Medical multi-agent oriented multi-objective dynamic reasoning method and system, computer device and storage medium

Abstract

This application discloses a multi-objective dynamic reasoning method, system, computer device, and storage medium for medical multi-agent systems. The method includes: generating a biased and incomplete current simulated representation based on current medical facts and persona information related to the patient using a patient agent; verifying this representation through a verification agent; updating the current multi-objective optimization set based on the progress of objective achievement, the current multi-objective optimization set including multi-level hierarchical optimization dimensions; performing multi-path reasoning based on the current multi-objective optimization set for mixed retrieval of the current simulated representation; generating a response dialogue based on the current reasoning result and updating the objective achievement progress; and ending the dialogue. This application constructs a hierarchical objective architecture to achieve autonomous updating of reasoning objectives, adapting to dynamic interaction scenarios with incomplete input information. Independent multi-path reasoning decouples the reasoning path from the optimization objective. This solves the technical problem of poor reasoning performance in existing medical human-computer dialogue processes.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, to intelligent interaction scenarios of medical human-computer dialogue, and in particular to multi-objective dynamic reasoning methods, systems, computer devices and storage media for medical multi-agent systems. Background Technology

[0002] With the rapid development of natural language processing and large language model technologies, intelligent dialogue reasoning methods are constantly iterating and upgrading, and have been applied in human-computer interaction scenarios across various industries. Medical human-computer dialogue is an important application scenario for intelligent dialogue reasoning methods. User input information is easily influenced by subjective expression, cognitive level, emotional state, and other factors, generally resulting in problems such as missing information, ambiguous expression, and content deviation. The business scenario of medical human-computer dialogue is a typical multi-objective intelligent dialogue scenario under incomplete input information. This type of scenario places clear technical requirements on intelligent dialogue reasoning methods: they must be able to autonomously adjust reasoning logic and execution steps based on dynamically changing interaction information, while simultaneously considering multiple optimization objectives such as information acquisition completeness, dialogue interaction efficiency, response content adaptability, and interaction process stability.

[0003] Currently, a series of research results have been achieved both domestically and internationally regarding multi-objective intelligent dialogue reasoning methods under incomplete input information. These methods are mainly divided into three categories: general-purpose large language model intelligent dialogue reasoning methods, medical large language model-specific intelligent dialogue reasoning methods, and multi-agent collaborative intelligent dialogue reasoning methods. The core of all these methods revolves around the reasoning logic, execution steps, and result generation of intelligent dialogue. Existing reasoning methods all rely on predefined static optimization objective systems. Whether it's the single objective of responsiveness in a general-purpose large language model or the domain language adaptability objective in a medical-specific model, these are all pre-set during the development phase. The objective management logic of existing reasoning methods is rigid and lacks dynamic adaptation. This deficiency leads to an inability to cover the full-dimensional optimization needs of multi-objective intelligent dialogue, making it difficult to adapt to scenarios in medical human-computer dialogue where user information is incomplete and interactive content dynamically changes. This easily results in problems such as incomplete information acquisition or low interaction efficiency during the reasoning process. When facing multi-objective optimization needs, the reasoning logic of different objectives constrains and conflicts with each other, forcing the method to passively compromise among multiple objectives, leading to a decline in overall reasoning performance. Simultaneously, the single-agent, single-path reasoning steps are prone to confirmation bias, further reducing the adaptability of the reasoning results to multi-objective requirements. This means that existing technologies suffer from poor reasoning performance during medical human-computer dialogue. Summary of the Invention

[0004] This application provides a multi-objective dynamic reasoning method, system, computer device, and storage medium for medical multi-agent systems, which can solve the technical problem of poor reasoning performance in medical human-computer dialogue.

[0005] In a first aspect, embodiments of this application provide a multi-objective dynamic reasoning method for medical multi-agent systems. This method is applied to a multi-objective dynamic reasoning system for medical multi-agent systems, which includes patient agents, examination agents, and doctor agents as subsystems, comprising:

[0006] The patient-related current input is processed by the patient-intelligent agent to generate a current simulated statement in the current patient role dialogue scenario. The current simulated statement is an incomplete statement with bias. The current input includes the patient's current medical facts and the patient's persona information.

[0007] The current simulation statement is verified by the verification agent;

[0008] If the verification of the testing agent passes, the current multi-objective optimization set is updated by the doctor agent subsystem based on the progress of the goal achievement. The current multi-objective optimization set includes multi-level hierarchical optimization dimensions.

[0009] The doctor intelligent agent subsystem performs multi-path reasoning based on the current multi-objective optimization set to perform hybrid retrieval on the current simulated expression, and obtains the current reasoning result.

[0010] Based on the current reasoning result, generate a response dialogue and update the progress of the goal achievement;

[0011] In response to the end-of-dialogue command, the dialogue ends.

[0012] In some embodiments, the doctor intelligent agent subsystem includes an assistant intelligent agent, and if the verification intelligent agent passes the verification, the doctor intelligent agent subsystem updates the current multi-objective optimization set based on the goal achievement progress, including:

[0013] If the verification agent passes the test, then the initial multi-objective optimization set is determined;

[0014] The assistant agent detects the progress of each objective in the current multi-objective optimization set.

[0015] The assistant agent identifies obstacles in the progress of each target to obtain a target blockage set.

[0016] Based on the target blocking set, the current multi-objective optimization set, and external input knowledge, candidate targets are generated at multiple levels according to the multi-level hierarchical optimization dimension.

[0017] The candidate objectives and the initial multi-objective optimization set are merged to obtain a new current multi-objective optimization set;

[0018] The current multi-objective optimization set is updated based on the new current multi-objective optimization set.

[0019] In some embodiments, the multi-path reasoning performed by the doctor intelligent agent subsystem on the current simulated representation based on the current multi-objective optimization set to obtain the current reasoning result includes:

[0020] The current multi-objective optimization set is prioritized, and multiple core optimization objectives are selected from it based on the priority ranking.

[0021] For each of the core optimization objectives, a hybrid knowledge retrieval is performed in a preset experience base to obtain matching knowledge items;

[0022] By combining each core optimization objective with the corresponding knowledge matching item, multi-path reasoning is performed on the current simulated expression to obtain multiple reasoning chains in sequence;

[0023] The current inference result is obtained by combining multiple inference chains.

[0024] In some embodiments, the doctor intelligent agent subsystem further includes a doctor intelligent agent, wherein generating a response dialogue and updating the progress of the goal achievement based on the current inference result includes:

[0025] The assistant agent retrieves historical cases from the experience base and filters them to obtain target historical cases.

[0026] The target historical cases, historical dialogues, and the current simulated statements are processed into reference experience;

[0027] The doctor's intelligent agent integrates the current reasoning result with the reference experience to generate a response dialogue to the current simulated expression in the current patient role dialogue scenario;

[0028] The assistant agent evaluates the progress of achieving the core optimization goal based on all historical dialogues and reasoning processes, and obtains the progress of achieving the core optimization goal.

[0029] Based on the core optimization objective, the progress score is updated, and an explanatory text is attached.

[0030] In some embodiments, the response to the dialogue end instruction, after ending the dialogue, includes:

[0031] The progress of achieving each target is compared with a preset progress threshold, and the number of targets that have reached the progress threshold is counted.

[0032] If the target number of the current case reaches the preset case selection criteria, then the current case will be designated as a classic case.

[0033] The classic cases are then structured.

[0034] Structured classic cases are stored in the experience base for later retrieval.

[0035] In some embodiments, after the verification agent verifies the current simulation representation, it includes:

[0036] If the verification of the testing agent fails, a correction instruction is generated and sent to the patient agent.

[0037] Identify all historical dialogues;

[0038] The patient's intelligent agent processes the correction instructions, the historical dialogue, and the current input to obtain a secondary simulated expression in the current patient role's dialogue scenario;

[0039] The current simulation statement is updated based on the second-order simulation statement.

