A method and system for post-position matching based on multi-agent collaborative reasoning

Abstract

The present application relates to the technical field of human resource information processing, natural language processing and artificial intelligence aided decision-making, and discloses a post-position matching method and system based on multi-agent collaborative reasoning. The method obtains post description, enterprise historical recruitment, team structure, post upstream and downstream cooperation, candidate data and feedback data, performs authorization verification, desensitization and sensitive attribute filtering on the candidate data; a post demand object is generated by a post analysis agent, a candidate portrait object is generated by a portrait construction agent, an initial matching result and an evidence object are generated by a matching reasoning agent; an evidence verification agent calculates an evidence correction factor, and combines a confidence and a time limit parameter to correct the matching result; a bounded quadratic matching is triggered through a shared state for a to-be-verified dimension, and finally a matching report is generated and the weight is updated according to the effective feedback.

Description

Technical Field

[0001] This invention relates to the fields of human resource information processing, natural language processing, and artificial intelligence-assisted decision-making, and in particular to a method and system for matching people to jobs based on multi-agent collaborative reasoning. Background Technology

[0002] With the development of natural language processing, text vector retrieval, and large language model technologies, intelligent job matching is increasingly being applied to scenarios such as recruitment screening, job recommendation, internal transfers, and talent inventory. Existing methods typically assess the suitability between candidates and positions through keyword matching, resume field rule filtering, text embedding similarity calculation, or overall scoring by a single model.

[0003] However, job requirements typically include not only explicit criteria such as education, years of experience, skills, and certifications, but also implicit or risk-related requirements such as teamwork style, job motivation, stability, stress tolerance, and willingness to travel. Candidate data may also come from various data sources, including resumes, certifications, project experience, assessment results, interview records, communication texts, and behavioral feedback. These different data sources may have issues such as insufficient evidence, conflicting information, time-lapse, or lack of traceability.

[0004] Existing job-person matching methods often rely on text similarity or a single model to implicitly capture the correspondence between jobs and candidates during training or inference. This makes it difficult to provide clear evidence to support each matching conclusion, and also to identify situations such as insufficient evidence, conflicting information, inconsistencies across requirements, and unreliable feedback quality. Therefore, in practical use, problems may arise such as similar wording but high job risk, high model scores without supporting evidence, and contradictory interpretations of the same candidate's characteristics across different requirement dimensions, affecting the reliability and auditability of the matching conclusions.

[0005] Therefore, there is a need for a person-job matching method and system that can structure the expression of job requirements, candidate profiles, matching evidence and feedback signals, and perform evidence constraints, confidence propagation and collaborative verification on the matching conclusions. Summary of the Invention

[0006] This invention aims to provide a method and system for matching job requirements and candidates based on multi-agent collaborative reasoning. By having a job analysis agent, a profile building agent, a matching reasoning agent, an evidence verification agent, and a job optimization agent collaboratively reason in a shared state, the method represents job requirements, candidate profiles, matching evidence, matching results, and feedback signals as structured objects. The method then corrects and outputs the job matching results through evidence correction, confidence propagation, bounded collaborative loops, feedback validity evaluation, and compliance processing.

[0007] To address the aforementioned technical problems, this invention provides a method for matching personnel to positions based on multi-agent collaborative reasoning, comprising: Acquire job description data, historical recruitment data of enterprises, team structure data, upstream and downstream collaboration data of positions, candidate data, and feedback data, and perform authorization verification, de-identification and anonymization, and sensitive attribute filtering on the candidate data to obtain a compliant input dataset.

[0008]

[0009] in, For job description data, For the company's historical recruitment data, For team structure data, For data on upstream and downstream collaboration in the job, For candidate data, For feedback data, For authorization verification function, This is a desensitization function. For filtering functions of sensitive attributes, Input the dataset for compliance.

[0010] Secondly, the job-side data in the compliant input dataset is input into the job parsing agent to generate a set of job demand objects.

[0011]

[0012] Each job requirement includes a requirement identifier, requirement type, requirement description, requirement weight, requirement confidence level, and requirement source, represented as follows:

[0013] in, For the first Target candidates for each position As a demand identifier, For demand type, This is a description of the requirements. As demand weight, For demand confidence level, This is the source of demand. Demand types include hard thresholds, capability requirements, scenario requirements, implicit requirements, and risk requirements.

