LLM-based multi-agent professional path planning method and system

Abstract

This invention discloses a career path planning method and system based on LLM (Multi-Agent Learning). The method first generates a set of target constraints through intent and constraint parsing and requirement clarification. Then, it constructs individual ability profiles and job-specific quantitative profiles. Next, based on ability gaps, it generates task packages level by level along stage boundaries and supplements prerequisite dependencies, forming a set of candidate task packages. Finally, addressing the problem that traditional sorting or traversal methods struggle to balance search efficiency and path quality due to the large candidate task package combination space, complex dependencies, and coexisting time and risk constraints, an improved Monte Carlo tree search algorithm is employed to filter invalid search branches in real time, efficiently searching and outputting career goal paths with better comprehensive scores. This invention improves the search efficiency, path quality, and interpretability of career goal path generation under complex constraints and can be widely applied to career development scenarios such as internships, campus recruitment, social recruitment, and entrepreneurship.

Description

Technical Field

[0001] This invention relates to the field of intelligent career planning technology, and in particular to a career path planning method and system based on LLM multi-agent. Background Technology

[0002] With increasingly diverse career development scenarios such as internships, campus recruitment, social recruitment, and entrepreneurship, users typically need to consider multiple constraints, including target positions, location, salary, time budget, and risk appetite, when formulating career plans to create a structured development path from their current skill level to the requirements of their target position. However, existing career planning tools mainly focus on job recommendations, course recommendations, or experience suggestions, making it difficult to provide refined and actionable paths to specific career goals.

[0003] Currently, career planning methods can be mainly categorized as follows: planning methods based on static rules or expert experience, recommendation methods based on job text matching and competency gap analysis, and path planning methods combining knowledge graphs, graph structure modeling, or search strategies. While these methods have improved the targeting and personalization of recommendations to some extent, they still have the following shortcomings when dealing with complex career goal scenarios:

[0004] First, existing methods generally lack a unified benchmark for quantifying capabilities. Users' current capabilities are usually derived from resumes, project experience, course experience, or personal descriptions, while job requirements are derived from job descriptions, recruitment requirements, or industry knowledge. There are significant differences between the two in terms of source, expression, and evaluation criteria, making it difficult to establish a consistent and comparable alignment and mapping relationship. This results in inaccurate identification of capability gaps, which in turn affects the determination of the starting point and the decomposition of goals in path planning.

[0005] Second, existing methods mostly generate recommended content directly based on ability items, course items, or experience items, lacking a mechanism to break down tasks step by step according to the boundaries of ability stages. This makes the recommended tasks too coarse-grained, making it difficult to reflect the progressive pattern of ability improvement and the stage characteristics of the learning path, which is not conducive to users' orderly progress in stages.

[0006] Third, existing methods are insufficient in modeling the pre-dependent relationships between tasks, making it difficult to balance the executability of task packages with the overall rationality of path arrangement. Especially when there are a large number of candidate task packages, complex dependencies, and both time and risk constraints, existing methods often employ simple sorting, rule-based filtering, or inefficient traversal strategies, making it difficult to efficiently search for career goal paths with better overall scores based on a comprehensive consideration of benefits, time consumption, and risks.

[0007] In summary, there is an urgent need to propose a new method for generating career goal paths, in order to establish a consistent quantitative representation between users' current abilities and job requirements, to achieve step-by-step decomposition of task packages according to stage boundaries, and to efficiently search for career goal paths with better comprehensive scores under time and risk constraints. Summary of the Invention

[0008] The problem to be solved by this invention is to provide a career path planning method and system based on LLM multi-agent, which solves the technical problems in the prior art such as the lack of a unified quantitative benchmark between the user's current ability and job requirements, the lack of a step-by-step splitting mechanism along the ability stage boundary in the task generation process, and the difficulty in efficiently searching for a career goal path with better comprehensive score under the conditions of complex task dependencies and the coexistence of time constraints and risk constraints.

[0009] This invention adopts the following technical solution: a career path planning method based on LLM multi-agent, comprising:

[0010] The agent receives the user's natural language description of occupational needs through intent and constraint parsing, and obtains the structured target object and constraint object.

[0011] The structured target object and constraint object are verified by a constraint clarification agent to determine the target constraint set;

[0012] The personal profile intelligent agent processes users' personal information to generate a personal capability profile;

[0013] The job profiling agent retrieves the target job description text from at least one job data source based on the target job identifier, target region identifier, and target salary range parameters, and generates a quantitative job profile.

