AI expert training method based on multi-team competition and knowledge genome repository

Abstract

This invention relates to an AI expert training method based on multi-team competition and a knowledge genome repository, comprising: receiving an input task and retrieving historical genomes from the knowledge genome repository according to vector semantic similarity; configuring solution teams with different seed strategies to solve the input task in parallel; ranking the output solutions of each solution team according to a comprehensive confidence score; writing the winning solution and candidate solutions that meet preset inclusion conditions into the knowledge genome repository in a structured genome format; calculating the maximum cosine similarity between the vector embedding of the winning solution's corresponding genome and the vector embedding of existing genomes in the knowledge genome repository; and updating the professional maturity score of the workspace. The beneficial effects of this invention are: it can simultaneously cover three different solution paths—historical experience utilization, differentiated path exploration, and solution without historical constraints—in a single task processing, thereby reducing the additional inference waiting time caused by repeated retries of a single path.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and in particular relates to an AI expert training method based on multi-team competition and knowledge genome repository. Background Technology

[0002] The rapid development of Large Language Models (LLMs) has spurred the emergence of automated task-solving systems with LLMs as their core inference engine. These systems (such as multi-agent frameworks like AutoGPT and MetaGPT) enable the large language model to autonomously analyze, plan, and solve user-input tasks by configuring it with tool-calling capabilities and multi-round inference mechanisms, demonstrating capabilities in general tasks such as code generation, document processing, and data analysis.

[0003] Existing large language model (LLM) driven automatic task solving systems suffer from the following technical problems when dealing with domain-specific continuous tasks:

[0004] The knowledge gained from solving a single task is not cumulative. In existing systems, each task is executed independently, and historical successful solutions cannot be persistently stored in a structured format. Therefore, the system's solving capability does not increase with the number of tasks.

[0005] Systematic biases of single agents. Single models or single-prompt-word strategies have blind spots in specific domains, failing to discover diverse solutions through internal competition mechanisms, thus limiting the upper limit of output quality.

[0006] There is a lack of a mechanism to quantify the level of expertise. Existing systems cannot quantitatively determine the level of expertise achieved in a specific field, nor can they determine when an expert-level capability has been acquired.

[0007] The knowledge silo effect across workspaces. Expertise accumulated in one domain (including historical solutions, heuristics, and evaluation use cases) cannot be transferred to a new work environment, and each deployment must start from scratch.

[0008] The ineffective accumulation of redundant knowledge from historical tasks. The lack of a mechanism to judge the novelty of new task solutions relative to the existing knowledge base leads to the continuous addition of duplicate or similar solutions to the knowledge base, reducing knowledge density. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the prior art and provide an AI expert training method based on multi-team competition and knowledge genome repository.

[0010] Firstly, it provides an AI expert training system based on multi-team competition and a knowledge genome repository, including:

[0011] The basic processing module is configured to receive input tasks and drive the large language model agent to solve the tasks.

[0012] The multi-team parallel competition module is configured to simultaneously launch at least two types of solution teams with different seed strategies to solve the same input task in parallel, and sort the output solutions of each solution team according to the comprehensive confidence score, and determine the solution with the highest ranking as the winner.

[0013] The knowledge genome repository is configured to: persistently store schemes that meet preset entry conditions and are output by multiple teams' parallel competition modules in a structured genome format, and perform vector embedding encoding on each genome to support semantic similarity retrieval;

[0014] The post-task knowledge growth module is configured to: after each task is completed, based on the comparison result of the maximum cosine similarity between the vector embedding of the winning solution's corresponding genome and the existing genome vector embedding in the knowledge genome repository and the preset novelty threshold, determine whether to trigger a knowledge growth operation including heuristic rule extraction.

[0015] The maturity scoring module is configured to calculate the professional maturity score of the workspace after each task is completed, based on the number of genomes in the knowledge genome repository, the proportion of task types covered by the accumulated genomes, and the historical evaluation results after time decay weighting.

[0016] Preferably, the solution teams with at least two different seed strategies include:

[0017] An elite team, configured to solve problems using the historical best-ranked genome retrieved from the knowledge genome repository as the knowledge seed;

[0018] Orthogonal teams, configured to solve problems using historically non-optimal ranked genomes as knowledge seeds; and

[0019] Blank team, which is configured to solve without injecting any historical knowledge seeds;

[0020] The structured genome format includes at least a core logical description field group, an optimization strategy field group, a preprocessing method field group, and an evaluation method field group.

[0021] Preferably, the multi-team parallel competition module further includes:

[0022] An adversarial red team unit, configured to generate a list of adversarial challenges, including challenge descriptions and severity levels, for each solver team's output solution; and

[0023] An iterative repair unit is configured to drive the engineering agent to repair the output solution when there is a fatal challenge in the adversarial challenge list, and to call the adversarial red team unit again after the repair is completed, until there are no fatal challenges or the number of repairs reaches a preset limit.

[0024] Preferably, the comprehensive confidence score is calculated by weighting the following components:

[0025] The theoretical support component is calculated based on the degree to which the solution scheme satisfies the preset rule constraints, input integrity constraints, inference step integrity requirements, and output format requirements;

[0026] The empirical validation score component is calculated based on the solution's performance in pre-defined test cases, sandbox execution environments, or benchmark datasets, and its weight is greater than that of other components; and

[0027] The red team pass rate component is calculated based on the proportion of challenges that were judged as passed or passed after repair in the adversarial red team unit to the total number of challenges.