[0014] Next, the candidate-side data input profiles in the compliant input dataset are used to construct an intelligent agent, generating a set of candidate profile objects.

[0015]

[0016] Each candidate profile includes a feature identifier, profile type, feature value, feature source, feature confidence level, feature timeliness parameter, and feature observation time, represented as:

[0017] in, For the first A candidate profile object, For feature identification, For portrait type, For eigenvalues, As a source of characteristics, For feature confidence, For characteristic time-sensitivity parameters, The characteristic observation time; the portrait types include explicit portrait, implicit portrait and dynamic portrait.

[0018] Then, the set of job requirements and the set of candidate profiles are input into the matching reasoning agent to generate an initial matching result and a set of evidence objects to support the initial matching result.

[0019]

[0020]

[0021] in, For the first The initial matching results for each job opening. The set of evidence objects supporting this initial matching result, For the number of pieces of evidence; for each piece of evidence It should include at least the related demand identifier, related feature identifier, evidence content, supporting strength, conflict marker, and traceability marker.

[0022] The matching reasoning includes semantic recall and keyword overlap calculation; the semantic recall includes: matching job requirements with... and candidate portrait objects Encode into vectors respectively and And calculate semantic similarity.

[0023]

[0024] in, For job requirements The text vector, Profiling candidates The text vector, For semantic similarity; The keyword overlap calculation includes: targeting job requirements. and candidate portrait objects Weighting of keywords in the computing field to determine their relevance.

[0025]

[0026] in, For job requirements The set of keywords in the target domain. Profiling candidates The set of keywords in the target domain. For the common keywords, Keywords Domain weights To increase the weighting of appropriateness, the weights of core job competency terms, business scenario terms, and risk terms are greater than the weights of general terms.

[0027] Subsequently, the initial matching results and the set of evidence objects are input into the evidence verification agent, which performs sufficiency verification, conflict verification, consistency verification, and cross-requirement consistency verification on the evidence objects, and calculates the evidence correction factor.

[0028]

[0029] in, For the first Evidence correction factor corresponding to the requirements of each job position As a factor for sufficiency of evidence, As a factor of evidence conflict, This is the evidence consistency factor.

[0030] Furthermore, regarding the target audience for the positions... Corresponding set of evidence Calculate the average support strength, the proportion of conflicting evidence, and the proportion of traceable evidence, and calculate the sufficiency factor, conflict factor, consistency factor, and evidence correction factor respectively.

[0031]

[0032]

[0033]

[0034]

[0035] in, For average support strength, For the proportion of conflicting evidence, The percentage of traceable evidence. To support the strength threshold. When the evidence set is empty, let .

[0036] Then, based on the confidence level of the job requirements, the confidence level of the candidate profile, the feature timeliness parameter, and the evidence correction factor, the matching confidence level is calculated and the initial matching result is corrected to obtain the final matching result.

[0037]

[0038]

[0039] in, The base matching confidence is derived from the propagation of demand confidence and feature confidence. To match the confidence level, For the initial matching results, This is the final matching result.

[0040] The basic matching confidence is derived from the demand confidence of the job requirements and the feature-effective confidence of the candidate profile after time-adjusted verification; based on the candidate profile... The observation time and characteristic time-dependent parameters are used to calculate the characteristic age, time decay factor, and effective confidence level of the characteristic.

[0041]

[0042]

[0043]

[0044]

[0045] in, For the current matching time, For the characteristic observation time, For time unit conversion parameters, Characteristic age, For characteristic time-sensitivity parameters, The time decay factor, The effective confidence level of the feature after time-adjusted. Based on the confidence level of the matching.

[0046] When the evidence object does not exist, there is a conflict in the evidence, the evidence is untraceable, there is inconsistency across requirements, or the evidence correction factor is lower than the preset threshold, the corresponding matching dimension is marked as a dimension to be verified. When there is a dimension to be verified and the current collaborative round has not reached the preset maximum round, the dimension to be verified and its cause are written into the shared state, and the matching reasoning agent re-executes the secondary matching based on the shared state.