[0014] Based on the individual ability profile and job quantitative profile, an intelligent agent is constructed by using task packages to build a set of candidate task packages;

[0015] The MCTS path planning agent constructs a candidate state tree based on the candidate task package set, and performs an improved Monte Carlo tree search under time and risk constraints to output a career goal path with a better comprehensive score.

[0016] The characteristic feature is that the constraint clarification agent performs integrity verification, conflict verification, and rationality verification; wherein, integrity verification is used to determine whether the structured target object and constraint object contain necessary fields, conflict verification is used to determine whether there is a contradiction between the constraints, and rationality verification is used to determine whether the value of the constraint is within a preset reasonable range.

[0017] Optionally, the personal profile intelligent agent generates a personal capability profile using a "feature extraction—stage determination—numerical mapping—risk quantification" approach, specifically including:

[0018] Feature extraction: Extract structured capability feature descriptions by calling a large language model;

[0019] Stage Determination: Determine the current capability stage based on preset capability stage division rules;

[0020] Numerical mapping: Determining the current capability value within the corresponding numerical range using a machine learning model;

[0021] Risk quantification: Based on the user's personality test information and risk preference parameters, risk is quantified, risk coefficients are calculated, and a personal ability profile is generated.

[0022] Optionally, the preset capability stage division rules are predetermined by domain experts and are used to define multiple progressive capability stages and corresponding characteristics and numerical ranges for each capability item, so as to establish a comparable benchmark between individual capabilities and job requirements.

[0023] The risk coefficient serves as the risk constraint threshold in the MCTS path planning agent, satisfying the following:

[0024]

[0025] in, Indicates the risk coefficient. This represents the comprehensive risk aversion index, calculated using a preset weighted model based on user-provided personality test information. Indicates the calibration coefficient; This represents the mapping value of the risk preference parameter.

[0026] Optionally, the job profile intelligent agent processes the following:

[0027] Based on the target job identifier, target region identifier, and target salary range parameters, retrieve job description text from at least one job data source; identify and extract the skill requirements, experience requirements, and general competency requirements required for the job.

[0028] The job requirement is determined to be in a certain stage based on the preset capability stage division rules, and the upper limit of the corresponding stage value range is used as the target value of the job requirement.

[0029] At the same time, for each competency item, the number of texts in the retrieved job description texts that identify it as a required or important competency for the job is counted, and the proportion of the number of such texts to the total number of retrieved job description texts is used as the importance weight of the competency item to generate a quantitative job profile.

[0030] Optionally, the task package constructs an intelligent agent, and the processing includes:

[0031] Compare individual competency profiles with job-specific quantitative profiles, calculate the competency gaps for each competency item, and identify competency items that need improvement.

[0032] For each skill to be improved, based on its current skill value and the target value required for the position, and in conjunction with the preset skill stage division rules, the stage boundary between the two is identified.

[0033] An initial task package is generated for each of the aforementioned stage boundaries, so that the capability enhancement process can proceed step by step along the stages.

[0034] Optionally, after generating the initial task package, the task package building agent determines the corresponding set of prerequisite dependencies for each initial task package based on the capability dependency graph, forming a set of candidate task packages;

[0035] The capability dependency graph is a pre-defined knowledge structure with capability items as nodes and pre-existing dependencies between capabilities as directed edges. Each candidate task package includes at least: capability items to be improved, job requirement target values, starting capability values ​​or stage starting point values, stage target capability values, recommended learning content, basic estimated time consumption, and a set of pre-existing dependencies.

[0036] Optionally, the MCTS path planning agent processes the following:

[0037] Construct the search root node with an initial state, which includes at least the current capability state, the set of satisfied capability items, an empty sequence of executed task packages, the cumulative revenue with an initial value of 0, the cumulative time consumed, and the cumulative risk.

[0038] Based on the initial state and preset constraints, the improved Monte Carlo tree search steps are executed cyclically to output a path with a better overall score.

[0039] The improved Monte Carlo tree search includes the following steps in sequence:

[0040] (1) Node recursive selection: Starting from the root node, the improved UCT formula is used to recursively select child nodes until the node that is not fully expanded is reached;

[0041] (2) Thought decision: The large language model calculates and judges whether the dependency coefficient of the task package to be expanded meets the standard based on the current node state, the stage progression rules and the coverage of the prerequisites, and at the same time judges whether the stage prerequisites are met and whether the time and risk constraints are met, and gives the current optimal choice.