[0028] Preferably, it also includes an expert mode activation unit, which is configured to: when the professional maturity score is not lower than a preset activation threshold, switch the task solving path selection strategy and the knowledge retrieval injection strategy, and write the verified heuristic rules in the workspace into the context control data area used to control the execution order and calling constraints of the solving agent;

[0029] A capsule serialization module, configured to: serialize validated heuristics in the workspace, the best-ranked genomes and their vector embeddings in the knowledge genome repository, reusable tool functions, and evaluation use cases into professional knowledge capsule files, wherein the professional knowledge capsule files record the professional maturity scores of the source workspace; and

[0030] A cold start guidance unit is configured to: when the target workspace completes the import of the professional knowledge capsule file, if the semantic similarity retrieval of the knowledge genome repository does not hit any genome when a new task is input, then the best-ranked genome in the professional knowledge capsule file is directly used as the knowledge seed for the new task.

[0031] Secondly, a method for training AI experts based on multi-team competition and a knowledge genome repository is provided, executed by any of the systems described in the first aspect, including:

[0032] S1. Receive the input task and retrieve the historical genome from the knowledge genome repository based on vector semantic similarity;

[0033] S2. Configure at least two types of solution teams with different seed strategies to solve the input task in parallel, wherein the solution team uses the historical genome or empty knowledge as knowledge seeds;

[0034] S3. Rank the output solutions of each solution team according to the comprehensive confidence score, and determine the solution with the highest ranking as the winning solution.

[0035] S4. Write the winning solution and the candidate solutions that meet the preset entry conditions into the knowledge genome repository in a structured genome format.

[0036] S5. Calculate the maximum cosine similarity between the vector embedding of the winning scheme's genome and the existing genome vector embeddings in the knowledge genome repository. When the maximum cosine similarity is lower than a preset novelty threshold, write the winning scheme's genome into the knowledge genome repository and perform a heuristic rule extraction operation; otherwise, skip the heuristic rule extraction operation.

[0037] S6. Update the professional maturity score of the workspace based on the total number of genomes in the knowledge genome repository, the proportion of task types covered by the accumulated genomes, and historical evaluation results.

[0038] As a preferred option, the at least two types of solution teams configured in S2 include: an elite solution team that uses the best-ranked genome in the search results as the knowledge seed, an orthogonal solution team that uses non-best-ranked genomes as the knowledge seed, and a blank solution team that does not inject historical knowledge seeds.

[0039] In S6, the professional maturity score is updated by weighted summation of three components. The first component is determined based on the total number of genomes in the knowledge genome repository, the second component is determined based on the proportion of task types covered by the accumulated genomes, and the third component is determined by weighted average of historical evaluation results with time decay weight.

[0040] The weighting of the timeliness difference of historical evaluation results is set as follows: the weight of evaluation results whose generation time is no more than a first preset number of days from the current time is the first weight value, the weight of evaluation results that exceed the first preset number of days but do not exceed the second preset number of days is the second weight value, the evaluation results that exceed the second preset number of days are not included in the calculation, and the first weight value is greater than the second weight value.

[0041] Preferably, the method also includes S7, where, when the professional maturity score is not lower than a preset activation threshold, the expert mode is activated, and the validated heuristic rules in the workspace are injected into the system prompts of the planning agent.

[0042] Thirdly, a method for exporting and transferring AI expertise capsules is provided, the method being executed based on the system described in any one of the first aspects, including the following steps:

[0043] S1. Extract the optimal ranking genome and its vector embedding from the knowledge genome repository of the system, and extract the validated heuristic rules, reusable tool functions and evaluation use cases from the system;

[0044] S2. The extracted optimal ranking genome and its vector embedding, heuristic rules, tool functions and evaluation use cases, as well as the professional maturity score of the system, are serialized into a professional knowledge capsule file;

[0045] S3. Import the expertise capsule file into the target workspace, reconstruct the extracted knowledge components in the target workspace, and set the initial expertise maturity score of the target workspace to the expertise maturity score recorded in the expertise capsule file; and

[0046] S4. When the target workspace receives a new task and its knowledge genome repository does not match any genome in the vector semantic similarity search, the best-ranked genome reconstructed in the professional knowledge capsule file is directly used as the knowledge seed of the new task.

[0047] Fourthly, a computer storage medium is provided, wherein a computer program is stored therein; when the computer program is run on a computer, the computer causes the computer to perform any of the methods described in the second aspect.

[0048] The beneficial effects of this invention are:

[0049] 1. This invention adopts a multi-team heterogeneous seed parallel competition structure. Solving teams with different seed strategies execute task solving in parallel within the same computation batch, enabling the system to simultaneously cover three different solution paths in a single task processing, namely historical experience utilization, differentiated path exploration, and solution without historical constraints. This reduces the additional inference waiting time caused by repeated retries of a single path and improves the coverage of candidate solutions for complex proprietary domain tasks.

[0050] 2. This invention employs a structured knowledge genome repository, which enables solution schemes that meet the inclusion criteria to be persistently stored and semantically retrieved using a unified field structure. This allows subsequent tasks to directly retrieve and reuse preprocessing methods, core logic, optimization strategies, and evaluation methods from historical solution schemes, thereby reducing the computational overhead caused by repeatedly generating similar solution structures.

[0051] 3. This invention employs a novelty gating mechanism that occurs before the core knowledge genome repository is written. Only when the proposed solution has a substantial difference from the existing knowledge in the repository is it allowed to be written into the core knowledge genome repository and trigger heuristic rule extraction and tool extraction. For solutions that are highly similar to existing knowledge, they are only saved as competition process records and are not written into the core knowledge genome repository, thereby suppressing the redundant growth of the core knowledge repository, reducing the write I / O load of the core database and the memory occupation pressure of the similarity retrieval stage.