[0047]

[0048] in, The set of job requirements to be verified. In a shared state, This is a secondary matching result; the secondary matching result replaces the original matching result only when the amount of evidence increases or the strength of the evidence is improved.

[0049] Finally, the final matching results are grouped according to six dimensions: hard requirements, abilities, scenarios, motivation, potential, and risks, and the overall match between candidates and positions is calculated.

[0050]

[0051] in, Scores are given for each of the following dimensions: hard skills, abilities, context, motivation, potential, and risk. These are the weights for the corresponding dimensions. For the overall matching score, the risk dimension participates in the calculation of the overall matching score with negative weights.

[0052] The total matching score, scores for each dimension, final matching result, evidence chain, risk points, and dimensions to be verified are used to generate a matching report.

[0053]

[0054] in, Person-job matching report For the score set of each dimension, For the final set of matching results, This is the chain of evidence.

[0055] The weights of job requirements or candidate profiles are updated based on the effectiveness of subsequent feedback signals.

[0056]

[0057] in, For the weights before the update, To provide feedback on the target weight, Based on the learning rate, To provide feedback on effectiveness, This is the updated weight.

[0058] This invention also provides a job matching system based on multi-agent cooperative reasoning, comprising: The data access module is used to acquire job description data, historical recruitment data of enterprises, team structure data, upstream and downstream collaboration data of positions, candidate data, and feedback data; The compliance module is used to perform authorization verification, de-identification and masking, sensitive attribute filtering, and audit log generation on candidate data; The job analysis agent is used to generate a set of job requirement objects based on job-side data. A profile building agent is used to generate a collection of candidate profile objects based on candidate-side data. A matching reasoning agent is used to generate an initial matching result and a set of evidence objects based on the set of job requirement objects and the set of candidate profile objects. The evidence verification agent is used to perform sufficiency verification, conflict verification, consistency verification, and cross-requirement consistency verification on evidence objects, and generate evidence correction factors, matching confidence scores, and final matching results. The job optimization agent is used to generate a matching report based on the total matching degree, scores of each dimension, final matching results, evidence chain, risk points, and dimensions to be verified. The feedback update module is used to update the job requirement weight or candidate profile weight based on the effectiveness of the feedback. An orchestrator is used to organize the job analysis agent, profile building agent, matching reasoning agent, evidence verification agent, job optimization agent, and feedback update module to perform the above methods through shared state; Specifically, when there is a dimension to be verified and the current collaborative round has not reached the preset maximum round, the orchestrator controls the matching reasoning agent to perform a secondary matching, and replaces the original matching result only when the number of evidences increases or the strength of evidence support increases.

[0059] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: By recording job requirements, candidate profiles, and evidence objects through structured objects, the matching results can be associated with evidence sources; by correcting the matching results through evidence correction factors and feature timeliness parameters, matching dimensions with insufficient evidence, conflicting evidence, untraceable evidence, or inconsistencies across requirements are placed into a state of pending verification; by performing secondary matching on dimensions to be verified through bounded collaborative loops, the impact of matching results lacking evidence support on the overall matching degree is reduced; and by updating the weights after evaluating the effectiveness of feedback, the impact of low-quality feedback on subsequent matching results is reduced. Attached Figure Description

[0060] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 A flowchart of the person-job matching method based on multi-agent collaborative reasoning provided in this embodiment of the invention; Figure 2 This is a structural block diagram of a human-job matching system based on multi-agent collaborative reasoning provided in an embodiment of the present invention. Figure 3 A flowchart of evidence verification and bounded quadratic matching provided for embodiments of the present invention. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] This embodiment uses the "B2B SaaS Sales Manager" position in enterprise recruitment as an example to illustrate the method of the present invention. The system first accesses job description data, historical recruitment data of the enterprise, team structure data, upstream and downstream collaboration data for the position, candidate data, and feedback data. Specifically, the job description data includes the job title, job responsibilities, job requirements, sales scenarios, customer types, and performance targets; the historical recruitment data includes historical hiring records, interview evaluations, reasons for offer rejection, reasons for leaving, and probationary period results; the team structure data includes team structure information and job collaboration relationships; the upstream and downstream collaboration data includes collaboration relationships with pre-sales, implementation, customer success, product, and marketing departments; the candidate data includes resumes, certificates, project experience, assessment results, interview records, communication texts, and behavioral feedback; and the feedback data includes interview feedback, hiring feedback, reasons for rejection feedback, and subsequent performance feedback.