[0042] (3) Action expansion: Based on the decision-making process and verification results, expand the task package and generate new child nodes;

[0043] (4) Path simulation: Starting from the new child node, based on the stage progression rules and prerequisite dependency conditions, select compliant task packages for constrained random walks, and deduce until the path ends;

[0044] (5) Multi-dimensional scoring: Calculate the comprehensive score of the path's benefits, time consumption, and risks, and combine it with the path rationality score and the confidence score of goal achievement from the LLM assessment to obtain the final return;

[0045] (6) Value feedback: The final reward is propagated back along the search path to update the number of visits and cumulative reward of each node.

[0046] Optionally, the recursive selection of nodes adopts an improved UCT rule, wherein the improved UCT value satisfies:

[0047] ;

[0048] in, Represents a node Cumulative returns Represents a node Number of visits, Represents a node The number of visits to the parent node. Indicates the exploration coefficient;

[0049] The MCTS path planning agent calculates the prerequisite dependency coverage rate based on the prerequisite dependency set of the task package and the set of capabilities already satisfied by the current node. The dependency coefficient is obtained by combining the coverage of prior dependencies and the degree of ability mastery with a large language model. ;

[0050] The dependency coefficient is used to dynamically adjust the revenue value, standardized time consumption estimate, and risk value of the candidate task package, as shown in the following formula:

[0051] ;

[0052] in, Indicates the first The revenue value of each task package, Indicates the first The importance weight of each ability in the job quantitative profile Indicates the first The task package is for the first The extent to which the ability of each skill is improved; Indicates the first Standardized time estimate for each task package Indicates the basic estimated time. Indicates the total time budget. Indicates the time-consuming adjustment weight; Indicates the first Risk value of each task package This represents the stage target capability value. This indicates the current ability value. This indicates the smallest positive number whose denominator is zero. Indicates the risk-adjusted weight;

[0053] After the simulation phase ends, a comprehensive score is calculated based on the complete path from the root node to the simulation termination node. As the base return value, combined with the path rationality score of the LLM assessment. Confidence score of achieving the goal Weighted average yields the final return. ,in:

[0054] ;

[0055] in, and All represent preset weights;

[0056] ;

[0057] This updates the number of visits and cumulative rewards for each node.

[0058] The present invention also provides: an LLM-based multi-agent career path planning system for implementing the aforementioned method, comprising:

[0059] Input layer: Inputs user's career needs, personal information, and job data.

[0060] Multi-agent collaborative processing layer: includes an intent and constraint parsing agent, a constraint clarification agent, a personal profile agent, a job profile agent, a task package construction agent, and an MCTS path planning agent; these are respectively used to obtain structured target objects and constraint objects, determine the target constraint set, generate personal ability profiles, generate job quantitative profiles, construct candidate task package sets, and perform Monte Carlo tree search and job objective path planning under time constraints and risk constraints;

[0061] Output layer: Outputs intermediate results and a better career goal path with a comprehensive score. The intermediate results include: a set of goal constraints, an individual ability profile, a job quantitative profile, and a set of candidate task packages.

[0062] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0063] 1. The method of this invention can establish a consistent quantitative representation between the user's current ability and job requirements, thereby improving the accuracy of ability gap identification; it can generate task packages step by step according to the boundary of ability stages, thereby enhancing the progressiveness of the ability improvement process and the operability of task execution; it can improve the rationality of task package combination by combining ability dependency relationship graphs; at the same time, by participating in dependency coefficient evaluation and path rationality judgment through a large language model, it enhances the interpretability of the path planning process and results, so that each step of the career development path selection has a clear basis and explanation.

[0064] 2. This invention is based on the constrained pruning Monte Carlo tree search algorithm. By performing risk and time constraint filtering in real time during the search process, it significantly compresses invalid search branches, thereby achieving efficient search for career goal paths with better comprehensive scores in a high-dimensional path combination space. This substantially improves the pertinence, feasibility, search efficiency, and path quality of career planning solutions. Attached Figure Description

[0065] Figure 1 This is a diagram illustrating the overall architecture of the LLM-based multi-agent career path planning system in this embodiment of the invention.

[0066] Figure 2 This is a flowchart illustrating the construction process of individual competency profiles, job-specific quantitative profiles, and task packages based on stage boundaries in this embodiment of the invention.