[0052] 4. This invention adopts a three-component time decay maturity scoring mechanism. The system can objectively calculate the professional maturity score of the current workspace based on the scale of core knowledge accumulation, the degree of task type coverage, and the quality of historical evaluation. It can also switch the task solution path selection strategy and knowledge retrieval injection strategy according to the maturity score, so that the system resource scheduling and knowledge call method match the current knowledge accumulation level.

[0053] 5. This invention employs a professional knowledge capsule serialization and cold start guidance mechanism. The structured knowledge components accumulated in the source workspace can be imported into the target workspace and directly participate in the solution during the initial task phase of the target workspace. This reduces the inference computing power required for repeated accumulation of similar knowledge in the cold start phase of the newly deployed workspace and the storage resource occupation caused by repeated generation of vector embedding. Attached Figure Description

[0054] Figure 1 This is a schematic diagram of the overall architecture of the AI ​​expert training system 100 for a specific field according to the present invention;

[0055] Figure 2 This is a schematic diagram of the three-component weighted calculation and expert mode activation of the maturity scoring module of this invention;

[0056] Figure 3 This is a flowchart of the six types of parallel extraction and novelty gating of the post-task knowledge growth module of this invention;

[0057] Figure 4 This is a flowchart of the task solving and genome input process of the multi-team parallel competition module of this invention;

[0058] Figure 5 This is a flowchart of the export, import, and cold start guidance process of the capsule serialization module of this invention;

[0059] Figure labeling: 100 - Expert training system; 110 - Basic processing module; 120 - Multi-team parallel competition module; 121 - Elite team; 122 - Orthogonal team; 123 - Blank team; 124 - Adversarial red team unit; 125 - Iterative repair unit; 130 - Knowledge genome repository; 140 - Post-task knowledge growth module; 150 - Maturity scoring module; 151 - Expert mode activation unit; 152 - Maturity calculation unit; 160 - Capsule serialization module; 161 - Capsule export unit; 162 - Cold start guidance unit; 511 - Source workspace; 512 - Knowledge component; 513 - Professional knowledge capsule file; 514 - Target workspace. Detailed Implementation

[0060] The present invention will be further described below with reference to embodiments. The description of the embodiments below is only for the purpose of helping to understand the present invention. It should be noted that those skilled in the art can make several modifications to the present invention without departing from the principle of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

[0061] Example 1:

[0062] To address the problems of existing technologies, Embodiment 1 of this invention provides an AI expert training system 100 based on multi-team competition and a knowledge genome repository, such as... Figure 1 As shown, it includes:

[0063] The basic processing module 110 is configured to receive input tasks and drive the large language model agent to solve the tasks.

[0064] Specifically, the basic processing module 110 is responsible for receiving the externally input task description, performing domain identification, complexity assessment, and task DNA extraction, and then transmitting the task DNA to the multi-team parallel competition module 120. The task DNA includes domain labels, task target types, data structure features, and task characteristic key-value pairs.

[0065] The multi-team parallel competition module 120 is configured to simultaneously launch at least two types of solution teams with different seed strategies to solve the same input task in parallel. Each solution team contains an independent planning agent, engineering agent, and task management agent. The output schemes of each solution team are ranked according to the comprehensive confidence score, and the scheme with the highest ranking is determined as the winning scheme.

[0066] Among them, at least two types of solution teams with different seed strategies include:

[0067] Elite team 121 is configured to solve the problem using the historical best-ranked genome retrieved from the knowledge genome repository 130 as the knowledge seed.

[0068] Orthogonal team 122, which is configured to solve the problem using historically non-optimal ranked genomes as knowledge seeds;

[0069] Blank team 123, which is configured to solve without injecting any historical knowledge seeds;

[0070] The structured genome format includes at least a core logical description field group, an optimization strategy field group, a preprocessing method field group, and an evaluation method field group.

[0071] The multi-team parallel competition module 120 also includes:

[0072] Adversarial red team unit 124, configured to generate a list of adversarial challenges, including challenge descriptions and severity levels, for each solver team's output solution; and

[0073] The iterative repair unit 125 is configured to drive the engineering agent to repair the output scheme when there is a fatal challenge in the adversarial challenge list, and to call the adversarial red team unit 124 again after the repair is completed, until there is no fatal challenge or the number of repairs reaches a preset limit.

[0074] This invention also provides the workflow of the multi-team parallel competition module 120, such as... Figure 4 As shown, it specifically includes:

[0075] Step 201: The basic processing module 110 receives the input task, completes the task profile extraction, and then triggers the multi-team parallel competition module 120.

[0076] Step 202: The multi-team parallel competition module 120 initiates a vector semantic similarity retrieval to the knowledge genome repository 130. During the retrieval, the text description of the input task is converted into a high-dimensional vector embedding through a vector encoding model. Using cosine distance as a metric, the module retrieves several historical genomes with the highest similarity in the knowledge genome repository 130 and returns them grouped into best-ranked / alternative / not recommended groups.

[0077] Step 203: Based on the search results of Step 202, three types of solution teams are configured simultaneously: Elite Team 121 uses several entries from the optimal ranking genome and candidate ranking genomes returned by the search as knowledge seeds, and injects them into the context of the planning agent of Elite Team 121; Orthogonal Team 122 uses several entries from the candidate ranking and non-recommended ranking genomes returned by the search that do not overlap with the optimal ranking genome as knowledge seeds, and guides the planning agent to explore a solution path that is different from that of Elite Team 121; Blank Team 123 does not inject any historical knowledge seeds, and the planning agent solves independently under the condition of no historical information constraints, so as to prevent the whole from getting trapped in local optima.