[0063] Before data enters the matching process, the system performs authorization verification on candidate data through the compliance module to confirm that the candidate data can be used for job matching, recruitment assessment, or subsequent feedback updates. For personally identifiable information such as name, mobile phone number, email address, and ID number, the system performs de-identification and anonymization processing; for sensitive attributes such as gender, age, ethnicity, household registration, marital status, and religion, the system performs sensitive attribute filtering to prevent them from entering the job analysis, profile building, matching reasoning, and scoring stages. After the above processing, the system forms a compliant input dataset and records the data source, processing time, and audit flags for data entering subsequent processing.

[0064] In one specific embodiment, this method employs a multi-agent state orchestration mechanism, organizing the job analysis agent, profile building agent, matching reasoning agent, evidence verification agent, and job optimization agent into a collaborative reasoning process. Each agent exchanges structured objects through a shared state. These structured objects include job requirement objects, candidate profile objects, evidence objects, matching result objects, feedback signal objects, causes to be verified, and audit logs. The orchestrator schedules the operation of each agent in the order of job analysis, profile building, matching reasoning, evidence verification, collaborative loop, job optimization, and feedback update. When the evidence verification result indicates insufficient evidence, evidence conflict, untraceable evidence, or inconsistent interpretation across requirements, the orchestrator writes the corresponding dimension into the fields to be verified in the shared state and controls the process to enter a secondary matching loop.

[0065] During the job requirement construction phase, the job analysis agent receives job description data, historical recruitment data from the company, team structure data, and upstream and downstream collaboration data for the job, breaking down the job requirements into multiple job requirement objects. Each job requirement object includes at least a requirement identifier, requirement type, requirement description, requirement weight, requirement confidence level, and requirement source. For requirements such as "Bachelor's degree or above," "more than three years of enterprise software sales experience," and "familiarity with CRM or ERP software sales processes," which can be directly obtained from the job description, the job analysis agent marks them as hard thresholds or competency requirements and records their source as job description data. For requirements such as "ability to conduct long-term follow-up with key account (KA) clients" and "ability to collaborate with pre-sales, implementation, and customer success teams to advance projects," the job analysis agent marks them as competency requirements or scenario requirements and records their source as job description data and upstream and downstream collaboration data for the job. For requirements such as "stress resistance," "stability," "acceptance of business trips," and "willingness to follow up with long-term clients," which may not be explicitly stated in the job description but can be inferred from historical hiring records, reasons for leaving previous jobs, reasons for rejecting offers, or reasons for failing the probationary period, the job analysis AI marks them as implicit or risky requirements and records the source of the inference and the confidence level of the requirement. The confidence level of requirements directly extracted from job description data is higher than the confidence level of implicit requirements inferred from historical recruitment data, team structure data, or upstream and downstream collaboration data.

[0066] During the candidate profile construction phase, the profile building agent constructs multiple candidate profile objects from candidate data. Explicit profiles are directly extracted from resumes, certificates, educational backgrounds, work experience, and project experience, describing the candidate's objective experience and hard skills. Implicit profiles are inferred from assessment results, interview records, communication texts, and behavioral data, describing the candidate's communication style, collaboration methods, job motivation, stress tolerance, and risk appetite. Dynamic profiles are generated from recent job-seeking intentions, job preferences, application behavior, communication feedback, and interview feedback, and decay or update over time. When candidate data includes MBTI personality types or Holland Occupational Interests (OOC) assessment results, the system maps the assessment results to a set of implicit profile features using an assessment mapping function. For example, the extroversion / introversion, sensing / intuition, thinking / feeling, and judging / perceiving dimensions in the MBTI personality type are mapped to communication orientation, information processing style, decision-making style, and execution preference characteristics, respectively; the Holland OOC type is mapped to occupational interest, job motivation, or occupational inclination characteristics. The implicit profile obtained from the assessment mapping is only used as an auxiliary feature. Its feature confidence is lower than that of the explicit profile features obtained directly from resumes, certificates or project experience, and corresponding feature timeliness parameters are set.