[0040] In some embodiments, the verification of the current simulation representation by the verification agent includes:

[0041] Perform format consistency verification on the current simulation representation;

[0042] If the format consistency verification passes, then based on the persona information, the information fit of the current simulated expression is verified.

[0043] If the information fit verification passes, then the authenticity of the current simulated statement is verified based on the current medical facts.

[0044] If the authenticity verification passes, then all verifications are considered successful.

[0045] The verification pass status is transmitted to the doctor's intelligent agent subsystem for further reasoning.

[0046] Secondly, embodiments of this application also provide a multi-objective dynamic reasoning system for medical multi-agent systems, comprising:

[0047] The patient agent is used to process the current input related to the patient and generate a current simulated statement in the current patient role dialogue scenario. The current simulated statement is an incomplete statement with bias. The current input includes the patient's current medical facts and the patient's persona information.

[0048] A verification agent is used to verify the current simulated representation;

[0049] The doctor intelligent agent subsystem is used to update the current multi-objective optimization set based on the progress of target achievement if the verification of the intelligent agent passes the test. The current multi-objective optimization set includes multi-level hierarchical optimization dimensions.

[0050] The doctor intelligent body subsystem is also used for multi-path reasoning to perform hybrid retrieval of the current simulated expression based on the current multi-objective optimization set, so as to obtain the current reasoning result;

[0051] The doctor's intelligent body subsystem is also used to generate a response dialogue based on the current reasoning result and update the progress of the goal achievement;

[0052] The doctor's intelligent body subsystem is also used to end conversations in response to end-of-conversation commands.

[0053] Thirdly, embodiments of this application also provide a computer device for multi-objective dynamic reasoning for medical multi-agent systems, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0054] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a computer program, the computer program including program instructions that, when executed by a processor, can implement the above-described method.

[0055] This application provides a multi-objective dynamic reasoning method, system, computer device, and storage medium for medical multi-agent systems. The method includes: processing current input related to the patient through a patient agent to generate a current simulated representation in the current patient role dialogue scenario, wherein the current simulated representation is an incomplete representation with biases, and the current input includes the patient's current medical facts and the patient's persona information; verifying the current simulated representation through a verification agent; if the verification passes, updating the current multi-objective optimization set based on the goal achievement progress through a doctor agent subsystem, wherein the current multi-objective optimization set includes multi-level hierarchical optimization dimensions; performing multi-path reasoning based on the current multi-objective optimization set using the doctor agent subsystem to perform mixed retrieval on the current simulated representation to obtain a current reasoning result; generating a response dialogue based on the current reasoning result and updating the goal achievement progress; and ending the dialogue in response to a dialogue end command. This application constructs a hierarchical goal architecture to achieve autonomous updating of reasoning goals, adapting to dynamic interaction scenarios with incomplete input information. Independent multi-path reasoning decouples the reasoning path from the optimization goal. This solves the technical problem of poor reasoning performance in medical human-computer dialogue processes in existing technologies. Attached Figure Description

[0056] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 A flowchart illustrating the multi-objective dynamic reasoning method for medical multi-agent systems provided in this application embodiment;

[0058] Figure 2 A schematic block diagram of a multi-objective dynamic reasoning system for medical multi-agent systems provided in this application embodiment;

[0059] Figure 3 A schematic block diagram of a computer device provided in an embodiment of this application. Detailed Implementation

[0060] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0061] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the stated features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0062] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the application. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0063] It should also be further understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0064] With the rapid development of natural language processing and large language model technologies, intelligent dialogue reasoning methods are constantly iterating and upgrading, and have been applied in human-computer interaction scenarios across various industries. Medical human-computer dialogue is an important application scenario for intelligent dialogue reasoning methods. User input information is easily influenced by subjective expression, cognitive level, emotional state, and other factors, generally resulting in problems such as missing information, ambiguous expression, and content deviation. The business scenario of medical human-computer dialogue is a typical multi-objective intelligent dialogue scenario under incomplete input information. This type of scenario places clear technical requirements on intelligent dialogue reasoning methods: they must be able to autonomously adjust reasoning logic and execution steps based on dynamically changing interaction information, while simultaneously considering multiple optimization objectives such as information acquisition completeness, dialogue interaction efficiency, response content adaptability, and interaction process stability.

[0065] Currently, a series of research results have been achieved both domestically and internationally regarding multi-objective intelligent dialogue reasoning methods under incomplete input information. These methods can be mainly divided into three categories: general-purpose large language model intelligent dialogue reasoning methods, medical large language model-specific intelligent dialogue reasoning methods, and multi-agent collaborative intelligent dialogue reasoning methods. The core of each method revolves around the reasoning logic, execution steps, and result generation of intelligent dialogue.

[0066] The first category is intelligent dialogue reasoning methods based on general large language models;

[0067] The intelligent dialogue reasoning methods employed by general-purpose large language models, such as GPT-5, Gemini-3-Flash, and Claude-Sonnet 4.5, are the foundational methods for current intelligent dialogue reasoning. They involve pre-training the model on large-scale general-purpose corpora to learn general rules of language understanding and semantic reasoning, followed by targeted fine-tuning and optimization of the response generation steps.

[0068] The method's execution steps are as follows: First, features are extracted from the user-input dialogue information. Then, the extracted features are used as the sole basis for inference, and inference is performed through pre-set logic within the model. Finally, the dialogue response content is directly generated. Its optimization objective focuses solely on the linguistic fluency and semantic relevance of the response content; the inference steps are fixed and lack dynamic adjustment capabilities.

[0069] The second category is intelligent dialogue reasoning methods specifically designed for medical large language models;

[0070] The intelligent dialogue reasoning methods for medical scenarios, represented by HuatuoGPT[1], MedPaLM 2[2], and DoctorAgent-RL[3], are technical solutions that are adapted and optimized for medical human-computer dialogue scenarios based on general large language model reasoning methods.

[0071] Among them, HuatuoGPT adds a medical knowledge retrieval execution step to the general reasoning method, integrating external medical knowledge into the reasoning basis to reduce information bias in response generation; DoctorAgent-RL introduces multi-agent collaborative reasoning logic, optimizing the reasoning execution steps of each agent through reinforcement learning to improve adaptability to medical scenario interaction language. The core optimization direction of this type of method is to enhance the understanding and response adaptability of the reasoning method to medical domain corpora, but its reasoning logic and execution steps still revolve around the static goals predefined in the development phase.

[0072] The third category is multi-agent collaborative intelligent dialogue reasoning methods;

[0073] The multi-agent collaborative intelligent dialogue reasoning method represented by MedAgents[4] and MAI-DxO[5] is the existing technology that is closest to the technical essence of this invention. Its core implementation logic is to decompose the reasoning task into sub-tasks, and different agents perform sub-task reasoning respectively. It also designs the execution steps of information interaction between agents. Some methods also introduce the steps of domain knowledge retrieval enhancement, in an attempt to solve the limitations of single agent single-path reasoning.

[0074] The execution steps of this method are as follows: First, the multi-turn dialogue is defined differently. Then, according to the definition, the dialogue task is broken down into sub-tasks such as semantic understanding and departmental triage. These sub-tasks are then assigned to corresponding professional intelligent agents, each independently executing the sub-task reasoning. Finally, the reasoning results of the sub-tasks are integrated through preset information interaction steps to generate the final dialogue response. This method has the core technical features of multi-agent collaboration and retrieval enhancement, but it lacks the relevant logic and execution steps for dynamic goal management. The reasoning process always revolves around a predefined static goal.

[0075] In summary, existing technologies have core flaws and technical pain points.