[0078] Step 204: The elite team 121, the orthogonal team 122, and the blank team 123 perform the solution in parallel. The planning agent, engineering agent, and task management agent within each team cooperate to complete the solution planning, implementation, and self-checking, and output their respective solution solutions.

[0079] Step 205: The adversarial red team unit 124 receives the solution proposals from each team and generates an adversarial challenge list for each proposal. Each challenge includes a challenge description and a severity level, which is divided into three levels: fatal, major, and minor.

[0080] Step 206: The iterative repair unit 125 performs repairs on solutions with critical challenges: Repair prompts are generated based on the challenge description, driving the corresponding team's engineering agent to repair the solution. After repair, the adversarial red team unit 124 is invoked again to generate a new challenge list. If critical challenges still exist in the new challenge list and the number of repairs performed has not reached the preset maximum number of repairs, then repair continues; otherwise, the repair loop exits. In one embodiment, the preset maximum number of repairs is 2.

[0081] Step 207: The multi-team parallel competition module 120 calculates a comprehensive confidence score for the final solutions of all teams. The comprehensive confidence score is calculated by weighting three components: a theoretical support component reflecting the theoretical reliability of the solution, an empirical verification component reflecting the empirical effectiveness of the solution, and a red team pass rate component reflecting the degree to which the solution passes adversarial review. The weight of the empirical verification component is greater than the weight of the theoretical support component and the weight of the red team pass rate component. All team solutions are ranked from highest to lowest according to the comprehensive confidence score, and the solution with the highest score is determined as the winning solution, and the corresponding team is determined as the winning team.

[0082] Step 208: The winning solution and candidate solutions that meet the preset inclusion criteria are written into the knowledge genome repository 130 in a structured genome format. The genome marker corresponding to the winning solution ranked first is the optimal ranking; the remaining candidate solutions that meet the preset inclusion criteria can be marked as alternative rankings or other candidate markers; the remaining team solutions that do not meet the preset inclusion criteria are stored separately as competition process records in the process record area or review record area, and are not written into the knowledge genome repository 130. The structured genome format includes a core logic description field group, an optimization strategy field group, a preprocessing method field group, an evaluation method field group, performance indicator JSON records, and a failure boundary description field group.

[0083] In addition, the overall confidence score is calculated by weighting the following components:

[0084] The theoretical support component is calculated based on the degree to which the solution scheme satisfies the preset rule constraints, input integrity constraints, inference step integrity requirements, and output format requirements;

[0085] The empirical validation score component is calculated based on the solution's performance in pre-defined test cases, sandbox execution environments, or benchmark datasets, and its weight is greater than that of other components; and

[0086] The red team pass rate component is calculated based on the proportion of challenges that were judged as passed or passed after repair among the challenges generated by the adversarial red team unit 124 to the total number of challenges.

[0087] For example, the formula for calculating the composite confidence score is:

[0088] Confidence score = W 理论 × Theoretical support + W 实证 × Empirical verification degree + W 红队 × Red Team's test pass rate.

[0089] Among them, W 理论 W 实证 and W 红队 Let Wempirical represent the weights corresponding to theoretical support, empirical validation, and red team test pass rate, respectively, where Wempirical > Wtest. 理论 W 实证 >W 红队 In a specific example, Wtheoretical = 0.3, Wempirical = 0.4, and Wred team = 0.3.

[0090] The AI ​​expert training system 100 also includes:

[0091] The knowledge genome repository 130 is connected to a vector database and stores the structured genomes and their vector embeddings of all historical task solution schemes. Specifically, the knowledge genome repository 130 is configured to persistently store the solutions output by the multi-team parallel competition module 120 that meet the preset entry conditions in a structured genome format, and to perform vector embedding encoding on each genome to support semantic similarity retrieval.

[0092] The post-task knowledge growth module 140 performs genome embedding into the repository, heuristic rule extraction, tool function extraction, failure rule learning, and process memory extraction in parallel after each task is completed, and writes the extraction results to persistent storage in the workspace. Specifically, the post-task knowledge growth module 140 is configured to: after each task is completed, based on the comparison result of the maximum cosine similarity between the vector embedding of the winning genome and the existing genome vector embeddings in the knowledge genome repository 130 and a preset novelty threshold, determine whether to trigger knowledge growth operations, including heuristic rule extraction.

[0093] This embodiment of the invention also provides a detailed workflow for the post-task knowledge growth module 140, such as... Figure 3 As shown, it includes:

[0094] Step 301: Obtain the structured genome of the winning scheme determined in step 208.

[0095] Step 302: Generate vector embeddings for the winning genome of the proposed genome to be added to the database. These vector embeddings are then compared for similarity with the vector embeddings of existing genomes in the knowledge genome repository 130.

[0096] Step 303: Search the knowledge genome repository 130 for the existing genome with the highest cosine similarity to the genome vector embedding of the winning scheme to be included in the repository, and obtain the highest cosine similarity value.

[0097] Step 304: Compare the highest cosine similarity value with a preset novelty threshold. If the highest cosine similarity value is lower than the preset novelty threshold, the winning solution is determined to be novel relative to existing knowledge in the repository, and proceed to step 305; if the highest cosine similarity value is not lower than the preset novelty threshold, the winning solution is determined to be highly similar to existing knowledge in the repository, and proceed to step 308. For example, the preset novelty threshold is a value between 0.85 and 0.95.

[0098] Step 305: Write the winner's genome into the knowledge genome repository 130.

[0099] Step 306: After step 305 is completed, perform heuristic rule extraction and utility function extraction.