[0067] In the initial matching and inference phase, the matching and inference agent uses job requirement objects as the query unit and retrieves relevant profile objects from the candidate profile object set. For each job requirement object, the matching and inference agent first calculates semantic similarity based on the job requirement description and candidate profile feature values, then calculates weighted similarity based on job domain keywords, skill terms, business scenario terms, and risk terms, and obtains the initial matching result based on semantic similarity and keyword weighted similarity. The weights of core job skill terms, business scenario terms, and risk terms are greater than the weights of general terms.

[0068] For example, for a job requirement of "having CRM or ERP software sales experience," the matching inference agent retrieves explicit profile objects such as "the candidate was responsible for CRM customer development" or "the candidate has three years of SaaS sales experience" from the candidate profile object set. It then generates the corresponding initial matching result and simultaneously generates an evidence object. The related requirement identifier of this evidence object points to the requirement object of "having CRM or ERP software sales experience," the related feature identifier points to the relevant candidate profile object, the evidence content records the corresponding fragments from the resume or project experience, the support strength is determined according to the relevance of the evidence to the requirement, the conflict marker is initially set to no conflict, and the traceability marker points to the original resume or project experience source.

[0069] Regarding the "acceptance of frequent business trips" for job requirements, if the matching inference agent retrieves information from communication feedback indicating that "the candidate prefers local positions" or "does not accept long-term frequent business trips," then it generates evidence objects corresponding to the risk requirements. These evidence objects are used to support the risk dimension judgment, not as positive evidence for sales experience matching. Therefore, the same candidate may have a high matching result in the sales experience dimension, but generate a deduction basis in the risk dimension.

[0070] During the evidence verification phase, the evidence verification agent reviews each initial matching result and its set of evidence objects. First, the agent calculates the average support strength of the evidence objects under a specific job requirement to determine the sufficiency of the evidence. Second, the agent calculates the proportion of conflicting evidence to determine if there are contradictions between different data sources. For example, if a resume shows a candidate has long been engaged in sales to corporate clients, but communication feedback indicates they recently explicitly rejected a corporate client position, this evidence can be marked as conflicting evidence. Third, the agent calculates the proportion of traceable evidence to determine if the evidence can be traced back to its original data source. Based on this, the system generates an evidence correction factor.

[0071] When the set of evidence objects is empty, the evidence correction factor is set to 0, and the corresponding matching dimension must not form a high-confidence matching conclusion. When evidence objects exist but the proportion of conflicting evidence is high, the proportion of traceable evidence is low, or there is inconsistency in cross-demand interpretation, the evidence correction factor is lowered. The system uses demand confidence, feature confidence, feature timeliness parameters, and evidence correction factor to calculate the matching confidence, and corrects the initial matching results through the evidence correction factor, so that conclusions with insufficient evidence, conflicting evidence, or untraceable evidence do not have an excessive impact on the overall matching score.

[0072] For dynamic profiles, the system performs time decay based on the interval between the feature observation time and the current matching time, reducing the influence of older job intentions, communication feedback, or behavioral preferences on the current match. The final matching result is obtained by correcting the initial matching result with an evidence correction factor. When the evidence object does not exist, the evidence is conflicting, the evidence is untraceable, there is inconsistency across requirements, or the evidence correction factor is lower than a preset threshold, the corresponding matching dimension is marked as a dimension to be verified.

[0073] In the secondary matching phase of the bounded collaborative loop, when there is a dimension to be verified and the current collaborative round has not reached the preset maximum round, the orchestrator writes the dimension to be verified and its cause into the shared state. For example, the cause to be verified can be "no supporting evidence," "evidence conflict," "evidence untraceable," or "inconsistent interpretation across requirements." After reading the above causes, the matching inference agent only re-searches for candidate profile evidence for the dimension to be verified and performs secondary matching. Secondary matching must not fabricate evidence; the original matching result is only replaced when the newly added evidence has a traceable marker and the number of evidences increases or the strength of evidence support improves; otherwise, the original conclusion is retained and the state to be verified is maintained. This loop ensures the convergence of the inference process through the constraints of the preset maximum round, non-backback merging, and no fabrication of evidence.