[0067] Figure 3 This is a flowchart of career goal path planning based on Monte Carlo tree search in an embodiment of the present invention. Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the application will be further described in detail below with reference to the accompanying drawings. The described embodiments are only a part of the embodiments involved in this invention. All non-innovative embodiments based on these embodiments by other researchers in the art are within the protection scope of this invention. Furthermore, the step numbers in the embodiments of this invention are only set for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0069] In one embodiment of the present invention, a career goal path generation system based on multi-agent cooperation and constrained pruning Monte Carlo tree search is provided, the overall structure of which is as follows: Figure 1 As shown, it includes: an input layer 10, a multi-agent collaborative processing layer 20, and an output layer 30.

[0070] The data input in the input layer 10 includes: user career needs input 101, user personal materials input 102, and job data source input 103.

[0071] The multi-agent collaborative processing layer 20 includes: intent and constraint parsing agent 201, constraint clarification agent 202, personal profile agent 203, job profile agent 204, task package construction agent 205, and MCTS path planning agent 206. Each agent calls the large language model through a preset application programming interface and combines prompt word templates, rule base, capability stage division rules, capability dependency graph, and machine learning model to complete the corresponding processing task.

[0072] The final output layer 30 outputs include: intermediate result output 301 and final career goal path output 302; among which, intermediate result output 301 includes: goal constraint set, personal ability profile, job quantitative profile, and candidate task package set.

[0073] Furthermore, in this embodiment, the following career goal scenario is used as an example: A user hopes to become a senior data analyst in Shanghai within 3 years, can dedicate 15 hours per week to learning, has a medium risk tolerance, and expects a monthly salary of over 20,000 yuan. After receiving the user's natural language career demand description, the system sequentially performs intent and constraint parsing, constraint clarification, personal ability profile construction, job quantitative profile construction, candidate task package construction, and improved MCTS path planning to output a career goal path with a better comprehensive score.

[0074] I. Intent and Constraint Analysis and Constraint Clarification

[0075] The system first receives the user's natural language description of occupational needs through the intent and constraint parsing agent, and then parses it to obtain structured target objects and constraint objects.

[0076] The structured target object includes a target job identifier, a target region identifier, and a target period parameter; the constraint objects include a weekly time budget parameter, a target salary range parameter, and a risk preference parameter. In this embodiment, the structured target object can be represented as: a target job identifier of "Senior Data Analyst," a target region identifier of "Shanghai," and a target period parameter of "3 years"; the constraint objects can be represented as: a weekly time budget parameter of "15 hours / week," a target salary range parameter of "≥20,000 RMB / month," and a risk preference parameter of "medium risk."

[0077] Subsequently, the constraint clarification agent performs integrity checks, conflict checks, and rationality checks on the structured target object and constraint object.

[0078] Completeness verification determines whether the target position, target region, target period, time budget, salary range, and risk appetite are all present. Conflict verification determines whether there are obvious contradictions between the target period and the time budget. Reasonableness verification determines the user's achievability within the target period based on the target position's competency requirements. After verification, a set of target constraints is obtained, which serves as input for subsequent personal competency profile construction, job quantitative profile construction, candidate task package construction, and path planning.

[0079] II. Personal Competency Profile Construction

[0080] The personal profile intelligent agent processes the personal materials submitted by users, including but not limited to resumes, educational background, project experience, internship experience, self-evaluations, and personality test results. The generation of the personal competency profile adopts a "feature extraction - stage judgment - numerical mapping - risk quantification" approach.

[0081] First, a large language model is used to identify skill entities, experience entities, and general ability entities from the personal materials, and structured ability feature descriptions are extracted. Taking this embodiment as an example, the identified skill entities may include: Python, SQL, Excel, and data visualization; experience entities may include data analysis project experience and internship experience; and general ability entities may include logical analysis, communication and collaboration, and learning abilities.

[0082] Secondly, the system determines the stage of each capability item according to preset capability stage division rules. The capability stage division rules are preset unified quantitative standards used to define multiple progressive capability stages and corresponding characteristics and numerical ranges for each capability item.

[0083] For example, the skill levels can be divided into basic, intermediate, and advanced stages. In this embodiment, the system can determine that the user's Python data analysis skills are at the advanced stage, their SQL skills are at the basic stage, and their data visualization skills are at the basic stage.