[0076] The three existing technologies mentioned above all revolve around the reasoning logic, execution steps, and result generation of intelligent dialogue. They can be applied to complex scenarios with incomplete input information, such as medical human-computer dialogue. However, from the perspective of the implementation logic and execution step design of the intelligent dialogue reasoning method itself, they all have certain defects and cannot meet the needs of practical applications. The specific defects and their causes are as follows:

[0077] First, the logic of goal management is rigid;

[0078] Existing inference methods all rely on predefined static optimization objective systems. Whether it's the single objective of "response fluency" in general-purpose large language models or the objective of "domain language adaptability" in medical-specific models, these are all pre-set during the development phase. This deficiency prevents these methods from covering the full-dimensional optimization needs of multi-objective intelligent dialogue, making it difficult to adapt to scenarios in medical human-computer dialogue where user information is incomplete and interaction content is dynamically changing. This can easily lead to problems such as "incomplete information acquisition" or "low interaction efficiency" during the inference process.

[0079] Second, the reasoning path is strongly coupled with the optimization objective, and there are no effective execution steps for multi-objective reasoning.

[0080] Existing inference methods all adopt a design pattern of single path corresponding to a single goal or multiple paths surrounding a static goal. When faced with multi-goal optimization requirements, the inference logic of different goals is mutually constrained and conflicting, and the method can only passively compromise among multiple goals, resulting in a decrease in overall inference performance. At the same time, the inference steps of a single agent and a single path are prone to confirmation bias, which further reduces the adaptability of the inference results to multi-goal requirements.

[0081] This means that existing technologies suffer from poor reasoning performance during medical human-computer dialogue.

[0082] In multi-objective intelligent dialogue scenarios with incomplete input information, such as medical human-computer dialogue, the adaptability, reasoning efficiency, and response adaptability are all difficult to meet the needs of practical applications. There is an urgent need for a new dynamic reasoning method for multi-objective intelligent dialogue with incomplete input information to solve the above-mentioned technical problems.

[0083] To address the aforementioned issues, embodiments of this application provide a multi-objective dynamic reasoning method, system, computer device, and storage medium for medical multi-agent systems.

[0084] Unlike existing technologies that pre-set targets during the development phase, this application constructs a hierarchical target architecture and dynamic generation logic, and designs corresponding execution steps. While ensuring target diversity, it enables the autonomous generation, filtering, and priority ranking of inference targets, adapting to dynamic interactive scenarios with incomplete input information.

[0085] Unlike existing inference methods that employ a single path to a single objective or multiple paths around a static objective, this application also designs a decoupling mechanism between inference paths and optimization objectives, constructs independent execution steps for multi-path inference, and formulates reasonable rules and steps for integrating inference results to avoid mutual constraints and conflicts in the multi-objective inference process.

[0086] This application achieves deep integration between the hybrid retrieval system and dynamic target reasoning, designs targeted retrieval and result fusion execution steps based on reasoning targets, and constructs structured historical interaction experience storage and retrieval steps to form a closed-loop optimization capability for intelligent dialogue reasoning methods.

[0087] This application provides a dynamic reasoning method for multi-objective intelligent dialogue with incomplete input information. It aims to achieve dynamic generation, filtering, and priority adjustment of reasoning objectives, decouple reasoning paths from optimization objectives to avoid multi-objective reasoning conflicts, realize deep linkage between the retrieval process and reasoning objectives, form a closed-loop optimization capability for the reasoning method, standardize the goal-oriented execution steps of multi-agent collaborative reasoning, improve the adaptability of the reasoning method to dynamic interaction scenarios with incomplete input information, optimize the reasoning efficiency and response content adaptability of multi-turn intelligent dialogue, and is applicable to multi-objective intelligent dialogue interaction scenarios such as medical human-computer dialogue.

[0088] In some embodiments, this multi-objective dynamic reasoning method for medical multi-agent systems is applied to a multi-objective dynamic reasoning system for medical multi-agent systems. The method is applied to the system, which includes multiple units, namely, a patient agent, a testing agent, and a doctor agent subsystem. Each unit of the multi-objective dynamic reasoning system for medical multi-agent systems can be deployed in a computer device for multi-objective dynamic reasoning for medical multi-agent systems. This computer device can be a terminal or a server, such as a smartphone, tablet, PDA, or laptop computer.

[0089] This technical solution addresses multi-objective reasoning with incomplete input information and is applicable to multi-objective intelligent interaction scenarios such as medical human-computer dialogue. Its core functionality utilizes a three-agent collaborative architecture, dual-core technology, and a closed-loop process to achieve dynamic adaptation and efficient reasoning under incomplete input. The method is applied to a multi-objective dynamic reasoning system for medical multi-agent systems. The system comprises multiple units, such as patient agents, laboratory agents, and doctor agents. The functions and interrelationships of each unit are as follows:

[0090] Patient intelligent agent: As an input source, it generates incomplete, vague or biased dialogue statements based on preset medical facts and personality information, providing input data that fits the actual scenario for reasoning.

[0091] Verification agent: As a supervisory component, it verifies the consistency of the format, the fit of the persona, and the authenticity of the medical facts in the statements generated by the patient agent, outputs the verification results, and provides correction instructions to prompt regeneration if the statements are not up to standard.

[0092] Doctor Intelligent Agent Subsystem: The core execution component, consisting of a doctor intelligent agent and an assistant intelligent agent, receives verified statements and combines core technologies to generate a "continue interaction" or "end interaction" response and the final interaction result.

[0093] Figure 1 This is a flowchart illustrating the multi-objective dynamic reasoning method for medical multi-agent systems provided in an embodiment of this application. Figure 1 As shown, the method includes the following steps S110-S160:

[0094] S110. The patient-related current input is processed by the patient intelligent agent to generate a current simulated statement in the current patient role dialogue scenario, wherein the current simulated statement is an incomplete statement with bias, and the current input includes the patient's current medical facts and the patient's persona information.

[0095] The patient agent means simulating a virtual patient with a specific background and personality.

[0096] The current input includes the patient's current medical facts and the patient's personal information. Current medical facts include objective medical data such as the patient's specific disease, disease course, and examination results. The patient's personal information includes age, occupation, personality traits (such as anxiety, stoicism, and paranoia), health literacy, and socioeconomic status.

[0097] Currently, in multi-objective intelligent dialogue scenarios, user input is easily influenced by factors such as subjective expression, cognitive level, and emotional state, resulting in common problems such as missing information, vague expression, and content deviation. In this application embodiment, user input is uniformly categorized into incomplete information. Such scenarios place clear technical requirements on intelligent dialogue reasoning methods: they must be able to autonomously adjust reasoning logic and execution steps based on dynamically changing interaction information, while simultaneously considering multiple optimization objectives such as information acquisition completeness, dialogue interaction efficiency, response content adaptability, and interaction process stability.

[0098] Because the information input by the user is incomplete, the simulated representation generated by the patient agent in this application, based on the current patient role dialogue scenario, should also be a biased and incomplete representation when processing the current input related to the patient. This truly simulates the input volatility and subjectivity in real-world application scenarios. This creates a realistic reasoning premise environment for subsequent multi-objective reasoning, increasing the credibility of the reasoning results.

[0099] S120. The current simulation statement is verified by the verification agent;

[0100] S120 includes S1201-S1205:

[0101] S1201. Perform format consistency verification on the current simulation representation;

[0102] The testing agent performs format consistency verification on the current simulated expression to ensure it conforms to preset text formats, such as length and perspective. The output text length should be moderate, for example, approximately 500-800 words. For instance, the dialogue response length should be controlled to 1-3 sentences to meet the efficiency requirements of multi-turn interactions and avoid information overload. The perspective should be that of the patient.

[0103] S1202. If the format consistency verification passes, then based on the persona information, perform information fit verification on the current simulated expression;

[0104] If the format consistency verification passes, it is necessary to ensure that the expression conforms to the preset persona. Based on the persona information, the information fit of the current simulated expression is verified.

