reaching mechanism.

2. The system of claim 1, wherein, The verification committee module specifically includes a conflict detection unit, an arbitration triggering unit, and a self-repair unit; The conflict detection unit is used to determine whether the reasoning conclusions output by each expert agent satisfy logical equivalence or implication relations; The arbitration triggering unit is used to automatically call the common sense verification expert agent or external knowledge base for third-party verification when the difference in confidence of the reasoning conclusions output by each expert agent exceeds the threshold or when contradictory conclusions occur. The self-repair unit is used to backtrack the erroneous reasoning node of the expert agent with reasoning errors based on the arbitration result output by the arbitration trigger unit, and drive the expert agent to recalculate and obtain a new reasoning conclusion.

3. The system according to claim 1, characterized in that, The synthesizer module maintains a dynamically updated capability score table for each expert agent. The capability score table is used to record the accuracy, response speed and user feedback of the corresponding expert agent in historical tasks. When performing fusion using the confidence-weighted fusion algorithm, the reasoning conclusions of expert agents with high capability scores are given higher weight coefficients, while low-weight but novel reasoning conclusions are retained as supplementary explanations.

4. The system according to claim 1, characterized in that, The expert agents in the expert pool are deployed using a modular microservice architecture, supporting hot-swappable expansion. New expert agents can be automatically connected to the system by registering capability tags and inference interfaces.

5. The system according to claim 1, characterized in that, The system also includes a training management module, which is used for: Record the deviation between the actual output of each expert agent and the final output answer in each task; Regularly perform incremental training or parameter fine-tuning on each expert agent to achieve individual capability evolution; By leveraging a federated learning framework, joint optimization of expert agents across devices can be achieved while protecting data privacy.

6. A multi-agent cooperative reasoning method using the system described in any one of claims 1-5, characterized in that, Includes the following steps: S1. Receive user-submitted question requests through the user input interface; S2. The scheduler model uses a lightweight semantic analysis model to perform multimodal semantic parsing on the question request, extract keywords, graphic information and implicit assumptions, and identify the task type and required capabilities. Then, according to the preset expert capability index table, it matches the required combination of expert agents from the expert pool and generates an expert agent invocation priority sequence. S3. Start the collaborative reasoning process: The collaborative reasoning engine drives the activated expert agents to perform parallel and / or pipelined collaborative reasoning operations based on the expert agent combination and expert agent invocation priority sequence selected in step S2. S4. When the current task has multiple parallel reasoning sequences, the expert agent that outputs the reasoning conclusion submits its own reasoning conclusion to the verification committee module. The verification committee module uses a consensus model or a multi-expert voting mechanism to verify the consistency of the reasoning conclusions submitted by each expert agent. When a logical conflict or abnormal conclusion is detected, the review process is triggered and common sense verification expert agents are dynamically introduced to participate in decision correction. S5. After verification by the verification committee module, the synthesizer module assigns weights to the reasoning conclusions output by each expert agent based on their historical performance, and fuses the reasoning conclusions of each expert agent through a confidence-weighted fusion algorithm to generate a coherent, readable answer with an explanation path. S6. The final answer is returned in a streaming manner through an interpretable output interface. Users can choose to display a reasoning tracing view to show the identity of the expert agents involved in the reasoning, a summary of the reasoning path at each stage, and an explanation of the conclusion achievement mechanism.

7. The method according to claim 6, characterized in that, In step S4, if there is an irreconcilable logical contradiction in the reasoning conclusions of two or more expert agents, the verification committee module initiates a deep tracing mechanism to re-evaluate whether the initial assumptions in the question request are valid and returns the ambiguous information to the user requesting clarification.