Method for evaluating marketing training of AI agent based on capability evolution

Abstract

The application discloses an AI agent marketing training evaluation method based on capability evolution, and particularly relates to the technical field of artificial intelligence marketing training evaluation; a multi-stage capability evaluation model is constructed, marketing tasks are divided into phased sub-tasks, and corresponding capability evaluation factors are extracted; based on the mapping relationship between observable behavior variables and capability factors, a capability state vector sequence of an AI agent is generated; a capability evolution path model is constructed through a gated recurrent network and periodic position coding to capture the capability growth trend; further based on the nonlinear mapping relationship between the capability path and the task completion stage, an evolution evaluation curve reflecting the growth speed and inflection point characteristics is generated; and the derivative analysis and capability contribution evaluation mechanism are used to identify the capability bottleneck point, and the corresponding training strategy adjustment instruction is generated; the method can realize dynamic evaluation, bottleneck positioning and strategy optimization in the AI training process, and improve the efficiency of agent training and the quality of task completion.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence marketing training evaluation, and particularly relates to an AI agent marketing training evaluation method based on capability evolution. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, AI agents are widely used in intelligent customer service, automatic recommendation, content generation and other scenarios, and particularly show great commercial potential in the marketing field. AI agents continuously improve the accuracy and personalized response ability of their marketing techniques through dialogue learning, user feedback analysis and other methods. However, the commonly used AI marketing training evaluation methods at present mostly rely on static index evaluation or test feedback based on a single scene, and cannot fully reflect the dynamic adaptation ability of AI agents to user intentions, behavior preferences and other factors in a changing market environment.

[0003] In addition, the existing evaluation mechanism generally lacks quantitative modeling and analysis support for the capability growth process of AI agents, making it difficult to accurately track their long-term learning path and strategy evolution process, thereby affecting the model optimization efficiency and deployment effect. In particular, in complex marketing scenarios involving multiple rounds of interaction and task-type goal achievement, the existing methods are difficult to identify the strategy evolution ability and performance bottleneck of AI agents in a specific task stage, which restricts the further optimization of the agents.

[0004] Therefore, there is an urgent need for a method that can dynamically model and evaluate the training process of AI agents in marketing tasks based on a capability evolution mechanism, in order to improve the relevance of the evaluation and the effectiveness of the strategy evolution of the agents. SUMMARY

[0005] The purpose of the present application is to provide an AI agent marketing training evaluation method based on capability evolution, to solve the problems in the background art.

[0006] In order to achieve the above-mentioned purpose, the present application provides the following technical solution: an AI agent marketing training evaluation method based on capability evolution, comprising:

[0007] According to a predetermined multi-level marketing task structure, a multi-stage capability evaluation model composed of task objectives, stage-based sub-task indicators and capability evaluation factors is constructed, wherein each capability evaluation factor is associated with at least one observable behavior variable;

[0008] Interaction behavior data of the AI agent in the marketing scenario is collected, and based on the mapping relationship between the behavior variables and the capability factors, a corresponding capability state vector sequence is generated;

[0009] The capability state vector sequence is input into a capability evolution modeling module to construct a capability evolution path model thereof across training cycles;

[0010] establish a nonlinear mapping relationship between the capability evolution path and the task completion stage to generate an evolution evaluation curve for evaluating the comprehensive growth state of the AI agent;

[0011] Based on the slope change, inflection point identification and relative gradient change of the capability sub-dimension in the evolution evaluation curve, the capability bottleneck point of the AI agent in the training stage is determined, and the training strategy adjustment instruction is output combined with the corresponding task index feedback.

[0012] Preferably, the mapping relationship between the behavior variable and the capability factor generates a corresponding capability state vector sequence, including:

[0013] The original behavior data of the AI agent in the marketing interaction process is subjected to multi-channel feature extraction, and the behavior data includes reply delay, intent recognition confidence, dialogue round number, user retention rate and key phrase hit frequency, and is mapped to a unified dimensional behavior representation vector through a behavior embedding network;

[0014] The behavior representation vector is subjected to nonlinear mapping using the capability factor mapping matrix trained by task stage label supervision to obtain initial capability scores in multiple capability sub-dimensions;

[0015] The initial capability scores are weighted and processed through a time attention mechanism, and the historical behavior influence is dynamically fused, and finally the capability state vector sequence evolving over time is output.