[0105] For example, verify the fit of age and occupation in the character profile. Capture personality traits from the current simulated statements and verify whether the captured personality traits match the character profile. Verify whether the economic status in the statements is consistent with the socioeconomic status in the character profile. If all are consistent, the information fit verification is successful.

[0106] S1203. If the information fit verification passes, then the authenticity of the current simulated statement is verified based on the current medical facts.

[0107] If the information fit verification passes, it ensures that the statement conforms to the preset medical facts and does not contain contradictory or fictitious structured information.

[0108] S1204. If the authenticity verification passes, then all verifications are confirmed to have passed.

[0109] If the authenticity verification passes, then all verifications are considered successful. A feedback status is output after verification, and the statement is transmitted to the doctor's AI agent.

[0110] If any of the above verification steps fails, the verification is considered unsuccessful. If the verification of the current simulation statement by the testing agent fails, it indicates that the current simulation statement does not closely match reality and needs improvement.

[0111] S1205. The verification pass status is transmitted to the doctor's intelligent agent subsystem for further reasoning.

[0112] The verified status is transmitted to the doctor's intelligent body subsystem. The doctor's intelligent body subsystem receives the verified statement and, in conjunction with the core technology, generates a "continue interaction" or "end interaction" response and a final interaction result.

[0113] Following S120, steps A1-A4 are included:

[0114] A1. If the verification of the testing agent fails, a correction instruction is generated and sent to the patient agent.

[0115] If any of the above verification steps fails, it indicates that the verification of the testing agent has failed. The testing agent generates a correction instruction δ, which is fed back to the patient agent to regenerate the current simulation statement in a single round.

[0116] A2. Identify all historical dialogues;

[0117] A3. The patient's intelligent agent processes the correction instruction, the historical dialogue, and the current input to obtain a secondary simulated expression in the current patient role dialogue scenario;

[0118] The patient agent, acting as the source of input information, generates incomplete, ambiguous, or biased statements that conform to the actual scenario based on preset patient information. Its technical implementation logic is as follows: it receives a prompt constructed from preset patient information and historical doctor-patient dialogue, and generates the input text for the t-th round of dialogue through a large language model. Here, δ represents feedback correction information, D_{t-1} represents the historical dialogue data from the previous t-1 rounds, and P_{feat} represents the input feature parameters, which are the current input containing patient information, including the patient's current medical facts and persona information.

[0119] In some embodiments, the patient agent processes δ, D_{t-1}, and P_{feat} to regenerate a secondary simulation representation of the current patient role's dialogue scenario.

[0120] Precise correction is achieved using correction instructions. In the initial simulation, the patient agent may omit or deviate from the character design details. Explicit correction instructions allow for targeted supplementation, correcting only missing items, avoiding a complete rewrite, and saving computational resources. Reasonable parts are retained, and existing expressions consistent with the character design are continued to prevent overcorrection.

[0121] By integrating historical dialogue to maintain scene continuity, the initial scene description has passed format verification, and information fit verification results have identified omissions. These are used as input for secondary processing to ensure narrative coherence. The secondary representation will not contradict the patient's identity, treatment environment, etc., already established in the past. The agent remembers which content has been accepted and which needs modification, avoiding repetitive errors.

[0122] A4. Update the current simulation statement based on the second simulation statement.

[0123] The secondary simulation statement is used as the current simulation statement to update the current simulation statement.

[0124] S130. If the verification of the intelligent agent passes, the current multi-objective optimization set is updated by the doctor intelligent agent subsystem based on the progress of the goal achievement, wherein the current multi-objective optimization set includes multi-level hierarchical optimization dimensions.

[0125] This system supports multi-agent, multi-objective dynamic reasoning in healthcare. It optimizes dimensions through a hierarchical objective architecture, dynamically generating three categories of objectives: progressive, opposing, and supplementary objectives, which are then selected and prioritized. Independent reasoning paths are constructed for each core objective, and these are integrated according to priority, decoupling objectives from reasoning. The progress of each objective is evaluated to obtain its achievement progress. The current multi-objective optimization set is updated based on the achievement progress. If the verification agent passes the validation, the doctor agent subsystem updates the current multi-objective optimization set based on the achievement progress. During updates, real-time semantic deduplication is used to avoid duplication with existing objectives. Redundant objectives are eliminated after generation using semantic similarity calculation, with a semantic similarity threshold of γ=0.8, ensuring the diversity and effectiveness of the objective set.

[0126] The doctor intelligent agent subsystem includes an assistant intelligent agent, and S130 includes S1301-S1305:

[0127] The physician intelligent agent subsystem consists of a physician intelligent agent and an assistant intelligent agent. This subsystem represents a common and efficient multi-agent collaborative architecture. The core purpose of this design is to separate core decision-making from auxiliary execution, allowing each agent to focus on different tasks, thereby improving the overall system's robustness, scalability, and clinical applicability.

[0128] The core responsibility of the physician agent is clinical reasoning and critical decision-making, providing a preliminary diagnosis and formulating a treatment plan based on medical history, physical examination, and test results. This preliminary diagnosis and treatment plan formulation includes medication, surgical information, and follow-up data. The physician agent is also responsible for answering highly complex medical questions and handling emotional and ethical dilemmas in doctor-patient communication.

[0129] The core responsibilities of the assistant agent are information management and process assistance. It extracts structured medical records from dialogues, including information such as chief complaint, present illness, and allergy history. The assistant agent also queries knowledge bases, drug instructions, and clinical guidelines to generate examination requests, prescriptions, and medical record summaries, reminding doctors of overlooked steps such as vaccinations and complication screenings.

[0130] S1301. If the verification agent passes the verification, then the initial multi-objective optimization set is determined.

[0131] The multi-objective optimization set is a hierarchical target architecture for optimization dimensions. As an example, a three-level hierarchical structure of technical objectives is used to define the technical optimization dimensions of intelligent dialogue reasoning, specifically:

[0132] (1) Macro-foundation layer, Level-1 objective: Define the core technical indicators of dialogue reasoning, such as the completeness of information acquisition and the adaptability of response content, with an importance score of 3;

[0133] (2) Strategy execution layer, Level-2 objective: Define the execution rules for reasoning and interaction, such as the coherence of reasoning logic and the accuracy of retrieval matching, with an importance score of 2;

[0134] (3) Scene adaptation layer, Level-3 goal: Define the adaptation direction for the input expression, such as incomplete information completion and clarification of ambiguous expression, with an importance score of 1;

[0135] The target data structure of the initial multi-objective optimization set is unified as G=(name, desc, S{imp}), where name is the target name, desc is the technical execution logic description to ensure the quantifiability and executability of the target, and S{imp} is the corresponding importance score.

[0136] S1302. The assistant agent identifies the blockages in the progress of each target to obtain a target blockage set.

[0137] The assistant agent detects the progress of each objective in the current multi-objective optimization set, identifies blockages in the progress of each objective, and obtains a target blockage set.

[0138] S1303. Based on the target blocking set, the current multi-objective optimization set, and external input knowledge, generate corresponding multi-level candidate targets according to the multi-level hierarchical optimization dimension.

[0139] Based on the target hierarchical structure of the above three-level optimization dimensions, candidate targets are generated based on the target blocking set obtained from blocking identification, the current multi-target optimization set, and external input knowledge. The candidate targets are multiple targets at different levels.

[0140] As an example, candidate objectives can be broadly categorized into three types: progressive objectives, incremental objectives, and supplementary objectives. The logic for generating candidate objectives is as follows:

[0141] Asymptotic Goals: Blocking goals are those where the progress gain is less than a preset threshold Δ=0.1 for t consecutive rounds. When there are blocking goals, progressive goals are generated to overcome the blockage through more feasible progressive objectives;

[0142] Incremental objectives: To enhance comprehensiveness and practicality, reverse objectives that are opposite to existing objectives are generated;

[0143] Supplementary objectives: To bridge functional gaps, generate complementary objectives that supplement the existing objective set and broaden the decision-making perspective.