[0100] Step 307: Record the evidence of the solution method for this task; when iterative repairs occur during the solution process, extract the process memory record; when the workspace maturity score reaches the preset lower limit, execute the failure rule learning.

[0101] Step 308: Do not write the winning solution into the core knowledge genome repository 130, but save it only as a competition process record or review record, and skip the heuristic rule extraction and tool function extraction, and then proceed to step 307.

[0102] The maturity scoring module 150 includes a maturity calculation unit 152 and an expert mode activation unit 151. Specifically, the maturity calculation unit 152 is configured to calculate the professional maturity score of the workspace after each task is completed, based on the number of genomes in the knowledge genome repository 130, the proportion of task types covered by accumulated genomes, and the historical evaluation results after time decay weighting. The expert mode activation unit 151 determines whether to activate the expert mode based on the professional maturity score and the historical evaluation pass rate. Specifically, the expert mode activation unit 151 is configured to switch the task solution path selection strategy and the knowledge retrieval injection strategy when the professional maturity score is not lower than a preset activation threshold, and write the verified heuristic rules in the workspace into the context control data area used to control the execution order and call constraints of the solving agent.

[0103] This embodiment of the invention also provides the working principle of the maturity scoring module 150. In this embodiment, the maturity scoring module 150 is used to calculate the professional maturity score M of the workspace. The professional maturity score M is obtained by weighted summation of the log-normalized component of genome quantity, the task type coverage component, and the time decay assessment confidence component, as shown below. Figure 2 As shown.

[0104] The log-normalized component of the genome number (i.e., the first component) can be expressed as:

[0105]

[0106] Where N represents the total number of genomes stored in the knowledge genome repository 130, and W1 represents the weight coefficient of the first component. The first component increases with the number of genomes and reaches saturation when N≥100. In a specific embodiment, W1=30, and It is approximately 6.658, so when N=100, the first component reaches its maximum value of 30.

[0107] The task type coverage component (i.e., the second component) can be represented as:

[0108]

[0109] Where T represents the set of typical task types preset in the workspace, G represents the set of genomes already stored in the knowledge genome repository 130, and g.coreLogic represents the core logic description field group of genome g. W1 represents the number of elements in the set, and W2 represents the weight coefficient of the second component. The second component is determined by the coverage ratio of the core logical description fields of the accumulated genome to the typical task type set T. In a specific embodiment, W2 = 40. When T is an empty set, the second component takes full marks.

[0110] The confidence component 403 (i.e., the third component) of the time decay assessment can be expressed as:

[0111]

[0112] Where E represents the set of historical evaluation results of the workspace, e.score represents the score of a single evaluation result, which is normalized to the range of 0 to 100, w(e) represents the time decay weight function, and W3 represents the weight coefficient of the third component.

[0113] The time decay weight function w(e) assigns different weights based on the time interval between the generation time of the evaluation result and the current time. Specifically, if the generation time of the evaluation result is no more than 60 days from the current time, the weight is 1.0; if it is more than 60 days but no more than 120 days, the weight is 0.5; if it is more than 120 days, the weight is 0 and is not included in the calculation. In a specific embodiment, W3=30.

[0114] Based on the above three components, the final professional maturity score M can be expressed as:

[0115]

[0116] The value of M ranges from 0 to 100.

[0117] Furthermore, the expert mode activation unit 151 determines whether to activate expert mode 405 based on the following conditions:

[0118] First, determine whether the professional maturity score M is not lower than a preset activation threshold. If M is lower than the preset activation threshold, maintain the normal mode. If M is not lower than the preset activation threshold, further determine whether the number of historical evaluation results is less than a preset minimum number of evaluations. When the number of historical evaluation results is less than the preset minimum number of evaluations, directly activate the expert mode. When the number of historical evaluation results is not less than the preset minimum number of evaluations, continue to determine whether the historical evaluation pass rate is not lower than a preset pass rate threshold. If it is not lower than the preset pass rate threshold, activate the expert mode; otherwise, maintain the normal mode.

[0119] After activating the expert mode, the expert mode activation unit 151 performs the following operations: switches the task solution path selection strategy and the knowledge retrieval injection strategy, and writes the heuristic rules of the confirmed state in the workspace into the context control data area used to control the execution order and call constraints of the solving agent; at the same time, it switches the review intensity level of the adversarial red team unit to the strict level to strengthen the constraint verification of the solution scheme.

[0120] In a specific example, the preset activation threshold is 70, the preset minimum number of evaluations is 30, and the preset pass rate threshold is 50%.

[0121] After activating expert mode, the expert mode activation unit 151 performs the following operations:

[0122] First, inject the heuristic rules that are in the confirmed state in the workspace into the system prompts of the planning agent, so that the planning agent will give priority to referring to the heuristic rules in subsequent tasks to solve the planning problem;

[0123] Second, the intensity level of the red team's review will be raised to the strict level to strengthen the adversarial review of the solution.

[0124] The capsule serialization module 160 includes a capsule export unit 161 and a cold start boot unit 162. The capsule export unit 161 serializes workspace knowledge into professional knowledge capsule files; the cold start boot unit 162 handles the initial task seed injection problem of the target workspace after capsule import.

[0125] Specifically, the capsule export unit 161 is configured to: serialize validated heuristics in the workspace, the best-ranked genomes and their vector embeddings in the knowledge genome repository 130, reusable tool functions, and evaluation use cases into a professional knowledge capsule file, wherein the professional knowledge capsule file records the professional maturity score of the source workspace; and

[0126] The cold start guidance unit 162 is configured such that: after the target workspace completes the import of the professional knowledge capsule file, if the semantic similarity retrieval of the knowledge genome repository 130 does not hit any genome when a new task is input, the best-ranked genome in the professional knowledge capsule file is directly used as the knowledge seed for the new task.