[0016] Preferably, the capability state vector sequence is input into a capability evolution modeling module to construct a capability evolution path model, including:

[0017] An evolution modeling network based on a gated recurrent unit and fusing periodic position encoding is constructed to receive the capability state vector sequence as input, and the capability change trend of the agent in each training period is time-series modeled;

[0018] A multi-scale residual connection mechanism is introduced in the gated recurrent unit network for simultaneously retaining short-term behavior fluctuations and long-term capability accumulation effects;

[0019] The capability evolution path model is output, which is represented as a set of vector trajectories in a time series, each time corresponding to a capability state prediction result and its change gradient.

[0020] Preferably, the nonlinear mapping relationship between the capability evolution path and the task completion stage is established to generate an evolution evaluation curve for evaluating the comprehensive growth state of the AI agent, including:

[0021] The capability state vector trajectory output by the capability evolution path model is aligned with the stage performance indicators in the task completion log, a stage label code and multi-dimensional task performance features are introduced for each stage, and a capability-task joint mapping sample set is constructed;

[0022] A nonlinear mapping model based on a double-branch neural network is used, one branch processes the continuous vector sequence of the capability trajectory, and the other branch processes the task stage discrete event features, and after fusion, a stage growth score is generated and regression fitting is performed;

[0023] All stage growth score sequences are normalized to form a continuous evolution evaluation curve, which can reflect the capability growth speed in numerical value.

[0024] Preferably, the capability bottleneck point of the AI agent in the training stage includes:

[0025] First and second derivative operations are performed on the evolution evaluation curve in the continuous training period to generate a slope change sequence and an acceleration sequence; when the first derivative change in the continuous n periods is less than a set threshold δ1, and the corresponding second derivative is negative or close to zero, the section is marked as a potential growth stagnation section; in the identified potential growth stagnation section, the period increment of each capability sub-dimension in the capability state vector is extracted, and its relative speed is calculated to construct a sub-dimension relative speed matrix; if the relative speed of m sub-dimensions is less than a threshold δ2, or shows a negative growth trend, it is determined that the capability factor set is in a weak response or degradation state; and perform partial derivative influence analysis to quantify the sensitivity of each factor to the total growth score in the current period; if the capability factor is in a low speed state while having a high contribution degree, it is marked as a main bottleneck factor, and a capability bottleneck identification map is generated;

[0026] The local partial derivative absolute value of the overall growth score of the evolution evaluation curve is higher than the pth percentile of the absolute value of the partial derivative of all capability factors, which is called a high sensitivity factor, where p is in the range of 80% to 95%;

[0027] The relative speed in the continuous q training periods is lower than the median minus the standard deviation σ of the sub-dimension speed distribution, or shows a continuous negative growth trend, which is called a low speed state.

[0028] Preferably, the generation method of the training strategy adjustment instruction includes:

[0029] For the capability dimension marked as a bottleneck main factor, based on the historical performance records and capability growth path of the corresponding task sub-stage, a historical training sample set with a similarity to the current capability structure exceeding a threshold η is queried from a preset strategy template library;

[0030] The sample set is subjected to strategy effect clustering analysis, a multi-objective evaluation function is used to weight the growth rate, stability and sub-task completion rate after the strategy intervention, and the optimal strategy template is screened out;

[0031] The optimal strategy template is parameterized and converted into specific strategy adjustment instructions, including adjusting the training sample sampling mechanism, adding a capability factor directional intervention module and adjusting the reinforcement learning incentive weight, in combination with the AI agent behavior pattern characteristics of the current training cycle, and the instructions are pushed to the agent training system for execution.

[0032] In the above technical solution, the technical effects and advantages provided by the present application are:

[0033] 1. The present application breaks through the technical bottlenecks of "coarse evaluation granularity, strategy feedback lag and unexplainable growth process" in the existing AI marketing training evaluation method by introducing capability state vector modeling, capability evolution path analysis and nonlinear mapping mechanism with task completion stage. The multi-stage capability evaluation model and capability sub-dimension analysis method can realize dynamic modeling and fine-grained tracking of the growth process of the agent, and improve the accuracy, continuity and behavior explanation ability of the training evaluation.