[0144] In this embodiment, the external input knowledge originates from retrieved authoritative external knowledge. The method enhances the retrieval by blending external input knowledge with historical interaction experience, and uses the retrieval results as the basis for reasoning, providing evidentiary support for the intelligent reasoning process.

[0145] The assistant agent then utilizes a multi-objective dynamic reasoning mechanism for medical multi-agent systems to generate three types of candidate targets. All newly generated targets undergo a dual deduplication process, including real-time name verification and post-generation semantic filtering. After filtering, the final set of candidate targets is obtained.

[0146] S1304. Merge the candidate objectives and the initial multi-objective optimization set to obtain a new current multi-objective optimization set;

[0147] The final candidate target set With the initial multi-objective optimization set Merge to obtain a new current multi-objective optimization set.

[0148] S1305. Update the current multi-objective optimization set based on the new current multi-objective optimization set.

[0149] The current multi-objective optimization set is updated based on the new current multi-objective optimization set to obtain the updated current multi-objective optimization set.

[0150] S140. The doctor intelligent body subsystem performs multi-path reasoning based on the current multi-objective optimization set to perform hybrid retrieval on the current simulated expression to obtain the current reasoning result;

[0151] Hybrid knowledge retrieval refers to the integration of external authoritative knowledge retrieval and historical experience retrieval to provide accurate evidence support for dynamic reasoning. This application continuously updates the experience base and target set through structured storage and closed-loop optimization, improving the adaptability of reasoning. The structured storage constrains the data format, with all interaction data using JSON structured format to ensure compatibility between various agents and modules.

[0152] Multi-objective dynamic reasoning for medical multi-agent refers to optimizing dimensions through hierarchical objective architecture definition technology, dynamically generating three types of candidate objectives—progressive, opposing, and complementary—and sorting them. After constructing independent reasoning paths for each core objective, these objectives are integrated according to priority, thus achieving decoupling between objectives and reasoning.

[0153] This application achieves dynamic target adaptation, autonomously generating, filtering, prioritizing and updating reasoning targets based on changes in input information during the interaction process, and adjusting the reasoning execution steps accordingly.

[0154] By combining hybrid knowledge retrieval and multi-objective dynamic reasoning for medical multi-agent systems, the physician agent subsystem can generate responses that meet the requirements.

[0155] S140 includes S1401-S1404:

[0156] S1401. Prioritize the current multi-objective optimization set and select multiple core optimization objectives from it based on the priority ranking.

[0157] We continue to apply the multi-objective dynamic reasoning mechanism for medical multi-agent systems to prioritize the current set of objectives, and select the top K objectives based on the priority ranking. As a core dimension of reasoning, K=2 has been experimentally verified to be the optimal value for both performance and efficiency.

[0158] S1402. For each of the core optimization objectives, perform hybrid knowledge retrieval in a preset experience base to obtain matching knowledge items;

[0159] Hybrid knowledge retrieval refers to the simultaneous use of multiple retrieval methods, such as keyword search, vector semantic matching, knowledge graph query, and rule reasoning, to cover different knowledge forms.

[0160] Separate searches are performed for each core optimization objective, ensuring that each inference path focuses on optimizing a single objective and avoiding mutual constraints among multiple objectives. The search is precise rather than generalized, as different optimization objectives focus on different knowledge dimensions. Searching together would introduce noise. Separate searches guarantee that each objective obtains the most relevant knowledge matches. Hybrid retrieval combines multi-path recall with fusion ranking, leveraging the advantages of various methods to improve the coverage and accuracy of knowledge matches.

[0161] Hybrid knowledge retrieval yields knowledge matching items, which can be used by the agent to generate decision-making basis, provide dialogue material, and trigger warnings. For example, if a drug interaction rule is matched, a warning can be issued to be cautious about combining medications.

[0162] When dynamic target generation or multi-path reasoning is initiated, the system automatically generates structured queries. Where D represents historical dialogue data. For the current multi-objective optimization set, structured queries are used to ensure that the query and reasoning objectives are highly correlated.

[0163] In some embodiments, relevant literature data can be retrieved by calling authoritative domain knowledge sources such as PubMed through external knowledge interfaces; and core evidence fragments from the literature can be extracted. After sorting by target relevance, the reasoning path associated with the corresponding target provides authoritative support for the reasoning process.

[0164] The pre-defined experience base stores historical experiences, which are encapsulated in round-based units. Each unit contains five fields: dialogue round ID, previous summary, current dialogue content, integrated reasoning result, and target progress assessment. This ensures the semantic continuity and retrieval of experience data, avoiding the context fragmentation problem caused by traditional storage methods.

[0165] S1403. Combining each core optimization objective with the corresponding knowledge matching item, perform multi-path reasoning on the current simulation statement to obtain multiple reasoning chains in sequence;

[0166] for For each high-priority objective, the assistant agent independently performs reasoning. Independent reasoning paths are constructed, and hybrid retrieval enhances the acquisition of evidence supporting objective matching, ensuring that each reasoning path focuses on optimizing a single objective and avoiding mutual constraints among multiple objectives. During this process, hybrid knowledge retrieval is used to retrieve matching terms from the experience base E to assist in objective reasoning, and the resulting K reasoning chains are synthesized into a comprehensive reasoning result.

[0167] In some embodiments, a two-stage retrieval mechanism is employed:

[0168] Retrieval during inference: For each core objective, 2K candidate experiences are first filtered by vector similarity, then further refined by target matching degree, and finally a comprehensive score is calculated. The top K experience fragments are returned and injected into the inference process.

[0169] Retrieval in Response: Generate a query based on the current dialogue history, filter 3K candidate experiences, sort them by comprehensive score, and return the top K experience fragments for response generation.

[0170] In this application, experiments have verified that K=2 is the optimal value for both performance and efficiency.

[0171] S1404. Combine the multiple inference chains to obtain the current inference result.

[0172] The inference results are integrated according to rules, and the integration process uses a large language model to verify logical consistency. The inference chains are integrated in the order of progressive goal → opposing goal → supplementary goal: the inference chains corresponding to progressive goals are merged first, and these chains serve as the core foundation of the inference results. The inference chains corresponding to opposing goals are used to balance different optimization directions. Finally, the inference chains corresponding to supplementary goals are incorporated to expand the functional coverage of the inference results.

[0173] S150. Based on the current reasoning result, generate a response dialogue and update the progress of the goal achievement;

[0174] The assistant agent retrieves high-quality past dialogue examples from E for reference. The doctor then integrates the reasoning results with these experience references based on the historical dialogue data from previous rounds and the current simulated expression to generate a response.

[0175] S150 includes S1501-S1505:

[0176] S1501. The assistant agent retrieves historical cases from the experience base and filters them to obtain target historical cases.

[0177] By leveraging experience bases to reuse successful models, historical cases containing typical symptoms, diagnostic pathways, and communication skills, it is possible to avoid repeating mistakes or overlooking key steps.

[0178] The assistant agent should handle the retrieval task, while the doctor agent should focus on reasoning and decision-making. The retrieval task should be delegated to the assistant, thus achieving a separation of concerns.

[0179] The screening mechanism ensures relevance, but it is not applicable to all historical cases. Similarity matching (such as vector retrieval or tag filtering) should be performed based on the current patient's chief complaint, persona, and medical facts.

[0180] S1502. Process the target historical cases, historical dialogues and the current simulated statements into reference experience;

[0181] A single case may not be complete; only by combining the current dialogue context and historical interactions can actionable reference experiences be formed. The original case may be unstructured and needs to be reconstructed into a format that is easy for the doctor's intelligent agent to understand.

[0182] Historical cases may contain details that are not applicable to the current scenario. The assistant needs to extract and generalize these details, retaining the most relevant experience patterns to remove redundant noise.