[0074] In the multi-dimensional scoring and report output stage, the system groups the final matching results according to six dimensions: hard skills, capabilities, scenario, motivation, potential, and risk. It calculates the score for each dimension and then aggregates them according to their respective weights to obtain the total matching degree. The risk dimension participates in the total matching degree calculation with negative weights, reducing the contribution of low-confidence or unverified conclusions to the dimension score. Based on the total matching degree, scores for each dimension, the final matching results, the evidence chain, risk points, and unverified dimensions, the job optimization agent generates recommendation conclusions, evidence explanations, risk warnings, interview follow-up questions, job description optimization suggestions, and decision support statements.

[0075] During the feedback update phase, after receiving at least one of the following: interview feedback, hiring feedback, rejection reason feedback, or performance feedback, the system first calculates the feedback validity based on sample size, source reliability, and historical consistency. Then, based on the feedback validity, the baseline learning rate, the weight before the update, and the feedback target weight, the system weights the job requirement weights or the candidate profile weights are weighted and updated. This reduces the contamination of subsequent matching results by low-quality or incidental feedback.

[0076] The core of this embodiment lies in breaking down the job matching task into five collaborative sub-tasks: job analysis, profile construction, matching reasoning, evidence verification, and job optimization. A structured reasoning loop is achieved through shared state. Compared to methods that rely solely on keywords, vector similarity, or a single model for overall scoring, this invention can explicitly model implicit job requirements and risk requirements, ensuring that each matching conclusion is supported by a chain of evidence. Through evidence correction and confidence propagation, it avoids drawing overly strong conclusions when evidence is insufficient. It improves the reliability and interpretability of matching results by performing secondary verification on the dimensions to be verified through bounded collaborative loops. Through feedback validity assessment, sensitive attribute filtering, de-identification and anonymization processing, and audit records, the system achieves stable, traceable, and compliant job matching in scenarios such as recruitment screening, job recommendation, internal transfers, and talent inventory.

[0077] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for matching people to positions based on multi-agent collaborative reasoning, characterized in that, include: Acquire job description data, historical recruitment data of enterprises, team structure data, upstream and downstream collaboration data of positions, candidate data, and feedback data, and perform authorization verification, de-identification and anonymization, and sensitive attribute filtering on the candidate data to obtain a compliant input dataset; The job-side data in the compliance input dataset is input into the job parsing agent to generate a set of job requirement objects. The job requirement objects include requirement identifier, requirement type, requirement description, requirement weight, requirement confidence and requirement source. The requirement type includes hard threshold, ability requirement, scenario requirement, implicit requirement and risk requirement. The candidate-side data in the compliant input dataset is used to construct an intelligent agent for the profile, generating a set of candidate profile objects. The candidate profile objects include feature identifiers, profile types, feature values, feature sources, feature confidence levels, feature timeliness parameters, and feature observation times. The profile types include explicit profiles, implicit profiles, and dynamic profiles. The shared state organizes the job analysis agent, profile building agent, matching reasoning agent, evidence verification agent, and job optimization agent to exchange structured objects, which include job requirement objects, candidate profile objects, evidence objects, matching result objects, and feedback signal objects. The set of job requirements and the set of candidate profiles are input into the matching reasoning agent to generate an initial matching result and a set of evidence objects to support the initial matching result. The evidence objects include at least the associated requirement identifier, associated feature identifier, evidence content, support strength, conflict marker, and traceability marker. The initial matching results and the set of evidence objects are input into the evidence verification agent, which performs sufficiency verification, conflict verification, consistency verification and cross-requirement consistency verification on the evidence objects, and calculates the evidence correction factor based on the evidence sufficiency factor, evidence conflict factor and evidence consistency factor. Based on the confidence level of the job requirements, the confidence level of the candidate profile, the feature timeliness parameter, and the evidence correction factor, the matching confidence level is calculated and the initial matching result is corrected to obtain the final matching result. When the evidence object does not exist, there are conflicts in the evidence, the evidence is not traceable, there is inconsistency across requirements, or the evidence correction factor is lower than the preset threshold, the corresponding matching dimension will be marked as a dimension to be verified. When there is a dimension to be verified and the current collaborative round has not reached the preset maximum round, the dimension to be verified and its cause are written into the shared state, and the matching reasoning agent re-executes the secondary matching based on the shared state. The result of the secondary matching only replaces the original matching result when the number of evidences increases or the strength of evidence support increases. The final matching results are grouped according to six dimensions: hardness, ability, scenario, motivation, potential and risk. The overall match between candidates and positions is calculated, and a matching report is generated based on the overall match, scores of each dimension, final matching results, evidence chain, risk points and dimensions to be verified. The weights of job requirements or candidate profiles are updated based on the effectiveness of subsequent feedback signals.