[0034] 2. The present application can accurately identify the main bottleneck factor and generate individualized training adjustment instructions in the training stagnation stage by constructing a capability bottleneck identification map and combining historical strategy samples for strategy matching and effect clustering, and realizes automatic optimization of the training path. Compared with the traditional method, the present application is superior in terms of capability growth rate, task completion rate and output stability, and has significant technical improvement effect. BRIEF DESCRIPTION OF DRAWINGS

[0035] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings described below are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0036] Figure 1 The method mind map of the present application. DETAILED DESCRIPTION

[0037] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, the technical scheme in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0038] Embodiment 1, please refer to Figure 1 As shown in the embodiment, the AI agent marketing training evaluation method based on capability evolution includes:

[0039] According to the preset multi-level marketing task structure, a multi-stage capability evaluation model composed of task objectives, phased sub-task indicators and capability evaluation factors is constructed, wherein each capability evaluation factor is associated with at least one observable behavior variable;

[0040] Collect the interaction behavior data of the AI agent in the marketing scene, and based on the mapping relationship between the behavior variable and the capability factor, generate the corresponding capability state vector sequence;

[0041] The capability state vector sequence is input into the capability evolution modeling module to construct the capability evolution path model across the training period;

[0042] Establish a nonlinear mapping relationship between the capability evolution path and the task completion stage to generate an evolution evaluation curve for evaluating the comprehensive growth state of the AI agent;

[0043] Based on the slope change, inflection point identification and relative gradient change of the capability sub-dimension in the evolution evaluation curve, the capability bottleneck point of the AI agent in the training stage is determined, and the corresponding training adjustment suggestion or strategy reconstruction instruction is output in combination with the task indicator feedback.

[0044] In this embodiment, a method for constructing a multi-stage capability evaluation model is provided to support the training effect quantification and growth state tracking of the AI agent in complex marketing tasks.

[0045] First, a set of multi-level marketing task structure is set, which is divided into three main levels according to the actual application scene: overall marketing target, phased sub-task target, and task execution indicator. Among them, the overall target can include macro indicators such as "improve conversion rate" and "enhance user stickiness", and the sub-task target can include "identify potential intent customers", "accurately recommend goods", and "efficiently complete multi-round interaction closed loop".

[0046] Based on the above task structure, a set of capability evaluation factor set is defined, n is the total number of capability evaluation factors, each capability evaluation factor is used to quantify the execution capability of the AI agent in a specific sub-task, and the capability factor includes but is not limited to: intent recognition accuracy, user emotion response recognition accuracy, response language diversity, personalized recommendation hit rate, and user active interaction rate.

[0047] Further, in order to realize quantifiable evaluation, each capability evaluation factor c i is associated with at least one observable behavior variable A mapping relationship is established, i.e., a mapping function f exists: For example:

[0048] (intention recognition capability) <- confidence change rate, recognition delay time;

[0049] (dialogue advancing capability) <- user continuation response rate, round control rate;

[0050] (emotion perception capability) <- emotion category recognition accuracy, emotion mutation response time.

[0051] The above behavior variables v j can be collected in real time during the interaction between the AI agent and the user, and have the characteristics of high frequency, quantifiability and calculability. Through the mapping relationship, the system can convert the behavior variables into a structured capability vector, thereby supporting subsequent capability state modeling and growth path construction.

[0052] In order to adapt to different types of marketing scenarios, the model also supports a task phase dynamic configuration mechanism, which can set different phase division and weight strategies for short-term promotion tasks and long-term relationship marketing tasks, and realize flexible adaptability.

[0053] Through the method described in this embodiment, a four-level evaluation system from task target -> subtask indicator -> capability factor -> behavior variable can be established, which has a clear data-driven structure and executability, provides a fine-grained training effect feedback mechanism for the AI agent, and significantly improves the accuracy and interpretability of model evaluation.

[0054] In this embodiment, a capability state vector construction method is provided for the behavior performance of the AI agent in the marketing training scene, which is used to support subsequent capability evolution modeling and bottleneck identification.