[0127] Specifically, such as Figure 5 As shown, the complete workflow of the capsule serialization module 160 is as follows:

[0128] Capsule export process:

[0129] Step 501: Confirm that the source workspace 511 is in an exportable state, and read the current professional maturity score calculated by the maturity scoring module 150 as the capsule maturity record value.

[0130] Step 502: Extract the following knowledge components 512 from the source workspace 511: the optimal ranking genome and its vector embedding in the knowledge genome repository 130; heuristic rules in the workspace heuristic rule base that are in the confirmed state; reusable tool functions recorded in the workspace tool function library; and evaluation use cases and historical evaluation results stored in the workspace evaluation use case library.

[0131] Step 503: Serialize the knowledge components extracted in Step 502 into a professional knowledge capsule file 513 according to a preset compression level. The compression level is divided into three levels: at the minimum compression level, the professional knowledge capsule file 513 includes genome description content and heuristic rules, but does not include vector embeddings; at the standard compression level, it further includes vector embeddings on top of the minimum compression; at the full compression level, it further includes knowledge graph nodes, knowledge graph edges, and process memory records on top of the standard compression. The structured format of the professional knowledge capsule file 513 records the compression level, source workspace domain label, vector encoding model information, capsule mode version, and export timestamp, and records the maturity score obtained in Step 501 as the capsule maturity record value.

[0132] Capsule delivery process:

[0133] Step 504: Read the expertise capsule file 513 in the target workspace 514.

[0134] Step 505: Reconstruct the genome, heuristic rules, tool functions, evaluation use cases, and, if the compression level is a complete knowledge graph, and process memory records contained in the professional knowledge capsule file 513 in the target workspace 514. If the genome in the professional knowledge capsule file 513 does not contain the minimum compression level of vector embeddings, then regenerate the vector embeddings for the reconstructed genome and store them in the knowledge genome repository of the target workspace 514. Set the initial professional maturity score of the target workspace 514 to the capsule maturity record value. If the capsule maturity record value is not lower than a preset activation threshold, then simultaneously activate the expert mode of the target workspace 514.

[0135] Cold start boot process:

[0136] Step 506: After receiving the first new task, the target workspace 514 initiates a vector semantic similarity retrieval 516 to the reconstructed knowledge genome repository according to the conventional process.

[0137] Step 507: Determine whether the search in step 506 hit any genomes. If the search hits (517 - No), configure the solution team seed using the search results according to the normal process; if the search does not hit (517 - Yes), i.e., in the cold start state, the cold start guidance unit 161 directly loads several genomes with the best ranking from the genome reconstructed in step 505 as the knowledge seed for the new task, bypassing the vector semantic similarity threshold, ensuring that the target workspace 514 can utilize the historical knowledge carried in the capsule when receiving the first task. In a specific embodiment, the maximum number of genomes with the best ranking loaded by the cold start guidance is 6.

[0138] To verify the effectiveness of the technical solution of this invention, a comparative experiment was also conducted on the structured clinical medical computing task set MedCalc. Under the same task set, the same cold start, and no initial knowledge injection, the task solving performance of using the technical solution of this invention and not using the technical solution of this invention was compared.

[0139] In one example, using qwen3-max as the base model, the task solving accuracy was 68.5% without adopting the technical solution of this invention; after adopting the technical solution of this invention and completing the preset training task and accumulating knowledge, the task solving accuracy increased to 77.2%.

[0140] In another example, using qwen2.5-7B as the base model, the task solving accuracy was 16.4% without importing the expertise capsule; after importing the expertise capsule, the task solving accuracy improved to 61.8%.

[0141] The above results show that the present invention can significantly improve the solution performance of domain-specific tasks in the cold start phase by forming and accumulating a structured knowledge genome through multi-team competition and then migrating it to a new workspace through a knowledge capsule. This reduces the task trial-and-error cost and repeated reasoning overhead required to repeatedly accumulate equivalent knowledge from scratch in the target workspace.

[0142] Furthermore, the multi-team parallel competition module 120 can dynamically adjust the type and number of solution teams launched based on the computing resource budget. When computing resources are sufficient, three types of teams—elite team 121, orthogonal team 122, and blank team 123—are launched simultaneously. When computing resources are limited, only elite team 121 and blank team 123, or only elite team 121, can be launched to complete a single round of solving with fewer resources, while subsequent tasks can still benefit from the accumulated genome.

[0143] In another example, the preset novelty threshold can be dynamically configured by the workspace administrator based on the domain knowledge density requirements. In knowledge-intensive proprietary domains (such as software engineering), the threshold can be set to a higher value (e.g., 0.90 to 0.95) to more strictly filter redundancy; in emerging domains where knowledge is relatively sparse, the threshold can be set to a lower value (e.g., 0.85 to 0.88) to retain more nuanced knowledge records.

[0144] In yet another example, the capsule sequencer module 160 can support incremental capsule export, that is, exporting only the genome, heuristic rules and tool functions added since the last capsule export, in order to reduce the size of the capsule file and facilitate incremental knowledge synchronization across workspaces.

[0145] Example 2:

[0146] Based on Example 1, Example 2 of the present invention provides an AI expert training method based on multi-team competition and a knowledge genome repository, including:

[0147] S1. Receive the input task and retrieve the historical genome from the knowledge genome repository 130 according to vector semantic similarity.

[0148] S2. Configure at least two types of solution teams with different seed strategies to solve the input task in parallel, wherein the solution teams use the historical genome or empty knowledge as knowledge seeds.