2. The method as described in claim 1, characterized in that, The compliant input dataset is represented as follows: ； in, For job description data, For the company's historical recruitment data, For team structure data, For data on upstream and downstream collaboration in the job, For candidate data, For feedback data, For authorization verification function, This is a desensitization function. For filtering functions of sensitive attributes, Input dataset for compliance; The set of job requirements is represented as follows: ； The required candidates for each position are represented as follows: ； in, For the first Target candidates for each position As a demand identifier, For demand type, This is a description of the requirements. As demand weight, For demand confidence level, As a source of demand; The set of candidate profile objects is represented as follows: ； Each candidate profile is represented as follows: ； in, For the first A candidate profile object, For feature identification, For portrait type, For eigenvalues, As a source of characteristics, For feature confidence, For characteristic time-sensitivity parameters, The time of the characteristic observation.

3. The method as described in claim 1, characterized in that, The job analysis agent includes an explicit requirement extraction unit and an implicit risk inference unit; The explicit requirement extraction unit is used to extract hard thresholds, ability requirements, and scenario requirements from job description data; The implicit risk inference unit is used to infer implicit requirements and risk requirements based on the company's historical recruitment data, commonalities among high-performing employees in the same position, team structure data, upstream and downstream collaboration relationships of the position, reasons for past departures, reasons for rejecting offers, or reasons for failing the probation period, and to mark the source and confidence level of the implicit requirements and risk requirements. Among them, the confidence level of explicit requirements directly extracted from job description data is higher than that of implicit requirements inferred from historical recruitment data, team structure data, or upstream and downstream collaboration data of the job.

4. The method as described in claim 1, characterized in that, The explicit profile includes features directly extracted from resumes, certificates, project experience, educational background, or work experience; The implicit profile includes abilities, motivations, personality traits, collaboration styles, or risk preferences inferred from assessment results, interview records, communication texts, or behavioral data. The dynamic profile includes the candidate's recent intentions, job-seeking status, job preferences, behavioral feedback or communication feedback, and decays or updates over time based on the feature timeliness parameters; When candidate data includes MBTI personality type or Holland Occupational Interests Assessment results, a set of latent profile features is generated using the assessment mapping function: ； in, To evaluate the mapping function, For personality type assessment results, Based on Holland's career interest assessment results, This is a set of latent profile features generated from the assessment results; The latent profile features obtained by mapping the results of psychological assessments are assigned a lower feature confidence level than the explicit profile features, and corresponding feature timeliness parameters are set.

5. The method as described in claim 1, characterized in that, The matching reasoning includes semantic recall and keyword overlap calculation; The semantic recall includes: retrieving job demand objects. and candidate portrait objects Encode into vectors respectively and And calculate semantic similarity: ； in, For job requirements The text vector, Profiling candidates The text vector, For semantic similarity; The keyword overlap calculation includes: targeting job requirements. and candidate portrait objects Weighted relevance of keywords in the computing domain: ； in, For job requirements The set of keywords in the target domain. Profiling candidates The set of keywords in the target domain. For the common keywords, Keywords Domain weights To increase the weighting of the degree of fit; Among them, the weight of core job competency terms, business scenario terms, and risk terms is greater than that of general terms.

6. The method as described in claim 1, characterized in that, The calculation of the evidence correction factor includes: Target candidates for the position Corresponding set of evidence Calculate the average support strength, the proportion of conflicting evidence, and the proportion of traceable evidence, and then calculate the sufficiency factor, conflict factor, consistency factor, and evidence correction factor respectively: ； ； ； ； in, For average support strength, For the proportion of conflicting evidence, The percentage of traceable evidence. To support the strength threshold, For the first Evidence correction factor corresponding to the requirements of each job position As a factor for sufficiency of evidence, As a factor of evidence conflict, As a factor of consistency of evidence; When the set of evidence is empty, let .