[0055] First, the original behavior data of the AI agent in the actual marketing interaction with the user is collected, which includes but is not limited to:

[0056] reply delay (in milliseconds);

[0057] intention recognition confidence (probability value between 0 and 1);

[0058] number of dialogue rounds in a single session;

[0059] user retention rate (session completion rate / interruption rate);

[0060] key phrase hit frequency (high-weight word trigger frequency per unit time).

[0061] To capture the multi-dimensional characteristics of behaviors, the embodiment adopts a multi-channel feature extraction mechanism, sends the original behavior variables into multiple feature transformation channels such as normalization channels, sliding window statistical channels and position embedding channels, and uses a lightweight convolution module for time series local feature extraction. All channel outputs are uniformly input into a set of shared parameter behavior embedding networks after splicing, which encodes behavior features of different dimensions into behavior representation vectors of consistent length ; is a real number set.

[0062] Then, the constructed ability factor mapping matrix (where k is the number of ability sub-dimensions, and d is the behavior representation dimension) is used to perform nonlinear mapping on z, specifically using a multi-layer perception structure (MLP) and introducing ReLU activation function and Dropout regularization to obtain an initial ability score vector To dynamically model the ability evolution trend of the AI agent in the continuous training period, a time attention mechanism is further introduced to construct a historical ability memory unit with a time window. This module uses position encoding to fuse the historical score sequence and the current score to participate in attention weight calculation and weighted summation, and finally outputs a time-weighted ability state vector and generates an ability state vector sequence. As the ability expression of the AI agent in different training periods, it is used in subsequent ability evolution path modeling, bottleneck identification and strategy generation modules, and has good time series expression ability and evaluation stability.

[0063] Through the modeling method in the embodiment, the original behavior data can be effectively converted into structured ability representation, with end-to-end learning ability, which significantly improves the accuracy and dynamics of AI agent training effect analysis.

[0064] In the embodiment, a method for constructing an AI agent ability evolution path model based on an ability state vector sequence is provided, which is used to model the ability growth trend and state transition of the AI agent in the training process.

[0065] First, the ability state vector sequence generated by the previous module is input into the ability evolution modeling module, which adopts a recurrent neural structure with time modeling mechanism, used to simulate the cross-period ability evolution trajectory of the agent.

[0066] The core of the modeling structure is a set of gated recurrent units, each unit accepting the ability vector of the current period as input and outputting the hidden state To capture the dynamic change characteristics of the ability, periodic position encoding is introduced in the GRU unit to add time position information to the input vector through a sine function and a cosine function, so that the model has a periodic perception ability when expressing the ability of different training stages.

[0067] In addition, in order to consider both short-term behavior fluctuations and long-term trend accumulation in ability evolution modeling, a multi-scale residual connection mechanism is designed and introduced in this embodiment. The specific method is as follows:

[0068] A plurality of scale residual paths (for example, 1 step, 3 steps, and 5 steps) are constructed on the GRU hidden state propagation path, and a gating weighting strategy is used to fuse information of each scale;

[0069] The residual path is used to return the ability state information of the earlier period, so as to ensure that the key growth point will not be forgotten in the long-term propagation.

[0070] After the above processing, the model outputs an ability evolution path model, which is in the form of the ability prediction state of each period in the time sequence and the change gradient , that is,

[0071] Ability path model output ;

[0072] wherein, is the ability prediction vector of the current period, is the change rate of the ability state of the period, reflecting the acceleration, deceleration or stagnation trend of the AI agent in the ability growth process.

[0073] Through the modeling method described in this embodiment, high-precision prediction and structured expression of the ability growth trajectory of the AI agent in the training process can be realized, which not only retains the sensitivity of local behavior changes, but also has long-term trend modeling ability, providing strong support for subsequent growth evaluation curve generation and ability bottleneck analysis.

[0074] In this embodiment, a method for generating an AI agent growth evaluation curve based on a nonlinear mapping model is provided, which is used to combine the ability evolution trajectory and the task performance result to realize the structured measurement of the comprehensive growth state of the agent.