[0084] Then, the system uses a machine learning model to determine the current ability value within the corresponding numerical range. For example, when a certain ability item is determined to be in the advanced stage, the machine learning model can combine features such as project complexity, application scenario, and outcome indicators of that ability item to output a continuous numerical value within the corresponding range as the current ability value. By performing the above processing on all relevant ability items, skill vectors, experience vectors, and general ability vectors are obtained.

[0085] Simultaneously, the system quantifies risk based on the user's personality test information and risk preference parameters, calculating a risk coefficient. This risk coefficient serves as the risk constraint threshold in subsequent MCTS path planning and can be calculated using the risk coefficient formula described above.

[0086] The comprehensive risk aversion index in the risk coefficient is calculated based on the personality test information provided by the user through a preset weighted model, while the risk preference parameter mapping value is obtained by the risk preference parameter output by the constraint clarification agent through a preset mapping rule.

[0087] In this embodiment, the risk preference parameter mapping value is obtained by the risk preference parameter output by the constraint clarification agent through a preset mapping rule.

[0088] Preferably, the risk preference parameter includes three levels: low risk, medium risk, and high risk, and is converted into corresponding risk preference parameter mapping values ​​according to a preset mapping rule; the higher the risk preference level, the larger the corresponding risk preference parameter mapping value. In this embodiment, the user's "medium risk" preference is mapped to the corresponding intermediate risk preference parameter mapping value.

[0089] The comprehensive risk aversion index is calculated from user-provided personality test information using a preset weighted model. The personality test information includes scores for at least one or more dimensions of conscientiousness, emotional stability, and openness.

[0090] Preferably, after standardizing the scores of each dimension, the scores are weighted and summed according to preset weights to obtain the comprehensive risk aversion index.

[0091] For example, the calculation can be performed in the following way:

[0092] ;

[0093] Among them, w1, w2 and w3 are preset weights, and the sum of each preset weight is 1.

[0094] Then, the mapping value of the comprehensive risk aversion index and the risk preference parameter is substituted into the risk coefficient formula to obtain the risk coefficient used for subsequent path planning risk constraint judgment.

[0095] As a result, the system outputs a complete personal capability profile, which includes a skill vector, an experience vector, a general capability vector, and a risk coefficient.

[0096] III. Construction of Job Quantitative Profiles

[0097] The job profiling agent retrieves job description text from at least one job data source based on parameters such as target job identifier, target region identifier, and target salary range. The job data source may include job information from recruitment platforms, job knowledge bases, and industry competency model libraries. In this embodiment, the job profiling agent retrieves job description texts with conditions such as "Shanghai," "Senior Data Analyst," and "Monthly Salary Over 20,000 RMB," and extracts the required skills, experience, and general competency requirements for the job.

[0098] Subsequently, the system uses the same preset ability stage division rules as the individual ability profile to determine the stage of job requirements, and uses the upper limit of the corresponding stage's numerical range as the target value for the job requirements. At the same time, it counts the number of texts in the retrieved job descriptions that identify each ability item as a required or important ability for the job, and uses the proportion of this number to the total number of retrieved job descriptions as the importance weight of that ability item.

[0099] For example, if 60 out of 100 job descriptions mention SQL as a required or important skill for the position, then the importance weight of SQL is determined to be 0.6. This generates a quantitative job profile that includes the target value of job requirements and the importance weight of skill items.

[0100] IV. Construction of Phase Boundary Task Packages

[0101] The task package construction module compares individual competency profiles with job-specific quantitative profiles, calculates the competency gaps for each competency item, and identifies competency items that need improvement. Unlike existing methods that directly generate recommended content based on competency items, course items, or experience items, this embodiment does not generate a coarse-grained task for a particular competency all at once. Instead, it generates initial task packages step by step based on the stage boundaries crossed between the current competency value and the job target value, so that the competency improvement process conforms to a progressive stage pattern.

[0102] For each skill to be improved, the system identifies the stage boundary between its current skill value and the job requirement target value, combined with the preset skill stage division rules, and generates a corresponding initial task package for each stage boundary.

[0103] Taking this embodiment as an example, if a user's SQL skills are currently at the basic level, while the target position requires an advanced level, the system can identify two stage boundaries: "basic to intermediate" and "intermediate to advanced," and generate two initial task packages for each. The first initial task package is used to improve basic SQL query, data cleaning, and simple statistical analysis capabilities; the second initial task package is used to improve complex query optimization, multi-table join analysis, and data modeling capabilities.