7. The method as described in claim 1, characterized in that, The basic matching confidence is obtained by propagating the demand confidence of the job requirements and the feature effective confidence of the candidate profile after time-revision correction. Based on the candidate's portrait Given the observation time and characteristic time-sensitivity parameters, calculate the characteristic age, time decay factor, and effective confidence level of the characteristic: ； ； ； ； in, For the current matching time, For the characteristic observation time, For time unit conversion parameters, Characteristic age, For characteristic time-sensitivity parameters, The time decay factor, The effective confidence level of the feature after time-adjusted. Based on the confidence level of the matching; The matching confidence level is obtained based on the basic matching confidence level and the evidence correction factor: ； The initial matching results are then corrected based on the evidence correction factor to obtain the final matching results: ； in, Based on the job requirements The base matching confidence level corresponding to the associated candidate profile objects is determined. To match the confidence level, For the initial matching results, This is the final matching result.

8. The method as described in claim 1, characterized in that, The shared state is used to save job requirement objects, candidate profile objects, evidence objects, matching result objects, feedback signal objects, causes to be verified, and audit logs; The job analysis agent, profile building agent, matching reasoning agent, evidence verification agent, and job optimization agent exchange structured objects through the shared state; The causes of the dimension to be verified include at least one of the following: lack of supporting evidence, conflicting evidence, untraceable evidence, or inconsistent interpretation across requirements. The secondary matching is controlled and executed by the orchestrator and is subject to the constraints of a preset maximum number of rounds, non-back-off merging, and no fabricated evidence. The non-back-down merging refers to retaining the original matching result when the secondary matching result does not increase the number of evidences or improve the strength of the evidence support. The aforementioned non-fictional evidence refers to the fact that the new evidence on which the secondary matching results are based has a traceable marker and can be linked to the candidate profile object, the job requirement object, or the original data source.

9. The method as described in claim 1, characterized in that, The total matching degree is expressed as: ； in, Scores are given for each of the following dimensions: hard skills, abilities, context, motivation, potential, and risk. These are the weights for the corresponding dimensions. Total match score; The feedback signal includes at least one of interview feedback, hiring feedback, rejection reason feedback, or performance feedback; Before updating the job demand weight or candidate profile weight based on the feedback signal, the validity of the feedback is calculated based on the sample size, source reliability, and historical consistency. The update of the job requirement weight or candidate profile weight is represented as follows: ； in, For the weights before the update, To provide feedback on the target weight, Based on the learning rate, To provide feedback on effectiveness, The updated weights; The matching report includes recommendation conclusions, evidence chains, risk warnings, follow-up interview questions, suggestions for optimizing job descriptions, and statements to support decision-making.

10. A person-job matching system based on multi-agent collaborative reasoning, characterized in that, include: The data access module is used to acquire job description data, historical recruitment data of enterprises, team structure data, upstream and downstream collaboration data of positions, candidate data, and feedback data; The compliance module is used to perform authorization verification, de-identification and masking, sensitive attribute filtering, and audit log generation on candidate data; The job analysis agent is used to generate a set of job requirement objects based on job-side data. A profile building agent is used to generate a collection of candidate profile objects based on candidate-side data. A matching reasoning agent is used to generate an initial matching result and a set of evidence objects based on the set of job requirement objects and the set of candidate profile objects. The evidence verification agent is used to perform sufficiency verification, conflict verification, consistency verification, and cross-requirement consistency verification on evidence objects, and generate evidence correction factors, matching confidence scores, and final matching results. The job optimization agent is used to generate a matching report based on the total matching degree, scores of each dimension, final matching results, evidence chain, risk points, and dimensions to be verified. The feedback update module is used to update the job requirement weight or candidate profile weight based on the effectiveness of the feedback. An orchestrator is used to organize the job parsing agent, the profile building agent, the matching reasoning agent, the evidence verification agent, the job optimization agent, and the feedback update module to perform the method of any one of claims 1 to 9 by sharing a state; Specifically, when there is a dimension to be verified and the current collaborative round has not reached the preset maximum round, the orchestrator controls the matching reasoning agent to perform a secondary matching, and replaces the original matching result only when the number of evidences increases or the strength of evidence support increases.

Images