[0075] First, the ability evolution path model result output by the previous module is obtained, including the ability state prediction sequence under continuous training periods, and each ability vector contains a plurality of sub-dimension indicators. At the same time, the task completion log data in the training process is called to obtain a task performance indicator set T corresponding to each training period, such as user response rate, conversion rate, recommendation click rate, and average session satisfaction.

[0076] To construct the training sample for supervised learning, the ability path and the task log are aligned by period in the embodiment, and a stage label code (such as task type, execution time period, intervention strategy number, etc.) is added for each training stage to form a mapping input with task semantics. Finally, the ability: task joint sample set is constructed: ; wherein is the ability trajectory, is the task performance feature, is the stage label.

[0077] To mine the nonlinear relationship between the ability vector and the task performance, the embodiment adopts a double-branch neural network structure for processing heterogeneous input features:

[0078] The first branch receives the sequence of ability state vectors , and uses a stacked GRU module to capture the time sequence structure;

[0079] The second branch processes the task stage features , and uses a multi-layer fully connected network to extract its semantic representation;

[0080] The outputs of the two branches are spliced in the fusion layer, and after fusion activation and residual connection, they are sent to the regression layer to output the growth score of the training stage ∈[0,1].

[0081] Through the model, the growth score sequence S of all training stages can be obtained, and further normalization and smoothing processing is performed on the sequence to construct a continuous evolution evaluation curve G(t): wherein G(t) can reflect the growth speed, stability and inflection point change of the agent in the entire training period, supporting subsequent growth trend analysis and bottleneck identification tasks.

[0082] The method described in the embodiment successfully connects the causal path of "behavior ability performance-task output effect" through double-modal modeling and nonlinear fitting, avoids the limitations of traditional single scoring methods, and provides a higher-dimensional and more timely comprehensive growth measurement tool for AI training evaluation systems.

[0083] In the embodiment, to determine the growth bottleneck point of the AI agent in the training process, the system introduces a comprehensive diagnostic mechanism based on derivative analysis and sub-dimension speed statistics to identify the key stages and influencing factors that limit the growth of ability.

[0084] First, the first-order derivative and the second-order derivative of the evolution evaluation curve G(t) generated above are calculated to obtain the growth slope change sequence G'(t) and the growth acceleration sequence G''(t), respectively. If in the continuous n training periods, it satisfies:

[0085] |G'(t)| < δ1 (weak change in growth slope);

[0086] G''(t) < 0 or near zero (growth curve enters a plateau or decaying trend), then mark this time segment as a potential growth stagnation segment, indicating that the speed of capability growth is significantly reduced or tends to be saturated.

[0087] Within this segment, extract the capability state vector sequence C, calculate the relative growth rate of each capability sub-dimension in each cycle, and construct the sub-dimension relative growth rate matrix . If there are m capability sub-dimensions whose growth rates satisfy: relative growth rate < δ2, or continuous negative growth, then the set of sub-dimensions is determined as a weak response factor set, indicating that the capability growth is weak or shows a degenerative trend to training input feedback.

[0088] Next, the system performs partial derivative influence analysis on the above set of capability sub-dimensions, and for each factor c i calculates its local partial derivative of the total growth score S(t) . If a factor satisfies the following two conditions:

[0089] > the p-th percentile (high sensitivity factor), p ∈ [80%, 95%]; in the last q cycles, its growth rate is lower than the median minus the standard deviation σ of the historical growth rate distribution of this dimension, or there is continuous negative growth, then mark this factor as a main bottleneck factor, and construct a capability bottleneck identification atlas containing main bottleneck dimensions, contribution degree, trend change rate, etc. as input basis for strategy optimization.

[0090] After identifying the main bottleneck capability factor, the system starts the strategy optimization module and automatically generates training adjustment instructions based on historical strategy data and capability structure matching mechanism.

[0091] First, call the capability bottleneck identification atlas to extract the main bottleneck dimensions, corresponding task sub-stages, and capability growth trajectories in the current cycle. The system queries the pre-set strategy template library and calculates the similarity between the current capability structure (capability vector distribution, growth trend, etc.) and the capability structure of the historical training samples in the template, selecting the sample set with a similarity greater than a set threshold η.