[0149] The at least two types of solver teams configured in S2 include: an elite team 121 that uses the best-ranked genome in the search results as the knowledge seed, an orthogonal team 122 that uses non-best-ranked genomes as the knowledge seed, and a blank team 123 that does not inject historical knowledge seeds.

[0150] S3. Rank the output solutions of each solution team according to the comprehensive confidence score, and determine the solution with the highest ranking as the winner.

[0151] S4. Write the winning scheme and the candidate schemes that meet the preset entry conditions into the knowledge genome repository 130 in a structured genome format.

[0152] S5. Calculate the maximum cosine similarity between the vector embedding of the winning scheme's corresponding genome and the existing genome vector embeddings in the knowledge genome repository 130. When the maximum cosine similarity is lower than a preset novelty threshold, write the winning scheme's genome into the knowledge genome repository 130 and perform a heuristic rule extraction operation; otherwise, skip the heuristic rule extraction operation.

[0153] S6. Update the professional maturity score of the workspace based on the total number of genomes in the knowledge genome repository 130, the proportion of task types covered by the accumulated genomes, and historical evaluation results.

[0154] In S6, the professional maturity score is updated by weighted summation of three components. The first component is determined based on the total number of genomes in the knowledge genome repository 130, the second component is determined based on the proportion of task types covered by the accumulated genomes, and the third component is determined by weighted average of historical evaluation results with time decay weights.

[0155] The weighting of the timeliness difference of historical evaluation results is set as follows: the weight of evaluation results whose generation time is no more than a first preset number of days from the current time is the first weight value, the weight of evaluation results that exceed the first preset number of days but do not exceed the second preset number of days is the second weight value, the evaluation results that exceed the second preset number of days are not included in the calculation, and the first weight value is greater than the second weight value.

[0156] Specifically, the steps for updating the professional maturity score include: calculating the professional maturity score M using the weighted formula M = W1×f(N) + W2×Coverage(T) + W3×g(E), where f(N) is the log-normalized function of the total number of genomes N, Coverage(T) is the proportion of accumulated genomes covering task types, g(E) is the weighted average function of historical evaluation results weighted by time difference, and W1, W2, and W3 are the weight coefficients of each component; when the professional maturity score M is not lower than a preset activation threshold, the expert mode is activated, and the validated heuristic rules in the workspace are injected into the system prompts of the planning agent.

[0157] S7. When the professional maturity score is not lower than the preset activation threshold, activate the expert mode and inject the verified heuristic rules in the workspace into the system prompt words of the planning agent.

[0158] It should be noted that the method provided in this embodiment is the same as the method of the system provided in embodiment 1. Therefore, the parts that are the same as or similar to those in embodiment 1 can be referred to each other and will not be repeated in this application.

[0159] Example 3:

[0160] Based on Example 1, Example 3 of the present invention provides a method for exporting and transferring AI expertise capsules, including the following steps:

[0161] S1. Extract the optimal ranking genome and its vector embedding from the knowledge genome repository 130 of the expert training system 100, and extract the validated heuristics, reusable tool functions and evaluation use cases from the expert training system 100.

[0162] S2. The extracted optimal ranking genome and its vector embedding, heuristic rules, tool functions and evaluation use cases, as well as the professional maturity score of the expert training system 100, are serialized into a professional knowledge capsule file.

[0163] S3. Import the knowledge capsule file into the target workspace, reconstruct the extracted knowledge components in the target workspace, and set the initial knowledge maturity score of the target workspace to the knowledge maturity score recorded in the knowledge capsule file.

[0164] S4. When the target workspace receives a new task and its knowledge genome repository 130 fails to find any genome through vector semantic similarity retrieval, the optimal ranking genome reconstructed in the professional knowledge capsule file is directly used as the knowledge seed for the new task.

[0165] It should be noted that the method provided in this embodiment is the same as the method of the system provided in embodiment 1. Therefore, the parts that are the same as or similar to those in embodiment 1 can be referred to each other and will not be repeated in this application.

Claims

1. An AI expert training system based on multi-team competition and a knowledge genome repository, characterized in that: include: The basic processing module is configured to receive input tasks and drive the large language model agent to solve the tasks. The multi-team parallel competition module is configured to simultaneously launch at least two types of solution teams with different seed strategies to solve the same input task in parallel, and sort the output solutions of each solution team according to the comprehensive confidence score, and determine the solution with the highest ranking as the winner. The knowledge genome repository is configured to: persistently store schemes that meet preset entry conditions and are output by multiple teams' parallel competition modules in a structured genome format, and perform vector embedding encoding on each genome to support semantic similarity retrieval; The post-task knowledge growth module is configured to: after each task is completed, based on the comparison result of the maximum cosine similarity between the vector embedding of the winning solution's corresponding genome and the existing genome vector embedding in the knowledge genome repository and the preset novelty threshold, determine whether to trigger a knowledge growth operation including heuristic rule extraction. The maturity scoring module is configured to calculate the professional maturity score of the workspace after each task is completed, based on the number of genomes in the knowledge genome repository, the proportion of task types covered by the accumulated genomes, and the historical evaluation results after time decay weighting.

2. The AI ​​expert training system based on multi-team competition and knowledge genome repository according to claim 1, characterized in that, The solution teams with at least two different seed strategies include: An elite team, configured to solve problems using the historical best-ranked genome retrieved from the knowledge genome repository as the knowledge seed; Orthogonal teams, configured to solve problems using historically non-optimal ranked genomes as knowledge seeds; and Blank team, which is configured to solve without injecting any historical knowledge seeds; The structured genome format includes at least a core logical description field group, an optimization strategy field group, a preprocessing method field group, and an evaluation method field group.