[0104] Furthermore, such as Figure 2 As shown, forming a candidate task package set includes the following sub-steps:

[0105] S201. Generate a personal capability profile, including: current capability value, vectors of each dimension, and risk coefficient;

[0106] S202. Generate a quantitative profile of the job, including: target values ​​for job requirements and importance weights for competency items;

[0107] S203. Conduct competency gap analysis based on individual competency profiles and job quantitative profiles to identify competency items that need improvement;

[0108] S204. Identify the stage boundaries that are crossed between the current capability value and the target value required for the position;

[0109] S205. Generate the corresponding initial task package for each stage boundary step by step;

[0110] S206. Supplement the set of prerequisite dependencies for each initial task package based on the capability dependency graph;

[0111] S207. Output a set of candidate task packages, including: capabilities to be improved, stage starting point / target value, target capability value of the task package, recommended learning content, basic estimated time consumption, and set of prerequisites.

[0112] For example, if the "Advanced Data Analysis Modeling" task package depends on "Statistical Fundamentals" and "SQL Data Processing" capabilities, the system will write the aforementioned capabilities into the prerequisite dependency set of the task package. Each candidate task package shall include at least: the capability to be improved, the target value of the job requirements, the starting capability value or the starting point value of the stage, the target capability value of the stage, the recommended learning content, the basic estimated time consumption, and the prerequisite dependency set.

[0113] Compared to solutions that simply list recommended content based on capabilities, this embodiment uses a "stage boundary splitting + pre-dependency supplementation" approach to make task packages both progressive and executable. On one hand, the granularity of the task packages matches the capability growth stages; on the other hand, the dependencies between task packages are explicitly written, which is beneficial for determining the degree of dependency satisfaction and optimizing path quality in the subsequent search stage.

[0114] V. MCTS Path Planning

[0115] In this embodiment, as Figure 3 As shown, the MCTS path planning module constructs a candidate state tree based on the candidate task package set, and searches for a career goal path with a better comprehensive score under time and risk constraints by improving Monte Carlo tree search. The specific method is as follows:

[0116] S301. Initialize the search root node:

[0117] The search root node corresponds to the initial state, which is an initial value of the node state. The initial state includes at least the current capability state, the set of capabilities that have been satisfied, an empty sequence of executed task packages, the cumulative revenue with an initial value of 0, the cumulative time consumed, and the cumulative risk.

[0118] S302, Node Recursive Selection

[0119] During the node selection phase, the system starts by searching the root node and recursively selects child nodes using a modified UCT formula until it reaches a node that has not been fully expanded. The modified UCT formula is:

[0120] ;

[0121] in, Represents a node Cumulative returns Represents a node Number of visits, Represents a node The number of visits to the parent node. Indicates the exploration coefficient;

[0122] S303, Thought Decision-Making:

[0123] Upon reaching a node that is not fully expanded, the large language model calculates and determines whether the dependency coefficient of the task package to be expanded meets the standard based on the current node status, the stage progression rules, and the coverage of prerequisite dependencies. At the same time, it verifies whether the stage prerequisite conditions, time and risk constraints are met, and gives the optimal choice.

[0124] S304, Action expansion generates new child nodes:

[0125] Based on the optimal task package determined in the Thought decision-making phase, a corresponding new child node is generated under the current node; wherein, the new child node represents a new state formed by adding the optimal task package to the current node state.

[0126] S305, Path Simulation (Constrained Random Walk):

[0127] During the path simulation phase, the system starts from the new child node and, based on the phase progression rules and prerequisite dependency conditions, selects compliant task packages for constrained random walks until no scalable task packages exist, the target capability requirements are met, or a preset stopping condition is reached. During the simulation, the system calculates the prerequisite dependency coverage rate based on the current node state and the degree to which the prerequisite dependencies of the task packages are satisfied. A dependency coefficient is then obtained by combining the prerequisite dependency coverage rate with the capability mastery level using a large language model. This dependency coefficient is used to dynamically adjust the revenue value, standardized time consumption estimate, and risk value of candidate task packages, using the following adjustment method:

[0128] ;

[0129] in, Indicates the first The revenue value of each task package, Indicates the first The importance weight of each ability in the job quantitative profile Indicates the first The task package is for the first The extent to which the ability of each skill is improved; Indicates the first Standardized time estimate for each task package Indicates the basic estimated time. Indicates the total time budget. Indicates the time-consuming adjustment weight; Indicates the first Risk value of each task package This represents the stage target capability value. This indicates the current ability value. This indicates the smallest positive number whose denominator is zero. Indicates the risk-adjusted weight;