[0092] Cluster analysis is performed on the above historical sample set, and the index performance after strategy intervention is compared in each class, including the capability growth rate improvement value, the training output stability improvement amplitude, and the sub-task completion rate change. The system uses a multi-objective weighted scoring function to evaluate the effectiveness of each strategy template, and selects the optimal strategy template.

[0093] Finally, the system combines the behavior characteristics of the current cycle AI agent (such as behavior variable activity, strategy response preference, etc.), parameterizes the optimal template, and generates executable training strategy adjustment instructions, including:

[0094] Adjusting the training sample sampling mechanism (such as biasing key sub-tasks and strengthening bottleneck scenarios);

[0095] Adding a capability factor directed intervention module (such as intent recognition special reinforcement learning);

[0096] Adjusting the reinforcement learning incentive weight (improving the learning priority of high-contribution low-growth factors).

[0097] The instructions are automatically pushed to the agent training system, implementing a strategy feedback loop to promote the precise optimization and continuous growth of agents in the capability bottleneck stage.

[0098] Example 2: To verify the actual application effect of the "AI agent marketing training evaluation method based on capability evolution" proposed in the present application in terms of capability recognition accuracy and training strategy effectiveness, this embodiment constructs an experimental environment in a simulated marketing dialogue scenario and compares it with existing mainstream training evaluation mechanisms.

[0099] The experimental subjects are two groups of AI agents using the same basic model architecture (Transformer+RL), trained using the following two training evaluation strategies:

[0100] Comparison group (Baseline): Use traditional static evaluation methods, only evaluate feedback based on task completion rate and dialogue length;

[0101] Experimental group (the present application): Use the capability state vector modeling + capability evolution modeling + nonlinear evaluation curve analysis + bottleneck identification + strategy feedback mechanism proposed in the present application.

[0102] The training period is 30 rounds, involving 300 real marketing intent data (including recommendation, guiding transactions, user retention, etc.) per round, and evaluating the capability growth changes after each round of training.

[0103] The following three core technical indicators are selected in the experiment

[0104]

[0105] The above indicators are normalized according to the aforementioned standards, with model structure and samples as control variables and evaluation feedback methods as independent variables.

[0106] Experimental results

[0107]

[0108] In addition, the experimental group enters a jump growth interval after the tenth round, and the bottleneck recognition rate reaches 93.5%, which is increased by 28% compared with the comparative group.

[0109] From the above experimental results, it can be known that the method can more finely and dynamically recognize the ability growth trajectory, and realize a more efficient training process; through the ability bottleneck recognition mechanism, the growth stagnation point is accurately found out, and a training optimization strategy is automatically generated; and the training stability, the sub-task completion effect and the overall growth speed are significantly improved.

[0110] The embodiment verifies that the method is superior to the prior art in the timeliness of training feedback, the accuracy of strategy optimization, the delicacy of ability modeling and the like, and achieves a significant technical improvement effect.

[0111] The above merely illustrates the specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. An AI agent marketing training and evaluation method based on capability evolution, characterized by: include: The task process is divided into stages according to the multi-level marketing task structure, and at least one capability evaluation factor is set in each task stage, and at least one behavioral variable associated with each capability evaluation factor is set based on preset rules; Based on the mapping relationship between the AI ​​agent's behavioral variable data and capability factors in marketing scenarios, a corresponding capability state vector sequence is generated; Specifically, the process includes: extracting multi-channel features from the original behavioral data of the AI ​​agent during the marketing interaction process. The behavioral variable data includes confidence change rate, recognition delay time, user continued response rate, turn control rate, emotion category recognition accuracy, and emotion mutation response time, and mapping them into a unified-dimensional behavior representation vector through a behavior embedding network; using the ability factor mapping matrix obtained from task-stage label supervision training, nonlinearly mapping the behavior representation vector to obtain initial ability scores in multiple ability sub-dimensions; weighting the initial ability scores through a temporal attention mechanism, dynamically integrating the impact of historical behaviors, and ultimately outputting a sequence of ability state vectors that evolve over time; Inputting the capability state vector sequence into the capability evolution modeling module to construct its capability evolution path model across training cycles; Establish a nonlinear mapping relationship between capability evolution paths and task completion stages to generate an evolutionary evaluation curve for evaluating the comprehensive growth status of AI agents; Based on the slope change, inflection point identification, and relative gradient change of the capability sub-dimensions in the evolutionary evaluation curve, the capability bottleneck of the AI ​​agent during the training phase is determined, and training strategy adjustment instructions are output in combination with the corresponding task indicator feedback; The method for generating the training strategy adjustment instruction includes: For the capability dimension marked as the bottleneck factor, based on the historical performance records and capability growth paths of the corresponding task sub-stages, the preset strategy template library is queried to extract the historical training sample set whose similarity with the current capability structure exceeds the threshold η; Conduct a cluster analysis of the strategy effectiveness of the sample set, use a multi-objective evaluation function to weightedly score the growth rate, stability, and subtask completion rate after the strategy intervention, and screen out the optimal strategy template; Combined with the behavioral pattern characteristics of the AI ​​agent in the current training cycle, the optimal strategy template is parameterized and converted into specific strategy adjustment instructions. The instructions include adjusting the training sample sampling mechanism, adding a capability factor targeted intervention module, and adjusting the reinforcement learning incentive weights, and are pushed to the agent training system for execution.