[0183] S1503. The doctor's intelligent agent integrates the current reasoning result with the reference experience to generate a response dialogue to the current simulated expression in the current patient role dialogue scenario.

[0184] The doctor's agent, based on medical knowledge, arrives at preliminary conclusions that need to be compared with relevant experiences, such as similar patients refusing hospitalization due to financial constraints, to adjust communication strategies or diagnostic priorities. Directly outputting diagnostic conclusions is often insufficient; persuasive and empathetic dialogue must be generated by incorporating the patient's persona. Reasoning results may exceed the scope of historical experience; in such cases, reasoning takes precedence. If reasoning conflicts with experience, experience takes priority.

[0185] S1504. The assistant agent evaluates the progress of the core optimization goal based on all historical dialogues and reasoning processes to obtain the progress of the core optimization goal.

[0186] Core optimization goals, such as blood sugar control, often require multiple rounds of interaction to achieve. The assistant needs to assess in real time how much progress has been made.

[0187] A single message cannot reflect the overall progress; all historical dialogues and reasoning chains must be considered to determine whether any key steps have been missed.

[0188] Transforming abstract goals into measurable indicators provides a basis for the next steps.

[0189] S1505. Update the progress score based on the core optimization objective and add explanatory text.

[0190] Update the progress towards the goal. At the end of round t, the assistant agent evaluates the progress based on the complete dialogue history and reasoning process. For each objective, progress is tracked, its progress score is updated, and a concise textual explanation is generated.

[0191] Progress scores provide a visual representation of system performance, facilitating human oversight or automatic adjustment of subsequent strategies. For example, when scores are low, further follow-up questions can be asked. Simply providing a score is insufficient to explain the underlying reasons; accompanying written explanations, such as those addressing failure to meet economic assessment goals (since patients often don't mention income details), helps in establishing trust and adjustment. Accumulated progress scores and explanations can feed back into an experience base, training more accurate progress prediction models and supporting long-term learning.

[0192] S160. In response to the dialogue end command, end the dialogue.

[0193] When the conversation reaches the preset number of rounds or the "End Interaction" action is triggered, a conversation end command is issued. The process then proceeds to the subsequent processing stage. In response to the conversation end command, the conversation ends.

[0194] Following S160, steps B1-B4 are included:

[0195] After the dialogue ends, this application supports updating the target set, which is an offline optimization process following the dialogue's conclusion. Through target feasibility assessment in subsequent processing stages, highly feasible targets are included in the initial target set, enabling dynamic updates to the target set and improving the adaptability of the initial targets in subsequent dialogues. By using high-quality experience filtering and structured storage, the historical experience storage pool is continuously expanded, improving the relevance and effectiveness of search results; simultaneously, semantic deduplication of the experience pool is performed periodically to retain high-value experience fragments and optimize search efficiency.

[0196] B1. Compare the progress of each target with a preset progress threshold, and count the number of targets that have reached the progress threshold.

[0197] A case study may involve multiple core optimization goals such as "blood sugar control," "complication prevention," and "improved adherence." Success is considered achieved only when a pre-defined threshold is reached, such as progress ≥ 80%. Statistical analysis of the number of cases achieving these goals measures the overall success rate. These thresholds are objective, pre-defined values, avoiding the arbitrariness of manual evaluation and subjective judgment. This provides a numerical basis for selection.

[0198] Based on the progress record of the objectives, a multi-objective dynamic reasoning mechanism for medical multi-agents is executed to include highly feasible objectives into the initial objective set.

[0199] High-quality experience screening. If the number or proportion of dialogue data for a case that reaches or exceeds the progress threshold meets the specified screening criteria, then the case is designated as a high-quality experience.

[0200] Experience structuring and storage. The system will structure the selected high-quality cases according to a set format and store them in the experience database E for later retrieval.

[0201] B2. If the target number of the current case reaches the preset case selection criteria, then the current case will be designated as a classic case.

[0202] Not all historical cases are worth preserving. Only those cases that excelled in a number of core objectives, such as achieving ≥80% of the total objectives, are marked as classics to ensure a high-value density of experience in the database.

[0203] Low-quality cases can pollute the experience database, leading to irrelevant content being retrieved during subsequent searches. Rigorous screening can improve the efficiency of the experience database and reduce storage and retrieval noise.

[0204] Case selection criteria can be flexibly adjusted according to the field. For example, the requirements are higher for emergency scenarios, while they can be slightly more lenient for chronic disease management. The threshold can be configured.

[0205] B3. The classic cases are structured.

[0206] It supports efficient retrieval, as raw dialogues or free text are difficult for algorithms to match directly. Structured processing allows extraction of fields such as: patient profile (age, occupation, disease), core question list, key dialogue rounds, treatment actions taken, progress of each goal, and summary of success factors.

[0207] Facilitates subsequent hybrid knowledge retrieval: Structured cases can be used simultaneously for various retrieval methods such as tag-based keyword matching, semantic vectorization, and knowledge graphs.

[0208] The purpose of structuring classic cases is to make them readable by both humans and machines, so that they can be understood by the doctor's AI agent and are also easy for humans to review and debug.

[0209] B4. Store structured classic cases in the experience base for later retrieval.

[0210] By storing classic cases, continuous learning is achieved. Each successful diagnosis and treatment can be transformed into experience, allowing the experience database to grow and be updated continuously to adapt to new clinical models or patient groups. When encountering similar patients, classic cases can be directly retrieved as reference experience, avoiding reasoning from scratch and shortening future decision-making time.

[0211] Multiple intelligent agent instances or different departments can share the same experience base, forming an institutional-level collective intelligence.

[0212] The stored classic cases are accompanied by progress scores and explanatory text, allowing you to trace back why a certain strategy is effective and improving interpretability.

[0213] This application provides a multi-objective dynamic reasoning method, system, computer device, and storage medium for medical multi-agent systems. The method includes: processing current input related to the patient through a patient agent to generate a current simulated representation in the current patient role dialogue scenario, wherein the current simulated representation is an incomplete representation with biases, and the current input includes the patient's current medical facts and the patient's persona information; verifying the current simulated representation through a verification agent; if the verification passes, updating the current multi-objective optimization set based on the goal achievement progress through a doctor agent subsystem, wherein the current multi-objective optimization set includes multi-level hierarchical optimization dimensions; performing multi-path reasoning based on the current multi-objective optimization set using the doctor agent subsystem to perform mixed retrieval on the current simulated representation to obtain a current reasoning result; generating a response dialogue based on the current reasoning result and updating the goal achievement progress; and ending the dialogue in response to a dialogue end command. This application constructs a hierarchical goal architecture to achieve autonomous updating of reasoning goals, adapting to dynamic interaction scenarios with incomplete input information. Independent multi-path reasoning decouples the reasoning path from the optimization goal. This solves the technical problem of poor reasoning performance in medical human-computer dialogue processes in existing technologies.

[0214] Figure 2 This is a schematic block diagram of a multi-objective dynamic reasoning system for medical multi-agent systems provided in an embodiment of this application. Figure 2As shown, corresponding to the above-described multi-agent dynamic reasoning method for medical applications, this application also provides a multi-agent dynamic reasoning system 600 for medical applications. This multi-agent dynamic reasoning system 600 includes units for executing the above-described multi-agent dynamic reasoning method for medical applications, and can be configured in terminals such as desktop computers, tablet computers, and laptops. For details, please refer to... Figure 2 The multi-objective dynamic reasoning system 600 for medical multi-agent systems includes a patient agent 601, a testing agent 602, and a doctor agent subsystem 603, wherein:

[0215] The patient agent 601 is used to process the current input related to the patient and generate the current simulated expression in the current patient role dialogue scenario. The current simulated expression is an incomplete expression with bias. The current input includes the patient's current medical facts and the patient's persona information.