[0130] S306. Multi-dimensional scoring calculation of final return:

[0131] After the simulation phase ends, a comprehensive score is calculated based on the complete path from the root node to the simulation termination node. As the base return value, combined with the path rationality score of the LLM assessment. Confidence score of achieving the goal Weighted average yields the final return. ,in:

[0132] ;

[0133] S307, Value Backhaul Update Node Information:

[0134] During the value feedback phase, the system will provide the final reward for the entire path. The data is uploaded back along the current access path, updating the access count and cumulative reward for each node along the way. Specifically, the node access count is incremented by 1, and the cumulative reward is added to the final reward of the current path. .

[0135] S308, Iteration and Output Path:

[0136] In this embodiment, the final output career goal path includes multiple candidate task packages arranged in order. The career goal path includes at least: task package execution order, target ability items corresponding to each task package, recommended learning content, stage target ability value, basic estimated time consumption, cumulative benefits, cumulative time consumption, cumulative risk, and comprehensive path score.

[0137] As a result, users can obtain a career development path that takes into account the progressiveness of skill improvement, the rationality of task dependence, time budget, and risk appetite.

[0138] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A career path planning method based on LLM multi-agent model, characterized in that, include: The agent receives the user's natural language description of occupational needs through intent and constraint parsing, and obtains the structured target object and constraint object. The structured target object and constraint object are verified by a constraint clarification agent to determine the target constraint set; The personal profile intelligent agent processes users' personal information to generate a personal capability profile; The job profiling agent retrieves the target job description text from at least one job data source based on the target job identifier, target region identifier, and target salary range parameters, and generates a quantitative job profile. Based on the individual ability profile and job quantitative profile, an intelligent agent is constructed by using task packages to build a set of candidate task packages; The MCTS path planning agent constructs a candidate state tree based on the candidate task package set, and performs an improved Monte Carlo tree search under time and risk constraints to output a career goal path with a better comprehensive score.

2. The career path planning method according to claim 1, characterized in that, The constraint clarification agent performs integrity checks, conflict checks, and rationality checks. The integrity check is used to determine whether the structured target object and the constraint object contain necessary fields; the conflict check is used to determine whether there are contradictions between the constraints; and the rationality check is used to determine whether the values ​​of the constraints are within a preset reasonable range.

3. The career path planning method according to claim 1, characterized in that, The personal profile intelligent agent processes the following: Feature extraction: Extract structured capability feature descriptions by calling a large language model; Stage Determination: Determine the current capability stage based on preset capability stage division rules; Numerical mapping: Determining the current capability value within the corresponding numerical range using a machine learning model; Risk quantification: Based on the user's personality test information and risk preference parameters, risk is quantified, risk coefficients are calculated, and a personal ability profile is generated.

4. The career path planning method according to claim 3, characterized in that, The preset capability stage division rules are determined in advance by domain experts and are used to define multiple progressive capability stages and corresponding characteristics and numerical ranges for each capability item, so as to establish a comparable benchmark between individual capabilities and job requirements. The risk coefficient serves as the risk constraint threshold in the MCTS path planning agent, satisfying the following: ； in, Indicates the risk coefficient. This represents the comprehensive risk aversion index, calculated using a preset weighted model based on user-provided personality test information. Indicates the calibration coefficient; This represents the mapping value of the risk preference parameter.

5. The career path planning method according to claim 1, characterized in that, The job profile intelligent agent processes the following: Based on the target job identifier, target region identifier, and target salary range parameters, retrieve job description text from at least one job data source; identify and extract the skill requirements, experience requirements, and general competency requirements required for the job. The job requirement is determined to be in a certain stage based on the preset capability stage division rules, and the upper limit of the corresponding stage value range is used as the target value of the job requirement. For each competency item, the number of texts in the retrieved job description texts that identify it as a required or important competency for the job is counted, and the proportion of the number of such texts to the total number of retrieved job description texts is used as the importance weight of the competency item to generate a quantitative job profile.

6. The career path planning method according to claim 1, characterized in that, The task package constructs an intelligent agent, and the processing includes: Compare individual competency profiles with job-specific quantitative profiles, calculate the competency gaps for each competency item, and identify competency items that need improvement. For each skill to be improved, based on its current skill value and the target value required for the position, and in conjunction with the preset skill stage division rules, the stage boundary between the two is identified. An initial task package is generated for each of the aforementioned stage boundaries, so that the capability enhancement process can proceed step by step along the stages.