3. The AI ​​expert training system based on multi-team competition and knowledge genome repository according to claim 2, characterized in that, The multi-team parallel competition module also includes: An adversarial red team unit, configured to generate a list of adversarial challenges, including challenge descriptions and severity levels, for each solver team's output solution; and An iterative repair unit is configured to drive the engineering agent to repair the output solution when there is a fatal challenge in the adversarial challenge list, and to call the adversarial red team unit again after the repair is completed, until there are no fatal challenges or the number of repairs reaches a preset limit.

4. The AI ​​expert training system based on multi-team competition and knowledge genome repository according to claim 3, characterized in that, The overall confidence score is calculated by weighting the following components: The theoretical support component is calculated based on the degree to which the solution scheme satisfies the preset rule constraints, input integrity constraints, inference step integrity requirements, and output format requirements; The empirical verification degree component is calculated based on the achievement of the target indicators of the solution in the pre-set test cases, sandbox execution environment or benchmark dataset, and the weight of the empirical verification degree component is greater than the weight of other components. as well as The red team pass rate component is calculated based on the proportion of challenges that were judged as passed or passed after repair in the adversarial red team unit to the total number of challenges.

5. The AI ​​expert training system based on multi-team competition and knowledge genome repository according to claim 4, characterized in that, Also includes: The expert mode activation unit is configured to: when the professional maturity score is not lower than a preset activation threshold, switch the task solving path selection strategy and the knowledge retrieval injection strategy, and write the verified heuristic rules in the workspace into the context control data area used to control the execution order and calling constraints of the solving agent; A capsule serialization module is configured to serialize validated heuristic rules in the workspace, the best-ranked genome and its vector embedding in the knowledge genome repository, reusable tool functions and evaluation use cases into a professional knowledge capsule file, wherein the professional knowledge capsule file records the professional maturity score of the source workspace; as well as A cold start guidance unit is configured to: when the target workspace completes the import of the professional knowledge capsule file, if the semantic similarity retrieval of the knowledge genome repository does not hit any genome when a new task is input, then the best-ranked genome in the professional knowledge capsule file is directly used as the knowledge seed for the new task.

6. An AI expert training method based on multi-team competition and knowledge genome repository, characterized in that, Performed by the system according to any one of claims 1 to 5, comprising: S1. Receive the input task and retrieve the historical genome from the knowledge genome repository based on vector semantic similarity; S2. Configure at least two types of solution teams with different seed strategies to solve the input task in parallel, wherein the solution team uses the historical genome or empty knowledge as knowledge seeds; S3. Rank the output solutions of each solution team according to the comprehensive confidence score, and determine the solution with the highest ranking as the winning solution. S4. Write the winning solution and the candidate solutions that meet the preset entry conditions into the knowledge genome repository in a structured genome format. S5. Calculate the maximum cosine similarity between the vector embedding of the winning scheme's genome and the existing genome vector embeddings in the knowledge genome repository. When the maximum cosine similarity is lower than a preset novelty threshold, write the winning scheme's genome into the knowledge genome repository and perform a heuristic rule extraction operation; otherwise, skip the heuristic rule extraction operation. S6. Update the professional maturity score of the workspace based on the total number of genomes in the knowledge genome repository, the proportion of task types covered by the accumulated genomes, and historical evaluation results.

7. The AI ​​expert training method based on multi-team competition and knowledge genome repository according to claim 6, characterized in that, The S2 configuration includes at least two types of solution teams: an elite solution team that uses the best-ranked genome in the search results as the knowledge seed, an orthogonal solution team that uses non-best-ranked genomes as the knowledge seed, and a blank solution team that does not inject historical knowledge seeds. In S6, the professional maturity score is updated by weighted summation of three components. The first component is determined based on the total number of genomes in the knowledge genome repository, the second component is determined based on the proportion of task types covered by the accumulated genomes, and the third component is determined by weighted average of historical evaluation results with time decay weight. The weighting of the timeliness difference of historical evaluation results is set as follows: the weight of evaluation results whose generation time is no more than a first preset number of days from the current time is the first weight value, the weight of evaluation results that exceed the first preset number of days but do not exceed the second preset number of days is the second weight value, the evaluation results that exceed the second preset number of days are not included in the calculation, and the first weight value is greater than the second weight value.

8. The AI ​​expert training method based on multi-team competition and knowledge genome repository according to claim 7, characterized in that, It also includes S7, which activates the expert mode when the professional maturity score is not lower than the preset activation threshold, and injects the verified heuristic rules in the workspace into the system prompt words of the planning agent.

9. A method for exporting and transferring AI expertise capsules, characterized in that, The method is performed based on the system according to any one of claims 1 to 5, and includes the following steps: S1. Extract the optimal ranking genome and its vector embedding from the knowledge genome repository of the system, and extract the validated heuristic rules, reusable tool functions and evaluation use cases from the system; S2. The extracted optimal ranking genome and its vector embedding, heuristic rules, tool functions and evaluation use cases, as well as the professional maturity score of the system, are serialized into a professional knowledge capsule file; S3. Import the expertise capsule file into the target workspace, reconstruct the extracted knowledge components in the target workspace, and set the initial expertise maturity score of the target workspace to the expertise maturity score recorded in the expertise capsule file; and S4. When the target workspace receives a new task and its knowledge genome repository does not match any genome in the vector semantic similarity search, the best-ranked genome reconstructed in the professional knowledge capsule file is directly used as the knowledge seed of the new task.

10. A computer storage medium, characterized in that, The computer storage medium stores a computer program; when the computer program is run on the computer, it causes the computer to perform the method described in any one of claims 6 to 8.

Images