[0216] The testing agent 602 is used to verify the current simulated representation;

[0217] The doctor intelligent agent subsystem 603 is used to update the current multi-objective optimization set based on the progress of target achievement if the verification of the testing intelligent agent passes, wherein the current multi-objective optimization set includes multi-level hierarchical optimization dimensions;

[0218] The doctor intelligent body subsystem 603 is also used for multi-path reasoning to perform hybrid retrieval of the current simulated expression based on the current multi-objective optimization set, so as to obtain the current reasoning result;

[0219] The doctor intelligent body subsystem 603 is also used to generate a response dialogue based on the current reasoning result and update the progress of the goal achievement;

[0220] The doctor's intelligent body subsystem 603 is also used to end the dialogue in response to a dialogue end command.

[0221] In some embodiments, the doctor intelligent agent subsystem includes an assistant intelligent agent. If the verification of the verification intelligent agent passes, the doctor intelligent agent subsystem 603 updates the current multi-objective optimization set based on the progress of goal achievement, specifically for:

[0222] If the verification agent passes the test, then the initial multi-objective optimization set is determined;

[0223] The assistant agent detects the progress of each objective in the current multi-objective optimization set.

[0224] The assistant agent identifies obstacles in the progress of each target to obtain a target blockage set.

[0225] Based on the target blocking set, the current multi-objective optimization set, and external input knowledge, candidate targets are generated at multiple levels according to the multi-level hierarchical optimization dimension.

[0226] The candidate objectives and the initial multi-objective optimization set are merged to obtain a new current multi-objective optimization set;

[0227] The current multi-objective optimization set is updated based on the new current multi-objective optimization set.

[0228] In some embodiments, the physician intelligent body subsystem 603 performs multi-path reasoning based on the current multi-objective optimization set to perform hybrid retrieval of the current simulated representation, and obtains the current reasoning result, specifically for:

[0229] The current multi-objective optimization set is prioritized, and multiple core optimization objectives are selected from it based on the priority ranking.

[0230] For each of the core optimization objectives, a hybrid knowledge retrieval is performed in a preset experience base to obtain matching knowledge items;

[0231] By combining each core optimization objective with the corresponding knowledge matching item, multi-path reasoning is performed on the current simulated expression to obtain multiple reasoning chains in sequence;

[0232] The current inference result is obtained by combining multiple inference chains.

[0233] In some embodiments, the doctor intelligent agent subsystem 603 further includes a doctor intelligent agent, which, when performing actions based on the current inference result, generates a response dialogue and updates the progress of the goal achievement, specifically for:

[0234] The assistant agent retrieves historical cases from the experience base and filters them to obtain target historical cases.

[0235] The target historical cases, historical dialogues, and the current simulated statements are processed into reference experience;

[0236] The doctor's intelligent agent integrates the current reasoning result with the reference experience to generate a response dialogue to the current simulated expression in the current patient role dialogue scenario;

[0237] The assistant agent evaluates the progress of achieving the core optimization goal based on all historical dialogues and reasoning processes, and obtains the progress of achieving the core optimization goal.

[0238] Based on the core optimization objective, the progress score is updated, and an explanatory text is attached.

[0239] In some embodiments, the multi-objective dynamic reasoning system 600 for medical multi-agent systems further includes a storage unit after the dialogue ends in response to a dialogue termination instruction. The storage unit is specifically used for:

[0240] The progress of achieving each target is compared with a preset progress threshold, and the number of targets that have reached the progress threshold is counted.

[0241] If the target number of the current case reaches the preset case selection criteria, then the current case will be designated as a classic case.

[0242] The classic cases are then structured.

[0243] Structured classic cases are stored in the experience base for later retrieval.

[0244] In some embodiments, after the current simulation representation is verified by the verification agent, the multi-objective dynamic reasoning system 600 for medical multi-agent systems further includes an update unit, which is specifically used for:

[0245] If the verification of the testing agent fails, a correction instruction is generated and sent to the patient agent.

[0246] Identify all historical dialogues;

[0247] The patient's intelligent agent processes the correction instructions, the historical dialogue, and the current input to obtain a secondary simulated expression in the current patient role's dialogue scenario;

[0248] The current simulation statement is updated based on the second-order simulation statement.

[0249] In some embodiments, the verification agent performs verification of the current simulated representation, specifically for:

[0250] Perform format consistency verification on the current simulation representation;

[0251] If the format consistency verification passes, then based on the persona information, the information fit of the current simulated expression is verified.

[0252] If the information fit verification passes, then the authenticity of the current simulated statement is verified based on the current medical facts.

[0253] If the authenticity verification passes, then all verifications are considered successful.

[0254] The verification pass status is transmitted to the doctor's intelligent agent subsystem for further reasoning.

[0255] In summary, the multi-objective dynamic reasoning system 600 for medical multi-agent systems in this application embodiment processes the current input related to the patient through the patient agent to generate a current simulated statement in the current patient role dialogue scenario. This current simulated statement is an incomplete statement with biases. The current input includes the patient's current medical facts and the patient's persona information. The current simulated statement is verified by the verification agent. If the verification passes, the current multi-objective optimization set is updated by the doctor agent subsystem based on the goal achievement progress. This current multi-objective optimization set includes multi-level hierarchical optimization dimensions. The doctor agent subsystem performs multi-path reasoning based on the current multi-objective optimization set to perform mixed retrieval on the current simulated statement, obtaining the current reasoning result. Based on the current reasoning result, a response dialogue is generated and the goal achievement progress is updated. The dialogue ends in response to a dialogue end command. This application constructs a hierarchical goal architecture to achieve autonomous updating of reasoning goals, adapting to dynamic interaction scenarios with incomplete input information. Independent multi-path reasoning decouples the reasoning path from the optimization goal. This solves the technical problem of poor reasoning performance in medical human-computer dialogue processes in existing technologies.

[0256] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the above-mentioned multi-objective dynamic reasoning system for medical multi-agents and each unit can be referred to the corresponding descriptions in the foregoing method embodiments. For the sake of convenience and brevity, these descriptions will not be repeated here.

[0257] The aforementioned multi-objective dynamic reasoning system for medical multi-agent systems can be implemented as a computer program, which can, for example... Figure 3 It runs on the computer device shown.

[0258] Please see Figure 3 , Figure 3 This is a schematic block diagram of a computer device 700 provided in an embodiment of this application. The computer device 700 can be a terminal or a server. The terminal can be an electronic device with communication functions, such as a smartphone, tablet, laptop, desktop computer, personal digital assistant, or wearable device. The server can be a standalone server or a server cluster composed of multiple servers.

[0259] See Figure 3The computer device 700 includes a processor 702, a memory, and a network interface 705 connected via a system bus 701. The memory may include a non-volatile storage medium 703 and internal memory 704.

[0260] The non-volatile storage medium 703 may store an operating system 7031 and a computer program 7032. The computer program 7032 includes program instructions that, when executed, cause the processor 702 to perform a multi-objective dynamic reasoning method for medical multi-agent systems.

[0261] The processor 702 provides computing and control capabilities to support the operation of the entire computer device 700.

[0262] The internal memory 704 provides an environment for the execution of the computer program 7032 in the non-volatile storage medium 703. When the computer program 7032 is executed by the processor 702, the processor 702 can execute a multi-objective dynamic reasoning method for medical multi-agent systems.