7. The career path planning method according to claim 6, characterized in that, After generating the initial task package, the task package building agent determines the corresponding set of prerequisite dependencies for each initial task package based on the capability dependency graph, forming a set of candidate task packages; The capability dependency graph is a pre-defined knowledge structure with capability items as nodes and pre-existing dependencies between capabilities as directed edges. Each candidate task package includes at least: capability items to be improved, job requirement target values, starting capability values ​​or stage starting point values, stage target capability values, recommended learning content, basic estimated time consumption, and a set of pre-existing dependencies.

8. The career path planning method according to claim 6, characterized in that, The MCTS path planning agent processes the following: Construct the search root node with an initial state, which includes at least the current capability state, the set of satisfied capability items, an empty sequence of executed task packages, the cumulative revenue with an initial value of 0, the cumulative time consumed, and the cumulative risk. Based on the initial state and preset constraints, the improved Monte Carlo tree search steps are executed iteratively to output a path with a better overall score; The improved Monte Carlo tree search includes the following steps in sequence: Node recursive selection: Starting from the root node, the improved UCT formula is used to recursively select child nodes until an incompletely expanded node is reached; Thought Decision: Based on the current node state, stage progression rules, and prerequisite dependency coverage, the large language model calculates and judges whether the dependency coefficient of the task package to be expanded meets the standard, and at the same time judges whether the stage prerequisite conditions and time and risk constraints are met, and gives the current optimal choice. Action Expansion: Based on the decision-making process and verification results, the task package is expanded to generate new child nodes; Path simulation: Starting from the new child node, based on the phase progression rules and prerequisite dependencies, select compliant task packages for constrained random walks, and deduce the path until it ends; Multi-dimensional scoring: Calculate the comprehensive score of the path's benefits, time consumption, and risks, and combine it with the path rationality score and goal achievement confidence score from the LLM assessment to obtain the final return. Value propagation: The final reward is propagated back along the search path, updating the number of visits and cumulative rewards for each node.

9. The career path planning method according to claim 8, characterized in that, The recursive node selection uses an improved UCT rule, and the improved UCT value satisfies: ； in, Represents a node Cumulative returns Represents a node Number of visits, Represents a node The number of visits to the parent node. Indicates the exploration coefficient; The MCTS path planning agent calculates the prerequisite dependency coverage based on the set of prerequisite dependencies of the task package and the set of capabilities already satisfied by the current node. The dependency coefficient is obtained by combining the coverage of prior dependencies and the degree of ability mastery with a large language model. ; The dependency coefficient is used to dynamically adjust the revenue value, standardized time consumption estimate, and risk value of the candidate task package, as shown in the following formula: ； ； ； in, Indicates the first The revenue value of each task package, Indicates the first The importance weight of each ability in the job quantitative profile Indicates the first The task package is for the first The extent to which the ability of each skill is improved; Indicates the first Standardized time estimate for each task package Indicates the basic estimated time. Indicates the total time budget. Indicates the time-consuming adjustment weight; Indicates the first Risk value of each task package This represents the stage target capability value. This indicates the current ability value. This indicates the smallest positive number whose denominator is zero. Indicates the risk-adjusted weight; After the simulation phase ends, a comprehensive score is calculated based on the complete path from the root node to the simulation termination node. As the base return value, combined with the path rationality score of the LLM assessment. Confidence score of achieving the goal The final return is obtained by weighting. ,in: ； ； in, and All represent preset weights; ； This updates the number of visits and cumulative rewards for each node.

10. A multi-agent career path planning system based on LLM, used to implement the method according to any one of claims 1 to 9, characterized in that, include: Input layer: Inputs user's career needs, personal information, and job data. Multi-agent collaborative processing layer: including intent and constraint parsing agent, constraint clarification agent, personal profile agent, job profile agent, task package construction agent, and MCTS path planning agent; These functions are respectively used to obtain structured target objects and constraint objects, determine the target constraint set, generate individual ability profiles, generate job quantitative profiles, construct candidate task package sets, and perform improved Monte Carlo tree search and career goal path planning under time and risk constraints. Output layer: Outputs intermediate results and a better career goal path with a comprehensive score. The intermediate results include: a set of goal constraints, an individual ability profile, a job quantitative profile, and a set of candidate task packages.

Images