2. The AI ​​agent marketing training and evaluation method based on capability evolution according to claim 1 is characterized by: in, Inputting the capability state vector sequence into the capability evolution modeling module to construct the capability evolution path model includes: Construct an evolutionary modeling network based on gated recurrent units and integrated with periodic position encoding, which receives a sequence of capability state vectors as input and performs temporal modeling of the capability change trend of the agent during each training cycle; Introducing a multi-scale residual connection mechanism into the gated recurrent unit network to simultaneously preserve short-term behavioral fluctuations and long-term ability accumulation effects; The output capability evolution path model is represented as a set of vector trajectories in a time series, where each moment corresponds to a capability state prediction result and its change gradient.

3. The AI ​​agent marketing training and evaluation method based on capability evolution according to claim 1 is characterized by: in, The establishment of a nonlinear mapping relationship between the capability evolution path and the task completion stage to generate an evolutionary evaluation curve for evaluating the comprehensive growth status of the AI ​​agent includes: The capability state vector trajectory output by the capability evolution path model is aligned with the stage performance indicators in the task completion log. Stage label encoding and multi-dimensional task performance characteristics are introduced for each stage to construct a capability-task joint mapping sample set. A nonlinear mapping model based on a two-branch neural network is used. One branch processes the continuous vector sequence of ability trajectories, while the other processes the discrete event characteristics of the task phase. After fusion, a stage-by-stage growth score is generated and regression fitting is performed. All stage growth score sequences are normalized to form a continuous evolutionary evaluation curve, which can reflect the ability growth rate in terms of numerical value.

4. The AI ​​agent marketing training and evaluation method based on capability evolution according to claim 1 is characterized by: in, The determination of the AI ​​agent's capability bottlenecks during the training phase includes: Performing first-order and second-order derivative operations on the evolutionary evaluation curve in consecutive training cycles to generate a slope change sequence and an acceleration sequence; when the first-order derivative change in n consecutive cycles is less than a set threshold value δ1, and the corresponding second-order derivative is negative or near zero, the segment is marked as a potential growth stagnation zone; within the identified potential growth stagnation segment, extract the periodic increments of each capability sub-dimension in the capability state vector, calculate their relative growth rates, and construct a sub-dimension relative growth rate matrix; if there are m sub-dimensions whose relative growth rates are less than the threshold value δ2, or show a negative growth trend, then the capability factor set is determined to be in a weak response or degradation state; and performing partial derivative impact analysis on it to quantify the sensitivity of each factor to the total growth score in the current cycle; if the capability factor is in a low growth rate state while having a high contribution, it is marked as a main bottleneck factor, and a capability bottleneck identification map is generated; The absolute value of the local partial derivative of the overall growth score of the evolutionary evaluation curve is higher than the pth percentile of the absolute value of the partial derivatives of all ability factors, which is recorded as a high-sensitivity factor, where the p value range is 80% to 95%; If the relative growth rate in q consecutive training cycles is lower than the median of the sub-dimension growth rate distribution minus the standard deviation σ, or shows a continuous negative growth trend, it is recorded as a low growth rate state.

Images