[0263] This network interface 705 is used for network communication with other devices. Those skilled in the art will understand that... Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 700 to which the present application is applied. The specific computer device 700 may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0264] The processor 702 is used to run a computer program 7032 stored in the memory to perform the following steps: The patient agent processes the current input related to the patient to generate a current simulated representation in the current patient role dialogue scenario, wherein the current simulated representation is an incomplete representation with bias, and the current input includes the patient's current medical facts and the patient's persona information; the verification agent verifies the current simulated representation; if the verification agent passes, the doctor agent subsystem updates the current multi-objective optimization set based on the goal achievement progress, wherein the current multi-objective optimization set includes multi-level hierarchical optimization dimensions; the doctor agent subsystem performs multi-path reasoning based on the current multi-objective optimization set to perform mixed retrieval on the current simulated representation to obtain a current reasoning result; based on the current reasoning result, a response dialogue is generated and the goal achievement progress is updated; in response to a dialogue end command, the dialogue ends.

[0265] It should be understood that in the embodiments of this application, the processor 702 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0266] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0267] Therefore, this application also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program, wherein the computer program includes program instructions. When executed by a processor, the program instructions cause the processor to perform the following steps: processing current input related to the patient through the patient agent to generate a current simulated representation in the current patient role dialogue scenario, wherein the current simulated representation is a biased and incomplete representation, and the current input includes the patient's current medical facts and the patient's persona information; verifying the current simulated representation through the verification agent; if the verification by the verification agent passes, updating the current multi-objective optimization set based on the target achievement progress through the doctor agent subsystem, wherein the current multi-objective optimization set includes multi-level hierarchical optimization dimensions; performing multi-path reasoning on the current simulated representation based on the current multi-objective optimization set through the doctor agent subsystem to obtain a current reasoning result; generating a response dialogue based on the current reasoning result and updating the target achievement progress; and ending the dialogue in response to a dialogue end instruction.

[0268] The storage medium can be any computer-readable storage medium that can store program code, such as a USB flash drive, external hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0269] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0270] In the embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For example, the division of each unit is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0271] The steps in the methods of this application embodiment can be adjusted, merged, or deleted according to actual needs. The units in the system of this application embodiment can be merged, divided, or deleted according to actual needs. Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0272] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application.

[0273] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A multi-objective dynamic reasoning method for medical multi-agent systems, characterized in that, The method is applied to a multi-objective dynamic reasoning system for medical multi-agent systems, the system comprising patient agent, examination agent, and doctor agent subsystems, and the method includes: The patient-related current input is processed by the patient-intelligent agent to generate a current simulated statement in the current patient role dialogue scenario. The current simulated statement is an incomplete statement with bias. The current input includes the patient's current medical facts and the patient's persona information. The current simulation statement is verified by the verification agent; If the verification of the testing agent passes, the current multi-objective optimization set is updated by the doctor agent subsystem based on the progress of the goal achievement. The current multi-objective optimization set includes multi-level hierarchical optimization dimensions. The doctor intelligent agent subsystem performs multi-path reasoning based on the current multi-objective optimization set to perform hybrid retrieval on the current simulated expression, and obtains the current reasoning result. Based on the current reasoning result, generate a response dialogue and update the progress of the goal achievement; In response to the end-of-dialogue command, the dialogue ends.

2. The method according to claim 1, characterized in that, The doctor intelligent agent subsystem includes an assistant intelligent agent. If the verification of the verification intelligent agent passes, the doctor intelligent agent subsystem updates the current multi-objective optimization set based on the progress of goal achievement, including: If the verification agent passes the test, then the initial multi-objective optimization set is determined; The assistant agent detects the progress of each objective in the current multi-objective optimization set. The assistant agent identifies obstacles in the progress of each target to obtain a target blockage set. Based on the target blocking set, the current multi-objective optimization set, and external input knowledge, candidate targets are generated at multiple levels according to the multi-level hierarchical optimization dimension. The candidate objectives and the initial multi-objective optimization set are merged to obtain a new current multi-objective optimization set; The current multi-objective optimization set is updated based on the new current multi-objective optimization set.

3. The method according to claim 2, characterized in that, The multi-path reasoning, which involves hybrid retrieval of the current simulated expression based on the current multi-objective optimization set by the doctor's intelligent body subsystem, to obtain the current reasoning result includes: The current multi-objective optimization set is prioritized, and multiple core optimization objectives are selected from it based on the priority ranking. For each of the core optimization objectives, a hybrid knowledge retrieval is performed in a preset experience base to obtain matching knowledge items; By combining each core optimization objective with the corresponding knowledge matching item, multi-path reasoning is performed on the current simulated expression to obtain multiple reasoning chains in sequence; The current inference result is obtained by combining multiple inference chains.

4. The method according to claim 3, characterized in that, The doctor intelligent agent subsystem also includes a doctor intelligent agent, which generates a response dialogue based on the current inference result and updates the progress of the goal achievement, including: The assistant agent retrieves historical cases from the experience base and filters them to obtain target historical cases. The target historical cases, historical dialogues, and the current simulated statements are processed into reference experience; The doctor's intelligent agent integrates the current reasoning result with the reference experience to generate a response dialogue to the current simulated expression in the current patient role dialogue scenario; The assistant agent evaluates the progress of achieving the core optimization goal based on all historical dialogues and reasoning processes, and obtains the progress of achieving the core optimization goal. Based on the core optimization objective, the progress score is updated, and an explanatory text is attached.

5. The method according to claim 3, characterized in that, The response to the dialogue end command, after ending the dialogue, includes: The progress of achieving each target is compared with a preset progress threshold, and the number of targets that have reached the progress threshold is counted. If the target number of the current case reaches the preset case selection criteria, then the current case will be designated as a classic case. The classic cases are then structured. Structured classic cases are stored in the experience base for later retrieval.

6. The method according to claim 1, characterized in that, After the verification agent verifies the current simulated representation, the process includes: If the verification of the testing agent fails, a correction instruction is generated and sent to the patient agent. Identify all historical dialogues; The patient's intelligent agent processes the correction instructions, the historical dialogue, and the current input to obtain a secondary simulated expression in the current patient role's dialogue scenario; The current simulation statement is updated based on the second-order simulation statement.

7. The method according to claim 1, characterized in that, The verification of the current simulation statement by the verification agent includes: Perform format consistency verification on the current simulation representation; If the format consistency verification passes, then based on the persona information, the information fit of the current simulated expression is verified. If the information fit verification passes, then the authenticity of the current simulated statement is verified based on the current medical facts. If the authenticity verification passes, then all verifications are considered successful. The verification pass status is transmitted to the doctor's intelligent agent subsystem for further reasoning.

8. A multi-objective dynamic reasoning system for medical multi-agent systems, characterized in that, The method of any one of claims 1-7 is applied to the system, the system comprising: The patient agent is used to process the current input related to the patient and generate a current simulated statement in the current patient role dialogue scenario. The current simulated statement is an incomplete statement with bias. The current input includes the patient's current medical facts and the patient's persona information. A verification agent is used to verify the current simulated representation; The doctor intelligent agent subsystem is used to update the current multi-objective optimization set based on the progress of target achievement if the verification of the intelligent agent passes the test. The current multi-objective optimization set includes multi-level hierarchical optimization dimensions. The doctor intelligent body subsystem is also used for multi-path reasoning to perform hybrid retrieval of the current simulated expression based on the current multi-objective optimization set, so as to obtain the current reasoning result; The doctor's intelligent body subsystem is also used to generate a response dialogue based on the current reasoning result and update the progress of the goal achievement; The doctor's intelligent body subsystem is also used to end conversations in response to end-of-conversation commands.

9. A computer device for multi-agent, multi-objective dynamic reasoning in medical settings, characterized in that: The method includes a memory, a processor, and a multi-objective dynamic reasoning program for medical multi-agent systems stored in the memory and executable on the processor. The processor executes the multi-objective dynamic reasoning program for medical multi-agent systems to implement the steps of the multi-objective dynamic reasoning method for medical multi-agent systems according to any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores a program for implementing a multi-objective dynamic reasoning method for medical multi-agent systems, which is executed by a processor to implement the steps of the multi-objective dynamic reasoning method for medical multi-agent systems as described in any one of claims 1 to 7.

Images